A hybrid GA-ANN machine learning with novel DBSTLink data balancing approach for driving action detection: solution to road crash prediction

Introduction

Road accidents pose a crucial dilemma for public health and safety and are continually increasing. Accidents are recognized as one of the major concerns, as every year road traffic accidents cause the deaths of around 1.19 million people, according to the World Health Organization reports (WHO, 2023). Accidents cause non-fatal injuries to an extra 20 to 50 million individuals, resulting in long-term disability. Research has proven that the influence of driver actions on vehicle kinematics significantly affects road crashes. According to Lu et al. (2020), well over 1 million people worldwide lose their lives in traffic accidents every year, and well over 50 million suffer injuries. Most crashes are caused by human factors, according to an analysis of the factors that cause them, which includes variables affecting the vehicle itself, the roadway, or a combination of these variables (Rahman et al., 2022).

Many researchers have conducted investigations to determine state-of-the-art road safety analysis and have developed systems to enhance road safety. For road safety research, analysis of driver behavior (DB) (Albert et al., 2018; Gupta, Choudhary & Parida, 2021) should be a major concern as traffic accidents are often caused by human factors (Luk et al., 2017; Wang et al., 2022). Researchers have utilized driving performance measures to explore the impact of drivers’ behavioral traits on driving quality, presuming that deviations from standard driver actions signify hazardous circumstances (Yadav & Velaga, 2019; Choudhary & Velaga, 2019). According to Osman et al. (2019), they used machine learning techniques such as K-nearest neighbor (KNN), random forest (RF), support vector machine (SVM), decision trees (DT), Gaussian naïve Bayes (Gaussian NB), and adaptive boost (AdaBoost) to predict close crashes based on data from inside the vehicle. Artificial neural networks (ANN) and decision trees (DT) were utilized by Zouhair, Mousannif & Al Moatassime (2020) to analyze crash events and assess the effects of weather on traffic safety. Researchers (Lee & Li, 2015) use the boosted regression tree to predict the severity of driver injuries. However, they noted a drawback when using ensemble approaches to leverage the advantages of decision trees. In parallel, Parsa et al. (2020) applied extreme gradient boosting (XGBoost) for real-time accident detection and feature analysis, with the aim of improving highway safety. Their study demonstrated the potential of these methods for identifying key factors contributing to accidents.

Applications of machine learning (ML) models in finding DA are still limited, but more research should be done by the researcher to find out the major DA that causes traffic accidents. The literature, which includes the aforementioned studies and others, clarifies it that the effectiveness of produced models depends on the datasets and data mining algorithms used, with these factors varying between investigations. The RF method was utilized by Mafi, Abdelrazig & Doczy (2018) to analyze gap acceptance behavior in a driving simulation approach. Road accidents are the result of a complex interaction of variables involving the driver, the vehicle, and the environment. Road accidents involve a complicated human-machine system, as shown by Aljanahi, Rhodes & Metcalfe (1999). According to Treat et al. (1979), who ascribed around 90% of accidents to driver errors, driver actions play a significant role. Although research has mostly focused on vehicle data, it is essential to include driver response elements in prediction models. We want to find out how well supervised learning models, more specifically hybrid machine learning approaches, can predict crashes by using variables from a driving simulator database that show what the driver did.

The study by Martinez et al. (2017) investigates how a driver’s actions might actively influence and alter the context and dynamics of the driving environment. Simultaneously, Xu et al. (2024) conducted comprehensive studies on the impact of drivers’ psychological and physical characteristics. Their research focused on gaining a thorough grasp of how these specific driver characteristics influence other driving-related factors. These characteristics encompass reaction times, decision-making processes, and overall driving performance, making a substantial contribution to the field of driver behavior analysis and road safety research. This field is new for the researcher, and they try to analyze driver behavior using ML. A promising area of research is to achieve a better solution for predicting the risk of road accidents based on DB and other factors associated with traffic crashes.

Recent research by Xu et al. (2024) assessed the interaction between DA and autonomous driving systems (ADS) in a virtual reality environment, simulating over 10,000 road scenarios. Wan et al. (2023) assess drivers’ physiological and psychological responses in expressway tunnels using machine learning methods to identify dangerous behaviors. Ahmad et al. (2024) analyze driver behavior in the context of wildfire evacuations, based on braking and acceleration patterns from connected car data. Mazengia et al. (2024) investigated public transport operators in Ethiopia and found that increasing awareness of road safety can decrease risky behavior. Kontaxi et al. (2023) explored the effects of mobile phone use and trip characteristics on speeding, and the authors argue that interventions are necessary to address mobile-driven driving violations. These studies collectively highlight the impact of various environmental factors on driving behavior and the effectiveness of DA systems.

