Predicting physician burnout using clinical activity logs: Model performance and lessons learned.

Area of Science:

• Medical Informatics
• Public Health
• Machine Learning

Background:

• Burnout affects over half of the healthcare workforce, posing a significant public health challenge.
• Existing burnout screening tools are often passive and lack efficiency.
• Electronic health record (EHR) audit logs offer a potential source for passive burnout detection.

Purpose of the Study:

• To investigate the efficacy of machine learning (ML) techniques in identifying physician burnout.
• To assess the predictive capability of passively collected EHR audit log data for burnout.
• To explore the utility of clinical workload and EHR usage patterns in burnout prediction.

Main Methods:

• A longitudinal study involving 88 physician trainees who completed monthly burnout surveys.
• Collection of over 10 million EHR-based audit log actions per participant.
• Development and comparison of five ML models (penalized linear regression, support vector machine, neural network, random forest, gradient boosting machine) using workload, temporal EHR use, and baseline burnout features.

Main Results:

• ML models using only workload or temporal EHR features showed limited success in predicting burnout scores (MAE ~0.60, AUROC ~0.59).
• Incorporating baseline burnout scores with workload features significantly improved prediction performance (AUROC ~0.83, accuracy ~0.78).
• The enhanced performance was comparable to using the baseline burnout score alone, suggesting limitations of EHR data in isolation.

Conclusions:

• Predicting burnout solely based on EHR clinical work activities is complex due to its multi-factorial and individualized nature.
• Future burnout prediction models should integrate individual factors (e.g., resilience, sleep) and system-level factors (e.g., leadership).
• Passive screening for burnout using EHR data requires further refinement to capture the full spectrum of contributing elements.