Reinforcement learning rTMS for depression treatment

A new npj Digital Medicine study tests whether a reinforcement learning rTMS model, trained on routine clinic records, can recommend stimulation protocols that go with better symptom improvement at the next session.

The retrospective cohort drew on records from 1,609 adults with treatment-resistant depression who received repetitive transcranial magnetic stimulation between January 2023 and June 2025 at 19 clinic groups of Discovery Behavioral Health. The model was evaluated only at sites it had not encountered in training.

How the model works

Each rTMS course usually runs about 36 daily sessions over six to seven weeks. At each session the clinician selects from 12 possible treatments, defined by stimulation type (standard 10-Hz rTMS or intermittent theta-burst stimulation, iTBS), intensity relative to motor threshold, and laterality. The authors applied a double deep Q-network with a dueling architecture. It estimates the cumulative expected reward of each treatment for a given patient state, then selects the treatment with the highest value.

Ad

The reward rested on next-session PHQ-9 improvement, with added weight for gains from more severe symptoms. Remission (PHQ-9 below 5) carried a bonus, response (a reduction of at least 50%) carried a smaller one, and worsening or suicidal ideation at the next session carried penalties. The authors note that these reward weights were chosen by investigators, not identified empirically as the best definition of treatment value.

Reinforcement learning rTMS: agreement and next-session improvement

On the held-out test clinics (6 clinic groups, 370 patients, 14,074 treatment decisions), the algorithm’s recommendation matched the clinician’s choice in 82.1% of sessions. In sessions where the two matched, PHQ-9 improvement at the next session was 0.788 points greater than in sessions where they differed (95% CI: 0.634 to 0.940). The gap was largest early in treatment and in patients with moderately severe depression.

Source: Gibbons JB et al., npj Digital Medicine, Article in Press, DOI 10.1038/s41746-026-03417-4, Results and Supplementary Table 6 as reported in the text. CIs were computed by resampling whole patients (2,000 replicates).

The authors ran several checks on the main association. Adjusting for current PHQ-9, recent change and treatment phase gave 0.562 points (95% CI: 0.434 to 0.718). Giving each patient equal weight gave 0.850 points. Point estimates were positive in all ten calendar quarters, and the confidence interval excluded zero in nine of them.

Comparison with simple policies

A policy that always recommended the most frequently observed treatment produced 0.272 points. A policy that chose treatments at random produced 0.011 points. Retrained models with alternative reward definitions gave associations between 0.706 and 1.172 points. Removing the suicidal ideation penalties gave 0.706 points (95% CI: 0.559 to 0.872).

The authors estimated how strong an unmeasured confounder would need to be to explain the result. The E-value was 1.71 for the point estimate and 1.61 for the confidence limit. The authors describe this as moderate to severe confounding, which they say would be needed to explain the results entirely.

Projected gains across a treatment episode

The authors also projected what would happen if clinicians followed every recommendation. In the first observed treatment episode for each patient, the projection was 3.22 additional PHQ-9 points (95% CI: 2.72 to 3.75). That is close to a published average minimal clinically important difference of 3.7 points. Under observed care, 190 of 370 patients (51.4%) met their baseline-specific threshold. The projection put that figure at 236 patients (63.8%), an increase of 46 patients or 12.4 percentage points (95% CI: 8.6 to 17.6).

In 102 later observed episodes from 74 patients, the association stayed positive at 0.508 points. Across all 472 observed episodes, the projected additional improvement was 3.15 points (95% CI: 2.68 to 3.60). These projections assume the association holds in the same patients, and the authors present them as exploratory.

What the model prioritised

The model’s largest source of influence was the previous treatment, at 17.0% of total SHAP importance. Motor threshold and depression-severity measures together accounted for more than 25%. The authors read this as evidence that the model learned treatment sequencing, and that it also responded to dosing physiology and symptom trajectory.

The model recommended iTBS more often than clinicians did: 7.0% of treatment decisions against 4.5%. In sessions 1 to 5 the gap was wider, at 19.2% against 3.3%. From session 6 onward the rates were 4.9% and 4.7%. The policy also changed its recommendation more often after a disappointing session, switching after 13.0% of decisions against 7.8% for clinicians.

Why it matters for providers and payers

For health economists and market access teams, the value of this work lies in the estimates of heterogeneity and projected benefit, not in a ready-to-use tool. The study cannot yet show that following the algorithm improves outcomes, because the association is observational. Our coverage of non-invasive neuromodulation places rTMS among a wider set of devices now moving from invasive to non-invasive approaches. For the question of whether AI support can reach routine use, see our piece on evaluating AI in healthcare, which notes that only a small share of clinical decision support models progress beyond prototyping. Digital tools for depression are also reaching patients, as our look at transdiagnostic mental health apps shows.

The study used one provider’s routine data, so its cost and reimbursement implications remain untested. Any payer assessment would need the prospective trial the authors call for, along with an explicit view on how much a point of PHQ-9 improvement is worth.

Source: Gibbons JB, Leenaerts N, Siddiqi SH, Murphy SA, et al. Reinforcement learning for repetitive transcranial magnetic stimulation protocol selection in treatment-resistant depression. npj Digital Medicine, Article in Press, received 4 June 2026, accepted 5 October 2026, DOI 10.1038/s41746-026-03417-4.