S. Diaz1, A. Sinha1, E. Baker1, R. Karimi1, S. Maloney1, A. Victory1, R.A. Jean1 1University Of Michigan, Acute Care Surgery, Ann Arbor, MI, USA
Introduction: Despite improved safety measures, motor vehicle crashes (MVCs) remain a source of preventable death. This may be in part due to the inability of emergency services to rapidly assess, stabilize, and transport patients to high-level trauma centers because the decision tools available in the prehospital setting are severely limited. Machine Learning(ML)-enabled decision support systems offer a significant promise in guiding field triage after MVCs. ML-based prediction models incorporate clinical, demographic, environmental, and other information to estimate the probability of important clinical events and have demonstrated efficacy across multiple clinical situations and environments.
Methods: We performed a retrospective study using the National Highway Traffic Safety Administration’s Crash Investigation Sampling System (NHTSA-CISS) dataset from 2016-2022. The NHTSA-CISS dataset contains representative sample of MVCs for safety analysis, including details about the vehicle and occupants. We compared model performance between logistic regression and ML methods for predicting either the presence of any injury, defined as an injury severity score greater than 0 (ISS>0) or severe injury (ISS>15), for any occupant in the vehicle using patient and crash-specific parameters. Performance was compared using area under the receiver operator curve (AUC).
Results: A total of 11,006 occupants involved in MVCs were identified over the study period. Of this group 5,388 (48.9%) had some injury, and 324 (2.9%) had severe injuries. Both the LR and ML models identified several patient a vehicle factors that were strongly predictive for injury, including belt use, occupant sex, left-sided or frontal crashes, crash velocity, and occupant body mass index. Overall the ML model performed similarly to the logistic regression model in predicting severe injury (AUC 0.671 vs 0.678), however had slightly better performance in detecting ISS>15 (AUC 0.683 vs 0.643).
Conclusion: There exists significant potential in developing automated ML regression models for predicting patient injury after car crashes. Further investigation is necessary to optimize these novel prediction models.