Fig 1.
Electrode placements: The electrode locations were determined based on prior foundational investigations, designating the superior auricular region for the channel, a transitional position for the ground, and the mastoid area as the reference point (a) Application of the measurement device (b) BTE EEG measurement device.
Fig 2.
Overview of Model architectures (a) Single-column architecture: a method for training the model using either EEG or GSR data (b) Multiple-column architecture: the presented model comprises two single-column models, each equipped with dedicated model backbones(either CNN or transformer-based) for their respective modalities. GSR and EEG features are extracted independently from each model backbone and then concatenated.
Fig 3.
Experiment result: An analysis of the segment-specific scores and the averaged values of respective physiological signals for rest, Session 1, Session 2, and recovery, as indicated in Fig 4.
(paired t-test, *p < 0.05, **p < 0.005).
Fig 4.
Experimental sequence: The experimental sequence, as depicted in Fig 1, began with a 30-minute period of rest and relaxation, followed by two 15-minute sessions of VR interviews intended to induce stress.
The final interview session was followed by a 20-minute recovery interval. During the rest and recovery phases, participants utilized VR equipment (MintPot Co., Ltd.) to view serene natural landscapes. The VR interviews were conducted using the “God of interview” program (MintPot Co., Ltd), simulating a virtual mock interview experience. Stress levels were assessed using a stress visual analogue scale (VAS) questionnaire administered both before and after each session.
Table 1.
The average evaluation metrics of each model for stress detection with 5-fold cross validation.
Fig 5.
AI predicted outcome analysis: Each graph’s x-axis signifies time, with green and orange markings above or below the x-axis corresponding to actual normal and stress data used for training the deep learning model (a) This denotes the GSR raw data, where the blue segment indicates the phasic component, and the gray signifies the tonic component, as determined using LEDALAB tool. (b) Analyzing the stress probability throughout the entire experimental process was conducted using a model trained with the multiple-column architecture (c) The outcomes after applying a smoothing technique, which is a post-processing method that calculates the average probability of the 5-minute window to mitigate abrupt misclassifications. In each graph, the color of the bars indicates whether the probability exceeds 0.5 (depicted in red) or falls below 0.5 (shown in blue).
Fig 6.
AI variable outcome analysis: A comprehensive outcome.
Fig 6(a)-6(c) correspond to Fig 5, sharing identical content.
Table 2.
Stress detection results after applying the smoothing technique with various time windows.
Fig 7.
Grad-CAM visualizations of GSR and EEG spectrograms associated with stress responses.
The x-axis in the images represents time intervals of 30 seconds, while the y-axis signifies frequency ranges: 1-15Hz for GSR and 1-30Hz for EEG.