Prediction of Antidepressant Treatment Response Using Machine Learning Classification Analysis

Major depressive disorder (MDD) is a highly prevalent, chronic disabling condition with substantial morbidity and mortality. Depression currently is the fourth leading cause of global burden of disease (DALYs) and disability worldwide, and is expected to be second by 20201. Around one in eight people in Canada will develop depression during their lifetime, with the total cost to the Canadian economy estimated at $51 billion per year2. The costs of treating MDD are high in part due to limitations in effectiveness of antidepressant treatment. Approximately 60% of patients fail to remit to the first antidepressant prescribed3 and the subsequent selection of antidepressants remains a matter of trial and error. Using this trial and error approach, it may take a year or more to find the successful treatment for a patient4,5. The protracted ineffective treatment results in prolonged suffering, substantial morbidity, loss of productivity and an increased burden on patient's family. Brain-based biomarkers could assist in predicting clinical response to treatment intervention and in tailoring treatment for individual patients. The results of previous neuroimaging studies that examined brain markers of treatment response were derived from group averages 6,-9 and have limited predictive value at individual level. Another limitation of these studies is that predictors derived from the pretreatment baseline brain scans could be influenced by many personal (personality, childhood trauma, genotypes) and clinical (course, duration of illness, episodes, symptom clusters, and severity of symptoms and past medication exposure) characteristics which may limit the generalizability. On the other hand, there is growing evidence that the early clinical response within 2 weeks of antidepressant treatment and EEG changes in the first week of treatment can predict later outcomes. Furthermore, early treatment changes in brain function may provide crucial information on the brain's capacity to change with treatment and on the interactive effects between personal/ clinical characteristics and pharmacological factors, which may help differential prediction of treatment responses to two antidepressants. Hence, examining the predictive value of dynamic brain changes during the first two weeks of treatment in individual patients would improve statistical reliability and predictive accuracy and minimize the confounding effect inherent to pretreatment scans. In this study, we propose to investigate the predictive value of resting state and task related fMRI data collected at the pretreatment baseline and 2 weeks after treatment to predict remitters and non-remitters to desvenlafaxine antidepressant treatment at week 12 using machine learning classifier. Desvenlafaxine is a serotonin norepinephrine reuptake inhibitor (SNRI) with proven efficacy, and safety and is easy to administer in single daily dose. It has limited sedative and cognitive side effects such as drowsiness, lack of alertness and poor attention, which may confound early brain changes with treatment. This study will provide brain-based predictive biomarkers that can be tested prospectively in clinical trials and eventually in clinical practice for accuracy. Machine Learning Classification (Support Vector Machine): The support vector machine (SVM) is a computer based analytical technique designed for high dimensional biological data such as fMRI data and provides the best classification of individual observations into distinct groups 38. Diagnostic classification (depression diagnosis and healthy control) and classification of treatment responsiveness (responders and non-responders) have been examined in a clinical population with fMRI data using SVM 39-41. This technique consists of two phases: training phase and testing phase. During the training phase an SVM is trained to develop a decision function or hyperplane that separates the data into two groups according to a class label. In the testing phase, this decision function can be used to predict the class label of a new subject as being a responder or non-responder. The accuracy of prediction by SVM depends on its specificity (identification of true negatives) and sensitivity (identification of true positives). In recent years a few neuroimaging studies have employed SVM to structural and functional MRI data in order to predict the MDD patients who improved with treatment and who did not. Fu et al (2008) showed that applying SVM on emotional task-related fMRI data, 62% of patients who achieved remission (sensitivity) and 75% of patients who did not achieve remission (specificity) following 8 weeks of fluoxetine treatment could be predicted. But these results were not statistically significant due to small sample sizes (remitters =8, non-remitters=10). Similarly, Costafreda et al (2009) applied SVM to pretreatment structural scans and showed prediction with a sensitivity of 88.9 % and a specificity of 88.9% and accuracy of 88.9% in a small sample comprised of 18 patients 40. In a recent study involving 61 MDD patients, SVM analysis of pretreatment white matter data predicted clinical outcome of refractory and non- refractory depression with an accuracy of 65.22%, sensitivity of 56.2% and specificity of 73.91% 41. Although the results of the later study were statistically significant, the low sensitivity and accuracy may limit its clinical use. Moreover, the structural imaging may not be useful to examine predictive value of early treatment changes in the brain function. In summary, there are no studies, to date that have applied SVM to functional data generated from a large sample for use in evaluating predictive accuracy at the individual level. Main objective : Using machine learning method to examine the predictive value (sensitivity, specificity, accuracy) of resting state and emotional task-related fMRI data collected at the pretreatment time (week 0) and at the early phase of antidepressant treatment (week 2) in the classification of remitters and non-remitters in patients with MDD after 12 weeks of treatment. Secondary objective: To compare the predictive value of pretreatment baseline brain activity (week 0) with early treatment brain activity (week 2). Primary hypothesis: By employing a machine learning method to pretreatment and 2 week post-treatment fMRI data, we hypothesize that it is possible to predict with significant accuracy whether an individual patient with MDD could be classified as remitter or non-remitter at the end of 12 weeks of antidepressant treatment. Secondary hypothesis (Exploratory): Based on previous EEG studies and our preliminary data of standard