cs.LGOct 29, 2026

ACAM-J-MARS: Multi-Atlas Riemannian Spectral Representations for Decoding Advanced Meditation States from fMRI

Authors: Jian Wang, Winson F. Z. Yang, Wugeng Zheng, Matthew D. Sacchet

Organizations: Meditation Research Program, Psychiatry, MGH and Harvard Medical School, Boston, MA, USA · Martinos Center for Biomedical Imaging, Radiology, MGH and Harvard Medical School, Boston, MA, USA · College of Engineering, Northeastern University, Boston, MA, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Inferring jhana-type advanced concentration absorption meditation (ACAM-J) states from fMRI data is challenging due to the spatially distributed nature of meditation-related effects, substantial inter-individual variability that can exceed within-subject state differences, and limited sample sizes. Conventional machine-learning approaches often formulate this problem as unimodal, Euclidean, atlas-based classification, thereby neglecting the non-Euclidean geometry of functional connectivity matrices and underrepresenting temporal information in functional magnetic resonance imaging (fMRI); these limitations can lead to suboptimal classification performance. We propose ACAM-J-MARS (Multi-Atlas Riemannian Spectral representation), which computes functional connectivity using multiple atlases, represents symmetric positive definite (SPD) connectivity matrices on a Riemannian manifold, projects them into a log-Euclidean tangent space, and augments connectivity-derived features with low-frequency BOLD spectral energy to better resolve subtle state differences under heterogeneous variability. Evaluated on 204 segments from 22 practitioners, ACAM-J-MARS improves performance across multiple backbone models, achieving a balanced accuracy of 0.757 for early versus late ACAM-J classification and 0.545 for three-class classification, corresponding to a 10.8-point absolute improvement over the strongest baseline. These findings suggest that the proposed representation improves the decoding of advanced meditative brain states and may enable downstream applications such as closed-loop neurofeedback and real-time psychophysiological brain-computer interfaces.

Figures & tables

Figure 1: Deep neural networks cannot decode ACAM-J states across subjects. Under subject-wise cross-validation, deep models are near chance (0.44–0.55) on raw BOLD or ACAM-J-MARS features, whereas simpler classifiers perform well (nearest class mean, 0.745; linear SVM, 0.766).
Figure 2: ACAM-J-MARS. Regional BOLD signals from four atlases are converted to shrinkage-regularised SPD connectivity matrices, mapped to the log-Euclidean tangent space at the training-set Fréchet mean, and concatenated with low-frequency spectral energy. After per-subject mean removal, the resulting features can be decoded by classifiers ranging from nearest class-mean to deep neural networks.
 Representation Binary (early vs. late) Ternary (3-level)
MethodAtlasGeometrySpectralClassifier Acc.BAF1 Acc.BAF1
Unimodal Euclidean connectivity
S200 FCS200Euc.✗NCM 0.5760.5790.574 –––
S200 FCS200Euc.✗Ridge 0.6680.6490.645 –––
All-atlas FCAllEuc.✗Ridge 0.6590.6410.637 –––
Partial geometric representations
Table 1: Subject-level cross-validation for ACAM-J decoding. Rows cumulatively add ACAM-J-MARS components. Atlases: S200 = Schaefer200; All = A\mathcal{A} (four parcellations). Geometry: Euc. = Euclidean vectorisation; LE = log-Euclidean tangent space. Classifiers: NCM = nearest class mean (cosine); Ridge = ℓ2\ell_{2} -regularised linear regression. Chance: 0.500 (two-class), 0.333 (three-class). – = not tested. Best in red. Right: stage-specific regional differences vs. ACAM-J1 (MNI z=24z{=}24 mm; row-wise scale).
Figure 3: Component contributions. Subject-wise cross-validated balanced accuracy. a Binary (early vs. late): log-Euclidean tangent features beat Euclidean FC by 11.7 points. b Ternary: spectral+tangent features combine; extra atlases help only in tangent space.
Table 2: Classifier Capacity and ACAM-J-MARS Features. Deep networks: seed-averaged means. Chance is 0.333 (ternary) and 0.500 (binary).
Table 3: Robustness to ACAM-J contrasts. Subject-wise cross-validation. S200 FC: single-atlas Euclidean baseline; Ours (tangent): four-atlas log-Euclidean connectivity only; Ours (full): full ACAM-J-MARS with spectral energy and subject alignment. NCM: nearest class mean; SVM: linear support vector machine. Bars show AUC above chance (0.5).
More in this topic

Neural Decoding

Neural decoding maps brain activity or error-correction syndromes to stimuli, behavior, language, or corrective actions.

See all 187 papers

Explore similar work

CardsList

  1. Machine Learning-Based Classification of Jhana Advanced Concentrative Absorption Meditation Using 7 Tesla Functional Magnetic Resonance Imaging

    Feb 13, 2026Puneet Kumar, Winson F. Z. Yang, Alakhsimar Singh +2Functional Magnetic Resonance ImagingNeural Decoding

    School of AI and Data Engineering, Indian Institute of Technology Ropar 140001, India · Meditation Research Program, Department of Psychiatry, Massachusetts General Hospital, Harvard Medical School, Boston, MA 02129, USA · Athinoula A. Martinos Center for Biomedical Imaging, Department of Radiology, Massachusetts General Hospital, Harvard Medical School, Boston, MA 02129, USA +2

  2. Natural Image Autoencoder-Based fMRI Representations for Trait and State Prediction

    Sep 28, 2026Juhyeon Park, Yeonwoo Kim, Peter Yongho Kim +5Medical Imaging Foundation ModelsfMRI Foundation Models

    IPAI, Seoul National University · Microsoft Research · ECE, Seoul National University +1

  3. Learning Cross-Atlas Consistent Brain Disorder Representations via Disentangled Multi-Atlas Functional Connectivity Learning

    May 7, 2026Minheng Chen, Chao Cao, Jing Zhang +2Multi-View ConsistencyDisentangled Representation Learning

    Department of Computer Science and Engineering, University of Texas at Arlington, United States · School of Computing, University of Georgia, United States