Development and multi-site external validation of a generalizable risk prediction model for bipolar disorder

Area of Science:

• Psychiatry and Mental Health
• Computational Medicine
• Health Informatics

Background:

• Bipolar disorder significantly contributes to disability, premature mortality, and suicide.
• Early identification of bipolar disorder risk is crucial for targeted interventions and efficient resource allocation.
• Existing predictive models often lack generalizability across diverse populations.

Purpose of the Study:

• To develop and validate generalizable predictive models for bipolar disorder risk.
• To assess the feasibility of using electronic health record (EHR) data for prediction across diverse sites.
• To compare the performance of various machine learning algorithms for bipolar disorder risk prediction.

Main Methods:

• An observational case-control study was conducted within the PsycheMERGE Network.
• Predictive models were developed using random forests, gradient boosting machines, penalized regression, and stacked ensembles.
• Data from three diverse academic medical centers with linked EHRs were utilized, including demographics, diagnostic codes, and medications.

Main Results:

• The study included over 3.5 million patient records, with 12,533 cases of bipolar disorder.
• Stacked ensemble models demonstrated optimal performance, with AUC ranging from 0.82-0.87.
• High-risk quantiles showed positive predictive values above 5% across all study sites.

Conclusions:

• Generalizable predictive models for bipolar disorder risk can be developed across diverse healthcare settings.
• Machine learning, particularly ensemble methods, shows strong potential for precision medicine in mental health.
• These models, to be disseminated via the PsycheMERGE Network, can aid in early identification and resource allocation.