Weng et al. (2020) — Focus on the Breath: brain decoding of internal attention during meditation
The primary source for a method the wiki has been carrying secondhand since the Weng et al. (2021) ingest, where “brain decoding of interoceptive attention” appeared as one line in a review and the wiki flagged, on helen-weng and interoceptive-attention-task, that it “has no primary source to check it against.” This is that source. It is the paper behind the recurring claim that a classifier can estimate the percentage of time a meditator spends attending to the breath without asking them and without counting a heartbeat.
Read here for what it actually establishes, which is narrower and more interesting than the one-line citation suggests: not that we can read a meditator’s mind, but that we can build a per-person brain-pattern detector for a handful of instructed internal states, and that the detector, turned loose on an uninstructed meditation, produces attention statistics that behave the way sustained breath-focus should.
What EMBODY is
The framework — Evaluating Multivariate Maps of BODY awareness — is a three-step pipeline, and the three steps are worth keeping distinct because they carry very different evidential weight. See embody-framework for the method page.
| step | what happens | how well-supported |
|---|---|---|
| 1. Train | Participant-specific MVPA classifiers learn five internal-attention states from a directed IA task (breath, MW, self-referential, feet, sounds), with eyes closed so the only thing changing is the focus of attention | Strong — cross-validated, above chance in 14/16, ratings-correlated |
| 2. Decode | The Step-1 classifiers are applied across task to a separate 10-min breath meditation, labelling each second as breath / MW / self | Feasibility only — no ground truth for what was actually attended |
| 3. Quantify | The decoded time-series is turned into attention metrics: % time per state, number of events, mean and variance of event durations | Preliminary, group-averaged |
The design’s cleverness is entirely in Step 1’s control conditions. To make “attention to the breath” a specific neural target rather than a generic interoceptive one, the training set includes attention to the feet — interoception directed at another body region — and attention to ambient scanner sounds — a constant external distractor. Breath-classification specificity is bought by teaching the classifier what breath-attention is not.
Why this matters for the wiki
The wiki’s standing complaint, stated on interoceptive-attention-task and does-mindfulness-enhance-interoceptive-accuracy, is that interoceptive attention can only be assessed two ways, both compromised: ask (questionnaires like the maia, contaminated by belief and demand) or run a cardiac task (whose validity is a live debate and which measures accuracy, not attention). EMBODY is a genuine third route — it measures where internal attention is pointed, continuously, without a probe interrupting the practice.
Two consequences the wiki should hold:
- It gives experience-sampling during meditation a neural analogue, and produces a quantity — “percentage time of interoceptive focus” — that no self-report instrument can deliver, because self-report during meditation destroys the state it measures.
- Orienting interoceptive attention to different body locations (breath vs feet) is decodable, which makes attentional topography an empirical question rather than an instruction. That bears on bodily-sensation-maps and on the topographic claims of body-scan practices and MABT.
The result also lands, quietly, on the wiki’s dual-mode thread. EMBODY’s three meditation-relevant states are breath (interoception), mind-wandering, and self-referential processing — and the self-referential state is operationalized (think about the past week / the present moment regarding yourself) as exactly the narrative self-reference pole Farb et al. (2007) built the corpus on. Weng et al. cite Farb’s interoception and DMN work as load-bearing. So the classifier is, in effect, decoding movement between the sensory/experiential pole (breath) and the narrative pole (self), plus the DMN-associated wandering that Christoff et al. treat as movement between mental states. The two literatures are measuring the same dyad from opposite ends — Farb by contrast, Weng by decode.
The honest boundary: what Step 2 does and does not show
This is a feasibility paper and says so. The single most important thing to keep straight is that the meditation-period decoding has no ground truth. Step 1 is validated: the classifier is trained and tested on trials where the participant was told to attend to X, so cross-validated accuracy means something. Step 2 takes that trained classifier and applies it to a meditation in which nobody was told anything and nothing was recorded about what was actually attended. The decoded “breath: 40% of the time” is therefore an estimate from a validated-elsewhere classifier, not a verified measurement of the meditation.
What makes it more than a guess is that the Step-3 metrics come out the way a real sustained-attention practice should: more time on the breath than on distractors, and — the specific result — the breath advantage is carried by longer breath events, not more frequent ones (10.9 s vs 8.1 s for MW; equal event counts). A classifier decoding noise would not preferentially lengthen the instructed-focus state. That is convergent validity, not verification, and the paper is careful to call the profiles preliminary.
The decoder’s other limits are the ordinary ones for MVPA and the wiki should carry them: classifiers are participant-specific (no group model), the importance maps are idiosyncratic and mostly cortical (so there is no “interoceptive attention region” to point at), and what gets decoded is bounded by the five labels the experimenter supplied — the MW class in particular is almost certainly a bag of several states.
Relation to the rest of the wiki
- The measurement contribution the wiki attributed to helen-weng is now sourced first-hand; the “no primary source to check it against” caveat on that page is resolved (the caveats that replace it are the ones above, which are the paper’s own).
- On does-mindfulness-enhance-interoceptive-accuracy, this is the instrument that could, in principle, test the Farb-side claim that training changes attention deployment rather than accuracy — by measuring deployment directly. As of this paper it has not been used that way; it is demonstrated, not applied to the debate.
- Companion to Farb et al. (2013) and Farb, Segal & Anderson (2013): where Farb uses attention to decide which cortex represents the breath (a localizer with recorded respiration), Weng uses the distributed pattern to decide what a meditator is attending to (a decoder with no respiration recorded). Same paradigm family (interoceptive-attention-task), opposite inferential direction — Farb reads the map, Weng reads the mind.
- Banellis et al. (2026) approach the same target — what the mind turns to during rest/practice — from self-report (multidimensional-experience-sampling) rather than decoding; the two are the report and the neural analogue of “what is attention doing when unprobed.”