category: literaturenote citekey: wangcontextdependenteffectsbuilt2025 title: Context-dependent effects of built environment factors on pedestrian-injury severities with imbalanced and high dimensional crash data authors: "Wang, Zehao; Fan, Wei" year: 2025 date: 2025-00-00 2025 doi: 10.1016/j.aap.2025.108104 publication: "Accident Analysis & Prevention" url: "https://doi.org/10.1016/j.aap.2025.108104" zotero_key: KQQ7S29I zotero_storage: BS3QHPPR collections: transpordi_planeerimine folder: 001_artiklid/Road Safety Indicators firstAuthor: "Wang, Zehao"

status: converted

Contents lists available at ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

Context-dependent effects of built environment factors on pedestrian-injury severities with imbalanced and high dimensional crash data

Zehao Wang , Wei Fan *

USDOT Center for Advanced Multimodal Mobility Solutions and Education (CAMMSE), Department of Civil and Environmental Engineering, University of North Carolina at Charlotte, 9201 University City Boulevard, Charlotte, NC 28223-0001, United States of America

ARTICLE INFO

Keywords: Pedestrian injury severities Built environment Heterogeneity Imbalanced data Feature selection Machine learning

ABSTRACT

Built environment is an important component that influences pedestrian injury severities in pedestrian-vehicle crashes. Previous studies indicated that the effects of various built environment factors on pedestrian injury severities are heterogeneous, meaning the direction of their effects varies under different conditions. This heterogeneity poses a challenge in developing effective countermeasures. Therefore, this study aims to explore the context-dependent effects of built environment factors and uncover the potential sources of this heterogeneity. Using North Carolina crash data (2017–2022) as a case study, an enhanced machine learning framework that incorporates a conditional tabular generative adversarial network and wrapper-based feature selection techniques is designed to investigate the context-dependent effects. The enhancements seek to mitigate the adverse impacts associated with imbalanced and high dimensional crash data. Model results show that 13 out of the top 19 important factors belong to built environment factors, with 11 of these factors (i.e., restaurants present and road class) being heterogeneous. The context-dependent effects analysis can help identify conditions for the positive effects of heterogeneous factors. For example, restaurants near the local routes or in urban areas, highergrade roads (i.e., US, Interstate, and NC routes) in better economically developed counties, and signs or signals control roads under dark (with or without) roadway light conditions have clear positive effects on pedestrian injury severities. These findings can provide guidance for urban planning and policy development, thereby promoting the sustainable development of cities, regions, and transportation systems.

1. Introduction

Walking, as a sustainable transportation mode, has many benefits such as lowering obesity rates, mitigating traffic congestion, and reducing vehicle emissions (Su et al., 2021; Baek and Lim, 2024). However, during rapid urbanization and motorization, some adversely built environments may discourage walking and expose pedestrians to danger. According to the records of the National Highway Traffic Safety Administration (NHTSA) and Smart Growth America (SGA), in the United States, a total of 7,522 pedestrians were killed in vehicular crashes in 2022 and 30 percent of them were from low-income areas with poor built environments (NHTSA, 2024; SGA, 2024). Previous studies have also confirmed that built environments, such as land use patterns, sociodemographic features, and road design characteristics, play a key role in pedestrian safety (Ma et al., 2022; Mirhashemi et al., 2022). Thus, an in-depth understanding of their mechanisms is essential to improving the safety of the walking environment.

Prior studies have indicated that the effects of built environment factors on pedestrian injury severities are heterogeneous. For instance, across all road types in North Carolina (NC), NC routes increased the probability of severe injuries for 54.7 % of pedestrians, while demonstrating negative effects in the remaining observations (Li et al., 2021); primary roads increased the probability of severe injuries for 60.26 % of the vulnerable road users, while demonstrating negative effects in the remaining observations (Sun et al., 2023). The reason is that these factors exerted varying impacts on injury severities under different conditions (Mannering et al., 2016; Azimi et al., 2020). However, most studies examined the presence of heterogeneous factors but ignored analyzing the possible sources of heterogeneity, making it difficult to provide effective safety countermeasures (Azimi et al., 2020). The contextdependent effect refers to the phenomenon that the influence of built environment variables on crash injury severity may vary depending on different contexts (Merlin et al., 2020). These variations arise from underlying heterogeneity in population characteristics, infrastructure, and

E-mail addresses: zwang54@uncc.edu (Z. Wang), wfan7@uncc.edu (W. Fan).

^* Corresponding author.

travel behavior patterns. For instance, urban areas were more likely to suffer fatal crashes during summer, but the probability of fatal crashes decreased in the winter (Sun et al., 2023); road segments had a positive effect on severe injuries in commercial land use but a negative effect on residential land use (Yang et al., 2023); expressways were associated with severe injuries in hazardous terrain and with minor injuries in nonhazardous terrain (Wei et al., 2024); intersections increased the probability of fatal injuries in morning rush hours but decreased the probability in evening rush hours (Abrari Vajari et al., 2020); many road-level factors influenced safety through their varying effects across different regions (Han et al., 2018; Lym and Chen, 2021). Therefore, exploring the context-dependent effects of heterogeneous factors can identify the possible sources of heterogeneity and conditions for the positive impact of heterogeneous factors.

To fill this gap, this research presents an explainable machine learning model to analyze the context-dependent effects of heterogeneous factors and uncover the potential sources of heterogeneity regarding their impact on pedestrian injury severities. However, the crash data collected from North Carolina between 2017 and 2022 is imbalanced and high dimensional, posing challenges to the fitting ability of the explainable machine learning model. Thus, some enhancements are developed. Specifically, a conditional tabular generative adversarial network is designed to overcome the adverse effects of imbalanced crash data while preserving the original distribution and avoiding the generation of noisy samples. In addition, an improved wrapper-based feature selection method is employed to remove the noisy features. Finally, scenario-specific countermeasures are proposed to mitigate serious injury severity and promote pedestrian-friendly development in cities and regions.

2. Literature review

2.1. Pedestrian injury severity studies considering built environment

Table 1 summarizes the literature on pedestrian injury severity considering built environment factors. According to existing studies, pedestrian and driver characteristics, environment and temporal characteristics, road infrastructure and vehicle characteristics have been well investigated (Ma et al., 2022; Mirhashemi et al., 2022). Except for the road infrastructure factors that can be directly recorded by police, other built environment factors need to be supplemented by multisource data. However, the effects of some built environment factors are heterogeneous. For example, 54.7 % of pedestrians were more likely to suffer serious injury on NC roads (Li et al., 2021), making it difficult to develop a unified strategy for management. Due to the heterogeneity, the effects of some facilities (i.e., restaurants) were contradictory in different studies (Park and Ko, 2020; Dumbaugh et al., 2023). Besides, sociodemographic and economic characteristics factors were usually used in crash risk and crash exposure modeling (Merlin et al., 2020), thus their impacts on pedestrian injury severities are not well documented.

For the analyzing methods, statistical models, explainable machine learning models and hybrid approaches integrating traditional statistical models with machine learning techniques have been widely used in road safety modeling (Mannering et al., 2016; Chang et al., 2022; Adanu et al., 2023; Agheli and Aghabayk, 2024). However, the built environment factors are interrelated, for example, restaurants and shopping centers are typically present in commercial land uses (Wang et al., 2024a). Thus, statistical models that assume independence between factors may not reflect real-world conditions. Besides, the effects of built environment factors may vary across observations (Li et al., 2021; Sun et al., 2023). The advanced heterogeneity models (i.e., the random parameter logit model and its advanced variants) have been used to examine the presence of heterogeneity and improve the accuracy of parameter estimates (Mannering et al., 2016; Khan and Habib, 2022; Adanu et al., 2023). To elucidate the possible sources of heterogeneity,

Table 1

Authors	Methodology	Key findings
Clifton et al. (2009)	Ordered probit model	Positive factors: not in a crosswalk area, dark, children,
		older people. Negative factors: transit access
		and greater pedestrian
Mohamed	Latent class with ordered	connectivity. Positive factors: heavy vehicles,
et al.	probit model	dark lighting conditions, mixed
(2013)		land use, and major roads.
		Negative factors: crossing at
		intersections.
Xin et al. (2017)	Random parameters generalized ordered	Positive factors: middle-aged and elderly pedestrians, teen
	probability model with	drivers.
	heterogeneity in means and	Negative factors: the presence
	variances	of a school and bus stop.
Prato et al.	Linearized spatial logit model	Positive factors: older
(2018)		pedestrians and intoxicated pedestrians, heavy vehicles,
		higher speed limits.
		Negative factors: shopping
		areas, residential areas, and
		walking traffic density.
Zhai et al. (2019)	Random parameter logistic regression	Positive factors: crossing from offside, negligently driving,
		buses involved crashes.
		Negative factors: one way road
		type, 10:00–16:00 period,
		interaction between raining and footpath overcrowding.
		Heterogeneous factors:
		carriageway (positive in 35.87
		% of observations)
Park and Ko	Hierarchical model	Positive factors: signalized
(2020)		crosswalks and intersections.
		Negative factors: parks, convenience stores, and
		restaurants.
Li et al.	Random parameters logit	Positive factors: Ambulance
(2021)	approach with heterogeneity	rescue, urban areas, curved
	in means and variances	roadway.
		Heterogeneous factors: NC routes (positive in 54.7 % of
		observations), State secondary
		routes (positive in 8.06 % of
		observations).
Khan and Habib	Latent segmentation-based ordered logit model	Positive factors: arterial roads, curved roads, sloped roads, and
(2022)		roundabouts.
		Negative factors: higher mixed
		land use, length of the sidewalk,
		near the schools, interaction
		effects between the communication device and
		straight roads.
Chang et al.	XGBoost	Positive factors: commercial
(2022)		and residential land use density,
		road segment betweenness and
		closeness, local physical conditions, presence of
		pedestrian crossings, and
		interaction effects between high
		residential density and low
Adanu et al.	Random parameter	intersection density. Positive factors: walking along
(2023)	multinomial logit model	the road, crossing the road, male
		drivers, and urban areas.
Barbour et al.	Partially temporally	Positive factors: Four-way
(2024)	constrained random	intersection, road with curb.
	parameters logit model	Heterogeneous factors: urban
		environment (positive in 33.36 % of observations), marked
		crosswalk (positive in 68.05 % of
		observations)

Table 1 (continued )

Authors	Methodology	Key findings
Khan et al. (2024)	Deep pre-trained convolutional neural network	Positive factors: Gender of the pedestrian, driver and pedestrian impairment, pedestrian age and urban environments.

some studies have further introduced interaction terms in statistical models. For example, Azimi et al. (2020) added three interaction variables for factors with random coefficients in the random parameter ordered logit model. Chen et al. (2021) constructed four interaction terms in the Bayesian random-parameter Tobit model and the effects of high intersection density changed under the impact of medium- and heavygoods vehicle proportion and Taxi vehicle proportion. Khan and Habib (2022) investigated the interaction between distracted driving type and built environment characteristics on pedestrian injury severity by constructing interaction terms in a latent segmentation-based ordered logit model. However, it is difficult to implement in multi-dimensional built environment factors due to the large number of combinations (Agheli and Aghabayk, 2024). Unlike constructing interaction terms explicitly in statistical models, explainable machine learning models can automatically capture all interactions between factors during training, thereby facilitating the precise identification of potential context-dependent effects with relatively low computational costs (Agheli and Aghabayk, 2024; Samerei and Aghabayk, 2024). For example, Zhang et al. (2024) used an XGBoost model and bivariate partial dependence plots to quantify the context-dependent effects and found that the effects of primary and secondary roads on speeding risk are different under specific conditions such as bike lane width, land use density, and rainfall weather. Chen et al. (2025) applied a Random Forest model and SHAP interaction value to uncover the context-dependent effects and found that the effects of posted speed limit on crash severity are influenced by driver gender. Besides, some hybrid approaches have also been employed. For example, Sun et al. (2023) used the random parameters logit model to identify the heterogeneity factor and employed the Random Forest-SHAP interaction value to explore the heterogeneity sources of heterogeneous factors. Agheli and Aghabayk (2024) used the CatBoost-SHAP method to identify the important variables for advanced heterogeneity models and explore the interaction effects between the variables. Azmeri Khan et al. (2023) employed a decision tree to generate higher-order interactions for random parameters hierarchical ordered probit model. However, constructing complex heterogeneity models on large-scale crash datasets is time-consuming and can consider only a limited number of interaction terms. Moreover, some studies have demonstrated that explainable machine learning can capture heterogeneity (Chang et al., 2022; Jin et al., 2023). Based on these considerations, explainable machine learning is used in this research.

