EEG Signal Classification between Audio and Visual Stimulation and Multi-level Stress Detection

Abstract

Stress in the mind and body is a reaction to perceived difficulties or dangers. It can have significant impacts on mental and physical health, affecting cognitive abilities, emotional stability, and physical well-being. To control stress and avoid its long-term impacts, it is essential to comprehend these effects. Mental stress can be detected in many ways and Electroencephalogram (EEG) is one of them. EEG is the graphical representation of Brain's electrical activity. Regular mental stress gives rise to many mental disorders and it may cause various physiological and psychological diseases. As a result, early-stage detection of stress is very important. In this research, brain activity was recorded through EEG headset during inducing different levels of stress from audio-visual stimulus. To better evaluate visual and auditory stress, an automated system is designed to differentiate among various audio and visual evoked potentials. This may further help for designing different assistive devices for the people having visual and hearing disability. Again, many researchers worked on different level of stress detection, but four level stress detection is still unchecked. In this thesis, mental arithmetic tasks were used as both audio and visual stimuli. Here, a framework was proposed to classify different levels of stress in response to audio and visual stimuli and classification was done between these two stimuli by analyzing EEG signals. A proper pre-processing pipeline selection is essential to the effectiveness of EEG-based mental stress detection systems, as noisy EEG data are unlikely to produce improved findings. An ideal EEG pre-processing pipeline was suggested in this work and compared it to the conventional approach now in use. Using the two pre-processing pipelines, raw EEG data was pre-processed which was collected at lab environment. By extracting robust features from the denoised audio and visual data, binary and multi-level stress were classified. Four machine learning (ML) classification models were used for this purpose. On top of that, the audio and visual stimuli was classified with highest 98.37% accuracy using SVM. When the suggested pipeline for pre-processing EEG data was put into practice, a noticeable increase was found in classification accuracy over the traditional approach. With SVM, the best accuracy of 97.14%, 89.01%, 89.59% were achieved for two, three and four level stress detection using visual stimulation. By contrast, the conventional pipeline produced results of 85%, 77.67% and 66.6%. Again, for auditory stimulation, the highest accuracy of 94.51%, 87.7% and 82.63% was found for two, three and four level stress detection using SVM. On the other hand, the traditional pipeline produced results of 90%, 77.46% and 71.69% respectively.