🤖 machine learning

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

This study demonstrates that machine learning classifiers trained on 7T fMRI-derived regional homogeneity patterns can successfully distinguish advanced concentrative absorption meditation (ACAM-J) from non-meditative states, with prefrontal and anterior cingulate regions identified as key neural markers.

Original authors: Puneet Kumar, Winson F. Z. Yang, Alakhsimar Singh, Xiaobai Li, Matthew D. Sacchet

Published 2026-02-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Puneet Kumar, Winson F. Z. Yang, Alakhsimar Singh, Xiaobai Li, Matthew D. Sacchet

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Reading Minds in Deep Meditation

Imagine your brain is a bustling city. Usually, traffic is chaotic: thoughts are cars zooming everywhere, worries are construction zones, and distractions are street performers. This is your normal, everyday mind.

Now, imagine a special group of people who have trained for decades to turn that chaotic city into a perfectly still, silent library. In Buddhist tradition, this state of deep, laser-focused silence is called Jhana (or in this study, ACAM-J). It's a state where you are fully awake but completely free from distraction, inner chatter, and even the sense of "self."

For a long time, scientists could only guess what was happening in the brain during these states because they are so rare and hard to describe. This study asked a bold question: Can we use a computer to "read" the brain and tell the difference between a normal mind and a mind in deep Jhana?

The answer is yes.

The Experiment: The 7T Super-Microscope

To do this, the researchers used a 7 Tesla MRI scanner. Think of a standard MRI as a regular camera, but this one is a super-microscope. It's so powerful it can see tiny details in the brain that normal scanners miss.

They scanned 20 expert meditators (people with thousands of hours of practice) and one "super-practitioner" who had meditated for over 20,000 hours.

The Setup:

The "Deep" State: The meditators went into their deep Jhana states.
The "Normal" State: They did normal mental tasks, like counting backward from 10,000 or remembering what they did last week.

The researchers wanted to see if the computer could look at the brain scans and say, "Ah, this is the deep meditation state," versus "This is just counting."

The Secret Sauce: "Brain Synchronization" (ReHo)

The researchers didn't just look at which parts of the brain were active. Instead, they looked at how well the neighbors were talking to each other.

Imagine a neighborhood.

Normal Brain: The neighbors are all doing different things. One is mowing the lawn, another is watching TV, and another is arguing. They aren't synchronized.
Jhana Brain: The whole neighborhood suddenly starts humming the exact same tune at the exact same time.

The researchers measured this "neighborhood harmony" (called Regional Homogeneity or ReHo). They found that during deep meditation, specific neighborhoods in the brain (like the Prefrontal Cortex, which handles focus, and the Anterior Cingulate, which handles self-control) started humming in perfect unison.

The Machine Learning Detective

This is where the Machine Learning (AI) comes in.

The researchers fed the brain data into a computer program. Think of the AI as a detective trying to learn the "fingerprint" of deep meditation.

They showed the AI thousands of examples of "Normal Brain" and "Jhana Brain."
The AI had to learn the pattern: "When the prefrontal cortex and the brainstem are humming together like this, it's Jhana."

The Results:
The AI detective got it right 67% of the time.

In the world of brain science, where signals are messy and complex, getting 67% accuracy is like a detective solving a mystery with a 2/3 chance of being right every time. It's not perfect, but it's statistically significant (meaning it's definitely not just luck).
The AI learned that the most important clues were in the areas of the brain responsible for focus and ignoring distractions.

Why This Matters: The "Neural Fingerprint"

The study found something fascinating about how the brain changes as you go deeper into meditation:

Early Stages: The brain is still dealing with some "sights and sounds" (like a quiet room with a ticking clock).
Deep Stages: The brain shuts down the "self-talk" center. The part of the brain that usually says "I am doing this" goes quiet. The brain becomes a vast, open space.

The AI could tell the difference between the "early quiet" and the "deep void." It's like the AI can tell the difference between a library where people are whispering and a library where everyone is holding their breath.

