Enhancing Driver Safety: Integrating ECG and CAN-Bus Data for Accurate Drowsiness Detection

Turkey road network frequently faces with severe traffic accidents, resulting in thousands of fatalities, injuries, and material damage annually. According to TSI 2022 – Road Traffic Accident Statistics [1], a total of 1,232,957 traffic accidents occurred, of which 235,176 led to deaths or injuries. These accidents resulted with 288,696 injuries and 5,229 deaths (the death toll includes those who passed away from their injuries within 30 days following the incident). These alarming statistics once again highlight significant issues related to traffic safety. Based on TSI 2022 – Road Traffic Accident Statistics [1], it was reported that driver errors caused 87% of such accidents in 2022. Although no specific study has been conducted for Turkey, it is generally known that drowsiness while driving contributes to 3% to over 30% of accidents [2]. According to data from the National Highway Traffic Safety Administration (NHTSA) in the United States, drowsiness is reported as the most common critical cause among non-performance-related faults, which account for 7% (±1.0%) of driver-related accidents [3]. Therefore, it is crucial to examine accidents caused by driver drowsiness and discuss the solutions developed to address these issues.

Drowsiness can be simply defined as the tendency to fall asleep. Sleep is generally divided into three phases: wakefulness, non-rapid eye movement (NREM) sleep, and rapid eye movement (REM) sleep. The NREM phase can be further subdivided into three stages such as: Transition from wakefulness to sleep; Light sleep and Deep sleep. According to Sahayadhas et al. [4], NREM sleep is the first stage that is commonly referred to be "driver drowsiness." According to reports of the National Highway Traffic Safety Administration (NHTSA) in the United States, drowsiness in the drivers can lead to slowed reaction time; reduced attention and decreased brain processing capacity [5]. Therefore, it has become essential to monitor drivers during their journeys to prevent accidents caused by driver drowsiness.

The topic of driver drowsiness has gained significant attention in academic literature. A bibliometric analysis of publications from 1998 to 2023 indicates that a substantial portion of this research is related to computer science disciplines, reflecting the growing interest in sensor technologies, data mining, deep learning, and facial recognition, especially after 2010 [6].

In the literature on driver drowsiness studies, the measurement methods developed for driver monitoring are categorized into three groups as vehicle-based, behavioral and physiological measurements [4, 7-13]. In vehicle-based measurements, driver drowsiness is assessed through various metrics during driving, such as lane position deviations, steering wheel movement, driving speed, accelerator pedal pressing, and vehicle yaw angle [14, 15]. These studies can be conducted using data collected from real drivers or through driving simulators. However, studies using simulators may provide misleading results as they do not fully replicate real-life conditions. For instance, when driving in bad weather or on poorly maintained roads, a driver may deviate from their lane more frequently. Similarly, a vehicle driven over rough terrain may exhibit greater variance in steering wheel movements than a simulator. As a result, it can be argued that the findings from simulator research may not accurately represent driving situations in real life and could therefore lead to inaccurate conclusions [10]. Behavioral measurements mostly rely on facial recognition technologies. Although the measurements obtained from these methods are generally reliable, the accuracy of results can vary depending on the quality of the equipment used and specific characteristics of the driver. For instance, results can be misleading when typical real-world variables like the use of sunglasses or masks, facial hair, and race-specific facial features may lead to misleading results. Eye closure analysis [13, 16, 17], Blink rate [18], Yawning analysis [19], and Facial expression analysis [11, 13, 20, 21], are frequently employed in research utilizing behavioral measurements. The physiological measurements used to assess drowsiness are most commonly based on the Electroencephalography (EEG) [22], Electrocardiography (ECG) [23], Electrooculogram (EOG) [24], Electromyogram (EMG) [25], Galvanic Skin Response (GSR) [26], and Skin Temperature [27].

Drivers' degrees of drowsiness can be accurately assessed using physiological measures. It is crucial to take into account the high cost of the tools needed for these measurements, how challenging it is to utilize them in everyday tasks, and the possibility of inaccurate data resulting from external factors.

Hybrid approaches that incorporate behavioral, physiological, and vehicle-based measures are also observed in the literature. In particular, several researches have combined ambient factors, EEG, and drowsiness detection with driver monitoring using a camera while gathering vehicle-based measures [28, 29]. However, there are currently numerous more studies on this topic due to recent developments in sensor and large data analysis techniques. Despite this, there are still gaps in the literature regarding hybrid studies. On the other hand, errors that may occur in each method can be mitigated by hybrid methods that combine multiple approaches [10]. Therefore, the primary motivation of this study is to create a highly reliable drowsiness measurement system that can be easily applied to real-life drivers and makes use of various parameters derived from the vehicle.

