Beyond Treatment Failure

Objective. Treatment-resistant depression (TRD) affects 30–50% of patients with major depressive disorder, representing the most disabling and economically costly subgroup in psychiatry. Clinical decision-making remains reactive, guided by iterative pharmacological escalation rather than prospective biomarker-guided stratification. Existing predictive models are predominantly unimodal and lack the interpretability and privacy guarantees required for real-world clinical deployment.

Methods. We developed a Multimodal Explainable Federated Stacking Ensemble (MEFSE) integrating structured clinical, quantitative EEG, and digital phenotyping features across 1,200 synthetic patients at three simulated clinical sites. Random Forest and XGBoost base learners were combined via out-of-fold stacking with Transformer and logistic regression meta-learner alternatives. SHAP TreeExplainer provided joint feature attribution. Federated training applied FedAvg with calibrated Gaussian differential privacy across five epsilon levels. Clinical utility was quantified via Brier score and decision curve analysis.

Results. MEFSE achieved AUC = 0.922 (95% CI: 0.862–0.972), F1 = 0.831, and the lowest Brier score (0.082) among all competitors. Decision curve analysis confirmed superior net benefit across threshold probabilities 0.25–0.60. McNemar's test revealed significant classification advantages over logistic regression (p = 0.009) and Random Forest (p = 0.016). Federated training preserved 99.2% of model utility at strong privacy (ε = 0.1).

Conclusions. MEFSE provides a rigorous, interpretable, and privacy-preserving benchmark for TRD risk stratification. The synthetic cohort constitutes a validated methodological template; the framework is architecturally complete and ready for extension to real-world multi-site clinical data.