Predicting Segment-Level Road Traffic Injury Counts Using Machine Learning Models: A Data-Driven Analysis of Geometric Design and Traffic Flow Factors

Abstract

Accurate prediction of road traffic crash severity is essential for developing data-driven safety strategies and optimizing resource allocation. This study presents a predictive modeling framework that utilizes Random Forest (RF), Gradient Boosting (GB), and K-Nearest Neighbors (KNN) to estimate segment-level frequencies of fatalities, serious injuries, and slight injuries on Hungarian roadways. The model integrates an extensive array of predictor variables, including roadway geometric design features, traffic volumes, and traffic composition metrics. To address class imbalance, each severity class was modeled using resampled datasets generated via the Synthetic Minority Over-sampling Technique (SMOTE), and model performance was optimized through grid-search cross-validation for hyperparameter optimization. For the prediction of serious- and slight-injury crash counts, the Random Forest (RF) ensemble model demonstrated the most robust performance, consistently attaining test accuracies above 0.91 and coefficient of determination (R²) values exceeding 0.95. In contrast, for fatalities count prediction, the Gradient Boosting (GB) model achieved the highest accuracy (0.95), with an R² value greater than 0.87. Feature importance analysis revealed that heavy vehicle flows consistently dominate crash severity prediction. Horizontal alignment features primarily influenced fatal crashes, while capacity utilization was more relevant for slight and serious injuries, reflecting the roles of geometric design and operational conditions in shaping crash occurrence and severity. The proposed framework demonstrates the effectiveness of machine learning approaches in capturing non-linear relationships within transportation safety data and offers a scalable, interpretable tool to support evidence-based decision-making for targeted safety interventions.

1. Introduction

Road traffic crashes (RTCs) remain a persistent global challenge, exacting substantial human, social, and economic tolls. These incidents result in millions of fatalities and injuries annually, straining healthcare systems and disrupting economic productivity due to emergency responses, long-term rehabilitation, and lost labor contributions [1,2,3]. From an infrastructure planning perspective, the ability to accurately predict the severity of road traffic incidents at a granular level, such as the roadway segment, is fundamental to enabling data-driven safety interventions. Effective prediction informs not only where but also how safety measures such as geometric redesign, enforcement strategies, and traffic control systems should be deployed to maximize reductions in crash severity and frequency [4,5,6]. Conventional statistical approaches, including Poisson and negative binomial regression models, have historically served as the analytical backbone for crash frequency and severity modeling [7,8]. While these models offer transparency and statistical rigor, they are frequently limited in their capacity to account for the complex, non-linear interactions that characterize real-world crash dynamics [9,10]. Variables such as road geometry, traffic flow composition, and temporal patterns often interact in ways that violate the assumptions underpinning traditional parametric techniques. In response to these limitations, the transportation safety field has increasingly adopted machine learning (ML) methodologies, which offer enhanced flexibility in uncovering hidden patterns within high-dimensional datasets [11]. Techniques such as support vector machines, artificial neural networks, and ensemble methods like gradient-boosted trees have shown improved performance in classifying crash severity [12,13,14,15]. However, a substantial portion of existing ML-based research tends to either aggregate crash outcomes or divide severity levels into simplistic binary categories (e.g., fatal vs. non-fatal) [16,17,18]. This oversimplification undermines the specificity needed for granular policy design and weakens the alignment between predictive insights and real-world safety interventions.

To address these gaps, the present study introduces a comprehensive ensemble-learning framework utilizing three machine-learning algorithms, principally Random Forest, Gradient Boosting, and K-Nearest Neighbors classifiers, to model segment-level counts of three crash severity outcomes: fatalities, serious injuries, and slight injuries. The predictive models are constructed using a rich set of cross-sectional covariates, incorporating roadway design characteristics (e.g., horizontal and vertical alignment), traffic-volume attributes, and disaggregated vehicle-class flow metrics [19]. To counter the pronounced class imbalance that often characterizes crash data, the study implements the Synthetic Minority Over-sampling Technique (SMOTE) [20,21]. Furthermore, a stratified sampling strategy is adopted to maintain proportional distributions of severity classes within both the training and testing subsets, thereby improving the statistical robustness and reliability of model evaluation. Hyperparameter optimization is achieved through grid-search cross-validation to enhance generalizability and predictive robustness [22,23]. Model performance is comprehensively assessed using a suite of evaluation metrics, including accuracy, precision, F1-score, the Matthews correlation coefficient (MCC), and the geometric mean (G-Mean), to provide a rigorous and balanced appraisal of predictive quality across severity classes. Unlike prior efforts that prioritize model interpretability or predictive accuracy, this study seeks to harmonize both objectives. By leveraging the inherent feature-ranking capability of ensemble algorithms, particularly Random Forest and Gradient Boosting, the developed framework not only achieves high classification performance but also produces interpretable insights regarding the relative importance of predictor variables. This dual capacity ensures that the framework remains both practically applicable and theoretically grounded.

The remainder of the paper is organized as follows: Section 2 presents a review of related literature, situating this work within the broader context of crash severity modeling. Section 3 details the data sources, preprocessing procedures, and modeling methodology. Section 4 reports the empirical findings and discusses the implications of the results. Finally, Section 5 outlines practical applications, policy considerations, and potential directions for future research.

2. Related Work

The modeling of road traffic injury severity has traditionally relied on classical count-data regression techniques, particularly Poisson and negative binomial models. These methods have been extensively used to estimate crash frequency and severity as functions of roadway geometry, traffic exposure, and environmental conditions [24,25]. Consequently, such models often struggle to capture the complex, non-linear interactions and threshold effects that characterize real-world traffic incidents, particularly when modeling varying levels of injury severity.

In response to these limitations, the traffic safety literature has increasingly embraced machine learning (ML) techniques as more flexible, data-driven alternatives. Numerous studies employing algorithms such as support vector machines (SVMs), gradient-boosted decision trees (GBDTs), and other ensemble methods have demonstrated enhanced capabilities in uncovering latent patterns and modeling non-linear relationships in crash data [26,27,28]. Despite their improved predictive performance, most existing ML-based models approach crash severity as a binary classification task, typically distinguishing fatal from non-fatal crashes, thereby oversimplifying a multifaceted outcome and limiting the granularity required for targeted policy interventions [29].

