Predicting remission following repetitive transcranial magnetic stimulation in treatment-resistant depression: a comparison of parsimonious and ensemble machine learning models.

Repetitive transcranial magnetic stimulation (rTMS) is resource-intensive, and prognostic models may estimate remission probability after treatment in treatment-resistant depression (TRD). Whether algorithmic complexity improves remission prediction over a parsimonious clinical model remains unclear. In a retrospective cohort of 263 adults treated with rTMS who met a baseline data-availability criterion, we compared a simple logistic regression model using age, sex, and baseline MADRS with tuned multivariable regression and SuperLearner ensembles across feature sets ranging from routine clinical variables to questionnaire scores, symptom dimensions, and individual items. Models were evaluated using 10 repeats of stratified 10-fold cross-validation; remission (follow-up MADRS ≤10) was the primary outcome. Seventy-one participants (27.0%) remitted. Across the 12 feature-set/model-strategy configurations, mean calibrated ROC-AUCs ranged from 0.685 to 0.711; LR_Simple yielded 0.707 (across-repeat SD 0.005), and the MADRS-only benchmark yielded 0.708 (0.004). Across eight paired comparisons of LR_Simple with richer or ensemble strategies, ROC-AUC differences ranged from -0.008 to 0.018, every 95% CI included zero, and none differed significantly after BH correction. Random-forest imputation produced similar remission discrimination, whereas models discriminated MADRS response less well than remission. At a 25% threshold, the Base-LR_Simple and Dimensions-SuperLearner configurations yielded 6.0 (95% CI 1.8-10.0) and 7.1 (3.3-10.9) net true-positive decisions per 100 versus treat-all. In this internally validated cohort, richer and ensemble strategies did not demonstrate improved remission discrimination over the a priori three-variable reference model; these findings do not establish equivalence or transportability.