A Comparative Analysis of EEG-based Stress Detection Utilizing Machine Learning and Deep Learning Classifiers with a Critical Literature Review

Abstract

Background: Mental stress is considered to be a major contributor to different psychological and physical diseases. Different socio-economic issues, competition in the workplace and amongst the students, and a high level of expectations are the major causes of stress. This in turn transforms into several diseases and may extend to dangerous stages if not treated properly and timely, causing the situations such as depression, heart attack, and suicide. This stress is considered to be a very serious health abnormality. Stress is to be recognized and managed before it ruins the health of a person. This has motivated the researchers to explore the techniques for stress detection. Advanced machine learning and deep learning techniques are to be investigated for stress detection. 
 Methodology: A survey of different techniques used for stress detection is done here. Different stages of detection including pre-processing, feature extraction, and classification are explored and critically reviewed. Electroencephalogram (EEG) is the main parameter considered in this study for stress detection. After reviewing the state-of-the-art methods for stress detection, a typical methodology is implemented, where feature extraction is done by using principal component analysis (PCA), ICA, and discrete cosine transform. After the feature extraction, some state-of-art machine learning classifiers are employed for classification including support vector machine (SVM), K-nearest neighbor (KNN), NB, and CT. In addition to these classifiers, a typical deep-learning classifier is also utilized for detection purposes. The dataset used for the study is the Database for Emotion Analysis using Physiological Signals (DEAP) dataset.
 Results: Different performance measures are considered including precision, recall, F1-score, and accuracy. PCA with KNN, CT, SVM and NB have given accuracies of 65.7534%, 58.9041%, 61.6438%, and 57.5342% respectively. With ICA as feature extractor accuracies obtained are 58.9041%, 61.64384%, 57.5342%, and 54.79452% for the classifiers KNN, CT, SVM, and NB respectively. DCT is also considered a feature extractor with classical machine learning algorithms giving the accuracies of 56.16438%, 50.6849%, 54.7945%, and 45.2055% for the classifiers KNN, CT, SVM, and NB respectively. A conventional DCNN classification is performed given an accuracy of 76% and precision, recall, and F1-score of 0.66, 0.77, and 0.64 respectively.
 Conclusion: For EEG-based stress detection, different state-of-the-art machine learning and deep learning methods are used along with different feature extractors such as PCA, ICA, and DCT. Results show that the deep learning classifier gives an overall accuracy of 76%, which is a significant improvement over classical machine learning techniques with the accuracies as PCA+ KNN (65.75%), DCT+KNN (56.16%), and ICA+CT (61.64%).