Another persistent methodological challenge concerns class imbalance. Severe and fatal crashes comprise a small proportion of total incidents, which are typically dominated by minor injuries or property-damage-only events. This imbalance skews model training, often resulting in biased classifiers that underperform in predicting rare but critical outcomes. To mitigate this, techniques such as the Synthetic Minority Over-sampling Technique (SMOTE) have been applied with success in binary classification contexts, enabling improved sensitivity to minority classes without distorting the underlying data distribution [30,31]. However, their use in multi-class or count-based severity prediction remains limited, thereby constraining broader applicability in multi-output modeling frameworks. Among ensemble learning methods, Random Forests (RF) has emerged as a particularly robust and interpretable approach for traffic crash modeling. As a bagging-based technique, RF constructs an ensemble of decision trees using bootstrapped subsets of the data and aggregates their outputs to improve generalization performance. The algorithm is well-suited for handling high-dimensional inputs, non-linear relationships, multicollinearity, and noisy data, all common characteristics of transportation safety datasets [32,33]. Furthermore, RF models offer intrinsic measures of variable importance, facilitating post hoc interpretability and yielding in-sights into the relative influence of predictor variables. Comparative studies have consistently highlighted the superior performance of Random Forests relative to both traditional statistical models and other ML algorithms in crash severity classification tasks. Gains in key evaluation metrics, including accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic (ROC) curve, have been reported even under conditions of class imbalance and data heterogeneity [34,35]. Nevertheless, much of the existing literature continues to treat crash severity as a classification problem and either aggregates injury categories or disregards the need for predicting real-valued severity counts. This limits the operational utility of such models in segment-level resource allocation and proactive safety planning. Moreover, the integration of detailed roadway design features and high-resolution traffic composition variables in machine learning models remains underdeveloped. While individual geometric attributes such as curvature or horizontal alignment have been explored in isolation, the combined use of geometric variables and vehicle-type-specific flow metrics is still underrepresented in the literature, despite their demonstrated relevance in determining injury severity outcomes.

Table 1. Previous Relevant Studies.

Authors	Description	Methods	Key Findings
Dadashova et al. (2020) [36]	Investigated crash injury severity using discrete-choice models and Random Forest on Spanish trans-European corridors.	Logistic regression and RF, with crash types disaggregated by roadway, driver, and environmental variables.	Roadway design elements (curvature, super elevation, lane width) were significant predictors. Logistic regression highlighted conditional effects by crash type, whereas RF identified critical factors across crash categories.
Yassin (2020) [37]	Developed a hybrid framework integrating K-means clustering with Random Forest for crash severity prediction	K-means was used to extract hidden features; RF was employed for classification against alternative classifiers.	The hybrid approach achieved outstanding accuracy (99.86%). Driver experience, lighting conditions, driver age, and vehicle service year emerged as dominant factors in predicting severity outcomes.
Atumo et al. (2021) [38]	Applied spatial statistics and Random Forest to traffic crash hot spot identification.	Getis-Ord statistics for spatial clustering complemented with RF-based crash prediction using 2010–2017 data.	Identified crash hot spots on interstate routes; RF achieved validation and prediction accuracy of 76.7% and 74%, respectively. Results highlighted the spatial dependence of crash distributions and confirmed predictive robustness.
Akin et al. (2022) [39]	Identified behavioral causes of crashes on Saudi Arabian highways using supervised ML.	Logistic regression, RF, and KNN applied to driver error-related crashes on the highway.	RF and logistic regression achieved the highest accuracy (78.7%), with RF attaining the largest AUC (0.712). Findings revealed that traffic flow speed and lane count reduced driver error–related crashes, while higher AADT and curve sections increased risk.
Gatera et al. (2023) [40]	Compared Random Forest and Support Vector Machine regression models for short-term road crash forecasting	RF and SVM were evaluated using error indices (MAE, MSE, RMSE) and R2 values.	RF demonstrated superior predictive capacity (R2 = 0.91) compared to SVM (R2 = 0.86), reinforcing the promise of ML in traffic crash forecasting.
Nikolaou et al. (2023) [41]	Compared ML techniques for motorway segment crash risk prediction	Logistic Regression, Decision Tree, RF, SVM, and kNN applied to road design and naturalistic driving datasets.	RF achieved the highest accuracy (89.3%) and superior precision-recall-F1 balance. Shapley additive explanations enhanced interpretability, highlighting RF’s effectiveness for crash risk assessment.
Yang et al. (2023) [33]	Proposed a Random Forest based framework for predicting crash severity using enriched feature sets	RF compared with BP neural network, SVM, and radial basis neural network; feature importance ranking applied to 12 variables.	RF outperformed other models, achieving higher recall (0.83) and F1 scores, with a lower false alarm rate. Results confirmed its reliability and stability in severity prediction.
Ahmed et al. (2023) [42]	Examines crash prediction and contributing factors using explainable ensemble machine learning models	Random Forest (RF), Decision Jungle, AdaBoost, XGBoost, LightGBM, and CatBoost, with interpretability through SHAP analysis.	RF achieved the highest predictive accuracy (81.45%), with road category and number of vehicles identified as the most influential factors affecting injury severity.
Wang et al. (2024) [43]	Developed a meta-heuristic optimized Random Forest model for traffic crash severity prediction.	Compared nine meta-heuristic RF variants (e.g., CPO-RF, SSA-RF) against standard ensemble and single classifiers using U.S. crash data.	CPO-RF yielded superior accuracy (95.2%) and F1 scores exceeding 90%. Application of inverse SMOTE improved accuracy to 99.6%. Key predictors included temperature, weather, pressure, GDP, population density, and time of day.
Uzunov et al. (2025) [44]	Comparative analysis of proportional distribution methods and Random Forest for pedestrian crash risk prediction	Proportional risk distribution and RF applied to data derived from court cases, with risk factors quantified through expert evaluation.	Both approaches were valid; however, RF provided superior accuracy and robustness. Significant correlation between methods confirmed validity, with graphical visualizations aiding the interpretability of risk severity.
Daoud et al. (2025) [45]	Examined nighttime crash severity and the role of roadway illumination	Developed and compared seven machine learning models (logistic regression, k-NN, naïve Bayes, random forest, ANN, XGBoost, LSTM) using TxDOT crash data.	The random forest model produced the most promising results by predicting severe crashes with 97.6% accuracy.

summarizes the methods and the key findings in the relevant studies.

To address these research gaps, this study develops a comprehensive ensemble-learning framework comprising Random Forest, Gradient Boosting, and K-Nearest Neighbors models to predict segment-level counts of fatalities, serious injuries, and slight injuries. A key limitation in prior crash-severity research is the restricted resolution and geographic diversity of commonly used datasets, which often rely on aggregated traffic measures and coarse roadway classifications [18,19,26,46,47,48,49]. In contrast, this study introduces a novel contribution by utilizing a uniquely detailed national dataset from Hungary that integrates high-resolution geometric attributes with disaggregated traffic-flow indicators for multiple vehicle classes, enabling a more precise examination of how vehicle-type composition interacts with segment-level design features. Building on this rich feature set, the modeling framework incorporates roadway geometry, traffic volume, and vehicle-class distributions, while addressing class imbalance through SMOTE and ensuring statistical consistency via stratified sampling. Hyperparameter tuning through grid-search cross-validation further enhances model robustness, resulting in an interpretable and scalable approach to segment-level crash severity counts prediction.

