2.1. Overall framework

This study proposes a safety-oriented supervised machine learning framework for predicting severe traffic crash outcomes on the UK road network. Instead of conventional multi-class severity classification, crash severity is reformulated as a binary classification task distinguishing slight injury crashes from crashes resulting in KSI outcomes. This formulation aligns with practical road safety objectives, where preventing severe human injury is the primary concern.

As illustrated in Fig 1, the proposed framework provides an integrated pipeline that includes data preprocessing, feature engineering, binary target construction, imbalance handling, model training with decision threshold optimization, and probabilistic reliability and explainability analysis. The detailed implementation of this workflow is presented in Algorithm 1. This unified framework ensures methodological consistency and supports risk-based interpretation for practical traffic safety applications.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Safety-oriented supervised machine learning framework.

https://doi.org/10.1371/journal.pone.0347873.g001

2.2. Software and implementation environment

All experiments were conducted within a unified computational environment to ensure reproducibility and fair comparison. Table 1 summarizes the software, libraries, and execution platform used in this study.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Summary of software and implementation environment.

https://doi.org/10.1371/journal.pone.0347873.t001

For transparency, the overall learning procedure is summarized in Algorithm 1.

Algorithm 1. Safety-oriented KSI crash prediction framework.

Input:

Crash dataset , where (0: Slight injury, 1: KSI)

Output:

Calibrated KSI probabilities and binary predictions

  1. Remove identifiers and post-crash variables to prevent data leakage.

  2. Encode categorical features, extract temporal components, and impute missing values.

  3. Construct binary target variable (KSI vs. slight injury).

  4. Split D into training (80%) and testing (20%) sets using stratified sampling.

  5. Apply SMOTE to the training set to balance KSI and non-KSI classes.

  6. Train a LightGBM classifier with early stopping on the balanced training data.

  7. Predict KSI probabilities on the test set.

  8. Optimize the decision threshold by maximizing the Operational Utility Score (OUS), which balances Recall (KSI) and Precision under operational constraints.

  9. Evaluate performance using ROC–AUC, PR-AUC, F1-Score, Precision, and Recall metrics.

  10. Assess probability calibration using the Brier Score and calibration curve.

  11. Interpret model predictions using SHAP values.

While Algorithm 1 summarizes the proposed safety-oriented learning pipeline, it is important to clarify how this strategy differs conceptually from default accuracy-oriented learning commonly adopted in crash severity prediction. To this end, Algorithm 2 contrasts the default configuration with the proposed safety-oriented framework.

Algorithm 2. Default accuracy-oriented vs. safety-oriented learning paradigm.

Input:

Crash dataset , binary target (0: Slight injury, 1: KSI)

Output:

Comparative performance of learning strategies

Default configuration (accuracy-oriented learning)

  1. Split D into training and testing sets using stratified sampling.

  2. Train a LightGBM classifier on the original imbalanced training data.

  3. Optimize model parameters by minimizing the global log-loss.

  4. Apply the default decision threshold .

  5. Evaluate performance using Accuracy, ROC–AUC, and Recall (KSI).

Safety-oriented configuration (proposed framework)

  6. Apply SMOTE to the training set to balance KSI and non-KSI classes.

  7. Train the proposed classifier (LightGBM) with early stopping on the balanced data.

  8. Predict KSI probabilities on the test set.

  9. Optimize the decision threshold by maximizing the OUS, ensuring a balance between detection sensitivity and false-alarm control.

  10. Evaluate performance using Recall (KSI), Precision–Recall curve, Brier score, and calibration curve.

2.3. Data preprocessing and feature engineering

The crash dataset was obtained from the UK Department for Transport and covers five consecutive years (2020–2024), comprising 503,475 crash records with 44 variables [20]. A rigorous preprocessing pipeline was applied to ensure data quality and prevent information leakage (Fig 2).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Data cleaning pipeline and feature extraction strategy.

https://doi.org/10.1371/journal.pone.0347873.g002

Feature selection was conducted prior to model training using a domain-informed filtering strategy. This filtering process was designed to eliminate variables that may introduce information leakage and to retain interpretable predictors relevant to traffic safety analysis.

