Optimizing Traffic Accident Severity Prediction with a Stacking Ensemble Framework

Abstract

Road traffic crashes (RTCs) have emerged as a major global cause of fatalities, with the number of accident-related deaths rising rapidly each day. To mitigate this issue, it is essential to develop early prediction methods that help drivers and riders understand accident statistics relevant to their region. These methods should consider key factors such as speed limits, compliance with traffic signs and signals, pedestrian crossings, right-of-way rules, weather conditions, driver negligence, fatigue, and the impact of excessive speed on RTC occurrences. Raising awareness of these factors enables individuals to exercise greater caution, thereby contributing to accident prevention. A promising approach to improving road traffic accident severity classification is the stacking ensemble method, which leverages multiple machine learning models. This technique addresses challenges such as imbalanced datasets and high-dimensional features by combining predictions from various base models into a meta-model, ultimately enhancing classification accuracy. The ensemble approach exploits the diverse strengths of different models, capturing multiple aspects of the data to improve predictive performance. The effectiveness of stacking depends on the careful selection of base models with complementary strengths, ensuring robust and reliable predictions. Additionally, advanced feature engineering and selection techniques can further optimize the model’s performance. Within the field of artificial intelligence, various machine learning (ML) techniques have been explored to support decision making in tackling RTC-related issues. These methods aim to generate precise reports and insights. However, the stacking method has demonstrated significantly superior performance compared to existing approaches, making it a valuable tool for improving road safety.

1. Introduction

Road traffic crashes (RTCs) represent a significant global health crisis, ranking as the fifth-leading cause of death worldwide, following cardiovascular diseases, cancers, respiratory diseases, and digestive diseases. Annually, approximately 1.3 million lives are lost in RTCs, with an additional 20 to 50 million individuals sustaining injuries, often leading to long-term disabilities. These incidents also impose substantial economic burdens on individuals, families, and nations.

The severity of this issue is underscored by statistics from the Moroccan Ministry of Equipment, Transport, Logistics, and Water (METLW) for 2018, which reported 96,133 accidents (+6.82%), including 3066 fatal accidents (−0.62%), resulting in 3485 deaths (−0.40%), 8725 serious injuries (−4.90%), and 128,249 minor injuries (+7.65%). Globally, RTCs account for nearly one in ten deaths, with varying degrees of severity [1,2,3,4,5,6,7,8,9]. Factors contributing to casualty severity are complex and multifaceted, encompassing road conditions [10], as well as driver-related attributes such as age and gender, and environmental factors like traffic congestion [11,12,13,14,15]. Consequently, researchers worldwide are actively developing prediction models and prevention strategies, with artificial intelligence (AI) playing a crucial role. This study focuses on predicting traffic accident severity using a three-class classification system: fatal, serious, and slight [16,17,18,19,20,21]. We utilize the TRAFFIC ACCIDENTS_2019_LEEDS (TAL19) dataset, comprising 7628 instances with 18 attributes each, sourced from the Leeds City Council ML directory. The dataset includes 88 fatal, 1336 serious, and 6204 slight accidents. Model performance is evaluated using a comprehensive suite of metrics, including the confusion matrix, false positive rate, true positive rate, accuracy, sensitivity, precision, recall, F1-score, root mean square error (RMSE), mean square error (MSE), mean absolute error (MAE), coefficient of determination, balanced accuracy, and receiver operating characteristic (ROC) analysis [22,23,24,25,26,27,28,29]. This research investigates the effectiveness of the stacking ensemble method, which combines the predictive power of multiple base models to enhance the classification of road traffic accident severity. Stacking has demonstrated superior performance compared to individual models like logistic regression, K-Nearest Neighbors, random forest, and support vector machines (SVMs), motivating its further exploration. This research demonstrates the potential of the stacking method for improving road traffic accident severity classification, offering a valuable tool for enhancing decision making and resource allocation in road safety initiatives. The structure of this article is as follows: Section 2 presents related work and the recent literature. Section 3 describes the dataset. Section 4 details the performance evaluation metrics. Section 5 presents the experimental results. Section 6 provides a discussion of the findings, and Section 7 concludes the paper.