3. Materials and Methods

This study implements a data-driven ensemble-learning framework to model segment-level road traffic injury counts using multiple supervised machine-learning algorithms. Specifically, three independent predictive models are developed—Random Forest (RF), Gradient Boosting (GB), and K-Nearest Neighbors (KNN), as shown in Figure 1. Each model is tailored to predict the number of injury outcomes per road segment across three distinct severity categories: (1) fatalities, (2) serious injuries, and (3) slight injuries. The models are trained on a rich set of covariates that characterize roadway geometric design, traffic volume, and vehicle composition, enabling nuanced, high-resolution injury forecasting grounded in empirical roadway conditions.

3.1. Data Description and Feature Engineering

Crash severity classification in this study follows the Hungarian national road crash database, which adheres to the European Union CARE (Community Road Accident Database) definition. In this classification, a fatal crash is one in which at least one person dies within 30 days of the crash; a serious (severe) injury crash involves at least one person who sustains injuries requiring hospital treatment for more than 24 h; and a slight injury crash involves injuries requiring medical attention but less than 24 h of hospitalization [50]. Furthermore, in this study, a road segment is defined according to the specifications provided by the Hungarian Institute for Transport Sciences and Logistics Non-profit Ltd. (KTI). Each segment is uniquely identified by the road number and precise start and end locations, including kilometer and meter coordinates, as well as segment codes. This definition provides spatial consistency and enables accurate aggregation of traffic, geometric, and crash data for modeling purposes. This research utilizes a detailed segment-level crash dataset compiled from the Hungarian national road network for the year 2023. Each record represents the actual number of persons injured or fatally wounded in traffic crashes along a specific roadway segment. dataset integrates geometric design characteristics, traffic exposure indicators, and vehicle composition metrics, forming a robust basis for developing machine learning-based models of crash injury severity.

The dataset comprises 12,216 crash-related injury records representing the actual number of individuals with severe injuries or fatal outcomes across the Hungarian road network. Of these, 9571 correspond to slight injuries (≈67%), 4071 to serious injuries (≈29%), and 626 to fatal outcomes (≈4%). This proportional distribution reflects the empirical injury burden within the Hungarian network, where slight injuries dominate, yet fatal and serious crashes remain concentrated on specific high-risk corridors [51,52]. The inclusion of segment-level geometric and operational indicators enables a comprehensive representation of crash exposure conditions across diverse roadway environments.

To ensure analytical rigor and reproducibility, multiple preprocessing steps were undertaken. Alphanumeric road identifiers were numerically encoded using a LabelEncoder, ensuring consistency across all feature types and eliminating non-numeric symbols from feature names to facilitate model processing. Missing entries were imputed using median substitution, while extreme values were filtered using the interquartile range (IQR) method to reduce statistical distortion [53,54].

A summary of key variables is presented in

Table 2. Summary of Key Segment-Level Variables by Injury Severity.

Variable Category	Representative Variable	Fatal	Serious	Slight
Track Code	Undivided	83%	84%	85%
	Left track	9%	9%	8%
	Right track	8%	7%	7%
Road Category	Motorway	18%	16%	14%
	Expressway	4%	5%	5%
	Primary Main Road	22%	25%	26%
	Secondary Main Road	57%	54%	55%
Section Type	Rural	81%	62%	63%
	Urban	19%	38%	37%
Number of Lanes	Two lanes	85%	83%	83%
Horizontal Alignment	Straight	48%	44%	46%
	Right/Left	34%	36%	35%

, highlighting the geometric and operational distinctions across injury severity levels. Fatal and serious crashes were predominantly concentrated on rural, undivided two-lane roads (over 80%), emphasizing the influence of high-speed, bidirectional traffic with limited separation. Primary and secondary main roads accounted for the majority of severe outcomes, while expressways exhibited a relatively higher share of slight injuries, reflecting improved geometric standards and access control.

Traffic exposure indicators further underscore the risk disparity. The Annual Average Daily Traffic (AADT) averaged approximately 13,000 veh/day for fatal segments, marginally higher than for serious and slight injury segments (12,640 and 11,849 veh/day, respectively). Heavy vehicle presence was a notable differentiator, with truck and bus flows constituting a higher proportion of traffic on high-severity segments.

Table 3. Dataset Variables Overview.

Category	Variable Name	Description
Road Attributes	Track Code	(0: undivided, 1: left track, 2: right track).
	Road Category	(1: motorway, 2: expressway, 3: primary main road, 4: secondary main road).
	Section Type	(1: rural or 2: urban).
	Number of Traffic Lanes	Numeric
	Radius of Horizontal Curve	(In meters).
	Direction of Horizontal Curve	(1: right, 2: left, 3: straight).
	Slope/Gradient	(In %.).
	Type of Vertical Curve	Type of vertical alignment curve (e.g., crest, sag).
	Radius of Vertical Curve	(In meters).
Traffic Attributes	Capacity Utilization	Ratio of AADT to road capacity (%).
	Heavy Truck Traffic	Numeric
	Medium Heavy (2-Axle) Truck Traffic	Numeric
	Truck with Trailer or Semi-Trailer Traffic	Numeric
	Tractor-Trailer with Semi-Trailer Traffic	Numeric
	Light Truck Traffic	Numeric
	Total Bus (Single) Traffic	Numeric
	Bus (Articulated) Traffic	Numeric
	Motorcycle and Moped Traffic	Numeric
	Bicycle Traffic	Numeric
	Slow Vehicle and Agricultural Tractor Traffic	Numeric
Derived Variables	Severity	Target variable: crash severity counts (slight injuries, serious injuries, fatalities).

provides an overview of the dataset variables used in this study.

3.2. Machine Learning Models

Three machine learning classification algorithms were selected based on their proven utility in transportation safety modeling and ability to handle mixed-type, high-dimensional datasets:

3.2.1. Random Forest (RF)

Random Forest (RF) is a popular ensemble learning method for classification tasks, it was first introduced by Ho (1995) [55] and later improved by Breiman (2001) [56]. This method leverages the bagging technique, which builds multiple independent decision trees during the training phase by utilizing feature randomness to create a diverse set of uncorrelated trees. Each tree in the forest is trained on a different bootstrap sample, where bootstrapping refers to the sampling of data with replacement, ensuring that each tree is exposed to a slightly different subset of the data [57]. The underlying principle of this ensemble method is that combining multiple models leads to improved generalization and consequently, higher prediction accuracy. Mathematically, the prediction for a given instance $X$ is determined by Equation (1):

$$\widehat{y}={\displaystyle \frac{1}{T}}\sum _{t=1}^T{f}_t\left(X\right)$$

(1)

where

$T$ represents the total number of trees in the forest

$f_t\left(X\right)$ the prediction of the t-th tree for the input X

$\widehat{y}$ is the final predicted output, typically determined by majority vote in classification tasks

For classification tasks, each tree in the forest produces a class label $\widehat{y}$, and the final prediction is given by: $\widehat{y}$ = mode ($y_1,y_2,\dots ,y_T)$, where the mode function returns the most frequent class label from the predictions of all the trees. This ensemble approach allows the random forest to significantly improve the model’s accuracy compared to individual decision trees, especially in terms of generalization to unseen data.