The research questions for this study are:

Research Question 1: Is it possible to quickly and accurately collect physiological data from drivers during driving?

Research Question 2: Can a drowsiness measurement be proposed that simultaneously considers both physiological data and vehicle-based measurements of drivers?

Research Question 3: Besides the parameters identified in previous studies on vehicle-based measurements, are there other parameters that influence driver drowsiness?

Regarding Research Question 1, physiological data will be collected through ECG data from drivers. ECG measures the bioelectrical activity of the heart. The ECG signal is the result of very low-level electrical voltages generated on the body's surface due to the depolarization of heart muscles during heartbeats. There are specific names for the waves and parameters observed in an ECG output, such as Q, R, S, T, QT, ST, RR, and QRS [30]. However, for fatigue detection, HRV (Heart Rate Variability) derived from ECG data should be considered. HRV measures the variation in time intervals between heartbeats, which reflects the regulation of the sympathetic and parasympathetic nervous systems (PSNS) and has been used in previous studies to detect fatigue [31]. HRV characteristics are examined by looking at the series of peak points of the R wave and the intervals between the RR waves. Physiological measurements will be conducted by extracting HRV values from the gathered ECG data in order to answer this research question.

To precisely determine the degree of driver drowsiness, gathering data from both physiological and vehicle-based sources while driving is the main goal in addressing Research Question 2. The study intends to use advanced classification algorithms to combine data from two different sources in order to create a system that more accurately models and predicts driver drowsiness. Because electrocardiogram (ECG) data is readily available and dependable, it will be used for physiological measures. Concurrently, information regarding to the vehicle will be gathered through the use of CAN-Bus (Controller Area Network - Bus), a communication standard found in vehicles that enables instantaneous communication between different electrical parts, including the engine, brakes, and sensors. The integration of vital vehicle components is guaranteed by CAN-Bus, which improves overall performance and safety. The present study seeks to close the current gap in the literature by utilizing a hybrid methodology for both data gathering and analysis.

To address Research Question 3, the research will collaborate with an automotive company to gather CAN-Bus data from actual drivers, which will be used to create new drowsiness detection parameters. Machine learning techniques will be employed to examine this data with the aim to find factors that exhibit a significant association with driver drowsiness.

The originality of this study lies in its potential to address gaps in the literature and contribute to driving safety. It provides a practical and applicable solution in real-life circumstances via rendering it straightforward to collect drivers' physiological data while minimizing error rates. The accuracy of techniques for identifying driver drowsiness will also be improved by the inclusion of additional parameters and the enrichment of vehicle-based measurements. Road safety will be significantly improved as a result of the development of more effective accident prevention strategies.

This study is being carried out using real-time data gathered from a broad set of drivers in a variety of environmental situations in partnership with the R&D center of a leading automotive company in Turkey. The focus of this study is on real-time data collecting from real drivers, which distinguishes it apart from previous studies in the literature that frequently use data from driving simulators. The study intends to yield more significant and accurate findings by including drivers from a variety of age groups and collecting data in a range of environmental circumstances. Consequently, the assessment outcomes are anticipated to provide more practical and effective solutions for real-world applications. Furthermore, the alert algorithm developed in this study will detect drowsiness—categorized into three classes—by analyzing specific parameters in CAN-Bus data, triggering appropriate warning systems. The findings of this study will be directly applied by the partnering R&D center to enhance driver safety technologies.

To identify the gap in the literature regarding driver drowsiness detection based on ECG values, the most recent articles from the Web of Science database, published between 2020 and 2024, have been thoroughly examined (Table 1). The results of the studies examined from the WOS database and presented in detail are assessed in the discussion section with the findings obtained from this study.

In the first part of the study, driver drowsiness and measurement methods were explained, and research questions were determined. Then, previous studies related to the research questions were reviewed and the gap in the literature was revealed. In the second part of the study, data characteristics and methods were included. The third part included the results of the study, and the fourth part included the discussion. The last part of the study consisted of the conclusion.

Table 1. Literature regarding driver drowsiness detection using ECG values

This study used a variety of machine learning models to conduct a classification analysis on a dataset gathered for a driver fatigue detection system. XGBoost, K-Nearest Neighbor (K-NN), Decision Tree, Support Vector Machine (SVM), and Random Forest were among the models that have been tested to the test. The performance of each model was assessed and contrasted with the others based on the best possible parameter combinations.