2. Related Work

This section presents a review of the relevant literature concerning the classification of traffic accident severity. Several studies have been conducted using various learning models, with comparisons made to recent approaches in the field. In 2022, Md. Ebrahim Shaik et al. [3] provided an overview and summary of numerous neural network models for road traffic crashes (RTCs). They discussed future research directions and perspectives, including the use of the radial basis function (RBF), multilayer perceptron (MLP), and single-layer perceptron (SLP). Rezapour et al. [4] implemented multiple neural network models in 2020 to predict the intensity and frequency of RTCs. They utilized SLP, MLP, and recurrent neural network (RNN) models and evaluated their results. In 2016, Sameen et al. [5] estimated that traffic accidents would become the fifth-leading cause of death worldwide by 2030 based on a recent study. Chakraborty et al. [6] employed artificial neural network (ANN) models in 2019 to predict RTC severity. Their approach demonstrated higher accuracy and precision compared to conventional methods. Lee et al. and Ebrahim et al. conducted studies in 2020 and 2018, respectively [7,8], focusing on understanding and evaluating the importance of traffic accident severity and devising strategies to mitigate it. Behbahani et al. [9] developed an unsupervised learning method in 2018 using a radial basis function neural network (RBFNN), incorporating the least squares strategy and the recurrent least squares method (RLS). In 2017, Taamneh et al. [10] employed clustering-based classification of RTC severity, utilizing an ANN model and hierarchical clustering. Their approach aimed to improve decision making for severity mitigation compared to conventional methods. Behbahani et al. [11] proposed a novel approach in 2018, dividing the dataset into three clusters using the k-means method. They then applied MLP to predict RTC severity and improve accuracy. Sheng Dong et al. [12] analyzed and predicted RTCs in 2022 using an ensemble learning model with explanations such as Additive and Shapley. The motivation for enhancing road traffic accident severity classification using the stacking method stems from the critical need for accurate accident severity prediction, which plays a vital role in emergency response, resource allocation, and policy planning. While several methods have been proposed in the literature, existing limitations must be addressed. For instance, traditional models like logistic regression may struggle to capture complex nonlinear relationships in the data, while K-Nearest Neighbors may be sensitive to the choice of k and computationally expensive for large datasets. Random forest, although powerful, may not generalize well to unseen data, and SVMs can be highly sensitive to hyperparameter tuning. Considering these existing studies, the stacking method presents a novel approach to addressing the limitations of traditional models while exploring the benefits of combining multiple models for road traffic accident severity classification. By systematically evaluating and comparing the stacking method against traditional models such as logistic regression, K-Nearest Neighbors, random forest, and SVMs, this research aims to demonstrate its superior performance and provide a robust framework for accurate accident severity prediction [28,29,30,31,32,33,34].

Table 1. This table summarizes the methodologies, results, and limitations of related works.

SL	A Summary of Related Works	Techniques and Methods Applied	Problem and Approach Attacked	Obtained Result	Limits of Related Works and Perspective
1.	Md. Ebrahim Shaik et al. [3]	SLP, MLP, RBF, SLP	Prediction of RTC severity	Decision aids in RTC prediction	Needs experimental validations
2.	Rezapour et al. [4]	SLP, MLP, RNN	Prediction of intensity and frequency of motorbike accidents	Decision aids in motorbike conduct	Needs to compare their confusion matrices and ROC
3.	Sameen et al. [5]	Deep learning approach (RNN and CNN)	Prediction of road traffic accidents for road safety assessment	Estimation of RTC and predicts it to be the fifth-leading cause of death worldwide in 2030	Classification accuracy requires improvements
4.	Chakraborty et al. [6]	ANN, MLP	Predict traffic accident RTC and the severity of transport	Prediction in injuries and deaths	Data normalization, standardization, and transformation
5.	Lee et al. [7]	Machine learning algorithms	Anticipation of RTC severity in rainy seasons	Decision aids in RTC severity	Anticipation of RTC severity in spring, fall, and winter seasons
6.	Taamneh et al. [8]	ANN	Severity prediction of traffic accidents	Prediction accuracy	Needs to compare their confusion matrices and ROC
7.	Behbahani et al. [9]	Implement an activation function RBF, and compare it with FINN and RBFNN using a radial function	Anticipation of the frequency of RTCs	Prediction frequency	Needs to compare their confusion matrices
8.	Dong et al. [10]	Deep learning, FFNN	Traffic crash prediction	Prediction accuracy	Needs to compare their confusion matrices and ROC
9.	Behbahani et al. [11]	MLP	Prediction of traffic accidents severity	Prediction accuracy	Needs to compare their confusion matrices and ROC
10.	Sheng Dong et al. [12]	Ensemble learning based on boosting models such as SHAPley and exPlanations	Analyzing and predicting RTC severity	Prediction accuracy	Needs optimization model

provides a summary of the related works from the literature on traffic accident severity; the reviewed studies on traffic accident severity classification employ various machine learning and deep learning techniques, including neural networks (ANN, RNN, CNN), ensemble methods, and activation functions like RBF.