2.2. Research on enhanced machine learning framework

As summarized in Table 2, explainable machine learning models have been widely used in traffic safety modeling, including LightGBM and Random Forest. Among all the explanation strategies, SHapley Additive exPlanations (SHAP) can decompose the feature contribution to each sample, which helps to find the heterogeneity factor, and accurately explain the interaction effects of any two features, which helps to explore the context-dependent effect (Li et al., 2023; Sun et al., 2024). To improve the fitting ability of these models, several issues need to be addressed when developing and applying explainable machine learning models.

Firstly, due to the rare nature of serious injury crashes (i.e., incapacitating injury and fatal injury), their proportions are usually small in all crashes (Li et al., 2023; Chen et al., 2024). Therefore, crash severity data is usually imbalanced. Machine learning models trained in

Table 2 Summary of literature on enhanced machine learning.

Authors	Machine learning model	Enhanced technique	Explanation method
Shi et al.	XGBoost	Data enhancement:	−
(2019)		Under-samplings by
		repeated ENN Hyperparameter
		search: Grid Search
		Feature selection:
		Recursive feature
		elimination
Chang et al.	XGBoost, GBDT,	Data enhancement:	SHAP
(2022)	Logistic regression	Random under sampling
		Hyperparameter
		search: Grid search
Wang et al.	Stacking	Hyperparameter	Permutation
(2023b)		search: Bayesian	feature
		optimization	importance
		Feature selection: NSGA-II wrapper
		algorithm
Li et al.	Logistic regression,	Data enhancement:	SHAP
(2023)	Multilayer	Wasserstein generative
	perceptron, Random	adversarial network
	Forest, AdaBoost	with gradient penalty
	and XGBoost
Ashraf et al. (2023)	XGBoost, AE-based DNN	Data enhancement: ADASYN	SHAP
		Hyperparameter
		search: Grid search
Sun et al.	LightGBM, XGBoost,	Data enhancement:	SHAP; LIME
(2024)	CatBoost, Random	ADASYN
	Forest	Hyperparameter
		search: Bayesian optimization
		Feature selection:
		NSGA-II wrapper
		algorithm
Ziakopoulos	XGBoost, SVM,	Data enhancement:	SHAP
(2024)	Random Forest,	SMOTE
	Deep NN	Hyperparameter search: Random search
Zhu et al.	Convolutional	Data enhancement:	SHAP
(2024b)	neural network	Variational
		autoencoder
		generative adversarial
		network
		Feature selection: Average gain
		importance
Khan et al.	Random forest,	Data enhancement:	Boruta
(2024)	SVM, GBDT, Naíve	Both under and over	algorithm
	Bayes, KNN, CART,	sampling	sorting
	Inception-v3	Hyperparameter	features
		search: Grid search Feature selection:
		Boruta algorithm
Xue et al.	Random Forest,	Data enhancement:	SHAP
(2024)	LightGBM, GBDT,	Cluster-centroid based
	XGBoost, and Extra	under-sampling
	trees	Hyperparameter
		search: Bayesian
		optimization Feature

		selection: Recursive
Chen et al.	Logistic regression	feature addition Data enhancement:	−
(2024)		Conditional Tabular GAN-random

imbalanced data are prone to learn majority class samples (i.e., no injury or possible injury crashes), leading to biased models and potentially inaccurate interpretations (Khan et al., 2024). To address this issue, resampling methods are widely used, such as undersampling (Shi et al., 2019; Chang et al., 2022) and oversampling (Schlogl, 2020; Wen et al., 2021a). However, undersampling the majority class data may lead to an underrepresentation of the feature space distribution (Johnson and Khoshgoftaar, 2019; Wan et al., 2024). Oversampling methods can avoid information loss, but duplicating samples fails to capture the correlations between variables and can easily generate overfitting issues (Chen et al., 2025). Therefore, these methods have limited effectiveness in data augmentation and may even degrade the performance of subsequent model training. The generative adversarial network (GAN)-based data augmentation methods, such as Wasserstein GAN with gradient penalty (WGAN-GP), conditional tabular GAN (CTGAN), and Variational autoencoder GAN, are proposed to augment minority class with the advantage of increasing diversity and generating fewer noisy samples (Chen et al., 2024; Li et al., 2023; Zhu et al., 2024b). However, GAN-based models are more suitable for data with continuous variables, such as image and video data (Xu et al., 2019; Chen et al., 2025). The conditional tabular GAN is designed to solve data containing numerous categorical variables and is better suited to crash data that are usually described using discrete variables (Chen et al., 2024).

Secondly, the inclusion of numerous built environment-related factors into raw crash data resulted in high dimensional crash data (Nasrollahzadeh et al., 2021; Sun et al., 2024). Although high dimensional features can describe crash characteristics more accurately, they also present challenges for analysis. For example, some features may be irrelevant or redundant for pedestrian injury severity. The model still processes these variables, resulting in lower computational efficiency and increased training costs (Khan et al., 2024). Furthermore, some features may contain excessive noise due to data acquisition quality. These features can mislead the model, resulting in degraded performance and reduced generalization ability (Ali et al., 2024). Moreover, high dimensional features may diminish the model's interpretability (Shi et al., 2019). In crash management, countermeasures should be developed based on the most critical factors rather than considering all possible ones. Wrapper-based feature selection, such as recursive feature elimination, recursive feature addition, and Boruta algorithm, have been widely used to identify the key features (Khan et al., 2024; Rangam et al., 2024). However, existing research usually uses average gain or split weight to estimate feature importance, leading to an overestimation of high-cardinality categorical variables (Huang et al., 2024). Besides, the hyperparameters of machine learning are initialized using default values or based on all features, resulting in a suboptimal feature subset (Chen et al., 2022). Additionally, the effectiveness of different feature selection strategies has also not been adequately compared (Ali et al., 2024).

2.3. Research gaps

This study aims to address the research gaps identified in the existing literature, which are outlined below:

(1) Previous studies have demonstrated that the built environment is a key component affecting pedestrian injury severities. However, most of their effects are heterogeneous, posing challenges for policymaking. Thus, simply examining the presence of heterogeneity is insufficient, and exploring its potential sources is necessary. Context-dependent effects can greatly help uncover the effects of a specific factor under different conditions or contexts. Analyzing the context-dependent effects of heterogeneous factors can determine their positive or negative effects under the influence of certain factors.
(2) Although the explainable machine learning models have many advantages in illustrating the context-dependent effects, the imbalanced and high-dimensional crash data raises some issues. Previous studies usually used sampling strategies to enhance the minority samples (i.e., serious injury crashes) but these methods ignored the original distribution of the variables and thus introduced the variance issue. Previous studies usually used traditional wrapper algorithms to remove noisy features; however, improving the accuracy of feature importance estimation and comparing the effectiveness of different algorithms are still

insufficient. Thus, conditional tabular GAN and three improved wrapper-based feature selection methods are developed, compared and used to enhance the machine learning model in this study.

3. Data collection and description

This study uses the police-reported pedestrian-vehicle crash data collected from 2017 to 2022 in North Carolina. After directly removing the missing or incomplete crash records, the data contains a total of 13,152 observations, with each observation involving only one pedestrian and one vehicle. In the original data, the injury severity of pedestrians is categorized into five categories, including no–, possible-, non-incapacitating, incapacitating, and fatal injury. Due to the fact that incapacitating and fatal crashes can bring the worst consequences, identifying the factors affecting these crashes is critical (Li et al., 2023). Moreover, as no-injury crashes represent the least serious type, their underlying mechanisms differ from those of serious injury crashes (Se et al., 2022; Li et al., 2023). Including these cases can help the model more accurately distinguish between different levels of crash severity. Thus, the five levels of data are aggregated into two classes: non-serious injury (i.e., no injury, possible injury, and non-incapacitating injury) and serious injury (i.e., incapacitating injury and fatal injury), which account for 75.87 % (9,979) and 24.13 % (3,173) of the total crashes, respectively. To assess the degree of imbalance, the imbalance ratio is calculated as the ratio of the sample size of the majority class/nonserious injury crashes to that of the minority class/serious injury crashes (Zhu and Meng, 2022; Chen et al., 2025). The computed imbalance ratio is 3.14:1, indicating that the dataset is imbalanced.

Based on previous research and the "5D" principle (Ewing and Cervero, 2010), the built environment factors are collected and presented as follows:

(1) Sociodemographic and economic characteristics, such as proportion of different age groups (Su et al., 2021), the proportion of different gender groups (Wu et al., 2021) and poverty rate (Ding et al., 2018), correlate with safety culture and the intensity of enforcement measures, and thus can raise the potential inequities issues in road safety (Su et al., 2021; Zhai et al., 2024). These factors based on country units are collected from the North Carolina Office of State Budget & Management.
(2) Land use characteristics can reflect the traffic demand and travel behavior (Wang et al., 2024b). Except for the type of regional development and functional zones (rural or urban areas), the points of interest (POI) are also correlated with the land use, such as shops, churches, schools, restaurants, and bar (Wang et al., 2024a). To obtain these factors, a common 1-km buffer zone is created for each crash point by the ArcGIS platform (Laviolette et al., 2022) and POI obtained from Open Street Map is used to count whether it appears in the created buffer.
(3) Location characteristics can be described by the distance to the nearest government building, such as the police department and Emergency Medical Services (EMS) agency. These factors can reflect the impact of the response time after a crash has occurred (Fu et al., 2022). The location information of these two government institutions is obtained from Open Street Map and a statewide GIS inventory of North Carolina OneMap, respectively.
(4) Road infrastructure characteristics, such as traffic control type and road construction, can directly determine conflict situations between road users and the North Carolina Department of Transportation (NCDOT) records these features in detail after a crash occurs (Li et al., 2021).

It is worth noting that some continuous variables, such as the sociodemographic and economic characteristics and distance to emergency medical services and police station, are transformed into categorical variables (i.e., low, medium, and high) using the Natural Breaks (NB) method. This method calculates the optimal breakpoints to minimize variance within intervals and maximize variance between intervals. (Cao et al., 2013). The discretized variables help reduce the influence of outliers, enhance the model's robustness, and develop scenario-specific policies. Besides these factors, some non-built environment factors are also collected from NCDOT, such as human-related factors, vehiclerelated factors, and crash characteristics. In addition, temporal instability may also influence the performance of machine learning models (Ali et al., 2024). Thus, several temporal characteristics, including peak hour, workday, season, and period (before or after the COVID-19 outbreak), are extracted from the timestamp of each crash. This can help machine learning models to learn the interactive effects of the temporal factor on other factors during the training phase. Finally, a total of 43 influencing factors are collected from seven aspects, as shown in Table 3. Compared with other studies (Kuo et al., 2024; Rangam et al., 2024), the number of these features suggests that the crash data is high dimensional. Therefore, there is a need to use advanced feature selection techniques to deal with the adverse effects of high dimensional data.

