Predicting Dimensional Antidepressant Response to Repetitive Transcranial Magnetic Stimulation using Pretreatment Resting-state Functional Connectivity

Abstract Repetitive transcranial magnetic stimulation (rTMS) is an effective treatment for depression which modulates resting-state functional connectivity (RSFC) of depression-relevant neural circuits. To date, however, few studies have investigated whether individual treatment-related symptom changes are predictable from pretreatment RSFC. We use machine learning to predict dimensional changes in depressive symptoms using pretreatment RSFC. We hypothesized that changes in dimensional depressive symptoms would be predicted more accurately than scale total scores. Patients with depression (n = 26) underwent pretreatment RSFC MRI. Depressive symptoms were assessed with the 17-item Hamilton Depression Rating Scale (HDRS-17). Random forest regression (RFR) models were trained to predict treatment-related symptom changes captured by the HDRS-17, HDRS-6 and three previously identified HDRS subscales: core mood/anhedonia (CMA), somatic disturbances, and insomnia. Changes along the CMA, HDRS-17, and HDRS-6 were predicted significantly above chance, with 9%, 2%, and 2% of out-of-sample outcome variance explained, respectively (all p < 0.01). CMA changes were predicted more accurately than the HDRS-17 (p < 0.05). Higher baseline global connectivity (GC) of default mode network (DMN) subregions and the somatomotor network (SMN) predicted poorer outcomes, while higher GC of the right dorsal attention (DAN) frontoparietal control (FPCN), and visual networks (VN) predicted reduced CMA symptoms. HDRS-17 and HDRS-6 changes were predicted with similar GC patterns. These results suggest that RSFC spanning the DMN, SMN, DAN, FPCN, and VN subregions predict dimensional changes with significantly greater accuracy than syndromal changes following rTMS. These findings highlight the need to assess more granular clinical dimensions in therapeutic studies and echo earlier studies supporting that dimensional outcomes improve model accuracy.