Figure 1 illustrates a detailed block diagram of the stacking method for traffic accident severity classification. It begins with data collection and preprocessing, followed by data distribution analysis, which reveals class imbalances (slight, serious, fatal). To address this, resampling techniques are applied to balance the dataset before training. The methodology incorporates multiple machine learning models, including random forest, logistic regression, SVM, and K-Nearest Neighbors, combined through a stacking method to enhance classification performance. The best-performing model is then selected for further refinement. Feature analysis using explainable machine learning identifies key factors influencing accident severity, and the models are re-trained using these selected features to improve accuracy and interpretability. This comprehensive approach ensures better prediction performance by handling data imbalance, optimizing model selection, and enhancing explainability.

3. Details of the Dataset

3.1. Dataset Description

The dataset used in this study is the Traffic Accidents 2019 Leeds (TAL19) dataset, which serves as the primary source of information for road traffic accident severity classification. This section will discuss the details of the dataset, including its characteristics, variables, and data preprocessing steps. The TAL19 dataset comprises a comprehensive collection of road traffic accidents that occurred in Leeds during the year 2019. It contains a diverse range of accidents, including different severity levels and various contributing factors. The dataset consists of both numerical and categorical variables, providing valuable information for accident severity classification. The dataset includes several key variables, such as the following:

Accident ID: a unique identifier for each accident in the dataset.

Accident Date and Time: the date and time at which the accident occurred.

Accident Severity: the severity level of the accident, categorized as minor, serious, or fatal.

Location: the geographical location of the accident, represented by coordinates or street addresses.

Weather Conditions: describes the prevailing weather conditions at the time of the accident.

Road Surface Conditions: specifies the condition of the road surface during the accident, such as dry, wet, icy, etc.

Contributing Factors: indicates the factors contributing to the accident, including driver behavior, road conditions, and vehicle-related issues. Before applying machine learning models and the stacking method, the TAL19 dataset requires preprocessing steps to ensure data quality and compatibility. This may involve handling missing values, encoding categorical variables, normalizing numerical features, and addressing class imbalance if present. The dataset can then be split into training and testing sets for model development and evaluation.

The TRAFFIC ACCIDENTS_2019_LEEDS (TAL19) dataset employed for this empirical analysis was downloaded from the traffic accidents in the Leeds City Council machine learning repository. This dataset contains a total of 7628 instances, and each instance consists of 18 attributes. There are 88 fatal, 1336 serious, and 6204 slight instances. The brief dataset feature information is shown in

Table 2. Encoding RTC methods and features used.

SN	Attributes	Abbreviation
1	Reference Number	Reference_Number
2	Grid Ref: Easting	Easting
3	Grid Ref: Northing	Northing
4	Number of Vehicles	Number_of_Vehicles
5	Accident Date	Accident_Date
6	Time (24 h)	Time
7	1st Road Class	1_Road Class
8	1st Road Class & No	1st_Road_Class_No
9	Road Surface	Road_Surface
10	Local Authority	Local_Authority
11	Type of Vehicle	Type_of_Vehicle
12	Road Surface	Road_Surface
13	Lighting Conditions	Lighting_Conditions
14	Weather Conditions	Weather_Conditions
15	Age of Casualty	Age
16	Type of Vehicle	Type_Vehicle
17	Sex of Casualty	Sex
18	Casualty Severity: 1, Fatal; 2, Serious; 3, Slight	Casualty_Severity Class

. All the instances in the dataset with the absence of TAL19 were taken as the slight class, and the instances with the existence of TAL19 were taken as the serious class for investigation purposes. It is evident from Figs. 3,4,5 that there exists a strong degree of correlation among all the 18 attributes for the slight as well as the serious or fatal class. Among all the attributes, ‘Lighting_Condition’ and ‘Weather_Conditions’ have the highest degree of correlation (0.65) between them for the TAL19 absence class.

3.2. Machine Learning Techniques

3.2.1. Logistic Regression

Logistic regression is a classification algorithm used to model the relationship between input features and the probability of an event occurring. It is based on the logistic function and is widely used for binary classification tasks. However, in the context of the Traffic Accidents 2019 Leeds dataset, the stacking method has been found to outperform logistic regression in terms of classification accuracy and predictive power.

3.2.2. K-Nearest Neighbors

KNN is a non-parametric classification algorithm that assigns a class label to an instance based on its proximity to the K-Nearest Neighbors in the training dataset. KNN relies on the assumption that instances with similar features tend to belong to the same class. However, the stacking method has demonstrated superior performance compared to KNN in the classification of road traffic accident severity in the Traffic Accidents 2019 Leeds dataset.

