big data solutions omics and ehr · • bristol-meyers squibb (paid to institution) • orion...
TRANSCRIPT
Anthony GordonProfessor of Anaesthesia and Critical Care
Imperial College London
NIHR Research Professor
Big data solutions – Omics and EHR
Disclosures / Acknowledgements
Grant funding / support
• NIHR
• Tenax Therapeutics
• Orion Pharma
• HCA International
Consulting / speaker fees
• Ferring Pharmaceuticals (paid to institution)
• Tenax Therapeutics (paid to institution)
• GSK (paid to institution)
• Bristol-Meyers Squibb (paid to institution)
• Orion Pharma
• Amomed Pharma
The views expressed are those of the author(s) and not necessarily those of the NHS, the
NIHR or the Department of Health.
Lessons learnt
1. Need to work with real experts
4
Environment
Proteomics
Metabonomics
Transcriptomics
Pathogen
Host
Genomics
Community acquired pneumonia
Consistent signal across validation cohorts
-lo
g1
0 p
-va
lue
8
6
4
2
Combined cohort
Lessons learnt
1. Need to work with real experts
2. Genomics
– (generally) need large numbers
– pathogen may be important
Metabonomics
Chemical shift Mass to charge ratio
NMR Mass Spec
Benefits
• Large data set
• Very sensitive
• Final downstream marker
Metabonomics
Disadvantages
• Large data set
• Very sensitive
• Final downstream marker
10
Serum - NMR
R2Y = 0.91, Q2Y = 0.28, p = 0.02
Ventilator associated pneumonia
VAP
Brain injury
R2Y = 0.94, Q2Y = 0.27, p = 0.05
AUROC = 0.91
Lessons learnt
1. Need to work with real experts
2. Genomics
– (generally) need large numbers
– pathogen may be important
3. Metabonomics
– promising but challenging
Unsupervised hierarchical cluster analysis Principal component analysis
41%
59%
Derivation cohort Validation cohortSRS1
SRS2
7 genes could predict SRS group membership - 3.8% misclassification
Unsupervised hierarchical cluster analysis Principal component analysis
Interaction p = 0.02
SRS1 SRS2
HC Placebo HC Placebo P-value
interaction
Time to shock
reversal (hrs)
30.6
(18.1 – 77.7)
43·8
(21·5 - 91·5)
58·9
(36·1 - 82·3)
89·5
(31·5 - 122·0)0.60
Possible explanation? – downregulation of MHC II by steroids
Lessons learnt
1. Need to work with real experts
2. Genomics
– (generally) need large numbers
– pathogen may be important
3. Metabonomics
– promising but challenging
4. Transcriptomics
– ready for use in RCTs?
Using clinical data - EHR
Reinforcement learning
Category of artificial intelligence
A virtual agent learns from trial and error an optimized set of rules – a
policy - that maximises a return / reward
Similar to a clinician?
But
Doesn’t suffer from recall bias
Can learn from huge numbers
We can select a delayed reward
Markov Decision Process
A general framework used for modelling sequential decision
making.
Most useful in problems involving complex, stochastic and
dynamic decisions, for which they can find optimal
solutions.
Environment
Agent
Markov Decision Process
Clinicianpolicy π
Patient
action, astate, s reward, r
= prescription= patient condition = mortality risk
Q(s,a): the state-action
value function: value
action a taken in patient
state s
1. Evaluate clinicians’ policy
2. Learn optimal policy
Markov Decision Process
• Defined by 𝑆, 𝐴, 𝑇, 𝑅• 𝑆: a finite set of states
• 𝐴: a finite set of actions
• 𝑇 𝑠𝑡+1, 𝑠𝑡 , 𝑎𝑡 : the transition matrix
• 𝑅: the immediate reward = {-100, +100}
Medical Information Mart for Intensive Care version III (MIMIC-III)
openly available dataset
developed by the MIT Lab for Computational Physiology
deidentified health data
~60,000 critical care patients (2001-2012, 6 ICUs Beth Israel Deaconess Medical
Center)
Includes
– demographics
– vital signs
– laboratory tests
– medications
– outcomes (Social Security Administration Death Master File)
Suspected Sepsis-3 criteria
Antibiotic + micro sample
Increase SOFA >2
States – 48 variables
Category Items Type
Demographics Age
Gender
Weight
Readmission to ICU
Elixhauser score
(premorbid status)
Cont
Binary
Cont.