3.2.2. Gradient Boosting Machines (GBM)

Constitute a class of powerful ensemble learning algorithms that construct predictive models through the sequential integration of weak learners, most commonly shallow decision trees, via gradient-based optimization in function space [58]. At each iteration, the algorithm fits a new tree $h_m\left(x\right)$ to the negative gradient of the selected loss function $\mathcal{L}$ (y, F($x$)), yielding a pseudo-residual defined as:

$$r_{im}=- {\displaystyle \frac{\partial \mathcal{L}(y_i,F(x_i\left)\right)}{\partial F\left(x_i\right)}}$$

(2)

The ensemble model is then updated using:

$$F_m\left(x\right)= F_{m-1}\left(x\right)+\nu ·{\gamma }_m{h}_m\left(x\right)$$

(3)

where ν ∈ (0, 1] is the learning rate, and ${\gamma }_m$ is a step-size parameter optimized at each iteration by minimizing the loss function:

$${\gamma }_m=\mathit{arg}\underset{\gamma }{\mathit{min}}\sum _{i=1}^N\mathcal{L}(y_i,F_{m-1}\left(x_i\right)+ {\gamma }_m{h}_m\left(x_i\right))$$

(4)

To enhance generalization and reduce overfitting, GBM incorporates regularization through shrinkage (via the learning rate ν) and stochastic subsampling of the training data at each boosting iteration. In this study, GBM was applied to the binary classification problem of predicting pedestrian priority at unsignalized intersections, capturing complex nonlinear dependencies among geometric, traffic, and behavioral variables. The logistic loss function was employed as the objective criterion for model optimization.

3.2.3. K-Nearest Neighbors (KNN)

The K-Nearest Neighbors (KNN) method [59] determines the class label of a test sample based on the labels of its $\mathrm{k}$ nearest neighbors. Suppose a distance metric is predefined (such as Euclidean distance, Mahalanobis distance, etc.). For any test sample $x$, its $k$ nearest neighbors can be identified and denoted as $N_k\left(x\right)$. The class label of $x$ is then determined by the labels of the training samples in $N_k\left(x\right)$, which can be expressed as:

$$f\left(x\right)=\underset{c_j\in C}{\arg \max }\sum _{x_i\in N_k\left(x\right)}I(y_i=c_j)$$

(5)

where C = {c1, c2, …, cm}, $y_i$ is the class label of $x_i$, m is the number of classes.

Here $f\left(x\right)$, represents the decision function, and $I\left(.\right)$ is the indicator function. For each class $c_j$, the indicator function is defined as:

$$(y_i=c_j)=\left\{\begin{array}{cc}1,& if y_i=c_j\\ 0,& otherwise\end{array}\right.$$

(6)

In the KNN model, the form of the decision function may vary depending on the chosen strategy. In this study, the decision function is derived according to a voting strategy, where the majority label among the $k$ nearest neighbors determines the predicted class.

3.3. Hyperparameter Tuning and Model Optimization

For each injury severity category, feature matrices and label vectors were constructed by isolating the corresponding target count variable and excluding it from the predictor set. Due to the inherent class imbalance, particularly for fatal and serious injury outcomes, the Synthetic Minority Over-sampling Technique (SMOTE) was applied. This procedure generates synthetic observations for minority classes in the training data, thereby improving model sensitivity without overfitting to overrepresented outcomes [60,61].

To preserve outcome distribution proportions, an 80/20 stratified train/test split was implemented using the StratifiedShuffleSplit function. This ensured that the severity classes remained proportionally represented in both training and testing datasets, a crucial consideration for reliable model evaluation. All computational procedures were conducted using Python 3.10. Data preprocessing and manipulation were performed using the pandas and numpy libraries. Model development and evaluation leveraged scikit-learn, while oversampling of minority classes was handled with imblearn. The complete modeling pipeline, including, feature engineering, class rebalancing via SMOTE, model training, hyperparameter tuning, and performance evaluation, was encapsulated within a modular and fully reproducible script. Final outputs, including performance metrics and ranked feature importance, were programmatically exported into structured Excel workbooks using pandas. ExcelWriter ensured transparency, auditability, and ease of downstream analysis. Each (Random Forest, Gradient Boosting, and K-Nearest Neighbors) model was instantiated using corresponding scikit-learn class from the scikit-learn library, with a fixed random seed for reproducibility. To optimize model complexity and generalization, a grid search over the hyperparameter n_estimators was conducted. Five-fold cross-validation on the training set was used to identify the hyperparameter combination that maximized classification accuracy. This method was selected for its balance between computational efficiency and statistical reliability, providing a stable estimate of each model’s performance. By averaging results over multiple folds, it mitigates overfitting and enhances the robustness and generalizability of the final models [62,63,64].

The hyperparameter grids are summarized in

Table 4. Hyperparameters optimization.

Model	Hyperparameter	Type	Search Range/Values
Random Forest	n_estimators	Discrete	{100, 200, 300}
	max_depth	Discrete	{10, 15, 20, 25}
	min_samples_split	Discrete	{2, 5, 10}
	min_samples_leaf	Discrete	{2, 3, 4}
	max_features	Categorical	{‘sqrt’, ‘log2’}
K-Nearest Neighbors	n_neighbors	Discrete	{3, 5, 7}
Gradient Boosting	n_estimators	Discrete	{100, 150, 200}
	learning_rate	Continuous	{0.01, 0.05, 0.1}
	max_depth	Discrete	{2, 3, 5}
	min_samples_split	Discrete	{2, 5, 10}
	subsample	Continuous	{0.7, 0.9, 1.0}

. Grid-search cross-validation was thus instrumental in reducing overfitting risk and enhancing model robustness across all severity categories [22,65].

3.4. Model Evaluation and Performance Metrics

Model performance was assessed using accuracy, weighted precision, F1-score, G-Mean, MCC, and the coefficient of determination (R²) [66,67]. The accuracy score provides the proportion of correct predictions, while precision and F1-score offer a more detailed view of how well each model performs across different classes [68]. These metrics are computed for each classifier. Equations (7)–(9) represent mathematical formulations used for metrics evaluation, as follows:

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(7)

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(8)

$$F1=2·{\displaystyle \frac{precison·recall}{precison+recall}}$$

(9)

where true positive (TP) denotes positive samples predicted by the model as positive classes, and false positive (FP) denotes negative samples predicted by the model as a positive class; false negative (FN) denotes positive samples predicted by the model as a negative class.

