EMBODY framework (brain decoding of internal attention)

Method at a glance

• Modality: multimodal (whole-brain fMRI decoding of internal attention; applied to breath meditation)
• Measures: The moment-to-moment focus of a person’s INTERNAL attention — whether they are attending to the breath, mind-wandering, or engaged in self-referential processing — read out from distributed fMRI patterns via participant-specific machine-learning classifiers, without asking the participant and without any physiological ground truth. It measures interoceptive ATTENTION (the taxonomies’ attention construct), estimated continuously, not accuracy, detection, or sensibility.
• Procedure: Three steps (Weng et al. 2020). STEP 1 (train): with eyes closed, participants are cued by 2-s auditory instructions to engage one of five internal-attention states — breath, mind-wandering (MW), self-referential processing (Self), attention to the feet, and attention to ambient scanner sounds — in a directed Internal Attention (IA) task; a penalized-logistic-regression MVPA classifier (L2, condition labels shifted 6 s for haemodynamic lag) is trained per participant with leave-one-block-out cross-validation over six blocks (2,160 brain patterns, 432/condition). STEP 2 (decode): the trained classifier is applied ACROSS TASK to a separate 10-min breath-focused meditation (600 s), assigning each second to breath/MW/Self; a ‘mental event’ is defined as three or more consecutive same-category seconds, with single incongruent decisions smoothed to the surrounding category. STEP 3 (quantify): the decoded time-series yields attention metrics — percentage time per state, number of events, mean and variance of event durations. Code open-sourced at github.com/HelenWengPhD/embodystudy.
• Also called: EMBODY, Evaluating Multivariate Maps of BODY Awareness, brain decoding of meditation states, MVPA decoding of interoceptive attention, EMBODY Task

Strengths

• Measures interoceptive ATTENTION directly and continuously, without a probe. The two standing alternatives — questionnaires (Multidimensional Assessment of Interoceptive Awareness (MAIA)) and cardiac accuracy tasks — either interrupt the practice, depend on belief, or measure a different construct. EMBODY reads attention deployment off the brain while the practice continues undisturbed.
• Produces a quantity self-report cannot: ‘percentage time in interoceptive focus’ during meditation. Experience-sampling during meditation destroys the state it samples; a decoder does not.
• Attentional TOPOGRAPHY is decodable — orienting to the breath vs to the feet gives separable patterns — so where internal attention is directed becomes an empirical question rather than an instruction (Bodily sensation maps (bodily fingerprints of emotion), body-scan claims).
• Construct validity within Step 1: trial-level classifier accuracy correlates with subjective attention ratings (breath Z = 0.41), so the decoded pattern tracks a subjectively real state.
• Participant-specific, which concentrates statistical power at the individual level and lets the framework be applied cross-sectionally or longitudinally (Smith & Little 2018 small-N logic).

Limitations

• NO GROUND TRUTH for the decoding step. Step 1 is validated on INSTRUCTED trials; Step 2 applies the classifier to an UNINSTRUCTED meditation where nothing records what was actually attended. Decoded percentages are estimates from a validated-elsewhere classifier, not verified measurements of the meditation. No concurrent experience-sampling has yet been run to check the read-out — the method’s central open problem.
• What is decoded is bounded by the labels the experimenter supplied. The five training states define the output space; the MW class in particular is likely a conglomerate (mental movement, creative thinking, other stimuli), and real practice may contain states absent from the training set.
• Classifiers are participant-specific with idiosyncratic, mostly-cortical importance maps (no region important for >10/14 participants), so there is no group model and no localizable ‘interoceptive attention’ signature — a distributed per-person fingerprint, not a coordinate.
• Decodability is partly a property of the sample: reliable in all experienced meditators but only 6/8 novices in the founding study. A population with less distinct internal states may not yield usable classifiers.
• Demonstrated on the breath. Whether the same pipeline generalizes to other interoceptive channels, other practices, or a resting baseline is untested — the founding study decoded only meditation, with no rest condition to contrast.
• Expensive and low-temporal-resolution (fMRI at TR = 1 s); the authors themselves flag EEG/MEG extension as needed for practical use.

