Stress is defined as the reaction of the sympathetic nervous system to any type of threat, which generates a sudden release of hormones such as adrenaline and cortisol into the body [
1]. These hormones induce to the body to a state of emergency or alert, which can cause an increase in the heart rate, muscle tension, increased blood pressure, accelerated breathing, and an increased acuity of the senses [
2], among other reactions. Hence, this condition can negatively affect the daily life as well as the wellness of a person that experiences frequent stress events [
3]. In particular, automobile drivers can be affected by this condition because of negative mood, lane departure, running red lights, traffic noise, congestion, heavy traffic, lack of sleep, driving phobia, impatience, curved narrow roads, and fatigue, among other causes [
4,
5,
6,
7,
8], which can limit their concentration and ability to make reasonable decisions during any event. In consequence, stress, along with the abovementioned factors, could be an additional factor in certain car accidents that can inflict serious injuries upon those who are involved, even causing deaths [
9,
10]. In this regard, the World Health Organization reports that each year, about 1.35 million deaths are caused by road traffic crashes, where people from 5 to 29 years old are mainly involved [
11]. Hence, it is of imperative importance to develop strategies or methodologies with the capability of detecting stress in drivers in a timely manner, allowing them to take preventive actions in order to avoid car accidents and injuries, which can negatively affect the driver’s life quality as well as that of the people involved in the accident.
In recent years, machine learning algorithms have been used for detecting stress in drivers. This method has two main steps: (1) feature extraction and (2) classification or pattern recognition, as shown in
Figure 1 [
12,
13,
14,
15,
16,
17,
18,
19]. For feature extraction, the measures of the physiological signals are extracted through different methods in order to find a particular pattern that can be associated with presence of stress events in drivers, so the classification algorithm of the extracted features are used for designing and training various algorithms that can automatically recognize stress in drivers [
13]. In this sense, several researchers worldwide have presented different methods or methodologies for detecting stress in automobile drivers, which are based mainly on physiological signals such as electrocardiogram (ECG), galvanic skin response (GSR), electromyogram (EMG), or breathing rate, among others [
14,
15,
16,
17,
18,
19,
20], using machine learning-based classifiers. For example, Munla et al. [
14] developed a methodology based on wavelet transform; statistical machine learning (i.e., maximum, minimum, mean, among others) features of the heart rate variability (HRV) were obtained from ECG signals with a support vector machine (SVM) as classifier to detect whether the driver is stressed or not. The proposal used the information of 16 participants provided by the database
Stress Recognition in Automobile Drivers (SRAD) [
10], which includes recordings of ECG, GSR, and EMG signals. The authors reported that their proposal can determine when the driver suffers from stress with an accuracy of 83.33%. In this study, a preprocessing step is applied with the wavelet transform, which increases the computational load to the methodology. It should be noted that the following papers reviewed in this section use the same database. Rizwan et al. [
15] analyzed the QT (ventricular depolarization and repolarization interval) and RR intervals (measures the heartbeats rate between two cardiac cycles) of ECG signals and ECG derived respiration combined with a machine learning-based classifier, specifically the SVM, for stress detection in automobile drivers [
21]; but, in this study, the authors concluded that the more features are used, the more accurate the methodology. This condition makes the implementation in real time more difficult and it may take more time to process for delivering the stress alert if an episode is detected. Wang and Guo [
16] proposed an autoencoder classification model for driving stress recognition using ECG, HRV, foot (FGSR), and hand galvanic skin response (HGSR) signals. The obtained results show that their proposal combines the four signals for determining stressed automobile drivers with an accuracy of 97%. Chui et al. [
17] combined the convolution and cross-correlation methods with a multiple-objective genetic algorithm optimized deep multiple kernels learning SVM for recognizing automobile drivers stressed by using ECG signals, achieving an accuracy of 96.9%. Despite the obtained results, the authors mention that their proposal presents a high computational load, which would limit its application in real-time due to their long delays in the diagnosis. They reported that an accuracy of 98% is reached, distinguishing both groups. Recently, Cruz et al. [
18] investigated three features of ECG signals (i.e., ECG-derived Respiration, Respiration Rate, and QT interval) fused with the wavelet transform and machine learning based on SVM (used as classifier) for identifying automobile drivers with stress, reaching an accuracy of 96.3%. On the other hand, Wang et al. [
19] acquired 10 signal sets which include ECG, EDA, and breathing rate signals, the authors used two wearable devices to obtain these signals. They used four signals: (1) Heart Rate (HR), (2) Breathing Rate (BR), (3) HRV, and (4) Galvanic Skin Response, from which four statistical features were extracted (mean, median, first and third quartile). In this study, a preprocessing step was employed, consisting of applying a low-pass filter and a change point detection algorithm. Once this step was executed, a convolutional neural network was used to detect stress in drivers, obtaining an accuracy of 89% using the leave-one-out validation. Zontone et al. [
