A novel hybrid supervised machine learning model with metaheuristic optimization algorithm for prediction of driver concentration levels

Abstract

The rapid expansion of urban living has led to increased road congestion, posing significant challenges for maintaining driver attention and ensuring road safety. This study investigates advanced machine learning techniques to assess and predict driver focus by analyzing key metrics such as concentration scores, reaction times, and stress management capabilities. A comprehensive evaluation of various machine learning models was conducted, including Logistic Regression (LR), Linear Discriminant Analysis (LDA), Ridge Classifier, Decision Tree (DT), Light Gradient Boosting Machines (LightGBM), Naive Bayes (NB), Support Vector Machines (SVM), Random Forest (RF), AdaBoost, XGBoost, CatBoost, K-Nearest Neighbors (KNN), and Stacking. To optimize performance, a metaheuristic optimization algorithm was employed for precise hyperparameter tuning. The novelty of this research lies in the introduction of a novel hybrid machine learning algorithm, “CatBoost+Stacking.” By integrating the strengths of CatBoost within a stacking framework, this new model significantly enhances the accuracy of predicting driver concentration levels. Experimental results demonstrate that the “CatBoost+Stacking” model outperforms existing baseline models, offering a more effective approach for monitoring driver behaviour. These findings provide practical insights for developing proactive road safety strategies and reducing accidents, highlighting the transformative potential of hybrid models in creating safer driving environments.

Introduction

Bus drivers are essential for safe and efficient passenger transportation, often navigating complex road networks in urban areas or over long distances. Their ability to maintain high levels of concentration is essential for adhering to traffic rules, anticipating potential hazards, and managing unpredictable road conditions while driving. Response time, a critical factor for safe driving, reflects how quickly a driver reacts to unforeseen situations such as abrupt stops, erratic behavior from other road users, or sudden changes in weather and traffic patterns. Factors such as fatigue, distraction^1–4, and the driver’s mental and physical well-being significantly impacted the response time.

The demanding nature of a bus driver’s job, especially in congested urban environments, adverse weather conditions, or under tight schedules, adds to the stress, which can impair their focus and decision-making abilities. Elevated stress levels often lead to slower reaction times and a higher likelihood of errors, increasing the risks associated with driving⁵. Identifying the factors that affect concentration and response times is crucial for developing interventions to enhance road safety⁶.

The incorporation of machine learning techniques has transformed several industries in recent years, including healthcare, transportation, and recruitment, by providing data-driven insights and predictive skills. By examining enormous datasets and identifying significant patterns, machine learning algorithms have demonstrated their efficacy in resolving challenging problems. Predictive modeling techniques, including KNN, DT, RF, AdaBoost, XGBoost, CatBoost, NB, SVM, and LR, have been widely used in various industries.

This study leverages these advanced machine learning techniques to analyze driver performance data, including metrics such as concentration level, response time, and stress management capability⁷. By evaluating and comparing the predictive accuracy of these algorithms, this study aimed to identify the most effective hybrid model for assessing driver concentration and response times. The findings are expected to contribute to the development of strategies and tools that improve driver safety and reduce the risk of accidents on busy roadways.

The primary objective of this study is to develop and evaluate advanced machine learning models for predicting driver concentration levels by analyzing key metrics such as concentration scores, reaction times, and stress management capabilities, addressing urban road congestion challenges. Specifically, it aims to:

• Conduct a comprehensive performance comparison of 14 machine learning algorithms, including Logistic Regression, Linear Discriminant Analysis, Ridge Classifier, Decision Tree, LightGBM, Naive Bayes, SVM, Random Forest, AdaBoost, XGBoost, CatBoost, KNN, and Stacking.

• Apply metaheuristic optimization algorithms for hyperparameter tuning to enhance model accuracy and robustness.

• Introduce and validate a novel hybrid "CatBoost+Stacking" model that integrates CatBoost’s strengths within a stacking ensemble for superior prediction of driver focus.

• Demonstrate the practical implications of the top-performing model for proactive road safety strategies, accident reduction, and safer urban driving environments.

Literature review

More than 1.2 million people are killed and 50 million injured in motor vehicle collisions every year⁸. Driving behaviour plays a significant role in the probability of a car crash^9–11. It is estimated that almost 90% of all car accidents are caused by driver error¹², including driving under the influence, error, and reckless actions^13,14. Some common human factors that cause critical incidents on the road are distraction, drowsiness, intoxication, aggression, and speeding. In the United States alone, about 35% of transportation-related fatalities are attributed to a driver who is impaired, accounting for 10,497 lives lost in 2016 alone¹⁵. Impaired drivers have impaired perception, take more time to react, and their ability to steer or apply the brakes would be compromised¹⁶.

Vehicle sensor data provides a rich source of insight into driver actions, as the resultant motion response to inputs from steering, acceleration, and braking is captured. This can reveal whether drivers are acting appropriately to potential hazards. Some studies have suggested enhancing driver performance by adjustments to the vehicle system when capability is impaired by the likes of fatigue, distraction, or impairment^17,18. Others still have focused on conditional automation, such as detecting, under complex driving situations, the time required by a driver to take back control following a takeover request by the vehicle¹⁹. Driver readiness has also been assessed by other research using wearable sensors during real-road tests.

Most of the literature explains driver behavior and investigates the relationships between age, fatigue, sleepiness, and driving risk^11,20–23. Crash risk is also estimated using various methods, including probabilistic models incorporating diverse data sources. Research finds that distraction degrades driving performance, increases collision risk²⁴ and driver injury²⁵. For example, several methods utilize in-vehicle position and speed data within driving datasets for iterative crash energy and risk predictions within given trips.

Machine learning has demonstrated considerable potential in driver monitoring systems for predicting concentration levels and preventing accidents²⁶. Logistic Regression has been effectively employed in real-time detection of driver drowsiness using heart rate variability derived from physiological signals, camera-based classification of drowsiness states, and early prediction of fatigue^27–29. Support Vector Machines (SVM) have excelled in assessing driver attention through eye-tracking, detecting real-time distractions from driving performance data, classifying cognitive distractions in metro drivers, and conducting ensemble risk assessments with road object recognition^30–33. Naive Bayes (Bayesian Networks) has shown robust performance in predicting driver distraction from simulator behavioral data³⁴. Random Forest has proven superior for drowsiness detection compared to logistic regression, integrated distraction-risk classification, and trip-wise driver behavior analysis^35–37. AdaBoost has enhanced sequential models of driver fatigue progression. XGBoost has delivered high accuracy in predicting driving risk from vehicle telemetry and HRV-based concentration assessment³⁸. K-Nearest Neighbors (KNN) has been utilized for detecting gaze anomalies in distraction monitoring and early signals of drowsiness^39,40. Rotation Forest and other ensembles have advanced multi-target driver state prediction from multimodal data, with recent hybrids achieving state-of-the-art performance in fatigue monitoring^41,42.

In the context of road safety, prediction of driver behaviour is crucial for mitigating accidents which can be done using various machine learning techniques⁴³. To assess and forecast driver concentration levels in urban traffic situations, this study proposes a comparative machine learning-based method.