Binary
Cont.
Vital signs Modified SOFA*
SIRS
GCS
Heart rate,
systolic, mean and
diastolic BP,
shock index
Respiratory rate, SpO2
Temperature
Cont.
Cont.
Cont.
Cont.
Cont.
Cont.
Cont.
Cont.
Category Items Type
Lab values Potassium, sodium, chloride
Glucose, BUN, creatinine
Magnesium, calcium, ionized calcium,
carbon dioxide
SGOT, SGPT, total bilirubin, albumin
Hemoglobin
White blood cells count, platelets
count, PTT, PT, INR
pH, PaO2, PaCO2, base excess,
bicarbonate, lactate, PaO2/FiO2 ratio
Cont.
Cont.
Cont.
Cont.
Cont.
Cont.
Cont.
Cont.
Cont.
Cont.
Ventilation
parameters
Mechanical ventilation
FiO2
Binary
Cont.
Medications
and fluid
balance
Current IV fluid intake over 4h
Max dose of vasopressor over 4h
Urine output over 4h
Cumulated fluid balance since
admission (includes preadmission
data when available)
Cont.
Cont.
Cont.
Cont.
Outcome Hospital mortality
90-day mortality
Binary
Binary
States
4 hour windows
K-means clustering
750 + 2 states
(death, successful discharge)
Actions (treatments)
Vasopressors
norepinephrine equivalents (Brown et al, Chest 2013)
Fluids
boluses & background infusions
crystalloids, colloids and blood products,
normalised by tonicity (Waechter et al. CCM, 2014)
5 possible dose ranges for each,
zero drug given
Four quartiles of actual doses given
Actions (treatments)
Discretized
action
IV fluids (mL in 4 hours) Vasopressors (µg/kg/min)
Range Median dose Range Median dose
1 0 0 0 0
2 ]0-50] 30 ]0-0.08] 0.045
3 ]50-180] 86 ]0.08-0.22] 0.135
4 ]180-530] 324 ]0.22-0.45] 0.30
5 >530 974 >0.45 0.90
Trajectories
Reinforcement Learning algorithms
1.Clinicians’ policy evaluation
2.Optimal policy estimation
Is the AI policy better?
Internal validation in MIMIC-III – 20%
Off-policy policy evaluation
Clinicians’ policy value Model-based AI policy value
90-day mortality 53.7 (53.0-55.2) 82.3 (81.8-82.7)
Clinicians’ policy value IS-based AI policy value
90-day mortality 51.9 (50.7-53.1) 87.7 (85.2-88.9)
Independent external validation
Philips Research Institute - eICU database
20,000 patients publicly available
3.3M in total!!
459 ICUs in USA
2008 – 2016
Similar data to MIMIC-III but data quality “variable”
Independent validation
Clinicians’ policy value IS-based AI policy value
Hospital mortality 56.9 (54.7 – 58.8) 84.5 (84.3 – 87.7)
Independent validation
Comparing policies
17% of patients actually received vasopressors
AI recommended vasopressors for 30%
When doses differed
Equal proportion given too much / too little fluid
– On average too much fluid (~80ml/h)
75% given too low vasopressor dose
– median dose deficit 0.13 µg/kg/min
Comparing policies
Next steps
Real-time analysis – as decision support system
Prospective testing
• “hidden”
• RCT
Further AI policy development
@agordonICU
What determined the policies?