Circuit Dynamics and Computational Neuroscience I.1.l Deep and machine learning Monday AM + Wednesday AM

2475 - EEG analysis without data preprocessing using convolutional neural network

Topic / Sub Topic
I.1.l Deep and machine learning
My link to connect:
https://us04web.zoom.us/j/2068848217?pwd=L0VJOHpiOVZUbW1uUWd3ZnFLeFFvQT09

Abstract

Abstract Body

Brain Computer Interface (BCI) is a system which conveys brain signals to computer and expected to control a prosthesis limb or to type words with high accuracy. Analysis of motor electroencephalography (EEG) is drawing increasing attention in BCI study because it could help people with movement disability non-invasively. To analyze motor imagery EEG, previous studies extensively employed feature extraction and/or data preprocessing which have yielded significant impacts on task classification accuracy. However overlooked information during these processes would contribute further improvements of the accuracy.

To scrutinize these possibilities, we tried to utilize machine learning methods including deep learning, which is rapidly developing in the fields of data analysis.

We adopted convolutional neural network (CNN), a widely used and very successful network for image recognition tasks. CNN enables end-to-end learning, allowing us to extract features of motor imagery EEG without data preprocessing. In current study we first examined the effective architecture of CNN for EEG decoding. Obtained CNN model achieved high accuracy (97%) with motor imagery EEG datasets recorded in the lab. We then attempted to identify the regions in EEG signal which contributed to prediction accuracy in the model, finding that the network precisely captured event related desynchronization (ERD) in the tasks. Our results suggest that CNN could be an efficient alternative to conventional signal processing in the motor imagery EEG analysis.

Hide