c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15

j o ur nal homep age : w ww.int l .e lsev ierhea l th .com/ journa ls /cmpb

An integrated computer-controlled system for assistingresearchers in cortical excitability studies by usingtranscranial magnetic stimulation

D. Giordanoa,∗, I. Kavasidisa, C. Spampinatoa, R. Bellab, G. Pennisib, M. Pennisi c

a Department of Electrical, Electronics and Computer Engineering, University of Catania, Viale Andrea Doria 6, 95125 Catania, Italyb Department of Neuroscience, University of Catania, Via Santa Sofia 78, 95123 Catania, Italyc Department of Chemistry, University of Catania, Viale Andrea Doria 6, 95125 Catania, Italy

a r t i c l e i n f o

Article history:

Received 23 March 2011

Received in revised form

13 October 2011

Accepted 17 October 2011

Keywords:

TMS paradigms

Neurodegenerative and

cerebrovascular diseases

Cortical excitability curve

RDF semantic storage

a b s t r a c t

Transcranial magnetic stimulation (TMS) is the most important technique currently avail-

able to study cortical excitability. Additionally, TMS can be used for therapeutic and

rehabilitation purposes, replacing the more painful transcranial electric stimulation (TES).

In this paper we present an innovative and easy-to-use tool that enables neuroscientists to

design, carry out and analyze scientific studies based on TMS experiments for both diagnos-

tic and research purposes, assisting them not only in the practicalities of administering the

TMS but also in each step of the entire study’s workflow. One important aspect of this tool

is that it allows neuroscientists to specify research designs at will, enabling them to define

any parameter of a TMS study starting from data acquisition and sample group definition

to automated statistical data analysis and RDF data storage. It also supports the diagnos-

ing process by using on-line support vector machines able to learn incrementally from the

diseases instances that are continuously added into the system. The proposed system is

Online-SVM

Scientific research workflow

Decision support systems

a neuroscientist-centred tool where the protocols being followed in TMS studies are made

explicit, leaving to the users flexibility in exploring and sharing the results, and providing

assistance in managing the complexity of the final diagnosis. This type of tool can make the

results of medical experiments more easily exploitable, thus accelerating scientific progress.

1. Introduction

Transcranial magnetic stimulation (TMS) is a noninvasive andpainless technique for the evaluation of corticospinal tract

function as well as of motor cortex excitability of the humanbrain and it is used to investigate the central motor path-ways of several neurological and psychiatric diseases. Morespecifically, TMS is the most important technique currently

∗ Corresponding author. Tel.: +39 095 7382371.E-mail addresses: dgiordan@dieei.unict.it (D. Giordano), ikavasidis@

cspampin@dieei.unict.it (C. Spampinato), rbella@unict.it (R. Bella), pen0169-2607/$ – see front matter © 2011 Elsevier Ireland Ltd. All rights resdoi:10.1016/j.cmpb.2011.10.008

© 2011 Elsevier Ireland Ltd. All rights reserved.

available to study cortical excitability [1], and can be usedfor therapeutic and rehabilitation purposes [2,3], replacing themore painful transcranial electric stimulation (TES). In the lasttwenty years, TMS has been applied to explore the pathophys-iology of many neurological and psychiatric diseases [4], such

dieei.unict.it (I. Kavasidis),nigi@unict.it (G. Pennisi), mpennisi@unict.it (M. Pennisi).

as multiple sclerosis [5], stroke [6], dementia [7], Parkinson’sdisease [8], myelopathies [9], depression [10], schizophrenia[11], and as a possible therapeutic tool for some of thesedisorders [9].

erved.

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15 5

Fig. 1 – Magnetic field generated by the different coils: (a) magnetic field by a “figure-of-eight” coil and (b) magnetic field bya “circular” coil.

pgfiee(

apm[

gTmim

1

2

3

4

vtldp

TMS produces a modification of the neuronal activity of therimary motor cortex stimulated by the variable magnetic fieldenerated by a coil placed on the scalp. This variable magneticeld, produced by the current flowing in the coil, induces anlectric current in the underlying brain tissue. The “figure-of-ight” or “butterfly” coil can stimulate a relatively focal areaFig. 1), whereas the circular coil a more diffuse one [12].

When TMS is applied to the primary motor cortex, atppropriate magnetic field intensity, it induces motor evokedotentials (MEP), recorded with an electromyograph, in theuscles that are contralateral to the stimulated motor cortex

13].In clinical practice, TMS may be delivered as either sin-

le or paired pulses or regularly repeating pulses (repetitiveMS) in order to assess different parameters about theotor system. The single pulse TMS is used to evaluate the

ntegrity of motor pathways and motor cortex excitability byeasuring:

the MEP amplitude (defined as the distance between thelowest negative peak and the highest positive peak andexpressed in mV);

the motor threshold (defined as the minimum TMS inten-sity necessary to evoke small-amplitude MEPs, larger than50 �V in amplitude);

the central motor conduction time (i.e. the latency differ-ence between the MEPs induced by stimulation of the motorcortex and those evoked by spinal stimulation);

the cortical silent period (cSP, defined as a period of elec-tromyographic suppression after a MEP).