Connections

• Key studies: Weng et al. (2020), Weng et al. (2021)
• Related methods: Interoceptive attention task, body-focused contemplative training, Ecological Momentary Assessment (EMA), Multidimensional Experience Sampling (MDES), Multidimensional Assessment of Interoceptive Awareness (MAIA)

Name collision, read carefully. This is not emBODY (topographical bodily sensation mapping), Nummenmaa’s emBODY body-colouring tool for felt bodily-sensation maps. Weng’s EMBODY (“Evaluating Multivariate Maps of BODY Awareness”) is an fMRI brain-decoding pipeline for internal attention during meditation. Same acronym, unrelated methods, different literatures. The wiki keeps them on separate pages; if you arrived looking for the emotion body-maps instrument, see emBODY (topographical bodily sensation mapping).

The measurement framework introduced by Weng et al. (2020) and open-sourced at github.com/HelenWengPhD/embodystudy. It is the wiki’s first — and so far only — method that estimates what a person’s internal attention is doing from brain activity, continuously, without asking them.

The design in one line

Teach a per-person classifier to recognise five instructed internal-attention states; then let it label, second by second, an uninstructed meditation it never saw.

Why it exists

The wiki records the same measurement impasse on Interoceptive attention task, Helen Y. Weng and contemplative training increase objective interoceptive accuracy?: interoceptive attention can be assessed only by questionnaire (belief- and demand-contaminated) or by cardiac task (a validity dispute, and the wrong construct — accuracy, not attention). Both also share a deeper problem specific to meditation: the act of measuring interrupts the thing measured. Probe a meditator to ask where their attention is and you have moved it. A decoder reads the state off the brain and leaves the practice intact.

EMBODY is the response. Its output — a continuous estimate of the focus of internal attention — is a quantity neither questionnaire nor probe can produce, and it is the taxonomies’ attention construct measured on its own terms rather than inferred from an accuracy score.

Where the evidential weight sits, and where it doesn’t

The framework’s three steps are not equally supported, and citing “EMBODY decodes interoceptive attention” flattens a real distinction:

• Step 1 is validated. Classifiers are trained and cross-validated on trials where the participant was instructed what to attend to, so above-chance accuracy (14/16 participants in the founding study) is a genuine result, and it correlates with subjective attention ratings.
• Step 2 is feasibility. The trained classifier is applied across-task to a meditation with no instruction and no recording of what was attended — so the decoded time-course has no ground truth. It is convergently plausible (the Step-3 metrics behave as sustained breath-focus should — longer breath events, not more of them) but it is not verified. The single most valuable next study is to run concurrent Ecological Momentary Assessment (EMA) against the decode.
• Step 3 is preliminary. The attention metrics are group-averaged and drawn from a small feasibility sample.

The specificity of Step 1 is engineered through its control conditions: attention to the feet (interoception, other body region) and ambient sounds (constant external distractor) are trained alongside the breath so that “breath attention” is a specific target defined partly by what it is not.

Relation to the interoceptive-attention paradigm

EMBODY’s Step 1 is a variant of the Interoceptive attention task — cue a target, hold attention on it — but the framework’s contribution is Steps 2–3, which convert a localizer (what lights up during instructed attention) into a measure (what a person is attending to when uninstructed). The Interoceptive attention task page records this as the paradigm’s most promising exit from its central weakness (no ground truth): a decoder supplies the verification an instruction cannot — for the training task. The irony the wiki should keep visible is that the decoder then inherits a new no-ground-truth problem one level up, at the meditation it is applied to.

Status

Held as a demonstrated capability, not an evidence base: one feasibility study, N = 16, breath meditation only, no resting baseline, no subjective validation of the decode. Its promise is real and specific — see Helen Y. Weng — and its cash value for the wiki’s open debates (contemplative training increase objective interoceptive accuracy?, Is more interoceptive awareness better?) is entirely prospective: nobody has yet used it to test whether contemplative training changes attention deployment.