Two‑tier prediction framework with OR‑voting for real‑time freeway secondary-crash prediction: Balancing sensitivity...

Secondary crashes on freeways pose significant safety risks and are often preventable with timely intervention. This study aims to develop a real-time prediction framework for secondary-crash risk using traffic flow precursor characteristics, enabling proactive traffic safety management.

A novel secondary-crash identification method based on a crash buffer and speed contour map was first proposed to accurately determine the spatiotemporal influence range of primary crashes. Using historical crash data from Interstate I-405, traffic flow features (including flow rate, speed, and occupancy) were extracted. Four machine learning algorithms-Support Vector Machine (SVM), Random Forest (RF), XGBoost, and CatBoost-were employed to construct prediction models. A two-tier prediction framework was developed, consisting of a primary-crash risk prediction sub-model and a secondary-crash risk prediction sub-model. The outputs of the two sub-models were integrated using a voting strategy. Model performance was evaluated using Accuracy, Precision, Recall, F1-score, and AUC.

The proposed identification method effectively distinguished secondary crashes from historical data. Among the tested algorithms, CatBoost demonstrated the best overall performance in both sub-models. The two-tier prediction framework outperformed single‑tier models, achieving higher recall and AUC values for secondary‑crash detection. Although the integrated model slightly reduced precision, it significantly improved the identification of secondary-crash events, which is critical for safety-sensitive applications. The model also provided probabilistic risk outputs, enhancing interpretability for traffic management decision-making.

This study presents an effective and practical framework for real-time prediction of freeway secondary-crash risk. The proposed two-tier prediction framework, combined with traffic flow precursor features and machine learning techniques, improves prediction accuracy and enhances early warning capabilities. The findings provide theoretical and methodological support for proactive traffic safety management and secondary-crash prevention. Future work should incorporate more diverse datasets and external factors (e.g., weather and road conditions) to further improve model generalizability.