Usually, the cortical excitability and intracortical circuits inarious diseases are studied by a paired pulse TMS paradigm

hat couples a subthreshold stimulus (the amplitude is setower than the patient motor threshold and it is called a con-itioning pulse) and a suprathreshold stimulus (called a testulse), at different interstimulus intervals (ISIs) through the

same coil. The effects of the conditioning pulse on the sizeof the MEP depend on the duration of the ISIs. Indeed, at ISIswithin the range 1–4 ms there is a strong inhibitory effect onthe MEP (in the form of a reduced amplitude) [14], while at ISIswithin the range 7–20 ms there is a facilitatory effect on theMEP (in the form of increased amplitude) [15].

Since there is an extensive use of TMS in different researchfields and for each use of TMS several different factors arecrucial, a data acquisition and processing system is requiredto create more standardized conditions and to reduce the highintra- and inter-rater variability in the execution of the clinicalexperiments (typically due to coil positioning and to the timeinterval between each pulse administration).

As far as we know, very few software-based approacheshave been proposed for supporting neuroscientists in per-forming TMS experiments. The first attempt was developedin 2000 by Kaelin-Lang and Cohen [16] who tried to help neu-roscientists in the execution of TMS experiments, but thesystem was designed only for data acquisition and for datapost-processing, and not for supporting researchers in thewhole life-cycle of a research study. In order to improve thefunctionalities of this system, we have recently proposed aflexible TMS data acquisition and processing system affordingthe scientists an easy and customizable interaction with theTMS hardware, for more efficient and accurate data record-ing and analysis [17]. In this paper we expand this workby presenting a system that, beyond the customization ofthe TMS experiments, uses machine learning techniques toassist scientists in the diagnosing process. In detail, here wepropose an easy-to-use tool that enables neuroscientists todesign, carry out and analyze scientific studies based on TMSexperiments for both diagnostic and research purposes, andassists neuroscientists in each step of the entire study’s work-

flow. The tool allows neuroscientists to specify any researchdesign, by defining any parameter of a TMS study startingfrom data acquisition to sample group definition to statisticaldata analysis. All the data used in the proposed tool, including

6 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15

S is

in the data storage and experiment management and in thediagnosis. The architecture of the proposed system is shown

Fig. 2 – Example of: (a) MEP response when a paired pulse TM

experiment protocol data, is also stored in RDF, thus they canbe shared with other systems compliant to semantic web stan-dards. Finally, the tool is also provided with on-line supportvector machines (SVM) to help neuroscientists in the diagnosisprocess.

The remainder of the paper is as follows: the next sectionintroduces the signals and the parameters involved in a TMSexperiment. In Section 3 the proposed tool is presented, fol-lowing each step of the workflow carried out by scientistsfor TMS experiments, from hardware interfacing to proto-col definition, to experiment execution, to statistical analysisand RDF data storage. In the same section, the proposed on-line SVM approach for supporting scientists in the diagnosisis described, pointing out its advantages. Finally, concludingremarks are given.

2. Transcranial magnetic stimulation: mainconcepts

As mentioned in Section 1, TMS may be administered aseither single or paired pulses or regularly repeating pulses(repetitive TMS). Single and paired pulses TMS are used fordiagnostic purposes in order to assess different parametersabout the motor cortex excitability, whereas repetitive TMS isused for therapeutic purposes. Investigating the motor cor-tex excitability involves measuring MEP amplitudes, motorthreshold and silent period by using the single pulse TMSand the intracortical inhibition (ICI) and facilitation (ICF) byusing the paired pulses TMS. The single pulse TMS consists ofadministering a single pulse and of recording the electromyo-graphic (EMG) response, whereas TMS paired pulses consistsof the administration of two pulses (a conditioning one and atest one) with a certain delay, called Inter-Stimulus Interval ISI.Fig. 2a shows the MEP response when a paired pulse stimulusis administered to a patient.

In such signals it is possible to identify:

• The latency, which is the time interval between the instantwhen the stimulation is administered to the subject and the

administered to a patient and (b) cortical excitability curve.

instant when the muscle starts to move. Latency tends toincrease with age and height.

• The amplitude of the muscular response, which is the peak-to-peak excursion expressed in volts of the instrument thatmeasures the muscle response.

The intracortical inhibition (ICI) and facilitation (ICF) are,instead, related to the cortical excitability that is estimatedby a graph that describes the obtained amplitudes of themuscular responses at varying of the ISIs with respect tothe amplitude obtained at ISI = 0. An example of a corticalexcitability curve is shown in Fig. 2b. Currently all the TMSexperiments are carried out by interacting manually with theTMS hardware, hence by setting only one ISI per time, whereasthe number of repetitions for each ISI is performed by theexperimenter by clicking a button on the coil as many timesas the number of repetitions. Indeed, although the availableTMS equipment is provided with tools allowing the automaticparameters’ setting, such tools use proprietary script lan-guages (similar to programming languages, e.g. the softwareSignal1 of the Cambridge Electronic Design) that make the taskof designing TMS experiments very difficult and tedious formedical doctors.

In the next section the proposed customizable acquisitionand processing system that permits the full customizationof all currently used TMS paradigms (single pulse and pairedpulse TMS) is described.

3. The proposed tool

In this paper we propose a customizable data acquisition andprocessing tool that supports neuroscientists in the automati-zation and customization of all currently used TMS paradigms,

in Fig. 3 and consists of three main modules:

1 http://www.ced.co.uk/pru.shtml.

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15 7

pose

•

•

•

3

Atusic

Cifatbaa(ccgtct

Fig. 3 – The pro

Hardware interaction module: it handles the interactionwith the hardware equipment for executing TMS experi-ments;Experiment data management module: it allows neurosci-entists, through an intuitive interface, to store patient datain RDF format, to set the parameters of TMS experiments, toprocess the acquired data, to define research studies involv-ing several patients and to analyze data from such studieswith statistical tests.

