Random Forest Machine Learning Matches Human Expert Accuracy in Trauma Severity Scoring.

Abstract

Background: Accurate Abbreviated Injury Scale (AIS) and Injury Severity Score (ISS) are essential for trauma care and research, yet manual scoring often yields incomplete data due to omissions. The hybrid electronic medical registry (HEMR) is used by our Level 1 trauma service for recording AIS and ISS.

Methods: We analyzed 21,704 patients with trauma records from the HEMR. Four machine learning (ML) algorithms predicted missing AIS scores per body region, from which ISS was derived mathematically. Performance was evaluated using coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), sensitivity (true high-severity cases correctly identified), specificity (true low-severity cases correctly excluded), and Cohen's kappa. Statistical significance was set at p < 0.05.

Results: Random forest models achieved R² = 0.847, RMSE = 2.31, MAE = 1.87, sensitivity = 87.1%, specificity = 100.0%, and Cohen's kappa = 0.893 (p < 0.001), demonstrating reliable prediction of omitted AIS and ISS scores. Data completeness improved from 75.3% (16,343/21,704) to 88.3% (19,158/21,704; p < 0.001), recovering 2815 missing scores.

Conclusion: Random forest ML algorithms accurately predict missing AIS and ISS scores, significantly improving trauma registry data completeness while maintaining clinical accuracy equivalent to human expert scoring.