Sattar et al. (2024) developed a graph-based framework for predicting road crash injury severity using Graph Neural Networks (GNNs), specifically the GraphSAGE architecture. Their study addressed the limitations of tabular-based machine learning by converting traditional crash datasets into graph structures using k-nearest neighbor (kNN) relationships based on feature similarity. This transformation allowed the model to aggregate contextual information from neighboring crash records, enabling better generalization to unseen data. The system achieved an impressive accuracy of 85.55%, outperforming baseline classifiers such as Random Forest and XGBoost. However, the approach lacked integration with interpretability tools like Explainable AI (XAI), making it difficult for stakeholders to trust model predictions. Furthermore, issues like real-time deployment readiness and class imbalance were not directly addressed, limiting its utility in high-risk, real-world applications.

Huang et al. (2023) investigated the consequences of different types of injuries sustained during multi-vehicle motorcycle collisions using a mixed-effects statistical model.By adding random parameters, the model was able to over come the ‘observed heterogeneity’ problem, which is a common problem in crash model data, thus adding robust ness and dependability to the model. The analysis identified age as one of the vital factors affecting the severity of injuries sustained. While the model was methodologically robust and provided useful implications for policies,it still demonstrated low predictive power when applied to more extensive or sophisticated datasets. The method was also exclusively ‘flat’, failing to capture inter-instance structures, which correspondingly reduced the speed and flexibility of the method in actual-use scenarios. Zong et al. (2019) looked into combining information entropy and Bayesian networks to model crash severity. Their approach used entropy measures to quantify uncertainty and then applied Bayesian networks for probabilistic reasoning to model how crash variables depended on each other. This method provided clear insights by showing how different factors had an impact on injury outcomes in terms of probability. While it helped to understand cause-and-effect chains and interactions, the technique had a few drawbacks: it needed preset network structures making it inflexible; it also struggled with large or graph-based datasets and didn’t consider time or space patterns in crash data—key aspects for real-world use.

Manzoor et al. (2021) created a hybrid system named RFCNN, which blends Random Forest classifiers with Convolutional Neural Networks (CNNs) to boost injury severity classification.⁴ They aimed to combine Random Forest’s power in decision tree ensemble learning with CNN’s knack for learning complex features on its own—this proves helpful when crash data includes both structured records and visual info (like CCTV or vehicle sensor images). The mixed model showed better accuracy than single algorithms confirming the worth of joining multiple models. Yet, it didn’t have any way to explain itself such as SHAP or LIME, which makes it hard to use in key applications. Also, CNNs were first made for image inputs, and no one looked into or explained how well they work with table-like crash data making them less effective for structured data.⁵

Aboulola et al. (2024) proposed an advanced model that integrates Transformer architectures with Explainable AI (XAI) tools for predicting the severity of traffic accidents. Their approach employed attention mechanisms to dynamically focus on the most critical features influencing crash outcomes. Using feature attribution maps, the model provided transparency regarding which variables—such as road lighting, time of day, or number of vehicles—contributed most to injury predictions. This study stands out by combining accuracy and model transparency, making it highly relevant for policy-makers and traffic management teams. However, the model did not use graph-based representations, which could have added contextual learning capacity. Moreover, it did not implement hybrid learning mechanisms nor address the issue of data imbalance, both of which are crucial for real-world model robustness.

The first module is responsible for preparing raw traffic accident data for modeling. The dataset, obtained from the UK Department for Transport, contains detailed crash records including road type, environmental conditions, vehicle characteristics, and temporal information.⁷ To ensure consistency and quality, the data undergoes cleaning procedures to remove null, noisy, or redundant entries. All categorical variables are encoded using Principal Component Analysis (PCA)-based embedding to reduce dimensionality and enhance feature representation. PCA was chosen to capture correlations among high-cardinality categorical variables (e.g., vehicle type, road class), unlike one-hot encoding which results in sparsity. Comparative tests confirmed stability across 30–70 PCA dimensions.