4. Methodology

4.1. Conceptual framework

In this study, we focus on exploring the context-dependent effects of built environment factors affecting pedestrian injury severities. To construct a comprehensive set of risk factors, the pedestrian crashes are first geocoded in the North Carolina county map, and built environment factors collected from multiple GIS data sources are supplemented for each crash record. Then, an enhanced machine learning framework is designed to accurately capture the relationships between pedestrian injury severities and various risk factors, as shown in Fig. 1.

Specifically, the advanced LightGBM model is enhanced by two state-of-the-art technologies. To avoid the biased estimation (usually bias to the majority class) caused by imbalanced crash data, a conditional tabular generative adversarial network incorporating the raw data distribution is designed to generate minority class samples (i.e., serious injury crashes) with less noise. To reduce the model overfitting on high dimensional crash data, wrapper-based feature selection methods (i.e., recursive feature elimination, recursive feature addition, and Boruta algorithm) improved by Bayesian hyperparameter optimization and SHAP values-based feature importance estimation are used to remove the noisy features. The effectiveness of the enhanced model is confirmed through a series of evaluation metrics, such as sensitivity, specificity, F1-score, G-means, and AUC (Li et al., 2023; Ali et al., 2024).

Furthermore, the enhanced model is interpreted by the SHAP approach. The feature importance rank and cumulative importance are first conducted to identify the core factors. Then, the fixed effects factors and heterogeneous factors are identified. Fixed effects factors can be used directly in policy formulation. The context-dependent effects among heterogeneous built environment factors and the contextdependent effects between heterogeneous built environment factors and non-built environment factors are examined to provide insights into potential sources of heterogeneity.

4.2. Conditional tabular generative adversarial network

Conditional tabular generative adversarial network (CTGAN) is developed based on generative adversarial networks (GAN) for modeling the distributions of tabular data and sampling rows from the distributions (Xu et al., 2019). CTGAN has two neural network models: the conditional generator network G and the discriminator network D, as shown in Fig. 2. In the training process, the generator produces synthetic data, and the discriminator tries to distinguish the generated fake data and the real data (Li et al., 2023). These two networks compete with each other and enhance themselves through a loss function that can be expressed as Eq. (1).

$$L = E{x \ p{\text{data}}}[D(x)] - E{x \ p{G}}[D(x)] + \lambda E{x \ p{\text{pendity}}} \left[ \left( \left| \nabla{x} D(x) \right|{2} - 1 \right)^{2} \right]$$ (1)

Table 3 Descriptive statistics of variables.

Variable	Description	Total	Pedestrian-Injury Severity
			Non serious injury	Serious injury
		13152	9979 (75.87 %)	3173 (24.13 %)
Sociodemographic and economic	Built Environment related Characteristics
characteristics Male Rate	1 = Low (≤0.4788)	3473	2731 (78.64 %)	742 (21.36 %)
	2 = Medium (0.4788,0.5023)	9217	6920 (75.08 %)	2297 (24.92 %)
	3=High (≥0.5023)	462	328 (71.00	134 (29.00
Youth Population Rate (age <	1 = Low (≤0.2930)	2009	%) 1391 (69.24 %)	%) 618 (30.76 %)
25)	2 = Medium (0.293,0.3583)	9770	7665 (78.45 %)	2105 (21.55 %)
	3=High (≥0.3583)	1373	923 (67.23 %)	450 (32.77 %)
Elderly Population Rate (age >	1 = Low (≤0.1479)	6345	5002 (78.83 %)	1343 (21.17 %)
64)	2 = Medium (0.1479,0.2233)	6109	4504 (73.73 %)	1605 (26.27 %)
	3=High (≥0.2233)	698	473 (67.77	225 (32.23
Unemployment Rate	1 = Low (≤0.0450)	8112	%) 6344 (78.21	%) 1768 (21.79
	2 = Medium (0.0450,0.0660)	3297	%) 2357 (71.49	%) 940 (28.51
	3=High (≥0.0660)	1743	%) 1278 (73.32	%) 465 (26.68
Poverty Rate	1 = Low (≤0.1411)	6410	%) 5053 (78.83	%) 1357 (21.17
	2 = Medium (0.1411,0.2242)	5636	%) 4144 (73.53 %)	%) 1492 (26.47 %)
	3=High (≥0.2242)	1106	782 (70.71 %)	324 (29.29 %)
Land use characteristic Shop	1=No	8268	6074	2194
			(73.46 %)	(26.54 %)
	2 = Yes	4884	3905 (79.95 %)	979 (20.05 %)
Church	1=No	7261	5491 (75.62 %)	1770 (24.38 %)
	2 = Yes	5891	4488 (76.18	1403 (23.82
			%)	%)

(continued on next page)

Table 3 (continued )

Variable	Description	Total	Pedestrian-Injury Severity		Variable	Description	Total	Pedestrian-Injury Severity
			Non serious injury	Serious injury				Non serious injury	Serious injury
		13152	9979 (75.87 %)	3173 (24.13 %)			13152	9979 (75.87 %)	3173 (24.13 %)
	2 = Yes	4106	3270 (79.64 %)	836 (20.36 %)		3=NC roads	2726	1728 (63.39 %)	998 (36.61 %)
Restaurants	1=No	6646	4829 (72.66 %)	1817 (27.34 %)		4 = Local roads	8797	7393 (84.04 %)	1404 (15.96 %)
	2 = Yes	6506	5150 (79.16 %)	1356 (20.84 %)	Road Configuration	1=One-Way, Not Divided	827	707 (85.49 %)	120 (14.51 %)
Bar	1=No	10,993	8183 (74.44 %)	2810 (25.56 %)		2 = Two-Way, Not Divided	9858	7472 (75.80 %)	2386 (24.20 %)
	2 = Yes	2159	1796 (83.19 %)	363 (16.81 %)		3 = Two-Way, Divided	2467	1800 (72.96 %)	667 (27.04 %)
Rural or Urban	1 = Rural	3280	1955 (59.60 %)	1325 (40.40 %)	Road Surface	1=Smooth Asphalt/Coarse Asphalt	12,758	9673 (75.82 %)	3085 (24.18 %)
	2 = Urban	9872	8024 (81.28	1848 (18.72		2=Concrete/Grooved Concrete	394	306 (77.66	88 (22.34
Development	1 = Residential	4709	%) 3792 (80.53	%) 917 (19.47	Road Condition	1 = Dry	11,193	%) 8510 (76.03	%) 2683 (23.97
	2=Commercial	6214	%) 4852 (78.08	%) 1362 (21.92		2 = Wet	1858	%) 1382 (74.38	%) 476 (25.62
	3 = Industrial	66	%) 49 (74.24	%) 17 (25.76		3=Others (e.g. Ice, Snow, Water)	101	%) 87 (86.14	%) 14 (13.86
	4 = Institutional	330	%) 299 (90.61	%) 31 (9.39	Number of Lanes	1 = Low (≤2)	6838	%) 5367 (78.49	%) 1471 (21.51
	5 = Farms, Woods, Pastures	1833	%) 987 (53.85	%) 846 (46.15		2 = Medium [3,4]	4110	%) 3074 (74.79	%) 1036 (25.21
Road infrastructure characteristics			%)	%)		3=High (≥5)	2204	%) 1538 (69.78	%) 666 (30.22
Work zone	1=No	12,871	9784 (76.02 %)	3087 (23.98 %)	Speed Limit	1 = Low (≤30 mph)	3335	%) 2972 (89.12	%) 363 (10.88
	2 = Yes	281	195 (69.40 %)	86 (30.60 %)		2 = Medium [30 mph, 50 mph]	7800	%) 6019 (77.17	%) 1781 (22.83
Road Curve	1=Straight	12,452	9527 (76.51 %)	2925 (23.49 %)		3=High (≥50 mph)	2017	%) 988 (48.98	%) 1029 (51.02
	2=Curve	700	452 (64.57 %)	248 (35.43 %)	Traffic control	1=No Control Present	7605	%) 5626 (73.98	%) 1979 (26.02
Road Grade	1 = Level	11,052	8394 (75.95 %)	2658 (24.05 %)		2=Signs	1162	%) 1012 (87.09	%) 150 (12.91
	2 = Grade	1572	1175 (74.75 %)	397 (25.25 %)		3=Signal	2604	%) 2233 (85.75	%) 371 (14.25
	3=Hillcrest	528	410 (77.65 %)	118 (22.35 %)		4 = Double Yellow Line, No Passing Zone	1781	%) 1108 (62.21	%) 673 (37.79
Road Class	1 = US roads	1213	690 (56.88 %)	523 (43.12 %)	Crash Location	1=Non-intersection	7587	%) 5290 (69.72	%) 2297 (30.28
	2 = Interstate roads	416	168 (40.38	248 (59.62		2 = Intersection	5565	%) 4689	%) 876

Table 3 (continued )

Variable	Description	Total	Pedestrian-Injury Severity		Variable	Description	Total	Pedestrian-Injury Severity
			Non serious injury	Serious injury				Non serious injury	Serious injury
		13152	9979 (75.87 %)	3173 (24.13 %)			13152	9979 (75.87 %)	3173 (24.13 %)
Pedestrian Position	1=Crosswalk Area	3140	2816 (89.68 %)	324 (10.32 %)	Driver Age	1 = Young (<25)	1980	1417 (71.57 %)	563 (28.43 %)
	2 = Travel Lane	8156	5703 (69.92	2453 (30.08		2 = Middle age [25,65)	9743	7454 (76.51	2289 (23.49
	3=Others	1856	%) 1460 (78.66 %)	%) 396 (21.34 %)		3=Old (≥65)	1429	%) 1108 (77.54 %)	%) 321 (22.46 %)
					Driver Gender	1 = Male	8887	6647 (74.79	2240 (25.21
Location characteristics								%)	%)
Region	1 = Mountains	1146	838 (73.12 %)	308 (26.88 %)		2 = Female	4265	3332 (78.12	933 (21.88
	2=Piedmont	8675	6849 (78.95	1826 (21.05	Driver Alcohol	1=No	12,753	%) 9777 (76.66	%) 2976 (23.34
	3=Coastal	3331	%) 2292 (68.81	%) 1039 (31.19		2 = Yes	399	%) 202 (50.63	%) 197 (49.37
Distance to nearest	1 = Low (≤1.8205 km)	8175	%) 6418 (78.51	%) 1757 (21.49	Driver Vehicle Type	1=Passenger Car	6364	%) 4905 (77.07	%) 1459 (22.93
Emergency Medical Services	2 = Medium (1.8205 km,4.4262 km)	4100	%) 2965 (72.32	%) 1135 (27.68		2 = Large vehicles (Sport Utility, Pickup, Van/	5175	%) 3818 (73.78	%) 1357 (26.22
	3=High (≥4.4262 km)	877	%) 596 (67.96	%) 281 (32.04		minivans and Panel trucks) 3=Others	1613	%) 1256 (77.87	%) 357 (22.13
Distance to nearest Police	1 = Low (≤8.0282 km)	8490	%) 6513 (76.71	%) 1977 (23.29				%)	%)
Station			%)	%)	Crash characteristics
	2 = Medium (8.0282 km,19.8185 km)	2974	2225 (74.82 %)	749 (25.18 %)	Ambulance Rescue	1=No	3544	3143 (88.69 %)	401 (11.31 %)
	3=High (≥19.8185 km)	1688	1241 (73.52 %)	447 (26.48 %)		2 = Yes	9608	6836 (71.15 %)	2772 (28.85 %)
	Built Environment unrelated Characteristics				Hit and Run	1=No	10,039	7484 (74.55	2555 (25.45
Pedestrian characteristics								%)	%)
Pedestrian Age	1 = Young (≤25)	3478	2879 (82.78 %)	599 (17.22 %)		2 = Yes	3113	2495 (80.15 %)	618 (19.85 %)
	2 = Middle age (25,65)	8559	6349 (74.18	2210 (25.82	Crash Group	1=Pedestrian Crossing	5424	4320 (79.65 %)	1104 (20.35 %)
	3=Old (≥65)	1115	%) 751 (67.35	%) 364 (32.65		2=Pedestrian in Roadway	2888	1976 (68.42 %)	912 (31.58 %)
Pedestrian Gender	1 = Male	8664	%) 6401 (73.88	%) 2263 (26.12		3 = Walking Along Roadway	2210	1622 (73.39 %)	588 (26.61 %)
	2 = Female	4488	%) 3578 (79.72	%) 910 (20.28		4=Others	2630	2061 (78.37 %)	569 (21.63 %)
Pedestrian Alcohol	1=No	11,581	%) 9035 (78.02	%) 2546 (21.98		Environment and temporal characteristics
	2 = Yes	1571	%) 944	%) 627	Peak Hour	1=No	9148	6765 (73.95	2383 (26.05
			(60.09 %)	(39.91 %)		2 = Yes (6:00–9:00&17:00–19:00)	4004	%) 3214 (80.27	%) 790 (19.73
Driver and vehicle characteristics								%) (continued on next page)	%)