The following are the main contributions of this study:

• An automated system for assessing driver focus using behavioural and physiological metrics, such as concentration scores, response times, and stress management indicators.

• A data-driven solution for identifying the most accurate machine learning model to support proactive road safety strategies and reduce the risk of traffic accidents.

Research gap

Previous works primarily focused on single-modality data such as eye-tracking or vehicle metrics only, which often fail to capture complex human nature involving distraction and fatigue. These methods tend to be effective only after a dangerous event has begun, rather than proactively identifying deteriorating concentration of drivers. Moreover, existing machine learning models tend to overfit on limited datasets, resulting in lack of generalization for diverse real-world driving conditions. This creates an urgent need for robust, predictive system that integrates multiple data streams to forecast driver attentional lapses before any accident occurs.

Novelty of the work

The novelty of this work lies in the development of a novel hybrid machine learning framework that integrates 4 machine learning models—CatBoost, LR, SVM and GBM resulting in better accuracy as compared to the previous works. The model’s architecture and hyperparameters are finely tuned using an advanced metaheuristic optimization technique known as Mountain Gazelle Optimization to enhance accuracy. This integrated, optimized approach enables more reliable prediction of driver concentration levels ensuring the low-risk driving resulting in road safety.

Methodology

The driver concentration dataset was taken from the Figshare open access data repository and link is provided in the data availability statement. To evaluate and enhance driver concentration, a series of tests were administered to a sample of 204 drivers as given in Fig. 1. These tests were designed to assess key attributes related to driving performance, including concentration, response time, and composure under pressure, which are critical for ensuring safe and efficient driving performance. Training, validation and testing sets were created from the collected dataset. Different machine learning models were trained on the training set, model optimization was validated using the validation set and their prediction accuracies were assessed using the testing set. Model hybridization technique was employed to enhance model accuracies. To maximise driver training and safety procedures and enhance overall road safety and performance, this study aimed to determine the best hybrid model for predicting a driver’s degree of concentration.

Fig. 1.

Methodology flowchart.

The significance of driver concentration

Driver concentration is a crucial factor for safe driving, as it directly impacts the response time, decision-making, and ability to remain calm under stressful conditions. Drivers with high concentration levels are less likely to be involved in accidents, react more quickly to changes in road conditions, and are better equipped to handle challenging driving scenarios. By accurately assessing and enhancing driver concentration, organisations can improve safety, reduce accident rates, and enhance the efficiency of transportation systems. Identifying drivers with high concentration potential is essential for optimizing safety measures and improving driving performance.

Dataset description

The dataset utilised in this study contained 204 records that assessed driver concentration levels using three important performance metrics: concentration, response time, and composure under pressure. Each record includes 12 attributes: three concentration scores (concentration_score_1, concentration_score_2, concentration_score_3) assessing the drivers’ ability to sustain focus, four response time scores (response_time_score_1 to response_time_score_4) measuring their reaction speed under various conditions, and three keep cool scores (keep_cool_score_1, keep_cool_score_2, keep_cool_score_3) evaluating their ability to remain calm during stressful scenarios. The target variable, Category, classifies drivers into three groups: High Concentration, Moderate Concentration, and Low Concentration. The data were carefully collected, cleaned, and structured in CSV format to facilitate analysis. The descriptive statistics of the collected dataset is given in Table 1. This study aims to identify the key factors influencing driver concentration and develop a robust classification model using advanced analytical techniques to enhance driver performance and road safety.

Table 1.

Descriptive statistics of the collected dataset.

Attribute	Count	Mean	Std Dev	Min	25%	Median (50%)	75%	Max
Concentration_score_1	204	54.26	16.79	33	39.00	45.0	65.00	85
Concentration_score_2	204	54.90	17.11	34	40.75	45.5	66.00	86
Concentration_score_3	196	54.93	17.11	35	39.00	55.0	74.25	85
Response_time_score_1	204	59.52	25.64	34	41.00	56.0	76.00	150
Response_time_score_2	195	59.95	24.20	34	39.00	57.0	76.50	135
Response_time_score_3	204	54.61	16.98	34	39.75	44.5	65.00	85
Response_time_score_4	194	59.94	24.28	34	41.00	57.0	76.00	140
Keep_cool_score_1	204	59.13	23.12	34	41.00	56.0	76.25	130
Keep_cool_score_2	204	54.56	16.87	34	39.00	45.0	65.00	85
Keep_cool_score_3	195	55.62	16.74	34	41.00	54.0	74.00	85
Category	204	1.30	0.76	0	1.00	1.0	2.00	2

Data dictionary

The dataset includes ten essential attributes, each capturing a significant aspect of driver concentration and performance.

• Concentration Score 1: Reflects the driver’s ability to maintain attention and focus while driving.

• Concentration Score 2: Measures the driver’s alertness and ability to avoid distractions under varying driving conditions.

• Concentration Score 3: Assesses the driver’s overall concentration level, indicating sustained attention over time.

• Response Time Score 1: Evaluates the driver’s reaction speed to sudden changes in the driving environment, such as road hazards or traffic fluctuations.

• Response Time Score 2: Measures the driver’s quickness in reacting to unexpected events, such as pedestrians or other vehicles entering the driving space.

• Response Time Score 3: Indicates the driver’s ability to respond efficiently to unforeseen situations and potential dangers.

• Response Time Score 4: Assesses the driver’s overall response time, capturing their ability to make timely decisions at critical moments.

• Keep Cool Score 1: Reflects the driver’s capability to remain calm and composed in stressful driving situations while maintaining focus.

• Keep Cool Score 2: Measures the driver’s ability to remain resilient under pressure, ensuring effective decision-making in tense driving conditions.

• Keep Cool Score 3: Evaluates the driver’s emotional regulation and composure in high-stress environments, contributing to safe driving behaviors.

The data were organized into a well-structured dataset, forming the basis for analysis and modelling to assess and enhance driver concentration and safety.

Data pre-processing

Data pre-processing is essential to prepare the dataset for analysis and model building. To prevent bias, the missing values of the driver concentration data were initially handled by imputing them using the means of the corresponding columns. Interquartile Range (IQR) approach was used to check for outliers. It results show that there were no outliers present in the data. To ensure uniform input ranges for algorithms that are sensitive to scale, including SVM and KNN, feature scaling was used to normalize the numerical properties. The data was then divided into 70% training, 20% validation and 10% testing sets. Training set was utilized for model fitting and training, validation set was used for model optimization and testing set was used for predictions. SMOTETomek method was utilized to create synthetic data for under-represented classes to rectify the class imbalance and promote a balanced dataset. By following these procedures, the data was guaranteed to be balanced, clean, and prepared for the creation of precise models.

Data imputation

In real-world datasets, missing data is a frequent problem that can seriously affect machine learning models’ accuracy and dependability. We used a strong imputation method called Miss Forest to solve this problem. This approach makes use of RF models to forecast missing data with accuracy. The Miss Forest algorithm operates in an iterative manner:

1. Initial Imputation: A placeholder value, such as the feature’s mean or median, is first used to fill in missing values.

