Uncovering Predictors of Bipolar II Conversion to Bipolar I: A Machine Learning Analysis of National Health Records in Taiwan

Authors

• Chih-Wei Hsu
• Yang-Chieh Brian Chen
• Liang-Jen Wang
• Kuo-Chuan Hung
• Mu-Hong Chen
• Eduard Vieta
• Chien-Yuan Chen
• Chih-Sung Liang
• Andre F Carvalho

Language

English

Publication Date

2-1-2026

Journal

Journal of Affective Disorders

DOI

10.1016/j.jad.2025.120518

PMID

41138948

Abstract

Background: Bipolar II disorder (BD-II) can progress to bipolar I disorder (BD-I), carrying profound psychosocial consequences. However, limited research has investigated the transition from BD-II to BD-I. Our study uses machine learning to predict the conversion using real-world data.

Methods: We conducted a retrospective cohort study using the Taiwan National Health Insurance Research Database (2000-2013) to identify adults diagnosed with BD-II. A predictive model for BD-I conversion was developed using extreme gradient boosting (XGBoost), incorporating demographic factors, healthcare utilization, comorbidities, and psychotropic medication use. The dataset was randomly divided into development (80 %) and test (20 %) sets. Model performance was evaluated by accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC). Feature importance was interpreted using Shapley Additive Explanations (SHAP).

Results: Among the 1890 individuals diagnosed with BD-II, 14 % subsequently converted to BD-I. The final machine learning model demonstrated strong predictive performance, achieving an accuracy of 86 %, sensitivity of 77 %, specificity of 87 %, and an area under the receiver operating characteristic curve (AUC) of 0.91 in the external validation cohort. Younger age, female sex, fewer outpatient visits before and more after diagnosis were key predictors. Lower incidence of new-onset medical (e.g., respiratory, gastrointestinal) and psychiatric (e.g., anxiety, insomnia) comorbidities also indicated higher risk. Increased use of lithium, anticonvulsants, second-generation antipsychotics, and hypnotics further characterized high-risk profiles.

Conclusion: Our findings highlight the utility of machine learning in stratifying BD-II patients at elevated risk for BD-I conversion, offering a data-driven framework to support early clinical intervention and personalized care.

Keywords

Humans, Bipolar Disorder, Taiwan, Machine Learning, Male, Female, Adult, Retrospective Studies, Middle Aged, Young Adult, ROC Curve, Artificial intelligence, Bipolar disorder, Feature, Hypomania, Interpretation, Prediction

Published Open-Access

yes

Recommended Citation

Hsu, Chih-Wei; Chen, Yang-Chieh Brian; Wang, Liang-Jen; et al., "Uncovering Predictors of Bipolar II Conversion to Bipolar I: A Machine Learning Analysis of National Health Records in Taiwan" (2026). Children’s Nutrition Research Center Staff Publications. 414.
https://digitalcommons.library.tmc.edu/staff_pub/414