3.2.3. Random Forest

Random forest is an ensemble learning method that constructs multiple decision trees and combines their predictions through voting or averaging. It is known for its robustness against overfitting and its ability to handle high-dimensional datasets. Despite its strengths, the stacking method has been shown to outperform random forest in accurately classifying the severity of road traffic accidents in the Traffic Accidents 2019 Leeds dataset.

3.2.4. Support Vector Machines

SVM is a supervised learning algorithm that aims to find an optimal hyperplane that separates instances of different classes. It works well for linearly separable data and can be extended to nonlinear classification using kernel functions. However, the stacking method has been found to outperform SVM in the classification of accident severity, suggesting that the combination of multiple models in the stacking ensemble leads to better results. The significance of the stacking method lies in its ability to leverage the strengths of multiple base models and create a meta-model that improves classification performance. By combining the predictions of different models, the stacking method can capture diverse patterns and nuances present in the Traffic Accidents 2019 Leeds dataset, leading to enhanced accuracy and predictive power for accident severity classification.

3.2.5. Stacking Method

The stacking method is an ensemble learning technique that aims to improve classification performance by combining the predictions of multiple base models. It involves training several individual models on the same dataset and then utilizing a meta-model, often referred to as a “stacker” or a “blender”, to learn how to best combine the predictions of these base models. The stacking method is particularly relevant in tackling road traffic accidents (RTCs) because accident severity classification is a complex task that often requires capturing various factors and patterns. By leveraging the strengths of different models and combining their predictions, the stacking method can enhance the overall predictive power and accuracy in RTC severity classification. The process of implementing the stacking method involves the following steps:

Dataset Split: The available data are typically divided into a training set and a holdout set. The training set is used to train the base models and create a new dataset for the stacker, while the holdout set is used to evaluate the performance of the final stacked model.

Base Model Training: Several different base models are trained using the training set. These models can be diverse in nature, such as logistic regression, random forest, support vector machines, or neural networks. Each model learns to make predictions based on the input features and the known accident severity labels.

Stacker Training: The predictions from the base models are then used as input features to train the stacker model. This meta-model learns to combine the predictions of the base models and generate a final prediction for accident severity classification. The stacker can be any classification algorithm, such as logistic regression or random forest.

Prediction: Once the stacker is trained, it can be used to make predictions on new, unseen instances. The stacker takes the predictions of the base models as input and generates the final prediction for accident severity.

The stacking method is relevant in tackling RTCs because it leverages the diversity and complementary strengths of multiple models. Different base models may excel in capturing specific patterns or relationships in the data, and by combining their predictions, the stacking method can provide more accurate and robust predictions for accident severity classification.

By employing the stacking method, researchers and practitioners can benefit from improved accuracy in predicting accident severity, leading to better decision making in emergency response, resource allocation, and policy planning. The stacking method’s ability to harness the collective intelligence of multiple models makes it a valuable tool in tackling the complexities of RTC severity classification.

4. Metrics of Evaluation for the Performance of Different Learning Models

Several metrics can be utilized to assess the classification accuracy and predictive power of different learning models for the TAL19 dataset [24,25]. The most common ones are summarized in

Table 3. Confusion matrix.

	Predicted Positive	Predicted Negative
Actual Positive	True Positive (TP): The actual class is positive, and the model correctly predicts it as positive.	False Negative (FN): The actual class is positive, but the model incorrectly predicts it as negative.
Actual Negative	False Positive (FP): The actual class is negative, but the model incorrectly predicts it as positive.	True Negative (TN): The actual class is negative, and the model correctly predicts it as negative.

and Equations (1)–(8).

The precision represents the quality can be calculated in Equation (4) as

$$Precision={\displaystyle \frac{True positive}{True positive+False positive}}$$

(1)

$$Recall={\displaystyle \frac{True positive}{True positive+False negative}}$$

(2)

$$F-Score={\displaystyle \frac{2\times \left(precision\times recall\right)}{precision+recall}}$$

(3)

$$Accuracy={\displaystyle \frac{True positive+True negative}{True positive+True negative+False positive+False negative}}$$

(4)

$$RMSE=\sqrt{{\displaystyle \frac{{\sum }_{cϵC}{\sum }_{iϵTtestC}{\left(reco\left(C,i\right)-r_{C,i}\right)}^2}{{\sum }_{cϵC}{Itest}_C}}}$$

(5)

$$MSE={\displaystyle \frac{1}{N}}\sum _{i=1}{\left(Y-\widehat{Y}\right)}^2$$

(6)

$$MAE=\sqrt{{\displaystyle \frac{{\sum }_{cϵC}{\sum }_{iϵItestC}\left|reco\left(C,i\right)-r_{C,i}\right|}{{\sum }_{cϵC}{Itest}_C}}}$$

(7)

$$BA=C\times {\displaystyle \frac{True Positive}{P}}+{\displaystyle \frac{True Negative}{N}}$$

(8)

5. Experiments

Table 4. Report on linear correlation for the fatal class.

