Data-Driven Calibration of Vehicle Pulse Index Parameters: Optimizing Restraint Stiffness for Improved Crash Severity Assessment in Frontal Crashes

Paper i proceeding, 2026

Advanced Automatic Collision Notification (AACN) systems dispatch emergency services automatically using on-board Event Data Recorder (EDR) data, where timely identification of serious occupant injury is paramount. The most widely used severity metric, delta-V (∆V), captures only the peak velocity change and ignores both crash-pulse waveform shape and occupant-vehicle dynamics. The Vehicle Pulse Index (VPI), standardized in ISO/TR 12353-3, models the occupant-restraint system as a spring-mass system driven by the complete crash pulse, but its fixed stiffness (k = 2500 N/m) and slack (s = 0.03 m) do not represent the diverse restraint configurations encountered in practice. This paper proposes a data-driven calibration framework for an EDR-complete subset of frontal-impact crashes that derives case-specific slack distances from EDR restraint-activation timestamps and optimizes spring stiffness k separately for each of four restraint groups (belted and unbelted occupants, each further subdivided by whether restraint systems activated) using a Somers’ D rank-correlation objective. Applied to 1,703 frontal-impact cases from the Crash Investigation Sampling System (CISS), the framework yields optimal stiffness values of 4600N/m, 1900N/m, 1300N/m, and 7300N/m, all substantially different from the ISO default of 2500N/m, and improves Somers’ D with Maximum Abbreviated Injury Scale (MAIS) by 0.5% to 13.3% over the standard VPI across all groups. Bootstrap validation across 100 stratified resamples confirms robust predictive improvement, with mean Somers’ D improvements of 21.7% to 46.7% over ∆V across all groups, while indicating that the optimized stiffness estimates are more precise in larger groups than in smaller groups. In the context of MAIS ≥ 3 prediction for AACN dispatch, calibrated VPI alone reduces the AUUOC under-triage metric by 21.4% and 15.7% at the 5% and 10% false-positive-rate thresholds relative to ∆V alone; the combined ∆V + calibrated VPI model achieves the highest overall AUC. These results suggest that scenario-specific VPI calibration may reflect injury-relevant information related to crash-pulse waveform and restraint compliance beyond peak ∆V and may offer threshold-specific operational value for AACN serious-injury triage in EDR-complete frontal crashes.