Subaru has developed vehicle-based Injury Severity Predictions (ISP) models using
data from the National Automotive Sampling System Crashworthiness Data System
(NASS-CDS) covering calendar years 1999–2015, for integration into Advanced
Automatic Collision Notification (AACN) systems. This study evaluates the
accuracy of these ISP models by comparing predictions derived from Subaru
vehicle telemetry with actual Injury Severity Scores (ISS) of transported
occupants. Two crash databases were utilized: Subaru Telematics Assisted
Accident Research (STAAR) data for calendar years 2021–2024, which includes
Automatic Collision Notification (ACN) data, police reports, emergency medical
services (EMS), and medical records from the medical centers across Michigan;
and the Fatality Analysis Reporting System (FARS) data for calendar years
2021–2023, matched with ACN data to supplement serious injury cases. ISS values
were obtained from medical records in STAAR, while fatal cases in FARS were
assigned as fatal injury. Four ISP models were evaluated, grouped into two main
approaches: (1) models using categorical impact directions (Front, Right, Rear,
Left), (2) models applying functional data analysis with cyclic spline modeling
of Principal Direction of Force (PDOF). The presence of a right-front passenger
was also considered as an interaction factor. Among 56 STAAR cases, only one
involved serious injury (ISS ≥ 15), limiting sensitivity analysis. All models
demonstrated specificity above 90%. In 102 FARS cases, 44 were fatal, yielding
sensitivity between 52% and 57%. Models using PDOF splines performed similarly
to directional models. When multiple impacts were excluded, sensitivity improved
to 63%–74%, suggesting that in such crashes, PDOF may not be clearly
identifiable from vehicle telemetry data. Although functional data analysis was
expected to enhance sensitivity, this improvement was not confirmed. Additional
data collection is needed to improve ISP accuracy. Vehicle telemetry remains a
rapid and cost-effective method for acquiring crash data to advance vehicle
safety.