The R² is a statistical metric used to measure how much of the outcome is expected (crashes). The R² values range from zero to one [0, 1]. Zero (0) illustrates that the crashes cannot be predicted based on the historically recorded crashes cases, while One (1) implies the perfect prediction and is given by Equation (10).

$${\mathrm{R}}^2=1-{\displaystyle \frac{\sum _{i=1}^n{(y_i-{\widehat{y}}_l)}^2}{\sum _{i=1}^n{(y_i-{\overline{y}}_l)}^2}}$$

(10)

where ${\widehat{y}}_l$ is the predicted value of the i th sample and $y_i$ is the corresponding true value for the total n samples.

G-mean score, as shown in (11), is also important in the context of imbalanced data, as it measures the geometric mean of sensitivity and specificity, providing a balanced assessment of the model’s performance across classes.

$$\mathrm{G}‑\mathrm{mean}=\sqrt{\left({\displaystyle \frac{TP}{TP+FN}}\right)\times \left({\displaystyle \frac{TN}{TN+FP}}\right)}$$

(11)

Matthew’s Correlation Coefficient (MCC) is highly valuable for imbalanced datasets as it considers all four quadrants of the confusion matrix (TP, TN, FP, FN), providing a more comprehensive evaluation of predictive performance that is robust to class distribution, as depicted in (12).

$$MCC={\displaystyle \frac{\left(TP\times TN\right)-(FP\times FN)}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}}$$

(12)

The MCC quantifies the correlation between predicted and actual classifications, yielding a value between (−1) and (+1). A coefficient of (+1) signifies an ideal prediction, while (0) represents an average performance, and (−1) indicates a poor prediction. To enhance interpretability, feature importance scores were extracted from each trained model. These scores quantify the relative contribution of each predictor to the model’s output and provide practical insights into which geometric or traffic flow characteristics most influence injury counts at the segment level.

4. Results and Discussion

This section presents a comprehensive evaluation of the predictive performance of the models developed for estimating road traffic injury counts at the segment level across three severity categories: fatal, serious, and slight injuries. The analysis emphasizes both predictive fidelity and interpretability by examining model evaluation metrics, including accuracy, precision, F1-score, MCC, G-Mean, MSE, RMSE, and coefficient of determination (R²), as well as feature importance rankings derived from the trained models. These insights collectively inform a thematic synthesis of risk factors and policy implications for data-driven road safety interventions.

4.1. Model Performance Overview

The analysis that follows delineates the predictive behavior of the developed models across the three crash-severity categories, providing a coherent and rigorous examination of model performance for fatal, serious, and slight injury counts.

4.1.1. Fatal Injury Model

The optimal hyperparameter configurations were obtained through the grid-search cross-validation process. For the fatal-injury model, the Random Forest achieved its best performance with max_features = sqrt, max_depth = 10, min_samples_leaf = 3, min_samples_split = 5, and n_estimators = 300, while Gradient Boosting reached optimal results with max_depth = 3, min_samples_split = 5, n_estimators = 200, learning_rate = 0.05, and subsample = 0.9. The KNN model yielded its best predictive performance with n_neighbors = 3. These optimal settings reflect a balance between model complexity and generalization capacity, ensuring that the models capture the nonlinear relationships inherent in fatal crash occurrence data without overfitting. The Random Forest (RF) model demonstrated a robust predictive capability for segment-level fatal crash occurrence, outperforming both the Gradient Boosting (GB) and K-Nearest Neighbors (KNN) models across all evaluated performance metrics (

Table 5. Performance metrics of the predicting models for the counts of fatalities.

	Train			Test
Metric	RF	KNN	GB	RF	KNN	GB
Accuracy	0.9918	0.8499	0.9895	0.9373	0.7224	0.9463
Precision	0.9920	0.8504	0.9897	0.9381	0.7296	0.9471
F1-Score	0.9918	0.8487	0.9895	0.9371	0.7168	0.9460
MCC	0.9878	0.7762	0.9844	0.9065	0.5899	0.9200
G-Mean	0.9938	0.8866	0.9922	0.9528	0.7886	0.9596
MSE	0.0082	0.2801	0.0149	0.1075	0.4836	0.0896
RMSE	0.0906	0.5292	0.1222	0.3278	0.6954	0.2993
R2	0.9877	0.5801	0.9776	0.8386	0.2735	0.8655

). In the training phase, RF achieved exceptionally high accuracy (0.991), precision (0.992), and F1-score (0.992), all accompanied by a Matthews Correlation Coefficient (MCC) of 0.988 and a G-Mean of 0.988. These results indicate a high level of internal consistency and model stability, with minimal bias-variance tradeoff effects. While, in the test set, Gradient Boosting, achieves the higher accuracy (0.946), precision (0.0.947), and F1-score (0.946), outperforming the RF model (accuracy = 0.937) and substantially exceeding KNN (accuracy = 0.722).

The GB model also yielded superior statistical indicators of fit, including a test (R²) of 0.867, signifying that approximately 87% of the variance in fatal crash frequency was successfully explained by the model. Its error measures, namely MSE (0.0896) and RMSE (0.2993), were notably lower than KNN (MSE = 0.4836, RMSE = 0.6954), further reinforcing the model’s precision in capturing nonlinear relationships within the data. The RF model exhibited comparable results in terms of (R²) (0.839) and RMSE (0.328), suggesting competitive performance, though the GB model’s interpretability and stability favored its selection as the optimal classifier.

The feature-importance assessment yielded substantive insights into the variables contributing most strongly to the predictive performance of the fatality-count model (Figure 2). Geometric design characteristics emerged as the most influential predictors within the model structure, with the direction of horizontal alignment (0.0988), the radius of horizontal curvature (0.0962), and the urban–rural section classification (0.0938) exhibiting the highest contributions to classification accuracy. This ordering indicates that roadway alignment and the contextual operating environment provide substantial discriminative information for distinguishing among fatality count levels, underscoring the relevance of geometric–operational context in shaping the model’s internal decision structure. The prominence of alignment direction and curvature is consistent with established safety literature, although in this case the model’s importance ranking reflects their predictive value rather than their causal effect on crash outcomes.

Bicycle traffic (0.0805) also demonstrated considerable importance to the model’s predictive capability. The relatively high contribution of this variable suggests that segments with elevated bicycle exposure display distinctive patterns that assist the classifier in differentiating fatal-count categories. Within motorized traffic classes, total single-unit bus flow (0.0788), tractor-trailer activity (0.0744), and roadway category classification (0.0744) collectively exhibited strong predictive contributions. These findings indicate that heavy-vehicle activity and functional roadway hierarchy provide meaningful signals to the model regarding the likelihood of observing higher fatality counts, aligning with broader empirical trends but remaining interpretive rather than causal within this modeling framework.

Additional predictors—including the number of traffic lanes (0.0633), articulated bus traffic (0.0550), motorcycle and moped volumes (0.0479), and trailer-based truck flows (0.0463)—further enhanced model accuracy through their representation of cross-sectional characteristics, traffic heterogeneity, and vulnerable-road-user dynamics. Heavy-truck (0.0410) and light-truck traffic (0.0391) displayed moderate importance, suggesting that while they contribute meaningfully to the predictive task, their informational value is secondary relative to the higher-ranked geometric and heavy-vehicle variables.

