Optimizing EEG Signal Classification in BCI Systems: A Comparative Study of Hybrid Transfer Learning and Advanced Machine Learning Approaches

Abstract

 Electroencephalogram (EEG) data can be challenging to classify for motor imagery (MI) tasks in brain-computer interfaces (BCI) due to low signal-to-noise ratios, complex patterns, and subject variability. This study aims to overcome these issues by evaluating various Transfer Learning (TL) and advanced machine learning models to improve EEG data classification accuracy. We pre-processed raw EEG signals into scalogram images by Continuous Wavelet Transform (CWT) and fed them into TL models like DenseNet, VGG19, ResNet, and InceptionV3. Features from these models were classified using advanced machine learning classifiers, including Random Forest, K-Nearest Neighbours, Decision Tree, and XGBoost. Using the BCI Competition IV 2a raw dataset, DenseNet combined with XGBoost achieved 99.2% classification accuracy on both training and validation datasets. According to the study, TL-based architecture can be used for controlling rehabilitation devices using EEG data for post-stroke patients, improving their quality of life and facilitating a more convenient way for individuals with severe physical disabilities to manage their healthcare via hybrid TL and advanced ML integrated BCI systems.