Human driving parameters are an important concern in road accident research. Driver actions are a significant concern because drivers play a leading role in operating vehicles in the environment and driving scene (Li et al., 2022). When road crashes occur, the primary focus of investigation is on the database factors that directly contribute to them. Some studies focus on the relationship between driving behavior and the surrounding environment (Elamrani Abou Elassad, Mousannif & Al Moatassime, 2020a). ML is a prominent topic today; the researcher proposes many ML-based models to predict the risky driving parameter for traffic crashes and discover some driving factors that may contribute to road safety (Gálvez-Pérez, Guirao & Ortuño, 2023). However, identifying appropriate machine learning algorithms that can accurately identify driver actions with minimal error and high accuracy is not a simple task. Moreover, based on the weather conditions, driver actions, such as vehicle kinematics and driver input, play an important role in road crashes. These factors have led to the integration of various intelligent systems into mass transport systems, enabling automatic actions or directions that could potentially reduce traffic crashes.

The National Highway Traffic Safety Administration (NHTSA) provides data indicating that driver error accounts for a significant majority of road accidents, specifically 94%. They use historical data for examination, which reveals a consistent association between motor vehicle accidents and human errors. Comprehensive research has investigated the primary factors contributing to vehicular crashes, leading to this conclusion. Therefore, it is imperative to continue conducting research on driver actions (Sarsangi et al., 2023) to address these faults and enhance safety protocols, according to reports by Ayers, Whitlow & Dressler (2018). According to Cai & Lin (2011), the primary determinant of traffic safety is the driver’s behavior.

The road accident datasets usually have a lower frequency of accident occurrence compared to non-accidents. A prevailing recommendation among researchers is to maintain a ratio of four non-accident cases for every accident case (Yusoff et al., 2024; Xu & Prozzi, 2023; Gálvez-Pérez, Guirao & Ortuño, 2023). Nevertheless, this ratio sometimes gives rise to data imbalances, since models may exhibit a bias towards the more common non-accident class, producing an over prediction of that class (Elamrani Abou Elassad, Mousannif & Al Moatassime, 2020b). In real-time crash analysis, ML approaches have proven to be helpful in identifying the relationship between accident occurrences and their factors. Among the frequently used ML models are support vector machines(Govinda, Raju & Shankar, 2022; You, Wang & Guo, 2017), neural networks (Liu, Boyle & Banerjee, 2018), and Bayesian networks (Park et al., 2018). Artificial neural networks (ANNs) prove their usefulness (Abou Elassad, Mousannif & Al Moatassime, 2020) to handle noisy data and carry out quick real-time computations with resilient efficiency. According to Hussain et al. (2022), the hybrid models that use isolation forest and minimum covariance determinant for sampling worked better than other methods and traditional models. Another significant discovery from this study is that most of the hybrid models performed better than conventional models.

Traditional statistical learning-based methods, such as linear regression, Gaussian regression, and discriminant analysis, have played a fundamental role in crash prediction and road safety analysis (León & Aoyama, 2022; Ahmadi et al., 2014; Sam, Velanganni & Evangelin, 2016). Nevertheless, these methodologies frequently encounter challenges pertaining to the quality of data, necessitate a significant amount of historical data, and pose difficulties when dealing with attributes that encompass a diverse array of categories. The utilization of statistical techniques alone for crash event analysis sometimes leads to equivocal conclusions due to the presence of these obstacles. On the other hand, Tang et al. (2018) studies have shown that machine learning-based methods exhibit enhanced capacities in forecasting future occurrences, specifically within the domain of transportation systems. ML models have many advantages: they can solve complex problems on their own using data from different sources, they can update themselves with new data to perform better, and they can provide predictions and explanations by finding patterns. Significantly, crash event prediction has been effectively addressed by the utilization of notable strategies such as Adaptive Boosting (AdaBoost) (Freund & Schapire, 1997), eXtreme Gradient Boosting (XGBoost) (Ke et al., 2019), and RF (Ragab et al., 2014), which have demonstrated noteworthy outcomes.

