finite element analysis for surgical decision support …
TRANSCRIPT
The Pennsylvania State University
The Graduate School
Department of Engineering Science and Mechanics
FINITE ELEMENT ANALYSIS FOR SURGICAL DECISION SUPPORT OF
INTRAMEDULLARY NAIL FIXATION FOR PROXIMAL FEMUR FRACTURES
A Dissertation in
Engineering Science and Mechanics
by
Scott M. Tucker
© 2019 Scott M. Tucker
Submitted in Partial Fulfillment
of the Requirements
for the Degree of
Doctor of Philosophy
May 2019
The dissertation of Scott M. Tucker was reviewed and approved* by the following:
Gregory S. Lewis
Assistant Professor, Department of Orthopaedics and Rehabilitation, Department
of Engineering Science and Mechanics
Dissertation Advisor
Chair of Committee
Lisa Shantz
Associate Professor, Department of Cellular and Molecular Physiology
John Elfar
Michael and Myrtle Baker Professor and Vice Chair of Research Director, Center for Orthopaedic Research and Translational Science Professor of Orthopaedics and Rehabilitation, Public Health Sciences, and Neural
and Behavioral Sciences
Judith Todd
P.B. Breneman Chair, Department of Engineering Science and Mechanics
Professor of Engineering Science and Mechanics
Head of the Department of Engineering Science and Mechanics
*Signatures are on file in the Graduate School
J. Spence Reid
Special Member
Professor and Chief
Division of Orthopaedic Trauma
iii
ABSTRACT
Intramedullary nails are the most commonly used implant for surgical fixation of proximal
femur fractures. Although these fracture fixation procedures are generally successful with fewer
than 15% of intramedullary nailing cases progressing to nonunion,1 hip fracture cases represent
nearly 75% of the annual $19 billion spent on fracture management in the United States.2 Clinical
and research efforts to reduce these costs focus on mitigating failure risk and optimizing patient
outcomes to reduce the incidence of costly hospital readmissions and secondary surgeries. Surgeons
play a role in optimizing nailing constructs through selection of appropriate implant features as
well as recommendation of postsurgical return-to-function limitations. However, quantitative data
to support implant selection criteria are generally lacking.
Chapter 1 provides a broad overview of the biomechanics and biomaterials of fracture
fixation, including clinical considerations, fracture healing biology, biomechanics, biomaterials,
and experimental approaches to fracture biomechanics. Chapter 2 narrows the reader’s focus to a
review of intramedullary nail-fixed long bone fractures with emphasis on the femur. Relevant
literature was collected, reviewed, and organized to provide a comprehensive picture of recent
discoveries regarding intramedullary nail biomechanics. A discussion of some limitations of
application of biomechanical results to clinical practice is presented in Chapter 2 to remind readers
that the laboratory environment often requires simplification of authentic scenarios for proper
variable control.
Finite element computational models of proximal femur fractures fixed by intramedullary
nailing were developed, validated, and interpreted in Chapters 3 and 4. Specifically, a wide array
of clinically relevant intramedullary nail design variable combinations was simulated for nine
common femur fracture types under peak gait loading conditions with a generic femur model.
Biomechanical influences of implant features such as nail diameter, nail length, use or disuse of
iv
distal fixation screws, and material stiffness were assessed for each fracture type. Custom-written
code was developed to automate model assembly for a commercial finite element solver. Model
validation was achieved through comparison of predicted construct stiffness to previously reported
mechanical testing data with similar design variables. Reported outcome measures of peak implant
stress and motion at the fracture site are surrogates of relative construct failure risk and offer
insights into optimizing construct selection to reduce failure rates and promote fracture healing.
The results of parametric modeling are summarized here and presented in full detail in
Chapter 3 of this work. Filling the reamed endosteal canal with the largest fitting nail diameter
reduced axial and shear interfragmentary motion for all modeled fracture types. Nail length was
less predictive of shear interfragmentary motion for most simulated fracture types than other
construct variables such as nail diameter, distal fixation screws, and nail material properties.
Furthermore, gapping at the fracture site predisposed the construct to higher implant stresses and
larger interfragmentary motions. The biomechanical outcomes from this computational study can
directly aid in surgical decision-making for optimizing hip fracture fixation with intramedullary
nailing.
The work presented in Chapter 4 utilizes a subset of gapped fracture models from Chapter
3 to provide insight into safe loadbearing levels based on shear motion at the fracture site.
Successful union following fracture fixation surgery is partially dependent on how the fixation
construct resists shear motion as the patient returns to functional activities. Orthopaedic surgeons
attempt to optimize these biomechanics via implant selection as well as the ability to recommend
specific loadbearing limits for the patient postoperatively. We show the loadbearing percentage
that maintains shear motions at the fracture site below a literature-reported threshold for each
construct. Specifically, the data show that full weightbearing achieves the shear motion threshold
for short nails with distal fixation screws for all tested nail diameters in a pertrochanteric and an
v
intertrochanteric fracture type but not for a subtrochanteric fracture. Increasing nail diameter
increases the safe loadbearing limit for nearly all tested configurations, while removing distal
fixation screws has mixed results depending on if the nail is short or long.
Some assumptions of nonspecific modeling are then challenged in Chapter 5 though
introduction of patient-specific fracture type to the otherwise generic simulations. A patient image
based fracture shape-specific model was generated and demonstrated a clinically relevant decrease
in peak implant stress of over 50%, suggesting that simulation results are indeed sensitive to
patient-specific modeling. This Chapter 5 pilot study provides motivation for future investigation
into how results of generic models can be utilized, offers suggestions for directing future modeling
efforts, and questions the functionality of simplistic fracture classification for certain applications.
Results from such efforts have direct implications on surgical resident training, offering
previously unavailable detail on the specific biomechanical influences of implant selection in
parametric fashion, as well as utility in presurgical planning. Overall, the work presented in this
dissertation provides a clinically oriented comprehensive understanding of the biomechanics of
intramedullary nailing of proximal femur fractures.
vi
TABLE OF CONTENTS
LIST OF FIGURES ................................................................................................................. viii
LIST OF TABLES ................................................................................................................... xi
ACKNOWLEDGEMENTS ..................................................................................................... xii
Chapter 1 Introduction to Fracture Fixation Biomechanics and Biomaterials ........................ 1
Overview .......................................................................................................................... 1
Clinical Aspects ............................................................................................................... 2
Introduction .............................................................................................................. 2
Types of Implants ..................................................................................................... 3 Anatomical Constraints ............................................................................................ 8
Fracture Healing Biology ................................................................................................. 9
Fracture Healing ....................................................................................................... 9 Infection ................................................................................................................... 13
Biomechanics ................................................................................................................... 13
Implant Loading ....................................................................................................... 13 Implant Stress and Failure ........................................................................................ 15 Fracture Gap Strain .................................................................................................. 18 Biomechanical Variables .......................................................................................... 20
Biomaterials ..................................................................................................................... 21
Stainless Steel vs. Titanium Alloys & Other Materials ............................................ 21 Biocompatibility ....................................................................................................... 22
Corrosion .................................................................................................................. 24
Experimental and Computational Modeling of Fracture Fixation Mechanics ................. 25
Experimental ............................................................................................................ 26 Computational .......................................................................................................... 28
Internal Plating ................................................................................................................. 30
Intramedullary Nailing ..................................................................................................... 32
Summary/Perspective ....................................................................................................... 37
Acknowledgments ............................................................................................................ 37
Chapter 2 A Review of Finite Element Modeling within Orthopaedics with Focus on
Femur Intramedullary Nails ............................................................................................. 38
Overview .......................................................................................................................... 38
Experimental and Computational Modeling of Femur Intramedullary Nails .................. 40
Applied Loads, Boundary Conditions, and Constraints ........................................... 40
Interfacial Modeling ................................................................................................. 43 What Biomechanics Research Has Taught us About Femur Intramedullary
Nailing .............................................................................................................. 43
Selecting an Implant ......................................................................................... 45 To Ream or Not to Ream.................................................................................. 48 Nail Diameter and Material .............................................................................. 49
vii
Static vs. Dynamic Distal Locking Screw Fixation .......................................... 50
Techniques Accounting for Biological Variability .......................................................... 53 Finite Element Mesh Generation ..................................................................................... 54 Modeling Applications ..................................................................................................... 55 Coordinate Systems for Description of Motion ............................................................... 57
Chapter 3 Parametric Finite Element Analysis of Intramedullary Nail Fixation of
Proximal Femur Fractures ................................................................................................ 58
Abstract ............................................................................................................................ 58
Introduction ...................................................................................................................... 59 Materials and Methods ..................................................................................................... 61
Bone and Fracture Model ......................................................................................... 61 Implant, Reaming, and Construct Configuration ..................................................... 62 Loading and Boundary Conditions .......................................................................... 63 Finite Element Simulation Parameters ..................................................................... 65 Validation, Outcome Measures, and Statistical Methods ......................................... 65
Results .............................................................................................................................. 66 Discussion ........................................................................................................................ 71
Chapter 4 Return to Function After Intramedullary Nail Proximal Femur Fracture
Fixation: Effects of Fixation Construct and Weightbearing on Shear Motion at the
Fracture Site ..................................................................................................................... 76
Abstract ............................................................................................................................ 76
Introduction ...................................................................................................................... 77 Materials and Methods ..................................................................................................... 78
Results .............................................................................................................................. 80 Discussion ........................................................................................................................ 82
Chapter 5 Examples of Precision Medicine in Orthopaedics.................................................. 84
Precision Finite Element Modeling for Intramedullary Nail Femur Fracture Fixation.... 84
Custom 3D Printed Implant Desing and Modeling for Pediatric Orthopaedic
Oncology .................................................................................................................. 88
Chapter 6 Conclusions, Discussion, and Future Work ........................................................... 93
Appendix A An Unsupervised Machine Learning Method for Discovering Patient
Clusters Based on Genetic Signatures ............................................................................. 97
Reference List .......................................................................................................................... 117
viii
LIST OF FIGURES
Figure 1-1: Applications of internal plating (B) and combined internal plating and
intramedullary nailing (D) to fix complex femur fractures (A, C). (Images from
cases in Hershey Medical Center). ................................................................................... 4
Figure 1-2: Bone screws used to stabilize a femoral neck fracture. Two of the screws
employ partial threading to achieve compression of the bone fragments. (Images
from cases in Hershey Medical Center).. ......................................................................... 6
Figure 1-3: (A) 3D CT-based models created in the Lewis laboratory of proximal
humerus fracture patients. (B) 3D-printed models were anatomically reduced to
simulate intraoperative reduction, prior to (C) computer modeling of stresses and
biomechanical stability with plate and screw fixation. (Images from the Lewis
laboratory.). ...................................................................................................................... 6
Figure 1-4: (A) Periprosthetic internal plate fixation on a synthetic bone model. (B-D)
Intramedullary nail implant with helical blade demonstrating nail curvature and
screw holes. (Images from the Lewis laboratory.). .......................................................... 7
Figure 1-5: We are currently conducting a clinical study in which the forces in vertical
struts of ring-type external fixators are measured. Force data from the struts, as well
as foot forces, are transmitted wirelessly (A) during the patient’s gait. A decrease
over time in normalized strut forces was observed in this patient (B), and in other
patients, consistent with offloading due to bone healing. In (B), ‘W37D’ indicates
dynamization of the frame. (Image from Hershey Medical Center.). .............................. 7
Figure 1-6: Careful considerations of biomechanics is needed when treating difficult
fractures. Surgeons attempt to create conditions at the fracture site that will promote
healing while allowing functional use of the extremity. Radiographs from our center
show a loss of fixation, plate, and/or screw failure in the (left) proximal tibia,
(middle) proximal humerus, and (right) distal dibia. (Images from our medical
center.).............................................................................................................................. 18
Figure 1-7: Plot to quantify the race between bone healing (callus stiffness: blue) and
implant fatigue (stresses: orange), with data based on simulations with callus. The
red line represents the fatigue strength of the material. (Data from our laboratory). ....... 28
Figure 1-8: Examples of finite element models of subtrochanteric fracture fixation with a
lateral plate and screws subject to axial loading, showing effects of screw
configuration and fracture gap size on resulting implant stresses (colors) and fracture
gap strains. (Right) Callus was represented with springs having various stiffness.
Blue represents low stress, red represents high stress. (Data from our laboratory.). ....... 32
Figure 1-9: Finite element models of intramedullary nailed long bones subject to axial
force and bending. Effects on fracture gap motions and stresses due to construct
ix
configuration are evident. Blue represents low stress, red represents high stress.
(Data from our laboratory.). ............................................................................................. 36
Figure 1-10: Stress color map of an implanted IM nail within an unfractured femur under
loading simulating the stance phase of gait. Blue represents low stress, red
represents high stress. (Data from our laboratory.). ......................................................... 36
Figure 2-1: Nail dynamization by removing distal locking screws allows for closure of
the fracture gap under axial loading. Image from Shih et al.3 .......................................... 51
Figure 2-2: Finite element models of a femoral head trabecular specimen created at a
voxel resolution of 84 μm (left) and 168 μm (right) using either voxel-based
meshing (top) or a tetrahedral meshing approach (bottom) from Ulrich et al.4. .............. 55
Figure 2-3: Cadaveric femur instrumented with a host of strain gauges for computational
model validation from Huiskes et al.5 .............................................................................. 56
Figure 3-1: Simulated proximal femur fracture types (type 31A-X.X). Types 31A-1.1,
2.1, and 3.2 were additionally modeled with gaps that remained open under the
applied load. Point pairs used to calculate interfragmentary motions are shown. ........... 62
Figure 3-2: Two design configurations under maximum applied load with the A32 (top)
and A11 (bottom) fracture types demonstrating a long, 10mm diameter nail with
distal fixation screws (top) and a short, 13mm diameter nail without distal fixation
screws (bottom). Applied loads and boundary conditions at the center of the femoral
head and implant meshes are shown. A sphere of bone is visible in the medial
femoral condyle to indicate where femoral boundary conditions are applied. All
bone models were pre-drilled for both long and short nail distal fixation screws,
regardless of configuration. .............................................................................................. 64
Figure 3-3a: Simulation outcomes for titanium implant material configurations for
reduced fractures. Maximum screw stress results are shown in the superior fixation
screw for cases that included fixation screws. ................................................................. 68
Figure 3-3b: Simulation outcomes for titanium implant material configurations for
gapped fractures. Maximum screw stress results are shown in the superior fixation
screw for cases that included fixation screws. Note vertical axis scales have a larger
range than in Figure 3-3a. ............................................................................................... 69
Figure 4-1: IM nails included in this study ranged from 10mm in diameter (left) up to
13mm and nail length was modeled as either short (26.9cm) or long (43.3cm) .............. 79
Figure 4-2: Gapped fractures modeled for analysis at intermediate load levels. Of note is
the difference in fracture gap size between the subtrohanteric 3.2 type (10mm) as
opposed to the others (5mm). Black circles represent point pairs at which
interfragmentary motions are determined. ....................................................................... 80
x
Figure 5-1: Patient fracture x-radiograph (left) superimposed over generic 3D bone model
(right). The fracture line at the superior aspect of the distal fragment is highlighted
with a spline which was used to create the fracture. ........................................................ 86
Figure 5-2: Solid model reconstructed from 3D CT data (left) shown as a gold standard
alongside patient-specific fracture from 2D x-ray alignment (right) demonstrating
good agreement in fracture shape. Reconstruction was graciously completed by
medical student Marcus Erdman. ..................................................................................... 87
Figure 5-3: Tumor volume visualization for the proximal tibia case with segmented
medical image (left) and resulting 3D reconstruction (right). Tumor is highlighted by
the red circle. .................................................................................................................... 90
Figure 5-4: Workflow for custom implant design for distal femur and proximal tibia
osteosarcomas. Patient imaging data (A, G) is reconstructed into a solid model with
virtual tumor removal (B, H) and an implant is designed to fill the bone defect (C, I).
Implant fit is then assessed (D, J). Finite element simulation stress results of implant
mechanical loading from low (blue) to high (red) under compressive loads (E, K).
Additively manufactured Ti-6Al-4V implants (F, L). ...................................................... 91
Figure 5-5: Detailed views of the proximal tibia articular-sparing implant design
demonstrating porous architecture and cortical contour. ................................................. 92
Figure 6-1: Example system load-displacement response shown (green). The piecewise
nature of the green curve demonstrates a change in system stiffness at each of the
two vertices, such as what could happen if bodies came into contact and
subsequently separated. The curve drawn in red indicates the behavior of the
simulation when contact controls is engaged. Dampening, which helps to smooth the
load-displacement relationship, introduces error in the approximated solution but
aids numerical convergence. ............................................................................................ 95
xi
LIST OF TABLES
Table 3-1: Fracture fixation construct configurations modeled. All combinations of
depicted design variables were simulated. ....................................................................... 63
Table 3-2: Results of multiple linear regression of all simulations for maximum implant
stress, axial interfragmentary motion, and shear interfragmentary motion grouped by
fracture type. Design variables with statistically significant predictive value (p<0.05)
are shaded green. Average results ± standard deviation for each outcome variable are
presented in the rightmost column. .................................................................................. 70
Table 4-1: Loadbearing percentiles (shaded) that did not induce greater than 1.5mm of
average shear interfragmentary motion. Data are separated by fracture type, nail
length, nail diameter, and presence or absence of distal screw fixation. Blank space
indicates that the shear motion cutoff is exceeded for a configuration at the
corresponding loadbearing level. ..................................................................................... 81
xii
ACKNOWLEDGEMENTS
I am grateful for my mentor Greg – I would not have been successful without his
commitment to excellence in science, unwavering support, friendship, and belief in my personal
mission. Hwa Bok Wee sustained my research efforts by teaching me many of the tools required
to conduct my thesis work, and is a valued friend.
To my wife Abby, who I feel has earned this degree just as much as I have. Thank you for
believing in me and including my journey in yours.
To my parents, for their wonderful example.
To my sister and friends for their endless support.
I am thankful for the Penn State Medical Scientist Training Program, Department of
Engineering Science and Mechanics, and Department of Orthopaedics and Rehabilitation for their
time, attention, and assistance in helping me achieve this degree. Furthermore, I thank my thesis
committee members for their guidance, time, and commitment to my success. I am proud of the
multidisciplinary path we have forged together and am encouraged that it remains open for future
students to explore.
Chapter 1
Introduction to Fracture Fixation Biomechanics and Biomaterials
Reprinted by permission from Springer Nature Customer Service Center GmbH: Springer Nature
eBook: Tucker SM, Reid, JS, Lewis, GS, Fracture Fixation Biomechanics and Biomaterials,
Springer, 15, 401-26 (2018).
Overview
Surgical fracture fixation is an important part of modern orthopaedic care. Implants
are designed by engineers, and selected and applied by surgeons, with careful consideration
of clinical, biological, biomechanical, and biomaterials principles. Clinically, a large
variety of screws, plates, intramedullary nails, and external fixation devices are used.
Fracture healing is a biologically complex process that may proceed down one of multiple
possible pathways. Successful fracture healing, as well as implant survival, is dependent
on three-dimensional biomechanics as the patient resumes activity. These biomechanics
are dependent on patient variables as well as the fracture fixation construct chosen by the
surgeon. Implant biomaterials must satisfy stringent biomechanical and biocompatibility
requirements. Experimental and computational models enable advances in implant design,
as well as our understanding of how surgeons may best apply these implants for each
patient.
2
Clinical Aspects
Introduction
Fracture management constitutes a very large portion of modern orthopaedic care. Many
fractures can be managed by non–surgical methods such as splinting, casting, and functional
bracing. This type of fracture management is particularly applicable in nondisplaced or minimally
displaced fractures – often in the pediatric population whose healing times are quite short and the
remodeling potential after healing can correct residual deformity. Nonetheless, operative fracture
care is being used increasingly as it has been shown to reduce disability, improve outcome, and
improve the quality of life following a significant fracture. There are three areas in which surgical
facture care has been shown to make a major impact on patient recovery from injury, and is pivotal
in the restoration of function and quality of life.
Displaced fractures of the femur and tibia in the adult are now usually treated surgically as it
has been well-demonstrated that this method is safe, promotes the rapid return to function including
ambulation, preserves joint function, and avoids the condition associated with non-operative
treatment previously coined “fracture disease” which is a complex of problems such as joint
stiffness, pressure ulcers, disuse osteopenia, formation of deep vein thrombosis, and muscle
atrophy.
The second area in which operative fracture care has resulted in dramatic improvement in
clinical outcomes is the fracture of the articular surfaces of joints. The surfaces of most joints (e.g.
hip, knee, ankle, and elbow) are highly congruent with each other such that distortion of the shape
of either side of a joint from a fracture can dramatically alter the load transmission across the
surface and raise cartilage contact pressures to the point that the cartilage will breakdown and lead
to post-traumatic degenerative arthritis. Often, this will result in joint destruction within several
3
years of the injury. Modern orthopaedic care of the injured joint rests on the concept of restoration
of the shape of bones on both sides of the joint as well as restoration of axial alignment and
ligamentous stability of the joint. This approach restores the normal pressures across the joint
during functional use of the limb. A joint that has been restored surgically in this way can undergo
very early motion which prevents stiffness and optimally promotes recovery of the injured articular
cartilage.
The final arena in which surgical treatment of fractures has made a dramatic improvement in
outcome is the multiple injury trauma patient. In this group, rapid stabilization of long bone and
pelvis fractures is viewed as part of the overall resuscitation of the patient since it helps control
bleeding, pain, and facilitates overall management of the multisystem injuries. Often, this is
accomplished initially using external fixation in a “damage control” mode. Then, after the patient
is stable and through the initial period of bleeding, these external fixation frames are replaced with
more definitive internal fixation to allow progressive mobilization and ambulation.
Types of implants
The common element in the entire spectrum of surgical fracture care is the use of orthopaedic
implants to restore skeletal stability that has been lost by the fracture or joint dislocation. It is the
thoughtful application of these devices that allows the modern fracture surgeon to create conditions
at the fracture site that will promote healing as well as allow functional use of the extremity during
the period of healing (Figure 1-1).
4
Four basic types of implants are used in the surgical treatment of fractures: screws, plates,
nails, and external fixators. The most basic type of orthopaedic implant is the bone screw (Figure
1-2). Modern orthopaedic bone screws are available in a vast variety of sizes (1-8mm diameter)
and thread designs which allow use in a wide variety of clinical situations. Screws can be used as
stand-alone fracture fixation devices which are placed across fracture fragments to generate
compression and create stability and promote fracture union. These “compression screws” employ
two different design techniques to achieve compression. The partially threaded bone screw gains
purchase in the bone fragment away from where the screw is placed with no screw threads in the
near fragment. Compression is generated when the head of the screw contacts the near bone
fragment and the threads engage the far bone fragment in essence “pulling” them together. A fully
threaded bone screw can also generate compression by the method of application (compression by
technique). A hole is drilled in the near bone fragment that is slightly larger than the outer diameter
of the screw (gliding hole) and the far bone fragment is drilled with a size equal to the core diameter
of the screw (threaded hole). When the screw is placed, the threads gain purchase only in the far
fragment and as the screw head contacts the bone surface, compression is generated. The same
fully threaded screw can be placed across bone fragments without the use of a gliding hole. In this
Figure 1-1: Applications of internal plating (B) and combined internal plating and intramedullary
nailing (D) to fix complex femur fractures (A, C). (Images from cases in Hershey Medical Center.)
5
situation, the screw engages threads on both sides of the fracture. Compression cannot be generated,
and this construct is termed a “static” or “position” screw.
Plates are placed on the outer aspect of the bone surface and are mechanically linked to the bone
via the use of screws. These bone screws can be mechanically joined to the plate either by a
frictional method (non-locked), or designed in such a way to allow the screw head to engage the
holes in the plate via a series of threads in the head of the screw (locked) such that movement of
the screw-plate interface is highly constrained. A plate designed to stabilize a fracture must fit
within the local anatomy such that muscles and tendons and the adjacent joints can function
normally while the plate is in place (Figures 1-3, 1-4A).
The third major type of fixation device is the intramedullary (IM) nail (Figure 1-4B-D). These
devices are placed into the medullary cavities of diaphyseal bones (e.g. femur, tibia, and humerus),
and act as internal splints across the fractures. IM nails can span the length of the bone and be large
devices, up to 50 cm. in length. Modern IM nails allow screws to be placed perpendicular to the
nail through the ends of the bone and the ends of the nail on either side of the fracture to control
motion and displacement. This type of device is termed an “interlocked” nail and is the workhorse
device in the treatment of adult femur and tibia fracture.
The fourth category of fixation device is the external fixator (Figure 1-5Error! Reference source
not found.). In this type of device, threaded bone pins (3-6 mm in diameter) are percutaneously
placed into the bone on either side of the fracture and then connected outside the body using clamps
and bars. External fixation can be constructed across an entire spectrum of stability (under the
surgeon’s control), to solve many clinical problems in orthopaedic fracture care up to the creation
of stability to allow definitive fracture healing. By its very design, external fixation is temporary
and will have to be removed. Screws, plates and intramedullary nails do not have to be removed
per se, but may be at a secondary procedure for a large variety of clinical reasons.
6
Figure 1-2: Bone screws used to stabilize a femoral neck fracture. Two of the screws employ partial
threading to achieve compression of the bone fragments. (Images from cases in Hershey Medical
Center.).
Figure 1-3: (A) 3D CT-based models created in the Lewis laboratory of proximal humerus fracture
patients. (B) 3D-printed models were anatomically reduced to simulate intraoperative reduction,
prior to (C) computer modeling of stresses and biomechanical stability with plate and screw
fixation. (Images from the Lewis laboratory.)
B C A
7
Figure 1-4: (A) Periprosthetic internal plate fixation on a synthetic bone model. (B-D)
Intramedullary nail implant with helical blade demonstrating nail curvature and screw holes.
(Images from the Lewis laboratory.)
Figure 1-5: We are currently conducting a clinical study in which the forces in vertical struts of
ring-type external fixators are measured. Force data from the struts, as well as foot forces, are
transmitted wirelessly (A) during the patient’s gait. A decrease over time in normalized strut forces
A B
Ratio of Total Strut
Forces to Injured
Limb Foot Force
Week (W) Clinical Visit
8
Variations in fixation implant design and surgical application may enhance overall mechanics
and promote healing. Features such as improved stability, reduced anatomical interference, and
more robust fixation to bone may be achieved through implant design modification, and careful
planning by the surgeon.
Anatomical Constraints
The biomechanical and anatomic design constraints on orthopaedic fracture implants are quite
rigorous. They must be biologically compatible, and meet the biomechanical requirements of the
situation in which they are used. Since they are used to support the bone and not replace it during
the period of healing, the implants must fit the local anatomy which varies among patients. IM
nails must fit within the intramedullary canal. Plates must fit on the external surface of bone and
not impede important muscles, tendons, and ligaments surrounding them. Implants are also
designed with consideration of the surgical exposure needed to apply them. Although many plates
and screws are applied with an open approach that exposes the bone fragments, IM nails, external
fixators and certain plates can be applied with a minimally invasive approach that avoids disturbing
the soft tissues surrounding the fracture that are important for blood supply and the healing process.
Plates designed to reduce the area of contact with bone via geometric modification of the contact
surface are termed minimum contact plates. Reducing contact area at the bone-plate interface may
enhance fracture healing by mitigating disruption to the periosteum and bone blood supply.6
As a result of the aforementioned stringent geometric design requirements, fracture fixation
implants typically do not have an infinite fatigue life. Since the fixation device may initially (at
was observed in this patient (B), and in other patients, consistent with offloading due to bone
healing. In (B), ‘W37D’ indicates dynamization of the frame. (Image from Hershey Medical
Center.)
9
the time of the acute fracture and before bone healing) be load bearing to allow the patient to begin
functional use of the extremity or ambulation, there will be accumulated damage in the
microstructure of the implant. It is well understood by both orthopedic clinicians and implant
designers that there is a “race” between bone healing and implant fatigue fracture. The clinical
presentation of a fatigue fracture of an orthopedic implant may imply some pathology of bone
healing.
Fracture Healing Biology
Fracture Healing
Fracture healing is a complex biomechanical process with multiple possible endpoints. The
most desirable endpoint, termed ‘union’ in the medical vernacular, is complete fusion of the
fractured components into one continuous osseous structure. A union may be achieved through one
of two well-described mechanisms.7 Primary or intramembranous bone healing occurs through
Haversian remodeling which requires absolute stability between bony fragments. On a cellular
level, primary healing involves direct healing of bone without the presence of intermediary tissues.
A very stable interface between two bone fragments is necessary because there is creation of
bridging capillaries as part of the Haversian remodeling which is intolerant of shear forces due to
its mostly linear morphology.8 Primary healing also requires contact between the fracture
fragments, or only a very small gap (on the order of 1 mm). The degree of stability required for
primary bone healing can usually only be achieved via compression between fragments. Thus,
primary bone healing usually only occurs for simpler fractures in which the fragments can readily
be fixed back in their anatomical position, usually with plates and/or lag screws.
10
Secondary healing or enchondral ossification is characterized by formation and evolution of
progressively stiffening intermediary tissues in a relative stability mechanical environment.
Secondary fracture healing progresses through three biological stages. The inflammation phase
begins the moment fracture occurs and can last up to five days. This first phase is characterized by
an immune response in which damaged, necrotic tissues are scavenged by immune cells and the
fibrinous hematoma formed during injury is organized into granulation tissue. Fibroblasts produce
type III collagen. Inflammatory mediators are released resulting in pain, swelling, and an increase
in local blood flow. A chondroid stage then begins enhancing the vascularity and cellularity of the
granulation tissue via mesenchymal stem cells from the nearby periosteum. The chondroid matrix
is then replaced by osteoblasts creating type I collagen. The replacement and removal of the
cartilage intermediate is the hallmark of endochondral ossification. At this point in the healing
process, mechanical loads can be transmitted from one fragment to the other, but there is no internal
organization of the callus. As normal loading returns, the woven bone characteristic of early callus
is remodeled into lamellar bone in response to the loading pattern. This remodeling process can
take months to years and is accompanied by a decrease in the size of the callus mass often closely
re-approximating the diameter of the normal surrounding bone. Coincident with the decrease in
size during remodeling, is an optimization of the mechanical properties of the bone.