Diagnosis support system module for supporting neurosci-entists especially in the differential diagnosis. This moduleperforms on-line training from data to handle uncertaincases.

.1. Hardware interaction module

hardware-interface communicates with the TMS equipmenthat interacts with a real-time data acquisition unit. This mod-le implements a common programming interface in order toupport different data acquisition systems. It is sufficient tomport a library (specific for the hardware) for enabling theommunication with the TMS hardware.

To date, only the library for communicating with theED 14012 is present in our system. The CED 1401 A/D

s one of the most common signal acquisition systemsor TMS response acquisition and stimulation synchronizernd it usually comes with MagStim3 stimulators. It fea-ures 4 analog inputs capable of acquiring signals with 16it resolution at a 500 kHz sampling rate, 2 digital inputsnd 2 digital outputs. One of the analog inputs is used tocquire the response signals through a small signal amplifierCED 1902). Therefore, the CED 1401 receives the user-ommands, and synchronizes two stimulators MagStims 200,onnected on its digital outputs, for the creation of the sin-le pulses, which are further combined in a paired pulse by

he Magstim BiStim and are administered to the patient’sortex through the coil. After the TMS stimulus adminis-ration, the muscular response (MEP) is registered by using

2 http://www.ced.co.uk.3 http://www.magstim.com.

d architecture.

single-use, low-noise, high conductivity electrodes. Suchmotor responses are then amplified, using the CED 1902, witha gain ranging from 100 to 1,000,000 (V/V) and a maximumvoltage input range ±10 V.

3.2. Experiment data management module

To assist the neuroscientists in the entire life-cycle of a TMSbased research, the proposed tool provides the users witha set of flexible functionalities for setting all the necessaryparameters, for processing the acquired data and for storingthe information in order to be processed by other semantic-based applications or to be shared with other researchers. Thismodule consists of four sub-modules:

• Experiment setting sub-module, for establishing the param-eters of a TMS paradigm (ISI, number of repetitions, etc.),the criteria for patients’ enrollment and the variables (clin-ical, neuropsychological, etc.) of the patients that should beinvestigated for the specific scientific research;

• Signal post processing sub-module, for processing theacquired muscular responses in order to remove noise andother inconsistencies that may affect the quality of theacquired data;

• Statistical analysis sub-module, for assessing the results ofthe performed studies;

• Data storage sub-module, for handling the storage of anydata produced in the system, from the patient’s data, tostatistical analysis results, to classifier’s parameters. It isprovided with different RDF repositories for each type ofproduced data.

3.2.1. Experiment setting moduleUsually, a research study starts with the definition of a pairedTMS protocol that involves the specification of the protocolvariables to be analyzed (clinical, psychiatric, neurophysio-logical, etc.) that are strictly related to the disease/diseasesunder investigation, and the TMS parameters, namely the

ISIs to administer, the number of repetitions for each ISI andthe modality of administration (random or sequential). Theschema of this module is shown in Fig. 4 and the graphicaluser interface for protocol definition is shown in Fig. 5.

8 c o m p u t e r m e t h o d s a n d p r o g r a m s i

Fig. 4 – Experiment data management module’s

architecture.

After the protocol definition for a paired-pulse TMS, thedata of each patient belonging to a specific study can beacquired. Among the variables specified in the protocol foreach study, the common neurophysiological parameters suchas motor threshold, silent period must be estimated using thesingle pulse TMS. After entering relevant demographic/clinicaldata of the patient under investigation, the paired-pulse TMSwith the parameters set during the protocol definition can beadministered to the patient. Fig. 6 shows the user interfacewhile administrating paired-pulse TMS (with MEP responses)according to a specific protocol: in the left side the plots of MEPresponses for a specific ISI are shown, whereas in the right sidethe TMS protocol settings are listed and the monitoring of thesubject’s relaxation status is displayed.

3.2.2. Signal post processing moduleAfter protocol setting and execution, the acquired data areprocessed by the “signal post processing” module. Indeed,

the acquired MEP responses to the administered TMS stimulirarely respect the quality criteria imposed by the experimenterbecause of both the variability of the MEP signals during therecording and the noise affecting such signals. MEP signals

Fig. 5 – The graphical user inter

n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15

show, typically, a high variability in the values depending ontwo main factors: (1) the misalignment of the coil over thepatient’s head, which can be corrected by adjusting the coil’sposition, and (2) the stimulus administration when the relax-ation level of the patient invalidates the muscular response;indeed, if the patient is relaxed the muscular response is gen-erally accurate, whereas if the patient is nervous, suffers froma disease or is on medications that alter the electrical signalsthat the brain sends to the peripheral nerves, the acquisition ofmuscular responses is difficult, and often not possible. To dealwith this problem, the proposed system includes an on-linemonitoring module (right side in Fig. 6) that continuously eval-uates the relaxation level of patients. This module checks thepatient’s relaxation level in real-time and eventually informs,in case of inappropriate levels, the experimenter, who can dis-card manually the acquired signals. Moreover, the system canbe set to discard automatically the MEP responses accordingto the evaluated relaxation condition. The automatic MEP sig-nal elimination is implemented by estimating if, at the timeof the pulse administration, the relaxation level (computed asthe area under the muscular response detected by the EMG,e.g. the curve of the MEP signal shown in the right side ofFig. 6) is in the range � ± � where � and � are, respectively, themean and the standard deviation of the previously evaluatedrelaxation level.

