2.1. Progress in research on the relationship between the built environment and traffic safety

The relationship between the built environment and traffic safety has become an important and rapidly growing topic at the intersection of transport geography, urban planning, and road safety research. Stoker et al. [7] provided one of the earlier systematic reviews of pedestrian safety and the built environment, indicating that land use, roadway cross-sectional design, intersection form, pedestrian facilities, and traffic control conditions may all affect pedestrian safety. Subsequently, Saha et al. [8] proposed a conceptual framework for understanding how the built environment influences traffic safety, arguing that the built environment does not directly cause crashes but instead affects safety outcomes through intermediary mechanisms such as traffic exposure, conflict frequency, and operating speed. Merlin et al. [9] further pointed out that one key reason for inconsistent findings in the literature is the failure to distinguish between increased exposure and elevated risk per unit of exposure. In other words, although crashes may be more frequent in high-density and mixed-use areas, this does not necessarily imply that the risk per trip or per traveler is higher in those areas.

Empirical studies have examined the relationship between built environment factors and traffic safety across different spatial scales and user groups. Dai et al. [10], using network GIS analysis, found a significant spatial correspondence between built environment characteristics and pedestrian crash hotspots. Yin and Zhang [11] further incorporated D variables commonly used in walkability research and found that the number of intersections and pedestrian environment attributes exhibited spatially varying effects on pedestrian injury, indicating that the association between the built environment and pedestrian safety is not simply linear. Asadi et al. [12], in a study of urban areas in the Netherlands, showed that land-use density, functional mix, destination proximity, and road network characteristics jointly influenced crash probability, frequency, and severity. Their findings provide strong support for introducing the 5D framework into traffic safety analysis.

Studies focusing on specific population groups have further demonstrated that the safety effects of the built environment are highly heterogeneous. Amiour et al. [13], in a systematic review of children’s objective and perceived traffic safety, reported that factors such as road facilities around schools, traffic flow conditions, intersection density, and continuity of walking space affect not only children’s actual injury risk but also the subjective safety perceptions of both children and parents. Lee et al. [14] and Gálvez-Pérez et al. [15], in studies of older pedestrians in Seoul and Madrid, respectively, found that sidewalk width, signal density, and certain land-use and socioeconomic characteristics were closely associated with older pedestrian crashes, and that these mechanisms differed from those observed in the general population. Rothman et al. [16], in a multicity Canadian study, further showed that child pedestrian and cycling injuries are strongly associated with the combined effects of built environment and social environment factors, suggesting that a purely roadway-engineering perspective is insufficient to fully explain child traffic injuries.

In addition, studies focusing on specific spatial units have indicated that the effects of the built environment are also local and scale-sensitive. Okaidjah et al. [17], at the intersection level, found that the urban built environment and average income jointly influence intersection safety performance. Such findings suggest that the spatial distribution of traffic safety is shaped not only by road infrastructure, but also by the broader socio-spatial structure of cities. Therefore, examining injury crashes from the perspective of the 5D built environment is valuable not only for systematically organizing built environment variables at the theoretical level, but also for explaining differences in traffic injury risk through the logic of urban spatial organization.

2.2. Micro-scale streetscape, spatial heterogeneity, and explainable analysis

With the development of multisource data and computational methods, research on the built environment and traffic safety has gradually moved beyond conventional areal-unit-based statistical analyses toward more fine-grained, dynamic, and interpretable analytical approaches. On the one hand, technologies such as street-view imagery and semantic segmentation have enabled researchers to extract micro-scale streetscape characteristics at the road-segment level, including greenery visibility, building frontage, parking occupancy, and street openness, thereby allowing a more refined characterization of how street environments affect crash risk. Hu et al. [18] used street-view imagery to identify street-level built environment features and showed that fine-grained streetscape information can complement the limitations of conventional land-use and road network indicators in crash research. On the other hand, the integration of spatiotemporal hotspot analysis with geographically weighted models has enabled researchers to identify the spatial nonstationarity of built environment effects on crash trends. The coupled spatiotemporal hotspot trend analysis and geographically weighted logistic regression proposed by Qu et al. [19] demonstrated that the effects of built environment factors on crash evolution vary significantly across space.