The biological stages of secondary bone healing are intimately linked to the local mechanical
environment. The interfragmentary strain theory, first described by Stephen Perren in 1991,9 states
that a tissue type cannot exist in a region in which the mechanical forces exceed its strain tolerance.
The elegant interplay of biology and mechanics described in the strain theory provides the fracture
surgeon and implant designer a framework to understand these seemingly unrelated concepts. For
example, recent research on bone healing has found immature myeloid cell-mediated angiogenic
cascade to enhance bone healing in a mouse model.10
11
By definition, a fracture involves damage to the local blood supply of the bone often extending
to the local tissue. As a result, bleeding occurs around the bone ends creating a fracture hematoma.
As might be expected, conditions within this clotted hematoma are acidotic, hypercarbic, and have
a low oxygen tension. The first tissue to appear histologically in this region is type III collagen
made by fibroblasts. The syncytia of loosely organized collagen fibers and fibroblasts within the
clotted fracture hematoma is called granulation tissue. Mechanically, granulation tissue has a strain
tolerance of about 100%. Metabolically, granulation tissue survives in this region because it has a
low oxygen requirement which can be met via diffusion. The granulation tissue organizes into an
early fracture callus and conditions for the second stage of fracture healing are created. The second
tissue type to appear is cartilage, which is composed of type II collagen, is made by chondroblasts.
The presence of chondrocytes implies that the region has become mechanically stiffer as the strain
tolerance of cartilage is about 10%. At this time, the local blood supply is improving in response
to the release of angiogenic pyrogens. This larger more organized callus composed of granulation
tissue and cartilage further stiffens the region to the point where enchondral ossification can occur.
In regions of exceptionally low strain, osteoblasts will differentiate from mesenchymal stem cells
and begin production of an osteoid. The histological appearance of osteoblasts implies significant
mechanical stability has been achieved, as this tissue type has a strain tolerance of only about 2%.
An important concept is that osteoblasts are quite metabolically active and have a high oxygen
requirement. This requires the co-development of a rich capillary network. Bone cannot form in
regions of high strain because the supplying capillaries cannot survive. The strain theory informs
that each tissue type that appears in a healing fracture prepares the region both mechanically and
biologically for the next tissue. Bone formation implies that stability to maintain a capillary network
has been achieved. A histological hallmark of fracture healing is capillaries crossing the fracture
gap. In the case of primary bone healing, a surgical procedure has created stability between bone
fragments such that a capillary can cross the gap without the appearance of precursor tissues.
12
The different stability requirements between primary and secondary fracture healing are
functions of the strain tolerance of the local biological tissues. Although the strain tolerances for
relative stability are less defined, secondary healing has been shown to be induced by
interfragmentary motion on the order of millimeters.11
Impaired fracture healing can have biological and mechanical origins. Focusing on mechanical
causes, nonunions or delayed unions are generally a result of insufficient stability at the healing
site. Excessive strain during healing can rupture nascent capillaries and prevent nutrient transport
to the metabolically active healing site. Poor stability can have multiple origins including
inadequate initial surgery, implant hardware failure, or an unforeseen traumatic mechanical event.
Biologically, a patient may have decreased bone density that will affect the stability that can be
achieved at surgery. Tissue damage, infection12, medications, smoking13, nutrition, and genetic
factors14 influence the healing response following a fracture. Surgically, a risk of nonunion is
created if the local tissues are damaged by the procedure especially if secondary bone healing is
relied upon. In this situation, callus formation is impaired as a result of the tissue damage, and the
healing may not be able to proceed to union. In nonunion cases, stresses on the implant are never
relieved by bone unions, eventually resulting in a fatigue failure of the implant.
Residual bone deformity can result from a failed fracture fixation or improper bone healing.
Residual deformity can occur when the fracture is not reduced to its original anatomical position
or orientation, bone migrates through partial implant loosening, or implants partially fail through
yield. Most common in comminuted and bone loss fractures, residual deformity can leave a fracture
patient with pain, joint stiffness, limb-length discrepancies, and posttraumatic arthritis (in addition
to the deformity) associated with insufficient reduction of fractures affecting a joint. Residual
deformity outcomes may require medical or even surgical attention and reduce a patient’s quality
of life. The close relationship of local mechanics to successful bone healing necessitates that proper
care and consideration be given to all fractures.
13
Infection
Infection is a devastating complication following surgery, and implanted materials lead to
greater risk of infection. Open fractures where the skin is broken increase the risk for infection.
This risk can reach 30% in certain high grade open fractures with severe contamination and damage
to muscle and bone. Surgery further increases the risk for infection, although most infection rates
during surgery on non-immunocompromised patients are below 1%. Diabetes mellitus, HIV, and
rheumatoid arthritis are common examples of chronic conditions which can also increase the risk
for infection during fracture fixation. Additionally, lifestyle characteristics such as smoking,
obesity, and poor nutrition can also predispose patients to higher infection risks. Novel approaches
for reducing bacterial colonization and biofilm formation on fracture fixation biomaterials are an
active area of research.
Biomechanics
The biomechanics of the fracture fixation construct are integral to understanding how to
develop an optimal implant and surgical approach to fracture repair. The term ‘implant’ herein will
refer to the large fixation component that stabilizes the fracture often via attachment with screws.
The term ‘construct’ refers to the entire system encompassing the implant, screws, fracture site,
bone, and healing tissues. ‘Configuration’ will refer to surgical variables, i.e. number of screws
used, implant position, type of screw fixation, etc.
Implant Loading
The initial stress born by a fracture implant following surgery is quite variable across patients.
At one end is the scenario in which bone fragments have been anatomically reduced and
14
compressed (absolute stability). Load transmission will occur from one bone fragment to the other
directly and result in low implant stresses. At the other end of the stability spectrum is the situation
in which there is no initial contact between bone fragments and the implant carries the entire initial
load of the construct. A bridge plating across a large zone of comminution would be an example of
this.
Fracture fixation constructs are subjected to a variety of loading conditions ranging from
singular high force loading, such as due to an accidental patient fall, to repeated functional
activities. These load patterns are classified as either static or cyclic, respectively, and may lead to
either material yield or fatigue failure mechanisms. Orthopaedic implants can be subject to
surprisingly large mechanical loads within the body. For example, instrumented arthroplasty
implant studies show that implant reaction forces in the knee, hip, and spine can reach nearly three
times the bodyweight during simple gait.15 Furthermore, instrumented fracture fixation implant
studies can show the strain in an implant and use strain data as a surrogate for fracture healing over
time.16 A study using an instrumented femoral nail demonstrated that the load in that implant was
almost exclusively parallel to the implant’s primary axis in four different postures.17
Before a fractured bone has begun healing, the fixation construct may be responsible for up to
all of the load transferred across the fracture gap. This is particularly true in situations in which
there is no contact between the major bone fragments of the fracture. As time progresses,
intermediate bone tissues form across the fracture site which gradually reduce the load carried by
the implant. The construct must maintain mechanical integrity during the course of bone healing as
the mechanical environment at the fracture site plays a large role in achieving union. As discussed
previously, excessive motion at the fracture gap can result in local tissue strain that exceeds the
tolerance of nascent capillary networks that are attempting to bridge these gaps.3 Thus, we must
understand the mechanical responses of fixation constructs to physiological loads within the scope
of the stability tolerance required for a biological fracture healing process.
15
Bone, as a living material, adapts its composition and shape over time in response to its
mechanical stimulus history. Stress shielding refers to decreased bone density in regions that are
subjected to lower stress levels due to loads borne by the implant. The fixation implant effectively
offloads some of the forces that normally pass through the bone. Such a decrease in stress is
believed to engage biomechanical signals and either stimulate bone resorption or inhibit bone
formation.
Although limited or non-weightbearing is often appropriate in the early stages of fracture
healing, sometimes to protect the implant itself, prolonged limited weightbearing in a patient can
result in disuse osteopenia. Disuse osteopenia is a decrease in bone mineral density as a result of
prolonged periods of unloading in the bone. Bone remodeling processes, perhaps with sensing by
osteocyte cells, detect the relatively low loading profile of bone and favor resorption over
deposition, resulting in a bone density that decreases with time. Osteopenia decreases bone mineral
density and increases the risk of implant-bone loosening or bone fracture during activity.18,19
Implant Stress and Failure
Stress is a measure of internal forces in a localized region of material and has units of force
divided by area. Most often in orthopedic biomaterials stresses are generated in an implant in
response to applied external loads. Stress causes a material to deform, creating strain which is a
measure of internal displacement in a localized region of material. Strain is a unitless ratio of final
length to initial length. Basic yield and fatigue failure theories for engineering materials such as
orthopaedic implants primarily utilize stress to predict failure. Most current fracture fixation
implants are usually considered as linear elastic materials, meaning that while they undergo
deformation below failure thresholds, a linear relationship between stress and strain is maintained.
Young’s Modulus or elastic modulus is equivalent to the ratio of stress to strain. Examples of linear
16
elastic materials include steel, carbon fiber, and glass. Once the yield stress is reached, the material
begins to fail plastically and the Young’s Modulus no longer accurately describes the stress-strain
relationship.
Stress in a fracture fixation construct, like in other mechanical applications, is subject to
concentration effects. Stress concentration describes the localized magnification of stress due to
geometric features such as holes and sharp corners. Changes in cross sectional area are common
along fracture fixation implants and are frequently caused by variable implant design, holes in the
implant intended for screw placement, tapers, fillets, and edge characteristics. Most fracture
fixation plates have unique underside geometries that limit areas of contact with the periosteum in
order to reduce damage to blood supply. Variable cross sections are introduced in some cases to
avoid interference with anatomical structures and reduce the weight of the implant. In most modern
plates, the removal of material under the plate to limit bone contact area is done in such a way as
to keep the cross-sectional moment area of inertia of the implant the same throughout the plate
length. This allows the surgeon to easily bend the plate if needed during the surgical procedure.
Fracture fixation construct failure can occur within the implant. If the failure occurs under a
static load, yield can occur at the location of maximum shear stress in ductile materials. The yield
strengths for common fracture fixation implant materials such as cold-worked stainless steel, hot-
forged CoCr alloy, and forged, heat treated titanium are 792, 1600, and 1034 MPa, respectively.
Bone screws are typically made of annealed stainless steel which has a yield strength of 331 MPa.
Typically, shear stress is maximized in the implant region near an unsupported fracture gap where
there are fewer points of fixation and where the largest bending moments are likely to occur. When
the applied load is not sufficient to cause yield failure, the implant may still accumulate damage in
the form of microcracks. Cyclic loading at sub-yield stress levels causes fatigue failure via
progressive crack propagation and eventual brittle fracture within the implant. At stresses above a
material’s endurance limit (a cyclic stress amplitude below which failure will never occur), one can
17
calculate the number of fully reversed loading cycles required to reach fatigue failure. The resulting
data are plotted on an S-N curve to characterize the material’s response to cyclic loading. A
Goodman Diagram may be used to consider the fact that implant mechanical loading is often not
fully reversed, and is instead repeated with a non-zero mean stress.
However, the amplitude of cyclic loading of fracture fixation implants is often variable,
dependent on a variety of functional activities (e.g. walking, climbing stairs, jumping, etc.). Under
Miner’s rule, the total damage is defined as the sum of incremental damages calculated for each
load amplitude.20 Each incremental damage can be calculated as the ratio of the number of cycles
at a given stress or strain range to the total number of cycles required for failure at that range. Some
commercial finite element software include more advanced algorithms for predicting fatigue, e.g.
‘fe-safe’, a durability finite element software suite from Dassault Systemes. Fixation and contact
points between the implant, screws, and bone are of special interest for biomechanical
characterization as these locations are common sites of construct failure. Screw head shearing has
been documented to occur. In this case, screw fractures typically occur near the screw-implant-
bone junction (Figure 1-6). Shear stresses in this region can be influenced by decreasing the
distance between the fracture gap and the closest screws (working length), increasing the screw
count, and biological stiffening across the fracture site via healing.21 Local bone micro-fracture at
the screw-bone interface can also occur, typically leading to screw pullout from the bone.22
18
Fracture Gap Strain
Bone is remarkable in its ability to regenerate following a fracture. Unlike most other
tissues which produce scar tissue, a fractured bone, adequately stabilized and with sufficient
vascularization and other biological factors, is able to ultimately regain its near-original form.23
Nonoperative treatment relies on some degree of stability from soft tissues including muscles,
ligaments, fat, and skin, whereas operative treatment increases stability through implant fixation.
Various theories have been proposed for how mechanical stimulus affects differentiation
and adaptation of healing tissues. These theories are based on mechanical stimulus in the form of
hydrostatic pressure, deviatoric stress, and even fluid flow velocity, the latter viewing the tissue as
a poroelastic material.
Interfragmentary displacement or strain (displacement divided by fracture gap width)
measures provide a useful and intuitive means of characterizing mechanics at the fracture gap.
Based on available evidence, moderate strains of approximately 10-50% have been proposed as
Figure 1-6: Careful considerations of biomechanics is needed when treating difficult fractures.
Surgeons attempt to create conditions at the fracture site that will promote healing while allowing
functional use of the extremity. Radiographs from our center show a loss of fixation, plate,
and/or screw failure in the (left) proximal tibia, (middle) proximal humerus, and (right) distal
dibia. (Images from our medical center.)
19
being beneficial to healing.24,25 This strain is often split into two components: axial or longitudinal,
acting along the long bone axis, and shear, acting perpendicular to the long bone axis. There is
some consensus that shear strains can be detrimental to fracture healing, whereas moderate axial
strains are beneficial.26,27 Dynamized intramedullary nails and certain new plate technologies
promote additional axial strain.28 Excessively high strains are inhibitory to fracture healing and may
result in nonunions.29–31 Compressive interfragmentary displacements are superior to distractive
displacements in creating callus.32 Bone formation is linked with vascularization, and
interfragmentary movements early in fracture healing to promote revascularization, but in the later
stages of healing can inhibit blood flow.29,33 Some investigators have demonstrated clinically the
influence of biomechanical and biomaterial choices on fracture healing and callus formation.34
However much of the evidence supporting these theories are from studies in animals, especially
sheep. In these studies, fractures, often in the tibia, were simulated with osteotomy cuts leaving a
gap. The fractures were stabilized with external fixators. In one such study,26 five sheep were fixed
with a device that allowed only axial interfragmentary displacements, and five sheep were fixed
with a device that allowed only shear displacements. Displacement magnitudes were 1.5 mm with
a 3 mm fracture gap (50% strain) in both groups. The group with shear displacements experienced
significantly delayed healing of the osteotomies, with one third stiffness of the healing site after
eight weeks.
Although there is some consensus that axial and shear strain influence fracture healing
differently, unfortunately control of these two strain components by changes in fracture fixation
construct are not always intuitive. For example, increasing the working length (or bridge span)
between the inner-most screws in a plate construct seems that it would mostly affect axial
interfragmentary strain; however recent predictions from finite element models of 66 supracondylar
femoral fracture fixations in patients (with supporting data from synthetic bone experiments)
revealed that increasing working length primarily affected interfragmentary shear motions, not
20
axial motions.27 Changing the plate from stainless steel to titanium increased both types of motion.
Callus formation in the patients was not associated with comorbidities including smoking or
diabetes, but was promoted by longitudinal motions and inhibited by shear motions at the fracture
site.27 Additionally, improvements in one aspect of biomechanics may lead to concerns in another
aspect; the scenario of large longitudinal strains combined with small shear strains may be achieved
with smaller plate bridge spans, but unfortunately these smaller spans lead to larger plate stresses
and a risk of plate failure.27,35
Biomechanical Variables
Bone strain and displacement at the fracture gap are influenced by the stiffness of the construct,
interfragmentary contact, and load transfer at the fracture gap. The stiffness of a fixation construct
is its resistance to deformation in response to applied loads. Stiffness is a function of several
construct variables: design, orientation, material composition, working length, screw count, and
biological material properties in the fracture gap. Construct stiffness can be measured in multiple
modes: bending, torsion, and compression. The bending stiffness for plate constructs in long bones
often dictates most of the displacement at the fracture gap, whereas circular-type external constructs
rely more on the compression stiffness mode. Torsional stiffness can become important clinically
in all types of fracture repair constructs (plate, intramedullary nails, and external fixation).
Various fracture fixation devices provide stiffness to the fracture. For example, a bridge plate is
a construct in which a plate implant acts as an extramedullary splint spanning a complex fracture
and bends and compresses in response to axial loads on the bone. The bending stiffness of a bridge
plate can be calculated as an applied bending moment divided by lateral deflection. Similarly, the
compressive stiffness of the bridge plate can be determined by dividing the applied axial load by
axial displacement. All loading modes (compression, bending, and torsion) work in unison to
21
dictate the bony displacement and strain at the fracture site, critical for fracture healing, under a
given load scenario.
Graded-stiffness and composite material compositions can modify construct stiffness although
currently they do not see much clinical use. The working length of a fixation construct is the
distance between the two fixation points closest to the fracture gap. The stiffness of the construct
can be approximately inversely proportional to the working length.36 Surgical decisions such as
screw placement can cause construct stiffness to be too extreme in either direction (too stiff or too
compliant) to promote healing.
Use of either locking or nonlocking screws will influence the fracture fixation biomechanics.
Locking screws maintain a fixed angle with the plate and do not enable motion at the screw-plate
interface. Nonlocking screws compress the plate and the bone together creating a frictional
interface between them. Over time, especially in unicortical screw fixation, normal stresses in the
cancellous bone surrounding screws may cause loosening at the bone-screw interface, allowing the
screws to toggle at the implant-screw junction. Thus, the biomechanics of nonlocking screws allow
for some motion at the screw-plate interface, potentially resulting in a decreasing construct stiffness
over time.37
Biomaterials
Stainless Steel vs. Titanium Alloys and Other Materials
Although a vast array of metallic alloys are available for titanium and stainless steel, the
mechanical, chemical, and biocompatibility properties dictate that a small subset of these alloys are
appropriate for use in fracture fixation applications. Metallurgical alloying can influence multiple
material properties of the base metal. Specifically relevant to fracture fixation are the toughness,
22
manufacturability, oxidative potential, corrosion resistance, and biocompatibility. Most modern
implants are manufactured through forging.38
The current most common stainless steel alloy used in orthopaedic fracture fixation
applications is designated as 316L. This low carbon content alloy is manufactured to follow ASTM
F138 and F139 standards which specify: ≤ 0.03% carbon, ≤ 2% manganese, ≤ 0.03% phosphorous,
≤ 0.75% silicon, 17-20% chromium, 12-14% nickel, 2-4% molybdenum, with the remainder being
iron. The carbon and chromium content contribute to corrosion resistance while nickel and
molybdenum improve the mechanical toughness of the alloy. 316L is amenable to cold-working
strengthening which can potentially increase the yield strength from 330 MPa to a maximum of
1200 MPa.38 The elastic modulus for stainless steel is 190 GPa.
Titanium and titanium alloys have an elastic modulus of 110 GPa, considerably lower than that
of stainless steel and closer to that of bone, which explains why titanium alloy implants mitigate
the risk of bone loss associated with stress shielding. Titanium is commonly alloyed with aluminum
and vanadium, the most common being Ti-6AL-4V (5.5-6.5% aluminum and 3.5-4.5% vanadium).
Biocompatibility
All components of internal fracture fixation implants and percutaneous components of external
implants must be biocompatible. Because fracture healing typically occurs on the order of weeks
to months, biocompatibility must be maintained long after the materials have been implanted.
Although a complex topic, the biocompatibility of an implant refers to the local and systemic
physiological changes undertaken by host tissues in response to the implant’s presence. Potential
biocompatibility hazards of orthopedic implants include toxicity, immunogenicity, mutagenicity,
and infection propensity. Most assessments of biocompatibility of fracture fixation constructs focus
on immune responses to the implant. Immune responses can be triggered by material composition,
23
surface topography, as well as size, depending on the local implant environment, and can activate
traditional innate, complement, and humoral immunological pathways.
The implant surface-tissue interface is the single most important driver of biocompatibility in
metallic implants.39 Morphological features such as porosity, surface modification, and smoothness
as well as chemical characteristics such as hydrophobicity, wettability, surface charge, and polarity
influence an implant’s immunogenicity.40 Furthermore, corrosion (see below) may modify an
implant’s morphology, produce small particles from the implant, and ultimately reduce long term
biocompatibility. Soluble corrosion products can be immunogenic and can also be associated with
local cytotoxic and tissue toxicity reactions.
Biomaterials may be classified as inert, interactive, or viable.41 Inert biomaterials such as
cobalt-chromium alloys generate little or no biological response in the absence of wear and
corrosion. Titanium and stainless steel are classified as inert biomaterials. Titanium and titanium
alloys typically exceed the biocompatibility of stainless steel42 as well as have lower infection rates
over steel.43 Both titanium and stainless steel alloys commonly used in fracture fixation have very
good biocompatibility modulated by an inert, insoluble oxide layer that forms on the surface and is
chemically impermeable.
Interactive biomaterials differ in that they are designed to trigger biological responses such as
osteoinduction or osteointegration. Viable biomaterials contain a biological component and may
be resorbed or biodegraded. Fracture fixation implants are currently most commonly comprised of
inert biomaterials, but a construct that can adapt its mechanical properties as a fracture heals may
be desired. Early investigation into resorbable plates and screws shows potential in a rising field of
research.44,45
Biodegradable polymeric biomaterials, such as synthetic polyesters, are emerging as candidates
for orthopedic implant applications as biodegradation may eliminate the need for surgical removal
24
of implants. Degradation of synthetic polymeric biomaterials causes foreign-body reactions that
are measured using histopathology and currently their immunogenic hazards have little clinical
significance.46 Current research finds biomaterials such as polyetheretherketone (PEEK) to have
good strength and radiolucent properties that enable improved visualization of fracture healing by
computed tomography and standard radiographs. However, PEEK suffers from inferior fixation
strength and stability compared to titanium.47 Advances in polymer manufacturing and fracture
fixation technology may yet have clinical applications in fracture fixation.
Corrosion
Metallic materials used for fracture fixation in the body are susceptible to chemical attack,
termed corrosion. Corrosion is a collection of degradative processes resulting from metals reacting
with charged particles in solution. An oxidation reaction in which the metal loses electrons to the
surrounding solution starts a corrosive process. Metallic surface particles then either dissolve or
form an oxide on the material surface. Corrosion initiation is driven by the thermodynamics of
redox reactions and inhibited by kinetic barriers. For a given implant material, the thermodynamics
of reduction/oxidation reactions are fixed within a biological milieu. Therefore, a metallic implant’s
defense against corrosion is through development of kinetic barriers through passivation.
Passivation typically characterizes formation of a passive metal-oxide layer on the outer surface of
the metal. A continuous passive film presents a physical obstacle for the ion transfer necessary to
begin a corrosive process.
Broadly, corrosion can occur through several main mechanisms described as uniform attack,
galvanic corrosion, intergranular corrosion, crevice corrosion, and fretting corrosion.48 Fracture
fixation implants, however, are typically susceptible to galvanic, crevice, and fretting corrosion
pathways.
25
Galvanic corrosion is driven by an electrochemical potential gradient of metals in contact with
each other. While a larger problem for modular implants, galvanic corrosion can occur in fracture
fixation implants, for example, if the screws are made of a metal with a different electrochemical
potential than that of the implant metal. For this reason it is uncommon for a fracture fixation
construct to vary in composition between screws and implant.
Stress corrosion cracking is a subset of crevice corrosion in which pits grow on the surface of
the material. The ionic microenvironment in pits serves as a barrier between the local implant
surface and the surrounding tissues. The net effect of the pit microenvironment is to facilitate
corrosion through ionic exchange between the implant metal and nearby tissue. Additionally pit
geometry concentrates stress and promotes crack formation. When the stress intensity in a pit
reaches the critical value for crack propagation under corrosive conditions, a crack can propagate
and cause component failure. Unfortunately, the critical value for crack propagation under
corrosive conditions is typically lower than the fracture toughness of the material, making stress
corrosion cracking difficult to predict.
Fretting corrosion is corrosive damage resulting from cyclic motion between two opposing
surfaces. In this mechanism, grooves and oxide debris develop on the surface as a result of toggling
between contacting components. Fretting corrosion exposes the material beneath the passive oxide
layer. For example, nonlocking fracture fixation designs are at risk for fretting corrosion whereas
locking screws do not experience motion at the screw-plate interface and are less susceptible.
Experimental and Computational Modeling of Fracture Fixation Mechanics
The growth of orthopedic fracture knowledge facilitates a shift from hypothesis-based
medicine to evidence based medicine. This change in approach accompanies technological
developments in computer science, manufacturing, materials research, and mathematical methods
26
which combine to create more accurate laboratory models of interactions between biology and
medical devices than past iterations. Medical decisions should be made based on the best known
available models which, sometimes, are still clinically observation-based. As research efforts
expand to meet the evidence needs of modern fracture fixation, two distinct modalities emerge.
Experimental research in this context will refer to physical experiments, often involving mechanical
loading, kinematics, kinetics, or chemical phenomena. Computational experiments will encompass
those experiments conducted predominantly in silica, although experimental research accompanies
them as validation frequently.
Mechanical and biological environments involved in clinical fracture fixation are very
complex. Despite substantial advances in lab-based experimental and computational models,
clinical studies remain a vital part of modern investigation.
Experimental
Most lab-based experimental research in fracture fixation biomechanics and biomaterials
focuses on mechanical testing and/or bone healing. Mechanical testing involves a direct application
of load or displacement and is valuable because it enables direct measurement of a construct’s
response. The mechanical testing setup includes a loading apparatus, implant hardware, a fixation
mechanism, and a bone model. The most commonly used loading machines are either uniaxial or
biaxial standard mechanical testing frames retrofitted with fixtures to accommodate a fracture
construct. Implant hardware used during testing is commercially available hardware or, in the case
of testing new devices, custom made. Regardless of hardware, most fixation mechanisms selected
for testing are those that would be also used in clinical cases analogous to the experimental model.
Depending on the research question, the bone model can range in complexity from cylindrical
plastic pipe to compound synthetics to human cadaveric specimens. Such studies are typically
27
focused on the ‘time-zero’ mechanical response before any fracture healing occurs, unless the
fracture callus or bone healing is artificially simulated. Advanced composite fiber glass, epoxy
resin, and polyurethane synthetic bones can adequately represent certain bone mechanical
properties, but adequacy in representation of screw or implant fixation and loosening is still
controversial. A question regarding specific implant stresses and strains may be answered without
the need for expensive cadaveric bone specimens but a question focused on screw-bone fixation or
involving a complex fracture pattern may be best answered by recreating the fracture in a true bone
in the laboratory setting. These cadaver bones are typically stored frozen, not formalin-fixed
because of potential changes in mechanical properties associated with chemical preservation.
Experimental research can be used to test fracture construct performance during functional
activities or to simulate traumatic construct yield events. Generally of interest are failure scenarios
in the implant, the fixation mechanism, or in the bone. Static loading patterns are effective ways to
test yield failure modes, whereas cyclic loads can be applied to simulate fatigue failure. Both failure
modes are usually possible in a patient; yield if a patient has an accidental fall or other high load
event, and fatigue associated with long term physiological loading, especially if the fracture does
not heal quickly and the implants bear larger loads for longer time (Figure 1-7). Outcomes such as
construct stiffness and strength are commonly measured under loading conditions including axial
compression, torsion, bending, or a physiological combination thereof.49 Motion of bone fragments
or implant components during mechanical testing can be measured using infrared cameras, DVRTs,
or simple optical measurement methods using a digital camera.
28
Lab-based experimental research aimed at fracture healing is often conducted in animal
models. Animal models can be used to assess the impact of novel fixation constructs or biological
therapies on the healing response. For example, a recent study demonstrates a faster and stronger
healing response with active plating over conventional compression plating in sheep.50 Although
larger animal models are costly and time consuming they have the ability to simulate a biological
environment with similarities to that of a fractured human bone. Animal models can be exposed to
loading protocols in custom loading devices to simulate the mechanics of activities with ties to
human function.51,52 Furthermore, studies in smaller species enable the use of genetic modification
to understand mechanotransduction of fracture healing on a cellular and molecular level,53,54
although the metabolism and anatomy of these animals is often very different than humans, a hurdle
to clinical translation.
Computational
Computational fracture fixation research aims to simulate the mechanical environment of a
fixation construct. Due to limitations in computational power and modern characterizations of
biological materials it is not yet possible to accurately model all biological, chemical, and physical
Figure 1-7: Plot to quantify the race between bone healing (callus stiffness: blue) and implant
fatigue (stresses: orange), with data based on simulations with callus. The red line represents the
fatigue strength of the material. (Data from our laboratory)
bone
cartilage
granulation Cal
lus
Mat
eria
l Sti
ffn
ess
29
properties of implant hardware and biological tissue, however these features may have limited
impact on specific aspects of construct behavior. Therefore, major assumptions are made to
simplify development and run time of computational models of fracture fixation. The most
commonly used computational approach is the finite element method.
Computational models from previous studies focus on the effects of varying implant design,55
implant material composition,56–58 implant alignment and positioning,59,60 implant fixation,3,61,62 and
bone and fracture characteristics.63–65 Similar to experimental models, computational models often
focus on predicting implant failure, fracture gap strain, and implant load transmission.
Computational models have an important advantage of being able to predict displacement, stress,
and strain throughout all 3D points of a construct, whereas experimental measurements are often
limited to selected external points of a construct. Computational models may also simulate the time-
dependent effects of fracture healing on local mechanics, a challenging scenario to recreate in an
experimental model. Computational models enable comparison of results associated with variation
in a single model parameter, which may not be easily facilitated by experimental models, especially
if expensive biological specimens are involved. Computational models allow for multiple levels of
data interrogation while the physical nature of experimental models limit analysis to specifically
measured outcomes. However, computational models depend on user-defined inputs to run
simulations whereas experimental models allow for direct measurement of a physical system. A
computational model developer must understand and justify all inputs to the model. Experimental
models generally can provide better realism, especially when using cadaver tissues, and are thus
often used to validate computer models. In some cases though, experimental setups are less
accurate such as when measurements are sensitive to boundary conditions that are difficult to
precisely control experimentally but not computationally. Both computational and experimental
modeling modes can be very time consuming depending on setup and research question.
