Electroencephalogram stress level classification using k-means clustering and support vector machine

Abstract

Stress is the body’s natural reaction to life events and chronic stress disrupts the physiological equilibrium of the body which ultimately contributes to negative impact on physical and mental health. For this reason, an endeavour to develop stress level monitoring system is necessary and important to clinical intervention and diseases prevention. Various standardized questionnaires for assessing stress are available, yet they are based on individual perceptions and not subtle enough to capture mental state. Since an array of stress responses is initiated by the brain, thus, it is highly desirable to capture the stress non-invasively through neuroimaging technique, specifically electroencephalography (EEG) acquisition tool. The EEG acquisition tool was exploited in this study to capture the brainwave signals at prefrontal cortex from 50 participants and to investigate the brain states related to stress induced by virtual reality (VR) horror video and intelligence quotient (IQ) test in order to provide objective inspection of the brain functions. The collected EEG signals were pre-processed to remove artifacts and divided into four frequency bands including Delta (0.5 – 4 Hz), Theta (4 – 8 Hz), Alpha (8 – 13 Hz) and Beta (13 – 30 Hz) respectively. This was followed closely by extracting power spectral density (PSD) features from EEG frequency domain using Welch’s fast Fourier transform (FFT). In particular, absolute power of Theta, Alpha, Beta frequency bands, Alpha asymmetry and Theta/Beta power ratio were further analysed. Wilcoxon signed-rank test was carried out to find out the statistically significant features that react sensitively to stress-related changes. The results showed that Theta absolute power was significantly increased at Fp1 electrode (p<0.001) and Fp2 electrode (p<0.015) during post-IQ. Whereas Beta absolute power at Fp2 electrode was observed to significantly increase during both conditions, the post-VR (p<0.024) and post-IQ (p<0.011) respectively. However, Alpha asymmetry and Theta/Beta ratio did not significantly differ from the resting baseline. Evidently, these two parameters were indeed a good indicator of underlying bioregulatory responses especially the emotional regulation, behavioural motivation and attentional control. Following this, the significant features were selected for k-means clustering to assign the features into three groups of stress levels according to their inherent homogeneity whereby each group share similar patterns of stress response and finally, the labelled data based on clustering method were fed into support vector machine (SVM) to classify the stress level. The performance of SVM classifier was validated by 10-fold cross validation method and the result affirmed the highest performance of 98% accuracy by using only the feature of Beta-band absolute power (Fp2) on account of the significant changes of Beta activity during pre- and post-stimuli. In essence, stress pattern has been found in brain activity of Beta frequency band within right prefrontal cortex that has shown to be significantly more active under stimuli. The hybrid approach of classification using k-means clustering and SVM has been proven to be effective methods in lieu of pre-labelling the stress level to reduce individual differences in stress response, and in turn to improve the reliability and detection rate of mental stress. More future studies can be conducted to further validate and implement a stress level classification system. The system can be of assistance to support the current practice of stress diagnosis as well as be a beneficial future health indicator to improve stress management.