Depression rTMS is adjusted session by session: a reinforcement-learning policy trained on real records still agrees at held-out clinics

Each session's choice of stimulation type, intensity and laterality in an rTMS course can now be recommended session by session by a reinforcement-learning policy learned from routine electronic health records, rather than left to experience alone. Clinical guidelines offered no evidence-based guidance for adapting rTMS within a course, so these decisions were largely empirical and practice patterns and outcomes varied substantially. Across 14,074 treatment decisions from 370 patients at six held-out clinic groups, sessions where the clinician chose exactly the policy's treatment showed 0.788 points greater next-session PHQ-9 improvement than sessions where choices differed (95% CI 0.634–0.940).

The policy treats each treatment decision as a choice and learns the sequence of choices with the best cumulative outcome over the whole course, not the best single-session change. It uses a dueling double deep Q-network that estimates a cumulative expected reward for each of 12 treatment combinations and recommends the highest; conservative Q-learning regularization keeps recommendations near treatments with enough recorded use. Training covered 10 clinic groups, 914 patients and 31,804 decisions, with 3 clinic groups for validation and 6 held out entirely for test; patient state included current PHQ-9, episode-start score, prior treatment and motor threshold, plus causal-forest estimates of individualized effects on three treatment dimensions.

The results come from retrospective electronic health records of one national behavioral health organization from January 2023 to June 2025, covering 19 clinic groups, 1609 adults and 56,939 treatment decisions, with evaluation on 370 patients and 14,074 decisions at six entirely held-out clinics. The intended readers are researchers in psychiatric neuromodulation and clinical decision support; the authors state that prospective evaluation in a clinical trial is the next step and suggest a micro-randomized design to estimate proximal effects across repeated decisions.

The authors name unmeasured confounding as the primary limitation: unrecorded intolerance, patient preference or scheduling constraints could influence both clinician choice and later symptoms, and E-values of 1.71 (1.61 at the confidence limit) indicate moderate to severe confounding could explain the result. The reward combines investigator-chosen severity weights, milestone bonuses and suicidal-ideation penalties rather than an empirically identified definition of treatment value; missing PHQ-9 values were carried forward or interpolated, and when improvement counted only for directly recorded scores the association fell to 0.327 points. Data come from one health system, no prospective deployment occurred, and gender identity and race/ethnicity were not consistently recorded, so uniform performance across groups could not be assessed.