Deep learning for EEG data analysis

Cheah, Kit Hwa (2018) Deep learning for EEG data analysis. Final Year Project, UTAR.

Preview

PDF Download (999Kb) | Preview

Abstract

Electroencephalogram (EEG) is a multi-dimensional time-series brain signal that is highly information packed. While an EEG has high potential to serve in medicine (e.g. disease diagnosis, prognosis, pre-disease risk identification), psycho-physiology (e.g. mood classification, stress monitoring, alertness monitoring, sleep stage monitoring), brain-computer interface application (e.g. thought typing, prosthesis control), and many other areas, the classical design of EEG feature extraction algorithms and EEG classifiers is time-consuming and challenging to fully tap into the vast data embedded in the EEG. Deep learning (or deep neural network) which enables higher hierarchical representation of complex data has been strongly suggested by a wide range of recent research that these deep architectures of artificial neural network generally outperform the classical EEG feature extraction algorithms or classical EEG classifiers. In this project, deep neural network architectures have been constructed to perform binary classification on an EEG dataset that was shown by traditional EEG feature extraction methods to have no significant difference between its two data pools (resting EEG recorded before and recorded after listening to music). The convolutional neural network (CNN) model constructed in this project has achieved a validation accuracy of 75±1% using the same EEG dataset. Using the top performing CNN architectures, short duration of relaxing music listening is found to affect the EEG signals generated by the frontal lobe more than the other lobes of the brain; and also to affect the EEG generated by the left cerebral hemisphere more than the right hemisphere.

Actions (login required)