Method at a glance
- Modality: multimodal (whole-body subjective sensation)
- Measures: The subjectively felt topography of bodily activation and deactivation during an emotion or other state — where in the body a state is felt, at pixel resolution, not whether the felt change is physiologically real.
- Procedure: Subjects view two identical blank body silhouettes with a state term (e.g. an emotion word) between them. On the left body they colour regions whose activity they feel increasing; on the right, regions whose activity they feel decreasing. Painting is dynamic — repeated strokes raise opacity to a 255 maximum — and strokes are stored as a matrix of integers over ~50,000 body pixels (positive = activation, negative = deactivation). Activation and deactivation maps are combined per subject per emotion into a bodily sensation map (BSM), then submitted to mass-univariate pixelwise t-tests against zero with FDR correction.
- Also called: emBODY tool, bodily sensation mapping, body colouring task, topographical self-report of bodily sensation
Strengths
- High spatial resolution on the felt body, where psychophysiology typically records one or a few channels; the multivariate pattern is the measurement, not a derived summary.
- Reports the output layer — what is consciously felt — so it sidesteps the weak coupling between general-level autonomic recording and subjective feeling, at the price of physiological specificity (the authors’ own framing).
- Cheap, fast, and language-light enough to run online at population scale; enabled a 3954-subject, 101-country design (Volynets et al. (2020)).
- Word-cued maps are concordant with maps from nonverbal induction — vignettes, silent film, and photographs of faces all converge with word cues at rs = 0.71-0.82, and induction classifies better than words (Nummenmaa et al. (2014)). The main evidence that it is not purely a semantic task, and stronger than the single inherited citation this wiki took it for.
- Usable across the lifespan: applied to children (Hietanen et al. 2015) and to ages 18-90.
- Travels: replicated by an independent group, in another language, under picture rather than word/movie induction (Lyons et al. (2021)), and sensitive enough to dissociate two psychiatric populations (undifferentiated maps in schizophrenia, intact-but-deactivated maps in medicated depression).
Limitations
- Measures belief-about-sensation, not sensation. It cannot distinguish felt bodily change from conceptual knowledge about where an emotion ‘should’ be felt — the central interpretive risk, and sharpest when the cue is an emotion word.
- Typically cued by a verbal token in a single language, so shared vocabulary is confounded with shared experience in cross-cultural designs.
- The conceptual content the word-cue objection postulates has been measured and is large. Interoceptive content is the biggest single determinant of whether a word counts as an emotion word at all (Ferré et al. (2024)), so an emotion-word cue supplies a bodily expectation before the subject introspects.
- No physiological ground truth. Results say nothing directly about Autonomic specificity of emotion; the maps are subjective feeling maps.
- Online administration invites self-selection (Volynets et al.’s sample was 82% female and highly educated) and uncontrolled conditions.
- Painting behaviour has motor and interface confounds (tool diameter, opacity accumulation, screen size) that mass-univariate pixel statistics do not model.
- The deactivation half of a map is induction-dependent. Deactivation is a substantial part of the signature in Nummenmaa’s movie/word-cued maps (sadness and depression are carried primarily by it), yet under picture induction ‘almost no deactivation reached significance’ in healthy controls (Lyons et al. (2021)). If half the fingerprint depends on how it is elicited, the fingerprint is partly an artifact of the elicitation.
- Activation and deactivation are combined into a single map before analysis, so ‘more deactivation’ and ‘less painting overall’ are not separable — the confound that limits the medication finding in Lyons et al. (2021).
- Emotion induction, not the tool, is often the binding constraint: anger has defeated both movie induction (dropped outright, with surprise, in Nummenmaa et al. (2014)) and picture induction (below-chance classification in Lyons et al. (2021)). A null BSM may index a failed induction rather than an absent fingerprint.
- Different emotions’ maps already correlate at rs ~ 0.52 (Nummenmaa et al. (2014), across experiments; 0.36-0.40 across languages). Every concordance figure the paradigm reports sits on that floor, and the emotion-specific signal is the residual above it.
Connections
Name collision. This is Nummenmaa’s emBODY body-colouring tool. It is unrelated to Weng’s EMBODY (“Evaluating Multivariate Maps of BODY Awareness”), an fMRI brain-decoding pipeline for internal attention during meditation. Same acronym, different literatures.
The topographical self-report tool developed by Nummenmaa’s group (open-sourced at version.aalto.fi/gitlab/eglerean/embody) and the instrument behind the “bodily maps of emotions” literature. It produces Bodily sensation maps (bodily fingerprints of emotion).
What kind of measure this is
emBODY sits in an unusual place among this wiki’s methods, and placing it correctly is most of the interpretive work.
The Heartbeat detection task measures accuracy — performance against a physiological ground truth. Multidimensional Assessment of Interoceptive Awareness (MAIA) measures sensibility — beliefs about one’s own interoceptive style. physiology congruence measures coherence — the within-person correspondence between report and physiology. emBODY has no ground truth and no correspondence term: it measures the content and spatial structure of what is felt, taking the report at face value. On the Interoceptive taxonomy (Farb et al. Box 1), it is closest to sensibility, but sensibility about a state rather than about a trait.
