[Paper] Machine Learning-Based Classification of Jhana Advanced Concentrative Absorption Meditation (ACAM-J) using 7T fMRI

Source: arXiv - 2602.13008v1

Overview

This paper investigates whether patterns of brain activity captured with ultra‑high‑field (7 Tesla) functional MRI can be used to automatically recognize Jhana advanced concentration absorption meditation (ACAM‑J). By applying machine‑learning classifiers to regional homogeneity (ReHo) maps, the authors demonstrate that subtle neural signatures of deep meditative states can be distinguished from ordinary control tasks—opening a path toward objective, data‑driven monitoring of advanced meditation.

Key Contributions

• First ML‑based classification of ACAM‑J using 7 T fMRI‑derived ReHo features.
• Demonstrated 66.8 % accuracy (statistically significant, p < 0.05) for distinguishing ACAM‑J from control conditions with ensemble models.
• Identified pre‑frontal cortex and anterior cingulate cortex as the most informative regions via feature‑importance analysis, corroborating existing neuroscience literature on attention and metacognition.
• Provided a single‑case generalization test: a highly experienced meditator’s data were used to evaluate model transferability beyond the group‑level training set.
• Introduced a reproducible pipeline (preprocessing → ReHo computation → ROI feature extraction → stratified cross‑validation) that can be adapted to other altered‑consciousness states.

Methodology

1. Data collection

   • Group dataset: 20 expert meditators performed ACAM‑J and several control tasks while undergoing 7 T fMRI.
   • Single‑case dataset: One veteran practitioner completed the same tasks in a separate scanning session for out‑of‑sample testing.

2. Preprocessing & ReHo computation

   • Standard fMRI preprocessing (motion correction, slice‑time correction, spatial normalization).
   • Regional Homogeneity (ReHo): For each voxel, the Kendall’s coefficient of concordance with its 27 nearest neighbors was calculated, yielding a whole‑brain map of local synchrony.

3. Feature extraction

   • Predefined anatomical Regions of Interest (ROIs) (e.g., dorsolateral pre‑frontal cortex, anterior cingulate, insula).
   • Mean ReHo value within each ROI formed a feature vector per scan.

4. Machine‑learning pipeline

   • Tested several classifiers (logistic regression, SVM, random forest, gradient boosting, and stacking ensembles).
   • Stratified 5‑fold cross‑validation ensured balanced representation of meditation vs. control scans in each fold.
   • Model performance evaluated with accuracy, Cohen’s κ, and permutation‑based significance testing.

5. Interpretability

   • Feature‑importance scores (e.g., Gini importance for tree‑based models) highlighted which ROIs drove classification decisions.

Results & Findings

Overall, the study shows that local synchrony (ReHo) carries enough discriminative information to separate a highly trained meditative state from baseline activity, and that ensemble ML methods can exploit this signal.

Practical Implications

• Neurofeedback & training tools: Real‑time fMRI or EEG proxies could use the identified ROI patterns to give practitioners objective feedback on entering ACAM‑J, accelerating skill acquisition.
• Clinical wellness programs: Objective biomarkers for deep meditative states could help integrate meditation into mental‑health interventions, allowing clinicians to monitor adherence and efficacy.
• Brain‑computer interfaces (BCIs): The feature set (pre‑frontal & ACC ReHo) could be mapped onto more portable modalities (e.g., high‑density EEG) for on‑the‑go detection of concentration states.
• Research reproducibility: The pipeline offers a template for studying other altered‑consciousness phenomena (e.g., flow, psychedelic states) with ML‑driven classification.

Limitations & Future Work

• Sample size & diversity: Only 20 meditators and a single case were used; larger, more heterogeneous cohorts are needed to confirm generalizability.
• Temporal resolution: fMRI captures slow hemodynamic changes; integrating faster modalities (EEG/MEG) could improve real‑time detection.
• Feature scope: The study relied on ROI‑averaged ReHo; exploring whole‑brain voxel‑wise patterns or connectivity metrics may boost accuracy.
• Task design: Control conditions were relatively simple; adding more cognitively demanding baselines could test the specificity of the classifier.
• Interpretability depth: While pre‑frontal and ACC regions were highlighted, causal links between ReHo changes and subjective experience remain to be elucidated.

Future work should expand the dataset, test multimodal sensor fusion, and refine models toward clinical‑grade reliability for meditation‑based interventions.

Authors

• Puneet Kumar
• Winson F. Z. Yang
• Alakhsimar Singh
• Xiaobai Li
• Matthew D. Sacchet

Paper Information

• arXiv ID: 2602.13008v1
• Categories: cs.LG, cs.NE
• Published: February 13, 2026
• PDF: Download PDF