The "Phenomenology" Twist

One of the coolest parts of this study is that they didn't just rely on the brain scans. They asked the meditators to describe their experience (e.g., "How stable was your focus?" "How wide was your attention?").

They used these descriptions to clean up the data. It's like if you were trying to identify a song, but the recording had static noise. They used the meditator's description to filter out the "static" (personal quirks) and hear the pure "song" (the universal brain pattern of Jhana). This made the AI even better at spotting the real thing.

The Takeaway: What's Next?

This paper is a proof-of-concept. It's like the Wright Brothers' first flight. It proved that:

Deep meditation has a unique, measurable signature in the brain.
Computers can learn to recognize this signature.

Why does this matter for you?

Medical Breakthroughs: In the future, doctors might use this to help people with anxiety or depression. They could use a headset to see if a patient is actually entering a calm state during therapy, or use "neurofeedback" to train people to reach that state faster.
Understanding Consciousness: It gives us a scientific map of what happens when the human mind reaches its peak potential for focus and peace.

In short: The researchers built a digital "lie detector" for deep meditation. It's not perfect yet, but it successfully proved that when a human mind reaches a state of profound peace, the brain lights up in a way that a computer can actually see and understand.

1. Problem Statement

Advanced meditation states, specifically Jhana Advanced Concentrative Absorption Meditation (ACAM-J), represent profound, structured alterations in consciousness that are difficult to study using traditional neuroscientific methods. While previous studies have used univariate analyses (e.g., group-level differences) to map these states, they fail to capture the high-dimensional "neural fingerprint" required to distinguish subtle, progressive states of absorption from non-meditative controls or from each other.

The core challenges addressed in this study are:

Data Scarcity & Imbalance: Advanced meditators capable of entering deep ACAM-J states are rare, leading to small sample sizes and severe class imbalance between the few available meditation samples and control tasks.
Phenomenological Complexity: ACAM-J involves eight progressive stages with distinct phenomenological reports (e.g., shifting from form-based to formless absorption) that require precise decoding.
Generalization: Existing models struggle to generalize from group-level data to individual practitioners due to high inter-subject variability.

2. Methodology

The authors developed a tailored machine learning (ML) pipeline integrating 7T fMRI data, first-person phenomenology, and advanced data balancing techniques.

A. Data Collection & Preprocessing

Cohorts:
- Group Cohort: 20 advanced meditators (mean experience ~17.4 years, ~20,500 hours).
- Case Study: One expert practitioner (25+ years, ~20,000+ hours) providing intensive single-case data.
Imaging: 7T fMRI (SIEMENS MAGNETOM Terra) with 32-channel head coil.
Tasks:
- ACAM-J: Participants progressed through 8 stages (ACAM-J1 to J8) and an "afterglow" state.
- Control Tasks: A "Counting" task (mental countdown from 10,000) and a "Memory" task (recalling past events). Resting state was avoided to prevent inadvertent meditation.
Preprocessing: Included de-spiking, RETROICOR (physiological noise correction), motion correction, scrubbing, and band-pass filtering (0.01–0.1 Hz).
Feature Extraction: Regional Homogeneity (ReHo) was calculated to measure local functional connectivity. Data was parcellated into 498 Regions of Interest (ROIs) using four atlases (Schaefer, Tian, Bianciardi, MDTB).

B. Machine Learning Pipeline

The pipeline (Fig. 7) consists of five modules:

Data Balancing (SMOTE): To address the severe imbalance (approx. 27:1 ratio between case study and cohort data), the Synthetic Minority Over-sampling Technique (SMOTE) was applied to the group dataset. Synthetic samples were generated via interpolation, followed by controlled stochastic perturbations to ensure realistic variability without exact duplication.
Feature Selection: Recursive Feature Elimination (RFE) combined with a Random Forest model was used to prune the 498 ROIs, retaining only those with high predictive importance and theoretical relevance to attention and self-regulation.
Model Training: Six classifiers were trained using 5-fold stratified cross-validation (preserving class distribution in folds):
- Logistic Regression, Decision Trees, Random Forest, Linear SVM, k-Nearest Neighbors (kNN), and Artificial Neural Networks (ANN).
Ensemble Learning: The top three performing models were combined via probability averaging to create a robust ensemble classifier, reducing variance and overfitting.
Evaluation & Phenomenological Control:
- Models were trained on the SMOTE-balanced group data but tested on the unmodified case-study data to ensure true generalization.
- Residual Analysis: Linear regression was used to regress out variance associated with subjective phenomenological ratings (stability, width, intensity) from the ReHo values. The pipeline was re-run on these residuals to isolate neural signatures specific to the state rather than individual subjective reporting differences.

