Pre-Initiation Prediction of Antidepressant-Associated Cycle Acceleration in Bipolar Disorder: A Synthetic Proof-of-Concept Study

Antidepressant-associated cycle acceleration (AICA), encompassing antidepressant-related switching, rapid-cycling induction or acceleration, and mixed-state emergence, is an important treatment-safety concern in bipolar disorder (BD). Because these outcomes may be difficult to distinguish from spontaneous illness progression, pre-initiation risk stratification is clinically relevant. This proof-of-concept simulation study evaluates whether a multivariate machine-learning framework can recover prespecified AICA-risk structure from synthetic data; it does not test clinical effectiveness or establish a validated prediction tool. We developed an interpretable stacked ensemble using an entirely synthetic cohort of N = 850 simulated antidepressant-exposed patients with BD. No hospital records, registry observations, individual patient data, or hybrid clinical-synthetic records were used. The feature architecture integrated pharmacological history, episode chronology, and circadian digital biomarkers. Random Forest, XGBoost, and LightGBM were combined through a five-fold out-of-fold logistic-regression meta-learner and compared with five baseline models. Evaluation included held-out discrimination, calibration, decision-curve analysis, SHAP attribution, and exploratory model-fit summaries. On the held-out synthetic test set, the ensemble achieved AUC = 0.939 (95% CI: 0.889 - 0.976), F1 = 0.850 (95% CI: 0.754 - 0.929), sensitivity = 0.839 (95% CI: 0.704 - 0.949), and specificity = 0.946 (95% CI: 0.894 - 0.989). XGBoost had the lowest Brier score (0.077), followed closely by the ensemble (0.079). The ensemble showed the highest estimated net benefit across the reported threshold range in this synthetic test set. SHAP attribution ranked prior AICA history, interdaily stability, mood stabiliser adequacy, antidepressant-class risk, and mixed-episode fraction as the leading predictors, reflecting relationships embedded in the simulation design. The results demonstrate internal recovery of a synthetic risk-generating structure rather than prospective clinical validity. The model should therefore be regarded as a methodological and hypothesis-generating framework. Full disclosure of the simulator, robustness analyses under alternative label-generating assumptions, and validation in independently collected real-world cohorts are required before any prescribing recommendation, contraindication threshold, or clinical decision-support use can be considered.