2. RF Training: A RF model is trained utilizing the available data for every feature that has missing values.

3. Imputation: The target feature’s missing values are predicted using the trained RF model.

4. Iteration: Steps 2 and 3 are repeated multiple times, with each iteration refining the imputed values.

Miss Forest efficiently identifies intricate patterns and connections in the data through iterative training and prediction, producing more precise imputations. This approach helps to mitigate the potential biases and inaccuracies that can arise from simpler imputation techniques like mean or median imputation.

Outlier detection

We have examined the dataset for outliers in order to guarantee the correctness and dependability of our research. Data points that substantially depart from the general trend are known as outliers, and they have the potential to skew statistical analysis. In this work, we used the Interquartile Range (IQR) approach to find outliers. Data Points are considered outliers if they fall below or above the specified limit derived from the IQR. The IQR is calculated as:

Data augmentation and class imbalance handling using SMOTETomek

The Synthetic Minority Oversampling Technique with Tomek (SMOTETomek) is a technique that combines Synthetic Minority Over-Sampling Technique (SMOTE) with Tomek method. SMOTE creates new synthetic samples for the minority class while Tomek links reduces noisy samples from the majority class. The data’s highly imbalanced classes made it unsuitable for machine learning (ML) model fitting. SMOTETomek was used to solve this problem, producing an augmented dataset. Figure 2 illustrates the uneven distribution of the dataset, with approximately 48.4% of drivers assigned to one of three groups and the remaining 32.8% and 18.9% to two other categories. After SMOTETomek was used, the dataset had even distribution of classes, with 34.1%, 32.3% and 33.5% of drivers.

Fig. 2.

Data augmentation and class imbalance handling using SMOTETomek.

According to the following equation, SMOTE creates synthetic data points along the line segments connecting the minority class samples and their nearest neighbors⁴⁴ after choosing random samples from the minority class.

The SMOTE equation for generating a synthetic sample is given by:

where: is a sample from the minority class, is one of the nearest neighbors of x_i within the minority class, is a random integer between 0 and 1, which determines the position along the line segment between and .

By varying between 0 and 1, SMOTE generates synthetic points that lie between the original sample and its neighbors, thereby increasing the representation of the minority class in the dataset.

After applying SMOTE, the Tomek method is used for undersampling of majority class. A Tomek Link is a pair of samples, and , from different classes that are each other’s nearest neighbors. For two samples (minority class) and (majority class), the distance d(, ) should be the minimum possible distance between both points. No third sample should exist such that:

d(, ) < d(, ) or d(, ) < d(, )

When the Tomek Link is identified, the majority class sample is removed to reduce noise and overlapping of data points.

Model building

We built and evaluated a number of classification models, including LR, SVM, NB, RF, AdaBoost, XGBoost, KNN, LightGBM, CatBoost, and Stacking Classifier following thorough data cleaning. To correctly categorize drivers into three groups—high, moderate, and low—based on their concentration levels, these models were trained on a pre-processed dataset. The project intends to determine the best method for forecasting driver concentration and assisting in the creation of focused interventions to enhance road safety by contrasting the performance of several machine learning models.

Logistic regression

Logistic regression⁴⁵, a versatile statistical technique, finds extensive application in recruitment, particularly for both binary and multinomial classification problems. In binary classification, logistic regression models the probability of a driver being classified as “focused” or “distracted” based on features such as response time, stress levels, and concentration scores. For multinomial classification, it extends to predicting the likelihood of a driver falling into one of several concentration categories (e.g., "high," "moderate," or “low” concentration). Despite being straightforward and easy to understand, logistic regression’s efficacy depends on the presumption that the input features and the result have a linear relationship. It can assist in determining important elements influencing focus in the context of driver monitoring and guide the development of driver-assistance technologies. However, its limitations—such as an inability to capture complex, nonlinear patterns—may require the use of more sophisticated machine learning techniques for improved accuracy.

The model computes a linear combination , applies the sigmoid function  to yield probabilities between 0 and 1, and classifies using a threshold (usually 0.5):  if , else 0.

Training minimizes log-loss (binary cross-entropy):

via gradient descent, with gradients .

In driver monitoring, it identifies key factors affecting concentration from psychometric scores and response times, supporting safety interventions. While interpretable and distribution-agnostic, it assumes linear feature-outcome relationships and struggles with nonlinear patterns—necessitating advanced ML for complex cases. This study used augmented stochastic average gradient with L2 regularization for optimal performance.

Support vector machine

Support Vector Machines (SVM)⁴⁶ provide a potent classification tool for driver concentration study. SVMs excel in driver concentration classification, categorizing drivers into levels using reaction times, stress indicators, and concentration scores. Unlike logistic regression, SVMs handle nonlinear interactions via the kernel trick, finding optimal hyperplanes in high-dimensional spaces—ideal for behavioral and psychometric data.

For training data , SVM maximizes the margin by minimizing  subject to .

For non-separable data, slack variables  yield the soft-margin objective:

with constraints , .

Common kernels:

• Linear: 

• Polynomial: 

• RBF: 

Dual problem: Maximize

subject to , .

In this driver dataset, SVMs with polynomial kernel, L2 regularization, and one-vs-one multiclass strategy effectively capture nonlinear feature interactions without strong distributional assumptions.

Naive bayes

Naive Bayes⁴⁷, a probabilistic classifier, offers a simple yet effective method for categorizing drivers based on concentration-related features. Using Bayes’ theorem, it computes the probability of a driver falling into a specific concentration level by analyzing attributes such as stress indicators, response times, and psychometric scores. Despite its simplicity, Naive Bayes performs well in large-scale applications with categorical data, such as survey responses or standardized test outcomes. However, its core assumption of feature independence can limit its accuracy in real-world driving scenarios where many variables are often interrelated. While it serves as a fast and efficient baseline model for initial concentration assessments, more advanced models may be necessary to capture the complexity of driver behavior. One notable advantage of Naive Bayes is its ability to estimate parameters with relatively small training datasets, making it practical for early-stage or data-constrained studies.

Given a set of features , the Bayes theorem determines the likelihood of a class C.

Independence assumption: 

Classification rule: 

Variants:

• Gaussian NB: 

• Multinomial NB: For discrete counts,  from frequency tables

• Bernoulli NB: For binary features, 

In the driver concentration dataset, NB efficiently handles categorical variables for initial assessments, though advanced models better capture feature correlations.

Random forest classifier

Random Forest⁴⁸, a flexible ensemble learning method that works well for classifying drivers. RF efficiently classifies and forecasts driver concentration levels based on a wide range of characteristics, such as behavioral data and reaction times, by building several decision trees and combining their predictions. This ensemble approach is ideal for situations with many and possibly connected predictors since it is very resistant to overfitting and can describe intricate, nonlinear relationships between features. The strength of RF is its capacity to manage huge, noisy, high-dimensional datasets, which makes it an effective tool for spotting drivers who may be distracted or lack focus. RF builds a collection of decision trees in relation to the driving dataset, each of which adds to the final categorization. The model gains increased accuracy and stability by integrating these separate predictions, which helps to improve road safety interventions and promote the creation of more dependable driver monitoring systems.

Several decision trees are trained on arbitrary subsets of the data in RF classification, and each tree generates a classification result on its own. The majority vote from every tree in the forest determines the final prediction for a sample. For the input x, the final prediction, y, is provided by:

where:

• is the prediction from the ith decision tree, and

• B is the total number of trees in the forest.