Specifically, variables were systematically evaluated and removed based on three criteria. First, administrative identifiers (e.g., collision_index and collision_ref_no) were excluded because they are non-informative record labels that do not contain meaningful predictive information for crash severity modeling. Second, redundant coordinate systems (e.g., location_easting_osgr and location_northing_osgr) were removed to avoid duplicate spatial representations. Standard geographic coordinates (latitude and longitude) were retained as the primary spatial references, allowing the tree-based model to learn broader spatial patterns and regional safety variations. Finally, post-crash outcome variables (e.g., number_of_casualties and collision_adjusted_severity_serious) were excluded. These post-event variables represent consequences of the crash rather than predictive risk factors, and their inclusion would introduce data leakage into the modeling process.

After this domain-driven filtering, all remaining variables—representing roadway characteristics, environmental conditions, and temporal attributes—were retained without further feature elimination. To further enhance model learning, temporal variables were decomposed into interpretable components (month and hour of occurrence) to capture seasonal and diurnal traffic patterns. Missing values were subsequently addressed using mode imputation for categorical attributes and median imputation for numerical ones [22]. This preprocessing ensures data consistency and improves model robustness. To strictly prevent information leakage, data splitting was performed prior to any resampling operations [23]. The dataset was divided into training (80%) and testing (20%) subsets using stratified sampling to preserve class proportions [24]. Statistical integrity was verified by comparing class distributions and descriptive statistics between subsets. Ten-fold cross-validation was used within the training set to support model tuning and enhance generalization [25].

2.4. Binary target definition: KSI strategy

Crash severity in the STATS19 dataset is recorded as fatal, serious injury, or slight injury [20]. Multi-class prediction suffers from extreme imbalance and ambiguity between fatal and serious outcomes, which are often influenced by stochastic post-crash factors such as emergency response time [26].

Accordingly, this study adopts a binary KSI formulation defined as:

(1)

This strategy increases the effective sample size of the minority class and directly reflects the primary safety objective of reducing severe human injury.

2.5. Handling class imbalance

Despite the KSI reformulation, class imbalance remains pronounced, with KSI crashes accounting for approximately 23% of observations. To mitigate majority-class bias, data-level imbalance handling was performed using the Synthetic Minority Over-sampling Technique (SMOTE).

SMOTE was applied exclusively to the training set after data splitting, while the test set retained the original class distribution to reflect real-world operating conditions. The resampled training data achieved an approximately balanced 50:50 class distribution, strengthening the learning signal associated with severe crash outcomes.

In addition to SMOTE, class-weighted learning was considered as an alternative imbalance-handling strategy. Class-weighted learning operates at the algorithm level by modifying the loss function to penalize misclassification of the minority class, whereas SMOTE modifies the data distribution through synthetic sample generation.

In this study, SMOTE is adopted as the primary imbalance-handling strategy in the final model. Class-weighted learning is implemented as a comparative approach only, and is not used in the final model configuration. A quantitative comparison between these strategies is presented in the Results section to evaluate their impact on predictive performance and operational effectiveness. The final selection of SMOTE is therefore based on empirical evidence rather than methodological preference.

2.6. Model development using LightGBM

Light Gradient Boosting Machine (LightGBM) was selected as the primary predictive model for the proposed safety-oriented framework due to its scalability, robustness to heterogeneous features, and strong performance on large tabular datasets [27,28]. The model was trained using a binary logistic loss function.

Overfitting was controlled through:

• Early stopping based on validation loss (100 rounds);
• Regularization via constrained tree depth and number of leaves; and
• Validation-based monitoring using a hold-out validation subset distinct from the test data.

Categorical variables were handled using LightGBM’s native categorical feature support, avoiding extensive one-hot encoding and preserving computational efficiency.

