Adverse-effect burden as the dominant correlate of antidepressant non-adherence in major depressive disorder: insights from a multicentre machine-learning analysis.

BACKGROUND: Non-adherence to antidepressants contributes to relapse and recurrence in major depressive disorder (MDD). Existing risk-prediction models are drawn largely from chronic conditions in high-income settings; none exist for Pakistani outpatient psychiatry. METHODS: We analysed cross-sectional, patient-reported data from 2513 adults with MDD attending outpatient clinics in Lahore, Rawalpindi and Quetta. The outcome was an Urdu Morisky-Green-Levine score of 3 or higher, with 31 binary predictors from the Urdu Drug Attitude Inventory and Urdu Antidepressant Side-Effect Checklist. Six classifiers were compared within a pre-specified pipeline (leakage audit, nested cross-validation, bootstrap confidence intervals, calibration, SHAP, decision-curve analysis), reported per TRIPOD+AI. A sensitivity analysis re-fitted the primary model with available covariates. RESULTS: Non-adherence was present in 23.2% of patients. The four strongest models showed similar discrimination (AUC 0.771-0.779), with no significant DeLong difference. Penalised logistic regression, the primary model, achieved AUC 0.774 (95% CI 0.710-0.821), average precision 0.585, Brier score 0.136, calibration slope 1.17, requiring no recalibration. Permutation importance, SHAP, and standardized coefficients showed side-effect burden as the main predictor, with attitude items providing a smaller, mostly protective, contribution. Net benefit remained positive across thresholds 0.10-0.40. Adding covariates did not improve discrimination (AUC 0.759) and did not change adverse-effect coefficients. CONCLUSIONS: A parsimonious, well-calibrated logistic-regression model can identify MDD patients at risk of antidepressant non-adherence, and is to our knowledge the first TRIPOD+AI-compliant adherence model from South Asian psychiatry. Because predictors and outcomes were measured simultaneously and no external validation was performed, the findings only show correlations, not causation; the model isn't ready for deployment as a triage tool.