Approximately one-third of the data are excluded from each bootstrap sample because every tree is trained on a bootstrap sample. Out-of-bag (OOB) samples are what these are called. Using these samples for validation, the OOB error—a measure of the model’s accuracy—is computed. A RF classifier’s OOB error can be computed as follows:

where:

• N is the number of samples.

• is the true label of the ith sample.

• is the predicted label for the ith sample using only trees that did not include in their training data

• τ is an indicator function that is 1 if and 0 otherwise.

When splitting on a specific feature across all trees, RF provides a measure of feature relevance that is often determined by examining the average decrease in impurity (such as Gini impurity or entropy). Let denote a feature and the importance of , denoted as importance (), is given by:

where:

• is the decrease in impurity for feature in the ith tree.

• is the total number of trees in the forest.

RF is a versatile and powerful classifier, effective for high-dimensional data and capable of handling both classification and regression tasks. It is robust to overfitting, especially when a large number of trees are used.

AdaBoost

When it comes to driver concentration level classification, AdaBoost⁴⁹ is a potent ensemble learning technique can greatly improve the predicted accuracy of weak classifiers. The model’s overall performance is enhanced using AdaBoost, which iteratively focuses on misclassified cases and gives them more weight. When working with noisy data or when some candidate traits are hard to measure yet have a big influence on performance, this method is quite helpful.

Given a training dataset , where are the input features and are the class labels, the algorithm proceeds as follows. By iteratively modifying the weights of incorrectly classified data, the AdaBoost (Adaptive Boosting) algorithm combines several weak classifiers to create a powerful classifier. The detailed process for the AdaBoost algorithm is shown below.

AdaBoost may efficiently handle challenging classification decision-making tasks by combining several weak classifiers. This method may result in a more precise and trustworthy assessment of drivers’ levels of focus, which would eventually improve passenger safety.

XGBoost

XGBoost⁵⁰, a cutting-edge gradient boosting technique, is a powerful tool for predictive modeling. By iteratively optimizing decision trees, XGBoost effectively minimizes the loss function, resulting in highly accurate and efficient models. This technique is particularly valuable in recruitment scenarios where accurate predictions are critical.

Given a training dataset with samples and features , where and are the target labels, the model aims to predict outputs that are as close as possible to .

The prediction at the ith iteration is:

where is a function in the space of regression trees . The objective function to minimize is:

where l is a differentiable convex loss function (e.g., logistic loss for classification), and is a regularization term that penalizes the complexity of the model:

Here, T is the number of leaves in tree f, w represents the leaf weight vector, and γ and λ represent the regularization parameters. Using a second-order Taylor expansion of the loss function around the existing predictions, XGBoost makes the optimization tractable.

where

Discarding constant terms, the simplified objective function becomes:

For a tree model, we define as the leaf to which instance is assigned, and as the weight of that leaf. Therefore:

The objective function can then be expressed as:

where

and is the set of instances in leaf . Combining terms, we obtain:

To find the optimal weight for each leaf , we set the derivative of the objective function with respect to to zero:

Solving for :

Substituting back into the objective function, we obtain the optimal value:

When constructing the tree, the goal is to maximize the gain, which is the reduction in the objective function after a split. For splitting a leaf node into two child nodes (left L and right R), the gain is calculated as:

Gain = 

A split is accepted if the gain is positive, indicating an improvement in the model. XGBoost incorporates several features to enhance its performance:

• Regularization parameters and help control model complexity and prevent overfitting.

• Shrinkage (learning rate) increases the robustness of the model by scaling each tree’s contribution.

• Column subsampling introduces randomness, reducing overfitting and improving generalization.

• Handling of missing values and sparse data makes XGBoost suitable for real_world datasets.

XGBoost is a great option for predictive analytics because of its resilience to overfitting and capacity to manage big, complicated datasets. By leveraging XGBoost, this study enables data-driven decision-making to accurately identify drivers with high, moderate, or low concentration levels. Its gradient boosting framework enhances predictive accuracy by sequentially improving weak learners and minimizing errors. This makes XGBoost particularly effective for complex driver behavior data with nonlinear relationships. Ultimately, it supports the development of advanced driver monitoring systems and contributes to improving overall road safety through timely and precise concentration assessments.

K-nearest neighbours (KNN)

K-Nearest Neighbors (KNN)⁵¹ is a flexible classification technique that works well for analyzing driver concentration. By measuring the similarity between drivers based on multiple features such as response times, stress levels, and concentration scores, KNN assigns a level of focus to a new driver according to the majority class of their closest neighbors. This non-parametric method is especially helpful in situations where the underlying data distribution is complicated or uncertain. Despite providing a straightforward and understandable solution, KNN’s computational complexity may be a drawback, particularly when working with big datasets. Therefore, KNN is often used in conjunction with other techniques to enhance its performance and address its limitations.

Given a dataset , where represents the input features and displays the class labels for C classes. KNN’s objective is to forecast a new data point x’s class by using the dataset’s K nearest neighbors’ classes.

While Euclidean distance is commonly used, other distance metrics can also be applied in KNN depending on the data and the problem. Some popular distance metrics include:

Manhattan Distance:

Minkowski Distance (general form):

where is a parameter that determines the type of distance.

Cosine similarity (for similarity-based classification):

The choice of affects the performance of the KNN classifier. A small values makes the model sensitive to noise (overfitting), while a large value can smooth the decision boundary but may lead to overfitting. Cross-validation is commonly used to find the optimal value.

KNN is a simple and intuitive classifier, suitable for small datasets or when interpretability is essential. However, because distances between the test point and every training point must be calculated, it might be computationally costly for large datasets.

LightGBM

LightGBM, short for Light Gradient Boosting Machine⁵², is a framework for gradient boosting with good performance that uses decision tree techniques. It is made to be scalable and effective, particularly for high-dimensional features and big amounts of data. LightGBM’s speed and accuracy make it a popular choice for classification and regression problems.

LightGBM builds an additive model in a forward stage-wise fashion by fitting new decision trees to the residuals of the loss function. It employs innovative techniques such as Leaf-wise tree growth, Gradient-based One-Side Sampling (GOSS), and Exclusive Feature Bundling (EFB) to enhance performance.

Given a dataset with n samples and m features where and are the target labels, the goal is to build a model that predicts outputs close to

At iteration t, the prediction is:

where is a function in the space of regression trees . The objective function to minimize is:

where is a differentiable convex loss function (e.g., logistic loss for classification), and is a regularization term that penalizes the complexity of the model:

Here, is the number of leaves in the tree f, w is the vector of leaf weights, and , are regularization parameters.

To make the optimization tractable, LightGBM uses a second-order Taylor expansion of the loss function around the current predictions ;

where

Discarding constant terms, the simplified objective function becomes:

For a tree model, we define as the leaf to which instance is assigned, and as the weight of that leaf. Therefore:

The objective function can then be expressed as:

where

and is the set of instances in leaf . Combining terms, we obtain:

To find the optimal weight for each leaf , we set the derivative of the objective function with respect to to zero:

Solving for :

Substituting back into the objective function, we obtain the optimal value:

When constructing the tree, the goal is to maximize the gain, which is the reduction in the objective function after a split. For splitting a leaf node into two child nodes (left L and right R), the gain is calculated as:

Gain = 

A split is accepted if the gain is positive, indicating an improvement in the model. By combining these techniques, LightGBM achieves high efficiency and scalability without sacrificing accuracy. It is frequently used in machine learning contests and applications because it is especially well-suited for challenges involving huge datasets and high-dimensional data.