At the same time, the adoption of nonlinear analysis and machine learning has further advanced methodological innovation in this field. Ling et al. [20] found that street network structure and land-use characteristics generally have threshold effects and nonlinear relationships with crash density, and that these relationships vary with regional population age composition and income structure. Wu et al. [21], from a multiscale perspective, showed that streetscape features, POI characteristics, and road attributes contribute differently to crash risk, and that clear interactions exist among these variables. Correspondingly, in traffic injury research, Wu et al. [22] and Delen et al. [2] demonstrated that ensemble learning and predictive analytics can improve the identification of injury outcomes, while Se et al. [23] and Fang et al. [24] showed that even within parametric modeling frameworks, temporal instability and unobserved heterogeneity remain important issues that cannot be ignored.

Overall, the existing literature provides three important insights for the present study. First, research on traffic injury risk should move beyond simple crash frequency analysis and treat crash consequences as a central explanatory focus [1–6]. Second, the effects of the built environment on traffic safety are systematic, spatially differentiated, and heterogeneous across user groups, and the 5D framework offers a useful theoretical basis for organizing related variables [7–17]. Third, the mechanisms underlying traffic injury risk are characterized by substantial nonlinearity and interaction, and therefore require machine learning and explainable analytical methods to improve both identification accuracy and practical interpretability [18–24]. Nevertheless, a review of the existing literature indicates that relatively limited attention has been paid to studies that take injury occurrence as a binary outcome, systematically integrate 5D built environment factors with non-built-environment contextual factors, and further examine their mechanisms through explainable analysis. The present study is conducted in response to this research gap and seeks to provide a more interpretable identification and analysis of traffic injury risk from the coupled perspective of urban spatial environment and traffic operating conditions.

3.1. Study area and data source

Changsha, Hunan Province, China, was selected as the study area (Fig 1). Changsha, located in central China, is the provincial capital of Hunan and a core city within the Changsha–Zhuzhou–Xiangtan urban agglomeration. As an important regional transportation hub, the city has a complex road network and intensive traffic activity. Marked differences are observed between the central urban area and peripheral zones in land development intensity, functional mix, public transport accessibility, and roadway operating conditions. Such spatial and transportation heterogeneity provides an appropriate setting for investigating the relationships among built environment characteristics, roadway contextual factors, and traffic injury risk.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Geographic location of Changsha.

Administrative boundary data were obtained from the GADM database, version 4.1 (https://gadm.org), which permits the use of its data to create maps for academic research articles, including articles published under open licenses such as CC BY. The map was created by the authors.

https://doi.org/10.1371/journal.pone.0353205.g001

The data used in this study were obtained from road traffic crash records in Changsha from 2017 to 2019. The original database contained information on crash time and location, road conditions, traffic control type, environmental conditions, and crash outcomes. In accordance with the objectives of this study, the raw crash records were processed through variable screening, duplicate removal, missing-value treatment, and spatial variable matching. After data cleaning and integration, a total of 9,743 complete and valid samples were retained for subsequent modeling analysis.

To further illustrate the spatial distribution of the crash samples, kernel density estimation was applied to visualize crash locations, and a traffic crash heat map of Changsha was generated (Fig 2). From the spatial pattern, crash samples showed a certain degree of clustering in urban built-up areas, along major arterial road corridors, and in zones with relatively concentrated functional activities. This indicates that traffic crash risk is not evenly distributed across urban space, but is spatially coupled with road network structure, built environment characteristics, and the intensity of traffic activity. This spatial pattern also provides a foundation for the subsequent analysis of traffic injury risk from the dual perspectives of 5D built environment factors and non-5D contextual factors.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Heat map of traffic crashes in Changsha.

Crash locations were obtained from the study dataset. The base map and data were obtained from OpenStreetMap and OpenStreetMap Foundation. This figure contains information from OpenStreetMap and OpenStreetMap Foundation, which is made available under the Open Database License. The heat map was created by the authors.

https://doi.org/10.1371/journal.pone.0353205.g002

3.2. Spatial matching and feature construction

Spatial matching was conducted to construct the built environment variables surrounding each crash location. First, crash records were geocoded as point features and matched with transit stop, POI, road network, and population density data. The distances to the nearest bus stop and metro station were calculated for each crash point to represent transit accessibility. For destination accessibility and land-use diversity, a 1,000 m buffer was generated around each crash point, within which the numbers of school, hospital, shopping mall, and tourist attraction POIs were counted. POI mix was calculated using the Shannon diversity index based on the composition of POI categories within the same buffer.

