Classification of meditative phases through EEG analysis
Maurizio Palmieri, Alejandro Luis Callara, Enzo Pasquale Scilingo, Jampa Soepa, Ngawang Sherab, Jampa Khechok, Marco Avvenuti, Francesco Marcelloni, Bruno Neri, Alessio Vecchio
Research Square July 7, 2026 preprint DOI: 10.21203/rs.3.rs-9825296/v1 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Observational study using machine-learning classification of EEG data |
|---|---|
| Sample size | 35 |
| Population | Monks at Sera Jey Monastery |
| Duration | Three 15-minute meditative segments per participant |
| Measures | classification accuracy, macro-F1 |
| Key findings | The authors report that three meditative phases (Concentrative, Loving-Kindness, and Emptiness) can be discriminated from EEG recordings in a cross-subject setting above chance, with decision trees achieving 44.67% accuracy and Random Forest 52.44% accuracy. Inter-regional coherence features, especially delta-band coherence between mid-anterior and left-central regions early in meditation, were most discriminative. They suggest that temporally structured connectivity patterns encode reproducible cross-subject signatures of closely related contemplative phases. |
Abstract
Abstract Meditation has been associated with a range of psychological benefits. However, providing objective and subject-independent markers of distinct meditative phases remains challenging, especially when different meditation practices induce subtly different brain dynamics. This study investigates whether three meditative phases, namely Concentrative, Loving-Kindness, and Emptiness, can be discriminated from electroencephalographic (EEG) recordings in a cross-subject setting using interpretable machine-learning models. We analyze EEG data recorded from 35 subjects at Sera Jey Monastery during an experimental protocol comprising multiple rest and meditation blocks, and focus on three 15-minute meditative segments for each participant. From minute-wise EEG signals, we extract global spectral features, regional features, and functional connectivity measures across nine scalp regions and five canonical frequency bands. To capture slow temporal dynamics, we compute temporal derivatives between consecutive minutes and discretize them into categorical trend features. Classification with Gini-index decision trees under a leave-one-subject-out protocol yields a mean accuracy of 44.67 % over three approximately balanced classes (chance level: 33.33 %), indicating above-chance cross-subject generalization while preserving fully interpretable decision rules. Feature-subset analysis indicates that inter-regional coherence features are the most informative family, and that delta-band coherence between mid-anterior and left-central regions during the early phase of meditation is particularly discriminative. In the same protocol, a Random Forest achieves 52.44 % accuracy and 51.57 % macro-F1 (best with coherence-only features), highlighting the expected trade-off between predictive performance and interpretability. Overall, these findings indicate that temporally structured connectivity patterns encode reproducible cross-subject signatures of closely related contemplative phases, and suggest that simple, explainable models may provide a principled foundation for future subject-independent feedback tools in meditation research and training.