Abstract
Many previous intervention studies have used functional magnetic resonance imaging (fMRI) data to predict the antidepressant response of patients with major depressive disorder (MDD); however, practical constraints have limited many of those attempts to small, single centre studies which may not adequately reflect how these models will generalize when used in clinical practice. Not only does the act of collecting data at multiple sites generally increase sample sizes (a critical point in machine learning development) it also generates a more heterogeneous dataset due to systematic differences in scanners at different sites, and geographical differences in patient populations. As part of the Canadian Biomarker Integration Network in Depression (CAN-BIND-1) study, 144 MDD patients from six sites underwent resting state fMRI prior to starting escitalopram treatment, and again two weeks after the start. Here, we consider ways to use machine learning techniques to produce models that can predict response (measured at eight weeks after initiation), based on various parcellations, functional connectivity (FC) metrics, dimensionality reduction algorithms, and base learners, and also whether to use scans from one or both time points. Models that use only baseline (pre-treatment) or only week 2 (early-response) whole-brain FC features consistently failed to perform significantly better than default models. Utilizing the change in FC between these two time points, however, yielded significant results, with the best performing analytical pipeline achieving 69.6% (SD 10.8) accuracy. These results appear contrary to findings from many smaller single-site studies, which report substantially higher predictive accuracies from models trained on only baseline resting state FC features, suggesting these models may not generalize well beyond data used for development. Further, these results indicate the potential value of collecting data both before and shortly after treatment initiation.