Road network density was calculated using a 1 km × 1 km grid. Each crash point was assigned to the grid cell in which it was located, and the total road length within each grid cell was divided by the grid area to obtain road network density. The 1 km × 1 km grid was selected primarily because the smallest available spatial resolution of the population density data used in this study was 1 km. Using the same grid resolution helped ensure consistency among population-related information, road network density, and crash locations. In addition, the 1 km grid provides a reasonable neighborhood-scale unit for capturing local road network structure while maintaining sufficient statistical stability. A much smaller spatial unit may introduce excessive local noise and may not be compatible with the available population density data, whereas a larger unit may obscure local built-environment heterogeneity.

3.3. Variable setting and feature grouping

In this study, traffic injury risk refers to injury occurrence conditional on crash occurrence. In other words, the proposed model predicts whether a recorded crash resulted in injury, rather than estimating the overall probability or frequency of injury crashes across the road network. The dependent variable was therefore defined as a binary outcome: a value of 1 was assigned if the crash caused personal injury, whereas a value of 0 was assigned if no injury occurred.

Although injury severity can also be analyzed using multiclass outcomes, such as fatal, serious, and slight injuries, the binary specification was adopted in this study for three main reasons. First, the primary aim of this study was to identify whether the surrounding built environment and crash-context conditions increase the likelihood that a crash crosses the fundamental threshold from non-injury to injury. This distinction is directly relevant to urban traffic safety screening and early risk intervention. Second, the explanatory variables in this study mainly characterize 5D built environment attributes and selected crash-context factors, which are more suitable for identifying spatial and contextual patterns of injury occurrence than for explaining detailed biomechanical injury severity mechanisms. Third, the mechanisms leading to fatal or severe injuries are often strongly associated with detailed crash dynamics, impact speed, collision configuration, vehicle damage, road-user behavior, restraint use, and post-crash medical outcomes, which were not fully available in the present dataset. Therefore, a binary injury occurrence model was considered more consistent with the research objective, data structure, and interpretability requirements of this study.

Compared with multiclass injury severity classification, this binary setting places greater emphasis on whether a crash results in human injury, and is therefore more consistent with the practical needs of traffic safety governance for the rapid identification and intervention of high-risk crash environments.

With respect to the explanatory variables, and in accordance with the research objectives and final modeling framework, the variables were divided into two categories: 5D variables and non-5D variables. Based on this classification, three feature groups were further constructed to compare the ability of different variable systems to identify traffic injury risk.

The 5D variables included median barrier, traffic signal control type, intersection/road segment type, road alignment, road type, distance to the nearest bus stop, distance to the nearest metro station, road network density, POI mix, and the numbers of schools, hospitals, shopping malls, and tourist attractions within a 1,000 m buffer. These variables were used to characterize the built environment surrounding the crash location from the perspectives of roadway design, transit accessibility, road network density, land-use mix, and destination accessibility.

By contrast, the non-5D variables included cross-sectional location, pavement condition, weather, lighting condition, day of week, and month. These variables mainly captured the immediate operating context, environmental status, and temporal attributes at the time of crash occurrence, thereby complementing the short-term dynamic influences that cannot be adequately represented by built environment variables alone.

Based on the above variables, three feature groups were constructed:

1. (1). 5D feature group, including only 5D variables;
2. (2). non-5D feature group, including only non-5D variables;
3. (3). 5D + non-5D feature group, in which both categories of variables were jointly entered into the model.

This grouping strategy not only helps examine whether 5D built environment factors can independently identify traffic injury risk, but also makes it possible to assess the extent to which the joint inclusion of 5D and non-5D factors improves model discriminative performance. In this study, the 5D variables were defined using an extended operational interpretation of the framework. In particular, the Design dimension was not limited to street network form alone, but was extended to include relatively stable roadway design, traffic-control infrastructure, and road functional characteristics that shape the physical and operational environment of crash locations. Median barriers were included because they represent roadway cross-sectional design and the separation of traffic streams. Traffic control type was included because signals, signs, and pavement markings constitute built traffic-control facilities that regulate road-user interactions and conflict patterns. Road type was included because it reflects the functional hierarchy, design standard, operating environment, and expected speed conditions of the roadway. Therefore, these variables were treated as built-environment-related roadway design and infrastructure variables within the extended 5D framework. In addition to the conventional dimensions of Density, Diversity, Destination accessibility, and Distance to transit, variables reflecting roadway spatial design and road environment organization were also incorporated into the Design dimension, so as to provide a more comprehensive characterization of the built environment surrounding crash locations.