30
Internal Plating
Internal plating is a common method of treatment for many types of fractures. Plate
fixation can provide stability by transferring all of the load across a fracture or by sharing some of
the load burden with bone. Depending on the technique of application and design, plates can be
used for multiple functions: compression, bridging, buttress, protection, and tension band.66
Screw fixation between the plate and the bone must have sufficient strength to resist
failure. Screws can be locked or compression, uni- or bi- cortical, fixed angle or variable angle, and
can be inserted into as many or as few screw holes as desired. Locking screws are threaded into the
implant as well as the bone to maintain a gap between the two, whereas conventional screws rely
on frictional compression to hold the plate to the bone. Unicortical screws anchor in the cortex
closest to the implant and into any trabecular bone present beyond the cortex. Unicortical fixation
is necessary around articular joints and pre-existing prostheses, but is more susceptible to loosening
over time and may lead to construct failure via screw pullout from the bone.67 Bicortical fixation is
more rigid because the screws pass through the other dense cortex after crossing trabecular bone.
Variable angle locking screw holes allow for user-directed angulation of the screw relative to the
plate up to ~17o. Varying screw angle can allow the surgeon to accommodate for local anatomy
and other implants, or to avoid regions of poor bone quality for screw fixation. Unfortunately
variable angle locking plate-screw interfaces are not as strong as standard threaded locking
mechanisms,68 presenting an opportunity for additional improvements in plate systems.
Additionally, screw holes within a plate can be elastically suspended to enable dynamization of the
fracture in what are called active plates. Furthermore, the screw holes in the implant may be
designed to be slots instead of circular holes. Slot-shaped screw holes allow the plate to slide around
the screw which, like active plating, enables motion at the fracture.
31
Internal fixation plates are most commonly made of either stainless steel or titanium
alloys. It has been shown that stainless plates create a stiffer construct with a longer fatigue life
than titanium plates for applications at the distal femur.69 Other plate materials such as carbon fiber
PEEK composites have been shown to have comparable torsional stiffness to stainless steel but
inferior failure characteristics for distal fibula fractures.70 New hybrid materials such as
glass/flax/epoxy composites demonstrated higher ultimate strengths than conventional metals in
tension, compression, and bending while also having a lower axial stiffness,71 however, they are
not yet commonly used in the clinic.
Considering some of the subjective variables involved in internal plating construct
implantation (plate length, plate material, number of screws, screw fixation, fracture shape, plate
working length, etc.), it is evident that an optimization scheme could facilitate mechanically-
informed surgical decisions.72 Parametric models that iterate construct design variables have been
employed to observe the 3D mechanical effects of design changes under axial as well as combined
axial, torsional, and bending loads (Figure 1-8).21,73 Although the number of possible design
variable combinations is exponentially large, consideration of the surgical environment and local
anatomy and biology can focus the solution space and eliminate analysis of unlikely or impossible
construct designs such as one in which no screws are implanted. Model outputs such as implant
and screw stresses and fracture gap strain can also be used to predict construct failure and to observe
the mechanical influences of simulated healing. Thus, the utility of such models extends beyond
surgical decision support and preoperative planning into training applications for clinicians and
implant designers.
32
Intramedullary Nailing
Intramedullary (IM) nailing is a common method of treatment for adult shaft fractures
in diaphyseal bones. Nails are commonly used to provide relative stability for secondary
bone healing for fractures in the femur, tibia, and sometimes even the humerus. IM nailing
entails multiple decision points which may influence the healing course and even success
of a fixation construct. For example, the nail implant diameter can be selected to achieve a
tight or loose fit within the medullary canal. Additionally, the canal can be reamed to
accommodate a larger nail diameter and increase the contact area between the nail and the
bone. IM nails may be solid or cannulated. Solid nails were shown to have a lower risk of
infection in rabbits,74 whereas the more common cannulated nails may be surgically guided
by wire via their cannula during installation. IM nail length can dictate whether or not the
Figure 1-8: Examples of finite element models of subtrochanteric fracture fixation with a lateral
plate and screws subject to axial loading, showing effects of screw configuration and fracture gap
size on resulting implant stresses (colors) and fracture gap strains. (Right) Callus was represented
with springs having various stiffness. Blue represents low stress, red represents high stress. (Data
from our laboratory.)
33
nail engages bone in the distal metaphysis for stability. IM nails are often constrained
relative to the bone both proximal and distal to the fracture with screws (static locking).
Screws confer rotational as well as longitudinal stiffness to the construct.
Some femur nail designs accommodate two screws both proximal and distal to the
fracture. A common design to stabilize a proximal femur fracture incorporates a helical
blade which embeds within cancellous bone in the femoral head. Another design feature
common to femur nails is a dynamic interlocking screw slot. A dynamic screw slot allows
for guided translation of the fracture gap along the direction of the slot (inducing desirable
axial interfragmentary strain) while still conferring torsional shear stability with bicortical
screw fixation. Sometimes screws may not be installed on one side of the fracture to allow
for dynamic compression at the fracture site, although care must be taken in these cases to
ensure the nail does not perforate the bone on the unfixed side.
The first IM nail implants were made of stainless steel for their strength and
biocompatibility. However, stainless steel nails were too stiff to accommodate shape
mismatch between the nail and bone. These features were enhanced by a change to a Ti-
6AL-7NB titanium alloy which is now standard as a nail material. Titanium nails
demonstrate less slipping, more even stress distributions, and increased contact area within
the medullary canal over stainless steel nails in a computational model.56 Furthermore,
titanium nails reduce interfragmentary shear motion to better promote fracture healing over
stainless steel nails.75 Another study found that titanium nails are more stable than stainless
steel in torsion and axial compression, although both nail materials resisted failure at non-
weight-bearing loads.76
34
The mechanics of IM nails spanning a fracture are complex. The type and comminution
of the fracture dictates how much of the load applied through the bone must pass through
the nail. A fully comminuted fracture will require the nail to transfer all of the load across
the gap via the locking screws and represents the worst-case scenario in terms of implant
load bearing. Other factors that influence nail construct mechanics are the material
properties of the nail and screws, the cross sectional shape and anterior bow of the nail,
nail diameter, nail length, medullary canal reaming, and screw configuration. For example,
it has been shown that longer nails can generate higher contact stresses with the bone
medullary canal surface than shorter nails, and that the contact stresses can be mitigated by
increasing the flexibility of the distal end of the nail.77 An experimental model in cadavers
indicated that distal screws significantly increase maximum rotational load to failure in
unstable intertrochanteric fractures and recommended their use due to the improved
torsional strength.78
Contact within the medullary cavity is a significant mechanism for nail load transfer.79
The location and area of contact are influenced by nail size and shape, canal reaming, and
implant position and orientation. If the contact area is small, the load transferred to the
bone is concentrated and can generate high stress. This is thought to be a primary cause of
bone pain which some patients report after IM nail implantation. Increasing the area of
contact between the implant and the bone may reduce pain through design of the nail cross
section and radius of curvature, selecting an appropriate diameter nail for the bone, and
canal reaming. It has been shown that an 11mm-diameter nail with static interlocking
reduces motion at the fracture site up to 59% compared with a 9mm-diameter nail.80 Nail
implant modifications such as diameter, material, and cannulation as well as screw material
35
and area have been shown to reduce interfragmentary shear motion up to 54%.75 Prior work
combining computational and experimental modeling has validated finite element models
of an IM nailed femur at four stages of gait and suggested future improvements in implant
design and surgical implantation.81
Research in our laboratory using the finite element method has demonstrated
some effects of surgical and implant alterations on IM nail biomechanics (Figure 1-9).
These models utilize idealized bone and implant geometries. Additionally, more realistic
bone and IM nail geometries have been modeled for the femur. The femur model was built
using Mimics software suite from a patient CT scan. Separate materials are defined to
represent cortical bone and adjacent cancellous bone. The nail geometry is generated in
Solidworks software using the manufacturer’s design specifications. The nail implant is
aligned with the bone to represent surgical positioning and then the bone canal is reamed.
The implant and bone models are then imported into Abaqus software for finite element
analysis. The models are meshed and boundary conditions, contact constraints, and applied
loads are defined. Screws elements are positioned to align with screw holes in the nail and
are embedded within the surrounding bone mesh. A fracture can then be modeled by
defining a cutting plane and removing elements that intersect the plane. A complex 3D
model with physiological boundary conditions and applied loads can provide insight into
the location and magnitude of stresses in the implant, screws, and bone (Error! Reference
source not found. 1-10). Surgical variables such as nail positioning, bone quality, type and
number of fixation screws, applied loads, and fracture type may all be varied to view the
influence of these features on resultant stresses.
36
Figure 1-9: Finite element models of intramedullary nailed long bones subject to axial force and
bending. Effects on fracture gap motions and stresses due to construct configuration are evident.
Blue represents low stress, red represents high stress. (Data from our laboratory.)
Figure 1-10: Stress color map of an implanted IM nail within an unfractured femur under loading
simulating the stance phase of gait. Blue represents low stress, red represents high stress. (Data
from our laboratory.)
Addition of distal
screw
Single proximal
screw
Larger diameter
nail
Switch to dynamic
screw hole
Use of all four screws
37
Summary/Perspective
The 21st century has heralded advancement toward personalized medicine for the field
of orthopaedics. As computational technologies become more accessible and advanced for medical
applications, researchers identify novel challenges that can drive improvements in patient care.
Specifically for the field of fracture fixation, developed computational methods can be used to
iterate implant design concepts and suggest geometric, positional, and material variations to
achieve better optimization of construct mechanics. Furthermore, the detailed, accurate
approximations of mechanics now available from simulated fracture fixation experiments can be
used in conjunction with fabrication techniques such as additive manufacturing to create implant
concepts that were previously impractical to develop. As the one-implant-fits-all approach loses
traction in lieu of case-specific implant designs, it is expected that patient outcomes will continue
to improve, resulting in reduced time required for fixation, enhanced healed bone mechanics, and
fewer negative outcomes such as residual deformity and implant/fixation failure.
Acknowledgments
The authors gratefully acknowledge support from the AO Foundation, Switzerland (Project
S-15-196L), and the National Science Foundation/Penn State Center for Health Organization
Transformation. Hwabok Wee, PhD performed many of the finite element simulations shown in
figures. We also acknowledge contribution from April D. Armstrong, MD.
Chapter 2
A Review of Finite Element Modeling within Orthopaedics with Focus on
Femur Intramedullary Nails
Overview
Operative fracture fixation is being used increasingly to reduce disability, improve
outcomes, and improve quality of life following bony fracture. While fracture fixation is a generally
successful procedure, fracture management for the U.S. population over 50 years of age is estimated
to cost over $19 billion annually, with hip fractures accounting for 72% of these costs.2 Fracture
fixation failure occurs in up to 13.9% of patients and can result in nonunion, residual deformity,
and disuse osteopenia, with lasting impact on patients and their families.1,82 Revision of failed hip
fracture fixation is more costly than initial fixation, increases mortality risk, and significantly
lowers patient quality of life.83 These staggering numbers and debilitating morbidities motivated
development of research priorities for precision medicine in fracture management among leading
national orthopaedic focus groups.84
Surgical fracture fixation is achieved through the use of orthopaedic implants to stabilize
the fracture for bone healing.27,85 The fixation process involves reduction of bony fragments, when
possible, followed by implant positioning and fixation. Unfortunately, the surgical process is more
complex than this as fracture location and shape are driven by various causal injuries and local
bone quality. Furthermore, a surgeon’s toolkit includes multiple implants and fixation techniques
which must be narrowed to a single configuration for each case.86 Early fracture classification
strategies87 have evolved and informed the most recent AO Foundation Fracture and Dislocation
Compendium88 and can help to categorize common fracture patterns to assist in planning fracture
39
management. However, fracture fixation construct selection remains a largely subjective topic
among surgeons.
For example, most hip fractures are surgically fixed using a compression hip screw or an
IM nail. In such a case, the surgeon must select nail type and appropriate length, diameter, and
material, as well as decide whether or not to use fixation screws. Though there may be multiple
successful combinations of these design variables for a given case, each combination may have
relatively higher or lower risk of failure. In the case of the IM femur nail, failures are reported with
fixation screw bending and fatigue,89 screw cutout, and nail fatigue.90–92 Clinical outcomes of
fracture fixation are often binary: successful union or failure. However, an optimized fixation
construct might reduce the time required to achieve union or prove to be more capable of
withstanding unexpected loading than other constructs. Furthermore, early patient mobilization and
weightbearing is linked to better patient outcomes, especially for hip fractures in the elderly. Thus,
there is a need for systematic evaluation of fixation constructs as it could improve the financial,
physical, and emotional burden of failed fracture fixation in the U.S. healthcare system.93
Computational simulation provides an opportunity to determine specific treatment
recommendations based on individual case characteristics in numerous medical fields.94 Recent
advances in computational power and resource allocation have enabled multiple levels of
specificity in mechanical simulations, including bone geometry, fracture shape, and material
properties.95,96 However, there is not yet a process for taking patient imaging data and generating
model outputs on a clinically relevant timescale. Simplified computational models can be created
with relative ease, whereas precision modeling requires significant time and computational
resource allocation. There are various degrees of precision modeling that may or may not affect
clinically relevant outcomes. Therefore, the sensitivity of model outcomes to hierarchical levels of
precision modeling should be evaluated.
40
The most commonly used computational approach is the finite element method. The general
process of the finite element method for a mechanical problem is to discretize the system into a
finite number of smaller elements connected at nodes, approximate shape functions to describe
each element’s shape in all configurations of the system, implement boundary conditions, loading,
material properties, and contact formulations, develop the mass and stiffness matrices for the
problem, and solve a linear system of equations that represent the finite element approximation for
the system equations of motion. The resulting configuration of the system is defined in terms of
calculated values for degrees of freedom at nodes in the system. Once the configuration of the
system in both the reference and deformed configurations is known the desired output measures
can be calculated.
Experimental and Computational Modeling of Femur Intramedullary Nails
IM nails are the most commonly used femur fracture fixation implant and have been
studied in experimental and computational models since the early 2000s. Despite the numerous
variables present in a clinical femur fracture case, most research studies are limited by
computational power to assess one or only a few variables in an overly simplified framework.
Applied Loads, Boundary Conditions, and Constraints
It has been shown that FE femur nail simulation results are very sensitive to boundary
conditions and applied loads, and thus careful modeling choices are necessary.97,98 Because walking
is the most common functional activity for patients,99 simulations of femur fracture fixation often
use the maximum joint reaction force at the hip during gait as their applied load. A landmark paper
by Bergmann et al. established in vivo loads acting at the hip joint using instrumented total hip
41
arthroplasty implants during gait (as well as walking at variable speeds, stair ascent and descent,
knee bend, and standing from a chair) and expressed them as a percentage of patient bodyweight
for personalization in computational models.100 Most finite element femur models developed since
the Bergmann paper was published have used this loading convention.
Muscle forces are understood to significantly influence femoral loading101 but are
challenging to measure and model accurately, unlike joint reaction forces which can be measured
with implants instrumented with measuring devices.15 The most commonly used muscle loading
configuration in FE simulations of the femur was developed and simplified in a dynamic
musculoskeletal computational model of the hip by Heller et al.98 Their minimization scheme
systematically pared down the number of muscles contributing to their musculoskeletal model of
the hip until the calculated hip joint reaction force started to become ‘unphysiological.’
In addition to the influence of applied loads on simulation mechanics, some research
groups have focused on the mechanical influences of boundary conditions. For example, Speirs et
al. explored the sensitivity of motion in a finite element femur model to commonly used boundary
conditions.102 Their findings indicate that constraints that mimic closely the physiological
constraints on the biological system minimized large reaction forces near the points of constraint
application and minimized overall deflection of the hip. A more recent study evaluated the effect
of different constraint mechanisms on the femur during simulation.97 This study compared a
constraint in which the distal condyles are fixed in place (both with and without muscle loading) to
the physiological case suggested in the Speirs paper in which the center of the femoral head is
limited to one translational degree of freedom along the axis defining the applied load. This study
further confirmed the findings of Speirs et al., suggesting that physiologically similar constraints
enhance simulation validity and produce more realistic mechanics than nonphysiological
constraints.
42
The finite element modeling efforts described in Chapters 3 & 4 of this work are no
exception to the rule that applied loads and boundary conditions have great influence on simulation
behavior. For example, the mechanical testing experiments which are used for validation of the
models developed in this dissertation pot the distal femur in bone cement and fixturing such that
all translational degrees of freedom are locked and one rotational degree of freedom is limited.
However, the bone in the mechanical testing experiments was able to rotate in two degrees of
freedom about a pivot point within the medial condyle. The virtual nature of computational models
enabled implementation of a physical rotational constraint directly on this point in the finite element
model. However, on inspection the computational model appeared to be rotating in the locked
degree of freedom which was occurring because the boundary constraint was being applied to
trabecular bone elements which have low stiffness and were undergoing large deformations about
the constrained point. To remedy this, we kinematically chained a full circumference of points on
the nearby cortical bone to a reference point and enforced the boundary conditions at the reference
point. The resulting simulations were finally undergoing motion similar to the experiments on
which they would be validated after correction of this error.
Similarly, early simulations enforced the point of applied load within the center of the
femoral head because that is the location from which joint reaction forces were calculated in an
instrumented implant study.100 Anatomically, the hip relies on contact between the outer surface of
the femoral head and the acetabulum to transfer load to the femur. Thus, we adjusted our
simulations to kinematically chain a collection of points at the cortical surface of the femoral head
along the line of action of applied load to a reference point located at the center of the femoral head,
to which load was applied. However, it can be difficult to detect nuances such as these without
careful attention to boundary conditions and applied loads.
43
Interfacial Modeling
Choice of interfacial modeling strategies can also influence simulation outcomes. For
example, peak minimum principle strain in fixation screws decreased 75% when comparing a
thermal expansion/undersized pilot hole method to a simple cylindrical interface with friction.62
Because screw cutout from bone is a common mode of implant failure, especially in cases of poor
bone quality such as osteopenia and osteoporosis, it has been suggested that high-resolution finite
element models are needed to accurately model the local non-linear mechanical behavior at the
screw-bone interface.103 Furthermore, failure predictions at the screw-implant interface for plated
fracture fixation models are significantly influenced by interaction properties between the screw
head and the implant.104
However, sophisticated screw modeling approaches are computationally intensive and
may be unnecessary in some cases. Specifically, literature supports that complex screw modeling
such as including thread geometry or a simpler pseudo-threaded model both offer a more accurate
depiction of local bone strain than a bonded interface.105 However, the results of the same study by
Inzana et al. also indicate that relative comparisons of implant stability were not affected by the
modeling choice for the bone-screw interface, even when a simple bonding interaction was used.
Thus, the precision of interfacial modeling techniques can be selected based on desired outcome
measures and the degree of accuracy required to address the research questions.
What Biomechanics Research Has Taught us About Intramedullary Femur Nailing
Surgical femur fracture fixation involves multiple decision points, many of which are the
focus of current surgical debate. Specifically, surgeons first select the type of implant to be used in
the construct. The IM nail is the most commonly used implant for femur fractures and is often
44
preferred over other implants such as the distal hip screw due to better short term outcomes.106 If
an IM nail is to be used, however, the surgeon then needs to decide if they will ream the endosteal
canal to accommodate the nail. Data to support a preference for either reamed or unreamed nailing
is inconclusive and reaming remains an unanswered question.107 Furthermore, the size of the native
and/or reamed endosteal canal presents a challenge when selecting nail diameter which will dictate
the size of the gap between the implant and the endosteal canal. Femur nails are most commonly
made of titanium alloy but are also available in stiffer stainless steel. A nail can be long, spanning
nearly the full length of the femur, or short, extending only into the diaphyseal region. The nail can
be locked in place by distal fixation screws to constrain motion or can be dynamized through
removal of screws or use of slots rather than holes on the implant.
Most FE modeling studies of femur fracture fixation measure similar outcome variables:
maximum von Mises stress in the implant and screws, strain in the bone (often measured as
minimum principle strain), and a quantification of motion at the fracture site (measured as
displacement or strain). Von Mises stress is a theoretical predictor of material yielding for ductile
materials. The elastic limits of materials are tested under single dimensional load conditions under
which their stress-strain behaviors are described. However, the stress state on 3D bodies in complex
FE models is almost never one-dimensional. Thus, the von Mises stress is a criterion calculated
from the full 3D stress tensor that is approximated to predict yielding at a material point if the von
Mises stress equals or exceeds the simple tension yield limit stress.
Strain in the bone is commonly represented as minimum principle strain. The minimum
principle strain is the largest normal compressive strain at the material point being evaluated. Thus,
the minimum principle strain can be compared to the strain limit of a brittle material such as bone
as a predictor of yield.
Motion at the fracture site is a complex quantity to describe because there are many ways
it can be represented. The motion of individual point pairs can be measured but this can be
45
obfuscated by rotation of the two fragments relative to each other. A central virtual common point
between fracture fragments can be rigidly fixed to each fragment and motion of one fragment’s
point can be described as a translation from the other fragment’s point. In the case of gapped
fractures, fracture gap strain can be calculated and is thought to have a link to secondary fracture
healing (see Perren Strain Theory in Chapter 1). However, strain can be a confusing measure
because it is biased by the size of the initial gap (1mm of motion in a 10mm gap gives 10% strain,
whereas the same motion in a 5mm gap shows 20% strain). In cases with little or no fracture gap
and high stability between fragments that are expected to heal by primary bone healing, strain is
less meaningful and displacements are favored. In general, the description of motion used in
literature is customized from the above options based on the research question(s) being explored.
Selecting an Implant
Several FE studies have been conducted to assess biomechanical differences between
potential implants for femur fracture fixation. One study, focusing on distal diaphyseal femoral
fractures, compared the stresses between a laterally placed osteosynthesis plate and an IM nail
when used to fix the same fracture.108 Their findings indicated that the IM nail experienced higher
von Mises stress with distal femur fractures than the osteosynthesis plate but that this effect was
mitigated as the fracture gap was increased from 1 to 3mm. They further discuss that their model
indicated that the plate induced fracture healing through intramembranous ossification, maintaining
absolute stability at the fracture site, while the IM nail would induce endochondral ossification,
maintaining only relative stability at the fracture site. A later, similar study by a different research
group compared retrograde IM nailing to locking plate fixation for periprosthetic distal femur
fracture following total knee arthroplasty.109 This study by Shih-Hao et al. finds agreement that the
IM nail experiences higher stress than plate fixation and finds that both constructs sufficiently
46
stabilize a distal femur fracture. However, despite higher stresses in the nail, a biomechanical and
computational series of studies demonstrated that locking plate fixation had a higher probability of
fracture (21.8%) than IM nail fixation (0.019%).110,111 Conventional wisdom suggests that nails
have smaller stresses than plates for diaphyseal and proximal femur fractures have smaller peak
stresses than plates due to their positioning closer to the axis of load application.
Furthermore, if an IM nail has been decided upon for hip fracture fixation, there are still
several styles of nail available from each major implant manufacturer. As an example of this, a
study from 2009 evaluated bone strain and implant stress when a trochanteric fracture was virtually
fixed with a Gamma-Nail (Stryker), Gliding-Nail (Plus Orthopedics), Proximal Femoral Nail-A
(Synthes), and Targon-Proximal Femoral Nail (Aesculap) in a FE model.55 In this study similar
loads (including muscle forces) and boundary conditions were used as previously described by the
Bergmann and Duda research groups and the implants were compared to each other in ‘ideal’ and
slightly misaligned positions. The findings from the study indicate that, despite many geometric
similarities, each nail has a unique positioning error tolerance but that overall caudal alignment
should be preferred to cranial alignment to provide more favorable fracture healing conditions. This
finding is further supported by the results of a 2013 paper which evaluated the influence of implant
alignment in an intertrochanteric fracture finite element model.112 The study was limited to sliding
hip screw implants rather than IM nails and demonstrates that increasing the tip-apex distance
(TAD) of the lag screw (the distance between the tip of the screw and the apex of the femoral head),
decreased the risk for screw cut-out from the proximal femur. Although the risk of failure via screw
cutout is much higher in the sliding hip screw (80%) than in IM nails (8.5%),113 cutout remains an
important failure mode for all cephalomedullary devices.114
Thus, not only must an appropriate implant be selected, but it must also be implanted in an
optimized orientation and position. For this reason, a device called the variable angle nail was
developed which enables motion between the lag screw and the nail intraoperatively thus
47
influencing the TAD. A FE study compared the variable angle nail to a traditional nail design
(Gamma 3) and showed similar stiffness and fatigue characteristics between the two.115 However,
a 2018 retrospective case study of 83 patients with pertrochanteric femur fractures treated with
cephalomedullary nails showed that cases treated with a higher TAD did not increase the risk for
lag screw cutout.
Furthermore, surgically there are 2 common approaches for antegrade nailing – trochanter
tip and piriformis entry. Some nails are designed specifically for one type of entry or the other. The
surgical entry point used to install a cephalomedullary IM nail was studied in a FE model.60 It was
shown that the trochanter tip starting point exhibits higher bone strains than the piriformis entry
point, increasing the risk of iatrogenic fracture of the proximal femur when using a trochanter tip
entry point.
An additional consideration when selecting implant design features for fracture fixation is
the type of fracture. Specifically, it has been shown that “IM nails fail at a lower load in an unstable
fracture situation in the proximal femur than in a stable fracture.”116 Specifically, nails spanning
gapped fractures failed at 72% of the load at which they failed in reduced fractures. This study
further indicates the location of maximum stress is at the aperture of the lag screw, where fatigue
fracture of the nail is likely to begin. Another study by the same research group simulated a
pertrochanteric fracture, a lateral neck fracture, and a subtrochanteric fracture fixed by a
cephalomedullary nail.117 All three fractures were similar in global construct stiffness but differed
in maximum stress (665MPa for the subtrochanteric fracture, 621MPa for the pertrochanteric
fracture, and 480MPa for the lateral neck fracture). This further supports the idea from their
previous study that nails bear higher stresses in unstable fractures cases than in those that are stable.
In such cases, additional fixation techniques (i.e. supplemental plating) may be considered to
mitigate failure.
48
To Ream or Not to Ream
A question of constant clinical debate surrounding IM nailing for femur fracture fixation
is whether or not to ream the intramedullary canal.118 Reaming a canal can enable easier
implantation of an IM nail, reduce the risk of nonunion, and can facilitate usage of a stiffer, larger
diameter implant.119 However, reaming the bony canal decreases diaphyseal cortical thickness and
can increase the risk for bony fracture and nail punch-through. Reaming can also disrupt local
vascularization, essential for healing, in the innately poorly vascularized bone tissue.120 FE studies
are particularly well positioned to elucidate some of the biomechanical differences between reamed
and unreamed nailing but, unfortunately, canal reaming is typically not considered nor even
mentioned in most FE nailing studies.
One FE study published in the Journal of Orthopaedic Research in 2006 used an FE model
to predict interfragmentary strain at the fracture site, von Mises stress in an IM nail with locking
screws, and bony strain distributions for reamed and unreamed nailing of 3 different tibia
fractures.121 Their findings show that the unreamed nail is more likely to fatigue fail than the reamed
nail because the stresses at the interlocking bolts were higher in the unreamed case. The two cases
showed nearly equivalent maximum von Mises stresses in the locking screws with a proximal tibial
fracture but the mid-diaphyseal and distal fractures were both accompanied by higher screw stresses
in the unreamed case.
Another study specifically targeting the effect of reaming assesses how the mechanical
contact interface between the implant and the bone is affected by gap geometry and thickness.79
Specifically, their study used a simple transverse subtrochanteric fracture fixed by a
cephalomedullary nail under simulated maximum gait loading. Their model included a scenario
with perfect contact between implant and bone for the entire length of the nail (unreamed) as well
as a case in which there is initially no contact between the implant and bone (reamed). Their
49
findings indicate that larger endosteal reaming gaps lead to greater medial stress concentrations in
the implant while the unreamed case provides a more uniform stress distribution on the implant.
Nail Diameter and Material
IM nails are clinically available with multiple shaft diameters commonly between 9-12mm,
although ranging up to 17mm. Nail diameter influences the cross-sectional area moment of inertia
such that an increase in diameter increases the bending stiffness of the nail. One FE study using a
human tibia model demonstrated that shear motion at a diaphyseal gapped fracture site could be
reduced by adjusting nail alignment in the anteroposterior and mediolateral directions and by
increasing nail diameter and/or elastic modulus.75 Another FE study found a 51% decrease in
maximum nail stress in a subtrochanteric fracture model by increasing the diameter of a gamma
nail from 15.5mm to 17mm.122 It is intuitive that increasing nail diameter increases construct
stiffness and decreases motion at the fracture site. However, variations in reaming, patient bone
geometry, and bone quality can limit the opportunity for a large diameter nail. Under such
circumstances the two aforementioned computational studies evaluating nail diameter can provide
insight into the biomechanical costs of using smaller diameter nails.
Similar to nail diameter, nails made of materials with high modulus of elasticity such as
stainless-steel increase total construct stiffness. Early IM nails were nearly exclusively made of
stiff stainless steel (elastic modulus ~190GPa). However, this is much stiffer than healthy cortical
bone (~16 GPa) and thus these nails experienced high compressive and bending stresses.
Furthermore, the stiff stainless steel nails can stress shield the bone, thus reducing bony internal
and leading to bone resorption. Titanium alloy is now the market standard for nail material and has
been shown to decrease slipping, produce even stress distributions, and increase contact area within
the medullary canal over stainless steel nails in computational models.56,75 It has also been shown
50
that titanium nails are more stable than stainless steel in torsion and axial compression, although
both nail materials resisted failure at non-weight-bearing loads.76,123 Furthermore, titanium alloy is
more biocompatible than stainless steel, thus decreasing unintended immune responses. Additional
materials for intramedullary nails such as polymers and composites have been tested. A bovine
femur model demonstrated that polymeric femur nails resisted failure under static conditions,
however, maximum von Mises stress during walking exceeded the yield stress for each polymer.124
Polymeric materials are not yet strong and durable enough to withstand the dynamic loading of
fractured long bones.