XGBoost Algorithm Results: Various parameter combinations are used in the application of the XGBoost algorithm. These parameters are the number of estimators, learning rate, maximum depth, subsample, and colsample_bytree (which controls the proportion of features used for each tree). The optimal results were obtained with 450 estimators, a learning rate of 0.01, a maximum depth of 8, and a subsample and colsample_bytree both set to 0.9, yielding a training accuracy of 0.8077, a validation accuracy of 0.7836, and a cross-validation mean score of 0.7846.

Results indicate that the model's performance for Class 0 performs the best when compared to the other classes. Only 390 instances of Class 0 were misclassified out of 13,928 instances that were correctly classified. The high recall (97%) and precision (79%) values support this, yielding an F1 score of 87%. This suggests that the model is highly proficient in identifying and accurately categorizing Class 0. The model's performance for Class 1 was notably poor with 3,173 instances misclassified compared to 1,411 correctly classified. The primary issue was confusion with 2,721 instances of Class 0 and 452 instances of Class 2, leading to a recall of 31%, precision of 82%, and a weak F1-score of 45%. Class 2 also demonstrated moderate performance, with 1,704 correct and 1,205 incorrect classifications. The F1-score for Class 2 was 64%, with a recall of 59% and precision of 71%, affected by confusion with 1,026 instances of Class 0 and 179 instances of Class 1. Overall, the model's macro-average precision was 77%, recall 62%, and F1-score 65%, indicating notable limitations, particularly with Class 1.

Despite an overall accuracy of 78%, the weighted average F1-score fell to 75% due to classification errors in Class 1 and Class 2.

The learning curves for the XGBoost model are depicted in Figure 4, where the "Learning Curves (Score)" graph indicates a slight decline in training performance over time, despite the initial high performance. The model's successful adaptation to the training data is indicated by the training score, which stays high but gradually drops. On the other hand, with more data, the validation score increases dramatically from roughly 35% to about 80%, indicating improved generalization abilities. The "Learning Curves (Error)" graph illustrates how, as more data is added, the validation error rapidly drops, indicating improved model performance and generalization, while the training error stays low with a slight increase.

4a.png

4b.png

Figure 4. XGBoost learning curves

5a.png

5b.png

Figure 5. KNN learning curves

6a.png

6b.png

Figure 6. Decision tree learning curves

7a.png

7b.png

Figure 7. SVM learning curves

8a.png

8b.png

Figure 8. Random forest learning curves

K-NN (K-Nearest Neighbor) Algorithm Results: A variety of parameter combinations, such as the number of neighbors, distance metric, p value, and weights, are used to assess the KNN algorithm's results. For instance, the model obtained some of the best results when it was trained with uniform weights and the Manhattan distance metric with five neighbors. Its accuracy values were 0.7556, 0.7556 for validation, 0.8343 for training, and 0.7661 for cross-validation.

According to the results, the model performed well in classifying Class 0, with high precision (0.79), recall (0.92), and an F1-score (0.85). However, due primarily to significant confusion with Class 0, its performance for Class 1 was noticeably worse, with low recall (0.36), F1-score (0.46), and precision (0.63). Class 2 performed moderately, with an F1-score of 0.60, a recall of 0.57, and a precision of 0.64. With a precision of 0.69, recall of 0.62, and an F1-score of 0.64 overall, the macro-average metrics showed satisfactory results for Class 0 but difficulties with Class 1 and Class 2.

The model's learning curves are shown in Figure 5. The "Learning Curves (Score)" graph indicates that training performance is high at first (about 90%), but it declines as the dataset grows, possibly due to overfitting. On the other hand, the validation score indicates improved generalization; it begins low (about 40%) but gets better and stabilizes at about 80% with additional data. The "Learning Curves (Error)" graph illustrates how the validation error starts out high (about 65%) but dramatically decreases to about 20% as the dataset grows, while the training error stays low and only slightly increases with more data. These patterns demonstrate how the model performs best on training data at first and gets better at generalizing to validation data as the volume of data increases.

Decision Tree Algorithm Results: The Decision Tree algorithm's output is assessed using various splitters, criteria, and hyperparameter configurations. For instance, the model achieved among the highest scores of 0.911, 0.773, and 0.7834 for the cross-validation mean score and training accuracy when using the Gini criterion with a maximum depth of 18, a minimum samples split of 9, and no restrictions on min samples leaf and max features.

