On Scene Injury Severity Prediction (OSISP) machine learning algorithms for motor vehicle crash occupants in US

Abstract A significant proportion of motor vehicle crash fatalities are potentially preventable with improved acute care. By increasing the accuracy of triage more victims could be transported directly to the best suited care facility and be provided optimal care. We hypothesize that On Scene Injury Severity Prediction (OSISP) algorithms, developed utilizing machine learning methods, have potential to improve triage by complementing the field triage protocol. In this study, the accuracy of OSISP algorithms based on the “National Automotive Sampling System - Crashworthiness Data System” (NASS-CDS) of crashes involving adult occupants for calendar years 2010–2015 was evaluated. Severe injury was the dependent variable, defined as Injury Severity Score (ISS) > 15. The dataset contained 37873 subjects, whereof 21589 included injury data and were further analyzed. Selection of model predictors was based on potential for injury severity prediction and perceived feasibility of assessment by first responders. We excluded vehicle telemetry data due to the limited availability of these systems in the contemporary vehicle fleet, and because this data is not yet being utilized in prehospital care. The machine learning algorithms Logistic Regression, Ridge Regression, Bernoulli Naive Bayes, Stochastic Gradient Descent and Artificial Neural Networks were evaluated. Best performance with small margin was achieved with Logistic Regression, achieving area under the receiver operator characteristic curve (AUC) of 0.86 (95% confidence interval 0.82–0.90), as estimated by 10-fold stratified cross-validation. Ejection, Entrapment, Belt use, Airbag deployment and Crash type were good predictors. Using only a subset of the 5–7 best predictors approached the prediction accuracy achieved when using the full set (14 predictors). A simplified benefit analysis indicated that nationwide implementation of OSISP in the US could bring improved care for 3100 severely injured patients, and reduce unnecessary use of trauma center resources for 94000 non-severely injured patients, every year.