Table 3 (continued )

Variable	Description	Total	Pedestrian-Injury Severity
			Non serious injury	Serious injury
		13152	9979 (75.87 %)	3173 (24.13 %)
Workday	1=No	3161	2234 (70.67 %)	927 (29.33 %)
	2 = Yes	9991	7745 (77.52 %)	2246 (22.48 %)
Season	1=Spring	2936	2244 (76.43 %)	692 (23.57 %)
	2=Summer	2893	2167 (74.90 %)	726 (25.10 %)
	3 = Fall	3931	3010 (76.57	921 (23.43
	4 = Winter	3392	%) 2558 (75.41	%) 834 (24.59
Light Condition	1 = Daylight	6094	%) 5236 (85.92 %)	%) 858 (14.08 %)
	2 = Dawn/Dusk Light	584	465 (79.62 %)	119 (20.38 %)
	3 = Dark – Lighted Roadway	3336	2517 (75.45 %)	819 (24.55 %)
	4 = Dark – Roadway Not Lighted	3138	1761 (56.12 %)	1377 (43.88 %)
Weather	1=Clear	10,326	7892 (76.43 %)	2434 (23.57 %)
	2=Cloudy	1513	1117 (73.83 %)	396 (26.17 %)
	3 = Adverse weather (Rain, Snow, Sleet, Hail, Freezing Rain/Drizzle, Fog, Smog, Smoke)	1313	970 (73.88 %)	343 (26.12 %)
Period	1 = Before the COVID-19 outbreak	6882	5425 (78.83 %)	1457 (21.17 %)
	2 = After the COVID-19 outbreak a	6270	4554 (72.63 %)	1716 (27.37 %)

Notes: ^a On March 3, 2020, North Carolina reported its first case of COVID-19. Crashes occurring after this date are considered as after the COVID-19 outbreak crashes.

where pdata represents the distribution of real data; pG denotes the distribution of generated data; ppenalty represents the distribution of random interpolation sampling between real samples and generated samples; E means the expectation; λ is the gradient penalty parameter, which is usually set to 10 (Gulrajani et al., 2017).

After a crash occurs, law enforcement units typically use some characteristics, such as humans, vehicles, and road infrastructure, to record it. However, these features are often discrete variables (i.e., gender, alcohol consumption, and intersection or not), and the distribution of certain values in each feature is unequal. If synthetic data fails to obey this distribution, the model is prone to overfitting, meaning it performs well on the training dataset but poorly on the test dataset (Ali et al., 2024). To deal with the imbalanced discrete columns, the CTGAN model has a sophisticated property that the ordinary GAN model cannot account for. Specifically, this solution consists of three key elements: conditional generator, generator loss, and training-by-sampling (Chen et al., 2024). The conditional vector is a one-hot encoded representation of all discrete variables. In this vector, one category within the chosen discrete variables is assigned a value of 1, while all other categories are set to 0. This can force the generator to generate the data with a distribution similar to the true data. In generator loss, it adds cross-entropy to the loss term, which can enforce the generated crash data to obey the selected condition. The training-by-sampling strategy can help the CTGAN model uniformly explore all possible values in discrete variables.

4.3. LightGBM model and Shapley Additive exPlanations (SHAP)

The LightGBM is initially developed to tackle challenges related to efficiency and scalability encountered by GBDT and XGBoost in scenarios involving high-dimensional input features and large datasets (Ke et al., 2017). During the training process, it adopts a leaf-wise tree growth strategy, where the leaf with the highest information gain is split first (Chang et al., 2022; Ziakopoulos, 2024). Two innovative techniques, Gradient-based One-Side Sampling (GOSS) and Exclusive Feature Bundling (EFB), enable LightGBM to have higher training speed and prediction accuracy than other gradient boosting methods, such as XGBoost and CatBoost (Wen et al., 2021). Specifically, the GOSS method divides the samples into two subsets including subset A and subset B. The subset A contains the larger absolute values of gradients (i.e., top a*100%) while subset B randomly selects the remaining samples (i.e., b*(1 − a)*100%). By retaining samples within subset A and sampling samples within subset B, the LightGBM model can focus on poorly learned data, thereby improving computational efficiency. In terms of the EFB method, it bundles some exclusive features together (i.e., these features cannot exist at the same time) into a feature bundle and thus reduces the dimension of the data and speeds up the training process without decreasing performance. The detailed process can be seen in Algorithm 1. In addition, it is also robust to address multicollinearity problems as a decision tree-based model (Kotsiantis, 2013). Because of these advantages, this method is widely used in the field of traffic safety modeling (Wang et al., 2023b; Xue et al., 2024). Therefore, this approach is selected to model pedestrian injury severities.

Algorithm 1 LightGBM

Input:

Training dataset: ^D ⁼ { (*x1, y1),*(*x2, y2),* ⋯, xn, yn ) } . % xi is each crash record; yi ∈ {0, 1} is the pedestrian injury severity and 0 denotes non-serious injury, 1 denotes serious injury.

Iterations: T is the number of decision trees

The outputs of decision tree model:ht(x)

Process:

Combine the mutually exclusive features with the Exclusive Feature Bundling technique

Initialize ̂f(x) ⁼ argminfEy,sL y, ∑^T ^t=1ht(x)

For t = 1*,2,⋯,T*:

Calculate the gradient absolute values:

$$g_i = \left| \frac{\partial L(y, \widehat{f}(x_i))}{\partial \widehat{f}(xi)} \right|{\widehat{f}(x) = \widehat{f}_{i-1}(x)}, i = {1, \dots, n}$$

Resample of the dataset by Gradient-based One-Side Sampling:

topN = a × len(D); randN = b × len(D);

A = sorted[1 : topN]; B = RandomPick(sorted[topN : len(D)],randN );

Calculate variance gain:

$$V{J}(d) = \frac{1}{n} \left[ \left( \sum{x_{i_1,i_2}} gi + \frac{1-a}{b} \sum{x_{i_1,i_2}} gi \right)^2 + \left( \sum{x_{i_2,i_2}} gi + \frac{1-a}{b} \sum{x_{i_2,i_2}} g_i \right)^2 \right] \text{where}$$

^A¹ ⁼ { xi ∈ A : xij ≤ d , Ar ⁼ { xi ∈ A : xij > d , Bi ⁼ { xi ∈ B : xij ≤ d , Br ⁼ { xi ∈ B : xij > d }

Train a new decision tree ht(x) based on dataset D = A ∪ B

Update ̂ft(x) ⁼ ̂ft⁻ ¹(x) ⁺ ht(x)

(continued on next page)

(continued )

Algorithm 1 LightGBM

End

Output:̂ft(x)

SHAP is a popular post-hoc interpretation method especially for treebased ensemble models, which applies the game theory to explain the output of machine learning models in a consistent and accurate way (Lundberg and Lee, 2017). For a feature subset S ⊆ F (F stands all features in the dataset), calculating the contribution ϕⁱ of feature i needs to train two models. One model fS∪{i} xS∪{i} ) is trained with feature i. Another model fS(xs) is trained without feature i. The difference in outputs fS∪{i} ( xS∪{i} ) − fS(xs) between these two models is calculated for each subset S ⊆ F{i}. The Shapley value of feature i is calculated as follows:

$$\phii = \sum{S \subseteq F{i}} \frac{|S|!(|F| - |S| - 1)!}{|F|!} \left( f{S \cup {i}} \left( x{S \cup {i}} \right) - f_S(x_s) \right)$$ (2)

To reduce the computation cost in high-dimensional data, Lundberg et al. (2020) designed the TreeExplainer for tree-based ensemble models, which significantly increases the calculation efficiency of SHAP values. In addition, the interaction effects of two features can also be calculated by Eq (3), which reflects the difference between the SHAP values of factor i with and without factor j.

$$\phi{i,j} = \sum{S \subseteq F(i,j)} \frac{|S|!(|F| - |S| - 2)!}{|F|!} \left( f{S \cup {i,j}} \left( x{S \cup {i,j}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) + f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}} \right) - f{S \cup {i}} \left( x{S \cup {i}}$$

4.4. Improved wrapper-based feature selection methods

To overcome the adverse effects of high-dimensional crash data, the wrapper-based feature selection methods are improved by optimizing the hyperparameters of the LightGBM model and using a more accurate feature importance estimation method. In traditional wrapper-based feature selection methods, the assessment model (i.e., LightGBM and XGBoost) typically uses the default hyperparameters (Shi et al., 2019) or is initialized using all features (Chen et al., 2022), ignoring the coupling effect of the two issues. In this research, feature selection is implemented in each iteration of the Bayesian hyperparameter tuning until a predefined iteration number is reached, as shown in Algorithm 2. During the feature importance calculation process, traditional methods, such as average gain or split weight, tend to overestimate the importance of high-cardinality categorical variables (Huang et al., 2024). Thus, the SHAP value is used in current research, which can produce more consistent results (Lundberg and Lee, 2017). In addition, after each update of the feature subset, the model needs to be retrained to evaluate its performance with the current features. To improve the reliability of the evaluation, 5-fold cross-validation is applied which has been commonly recommended by many researchers (Chang et al., 2022; Chen et al., 2022; Khan et al., 2024). This method divides the data into five subsets, trains the model five times using four folds for training and one for validation, and then averages the results to assess overall performance.

Since the wrapper-based feature selection methods have different strategies, the feature subset may be different. Thus, three popular methods are compared to ensure that the optimal features subset can be obtained. Recursive Feature Elimination (RFE) algorithm, proposed by Guyon et al. (2002), is a backward search approach by recursively eliminating less important features. Recursive Feature Addition (RFA) algorithm, proposed by Hamed et al. (2014), is a forward search approach by sequentially adding the most important features. The Boruta algorithm, designed by Kursa and Rudnicki (2010), identifies important features by comparing their importance scores to those of randomized shadow features. Readers are encouraged to refer (Chai et al., 2023; Khan et al., 2024; Xue et al., 2024) to get details about these approaches.

Algorithm 2: Improved wrapper-based feature selection

Input:

Classification model: LightGBM.

Balanced training dataset: ^D ⁼ { (*x1, y1),*(*x2, y2),* ⋯, xn, yn ) } with feature set {S}.

Bayesian optimization iterations: N. The number of iterations R.