Table 1 summarizes all explanatory variables included in the final models and their corresponding coding schemes. Continuous variables were entered in their original form, whereas categorical variables were encoded according to predefined rules. By explicitly linking the theoretical classification of variables with their actual model inputs, the table helps ensure consistency between the variable framework and the subsequent modeling procedures.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Definition and coding of explanatory variables.

https://doi.org/10.1371/journal.pone.0353205.t001

3.4. Model construction and evaluation strategy

To systematically evaluate the performance of different variable systems and algorithms in identifying traffic injury risk, a dual analytical framework combining feature-group comparison and model comparison was adopted.

In the feature-group comparison, classification models were developed separately using the 5D feature group, the non-5D feature group, and the combined 5D + non-5D feature group, so as to examine both the independent and joint effects of different variable types on traffic injury risk discrimination. This design was intended to address two key questions: whether 5D variables alone could effectively identify traffic injury risk, and whether model performance could be further improved when 5D and non-5D variables were jointly incorporated.

In the model comparison, three algorithms—Random Forest (RF), eXtreme Gradient Boosting (XGBoost), and Categorical Boosting (CatBoost)—were implemented for each feature group. For RF and XGBoost, the training data were balanced using the Synthetic Minority Over-sampling Technique(SMOTE), followed by hyperparameter tuning. CatBoost, by contrast, was directly trained by leveraging its native capability to handle categorical features. Comparing multiple models in parallel helped reduce reliance on a single algorithm and improved the robustness of model selection.

3.5. Model training, imbalance handling, and evaluation

The dataset was divided into a training set and a testing set at a ratio of 8:2 using stratified random sampling, with a fixed random seed of 42 to ensure reproducibility. To address class imbalance, SMOTE was applied only to the training set after data splitting, while the testing set was kept unchanged to avoid information leakage.

For model comparison, RF, XGBoost, and CatBoost were implemented for each feature group. For RF and XGBoost, hyperparameter tuning was conducted on the training set using grid search combined with five-fold stratified cross-validation, with receiver operating characteristic-area under the curve (ROC-AUC) as the optimization criterion. CatBoost was trained without SMOTE by leveraging its native capability to handle categorical variables directly. Therefore, the final CatBoost model used for SHAP-based interpretation was not trained on synthetic samples generated by oversampling. This design helps ensure that the interpretation of key influencing factors is not dependent on SMOTE-generated observations.

Model performance was evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. In the robustness check related to imbalance handling, specificity was additionally reported to assess the models’ ability to identify non-injury crashes. The candidate parameter ranges and final hyperparameter settings for the models under the combined 5D + non-5D feature group are reported in S3 Table.

3.6. Random Forest and XGBoost models

Random Forest (RF) is an ensemble learning method based on bagging and decision trees. It constructs multiple decision trees using bootstrap samples and random subsets of predictors, and the final classification result is obtained by aggregating the predictions of all trees. RF can capture nonlinear relationships and variable interactions, and is relatively robust to noise and overfitting. In this study, RF was used as a representative bagging-based ensemble model for comparison with boosting-based models.

eXtreme Gradient Boosting (XGBoost) is a gradient boosting decision tree model that builds trees sequentially, with each new tree fitted to reduce the residual errors of the previous trees. Compared with conventional boosting methods, XGBoost incorporates regularization, shrinkage, column subsampling, and efficient tree optimization, which can improve predictive performance and reduce overfitting. In this study, XGBoost was used as a representative boosting-based model and compared with RF and CatBoost to evaluate the performance of different ensemble learning strategies in identifying injury occurrence among recorded crashes.

3.7. CatBoost model

Categorical Boosting (CatBoost) is an ensemble learning method developed within the gradient boosting decision tree framework. Its primary advantage lies in its ability to handle categorical variables directly while effectively capturing nonlinear relationships and high-order interactions among variables. For the traffic crash dataset used in this study, categorical variables account for a substantial proportion of the predictors. Variables such as median barrier, traffic control type, intersection/segment type, road alignment, road type, cross-sectional location, pavement condition, weather, and lighting condition are all discrete features. CatBoost therefore shows clear methodological suitability for the present analysis.

Compared with conventional gradient boosting decision tree (GBDT) methods, CatBoost adopts ordered target statistics and an ordered boosting mechanism for categorical feature processing, thereby effectively reducing target leakage and prediction shift and improving the model’s generalization ability. For binary classification tasks, CatBoost iteratively constructs weak learners and combines them through weighted aggregation, allowing the overall model to gradually approach the optimal classification boundary. The updating process can be expressed as follows:

(1)

where is the model output after the -th iteration, is the decision tree generated at iteration , and denotes the learning rate. The model progressively improves classification performance by fitting the residuals or negative gradients from the previous iteration.

