Interoception of breathing and its relationship with anxiety

Study at a glance

• Authors: Olivia K. Harrison, Laura Köchli, Stephanie Marino, Roger Luechinger, Franciszek Hennel, Katja Brand, Alexander J. Hess, Stefan Frässle, Sandra Iglesias, Fabien Vinckier, Frederike H. Petzschner, Samuel J. Harrison, Klaas E. Stephan
• Year: 2021
• Journal: Neuron 109(24):4080-4093
• DOI: 10.1016/j.neuron.2021.09.045
• N: 60
• Design: Two-group (low vs moderate trait anxiety) multi-modal study: questionnaires + psychophysical breathing task + novel associative-learning task during 7T fMRI, with computational modelling of both behavioural datasets; 15 held-out participants for model validation; pre-registered analysis plan

Key findings

• The first trial-by-trial computational model of interoceptive learning fitted to behaviour with concurrent brain imaging. A Rescorla-Wagner learning model over cue-to-inspiratory-resistance contingencies yields per-trial prediction certainty and prediction error; both are tracked by brain activity. This is the empirical test Petzschner et al. said the computational literature did not have.
• The anterior insula does both jobs. aIns deactivates in proportion to prediction certainty (with dlPFC, ACC, MFG) and activates in proportion to prediction error magnitude (with ACC, MFG and the PAG) — a single region carrying both terms, read as its role in representing and updating models of bodily state.
• Anxiety shifts the prediction term, not the error term. A valence x group interaction in bilateral aIns for prediction certainty: low-anxiety participants deactivated aIns more when predicting resistance than no-resistance; moderate-anxiety participants showed the opposite. No group or interaction effect anywhere in prediction error — contrary to several theoretical proposals (Barrett & Simmons; Brewer et al.; Paulus & Stein) that predicted the error side.
• Learning itself was unaffected. No group difference in learning rate (α: 0.24 vs 0.25) or inverse decision temperature. Anxiety changed the neural representation of predictions without changing the fitted learning parameters, physiology, or breathing difficulty ratings — only the post-task breathing-anxiety rating differed (0.0 vs 34.0, p<0.001).
• Moderate anxiety was associated with worse respiratory perceptual sensitivity on the FDT (threshold 3.0 vs 4.0 filters, Z=-2.4, p=0.01) and lower metacognitive bias — less confidence (6.7 vs 6.2, Z=2.0, p=0.02) — with equal metacognitive performance (M-Ratio 0.8 vs 0.8, ns) and equal decision bias.
• The anxiety-interoception link is strongest at the highest levels of the hierarchy. PCA over 16 measures: PC1 separated the groups at p<1e-11 and was loaded first by affective questionnaires, then breathing catastrophizing and (negative) MAIA interoceptive awareness, then metacognitive bias and performance, then perceptual sensitivity, and last by aIns activity. Subjective interoceptive beliefs carry the relationship; psychophysical sensitivity carries least.
• The interoceptive measures barely correlate with each other. Across the 16-measure matrix, cross-task correlations were sparse — supporting genuinely separable levels of breathing interoception (Critchley & Garfinkel 2017). The one strong cross-modal link: FDT metacognitive performance with peak aIns activity for negative (resistance) prediction errors, r=0.42.
• Anxiety sensitivity (ASI-3) tracked lower metacognitive bias (r=-0.30, p=0.02) and not perceptual threshold (r=0.20, p=0.13) — the wiki’s first measurement of the ASI against a non-cardiac interoceptive task.

Limitations

• One channel, one feature. Only inspiratory resistance within the respiratory domain; the authors say plainly that the results cannot be assumed to translate to cardiac or gastric interoception. Given the domain-generality null, that caveat is load-bearing rather than formulaic.
• No model won. Bayesian model selection gave no candidate a protected exceedance probability above 90% (RW:HGF2:HGF3 = 0.30:0.40:0.30), so the Rescorla-Wagner model was used by pre-registered fallback to Occam’s razor, not because it was selected. The whole computational-brain analysis rests on a model the data did not prefer.
• Prediction certainty is a transformation, not a model parameter. RW has no explicit precision or volatility term, so certainty had to be constructed post hoc as |v - 0.5|. The paper’s headline quantity — the one where the anxiety effect lives — is a distance from guessing, not an inferred precision. The models that do represent volatility explicitly (HGF2, HGF3) were the ones set aside.
• Extreme groups. Low anxiety was defined as STAI-T 20-25 and moderate as >=35, with the middle of the distribution excluded — the same selection design the wiki discounts on Pollatos et al. (2005) and Werner et al. (2009). Both groups are healthy young adults (mean age ~25); no clinical anxiety sample.
• Reduced field of view. The 7T functional scan was an axial-oblique slab over insula and midbrain (32 slices), so whole-brain claims are not available and null results outside the slab are uninformative.
• Correlational throughout: nothing was manipulated except the breathing stimulus itself, so the anxiety-to-aIns relationship is an association between a screening variable and a brain measure.
• Cross-modal correlations in the multi-modal matrix are mostly weak and exploratory (p<0.05 uncorrected, FDR applied); n=58 for the model-based analyses.