As described in Section 2.5, SMOTE was adopted as the primary imbalance-handling strategy, while class-weighted learning was used only for comparison. Accordingly, the final LightGBM model was trained using SMOTE-balanced data without class weighting. Class-weighted learning is retained as a baseline model, and its performance is analyzed in the Results section. To rigorously evaluate the effectiveness of the proposed LightGBM model, several widely used machine learning algorithms, namely Logistic Regression, Random Forest, and XGBoost, were implemented as baseline models for comparative analysis.

To ensure reproducibility and a fair comparison across all algorithms, the hyperparameters were explicitly configured. For the proposed LightGBM model, the main parameters included a learning rate of 0.03, 1000 boosting iterations, and a tree structure controlled by num_leaves = 40 with an unrestricted maximum depth (max_depth = −1).

For the baseline models, configurations were set as follows: Logistic Regression was implemented with L2 regularization and max_iter = 1000; Random Forest used 100 trees with a maximum depth of 20; and XGBoost was trained with 1000 boosting rounds, learning_rate = 0.03, max_depth = 6, and tree_method = “hist”.

These configurations follow commonly adopted settings in tabular machine learning tasks, ensuring both model stability and fair comparison across algorithms.

2.7. Decision threshold optimization

Instead of using the default probability threshold (0.50), the decision threshold was optimized to align with safety-oriented objectives, where detecting severe crashes is prioritized over minimizing false alarms. However, threshold selection inherently involves a trade-off between sensitivity (Recall) and predictability (Precision). As emphasized by Xu et al. (2016), lowering the threshold improves the detection of hazardous conditions but inevitably increases the false alarm rate, reflecting the fundamental trade-off between sensitivity and predictability [29]. From a human factors perspective, excessive false alarms can erode user trust and lead to alarm fatigue, thereby reducing operational effectiveness [30].

To avoid subjective decision-making, threshold selection was formulated as a constrained optimization problem using a penalty-based approach. Predicted KSI probabilities were evaluated across a range of candidate thresholds τ ∈ [0,1]. An Operational Utility Score (OUS) is defined to maximize Recall while applying a penalty coefficient (λ = 10) when Precision falls below an operational target ( = 0.33), corresponding to a true-to-false alarm ratio of approximately 1:2. This formulation penalizes excessive false alarms without rigidly discarding near-optimal solutions.

The objective function is defined as follows:

(2)

The optimal threshold is therefore selected as:

(3)

This formulation explicitly reflects the predictability–sensitivity trade-off described in Xu et al. (2016), where improving sensitivity inevitably increases the false alarm rate [29]. The proposed approach replaces heuristic threshold selection with a data-driven optimization strategy, ensuring a balanced trade-off between detection performance and false-alarm control.

2.8. Model evaluation metrics

Model performance was evaluated using ROC–AUC and Precision–Recall AUC (PR–AUC, commonly computed as Average Precision, AP), along with class-specific precision, recall, and F1-score, with particular emphasis on recall for KSI crashes. Precision–Recall curves were used to analyze the trade-off between recall and precision under class imbalance.

2.9. Probabilistic reliability and explainability

Probabilistic reliability was assessed using calibration curves and the Brier score, defined as:

(4)

where is the predicted probability of a KSI outcome, is the observed label, and N is the number of observations. Lower Brier scores indicate better probabilistic reliability, which is essential for risk ranking and intervention prioritization.

Model interpretability was addressed using SHAP. For a given prediction , SHAP expresses the model output as:

(5)

where represents the contribution of feature j to the predicted KSI risk. The magnitude and sign of SHAP values quantify the strength and direction of feature effects, enabling transparent and policy-relevant interpretation.

3.1. Dataset characteristics and experimental setup

After data cleaning and preprocessing, the final dataset comprised 503,475 crash records described by 30 explanatory variables, reduced from the original 44 variables to prevent data leakage and redundancy. The retained features represent roadway characteristics, traffic control attributes, environmental conditions, spatial context, and temporal factors.