An essential factor in the examination of road accident occurrences is effectively handling the disparity in datasets, which often consist of a smaller number of crash instances in comparison to non-crash data. To address this disparity, researchers employ resampling methods like the Synthetic Minority Oversampling Technique (SMOTE), as proposed by Chawla et al. (2002). The SMOTE method minimizes information loss and also helps in overfitting by generating artificial samples from the minority class (Fernández et al., 2008). The study by Arafa et al. (2022) addresses the notable difficulty of unbalancing datasets in machine learning. This is a common problem in which uneven class distributions result in inadequate categorization of minority classes. The suggested RN-SMOTE method merges the Synthetic Minority Over-sampling Technique (SMOTE) with Density-Based Spatial Clustering of Applications with Noise (DBSCAN) to enhance data preparation. Schlögl et al. (2019) used random under sampling (RUS), which removes examples from the majority class at random, is one of the techniques employed in random under-sampling where, until the number of examples in the majority class and the minority class are roughly equal, the process is repeated. While the RUS approach can distribute the class fairly, it may result in the loss of important information from the majority class. Random oversampling (ROS), used by Rocha et al. (2023), aims to increase the number of instances within the minority class. To achieve this, we randomly duplicate the samples from the minority class until their numbers approximate those of the dominant class. However, because the replicated samples are repetitious, this strategy may result in overfitting. There are different versions of SMOTE, such as SMOTE-Tomek Link (SMOTE-TL) (Batista, Prati & Monard, 2004), which creates synthetic samples and gets rid of overlapping data, and SMOTE-ENN (Edited Nearest Neighbor) (Xu et al., 2020), which uses both too many samples of the minority class and too few samples of the majority class. Another method, Adaptive Synthetic Sampling (ADASYN) (He et al., 2008), produces samples by considering the density distribution of the minority class. The complexity of the modeling requirements adjusts the number of synthetic samples.

Although machine learning approaches show promise in improving road safety, there are still certain limits that need to be addressed. The evaluated research frequently has limited scope, and there is a clear difference in the emphasis on analyzing and detecting driver action (DA) versus actually predicting driver action, as reviewed by Ameksa et al. (2024). Moreover, hybrid models (Ameksa et al., 2024), which have exhibited effectiveness in tackling intricate problems such as crash prediction, have garnered relatively less focus. Sangare et al. (2021) explores a method for predicting traffic accidents by fusing the advantages of the Gaussian mixture model (GMM) with the support vector classifier (SVC). He et al. (2008) suggests a transparent Deep Hybrid Network (DHN) modeling approach for extemporization crash prediction. It captures both geographical heterogeneity and temporal variability at the same time by combining the powers of convolutional neural network (CNN), deep neural network (DNN), and long short-term memory (LSTM). Future studies should consider more data sources that reflect human movement. Kashifi et al. (2022) applied a hybrid GA and ANN model to investigate the severity of fixed object crashes among older drivers in California in 2012. Their results indicated that the ANN model was more capable of predicting low-severity crashes, while the hybrid GA and ANN model was more accurate in predicting high-severity crashes. The existence of these gaps indicates important areas for future research, such as integrating a wider range of studies and concentrating on creating hybrid models to forecast driving errors. The ultimate goal is to improve the accuracy of predictions, anticipate driving errors in advance, and decrease road traffic accidents and fatalities.

To the best of our knowledge, little to no research has used a hybrid machine learning model that examined the impact of several combinations of four adopted features, namely driver actions, vehicular telemetry, tire conditions, and multiple weather conditions, for the prediction of road crashes. Furthermore, this work introduces a more balanced approach, combining DBSCAN and SMOTE TL (DBSTLink) with hybrid ML models GA and ANN. In this study, we develop and verify a system for anticipating road accidents for drivers by analyzing multiple parameters. To improve the study report’s comprehensiveness, we have included supplementary content.

The use of the hybrid machine learning method GA-ANN with a DBSTLink Balancing approach in examining driver action (DA) is a relatively new field with a small but expanding body of research. Hybrid machine-learning models amalgamate diverse algorithms or incorporate machine learning with other analytical techniques with the objective of harnessing the advantages of each constituent. Many researchers have proven that these models can provide more precise and detailed insights regarding DA by capturing intricate patterns and interactions that single-method models may overlook. Further experiments are required to determine the potential and constraints of the proposed approach. Using these kinds of models could greatly improve the accuracy and understanding of predicting driver behavior, leading to better road safety measures and interventions that use balancing data methods like under-sampling, over-sampling, and SMOTE to fix the common problem of data imbalance.

