Simulation-Based Bayesian Modelling of Antidepressant Discontinuation Risk in Bipolar Disorder: Integrating Side-Effect Profiles, Therapeutic Alliance Scores, and Synthetic Prescription Records

Antidepressant discontinuation in bipolar disorder is a clinically important contributor to treatment failure. Side-effect burden, therapeutic alliance, and prescription-adherence behaviour may jointly influence unplanned cessation. This proof-of-concept study evaluates whether those domains can support discontinuation-risk modelling in a fully synthetic cohort; it does not analyse real patient records or provide clinical validation. We generated a synthetic cohort of N = 900 antidepressant-exposed patients with bipolar disorder and a simulated 90-day discontinuation prevalence of 35.0%. Five feature domains represented side effects, therapeutic alliance, synthetic prescription records, patient-reported outcomes, and clinical variables. Random Forest, XGBoost, and LightGBM were combined through five-fold out-of-fold stacking with a logistic-regression meta-learner. A MAP-regularised Bayesian logistic regression served as the principal probabilistic reference. SHAP analyses and domain-specific visualisations were used for interpretation. BIC, WAIC, and Bayes-factor calculations for non-likelihood models were retained only as exploratory approximations. Bayesian logistic regression achieved the highest AUC (0.939) and lowest Brier score (0.095). The stacked ensemble achieved AUC = 0.913 (95% CI: 0.860 - 0.956), F1 = 0.816 (CI: 0.722 - 0.891), sensitivity = 0.817, specificity = 0.896, and Brier score = 0.109. Therefore, the ensemble was not superior in discrimination or calibration; its value lies in modelling non-linear interactions and supporting complementary interpretation. Within the simulator, prior gap count, refill adherence, patient-clinician concordance, medication-harm beliefs, and prior discontinuation history were the dominant SHAP predictors. These results demonstrate methodological feasibility in synthetic data only. The feature rankings and apparent clinical patterns may partly reflect the data-generation rules and require external validation using real dispensing, clinical, and therapeutic alliance data before any decision-support use.