Accuracy of Machine Learning Models in Predicting Clinical Outcomes in Bipolar Disorder: A Systematic Review.

Area of Science:

• Psychiatry and Mental Health
• Artificial Intelligence in Medicine
• Systematic Review Methodology

Background:

• Bipolar disorder (BD) is a leading cause of global disability, marked by significant functional impairments.
• The complex and heterogeneous nature of BD makes predicting clinical progress and outcomes challenging.
• Artificial intelligence (AI) methods offer potential for enhancing predictive accuracy in BD management.

Purpose of the Study:

• To systematically review the literature on the predictive accuracy of AI methods in forecasting clinical functioning, affective state, and relapse in patients with BD.
• To identify key predictors utilized by AI models for BD outcomes.
• To assess the performance of different AI approaches in BD prediction.

Main Methods:

• Systematic literature search across six electronic databases until July 2025, adhering to PRISMA and Cochrane guidelines.
• Inclusion of 40 articles reporting on AI applications for BD outcome prediction.
• Data summarization and analysis of predictive accuracy metrics, including Area Under the Curve (AUC).

Main Results:

• AI models demonstrated variable predictive accuracy for clinical functioning (AUC 0.59-0.72), affective state (AUC 0.57-0.97), and relapse (AUC 0.45-0.98).
• Supervised, tree-based algorithms generally exhibited superior performance.
• Effective predictive factors encompassed sociodemographic, clinical, psychological variables, wearable data, and speech/video recordings.

Conclusions:

• Machine learning methods show potential for predicting clinical progress and outcomes in BD, including functional status, affective state, and relapse.
• Further longitudinal studies are needed to validate predictive factors for early risk identification and improved BD management.