An HConvLSTM model with explainable feature weighting for traffic crash injury severity prediction

Traffic crash injury severity prediction provides an important basis for intelligent transportation risk management and emergency resource allocation. This study addressed three limitations in existing structured crash-data modeling: the lack of a stable and interpretable spatial organization for structured variables, the relative scarcity of severe-injury crash samples, and the limited incorporation of periodic context.

A hybrid convolutional neural network-long short-term memory model based on feature-weighted image transformation, named FWIT-HConvLSTM, was proposed to classify fatal and serious injury (FSI) crashes and slight injury (SI) crashes. Continuous feature weights were derived from CatBoost-SHAP explanations estimated exclusively on the training subset within each outer training fold, and structured crash variables were mapped into 16 × 16 semantic feature images. The CNN branch extracted local combinational features from the semantic images, while the periodic-context LSTM branch incorporated time-of-crash context using sine-cosine encodings of Hour, Weekday, and Month. A total of 6,016 crash records collected by the National Automobile Accident In-Depth Investigation System from 2018 to 2022 were used for model development and evaluation.

Under fixed stratified five-fold cross-validation, the five outer test folds yielded mean ± SD values of 0.926 ± 0.005 for Macro-F1, 0.977 ± 0.003 for FSI-class ROC-AUC, and 0.947 ± 0.004 for FSI-class average precision (AP). In 5 × 5 repeated stratified cross-validation, the proposed model obtained the highest mean values for all three evaluation metrics. After Holm correction across 18 pairwise tests, none of the three comparisons with CatBoost reached the 0.05 level; therefore, the small mean differences do not establish statistical dominance over CatBoost. Structural-variant and component comparisons, feature-layout sensitivity analysis, vehicle-damage-variable sensitivity analysis, and model-complexity analysis further indicated that the feature-weighted semantic image representation, CNN-based local feature extraction, and periodic context provided complementary information under the current dataset and experimental setting.

The proposed FWIT-HConvLSTM model showed consistent performance for FSI and SI crash classification under the present experimental setting. The results suggest that feature-weighted semantic image representation, local convolutional feature extraction, and periodic-context modeling provide complementary information for structured crash-data analysis. This study provides a traceable approach to representing structured crash data as interpretable semantic images and ranking FSI risk.