The current study employs a fusion decision system that includes several classifiers to examine merged data about driver actions, vehicle information, environmental factors, and personal details. After data pre-processing, we employ Our approach to handling imbalanced data from road crashes entails the use of SMOTE-TL and DBSCAN, named DBSTLink. Repeated cross-validation enables us to determine the optimal configuration for our model. Finally, we build crash detection models using a genetic algorithm combined with an artificial neural network called GA-ANN, considering the differences in data to ensure our predictive models are very accurate. Specifically, this study makes the following significant contributions:

• An DBSCAN+SMOTE TL (DBSTlink) data balancing method designed in this work helps to balance data for effective crash prediction.

• A hybrid GA-ANN classifier framework determines and analyzes the diverse parameters of a driver’s actions that are most indicative of road accident risks.

• The DBSTLink’s performance with a hybrid GA-ANN classifier was compared with the existing data-balancing approach and other models used by other researchers.

The main objective of this study is to determine statistically significant predictive relationships through a sound classification methodology. However, causal interpretations are not within the scope of this current approach. The remainder of this study is organized as follows. Methodology explains the experimental design and data collection, as well as the proposed methodology. The Experimental Results & Discussion section lists the obtained results and discussion. Finally, we summarize this work and offer directions for future research.

Materials and Methods

Variable description

Data was collected on various aspects of DA and vehicle kinematics. Table 2 defines the variable description. Driver actions refer primarily to the driver’s behavior and choices. The dependent variable is the crash event, coded as binary variable with a value of 1 if a crash was identified and 0 if not. Apart from the categorical features like weather conditions, all variables are with continuous variables. The “Driver Inputs” area contains variables like throttle, brake, steering, and RPM. ere the inputs directly represent the driver’s decisions and actions while operating the car. The movements and actions of a vehicle result from the vehicle kinematics, which are a combination of factors such as driving action, vehicle design, and environmental conditions. The driver influences the variables Speed, Lateral_g, Longitudinal_g, and the physical motion and dynamics of the vehicle, but not exclusively. Tire temperatures and weather conditions influence the driver’s inputs, but the driver does not directly control them.

Table 2:

Variable description of road crash.

Category	Variable	Description
Vehicle kinematics	Speed	The magnitude of the vehicle’s velocity.
	Lateral_g	Lateral (sideways) acceleration of the vehicle.
	Longitudinal_g	Acceleration of the vehicle along the direction of travel.
	Vertical_g	Vertical acceleration of the vehicle (upward or downward).
	Yaw_angle	Angle between the vehicle’s longitudinal axis and its line of travel.
	Drift_angle	Angle during a turn between the car’s orientation and its velocity direction.
	Spin_angle	Angle between the direction in which a wheel is pointing and the direction it is actually traveling.
Driver inputs	Throttle	Accelerator pedal position.
	Brake	Brake pedal position.
	Steering	Steering wheel angle.
	RPM	Revolutions per minute; number of rotations of the vehicle’s engine crankshaft.
Tire conditions	Tyre_temp_FL	Temperature of the front left tire.
	Tyre_temp_FR	Temperature of the front right tire.
	Tyre_temp_RL	Temperature of the rear left tire.
	Tyre_temp_RR	Temperature of the rear right tire.
Weather conditions	Weather	General weather conditions affecting driving (e.g., clear, fog, rain, snow).
Crash state	State	Crash or no crash

DOI: 10.7717/peerj-cs.3131/table-2

The other variables are either the consequences of driving behavior or external factors that have an influence. The major goal of this study is to design a hybrid machine learning model that can predict crash occurrences by taking into account the most important features of driving action and applying the designed hybrid machine learning methods. A binary classification model called “crash prediction” attempts to predict whether or not a driving behavior will be involved in an accident and compare it with non-ML techniques.

The main objective of this study is to develop a hybrid machine learning model that incorporates several facets of driving action. By evaluating these behaviors, the model seeks to identify specific activities or patterns that have a significant impact on car crashes. Figure 1 describes the proposed work’s overall process. It depicts a thorough machine learning workflow for accident prediction, commencing with data collection and concluding with model quality evaluation. At first, data is gathered, and then the data is splitted, then apply a balancing approach which helps to balance the data, where the number of ‘crash’ cases is considerably lower than that of ‘non-crash’ examples. Resampling techniques are employed to equalize the representation of crash and non-crash samples in the dataset, hence achieving balance in the classes. The balance data is subsequently employed to train a hybrid machine learning model, which exploits the collective advantages of many techniques to forecast collisions. The evaluation of the proposed system is conducted with a comprehensive set of measures such as precision, recall, accuracy, MCC, G-mean, and F1-scores. This procedure describes a methodical strategy that can be used for improving road safety. This approach emphasizes the significance of well-rounded datasets and thorough review in creating successful prediction models.

