Road traffic crashes cause significant fatalities and injuries globally, imposing substantial socio-economic costs. The UN aims to halve road traffic deaths and injuries by 2030. A shift from reactive to proactive safety approaches is underway, focusing on real-time crash prediction before incidents occur. While progress has been made in real-time crash likelihood prediction, challenges remain. Existing models often prioritize accuracy over computational efficiency and model size. Furthermore, there's a lack of integrated frameworks predicting both crash likelihood and severity simultaneously. This study addresses these limitations by proposing a calibrated confidence learning-based approach for real-time crash likelihood and severity prediction, incorporating deployment strategies for real-world application.

Existing literature primarily focuses on real-time crash likelihood prediction, with efforts to improve accuracy, address data imbalance, identify contributing factors, and develop models for specific roadway sections or traffic states. Research on real-time crash severity prediction has evolved since 2010, with studies conducted in various countries and on different roadway types. Data sources commonly used include crash data, traffic data (from loop detectors, AVI, RITIS, ATSPM, cameras, MVDS), weather data, and road geometric data. Data aggregation is common, often at 5-min intervals. Most studies model crash severity separately, often using two or three severity levels based on variations of the KABCO scale. Statistical methods (ordinal logistic regression, probit models, etc.) and machine learning techniques (SVM, XGBoost, Random Forest, etc.) have been employed. While ensemble learning has been used for crash likelihood prediction, its application to crash severity prediction, especially with considerations for model regularization arising from spatial ensemble learning, remains largely unexplored. This study aims to address these gaps.

This study uses data from a 67.6-mile section of I-75 in Florida, encompassing 97 segments, collected over three years. Data sources include crash data (S4A, SSOGIS), traffic data (MVDS detectors), and weather data (Visual Crossing). Data are aggregated to 5-min intervals. The methodology involves three layers: 1. **Spatial Ensemble Learning:** The heterogeneous, non-IID data is divided into zones with near-IID patterns. Individual segment-level models (lightweight multilayer perceptrons – MLPs) are trained for each zone. These models are then aggregated using importance weighting based on spatial relationships (Eq. 2), leveraging the zonal expert models. 2. **Confidence Calibration:** To address overfitting in local models, regularization techniques are applied during training: weight decay (Eq. 3), label smoothing (Eq. 4), and knowledge distillation (Eq. 5, Fig. 3). Knowledge distillation uses a smaller 3-layer MLP (student model) trained to mimic the output of the larger 10-layer MLP ensemble model (teacher model). 3. **Global Severity Post-Calibration:** Post-calibration refines model outputs to improve severity prediction. Temperature scaling (Eq. 6) adjusts confidence levels, mapping predicted crash likelihood to severity levels. Higher temperature values indicate higher severity predictions. The temperature scalers are determined from a validation set. This method leverages the crash prediction model's outputs to directly infer severity, enhancing efficiency and reducing training resources. Four benchmark methods are compared: binary classifier, SMOTE, undersampling, and ensemble. Performance is evaluated using accuracy, sensitivity, and false alarm rate (FAR).

The proposed calibrated confidence learning (CCL) approach, incorporating knowledge distillation, significantly outperforms benchmark methods in four-level crash severity prediction (Table 2). CCL achieves high sensitivity (91.7% for fatal crashes, 83.3% for severe, 85.6% for minor, and 87.7% for PDO) with low FAR (17.4%, 21.9%, 26.3%, and 28.7% respectively). Knowledge distillation proves superior to weight decay and label smoothing for regularization (Table 3). CCL demonstrates excellent performance across different crash types (rear-end, sideswipe/angle) (Table 4). The model shows better performance in urban areas and segments with less traffic fluctuation (Fig. 4). Ablation study (Table 5) highlights the importance of traffic volume data for accurate crash and severity prediction. Computational efficiency is significantly improved compared to training a single large model on the entire dataset, enabling faster training and inference times (reduced from over 24 h to approximately 3 min using parallel computing).

The results demonstrate the effectiveness of the proposed CCL approach in accurately predicting both crash likelihood and severity in real-time. The spatial ensemble learning effectively handles spatial heterogeneity, while the confidence calibration improves the reliability and accuracy of severity predictions. The high sensitivity and low FAR are significant for real-world implementation, minimizing false alarms and optimizing resource allocation for traffic management. The superior performance compared to benchmark methods and previous studies suggests that this approach offers a substantial improvement in real-time crash and severity prediction. The integration of crash prediction and severity modeling enhances the utility of the system for implementing proactive safety measures.

This study presents a novel, real-time crash and severity prediction framework using calibrated confidence learning with knowledge distillation. The approach outperforms existing methods in accuracy, sensitivity, and false alarm rate. The framework's efficiency, scalability, and adaptability make it suitable for real-world deployment in traffic management systems. Future research should focus on optimizing model updates, improving model robustness to noisy data, addressing communication latency, and evaluating the effectiveness of different active traffic management strategies in preventing or mitigating crashes based on the predicted severity levels.

The study focuses on a specific section of I-75 in Florida. Generalizability to other locations and roadway types needs further investigation. The model's performance may be affected by data quality and the accuracy of real-time data sources. Further research is needed to address potential challenges like model updates, robustness, and communication latency in real-world deployments. The study's data is not publicly available due to restrictions by original data sources.