CatBoost classifier

CatBoost, short for Categorical Boosting⁵³, is a framework for gradient boosting created by Yandex. By using strategies like ordered boosting and target statistics, it is intended to manage categorical features effectively and lessen overfitting. CatBoost excels in classification and regression tasks, especially when dealing with datasets that contain categorical variables.

CatBoost builds an ensemble of decision trees by fitting new trees to the residuals of the loss function in a forward stage-wise manner. It introduces innovative methods to handle categorical data and prevent prediction shift, which can occur due to the use of target statistics.

Given a dataset with n samples and m features where and are the target labels, the goal is to build a model that predicts outputs close to

At iteration t, the prediction is:

where is a function in the space of regression trees . The objective function to minimize is:

where l is a differentiable convex loss function (e.g., logistic loss for classification), and is a regularization term that penalizes the complexity of the model:

The number of leaves in the tree f is denoted by T, the vector of leaf weights by w, and the regularization parameters by γ and λ. CatBoost employs a second-order Taylor expansion of the loss function around the existing predictions in order to make the optimization tractable.

where,

Discarding constant terms, the simplified objective function becomes:

For a tree model, we define as the leaf to which instance is assigned, and as the weight of that leaf. Therefore:

The objective function can then be expressed as:

where,

and is the set of instances in leaf . Combining terms, we obtain:

To find the optimal weight for each leaf , we set the derivative of the objective function with respect to to zero:

Solving for :

Substituting back into the objective function, we obtain the optimal value:

When constructing the tree, the goal is to maximize the gain, which is the reduction in the objective function after a split. For splitting a leaf node into two child nodes (left L and right R), the gain is calculated as:

Gain = 

If the gain is positive—a sign that the model has improved—a split is approved. By combining these methods, CatBoost lowers the chance of overfitting and offers reliable performance on datasets with categorical features. It is a useful tool in machine learning applications since it can handle categorical data without requiring a lot of preparation.

Stacking

Stacked generalization (stacking)⁵⁴ is an ensemble learning method that is aimed to enhance predictive accuracy, by aggregating several heterogeneous classifiers via a meta-learning architecture. Unlike the traditional ensemble methods i.e. bagging and boosting which use resampling or successive error reduction methods to compute the combination strategy, stacking learns the appropriate combination strategy by leveraging the complementary behaviour between independent base models. The inherent rationale of stacking is that various classifiers possess dissimilar inductive biases and error modes, and a more trained learner can actually combine these projections to decrease each bias as well as error. A stacking architecture has several base learners that are trained simultaneously on the same input feature space.

The training dataset can be defined as , where represents the input feature vector and represents the class labels. Each base classifier , for , generates probabilistic predictions . These predictions represent different decision boundaries which have been trained on the respective models including linear separability with logistic regression or nonlinear margins with SVM.

The results of the base learners are then merged to create another feature representation also known as the meta-feature space. This is what is commonly referred to as horizontal stacking and is a process that involves joining the individual predicted probabilities of the underlying classifiers to create a meta-feature vector .

The resultant meta-dataset summarizes the high-level data regarding the degree of consensus and consistency of the base models and thus adds to the representational power at the disposal of the resultant classifier. The final prediction is then generated by training a meta-learner on the built meta-feature space. Formally, the meta model learns mapping , where the aim is to minimize an appropriate loss function, e.g. cross-entropy in a classification problem. The gradient boosting classifier is selected as the meta-learner in this research because of its capacity to establish complicated nonlinear interactions amidst the predictions of the base models and its resilience to overfitting by means of progressive residual learning and shrinkage.

Mountain Gazelle optimization

Mountain Gazelle Optimization (MGO)⁵⁵ is a swarm intelligence algorithm that is based on the behavioral patterns of the mountain gazelles in the wild. The algorithm is a mathematical model of three main survival mechanisms, which are migration to fertile grazing regions, territorial grazing behavior, and evasion mechanisms of predators. The mountain gazelles (Gazella gazella) are highly adapted to survive in the abstract sense of their survival behaviors being generalised into optimization principles:

• Migration over the long distance (global exploration) to the resource-rich regions

• Local grazing on the local territory (local exploitation)

• Emerging behavior when threatened (local optima avoidance)

These actions are related to the basic optimization requirements: diversification, intensification, and avoidance of stagnation.

MGO is inspired by the social hierarchy and territorial behavior of mountain gazelles, including: Territorial solitary males (TSM), Maternity herds (MH), Bachelor male herds (BMH), Migration to search for food (MSF).

The following steps are involved in MGO:

Algorithm.

Mountain Gazelle Optimization

Where,

: Best global solution (adult male).

: Young male herd coefficient vector.

: Random gazelle from population.

: Random coefficient vectors.

: Random integers ∈ {1, 2}.

: Random number ∈ [0, 1].

: Upper and lower bounds.

: Distance vector.

: control factor given as:

random number drawn from standard normal distribution of dimension D.

The Mountain Gazelle Optimizer (MGO) was effectively applied for hyperparameter tuning in machine learning models designed to predict driver concentration levels, where selecting optimal parameters is crucial for model accuracy and real-time applicability. The results show that the difference in training and validation accuracies have minimized after optimization of the models.

Model hybridization

Model Hybridization⁵⁶ is the methodical combination of two or more learning paradigms into a single system to maximize benefits of the complementary paradigms and to reduce the weaknesses of the individual paradigms. This approach is mostly useful in the situations of complicated classification tasks that have nonlinear relationships, overlapping classification or heterogeneous distributions of features that a single algorithm cannot sufficiently predict. In a hybrid ensemble architecture, a selection of different base models is done depending on their unique learning properties and they are either trained simultaneously or in stages. Linear classifiers (including logistic regression) can be used to embed global trends and readable relationships but nonlinear models (support vector machine or tree-based learners) can be used to capture more detailed patterns and local variations. These models are hybridized in a way that the framework exploits both the linear separability and nonlinear decision surfaces, which leads to more detailed representation of actual underlying data structure. The process of hybridization usually manifests itself in the fusion mechanism that integrates the outputs of models at the feature level, decision level or the meta-learning level. Learned representation or prediction results of various base learners are horizontally appended in hybrid stacking-based methods to create an enriched meta-feature space. This hybrid feature depiction retains complimentary predictive data of both models in addition to augmenting the expressive capacity of the ultimate classifier. This space is then trained to a meta-learner to learn the best combination rules instead of using a fixed one.

The proposed hybrid model is sequential hybridization where CatBoost is initially trained as an independent base learner and the predictions of the model are stacked as an ensemble using a horizontal feature augmentation (H-Stack). By making such a design, the model can learn higher level nonlinear representations using CatBoost and stacking on them, thus enhancing the discriminative ability of the downstream ensemble.

