Decoding meditation states from EEG signals across different people remains challenging due to individual brain differences. This work compares feature engineering methods with a deep learning approach (EEGNet) for cross-subject mindfulness meditation decoding, using EEG recordings from novice and expert meditators. Riemannian Space Data Alignment (RSDA) was applied session-wise and per subject to reduce variability. EEGNet applied to RSDA-aligned signals achieved the highest decoding performance with the shortest time segments, effectively extracting relevant features for meditation state classification. These findings suggest that EEGNet can enable more effective, calibration-free neurofeedback applications for meditation.
People with meditation training report greater focus and less mind wandering during meditation than those without training. EEG recordings from 29 experienced meditators and 29 non-meditators showed that these subjective differences correspond to distinct brain activity patterns. During meditation, meditators exhibited a decrease in individual alpha frequency and amplitude and a steeper 1/f slope compared to rest, changes not seen in controls. During mind wandering, controls showed increased alpha amplitude relative to focused breathing, while meditators did not. These findings indicate that meditation experience alters both the subjective experience and the underlying oscillatory and non-oscillatory EEG properties during meditation.