Subsequently, a k-nearest neighbor (kNN) graph is constructed to represent the relational structure among crash records. Each crash instance is modeled as a node in the graph, and edges are drawn between nodes that exhibit high feature similarity based on a Euclidean distance threshold in the embedded feature space. This graph transformation enables the application of Graph Neural Networks (GNNs) and supports contextual learning from neighboring crash instances. Mathematical Model is as follows,

Where H(l) refers to the node representations at layer l. H^(l+1) refers to the node representations at layer l+1, which are updated from the previous layer’s representations. AGG (HN(l)(v)) is an aggregation function that computes a summary (aggregation) of the neighboring nodes’ features for node v at layer l. The function could be mean, sum, or any other pooling mechanism that combines the information from neighboring nodes. W^(l) is a learnable weight matrix for layer l.

To ensure methodological rigor, all preprocessing steps—including PCA fitting, kNN graph construction, and SMOTE resampling—were performed strictly within each training fold during cross-validation. This prevents information leakage from validation or test sets.⁸

The second unit keeps the main design of GNN but changes the processing to exploiting relational connections inside the structure of the graph. It implements the model of GraphSAGE. This model functions by collecting and merging the attributes of adjacent nodes to update each node’s depiction. The capacity for inductive learning of GraphSAGE makes it be able to adjust quickly to new crash data. Therefore, it becomes very important to roll it out on-line to get the changes of the traffic pattern over time. In this work, the two-layer GraphSAGE architecture with mean aggregation is accepted (Figure 1). In the first layer, for each node a fixed number of neighbors are sampled, and a mean-pooled embedding is computed from those neighbors. The second layer goes on to combine the features of the given node with the summary of the neighborhood to get a complete representation. This final embedding is then forwarded to the softmax classification layer that is responsible for the decision of the crash severity category “severe” or “non-severe.” The adjacency matrix extracted from the kNN graph and the PCA-transformed feature matrix function as the input for the GNN.⁹

Although Graph Neural Networks (GNNs) excel at capturing relational dependencies, traditional ensemble methods are known for their stability and effectiveness in handling structured datasets. To assess the performance of the GNN model, this module also implements three widely-used ensemble and neural classifiers: Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Artificial Neural Networks (ANN). These models are trained using the same preprocessed dataset, without incorporating graph-based structures, and utilize one-hot encoding or embedding for categorical variables (Figure 2).

To fine-tune model performance, hyperparameter optimization is performed using GridSearchCV, adjusting parameters such as the number of trees, depth of the model, learning rate, and activation functions. Model predictions are evaluated using various performance metrics, including accuracy, precision, recall, F1-score, and the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). This comparative analysis enables the identification of each model’s strengths and limitations, with the GNN demonstrating superior results—particularly in modeling complex relationships among features.¹⁰

To promote transparency and enable human-centered decision-making, the final component of the framework integrates Explainable AI (XAI) methodologies. Specifically, SHapley Additive exPlanations (SHAP) are employed to measure the impact of each input variable on the model’s predictions. This interpretability layer offers clear visualizations that reveal how factors like lighting conditions, number of vehicles, and road surface type influence the severity of traffic accidents.

Moreover, feature importance scores derived from the Random Forest and XGBoost models are utilized to cross-validate the key predictors uncovered by the GNN. Incorporating these explanatory tools is crucial for enabling stakeholders—such as traffic management officials and emergency responders—to comprehend, trust, and effectively use the model’s outcomes. This layer of interpretability also supports the responsible and ethical application of AI technologies in critical public safety environments.

The final approach adopted for all evaluations is the hybrid model [H = {H, Z}], where H denotes GraphSAGE embeddings and Z the PCA-encoded features. Standalone GNN results are reported for comparison only (Table 1).

To validate the proposed hybrid framework for crash injury severity prediction, a comprehensive experimental setup was established. This section details the dataset characteristics, preprocessing techniques, model configurations, training procedure, and evaluation metrics used to assess performance across all models. The dataset used in this study is derived from the UK Department for Transport’s official road accident data repository, covering reported traffic accidents in the United Kingdom from 2011 to 2016. The dataset is made available in comma-separated values (CSV) format and includes structured records of vehicle-to-vehicle collisions. Each entry represents a unique accident incident characterized by multiple categorical and numerical features, enabling detailed traffic safety analysis and injury severity prediction.

The dataset was divided into 70% training, 15% validation, and 15% testing sets using stratified sampling to maintain class proportions. To avoid temporal leakage, the temporal ordering of crashes was respected by ensuring that training and testing instances came from non-overlapping time spans. All experiments were repeated 10 independent runs with different random seeds (42, 52, 62, 72, 82) to confirm robustness.¹¹

Table 2 shows the selected hyper parameters for each model after grid search.