CatBoost was ultimately selected as the final model not only because it delivered the best overall discriminative performance under the combined feature group, but also because it is able to process a large number of categorical variables directly, thereby avoiding the dimensional expansion and potential information loss caused by one-hot encoding. Moreover, CatBoost is highly compatible with interpretability methods such as SHAP, which provides a robust technical basis for identifying key influencing factors and explaining their underlying mechanisms in the subsequent analysis.

Although tree-based models such as XGBoost and CatBoost can provide model-intrinsic feature importance measures, these measures usually summarize the average contribution of variables to tree splitting, gain, or prediction loss reduction. They are useful for identifying important predictors at the global level, but they provide limited information on the direction, magnitude, and heterogeneity of variable effects. In particular, conventional feature importance measures do not directly show whether higher or lower values of a variable increase the probability of injury occurrence, nor do they reveal nonlinear response patterns across different value ranges.

Therefore, SHAP was further employed in this study to provide a more detailed and interpretable explanation of the final CatBoost model. Based on Shapley values, SHAP decomposes each model prediction into additive contributions from individual variables, thereby allowing both global feature ranking and local interpretation at the observation level. Compared with conventional tree-based feature importance, SHAP provides additional insights into the direction of influence, the distribution of variable effects across samples, nonlinear dependence patterns, and the relative contribution of 5D and non-5D variables. These advantages make SHAP particularly suitable for interpreting how built environment and crash-context factors jointly influence injury occurrence among recorded crashes.

4.1 Model comparison results

Table 2 presents the classification results for different feature groups and models. Overall, substantial differences were observed in the predictive performance of the different variable systems, and the three models did not perform consistently across the feature groups.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Comparative Classification Performance across Different Feature Groups and Models.

https://doi.org/10.1371/journal.pone.0353205.t002

From Table 2, clear differences were observed in the discriminative performance of the three feature groups. When only the 5D feature group was used, the overall performance of the three models was relatively limited. Among them, CatBoost achieved an accuracy of 0.912263, a recall of 0.990945, an F1-score of 0.953444, and a ROC-AUC of 0.773834, while the ROC-AUC values of RF and XGBoost were 0.769571 and 0.727510, respectively. These results indicate that although the 5D variables were able to identify a portion of injury-related crash samples, their overall ability to discriminate between positive and negative samples remained limited.

When only the non-5D feature group was used, model performance improved markedly. For non-5D CatBoost, the accuracy, recall, F1-score, and ROC-AUC reached 0.935351, 0.989247, 0.965213, and 0.924083, respectively. The ROC-AUC values of non-5D XGBoost and non-5D RF were 0.922845 and 0.916222, respectively. Compared with the 5D feature group, the non-5D feature group consistently achieved better performance across all evaluation metrics.

When the combined 5D + non-5D feature group was adopted, model performance improved further. Specifically, 5D + non-5D RF achieved the highest accuracy (0.939456), although its ROC-AUC was 0.926009. 5D + non-5D XGBoost yielded a ROC-AUC of 0.929375. By comparison, 5D + non-5D CatBoost achieved an accuracy of 0.938430, a precision of 0.946341, a recall of 0.988115, an F1-score of 0.966777, and the highest ROC-AUC of 0.941496 among all models. Although RF achieved a slightly higher accuracy under the combined feature group, CatBoost produced the strongest threshold-independent discriminative performance, as reflected by ROC-AUC, while maintaining competitive accuracy, precision, recall, and F1-score.

ROC-AUC was used as the primary model-selection criterion because it evaluates the model’s ability to distinguish injury from non-injury crashes across all possible classification thresholds and is less dependent on a specific decision threshold than accuracy or recall. This is particularly important given the imbalanced outcome distribution of the crash data. Therefore, CatBoost was selected as the benchmark model for subsequent SHAP-based interpretation based on its overall balance of ROC-AUC, classification performance, ability to directly handle categorical variables, and compatibility with SHAP analysis.

It should be noted that the improvement in accuracy from the non-5D CatBoost model to the combined 5D + non-5D CatBoost model was relatively small, increasing from 0.935351 to 0.938430. This indicates that the addition of 5D variables contributed only a limited improvement in accuracy. However, the ROC-AUC increased from 0.924083 to 0.941496, suggesting that 5D variables still provided additional threshold-independent discriminative information when combined with non-5D contextual factors. Therefore, the contribution of 5D variables should be interpreted cautiously: non-5D variables showed strong standalone predictive performance, whereas 5D variables provided complementary spatial-environmental information and contributed substantially to the model-based SHAP explanation within the combined CatBoost model.