The accuracy of the acquired signals may be also influencedby noise. For example, high amplitude 50–60 Hz alternatecurrents are commonly found in any intrinsically noisy envi-ronment such as hospitals. The 50–60 Hz AC noise is easilypredicted and it can be removed by using notch filters in theappropriate frequency range (49–51 Hz for Europe, 59–61 Hzfor the USA). Another type of environmental noise is thehigh frequency interference due to the usage of other electri-

cal/electronic devices near to the TMS acquisition equipment.Unfortunately, it is difficult to eliminate such noise withoutaltering substantially the base response signal thus our sys-tem is provided with a noise removal tool based on Fourier

face for protocol definition.

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15 9

nter

Spnvpr

3AjatbpfitSpaftspad(tae

• The MeSH controlled vocabulary for coding disease, symp-

Fig. 6 – The graphical user i

ignal Decomposition. This tool addresses only sinusoidal andredictable noise by analyzing the signal’s frequency compo-ents and therefore, the validation of the results is based onisual inspection carried out by the experimenters. The toolermits to re-administer a stimulus if the noise cannot beemoved.

.2.3. Statistical analysis modulefter completing the data acquisition phase from the sub-

ects sample, according to the designed protocol, the statisticalnalysis is performed. Usually, this step is done by a statis-ician, but often medical research centers are not providedy a statistics unit and this is a bottleneck. Therefore, theroposed system implements a statistics module that per-orms the most common tests for data statistical analysis,n an automatic and transparent way. This module exploitshe functionalities of the IBM SPSS4 software by using thepssClient API. Depending on the variables that the study’srotocol contains, the statistical analysis module is able toutomatically decide the appropriate statistical tests to per-orm. Moreover, according to the distribution of the values ofhe variables involved in a defined protocol, a specific test iselected. For example, in Fig. 7 we have the summary of a TMSrotocol carried out on two groups of patients: control patientsnd patients affected by vascular depression. The variablesefined in the protocol are Mini Mental State Examination

MMSE), Familiar History (F-Hyst), Personal History (P-Hyst) and

he average (averaged on the number of repetition of each ISI)mplitude at ISIs 1, 3, 5, 7, 10 and 15 extracted from the corticalxcitability curve.

4 http://www.spss.com/software/statistics/.

face for protocol execution.

According to the type of variable to be compared, our sys-tem checks if the variable is a numeric value or a percentageand also performs the normality test to decide if parametric ornon-parametric tests should be executed. In the case shownin Fig. 7 we have that the MMSE is a numeric variable and itis not normally distributed, therefore the Mann–Whitney testis performed, whereas since the variable P-Hyst is boolean,the comparison between the two groups is performed usingthe Chi-square test. Fig. 8 shows the output of the statisticalanalysis for the above described example.

3.2.4. Data storage moduleThe nature of the data processed by the proposed platformpermits the adoption of semantic repositories to be used asthe system’s storage servers. In fact, by using well estab-lished ontologies, like FOAF,5 and controlled vocabularies, likeMeSH,6 and by creating an appropriate schema to describe thewhole data structure, including data relationships, data canbe easily processed by intelligent medical systems, such asthe one proposed herein, and by semantic tools. In particular,the whole experiment workflow is enriched with informationfollowing an RDF schema that includes:

• The FOAF ontology to describe patient and neuroscientistsinformation;

toms and signs associated to diseases;

5 http://www.foaf-project.org/.6 http://www.ncbi.nlm.nih.gov/mesh.

10 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15

Fig. 7 – A subset of the patients’ group on which the “vascular depression” protocol has been executed. The comparisonbetween these two groups (controls, vascular depression) is performed automatically by means of statistical tests.

orme

Fig. 8 – Results of the statistical tests perf

• A set of RDF classes and properties that describe a TMSbased scientific study including protocols, variables andTMS technical parameters.

The proposed RDF schema for TMS studies is avail-able at http://i3s-lab.unict.it/semweb/ns/TMSSchema.xml.A complete description of the RDF schema is beyondthe aim of the paper, although we provide here somehighlights about the underlying design. The vari-ables used in the protocol definition are stored in RDFand are structured as a SKOS7 vocabulary (http://i3s-lab.unict.it/semweb/ns/VariableDictionary.xml). In detail,they are grouped in several categories and we have a SKOScollection for each category, e.g. for clinical variable, forneurophysiological variable, for neuropsychological variable,for medical imaging variable. We have also defined a class

TMSProtocolVariable, for describing the variables (differentfrom the ones above listed) that can be derived only from theTMS, which is also a subclass of SKOS:concept. This allows us to

7 http://www.w3.org/2004/02/skos/.

d on the patients’ group shown in Fig. 7.

create a collection of TMS variables and to add other features(such as the range of the variables) that are not includedin SKOS. In Appendix A an example of an RDF instance ofthe proposed schema and describing a generic TMS study isshown.

Personal information about the patient is inserted exclu-sively by the neuroscientist who carries out the examinationand, for privacy purposes, our semantic system replaces thepatient’s FOAF profile URI with an appropriate MD5 hashstring. The data storage has been implemented by semanticrepositories using SESAME servers (see Fig. 9) to make theseinformation available for other purposes. In detail, four dis-tinct RDF repositories are available:

• The patient master data store is the semantic databasewhere all the information about patients is stored, includingparameters for statistical analysis, like age, smoker, gender,etc.

