Related Experiment Video
Updated: Oct 3, 2025
10:14
Author Spotlight: Enhancing Neurorehabilitation Through EEG, Motor Imagery, and Virtual Reality
Published on: May 10, 2024
1.2K
A transfer learning-based CNN and LSTM hybrid deep learning model to classify motor imagery EEG signals
Zahra Khademi1, Farideh Ebrahimi1, Hussain Montazery Kordy1
1Faculty of Electrical and Computer Engineering, Babol Noshirvani University of Technology, Shariati Ave., Babol, Iran.
Computers in Biology and Medicine
|February 15, 2022
Summary
This study introduces hybrid deep learning models for motor imagery Brain Computer Interfaces (BCI), achieving superior classification accuracy by combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) with transfer learning. The Inception-v3 hybrid model demonstrated the highest performance, advancing BCI research.
Area of Science:
- Neuroscience and Biomedical Engineering
- Artificial Intelligence and Machine Learning
Background:
- Motor Imagery (MI)-based Brain Computer Interfaces (BCI) translate brain signals into control commands.
- Existing MI-EEG datasets are limited in size, and signal classification faces low signal-to-noise ratio challenges.
- Deep Learning (DL) methods offer potential solutions to overcome these limitations in MI-EEG signal classification.
Purpose of the Study:
- To propose and evaluate three hybrid deep learning models for classifying electroencephalogram (EEG) signals in MI-BCI.
- To investigate the effectiveness of transfer learning and data augmentation in addressing dataset limitations.
- To compare the performance of customized CNNs against pre-trained CNNs (ResNet-50, Inception-v3) integrated with LSTM.
Main Methods:
- Developed hybrid models combining Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM).
- Employed transfer learning using pre-trained ResNet-50 and Inception-v3 models, alongside data augmentation, to enhance performance on limited datasets.
- Utilized Continuous Wavelet Transform (CWT) to generate image representations of EEG signals for CNN input and evaluated models on the BCI Competition IV dataset 2a.
Main Results:
- Achieved mean accuracies of 86% (custom CNN), 90% (ResNet-50), and 92% (Inception-v3), with corresponding Kappa values of 81%, 86%, and 88%.
- The hybrid model integrating Inception-v3 with LSTM significantly outperformed other proposed models.
- The best result improved prior state-of-the-art classification accuracy by 7%, demonstrating the efficacy of the proposed approach.
Conclusions:
- Transfer learning with pre-trained CNNs (ResNet-50, Inception-v3) combined with LSTM represents a novel and effective method for MI-BCI.
- The Inception-v3 hybrid model shows superior performance in classifying MI-EEG signals.
- Future work can optimize computational time by identifying and utilizing the most effective EEG channels and brain regions for classification.