As given in Fig. 3, the first step involves training CatBoost on original input features to find complex nonlinear relationships and high-order interactions between features with the help of symmetric decision trees and ordered boosting. After training, CatBoost produces probabilistic predictions on a case-by-case basis, which represent deep information on decisions based on underlying data distribution. These predictions can be used as informative meta-features which can summarize the learnt knowledge of CatBoost.

Fig. 3.

Hybrid model workflow.

The next step involves the horizontal concatenation of the probabilistic outputs of the CatBoost with the original feature set to create an augmented feature set. This H-stacked description maintains the original input data but introduces nonlinear abstractions derived by the CatBoost which effectively fills the gap between feature-level and decision-level learning. A stacking ensemble of Logistic Regression and Support Vector Machine are then fed with the augmented space of features as base learners. Since in the stacking system the Logistic Regression and Support Vector Machine are trained through parallel, they are trained on the CatBoost-augmented feature space. Logistic Regression represents the world linear trends, but Support Vector Machine represents the nonlinear decision boundary using the kernel-based learning. The stacking ensemble is then combined by a Gradient Boosting classifier which acts as the meta-learner of these base learners. Gradient Boosting learns the complex interactions between the LR and SVM predictions at each iteration of fixing the misleading errors and generates the final classification prediction.

Evaluation metrices

To assess the effectiveness of our classification models, we utilized a combination of precision, recall, and F1-score metrics. These metrics provide a comprehensive evaluation of a model’s performance, considering both its ability to correctly identify positive instances (recall) and its accuracy in positive predictions (precision).

Confusion Matrix⁵⁷, to assess the performance of a classification model, we employed confusion matrix. This matrix categorizes predictions into four categories

		Assignment z
		+	−
Label l	+	TP	FN
	−	FP	TN

Where + and − represent relevant and non-relevant, TP (resp. T N) represents true positive (resp. negative), and FP (resp. F N) represents false positive (resp. negative).

• Accuracy⁵⁸ is the percentage of all cases that are correctly classified. It is provided by:Accuracy = 

• Precision ()⁵⁸, Measures the proportion of positive predictions that are actually correct. A high precision indicates a low false positive rate.

• Recall ()⁵⁸, Measures the proportion of actual positive instances that are correctly identified. A high recall indicates a low false negative rate.

  
• F1-score⁵⁸ is the harmonic mean of precision and recall, providing a balanced measure of the model’s overall performance.

  

• Cohen’s kappa score⁵⁹ is a metric for two raters’ agreement that accounts for chance agreement. It is provided by:

  K = 

  where: is the observed agreement between the raters and is the probability of chance agreement, calculated as:

  

• The ROC curve⁶⁰ shows, at different threshold settings, the True Positive Rate (TPR) against the False Positive Rate (FPR). The classifier’s level of separability is shown by the AUC (Area Under the Curve). AUC is provided by:

  AUC = 

  where:

• The PR curve⁶¹ shows the relationship between precision and recall at various thresholds. This curve is useful for evaluating classifiers on imbalanced datasets. Precision and recall vary with the decision threshold, which allows the PR curve to show trade-offs between these metrics. The area under the PR curve can also be calculated for performance evaluation.

  Area under PR curve = 

  We may choose the best model for our particular use case by examining these metrics, which provide insightful information about the advantages and disadvantages of each model. A model with a high F1-score is a good fit for deployment since it successfully strikes a balance between recall and precision.

• Stratified K-fold Cross Validation is a widely adopted resampling technique for evaluating the generalization performance of machine learning models, particularly in classification problems with imbalanced class distributions.

  In k-fold cross-validation, the dataset  is partitioned into  mutually exclusive subsets, or folds, of approximately equal size. The model is trained on  folds and evaluated on the remaining fold, and this process is repeated  times so that each fold serves as the validation set exactly once. The final performance is typically summarized as the mean and standard deviation of the chosen metric (e.g., accuracy, F1-score, AUC-ROC) across all folds. Stratification modifies this procedure by enforcing that the relative frequency of each class label is preserved in every fold, which is especially important when the minority class is rare and uneven sampling could lead to biased or unstable estimates.

  For this study, stratified k-fold cross-validation was implemented with  using the same stratification strategy and fixed random seed across all configurations to ensure comparability. With , each fold contains a relatively large portion of the data, which tends to reduce the variance of the cross-validated performance at the expense of a slightly higher bias, since the training sets are smaller relative to the overall dataset size. In contrast,  is commonly regarded as a practical default because it balances computational cost, bias, and variance, yielding performance estimates that are both stable and representative. When , the validation set in each iteration becomes smaller, increasing the number of training-validation cycles and potentially reducing bias in the error estimate, but also raising the variance of fold-wise scores and the overall computational burden.

Results and discussions

Table 2 displays the results of the accuracy measures of classification models used in this work, including Random Forest, Logistic Regression, Naive Bayes, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), AdaBoost, XGBoost, LightGBM, CatBoost, and Stacking classifier. Among the individual models AdaBoost performed best in training set with an accuracy of 89%, KNN, RF, GBM, LightGBM, XGBoost, CatBoost & Stacking Classifier performed well in validation set with an accuracy of 78% and LDA, Radius Neighbors, SVC, NuSVC, GaussianNB & Stacking Classifier attained highest accuracy of 80% in the testing set.

Table 2.

Classification report of individual and hybrid models.