• The variables data store is used for the variables definedduring the TMS protocol design.

• The experiment data store is a combination of a semanticrepository and a file server. The semantic database stores

c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15 11

dule

•

3

CcierdeibiopacctdrdtiaAsI

Fig. 9 – Interaction between the data storage mo

signal information, such as amplitude, latency, ISI. Thefile server retains the whole muscular responses in orderto extract the aforementioned values and to export theacquired signal in a human readable format (e.g. an image).

The classifier data repository, where the classifier’s param-eters for diagnosis support are stored.

.3. Diagnosis support system

urrently, the diagnosis of many neurodegenerative and vas-ular diseases is mainly based on clinical evidence and onmaging techniques such as MRI, PET and SPECT. Often,specially at a very early stage, the clinical evidence of neu-odegenerative disorders (e.g. Parkinson disease, Alzheimerisease, etc.) may be very similar to the one of vascular dis-ases. Medical imaging techniques (especially MRI) may helpn such cases: indeed the MRI shows mainly atrophy of therain in neurodegenerative disorders [18] and ischemic lesions

n vascular diseases [19]. The problem arises when both typesf evidence are present in an MRI, especially in elderly peo-le who may have brain’s atrophy due to the advanced age,lthough the main cause of their symptoms could be a vas-ular disease [20]. An example is the mixed dementia, i.e. thease when neurodegenerative dementia and vascular demen-ia occur at the same time [21]. The differential diagnosis isifficult not only in the above cases, but also among neu-odegenerative diseases (e.g. Alzheimer disease vs Lewy bodyisease [22] or Parkinson disease vs Lewy body disease [23])hat could exhibit similar features at early stages. Therefore,t is necessary to identify the main cause of the observed signs

nd symptoms in order to provide the appropriate treatment.s mentioned in the introduction, TMS-studies have demon-trated, by investigating motor threshold, cortical silent periodCF and ICI, that the various neurological diseases may involve

and the other modules of the proposed system.

motor pathways in different ways. Hence, given that TMSprovides detailed information about the motor system andsince motor system’s alterations have been identified inmany neurological diseases, an appropriate processing of MEPresponses may be used for supporting the diagnosis.

Under this scenario, a diagnosis support system may playa double role: first, assess if the obtained MEP responses areevidence of neurological disorders, and second, support neu-roscientists in differential diagnosis. To address the first need,two methods [24] and [25] have been proposed for classi-fying diseases such as Alzheimer and Subcortical ischemicvascular dementia by analyzing the MEP responses of a TMSparadigm. In particular a fuzzy system [25] and a neural net-work [24] were proposed and assessed for the differentialdiagnosis of Alzheimer and Vascular Dementia by using thefollowing features: latency, amplitude, max and min mod-ule of the Fourier Transform, max and min module of theHilbert transform of the MEP responses for ISI 1, 3, 5, 7, 10,and both of them achieved an average accuracy of about92%. However, since these approaches are disease-specific (thetraining is done off-line) they cannot be used in a dynamicresearch and clinical context, such as the one here fore-seen, where different TMS paradigms may be implementedfor analyzing different diseases. For all of the above reasons,the proposed system is provided with an on-line diagnosissupport system (DSS) that uses the above features extractedfrom a MEP response and it is based on a modified versionof a support vector machine for large-scale problems (typ-ically, about 1000 exams per year are executed in a singleneurophysiological unit), capable of learning incrementally

(averagely, between three and five exams per day are exe-cuted). Support vector machines (SVM) have been widely usedfor implementing classifiers because of their good generaliza-tion property [26]. Their main shortcoming is that training

12 c o m p u t e r m e t h o d s a n d p r o g r a m s i n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15

Fig. 10 – Graphical user interface for handling patients without diagnosis.

is time consuming, thus preventing their use in large-scaleproblems such as the one at hand. A solution is to resort toa modified SVM that supports on-line incremental learning.Several approaches for incremental learning have been pro-posed. The first attempts were developed by Syed et al. in[27] and by Ruping in [28] by re-training the SVM through newexamples combined with the already computed support vec-tors; however, these approaches are very memory demanding.Differently, to address large-scale issues, approaches basedon clustering techniques for down-sampling the size of theexamples and using the most representative ones for re-training have been proposed [29,30]. Therefore, the problemsto be solved for on-line SVM are: the on-line selection ofthe learning data and the re-use of the already computedsupport vectors. Our diagnosis support system relies on theon-line SVM proposed in [31] that implements on-line train-ing and, at the same time, solves the large scale problem. Adetailed evaluation of the achieved performance, in terms ofaccuracy, and the comparison with the existing on-line classi-fication systems are beyond the aim of this article, although wecan report that in 18 uncertain diagnosis cases, over a total-ity of about 70 patients, the DSS performed well identifyingthe four diseases these cases belonged to. The DSS moduleis, therefore, used when a new patient whose diagnosis isunknown is inserted into the system (see Fig. 10 for the relatedGUI).

