EEG CHANNELS SELECTION FOR PATTERN RECOGNITION OF FIVE FINGERS MOTOR IMAGERY ELECTROENCEPHALOGRAPHY SIGNALS

• Main Author: Nur Azizah, Resfyanti
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/58297
• Institution: Institut Teknologi Bandung

Brain-computer interfaces (BCI) utilize the brain's electrophysiological activities to control external devices. Brain-computer interface hand prosthetics can help people with hand disabilities to improve their quality of life. Hand prosthetic has limitation on finger movement. Research about pattern recognition on finger movement EEG-MI have been conducted as objected to being implemented on prosthetic hand BCI. However, the previous research still used tidak relevant channels to finger MI. So, channel selection algorithms are of great importance to identify relevant channels related to finger MI. It aims to support the development of a practical prosthetic hand BCI. The current study performs a channel selection method to a dataset published by Kaya et al. Dataset has EEG-MI signals from 19 EEG channels. EEG-MI has specific areas of brain activation in the SMA, part of frontal, and parietal lobes. Therefore, channel selection is performed using PCA and CSP-OVR, channel selection methods that utilize information on the spatial distribution of signals. The method is continued by a sequential search for the optimal combination of relevant channels. It aims to minimize accuracy degradation due to channel reduction. Thus, relevant channels were selected based on statistical parameters and their effect on EEG-MI pattern recognition. Each channel selection method provides different result. According to PCA method, the number of relevant channels to finger MI is eleven. Channel combinations consisting of C4, C3, Fz, Cz, T4, Fp1, P3, F7, F4, T3, dan F3. Meanwhile, according to CSP-OVR, the number of relevant channels is four, there are F8, F7, Fp2, and Fp1. Channel selection affects decreasing the accuracy of EEG-MI pattern recognition. Relevant channel combination from CSP-OVR provides an average accuracy of 46,1%. it is reduced by 0.6% compared to the classification using 19 channels. While the PCA method provides an average accuracy of 45,1%, reduced by 1% compared to the classification using 19 channels.