E-SAT: An extreme learning machine based self attention approach for decoding motor imagery EEG in subject-specific
Muhammad Ahmed Ahmed Abbasi1, Hafza Faiza Abbasi1, Xiaojun Yu1
1Northwestern Polytechnical University, 1 Dongxiang Road, Chang'an District,, Northwestern Polytechnical University, Xi'an,, Xi'an,, Shaanxi, 710129, CHINA.
Journal of Neural Engineering
|October 7, 2024
Summary
This study introduces an Extreme Learning Machine (ELM) based Self-Attention (E-SAT) mechanism to improve Brain-Computer Interface (BCI) performance for motor imagery (MI) tasks. The novel E-SAT approach significantly enhances subject-specific classification accuracy, outperforming existing state-of-the-art methods.
Area of Science:
- Neuroscience
- Machine Learning
- Biomedical Engineering
Background:
- Brain-Computer Interfaces (BCIs) enable direct communication between the brain and external devices, significantly impacting assistive technologies.
- Machine Learning (ML) and Deep Learning (DL) models have advanced BCI performance in decoding motor imagery (MI) tasks.
- Existing ML/DL models for BCIs face limitations like extensive training times and sensitivity to noise, hindering rapid development.
Purpose of the Study:
- To propose a novel Extreme Learning Machine (ELM) based Self-Attention (E-SAT) mechanism to enhance subject-specific classification performance in BCIs.
- To address limitations of existing models, including long training durations and susceptibility to noise and outliers.
- To improve the generalization ability and parameter initialization of self-attention modules within BCI systems.
Main Methods:
- Developed an E-SAT mechanism integrating ELM for improved self-attention generalization and parameter initialization.
- Employed ELM for feature extraction and classification within an end-to-end framework.
- Evaluated E-SAT performance on multiple Motor Imagery (MI) EEG signal datasets, including BCI Competition III and IV datasets.
Main Results:
- E-SAT demonstrated superior subject-specific classification performance across all tested datasets compared to state-of-the-art methods.
- Achieved high average classification accuracies: 99.8% (BCI Comp III IV-a), 99.1% (IV-b), 98.9% (BCI Comp IV 1), 75.8% (2a), 90.8% (2b), and 95.4% (3).
- E-SAT excelled in feature extraction, binary/multi-class classification, and performed robustly on noisy and non-noisy datasets.
Conclusions:
- The proposed E-SAT mechanism significantly enhances BCI performance for motor imagery tasks.
- E-SAT offers a robust and effective solution overcoming limitations of current BCI models.
- The findings highlight E-SAT's potential for advancing BCI technology in diverse applications.
Keywords:
Brain-Computer Interface (BCI)Electroencephalography (EEG)Extreme learning machine (ELM)Motor imagery (MI)Multiscale principal component analysis (MSPCA)
