# A novel channel reduction concept to enhance the classification of motor imagery tasks in brain-computer interface systems

## TL;DR

This paper introduces a new method to improve the accuracy of brain-computer interfaces by selecting important EEG channels and using deep learning for motor imagery tasks.

## Contribution

A novel hybrid channel reduction method combining statistical tests and deep learning for enhanced motor imagery classification in BCI systems.

## Key findings

- The proposed method achieved over 90% accuracy across all subjects in three EEG datasets.
- It outperformed seven existing algorithms by 3.27% to 42.53% in individual subject accuracy.
- The DLRCSPNN framework showed superior performance compared to traditional CSP and NN approaches.

## Abstract

Electroencephalogram (EEG) signals play a critical role in advancing brain-computer interface (BCI) systems, particularly for detecting motor imagery (MI) movements. However, analysing large volume of EEG datasets faces some challenges due to redundant information, and performance degradation. Irrelevant channels introduce noise, which reduces accuracy and slows system performance. To address these issues, this study aims to develop a novel channel selection method to enhance EEG-based MI task performance in BCI applications. Our proposed hybrid approach combines statistical t-tests with a Bonferroni correction-based channel reduction technique, followed by the application of a Deep Learning Regularized Common Spatial Pattern with Neural Network (DLRCSPNN) framework. This framework employs DLRCSP for feature extraction and neural network (NN) algorithm for classification. Our developed method excluded channels with correlation coefficients below 0.5, retaining only significant, non-redundant channels and tested on three real-time EEG-based BCI datasets. This study produces the highest accuracy score in the case of every subjects above 90% for all the applied datasets. In the first dataset, our method achieved the highest accuracy, improving by 3.27% to 42.53% in terms of individual subject compared to seven existing machine learning algorithms. In the second and third dataset, it outperformed existing approaches, with accuracy gains of 5% to 45% and 1% to 17.47% respectively. Comparisons with a CSP and NN framework confirmed DLRCSPNN’s algorithms superior performance. These results demonstrate the effectiveness of the approach, offering a new perspective on the identification of MI task performance in EEG based BCI technology. This proposed technique will enable rapid identification of motor-disabled individuals’ intentions, supporting patient rehabilitation and improving daily living.