To further examine whether the high recall values were driven by SMOTE, an additional robustness check was conducted for the combined 5D + non-5D feature group under different imbalance-handling strategies (S1 Table). The results showed that high recall values were also observed in models trained without SMOTE. For example, RF without SMOTE achieved a recall of 0.994907, which was comparable to that of RF with SMOTE. This suggests that the high recall values were not solely caused by synthetic samples generated by SMOTE, but were also related to the imbalanced distribution of the original testing set. In addition, the class-weight-adjusted XGBoost model achieved a much higher specificity of 0.631868, although its recall decreased to 0.838144. This result indicates a trade-off between sensitivity to injury crashes and discrimination of non-injury crashes. Therefore, the high recall values should be interpreted as strong sensitivity to injury crashes rather than as uniformly strong classification performance across both classes, and model performance should be evaluated using multiple metrics rather than accuracy and recall alone. It should be noted that S1 Table was designed as a sensitivity analysis under alternative imbalance-handling strategies and was not intended to reproduce the main model comparison results in Table 2.

To examine whether spatial dependence remained in the final model, Moran’s I was calculated for the residuals of the CatBoost model using an 8-nearest-neighbor spatial weight matrix and 999 permutations. As shown in S2 Table, Moran’s I was 0.073620, with an expected value of −0.000513 and a pseudo p-value of 0.001. This result indicates that the residuals still exhibited statistically significant but relatively weak positive spatial autocorrelation. Therefore, the SHAP-based importance of location-based 5D variables should be interpreted as model-based relative contributions rather than spatially unbiased causal effects.

4.2. SHAP analysis of overall contributions

After CatBoost was selected as the optimal model, SHAP (Shapley Additive Explanations) was further applied to interpret the model outputs and to identify differences in the contributions of variable categories and individual features to traffic injury risk. For clarity, only the top 10 ranked variables were displayed in the SHAP plots.

As shown in Fig 3, the 5D variables contributed approximately 75% of the total SHAP value, whereas the non-5D variables accounted for roughly 25%. This result suggests that the cumulative explanatory contribution of the 5D variables in the combined model was substantially greater than that of the non-5D variables. In other words, although the model comparison indicated that the non-5D feature group exhibited relatively strong predictive performance when used alone, built-environment-related variables contributed more information and exerted a stronger influence on model outputs within the joint interpretive framework.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. SHAP group-level contributions.

https://doi.org/10.1371/journal.pone.0353205.g003

To further identify the key determinants of traffic injury risk, the SHAP values of the optimal CatBoost model were summarized at the global level to generate the feature importance ranking.

As shown in Fig 4, the top 10 variables were, in descending order of importance, distance to the nearest metro station, lighting condition, road network density, POI mix within 1,000 m, distance to the nearest bus stop, weather condition, number of tourist attractions within 1,000 m, number of schools within 1,000 m, number of hospitals within 1,000 m, and number of shopping malls within 1,000 m. Among these variables, distance to the nearest metro station had the largest mean absolute SHAP value, indicating that it exerted the strongest marginal effect on traffic injury risk and served as the most influential explanatory variable in the current model.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. SHAP feature importance.

https://doi.org/10.1371/journal.pone.0353205.g004

From the perspective of variable type, eight of the top 10 variables belonged to the 5D category, whereas only lighting condition and weather condition were classified as non-5D variables. This finding is consistent with the previous group-level contribution analysis and further suggests that 5D variables not only dominated in terms of total contribution, but also occupied a leading position in the ranking of key features.

In terms of variable attributes, the most important factors were mainly concentrated in four aspects. The first was transit accessibility, as reflected in the high importance of both distance to the nearest metro station and distance to the nearest bus stop. The second was spatial structure, represented by road network density and POI mix. The third was environmental and operational conditions, mainly captured by lighting condition and weather condition. The fourth was destination accessibility, reflected in the numbers of schools, hospitals, tourist attractions, and shopping malls.

Based on the global feature importance ranking, the SHAP beeswarm plot was further used to examine the direction of variable effects and their heterogeneous patterns.

