Machine learning to predict passenger mortality and hospital length of stay following motor vehicle collision

Abstract

OBJECTIVE Motor vehicle collisions (MVCs) account for 1.35 million deaths and cost $518 billion US dollars each year worldwide, disproportionately affecting young patients and low-income nations. The ability to successfully anticipate clinical outcomes will help physicians form effective management strategies and counsel families with greater accuracy. The authors aimed to train several classifiers, including a neural network model, to accurately predict MVC outcomes. METHODS A prospectively maintained database at a single institution’s level I trauma center was queried to identify all patients involved in MVCs over a 20-year period, generating a final study sample of 16,287 patients from 1998 to 2017. Patients were categorized by in-hospital mortality (during admission) and length of stay (LOS), if admitted. All models included age (years), Glasgow Coma Scale (GCS) score, and Injury Severity Score (ISS). The in-hospital mortality and hospital LOS models further included time to admission. RESULTS After comparing a variety of machine learning classifiers, a neural network most effectively predicted the target features. In isolated testing phases, the neural network models returned reliable, highly accurate predictions: the in-hospital mortality model performed with 92% sensitivity, 90% specificity, and a 0.98 area under the receiver operating characteristic curve (AUROC), and the LOS model performed with 2.23 days mean absolute error after optimization. CONCLUSIONS The neural network models in this study predicted mortality and hospital LOS with high accuracy from the relatively few clinical variables available in real time. Multicenter prospective validation is ultimately required to assess the generalizability of these findings. These next steps are currently in preparation.