Process:

Initialization probabilistic model P

For n = 1*,2,⋯,N*:

Select a promising hyperparameter set Hn that yields the best performance on P function

Set the hyperparameter of LightGBM with Hn

Feature selection methods

[Optional] Recursive feature elimination-SHAP algorithm

Process:

While length{S} > 0:

Train the LightGBM model using D(S).and calculate performance F(S)

Calculate the SHAP values for each variable as feature importance φ^S Remove the least important feature s and renew {S}∕∈^s

End

Select the feature subset {S*}with the{F}max

[Optional] Recursive feature addition-SHAP algorithm

Process:

Train the LightGBM model using D(S).and calculate SHAP importance φ^S

While {S} ∕= ∅:

Select the most important feature s

Calculate performance F(S) of LightGBM model using D(S).

If F(S) > Fbest # initialize Fbest = − ∞

Retain s and renew Fbest = F(S), {S} = {S}∕∈^s

else: Remove s from {S}

End

Select the feature subset {S*} with the{F}max

[Optional] Boruta-SHAP algorithm

Process:

For r = 1, 2, ⋯ , R:

Create shadow feature set {SF} and shuffle randomly.

Train the LightGBM model using D(S,SF).and calculate SHAP importance φS,SF

Retain the real feature {S} with higher importance φ^s than the shadow features φ(SF)max

End

Build a binomial distribution *X∽(0.5, R*) and select significant important feature

subset {S*}

Output: Optimal {S*} Obtain the real performance F(Hn ) of LightGBM based on the optimal feature subset

{S*}

Update P function with F(Hn )

End Select the hyperparameter subset {H*}with the{F}max

Output: Optimal {H*}, {S*}

5. Results and discussion

5.1. Enhanced model classification performance analysis

5.1.1. Data balancing methods comparison

Based on numerous experiments, the detailed architecture of the CTGAN model is provided in Table 4 and the parameters in CTGAN are set as follows: The generator is designed with two hidden layers, each containing 256 neurons, to effectively capture intricate feature representations. This follows the default network architecture of the algorithm, which has been proven by other research to achieve promising sample generation results (Xu et al., 2019; He et al., 2024). The discriminator has the same structure and settings. The learning rate of the generator and discriminator is 2e-4 and 4e-4, respectively. Adam optimizer is used to train the generator and critic, and its weight decay is 1e-6. The number of iterations is set to 500 and the batch size is 500. The size of the random sample is 128. Besides, updating the discriminator multiple times per generator update helps stabilize GAN training but leads to slower convergence. In our dataset, the generator can learn

Fig. 1. Conceptual framework.

Fig. 2. Structure of the CTGAN model.

effectively without requiring additional discriminator updates. Therefore, we set the discriminator update frequency to 1 to balance training stability and efficiency.

Fig. 3. shows the training process of CTGAN model. The figure illustrates the loss values of the discriminator and generator over 500 training iterations. Initially, during the early stages of training, the loss values for both the discriminator and generator exhibit substantial fluctuations, reflecting the preliminary phase of model adjustment with a relatively monotonous convergence trend. As training proceeds, fluctuations in the loss values of the discriminator and generator are noticeable, indicating that the two components of the CTGAN model are beginning to compete, and there is no such situation where one part is too strong or too weak. Through this iterative fluctuation process, the discriminator and generator continually refine their performance, ultimately leading to the generation of high-quality samples.

To fairly verify the effectiveness of different methods in eliminating the effect of imbalanced data, we compare the performance of LightGBM, XGBoost, CatBoost, and Random Forest with the default hyperparameter settings on crash data balanced by the CTGAN method and two traditional sampling strategies (SMOTE and ADASYN), and on

Table 4 Architecture of the CTGAN model used in this research.

Generator architecture			Discriminator architecture
Layer	Layer type	Layer description	Layer	Layer type	Layer description
1	Fully connected + BN + ReLU	Neuron = 256, Output size = 256 × 1	1	Fully connected + LeakyReLU (0.2) + Dropout	Neuron = 256, Output size = 256 × 1
2	Fully connected + BN + ReLU	Neuron = 256, Output size = 256 × 1	2	Fully connected + LeakyReLU (0.2) + Dropout	Neuron = 256, Output size = 256 × 1

Notes: BN indicates Batch Normalization; ReLU indicates Rectified Linear Unit.

the original data, respectively. Metrics such as the area under the curve (AUC) value, F1-score, and G-means have been widely used for evaluating model performance (Chen et al., 2024; Xue et al., 2024). Besides, our goal is to prioritize reducing the number of fatal and incapacitating crashes as these crashes carry a substantially higher social cost compared to others. Thus, sensitivity, which represents the proportion of correctly identified serious injury crashes, is also frequently emphasized (Li et al., 2023; Chen et al., 2024). As shown in Table 5, CTGAN combined with LightGBM achieves the highest value in sensitivity, F1 score, G-means, and AUC metrics, outperforming CTGAN combined with XGBoost, CatBoost, and Random Forest. Besides, four machine learning models show the lowest sensitivity value on the original dataset, indicating that the model primarily emphasizes the majority class samples (non-serious injury crashes) while neglecting the minority class samples (serious injury crashes). Although two traditional sampling methods boost the four machine learning models in some metrics, the CTGAN method has more significant improvements on the model. Considering the extensive iterative training process involved in the feature selection stage, the LightGBM model is chosen to validate the effectiveness of the improved feature selection method due to its higher accuracy, faster training speed, and lower memory usage.

5.1.2. Improved wrapper-based feature selection methods comparison

Table 6 shows the performance metrics of the LightGBM model optimized by three feature selection methods. LightGBM enhanced by CTGAN + BO outperforms the model only enhanced by CTGAN in sensitivity, F1-score, G-means, and AUC metrics. Therefore, setting appropriate hyperparameters is the key to leveraging the full potential of a machine learning model in providing the highest predictive power. Among these feature selection methods, CTGAN + BO-RFE-SHAP outperforms all other optimization methods across four key metrics such as sensitivity, F1-score, G-means, and AUC, indicating the effectiveness of the proposed enhancement framework. This means that there are some redundant factors in the high-dimensional crash data, which not only fail to provide valid information but also disturb the LightGBM model to distinguish non-serious injury crashes and serious injury crashes. Therefore, removing these redundant features will further improve the

Fig. 3. Convergence trend of the CTGAN model in the training dataset.

Table 5 Model performance with different data balancing methods.

Model	Dataset	Precision	Specificity	Sensitivity	F1-score	G-means	AUC
LightGBM	Original dataset	0.6085	0.9265	0.3592	0.4518	0.5769	0.7792
	SMOTE	0.4602	0.8257	0.4674	0.4638	0.6212	0.7306
	ADASYN	0.4415	0.8086	0.4758	0.4580	0.6203	0.7221
	CTGAN	0.4853	0.8190	0.5368	0.5097	0.6631	0.7678
XGBoost	Original dataset	0.5541	0.9078	0.3603	0.4367	0.5719	0.7504
	SMOTE	0.4464	0.8293	0.4328	0.4395	0.5991	0.7132
	ADASYN	0.4319	0.8203	0.4296	0.4308	0.5936	0.7027
	CTGAN	0.4802	0.8206	0.5210	0.4997	0.6538	0.7492
CatBoost	Original dataset	0.6376	0.9362	0.3529	0.4544	0.5748	0.7767
	SMOTE	0.4511	0.8313	0.4359	0.4434	0.6020	0.7252
	ADASYN	0.4493	0.8203	0.4611	0.4552	0.6150	0.7194
	CTGAN	0.5081	0.8480	0.4937	0.5008	0.6470	0.7608
Random Forest	Original dataset	0.6747	0.9593	0.2658	0.3813	0.7758	0.5049
	SMOTE	0.4887	0.8641	0.4086	0.4451	0.5941	0.7374
	ADASYN	0.4850	0.8564	0.4254	0.4533	0.6035	0.7351
	CTGAN	0.4632	0.7982	0.5246	0.4920	0.6471	0.7523

Table 6 Model performance with different feature selection methods.

Feature selection method	Precision	Sensitivity	Specificity	F1-score	G-means	AUC
CTGAN + BO	0.4952	0.5462	0.8230	0.5195	0.6705	0.7726
CTGAN + BO-Boruta-SHAP	0.4868	0.5809	0.8053	0.5297	0.6839	0.7774
CTGAN + BO-RFA-SHAP	0.4833	0.6092	0.7929	0.5390	0.6950	0.7800
CTGAN + BO-RFE-SHAP	0.4838	0.6113	0.7926	0.5401	0.6961	0.7804

classification ability of the model.

Fig. 4 shows the feature subset at the best iterations for the three feature selection methods. The black mask represents the features selected by the method. Overall, improved RFA, RFE, and Boruta removed 5, 10, and 7 features, respectively. It is worth noting that the features removed by improved RFE have a high degree of overlap with the other feature selection methods. For example, the redundant features Workzone, RdGrad, RdSurface, and RdConditio removed by improved RFA and PedSex, DrvrVehTyp RdGrad, RdSurface, and PeakHour removed by improved Boruta are also identified by the improved RFE method. This suggests that improved RFE can identify redundant features more comprehensively.

5.2. Fixed and heterogeneous factors analysis

As shown in Fig. 5(a), we select the top 19 features as important features that can explain 80 % of the model's output, and 13 features of which (as shown in gray histogram) are built environment factors, which reflect that the built environment plays an important role in affecting pedestrian injury severities. As shown in Fig. 5(b), some factors show a clear relationship with pedestrian injury severities. For example, crashes with ambulance rescue (red dots) always show a positive effect.

Fig. 4. Feature subset of different feature selection methods.

Fig. 5. Feature importance and effects.

However, the effects of most factors are mixed. For example, the speed limit between 30 and 50 mph (purple dots) shows a positive effect in some observations and a negative effect on others. This phenomenon means that this factor is heterogeneous (Sun et al., 2023; Zhou et al., 2024). These fixed effect factors (i.e., dark with or without roadway light, the number of lanes over 2, and non-intersection) can be directly used to make countermeasures. For example, crashes with ambulance rescue are associated with serious injuries. The reason may be that the

Fig. 6. The effects of heterogeneous factors.

high cost of medical services may discourage the parties involved from calling for ambulance assistance unless the crash severity is serious (Li et al., 2021). Therefore, the government should provide partial or full subsidies for emergency medical services in crashes, especially for lowincome individuals.

Fig. 6 shows the effects of heterogeneity factors quantified by the SHAP values. Therefore, a positive SHAP value indicates that the factor contributes to the likelihood of a serious crash in the observed sample. To clearly illustrate the effects distribution of heterogeneous factors, the proportion of samples in which a specific factor exhibits a positive contribution to serious injuries is counted below each subfigure in Fig. 6. Interestingly, 11 of these 12 heterogeneity factors are built environment factors. For example, only 67.27 % of pedestrians are more likely to suffer serious injury on NC routes. Li et al. (2021) also found that this factor is a random parameter. The reason may be that NC routes usually connect local routes and interstate routes. Pedestrians at different road junctions are exposed to impacts from vehicles at different speeds. Besides, only 64.92 % of pedestrians are more likely to suffer a serious injury on medium speed limit roads. Previous studies have also found that the impact of medium speed limits (i.e., between 30mph and 50mph) is unclear. For example, Das et al. (2024) and Wang et al. (2021) found that the speed limit exceeding 45mph may also have a positive effect on fatal injury. Similarly, 26.62 % of pedestrians are susceptible to serious injuries in commercial areas. This may be related to the complex pedestrian behavior in this area. For example, dashing-out behavior can increase the probability of serious injury, but crossing roadway with vehicle turning is correlated with non-serious injury in commercial areas (Yang et al., 2023). These findings may confirm why much of the literature on built environment factors presents mixed or contradictory results and these heterogeneous effects pose a challenge in developing accurate countermeasures (Merlin et al., 2020). Thus, the contextdependent effects quantified by the SHAP interaction values are used to reveal the potential source of those heterogeneous factors and to identify the specific conditions under which these heterogeneous factors exert positive effects, thereby supporting the development of effective policies.

