Explainable machine learning for crash injury severity prediction

Purpose This study evaluates explainable machine learning for classifying injury severity conditional on a reported traffic crash.Design/methodology/approach The analysis uses 603,736 NHTSA Crash Report Sampling System records from 2018–2022 and three severity classes: no apparent injury, minor injury, and severe/fatal injury. Models are developed on 2018–2021 records and evaluated on 2022 records. Outcome-proximal variables are excluded from the primary models. A class-weighted random forest is compared with logistic, decision-tree, and majority-class baselines using per-class metrics, average precision, and calibration diagnostics. SHAP explains model predictions.Findings Random forest achieves 60.5% severe/fatal recall and 22.1% precision. The logistic baseline performs better on several aggregate and probability-based metrics, indicating that model selection depends on the evaluation criterion. Limited precision and calibration restrict practical interpretation.Originality/value The study combines temporal evaluation, predictor-timing screening, and class-specific reporting to support transparent crash-severity analysis. Explanations are model-based associations, not causal effects. Processed-data provenance, road-user heterogeneity, pandemic-period changes, and the absence of exposure/non-crash data limit generalizability and operational use.