4. Concluding remarks

Transcranial magnetic stimulation is an important non-invasive and painless diagnostic and research methodthat is progressively gaining ground against more painfulalternatives like transcranial electric stimulation. However,

application specific and easy-to-use tools integrating fea-tures such as automated statistical analysis and intelligentalgorithms that can support neuroscientists in the diagno-sis, currently do not exist. In this paper we proposed asoftware tool that covers the whole workflow of a TMS exper-iment. This tool is composed of four distinct modules, eachone addressing a specific aspect of a TMS-based experi-ment. In particular, the experiment data module managesthe patients and the results of the experiments, while thehardware interface module is responsible for the experi-ment execution and for the interaction with the acquisitionequipment. The diagnosis support system is based on anon-line SVM to automatically classify patients based on theMEP responses. The entire platform’s data is handled by thedata storage module that incorporates semantic web stan-dards to store the data in four distinct repositories for fasteraccess, sensitive data isolation and easier data sharing. More-over, the proposed RDF schema to describe TMS data allowsneuroscientists to share with the neuroscience communityboth single experiments and entire scientific research stud-ies (data sets and results) with the main aim to standardizethe method (i.e. the used variables and procedures/protocols)of studying cortical excitability using TMS. Future work onthe tool will regard enhancing the automatic signal correc-tion and denoising algorithms for more accurate results.Another important enhancement should be the integrationof an advanced dynamic feature selection module so thatthe DSS can use not only the features derived from MEP sig-nals but also the other patient’s variables. To achieve an eventighter integration between non TMS data collection proce-

dures and their joint analysis with TMS data, we are currentlyworking on adding to the proposed tool all the tests (neuro-physiological, neuropsychological, etc.) that can be performedusing a computer, and also on including available modules for

s i n

amcaFap

c o m p u t e r m e t h o d s a n d p r o g r a m

utomatic analysis of medical images, in particular the seg-entation approaches proposed in [32,33] and the MRI lesion

lassification module proposed in [34], to reduce the time

nd effort that the MRI-related variables calculation requires.inally, a personal health record management system, suchs the one in [35], is under development to make the system’satient records globally available.

Listing 1: RDF ins tance describing a TM S study

<? xml versio n ="1.0 " enc oding =" utf -8" ? ><rdf: RDF xmlns :tm s ="htt p :/ /i3 s - lab . unict .it/semwe b /n

xml ns :rd f ="htt p :/ / www .w3. org /1999/02/2 2 - rdf - syn taxxml ns :sko s ="htt p ://ww w .w3. org /2 004/0 2/ skos/cor e #"xml ns :rdf s ="htt p ://ww w .w3. org /2 000/0 1/ rdf -schem a #"xml ns :dc ="htt p :/ / pur l . org /dc/ ele ments /1.1 /"xml ns :foa f ="htt p :// xmlns . com /foa f /0 .1/" ><tms : TMSS tudy rd f : about ="htt p ://i3 s - lab . unict .it/s

Vascular Depress ionTest Study "><dc:des criptio n >A tes t proto col for assessin g Va <tms : Neuroph ysicia n rdf : res ource ="htt p ://i3 s - lab<tms :TMSPr otoco l >

<tms :TM SParamete rs ><tms : ISILis t >0 ,1 ,3 ,5 ,7 ,10 ,15 </tm s : ISIL ist ><tms :IBS >1 0 </ tms :IBS ><tms :Re pet ition s >15 </ tms :Rep etition s >

</ tms :TMSPar ameter s ></ tms :TMS Protoco l ><tms :TMSPr otoco l >

<tms :TMSVariableCollectio n ><sko s :membe r rdf :re source ="htt p :// i3s -la b . un<sko s :membe r rdf :re source ="htt p :// i3s -la b . un<sko s :membe r rdf :re source ="htt p :// i3s -la b . un<sko s :membe r rdf :re source ="htt p :// i3s -la b . un<sko s :membe r rdf :re source ="htt p :// i3s -la b . un

Brai nMRIL esi onLoa d "/ ></ tms :TMSVariableCollectio n >

</ tms :TMS Protoco l ><tms :TM SExperimen t >

<tms :Patie ntCollectio n ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

c4ca4238a0b923820dcc509a6f75849b "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

c81e728d9d4c2f636f067f89cc14862c "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

eccbc87e4b5ce2fe28308fd9f2a7baf3 "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

a87ff679a2f3e71d9181a67b7542122c "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

e4da3b7fbbce2345d7772b0674a318d5 "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

c5a880faf6f b5e6087e b1b2d c "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

f14e45fceea167 a5a36dedd4 bea254 3 "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

c9f0f895fb98ab9159f51fd0297e236d "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

c48cce2 e2d7f bdea1 afc5 1c7c6 ad2 6 "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

d3d9446802a44259755d38e6d163e820 "/ ><tms :Patien t rdf :re source ="htt p :// i3s -la b . un

bd43d9caa6 e02c990b0 a82652dca "/ ></ tms :PatientCol lectio n >

</ tms :TMSExp erimen t ></ tms :TMSStud y >

</ rdf :RDF >

b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15 13

Conflict of interest

None.