Lower-ranking features, such as capacity utilization (0.0327) and slow-vehicle or agricultural-tractor traffic (0.0271), still contributed to model performance but with diminished influence. These variables likely capture localized effects associated with variability in traffic flow and speed dispersion. Medium-heavy two-axle truck volumes (0.0237) and track configuration (0.0168) displayed similarly modest predictive contributions. In contrast, vertical-alignment features—including slope or gradient, type of vertical curve, and vertical-curve radius—exhibited minimal importance, indicating that these attributes provided only limited discriminative value for the fatality-count classification task within the context of this roadway network.

Overall, the feature-importance analysis highlights that the model predominantly relies on horizontal geometric configuration, roadway environment classification, and the composition and intensity of mixed traffic flows to differentiate between fatality-count levels. The relevance of heavy-vehicle exposure and vulnerable-road-user presence is evident in their predictive contributions, reflecting their influence on the underlying data patterns.

4.1.2. Serious Injury Model

In the serious-injury category, the Random Forest model achieved optimal performance with n_estimators = 300, max_depth = 20, max_features = sqrt, min_samples_leaf = 3, and min_samples_split = 5. The best-performing Gradient Boosting model used learning_rate = 0.05, max_depth = 3, min_samples_split = 5, n_estimators = 200, and subsample = 0.9, while KNN again demonstrated the highest stability with n_neighbors = 7.

Table 6. Performance metrics of the predicting models for the counts of serious injury c.

	Train			Test
Metric	RF	KNN	GB	RF	KNN	GB
Accuracy	0.9503	0.9093	0.9109	0.9377	0.8815	0.8976
Precision	0.9559	0.9170	0.9155	0.9428	0.8858	0.9022
F1-Score	0.9512	0.9114	0.9104	0.9385	0.8830	0.8967
MCC	0.9438	0.8969	0.8989	0.9294	0.8648	0.8839
G-Mean	0.9714	0.9474	0.9483	0.9641	0.9309	0.9404
MSE	0.1057	0.4293	0.2364	0.1859	0.5784	0.2890
RMSE	0.3251	0.6552	0.4862	0.4311	0.7605	0.5376
R2	0.9899	0.9591	0.9775	0.9823	0.9449	0.9725

summarizes the comparative results across the training and testing phases. During training, the RF model achieved an accuracy of 0.950, precision of 0.956, and F1-score of 0.951, with a Matthews Correlation Coefficient (MCC) of 0.944 and G-Mean of 0.971, indicating strong predictive coherence across class distributions. Its generalization capability remained robust in the test phase, with accuracy (0.938), precision (0.943), and F1-score (0.939) all exceeding those of KNN (accuracy = 0.882) and nearly matching GB (accuracy = 0.898).

Error-based measures further substantiate the Random Forest model’s superior performance. The test mean squared error (MSE = 0.186) and root mean squared error (RMSE = 0.431) were markedly lower than those of KNN (MSE = 0.578, RMSE = 0.761), reflecting more precise estimation of serious-injury counts. The coefficient of determination (R²) for the RF model was 0.99 on the test set, demonstrating that nearly 99% of the variability in serious-injury crash frequency was explained by the selected predictors. This degree of explanatory power emphasizes the model’s effectiveness in capturing nonlinear and interaction effects between geometric design and traffic flow variables.

Although Gradient Boosting produced slightly higher training (R²) (0.978) and test (R²) (0.973) values, its marginal advantage came at the cost of increased computational intensity. Therefore, considering predictive fidelity, generalization stability, and interpretability, the Random Forest model was identified as the most effective classifier for serious-injury crash prediction.

The feature-importance evaluation derived from the Gradient Boosting model (Figure 3) indicates that serious-injury counts are predicted through a multifaceted combination of traffic composition, operational characteristics, and roadway geometry. The model identified heavy truck traffic (0.080) as the most influential contributor to predictive performance, suggesting that segments with substantial high-mass vehicle presence exhibit distinctive data patterns that aid the classifier in differentiating serious-injury count levels. Light truck traffic (0.075) and trailer-based truck flows (0.071) also provided strong predictive value, reflecting the structural relevance of freight-related vehicle activity within the model, rather than implying a direct causal relationship with injury severity.

Road category (0.070) demonstrated similarly high importance, indicating that roadway functional classification offers meaningful discriminative information regarding the operating environment in which serious-injury crashes occur. Variables of moderate influence—including articulated bus traffic (0.063), slow-vehicle and agricultural-tractor flows (0.063), direction of horizontal alignment (0.062), tractor-trailer with semitrailer activity (0.060), total single-unit bus traffic (0.059), and capacity utilization (0.058)—collectively highlight the informational role of vehicle heterogeneity, horizontal geometry, and demand-related operational conditions in enhancing the model’s classification accuracy.

Additional predictors such as medium-heavy two-axle truck volumes (0.056), section type (0.056), motorcycle and moped traffic (0.056), bicycle traffic (0.050), and the radius of horizontal curves (0.045) contributed moderately to the model structure. Their importance indicates that roadway configuration, vulnerable road-user exposure, and curvature characteristics provide supplementary discriminative cues for distinguishing serious-injury levels, consistent with their roles in shaping the variability observed in the dataset.

Variables associated with administrative attributes and basic geometric features—including the number of lanes (0.036), slope or gradient (0.023), track code (0.018), and vertical-alignment parameters—exhibited minimal importance. Their limited contribution suggests that these features provided comparatively weak informational value for predicting serious-injury counts within the context of the studied network, though this result reflects model-based discrimination rather than the intrinsic safety relevance of the variables.

Overall, the feature-importance patterns indicate that the Gradient Boosting model relies predominantly on freight-related traffic activity, horizontal-alignment characteristics, operational demand, and vulnerable road-user presence to differentiate between serious-injury count levels. Conversely, vertical-alignment features and foundational cross-sectional descriptors played a secondary role in the predictive structure. These findings offer model-based insights into the types of roadway and traffic characteristics that most effectively support the classification of serious-injury outcomes.

4.1.3. Slight Injury Model

For the slight-injury model, the Random Forest model achieved optimal performance with n_estimators = 300, max_depth = 25, max_features = sqrt, min_samples_leaf = 2, and min_samples_split = 5, whereas Gradient Boosting performed best with max_depth = 3, n_estimators = 200, learning_rate = 0.05, min_sample_split = 5, and subsample = 0.9. The KNN classifier maintained its consistency with an optimal configuration of n_neighbors = 7.

The comparative evaluation of the three algorithms (

Table 7. Performance metrics of the predicting models for the counts of slight injury.