20] developed a methodology based on EDA and ECG signals to detect stress or no stress in car drivers. The authors obtain the ECG signal and hand EDA signal, eliminating the motion artifacts; next, the data are sent via a wireless connection to a computer. For the ECG signal, eight features were extracted, such as RR intervals, standard deviation of RR intervals, mean value of HR, and HR mean derivative value, among others. From the EDA signal, only five features were extracted, including energy, mean absolute value, mean absolute derivative, and max absolute derivative. Once the features were obtained, they were normalized for developing both SVM classifier and an Artificial Neural Network with a single input layer, two hidden layers, and a single output layer. The authors report 76.57% and 77.59% accuracy for each classifier, respectively. Despite obtaining promising results in the aforementioned works, there are some limitations that may be discussed: (1) a high computational cost, which can restrict the stress detection in real-time, and (2) they require the combination of diverse physiological signals. In addition, it is important to mention that most of these studies have been focused on analyzing ECG and EDA (electrodermal activity) signals, which can present diverse problems. For instance, the ECG signals can be affected by several heart conditions such as arrhythmias [
22]; on the other hand, EDA signals are strongly affected by sweat generated, which depends on the area of the body that is being studied [
23]. In this regard, EMG signals can offer an alternative to detect automobile drivers with stress since it presents several advantages in comparison with those mentioned above. For example, (1) EMG signals are safe, easy, and noninvasive sources of information, (2) they are capable of detecting changes in a particular muscle because of the correlation between the biochemical and physiological changes during the movements of the muscle, and (3) they are capable of isolating the muscles needed in the study (as related with stress condition) without the interaction of the nearby muscles or other muscles [
24,
25]. Nevertheless, the identification of relevant features into the EMG signals represent a challenging task in associating them with stress detection [
14]. For example, Katsis et al. [
26] investigated diverse nonlinear measurements of EMG signals (i.e., root mean square and mean value) as well as ECG, EDA, and respiration signals with a SVM for detecting automobiles drives with stress, reaching 79.3% accuracy. Fu and Wang [
27] integrated two nonlinear measurements (peak factor and maximum of cross-relation curve) with an SVM for recognizing automobile drivers with stress by using EMG and ECG signals. The authors report an accuracy of 86.7% for distinguishing automobiles drivers with stress. Recently, Wang and Guo [
16] combined a multilayer representation learning module and an ensemble classification module under the AdaBoost framework for distinguishing automobile drivers with stress using EMG signals. They mention that their proposal is capable of recognizing stressed automobile drivers with an accuracy of 58%.
On the other hand, it should be noted that with the development of deep learning frameworks, there is a new and updated version of machine learning [
12], different applications for detecting pedestrians [
28] or processing medical images for detecting mammographic lesions [
29] have been recently proposed. The former uses a body part-detector for a convolutional neural network-based classifier, reporting a reduction of the misclassification error of about 10%, whereas the latter used a modified version of the VGG neural network to perform the lesion detection using contrast, patient information, texture, and geometrical features, obtaining an area-under-the-curve (AUC) of 0.94 (the closer to 1, the better the classifier). Considering these promising results, Rastgoo et al. [
30] developed a convolutional neural network combined with a long short-term memory network to fuse an ECG with the brake, steering wheel, and gas pedals signals, as well as environmental data including distance from other vehicles and time of day for detecting stress in automobile drivers. They obtained an accuracy of 92.8%. The authors of these works indicate that their proposals require a significant amount of training data to generate a methodology with a reasonable generalization capability. In this sense, data augmentation, which is the procedure of generating new data training by slightly modifying the original data [
31], is used. Yet, to modify physiological signals, it is necessary to have a model, which in most cases is a challenging task due to the highly nonlinear nature of human organs [
32]. Considering that the most effective methodologies for detecting stress fuse the information of more than one physiological signal or, if a deep learning framework is employed, data augmentation is used to increase the database size, the necessity of developing methodologies that do not require the fuse of features extracted from physiological signals nor the increase in the database size remains an opportunity area. These methodologies, based on low-complex processing techniques, can detect features using EMG signals for achieving a high accuracy and a low-computational use classifier, allowing drivers take preventive actions in order to reduce their chances of having a car accident.
In this paper, a new methodology based on machine learning, including STFs and an SVM classifier, is presented for detecting automobile drivers who are experiencing stress using EMG signals. The methodology investigates the potential of 17 STFs (i.e., Root Mean Square (RMS), Shape Factor with RMS (SFrms), Square Mean Root (SMR), Shape Factor with SMR (SFsmr), Crest Factor (CF), Impulse Factor (IF), Latitude Factor (LF), Range (R), Mean (M), variance (Var), Standard Deviation (STD), Skewness (Sk), Kurtosis (K), 5th Moment, (5Mo), 6th Moment (6Mo), Median (Me), and Mode (Md)) for identifying relevant features or patterns in the EMG signals. Then, the calculated STFs are evaluated through the Kruskal–Wallis statistical analysis method (KWM) for determining the most discriminant STFs values. Once these values are selected, an SVM classifier is employed for determining stressed drivers automatically. In order to perform a comparative analysis between different machine learning techniques, an SVM classifier is tested with diverse kernels for getting the highest accuracy possible, and a multilayer perceptron (MLP) is also investigated, where SVM demonstrated to be the most efficient. It should be pointed out that this classifier is selected because SVM can be trained using few samples [
39]. The proposal effectiveness is validated using EMG signals acquired experimentally from 10 participants with diverse levels of stress [
10].