TITLE:
Deep Reinforcement Learning for Personalized Antidepressant Decision Support in Bipolar Spectrum Disorders: Simulated Randomized Trial Framework
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Bipolar Disorder, Reinforcement Learning, Precision Psychiatry, Treatment Optimization, Causal Inference, Machine Learning, Antidepressant, Mood Destabilization, Deep Learning, Personalized Medicine
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.5,
May
28,
2026
ABSTRACT: Treatment selection for bipolar depression remains largely trial-and-error, with substantial non-response to first-line strategies and clinically meaningful risk of mood destabilization. We developed a deep reinforcement learning (RL) framework to optimize treatment selection while explicitly penalizing destabilization events. We implemented RL-CADENCE, a simulated multi-centre experimental framework designed to emulate a parallel-group randomized trial across 12 virtual psychiatric centres. Using a combination of publicly available online data sources and clinically informed synthetic generation, we constructed a cohort of 2500 virtual participants representing bipolar spectrum disorders. Virtual participants were algorithmically allocated (3:3:3:1) to four treatment strategies: 1) lithium + SSRI, 2) quetiapine + lamotrigine, 3) lurasidone + mood stabilizer (lithium or valproate), or 4) RL-personalized treatment selection. The primary endpoint was the simulated change in Montgomery ?sberg Depression Rating Scale (MADRS) score over 12 months. Secondary outcomes included response, mood destabilization events, and quality-adjusted life years (QALYs). A causal machine learning pipeline estimated conditional average treatment effects (CATE) to characterize heterogeneity across subgroups within the synthetic cohort. In simulation, the RL-personalized strategy achieved greater MADRS improvement than pooled standard protocols (mean difference: ?5.6 points; 95% CI: ?7.6 to ?3.6; Cohen’s d = 0.78). Simulated response rates (≥50% MADRS reduction) were 95.7% versus 58.9%, and mood destabilization occurred in 4.8% versus 10.8% of synthetic patient-months. The RL policy network achieved an AUC-ROC of 0.89 for predicting the optimal treatment strategy under the simulated counterfactual evaluation. Heterogeneous effects were largest in mixed features (CATE = 15.2; 95% CI: 8.9 - 21.5) and bipolar I subtype (CATE = 12.3; 95% CI: 7.1 - 17.5). Within a simulated, synthetic-data evaluation, deep RL showed strong potential to personalize antidepressant-related treatment selection in bipolar spectrum disorders, improving depressive symptom outcomes while reducing destabilization risk. These findings provide proof-of-concept for RL-based precision psychiatry and motivate prospective validation in real-world clinical cohorts.