Following the binary KSI formulation, approximately 76.7% of crashes were classified as slight injury, while 23.3% belonged to the KSI category, confirming a substantial class imbalance typical of national crash datasets. The data were split into training (80%) and testing (20%) subsets using stratified sampling to preserve class proportions. SMOTE was applied exclusively to the training set, while all reported results correspond to the untouched test set.

3.2. Default versus safety-oriented learning strategy

To clarify the contribution of the proposed framework, a default learning configuration was first considered. The default configuration follows a conventional accuracy-oriented learning strategy: binary classification without imbalance handling and using the default decision threshold of 0.5. Under this configuration, model optimization is driven by global log-loss, implicitly favoring the majority class (slight injury).

In contrast, the proposed safety-oriented learning strategy integrates data-level imbalance handling (SMOTE) and explicit decision threshold optimization to prioritize the detection of severe crash outcomes. The conceptual differences between these two strategies are summarized in Algorithm 2, and their empirical performance is analyzed in the following subsections.

3.3. Overall discriminative performance

3.3.1. ROC–AUC analysis.

The LightGBM model achieved a ROC–AUC value of 0.664, indicating meaningful discriminative capability. Although moderate in absolute terms, this performance is consistent with values reported in crash severity prediction studies, where a substantial proportion of risk is driven by unobserved human behavior and stochastic factors.

Fig 3 presents the Receiver Operating Characteristic (ROC) curve of the proposed LightGBM model. The curve consistently lies above the diagonal reference line representing random classification, confirming that the model captures statistically significant risk patterns associated with severe crash outcomes. Importantly, ROC–AUC provides a threshold-independent assessment of discriminative performance, establishing a necessary but not sufficient condition for safety-critical deployment.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. ROC curve illustrating the discriminative ability of the LightGBM model for KSI crash risk prediction.

https://doi.org/10.1371/journal.pone.0347873.g003

Previous studies have reported ROC–AUC values in the range of approximately 0.60–0.72 for crash severity prediction under real-world conditions [31,32]. Within this context, the performance achieved in this study falls within the expected range and demonstrates that the proposed model captures statistically meaningful risk patterns despite the inherent uncertainty of crash outcomes.

While ROC–AUC confirms that the model possesses meaningful discriminative capability, it does not prescribe how predicted probabilities should be converted into operational decisions. In safety-critical applications such as KSI detection, the choice of decision threshold plays a decisive role in determining whether severe crashes are effectively identified or overlooked.

3.3.2. Precision–Recall characteristics under class imbalance.

Given the inherently imbalanced nature of KSI detection, the Precision–Recall (PR) curve provides a more informative performance assessment than ROC analysis. Unlike ROC–AUC, which can remain relatively insensitive to class imbalance, the PR curve directly characterizes the trade-off between detection capability (recall) and false-alarm rate (precision) for rare but safety-critical events.

Fig 4 illustrates the Precision–Recall curve, demonstrating that meaningful recall levels can be maintained across a wide range of precision values, with a PR–AUC of 0.35. This result confirms that the model retains substantial discriminative power for detecting KSI crashes despite the skewed class distribution.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Precision–Recall curve of the LightGBM model for KSI crash detection under class imbalance.

https://doi.org/10.1371/journal.pone.0347873.g004

Importantly, the PR curve reveals that increasing recall beyond conventional thresholds is feasible, albeit at the cost of reduced precision. This observation directly motivates the adoption of a safety-oriented decision threshold, where sensitivity to severe crashes is prioritized over overall accuracy.

The PR analysis highlights that the default probability threshold of 0.5 is not necessarily optimal for KSI detection. Instead, it reveals a continuum of operating points where recall can be substantially improved with an acceptable increase in false positives. This insight motivates the explicit optimization of the decision threshold from a safety-first perspective.

