An Injury Severity Prediction (ISP) algorithm was developed using a logistic regression model to predict the probability of sustaining an Injury Severity Score (ISS) 15+ injury. NASS-CDS (1999-2015) and model year 2000 or later were filters for new case selection criteria, which were based on vehicle body type, to match Subaru vehicle categories. In order to model the effect of PDOF (Principal Direction of Force) as a continuous curve, an analytical method called functional data analysis was employed to create a model of the PDOF curve with two different knot positions (ISP-f1R, ISP-f2R) defined for the periodic basis splines. The most significant variable identified in the current analysis was “Delta V”. “Left impact” was high risk compared to other impact directions in the models. “Belt use” shows a significant risk in the model and belt use was safer than unbelted for any impact direction. “Age” was a significant occupant variable in both models and the presence of a female was notable but not selected in the models. When a right-front passenger was present, the effect on side impact (farside) injury risk is larger than without a right-front passenger. To evaluate model performance, five-fold crossvalidation was performed within the training data (NASS-CDS 1999-2015). The area under the receiver operator characteristic curve (AUCs) was used as the metric to evaluate model performances, AUC was 0.847 with the ISPf1R model, 0.856 with the ISP-f2R model for cross-validation. This study utilizes the field triage of crash subjects and also evaluates the baseline vehicle safety performance of a representative Subaru vehicle categories in realworld crashes.
Keywords:
Advanced Automatic Collision Notification; Functional data analysis; NASS-CDS