Static vs. Dynamic Distal Locking Screw Fixation
IM nailing constructs can be dynamized by removing or adjusting the static fixation
enforced by distal fixation screws. This makes the distal fixation dynamic, allowing relative motion
between the implant and the distal fixation. In some cases dynamization is thought to be preferred
because bone heals in response to mechanical stimuli. Construct dynamization would promote
secondary bone healing with relative stability rather than primary bone healing which requires
absolute stability. Secondary healing is utilized especially in fractures with high comminution or
loss of bone in which absolute stability may not be attainable.
In 2004 a Synthes retrograde nailing system was evaluated in an unfractured femur model
under simplified gait loads with a 6 degree-of-freedom (DOF) constraint at the distal femur.81 The
initial focus of this study was to validate the FE model with mechanical testing in a composite bone
instrumented with strain gauges positioned along the outer cortical surface. Once agreement
between computationally predicted and mechanically tested strains was confirmed, the authors
assessed the effects of dynamization via screw removal and implant material on forces, moments,
and von Mises stress in the implant and screws. The study concluded that reamed, statically locked
51
retrograde IM nails should be used with all screws in place due to concerns about the effects of
screw removal on fracture healing that had been raised previously with a telemetrized (instrumented
with sensors capable of reporting force, pressure, and positional data) IM nail.17 However, the
unfractured nature of the bone model in their simulations is likely to have influenced their results,
particularly reducing stresses in the implant and screws due to decreased deformation relative to a
model which includes a fracture.
A more recent FE study also looked at nail dynamization but in the context of a fractured
bone model.3 While still employing the same distal boundary condition as the previously described
model, these simulations included an approximation for the gluteus medius muscle in addition to
the applied reaction force at the hip joint. Furthermore, the fracture models in this study were
diaphyseal with initial fracture gaps which closed when the system was dynamized by removing
the distal locking screws (Figure 2-1). The results show that dynamization increased contact area
and contact pressure at the fracture site in both proximal and distal diaphyseal femoral fractures,
eliminated potential failure at the distal screws, and decreased the maximum stress on the implant.
52
A 2013 biomechanical study challenged the findings of the above two presented FE studies
by assessing performance of static and dynamic distal locking screw fixation in long
cephalomedullary nails with unstable subtrochanteric fractures under torsional loading.125 Their
results showed, not surprisingly, that distal locking increases the rotational load to failure.
However, they found that the maximum torsional load to failure in the unlocked group falls within
the functional range of adult hip loading. Thus, they highly recommend use of the distal
interlocking screws due to the rotational failure not previously considered in the FE studies.
A clinical study in 2016 further demonstrated the complexity of the locking vs. unlocking
debate by evaluating removal of distal locking screws in the setting of short IM nails in stable
pertrochanteric fractures.126 The findings here indicate that removal of the distal locking screws in
stable fracture patterns was no different than the group treated with locking screws in terms of
walking function, residual pain, and overall satisfaction at one year followup. Furthermore, the
group without screws was associated with shorter operation time, less radiation from imaging,
shorter surgical incision length, and less blood loss and thigh pain than the locking group.
Thus, short nails spanning stable fractures appear to be preferred without distal locking
screws while long nails spanning unstable fractures likely require the use of distal locking screws
to avoid rotational failure. But what about long nails in stable fracture patterns? Short nails across
a gapped fracture? How would the recommendations from the FE studies differ if torsional loading
had also been considered? The disconnect between the early FE studies and the biomechanical
study is explained by the disregard for rotational loading in the FE studies but serves as an important
reminder to consider the context in which the results of both experimental and computational
Figure 2-1: Nail dynamization by removing distal locking screws allows for closure of the fracture
gap under axial loading. Image from Shih et al.3
53
studies will be considered. The results of the studies are validated within the experiments tested
but, when taken into a real-world clinical setting, will be extrapolated beyond the overly simplistic
conditions the experiments were conducted in. It is essential that the findings of clinically-relevant
work are presented within the constraints of their experimental framework so that their results are
not easily misinterpreted. To this end, it is essential that the gap between translational research and
clinical practice be shortened such that the research questions can be designed to have clinical
relevance and that clinical practice can be informed from research findings.
Techniques Accounting for Biological Variability
Biological tissues can be a challenge to model virtually due to variability in shape,
mechanical properties, and complex interactions between tissues. For example, although a majority
of long bone fractures fall into categories identified by the AO Foundation, nuanced geometry in a
specific fracture or implant can influence the mechanical outcomes of the system. Model geometry
for fracture fixation problems can be challenging to obtain due to variability and background noise
in medical imaging data. When available, 3D imaging data such as Magnetic Resonance Imaging
(MRI) or Computed Tomography (CT) can be used to generate a virtual description of the anatomy.
Several methods are available for image segmentation, during which efforts are made to delineate
signal from noise as well as distinguish tissues which may vary in mechanical properties (e.g.
muscle, bone, tendon, cartilage, etc.).95,127,128 Previous research into optimization strategies for
isolating refined geometric models from medical imaging data has yielded results, although no
single software or process is universally accepted.127,129,130
Due to radiation exposure with CT imaging and high cost and time inherent in both MRI
and CT imaging modalities, they may not be indicated for all patients. In such cases X-ray imaging
54
is the most commonly used diagnostic tool for fracture cases and thus only a 2D image set is
available. Strategies for developing 3D representations of geometry from plain film X-rays have
been developed but are more error-prone than MRI and CT.131,132
Recently, population-based methods have been developed for estimating patient-specific
features in computational models. For example, statistical shape modeling (SSM), which will be
discussed further in Chapter 5, can be used to estimate patient-specific femur geometry based on
demographic descriptors such as sex, age, BMI, and height.133–135
Finite Element Mesh Generation
Finite element meshes for musculoskeletal tissue geometry are sometimes generated in a
voxel-based manner, with resolution up to the resolution of available imaging data.136 Trabecular
bone has a complex structure that is sometimes modeled as a series of repeated units137 but has also
been represented with meshes generated from high-resolution MicroCT scans.138 All elements in a
voxel-based mesh are equal in size. Enhanced resolution in a voxel-based mesh can be achieved
via MicroCT scanning over traditional CT imaging. Voxel-based models typically require
additional processing to delineate boundaries and smooth surfaces. Small features may become lost
in a voxel-based mesh with poor resolution compared to relatively smoother tetrahedral meshes
(Figure 2-2).4 High-resolution voxel models may contain many elements and result in a very large
global stiffness matrix for the finite element formulation. One technique to obtain a
computationally efficient model from a dense voxel mesh is to use an element-by-element
technique which never assembles a global stiffness matrix.139 Several strategies have been
developed for automated mesh generation from object boundary data, voxel-based data, and hybrid
combinations of the two.140
55
Modeling Applications
The field of orthopedic medicine has many various applications for computational modeling.
Bone has been simulated in computational models for crack propagation141, plasticity142, and
microstructural143 analyses. As mentioned in the overview, orthopaedic joints have been modeled
extensively in native and medically-intervened states. In such models, stress and strain fields are
valuable for calculating and predicting outcomes such as translation and rotation under load,
contact pressure, contact area, and others in response to medical interventions such as joint
replacement, soft tissue repair, and fracture fixation. Cartilage and soft tissue have been modeled
in finite element formulations using a biphasic theory with fluid modeled without viscosity and
solids modeled as incompressible.144 Beyond biology, orthopedic implants themselves are also
Figure 2-2: Finite element models of a femoral head trabecular specimen created at a voxel
resolution of 84 μm (left) and 168 μm (right) using either voxel-based meshing (top) or a tetrahedral
meshing approach (bottom) from Ulrich et al.4
56
modeled for geometric design iteration, failure analysis, predicting stress and strain profiles, and
predicting mechanical interactions with biological tissues.
Painstaking validation of a femur bone model was conducted by Huiskes et al using 112
rosette strain-gauges glued to a cadaveric femur during mechanical loading.5 Previous work such
as this demonstrates why a computational model can be preferred to physical experimentation
(Figure 2-3).
Figure 2-3: Cadaveric femur instrumented with a host of strain gauges for computational model
validation from Huiskes et al.5
57
Coordinate Systems for Description of Motion
A common challenge for translating results from computational simulation to patient care
is to express the results in terms of anatomical coordinates. Human bones rarely conform to
common engineering geometries and our joints often undergo combined motions that are
sometimes not well approximated by simple definitions such as hinge and sliding joints. For this
reason Grood and Suntay, in a landmark paper, established recommendations for systematic
determination of coordinate systems for bones and joints.145 In this paper specific anatomic
landmarks are used to construct axes for describing translations and rotations relative to clinically
relevant orientations such as anteroposterior, mediolateral, and inferosuperior. For computational
models, anatomical landmarks are often digitized relative to each other or to fiducials and reference
frames to store relative position and orientation data between model components.
Further development of anatomic coordinate system concepts have been established for
specific joints such as the knee146, although modifications have been suggested recently to account
for errors in flexion angles lower than 0o and higher than 90o.147 An in-depth technical description
of the joint coordinate system convention published by MacWilliams and Davis 35 years after the
Grood and Suntay paper equates the convention to a x-y-z Cardan angle sequence, enforcing
sequence dependency and orthogonality.148 Understanding common conventions for describing
motion in medically relevant contexts is essential for computational model utility.
58
Chapter 3
Parametric Finite Element Analysis of Intramedullary Nail Fixation of
Proximal Femur Fractures
Text in this chapter has been submitted to the Journal of Orthopaedic Research and is
pending review (invited for resubmission following revisions).
Authors: Scott M. Tucker, Hwa Bok Wee, Edward Fox, J. Spence Reid, Gregory S. Lewis
Abstract
Proximal femur fracture fixation with intramedullary nailing relies on stability at the
fracture site and integrity of the fixation construct to achieve union. The biomechanics that dictate
fracture site stability and implant stress depend on fracture type as well as implant features such as
nail length, nail diameter, presence of distal fixation screws, and material composition of the
implant. When deciding how to fix a fracture, surgeons have choices in these implant-related design
variables. This study models all combinations of a range of implant variables for nine standard
AO/OTA proximal femur fractures using finite element analysis. Under simulated maximum load
during gait, the maximum stress in the implant and screws as well as interfragmentary motions at
the fracture site in the axial and shear directions were computed. The results were separated by
fracture type to show the influence of each design variable on measured biomechanical outcomes.
Filling the reamed canal with the largest fitting nail diameter reduced axial and shear
interfragmentary motion for all fracture types. Nail length was less predictive of shear
interfragmentary motion for most simulated fracture types than other construct variables.
Furthermore, gapping at the fracture site predisposed the construct to higher implant stresses and
larger interfragmentary motions. Clinical Significance: Biomechanical outcomes from this
59
computational study can directly aid in surgical decision-making for optimizing hip fracture
fixation with intramedullary nailing.
Introduction
Surgical fracture fixation is achieved through the use of orthopaedic implants to stabilize
the fracture for bone healing.27,85 The fixation process involves reduction of bony fragments
followed by implant positioning and fixation. Unfortunately, the biomechanics of fracture fixation
can be complex, in part due to variations in bone quality, anatomy, and local loading conditions.
Furthermore, a surgeon’s toolkit includes multiple implants and fixation techniques which must be
narrowed to a single configuration for each case.86 Despite a general understanding that fracture
geometry and fixation construct influence the biomechanics essential for proper and timely healing,
fracture fixation construct selection remains a largely subjective topic among surgeons and is often
linked to surgical training experience.149
At present, most hip fractures are surgically stabilized using an intramedullary (IM) nail.
The surgeon selects nail type, length, and diameter, as well as whether or not to use distal locking
screw fixation. Though there may be multiple successful combinations for a given case, each
configuration has a relatively higher risk of failure when compared to a theoretically optimal
construct.150 In the case of the intramedullary nail fixation, clinical failures are reported in up to
14% of patients1 via, most commonly, fixation screw bending and fatigue,89 cut-out, and nail
failure.90–92 Excessive interfragmentary motion at the fracture site resulting from instability can
result in high strain and disrupt healing, leading to nonunion.151 However, especially for patients
with unstable fractures or compromised healing capacity, an optimized fixation construct can
reduce the time required to achieve union and mitigate the risk of implant failure and nonunion
60
events. Thus, there is a need for systematic evaluation of fixation constructs to improve the
estimated $13.7 billion financial burden of hip fracture management in the U.S.93
Computational simulation provides an opportunity to determine specific treatment
recommendations based on individual case characteristics.94 Previous work evaluating
computational biomechanics of IM femur nails has been almost exclusively limited to variations in
a single construct variable such as fracture type,117 bone canal reaming,79,121 implant
positioning,60,152 nail material properties,153,154 nail size,155 and fixation screw use.3,125 A recent
study investigated multiple IM nail design variables but only evaluated transverse fracture types.123
Furthermore, physiological muscle loads may influence construct biomechanics but are only
sometimes considered in finite element femur fracture models.156,157 Thus, it is imperative that IM
femur fixation nail design variables be modeled with a holistic approach to evaluate synergistic
effects of combinations of fracture type and construct variables on biomechanics.
To our knowledge, this study provides the most comprehensive investigation of IM nail
biomechanics for proximal femur fractures to date. Utilizing a novel parametric high-throughput
modeling and analysis pipeline, this study investigates the effects of hip fracture type, IM nail
diameter, IM nail length, IM nail material, and use of distal fixation screws in a finite element
model of the heel strike phase of gait on maximum implant stress and interfragmentary motion.
The results of this study can be utilized toward optimizing clinical implant decisions based on
simple fracture classification.
61
Materials and Methods
Bone and Fracture Model
Simulations were developed using a previously validated generic virtual femur model.158
The femur was modeled as a composite of homogeneous isotropic linear elastic materials: cortical
bone (elastic modulus 17GPa, Poisson’s ratio 0.3) and trabecular bone (elastic modulus 155MPa,
Poisson’s ratio 0.3). Planar fractures were cut into the femur model using Solidworks (Dassault
Systèmes, Waltham, MA). Simulated fracture patterns were modeled after the AO/OTA fracture
classification scheme88 (Figure 3-1) and were selected to include common pertrochanteric,
intertrochanteric, and subtrochanteric fracture types. Specifically, fracture types 31A-1.1, 2.1, 2.2,
3.1, 3.2, and 3.3 were modeled. Fractures 31A1.1 and 31A2.1 were modeled as perfectly reduced
as well as with a 5mm gap simulating comminution. Subtrochanteric fracture type 31A3.2 was also
modeled with a 10mm gap. Gap sizes were selected such that the major fracture fragments did not
make contact under the maximum applied load.
62
Implant, Reaming, and Construct Configuration
A cannulated femur nail was modeled in Solidworks based on geometry of a popular IM
nail (Synthes Trochanteric Femoral Nail). The helical blade and lag screw (modeled as cylinders)
as well as the 15mm diameter proximal nail region were present in all cases. However, clinically
relevant variations in distal nail diameter (10, 11, 12, and 13mm) and nail length (long 43.3cm and
short 26.9cm) were implemented. The nail was centered relative to the femoral isthmus with the
proximal entry point located just medial to the superior tip of the greater trochanter. All bone
models were reamed to a distal diameter of 13.5mm, creating a concentric gap between the nail and
bone ranging between 0.25mm (for the 13mm diameter nail) and 1.25mm (for the 10mm nail).
Additional implant construct variations were modeled by adjusting the material properties to those
of titanium alloy (elastic modulus 110GPa, Poisson’s ratio 0.3) or of stainless steel (elastic modulus
Figure 3-1: Simulated proximal femur fracture types (type 31A-X.X). Types 31A-1.1, 2.1, and
3.2 were additionally modeled with gaps that remained open under the applied load. Point pairs
used to calculate interfragmentary motions are shown.
63
195 GPa, Poisson’s ratio 0.3) and by removing the two 5mm diameter distal fixation locking
screws. Fixation screw holes were present in the bone for both the long and short nail in all
simulations, even when fixation screws were not used (Figure 3-2). Model input file generation
was automated using custom Matlab (MathWorks, Natick, MA) code (Table 3-1).
Loading and Boundary Conditions
Finite element analysis was conducted using Abaqus (Dassault Systèmes, Waltham, MA)
software. Applied loads and boundary conditions were modeled for an 85.7kg person consistent
with previous FE simulation and mechanical testing experiments117 using a previously described
femur coordinate system.159 Briefly, a 2000N compressive load was applied at the center of the
femoral head oriented with a 13° abduction angle and directed 8° posteriorly from the long axis of
the femur to represent the maximum hip reaction force during human gait.100 The center of the
femoral head was constrained in translations orthogonal to the load vector. Distally, all translational
degrees of freedom as well as rotation about the epicondylar (knee flexion) axis were constrained
Table 3-1: Fracture fixation construct configurations modeled. All combinations of depicted
design variables were simulated.
Design Variable Options Cumulative
Combinations
Fracture Type/Gap
Status
31A-1.1, 2.1, 2.1, 2.2, 3.1, 3.2, 3.3,
and gapped fractures 31A-1.1, 2.1,
3.2
9
Nail Length Long, Short 18
Nail Diameter (mm) 10, 11, 12, 13 72
Distal Fixation Yes, No 144
Material Stainless Steel, Titanium 288
64
at a pivot point 23mm medial to the shaft axis in the medial condyle160 (Figure 3-2). Femur
internal/external rotation and rotation in the frontal plane were left unconstrained.
Based on pilot study simulations that demonstrated effects associated with ignoring muscle
forces, simulations included physiological muscle loads. Heller et al. determined a simplified
muscle-loading scheme for walking that maintains physiological hip-joint loading.98 Specifically,
body weight-scaled contributions from the gluteus medius, minimus, and maximus, vastus lateralis,
and the proximal and distal part of the tensor fascia latae were simulated. Muscle loads were applied
to single nodes translated onto the geometric femur model using the closest proximity nodes to the
muscle point-of-action coordinates described by Heller et al.
Figure 3-2: Two design configurations under maximum applied load with the A32 (top) and A11
(bottom) fracture types demonstrating a long, 10mm diameter nail with distal fixation screws (top)
and a short, 13mm diameter nail without distal fixation screws (bottom). Applied loads and
boundary conditions at the center of the femoral head and implant meshes are shown. A sphere of
bone is visible in the medial femoral condyle to indicate where femoral boundary conditions are
applied. All bone models were pre-drilled for both long and short nail distal fixation screws,
regardless of configuration.
65
Finite Element Simulation Parameters
Hard contact with Coulomb friction was simulated between IM nail and bone (µ=0.42), IM
nail and fixation screws (µ=0.20), and bony fragments at the fracture site (µ=0.46) based on
literature-reported values.112 Tie constraints were implemented between fixation screws and bone
to simulate the effect of screw threads and between the simplified cylindrical proximal head screw
and proximal trabecular bone. Quadratic tetrahedral elements were used to mesh all components
and totaled around 170,000 elements for each simulation. Mesh convergence analysis was
conducted by increasing the mesh to approximately 250,000, 300,000, and 400,000 elements.
Maximum implant stress differed by no more than 16MPa, axial motion differed by 0.17mm at
most, and shear interfragmentary motion differed by 0.07mm at most in convergence trials.
Validation, Outcome Measures, and Statistical Methods
The model validation was supported through comparison of global axial stiffness to
literature reported values from a study with very similar loads and boundary conditions, and small
differences in implant geometry.117 For pertrochanteric and subtrochanteric AO/OTA fracture types
31A1.1 and 31A3.2 the computational model resulted in a global axial stiffness 16% and 15%
stiffer than mechanical testing results, respectively.
The computational model outputs included maximum implant vonMises stress for the nail
and superior distal locking screw as well as interfragmentary motion. Interfragmentary motion was
calculated by averaging absolute displacements between 4 point pairs distributed about the fracture
plane in the anterior, posterior, medial, and lateral directions (Figure 3-1), a method adapted from
the literature.161 These displacements were transformed into axial and shear motions with respect
to the fracture type.
66
Statistical analysis was performed with JMP software (SAS, Cary, NC) using standard least
squares multiple linear regression for each fracture type with four design variables as regressors
(nail diameter, nail length, nail material, and distal screw fixation). Statistical significance was
determined at p<0.05.
Results
Individual simulation outcomes are shown in Figures 3-3a-b to demonstrate trends for
each design variable. Results for multiple linear regression for each outcome across all models
grouped by fracture type are presented in Table 3-2. Outcomes for configurations with stainless
steel nail material properties are omitted from the figure (available in supplemental figures) but
share similar trends as the shown configurations with titanium nail material properties.
Gapped fractures generally showed larger maximum implant stresses, as well as axial and
shear interfragmentary motions compared to their reduced counterparts. Stable fracture patterns
such as reduced 31A1.1, 31A2.1, and 31A3.1 showed smaller peak stresses and interfragmentary
motions than the gapped fracture patterns.
Increasing nail diameter decreased maximum implant stress and both axial and shear
interfragmentary motion, or had minimal effects in all simulated cases. In the A3.1, A3.2, and A3.3
fracture types adding distal fixation screws generally decreased axial and shear interfragmentary
motion, but increased maximum implant stresses due to the kinematic contact constraints imposed
by the screws on the nail. These effects were especially substantial in the A32 Gap fracture type.
Nail length had variable effects on outcomes which were dependent on fracture type.
From the regression analysis, nail diameter was statistically significant in predicting
maximum implant stress for the pertrochanteric and subtrochanteric reduced fractures but not for
the intertrochanteric fractures. Conversely, nail length was only statistically significant in
67
predicting implant stresses for the intertrochanteric fractures and was not a significant predictor of
maximum stress in the pertrochanteric and subtrochanteric fractures. Screw fixation was nearly
always a significant predictor of all outcome variables.
68
Figure 3-3a: Simulation outcomes for titanium implant material configurations for reduced
fractures. Maximum screw stress results are shown in the superior fixation screw for cases that
included fixation screws.
69
Figure 3-3b: Simulation outcomes for titanium implant material configurations for gapped
fractures. Maximum screw stress results are shown in the superior fixation screw for cases that
included fixation screws. Note vertical axis scales have a larger range than in Figure 3-3a.
70
Table 3-2: Results of multiple linear regression of all simulations for maximum implant stress,
axial interfragmentary motion, and shear interfragmentary motion grouped by fracture type. Design
variables with statistically significant predictive value (p<0.05) are shaded green. Average results
± standard deviation for each outcome variable are presented in the rightmost column.
Maximum Implant Stress
Regression
Fit (r2)
Nail
Diameter
Nail
Length
Nail
Material
Screw
Fixation
Average
(MPa)
31A1.1 0.53 503 ± 183
31A2.1 0.72 782 ± 320
31A2.2 0.71 830 ± 292
31A3.1 0.92 710 ± 348
31A3.2 0.92 894 ± 477
31A3.3 0.84 1013 ± 504
31A1.1 Gap 0.51 1654 ± 56
31A2.1 Gap 0.70 1364 ± 208
31A3.2 Gap 0.96 1346 ± 655
Axial Interfragmentary Motion
Regression
Fit (r2)
Nail
Diameter
Nail
Length
Nail
Material
Screw
Fixation
Average
(mm)
31A1.1 0.97 0.3 ± 0.0
31A2.1 0.96 1.3 ± 0.4
31A2.2 0.96 1.6 ± 0.3
31A3.1 0.86 0.9 ± 0.5
31A3.2 0.81 1.0 ± 0.6
31A3.3 0.91 1.2 ± 0.5
31A1.1 Gap 0.67 1.3 ± 0.3
31A2.1 Gap 0.93 1.8 ± 0.5
31A3.2 Gap 0.66 2.8 ± 2.8
Shear Interfragmentary Motion
Regression
Fit (r2)
Nail
Diameter
Nail
Length
Nail
Material
Screw
Fixation
Average
(mm)
31A1.1 0.88 0.4 ± 0.0
31A2.1 0.96 0.8 ± 0.3
31A2.2 0.96 1.4 ± 0.3
31A3.1 0.95 1.2 ± 0.4
31A3.2 0.80 1.5 ± 0.7
31A3.3 0.91 2.2 ± 0.6
31A1.1 Gap 0.96 1.3 ± 0.2
31A2.1 Gap 0.92 1.2 ± 0.6
31A3.2 Gap 0.80 6.0 ± 5.2
71
Discussion
To our knowledge, this study provides the most comprehensive investigation of IM nail
biomechanics for proximal femur fracture to date. This study demonstrates which implant
parameters are most important in determining construct biomechanics, and how these influences
vary among nine standardized fracture types. For some fractures such as 31A11, 31A11 Gap, and
31A22 Gap, biomechanics appear predetermined and under little control by the surgical construct.
However in the other fractures, certain decisions play a critical role in determining implant stresses
and interfragmentary motion (Figures 3-3a-b and Table 3-2).
The results suggest that nail diameter is an important predictor of both axial and shear
interfragmentary motion across all fracture types. Combining this with the inverse trends between
nail diameter and interfragmentary motions, the data suggest that larger diameter nails reduce
interfragmentary motions in both the shear and axial directions regardless of fracture type.
Furthermore, nearly all data lines in Figures 3-3a-b show downward slopes with increasing nail
diameter, demonstrating that increasing nail diameter tends to reduce stress in the implant, reduce
interfragmentary motions at the fracture site, as well as reduce stress in the fixation screws, when
present.
Excessive shear motion at a fracture site is thought to disrupt healing through rupture of
nascent capillaries that attempt to bridge the fracture151,162 and is important to mitigate. Nail length
was shown to be a statistically insignificant predictor of shear motion at the fracture site in all but
the 31A1.1 and 31A3.2 fracture types. This suggests that short nails can potentially be used without
deleterious effect on shear motion at the fracture site, an important consideration especially in
clinical cases where a long nail could interfere with distal hardware such as a total knee
arthroplasty. Furthermore, the 31A1.1 fracture type in which nail length was a significant predictor
is a stable fracture type that exhibited less than 1mm average shear motion in all cases, indicating
72
that optimization for shear motion has only a small effect in the 31A1.1 fracture type. In fact,
adjusting nail length had only small effects on axial interfragmentary motion (the vertical gap
between blue and orange lines in Figure 3-3a-b) in all cases which nearly always diminished as
nail diameter increased.
Distal screw fixation is an interesting design variable because of the competing effects it
had on maximum stress and interfragmentary motion. Specifically, adding the fixation screws
increased maximum implant stress for all fracture types but reduced interfragmentary motion in
both axial and shear directions in all cases except for shear motion in subtrochanteric fracture
31A3.1. Although our models suggest that screw fixation was a significant predictor of both axial
and shear interfragmentary motion, the differences in motion with and without screw fixation were
small except for in subtrochanteric fracture types. Furthermore, clinical data from 333 patients
indicated that stable pertrochanteric and intertrochanteric fractures can be treated with short
intramedullary nails without distal locking screws without increasing mortality nor hospital length
of stay, supporting the nail length and screw fixation findings of our model.126
The importance of nail diameter should be considered also in context of the gap between
the implant and the intramedullary canal, since prior work indicates that increasing canal fill
increases fracture site stability 163. Because all bones were reamed for a 13.5mm diameter canal,
the effect of nail diameter on stabilizing the fracture could depend on the gap between implant and
endosteal canal rather than implant diameter alone. Our data suggest at least that emphasis be
placed on filling the reamed canal as much as possible and may also indicate, pending further study,
that larger nail diameter itself reduces interfragmentary motions. A previous review of clinical
studies indicates that, by unknown mechanisms, union is more likely to occur in reamed cases than
in unreamed despite disruption in endosteal vascularity.164 It is thought clinically that reaming may
create a diffuse inflammatory response, or distribute bone debris into the fracture site and augment
union. Perhaps the unexplained healing mechanism is also linked to kinematics at the fracture site
73
which are predicted, in part, by nail diameter across all fracture types. Further investigation can
isolate effects of increasing nail diameter from filling the endosteal canal to provide further surgical
guidance on canal reaming.
Although nail failure by either yield or fatigue is reported in fewer than 2-5% of clinical
cases,92,165 the above analysis provides useful information on how fracture type and implant
construct influence the maximum implant stress. Specifically in cases where a clinician suspects
the patient may be slow to heal a fracture, such as with tobacco smokers166 and diabetics,167 the
yield strength and fatigue life of the implant can become factors in mitigating fixation failure under
the cyclic loads of gait. The location of maximum vonMises stress was at the superior aspect of the
junction between the cephalomedullary screw and the nail, consistent with previous
studies.55,116,117,122 The data available in Figures 3-3a-b and Table 3-2 can be used to optimize the
fixation construct to protect against high stress in the implant. Generally, increasing nail diameter
and omitting distal fixation screws leads to lower maximum stress in the implant. Observation of
simulation kinematics as load is increased to 2000N shows that addition of fixation screws limits
rotation and translation of the nail within the endosteal canal. The kinematic contact constraints
applied by the screws onto the nail result in a stiffer construct which concentrates stress on the
proximal nail more than when the fixation screws are omitted. The problem of IM nail fatigue is
quite complex. Ideally, a fracture construct that minimizes stress in the implant while optimizing
interfragmentary strain will result in rapid union of the fracture and minimal cycles on the implant.
Optimizing this combination for each patient, is the goal of fracture surgery and should be the
guiding principle.
Severely unstable 31A3.2 Gap fracture shape cases are not treated clinically without distal
fixation screws. In 31A3.2 simulations that omitted fixation screws, the translational boundary
conditions at the point of applied load artificially constrained motion of the proximal fragment.
Specifically, without the proximal boundary condition, rotation of the proximal fragment about the
74
long axis of the femur would only have been constrained by friction between the nail and the
endosteal canal, prohibiting simulation convergence. Therefore, the results reported for
interfragmentary motion in 31A3.2 Gap fracture models that do not include distal fixation screws
are underestimates. However, simulations confirm that addition of fixation screws in such cases
drastically reduces all interfragmentary motion, necessitating distal fixation with the 31A3.2 Gap
fracture type.
