Meditation Styles Are Highly Discriminable from EEG at the Subject Level With Limited Generalization Across the Population: A Machine-Learning Study

EEG-based machine learning can reliably distinguish between three meditation styles (Shamatha, Vipassana, and Metta) and mind-wandering in experienced meditators when models are trained on an individual's own data, achieving high intra-subject accuracy. However, performance drops substantially when models are applied across different people, especially for distinguishing meditation styles, due to large inter-individual variability in meditation-related EEG patterns. Neural distinctions between meditation states become more pronounced over time. These findings support personalized EEG-based assessment of meditative states but highlight the difficulty of creating generalizable, subject-independent classification systems.

Study at a glance

Abstract

ABSTRACT Meditation has been associated with improvements in attention, emotional regulation, and mental well-being, motivating increasing interest in objective methods for assessing meditative states. In this study, we investigate whether EEG-based machine learning can reliably distinguish between multiple meditation styles and mind-wandering states. EEG data were recorded from experienced meditators performing three meditation styles, Shamatha, Vipassana, and Metta, together with an eyes-closed mind-wandering condition. EEG signals were preprocessed to remove artifacts, and features were extracted from frequency, time-frequency, and time domains. Classification was evaluated using both intra-subject and inter-subject strategies with multiple machine learning classifiers. Results demonstrate high intra-subject classification accuracy across meditation-versus-mind-wandering and meditation-style comparisons, indicating strongly discriminative subject-specific neural signatures. In contrast, inter-subject performance decreased substantially, particularly for distinguishing meditation styles, suggesting considerable inter-individual variability in meditation-related EEG patterns. Furthermore, temporal analysis revealed that classification performance increase over time, indicating that the neural distinctions between meditation states become increasingly pronounced over time. Additionally, t-SNE visualization showed clear within-subject clustering but increased overlap across subjects, explaining the reduced inter-subject generalization. Overall, these findings highlight the potential of EEG-based machine learning for personalized assessment and monitoring of meditative states while emphasizing the challenges of developing subject-independent meditation classification systems.