3.3.3. Decision threshold optimization and confusion matrix analysis.

Building on the threshold-independent insights provided by the ROC and Precision–Recall analyses, the decision threshold was explicitly optimized to align model behavior with safety-oriented objectives. In crash severity prediction, the cost of missing a severe crash (false negative) is substantially higher than that of issuing a false alarm (false positive). Consequently, the default probability threshold of 0.5—typically optimized for overall accuracy—was deemed inappropriate for KSI detection.

A range of probability thresholds was systematically evaluated on the validation set. The optimal operating point was identified at τ* = 0.35, which provides a favorable balance between sensitivity to KSI crashes and practical precision. To provide a quantitative justification for the threshold selection, a sensitivity analysis was conducted across candidate thresholds ranging from 0.20 to 0.50. The results are summarized in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Quantitative threshold optimization based on sensitivity–predictability trade-off.

https://doi.org/10.1371/journal.pone.0347873.t002

As shown in Table 2, lower thresholds (τ ≤ 0.30) achieve high recall (0.699–0.867) but incur heavy penalties due to low precision, resulting in reduced OUS values and indicating excessive false alarm rates. In contrast, higher thresholds (τ ≥ 0.40) avoid penalties but exhibit a significant reduction in recall (down to 0.349–0.514), limiting the model’s ability to detect severe crashes.

The threshold τ* = 0.35 emerges as the optimal operating point. Its precision (0.328) is slightly below the target (0.33), incurring only a minimal penalty. In exchange, the model retains a relatively high recall of 0.605, leading to the maximum OUS (0.586). This result indicates that τ* = 0.35 provides an effective data-driven trade-off between detection capability and false-alarm control for practical deployment.

At the selected optimal threshold (τ* = 0.35), the LightGBM model achieves a Recall for KSI crashes of 0.605. This represents a substantial 73% improvement compared to the default threshold configuration (τ = 0.50), where the recall value is only 0.349. From a road safety perspective, this result implies that the proposed framework is capable of identifying approximately 60% of severe crashes, significantly enhancing the potential for proactive safety interventions.

The corresponding Precision (KSI) of 0.328 reflects an intentional and acceptable trade-off in a safety-first context. Many false-positive predictions may correspond to high-risk conditions that did not result in severe injury due to stochastic or contextual factors. Flagging such cases remains valuable for preventive planning, infrastructure improvement, and risk-based prioritization.

Table 3 summarizes the classification performance of the LightGBM model at the optimized decision threshold (τ* = 0.35), highlighting the trade-off between sensitivity and precision under the proposed safety-oriented decision strategy.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Classification performance of the LightGBM model at the optimized threshold (τ* = 0.35).

https://doi.org/10.1371/journal.pone.0347873.t003

The practical implications of this trade-off are further illustrated by the confusion matrix in Fig 5. Compared with the default threshold, the optimized configuration substantially reduces false negatives for the KSI class, confirming that threshold tuning is a critical mechanism for aligning machine learning outputs with real-world road safety priorities rather than purely statistical optimality.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Confusion matrix of the LightGBM model at the optimized decision threshold (τ = 0.35), illustrating the trade-off between increased KSI detection and higher false-positive rates under a safety-oriented learning strategy.

https://doi.org/10.1371/journal.pone.0347873.g005

3.3.4. Comparative performance with baseline models.

Before comparing different machine learning algorithms, it is important to clarify the impact of imbalance-handling strategies on model performance. To improve transparency, a comparative analysis between SMOTE and class-weighted learning was conducted under the same decision threshold (τ* = 0.35). The results are presented in Table 4.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Comparison of imbalance-handling strategies.

https://doi.org/10.1371/journal.pone.0347873.t004

As shown, class-weighted learning achieves a substantially higher Recall (0.904), indicating strong sensitivity to KSI crashes. However, this comes at the expense of a significantly reduced Precision (0.284), reflecting an excessive false alarm rate. In safety-critical applications, such behavior may lead to alarm fatigue and reduced operational trust.