As shown in Fig 5, the distributions of SHAP values were asymmetric across variables, and substantial differences were observed in the degree of dispersion among features, indicating pronounced nonlinearity and individual heterogeneity in their effects on traffic injury risk.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. SHAP beeswarm plot.

https://doi.org/10.1371/journal.pone.0353205.g005

For distance to the nearest metro station, observations with higher values were mainly distributed in the positive SHAP region, whereas lower-value observations were more concentrated around zero or in the negative SHAP region. This pattern suggests that as crash locations become farther from metro stations, traffic injury risk tends to increase; conversely, better metro accessibility is more likely to reduce injury risk. A similar pattern was observed for distance to the nearest bus stop, where areas with poorer bus stop accessibility were more likely to exhibit positive SHAP values and, in turn, a higher injury risk.

Road network density and POI mix showed directional patterns different from those of the above accessibility variables. Higher road network density and greater POI mix were more frequently associated with negative SHAP values, indicating a suppressing effect on traffic injury risk. By contrast, lower road network density and lower POI mix were more likely to shift the model output toward injury occurrence.

Although lighting condition and weather condition belonged to the non-5D category, they still exhibited relatively high importance and a wide spread of points in the beeswarm plot, suggesting that they had substantial effects on traffic injury risk.

Although the numbers of schools, hospitals, tourist attractions, and shopping malls were less important overall than transit accessibility and road network characteristics, all four variables were ranked among the top 10 features. This result indicates that destination concentration also provides a non-negligible explanation for traffic injury risk.

4.3. Nonlinear effects of key continuous variables

As shown in Fig 6, a clear nonlinear relationship was observed between distance to the nearest metro station and the corresponding SHAP value. Overall, the SHAP value exhibited an upward trend as the distance increased, indicating that crash locations farther from metro stations exerted a stronger positive effect on traffic injury risk. In particular, within the shorter distance range, SHAP values were mostly concentrated in the negative region. As the distance increased further, however, the SHAP values gradually shifted to the positive region and continued to rise, suggesting that reduced metro accessibility substantially increased traffic injury risk.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. SHAP dependence plot for distance to the nearest metro station.

https://doi.org/10.1371/journal.pone.0353205.g006

As shown in Fig 7, a relatively clear negative relationship was observed between POI mix and the corresponding SHAP value. As POI mix increased from lower to higher levels, the SHAP value gradually shifted from positive to negative, indicating that a higher degree of functional mix generally helped reduce traffic injury risk.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. SHAP dependence plot for POI mix Index within 1000 m.

https://doi.org/10.1371/journal.pone.0353205.g007

As shown in Fig 8, the relationship between the number of schools within 1,000 m and the corresponding SHAP value was not monotonic or linear, but instead exhibited noticeable fluctuations. Within the low-to-moderate range of school counts, SHAP values could be either positive or negative. Overall, however, the magnitude of the variable’s effect on model output became more pronounced as the number of schools increased, suggesting that the influence of school concentration on traffic injury risk was highly context-dependent.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. SHAP dependence plot for School POI Count within 1000 m.

https://doi.org/10.1371/journal.pone.0353205.g008

5.1. Joint effects of 5D and non-5D factors

The results of this study showed that the combined 5D + non-5D feature group consistently outperformed the models built with either the 5D or non-5D feature group alone, indicating that traffic injury risk is shaped not by a single category of determinants, but by the interaction between the spatial built environment and the immediate crash context. An important finding is that the non-5D feature group demonstrated relatively strong predictive performance when modeled independently, whereas the SHAP group-level contribution analysis indicated that 5D variables accounted for a substantially larger cumulative contribution in the combined model. These findings should not be interpreted as contradictory. Rather, they reflect two different dimensions of model understanding: These findings should be interpreted by distinguishing predictive performance from model-based explanation. The non-5D feature group achieved strong standalone performance, suggesting that immediate crash-context factors have substantial direct discriminative value for distinguishing injury from non-injury crashes. The addition of 5D variables produced only a marginal improvement in accuracy, indicating that their incremental contribution to overall predictive accuracy was limited. Nevertheless, the increase in ROC-AUC and the SHAP results suggest that 5D variables provided complementary spatial-environmental information and contributed substantially to the internal explanation of the combined CatBoost model. Therefore, the 5D variables should not be interpreted as the primary source of predictive performance, but rather as an important component for explaining spatial and built-environment-related heterogeneity in injury risk.

This pattern suggests that traffic injury risk is characterized by both immediate situational sensitivity and spatial embeddedness. On the one hand, variables such as weather, lighting condition, and pavement status directly influence visibility, vehicle handling, and conflict severity at the moment of crash occurrence, thereby exerting an immediate effect on injury outcomes. On the other hand, 5D variables—including transit accessibility, roadway design, road network density, and destination distribution—more fundamentally shape the spatial organization, activity intensity, and exposure structure of the crash environment. In this sense, non-5D factors may function more as proximate triggers of injury occurrence, whereas 5D factors provide the underlying contextual conditions within which injury risk is generated and differentiated.

