Exploiting pretrained CNN models for the development of an EEG-based robust BCI framework.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Automated brain-computer interface (BCI) systems require robust identification of electroencephalography (EEG) signals.
• Motor and mental imagery EEG signals are crucial for BCI applications.
• Existing methods may struggle with small or varied EEG datasets.

Purpose of the Study:

• To propose a novel automated framework for robust BCI systems using pretrained convolutional neural networks (CNNs).
• To evaluate the framework's performance with varying training data sizes (small to ample).
• To investigate the impact of different hyperparameters and data processing techniques on EEG signal identification.

Main Methods:

• A pretrained CNN-based framework was developed for motor and mental imagery EEG signal classification.
• Multiscale principal component analysis was used for denoising EEG datasets.
• Continuous wavelet transform generated time-frequency scalograms for feature extraction.
• Ten pretrained CNN models were evaluated, including ShuffleNet, with varying learning rates and optimizers.

Main Results:

• The proposed framework achieved a maximum average classification accuracy of 99.52% using ShuffleNet with an RMSProp optimizer and a low learning rate (0.0001).
• Lower learning rates demonstrated superior performance compared to higher learning rates.
• Denoised scalograms and features from tuned networks outperformed noisy scalograms and untuned networks.

Conclusions:

• Pretrained CNN models offer a robust approach for identifying EEG signals due to their ability to preserve time-frequency structures.
• The developed framework shows promising classification outcomes for BCI applications, even with limited training data.
• The study highlights the importance of data preprocessing and hyperparameter tuning for optimal BCI performance.