	Reference Number	Grid Ref: Easting	Grid Ref: Northing	Number of Vehicles	Accident Date	Time (24 h)	1st Road Class	1st Road Class & No	Road Surface	Lighting Conditions	Weather Conditions	Vehicle Number	Type of Vehicle	Casualty Class	Sex of Casualty	Age of Casualty
Reference Number	1	0.01	0.06	0.12	0.07	0.03	0.10	0.12	0.38	0.29	0.48	0.13	0.34	0.30	0.25	0.08
Grid Ref: Easting	0.01	1	0.13	0.17	0.35	0.03	0.18	0.00	0.06	0.00	0.35	0.04	0.25	0.38	0.24	0.39
Grid Ref: Northing	0.06	0.13	1	0.28	0.25	0.17	0.20	0.04	0.11	0.07	0.23	0.20	0.07	0.08	0.11	0.51
Number of Vehicles	0.12	0.17	0.28	1	0.11	0.30	0.19	0.50	0.37	0.44	0.24	0.10	0.06	0.30	0.17	0.22
Accident Date	0.07	0.35	0.25	0.11	1	0.02	0.29	0.01	0.20	0.05	0.45	0.33	0.40	0.28	0.14	0.24
Time (24 h)	0.03	0.03	0.17	0.30	0.02	1	0.41	0.18	0.05	0.27	0.11	0.10	0.07	0.30	0.16	0.25
1st Road Class	0.10	0.18	0.20	0.19	0.29	0.41	1	0.46	0.16	0.34	0.15	0.02	0.10	0.12	0.10	0.13
1st Road Class & No	0.12	0.00	0.04	0.50	0.01	0.18	0.46	1	0.23	0.32	0.35	0.22	0.21	0.39	0.06	0.12
Road Surface	0.38	0.06	0.11	0.37	0.20	0.05	0.16	0.23	1	0.03	0.49	0.36	0.23	0.22	0.11	0.08
Lighting Conditions	0.29	0.00	0.07	0.44	0.05	0.27	0.34	0.32	0.03	1	0.55	0.06	0.22	0.07	0.21	0.22
Weather Conditions	0.48	0.35	0.23	0.24	0.45	0.11	0.15	0.35	0.49	0.55	1	0.18	0.41	0.44	0.07	0.29
Vehicle Number	0.13	0.04	0.20	0.10	0.33	0.10	0.02	0.22	0.36	0.06	0.18	1	0.42	0.41	0.08	0.07
Type of Vehicle	0.34	0.25	0.07	0.06	0.40	0.07	0.10	0.21	0.23	0.22	0.41	0.42	1	0.76	0.33	0.47
Casualty Class	0.30	0.38	0.08	0.30	0.28	0.30	0.12	0.39	0.22	0.07	0.44	0.41	0.76	1	0.14	0.35
Sex of Casualty	0.25	0.24	0.11	0.17	0.14	0.16	0.10	0.06	0.11	0.21	0.07	0.08	0.33	0.14	1	0.70
Age of Casualty	0.08	0.39	0.51	0.22	0.24	0.25	0.13	0.12	0.08	0.22	0.29	0.07	0.47	0.35	0.70	1

and Figure 2 show strong correlation among all 13 attributes and both the fatal class and the others class. Notably, the attributes ‘Lighting Conditions’ and ‘Weather Conditions’ exhibit the highest degree of correlation within the fatal class. The visual representation in Figure 2, specifically the upper bars and positive values in Table 3, provides evidence of this high correlation among most attributes, with a few exceptions like ‘vehicle number’, ‘northing’, and ‘accident date’ that show relatively low values.

Furthermore, in the case of the serious class, it is apparent from Figure 3 and Table 4 that a significant correlation exists among all 11 attributes and both the serious class and the others class within the TAL19 dataset. Among these attributes, ‘Lighting Conditions’ and ‘time’ exhibit the highest degree of correlation. The visual representation in Figure 3, including the upper bars and positive values in Table 4, indicates a notable correlation among the responsible attributes, with only a few attributes displaying relatively small values.