Data resampling with balancing approach

For experiments, the generally accepted 80:20 ratio is that the training dataset receives about 80% of the dataset (Ahmed et al., 2023) and the other 20% is used for testing. Within the context, approximately 75,900 samples were recorded during simulations where the ratio of crash observations is about 3%, indicating that the data are extremely imbalanced, which has been found in similar studies related to crash prediction (Rocha et al., 2023; Jerez et al., 2010) through the adoption of the endorsed data balancing strategies. A novel balancing approach is designed in this work to resolve this issue by balancing class distribution in the data set. In this work, the DBSCAN technique along with SMOTE-Tomek Link, named DBSTLink, is proposed. Figure 2 depicts the DBSTLink algorithm’s flowchart. The two steps of the DBSTlink process are undersampling and oversampling. We used DBSCAN to extract clusters from the entire training set, successfully completing the undersampling portion. Then, each cluster had 50% of its negative instances removed. The undersampling technique reduces the number of negative examples by removing some of them from the dataset. SMOTE TL was used to add artificial instances of the positive class to the training set for the oversampling component. SMOTE-TL integrates the SMOTE capability to produce synthetic data for the minority class and the Tomek algorithm. Linking (two modifications of CNN 1976) enables the removal of data that are accurately classified as Tomek linkages from the dominant class. The Tomek linkages constitute data samples from the majority class that exhibit the closest proximity to the data from the minority class.

The algorithm employs a combination of of DBSCAN and SMOTE Tomek Links to successfully balance a dataset called D, which includes features X and labels y. Initially, DBSCAN operates by determining core points within D based on a specified radius $\mathrm{ϵ}$ and a minimum required number of neighbors. Min_samples points that meet these criteria are accumulated into Dclean, a subset of D containing only core points, thereby reducing noise and outliers. Following noise reduction, the algorithm applies the SMOTE to address class imbalance. For each important point Xi in Dclean, SMOTE creates new samples by blending Xi with its K-nearest neighbors that belong to the same class. The number of synthetic samples to generate for each Xi is determined by a predefined sampling strategy. We implement Tomek Links to further refine the class boundaries. This step entails identifying pairs of nearest neighbors (Xi, Xj) who are from opposite classes, one of which is synthetic. Once we identify these pairs, also known as Tomek Pairs, we remove the synthetic sample Xj from Dclean. This process enhances the separation between classes, leading to a more defined decision boundary in the resulting balancing dataset D′. By using DBSCAN to get rid of initial noise, SMOTE to fix class imbalance, and Tomek Links to fine-tune class boundaries, the algorithm makes sure that the data is well-prepared for future predictive modeling tasks. This could make the models trained on this balancing dataset more accurate and useful.

Model development

The proposed FusionNet GA-ANN approach has three steps:

• The balancing dataset was incorporated into the genetic algorithm (GA) as the initial population. The process of evolution was repeated until a predetermined termination condition was met. The main goal of this research was to determine the key characteristics that have the greatest impact on the probability of crash incidents in driving action.

• After selecting features from GA need to use ANN, where dropout will be used to prevent over-fitting.

• The effectiveness of the GA-ANN hybrid classifier will be performed using metrics like accuracy, precision, recall, g-mean, and F1-score compared to other classification models.

The GA-ANN hybrid classifier combines the adaptive learning skills of GA with the computing capacity of ANNs to form a strong classification tool. The GA adjusts the feature selection and network parameters, which then serve as inputs for the ANN.

The training procedure starts by utilizing a deep ANN structure on selected features of GA-ANN. In order to mitigate overfitting, dropout layers are selectively incorporated into the ANN to improve its capacity during the training session. A notable procedural transformation takes place once the ANN becomes proficient in identifying fundamental patterns and subtle intricacies in the data. The GA refines and optimizes the extracted characteristics, which are subsequently processed through the ANN for the final classification.