In contrast, the SMOTE-based model provides a more balanced performance, achieving a higher Precision (0.328) while maintaining a reasonable Recall (0.605). In addition, SMOTE yields a substantially lower Brier score (0.190 vs. 0.221), indicating more reliable probability estimates for risk-based decision-making. Importantly, when evaluated using the proposed Operational Utility Score (OUS), which explicitly incorporates both recall and precision constraints, SMOTE achieves the highest utility (0.586), outperforming class-weighted learning (0.445).

These results indicate that although class-weighted learning maximizes sensitivity, it does so at the cost of operational feasibility. SMOTE, on the other hand, provides a more effective trade-off between detection capability and false-alarm control, making it more suitable for real-world deployment in traffic safety systems.

Table 5 presents the comparative performance of the evaluated machine learning models at the safety-optimized decision threshold (τ* = 0.35). The comparison includes the proposed LightGBM model and three widely used baseline algorithms: Logistic Regression, Random Forest, and XGBoost.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Comparative performance of machine learning models for KSI crash prediction at the safety-optimized decision threshold (τ* = 0.35).

https://doi.org/10.1371/journal.pone.0347873.t005

As commonly observed in crash severity modeling, no single algorithm dominates all evaluation metrics due to the inherent trade-off between precision and recall. For example, Logistic Regression achieved the highest Recall (0.808), but this occurred at the expense of substantially lower Precision (0.282), implying a high rate of false alarms.

Among the evaluated models, LightGBM achieved the highest ROC–AUC (0.664), indicating the highest overall discriminative performance among the evaluated models. In addition, LightGBM achieved the highest Precision for KSI detection (0.328) while maintaining competitive Recall and F1-score values. The PR–AUC values are also comparable across the evaluated models, with XGBoost and LightGBM achieving the highest scores (approximately 0.35), indicating robust performance under class imbalance.

Importantly, LightGBM also produced the lowest Brier score (0.190), indicating superior probabilistic reliability compared with Random Forest (0.201) and XGBoost (0.192). These results suggest that the LightGBM model provides a balanced combination of discrimination and probabilistic reliability, making it particularly suitable for safety-oriented crash risk prediction.

The comparative analysis confirms that the proposed LightGBM model provides a strong balance between discriminative performance and probabilistic reliability at the safety-optimized threshold. Building on this foundation, the next section further investigates the model’s probabilistic reliability through calibration analysis, which is essential for risk-based traffic safety management.

3.4. Probability calibration and reliability analysis

Beyond classification accuracy, reliable probability estimates are essential for risk-based traffic safety management, where predicted risks are used to prioritize locations, conditions, or time periods for intervention rather than merely assigning class labels.

The probabilistic performance of the proposed LightGBM model was evaluated using the Brier score and calibration (reliability) curve. The Brier score measures the mean squared error between predicted probabilities and observed outcomes, and lower scores indicate better probabilistic reliability [33,34]. The model achieved a Brier score of 0.190, indicating a relatively low mean squared error between predicted KSI probabilities and observed outcomes. For rare-event prediction problems such as fatal and serious injury crashes, a Brier score below 0.20 is generally considered indicative of good probabilistic reliability.

Fig 6 presents the calibration curve of the LightGBM model. Each point represents the empirical frequency of KSI crashes within a predicted probability bin, plotted against the mean predicted probability. The dashed diagonal line denotes perfect calibration. Overall, the predicted probabilities closely follow the ideal diagonal across most probability ranges, indicating that the model’s outputs are well calibrated. Minor deviations are observed at higher probability levels, where the model exhibits slight under-confidence. Such behavior is expected in crash severity prediction due to the inherent stochasticity of severe injury outcomes and the limited number of extreme-risk observations.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Calibration curve of the LightGBM model for KSI prediction.