Connections

• Methods: Filter Detection Task (FDT), Breathing Learning Task (BLT), Multidimensional Assessment of Interoceptive Awareness (MAIA)
• Concepts: Respiratory interoception, Interoceptive inference, Predictive coding, Active inference, Interoceptive precision, Metacognitive efficiency and metacognitive bias, Anxiety sensitivity, Insular cortex, Periaqueductal gray (PAG), Anterior cingulate cortex (ACC), computational psychosomatics, Interoceptive taxonomy (Farb et al. Box 1), Interoceptive psychopathology, Forecasting (internal-state prediction), Interoceptive sensitivity
• Debates: inferential?, Is interoception domain-general or organ-specific?, Is more interoceptive awareness better?, Is the heartbeat-counting task a valid measure of cardiac perception?
• Related studies: Petzschner et al. (2021), Banellis et al. (2026), Allen (2026), Khalsa et al. (2018), Barrett (2017), Seth & Friston (2016), Quadt et al. (2018), Ehlers (1993), Zoellner & Craske (1999)

Neuron 109(24):4080–4093, open access (CC BY). From the Translational Neuromodeling Unit — Klaas Stephan’s group at UZH/ETH Zurich, with Petzschner on the author list — which is the same address that produced the computational review the wiki reads as the map of this quadrant.

The wiki has been waiting for this paper by name. Allen (2026)‘s page listed it as “the Filter Detection Task source. Still unprocessed”; the RRST page put it in a three-row table of respiratory instruments as the only one without a page of its own. It arrives carrying rather more than a task.

Why this is the ingest the computational thread needed

Petzschner’s page records her own verdict on the framework she works inside: “there is little empirical work testing such models’ predictions,” and the few tests she can cite include her own. computational psychosomatics repeats it. Forecasting (internal-state prediction) is filed as a gap page on the same grounds.

This is a test. Not of forecasting in the strict sense (nothing here models how a bodily state evolves under its own dynamics — the contingency between a cue and a resistance is what is learned), but of the claim one level down: that interoceptive predictions and prediction errors are quantities the brain computes trial by trial, and that a fitted learning model can find them. Sixty participants, eighty trials each, a generative model per person, and brain activity regressed on the model’s own per-trial estimates.

The result is affirmative and modest in the right places. Prediction certainty and prediction error magnitude both have neural correlates, they sit where the theory says (aIns, ACC, PAG), and the signs are the ones predictive coding requires — activity falls as predictions get more certain and rises with error. Before this, the wiki’s evidence for interoceptive predictive coding was cytoarchitecture (agranular visceromotor cortex), theory, and one mouse study (Livneh et al. via Berntson & Khalsa). Now there is a human, model-based, trial-resolved demonstration. See inferential?.

The design, in three layers

The paper’s structure is an argument: interoception spans levels, so measure several at once and ask which level anxiety attaches to.

level	instrument	what anxiety did
subjective belief	12 questionnaires (ASI-3, MAIA, BPQ, PCS-B, PVQ-B, …)	large, highly significant differences on nearly all
perceptual sensitivity + metacognition	Filter Detection Task (FDT)	worse sensitivity, lower confidence, equal insight
trial-by-trial learning + brain	Breathing Learning Task (BLT) at 7T	no behavioural difference; a valence x group interaction in aIns

Groups were sixty healthy adults pre-screened on the Spielberger trait scale — thirty scoring 20–25, thirty scoring ≥35 — matched for age and sex (15 women per group). Eight further participants supplied the model priors; fifteen more, not selected on anxiety, were held out to validate the fitted trajectory (β = 3.1, t = 12.1).

The finding the wiki should carry forward: it is the high levels that track anxiety

The PCA is the paper’s own summary of itself, and it inverts the direction the field’s measurement effort points.

