Central Information/ Signal Processing, and Medical Expert Prediction/Estimation Subsystem
Subjects were classified as “responders” if the 17- HamD score at six weeks showed at least a 50% improvement over the pre-treatment HamD score. The HamD percentage change value is discretized into two values (or classes), corresponding to responder (R) when it is larger than or equal to 50%, and nonresponder (NR) otherwise.
Quantitative Features:
pr ot
The set of Nc candidate features extracted from each data epoch consists of the anterior/posterior power ratios at various frequencies, and between various electrode pairs, in addition to some ratios involving more than two electrode pairs. Above features being calculated at all frequencies in 4Hz to 36Hz band with 1Hz resolution. Applied z-score normalization on extracted features, using EEG data of 91 normal i.e. healthy subjects. Nc = 1452 candidate features The regularized feature selection based on mutual information by Peng et al (2005) is used. Nr = 4 statistically discriminating features (out of total Nc) are selected.
ht
yr ig
ht p
ig
yr
op
C
s.
RESULTS
A list of most discriminating features,for non-responder (NR) and responder (R) groups:
s.
Po 00 .F 10
ed
Predicted R 3
Total No. 18
7 (77.8%) Avg. performance = 80%
9
27
The average correct prediction rate is 80% (specificity=83.3%, sensitivity=77.8%).
0.6
Fig. 1
NR R
0.4
Clustering of the Feature Space: Scatter
plot of 400 pre-treatment training samples (in Nr= 4 dimensional space) projected onto the two nonlinear principal components (PC) using the KPCA method. Fig. 1 shows one point per epoch. Fig. 2: Subject-wise scatter plot which is obtained from Fig. 1 by averaging all points (epochs) corresponding to each subject.
21 21 21
0.2
16 16
21
21 21 21
21
0
16 16 16 16 16 16 16 16 16 16 16
2121
21 2121 2121
-0.2
-0.4 -0.6
0.6
16 16
-0.4
-0.2
0 0.2 axis 1, (PC1)
0.4
0.6
Fig. 2
NR R
0.4
1
0.2
25
0
24
9
10
21
17 5
13 11 6 22
-0.2
-0.4 -0.6
-0.4
-0.2
7
23
15
12
3
27 14
19 2 18
0 0.2 axis 1, (PC1)
8
20
4
16
26
0.4
0.6
CONCLUSION
Our method automatically identifies predictive QEEG features from a large number of candidates, with no a priori conditions by which features or electrodes should be considered. This method could have clinical utility in assisting with individualized treatment planning by identifying patients with higher probability of responding to rTMS therapy. Through appropriate patient selection, the apparent effectiveness of rTMS can be enhanced. Our findings require validation in a larger set of subjects. These pilot data suggest that the ML process is effective in a Personalized Medicine context for predicting response of a particular individual to rTMS treatment for MDD.
References:
st er
d.
te
ec
ot
pr
ht
ig
All subjects also received concurrent SSRI antidepressant medication during rTMS therapy, and for an additional 4 weeks thereafter.
yr
ro te
op
C
The class likelihoods are estimated using a mixture of factor analysis (MFA) statistical model (Ghahramani and Hinton, 1996). For each subject, the prediction result is produced by averaging the probabilities over all corresponding epochs before quantization. To evaluate the response prediction performance, we used an iterative ‘leave2-subjects-out’ (L2O) cross-validation procedure.
st er
0
00
F1
pr
ht
ig
yr
C op
Treatments were administered using either a Dantec Magpro or a Magstim Superrapid daily for 10 sessions over 2 weeks to a site 5 cm anterior, parasagittally, to the activation site for the abductor pollicis brevis muscle.
Actual R
Classification and Performance Evaluation:
Po s
ot ec
All subjects were enrolled in a randomized sham controlled trial of rTMS. Only data from subjects receiving true rTMS [Left High frequency rTMS N=18, or left high frequency plus right low frequency rTMS N=9 to the dorsolateral prefrontal cortex (DLPFC)] were used in this analysis. High frequency rTMS = 10 Hz, 20 trains, 8 second duration, 52 second intertrain interval, figure of eight coil, 110% of motor threshold. Low frequency rTMS = 1 Hz, 2 trains, 60 second duration, 3 minute intertain interval using a round coil, 110% of motor threshold.
s.
ec te d. F1 00 0
ht
pr ot ec te d.
F1 00 0
Po
Resting awake EEG is measured after 10 days off medication and just before beginning rTMS treatment. Signals from 16 electrodes (configured according to the standard 10-20 system referenced to linked ears), consisting of Fp1, Fp2, F3, F4, F7, F8, T3, T4, C3, C4, T5, T6, P3, P4, O1 and O2 are recorded using a QSI-9500 system, fs=205Hz. For each patient, a maximum of 3 EEG data files each of 3.5 minutes duration were collected while the subject’s eyes were open (EO). De-artifacting: segments with saturation/clipping are discarded. The signals were then digitally bandpass filtered after recording between 2.5Hz and 39 Hz to partially mitigate the effects of eye movement and muscle artifacts. First 90 seconds of the de-artifacted portion of the EEG data are used. These segments are divided into 5 overlapping epochs of 30 sec. duration, to give a nominal 15 epochs per subject (15 = 3 files X 5 epochs per file).