Figure 3 presents the training loss trend and validation accuracy of the standalone Graph Neural Network (GNN) model through 100 epochs. The training loss drops showing that the model learned well and converged. At the same time, validation accuracy increases and levels off near 74.8%, which shows the model handled unseen data. Both curves smooth out, which suggests stable training without overfitting or performance issues. This shows how GNN-based designs can pick up valuable patterns from graph-organized crash data.

Figure 4 shows the performance results of the hybrid model called GNN + RF across 30 epochs. The graph on the left demonstrates how training and validation accuracy improve in the first few epochs. Validation accuracy climbs above 94% on and stays stable with little variation, which suggests the model generalizes well and avoids major overfitting issues. The right plot presents the training loss, which exhibits a sharp and consistent decline, converging to a low value early in the training process. Together, these curves confirm that the hybrid architecture not only accelerates convergence but also achieves superior predictive accuracy and robust learning dynamics compared to the standalone GNN model.¹²

The improved performance of the hybrid model can be attributed to several key architectural and methodological advantages. Firstly, Graph Neural Networks (GNNs), specifically GraphSAGE, provide relational context learning by capturing the neighborhood structures within crash data, enabling more nuanced and context-aware feature representations. This spatial awareness is critical in modeling road crash scenarios where relational dependencies between records influence prediction outcomes. Secondly, the integration of Random Forest (RF) into the hybrid model contributes robust classification capabilities. RF’s ensemble learning mechanism enhances the system’s ability to handle feature interactions and variance, leading to more stable predictions.

To comprehensively evaluate model performance, we employed 5-fold cross-validation, reporting mean ± standard deviation across folds. Statistical significance was verified using paired t-tests and Wilcoxon signed-rank tests, comparing the proposed GNN+RF model against baseline classifiers. In this work seven models has been benchmarked: SVM, Random Forest (RF), XGBoost, LightGBM, CatBoost, standalone Graph Neural Network (GNN), and the proposed hybrid GNN+RF architecture. These baselines were chosen as they represent widely used classical (SVM, RF), boosting-based (XGBoost, LightGBM, CatBoost), and neural (MLP) classifiers in transportation safety prediction.As shown in Table 3, the standalone GNN model demonstrated limited performance with an accuracy of 74.80%, indicating difficulty in capturing complex nonlinear relationships using only graph embeddings. While RF and SVM achieved higher accuracies (90.64% and 89.15%, respectively), they lacked the contextual relational learning that GNN provides.

Beyond conventional classification metrics, we computed additional probability and class-level evaluations. The hybrid GNN+RF achieved an AUC-PR of 0.89 for the Severe class, outperforming SVM (0.76), RF (0.80), and XGBoost (0.82). Macro-F1 reached 93.75%, indicating balanced performance under class imbalance. Figure 5 shows Precision–Recall curves for all models, and Figure 6 presents the normalized confusion matrix. Table 4 reports per-class precision/recall/F1. To assess robustness, we aggregated results over 10 runs × 5 folds; the hybrid model achieved 94.20% accuracy (95% CI: [93.95, 94.45]) and 93.75% macro-F1 (95% CI: [93.50, 94.00]), with AUC-PR 0.89 (95% CI: [0.87, 0.91])

The improvements of the proposed GNN+RF model over baseline classifiers were statistically significant (paired t-test and Wilcoxon signed-rank test, p < 0.05).

The hybrid GNN + RF model significantly outperformed all other models, achieving 94.20% accuracy, with balanced precision (93.85%), recall (93.70%), and F1-score (93.75%). This demonstrates the effectiveness of combining graph-based feature learning with ensemble decision trees to improve generalization and class-level performance.

Beyond conventional classification metrics, we also assessed calibration and probability estimation quality. The hybrid GNN+RF achieved an AUC-PR of 0.89, reflecting strong performance in detecting severe crashes despite class imbalance. Calibration analysis confirmed that predicted probabilities were well-aligned with observed frequencies (Figure 7), with a Brier score of 0.07, indicating well-calibrated risk estimates.

These results demonstrate that the model not only classifies accurately but also produces reliable probability estimates, which is essential for decision-making in safety-critical applications. The predicted probabilities (blue) closely follow the ideal diagonal (gray dashed), confirming good calibration.

