Predicting Serious Injury in ADAS-Involved Crashes: A Comparative Machine Learning Analysis of NHTSA Standing General Order Reports, 2021-2025

Abstract

Advanced Driver Assistance Systems (ADAS) are now standard equipment on most new vehicles, yet what happens to occupants once a crash does occur remains thinly documented outside simulation studies. This study models injury severity using crash reports filed under the National Highway Traffic Safety Administration (NHTSA) Standing General Order, covering January 2021 through June 2025. Records with unknown outcomes were dropped and duplicate filings consolidated, leaving 496 crashes. Severity was treated as binary: serious or fatal injury versus everything else. Twenty raw variables spanning environment, collision configuration, restraint deployment, vehicle attributes, and pre-crash motion were expanded to 80 encoded features, and five classifiers were compared under identical splits. Random Forest performed best, reaching 86.0% accuracy with an AUC-ROC of 0.859; five-fold cross-validation returned 84.6% ± 3.3%. Recall on the serious class was only 45.0%, however, meaning the model missed more than half of the severe outcomes it was built to catch. Manufacturer identity (Tesla, 15.26%) and model year (7.86%) outranked airbag status (11.04%) and collision configuration (8.86%) in the importance ranking. For screening use, the decision threshold must be lowered well below 0.5, trading precision for the sensitivity such applications require.