group comparisons showing early treatment response and associated brain changes, we hypothesize that prediction of antidepressant treatment outcome at an individual level, will be better using fMRI data obtained early in treatment (2 weeks) as compared with pretreatment fMRI data. Rationale: The current study is designed to evaluate the predictive value of early brain changes related to antidepressant treatment to classify remitters and non-remitters based on their clinical response at 12 weeks of treatment. Traditionally, the neuroimaging studies have used group comparisons of pretreatment scans for treatment outcome prediction, which has limited clinical value to make predictions at the individual level. Machine learning methods can provide prediction at the individual level, which can be prospectively used in clinical practice. As the meta-analysis of clinical trials indicate that early clinical response is a reliable predictor of later treatment outcome11,12, our study will examine brain scans at both pretreatment and early post-treatment (2 weeks) times. Experimental Design and Procedure: The eligible patients with MDD will enter into a single blind placebo treatment for two weeks. At the end of two weeks of placebo treatment, participants will be considered as placebo responders based on improvement in depression symptoms as measured by the MADRS scale ( \>50% decrease in MADRS scores from the baseline). The placebo responders will be excluded from the study. The end of two weeks of placebo treatment will be considered as week 0 for active treatment. The placebo non-responders who remain eligible with a score of 22 or higher in MADRS will receive desvenlafaxine 50mg/day for 14 days and the dosage will be increased to 100 mg /day at day 15 if the patient does not improve by 20% reduction in MADRS scores and the dosage determined at day 15 will be maintained until the end of 12 weeks. The first fMRI session will be performed at week 0 (pretreatment baseline) and the second session will be performed at the end of week 2 (14 th day). The participants will be assessed clinically at weeks 1,2,4,6,8,10 and 12 using MADRS, 17-item Hamilton Depression (HAM-D) rating Scale46, Hamilton anxiety (HAM-A) rating scale47 and clinical global impression severity of illness scale (CGI-S) and clinical global impression-improvement scale (CGI-I) 48. HAM-A (Hamilton 1959) will be used to rate anxiety symptoms. To evaluate the overall clinical improvement, CGI-S and CGI-I will be given. Quality of life measure (Q-LES-Q) 49 will be given at the baseline (week 0) and at week 12. Adverse effects will be recorded at each visit. MADRS scores at week 12 will be used to determine remitters and non-remitters. Patients who score less than 10 in MADRS at week 12 will be considered as remitters 50. fMRI Scanning Methods The first fMRI session will be performed at week 0 (pretreatment baseline) and the second session will be performed at the end of week 2. Images will be collected using a Discovery MR750 3T MRI system (GE Healthcare, Waukesha, WI, USA) at the Seaman Family MRI Research Centre at Foothills Hospital, Calgary. Anatomical images will include a 3D T1-weighted MPRAGE image (TR=9.2ms; TE= minimum; flip angle=20; FOV=25.6 cm; voxel size=1mm3. Resting state will consist of a 5-min resting-state scan during which the participants will be asked to keep their eyes closed and hold still ( TR=2000ms;TE=30ms;flip angle=75 degrees; FOV=24 cm; matrix size =64x64, number of slices=36; slice thickness=4mm) Two additional functional MRI scans will also be collected while participants perform an emotional stroop task (block design) using the same acquisition parameters as described for the fMRI resting scan. Machine learning Analysis We will use support vector machines (SVMs), as these have been successfully applied to predicting treatment outcomes in MDD from fMRI data. After preprocessing, SVM as implemented in PROBID software package (http://www.brain.map.co.uk/probid.htm) will be used to investigate the accuracy of whole brain resting and task-related Blood Oxygen Level Dependent (BOLD) data in predicting response to antidepressant treatment. Individual brain scans will be treated as points located in a high dimensional space defined by BOLD response values in the preprocessed images. During the training phase, a linear decision boundary in this high dimensional space will be defined by a "hyperplane" that separates the individual scans according to a class label. SVM classifier will be trained by providing examples of the form \<X,C\> where X represents the fMRI data and C represents the class label (C= 1 for remitters, and C = -1 for non-remitters). Once the hyperplane is determined from the training data, it will be used to predict the class label of a test sample. A linear kernel SVM will be used to extract the weight sector as an image (SVM discriminating map). A "leave-one-out" cross validation method will be used to validate the classifier. This procedure involves excluding a single subject from each group and the classifier will be trained using the remaining subjects. The subject pair excluded will be used to test the ability of the classifier to distinguish between remitters and non-remitters. The procedure will be repeated for each subject pair in the sample in order to assess the overall accuracy of SVM. To establish whether the classification accuracy is statistically significant, we will perform permutation testing. This will involve repeating the classification procedure 1000 times with a random permutation of the training group labels and counting the number of permutations that achieve higher sensitivity and specificity than the one observe with the true labels. The p value will be calculated by dividing this number by 1000. Bonferroni correction or false discovery rate will be used to correct for multiple testing. This SVM analysis and permutation testing will be performed on the pretreatment scan and 2-week post-treatment scan separately. The task-related data and resting state data will be analyzed separately. Sample size calculation The sample size calculation for a classification study is based on precision we want to achieve for sensitivity and specificity. The precision refers to the width of the 95% confidence intervals associated with the estimates. To achieve the 95% confidence interval of plus and minus of 0.16 for 85 % sensitivity and 85% specificity, we will need a total sample of 40 subjects. Having a sample size of 40 should achieve statistically significant classification accuracy and clinically meaningful sensitivity and specificity. Accounting a placebo response of 30% and drop out of 10%, we need to recruit a total of 61 subjects for the sample of 40.