	Train			Test
Metric	RF	KNN	GB	RF	KNN	GB
Accuracy	0.9706	0.9074	0.7425	0.9064	0.8391	0.7155
Precision	0.9761	0.9117	0.7362	0.9108	0.8374	0.7041
F1-Score	0.9718	0.9072	0.7294	0.9068	0.8352	0.7003
MCC	0.9683	0.8994	0.7210	0.8983	0.8252	0.6917
G-Mean	0.9839	0.9485	0.8515	0.9480	0.9093	0.8349
MSE	0.2103	1.0197	3.1294	0.6024	1.8831	3.3148
RMSE	0.4586	1.0098	1.7690	0.7761	1.3723	1.8207
R2	0.9824	0.9144	0.7374	0.9494	0.8419	0.7218

) highlights the Random Forest model’s notable ability to balance predictive accuracy and generalization stability. During training, RF achieved an accuracy of 0.971, precision of 0.976, and F1-score of 0.972, accompanied by a Matthews Correlation Coefficient (MCC) of 0.968 and a G-Mean of 0.984. These values reflect the model’s strong discriminatory power across both classes of slight-injury crash presence and absence.

When applied to the test dataset, the RF model maintained high generalization capability, with accuracy (0.906), precision (0.911), and F1-score (0.907). Its performance was consistently superior to both KNN (accuracy = 0.839) and GB (accuracy = 0.716), reaffirming the robustness of the ensemble-based approach in managing the inherent nonlinearity and multicollinearity present in the data.

In terms of goodness-of-fit measures, RF attained a test (R²) of 0.949, indicating that approximately 93% of the variance in slight-injury crash counts was explained by the model. This is a substantial predictive capability given the stochastic nature of crash occurrence data. The model also achieved relatively low error indicators (MSE = 0.602, RMSE = 0.776), markedly outperforming KNN (MSE = 1.883, RMSE = 1.372) and GB (MSE = 3.315, RMSE = 1.821). Collectively, these findings affirm the Random Forest model’s capacity to provide stable and interpretable predictions for the counts of slight-injury crashes, a category often characterized by higher data variability and less distinct patterning than fatal or serious-injury cases.

The feature-importance results derived from the Random Forest model (Figure 4) reveal that slight-injury counts are primarily differentiated by operational load indicators and traffic-composition variables, supplemented by contributions from horizontal geometric characteristics. Capacity utilization exhibited the highest importance value (0.087), indicating that variation in demand intensity and near-capacity flow conditions provides substantial discriminative information for predicting slight-injury levels. This prominence reflects the model’s sensitivity to operational states in which increased interaction frequency and reduced speed variance are characteristic of environments where lower-severity crash outcomes are more prevalent. It is important to note, however, that this relationship emerges from predictive structure rather than from a mechanistic or causal interpretation.

Traffic-composition attributes formed the next major tier of influential predictors. Total single-unit bus traffic (0.073), heavy truck volumes (0.071), light truck traffic (0.071), and medium-heavy two-axle truck flows (0.070) all contributed markedly to the model’s performance. Their importance suggests that roadway segments with pronounced heavy-vehicle presence exhibit distinguishable patterns in slight-injury crash occurrences, enabling the classifier to more accurately separate injury-count classes. Similar predictive value was observed for tractor-trailer with semitrailer activity (0.065), road-category classification (0.064), and truck-trailer or semitrailer traffic (0.063), underscoring the informational relevance of freight-vehicle heterogeneity and roadway functional hierarchy in describing the operational contexts associated with lower-severity outcomes.

Vulnerable-road-user variables, including bicycle traffic (0.059) and motorcycle or moped activity (0.057), displayed moderate yet meaningful importance. Their contribution reflects the classifier’s ability to extract discriminative patterns from exposure to unprotected users, even when such interactions tend to align with lower-energy collisions characteristic of slight-injury events. Horizontal geometric attributes—the radius of horizontal curves (0.059) and the direction of horizontal alignment (0.054)—also demonstrated notable predictive value. These results indicate that lateral curvature and alignment orientation continue to inform the classification process, though their relative importance is reduced compared with their influence in more severe crash models.

Operational and traffic-mix variables, such as articulated bus traffic (0.050) and slow-vehicle or agricultural-tractor flows (0.048), provided additional discriminative insight. Their importance suggests that speed differentials introduced by slow-moving vehicles contribute to the variability that the model uses to differentiate slight-injury levels. Roadway environment indicators—including section type (0.041), number of lanes (0.038), and track code (0.022)—offered moderate predictive contribution, highlighting the supporting role of cross-sectional configuration and administrative classification in the model’s internal decision structure. Slope or gradient displayed limited importance (0.009), consistent with the tendency of vertical grade to exert a comparatively minor influence on the classification of low-severity outcomes within this context.

Vertical alignment descriptors, including the radius and type of vertical curves, exhibited negligible importance. Their minimal predictive contribution suggests that vertical geometry provides little discriminative value for slight-injury count prediction compared with the dominant effects of operational demand, heavy-vehicle composition, vulnerable-user exposure, and horizontal curvature characteristics.

Overall, the feature-importance analysis demonstrates that the slight-injury model predominantly relies on operational conditions, freight-vehicle activity, and selected horizontal geometric factors to distinguish between injury-count categories.