5.2. Dominant role of transit accessibility and spatial structure

Among all variables, distance to the nearest metro station, distance to the nearest bus stop, road network density, and POI mix exhibited particularly high importance, indicating that transit accessibility and spatial structure play central roles in shaping traffic injury risk. Better public transport accessibility often reflects more mature roadway infrastructure, higher-quality street environments, and more organized traffic operations. Similarly, higher road network density and stronger functional mix are typically associated with finer-grained street networks, shorter travel distances, and a more balanced distribution of urban activities. Taken together, these characteristics may influence not only exposure patterns and conflict opportunities, but also the likelihood that a crash escalates into an injury event.

The association between proximity to metro stations and lower injury risk may be explained through several complementary mechanisms. First, metro station areas usually concentrate a large number of pedestrians and transit users. This may generate a “safety in numbers” effect, whereby higher pedestrian presence increases drivers’ awareness and encourages more cautious driving behavior. Second, areas with better metro accessibility are often characterized by transit-oriented and compact urban development, including denser street networks, more frequent intersections, shorter blocks, lower automobile dependence, and more complete pedestrian facilities. These built environment characteristics may contribute to lower vehicle operating speeds and reduce the probability that a crash results in injury.

However, this interpretation should be made cautiously. The present study did not include direct measures of pedestrian volume, traffic flow, or vehicle operating speed. Therefore, it is not possible to determine whether the observed association is mainly attributable to a “safety in numbers” effect, lower vehicle speeds in transit-oriented environments, or the combined influence of both mechanisms. Accordingly, proximity to metro stations should be interpreted as a location-based indicator associated with safer and more organized urban traffic environments, rather than as a direct causal factor reducing injury risk.

At the same time, the numbers of schools, hospitals, tourist attractions, and shopping malls were all ranked among the top 10 variables, suggesting that destination accessibility captures more than the intensity of local activities alone. It may also affect injury outcomes by reshaping the interaction patterns among pedestrians, non-motorized users, and motor vehicles, thereby altering the frequency and severity of potential traffic conflicts.

5.3. Comparison with previous studies

The findings of this study should be interpreted in relation to previous literature on the built environment and traffic safety. Previous studies have suggested that built environment characteristics may influence traffic safety through multiple pathways, including crash exposure, crash frequency, and crash severity or injury outcomes [7–12,25]. For example, Saha et al. [8] proposed that the built environment affects traffic safety indirectly through exposure, conflict frequency, and operating speed, while Merlin et al. [9] emphasized the importance of distinguishing crash exposure from crash risk. Asadi et al. [12] further showed that built environment factors are associated with crash probability, frequency, and severity, and Sung et al. [25] found that compact and mixed-use environments may have different associations with pedestrian crash occurrence and fatal crash risk.

However, the present study does not directly analyze crash frequency. Instead, it focuses on whether a recorded crash resulted in injury, that is, injury occurrence conditional on crash occurrence. Within this conditional framework, variables related to transit accessibility, road network structure, and functional mix showed important model-based explanatory contributions. Therefore, the present findings should be understood as evidence that built environment factors are associated with injury outcomes among recorded crashes, rather than as direct evidence that these factors affect the overall frequency of injury crashes.

It should be noted that the SHAP-based feature importance reflects the relative contribution of variables within the fitted CatBoost model. The Moran’s I test of model residuals indicated statistically significant but relatively weak positive spatial autocorrelation, suggesting that some spatial dependence was not fully captured by the current feature set. Therefore, the importance of location-based 5D variables, such as distance to transit, road network density, and POI mix, may partly reflect spatially clustered built environment characteristics. These results should therefore be interpreted as model-based associations rather than causal spatial effects.

5.4. Practical implications

From a practical perspective, the findings of this study suggest that traffic safety governance should not be limited to reactive responses after crashes occur, nor should it rely solely on isolated roadway engineering measures. Greater emphasis should instead be placed on the coordinated optimization of the built environment surrounding crash locations and the immediate contextual conditions under which crashes occur. On the one hand, regional spatial structure should be improved through better transit stop placement, road network organization, functional mix, and destination configuration. On the other hand, safety management and risk warning systems should be strengthened under immediate high-risk conditions related to lighting, weather, and pavement status. Only by integrating spatial environment optimization with contextual risk control can the likelihood that a crash develops into an injury outcome be reduced more effectively.