The computational nature of finite element studies requires validation. In direct comparison
our simulations were about 15% stiffer than literature-reported values based on mechanical testing,
as mentioned above in the methods section. Because of this we include additional comparisons of
vonMises stress data to previously validated studies. Eberle et al. report maximum nail stresses of
749MPa for the 31-A3.1 fracture with simplified muscle loading and 2367MPa for the 31-A3.3
fracture.122 The 11mm diameter, long, titanium alloy Gamma3 implant used with distal fixation
screws in the Eberle study as well as the applied muscle loads are not identical to those in our study,
however, our reported maximum stresses (1027MPa for the 31-A3.1 fracture and 1334MPa for the
31-A3.3 fracture) fall between their reported values.
Our study has several limitations. Although the AO/OTA fracture classification scheme
has utility in identifying similarities among common femur fracture patterns, the scheme itself is a
simplification of the complex 3D geometry of femur fractures. Our results indicate dependence of
outcomes on fracture type, and differences between simplified and complex fracture geometry
should be investigated in future work. However, we suggest that while the exact magnitudes
reported for our outcome measures would likely change with more accurate representations of
fracture type, the relative effects of design variable changes would likely persist. Similarly, the use
of a single femur geometry with a homogeneous material distribution for both cortical and
trabecular bone is a simplification that encourages analysis of relative effects rather than absolute
75
effects. Properly accounting for bone geometry and material heterogeneity may require use of
emerging statistical shape models of the femur.
Common to most other finite element modeling studies, our model used a single static load
representing heel strike in the gait cycle under full weightbearing. Clinically, patients are
commonly prescribed limited weightbearing following surgery. An additional limitation is that all
cases, regardless of fracture type and implant diameter, were reamed to a 13.5mm diameter canal
that was perfectly concentric with the implanted nail. This resulted in a perfect ream such that the
IM nails always began each simulation without distal endosteal contact. However, the friction and
contact formulations included in the model allowed for endosteal gap closure and normal and shear
reaction forces at sites of contact. Furthermore, our simulations assess fracture stability at ‘time
zero’ where no healing has occurred, thus modeling a worst-case scenario for biomechanics.
Chapter 4
Return to Function After Intramedullary Nail Proximal Femur Fracture
Fixation: Effects of Fixation Construct and Weightbearing on Shear Motion
at the Fracture Site
Abstract
Successful union following fracture fixation surgery is partially dependent on
biomechanics at the fracture site. The biomechanics are influenced by the stability of the fracture
fixation construct, specifically in how it resists shear motion, as well as the loads that pass across
the construct as the patient returns to functional activities. Orthopaedic surgeons attempt to
optimize these biomechanics via implant selection as well as the ability to recommend specific
loadbearing limits for the patient postoperatively. In this study we utilize finite element simulations
of 3 gapped or comminuted proximal femur fractures fixed by intramedullary nailing under gait
loads to explore the effects of construct selection on shear motion at the fracture site at multiple
levels of loadbearing. Construct variables include nail diameter, nail length, and presence of distal
fixation screws. We show the loadbearing percentage that maintains shear motions at the fracture
site below a literature-reported threshold for each construct. Specifically, the data show that full
weightbearing achieves the shear motion threshold for short nails with distal fixation screws for all
tested nail diameters in a pertrochanteric and an intertrochanteric fracture type but not for a
subtrochanteric fracture. Increasing nail diameter increases the safe loadbearing limit for nearly all
tested configurations, while removing distal fixation screws has mixed results depending on if the
nail is short or long. Clinical Significance: Biomechanical outcomes from this computational study
can build intuition about how construct selection and loadbearing prescriptions influence the
amount of shear motion at a proximal femur fracture fixed by an intramedullary nail.
77
Introduction
The primary goal of surgical fixation of fractures is to stabilize the bony fragments to
promote bone healing. However, surgeons have to make many choices regarding the fixation
construct from implant type to implant fixation which will influence biomechanics at the fracture
site. Furthermore, surgeons commonly recommend limited weightbearing following fracture
fixation to promote healing and reduce the risk of construct failure. To this end, it is important to
understand the relative stability of various fracture fixation constructs and how surgical design
variables combined with patient weightbearing levels influence this stability.
Secondary fracture repair is characterized by bridging of a fracture gap by progressively
stiffening healing tissues. While these tissues form, they must endure the local mechanical and
chemical environment. Fracture repair tissues are thought to have a strain tolerance beyond which
their vascular supply is too disrupted to promote normal biological healing responses.9 Gapped
fractures predispose the bone fragments to excessive motion which can delay the healing process.
An important metric has been developed that shear motion of 1.5mm significantly delayed healing
in an ovine diaphyseal tibia fracture model.26
This unsafe limit for shear motion can be combined with recent advances in computational
modeling technology to understand the relative safety of combinations of fixation construct
variables and prescribed weightbearing in terms of shear motion at the fracture site. Most proximal
femur fractures, for example, are fixed with an antegrade intramedullary (IM) nail. These implants
can vary in shape, length, diameter, and material stiffness, and also offer multiple screw fixation
strategies. Using a previously developed computational finite element model of IM nailing in
proximal femur fractures (Chapter 3), we compare a wide array of combinations of common IM
nail design variables under simulated gait loading to the preestablished shear motion limit to assess
the relative weightbearing safety limit for each construct. The findings of this study can be used to
78
build clinical intuition for how fracture fixation construct design variables influence shear motion
at gapped fracture sites, advance implant design, and customize weightbearing strategies based on
fixation features.
Materials and Methods
The 48 titanium nail simulations for gapped fracture types A11, A21, and A32 from
Chapter 3 are used in this section. However, rather than analyzing the results at the point of
maximum applied load, these data focus on the path of deformation as the applied load is ramped
up. Specifically, data were collected at 200N intervals between 0-2,000N of applied load. Because
of convergence considerations during the numerical solver each load level is measured within
±50N. The average shear interfragmentary motion of four pairs of points evenly spaced around the
fracture site was then calculated for each load level. The resulting shear motions were compared to
the literature-established 1.5mm shear motion limit and the maximum weightbearing percentile that
maintains shear interfragmentary motion below this limit is reported.
Experimental groups are defined by configuration name: short nail (S) vs. long nail (L) and
with screws (+) vs. without (-). It is worth repeating from Chapter 3 that the gap sizes were designed
such that they do not close under the maximum applied load of 2,000N. Specifically, a gap size of
5mm was used for the A11 and A21 fracture types while the A32 type had a 10mm gap (Figure 4-
2). The applied load and muscle forces were designed to represent the maximum load at the hip
during gait for an 85.7kg individual in the 100% load level.
79
Figure 4-1: IM nails included in this study ranged from 10mm in diameter (left) up to 13mm and
nail length was modeled as either short (26.9cm) or long (43.3cm)
80
Results
The S+ configuration consistently limited mean shear motion to less than 1.5mm at the
same or higher load level as other configurations for all fracture types and nail diameters (Table 4-
1). For fracture types A11 Gap and A21 Gap, the S+ group was within the shear motion threshold
at full weightbearing for all nail diameters. The L- configuration appears similar to the S+
Figure 4-2: Gapped fractures modeled for analysis at intermediate load levels. Of note is the
difference in fracture gap size between the subtrohanteric 3.2 type (10mm) as opposed to the others
(5mm). Black circles represent point pairs at which interfragmentary motions are determined.
81
configuration, with only small differences in weightbearing limit in the A11 and A32 fracture types.
Additionally, the L+ configuration shares similar trends with the S- configuration.
Implant diameter has a positive correlation with maximum loadbearing limit across all
configurations and fracture types. However, it is less noticeable in the A32 gapped fracture, in
which the safe loadbearing limit is often below the 25% cutoff. The A32 gapped fracture is unstable,
in general, with only 4 of the 16 design constructs achieving a loadbearing limit that maintains
mean shear motion below 1.5mm. Neither nail length nor screw fixation show independent trends
with maintaining below 1.5mm shear motion across fracture types.
Table 4-1: Loadbearing percentiles (shaded) that did not induce greater than 1.5mm of average shear
interfragmentary motion. Data are separated by fracture type, nail length, nail diameter, and presence
or absence of distal screw fixation. Blank space indicates that the shear motion cutoff is exceeded
for a configuration at the corresponding loadbearing level.
A11 Gap Fracture
Load
Level
Short Nail, No
Screws (S-)
Short Nail, Screws
(S+)
Long Nail, No
Screws (L-)
Long Nail, Screws
(L+)
100%
75%
50%
25%
Diameter
(mm)
10 11 12 13 10 11 12 13 10 11 12 13 10 11 12 13
A21 Gap Fracture
100%
75%
50%
25%
Diameter
(mm)
10 11 12 13 10 11 12 13 10 11 12 13 10 11 12 13
A32 Gap Fracture
100%
75%
50%
25%
Diameter
(mm)
10 11 12 13 10 11 12 13 10 11 12 13 10 11 12 13
82
Discussion
To our knowledge this is the first finite element study evaluating IM nail shear motion in
the context of partial weightbearing. This study demonstrates the load level at which mean shear
motion at the fracture site remains below a cutoff which was found to impact healing. Implant
diameter has a clear relationship with the safe load level for all tested fracture types, although
fracture A32 exceeds the shear motion threshold in most cases. Nail length and screw fixation
appear to have a more complex relationship with the safe load level and seem to be dependent on
each other for their effects.
Specifically, the S+ and L- are the most shear motion-stabilized constructs for higher nail
diameters at 100% load bearing in all 3 fracture types. In the A11 and A21 fracture types. It might
be expected that constructs matched in nail length or in screw fixation would behave similarly to
each other but actually the contrary is true in this data. The S+ scenario resists shear motion well
across all nail diameters at high loads because the nail is well-constrained by the fixation screws,
the bending moment on the nail is low due to its short length, and the resulting construct is stiff.
Similarly, the L- scenario enables the long nail to toggle within the medullary canal, thus enabling
axial displacement which starts to reduce the fracture gap and, therefore, the shear interfragmentary
motion. However, toggle in the S- case due to the lack of screw fixation results in high
displacements of the proximal fragment which are not so well compensated for by axial
displacement at the fracture site. The L+ case does not enable nail toggle but the moment arm is
long due to the nail length and thus bending deformation of the nail results in high shear motions
at the fracture site.
In fact, the complex interplay between nail length and screw fixation could explain why,
in Chapter 3, nail length was found to be an insignificant predictor of shear motion at the fracture
site. Perhaps the effect of nail length on interfragmentary shear motion is also dependent on screw
83
fixation and should not be considered separately. Previous clinical data indicate minimal
complications when stable femoral and tibial fractures were fixed using intramedullary nails
without interlocking, although the analysis did not consider nail length.168 Another study indicated
that short nails without distal interlocking demonstrated no increase in mortality nor hospital length
of stay when used to fix stable pertrochanteric and intertrochanteric fractures.126 It is likely that the
inherent stability of the fractures evaluated played a role in these outcomes and that similarly treated
unstable fracture types could behave differently.
In the literature, most findings indicate that shear motion delays or prohibits
healing.26,27,151,169 Although some studies have found that shear motion promotes healing,170,171
arguments have been made to suggest improperly controlled and measured results among studies
that found healing enhanced healing.172 A recent clinical study found no difference in outcomes of
healing and time to union in nail-fixed simple tibial shaft fractures but the non-gapped nature of
these fractures is a likely source of stability.173 Controversy aside, it is generally agreed that some
shear motion may be tolerated (or even beneficial) but also that there is an upper limit of shear
motion beyond which fracture healing is negatively impacted.
A limitation of this study is that the applied muscle loads are linearly scaled with the
applied reaction force at the hip. Due to the dynamic nature of these loads during gait it is difficult
to obtain an accurate representation of both the hip joint contact force and the applied muscle loads
under limited weightbearing. However, the applied muscle forces contribute to the overall hip joint
reaction force and it may be reasonable to assume they scale linearly together under limited
weightbearing.
84
Chapter 5
Examples of Precision Medicine in Orthopaedics
Precision medicine is a growing focus within the medical community which encourages a
detailed examination of individual clinical case details to inform treatment decisions, rather than
the formerly popular one-size-fits-all approach. Advances in technology have facilitated deeper
understanding of disease, improvements in diagnostic capabilities, and knowledge to potentially
treat two patients with the same condition differently, because their cases may be distinct. As
precision medicine continues to gain traction it creates a need to better understand and optimize
fracture fixation. Surgical plans can be optimized based on modeling results with a goal of reducing
failure and improving patient outcomes.
Precision Finite Element Modeling for Intramedullary Nail Femur Fracture Fixation
As mentioned previously, many assumptions and simplifications are employed to facilitate
finite element fracture model development and computation on a reasonable timescale. However,
recent advances in computational power and resource allocation have enabled multiple levels of
specificity in mechanical simulations, including bone geometry, fracture shape, and material
properties.95,96 However, there is not yet a process for taking patient imaging data and generating
model outputs on a clinically relevant timescale. Simplified computational models can be created
with relative ease, whereas precision modeling requires significant time and computational
resource allocation. There are various degrees of precision modeling that may or may not affect
clinically relevant biomechanical measures. Therefore, the sensitivity of model outcomes to
precision modeling was evaluated in a case study of one model.
85
Institutional Review Board approval was obtained to collect deidentified x-ray images from
patients with proximal femur fractures. A case was selected which had only 1 proximal femur
fracture line producing 2 bony fragments, similar to a fracture modeled in Chapter 3. Specifically,
the fracture type in the selected case was classified as an A21 fracture according to the AO/OTA
classification scheme.88 The generic bone model used for model development in Chapters 3 and 4
was again used but the model was initially unfractured. The 2D patient x-radiograph was
superimposed over the 3D generic bone model such that the bony cortices aligned (Figure 5-1).
The fracture line was then identified on the distal fragment from the x-ray and was used to create a
fracture in the 3D bone model.
86
Although the patient’s fracture was not perfectly recapitulated with a straight cut,
reconstruction of the same patient’s CT scan into a solid model and comparison to the cut generic
bone demonstrates good agreement in shape (Figure 5-2).
Figure 5-1: Patient fracture x-radiograph (left) superimposed over generic 3D bone model (right).
The fracture line at the superior aspect of the distal fragment is highlighted with a spline which was
used to create the fracture.
87
The patient-specific fracture model was then reamed to accommodate an IM nail and a
long, 13mm diameter titanium nail was implanted with distal fixation screws present. The model
was then run under the same simulation conditions as the previous generic models. The results from
the patient-specific fracture model were then compared to the A22 fracture simulation from the
generic modeling library in Chapter 3 with the same construct variables. Although interfragmentary
motions and global stiffnesses were very similar between the two simulations, the patient specific
model showed a 57.9% decrease in maximum nail stress over the generic model (228.2MPa vs
541.9MPa).
The drastic difference in maximum stress between the patient-specific fracture model and
the generic raises some interesting points. First, generic modeling is perhaps limited in scope to
discuss relative trends in design variables rather than specific magnitudes. However, there are still
multiple levels of patient specificity lacking from the patient specific fracture model. Specifically,
full bone geometry, customized applied loads that scale with patient body weight and activity level,
Figure 5-2: Solid model reconstructed from 3D CT data (left) shown as a gold standard alongside
patient-specific fracture from 2D x-ray alignment (right) demonstrating good agreement in fracture
shape. Reconstruction was graciously completed by medical student Marcus Erdman.
88
and nonhomogeneous bone material properties that match the patient’s bone mineral density are
examples of how model fidelity can be further enhanced. Second, perhaps the AO/OTA fracture
classification scheme could benefit from further parsing the fracture types. There could be non-
obvious differences in the shape of the patient specific fracture that drive fixation construct
biomechanics, along which a functional classification could be warranted. Further development of
patient specific models, including varying levels of patient specificity, could help identify which
levels of specificity are the most important to model.
Another approach for increasing the patient specificity of a computational model is called
statistical shape modeling.131 In this approach a parametric model of the femur is informed by a
large set of virtually reconstructed patient images such that individual patient demographics can be
used to predict specific femur shapes. A detailed description of this approach as well as a list of
potential applications is described by Heimann and Meinzer in a 2009 review article.174
Unfortunately, statistical shape models estimating fractured femur shapes are not currently
available but may be developed from intact femur statistical shape models.
Custom 3D Printed Implant Design and Modeling for Pediatric Orthopaedic Oncology
The following individuals contributed to work in this section: Scott Tucker, Maryam
Tilton, Ali Elakkari, Alex Preniczky, Guha Manogharan, Edward Fox, Gregory Lewis.
Due to the limitations of available implant hardware, current standards of care for pediatric
orthopaedic oncology reconstruction following bone resection often remove nearby articular joints.
Additive manufacturing, also known as 3D printing, enables rapid, cost-effective production of
custom implants that can be designed with porous features unattainable with traditional
manufacturing methods. Furthermore, personalized 3D printed constructs can improve bone-
implant fixation and be articular-sparing to preserve native, healthy anatomy. In this retrospective
89
IRB-approved study, we designed pediatric patient-specific implants based on medical imaging
data for proximal humerus, proximal tibia, midshaft femur, and distal femur osteosarcomas, and
fabricated the implants via electron beam melting additive manufacturing with Ti-6Al-4V powder.
Pediatric oncology patients were identified from our institution. Cases involving proximal
humerus, proximal tibia, midshaft femur, and distal femur were selected and MRI data were
obtained. Bones were reconstructed and rendered as solid models (Mimics software, Materialise).
Virtual resection surgery was performed (SolidWorks software, Dassault Systemes) with guidance
from an orthopaedic musculoskeletal oncology surgeon (Figure 5-3). Implants were designed to fill
post-surgical defects, and lattice structures were designed for each implant (ElementPro software,
nTopology). Specific attention was given to fixation points between implant and bone during the
design for 3D printing workflow using biocompatible titanium alloy. Finite element analysis
(Abaqus software, Dassault Systemes) was used to predict locations and magnitudes of maximum
von mises and principal stress. Loads and boundary conditions were specific to anatomic location
and based on well-established literature.102
90
Novel articular and growth plate-sparing implant designs were designed as alternatives to
joint- sacrificing endoprosthetic reconstruction following tumor resection. The proximal tibia and
distal femur implants shown (Figure 5-4) combine a non-stochastic lattice structure to fill the region
of resected bone, a stem for stabilization and fixation, and a boss for screw fixation to the spared
articular surface. The resulting designs have been 3D printed out of Ti-6Al-4V alloy for future
mechanical testing. Finite element analysis of the proximal tibia implant and lattice structure under
the maximum anticipated loads during gait demonstrated a maximum implant stress of 11.6MPa,
compared to literature-reported fatigue strength values of 200-250MPa at 107 cycles for electron
beam melted Ti-6Al-4V alloy.175
Figure 5-3: Tumor volume visualization for the proximal tibia case with segmented medical image
(left) and resulting 3D reconstruction (right). Tumor is highlighted by the red circle.
91
Thus, four unique custom implants have been designed for different applications in joint-
sparing tumor removal surgery in different anatomical regions. Custom implant design, taking
advantage of unique capabilities of 3D printing, can match defect geometry, improve implant
fixation, and provide porous scaffolds that likely promote bone integration while supporting
physiological loads and restoring function (Figure 5-5). Furthermore, extrapolation of the finite
element analysis stress data can be used to predict the fatigue life of the implant, although this
would likely be affected by any osteointegration that may occur as healing progresses.
Figure 5-4: Workflow for custom implant design for distal femur and proximal tibia osteosarcomas.
Patient imaging data (A, G) is reconstructed into a solid model with virtual tumor removal (B, H)
and an implant is designed to fill the bone defect (C, I). Implant fit is then assessed (D, J). Finite
element simulation stress results of implant mechanical loading from low (blue) to high (red) under
compressive loads (E, K). Additively manufactured Ti-6Al-4V implants (F, L).
92
With a well-established workflow, custom 3D printed implants can be designed and
manufactured on a shorter timescale than available through industry, can preserve healthy anatomy,
and can potentially promote osteointegration with porous architecture.
Figure 5-5: Detailed views of the proximal tibia articular-sparing implant design demonstrating
porous architecture and cortical contour.
Chapter 6
Conclusions, Discussion, and Future Work
The FE modeling work presented in this dissertation represents, to my knowledge, the most
comprehensive study of IM femur nail design variables to-date. The results presented have direct
clinical value in both educational and presurgical planning domains. Allegorical clinical data can
be replaced and surpassed by data from computational biomechanics studies such as those in this
work. Specifically, this work suggests in Chapter 3 that larger IM nail diameters reduce the implant
maximum stress and provide more stability over smaller nail diameters. Also shown is that nail
length is a poor independent predictor of shear motion at the fracture site for nearly all simulated
fracture shapes. Furthermore, in the results section of Chapter 4, I demonstrate that some design
variables have synergistic combination effects (the long nail without screws behaves similarly to
the short nail with screws in terms of shear stability). The value of these modeling efforts may
extend beyond surgery as the suggested weightbearing outcomes in Chapter 4 provide rationale for
prescribing postsurgical activity levels based on individual fracture and fixation construct
conditions. I also provide a brief example in Chapter 5 which demonstrates that introducing patient
specificity to generic modeling strategies does impact the results on a clinically relevant scale,
creating a call to action for development of model personalization strategies. Additionally, the near
60% reduction in maximum implant stress between corresponding generic fracture and patient
specific fracture models indicates that perhaps a more robust fracture classification scheme based
on functional biomechanics is needed. In addition to directly answering the research questions
asked in each chapter, this work in sum is aimed at setting a foundation upon which clinical tools
can be developed for both surgical resident education and presurgical planning purposes.
94
One benefit of computational studies is the ability to repeat the simulations under near-
identical conditions and with an alteration in only one design feature. A major advantage of this
capability is that computational models lend themselves well to parametric analyses and fully
exploring a design space. However, this feature of computational models complicates statistical
analysis because the design space is perfectly sampled and there is no variability in simulation
results. Natural variability in data enables most statistical analyses because of the normal
distribution. However, the natural variability in experimental data and the low sample sizes tested
in most mechanical testing experiments makes it difficult to interpret a validation study which
compares noisy experimental data with clean computational data. One often gets the feeling of
chasing one’s own tail when considering the cumulative errors in both mechanical testing and
computational experiments. In fact, computational model repeatability, validation, and verification
were the foci of a 90-minute panel discussion at the 2019 annual meeting of the Orthopaedic
Research Society in Austin, Texas. General consensus was reached that the broad research
community struggles with committing appropriate amounts of time and resources to repeatability,
validation, and verification in our ‘publish or perish’ research environment. However, rigor must
be taken to ensure that the results being reported are the most accurate results available as well as
provide some assessment of systematic errors and how they may influence results.
Initial FE model development for this work presented multiple challenges. Specifically,
due to the complex contact modeled between the fracture fragments, the endosteal canal, the nail,
and the screws, numerical convergence was not often reached. The limits for convergence used in
this work are the ABAQUS default values (5N for force and 10Nm for moment) and were
maintained to ensure accuracy of the Newton Raphson approximation method. However, an
alternative tool called ‘Contact Controls’ in ABAQUS was activated with a coefficient of 0.00001.
Contact controls enables force dampening during conditions when a contact condition opens
(bodies separate from each other) or closes (bodies come into contact with each other). This is
95
useful because the stiffness response of the system to applied load often changes when a contact
condition is altered. Although the analytical load vs displacement curve for that system would have
a sharp vertex when a contact condition changes, the damping enforced by contact controls will
smooth that vertex and help the solver locate an approximation that falls within the aforementioned
equilibrium tolerances (Figure 6-1). Errors introduced by this method can be measured using
energy methods. Specifically, one can compare the dissipated energy (the energy in the system lost
to nonconservative forces such as damping) to the total energy in the system. So long as the energy
dissipated was sufficiently small relative to the total energy (< 2%), errors introduced by damping
were deemed acceptable in this work.
Extensions of this work should expand the simulation library to include more implant
designs such as a retrograde nail and more fracture types such as distal and diaphyseal fractures.
An interesting point of discussion raised in Chapter 3 is whether the stabilizing and stress-reducing
Figure 6-1: Example system load-displacement response shown (green). The piecewise nature of
the green curve demonstrates a change in system stiffness at each of the two vertices, such as what
could happen if bodies came into contact and subsequently separated. The curve drawn in red
indicates the behavior of the simulation when contact controls is engaged. Dampening, which helps
to smooth the load-displacement relationship, introduces error in the approximated solution but
aids numerical convergence.
96
effects of large nail diameter are due to just having a larger diameter or if they are related to the
gap between the implant and the endosteal canal. A study evaluating the same diameter nail in
multiple different reaming scenarios would help identify which feature is driving the trends. In
example, outcome comparisons between simulations using a nail-matched, nail diameter +0.5mm,
and nail diameter +1mm reaming size would indicate if the reaming gap is having an effect. Perhaps
a more exciting extension of this work would focus on introducing varying levels of patient
specificity to the models. A sensitivity analysis on multiple independent and combined patient
specific modeling features could help focus future modeling efforts on the most impactful features.
Additionally, the simulations in this work tend to represent worst-case scenarios because no healing
tissues are accounted for. While this is still relevant as it models the patient immediately
postoperatively, over time we know that fractures undergoing secondary healing experience a
gradual hardening and calcification of tissue within the fracture gap which would provide additional
stability to the fracture and implant. Implant fatigue life predictions could be accurately made if
simulations were developed which could imitate the effects of fracture callus formation and
stiffening.
Appendix A
An Unsupervised Machine Learning Method for Discovering Patient Clusters
Based on Genetic Signatures
Text in this chapter is published in the Journal of Biomedical Informatics176
Authors: Christian Lopez, Scott Tucker, Tarik Salameh, Conrad Tucker
Abstract
Introduction: Many chronic disorders have genomic etiology, disease progression, clinical presentation, and response to
treatment that vary on a patient-to-patient basis. Such variability creates a need to identify characteristics within patient
populations that have clinically relevant predictive value in order to advance personalized medicine. Unsupervised machine
learning methods are suitable to address this type of problem, in which no a priori class label information is available to
guide this search. However, it is challenging for existing methods to identify cluster memberships that are not just a result
of natural sampling variation. Moreover, most of the current methods require researchers to provide specific input
parameters a priori.
Method: This work presents an unsupervised machine learning method to cluster patients based on their genomic makeup
without providing input parameters a priori. The method implements internal validity metrics to algorithmically identify
the number of clusters, as well as statistical analyses to test for the significance of the results. Furthermore, the method
takes advantage of the high degree of linkage disequilibrium between single nucleotide polymorphisms. Finally, a gene
pathway analysis is performed to identify potential relationships between the clusters in the context of known biological
knowledge.
Datasets and Results: The method is tested with a cluster validation and a genomic dataset previously used in the literature.
Benchmark results indicate that the proposed method provides the greatest performance out of the methods tested.
Furthermore, the method is implemented on a sample genome-wide study dataset of 191 multiple sclerosis patients. The
results indicate that the proposed method was able to identify genetically distinct patient clusters without the need to select
parameters a priori. Additionally, variants identified as significantly different between clusters are shown to be enriched
for protein-protein interactions, especially in immune processes and cell adhesion pathways, via Gene Ontology term
analysis.
Conclusion: Once links are drawn between clusters and clinically relevant outcomes, Immunochip data can be used to
classify high-risk and newly diagnosed chronic disease patients into known clusters for predictive value. Further
investigation can extend beyond pathway analysis to evaluate these clusters for clinical significance of genetically related
characteristics such as age of onset, disease course, heritability, and response to treatment.
1. Introduction
With advancements in genome-wide association study (GWAS) techniques and the advent of low cost
genotyping arrays, researchers have developed a significant interest in applying Machine Learning (ML) methods
98
to mine knowledge from patients’ genomic makeup 177,178. This knowledge has allowed researchers to improve
gene annotation and discover relationships between genes and certain biological phenomena 179,180.
The fields of personalized and stratified medicine benefits greatly from ML. For example, many cases in the
field of pharmacogenetics have identified genetic variants with clinically actionable impacts on drug response and
metabolism181,182. Moreover, many chronic disorders (e.g., asthma, diabetes, Crohn’s disease) have genomic
etiology, clinical presentation, and response to treatment that vary on a patient-to-patient basis. Such variability
reveals a need to identify characteristics within patient populations that have clinically relevant insights. For
example, Multiple Sclerosis (MS) is a chronic inflammatory disorder in which progressive autoimmune
demyelination and neuron loss occur in the central nervous system. MS varies from patient-to-patient in genomic
etiology, disease progression, clinical presentation, and response to treatment. Hence, MS patients, like other
chronic autoimmune patients, could benefit from ML methods that advance personalized medicine.
Machine learning methods are commonly classified into supervised and unsupervised methods. Supervised
methods, such as Support Vector Machines 183 and Random Forests 184,185, have been extensively used in the field
of bioinformatics. These methods classify new objects to a determinate set of discrete class labels while minimizing
an empirical loss function (e.g., mean square error). However, supervised methods require the use of a training set
that contains a priori information of several objects’ class labels. In contrast, unsupervised methods do not require
a training set that contains a priori information of objects’ class labels as input. Unsupervised methods are able to
detect potentially interesting and new cluster structures in a dataset. Moreover, they can be implemented when
class label data is unavailable. Hence, if the objective of a study is to discover the class labels that best describe a
set of data, unsupervised machine learning should be implemented in place of supervised methods 178. However, it
is challenging for existing unsupervised ML methods to identify object memberships that are due to the underlying
cluster structures in the dataset rather than the results of natural sampling variation 186. Moreover, most current
methods require researchers to provide certain input parameters a priori (e.g., number of clusters in the dataset),
which can limit their applicability.
In light of the limitations of existing methods and the need to advance personalized medicine, an unsupervised
machine learning method to cluster patients based on their genomic similarity is proposed. The method integrates
statistical analysis that accounts for family-wise-error rate, allowing the method to identify clusters resulting from
the underlying structure of the data and not just due to random chance. Moreover, the method takes advantage of
the high degree of linkage disequilibrium between Single Nucleotide Polymorphisms (SNP) by pruning correlated
nearby SNPs, which helps reduce redundant variants in the dataset. Finally, a gene pathway analysis shows the
potential relationships between the clusters in the context of known biological knowledge. The proposed method
is capable of clustering patients based on their genomic similarity without a priori information. Moreover, it is
capable of identifying the significant variants (i.e., SNPs) between patient sub-groups within a cohort with a
common disorder. Successfully identifying distinct genetic subtypes of patients within genomic datasets
demonstrates the potential of this method to advance personalized medicine of complex diseases with heritable
components, especially autoimmune disorders which have many shared susceptibility loci 187.