Appendix A. Example of an RDF instancedescribing a generic TMS study

s/ tms Schem a " -ns #"

emweb /TM SStudies /2011/03/ 17/

scular Depress ion in 10 pati ents </ dc:descrip tio n > . unict .it/semwe b /foa f /Jo hn_Smith "/ >

ict .it/semw eb /ns/ Varia ble Dicti ona ry #Smoke r "/ >ict .it/semw eb /ns/ Varia ble Dicti ona ry # LeftHan ded "/ >ict .it/semw eb /ns/ Varia ble Dicti ona ry # aMT "/ >ict .it/semw eb /ns/ Varia ble Dicti ona ry # CSP "/ >ict .it/semw eb /ns/ Varia ble Dicti ona ry #

ict .it/semw eb / TMSSt udies / Patients /

ict .it/semw eb / TMSSt udies / Patients /

ict .it/semw eb / TMSSt udies / Patients /

ict .it/semw eb / TMSSt udies / Patients /

ict .it/semw eb / TMSSt udies / Patients /

ict .it/semw eb / TMSSt udies / Patients /167 9091

ict .it/semw eb / TMSSt udies / Patients /8

ict .it/semw eb / TMSSt udies / Patients /

ict .it/semw eb / TMSSt udies / Patients /45

ict .it/semw eb / TMSSt udies / Patients /

ict .it/semw eb / TMSSt udies / Patients /6512

m s i

r

14 c o m p u t e r m e t h o d s a n d p r o g r a

e f e r e n c e s

[1] R. Chen, J. Classen, C. Gerloff, P. Celnik, E.M. Wassermann,M. Hallett, L.G. Cohen, Depression of motor cortexexcitability by low-frequency transcranial magneticstimulation, Neurology 48 (1997) 1398–1403.

[2] M.W. O’Dell, C.C. Lin, V. Harrison, Stroke rehabilitation:strategies to enhance motor recovery, Annu. Rev. Med. 60(2009) 55–68.

[3] L.A. Wheaton, F. Villagra, D.F. Hanley, R.F. Macko, L.W.Forrester, Reliability of tms motor evoked potentials inquadriceps of subjects with chronic hemiparesis afterstroke, J. Neurol. Sci. 276 (1–2) (2009) 115–117.

[4] R. Chen, D. Cros, A. Curra, V. Di Lazzaro, J.P. Lefaucheur, M.R.Magistris, K. Mills, K.M. Rosler, W.J. Triggs, Y. Ugawa, U.Ziemann, The clinical diagnostic utility of transcranialmagnetic stimulation: report of an IFCN committee, Clin.Neurophysiol. 119 (2008) 504–532.

[5] F. Mori, C. Ljoka, E. Magni, C. Codeca, H. Kusayanagi, F.Monteleone, A. Sancesario, G. Bernardi, G. Koch, C. Foti, D.Centonze, Transcranial magnetic stimulation primes theeffects of exercise therapy in multiple sclerosis, J. Neurol.(2011).

[6] R. Traversa, P. Cicinelli, M. Oliveri, M. Giuseppina Palmieri,M.M. Filippi, P. Pasqualetti, P.M. Rossini, Neurophysiologicalfollow-up of motor cortical output in stroke patients, Clin.Neurophysiol. 111 (2000) 1695–1703.

[7] G. Pennisi, R. Ferri, M. Cantone, G. Lanza, M. Pennisi, L.Vinciguerra, G. Malaguarnera, R. Bella, A review oftranscranial magnetic stimulation in vascular dementia,Dement. Geriatr. Cogn. Disord. 31 (2011) 71–80.

[8] R. Cantello, R. Tarletti, C. Civardi, Transcranial magneticstimulation and Parkinson’s disease, Brain Res. Brain Res.Rev. 38 (2002) 309–327.

[9] M. Kobayashi, A.P. Leone, Transcranial magnetic stimulationin neurology, Lancet Neurol. 2 (2003) 145–156.

[10] F. Fregni, A. Pascual-Leone, Transcranial magneticstimulation for the treatment of depression in neurologicdisorders, Curr. Psychiatry Rep. 7 (2005) 381–390.

[11] G. Chibbaro, M. Daniele, G. Alagona, C. Di Pasquale, M.Cannavo, V. Rapisarda, R. Bella, G. Pennisi, Repetitivetranscranial magnetic stimulation in schizophrenic patientsreporting auditory hallucinations, Neurosci. Lett. 383 (2005)54–57.

[12] W. Hu, X. Wang, Y. Yang, D. Liang, F. Zhao, Design of a halfsolenoid coil for optimization of magnetic focusing intrans-cranial magnetic stimulation, Sheng Wu Yi Xue GongCheng Xue Za Zhi 24 (2007) 910–913.

[13] P.M. Rossini, S. Rossi, Transcranial magnetic stimulation,Neurology 68 (7) (2007) 484–488,doi:10.1212/01.wnl.0000250268.13789.b2.

[14] T. Kujirai, M.D. Caramia, J.C. Rothwell, B.L. Day, P.D.Thompson, A. Ferbert, S. Wroe, P. Asselman, C.D. Marsden,Corticocortical inhibition in human motor cortex, J. Physiol.(Lond.) 471 (1993) 501–519.

[15] R. Chen, A. Tam, C. Butefisch, B. Corwell, U. Ziemann, J.C.Rothwell, L.G. Cohen, Intracortical inhibition and facilitationin different representations of the human motor cortex, J.Neurophysiol. 80 (1998) 2870–2881.

[16] A. Kaelin-Lang, L.G. Cohen, Enhancing the quality of studiesusing transcranial magnetic and electrical stimulation witha new computer-controlled system, J. Neurosci. Methods 102(2000) 81–89.

[17] A. Faro, D. Giordano, I. Kavasidis, C. Pino, C. Spampinato,M.G. Cantone, G. Lanza, M. Pennisi, An interactive tool forcustomizing clinical transacranial magnetic stimulation(TMS) experiments, in: R. Magjarevic, P.D. Bamidis, N.

n b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15

Pallikarakis (Eds.), XII Mediterranean Conference on Medicaland Biological Engineering and Computing 2010, vol. 29 ofIFMBE Proceedings, Springer, Berlin/Heidelberg, 2010, pp.200–203.

