OBJECTIVE: Advanced Automatic Collision Notification (AACN) systems rely on accurate prediction of serious occupant injuries to guide emergency response decisions. Current injury severity prediction (ISP) algorithms predominantly use logistic regression models that assume linear relationships in the log-odds space, potentially overlooking complex nonlinear patterns between crash characteristics and injury outcomes. This study aims to develop an improved ISP algorithm that can capture and explicitly represent these nonlinear relationships while maintaining model interpretability comparable to traditional approaches. METHODS: We developed a prediction model based on trainable B-spline functions using crash data from the US National Automotive Sampling System-Crashworthiness Data System (NASS-CDS, 2010-2015) and Crash Investigation Sampling System (CISS, 2017-2023). The final complete dataset comprised 17,045 crash-involved occupants representing 9,225,347 weighted occupants nationwide. In addition to developing the predictive model using the complete dataset, we also conducted imputation and resampling experiments to demonstrate the distribution of potential model outcomes. Beyond predictors commonly employed in existing AACN algorithms, we incorporated underutilized information related to collision objects, including crash type and hit object type. Model performance was evaluated using both traditional classification metrics and triage-specific measures designed for AACN applications. RESULTS: The proposed model outperformed existing AACN ISP algorithms across all evaluation metrics. Analysis of the trained model revealed that continuous risk factors exhibit distinct nonlinear relationships with serious injury in the log-odds space: delta-V follows an arctangent-like relationship, principal directions of force (PDOF) exhibit a distinct bimodal pattern, and both occupant age and BMI show a Gaussian-like relationship. Among categorical predictors, crash type and hit object type were identified as influential categorical predictors. CONCLUSIONS: Trainable B-spline functions enable effective modeling of complex nonlinear relationships in crash injury prediction while providing explicit mathematical formulations similar to traditional logistic regression. The identification of specific functional patterns for key risk factors enhances understanding of injury mechanisms and provides a foundation for more accurate AACN systems. These findings, including the importance of previously underutilized predictors such as crash type and hit object type, provide a reference for the future development of AACN prediction systems.
Mei et al. (Fri,) studied this question.