Road traffic injuries continue to be a leading cause of death around the world.
Rapid emergency response is a key factor in improving occupant outcomes. Over
the past ten years, Injury Severity Prediction (ISP) models have been developed
and deployed to assist in effective dispatch of emergency medical services
(EMS). Prior versions of ISP have relied on driver-based scenarios that are not
relevant in many of the possible autonomous vehicle (AV) contexts. This paper
describes the development and validation of occupant-based ISP models that
predict injury severity for specific vehicle seat positions. Models show
improved predictive performance, sensitivity 80% and specificity over 95%, for
front row occupants. Second row occupant models have similar specificity, but
sensitivity scores dropped due to occupant heterogeneity and small sample sizes
of seriously injured occupants.