Does preceding crash information improve classification of injury involvement in subsequent expressway crashes? The role of spatiotemporal linkage.

OBJECTIVES: To determine whether preceding-crash characteristics add information for classifying injury involvement in a later expressway crash after current-crash characteristics and explicit spatiotemporal linkage are considered. METHODS: Administrative crash records from the G2504 Hangzhou Ring Expressway (2020-2021) were used to construct 908 same-direction crash pairs, including 54 injury/fatal cases. Fourteen predictors were organized into current-crash, spatiotemporal, and preceding-crash blocks. Seven classifiers were each optimized over 30 Optuna trials, and the retained classifier was selected according to the mean five-fold training cross-validated precision-recall area under the curve (PR-AUC). A four-configuration ablation analysis assessed the contribution of each feature group, and SHapley Additive exPlanations (SHAP) were used to interpret the retained model. Robustness was examined using 30 strict connected-group-aware splits, temporal adjustment, and alternative pairing windows. RESULTS: XGBoost achieved the highest training cross-validated PR-AUC and was retained for subsequent analyses, although its repeated-split PR-AUC was similar to that of LightGBM. The ablation results showed that the main performance gain came from adding spatiotemporal variables, whereas preceding-crash variables made weak and inconsistent contributions when added alone. Their contribution became more apparent when combined with explicit spatiotemporal context, although the improvement of the Full model over the spatiotemporal configuration on the fixed test set was small and uncertain. SHAP ranked time interval as the most influential predictor, followed by preceding-crash duration, spatial distance, current truck involvement, current lane occupancy, and relative position. The dependence plots showed descriptive changes in fitted contributions around 20-30 min for time interval, 1.0-1.5 km for spatial distance, and 35-40 min for preceding-crash duration. Positive contributions were concentrated mainly in the early, short-distance region. CONCLUSIONS: Spatiotemporal linkage provided the largest and most consistent incremental predictive contribution across the evaluated configurations. Preceding-crash descriptors were weak and split-dependent in isolation and showed suggestive complementary value when linkage was represented. Because the study used completed administrative records and only 54 positive cases, the findings are associational and require prospective external validation before operational use.