3. Key Results

A. Classification Performance

ACAM-J vs. Control (Global): The ensemble model achieved 73.19% accuracy (Cohen's $\kappa$ = 0.2604, $p < 0.05$ ) in distinguishing ACAM-J from combined control tasks.
ACAM-J vs. Specific Controls:
- vs. Counting: 66.31% accuracy ( $\kappa$ = 0.2615).
- vs. Memory: 63.94% accuracy ( $\kappa$ = 0.1908).
Pairwise Stage Classification:
- Distinguishing adjacent stages (e.g., J1 vs. J2) yielded ~64% accuracy.
- Distinguishing extreme stages (J1 vs. J6) achieved 80.30% accuracy ( $\kappa$ = 0.5158), indicating that deeper states are more neurally distinct.
Impact of Residual Analysis: Removing phenomenological variance improved overall classification accuracy from 65.87% to 66.82%, confirming that subjective variability can obscure neural signatures.

B. Neural Correlates (Feature Importance)

Feature importance analysis and lesion overlap maps identified critical brain regions driving classification:

Prefrontal Cortex (PFC) & Anterior Cingulate Cortex (ACC): Crucial for attentional regulation and metacognition.
Salience Network (Anterior Insula): Involved in detecting relevant stimuli and maintaining contact with the meditative object.
Default Mode Network (DMN): Specifically, the attenuation of medial PFC activity in later stages (J4+) correlates with reduced self-referential thought.
Locus Coeruleus (LC): Brainstem region showing increased contribution in later stages, suggesting a state of "effortless alertness" and refined arousal regulation.
Parietal Lobes: Associated with spatial awareness and attentional orienting.

4. Key Contributions

First ML Decoding of ACAM-J: This is the first study to successfully use machine learning to decode progressive stages of Jhana absorption from 7T fMRI data.
Methodological Framework for Rare States: The paper proposes a robust pipeline for studying rare cognitive states, combining SMOTE for data imbalance, RFE for interpretability, and ensemble learning for stability, specifically designed for small-N, high-dimensional neuroimaging data.
Integration of Phenomenology: By explicitly regressing out subjective phenomenological variance, the study demonstrates how to isolate objective neural markers of advanced states, bridging the gap between first-person reports and third-person neural data.
Neural Model of Absorption: The findings support a network-level model where deep absorption is characterized by:
- Heightened alertness (LC activation).
- Strengthened object-based salience (Salience Network).
- Sustained attentional stability (Control Networks).
- Reduced self-generated cognition (DMN attenuation).

5. Significance and Future Directions

Clinical & Translational Potential: The identified neural biomarkers could enable closed-loop neurofeedback systems to help practitioners stabilize concentration or detect attentional lapses. They also offer targets for non-invasive neuromodulation (TMS/tDCS) to induce or study these states.
Consciousness Research: This work provides a scalable roadmap for mapping the "upper bounds" of human consciousness, moving beyond basic meditation to complex, structured states.
Limitations & Future Work: The study is limited by the small sample size of advanced practitioners. Future research should aim to:
- Expand datasets to include more participants.
- Integrate multimodal data (EEG/MEG for temporal resolution, diffusion MRI for structural connectivity).
- Explore deep learning architectures once larger datasets become available.

In conclusion, the paper successfully demonstrates that advanced meditative states are not merely "quiet" brain states but possess distinct, decodable, and progressive neural signatures that can be captured using rigorous machine learning approaches grounded in phenomenological precision.