EEG-based meditation decoding: tackling subject variability with spatial and temporal alignment.

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Electroencephalography (EEG)-based neurofeedback systems are crucial for mental well-being but struggle with subject generalization.
• Existing systems often fail to accurately classify meditation and rest states in new subjects without extensive retraining.

Purpose of the Study:

• To investigate the application of spatial and spectral alignment techniques to EEG data for improved subject-independent classification of meditation and rest states.
• To assess the effectiveness of unsupervised domain adaptation methods in reducing inter-subject variability in EEG recordings.

Main Methods:

• Employed unsupervised domain adaptation techniques: Riemannian Space Data Alignment (RSDA) for spatial domain and Convolutional Monge Mapping Normalization (CMMN) for spectral domain.
• Evaluated RSDA and CMMN individually, in combination, and with z-score normalization.
• Utilized EEGNet, a convolutional neural network, with leave-one-subject-out (LOSO) cross-validation on a dataset of 53 subjects.

Main Results:

• The combined RSDA+CMMN approach achieved a significant improvement in LOSO classification accuracy (66.6%) compared to non-aligned (55.7%) and z-score normalized (59.6%) baselines.
• Spectral analysis indicated significant contributions from Theta, Alpha, and Beta frequency bands.
• Spatial analysis identified Frontopolar and Temporal regions as critical for state discrimination.

Conclusions:

• Spatial and spectral alignment of EEG data significantly enhances cross-subject generalization for meditation state classification without model retraining.
• This approach offers a practical solution for real-time neurofeedback, reducing subject variability and enabling calibration-free neurotechnology.
• The findings pave the way for more accessible and effective neurofeedback tools supporting mental well-being.