5.3. Context-dependent effects analysis

5.3.1. Between heterogeneous built environment factors

Fig. 7(a), (b), and (c) show the effects of restaurants under different youth population rates, road class, and rural or urban areas,

Fig. 7. Context-dependent effects between heterogeneous built environment factors.

Fig. 7. (continued).

respectively. Briefly, there is a positive effect of restaurants present under the context of medium and high youth population rate counties, local roads, and urban areas. Restaurants can attract residents and visitors and thus can lead to large traffic volumes (Su et al., 2021). Besides, counties with medium and high youth populations are more active in terms of economy and entertainment. Young people frequently visit restaurants, but their impulsive nature makes them more prone to serious injuries in complex traffic environments (Luo et al., 2023). For road class, local roads usually have poor visibility and road design, which can easily cause serious injuries near restaurants with high pedestrian density (Sun et al., 2023). Meanwhile, restaurants are one of the preferred locations for alcohol consumption, which leads to more drinking drivers and pedestrians at nearby locations (Li et al., 2013). Residents in urban areas have higher incomes and thus are more likely to frequent restaurants, making them more likely to be involved in serious crashes. Therefore, in counties with a higher proportion of young populations, it may be appropriate to reallocate funding away from entertainment purposes and instead invest more in transportation programs. Besides, warning signs could be installed in popular dining areas to remind drivers to slow down (Zhai et al., 2019) and engaging safety education campaigns (e.g., short videos and street interviews) targeting young people could be implemented to enhance their awareness and

participation.

Fig. 7(d), (e), and (f) show the effects of road class under different unemployment, poverty, and elderly population rates. The unemployment rate and poverty rate show similar moderate effects in that local roads have more serious crashes in counties with low unemployment and poverty rates and higher-grade roads (i.e., NC, interstate and US routes) have more serious crashes in counties with medium unemployment and poverty rates. This may be because counties with better economic development typically have higher population densities and drivers are more likely to speed in these areas with better designed roads, resulting in higher fatalities (Wang et al., 2023a). Besides, local roads in counties with high poverty rates also tend to cause serious injuries. The reason may be that many low-income people in poverty counties cannot own a car due to the financial burden, thus they need to walk longer distances to reach their destinations, increasing the risk of serious injuries (Bernhardt and Kockelman, 2021). Moreover, due to the absence of dedicated pedestrian facilities, pedestrians are frequently forced to share road space with motor vehicles (Zhu et al., 2024a). To mitigate these risks, transportation authorities should implement lowcost, high-impact infrastructure enhancements in such areas. For instance, efforts should focus on refurbishment of clearly visible road markings and signage to enhance driver and pedestrian awareness (Khan and Habib, 2022). For the influence of demographics, counties with low elderly population rates are more likely to have serious crashes on the US routes and Interstate routes. The reason may be that these counties have more middle-aged or young people who are more confident crossing the wider and higher-graded roads, thus they are more likely to suffer serious injuries (Hossain et al., 2024).

In Fig. 7(g), high speed limit roads through residential and commercial areas can easily cause serious injuries. This result is reasonable because these areas have high population densities, making it difficult for high-speed vehicles to avoid suddenly appearing pedestrians (Song et al., 2020; Yang et al., 2023; Kutela et al., 2025). Therefore, urban planning and transportation authorities should integrate land use and traffic safety policies, ensuring that future roadway expansions or modifications in economically active and residential districts are guided by a "Vision Zero" (Law and Petric, 2024; Madarshahian et al., 2025). Besides, in Fig. 7(h), travel lanes near the low speed limit roads, and crosswalk areas on medium speed limit roads also have a higher probability of serious crashes. Similar to the findings in previous studies (Li et al., 2021; Campos Ferreira et al., 2022), severe injuries to pedestrians usually occur in facilities designed for them, such as crosswalks, pedestrians and cycle lanes. This is because the infrastructure provided for pedestrian travel is poorly designed, such as the lack of visible signs or markings at crosswalks and the lack of segregation of pedestrian paths from motor vehicle lanes.

5.3.2. Between heterogeneous built environment factors and non-built environment factors

In Fig. 8(a), counties with higher poverty rates are more likely to experience serious crashes in dark conditions without roadway light. A reasonable explanation is that walking infrastructures (e.g., traffic lights and traffic calming devices) are infrequently maintained in these counties, resulting in pedestrians engaging in unsafe behaviors, such as crossing the road in unlighted conditions (Zhai et al., 2024). One practical solution is the deployment of solar-powered LED streetlights or portable solar lamps, which offer energy-efficient and cost-effective lighting for areas lacking reliable electrical infrastructure. Fig. 8(b) shows that counties with higher unemployment rates are more likely to have serious crashes during the workday. The reason is higher unemployment rates are often associated with more risky driving behaviors, such as drug/alcohol-impaired driving and aggressive driving, and workday has a higher travel demand, increasing the likelihood of serious crashes (Dong et al., 2022; Yan et al., 2022). Fig. 8(c) and (d) show the impact of COVID-19. After the COVID-19 outbreak, the presence of restaurants also has a positive effect on serious injuries. This is because people drink more alcohol during the pandemic to relieve stress, increasing the likelihood of serious crashes near restaurants (Watson-Brown et al., 2024). Thus, customers could be reminded to use ridehailing services or walking escort services after drinking. Besides, the local roads have a positive effect on serious injuries after the COVID-19 outbreak. This finding is consistent with previous studies (Marshall et al., 2023; Stiles et al., 2023). It can be attributed to the fact that reduced road traffic flow during the COVID-19 pandemic makes people tend to walk for shorter distances through local roads, but increased vehicle speed increases their risk of exposure to serious injuries (Stiles et al., 2023). Besides, young and middle-aged pedestrians are more likely to suffer serious injuries in urban and rural areas, respectively, as shown in Fig. 8(e). The reason is that urban areas have more vehicles and interactions, making it difficult for young pedestrians to make the right decisions in such complex traffic conditions (Wang et al., 2023a). Middle-aged people are the main workers, and they have to travel frequently. However, inadequate pedestrian facilities in rural areas make them vulnerable to serious injury (Sun et al., 2023). As shown in Fig. 8(f), no traffic control roads under daylight conditions, and signs or signal control roads under the dark with and without roadway light conditions have a positive effect on serious injuries. The reason is that conflicts between pedestrians and vehicles are more pronounced in locations without traffic control and drivers tend to fail to yield to pedestrians (Sasidharan et al., 2015). Besides, signs and signal control areas are usually set up in areas with more conflict, but the protection they provide to pedestrians is still limited in poorly lighted conditions (Hossain et al., 2023).

6. Conclusion

This research develops and uses the enhanced machine learning model to investigate the context-dependent effects of built environment factors on pedestrian injury severities, aiming to identify the potential sources of heterogeneity and conditions for the positive impact of heterogeneous factors. However, the low percentage of serious injury crashes (i.e., incapacitating injury and fatal injury) and multidimensional built environment factors in North Carolina crash data (from 2017 to 2022) lead to imbalanced and high-dimensional issues. Thus, an enhanced machine learning method is designed in this paper. Results indicate that the conditional tabular generative adversarial network and improved RFE feature selection can significantly enhance the model's fitting ability. Besides, the results also confirmed the critical role of built environment factors in pedestrian injury severities. Specifically, 13 out of the top 19 important factors belong to built environment factors and 11 of them are heterogeneous.

For factors that demonstrate positive and consistent effects on serious injuries, we can readily make recommendations and/or develop countermeasures. For instance, dark with or without roadway light can increase the probability of serious injuries, thus government should encourage pedestrians to wear reflective material cloth when walking in dark environments (Chen and Shen, 2016); the number of lanes over 2 has positive effects on serious injuries, thus refuge islands should be provided in wide roadways to assist pedestrians crossing (Agheli and Aghabayk, 2024); non-intersection areas have a higher probability of serious crashes, thus warning signs or crossing facilities should be placed in long segment roads where pedestrians may suddenly appear (Khan and Habib, 2022).

In terms of heterogeneous built environment factors, it is not reasonable to design a uniform strategy for all observations. The context-dependent effects can uncover the potential sources of heterogeneity and identify the positive effects under different conditions. The main results are summarized as follows: (1) For human amenity infrastructure, restaurants have a positive effect on pedestrian injury severities when they are located near local roads or in urban areas. Therefore, appropriate traffic enforcement should be conducted to reduce alcoholinfluenced pedestrians and drivers (Li et al., 2021). Besides, physical separation such as barriers, curbs, and bollards should be installed in these heavy pedestrian and vehicular traffic volumes areas to prevent conflicts between them (Campos Ferreira et al., 2022). (2) For road infrastructure, higher-grade roads (i.e., US, Interstate, and NC routes) have a positive effect on injury severity in better economically developed counties. Therefore, the installation of speed cameras or increasing the frequency of police patrols in high pedestrian density areas can effectively reduce injury severities caused by speeding (Zhou et al., 2024). Besides, crosswalk areas on medium speed limit roads, and signs or signal traffic control roads under dark with or without roadway light conditions also put pedestrians at risk of serious injury. Therefore, more effective pedestrian crossing facilities should be installed, such as pedestrian bridges and tunnels. Meanwhile, the signs should be improved by incorporating flashing LED lights, and the traffic signals should be brightened, which can improve their effectiveness under dim conditions (Ashraf et al., 2023).

7. Limitations

This study still has some limitations. In terms of data, this research has constructed a relatively comprehensive set of influencing factors using multi-source data. However, there are still some valuable data

Fig. 8. Context-dependent effects between heterogeneous built environment factors and non-built environment factors.

sources (i.e., image data, sensor data, audio data, and meteorological data) worth collecting and extracting effective factors to gain a more comprehensive understanding of the impact of built environment factors. For instance, some street-level factors, such as vegetation, terrain, and buildings, can be extracted from image data obtained through traffic surveillance videos, vehicle-mounted camera images, and dronecaptured aerial images. Besides, the police-reported crash data extracted from NCDOT may be subject to reporting bias. To address this issue, integrating multi-source crash data, such as hospital records, crowdsourced data, and insurance claims data, can help improve data completeness and reliability.

From a methodological standpoint, hybrid approaches hold significant promise for improving predictive accuracy and effectively capturing unobserved heterogeneity in crash severity analyses. However, applying advanced heterogeneity models and incorporating numerous interaction terms into large-scale crash datasets remains challenging due to computational complexity. Therefore, the development of hybrid approaches that can effectively investigate potential sources of heterogeneity in large-scale datasets represents a valuable direction for future research. Moreover, the effects of various factors on injury severity may vary over time. Thus, temporal instability may affect the performance of machine learning models and the effectiveness of corresponding countermeasures. To account for temporal instabilities, this study constructed several time-related variables (i.e., period, season, and weekends) and investigated how the effects of key factors change under different temporal conditions. However, relying on individual temporal indicators may be insufficient to fully elucidate temporal variation, because machine learning models are non-parametric in nature and do not support statistical testing of parameter differences across time periods. A more rigorous method would involve developing separate models for different time periods and applying likelihood ratio tests to examine temporal instability. In the future, we plan to adopt a partially constrained temporal modeling approach, as proposed by Alnawmasi and Mannering (2023), to further investigate how the impact of built environment factors evolves over time.

CRediT authorship contribution statement

Zehao Wang: Writing – original draft, Methodology, Investigation, Formal analysis. Wei Fan: Methodology.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

The authors want to express their deepest gratitude to the financial support by the United States Department of Transportation, University Transportation Center through the Center for Advanced Multimodal Mobility Solutions and Education at The University of North Carolina at Charlotte (Grant Number: 69A3551747133).

Data availability

The authors do not have permission to share data.