Model	Train						Validation						Test
	Accuracy	Kappa	precision	recall	F1	support	Accuracy	Kappa	precision	recall	F1	support	Accuracy	Kappa	precision	recall	F1	support
Logistic regression	0.75	0.61	0.75	0.75	0.75	122	0.71	0.54	0.74	0.71	0.71	41	0.73	0.58	0.77	0.73	0.74	41
LDA	0.75	0.60	0.76	0.75	0.75	122	0.73	0.58	0.78	0.73	0.74	41	0.80	0.69	0.81	0.80	0.81	41
Ridge classifier	0.75	0.61	0.76	0.75	0.75	122	0.68	0.51	0.73	0.68	0.69	41	0.73	0.58	0.75	0.73	0.74	41
KNN	0.77	0.62	0.77	0.77	0.77	122	0.78	0.65	0.80	0.78	0.78	41	0.78	0.65	0.80	0.78	0.78	41
Radius neighbors	0.80	0.67	0.79	0.80	0.79	122	0.76	0.62	0.78	0.76	0.76	41	0.80	0.69	0.81	0.80	0.81	41
SVC	0.77	0.63	0.77	0.77	0.76	122	0.73	0.58	0.78	0.73	0.74	41	0.80	0.69	0.81	0.80	0.81	41
NuSVC	0.75	0.60	0.76	0.75	0.75	122	0.73	0.58	0.78	0.73	0.74	41	0.80	0.69	0.81	0.80	0.81	41
DT	0.78	0.64	0.78	0.78	0.78	122	0.76	0.62	0.78	0.76	0.76	41	0.78	0.65	0.80	0.78	0.78	41
RF	0.82	0.70	0.82	0.82	0.82	122	0.78	0.65	0.80	0.78	0.78	41	0.78	0.65	0.80	0.78	0.78	41
GBM	0.84	0.73	0.84	0.84	0.83	122	0.78	0.65	0.80	0.78	0.78	41	0.78	0.65	0.82	0.78	0.79	41
GaussianNB	0.75	0.60	0.76	0.75	0.75	122	0.73	0.58	0.78	0.73	0.74	41	0.80	0.69	0.81	0.80	0.81	41
AdaBoost	0.89	0.82	0.89	0.89	0.89	122	0.68	0.52	0.73	0.68	0.69	41	0.71	0.55	0.74	0.71	0.71	41
LightGBM	0.79	0.65	0.79	0.79	0.78	122	0.78	0.65	0.80	0.78	0.78	41	0.78	0.65	0.80	0.78	0.78	41
XGBoost	0.80	0.66	0.80	0.80	0.79	122	0.78	0.65	0.80	0.78	0.78	41	0.78	0.65	0.80	0.78	0.78	41
CatBoost	0.78	0.64	0.78	0.78	0.78	122	0.78	0.65	0.80	0.78	0.78	41	0.78	0.65	0.80	0.78	0.78	41
Stacking	0.73	0.57	0.73	0.73	0.72	122	0.78	0.66	0.83	0.78	0.79	41	0.80	0.69	0.81	0.80	0.81	41
LDA+Stacking	0.78	0.64	0.78	0.78	0.77	122	0.76	0.61	0.79	0.76	0.76	41	0.83	0.73	0.84	0.83	0.83	41
NuSVC+Stacking	0.78	0.64	0.78	0.78	0.77	122	0.76	0.61	0.79	0.76	0.76	41	0.83	0.73	0.84	0.83	0.83	41
Radius neighbors+stacking	0.80	0.67	0.80	0.80	0.79	122	0.76	0.61	0.78	0.76	0.76	41	0.83	0.73	0.83	0.83	0.83	41
GaussianNB+stacking	0.78	0.64	0.78	0.78	0.77	122	0.76	0.61	0.79	0.76	0.76	41	0.83	0.73	0.84	0.83	0.83	41
KNN+stacking	0.76	0.62	0.76	0.76	0.76	122	0.80	0.70	0.84	0.80	0.81	41	0.83	0.73	0.84	0.83	0.83	41
CatBoost+stacking	0.80	0.67	0.80	0.80	0.80	122	0.80	0.70	0.84	0.80	0.81	41	0.83	0.73	0.85	0.83	0.83	41

As the Stacking classifier performed best in both validation and testing set, it was considered as meta learner for the hybrid model. 15 combinations of hybrid models were trained where each individual model was considered as base learner and Stacking classifier as meta learner. Among the hybrid models, LDA+Stacking, NuSVC+Stacking, Radius Neighbors+Stacking, GaussianNB+Stacking, KNN+Stacking & CatBoost+Stacking demonstrated best performance with an accuracy of 83% in the test set.

The results show that CatBoost+Stacking hybrid model has attained highest accuracy as compared to other hybrid models in all three datasets which are train, validation and test. CatBoost has an accuracy of 78% in all three datasets while, Stacking classifier has attained 73%, 78% & 80% accuracy in train, validation and test respectively. The proposed hybrid model CatBoost+Stacking outperformed the individual models in all three datasets (Training, Test and Validation of the dataset). As compared to the individual models, the hybrid models perform competitively, achieving high metrics and making it suitable for prediction of driver concentration levels.

The cross-validation accuracy comparison given in Table 3 shows that CatBoost+Stacking is the best model, achieving 0.74 (5-fold), 0.74 (10-fold), and 0.76 (20-fold), indicating strong performance and robustness. KNN+Stacking also performs well with 0.73 (5 & 10-fold) and 0.75 (20-fold), followed by Radius Neighbors+Stacking with 0.74 (5-fold), 0.72 (10-fold), and 0.75 (20-fold). In contrast, LDA, NuSVC, and Gaussian Naive Bayes hybrids show lower performance of around 0.71–0.72 across folds. Overall, stacking improves accuracy, especially with more folds, but the performance gain remains moderate, with CatBoost+Stacking as the most reliable model.

Table 3.

Cross-validation scores of hybrid models.

Model	Cross-validation accuracy scores
	5—Fold	10—Fold	20—Fold
CatBoost+stacking	0.74	0.74	0.76
KNN+stacking	0.73	0.73	0.75
Radius neighbors+stacking	0.74	0.72	0.75
LDA+stacking	0.72	0.71	0.71
NuSVC+stacking	0.72	0.71	0.71
GaussianNB+stacking	0.72	0.71	0.71

ROC curve analysis

For each model, the trade-off between true positive and false positive rates is visually represented by the Receiver Operating Characteristic (ROC) curves, which are shown in Fig. 4 for individual models and Fig. 7 for hybrid models. Sensitivity and specificity are typically better balanced in models with a larger area under the curve (AUC).

Fig. 4.

ROC for the individual models for each class.

Fig. 7.

ROC for the hybrid models for each class.

RF, LightGBM and Stacking classifier have a maximum AUC of 0.96 for class 0, while SVC and CatBoost has performed better in class 1 with AUC of 0.91, whereas with AUC of 0.96 GBM shows superior performance for class 2. As compared to the individual models, all the hybrid models have attained better performance in terms of AUC values. KNN+Stacking and CatBoost+Stacking both achieved maximum AUC in 2 of the classes. The prior hybrid model has an AUC of 0.98 & 0.92 in class 0 & 1 respectively and the latter hybrid model has an AUC of 0.98 & 0.95 in class 0 & class 2 making it the most effective hybrid model for predicting driver concentration levels .

Precision and recall curve analysis

When it comes to the performance of different models, particularly when dealing with imbalanced data classes, the precision-recall curves displayed in Figs. 6 and 9 offer important insights. Stable precision-recall profiles were demonstrated by models including RF, GBM, CatBoost & Stacking classifier which continuously maintained high precision for Classes 0 and 2 even when recall rose. These models have good F1 scores, which indicates that they are stable when dealing with class imbalances. However, models such as Ridge Classifier & NuSVC showed more errors in recall and precision, especially for Class 1.

Fig. 6.

Precision-recall curve for the individual models for each output class.

Fig. 9.

Precision-recall curve for the hybrid models for each output class.

This indicates potential difficulties in maintaining precision at higher recall levels, likely due to the simpler nature of their linear or probabilistic approaches. As compared to the individual models, the hybrid models demonstrated a more superior performance for all classes. Overall, the precision-recall analysis highlights the stable performance of CatBoost+Stacking in all 3 classes, which are particularly effective in scenarios where achieving high precision is critical.

Confusion matrix analysis

Figures 5 and 8’s confusion matrices demonstrate how well various models performed when dividing the examples into three groups. LDA, Radius Neighbors, SVC, NuSVC, GaussianNB, AdaBoost & Stacking classifier have correctly predicted 7 occurrences in class 0. KNN, DT, GBM, LightGBM, XGBoost & CatBoost show 11 correctly predicted occurrences in class 1. RF predicts 17 occurrences correctly in class 2 with minor misclassifications in class 0 & class 1. The results show that among individual models, Logistic Regression and Ridge Classifier have a higher rate of misclassifications especially in class 2, suggesting that these models would have trouble with more intricate decision boundaries, which are necessary for precise class distinction. In comparison with the individual models, the hybrid models outperformed by correctly predicting a greater number of occurrences in all the 3 classes, as evidenced by their lower misclassification rates across all classes.

