Portable Drowsiness Detection through Use of a Prefrontal Single-Channel Electroencephalogram

Abstract

Drowsiness detection has been studied in the context of evaluating products, assessing driver alertness, and managing office environments. Drowsiness level can be readily detected through measurement of human brain activity. The electroencephalogram (EEG), a device whose application relies on adhering electrodes to the scalp, is the primary method used to monitor brain activity. The many electrodes and wires required to perform an EEG place considerable constraints on the movement of users, and the cost of the device limits its availability. For these reasons, conventional EEG devices are not used in practical studies and businesses. Many potential practical applications could benefit from the development of a wire-free, low-priced device; however, it remains to be elucidated whether portable EEG devices can be used to estimate human drowsiness levels and applied within practical research settings and businesses. In this study, we outline the development of a drowsiness detection system that makes use of a low-priced, prefrontal single-channel EEG device and evaluate its performance in an offline analysis and a practical experiment. Firstly, for the development of the system, we compared three feature extraction methods: power spectral density (PSD), autoregressive (AR) modeling, and multiscale entropy (MSE) for detecting characteristics of an EEG. In order to efficiently select a meaningful PSD, we utilized step-wise linear discriminant analysis (SWLDA). Time-averaging and robust-scaling were used to fit the data for pattern recognition. Pattern recognition was performed by a support vector machine (SVM) with a radial basis function (RBF) kernel. The optimal hyperparameters for the SVM were selected by the grind search method so as to increase drowsiness detection accuracy. To evaluate the performance of the detections, we calculated classification accuracy using the SVM through 10-fold cross-validation. Our model achieved a classification accuracy of 72.7% using the PSD with SWLDA and the SVM. Secondly, we conducted a practical study using the system and evaluated its performance in a practical situation. There was a significant difference (* p < 0.05) between the drowsiness-evoked task and concentration-needed task. Our results demonstrate the efficacy of our low-priced portable drowsiness detection system in quantifying drowsy states. We anticipate that our system will be useful to practical studies with aims as diverse as measurement of classroom mental engagement, evaluation of movies, and office environment evaluation.

Keywords: Karolinska sleepiness scale; drowsiness; electroencephalogram; portable system; power spectral density; single-channel; support vector machine.

Figure 1

(a) The portable electroencephalogram (EEG) device makes use of two electrodes, one of which is positioned on the left prefrontal region (Fp1) and the other of which is positioned on the left earlobe (A1). EEG data are transferred to a smartphone via Bluetooth. The portable device uses a lithium-ion rechargeable battery; (b) the international 10–20 system and the measurement points used by our device (Fp1–A1).

Figure 2

The experimental procedure, which participants repeated for 7 days. EEG recordings were obtained three times a day. A fixed cross was presented during measurement.

Figure 3

A histogram of Karolinska Sleepiness Scale (KSS) scores and number of answers.

Figure 4

Receiver operating characteristic (ROC) curves for the autoregressive (AR) model support vector machine (SVM) and step-wise linear discriminant analysis (SWLDA) SVM.

Figure 5

The practical experiment procedure, during which EEG recordings of participants were obtained throughout the simple counting task and Wisconsin Card Sorting Test (WCST) task.

Figure 6

Flow diagram of the drowsiness detection system. Parameter learning was performed on a PC. The selected parameters were loaded onto the iPad and show the predicted drowsiness level.

Figure 7

(a) Grid search result of the SVM. Classification accuracies are denoted by color. The optimal hyperparameters were $C=2^6$ and $\gamma =2^{-9}$; (b) grid search result of the SVM. Classification accuracies are denoted by color. The optimal hyperparameters were $C=2^9$ and $\gamma =2^{-3}$.

Figure 8

(a) The average drowsiness probability of the 20 subjects, as generated by the PSD (SWLDA) and SVM. Scores were significantly higher during the counting task (0.85 vs. 0.65, * p < 0.05); (b) the average drowsiness probability of the 20 subjects, as generated by the PSD (Theta, Alpha) and SVM. The average score was higher during the counting task (0.83 vs. 0.78).