Machine Learning-Enabled EEG Biomarkers Predict Divergent Antidepressant and Placebo Response in a Clinical Trial of Major Depression

Major depressive disorder (MDD) is a heterogeneous neuropsychiatric disorder with highly variable antidepressant outcomes. In randomized controlled trials (RCTs), low drug-placebo differences and high placebo response rates are persistent challenges. An objective biomarker that can prospectively identify which patients will respond to antidepressant or placebo could greatly enhance both clinical care and clinical trial outcomes. Baseline scalp EEG data from EMBARC, a multi-site RCT of the SSRI sertraline vs placebo in adult MDD, were analyzed using unsupervised machine learning to identify subtypes and compare these with their corresponding treatment response profiles. Subtypes response to sertraline versus placebo was evaluated by 8-week HAMD-17 outcomes (change from baseline). Of the 215 subjects, three EEG clusters yielded four response phenotypes. (1) Drug–Responders exhibited a large sertraline advantage over placebo (n = 124; d = 1.23; p < 0.0001). (2) Non–Responders derived no benefit from sertraline (n = 37; d = –0.07; p = 0.84). (3) Divergent–Responders shared a distinctive connectivity profile clearly separable from phenotypes 1 and 2. Within this group, participants randomized to placebo improved robustly (Placebo–Responders; n = 54; d = –1.52; p < 0.0001), whereas those receiving sertraline worsened (Adverse Drug–Responders; n = 31; d = -0.67; p = 0.004). Excluding Placebo–Responders more than tripled the overall drug–placebo effect size (d = 0.89 vs 0.28). Cluster membership was highly stable in 10–fold cross–validation (98–99 % consistency) and reproduced across three independent trial sites, underscoring generalizability. Scalp EEG activity analyzed with machine learning identified four biomarker-defined subtypes with strikingly distinct responses to an antidepressant and placebo. These results raise the possibility of using low-cost, noninvasive EEG to guide personalized treatment decisions, avoid ineffective or harmful medications, and improve clinical trial outcomes by identifying drug and high placebo responders in advance of initiating treatment.