Conditional VAE for personalized neurofeedback in cognitive training.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG) signals contain valuable data for neurological assessment.
• Machine learning (ML) shows promise for analyzing complex physiological signals like EEG.
• Automated tools are needed for efficient and accurate healthcare diagnostics.

Purpose of the Study:

• To explore the efficacy of a Conditional Variational Autoencoder (CVAE) for classifying EEG signals based on health status.
• To investigate the impact of various feature extraction techniques on CVAE performance.
• To develop a robust ML model for automated healthcare diagnostics using EEG data.

Main Methods:

• Utilized two public OpenNeuro datasets comprising healthy and orthopedic impairment groups.
• Extracted six channel-wise EEG signal descriptors: STFT, HE, DFA, CD, KS-proxy, and LLE.
• Implemented a CVAE model that incorporates health labels into the encoder and decoder, alongside extracted features.

Main Results:

• The CVAE model achieved 93% accuracy, 93% precision, 93% recall, and a 0.93 F1-score on an unseen test set.
• Performance was evaluated across different feature extraction methods, highlighting the importance of feature selection.
• The CVAE model outperformed re-trained Convolutional Neural Network (CNN) baselines.

Conclusions:

• Conditional Variational Autoencoders show significant promise for robust EEG classification.
• Effective feature extraction is crucial for optimizing ML model performance in healthcare applications.
• This approach could pave the way for advanced automated diagnostic tools in healthcare.