This intentional integration not only avoids the typical problem of early convergence in ANNs but also takes advantage of the global search power of GAs, thereby improving the overall accuracy and efficiency of the model. Algorithm 1 shows the pseudo code of the proposed DBSTlink approach with the hybrid GA-ANN classifier. It guarantees that the proposed model enhances the accuracy of the classification task and, in diverse datasets, may result in optimal categorization.

Algorithm 1::

Data processing with DBSCAN, SMOTE, Tomek Links, GA, and ANN.

Input:

•   D: Dataset with features X and labels y

•    $\epsilon$: DBSCAN radius

•   min_samples: DBSCAN minimum samples

•   sampling_strategy: SMOTE sampling strategy

•   K: Number of neighbors for SMOTE

•   Parameters for GA, ANN architecture

Output:

•   Trained ANN model

Step 1: Balanced Dataset Preparation

1.1 Initialize: D_clean ← ∅, D′ ← ∅

1.2 Apply DBSCAN for noise reduction

1.3 For each point X_i in D do

1. 1.3.1 Neighbors ← DBSCAN_NearestNeighbors(X_i, D, $\epsilon$, min_samples)

2. 1.3.2 If X_i is a core point then

• Add X_i to D_clean

End if

End for

Step 2: Apply SMOTE for Oversampling

2.1 For i = 1 to len(D_clean) do

3. 2.1.1 n ← KNearestNeighbors(X_i, D_clean, K)

4. 2.1.2 p ← sampling_strategy ÷ 100

5. 2.1.3 While p ≠ 0 do

• a. Neighbor ← randomly select from n

• b. X_smote ← X_i + rand(0, 1) × (Neighbor − X_i)

• c. Add X_smote to D_clean

• d. p ← p − 1

End while

End for

Step 3: Remove Tomek Links for Class Boundary Refinement

3.1 D′ ← D_clean\TomekLinks(D_clean)

Step 4: Feature Selection using Genetic Algorithm

4.1 F_selected ← ApplyGA(D′, GA parameters)

Step 5: Train Artificial Neural Network

5.1 ANN ← InitializeANN(ANN architecture)

5.2 Train ANN on D′ using F_selected

Return: Trained ANN model

DOI: 10.7717/peerj-cs.3131/table-12

Figure 3 illustrates the systematic procedure for analyzing crash data using a hybrid machine learning model that integrates the capabilities of ANN and GA.

Hybrid GA-ANN classifier

The proposed methodology uses a GA to determine the most important features for crash investigation. The GA represents each feature set as a chromosome and evaluates its usefulness through a fitness function that specifically measures its prediction ability in crash scenarios.

The genetic algorithm selects the features which are displayed in Fig. 4. The most effective traits that the GA identifies are used as input for an ANN to perform additional assessment. The method leverages the benefits of each algorithm to produce more accurate and reliable analysis results.

Results

Dealing with an imbalanced dataset, accuracy may suffer from the bias of the majority class. To address the imbalanced dataset, G-mean, MCC, and F1-score metrics are used. Table 6 presents the performance metrics of classifiers without any resampling techniques, highlighting the impact of class imbalance on model performance and the results of the proposed DBSTLink. Table 7 compares balancing approaches with other models. A comparison of the results depicted in Tables 6 and 7 demonstrates the major contribution of adopting DBSTLink as a resampling strategy to performance improvements in crash prediction.

DBSTLink emerges as the superior balancing technique across the four models—XGBoost, RF, SVM, and GA-ANN—when considering key performance metrics like F1-score, MCC, and G-mean. Starting with XGBoost, DBSTLink performs almost on par with SMOTE and SMOTE TL in terms of F1-score, achieving 0.79 compared to 0.93 for both SMOTE and SMOTE TL. However, the marginal drop in F1-score is compensated by the technique’s stability across the other metrics. Its G-mean of 0.72 is comparable to DBSM’s, which also scored 0.72, demonstrating that DBSTLink handles class imbalances in a way that preserves the harmonic balance between precision and recall, which is vital in imbalancing datasets.

