Quantitative Assessment of Stress Through EEG During a Virtual Reality Stress-Relax Session

Abstract

Recent studies have addressed stress level classification via electroencephalography (EEG) and machine learning. These works typically use EEG-based features, like power spectral density (PSD), to develop stress classifiers. Nonetheless, these classifiers are usually limited to the discrimination of two (stress and no stress) or three (low, medium, and high) stress levels. In this study we propose an alternative for quantitative stress assessment based on EEG and regression algorithms. To this aim, we conducted a group of 23 participants (mean age 22.65 ± 5.48) over a stress-relax experience while monitoring their EEG. First, we stressed the participants via the Montreal imaging stress task (MIST), and then we led them through a 360-degree virtual reality (VR) relaxation experience. Throughout the session, the participants reported their self-perceived stress level (SPSL) via surveys. Subsequently, we extracted spectral features from the EEG of the participants and we developed individual models based on regression algorithms to predict their SPSL. We evaluated stress regression performance in terms of the mean squared percentage error (MSPE) and the correlation coefficient (R ²). The results yielded from this evaluation (MSPE = 10.62 ± 2.12, R ² = 0.92 ± 0.02) suggest that our approach predicted the stress level of the participants with remarkable performance. These results may have a positive impact in diverse areas that could benefit from stress level quantitative prediction. These areas include research fields like neuromarketing, and training of professionals such as surgeons, industrial workers, or firefighters, that often face stressful situations.

Keywords: EEG; machine learning; regression; stress; virtual reality.

FIGURE 1

Experimental procedure of the study. Initially, we conducted the participants through a 2-min eyes-closed resting state phase. Secondly, we trained the participants to learn how to use the MIST interface. Then, we conducted them through the MIST, a test designed to induce psychosocial stress via arithmetical operations and social pressure. Subsequently, the participants relaxed for 5 min via a VR experience. Finally, they completed another 2-min eyes-closed resting state period. The entire procedure spanned for approximately 18 min. Throughout the session, we performed multiple surveys to acquire the SPSL of the participants (T1–T8).

FIGURE 2

Electroencephalography processing pipeline. First, we applied a notch filter to dismiss the power line artifact. Next, we used a bandpass filter to retain only the desired EEG spectral content. Then, we resampled the filtered signal to match the expected span of the phase under processing. Subsequently, we divided the EEG signals into 2-s epochs. We performed artifact removal, detrending, and standardization at epoch level, and finally we extracted spectral features and smoothed them through a moving-average filter.

FIGURE 3

Self-perceived stress level interpolation example. Blue dots represent the responses provided by the participant to the SPSL surveys (T1–T8). Dashed blue line corresponds to the interpolation of the SPSL values. Each graph represents one of the four interpolation schemes applied. (A) Linear interpolation, (B) pchip interpolation, (C) spline interpolation, and (D) nearest interpolation.

FIGURE 4

Participant’s data structures utilized for model development. Feature matrix X corresponds to the seven spectral features that we estimated from the EEG recordings. Target array y refers to the SPSL values that we obtained from the SPSL surveys of the participants through interpolation.

FIGURE 5

Model development pipeline. First, we extracted feature matrix (X) and target array (y) from EEG recordings and SPSL surveys, respectively. Later, for every participant, we splitted these data structures into training and test set. Then, we performed grid search cross-validation to find the best set of hyperparameters for the four regressors considered. Subsequently, we refitted the regressor with best hyperparameters on the training set, and we assessed the model on the test set.

FIGURE 6

Self-perceived stress level values predicted by the best regression model (random forest with linearly interpolated data) for each participant. Red square markers represent predicted SPSL values in T1–T8. Blue round markers represent SPSL values provided by the participant in T1–T8. MSPE and R² for each participant are reported in brackets in the graph titles. For simplicity, we have displayed Y-axis only in the first graph of each row, and we have used the same scale for all the participants.

FIGURE 7

Absolute error of SPSL predictions on each survey for the best regression model (random forest with linearly interpolated data). Each bar represents the average absolute error of the SPSL predictions across the participants. Error bars indicate the SEM.