Predicting severe injury using vehicle telemetry data

Patricia Ayoung-Chee,Christopher D Mack,Robert Kaufman,Eileen Bulger
DOI: https://doi.org/10.1097/TA.0b013e31827a0bb6
Abstract:Background: In 2010, the National Highway Traffic Safety Administration standardized collision data collected by event data recorders, which may help determine appropriate emergency medical service (EMS) response. Previous models (e.g., General Motors ) predict severe injury (Injury Severity Score [ISS] > 15) using occupant demographics and collision data. Occupant information is not automatically available, and 12% of calls from advanced automatic collision notification providers are unanswered. To better inform EMS triage, our goal was to create a predictive model only using vehicle collision data. Methods: Using the National Automotive Sampling System Crashworthiness Data System data set, we included front-seat occupants in late-model vehicles (2000 and later) in nonrollover and rollover crashes in years 2000 to 2010. Telematic (change in velocity, direction of force, seat belt use, vehicle type and curb weight, as well as multiple impact) and nontelematic variables (maximum intrusion, narrow impact, and passenger ejection) were included. Missing data were multiply imputed. The University of Washington model was tested to predict severe injury before application of guidelines (Step 0) and for occupants who did not meet Steps 1 and 2 criteria (Step 3) of the Centers for Disease Control and Prevention Field Triage Guidelines. A probability threshold of 20% was chosen in accordance with Centers for Disease Control and Prevention recommendations. Results: There were 28,633 crashes, involving 33,956 vehicles and 52,033 occupants, of whom 9.9% had severe injury. At Step 0, the University of Washington model sensitivity was 40.0% and positive predictive value (PPV) was 20.7%. At Step 3, the sensitivity was 32.3 % and PPV was 10.1%. Model analysis excluding nontelematic variables decreased sensitivity and PPV. The sensitivity of the re-created General Motors model was 38.5% at Step 0 and 28.1% at Step 3. Conclusion: We designed a model using only vehicle collision data that was predictive of severe injury at collision notification and in the field and was comparable with an existing model. These models demonstrate the potential use of advanced automatic collision notification in planning EMS response. Level of evidence: Prognostic study, level II.
What problem does this paper attempt to address?