TITLE:
Machine Learning Models for Predicting Antidepressant-Induced Mania in Bipolar Disorder: A Synthetic Proof-of-Concept Simulation Study
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
bipolar Disorder, Antidepressant-Induced Mania, Machine Learning, Precision Psychiatry, Predictive Modelling, Polygenic Risk Scores, Treatment Optimization
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.5,
May
29,
2026
ABSTRACT: Antidepressant-induced mania represents a significant clinical challenge in the management of bipolar depression, with incidence rates ranging from 5% - 20% in clinical populations. Despite the widespread use of antidepressants in bipolar disorder, reliable methods for predicting individual susceptibility to manic switches remain elusive. This study presents a synthetic proof-of-concept simulation to evaluate machine learning models to predict antidepressant-induced mania using comprehensive clinical, genetic, and pharmacological data. We generated a synthetic clinical dataset of 2000 patients with bipolar disorder based on established clinical risk factors and epidemiological parameters. Five machine learning algorithms Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine (SVM), and Neural Network were trained and validated using 5-fold cross-validation. Model performance was evaluated using AUC-ROC, precision-recall metrics, and calibration analyses. Feature importance analysis identified key predictive variables. The Gradient Boosting model achieved the highest predictive performance (AUC-ROC = 0.926, 95% CI: 0.901 ± 0.011), followed by Random Forest (AUC-ROC = 0.909). The model successfully stratified patients into four risk quartiles with observed mania rates ranging from 0% in the lowest risk group to 61% in the highest risk group. Key predictive features included bipolar disorder subtype (Type I), absence of concurrent mood stabilizer treatment, rapid cycling history, polygenic risk scores for mania vulnerability, and antidepressant class (tricyclic antidepressants and MAOIs). Machine learning models demonstrate excellent predictive accuracy for antidepressant-induced mania and enable clinically actionable risk stratification. These findings demonstrate methodological feasibility within a simulated environment and establish a principled framework for future validation in real-world clinical cohorts. No clinical deployment conclusions can be drawn from synthetic data alone.Subject AreasPsychiatry & Psychology