StressNet: a Deep Learning Framework for Distinguishing Different Stress Levels in the Human Brain

Stress is a pervasive health concern that affects people across all age groups and professions. The goal of this paper is to propose a novel human stress level detection technique using EEG (Electroencephalography) signals and pulse rate (PL). First, EEG and PL of fifty subjects are recorded when they were in relaxed (no stress), mildly stressed, moderately stressed, severely stressed and critically stressed conditions. Acquired EEG and PL data of the subjects are preprocessed for the removal of noise and artifacts. Then an Explainable AI (XAI) algorithm, SHAP (Shapley Addictive exPlanations) is employed to select the relevant EEG electrodes contributing to stress detection. In the feature extraction process, eight EEG features and seven PL features are extracted. Following that, Genetic Algorithm (GA) is utilized to select the optimal set of features from the extracted EEG and PL features. A deep learning-based model, named StressNet, is designed and trained using the GA-selected optimal features. The designed StressNet model consists of a CNN module, a multihead attention module, and a proposed PLSTM (Powerful Long-Short-Term Memory) module. The StressNet model obtains a stress detection efficacy of 95.66%. The novelty of the research lies in i) designing a novel EEG and PL-based stress level detection method for classifying human stress level into five classes, ii) proposing a new architecture of LSTM (Long-Short-Term Memory) named PLSTM and using it for stress level detection, iii) utilization of a multihead attention module in the designed StressNet model for enhancing its stress level detection efficacy, and iv) utilization of SHAP for the selection of optimal EEG electrodes relevant to human stress detection.