For the RF model, DBSTLink provides excellent scores across all metrics. The F1-score and G-mean for DBSTLink are both 0.94, which is higher than SMOTE (0.93) and SMOTE TL (0.91). MCC, which measures the correlation between the true and predicted classifications, also shows 0.94, tying with SMOTE and surpassing SMOTE TL (0.90) and DBSM (0.59). The result indicates that DBSTLink gives an optimal balance solution between over- and under-sampling. It provides a consistent and reliable model performance. For the SVM model, DBSTLink dramatically improves over the other techniques, achieving a significantly higher F1-score of 0.87, compared to 0.69 for SMOTE and SMOTE TL, and 0.64 for DBSM. Similarly, the MCC result is 0.60, which is considerably better than SMOTE and SMOTE TL (both at 0.47) and DBSM (0.31). G-mean is also higher at 0.86 compared to 0.85 for SMOTE and SMOTE TL. This performance indicates that DBSTLink handles imbalanced classes far more effectively in SVM compared to other balancing methods.

Finally, for the GA-ANN model, DBSTLink dominates with the highest scores across all metrics: F1-score (0.99), MCC (0.96), and G-mean (0.98). These values surpass those of SMOTE, SMOTE TL, and DBSM, which further cements DBSTLink’s position as the best technique. Its ability to provide consistently high values for precision, recall, and class balance, as reflected by G-mean, shows that DBSTLink not only mitigates class imbalance but also enhances the model’s generalization capabilities. Overall, DBSTLink proves to be the most robust balancing technique across models, delivering superior or comparable results in terms of F1-score, MCC, and G-mean. Its balancing handling of precision and recall, coupled with its ability to maintain high class-separation performance (G-mean), makes it a better choice than SMOTE, SMOTE TL, or DBSM in most scenarios. Results are illustrated for each balancing approach given in Fig. 5.

In Table 8, a performance evaluation of the FusionNet GA-ANN classifier was conducted using several balancing approaches, resulting in diverse results where random undersampling (RUS) achieved an accuracy of 0.84, a G-mean of 0.836, and a mean squared curve (MCC) of 0.67, which suggests a reasonable level of efficacy in addressing class imbalance. The implementation of random over sampling (ROS) significantly enhanced the accuracy to 0.98. However, the G-mean and MCC metrics demonstrated no changes, indicating that, although the accuracy is high, the model’s capacity to maintain a balanced class performance stays constant.t. The SMOTE showed notable enhancement to its performance, achieving an accuracy of 0.96, a G-mean of 0.96, and an MCC of 0.921, which define the efficacy of SMOTE in effectively managing imbalance of crash data. Furthermore, DBSM, which is a combination of DBSCAN and SMOTE, shows robust performance with similar metrics. The SMOTE with Tomek Links (SMOTE-TL) model shows a satisfactory performance with an accuracy of 0.95, a G-mean of 0.95, and an MCC of 0.90. But the proposed DBSTLink resampling procedure, which achieved an accuracy of 0.99, a G-mean of 0.98, and an MCC of 0.96, surpassed existing approaches.

Table 9 presents the parameter estimates of the logit regression model for crash events and the goodness-of-fit statistics. The statistical analysis demonstrates that Throttle (β = 0.0096, p < 0.001), Brake (β = 0.0048, p < 0.001), and Steering (β = 0.0045, p < 0.001) all contribute significantly to the predicted outcome, while Longitudinal_g (β = 0.4875, p < 0.001) and Vertical_g (β = 0.6827, p < 0.001) exhibit strong positive effects. Speed (β = -0.0164, p < 0.001), Tyre_temp_FL (β = −0.0299, p < 0.001), and Yaw_angle (β = −0.0006, p = 0.023) present significant negative effects. The results show that RPM (β = 2.306e−05, p = 0.140) and Lateral_g (β = −0.0066, p = 0.704) have no significant impact on the predicted outcome. The results demonstrate that essential driving actions together with environmental conditions have a major impact on the outcome. The 10-fold cross-validation produced an average accuracy rate of 98.79%. A 95% confidence interval calculation provided a statistical assessment of accuracy measurement reliability through a range from 98.24% to 99.33%, which is also presented in Table 9.

The findings emphasize the need to select resampling techniques for any model learning and guarantee equitable and efficient categorization across different class distributions. Figure 6 illustrates the results for each category of combined features. To further explore the findings visually, the results are presented in Fig. 7.