The close alignment between predicted probabilities and observed frequencies indicates good probabilistic reliability, supporting risk-based interpretation and prioritization of traffic safety interventions.

https://doi.org/10.1371/journal.pone.0347873.g006

Importantly, this level of calibration implies that the model’s probabilistic outputs are interpretable and trustworthy. For example, a predicted KSI probability of 0.6 can be meaningfully interpreted as an approximately 60% empirical risk, rather than an arbitrary confidence score. This property is critical for risk ranking, early warning systems, and evidence-based allocation of traffic safety resources. In contrast to uncalibrated boosting models, which often produce overconfident probabilities, the calibrated LightGBM framework supports policy-oriented decision-making, enabling authorities to move beyond binary predictions toward probabilistic, risk-informed safety management.

While the calibration analysis confirms that the proposed model provides reliable and interpretable probability estimates for risk-based decision-making, understanding why specific crashes are assigned higher KSI risk remains essential. Therefore, the following section employs SHAP-based explainability to uncover the key factors driving severe crash predictions and to translate model outputs into actionable safety insights.

3.5. Model explainability using SHAP

In the SHAP summary plot (Fig 7), input features are ranked in descending order according to their mean absolute SHAP values across all observations. This ranking reflects the global contribution of each variable to the model’s predictions, irrespective of whether its effect increases or decreases KSI risk. Variables appearing at the top of the plot therefore exert the strongest overall influence on crash severity prediction within the learned model.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. SHAP summary plot illustrating the encoded-feature contributions to KSI crash risk prediction.

Note: Feature colors indicate higher (red) and lower (blue) encoded values according to the STATS19 coding scheme. SHAP values represent conditional contributions relative to the model baseline and do not imply causality.

https://doi.org/10.1371/journal.pone.0347873.g007

It should be emphasized that this ordering represents model-based importance rather than causal dominance. SHAP values quantify conditional contributions learned from the data and are inherently dependent on the observed feature distributions and the underlying model structure.

Fig 7 visualizes the direction and magnitude of feature contributions to KSI risk across the entire dataset. To facilitate interpretation and translate these patterns into domain-relevant insights, Table 6 summarizes the SHAP patterns of the most influential explanatory variables, together with their corresponding STATS19 coding schemes and safety interpretations.

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Interpretation of SHAP patterns for selected explanatory variables (non-causal, risk-oriented interpretation).

https://doi.org/10.1371/journal.pone.0347873.t006

It is important to explicitly distinguish between statistical associations identified by the model and actionable safety mechanisms. SHAP values capture patterns present in the observed data, which may be influenced by reporting practices, exposure differences, or behavioral adaptations, rather than direct causal effects.

For example, variables such as police attendance or weather conditions may reflect post-event reporting processes or behavioral responses (e.g., reduced speed under adverse conditions), rather than intrinsic risk factors. Therefore, the observed SHAP patterns should be interpreted as associative signals that help identify high-risk conditions, rather than as direct evidence of causal relationships.

From a practical perspective, these insights are most appropriately used for risk screening and prioritization, while causal interpretation and policy intervention should be supported by domain knowledge and complementary analytical approaches.

Despite its advantages in enhancing model transparency, SHAP-based interpretation has several inherent limitations that should be acknowledged. First, SHAP values describe conditional associations learned from observational data rather than causal relationships, and therefore should not be interpreted as evidence of direct causality. Second, some observed patterns may be influenced by reporting biases or behavioral adaptations rather than underlying risk mechanisms. For example, variables related to police attendance or weather conditions may reflect differences in reporting practices or driver responses rather than intrinsic crash severity determinants. Accordingly, the SHAP results in this study are intended to support risk interpretation and prioritization, while causal inference and policy design should rely on complementary domain-specific analyses.

3.6. Integrated discussion and practical implications

This study demonstrates that effective traffic crash severity prediction depends less on algorithmic complexity than on the strategic alignment of the learning process with safety-critical objectives. Conventional accuracy-oriented machine learning models, even when achieving acceptable ROC–AUC values, tend to underperform in detecting fatal and serious injury crashes due to class imbalance, inappropriate decision thresholds, and unreliable probability estimates. The proposed safety-oriented framework addresses these limitations by integrating imbalance-aware learning, decision threshold optimization, probabilistic calibration, and explainable modeling within a unified pipeline.