Lastly, in the case of the slight class, it is evident from Figure 4 and

Table 5. Report on linear correlation for the serious class.

	Grid Ref: Easting	Grid Ref: Northing	Number of Vehicles	Time (24 h)	1st Road Class	Road Surface	Lighting Conditions	Weather Conditions	Vehicle Number	Type of Vehicle	Casualty Severity	Sex of Casualty	Age of Casualty
Grid Ref: Easting	1.00	0.13	0.17	0.03	0.18	0.06	0.00	0.35	0.04	0.25	0.38	0.24	0.39
Grid Ref: Northing	0.13	1.00	0.28	0.17	0.20	0.11	0.07	0.23	0.20	0.07	0.08	0.11	0.51
Number of Vehicles	0.17	0.28	1.00	0.30	0.19	0.37	0.44	0.24	0.10	0.06	0.30	0.17	0.22
Time (24 h)	0.03	0.17	0.30	1.00	0.41	0.05	0.27	0.11	0.10	0.07	0.30	0.16	0.25
1st Road Class	0.18	0.20	0.19	0.41	1.00	0.16	0.34	0.15	0.02	0.10	0.12	0.10	0.13
Road Surface	0.06	0.11	0.37	0.05	0.16	1.00	0.03	0.49	0.36	0.23	0.22	0.11	0.08
Lighting Conditions	0.00	0.07	0.44	0.27	0.34	0.03	1.00	0.55	0.06	0.22	0.07	0.21	0.22
Weather Conditions	0.35	0.23	0.24	0.11	0.15	0.49	0.55	1.00	0.18	0.41	0.44	0.07	0.29
Vehicle Number	0.04	0.20	0.10	0.10	0.02	0.36	0.06	0.18	1.00	0.42	0.41	0.08	0.07
Type of Vehicle	0.25	0.07	0.06	0.07	0.10	0.23	0.22	0.41	0.42	1.00	0.76	0.33	0.47
Casualty Severity	0.38	0.08	0.30	0.30	0.12	0.22	0.07	0.44	0.41	0.76	1.00	0.14	0.35
Sex of Casualty	0.24	0.11	0.17	0.16	0.10	0.11	0.21	0.07	0.08	0.33	0.14	1.00	0.70
Age of Casualty	0.39	0.51	0.22	0.25	0.13	0.08	0.22	0.29	0.07	0.47	0.35	0.70	1.00

that a strong correlation exists among all nine attributes and both the slight class and the others class within the TAL19 dataset. Among these attributes, ‘Road Surface’ and ‘Water Condition’ demonstrate the highest degree of correlation. Conversely, certain attributes such as ‘vehicle number’, ‘time’, and ‘grid ref easting’ exhibit a lower degree of correlation, although the extent of this decrease in correlation is minimal.

The experiments were conducted to predict the severity of road traffic crashes (RTCs) in the transport and logistics sector using the ACCIDENTS_2019_LEEDS dataset. Various classifiers were employed, with the dataset randomly split into 80% for training and 20% for testing. The sampling methods utilized included tenfold cross-validation, stratified shuffle split, and random sampling.

Based on the results presented in Table 5,

Table 6. Report on linear correlation for the slight class.