To assess the contribution of each component in the proposed framework, an ablation study was performed. Results in Table 5 show that while Random Forest and GNN individually achieved moderate accuracy, combining them significantly improved performance. The addition of SMOTE for class balancing and SHAP for interpretability further enhanced the hybrid model, demonstrating that each module contributed to the overall effectiveness of the system.SMOTE enhanced recall by addressing class imbalance, particularly for severe crashes. SHAP, while not improving accuracy, provided essential interpretability, making the framework more trustworthy for stakeholders.

For completeness, we also evaluated a Z-only configuration, where Random Forest was trained solely on PCA-encoded features without graph embeddings. This achieved 88.50% accuracy, confirming that PCA features alone provide moderate predictive power but fall short of the relational learning offered by GNN embeddings. The improvement from Z-only to GNN+RF demonstrates the added value of combining graph-based and tabular feature representations.

Additionally, the adoption of SMOTE for data resampling plays a pivotal role in ensuring balanced learning. By addressing the class imbalance, particularly for severe crash cases, SMOTE enables the model to generalize better across all severity levels. The ensemble generalization offered by combining GNN with RF allows the hybrid model to leverage both spatial and statistical feature learning strengths. This dual capacity results in a more accurate and generalizable system.¹³

To avoid bias and capture all aspects of the problem, besides classical GCN and GAT architectures, baselines from the recent deep tabular learning field, e.g. Logistic Regression, TabNet, TabTransformer, and FT-Transformer, were considered. These baselines reflect both deep tabular learning and graph-specific models, thus enabling a more thorough comparison. Besides, sensitivity analyses were performed for several important hyperparameters: the number of neighbors in the kNN graph (k ∈ {3, 5, 10, 15}), the number of PCA components (30, 50, 70), GraphSAGE hidden dimensions (64, 128, 256), and SMOTE resampling ratios (1:1, 1:2, 1:3). In all experiments with various parameter settings, the hybrid GNN+RF was the best performer all the time achieving accuracy in the range of 92%–94%. Thus it not only shows the advantage of the hybrid model over solo methods but also indicates the stability and the robustness, as well as the reliability for real-world applications, of the hybrid approach.

To evaluate the transparency of the proposed framework, both permutation-based feature importance and SHAP analysis were conducted. SHAP was applied to both original features (RF) and GNN embeddings (RF on H). This ensured interpretability across raw features and learned representations

Figure 8 shows global feature importance for the Random Forest on original features, where number of casualties, vehicles involved, and lighting conditions were dominant predictors.

Figure 9 highlights important dimensions from the GNN embeddings, confirming that learned representations capture diverse aspects of crash data.

Figure 10 presents SHAP global feature contributions, revealing consistent alignment with domain knowledge: casualty count, driver age, and road conditions significantly impacted severity predictions.

Finally, SHAP interaction plots (Figure 11) provide local interpretability, showing how specific factors such as driver age and time of accident jointly influence predictions.

These results demonstrate that the hybrid GNN+RF not only achieves strong predictive accuracy but also produces interpretable outputs, allowing stakeholders to validate and trust model recommendations.

Another important observation is the hybrid model’s superior convergence behavior and learning stability, as demonstrated by its smoother and more rapid improvement in accuracy across training epochs. Compared to the standalone GNN, the hybrid architecture shows reduced fluctuations in validation accuracy, suggesting better learning dynamics and less susceptibility to overfitting.

For deployment in the real world, the hybrid GNN+RF model was also evaluated. The system showed effective training with an average runtime of about 14 minutes for the combined GraphSAGE and Random Forest framework, while inference was made in ~18 milliseconds per crash record (which is the same as 55 frames per second). The memory footprint for the process of training was measured at 2.3 GB of GPU utilization, thus confirming the possibility of a real-time deployment in traffic monitoring centers. Moreover, fairness analyses along subgroups gave lower recall rates for night-time crashes (–3%) and motorcycle-related crashes (–2.5%). These differences, which are quite small in size, nevertheless, pinpoint the possibility of bias in the model's prediction. To lessen these differences, the next step will be to work on the re-weighted loss functions and subgroup-specific resampling strategies so that accident performance can be more equitable in all kinds of scenarios.