4.2. Cross-Model Insights and Thematic Synthesis

• A comparative assessment of the three severity-specific models reveals several consistent predictive patterns, alongside notable differences in the variables contributing to the discrimination of fatal, serious-injury, and slight-injury count levels. Across all models, heavy-vehicle flows—including buses, tractor-trailers, and various truck classes—emerged as dominant predictors. Their recurrent importance reflects the strong discriminative value that freight-related and large-vehicle traffic volumes provide in distinguishing severity. These findings are consistent with previous studies indicating that heavy vehicles contribute disproportionately to crash severity due to their mass and kinetic energy, corroborating the predictive trends observed in European and North American road networks [69,70,71,72].
• Horizontal alignment features, particularly curve radii, and alignment direction, exhibited varying degrees of importance across severity levels. These features demonstrated higher predictive contribution in the slight-injury model and progressively lower contributions in the serious-injury and fatal-injury models. This gradient suggests that the informational value of geometric complexity differs across severity types, enabling the models to extract distinct patterns from horizontal-alignment variability. This observation is in agreement with prior empirical studies that report sharper curves and alignment irregularities as key contributors to minor injury crashes, whereas severe outcomes tend to be influenced more by traffic composition and operational context [73,74].
• Capacity utilization displayed a comparatively strong contribution in the slight-injury and serious-injury models, indicating that interactions linked to traffic demand and congestion offer meaningful predictive information for moderate-severity outcomes. This result highlights the relevance of dynamic operational states—captured indirectly through demand-related variables—in providing supplementary discriminatory signals beyond those embedded in static geometric or traffic-volume indicators [62,65,75].
• Additional observations show that vulnerable road users, including motorcycles, mopeds, and bicycles, are of moderate importance across all models, with a greater influence in fatal crashes. These patterns align with previous research indicating that crashes involving unprotected road users disproportionately contribute to fatalities, while functional road hierarchy is consistently associated with exposure-related severity trends [18,76]. Road category likewise appeared as a consistently informative predictor across severity levels, suggesting that functional classification encapsulates multiple contextual attributes—such as access control, expected speed environments, and modal interactions—that collectively enhance predictive accuracy [77,78,79,80]. Furthermore, segment-level counts of articulated buses and slow-moving or agricultural vehicles exhibited a notable importance in both Random Forest and Gradient Boosting models, despite their limited prevalence in the traffic flow. These findings indicate that, although rare, these vehicle classes carry substantial predictive information for severe crash counts. Ensemble methods can exploit such low-frequency yet informative patterns, where a small number of high-information splits significantly enhance predictive performance [81,82]. This behavior is consistent with previous traffic safety studies, which show that even relatively infrequent vehicle types can be involved in severe crashes when operating in particular contexts. For instance, articulated buses account for 39% of bus-involved collisions with cyclists in Germany, second only to city buses [83]. Their unique articulated design also poses challenges for autonomous vehicles: in March 2023, a Cruise self-driving car rear-ended an articulated municipal bus in San Francisco because the AI mispredicted the motion of the bus’s rear section, despite detecting it with sensors [84]. These cases demonstrate that rare or less common vehicle types can still contribute disproportionately to severe outcomes [85,86]. Similarly, studies of rural corridors and arterial systems in developing regions show that agricultural vehicles, although infrequent, frequently co-occur with roadway conditions predictive of high-severity crashes [87,88]. These insights collectively affirm the ability of ensemble models (RF, and GB) to disentangle nuanced interactions among geometric, operational, and vehicular factors in the prediction of crash severity counts.

4.3. Policy and Planning Implications

The findings offer actionable insights for traffic safety policymakers and transportation engineers. The consistent association of heavy vehicle flows with all injury severities points to the need for:

• Dedicated freight lanes and time-of-day restrictions for heavy vehicles, particularly along high-capacity corridors and primary arterials. Similar measures implemented in Germany and the Netherlands, through dedicated freight corridors have demonstrated measurable reductions in heavy-vehicle conflict points and crash rates.
• Enhanced driver visibility infrastructure, particularly at curves and intersections, particularly at curves and intersections, where line-of-sight limitations exacerbate heavy-vehicle crash risks. Such interventions have proven effective in improving driver response times and reducing side-impact collisions on European rural arterials.
• Vehicle-type-specific speed enforcement, which has shown positive outcomes in several European contexts. For instance, differentiated speed limits for heavy vehicles on rural expressways in Sweden and Finland have effectively reduced crash frequency and injury severity by mitigating speed dispersion between heavy and light vehicles.

Additionally, the strong influence of horizontal alignment supports geometric enhancements such as wider shoulders, super-elevation adjustments, and better signage at curves [89,90,91,92]. These design-oriented countermeasures are consistent with proven practices in European Vision Zero programs, where localized geometric corrections have contributed to substantial safety gains. The sensitivity of the prediction models to capacity utilization underscores the importance of real-time traffic management strategies, such as adaptive signal control, incident response systems, and dynamic lane allocation, to mitigate mid-severity crashes. Similar systems, implemented along European TEN-T (Trans-European Transport Network) corridors, have shown demonstrable reductions in mid-severity crashes by stabilizing flow conditions during congestion periods [93]. Collectively, the high performance of the Random Forest, and Gradient Boosting models indicates their practical utility as decision-support tools in proactive road safety management, by integrating these evidence-based interventions within Hungary’s strategic safety framework, policymakers can enable data-informed resource allocation, enhance network resilience, and achieve measurable reductions in crash severity across diverse roadway environments.

5. Conclusions

This study developed and evaluated a data-driven predictive framework for estimating segment-level road traffic injury counts on the Hungarian national road network using three supervised machine learning algorithms: Random Forest (RF), Gradient Boosting (GB), and K-Nearest Neighbors (KNN). The framework integrated a comprehensive set of explanatory variables encompassing roadway geometric design, traffic flow characteristics, and vehicle-type composition, while addressing class imbalance using the Synthetic Minority Over-sampling Technique (SMOTE) and optimizing model configurations through grid-search cross-validation.

For the prediction of serious- and slight-injury crash counts, the Random Forest (RF) ensemble model demonstrated the most robust performance, consistently attaining test accuracies above 0.91 and coefficient of determination (R²) values exceeding 0.95. In contrast, for fatalities count prediction, the Gradient Boosting (GB) model achieved the highest accuracy (0.95), with an R² value greater than 0.87. The feature importance analysis revealed consistent and interpretable patterns across severity levels. Heavy vehicle flows, including buses, and truck combinations, emerged as key predictors in all models, reaffirming their central role in shaping severe crash outcomes. Horizontal alignment features were most influential for fatal crashes, indicating that geometry governs crash occurrence, while capacity utilization was more prominent for slight and serious injuries, reflecting operational and congestion-related risk. Collectively, these findings highlight that traffic composition determines crash severity, geometry influences crash frequency, and operational factors modulate crash exposure dynamics.

From a practical standpoint, the results underscore the applicability of ensemble learning approaches as decision-support instruments for proactive road safety management. The models provide interpretable outputs that can inform infrastructure design, network monitoring, and safety audits, particularly for high-risk segments with elevated heavy vehicle exposure or geometric complexity.

Despite its robust performance, this study is subject to several limitations. First, the analysis relied on segment-level aggregated data, which constrains the capacity to capture driver-level behavioral heterogeneity, weather effects, or temporal dynamics. Second, while the Hungarian national road network provides a diverse dataset, the transferability of the model to other countries may be affected by contextual differences in traffic composition, enforcement practices, geometric design standards, and reporting accuracy.

Future research should aim to enhance model generalizability and spatial transferability by integrating multi-regional or cross-country datasets, enabling comparative validation across distinct roadway and traffic contexts. Incorporating temporal and environmental variables such as weather conditions, time-of-day, and seasonal patterns could further refine crash risk prediction accuracy. The integration of real-time traffic sensing data, such as probe vehicle trajectories or connected vehicle telemetry, would also allow for the development of dynamic crash risk forecasting systems. Moreover, the application of explainable methods, such as SHAP, is recommended to elucidate causal pathways and enhance model transparency for policy interpretation.

In conclusion, this study demonstrates that machine learning-based predictive models, particularly ensemble techniques, provide an effective, scalable, and interpretable foundation for understanding and mitigating road traffic injuries. The proposed framework contributes both methodologically and empirically to the advancement of data-driven road safety analytics, supporting the global pursuit of Vision Zero and the Sustainable Development Goal of reducing road fatalities and serious injuries through intelligent, evidence-based transport system design.