Enriquez, 1 PREDICTING SEVERE INJURY IN MOTOR VEHICLE CRASHES

Research Question/Objective: The National Highway Traffic Safety Administration (NHTSA) is actively studying the implementation of Advanced Automatic Collision Notification (AACN) systems in motor vehicles. This technology allows motor vehicles to notify a Public Safety Answering Point (PSAP), such as a 911 call center, in the event of a severe crash. The system provides crash location, vehicle identification information, as well as a prediction of severe injury to occupants in the motor vehicle. This paper describes the development of a statistical model that predicts the presence of severely injured and fatal occupants in a motor vehicle involved in a crash. Methods and Data Source: A logistic regression model was developed using data from the 1999 – 2015 Crashworthiness Data System (CDS) of the National Automotive Sampling System (NASS). The binary response variable indicates whether or not a crashed vehicle contains a severely injured occupant or a fatally injured occupant, defined by an Injury Severity Score (ISS) of 16 or greater. The predictors are those recommended by the Centers for Disease Control and Prevention (CDC) National Expert Panel on Field Triage, which are delta-V, vehicle body type, multiple vs. single impact, seat belt usage, and principal direction of force. The final dataset is at the vehicle level. Results: The area under the receiver operator characteristic curve (AUC) was 0.843, indicating that the model was able to discriminate between vehicles with and without severely injured occupants. At the CDC recommended 0.20 risk threshold, the model produced a sensitivity rate of 26%, a specificity rate of 99%, and identified 41% of vehicles with a fatally injured occupant. Conclusion: The sensitivity rate at the CDC recommended 0.20 risk threshold missed 59% of vehicles with a fatally injured occupant. A preliminary cost-benefit analysis showed that the optimal threshold was close to 0.008 after considering the cost of lives saved versus the cost of overtriaging minor injured people using the AACN algorithm. At the 0.008 threshold, 92% of fatal occupants are predicted, the sensitivity is 91%, and the specificity is 60%, which comes close to the recommended levels by the American College of Surgeons. Limitations: An AACN system uses data from the event data recorder (EDR) of a vehicle; however, the model developed in this paper was trained with data collected from crash investigations, which may differ from EDR data. Also, this paper only considered the logistic regression model, whereas other data mining classifiers which may produce better results. The initial set of predictors was limited to those selected by the CDC Expert Panel.