[18] N.C. Fox, P.A. Freeborough, Brain atrophy progressionmeasured from registered serial mri: validation andapplication to alzheimer’s disease, J. Magn. Reson. Imaging 7(6) (1997) 1069–1075, doi:10.1002/jmri.1880070620.

[19] K.A. Jellinger, The pathology of ischemic-vascular dementia:an update, J. Neurol. Sci. 203-204 (2002) 153–157.

[20] G. Alagona, R. Ferri, G. Pennisi, A. Carnemolla, T. Maci, E.Domina, A. Maertens de Noordhout, R. Bella, Motor cortexexcitability in Alzheimer’s disease and in subcorticalischemic vascular dementia, Neurosci. Lett. 362 (2004) 95–98.

[21] K.M. Langa, N.L. Foster, E.B. Larson, Mixed dementia:emerging concepts and therapeutic implications, JAMA 292(2004) 2901–2908.

[22] C.F. Lippa, T.W. Smith, J.M. Swearer, Alzheimer’s disease andLewy body disease: a comparative clinicopathological study,Ann. Neurol. 35 (1994) 81–88.

[23] E.J. Byrne, G. Lennox, J. Lowe, R.B. Godwin-Austen, DiffuseLewy body disease: clinical features in 15 cases, J. Neurol.Neurosurg. Psychiatr. 52 (1989) 709–717.

[24] A. Faro, D. Giordano, M. Pennisi, G. Scarciofalo, C.Spampinato, F. Tramontana, Transcranial magneticstimulation (tms) to evaluate and classify mental diseasesusing neural networks, in: S. Miksch, J. Hunter, E. Keravnou(Eds.), Artificial Intelligence in Medicine, vol. 3581 of LectureNotes in Computer Science, Springer, Berlin/Heidelberg,2005, pp. 310–314.

[25] A. Faro, D. Giordano, M. Pennisi, G. Scarciofalo, C.Spampinato, F. Tramontana, A fuzzy model and tool toanalyze SIVD diseases using TMS, Int. J. Signal Process. 2 (1)(2006).

[26] M. Hearst, S. Dumais, E. Osman, J. Platt, B. Scholkopf,Support vector machines, IEEE Intell. Syst. Appl. 13 (4) (1998)18–28, doi:10.1109/5254.708428.

[27] N. Syed, H. Liu, K. Sung, Incremental learning with supportvector machines, in: International Joint Conference onArtificial Intelligence, IJCAI 99, Stockholm, Sweden, July31–August 6, 1999.

[28] S. Rüping, Incremental learning with support vectormachines, in: Proceedings of the 2001 IEEE InternationalConference on Data Mining, ICDM’01, IEEE ComputerSociety, Washington, DC, USA, 2001, pp. 641–642.

[29] I.W. Tsang, J.T. Kwok, P.-M. Cheung, Core vector machines:fast SVM training on very large data sets, J. Mach. Learn. Res.6 (2005) 363–392.

[30] I.W. Tsang, A. Kocsor, J.T. Kwok, Simpler core vectormachines with enclosing balls, in: Proceedings of the 24thInternational Conference on Machine Learning, ICML’07,ACM, New York, NY, USA, 2007, pp. 911–918.

[31] J. Zheng, H. Yu, F. Shen, J. Zhao, An online incrementallearning support vector machine for large-scale data, in:Proceedings of the 20th International Conference onArtificial Neural Networks: Part II, ICANN’10,Springer-Verlag, Berlin/Heidelberg, 2010, pp. 76–81.

[32] D. Giordano, R. Leonardi, F. Maiorana, G. Scarciofalo, C.Spampinato, Epiphysis and metaphysis extraction andclassification by adaptive thresholding and dog filtering forautomated skeletal bone age analysis, Conf. Proc. IEEE Eng.Med. Biol. Soc. 2007 (2007) 6552–6557,http://www.biomedsearch.com/nih/Epiphysis-metaphysis-extraction-classification-by/18003527.html.

[33] D. Giordano, C. Spampinato, G. Scarciofalo, R. Leonardi, An

automatic system for skeletal bone age measurement byrobust processing of carpal and epiphysial/metaphysialbones, IEEE Trans. Instrum. Meas. 59 (10) (2010) 2539–2553.

s i n

Personal Health Records, in: Proceedings of the 10th IEEE

c o m p u t e r m e t h o d s a n d p r o g r a m

[34] A. Faro, D. Giordano, C. Spampinato, M. Pennisi, Statisticaltexture analysis of MRI images to classify patients affectedby multiple sclerosis, in: R. Magjarevic, P.D. Bamidis, N.

Pallikarakis (Eds.), XII Mediterranean Conference on Medicaland Biological Engineering and Computing 2010, vol. 29 ofIFMBE Proceedings, Springer, Berlin/Heidelberg, 2010, pp.272–275.

b i o m e d i c i n e 1 0 7 ( 2 0 1 2 ) 4–15 15

[35] A. Faro, D. Giordano, I. Kavasidis, C. Spampinato, A web 2.0telemedicine system integrating TV-centric services and

International Conference on Information Technology andApplications in Biomedicine, Corfu, Greece, November 2–5,2010.