Predicting return to work after traumatic brain injury using machine learning and administrative data.

Area of Science:

• Medical Informatics
• Rehabilitation Medicine
• Public Health

Background:

• Accurate patient-specific predictions for return-to-work after traumatic brain injury (TBI) are crucial for clinical practice and policy.
• Machine learning (ML) applied to large administrative datasets offers a promising avenue for developing prognostic models.

Purpose of the Study:

• To assess the accuracy of predicting return-to-work one year post-TBI using administrative data.
• To explore how model performance and feature importance vary when distinguishing between mild and moderate-to-severe TBI.

Main Methods:

• Utilized a population-based dataset combining Belgian discharge, claims, and social security data for TBI patients hospitalized in 2016.
• Employed three ML algorithms: elastic net logistic regression, random forest, and gradient boosting.
• Evaluated model performance using accuracy, sensitivity, specificity, and ROC AUC.

Main Results:

• All algorithms yielded similar results, achieving 83% accuracy (ROC AUC 85%) for binary employed/unemployed classification.
• A multiclass operationalization of employment outcome reached 76% accuracy (ROC AUC 82%).
• Separate modeling of mild vs. moderate-to-severe TBI did not significantly alter model performance or feature importance; pre-injury employment was the primary predictor.

Conclusions:

• Administrative data provides valuable insights for return-to-work prediction post-TBI.
• Enhancing patient-specific prognoses necessitates supplementing administrative data with additional information.