Actual NR
Predicted NR 15 (83.3%) 2
axis 2, (PC5)
EEG Recordings:
yr ig
op
C
er s. st
Po
te
d.
F1
00
0
pr ot
ec
27 subjects diagnosed with MDD using the Diagnostic and Statistical Manual-IV (DSM-IV). All subjects had previously failed to respond to at least two courses of antidepressant medication therapy. Age at start of treatment [years]: avg.=46.3, std=9.85, min=23.9, max=65.8. Gender: 20 female (74%) and 7 male subjects (26%). Pre-treatment Hamilton depression rating scale (HamD) scores were: avg.=21.1, std=3.58, min=15, max=27.
The rTMS Therapy:
st er s. C
pr ot ec te d. F
yr ig
op
C
er s.
st
Po
00
10
.F te d
Participants:
Disorders Program, Centre for Mountain Health Services, St. Joseph Hospital, Hamilton, ON
Definition of Response:
ht
Po st er s. C
10 00
METHODS
3 Mood
axis 2, (PC5)
op yr ig ht pr ot ec te d.
INTRODUCTION
op
C
Electrical and Computer Engineering Dept.,
Emails:
[email protected] (Gary),
[email protected] (Ahmad),
[email protected] (Jim)
According to the STAR*D trial 33% with Major depressive disorder (MDD) do not reach remission with current psychotropic medications and Cognitive Behavioral Therapy. Treatment options for these resistant patients includes electroconvulsive therapy (ECT) or repetitive transcranial magnetic stimulation (rTMS). rTMS employs strong localized pulsed magnetic fields administered through a magnetic coil placed on the head of the subject, to induce electrical currents in the brain to change the activity of neuron populations. Although some studies show that rTMS is effective in many treatmentresistant patients, others have not been able to demonstrate a difference from sham rTMS. At our laboratory open label compassionate rTMS is only effective in 40% to 50% of the medication-resistant patients we typically treat. Often rTMS is chosen over ECT because of the significant side effects and logistical complexity of ECT. As a typical course of rTMS is currently 20 to 25 daily treatments, those who choose rTMS over ECT but do not respond experience at least a 4 to 5 week delay before ECT is initiated. A method to determine, in advance, the likelihood of response to rTMS would have clinical utility as it would 1) eliminate unnecessary delay in initiating alternative treatments, 2) allow rTMS team resources to be used for those with greater likelihood of showing robust response. Earlier pilot studies using conventional statistical approaches suggest that quantitative electroencephalography (QEEG or EEG) data can be used to predict response to rTMS. We propose that patient selection for rTMS using machine learning (ML) analysis of QEEG will lead to even more accurate response prediction.
er
2
rs .
yr ig
Patient
Dept. of Psychiatry and Behavioral Neurosciences,
10 00
1
Po st er s. C
Gary M. Hasey1,3, Ahmad Khodayari-Rostamabad1,2, James P. Reilly2, Hubert DeBruin2 and Duncan J. MacCrimmon1
Update, Adaptation and Improvement
ht p
Laboratory Test Results, Medical/Clinical Assessment and Examination Results, Presumptive Diagnosis and other Neuro-psychobiological Measurements
F1 00 0
Treatment and Diagnosis Databases
Clinician/Physician
op C y P r op ct i o g ed st h yr t e ig .F p r s ht r . o 1 C te 00 pr o c ot 0 p t e y ec Po d r i . g te s F h te d. 1 t 0 p r F1 s. 0 ht ro C te 00 pr o c ot 0 p t e y ec Po d r i . g te s F h te d. 1 t 0 p r F1 s. 00 ro C te 00 P o c o 0 py te s Po te d. rig r st F1 s. h tp er C 0 F1 s. 00 o ro C te 00 P o c o 0 py te s P t d r
,
ink tion L unica l Data m m o a C ction Bidire nsfer Tra
te
Local Communication, Interface, Utility and Data Security Subsystem
op yr ig ht pr
Po st er s.
ro te ct ed .F
A Personalized Approach Using Machine Learning Methods to Predict Response to rTMS Treatment for Major Depression
Central Communication, Utility and Data Security Subsystem
B. G. Druss, R. A. Rosenheck and W. H. Sledge, “Health and disability costs of depressive illness in a major U.S. corporation,” American Journal of Psychiatry, vol. 157, pp. 1274–1278, 2000. J. P. Lefaucheur, “Methods of therapeutic cortical stimulation,” Neurophysiologie Clinique/Clinical Neurophysiology, vol. 39, pp. 1–14, 2009. G. W. Price and J. W. Lee and C. Garvey and N. Gibson, “Appraisal of sessional EEG features as a correlate of clinical changes in an rTMS treatment of depression,” Clinical EEG and Neuroscience, vol. 39, pp. 131–138, 2008. Z. Ghahramani and G. E. Hinton, “The EM algorithm for mixtures of factor analyzers,” Department of Computer Science Technical Report, CRG-TR-96-1, University of Toronto, Toronto, Canada, 1996. H. Peng, F. Long and C. Ding, “Feature selection based on mutual information: Criteria of max-dependency, max-relevance, and min-redundancy,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 27, no. 8, pp. 1226–1238, Aug. 2005.