References

Abrari Vajari, M., Aghabayk, K., Sadeghian, M., Shiwakoti, N., 2020. A multinomial logit model of motorcycle crash severity at australian intersections. J. Saf. Res. 73, 17–24. Adanu, E.K., Dzinyela, R., Agyemang, W., 2023. A comprehensive study of child pedestrian crash outcomes in ghana. Accid. Anal. Prev. 189, 107146.

Agheli, A., Aghabayk, K., 2024. How does distraction affect cyclists' severe crashes? A hybrid catboost-shap and random parameters binary logit approach. Accid. Anal. Prev. 211, 107896.
Ali, Y., Hussain, F., Haque, M.M., 2024. Advances, challenges, and future research needs in machine learning-based crash prediction models: A systematic review. Accid. Anal. Prev. 194, 107378.
Alnawmasi, N., Mannering, F., 2023. An analysis of day and night bicyclist injury severities in vehicle/bicycle crashes: A comparison of unconstrained and partially constrained temporal modeling approaches. Analytic Methods in Accident Research
Ashraf, M.T., Dey, K., Mishra, S., 2023. Identification of high-risk roadway segments for wrong-way driving crash using rare event modeling and data augmentation techniques. Accid. Anal. Prev. 181, 106933.
Azimi, G., Rahimi, A., Asgari, H., Jin, X., 2020. Severity analysis for large truck rollover crashes using a random parameter ordered logit model. Accid. Anal. Prev. 135,
Azmeri Khan, S., Yasmin, S., Mazharul Haque, M., 2023. Effects of design consistency measures and roadside hazard types on run-off-road crash severity: Application of random parameters hierarchical ordered probit model. Analytic Methods in Accident Research 40.
Baek, J., Lim, L., 2024. Unveiling pedestrian injury risk factors through integration of urban contexts using multimodal deep learning. Sustainable Cities and Society 101.
Barbour, N., Abdel-Aty, M., Yang, S., Mannering, F., 2024. Pedestrian injury severities resulting from vehicle/pedestrian intersection crashes: An assessment of covidcontributing temporal shifts. Anal. Methods Accid. Res. 43.
Bernhardt, M., Kockelman, K., 2021. An analysis of pedestrian crash trends and contributing factors in texas. J. Transp. Health 22.
Campos Ferreira, M., Dias Costa, P., Abrantes, D., Hora, J., Felício, S., Coimbra, M., Galvao˜ Dias, T., 2022. Identifying the determinants and understanding their effect on the perception of safety, security, and comfort by pedestrians and cyclists: A systematic review. Transport. Res. F: Traffic Psychol. Behav. 91, 136–163.
Cao, F., Ge, Y., Wang, J.-F., 2013. Optimal discretization for geographical detectorsbased risk assessment. Giscience & Remote Sensing 50 (1), 78–92.
Chai, C., Li, J., Islam, M.M., Feng, R., Lu, M., 2023. Hierarchical automated machine learning approach for self-optimizable driving distraction recognition based on driver's lane-keeping performance. Transportation Research Record: Journal of the Transportation Research Board 2678 (6), 301–317.
Chang, I., Park, H., Hong, E., Lee, J., Kwon, N., 2022. Predicting effects of built environment on fatal pedestrian accidents at location-specific level: Application of xgboost and shap. Accid. Anal. Prev. 166, 106545.
Chen, J., Pu, Z., Zheng, N., Wen, X., Ding, H., Guo, X., 2024. A novel generative adversarial network for improving crash severity modeling with imbalanced data. Transp. Res. Part C Emerging Technol. 164.
Chen, J., Liu, P., Wang, S., Zheng, N., Guo, X., 2025. Prediction and interpretation of crash severity using machine learning based on imbalanced traffic crash data. J. Saf. Res. 93, 185–199.
Chen, P., Shen, Q., 2016. Built environment effects on cyclist injury severity in automobile-involved bicycle crashes. Accid. Anal. Prev. 86, 239–246.
Chen, T., Wong, Y.D., Shi, X., Wang, X., 2022. Optimized structure learning of bayesian network for investigating causation of vehicles' on-road crashes. Reliability Engineering & System Safety 224.
Chen, T., Sze, N.N., Chen, S., Labi, S., Zeng, Q., 2021. Analysing the main and interaction effects of commercial vehicle mix and roadway attributes on crash rates using a Bayesian random-parameter Tobit model. Accid. Anal. Prev. 154, 106089.
Clifton, K.J., Burnier, C.V., Akar, G., 2009. Severity of injury resulting from pedestrian–vehicle crashes: What can we learn from examining the built environment? Transp. Res. Part D: Transp. Environ. 14 (6), 425–436.
Das, S., Jafari, M., Hossain, A., Chakraborty, R., Mimi, M.S., 2024. Toll road crash severity using mixed logit model incorporating heterogeneous mean structures. Transport Science, Transportmetrica A, pp. 1–21.
Ding, C., Chen, P., Jiao, J., 2018. Non-linear effects of the built environment on automobile-involved pedestrian crash frequency: A machine learning approach. Accid. Anal. Prev. 112, 116–126.
Dong, N., Zhang, J., Liu, X., Xu, P., Wu, Y., Wu, H., 2022. Association of human mobility with road crashes for pandemic-ready safer mobility: A new york city case study.
Accid. Anal. Prev. 165, 106478. Dumbaugh, E., Stiles, J., Mitsova, D., Saha, D., 2023. The most vulnerable user: Examining the role of income, race, and the built environment on pedestrian injuries and deaths. Transportation Research Record: Journal of the Transportation Research Board 2678 (2), 743–752.
Ewing, R., Cervero, R., 2010. Travel and the built environment. J. Am. Plann. Assoc. 76 (3), 265–294.
Fu, X., Nie, Q., Li, X., Liu, J., Nambisan, S., Jones, S., 2022. The role of the built environment in emergency medical services delays in responding to traffic crashes. Journal of Transportation Engineering, Part a: Systems 148 (10).
Jin, J.L., Huang, H.L., Yuan, C., Li, Y., Zou, G.Q., Xue, H.L., 2023. Real-time crash risk prediction in freeway tunnels considering features interaction and unobserved heterogeneity: A two-stage deep learning modeling framework. Analytic Methods in Accident Research 40.
Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., Courville, A., Year. Improved training of wasserstein gans. In: Proceedings of the 31st Annual Conference on Neural Information Processing Systems (NIPS), Long Beach, CA.
Guyon, I., Weston, J., Barnhill, S., Vapnik, V., 2002. Gene selection for cancer classification using support vector machines. Mach. Learn. 46 (1), 389–422.
Hamed, T., Dara, R., Kremer, S.C., Year. An accurate, fast embedded feature selection for svms. In: Proceedings of the 13th International Conference on Machine Learning and Applications (ICMLA), Detroit, MI, pp. 135-140.
Han, C., Huang, H., Lee, J., Wang, J., 2018. Investigating varying effect of road-level factors on crash frequency across regions: A bayesian hierarchical random parameter modeling approach. Anal. Methods Accid. Res 20, 81–91.
He, B., Jahed Armaghani, D., Hin Lai, S., Samui, P., Tonnizam Mohamad, E., 2024. Applying data augmentation technique on blast-induced overbreak prediction. Resolving the Problem of Data Shortage and Data Imbalance. Expert Systems with Applications 237.
Hossain, A., Sun, X., Das, S., Jafari, M., Rahman, A., 2024. Investigating pedestrianvehicle crashes on interstate highways: Applying random parameter binary logit model with heterogeneity in means. Accid. Anal. Prev. 199, 107503.
Hossain, A., Sun, X., Shahrier, M., Islam, S., Alam, S., 2023. Exploring nighttime pedestrian crash patterns at intersection and segments: Findings from the machine learning algorithm. J. Saf. Res. 87, 382–394.
Huang, J., Peng, Y., Hu, L., 2024. A multilayer stacking method base on rfe-shap feature selection strategy for recognition of driver's mental load and emotional state. Expert Systems with Applications 238.
Ke, G.L., Meng, Q., Finley, T., Wang, T.F., Chen, W., Ma, W.D., Ye, Q.W., Liu, T.Y., Year. Lightgbm: A highly efficient gradient boosting decision tree. In: Proceedings of the 31st Annual Conference on Neural Information Processing Systems (NIPS), Long Beach, CA.
Johnson, J.M., Khoshgoftaar, T.M., 2019. Survey on deep learning with class imbalance. J. Big Data 6 (1).
Khan, M.N., Das, S., Liu, J., 2024. Predicting pedestrian-involved crash severity using inception-v3 deep learning model. Accid. Anal. Prev. 197, 107457.
Khan, N.A., Habib, M.A., 2022. Exploring the impacts of built environment on pedestrian injury severity involving distracted driving. J. Saf. Res. 80, 97–108.
Kotsiantis, S.B., 2013. Decision trees: A recent overview. Artif. Intell. Rev. 39 (4), 261–283.
Kursa, M.B., Rudnicki, W.R., 2010. Feature selection with the boruta package. J. Stat. Softw. 36 (11), 1–13.
Kuo, P.F., Hsu, W.T., Lord, D., Putra, I.G.B., 2024. Classification of autonomous vehicle crash severity: Solving the problems of imbalanced datasets and small sample size. Accid Anal Prev 205, 107666.
Kutela, B., Ngeni, F., Ruseruka, C., Chengula, T.J., Novat, N., Shita, H., Kinero, A., 2025. The influence of roadway characteristics and built environment on the extent of over-speeding: An exploration using mobile automated traffic camera data. Int. J. Transp. Sci. Technol 17, 120–130.
Laviolette, J., Morency, C., Waygood, E.O.D., 2022. A kilometer or a mile? Does buffer size matter when it comes to car ownership? J. Transp. Geogr. 104.
Law, J., Petric, A.T., 2024. Monitoring day and dark traffic collisions in toronto neighbourhoods with implications for injury reduction and vision zero initiatives: A spatial analysis approach. Accid. Anal. Prev. 207, 107728.
Li, Y., Song, L., Fan, W., 2021. Day-of-the-week variations and temporal instability of factors influencing pedestrian injury severity in pedestrian-vehicle crashes: A random parameters logit approach with heterogeneity in means and variances. Analytic Methods in Accident Research 29.
Li, Y., Yang, Z., Xing, L., Yuan, C., Liu, F., Wu, D., Yang, H., 2023. Crash injury severity prediction considering data imbalance: A wasserstein generative adversarial network with gradient penalty approach. Accid. Anal. Prev. 192, 107271.
Li, Y.C., Sze, N.N., Wong, S.C., 2013. Spatial-temporal analysis of drink-driving patterns in hong kong. Accid. Anal. Prev. 59, 415–424.
Lundberg, S.M., Lee, S.I., Year. A unified approach to interpreting model predictions. In: Proceedings of the 31st Annual Conference on Neural Information Processing Systems (NIPS), Long Beach, CA.
Lundberg, S.M., Erion, G., Chen, H., Degrave, A., Prutkin, J.M., Nair, B., Katz, R., Himmelfarb, J., Bansal, N., Lee, S.I., 2020. From local explanations to global understanding with explainable ai for trees. Nat. Mach. Intell. 2 (1), 56–67.
Luo, Y., Liu, Y., Tong, Z., Wang, N., Rao, L., 2023. Capturing gender-age thresholds disparities in built environment factors affecting injurious traffic crashes. Travel Behav. Soc. 30, 21–37.
Lym, Y., Chen, Z., 2021. Influence of built environment on the severity of vehicle crashes caused by distracted driving: A multi-state comparison. Accid. Anal. Prev. 150, 105920.
Ma, W., Kofi Alimo, P., Wang, L., Abdel-Aty, M., 2022. Mapping pedestrian safety studies between 2010 and 2021: A scientometric analysis. Accid. Anal. Prev. 174, 106744.
Madarshahian, M., Hawkins, J., Huynh, N., Siddiqui, C.K.A., 2025. Investigation of factors affecting crash severity of rear-end crashes with high collision speeds in work zones: A south carolina case study. Int. J. Transp. Sci. Technol. 17, 361–374.
Mannering, F.L., Shankar, V., Bhat, C.R., 2016. Unobserved heterogeneity and the statistical analysis of highway accident data. Anal. Methods Accid. Res 11, 1–16. Marshall, E., Shirazi, M., Ivan, J.N., 2023. Covid-19 and transport safety. Transp. Rev.
1–26. Merlin, L.A., Guerra, E., Dumbaugh, E., 2020. Crash risk, crash exposure, and the built environment: A conceptual review. Accid. Anal. Prev. 134, 105244.
Mirhashemi, A., Amirifar, S., Tavakoli Kashani, A., Zou, X., 2022. Macro-level literature analysis on pedestrian safety: Bibliometric overview, conceptual frames, and trends. Accid. Anal. Prev. 174, 106720.
Mohamed, M.G., Saunier, N., Miranda-Moreno, L.F., Ukkusuri, S.V., 2013. A clustering regression approach: A comprehensive injury severity analysis of pedestrian–vehicle crashes in new york, us and montreal, canada. Saf. Sci. 54, 27–37.
NHTSA, 2024. Traffic safety facts: Pedestrians. National Highway Traffic Safety Administration Report No.: DOT HS 813 590.
Nasrollahzadeh, A.A., Sofi, A.R., Ravani, B., 2021. Identifying factors associated with roadside work zone collisions using machine learning techniques. Accid Anal Prev 158, 106203.
Park, S., Ko, D., 2020. Investigating the factors influencing pedestrian–vehicle crashes by age group in seoul, south korea: A hierarchical model. Sustainability 12 (10).
Prato, C.G., Kaplan, S., Patrier, A., Rasmussen, T.K., 2018. Considering built environment and spatial correlation in modeling pedestrian injury severity. Traffic Inj. Prev. 19 (1), 88–93.
Rangam, H., Sivasankaran, S.K., Balasubramanian, V., 2024. Visual hazardous models: A hybrid approach to investigate road hazardous events. Accid. Anal. Prev. 200,
Samerei, S.A., Aghabayk, K., 2024. Analyzing the transition from two-vehicle collisions to chain reaction crashes: A hybrid approach using random parameters logit model, interpretable machine learning, and clustering. Accid Anal Prev 202, 107603.
Sasidharan, L., Wu, K.F., Menendez, M., 2015. Exploring the application of latent class cluster analysis for investigating pedestrian crash injury severities in switzerland. Accid. Anal. Prev. 85, 219–228.
Se, C., Champahom, T., Jomnonkwao, S., Karoonsoontawon, A., Ratanavaraha, V., 2022. Analysis of driver-injury severity: A comparison between speeding and non-speeding driving crash accounting for temporal and unobserved effects. Int J Inj Contr Saf Promot 29 (4), 475–488.
Schlogl, M., 2020. A multivariate analysis of environmental effects on road accident occurrence using a balanced bagging approach. Accid. Anal. Prev. 136, 105398.
Sga, 2024. Dangerous by design 2024. Smart Growth America.
Shi, X., Wong, Y.D., Li, M.Z., Palanisamy, C., Chai, C., 2019. A feature learning approach based on xgboost for driving assessment and risk prediction. Accid. Anal. Prev. 129, 170–179.
Song, L., Li, Y., Fan, W., Wu, P., 2020. Modeling pedestrian-injury severities in pedestrian-vehicle crashes considering spatiotemporal patterns: Insights from different hierarchical bayesian random-effects models. Analytic Methods in Accident Research 28.
Stiles, J., Kar, A., Lee, J., Miller, H.J., 2023. Lower volumes, higher speeds: Changes to crash type, timing, and severity on urban roads from covid-19 stay-at-home policies. Transportation Research Record: Journal of the Transportation Research Board 2677 (4), 15–27.
Su, J., Sze, N.N., Bai, L., 2021. A joint probability model for pedestrian crashes at macroscopic level: Roles of environment, traffic, and population characteristics. Accid. Anal. Prev. 150, 105898.
Sun, Z., Ai, Z., Wang, Z., Wang, J., Gu, X., Wang, D., Lu, H., Chen, Y., 2024. Considering multi-scale built environment in modeling severity of traffic violations by elderly drivers: An interpretable machine learning framework. Accid. Anal. Prev. 207, 107740.
Sun, Z., Wang, D., Gu, X., Abdel-Aty, M., Xing, Y., Wang, J., Lu, H., Chen, Y., 2023. A hybrid approach of random forest and random parameters logit model of injury severity modeling of vulnerable road users involved crashes. Accid. Anal. Prev. 192,
Wang, K., Shirani-Bidabadi, N., Shaon, R.R., M., Zhao, S., Jackson, E.,, 2021. Correlated mixed logit modeling with heterogeneity in means for crash severity and surrogate measure with temporal instability. Accid. Anal. Prev. 160, 106332.
Wan, L., Yan, Y., Zhang, C.A., Liu, C., Mao, T., Wang, W., 2024. Characteristics and identification of risky driving behaviors in expressway tunnels based on behavior spectrum. Int. J. Transp. Sci. Technol 16, 5–17.
Wang, S., Gao, K., Zhang, L., Yu, B., Easa, S.M., 2024a. Geographically weighted machine learning for modeling spatial heterogeneity in traffic crash frequency and determinants in us. Accid. Anal. Prev. 199, 107528.
Wang, X., Zhang, X., Pei, Y., 2024b. A systematic approach to macro-level safety assessment and contributing factors analysis considering traffic crashes and violations. Accid. Anal. Prev. 194, 107323.
Wang, Z., Jiao, P., Wang, J., Luo, W., Lu, H., 2023a. Contributing factors on the level of delay caused by crashes: A hybrid method of latent class analysis and xgboost based shap algorithm. Journal of Transportation Safety & Security 16 (2), 97–129.
Wang, Z., Jiao, P., Wang, J., Luo, W., Lu, H., 2023b. Improved two-layer stacking model for prediction of the level of delay caused by crashes: An empirical analysis of texas. Journal of Transportation Engineering, Part a: Systems 149 (2).
Watson-Brown, N., Nicolls, M., Mardani, A., Aveleira Fernandes, I., Truelove, V., 2024. A qualitative study of the first covid-19 pandemic lockdown: The impact on alcohol consumption and driving behaviours. Transport. Res. F: Traffic Psychol. Behav. 104, 449–460.
Wei, S., Yang, H., Li, Y., Xie, M., Wang, Y., 2024. Investigating the impact of temporal instability in smart roadway retrofitting on terrain-related crash injury severity. Accid. Anal. Prev. 207, 107757.
Wen, X., Xie, Y., Jiang, L., Pu, Z., Ge, T., 2021a. Applications of machine learning methods in traffic crash severity modelling: Current status and future directions. Transp. Rev. 41 (6), 855–879.
Wen, X., Xie, Y., Wu, L., Jiang, L., 2021b. Quantifying and comparing the effects of key risk factors on various types of roadway segment crashes with lightgbm and shap. Accid. Anal. Prev. 159, 106261.
Wu, P., Meng, X., Song, L., 2021. Bayesian space–time modeling of bicycle and pedestrian crash risk by injury severity levels to explore the long-term spatiotemporal effects. Physica a: Statistical Mechanics and Its Applications 581.
Xin, C., Guo, R., Wang, Z., Lu, Q., Lin, P.-S., 2017. The effects of neighborhood characteristics and the built environment on pedestrian injury severity: A random parameters generalized ordered probability model with heterogeneity in means and variances. Anal. Methods Accid. Res 16, 117–132.
Xu, L., Skoularidou, M., Cuesta-Infante, A., Veeramachaneni, K., Year. Modeling tabular data using conditional gan. In: Proceedings of the 33rd Conference on Neural Information Processing Systems (NeurIPS), Vancouver, CANADA.
Xue, H., Guo, P., Li, Y., Ma, J., 2024. Integrating visual factors in crash rate analysis at intersections: An automl and shap approach towards cycling safety. Accid. Anal. Prev. 200, 107544.
Yan, X., He, J., Wu, G., Zhang, C., Wang, C., Ye, Y., 2022. Differences of overturned and hit-fixed-object crashes on rural roads accompanied by speeding driving: Accommodating potential temporal shifts. Analytic Methods in Accident Research
Yang, T., Fan, W.D., Song, L., 2023. Modeling pedestrian injury severity in pedestrianvehicle crashes considering different land use patterns: Mixed logit approach. Traffic Inj. Prev. 24 (2), 114–120.
Zhai, X., Huang, H., Sze, N.N., Song, Z., Hon, K.K., 2019. Diagnostic analysis of the effects of weather condition on pedestrian crash severity. Accid Anal Prev 122, 318–324.
Zhai, G., Xie, K., Yang, D., Yang, H., 2024. Developing equity-aware safety performance functions for identifying hotspots of pedestrian-involved crashes. Accid. Anal. Prev. 207, 107759.
Zhang, X., Zhao, X., Bian, Y., Huang, J., Yin, L., 2024. Interactive effects analysis of road, traffic, and weather characteristics on shared e-bike speeding risk: A data-driven approach. Accid. Anal. Prev. 207, 107755.
Zhou, Y., Fu, C., Jiang, X., Liu, H., 2024. Analyzing the heterogenous effects of factors on high-range speeding likelihood of taxi speeders: Does explainable deep learning provides more insights than random parameter approach? Accid. Anal. Prev. 207,
Zhu, C., Dadashova, B., Lee, C., Ye, X., Brown, C.T., 2024a. Equity in non-motorist safety: Exploring two pathways in houston. Transp. Res. Part D: Transp. Environ. 132.
Zhu, W., Lü, C., Chen, X., 2024b. A crash occurrence risk prediction model based on variational autoencoder and generative adversarial network. Transportmetrica b:
Zhu, S.Y., Meng, Q., 2022. What can we learn from autonomous vehicle collision data on crash severity? A cost-sensitive cart approach, Accident Analysis and Prevention, p. 174.
Ziakopoulos, A., 2024. Analysis of harsh braking and harsh acceleration occurrence via explainable imbalanced machine learning using high-resolution smartphone telematics and traffic data. Accid. Anal. Prev. 207, 107743.