This is a deliberate trade, stated plainly by its authors: focusing on “the output layer of these functions — the subjectively felt bodily changes — circumvents [the weak self-report/physiology coupling] even though it sacrifices the physiological specificity.” The gain is that emotion-specific structure appears where general-level autonomic recording has struggled to find it. The cost is that finding structure in emBODY data says nothing about whether the body did anything.
The word-cue problem
The standard administration presents an emotion word. Any concordance across subjects, sexes, or cultures is therefore compatible with two very different explanations: they feel the same things, or they know the same things about where emotions belong.
The literature’s answer is Nummenmaa et al. (2014): word-cued maps match maps produced after nonverbal emotion induction. Having now read the source rather than the citations of it, this wiki records the claim as substantially stronger than it had assumed, and narrower than the papers citing it require.
Stronger: it is three experiments, not one finding. Vignettes that never name the emotion or any bodily sensation, silent films, and photographs of strangers’ faces all converge with word cues at rs = 0.71–0.82, and the nonverbal inductions classify better than words (48%, 50% and 31% vs 38% complete-classification accuracy). The silhouettes deliberately omit internal organs “to avoid triggering purely conceptual associations between emotions and specific body parts (e.g., love–heart).” This was designed against the objection, not retrofitted to it.
Narrower: what it licenses is that Finnish undergraduates produce concordant maps under verbal and nonverbal induction. Volynets et al. (2020) cites it to license word cues standing in for felt experience across 101 countries (see Are emotions culturally universal?). The datum does not reach that far, and no one has extended it.
And it leaves the general objection alive, which the paper concedes: a vignette about a dying child is not a word, but it is still a token of a shared emotion category. Subjects who recognise the scenario can paint from the same conceptual knowledge a word would have cued. Nummenmaa et al.’s reply is an argument from origins — where would culture-independent conceptual associations come from, if not from bodies? — and that is an argument, not a result.
Lyons et al. (2021) weaken the worry without dissolving it. Their induction was pictures (IAPS), not words, and emotion-specific maps came out anyway at accuracies matching the original — the first independent replication in this wiki that does not run through a verbal cue (Nummenmaa et al.’s own experiments 2–4 got there first, in-house). But a picture selected to depict disgust is still a token of a shared emotion category, so the constructionist can reply that conceptual knowledge was cued pictorially rather than lexically. What the result does rule out is the narrowest version of the objection, that the maps are an artifact of reading an emotion word. The general version — that subjects report where the category says the feeling belongs — survives it, exactly as it survives the 2014 induction experiments.
Anger is the test case nobody has run. Anger classifies well from words and is the best-recognised map in the 2014 observer study (58%) — yet it was dropped outright from the movie experiment for induction difficulty, and classified below chance under picture induction in Lyons et al. An emotion that produces a crisp map whenever it is named and fails whenever it must be induced is the sharpest available probe of whether these maps track feeling or category knowledge. See Nummenmaa et al. (2014).
Ferré et al. (2024) strengthen it, from outside the paradigm and without meaning to. The objection postulates that emotion concepts carry bodily content which a word cue could deliver in place of introspection. That postulate has now been measured: across 1,286 Spanish words, the single largest determinant of whether speakers judge a word to name an emotion is how far they judge it “related to internal body sensations” (R² increment .417, ahead of every other component of emotion). Being felt in the body is not incidental to the folk concept of emotion — it is most of what constitutes it.
So an emotion word is not a neutral pointer at an experience. It is a token of a category that is partly defined by where it is felt, and the interoceptive expectation arrives with the lexicon before any introspection occurs. This does not decide the question — see Ferré et al. (2024) for the symmetrical reading, on which the concept is interoceptive precisely because bodies are, making the lexicon a fossil of real bodily experience — but it retires the framing in which the objection is a merely-possible confound to be acknowledged and moved past. The confound is large and it has a number.
What still cuts the other way is the psychopathology strand below: shared lexical knowledge cannot fracture by diagnosis, and the maps do.
Beyond emotion
The same tool has been applied to social touch topography (Suvilehto et al. 2015), erogenous zones (Nummenmaa et al. 2016), subjective feelings generally (Nummenmaa et al. 2018), and psychopathology.
The psychopathology strand is the one with the most upside for the method, because it has produced a double dissociation on a single instrument: maps are undifferentiated between emotions in schizophrenia (Torregrossa et al. 2018) but well differentiated in depression, where what changes instead is the activation/deactivation balance (Lyons et al. (2021)). An instrument that separates two disorders in two different ways is doing more than recording folk knowledge about where emotions live — the strongest indirect answer yet to the word-cue problem above, since shared concepts of “sadness” should not fracture along diagnostic lines.