Discussion

The goal of this work is to solve the balancing problem by introducing a new method called DBSTLink, which uses a combination of genetic algorithms and artificial neural networks for crash prediction based on various machine learning techniques. To tackle the problem of imbalanced data, which is typical for crash-related observations where target classes are not equally represented, a new balancing strategy, DBSTLink, has been proposed. This approach is designed to equate the number of examples in each class to enhance the accuracy of the subsequent analyses. Furthermore, a comprehensive analysis has been carried out to determine the most significant predictors of crash events. Knowing these key predictors is important in the design of control measures. Lastly, a number of prediction methods have been explored and compared with the developed GA-ANN model. This comparison shows that driving action is a major cause of road accidents and that these behavioral attributes should be incorporated into active traffic safety systems to improve their efficiency. Table 11 presents a comparison between our system and other crash assessment studies. Based on modeling metrics, we compared several fusion models used in crash investigations. The performance of the proposed DBSTLink balancing approach is compared with various sets of other balancing approaches like SMOTE, SMOTE TL, and DBSM in reference to the values of F1-score, G-mean, and MCC, as these metrics are deemed to be highly acceptable for tackling the class imbalance issues. The combination of the proposed balancing approach with SVM, LR, XGBoost, and GA-ANN compares favorably with other balancing approaches. We obtained the highest scores for all prediction models using DBSTLink approaches, which consistently outperformed other models in most established combinations based on the individual learners applied. The proposed hybrid GA-ANN performed with the highest accuracy among other single machine learning models. Our model reached a predictive accuracy of 98.79%, which became more reliable because the 95% confidence interval was narrow, between 98.24% and 99.33%. The small range indicates that the model performs consistently when working with different data subsets. The model’s stability proves that the chosen predictors work well together with our modeling approach. Using 10-fold cross-validation makes our performance estimates trustworthy and realistic because it is a strong statistical method that reduces errors compared to other validation methods. The research results showcase the effectiveness of the proposed model in real-world predictive scenarios, especially in predicting crashes.

This research finds new ways to classify crash occurrences as a function of driving action so as to help in designing effective crash prevention strategies that entail issuing warning signals to drivers and encouraging safe behavior. The study establishes statistical relationships through validation and predictive modeling, but it does not establish cause-and-effect relationships. The methods we employed, such as the DBSTLink balanced approach with GA-ANN, prevent us from establishing cause-effect relationships. Nevertheless, in order to gain admission, there are certain restrictions that require attention. Simulator studies are applied to a controlled and modifiable environment to look at road crash incidents, while real trials are conducted in real road conditions and are risky; however, the driving simulator is unable to replicate the actual driving experience. The validity of the findings from a driving simulator depends on the specific tasks simulated in the environment. The proposed DBSTLink + GAANN method required evaluation through performance testing on simulated data and real-world data sets. The model demonstrated strong generalizability through its consistent results across different datasets which showed its ability to adapt to various data distributions. The method successfully detected meaningful spatio-temporal patterns together with contributing factors in real-world scenarios which supports its practical use for traffic incident analysis.

Conclusions

Global road safety is a critical issue. To mitigate accidents and enhance road safety, it is necessary to develop extremely proficient real-time crash prediction models that can identify the most significant antecedents contributing to collision incidents. A thorough examination is necessary to decrease road accidents and improve traffic safety for drivers through effective tactics. Previous studies concentrating on this domain employed machine learning. Models have demonstrated their effectiveness in enhancing road safety. The research findings propose innovative approaches to enhance our findings of crash occurrences by taking into account driver actions (DA). That is a crucial step in developing effective crash prevention strategies that warn drivers and promote safe driving habits for both experienced and non-experienced drivers. This research helps to propose a balancing approach that will tackle imbalance issues in a more efficient way and also propose a hybrid classifier for merging information and building fusion models. However, it is important to acknowledge certain limitations. This study is simulator-based, which offers a controlled environment, particularly in road crash approaches. Simulators do not fully replicate real driving experiences. The outcome of a simulator can vary, being either different or controlled based on the tasks executed. In this research, DA is based on driver input, vehicle kinematics, tire conditions, and weather conditions, which are simulated environments for the drivers. The simulator cannot replace real-world driving experiences while predicting crashes.

The proposed DBSTLink + GA-ANN framework received validation through testing with simulated data and publicly available real-life traffic data. The model demonstrated reliable performance because it achieved high results in all tested datasets. The research indicates that this method works effectively for theoretical modeling and real-world applications in traffic crash prediction and urban safety management. Future research must also more thoroughly examine potential biases associated with repeated crash scenarios and driver anticipation effects. Further, the capabilities of the simulator should be expanded to include advanced behavioral modeling and complex driving tactics to increase the robustness and practical utility of simulator-based driving research. Future research should use experimental or longitudinal designs to validate causal relationships. Besides, future research could expand the predictive models by using different techniques to address class imbalance and improve crash prediction.