Fig. 5.

Confusion matrix for the individual models.

Fig. 8.

Confusion matrix for the hybrid models.

Cross—validation plot analysis

The cross-validation plots given in Fig. 10 illustrate the performance of the top six hybrid models across folds ranging from 2 to 20. Among them, CatBoost+Stacking consistently demonstrates the best performance, with accuracy improving from around 0.71 at lower folds to a peak near 0.79 around 15–16 folds, followed by stable behavior. Similarly, KNN+Stacking shows a steady upward trend, reaching close to 0.78 at higher folds, indicating good generalization. Radius Neighbors+Stacking also performs competitively but exhibits moderate fluctuations before stabilizing around 0.74–0.76. In contrast, LDA, NuSVC, and Gaussian Naive Bayes combined with Stacking display lower accuracy levels (approximately 0.66–0.74) and higher variability, indicating less stability and weaker learning capability.

Fig. 10.

Cross-validation accuracy for the hybrid models.

Discussions

In order to determine the best categorization method for determining driver concentration levels, we examined a number of machine learning models in this study. In order to improve road safety, the main goal was to create a hybrid model that could reliably classify drivers into various concentration levels.

The initial preprocessing of data, including imputing missing values and addressing outliers, was crucial for ensuring that the models received clean and reliable data. Furthermore, the SMOTETomek technique was used to balance the distribution of classes, which improved the models’ ability to learn from the data.

Models such as Random Forest, K-Nearest Neighbors (KNN), Gradient Boosting Machines (GBM), LightGBM, XGBoost, CatBoost and Stacking demonstrated strong performances with an accuracy of 78% during validation. Among individual models, LDA, Radius Neighbors Classifier, SVC, NuSVC, Gaussian Naive Bayes and Stacking attained highest accuracy of 80% in testing. AdaBoost performed best during training with an accuracy of 88%. However, despite the strong training performance of AdaBoost, it did not match Stacking in terms of validation and test accuracy, indicating that Stacking classifier better handled the dataset’s complexity and noise.

As Stacking performed best in validation and testing, it was chosen as a meta learner for the hybrid models. The hybrid models reveal that the performance after model hybridization has improved as compared to the individual models. CatBoost+Stacking Classifier was the top-performing hybrid model when comparing the performance of different individual as well as hybrid classifiers. Despite the dataset’s complexity, this model demonstrated high generalization skills with training accuracy of 80%, validation accuracy of 80% and testing accuracy of 83%. The balanced precision, recall, and F1-scores for CatBoost+Stacking were also impressive, reflecting its consistent ability to identify both positive and negative instances accurately across training, validation and testing sets. The Kappa value of 0.73 during testing further reinforced the model’s reliability in making meaningful predictions. The proposed hybrid model CatBoost+Stacking, provide the most accurate and stable performance, with optimal results achieved at higher cross-validation folds.

Overall, the results clearly indicate that CatBoost+Stacking is the most suitable model for this application. Its ability to handle the intricacies of driver concentration data, including noise and class imbalance, makes it a powerful tool for assessing and categorizing driver behavior. By utilizing this model, stakeholders can gain valuable insights into driver concentration levels, which can lead to more effective safety interventions, potentially reducing the likelihood of accidents caused by driver inattention or fatigue.

In future work, incorporating real-time data for driver monitoring could open up possibilities for dynamic safety measures, offering a more proactive approach to improving road safety.

Conclusion

In order to evaluate driver concentration levels—which can be critical for enhancing road safety—we set out to create an effective classification model in this study. During the data preprocessing step missing values were handled by imputing with the mean using Miss Forest algorithm. IQR was used to detect outliers and there were none. The SMOTETomek approach was used to address the class imbalance. Mountain Gazelle Optimization algorithm was used for hyperparameter tuning of the machine learning models for better accuracy measures. Logistic Regression (LR), Linear Discriminant Analysis (LDA), Ridge Classifier, Decision Tree (DT), Light Gradient Boosting Machines (LightGBM), Naive Bayes (NB), Support Vector Machines (SVM), Random Forest, AdaBoost, XGBoost, CatBoost, K-Nearest Neighbors (KNN), and Stacking were among the machine learning models that were build followed by model hybridization and evaluation.

Among these models, CatBoost+Stacking became the best performer, attaining 80% accuracy in training & validation and 83% accuracy in testing. In the testing stages, the hybrid model demonstrated the best-balanced precision, recall, and F1-scores of 0.85, 0.83 and 0.83, respectively, with a Kappa value of 0.73. This demonstrates the hybrid model’s robust performance and its ability to handle the complexities of the dataset effectively.

On the other hand, Radius Neighbors+Stacking also demonstrated solid performance with training accuracy and training Kappa value of 80% and 0.67, respectively, but it showed lower validation performance. KNN+Stacking had positive outcomes as well, with validation and test accuracies 80% and 83% respectively but less train accuracy of 76% compared to CatBoost+Stacking.

Despite various models being explored and optimized, CatBoost+Stacking consistently outperformed the others, making it the most reliable and accurate solution for classifying driver concentration levels. This model can be employed to evaluate driver behavior, ultimately leading to improved safety measures on the road. By leveraging this hybrid model, the system can accurately identify levels of driver concentration, thus enabling timely interventions and improving driver safety outcomes.

The promising results of this study lay a foundation for future research linking driver concentration prediction to practical safety applications and behavior-based countermeasures.

Limitations & future work

The proposed methodology relies on a single optimization technique for model parameter tuning and performance enhancement. Although this approach yielded satisfactory and consistent results, it limits the ability to assess how alternative optimization strategies might influence model performance. Different metaheuristic optimization methods, such as Particle Swarm optimization, or Genetic Algorithm techniques often explore the solution space differently and may lead to improved convergence, stability, or generalization. The absence of a comparative analysis restricts the evaluation of the optimality of the chosen optimization technique and may affect the reproducibility of results across diverse datasets. Incorporating multiple optimization approaches and conducting a systematic comparison could provide deeper insights and strengthen the reliability of the predictive framework in future studies. Future work should evaluate the computational requirements and real-time feasibility of deploying the CatBoost+Stacking model in actual vehicle environments.

Author contributions

Sathya Narayana Sharma K.: Conceptualization; Formal analysis; Investigation; Methodology; Project administration; Supervision; Validation; Visualization; Roles/Writing—original draft; Writing—review & editing. Yashuvanthra Dhevi S., Prakash Sandhya, Madhu Venkataraman: Formal analysis; Methodology; Validation; Visualization; Roles/Writing—original draft; Writing—review & editing.

Funding

Open access funding provided by Vellore Institute of Technology. This research did not receive any specific grant from funding agencies.Open access funding provided by Vellore Institute of Technology .

Data availability

The data are taken for this study from the given link 10.6084/m9.figshare.31287400

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval

We collected secondary dataset. Hence the ethical committee approval is not required. We are confirming that the original data collection followed relevant guidelines and that informed consent was obtained by the primary investigators.