99
2. Literature review
In the last decade, the field of bioinformatics has seen a significant number of publications implementing
unsupervised machine learning methods, such as clustering algorithms [12-14]. Clustering algorithms partition
data objects (e.g., genes, patients) into groups (i.e., clusters), with the objective of exploring the underlying
structure on a dataset 191. In the medical field, these algorithms have been implemented to identify sets of co-
expressed genes 192, compare patients’ prognostic performance 193, cluster patients based on their medical records
194, and identify subgroups of patients based on their symptoms and other variables 195.
In previous work, genomic stratification of patients (i.e., stratified medicine) has been able to match specific
therapy recommendations to genetic subpopulations by predicting therapeutic response 181,182. However, most of
these studies implemented class label data (i.e., response to treatment) to cluster patients. In clinical datasets, class
label information is not widely available for convenient patient clustering. Unsupervised machine learning methods
can be used in such cases to identify clusters within the dataset. Further investigation of genetic subgroups within
a cohort of patients can offer a better clinical prediction of age of onset, disease course, heritability, and response
to therapy, leading to improved outcomes 196
2.1. Hierarchical clustering algorithms.
Agglomerative hierarchical clustering algorithms are one of the most frequently used algorithms in the
biomedical field 197,198. Researchers have found that hierarchical clustering algorithms tend to perform better than
other algorithms (e.g., k-means, partitioning around Medoids, Markov clustering) when tested in multiple
biomedical datasets 199. The objective of any agglomerative hierarchical clustering algorithm is to cluster a set of
n objects (e.g., patients, genes) based on an n x n similarity matrix. These clustering algorithms have grown in
popularity due to their capability to simultaneously discover several layers of clustering structure, and visualize
these layers via tree diagrams (i.e., dendrogram) 186. Even though these algorithms allow for easy visualization,
they still require preselecting a similarity height cut-off value in order to identify the final number of clusters. In
other words, it still requires researchers to know a priori the number of cluster in the dataset.
Agglomerative hierarchical clustering algorithms can be implemented with different linkage methods. For
example, Ahmad et al. (2016) 193 implemented the Ward’s linkage method to compare patients’ prognostics
performance; while Hamid et al. (2010) 195 implemented the Complete linkage method to identify unknown sub-
group of patients. Unfortunately, depending on the underlying structure of the data, different clustering results can
be obtained by implementing different linkage methods. Ultsch and Lötsch (2017) 200 demonstrated that neither
the Single nor Ward’s linkage methods provided similar clustering results when tested with the Fundamental
Clustering Problem Suite (FCPS) datasets 201. Their results reveal that these linkage methods were able to correctly
cluster all the objects in only a subset of the FCPS datasets. Similarly, Clifford et al. (2011) 202 discovered that
while testing multiple simulated GWAS datasets, the linkage methods of Median and Centroid were the only ones
to consistently be outperformed by the Single, Complete, Average, Ward’s, and McQuitty methods. In light of
these, Ultsch and Lötsch (2017) 200 proposed the use of emergent self-organizing map to visualize clustering of
high-dimensional biomedical data into two-dimensional space. Even though, their method allowed for better
100
visualization, it still required preselecting the number of clusters as well as other parameters to perform correctly
(e.g., toroid grid size) 200.
2.2. Parameter selection in clustering algorithms.
In order to avoid preselecting input parameters a priori (e.g., the number of clusters),
researchers have implemented cluster validation metrics. For example, Clifford et al. (2011) 202 proposed a method
that aimed to capture the clustering outcome of multiple combinations of linkage method and similarity metric
based on the Silhouette index 203. The Silhouette index was used to rank the results of the clustering combinations,
and select the best cluster set (i.e., cluster set with largest average Silhouette index). Similarly, Pagnuco et al.
(2017) 192 presented a method that implemented several linkage methods and implemented modified versions of
the Silhouette and Dunn indices 204 to select the final clustering results. Both the Silhouette and Dunn indices
served as internal cluster validation metrics (i.e., no external information needed) to guide the selection of the final
cluster set. However, the Silhouette index has been shown to have a stronger correlation with external cluster
validation metrics, such as the Rand Index, than the Dun index [28,30].
The methods of Clifford et al. (2011) and Pagnuco et al. (2017) did not require selecting the number of clusters
a priori due to the internal cluster validation metrics implemented. These metrics allow for algorithmic selection
of the number of clusters. Nonetheless, the computational complexity of testing all potential clusters increases
linearly with the number of objects in the dataset. Other studies have implemented model-based clustering methods
to overcome these limitations. For example, Sakellariou et al. (2012) 205 implemented an Affinity Propagation 206
algorithm to identify relevant genes in microarray datasets. Shen et al. (2009) 207 implemented an Expectation-
Maximization algorithm 208 to cluster genes based on an integration of multiple genomic profiling datasets.
However, models based methods make underlying assumptions that might not be applicable in certain datasets 209.
Recently, Khakabimamaghani and Ester (2016) 210 presented a Bayesian biclustering method to identify clusters
of patients. They benchmarked their method against the multiplicative Non-negative Matrix Factorization (NMF)
algorithm proposed by Lee and Seung (2001) 211. Their results revealed that their Bayesian biclustering method
was more effective in patient stratification than the NMF. While this Bayesian biclustering method did not require
selecting the number of clusters a priori, it did require selecting parameters for prior probability distributions. The
capability of biclustering algorithms to discover related gene sets under different experimental conditions have
made them popular within the bioinformatics community 212. One of the first works in this area was presented by
Cheng and Church (2000) 213. They proposed an iterative greedy search biclustering algorithm to cluster gene
expression data. Even though their method did not require selecting the number of clusters a priori, it did require
the selection of hyperparameters (e.g., maximum acceptable error).
2.3. Statistical significance of clustering results.
Even though the methods of Clifford et al. (2011) and Pagnuco et al. (2017) aimed to find the optimal clustering
outcome from multiple algorithms, which resembled the consensus clustering approach (i.e., approach in which a
solution is identified by validating multiple outcomes) 214, their methods did not account for possible clustering
101
memberships arising due to random variation. Whether identified clusters memberships are due to underlying
cluster structures in the data or are just a result of the natural sampling variation, is a critical and challenging
question that needs to be addressed when clustering high-dimensional data 186. To address this question, Suzuki
and Shimodaira (2013) 215 presented the pvclust R package, which calculates probability values for each cluster
using nonparametric bootstrap resampling techniques. Even though pvclust allows for parallelized computing, it
requires significant time (i.e., 480 mins) when implemented in genomic datasets. This is due to the large number
of resampling iterations (i.e., 10,000) required to reduce the error rate 215. In contrast, Ahmad et al. (2016) 193
applied a non-parametric analysis of variance (ANOVA) Kruskal-Wallis test to compare the clusters within a
hierarchical clustering method. Similarly, Bushel et al. (2002) 216 implemented a single gene parametric ANOVA
test to assess the effects of genes on hierarchical clustering results. Recently, Kimes et al. (2017) 186 proposed a
method based on a Monte Carlo approach to test the statistical significance of hierarchical clustering results while
controlling for family-wise-error rate. However, family-wise-error rate can also be controlled while applying
repetitive statistical tests by implementing a Bonferroni correction 217.
2.4. Integrating domain knowledge into clustering algorithms
Other frequently used clustering algorithms in the bioinformatics field are k-means and fuzzy c-means.
However, these algorithms require initial random assignments of the clusters, which can produce inconsistent
results 202. Hence, they might fail to converge to the same results, even after multiple initiations using the same
dataset 197. In light of these limitations, Tari et al.(2009) 197 proposed the “GO Fuzzy c-means” clustering algorithm.
Their method resembles the fuzzy C-mean algorithms 218 and implements Gene Ontology annotation 219 as
biological domain knowledge to guide the clustering procedure. Even though this method assigned genes to
multiple clusters, which could have improved the biological relevance of the results, it was not capable of
discriminating the cluster memberships that were assigned due to random chance. While the algorithm parameters
selected in this study might have been reasonable for the dataset analyzed, the authors highlighted that future
studies would need to experimentally determine these parameters. Similarly, Khakabimamaghani and Ester (2016)
210 integrated domain knowledge via the selection of parameters for prior probability distributions. However, their
results reveal that the selection of these parameters had a direct impact on their clustering results. When analyzing
the effects of priors, the authors indicate that “final selected priors favor better sample clustering over better gene
clustering” 210. These findings reveal that the parameters need to be carefully selected since they can bias their
method towards better sample clustering rather than better gene clustering results.
Researchers can implicitly integrate domain knowledge to their methods by judiciously selecting the input data
of their algorithms 178. Genomic datasets may include relevant features as well as correlated and non-informative
features. The presence of correlated and non-informative features might obscure relevant patterns and prevent an
algorithm from discovering the underlying cluster structure of a dataset 195. Genomic data is generally high-
dimensional because the number of features is frequently greater than the number of samples. Additionally, genetic
variants are commonly correlated with other variants in close proximity on DNA. Therefore, when clustering
genomic data it is important to prune non-informative and correlated features 178,185.
Highly correlated SNPs are said to be in Linkage Disequilibrium (LD). This characteristic makes it challenging
for unsupervised ML algorithms to discover relevant cluster structures in the dataset. GWA studies present
102
significant associations as tag SNPs, implying a true causal SNP can be found within the LD block of a tagged
location 187. LD pruning refers to removing highly correlated SNPs within LD blocks. For example, Yazdani et al.
(2016) 220 identified a subset of informative SNPs based on a correlation coefficient. Similarly, Goldstein et al.
(2010) 185, implemented several correlation coefficient cut-off values (e.g., 0.99, 0.9, 0.8, 0.5) to remove SNPs
with high LD. They achieved this by using the toolsets for Whole-Genome Association and Population-Based
Linkage Analyses (PLINK) 221, resulting in a reduction of up to 76% of the original dataset. This reduction
decreased the computational complexity of their method 185. However, researchers have not agreed yet on a
standard correlation coefficient cut-off value that can be applied to every genomic dataset to reduce complexity
without incurring in significant information loss.
Table 1. Summary of current methods
Papers LD Pruning
Automatic
selection of k *
Statistical tests performed No selection of parameters required ǂ
[9,11,19, 197 X
216 X X
192,202,205,207,210,
211,213
X
186, 193, 215, 222,
X
This work X X X X
*k is the parameter defining the number of cluster in the dataset.
ǂ No parameters or hyperparameters are required to be known or selected a priori by researchers (e.g., prior probability, toroid grid size).
Table 1 shows a summary of the current clustering methods in the field of bioinformatics applied to
genomics data. It can be shown that multiple methods prune the SNPs of their datasets based on the degree of LD
between nearby SNPs. This is done in order to guide their clustering search and remove potentially non-informative
features. However, the vast majority of existing methods still require preselecting the number of clusters and other
parameters a priori (e.g., prior probability distributions, toroid grid size). Moreover, the current methods do not
commonly implement statistical analysis to test for the significance of their results, or to account for possible
family-wise-error rates. In light of this, the authors propose an unsupervised machine learning method to identify
sub-groups within cohorts of patient afflicted with the same disease. This is done by clustering patients based on
their genomic similarity without the need to any a priori information or input parameter. The method presented in
this work takes advantage of LD between SNPs by pruning correlated SNPs. In addition, it automatically selects
the number of clusters by implementing an internal validation metric. The method ensembles the clustering
outcomes of multiple linkage methods via a majority vote approach. Subsequently, it tests for statistical
significance among results while accounting for family-wise-error rate. Finally, a gene pathway analysis is
performed to support the potential medical significance of the results.
103
3. Method
An unsupervised machine learning method is presented that does not require selection of input parameters a
priori. The method can help identify patient cluster structures within genomic data and potentially discover
valuable differences between them. This knowledge can be used to advance personalized medicine of complex
diseases with heritable components, especially autoimmune disorders which have many susceptibility loci. Fig. 1
shows an outline of the method presented in this work.
Figure 1. Outline Method
3.1 Linkage Disequilibrium Pruning
Pruning SNPs based on LD serves as a feature reduction step. Thus, in the proposed method, SNPs
that are strongly correlated to other nearby SNPs are pruned, as previously done in the literature. The degree of LD
between SNPs is assessed by calculating the correlation coefficients based on a sliding window method. In this
method, cut-off values of (i) 0.999, (ii) 0.99, (iii) 0.9, (iv) 0.8 and (v) 0.5 are employed. Previous studies have
shown these cut-off values provide a balance between error reduction and information loss 185. Hence, five subsets
of patients’ genomic data containing different sets of SNPs (i.e., features) are generated. The subsets generated
serve as input for the hierarchical clustering step.
3.2 Hierarchical Clustering
The objective of the unsupervised machine learning method proposed in this work is to cluster patients
based on their genomic similarity. Patients’ genomic similarity can be evaluated using a wide range of distance
metrics 202. The selection of the appropriate distance metric is driven by the type of data under analysis (e.g., ratio,
interval, ordinal, nominal or binary scale). For example, the Euclidian distance is appropriated for ratio or interval
scale data, while the Manhattan distance for ordinal scale data 223.
Subsequently, the method proposed in this work employs an agglomerative hierarchical clustering algorithm.
Hierarchical clustering algorithms are frequently used with only one linkage method, which can limit their ability
104
to identify underlying cluster structures in certain datasets 200. Hence, in this work, multiple linkage methods are
implemented. The linkage methods used in this work have been shown to consistently outperform other methods
when tested with simulated GWAS datasets 202. The cluster results obtained by implementing different linkage
methods are ensemble in the subsequent steps. This ensemble takes advantage of the performance of multiple
linkage methods. Moreover, it helps identify the underlying structure of the data, since the ensemble approach will
favor cluster structures identified by the majority (i.e., via a majority vote approach) of the linkage methods.
Specifically, the authors propose to implement:
(i) Single Linkage (or Minimum Linkage).
(ii) Complete Linkage (or Maximum Linkage).
(iii) Average Linkage (or Unweighted Pair Group Method with Arithmetic Mean, UPGMA).
(iv) Ward’s Linkage.
(v) McQuitty Linkage (or Weighted Pair Group Method with Arithmetic Mean, WPGMA).
3.3 Parameter Selection
Once the agglomerative hierarchical algorithm is implemented, the Silhouette index is employed as an internal
validity metric. This index has been used in previous studies to rank the results of multiple clustering algorithms
outcomes and guide the selection of final clusters 192,202. Nonetheless, in this method, the index is used to select the
number of clusters for all combinations of LD pruning data subsets (see section 3.1) and linkage methods (see
section 3.2). The number of clusters that provides the largest average Silhouette index value in each of the
combinations is selected.
The computational complexity of testing all possible numbers of clusters increases linearly as the number of
objects in a dataset increases. This can be a challenge in datasets that contain a large number of objects, even with
parallelized computing. In this work, an optimization approach is presented to identify the number of clusters that
maximizes the average Silhouette index. The mathematical formulation of this optimization problem is as follows:
𝑀𝑎𝑥𝑖𝑚𝑖𝑧𝑒 𝑆𝐼 (1)
𝑆𝐼 = 1
𝑘 ∑ [
1
𝜂𝑖 ∑ 𝑆(𝑥)𝑥∈𝑪𝑖
]𝑘𝑖=1 ∀ 𝑖 ∈ 𝑲{1, … , 𝑘} (2)
𝑆(𝑥) = 𝑏(𝑥)−𝑎(𝑥)
max {𝑏(𝑥),𝑎(𝑥)} ∀𝑥 ∈ 𝑪𝑖 (3)
𝑎(𝑥) = 1
𝜂𝑖−1 ∑ 𝑑(𝑥, 𝑦)𝑦∈𝑪𝑖
∀ 𝑥 ≠ 𝑦, 𝑥 and 𝑦 ∈ 𝑪𝑖 (4)
𝑏(𝑖) = 𝑚𝑖𝑛𝑤∈𝑲,𝑤≠𝑖 {1
𝜂𝑤 ∑ 𝑑(𝑥, 𝑔)𝑔∈𝑪𝑤
} ∀ 𝑔 ∈ 𝑪𝑤 , 𝑤 ∈ 𝑲 (5)
1 < 𝑘 ≤ 𝑛 (6)
Where,
𝑆𝐼: is the average Silhouette index of the clusters set K.
105
K: is the set of clusters obtained with the hierarchical clustering algorithm for a given number of k
disjoint clusters.
𝜂𝑖: is the number of objects that belongs to cluster 𝐂𝑖 , for 𝑖 ∈ 𝑠et of clusters {𝑲}.
𝑆(𝑥): is the Silhouette of object x, for x ∈ 𝐂𝑖 .
𝑎(𝑥): is the average similarity of object x with all other objects that belong to the same cluster of x
(i.e., 𝐂i).
𝑏(𝑥): is the average similarity of object x with the objects from the nearest clusters 𝐂𝑤, for 𝑤 ∈
𝑠et of cluster {𝑲}, 𝑖 ≠ 𝑤.
Eq. (1) represents the objective function that needs to be maximized (i.e., the average Silhouette index). Eq. (2),
shows the mathematical representation of the average Silhouette index, while Eq. (3) shows the silhouette of a
given object x. Both Eq. (4) and (5) represent the elements that constitute the Silhouette index of a given object x
203. Finally, Eq. (6) constrains the search for the number of clusters to be greater than 1 and less than the total
number of objects n (i.e., the maximum number of clusters). Since the objective function is non-linear with respect
to the parameter k (i.e., number of clusters), this optimization problem needs to be solved with a non-linear
optimization algorithm. In the literature, there are several algorithms suitable to solve this type of optimization
problem 224. Nonetheless, the proposed method is not constrained to any specific optimization algorithm.
Once the number of clusters is identified in all datasets combination, the results are ensemble into a final cluster
set via a majority vote approach. Table 2 shows an example of this consensus clustering approach in which patient
i is assigned to the final cluster 1 since the majority of the cluster results assigned that patient to that given cluster.
Similarly, patient n is assigned to the final cluster 2, since the majority of clusters assigned this patient to this
cluster.
Table 2. Example of consensus clustering
No. LD
Pruning
Linkage
Method
Patient
i cluster
…
Patient
n cluster
1 1 Single 1 2
2 0.99 Single 1 2
3 0.90 Single 1 1
4 0.8 Single 1 2
5 0.5 Single 2 1
… …
24 0.8 McQuitty 2 2
25 0.5 McQuitty 1 2
Final Cluster 1 2
3.4 Statistical Significance
106
After the final patient clusters are discovered, a single SNP ANOVA test is performed to reveal the SNPs that
are statistically significantly different between the clusters of patients. This step helps validate that the clusters
generated are different by at least one associated SNP. To account for family-wise-error-rate a Bonferroni
correction is applied by dividing the alpha value by the number of tested SNPs. In the case that no SNPs are found
to be statistically significantly different, it can be concluded that the resulting patients’ clusters might have arisen
due to random chance.
3.5 Gene Pathway Analysis
The set of SNPs significantly associated with differences between patient clusters can be explored via
Gene Ontology (GO) enrichment and mutational burden on molecular pathways. By assigning each SNP to a gene
and performing a gene network analysis, (e.g., via STRING-DB software 225), visualizations of gene networks and
estimations of significant enrichment along GO terms could provide evidence for potential biological significance.
The significance is assessed by comparing the number of evidence-based relationships between selected genes to
the number expected in a set of randomly selected genes. If an enrichment is established, the related genes are
examined by their molecular function, biological relevance, and known associations to the disease from GWA
studies. While pathway analysis does not provide the rigor of direct experiment or clinical trial, it remains valuable
in determining whether selected genes are functionally relevant to the disease studied, as opposed to being a
function of other factors such as ethnicity.
4. Application
The performance of the proposed method is first tested on the datasets presented in the Fundamental
Clustering Problem Suite (FCPS) 201. The FCPS contains 10 different datasets designed to assess the performance
of unsupervised machine learning algorithms on particular clustering challenges (e.g., outliers, undefined cluster
boundaries). The ground truth data of cluster membership are used to test the performance of the proposed method
in identifying clusters resulting from the underlying structures in the data and not just from random variation. To
measure this performance the Rand index 226 validation metric is implemented. Moreover, the performances of
other existing methods in the literature are benchmarked with the same datasets. All the benchmark analyses were
performed on a 12 Core i7 3.4 GHz IntelTM computer with 62.8 GB of RAM and Ubuntu 16.04 LTS. The
benchmark methods were implemented in R v.3.4 227 with the used of the packages mclust v.5.3 228, apcluster
v.1.4.4 229, DatabionicSwarm v.0.9.8 230,231, NNLM v.0.4.1 211, and biclust v.1.2.0 232.
Additionally, the proposed method and the benchmark methods that do not require providing the number of
cluster a priori (i.e., Clifford et al. (2011) 202: hierarchical clustering algorithm with silhouette index, Sakellariou
et al. (2012) 205: Affinity Propagation clustering algorithm, Shen et al. (2009) 207: Expectation Maximization
clustering algorithm, Cheng and Church (2000) 213: Iterative Greedy Search Biclustering algorithm.), are
implemented on two genomic datasets. Frist, the microarray gene expression data of patients with lymphoblastic
and acute myeloid leukemia from Golub et al. (1999) 233 was implemented. The dataset is publically available at
the Broad Institute and has been previously used to test the performance of clustering algorithms 199,234. The dataset
107
is composed of microarray gene expression data of 999 genes for 27 patients with acute lymphoblastic leukemia
and 11 patients with acute myeloid leukemia.
Lastly, a dataset of patients diagnosed with MS is employed. DNA samples from 191 Multiple Sclerosis (MS)
patients consented via the Pennsylvania State University PRIDE protocol at Hershey Medical Center were
subjected to the Immunochip assay (Illumina). Allelic variations were measured at previously described
susceptibility loci for multiple immune-mediated disorders 235,236. The Y chromosome data were filtered out of the
dataset to simplify comparisons in a predominantly female cohort. Mitochondrial markers were discarded for
analysis as well. Genotype calling was done with Illumina GenomeStudio v.2011.1 (www.illumina.com), and
genotype markers were excluded if their GenTrain score was less than 0.8, or if their call rate across the cohort
was less than 0.99. Finally, the MS dataset was filtered such that only variants within coding regions (i.e., exons),
were considered. Therefore, the MS dataset was composed of 191 patients and 25,482 SNPs.
With the MS dataset, a 10-fold cross-validation analysis was performed with the objective to test the performance
of the proposed and the benchmark methods, as well as to provide evidence regarding their propensity of overfitting
genomic datasets. In this cross-validation approach, the MS dataset was randomly partitioned into 10 subsets.
Subsequently, the methods were used to cluster the patients within these subsets. The clustering results obtained
from the 10 subsets were compared to those from the complete dataset. The agreement between the clusters
generated with the complete MS dataset and the 10-fold subsets is assessed with the Rand index metric. A match
between the clustering results (e.g., average Rand index of 1) will indicate that the proposed method was not
overfitting the MS dataset, thus, providing arguments of its generalizability. Moreover, it will support that the
method was identifying clusters due to underlying structures in the data and not just due to random variations.
Finally, the groups of SNPs identified by the proposed method to achieve statistical significance between clusters
generated were examined via gene pathway analysis.
4.1 Linkage Disequilibrium Pruning
For the MS dataset, the pruning of SNPs with a high LD was done based on the correlation-coefficient cut-off
values found in the literature, as proposed in section 3.1. LD pruning was performed using the widely used
genotype analysis toolset for Whole-Genome Association and Population-Based Linkage Analyses (i.e., PLINK)
221. This pruning resulted in a reduction of the original dataset as presented in Table 3. These percentages of SNPs
removed are consistent with the results found in previous studies.
Table 3. LD Pruning summary
R2
cut-off value
Number
of SNPs retained
Percentage
of SNPs removed
0.50 5,460 78.57%
0.80 6,849 73.12%
0.90 7,421 70.88%
108
0.99 8,666 65.99%
0.999 8,691 65.89%
4.2 Hierarchical Clustering
The FCPS and Golub et al. (1999) 233 datasets contain features that are in ratio scale. Hence, to measure
the similarity between the objects in the datasets the Euclidian distance is implemented. Genotype data can be
ordinal or additive scale, depending on whether heterozygous SNPs are treated as a label or as a half-dosage. While
additive models are more often used for GWA studies, in this work ordinal scale was used to demonstrate flexibility
in the described clustering method. Hence, the genomic similarity of MS patients based on different subsets of
pruned data is evaluated using the Manhattan distance metric. The similarity calculations and the agglomerative
hierarchical algorithm with multiple linkage methods were performed in R v.3.4 227.
4.3 Parameter Selection
The selection of the number of clusters k that maximized the average Silhouette index was performed with a
generalized simulation annealing algorithm. This algorithm was selected due to its underlying theory and proven
performance in problems with non-linear objective functions 237,238. The algorithm was implemented via the R
package GenSA v.1.1.6 239. Nonetheless, other non-linear optimization algorithms or greedy heuristics can also be
implemented. Once the number of clusters in every combination of LD pruned data and linkage method are
selected, the clustering results are ensemble via a majority vote approach (see section 3.3).
4.4. Statistical Significance
After the final clusters have been selected based on the average Silhouette metric and consensus clustering
approach the statistical significance of the results is evaluated. Clusters’ median values for each of the p features
in the MS dataset are evaluated via a single SNP non-parametric ANOVA Kruskal-Wallis test 222. To account for
family-wise-error rate, a Bonferroni correction is applied to the significance alpha level of 0.05 (i.e., Bonferroni
correction= 0.05/p, for p= 25,482).
4.5. Gene Pathway Analysis
Gene variants that show statistical significance are further analyzed via a gene pathway analysis to
explore their potential medical significance. Pathway analysis starts with generating a list of genes determined
from the set of SNPs with strong evidence of significance between patient clusters. Inputting the gene set via the
STRING-DB software algorithms 225 allows for convenient calculation of pathway enrichment hypothesis tests and
visualization of the gene network. STRING-DB determines gene relationships by aggregating several databases
into an evidence score. Experimental evidence comes from the BIND 240, GRID 241, HPRD 242, IntAct 243, MINT
109
244, and PID 245 databases. In addition, STRING-DB pulls from the curated databases KEGG 246, Gene Ontology
219, BioCarta 247, and Reactome 248. Interaction frequency is tested for enrichment compared to expectation from a
random sampling of genes, with p-values and false discovery rates reported for enrichment in specific cellular
processes, defined by Gene Ontology references. After statistical testing is done, the gene network is used as a
threshold for high confidence interaction and a k-means clustering algorithm is performed for visualization
purposes (see Fig. 7).
5. Results
5.1. FCPS Benchmark results
The majority of existing methods in the literature require the selection of parameters a priori (e.g., number of
clusters, see Table 1). Hence, to benchmark with multiple methods, the number of clusters provided by the FCPS
was used as input when testing these methods. Figure 2 shows the average Rand index obtained in the FCPS
datasets by the method proposed in this work (i.e., Proposed) and the methods benchmarked. This plot shows that
on average the proposed method outperformed other methods, with an average Rand index of 0.852. The
performance is statistically significantly greater than the results of the methods proposed by Cheng and Church
(2000), Sakellariou et al. (2012), Lee and Seung (2001), Ultsch and Lötsch (2017), and Clifford et al. (2011). Even
though these results indicate that, on average, the proposed method achieved the largest Rand index, there is not
enough evidence to conclude that it was statically significantly greater than the Rand index achieved by the
methods of Shen et al. (2009), Hamid et al. (2010), or Ahmad et al. (2016), at an alpha level of 0.05. This can be
attributed to the relatively small group of validation datasets provided in the FCPS (i.e., 10 datasets).
Note: p-value: <0.001***, <0.01**, <0.05*
Figure 2. Average Rand index for FCPS datasets
110
Note: p-value: <0.001***, <0.01**, <0.05*
Figure 3. The proportion of results with Rand index of 1 for FCPS datasets (i.e., perfect clustering)
Similarly, Figure 3 shows the proportion of the FCPS datasets that achieved a clustering result with a Rand index
of 1 (i.e., perfect clustering) for each of the given methods. The results reveal that the proposed method was able
to obtain a Rand index of 1 in 6 out of the 10 FCPS datasets. The results from the Wilcoxon tests indicate that
these results are statistically significantly greater than the results of the methods proposed by Ultsch and Lötsch
(2017), Cheng and Church (2000), Lee and Seung (2001), and Sakellariou et al. (2012). Even though the results
indicate the proposed method correctly clusters the largest percentages of datasets (i.e., 6/10), there is not enough
evidence to conclude that this proportion is statically significantly greater than the ones from the other methods
benchmarked, at an alpha level of 0.05. Nevertheless, these results provide evidence that the method proposed in
this work is able to identify true clusters in a wider range of datasets with different underlying structures.
5.2. Genomic dataset Benchmark results
Figure 4 presents the Rand index obtained on the Golub et al. (1999) dataset 233 by the method proposed in this
work and the benchmark methods that do not require providing the number of clusters a priori. Fig. 4 indicates
that the proposed method performed better than the methods presented by Clifford et al. (2011), Cheng and Church
(2000), and Sakellariou et al. (2012).
111
Figure 4. Rand index for Leukemia dataset
Figure 5 shows the average Rand index obtained with the MS dataset and the 10-fold cross-validation
approach by the proposed and benchmark methods. The iterative greedy search Biclustering algorithm proposed
by Cheng and Church (2000) was not able to find any cluster structure in the MS dataset; hence it was not included
in this plot. The plot shows that on average the proposed method outperformed the other methods, with an average
Rand index of 0.969. This is statistically significantly greater than the values obtained with the other methods
benchmarked. Moreover, the average Rand index obtained by the proposed method was not significantly different
than an average Rand index of 1 (t-value: -1.963, p-value=0.0812), at an alpha level of 0.05. This reveals that on
average the proposed method found a perfect match between the clusters of patients obtained with the complete
MS dataset and the cross-validation subsets.