	Reference Number	Grid Ref: Easting	Grid Ref: Northing	Number of Vehicles	Accident Date	Time (24 h)	1st Road Class	1st Road Class & No	Road Surface	Lighting Conditions	Weather Conditions	Vehicle Number	Type of Vehicle	Casualty Class	Sex of Casualty	Age of Casualty
Reference Number	1	0.01	0.05	0.02	0.06	0.02	0.07	0.03	0.21	0.10	0.07	0.00	0.03	0.02	0.07	0.01
Grid Ref: Easting	0.01	1	0.00	0.14	0.00	0.01	0.15	0.09	0.07	0.04	0.07	0.09	0.09	0.02	0.03	0.05
Grid Ref: Northing	0.05	0.00	1	0.07	0.02	0.01	0.22	0.13	0.03	0.04	0.03	0.06	0.06	0.03	0.01	0.02
Number of Vehicles	0.02	0.14	0.07	1	0.01	0.01	0.31	0.21	0.02	0.07	0.05	0.57	0.01	0.43	0.01	0.11
Accident Date	0.06	0.00	0.02	0.01	1	0.01	0.00	0.02	0.02	0.00	0.08	0.01	0.04	0.01	0.04	0.01
Time (24 h)	0.02	0.01	0.01	0.01	0.01	1	0.06	0.05	0.04	0.19	0.03	0.01	0.05	0.02	0.00	0.04
1st Road Class	0.07	0.15	0.22	0.31	0.00	0.06	1	0.60	0.04	0.03	0.05	0.16	0.14	0.22	0.02	0.11
1st Road Class & No	0.03	0.09	0.13	0.21	0.02	0.05	0.60	1	0.06	0.04	0.09	0.12	0.04	0.13	0.03	0.09
Road Surface	0.21	0.07	0.03	0.02	0.02	0.04	0.04	0.06	1	0.14	0.42	0.09	0.04	0.01	0.00	0.01
Lighting Conditions	0.10	0.04	0.04	0.07	0.00	0.19	0.03	0.04	0.14	1	0.08	0.09	0.02	0.05	0.04	0.04
Weather Conditions	0.07	0.07	0.03	0.05	0.08	0.03	0.05	0.09	0.42	0.08	1	0.04	0.02	0.00	0.06	0.00
Vehicle Number	0.00	0.09	0.06	0.57	0.01	0.01	0.16	0.12	0.09	0.09	0.04	1	0.10	0.35	0.03	0.09
Type of Vehicle	0.03	0.09	0.06	0.01	0.04	0.05	0.14	0.04	0.04	0.02	0.02	0.10	1	0.27	0.18	0.10
Casualty Class	0.02	0.02	0.03	0.43	0.01	0.02	0.22	0.13	0.01	0.05	0.00	0.35	0.27	1	0.13	0.20
Sex of Casualty	0.07	0.03	0.01	0.01	0.04	0.00	0.02	0.03	0.00	0.04	0.06	0.03	0.18	0.13	1	0.02
Age of Casualty	0.01	0.05	0.02	0.11	0.01	0.04	0.11	0.09	0.01	0.04	0.00	0.09	0.10	0.20	0.02	1

,

Table 7. Performance of the staking model.

	Precision	Recall	F1-Score	Support
Fatal	61%	18%	29%	32
Serious	27%	35%	41%	442
Slight	22%	47%	32%	17,773
Accuracy			81%	2247
Macro avg	49%	34%	32%	2247
Weighted avg	76%	79%	71%	2247

,

Table 8. Performance of the Logistic Regression model.

	Precision	Recall	F1-Score	Support
Fatal	66%	10%	29%	32
Serious	16%	17%	41%	442
Slight	20%	83%	30%	17,773
Accuracy			81%	2247
Macro avg	70%	49%	54%	2247
Weighted avg	79%	81%	79%	2247

,

Table 9. Performance of the KNN model.

	Precision	Recall	F1-Score	Support
Fatal	75%	53%	55%	32
Serious	15%	37%	14%	442
Slight	10%	20%	31%	17,773
Accuracy			88%	2247
Macro avg	89%	67%	74%	2247
Weighted avg	89%	89%	87%	2247

and

Table 10. Performance of the Random Forest model.

	Precision	Recall	F1-Score	Support
Fatal	76%	51%	44%	32
Serious	15%	40%	35%	442
Slight	9%	9%	31%	17,773
Accuracy			87%	2247
Macro avg	59%	68%	73%	2247
Weighted avg	81%	79%	70%	2247

and Figure 5, Figure 6, Figure 7, Figure 8, Figure 9 and Figure 10, the proposed classifiers exhibited favorable performance. The F1-score, which combines precision and recall ratios, was used as a measure of model ability.

Figure 7, Figure 8 and Figure 9, as well as Table 7, Table 8, Table 9, Table 10,

Table 11. Performance of the SVM model.

	Precision	Recall	F1-Score	Support
Fatal	56%	48%	34%	32
Serious	15%	12%	16%	442
Slight	29%	40%	50%	17,773
Accuracy			87%	2247
Macro avg	59%	68%	73%	2247
Weighted avg	81%	79%	70%	2247

and

Table 12. Performance of the Bayesian model.

	Precision	Recall	F1-Score	Support
Fatal	98%	88%	54%	32
Serious	91%	89%	63%	442
Slight	87%	98%	93%	17,773
Accuracy			88%	2247
Macro avg	92%	61%	77%	2247
Weighted avg	88%	88%	86%	2247

, provide valuable insight into the precision, recall, F1-score, fscore, and report rate offered by the improved classifiers. Similarly, the mean absolute error (MAE) and root mean square error (RMSE) performances are illustrated in Figure 10. The performance of the stacking model is reported in

Table 13. Performance of stacking model report.

	MSE	MAE	RMSE
Bayes	0.279039	0.252336	0.528241
SVM	0.253227	0.224744	0.503216
LR	0.249221	0.220739	0.464589
RF	0.215843	0.197152	0.464586
Stacking	0.133066	0.125946	0.364783
KNN	0.141077	0.125946	0.375602

.

