Scenario-first injury-risk phenotyping of autonomous vehicle crashes: distinguishing pre-crash interaction structures from ODD micro-contexts.

OBJECTIVES: Existing autonomous vehicle (AV) crash-injury studies are predominantly variable-centered and often separate scenario classification from injury modeling. This study developed a scenario-first framework to determine whether injury probability conditional on an observed crash is organized primarily by broad pre-crash interaction structures or by localized operational design domain (ODD) conditions. METHODS: A validated database of 2,946 AV crashes recorded from January 2015 to July 2024 was assembled from regulatory reports and verified public sources. Without using injury outcomes, crashes were mapped to five scenario skeletons and clustered within skeletons using Gower distance and k-medoids to derive 10 scenario-ODD phenotypes. Dual-baseline injury enrichment, Bayesian hierarchical models, model comparisons, bootstrap stability checks, and sensitivity analyses assessed adjusted injury differences and robustness. RESULTS: No phenotype met the strong dual-baseline enrichment criterion. S05-P01 had the highest observed injury proportion (51.9%) and global enrichment (RRg = 1.41) but no within-skeleton enrichment (RRss = 1.00). The two vulnerable-road-user phenotypes showed elevated global injury tendencies (RRg = 1.31-1.33), whereas the three rear-end/longitudinal phenotypes had lower injury proportions (RRg = 0.55-0.70). Seven of nine estimable phenotype deviations had 95% credible intervals excluding zero, although their magnitudes were modest. M2 had the lowest Brier score (0.227) and BIC (3939.5) and the highest AUROC (0.719); M5 had the highest AUPRC (0.431) and most favorable calibration slope (0.689). Seven of 10 phenotype classifications were unchanged across all threshold bundles. CONCLUSIONS: Pre-crash interaction structure provided the primary organization of conditional injury probability, while the phenotype layer identified localized within-skeleton variation and translated broad scenarios into operationally interpretable ODD micro-contexts. The two levels support complementary tasks in scenario testing, ODD review, and targeted safety validation.