From a methodological perspective, the results confirm that threshold optimization plays a decisive role in safety-critical classification. By shifting the decision threshold from the conventional value of 0.5 to an optimized value of 0.35, the model doubled the recall of KSI crashes without relying on more complex algorithms or additional data sources. This finding highlights that many previously reported limitations in crash severity prediction may stem from decision strategy design rather than intrinsic model capability.

The comparison of imbalance-handling strategies further highlights that maximizing recall alone is insufficient for safety-critical applications. A balanced trade-off between detection performance and false-alarm control is essential, reinforcing the importance of integrating data-level strategies such as SMOTE with decision-aware evaluation metrics.The SHAP-based analysis provides further insight into the mechanisms underlying severe crash risk. Dominant contributors such as speed limit, road classification, lighting conditions, and road geometry are consistent with established traffic safety theory. Meanwhile, several counterintuitive patterns—such as higher KSI risk under good weather or dry road conditions—warrant further scrutiny. These patterns have been widely discussed in the traffic safety literature and are commonly interpreted through behavioral adaptation and risk compensation mechanisms [35,36]. According to risk compensation theory, drivers may adjust their behavior to maintain an acceptable level of perceived risk. For instance, under adverse weather or slippery road conditions, drivers typically perceive higher risk and may compensate by driving more cautiously, often reducing travel speeds. As a result, although crash frequency may increase under such unfavorable conditions, previous studies have reported that crash severity may be mitigated by lower impact speeds [37]. Conversely, favorable environmental conditions can create a false sense of safety, encouraging higher travel speeds and reduced vigilance. When collisions occur under these conditions, the higher kinetic energy associated with greater travel speeds substantially increases the likelihood of severe outcomes, including KSI crashes. Importantly, these patterns reflect conditional risk relationships learned from observational data rather than direct causal effects, highlighting the importance of careful interpretation when translating model outputs into safety policies.

From a practical standpoint, the calibrated probability outputs enable the proposed model to function as a risk-ranking and early-warning tool, rather than a simple classifier. Transportation authorities can use predicted KSI probabilities to prioritize safety interventions across road types, time periods, or environmental conditions, supporting proactive and evidence-based decision-making. The explicit trade-off between recall and precision is particularly relevant for safety management, where missing severe crashes is typically more costly than issuing precautionary warnings for high-risk but non-severe events.

To operationalize these safety-oriented thresholds, transportation authorities could integrate the proposed framework into real-time traffic management systems. In practice, contextual crash information (e.g., location, roadway characteristics, weather conditions, and traffic environment) can be rapidly processed by the model to estimate the probability of a severe crash outcome. When the predicted probability exceeds a predefined high-risk threshold, the system may automatically trigger an alert within Traffic Management Centers (TMCs) or emergency coordination platforms.

Such threshold-based alerts can support proactive decision-making, including issuing driver warnings through variable message signs, prioritizing incident management resources, or coordinating emergency response services. By translating model predictions into operational alert levels, the proposed framework provides a practical mechanism for transforming predictive analytics into actionable traffic safety interventions.

Finally, it is important to acknowledge that the study period (2020–2024) partially overlaps with the COVID-19 pandemic, during which traffic demand and travel behavior experienced temporary disruptions in many regions. Reduced traffic volumes, changes in travel purposes, and altered mobility patterns during certain pandemic phases may have influenced the distribution of crash characteristics used for model training. However, the dataset spans multiple years and includes both pandemic and post-pandemic traffic conditions, which helps mitigate the influence of short-term anomalies. Consequently, the model is more likely to capture broader crash severity patterns rather than behaviors specific to pandemic-related travel conditions. Future research could further examine the potential impact of pandemic-induced mobility changes through more detailed temporal segmentation of crash data.