6. Discussion

Road traffic crashes are a major global concern, requiring advanced prediction methods for effective mitigation. The results in Table 13 show that traditional models like Bayesian regression, SVM, and linear regression have relatively high error rates, indicating their limited ability to capture complex patterns in the data. Random forest (RF) improves performance due to its ensemble nature, reducing errors compared to individual models. The K-Nearest Neighbors (KNN) model further lowers the error, but its performance is slightly inferior to stacking. The stacking model achieves the best overall results (MSE: 0.133066, RMSE: 0.364783), significantly outperforming all other models.

Table 14. Summary of previous studies that use RNN for traffic accident prediction.

Authors and Study Area	Input/Independent Variable	Output/Dependent Variable	Data Partitioning	Performances (%)	Severity Level
Md. Ebrahim Shaik (2021) [3]	Accident time, zone and location, collision type	Injury severity	Training = 80%, Validation = 20%	Accuracy of the RNN model was 71.77%, whereas the MLP and BLR models achieved 65.48% and 58.30%, respectively	Summarizes the different models
Rezapour, M., Nazneen, S., Ksaibati, K., (2020) [4]	2430 motorcycle crashes in a mountainous area in the United States over a 10-year period	Injury severity	Training = 80%, Validation = 20%	AUC ranges in value from 0 (100% wrong) vs. 1 (100% right)	Prediction of motorcycle crashes
Sameen et al. (2019), Malaysia [5]	Accident time, zone and location, collision type, surface and lighting condition, accident reporting	Injury severities	10-fold cross-validation	SD: RNN = 1.24 CNN = 0.53 FFNN = 2.21 Accuracy: RNN = 73.76 CNN = 70.30 FFNN = 68.79	PDO = 238 (last section), 209 (main route) Evident injury = 58 (last section), 155 (main route) Disabling injury = 82 (last section), 666 (main route)
Abhishek Chakraborty (2019) [6]	Pedestrian–vehicular interaction concerning ‘pedestrian-vehicular volume ratio’ and lack of ‘accessibility of pedestrian cross-walk’	Accidents, traffic/mortality	Training = 80%, Testing = 20%		Accidents, traffic/mortality
Lee et al., 2019 [7]	Road geometry data, precipitation data, and traffic accident data over nine years corresponding to the Naebu Expressway, which is located in Seoul, Republic of Korea	Severity of traffic accidents in Seoul City	Training = 75%, Testing = 25%	Accuracy of 1.6878, followed by curve length (CL) at 1.1213	Vehicle type (VT) showing a decrease of −1.2282, accident time (AT) at −2.9598, and super-elevation (Se) having the most negative impact at −3.8938.
Our Study	Traffic Accidents 2019 Leeds (TAL19) dataset	Slight, serious, fatal	Training = 80%, Testing = 20%	Precision = 98%, Recall = 60%, Score = 70%	Slight, serious, fatal, MSE = 0.133066, MAE = 0.125946, RMSE= 0.364783

depicts the previous study of road accident prediction based on deep learning techniques such as RNN and CNN. Some research has shown that the accuracy of the various machine learning/deep learning approaches is greatly affected by different hyperparameters, so it is necessary to change the variables of the models before presenting the output.

7. Conclusions

Road traffic crashes pose a significant risk to the transport and logistics sector and contribute to a considerable number of accidents. Predicting these crashes at an early stage can not only help reduce the cost of fatalities but also save lives. Therefore, the development of a robust prediction system is of utmost importance. In this study, we conducted an experimental setup to evaluate seven different supervised machine learning techniques in terms of precision, accuracy, recall/sensitivity/true positive rate, specificity, negative predictive value, false positive rate (FPR), F1-score, ROC curve, and rate of misclassification. The goal was to identify the most suitable learning model for predicting road traffic crashes. The experimental results demonstrated the effectiveness of the proposed stacking method, achieving a maximum accuracy rate. The findings highlight the advantages of the stacking method in predicting road traffic crashes and emphasize its potential for practical implementation. Future research will focus on developing improved classification models by incorporating ensemble learning and collaborative learning approaches, along with metaheuristic optimization techniques. This approach aims to enhance the overall performance and accuracy of the prediction system. In conclusion, this study has made significant progress in the prediction of road traffic crashes, providing valuable insights into the performance of different learning models. The proposed stacking method demonstrates superior performance compared to conventional approaches, paving the way for the development of more effective and efficient prediction systems in the future.