\ Note: p-value: <0.001***, <0.01**, <0.05*
Figure 5. Average Rand index for MS dataset
112
Table 4 shows the confusions matrix of the clusters obtained with the proposed method when implementing
the 10-fold cross-validation approach. The table indicates that the proposed method was able to group 96.33% of
the patients’ in the same clusters when both the complete dataset and the different data subsets were used (i.e.,
accuracy of 0.96), which is in line with the average Rand index of 0.969 shown in Fig. 5. The Rand index and
confusion matrix results indicate that the proposed method identified a similar cluster structure even with different
subsets of the MS dataset. This indicates that the proposed method was not overfitting the dataset. Furthermore, it
provides evidence that supports that the method was able to identify clusters due to the underlying structure of the
data and not just due to random change.
Table 4. MS dataset 10-fold cross-validation confusion matrix
Complete Dataset
Cluster
1
Cluster
2
Data
Subsets
Cluster
1 172 0
Cluster
2 7 12
The results from the pathway analysis on the set of statistically significant different SNPs between the
MS patient clusters are shown in Fig. 6. The cluster-defining SNPs show significantly more interactions than
expected among a random sampling of genes. Out of 515 genes, 1,463 interactions were found, with only 942
expected by chance (p-value: 1.04e-10), among a background set of 4,938 genes present on the Immunochip. The
gene interactions in the set shown in Fig. 6 demonstrate a high prevalence of cellular adhesion, cytokine response,
and general immune process pathways.
113
Figure 6. Gene Pathway Analysis results
Table 5 shows relationships between genes based on evidence from literature via STRING-DB 225. The highly
connected pathway depicted contains many genes known to be involved in cell adhesion and leukocyte physiology,
both of which are processes dysregulated in MS 249. Additionally, the genes selected show significant Gene
Ontology term enrichment in these categories, with false discovery rates less than 0.01. Taken together, pathway
analysis reveals that extracting significant features between clusters may be a valid feature reduction technique
for downstream analysis. Genes known to be relevant in MS pathophysiology (e.g., interleukin receptors, STAT
transcription factors, lymphocyte surface proteins from the CCR family) were highlighted despite no use of
supervised methods and label data, implying that the proposed unsupervised method’s value is not just
discovering patient clusters, but reducing the dimensionality by nearly 20-fold with few samples (i.e., from over
25,482 features to around 1,500, using 191 samples).
Pathway ID Pathway
Description
Count
in Gene Set
False
Discovery Rate
GO.0051249 Regulation of
lymphocyte activation
32 0.00641
GO.0002823 Negative
regulation of adaptive
response
9 0.00749
114
GO.0006952 Defense
response
73 0.00749
GO.0002694 Regulation of
leukocyte activation
33 0.00804
GO.0050865 Regulation of
cell activation
35 0.00804
GO.0002376 Immune
system process
93 0.00898
Table 5. Gene Pathway Analysis results
As a secondary observation, an analysis was done on the MS dataset after pruning samples which showed
greater than 0.2 similarity in PLINK’s Identity-By-Descent (IBD) algorithm 250. This was done to remove
potentially related patients from the analysis. IBD identified a total of 11 potentially related patients, from whom
10 were initially assigned to cluster number two. Consequently, after removing these potentially related patients
from the MS dataset and applying the proposed method, the number of patients in the second cluster were reduced
from 12 to 2, and no pathway enrichment was detected. However, the 120 genes detected still included T-cell
relevant proteins such as STAT and JAK, as well as members of the tumor necrosis factor and interleukin families,
supporting the claim that the method identified SNPs relevant to the disease process even if the sample size of the
smaller cluster (n=2) constrains the power of the pathway analysis. Furthermore, the cross-validation results
indicate that the average Rand index achieved after removing potentially related patients (i.e., 0.932) was not
significantly different than the initial cross-validation results (i.e., 0.969, see Fig. 5) (t-value: 1.52, p-value: 0.147).
This reveals that the proposed method was able to identify the same underlying cluster structure in the MS dataset,
and identify patients with similar genomic makeup after the removal of potentially related individuals. These
results provide evidence that supports that the method was able to identify clusters due to the underlying structure
of the data and not just due to random change.
6. Conclusion and future work
Many chronic disorders have genomic etiology, disease progression, clinical presentation, and response to
treatment that vary on a patient-to-patient basis. Such variability creates a need to identify characteristics within
patient populations that have clinically relevant predictive value. Unsupervised machine learning methods are
suitable to address this type of problem, in which no class label information is available to guide this search.
However, it is challenging for existing methods to identify cluster memberships that are due to the underlying
structures in the dataset and not just a result of natural sampling variation. Moreover, most current methods require
researchers to know and provide input parameters a priori. As a result of these limitations and the need to advance
115
personalized medicine, this work proposed an unsupervised machine learning method to identify genomically
distinct patients’ cluster. The proposed method integrates statistical analysis to test for significance of clustering
results and accounts for family-wise-error rate. Moreover, the method is capable of automatically identifying the
number of clusters by implementing an internal validity metric. Similarly, the method takes advantage of the degree
of linkage disequilibrium between SNPs by pruning correlated nearby SNPs, as well as implementing a post-
clustering gene pathways analysis.
The proposed method is tested with clustering validation datasets previously used in the literature. The
benchmark results reveal that proposed method provides, on average, the greatest performance (i.e., average Rand
index 0.852). Moreover, results indicate that it was able to obtain cluster results with a Rand index of 1 (i.e., perfect
clustering) in 6 out of the 10 Fundamental Clustering Problem Suite (FCPS) datasets. Similarly, the method is
applied to a dataset of 38 patients with leukemia, and subsequently to a dataset of 191 Multiple Sclerosis (MS)
patients. The results indicate that the proposed method is able to identify genetically distinct patient clusters without
the need to select the number of clusters or any input parameter a priori. Moreover, the cross-validation results
indicate that the proposed method outperformed the other methods when it comes to data overfitting, since the
average Rand index obtained was significantly greater than the benchmarked methods and not significantly
different than 1. This performance was maintained even after the removal of potentially related patients from the
dataset. This indicates that the method was identifying clusters due to the underlying structure of the data and
avoided overfitting the dataset. The identification of distinct genetic subtypes of patients demonstrates the potential
applicability of this process to advance personalized medicine of complex diseases with heritable components,
especially autoimmune disorders.
When applied to genomic data, the method also shows value as a feature reduction strategy. Out of over
25,482 exonic SNPs and 191 patient samples, the clustering of patients yielded a set of SNPs which significantly
vary between clusters. These variants represent 515 genes, several of which are known to be involved in MS
(CD69, CCRX5, IL-13, STAT3) and cell adhesion (ICAM1, LAMB4). The fact that many highlighted genes are
components of the immune system is not surprising due to the nature of the Immunochip assay, but the enrichment
of leukocyte-specific genes is evidence that the method can result in functionally relevant feature sets, even without
class labels. Notably, 57 genes representing over 10% of the network are involved in cytokine receptor processes.
This is greater than expected from random chance, as cytokine receptors constitute a small percentage of all
Immunochip genes. The evidence presented in this work alone is insufficient to define genetic subtypes of MS, but
the specific SNP set reaching significance may be a valuable resource in experimental studies examining immune
cell dynamics and genetics. For example, the hypothesis that these clusters represent different subtypes of MS can
be tested by evaluating clinical criteria such as image results and disease progression, as well as quantitative
cytokine profiling and gene expression studies for each cluster as a group, compared against random groupings of
patients.
This work demonstrates an iterative unsupervised machine learning method which identifies significant patient
clusters within a genomic dataset. Future research should explore the medical significance of the findings shown
in this work. Similarly, the method from this work should be implemented in studies collecting SNP array and
gene expression microarray data from additional disease cohorts to explore its potential benefits. Further
116
investigation can extend beyond pathway analysis to evaluate these clusters for clinical significance of genetically
related characteristics such as age of onset, disease course, heritability, and response to treatment. Once links are
drawn between clusters and clinically relevant outcomes, the Immunochip can be used to classify high-risk and
newly diagnosed chronic disease patients into clusters with predictive value.
Acknowledgments
The authors acknowledge the NSF I/UCRC Center for Healthcare Organization Transformation (CHOT), NSF
I/UCRC grant #1624727, and the Institute for Personalized Medicine at the Pennsylvania State University.
Additionally, the authors would like to acknowledge Dr. James R. Broach from the Institute for Personalized
Medicine at the Pennsylvania State University, for his valuable contributions. Any opinions, findings, or
conclusions found in this paper are those of the authors and do not necessarily reflect the views of the sponsors.
Reference List
1. Zura, R. et al. Epidemiology of Fracture Nonunion in 18 Human Bones. JAMA Surg 151,
e162775 (2016).
2. Brauer, C. A., Coca-Perraillon, M., Cutler, D. M. & Rosen, A. B. Incidence and mortality of
hip fractures in the United States. Jama 302, 1573–1579 (2009).
3. Shih, K.-S., Hsu, C.-C. & Hsu, T.-P. A biomechanical investigation of the effects of static
fixation and dynamization after interlocking femoral nailing: a finite element study. J
Trauma Acute Care Surg 72, E46-53 (2012).
4. Ulrich, D., van Rietbergen, B., Weinans, H. & Rüegsegger, P. Finite element analysis of
trabecular bone structure: a comparison of image-based meshing techniques. Journal of
biomechanics 31, 1187–1192 (1998).
5. Huiskes, R., Janssen, J. D. & Slooff, T. J. A detailed comparison of experimental and
theoretical stress analyses of a human femur. Mechanical properties of bone 45, 211–234
(1981).
6. Perren, S. & Buchanan, J. Basic concepts relevant to the design and development of the
point contact fixator (PC-Fix). Injury 26, 1–4 (1995).
7. Runyan, C. & Gabrick, K. Biology of bone formation, fracture healing, and distraction
osteogenesis. J Craniofac Surg Ahead of print, (2017).
8. Maggiano, I. et al. Three-dimensional reconstruction of haversian systems in human
cortical bone using synchotron radiation-based micro-CT: morphology and quantification of
branching and transverse connections across age. J Anatomy 228, 719–732 (2016).
9. Cheal, E. J., Mansmann, K. A., DiGioia, A. M., Hayes, W. C. & Perren, S. M. Role of
interfragmentary strain in fracture healing: ovine model of a healing osteotomy. J. Orthop.
Res. 9, 131–142 (1991).
10. Levy, S. et al. Immature myeloid cells are critical for enhancing bone fracture healing
through angiogenic cascade. Bone 93, 113–124 (2016).
11. Duda, G. N. et al. Interfragmentary motion in tibial osteotomies stabilized with ring
fixators. Clinical orthopaedics and related research 396, 163–172 (2002).
12. Garnavos, C. Treatment of aseptic non-union after intramedullary nailing without removal
of the nail. Injury 48, S76–S81 (2017).
13. Santiago, H., Zamarioli, A., Neto, M. & Volpon, J. Exposure to secondhand smoke impairs
fracture healing in rats. Clinical Orthopaedics and Related Research 475, 894–902 (2017).
14. Sathyendra, V. et al. Single Nucleotide Polymorphisms in Osteogenic Genes in Atrophic
Delayed Fracture-Healing: A Preliminary Investigation. J Bone Joint Surg Am 96, 1242–
1248 (2014).
15. Damm, P., Kutzner, I., Bergmann, G., Rohlmann, A. & Schmidt, H. Comparison of in vivo
measured loads in knee, hip and spinal implants during level walking. Journal of
Biomechanics 51, 128–132 (2017).
16. Kienast, B. et al. An electronically instrumented internal fixator for the assessment of bone
healing. Bone and Joint Research 5, 191–197 (2016).
17. Schneider, E. et al. Loads acting in an intramedullary nail during fracture healing in the
human femur. Journal of biomechanics 34, 849–857 (2001).
18. Schlecht, S. H., Pinto, D. C., Agnew, A. M. & Stout, S. D. Brief communication: The
effects of disuse on the mechanical properties of bone: What unloading tells us about the
adaptive nature of skeletal tissue. American Journal of Physical Anthropology 149, 599–
605 (2012).
19. Akhter, M. P., Alvarez, G. K., Cullen, D. M. & Recker, R. R. Disuse-related decline in
trabecular bone structure. Biomechanics and Modeling in Mechanobiology 10, 423–429
(2011).
20. Miner, M. A. Cumulative damage in fatigue. J Applied Mech 12, A159-64 (1945).
21. Wee, H., Reid, J. S., Chinchilli, V. M. & Lewis, G. S. Finite Element-Derived Surrogate
Models of Locked Plate Fracture Fixation Biomechanics. Annals of Biomedical Engineering
45, 668–680 (2017).
22. Lenz, M., Perren, S. M., Gueorguiev, B., Höntzsch, D. & Windolf, M. Mechanical behavior
of fixation components for periprosthetic fracture surgery. Clin Biomech (Bristol, Avon) 28,
988–993 (2013).
23. Fratzl, P. & Weinkamer, R. Nature’s hierarchical materials. Progress in Materials Science
52, 1263–1334 (2007).
24. Cowin, S. Pathology of functional adaptation of bone remodeling and repair in vivo. in
Bone mechanics handbook (Taylor and Francis Group, 2001).
25. Comiskey, D. P., MacDonald, B. J., McCartney, W. T., Synnott, K. & O?Byrne, J. The role
of interfragmentary strain on the rate of bone healing?A new interpretation and
mathematical model. Journal of Biomechanics 43, 2830–2834 (2010).
26. Augat, P. et al. Shear movement at the fracture site delays healing in a diaphyseal fracture
model. Journal of Orthopaedic Research 21, 1011–1017 (2003).
27. Elkins, J. et al. Motion Predicts Clinical Callus Formation: Construct-Specific Finite
Element Analysis of Supracondylar Femoral Fractures. The Journal of Bone and Joint
Surgery 98, 276–284 (2016).
28. Bottlang, M. et al. Far cortical locking can improve healing of fractures stabilized with
locking plates. J Bone Joint Surg Am 92, 1652–1660 (2010).
29. Claes, L., Recknagel, S. & Ignatius, A. Fracture healing under healthy and inflammatory
conditions. Nature Reviews Rheumatology 8, 133–143 (2012).
30. Augat, P. et al. Early, full weightbearing with flexible fixation delays fracture healing.
Clinical Orthopaedics and Related Research 194–202 (1996).
31. Schell, H. et al. Mechanical induction of critically delayed bone healing in sheep:
Radiological and biomechanical results. Journal of Biomechanics 41, 3066–3072 (2008).
32. Hente, R., Füchtmeier, B., Schlegel, U., Ernstberger, A. & Perren, S. M. The influence of
cyclic compression and distraction on the healing of experimental tibial fractures. Journal
of orthopaedic research 22, 709–715 (2004).
33. Claes, L., Eckert-Hübner, K. & Augat, P. The effect of mechanical stability on local
vascularization and tissue differentiation in callus healing. Journal of orthopaedic research
20, 1099–1105 (2002).
34. Lujan, T. J. et al. Locked plating of distal femur fractures leads to inconsistent and
asymmetric callus formation. Journal of orthopaedic trauma 24, 156–162 (2010).
35. Duda, G. N. et al. Interfragmentary motion in tibial osteotomies stabilized with ring
fixators. Clinical orthopaedics and related research 396, 163–172 (2002).
36. Nassiri, M., Macdonald, B. & O’Byrne, J. M. Computational modelling of long bone
fractures fixed with locking plates - How can the risk of implant failure be reduced? J
Orthop 10, 29–37 (2013).
37. Beltran, M. J., Collinge, C. A. & Gardner, M. J. Stress Modulation of Fracture Fixation
Implants: Journal of the American Academy of Orthopaedic Surgeons 24, 711–719 (2016).
38. Pruitt, L. A. & Chakravartula, A. M. Mechanics of Biomaterials: Fundamental principles
for implant design. (Cambridge University Press, 2011).
39. Morehead, J. & Holt, G. Soft-tissue response to synthetic biomaterials. Otolaryngol Clin
North Am. 27, 195–201 (1994).
40. Nuss, K. M. & von Rechenberg, B. Biocompatibility issues with modern implants in bone -
A review for clinical orthopedics. The Open Orthopaedics Journal 2, 66–78 (2008).
41. Wright, T. & Maher, S. Biomaterials. in Orthopaedic basic science 65–85 (American
Academy of Orthopaedic Surgeons, 2007).
42. Ungersboeck, A., Geret, V., Pohler, O., Schuetz, M. & Wuest, W. Tissue reaction to bone
plates made of pure titanium: a prospective, quantitative clinical study. Journal of Materials
Science: Materials in Medicine 6, 223–229 (1995).
43. Arens, S. et al. Influence of materials for fixation implants on local infection. J Bone Joint
Surg Br 78, 647–651 (1996).
44. Uckan, S., Veziroglu, F., Soydan, S. S. & Uckan, E. Comparison of stability of resorbable
and titanium fixation systems by finite element analysis after maxillary advancement
surgery. J Craniofac Surg 20, 775–779 (2009).
45. Marasco, S. F., Liovic, P. & Šutalo, I. D. Structural integrity of intramedullary rib fixation
using a single bioresorbable screw. J Trauma Acute Care Surg 73, 668–673 (2012).
46. Böstman, O. & Pihlajamäki, H. Clinical biocompatibility of biodegradable orthopaedic
implants for internal fixation: a review. Biomaterials 21, 2615–2621 (2000).
47. Schliemann, B. et al. PEEK versus titanium locking plates for proximal humerus fracture
fixation: a comparative biomechanical study in two- and three-part fractures. Archives of
Orthopaedic and Trauma Surgery 137, 63–71 (2017).
48. Pruitt, L. A. & Chakravartula, A. M. in Mechanics of Biomaterials: Fundamental principles
for implant design (Cambridge University Press, 2011).
49. Bottlang, M., Doornink, J., Fitzpatrick, D. C. & Madey, S. M. Far Cortical Locking Can
Reduce Stiffness of Locked Plating Constructs While Retaining Construct Strength: The
Journal of Bone and Joint Surgery-American Volume 91, 1985–1994 (2009).
50. Bottlang, M. et al. Dynamic Stabilization of Simple Fractures With Active Plates Delivers
Stronger Healing Than Conventional Compression Plating: Journal of Orthopaedic Trauma
31, 71–77 (2017).
51. Piccinini, M., Cugnoni, J., Botsis, J., Ammann, P. & Wiskott, A. Influence of gait loads on
implant integration in rat tibiae: Experimental and numerical analysis. Journal of
Biomechanics 47, 3255–3263 (2014).
52. Gardner, M. J., Ricciardi, B. F., Wright, T. M., Bostrom, M. P. & van der Meulen, M. C. H.
Pause Insertions During Cyclic In Vivo Loading Affect Bone Healing. Clinical
Orthopaedics and Related Research 466, 1232–1238 (2008).
53. Tsuji, K. et al. BMP2 activity, although dispensable for bone formation, is required for the
initiation of fracture healing. Nature Genetics 38, 1424–1429 (2006).
54. McBride-Gagyi, S. H., McKenzie, J. A., Buettmann, E. G., Gardner, M. J. & Silva, M. J.
BMP2 conditional knockout in osteoblasts and endothelial cells does not impair bone
formation after injury or mechanical loading in adult mice. Bone 81, 533–543 (2015).
55. Helwig, P. et al. Finite element analysis of four different implants inserted in different
positions to stabilize an idealized trochanteric femoral fracture. Injury 40, 288–295 (2009).
56. Perez, A., Mahar, A., Negus, C., Newton, P. & Impelluso, T. A computational evaluation of
the effect of intramedullary nail material properties on the stabilization of simulated femoral
shaft fractures. Med Eng Phys 30, 755–760 (2008).
57. Fouad, H. Assessment of function-graded materials as fracture fixation bone-plates under
combined loading conditions using finite element modelling. Med Eng Phys 33, 456–463
(2011).
58. Feerick, E. M., Kennedy, J., Mullett, H., FitzPatrick, D. & McGarry, P. Investigation of
metallic and carbon fibre PEEK fracture fixation devices for three-part proximal humeral
fractures. Med Eng Phys 35, 712–722 (2013).
59. Favre, P., Kloen, P., Helfet, D. L. & Werner, C. M. L. Superior versus anteroinferior plating
of the clavicle: a finite element study. J Orthop Trauma 25, 661–665 (2011).
60. Tupis, T. M., Altman, G. T., Altman, D. T., Cook, H. A. & Miller, M. C. Femoral bone
strains during antegrade nailing: a comparison of two entry points with identical nails using
finite element analysis. Clin Biomech (Bristol, Avon) 27, 354–359 (2012).
61. Brown, C. J., Sinclair, R. A., Day, A., Hess, B. & Procter, P. An approximate model for
cancellous bone screw fixation. Comput Methods Biomech Biomed Engin 16, 443–450
(2013).
62. MacLeod, A. R., Pankaj, P. & Simpson, A. H. R. W. Does screw-bone interface modelling
matter in finite element analyses? J Biomech 45, 1712–1716 (2012).
63. Wang, H. et al. Accuracy of individual trabecula segmentation based plate and rod finite
element models in idealized trabecular bone microstructure. J Biomech Eng 135, 044502
(2013).
64. Zdero, R., Olsen, M., Bougherara, H. & Schemitsch, E. H. Cancellous bone screw purchase:
a comparison of synthetic femurs, human femurs, and finite element analysis. Proc Inst
Mech Eng H 222, 1175–1183 (2008).
65. Leonidou, A. et al. The biomechanical effect of bone quality and fracture topography on
locking plate fixation in periprosthetic femoral fractures. Injury 46, 213–217 (2015).
66. Ruedi, T., Buckley, R. & Moran, C. AO principles of fracture management. 1, (AO
Publishing, 2007).
67. Lewis, G. et al. Tangential biocortical locked fixation improves stability in Vancouver B1
periprosthetic femur fractures: A biomechanical study. J Orthop Trauma 29, e364–e370
(2015).
68. Tidwell, J. E. et al. The biomechanical cost of variable angle locking screws. Injury 47,
1624–1630 (2016).
69. Kandemir, U. et al. Implant Material, Type of Fixation at the Shaft and Position of Plate
Modify Biomechanics of Distal Femur Plate Osteosynthesis: Journal of Orthopaedic
Trauma 1 (2017). doi:10.1097/BOT.0000000000000860
70. Wilson, W. K., Morris, R. P., Ward, A. J., Carayannopoulos, N. L. & Panchbhavi, V. K.
Torsional Failure of Carbon Fiber Composite Plates Versus Stainless Steel Plates for
Comminuted Distal Fibula Fractures. Foot & ankle international 37, 548–553 (2016).
71. Manteghi, S., Mahboob, Z., Fawaz, Z. & Bougherara, H. Investigation of the mechanical
properties and failure modes of hybrid natural fiber composites for potential bone fracture
fixation plates. J Mech Behav Biomed Mater 306–316 (2017).
72. Smith, W. R., Ziran, B. H., Anglen, J. O. & Stahel, P. F. Locking plates: tips and tricks.
JBJS 89, 2298–2307 (2007).
73. Wee, H., Reid, J. & Lewis, G. Parametric and surrogate modeling of internal fixation of
femur fractures demonstrate influence of surgical and patient variables. Annals of
Biomedical Engineering 44, 3719–3749 (2016).
74. Horn, J., Schlegel, U., Krettek, C. & Ito, K. Infection resistance of unreamed solid, hollow
slotted and cannulated intramedullary nails: An in-vivo experimental comparison. Journal
of Orthopaedic Research 23, 810–815 (2005).
75. Wehner, T., Penzkofer, R., Augat, P., Claes, L. & Simon, U. Improvement of the shear
fixation stability of intramedullary nailing. Clin Biomech (Bristol, Avon) 26, 147–151
(2011).
76. Mahar, A. T., Lee, S. S., Lalonde, F. D., Impelluso, T. & Newton, P. O. Biomechanical
comparison of stainless steel and titanium nails for fixation of simulated femoral fractures.
Journal of Pediatric Orthopaedics 24, 638–641 (2004).
77. Wang, C. J., Brown, C. J., Yettram, A. L. & Procter, P. Intramedullary nails: some design
features of the distal end. Med Eng Phys 25, 789–794 (2003).
78. Gallagher, D. et al. Is distal locking necessary? A biomechanical investigation of
intramedullary nailing constructs for intertrochanteric fractures. Journal of Orthopaedic
Trauma 27, 373–378 (2013).
79. Simpson, D. J., Brown, C. J., Yettram, A. L., Procter, P. & Andrew, G. J. Finite element
analysis of intramedullary devices: the effect of the gap between the implant and the bone.
Proc Inst Mech Eng H 222, 333–345 (2008).
80. Penzkofer, R. et al. Influence of intramedullary nail diameter and locking mode on the
stability of tibial shaft fracture fixation. Archives of Orthopaedic and Trauma Surgery 129,
525–531 (2009).
81. Cheung, G., Zalzal, P., Bhandari, M., Spelt, J. . & Papini, M. Finite element analysis of a
femoral retrograde intramedullary nail subject to gait loading. Medical Engineering &
Physics 26, 93–108 (2004).
82. Burge, R. et al. Incidence and Economic Burden of Osteoporosis-Related Fractures in the
United States, 2005-2025. Journal of Bone and Mineral Research 22, 465–475 (2007).
83. Swart, E., Makhni, E. C., Macaulay, W., Rosenwasser, M. P. & Bozic, K. J. Cost-
effectiveness analysis of fixation options for intertrochanteric hip fractures. J Bone Joint
Surg Am 96, 1612–1620 (2014).
84. American Academy of Orthopaedic Surgeons et. al. Unified Orthopaedic Research Agenda.
(2016). Available at: http://www.aaos.org/research/tools/ura/. (Accessed: 21st June 2018)
85. Epari, D. R. Timely Fracture-Healing Requires Optimization of Axial Fixation Stability.
The Journal of Bone and Joint Surgery (American) 89, 1575 (2007).
86. Bottlang, M. et al. Biomechanical concepts for fracture fixation. Journal of orthopaedic
trauma 29, S28–S33 (2015).
87. Seinsheimer, F. Subtrochanteric Fractures of the Femur. J. Bone Jt. Surg. Am. 3, 300–306
(1978).
88. AO Foundation. Revised AO/OTA fracture and dislocation classification compendium.
(2018). Available at: https://classification.aoeducation.org/. (Accessed: 21st June 2018)
89. Kneifel, T. & Buckley, R. A comparison of one versus two distal locking screws in tibial
fractures treated with unreamed tibial nails: a prospective randomized clinical trial. Injury
27, 271–273 (1996).
90. Bhat, A., Rao, S. & Bhaskaranand, K. Mechanical Failure in Intramedullary Interlocking
Nails. Journal of Orthopaedic Surgery 14, 138–141 (2006).
91. Griza, S., Zimmer, C. G., Reguly, A. & Strohaecker, T. R. A case study of subsequential
intramedullary nails failure. Engineering Failure Analysis 16, 728–32 (2009).
92. Wu, C.-C. & Shih, C.-H. Biomechanical analysis of the mechanism of interlocking nail
failure. Arch Orthop Trauma Surg 111, 268–272 (1992).
93. Santolini, E., West, R. & Giannoudis, P. V. Risk factors for long bone fracture non-union: a
stratification approach based on the level of the existing scientific evidence. Injury 46, S8–
S19 (2015).
94. Sadiq, S. K. et al. Patient-specific simulation as a basis for clinical decision-making.
Philosophical Transactions of the Royal Society A: Mathematical, Physical and
Engineering Sciences 366, 3199–3219 (2008).
95. Huang, H. et al. Patient-specific geometrical modeling of orthopedic structures with high
efficiency and accuracy for finite element modeling and 3D printing. Australasian Physical
& Engineering Sciences in Medicine 38, 743–753 (2015).
96. Poelert, S., Valstar, E., Weinans, H. & Zadpoor, A. A. Patient-specific finite element
modeling of bones. Proceedings of the Institution of Mechanical Engineers, Part H:
Journal of Engineering in Medicine 227, 464–478 (2013).
97. Bayoglu, R. & Okyar, A. F. Implementation of boundary conditions in modeling the femur
is critical for the evaluation of distal intramedullary nailing. Medical Engineering & Physics
37, 1053–1060 (2015).
98. Heller, M. O. et al. Determination of muscle loading at the hip joint for use in pre-clinical
testing. Journal of Biomechanics 38, 1155–1163 (2005).
99. Morlock, M. et al. Duration and frequency of every day activities in total hip patients.
Journal of Biomechanics 34, 873–881 (2001).
100. Bergmann, G. et al. Hip contact forces and gait patterns from routine activities. J Biomech
34, 859–871 (2001).
101. Duda, G. N. et al. Influence of muscle forces on femoral strain distribution. J Biomech 31,
841–846 (1998).
102. Speirs, A. D., Heller, M. O., Duda, G. N. & Taylor, W. R. Physiologically based boundary
conditions in finite element modelling. Journal of Biomechanics 40, 2318–2323 (2007).
103. Steiner, J. A., Ferguson, S. J. & van Lenthe, G. H. Computational analysis of primary
implant stability in trabecular bone. J Biomech 48, 807–815 (2015).
104. Zhang, Y.-K. et al. Biomechanical effect of the configuration of screw hole style on locking
plate fixation in proximal humerus fracture with a simulated gap: A finite element analysis.
Injury 47, 1191–1195 (2016).
105. Inzana, J. A., Varga, P. & Windolf, M. Implicit modeling of screw threads for efficient
finite element analysis of complex bone-implant systems. J Biomech 49, 1836–1844 (2016).
106. Calderón, A. et al. Proximal femoral intramedullary nail versus DHS plate for the treatment
of intertrochanteric fractures. A prospective analysis. Acta Ortop Mex 27, 236–239 (2013).
107. Canadian Orthopaedic Trauma Society. Reamed versus unreamed intramedullary nailing of
the femur: comparison of the rate of ARDS in multiple injured patients. J Orthop Trauma
20, 384–387 (2006).
108. Charles-Harris, M., Lacroix, D., Proubasta, I. & Planell, J. A. Intramedullary nails vs
osteosynthesis plates for femoral fracture stabilization: A finite element analysis. J Appl
Biomater Biomech 3, 157–167 (2005).