wangcontextdependenteffectsbuilt2025.md 147 KB История Исходник

status: converted

Accident Analysis and Prevention

Context-dependent effects of built environment factors on pedestrian-injury severities with imbalanced and high dimensional crash data

ARTICLE INFO

ABSTRACT

1. Introduction

2. Literature review

2.1. Pedestrian injury severity studies considering built environment

2.2. Research on enhanced machine learning framework

2.3. Research gaps

3. Data collection and description

4. Methodology

4.1. Conceptual framework

4.2. Conditional tabular generative adversarial network

4.3. LightGBM model and Shapley Additive exPlanations (SHAP)

Algorithm 1 LightGBM

Input:

Process:

Algorithm 1 LightGBM

4.4. Improved wrapper-based feature selection methods

[Optional] Recursive feature elimination-SHAP algorithm

[Optional] Recursive feature addition-SHAP algorithm

5. Results and discussion

5.1. Enhanced model classification performance analysis

5.1.1. Data balancing methods comparison

5.1.2. Improved wrapper-based feature selection methods comparison

classification ability of the model.

5.2. Fixed and heterogeneous factors analysis

5.3. Context-dependent effects analysis

5.3.1. Between heterogeneous built environment factors

5.3.2. Between heterogeneous built environment factors and non-built environment factors

6. Conclusion

7. Limitations

CRediT authorship contribution statement

Declaration of competing interest

Acknowledgments

Data availability

References

wangcontextdependenteffectsbuilt2025.md 147 KB

История Исходник