Mortality prediction of road traffic crash with artificial intelligence: a systematic review

Road traffic crashes cause substantial global mortality and disability. Conventional injury severity scores may not fully capture the complex interactions among demographic, clinical, crash and environmental factors. Artificial intelligence and machine learning may improve mortality prediction by modelling non-linear patterns in traffic crash data.

This systematic review followed Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidance. PubMed/MEDLINE, Web of Science, Scopus and IEEE Xplore were searched on 7 August 2025 for English-language peer-reviewed studies published from 2014 to 2025 that applied artificial intelligence or machine learning to predict mortality after road traffic crashes. Two reviewers screened records and extracted data on study characteristics, data sources, algorithms, predictors, validation, imbalance handling and performance. Methodological quality was assessed using the Qiao quality assessment tool. Because of substantial heterogeneity, findings were synthesised narratively.

18 studies met the inclusion criteria. Most were retrospective studies using structured tabular data. Common objectives were binary mortality prediction, multiclass injury severity prediction including death and death risk assessment. National or regional databases were the most frequent data sources, followed by hospital records and police or insurance datasets. Regression-based models and decision trees remained common, while ensemble methods including random forest and gradient boosting increased in recent years. Frequently reported predictors included age, Injury Severity Scores, body region injured, crash mechanism, temporal factors and geographical characteristics. Only two studies reported external validation.

Artificial intelligence and machine learning show promise for traffic crash mortality prediction, but clinical translation remains limited by insufficient external validation, inconsistent handling of class imbalance, incomplete reporting of tuning and missing data strategies and limited use of explainability methods. Future work should prioritise prospective, externally validated, interpretable models developed using standardised reporting frameworks.