TITLE:
AI-Driven Prediction of Antidepressant-Induced Mood Switching in Bipolar Disorder: A Synthetic Proof-of-Concept Machine Learning Study Using Clinical, Temporal, and Biomarker Data
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Bipolar Disorder, Antidepressants, Treatment-Emergent Affective Switching, Machine Learning, Predictive Modelling, Biomarkers, Precision Psychiatry, Clinical Decision Support
JOURNAL NAME:
Open Access Library Journal,
Vol.13 No.5,
May
29,
2026
ABSTRACT: The use of antidepressants in bipolar depression remains one of the most challenging decisions in psychiatric practice, with substantial risks of treatment-emergent affective switching (TEAS) to manic, hypomanic, or mixed states. Current clinical guidelines rely primarily on symptom history and clinical intuition, lacking objective, individualized biomarkers to predict treatment response and switching risk. We introduce a comprehensive machine learning framework that integrates clinical history, temporal patterns, and biomarker data to predict antidepressant-induced mood switching in bipolar disorder. Our approach employs multiple advanced algorithms including Logistic Regression, Random Forest, Gradient Boosting, and Support Vector Machines (SVM) trained on a carefully curated dataset of 2500 synthetic patients with realistic clinical characteristics. The framework incorporates established risk factors including bipolar subtype, rapid cycling status, mood state at prescription, antidepressant class and dosage, concomitant mood stabilizer use, and novel biomarkers including BDNF levels, inflammatory markers (CRP), cortisol, and polygenic risk scores. The Gradient Boosting classifier achieved superior performance (AUC-ROC = 0.861, F1-Score = 0.58) compared to other algorithms, with Random Forest demonstrating the highest precision (0.53) for identifying high-risk patients. Feature importance analysis revealed bipolar Type I diagnosis, rapid cycling, mixed mood states, and sleep deprivation as the strongest predictors of switching risk. Risk stratification analysis successfully categorized patients into five distinct risk tiers, with the highest risk group showing 85% observed switch rates versus 3% in the lowest risk group. Beyond prediction, we provide interpretable clinical decision support through partial dependence analysis, calibration plots, and decision curve analysis demonstrating clinical utility across threshold probabilities. Our framework addresses the critical need for precision psychiatry tools in bipolar disorder management, offering a pathway toward personalized antidepressant prescribing those balances therapeutic benefit against switching risk. This work constitutes a synthetic proof-of-concept study; all models were trained and evaluated exclusively on computationally generated data. Results demonstrate methodological feasibility and should not be interpreted as evidence of clinical deployability. External validation on real-world patient cohorts is required before clinical implementation.