109. Chen, S.-H., Chiang, M.-C., Hung, C.-H., Lin, S.-C. & Chang, H.-W. Finite element
comparison of retrograde intramedullary nailing and locking plate fixation with/without an
intramedullary allograft for distal femur fracture following total knee arthroplasty. Knee 21,
224–231 (2014).
110. Salas, C., Mercer, D., DeCoster, T. A. & Reda Taha, M. M. Experimental and probabilistic
analysis of distal femoral periprosthetic fracture: a comparison of locking plate and
intramedullary nail fixation. Part B: probabilistic investigation. Comput Methods Biomech
Biomed Engin 14, 175–182 (2011).
111. Salas, C., Mercer, D., DeCoster, T. A. & Reda Taha, M. M. Experimental and probabilistic
analysis of distal femoral periprosthetic fracture: a comparison of locking plate and
intramedullary nail fixation. Part A: experimental investigation. Comput Methods Biomech
Biomed Engin 14, 157–164 (2011).
112. Goffin, J. M., Pankaj, P. & Simpson, A. H. The importance of lag screw position for the
stabilization of trochanteric fractures with a sliding hip screw: a subject-specific finite
element study. J. Orthop. Res. 31, 596–600 (2013).
113. Geller, J. A., Saifi, C., Morrison, T. A. & Macaulay, W. Tip-apex distance of
intramedullary devices as a predictor of cut-out failure in the treatment of peritrochanteric
elderly hip fractures. Int Orthop 34, 719–722 (2010).
114. Gazzotti, G. et al. Causes and treatments of lag screw’s cut out after intramedullary nailing
osteosinthesis for trochanteric fractures. Acta Biomed 85, 135–143 (2014).
115. Matityahu, A. et al. The variable angle hip fracture nail relative to the gamma 3: a finite
element analysis illustrating the same stiffness and fatigue characteristics. Adv Orthop 2013,
143801 (2013).
116. Eberle, S., Bauer, C., Gerber, C., von Oldenburg, G. & Augat, P. The stability of a hip
fracture determines the fatigue of an intramedullary nail. Proc Inst Mech Eng H 224, 577–
584 (2010).
117. Eberle, S., Gerber, C., von Oldenburg, G., Hungerer, S. & Augat, P. Type of hip fracture
determines load share in intramedullary osteosynthesis. Clin. Orthop. Relat. Res. 467,
1972–1980 (2009).
118. Haddad, F. S., Desai, K., Sarkar, J. S. & Dorrell, J. H. The AO unreamed nail: friend or foe.
Injury 27, 261–263 (1996).
119. Boenisch, U. W., de Boer, P. G. & Journeaux, S. F. Unreamed intramedullary tibial nailing-
-fatigue of locking bolts. Injury 27, 265–270 (1996).
120. Schemitsch, E. H., Kowalski, M. J., Swiontkowski, M. F. & Senft, D. Cortical bone blood
flow in reamed and unreamed locked intramedullary nailing: a fractured tibia model in
sheep. J Orthop Trauma 8, 373–382 (1994).
121. Gómez-Benito, M. J. et al. Computational comparison of reamed versus unreamed
intramedullary tibial nails. J. Orthop. Res. 25, 191–200 (2007).
122. Eberle, S., Gerber, C., von Oldenburg, G., Högel, F. & Augat, P. A Biomechanical
Evaluation of Orthopaedic Implants for Hip Fractures by Finite Element Analysis and In-
Vitro Tests. Proceedings of the Institution of Mechanical Engineers, Part H: Journal of
Engineering in Medicine 224, 1141–1152 (2010).
123. Gabarre, S. et al. Influence of gap size, screw configuration, and nail materials in the
stability of anterograde reamed intramedullary nail in femoral transverse fractures. Injury
48, S40–S46 (2017).
124. Rodrigues, L. B. et al. A finite element model to simulate femoral fractures in calves:
testing different polymers for intramedullary interlocking nails. Vet Surg 41, 838–844
(2012).
125. Gallagher, D. et al. Is distal locking necessary? A biomechanical investigation of
intramedullary nailing constructs for intertrochanteric fractures. Journal of Orthopaedic
Trauma 27, 373–378 (2013).
126. Caiaffa, V. et al. Is distal locking with short intramedullary nails necessary in stable
pertrochanteric fractures? A prospective, multicentre, randomised study. Injury 47 Suppl 4,
S98–S106 (2016).
127. Dillard, S. I., Mousel, J. A., Shrestha, L., Raghavan, M. L. & Vigmostad, S. C. From
medical images to flow computations without user-generated meshes: IMAGE-TO-
COMPUTATION WITHOUT USER-GENERATED MESHES. International Journal for
Numerical Methods in Biomedical Engineering 30, 1057–1083 (2014).
128. Archip, N., Rohling, R., Dessenne, V., Erard, P.-J. & Nolte, L. P. Anatomical structure
modeling from medical images. Computer Methods and Programs in Biomedicine 82, 203–
215 (2006).
129. Magland, J. F., Zhang, N., Rajapakse, C. S. & Wehrli, F. W. Computationally-Optimized
Bone Mechanical Modeling from High-Resolution Structural Images. PLoS ONE 7, e35525
(2012).
130. Liu, B. et al. An automatic segmentation system of acetabulum in sequential CT images for
the personalized artificial femoral head design. Computer Methods and Programs in
Biomedicine 127, 318–335 (2016).
131. Baka, N. et al. 2D–3D shape reconstruction of the distal femur from stereo X-ray imaging
using statistical shape models. Med Imag Anal 15, 840–850 (2011).
132. Youn, K., Park, M. S. & Lee, J. Iterative approach for 3D reconstruction of the femur from
un-calibrated 2D radiographic images. Med Eng Phys Epub ahead of print, (2017).
133. Klein, K. F., Hu, J., Reed, M. P., Hoff, C. N. & Rupp, J. D. Development and Validation of
Statistical Models of Femur Geometry for Use with Parametric Finite Element Models.
Annals of Biomedical Engineering 43, 2503–2514 (2015).
134. Klein, K. F., Hu, J., Reed, M. P., Schneider, L. W. & Rupp, J. D. Validation of a parametric
finite element human femur model. Traffic Inj Prev 18, 420–426 (2017).
135. Nicolella, D. P. & Bredbenner, T. L. Development of a parametric finite element model of
the proximal femur using statistical shape and density modelling. Computer Methods in
Biomechanics and Biomedical Engineering 15, 101–110 (2012).
136. Prendergast, P. J. Finite element models in tissue mechanics and orthopaedic implant
design. Clin Biomech 12, 343–366 (1997).
137. Beaupre, G. S. & Hayes, W. C. Finite element analysis of a three-dimensional open-celled
model for trabecular bone. J Biomech Eng 107, 249–256 (1985).
138. Blanchard, R., Dejaco, A., Bongaers, E. & Hellmich, C. Intravoxel bone micromechanics
for microCT-based finite element simulations. J Biomech 46, 2710–21 (2013).
139. Hollister, S. J., Brennan, J. M. & Kikuchi, N. A homogenisation sampling procedure for
calculating trabecular bone effective stiffness and tissue level stress. J Biomech 27, 433–
444 (1994).
140. Kaminsky, J. et al. A universal algorithm for an improved finite element mesh generation.
Medical Engineering & Physics 27, 383–394 (2005).
141. Prendergast, P. J. & Huiskes, R. Microdamage and osteocyte-lacuna strain in bone: a
microstructural finite element analysis. J Biomech Eng 118, 240–246 (1996).
142. Mahanian, S. & Piziali, R. L. Finite element evaluation of the AIA shear specimen for bone.
J Biomech 21, 347–356 (1988).
143. Hollister, S. J., Fyhrie, D. P., Jepsen, K. J. & Goldstein, S. A. Application of
homogenisation theory to the study of trabecular bone mechanics. J Biomech 24, 825–839
(1991).
144. Halloran, J. P., Sibole, S. C. & Erdemir, A. The potential for intercellular mechanical
interaction: simulations of single chondrocyte versus anatomically based distribution.
Biomechanics and Modeling in Mechanobiology (2017). doi:10.1007/s10237-017-0951-1
145. Suntay, W. J., Grood, E. S., Noyes, F. R. & Butler, D. L. A coordinate system for
describing joint position. Adv in Bioeng 59–62 (1978).
146. Grood, E. S. & Suntay, W. J. A joint coordinate system for the clinical description of three-
dimensional motions: application to the knee. Journal of biomechanical engineering 105,
136–144 (1983).
147. Dabirrahmani, D. & Hogg, M. Modification of the Grood and Suntay Joint Coordinate
System equations for knee joint flexion. Medical Engineering & Physics 39, 113–116
(2017).
148. MacWilliams, B. A. & Davis, R. B. Addressing some misperceptions of the joint coordinate
system. J Biomech Eng 135, 1–3 (2013).
149. Tucker, S., Reid, J. & Lewis, G. Fracture Fixation Biomechanics and Biomaterials. in
Orthopedic Biomaterials: Progress in biology, manufacturing, and industry perspectives
401–428 (Springer, 2018).
150. Wee, H., Reid, J. S., Chinchilli, V. M. & Lewis, G. S. Finite Element-Derived Surrogate
Models of Locked Plate Fracture Fixation Biomechanics. Annals of Biomedical Engineering
45, 668–680 (2017).
151. Steiner, M., Claes, L., Ignatius, A., Simon, U. & Wehner, T. Disadvantages of
interfragmentary shear on fracture healing--mechanical insights through numerical
simulation. J. Orthop. Res. 32, 865–872 (2014).
152. Çelik, A., Kovacı, H., Saka, G. & Kaymaz, İ. Numerical investigation of mechanical effects
caused by various fixation positions on a new radius intramedullary nail. Comput Methods
Biomech Biomed Engin 18, 316–324 (2015).
153. Perez, A., Mahar, A., Negus, C., Newton, P. & Impelluso, T. A computational evaluation of
the effect of intramedullary nail material properties on the stabilization of simulated femoral
shaft fractures. Medical Engineering & Physics 30, 755–760 (2008).
154. Mahar, A. T., Lee, S. S., Lalonde, F. D., Impelluso, T. & Newton, P. O. Biomechanical
comparison of stainless steel and titanium nails for fixation of simulated femoral fractures.
Journal of Pediatric Orthopaedics 24, 638–641 (2004).
155. Penzkofer, R. et al. Influence of intramedullary nail diameter and locking mode on the
stability of tibial shaft fracture fixation. Archives of Orthopaedic and Trauma Surgery 129,
525–531 (2009).
156. Duda, G. N. et al. Influence of muscle forces on femoral strain distribution. J Biomech 31,
841–846 (1998).
157. Lu, T. W., Taylor, S. J., O’Connor, J. J. & Walker, P. S. Influence of muscle activity on the
forces in the femur: an in vivo study. J Biomech 30, 1101–1106 (1997).
158. Cristofolini, L., Viceconti, M., Cappello, A. & Toni, A. Mechanical validation of whole
bone composite femur models. J Biomech 29, 525–535 (1996).
159. Bergmann, G., Graichen, F. & Rohlmann, A. Is staircase walking a risk for the fixation of
hip implants? Journal of Biomechanics 28, 535–553 (1995).
160. Morrison, J. The mechanics of the knee joint in relation to normal walking. J Biomech 3,
51–61 (1970).
161. Märdian, S., Schaser, K.-D., Duda, G. N. & Heyland, M. Working length of locking plates
determines interfragmentary movement in distal femur fractures under physiological
loading. Clin Biomech (Bristol, Avon) 30, 391–396 (2015).
162. Perren, S. M. Physical and biological aspects of fracture healing with special reference to
internal fixation. Clin. Orthop. Relat. Res. 175–196 (1979).
163. Mahar, A., Sink, E., Faro, F., Oka, R. & Newton, P. O. Differences in biomechanical
stability of femur fracture fixation when using titanium nails of increasing diameter. J Child
Orthop 1, 211–215 (2007).
164. Brumback, R. J. & Virkus, W. W. Intramedullary Nailing of the Femur: Reamed Versus
Nonreamed. Journal of the American Academy of Orthopaedic Surgeons 8, 83 (2000).
165. Tucker, A., Warnock, M., McDonald, S., Cusick, L. & Foster, A. P. Fatigue failure of the
cephalomedullary nail: revision options, outcomes and review of the literature. European
Journal of Orthopaedic Surgery & Traumatology 28, 511–520 (2018).
166. Patel, R. A., Wilson, R. F., Patel, P. A. & Palmer, R. M. The effect of smoking on bone
healing. Bone Joint Res 2, 102–111 (2013).
167. Jiao, H., Xiao, E. & Graves, D. T. Diabetes and Its Effect on Bone and Fracture Healing.
Curr Osteoporos Rep 13, 327–335 (2015).
168. Kreb, D. L. et al. Intramedullary nailing without interlocking screws for femoral and tibial
shaft fractures. Arch Orthop Trauma Surg 133, 1109–1113 (2013).
169. Yamagishi, M. & Yoshimura, Y. The biomechanics of fracture healing. J Bone Joint Surg
Am 37-A, 1035–1068 (1955).
170. Aro, H. T. & Chao, E. Y. Bone-healing patterns affected by loading, fracture fragment
stability, fracture type, and fracture site compression. Clin. Orthop. Relat. Res. 8–17 (1993).
171. Sarmiento, A., Gersten, L. M., Sobol, P. A., Shankwiler, J. A. & Vangsness, C. T. Tibial
shaft fractures treated with functional braces. Experience with 780 fractures. J Bone Joint
Surg Br 71, 602–609 (1989).
172. McKellop, H., Hoffmann, R., Sarmiento, A. & Ebramzadeh, E. Control of motion of tibial
fractures with use of a functional brace or an external fixator. A study of cadavera with use
of a magnetic motion sensor. J Bone Joint Surg Am 75, 1019–1025 (1993).
173. Greenhill, D. A., Poorman, M., Pinkowski, C., Ramsey, F. V. & Haydel, C. Does weight-
bearing assignment after intramedullary nail placement alter healing of tibial shaft
fractures? Orthop Traumatol Surg Res 103, 111–114 (2017).
174. Heimann, T. & Meinzer, H.-P. Statistical shape models for 3D medical image segmentation:
A review. Medical Image Analysis 13, 543–563 (2009).
175. Hrabe, N., Gnäupel-Herold, T. & Quinn, T. Fatigue properties of a titanium alloy (Ti–6Al–
4V) fabricated via electron beam melting (EBM): Effects of internal defects and residual
stress. International Journal of Fatigue 94, 202–210 (2017).
176. Lopez, C., Tucker, S., Salameh, T. & Tucker, C. An unsupervised machine learning method
for discovering patient clusters based on genetic signatures. J Biomed Inform 85, 30–39
(2018).
177. Leung, M. K. K., Delong, A., Alipanahi, B. & Frey, B. J. Machine Learning in Genomic
Medicine: A Review of Computational Problems and Data Sets. Proceedings of the IEEE
104, 176–197 (2016).
178. Libbrecht, M. W. & Noble, W. S. Machine learning in genetics and genomics. Nature
Reviews. Genetics 16, 321–332 (2015).
179. Upstill-Goddard, R., Eccles, D., Fliege, J. & Collins, A. Machine learning approaches for
the discovery of gene-gene interactions in disease data. Briefings in Bioinformatics 14, 251–
260 (2013).
180. Yip, K. Y., Cheng, C. & Gerstein, M. Machine learning and genome annotation: a match
meant to be? Genome Biology 14, 205 (2013).
181. Ross, C. J. et al. A pharmacogenetic signature of high response to Copaxone in late-phase
clinical-trial cohorts of multiple sclerosis. Genome Medicine 9, (2017).
182. Kulakova, O. et al. Comparative pharmacogenetics of multiple sclerosis: INF-B versus
glatiramer acetate. Pharmacogenomics 15, 679–85 (2014).
183. Xu, W., Zhang, L. & Lu, Y. SD-MSAEs: Promoter recognition in human genome based on
deep feature extraction. Journal of Biomedical Informatics 61, 55–62 (2016).
184. Zhao, Y. et al. Exploration of machine learning techniques in predicting multiple sclerosis
disease course. PloS one 12, e0174866 (2017).
185. Goldstein, B. A., Hubbard, A. E., Cutler, A. & Barcellos, L. F. An application of Random
Forests to a genome-wide association dataset: Methodological considerations & new
findings. BMC Genetics 11, 49 (2010).
186. Kimes, P. K., Liu, Y., Neil Hayes, D. & Marron, J. S. Statistical significance for
hierarchical clustering. Biometrics (2017). doi:10.1111/biom.12647
187. Farh, K. K.-H. et al. Genetic and epigenetic fine mapping of causal autoimmune disease
variants. Nature 518, 337–343 (2015).
188. Lim, S., Tucker, C. S. & Kumara, S. An unsupervised machine learning model for
discovering latent infectious diseases using social media data. J Biomed Informat. 66, 82–94
(2017).
189. Xu, R. & Wunsch, D. C. Clustering algorithms in biomedical research: A review. IEEE
Reviews in Biomedical Engineering 3, 120–154 (2010).
190. Prelić, A. et al. A systematic comparison and evaluation of biclustering methods for gene
expression data. Bioinformatics 9, 1122–1129 (2006).
191. Jain, A. K., Murty, M. N. & Flynn, P. J. Data clustering: a review. ACM Computing Surveys
31, 264–323 (1999).
192. Pagnuco, I. A., Juan I. Pastorea;, Abras;, G., Brun;, M. & Ballarin;, V. L. Analysis of
genetic association using hierarchical clustering and cluster validation indices. Genomics 4–
11 (2017). doi:10.1016/j.pscychresns.2008.11.004
193. Ahmad, T. et al. Clinical implications of cluster analysis-based classification of acute
decompensated heart failure and correlation with bedside hemodynamic profiles. PloS one
11, e0145881 (2016).
194. Mei, K., Peng, J., Gao, L., Zheng, N. N. & Fan, J. Hierarchical Classification of Large-
Scale Patient Records for Automatic Treatment Stratification. IEEE Journal of Biomedical
and Health Informatics 19, 1234–1245 (2015).
195. Hamid, J. S., Meaney, C., Crowcroft, N. S., Granerod, J. & Beyene, J. Cluster analysis for
identifying sub-groups and selecting potential discriminatory variables in human
encephalitis. BMC infectious diseases 10, 364 (2010).
196. Redekop, W. & Mladsi, D. The faces of personalized medicine: a framework for
understanding its meaning and scope. Value in Health 6, S4–S9 (2013).
197. Tari, L., Baral, C. & Kim, S. Fuzzy c-means clustering with prior biological knowledge.
Journal of Biomedical Informatics 42, 74–81 (2009).
198. Bellazzi, R. & Zupan, B. Towards knowledge-based gene expression data mining. Journal
of Biomedical Informatics 40, 787–802 (2007).
199. Wiwie, C., Baumbach, J. & Röttger, R. Comparing the performance of biomedical
clustering methods. Nature Methods 12, 1033–1038 (2015).
200. Ultsch, A. & Lötsch, J. Machine-learned cluster identification in high-dimensional data.
Journal of Biomedical Informatics 66, 95–104 (2017).
201. Ultsch, A. Clustering with SOM: U*C. in In Proceedings of the 5th Workshop on Self-
Organizing Maps 75–82 (2005).
202. Clifford, H., Wessely, F., Pendurthi, S. & Emes, R. D. Comparison of clustering methods
for investigation of genome-wide methylation array data. Frontiers in Genetics 2, 1–11
(2011).
203. Rousseeuw, P. J. Silhouettes: A graphical aid to the interpretation and validation of cluster
analysis. Journal of Computational and Applied Mathematics 20, 53–65 (1987).
204. Bezdek, J. C. & Pal, N. R. Some new indexes of cluster validity. IEEE Transactions on
Systems, Man, and Cybernetics, Part B: Cybernetics 28, 301–315 (1998).
205. Sakellariou, A., Sanoudou, D. & Spyrou, G. Combining multiple hypothesis testing and
affinity propagation clustering leads to accurate, robust and sample size independent
classification on gene expression data. BMC Bioinformatics 13, 270 (2012).
206. Frey, B. J. & Dueck, D. Clustering by passing messages between data points. Science 315,
972–976 (2007).
207. Shen, R., Olshen, A. B. & Ladanvi, M. Integrative clustering of multiple genomic data
types using a joint latent variable model with application to breast and lung cancer subtype
analysis. Bioinformatics 25, 2906–2912 (2009).
208. Dempster, A. P., Laird, N. M. & Rubin, D. B. Maximum likelihood from incomplete data
via the EM algorithm. J Royal Stat Soc Series B 1–38 (1977).
209. Fraley, C. & Raftery, A. E. How many clusters? Which clustering method? Answers via
model-based cluster analysis. The Comp Journal 41, 578–588 (1998).
210. Khakabimamaghani, S. & Ester, M. Bayesian biclustering for patient stratification.
Biocomputing 2016: Proceedings of the Pacific Symposium 345–356 (2016).
211. Lee, D. & Seung, H. Algorithms for non-negative matrix factorization. Advances in neural
information processing systems 556–562 (2001).
212. Pontes, B., Giráldez, R. & Aguilar-Ruiz, J. Biclustering on expression data: A review.
Journal of biomedical informatics 57, 163–180 (2015).
213. Cheng, Y. & Church, G. Biclustering of expression data. Proceedings of the 8th
International Conference on Intelligent Systems for Molecular Biology, La Jolla, CA 93–
103 (2000).
214. Nguyen, N. & Caruana, R. Consensus clusterings. in Proceedings - IEEE International
Conference on Data Mining, ICDM 607–612 (2007). doi:10.1109/ICDM.2007.73
215. Suzuki, R. & Shimodaira, H. pvclust : An R package for hierarchical clustering with p-
values. Bioinformatics 22, 1–7 (2013).
216. Bushel, P. R. et al. Computational selection of distinct class- and subclass-specific gene
expression signatures. Journal of Biomedical Informatics 35, 160–170 (2002).
217. Cabin, R. J. & Mitchell, R. J. To Bonferroni or not to Bonferroni: when and how are the
questions. Bulletin of the Ecological Society of America 81, 246–248 (2000).
218. Bezdek, J. C. & Ehrlich, R. FCM: The fuzzy c-means clustering algorithm. Comp and
Geosci 10, 191–203 (1984).
219. Ashburner, M. et al. Gene Ontology: tool for the unification of biology. Nature Genetics 25,
25–29 (2000).
220. Yazdani, A., Yazdani, A., Samiei, A. & Boerwinkle, E. Generating a robust statistical
causal structure over 13 cardiovascular disease risk factors using genomics data. Journal of
Biomedical Informatics 60, 114–119 (2016).
221. Purcell, S. et al. PLINK: A Tool Set for Whole-Genome Association and Population-Based
Linkage Analyses. The American Journal of Human Genetics 81, 559–575 (2007).
222. Rédei, G. P. Kruskal-Wallis test. Encyclopedia of Genetics, Genomics, Proteomics, and
Informatics 1067–1068 (2008).
223. Everitt, B. S., Landau, S., Leese, M. & Stahl, D. Measurement of Proximity. Cluster
Analysis 43–69 (2011). doi:10.1002/9780470977811.ch3
224. Bazaraa, M. S., Sherali, H. D. & Shetty, C. M. Nonlinear programming: theory and
algorithms. (John Wiley & Sons, 2013).
225. Szklarczyk et al. STRING v10: protein-protein interaction networks, integrated over the tree
of life. Nucleic Acids Research D1, 447–452 (2015).
226. Rand, W. M. Objective criteria for the evaluation of clustering methods. J Amer Stat Assoc
66, 846–50 (1971).
227. R. R Development Core Team. R: A Language and Environment for Statistical Computing.
1, (2011).
228. Fraley, C., Raftery, A. E., Murphy, T. B. & Scrucca, L. mclust Version 4 for R: Normal
Mixture Modeling for Model-Based Clustering, Classification, and Density Estimation.
Technical Report No. 597, Dept of Statistics, University of Washington (2012).
229. Bodenhofer, U., Kothmeier, A. & Hochreiter, S. APCluster: an R package for affinity
propagation clustering. Bioinformatics 27, 2463–4 (2011).
230. Thrun, M. C., Lerch, F., Lotsch, J. & Ultsch, A. Visualization and 3D printing of
multivariate data of biomarkers. in Proceedings of international conference in Central
Europe on computer graphics, visualization, and computer vision (2016).
231. Thrun, M. C. Projection based clustering through self-organization and swarm intelligence:
combining cluster analysis with the visualization of high-dimensional data. (Springer
Fachmedien, 2018).
232. Kaiser, S. et al. biclust: BiCluster Algorithms. R package version 1 no. 1, (2015).
233. Golub, T. et al. Molecular classification of cancer: class discovery and class prediction by
gene expression monitoring. Science 5439, 531–7 (1999).
234. Monti, S., Tamayo, P., Mesirov, J. & Golub, T. Consensus clustering: a resampling-based
method for class discovery and visualization of gene expression microarray data. Machine
Learning 52, 91–118 (2003).
235. Cortes, A. & Brown, M. A. Promise and pitfalls of the Immunochip. Arthritis research &
therapy 13, 101 (2011).
236. Welter, D. et al. The NHGRI GWAS Catalog, a curated resource of SNP-trait associations.
Nucleic Acids Research 42, D1001–D1006 (2014).
237. Kirkpatrick, S., Gelatt, C. D. & Vecchi, M. P. Optimization by simulated annealing. Science
220, 671–80 (1983).
238. López, C. E. & Nembhard, D. Cooperative workforce planning heuristic with worker
learning and forgetting and demand constraints. in IIE Annual Conference Proceedings
380–85 (2017).
239. Xiang, Y., Gubian, S., Suomela, B. & Hoeng, J. Generalized simulated annealing for global
optimization: the GenSA Package. R J. 5, 13–28 (2013).
240. Bader, G. D., Betel, D. & Hogue, C. W. V. BIND: the Biomolecular Interaction Network
Database. Nucleic Acids Research 31, 248–50 (2003).
241. Chatr-Aryamontri, A. et al. The BioGRID interaction database: 2017 update. Nucleic Acids
Research (2016).
242. Prasad et al, T. S. K. Human Protein Reference Database - 2009 Update. Nucleic Acids
Research D767-72 (2009).
243. Hermjakob, H., Montecchi-Palazzi, L. & Lewington et al, C. IntAct: an open source
molecular interaction database. Nucleic Acids Research D452-5 (2004).
244. Licata, L. et al. MINT, the molecular interaction database: 2012 update. Nucleic Acids
Research (2012).
245. Schaefer, C. F., Anthony, K. & Krupa et al, S. PID: the Pathway Interaction Database.
Nucleic Acids Research D674-9 (2009).
246. Kanehisa, Furumichi, M., Tanabe, M., Sato, Y. & Morishima, K. KEGG: new perspectives
on genomes, pathways, diseases, and drugs. Nucleic Acids Research D353-61 (2017).
247. Nishimura, D. Biotech software and internet report. BioCarta, BIotech Software & Internet
Report: The Computer Software Journal for Scient 2(3) (2004). Available at:
https://doi.org/10.1089/152791601750294344.
248. Fabregat et al. The reactome pathway knowledgebase. Nucleic Acids Research D1, 481–487
(2015).
249. Larochelle, C., Alvarez, J. & Just, W. How do immune cells overcome the blood-brain
barrier in multiple sclerosis? FEBS Letters (2012).
250. Isobe, N. & et. al. An Immunochip study of multiple sclerosis risk in African Americans.
Brain 138, 1518–30 (2015).
Vita - SCOTT M. TUCKER
EDUCATION
Penn State University College of Medicine, Hershey, PA
MD/PhD Student 2013-Present
Penn State University, University Park, PA
PhD, Engineering Science and Mechanics 2015-2019
Cornell University, Ithaca, NY
MEng, Biomedical Engineering 2011
Cornell University, Ithaca, NY
BS, Biological and Environmental Engineering 2010
AWARDS AND HONORS
Penn State Fund for Medical Innovation Grant 2019-2020
I-CORPS Regional Short Course for Medical Innovation Spring 2019
Student Award for Excellence in Innovation 2018
NSF Center for Healthcare Organization Transformation Scholar 2017-2018
International Section for Fracture Repair Poster Award 2018
Daniel A. Notterman Physician Scientist Award 2018
Penn State Pediatric Research Day Young Investigator Award 2018
Penn State Cancer Institute Abstract Top 15 Finalist 2018
Finalist, Herodicus Award for Best Resident Paper (Moira McCarthy, MD) 2012
Cornell Tradition Fellow 2006-2010
Stephen Phillip Memorial Scholar 2006-2010
Northeast Atlantic BioEngineering Conference Student Design Competition 2010
Alpha Zeta Honors Fraternity National Officer of the Year 2008
REPRESENTATIVE PUBLICATIONS
Book Chapter
Tucker SM, Reid, JS, Lewis, GS, Fracture Fixation Biomechanics and Biomaterials,
Springer, 15, 401-26 (2018).
Peer Reviewed Manuscripts
1. Tucker SM, Wee HW, Reid S, Lewis G. Effects of fracture fixation construct and
weightbearing on shear motion in proximal femur fractures. 2019. In preparation
2. Tucker SM, Wee HW, Fox E, Reid S, Lewis G. Parametric finite element analysis of
intramedullary naild fixation of proximal femur fractures. J Orthop Res. 2019. In
review, invited for revisions.
3. Moncal K, Tigli A, Seda R, Abu-laban M, Heo D, Rizk E, Tucker S, Lewis G, Hayes
D, Ozbolat I. miR-148b enriched 3D printed hybrid scaffolds for critical-sized calvarial
bone defect repair. 2019. In review.
4. Smuin D, Tucker SM, Rothermel SD, Lewis G, Mason M. Should we be concerned
about acetabular implant stresses in standard THA and dual mobility constructs with
acetabular bone loss?: A finite element model. 2019. In review.
5. Moncal K, Heo D, Godzik K, Sosnoski D, Mrowczynski O, Rizk E, Ozbolat V, Tucker
SM, Gerhard E, Dey M, Lewis G, Yang J, Ozbolat I. 3D Printing of Poly (ε-
caprolactone)/ Poly (D, L-lactide-co-glycolide)/ Hydroxyapatite composite constructs
for bone tissue engineering. J Mat Res. 2018; 1-15.