With 13,050 accurate classifications and only 1,268 incorrect classifications on Class 0, the model performs well, yielding high precision (0.80), recall (0.93), and an F1-score of 0.86. Class 1 performs worse, with 2,084 accurate classifications but over 2,500 incorrect classifications, including 2,060 instances incorrectly classified as Class 0. Low recall (0.40), precision (0.67), and an F1-score of 0.50 result from this, indicating difficulties in correctly identifying Class 1. With 1,744 correct and 1,165 misclassified cases for Class 2, the model's performance is average; it produces a precision of 0.65, a recall of 0.58, and an F1-score of 0.61. This suggests that Class 2 is recognized more strongly than Class 1 but not as strongly as Class 0.

The model's learning curves are displayed in Figure 6, which also shows the model's consistent performance during cross-validation and training. As more samples were added, the model's training performance only slightly decreased to about 80%, as shown by the "Score" graph, indicating stable training success. After starting low at 30%, the validation score quickly increased to over 75%, indicating strong generalization and little overfitting. These results are further supported by the "Error" graph, which shows that while validation error initially started out high and eventually dropped to about 20% with additional data, training error remained low and only slightly increased. The model's strong generalization ability is highlighted by the narrow difference between training and validation errors.

SVM Algorithm Results: The results of the SVM algorithm are assessed with various gamma settings, C values, and kernel types. For instance, the model achieved the best performance among the tested configurations with the RBF kernel, C value of 2, and gamma set to scale, resulting in a training accuracy of 0.7189, validation accuracy of 0.7153, and cross-validation accuracy of 0.6581.

Model performance demonstrates remarkable performance in classifying Class 0 with 14,306 accurate classifications and only 12 incorrect classifications. Consequently, the F1-score is 0.82 and the recall is perfect at 1.00; however, the precision for Class 0 is only 0.69. On the other hand, the model does not work well for Class 1 or Class 2. Class 1 only has 403 correct classifications, a recall of 0.09, an F1-score of 0.15, and a precision of 0.56. Class 2 presents comparable problems, with 271 accurate classifications, a precision of 0.73, a recall of 0.09, and an F1-score of 0.17, demonstrating challenges in class distinction. Overall, there is minor class differentiation as evidenced by the macro-average precision, recall, and F1-score of 0.66, 0.39, and 0.38, respectively. The weighted average values indicate an overall accuracy of 0.69, but the model's performance on Class 1 and Class 2 has lowered the F1-score to 0.59.

It is evident from looking at the "Score" graph on the left side of Figure 7 that while the model performs better on the training set, its performance on the validation set is still lower than its training score. The validation score first rises quickly before decreasing at the same 70% level, while the training score stays stable at around 70%. The model fits the training data well, but it does not perform as well on the validation data, as shown by the difference in scores between training and validation. This implies that the model's ability to generalize beyond the validation data may be restricted.

Random Forest Algorithm Results: The number of estimators, max depth, min samples split, min samples leaf, and max features are among the parameter configurations for which the Random Forest algorithm results are evaluated. For instance, the model achieved some of the highest performance metrics in the dataset, including training accuracy of 0.9725, validation accuracy of 0.8346, and cross-validation mean score of 0.8427, with 350 estimators, max depth of 20, min samples split of 4, min samples leaf of 2, and max features set to sqrt.

The Random Forest model performs exceptionally well in classifying Class 0, with 13,896 correct classifications and only 422 misclassifications. This leads to a robust F1-score of 0.88, high recall (0.97), and high precision (0.81). Class 1's F1-score is 0.53, but despite having a high precision of 0.86, its recall is significantly lower at 0.39, suggesting that misclassifications occur frequently. With Class 2, the model exhibits superior performance, demonstrating a balanced recognition ability with an F1-score of 0.68, recall of 0.65, and precision of 0.73. The F1-score, recall, and macro-average precision are 0.80, 0.67, and 0.70, respectively, indicating areas that require improvement, especially for Class 1. With precision at 0.81, recall at 0.80, and an F1-score of 0.78, the weighted averages indicate acceptable overall performance. This is indicative of the model's high accuracy but poor performance for Class 1.

The Random Forest model performs exceptionally well on the training set, as shown in Figure 8. As more data is added, the model stabilizes at about 85%, indicating effective training data fitting. The model starts with a training score of about 95%. The validation score, which starts out low at roughly 30% and rises quickly to about 80%, indicates strong generalization. Figure 8's "Learning Curves (Error)" graph illustrates how validation error falls from a peak of 60% to a stable range of 20–25%, while training error stays low at 10%. These patterns suggest that there is little overfitting and good overall performance in the model, which fits the training data well and generalizes well with a small difference between training and validation scores.