Application of genetic algorithm based support vector machine in selection of new EEG rhythms for drowsiness detection

Highlights

• GA-SVM is used to find the optimal rhythms for drowsiness detection.

• The original EEG signals are decomposed using wavelet packet transform.

• The drowsiness detection effect of the new rhythm was evaluated using LOSO-CV.

Abstract

The electroencephalogram (EEG) signals are important for drowsiness detection. However, in some specific application scenarios, whether there is a more accurate rhythm for drowsiness detection is worth further study. Therefore, a method of finding the optimal EEG rhythm for drowsiness detection using the genetic algorithm based support vector machine (GA-SVM) has been proposed in this study. This study used the original EEG signals in the Sleep EDF [Expanded] database for analysis and experiments. First, the original signals were divided into several epochs, and the signals of each epoch were decomposed using db10 wavelet packet transform and haar wavelet packet transform, respectively. Then, the GA-SVM was used to select the most accurate rhythm for drowsiness detection. Finally, leave-one-subject-out cross-validation (LOSO-CV) was used to evaluate the performance of each rhythm for drowsiness detection. The results show that the gamma rhythm has the best detection efficiency in the five traditional rhythms, and the accuracy rate is 80.94%. The detection accuracy of the new rhythm Rhythm (III) (43.75–48.046875 Hz) proposed in this study is 89.52%. The new rhythm proposed in this study showed bast performance in drowsiness detection.

Introduction

Drowsiness is an intermediate state between alertness and sleep. At present, a large number of studies have discussed the impact of drowsiness on public health, public safety and productivity (Abdel-Rahman et al., 2015, Dawson et al., 2014). For example, driving a car in the drowsiness situation (fatigue driving) is a important factor in causing a traffic accident (Hashemi, Saba, & Resalat, 2014). At the same time, drowsiness is also a important issue in the fields of industry and medical care (Barua et al., 2019, Borghini et al., 2014).

Current researchers have developed several methods for detecting drowsiness or alertness. Such as facial movements (Aoi, Kamijo, & Yoshida, 2011), electromyography (EMG) (Sahayadhas, Sundaraj, Murugappan, & Palaniappan, 2015), electrocardiogram (ECG) (Zhang & Liu, 2013) and electroencephalography (EEG) (Olbrich et al., 2015, Mikkelsen et al., 2019, Jap et al., 2009, da Silveira et al., 2016). Among these methods for testing drowsiness or alertness, the test results using EEG signals are the best (Hashemi et al., 2014, Johnson et al., 2011), and the advantages of EEG are more obvious for tasks that require long-term continuous detection. However, continuous visual inspection of EEG signals is a difficult task because of the large number of artifacts in EEG signals and the large differences between individuals (Nakamura, Chen, Sugi, Ikeda, & Shibasaki, 2005). Therefore, it is necessary to study an effective method to identify the drowsiness.

Drowsiness is a local phenomenon and its duration is usually no more than 15 s (Skorucak et al., 2020, Hertig-Godeschalk et al., 2019, D’Ambrosio et al., 2019). Compared with drowsiness data, overnight EEG records with sleep stages are easier to obtain. Therefore, many researchers use the data in the overnight sleep database for drowsiness detection (Bajaj et al., 2020, da Silveira et al., 2016). It has been determined in the literature that there is a special frequency range (rhythm) in EEG signals. The subject's alertness and drowsiness can be detected by the EEG signals of these rhythms (Akerstedt, Kecklund, & Knutsson, 1991). The power of the delta (δ) rhythm (0.5–4 Hz) in the sleep phase is high; and the theta (θ) rhythm (4–8 Hz) in the early stage of sleep is significantly increased. The alpha (α) rhythm (8–13 Hz) is higher in the waking state, and the α rhythm is decreasing during the process from waking state to sleep state (Kiymik, Akin, & Subasi, 2004). The beta (β) rhythm (13–22 Hz) is associated with alertness levels and is reduced during waking state to sleep state (Eoh, Chung, & Kim, 2005). Torsvall and åAkerstedt (1987) believe that the α rhythm is the most reliable frequency band for detecting drowsiness, followed by the θ and δ rhythms.

Garcés Correa et al. (2014) developed an automatic detection method based on artificial neural network for drowsiness. 19 features based on the EEG signals of five traditional rhythms, δ, θ, α, β and γ, were proposed. At the same time, these 19 features are used to distinguish between alertness and drowsiness. The method gets 87.4% and 83.6% of alertness and drowsiness correct detection rates, respectively. Gurudath and Riley (2014) used Debauchies 3 (db3) wavelet basic to perform discrete wavelet transform on EEG signals, extracted EEG signals of five traditional rhythms, and calculated the mean, median, variance, standard deviation, etc. of each rhythm. At the same time, K-means clustering was used for drowsiness detection. da Silveira et al. (2016) used wavelet packet transform (WPT) to extract five traditional rhythms of EEG signals. According to the characteristics of the power spectrum of different rhythms, two indexes for drowsiness detection were proposed. The results show that the indexes based on the power spectrum of the five traditional rhythms for drowsiness detection have a good effect, and also show that drowsiness can be detected using a single channel EEG signal.

Currently, in drowsiness detection based on EEG signals, most researchers used five traditional rhythms, δ, θ, α, β, and γ, but few researchers have studied whether there are frequency bands with higher accuracy for drowsiness detection. Some researchers have found that in some cases the new rhythm has a good effect for drowsiness detection, but they do not have a perfect way to find the optimal rhythm (Belakhdar, Kaaniche, Djemal, & Ouni, 2018). Therefore, how to find the optimal rhythm for drowsiness detection requires further research.

This study proposed a method based on the genetic algorithm based support vector machine (GA-SVM) to find the optimal rhythm for drowsiness detection. Then, using EEG signals in the sleep EDF [Expanded] database as the training set, this method was used to find a high-accuracy new rhythm for drowsiness detection. Finally, the performance of each rhythm for drowsiness detection was evaluated using leave-one-subject-out cross-validation (LOSO-CV) (Altmann et al., 2016).

The remainder of this paper unfolds as follows. Section 2 introduces the algorithms and data used in this paper and proposes genetic algorithm based support vector machine. Section 3 shows the experimental results of this paper and discusses them. Section 4 discusses the limitations of the current approach. Section 5 concludes this paper.