PC1 separated the two anxiety groups at p < 1×10⁻¹¹. Its loadings, in order: depression, state anxiety, anxiety sensitivity, anxiety-disorder symptoms — then breathing catastrophizing and negative interoceptive awareness — then negative metacognitive bias, body perception, negative metacognitive performance — then perceptual threshold and decision bias — and last, aIns activity.

Read down that list. The relationship between anxiety and breathing interoception is carried overwhelmingly by what people believe and say about their bodies, secondarily by their confidence, and least by how well they actually detect a resistive load or by what their insula does. The authors are careful about the obvious rejoinder — higher levels may simply be measured less noisily than psychophysical thresholds, so the ordering may be detectability rather than strength — and the wiki should keep that caveat attached.

But taken at face value it is a direct empirical statement of something Interoceptive taxonomy (Farb et al. Box 1) has only been able to assert structurally: sensibility and accuracy are not two measurements of one thing, and the clinically interesting variance is in the first. It converges with Banellis et al.’s finding that the construct which travels between channels is confidence rather than competence — reached here from within a single channel, by asking which level of one hierarchy an affective trait attaches to.

Anxiety and respiratory sensitivity run opposite to anxiety and cardiac sensitivity

This is the most consequential collision the ingest produces, and it is not a contradiction.

The wiki’s panic literature is emphatic in one direction: panic patients detect their heartbeats better than controls (Ehlers 1993; replicated in analogue panickers by Zoellner & Craske 1999), and better cardiac perception predicts worse clinical outcome. “Anxious people are better at reading their bodies” is the premise the Cognitive model of panic and much of Is more interoceptive awareness better? run on.

Harrison et al. find the reverse in the lungs: moderate-anxiety participants needed more filters to reach threshold — they were less sensitive to an inspiratory load. And this is not an outlier; the authors cite it as replicating prior respiratory findings (Garfinkel et al. 2016a; Tiller, Pain & Biddle 1987).

Two readings, and the wiki should hold both:

1. Channel-specific, and evidence for organ-specificity. Anxiety relates to the heart and to the lungs in opposite directions, which is a stronger form of decorrelation than a null correlation — not merely “these axes do not covary” but “the same trait predicts them with opposite sign.” Filed on Is interoception domain-general or organ-specific?.
2. Instrument-specific, and evidence against the cardiac task. The counting task has a known failure mode in which anxiety manufactures accuracy (count faster in a population that undercounts; see Anxiety sensitivity and Van der Does’s artefact argument). The FDT is a two-alternative decision about a physically specified stimulus with no such route. On this reading the cardiac direction is the artefact and the respiratory direction is the finding.

Nothing here decides between them, and the discriminating study — both tasks, one anxious sample — has not been run. But reading 2 is the more parsimonious, and it makes the respiratory result the cleaner of the two.

What anxiety did to the brain, and what it did not

The one significant group effect in the imaging is an interaction between prediction valence and anxiety group in bilateral anterior insula, on prediction certainty. Low-anxiety participants showed greater aIns deactivation scaling with predictions of upcoming resistance than with predictions of no resistance; moderate-anxiety participants showed the opposite. So the shift is in how confidently-held predictions about threatening bodily states are represented — in line with Paulus & Stein’s insular account of anxiety, which the paper cites approvingly.

What is absent matters as much. No group difference and no interaction anywhere in prediction error — which several theories in this wiki predicted. Barrett & Simmons’s EPIC illness hypothesis, as computational psychosomatics records it, locates psychopathology in elevated interoceptive prediction error decoupled from priors. Brewer, Murphy & Bird’s atypical-interoception review and Paulus & Stein (2006, 2010) also point at the error side. Harrison et al. name the disagreement explicitly: their result “contrast[s] with some previously proposed hypotheses.”

That is a genuine empirical constraint on the wiki’s most-used clinical story, and it is the first one the wiki holds. It is not a refutation — anxiety is not depression, a resistive load is not dyshomeostasis, and a null in a reduced field of view at n=58 is a weak null. But every page here that explains a disorder by “elevated prediction error” now has one direct test to answer to, and the test found the effect on the other term.

The insula does both, and the PAG does one

Independent of anxiety, the whole-slab results are the clean part of the paper.

• Prediction certainty → deactivation of dlPFC, anterior insula, ACC and MFG. Certainty reduces belief updating, so activity falls as certainty rises. This is the sign predictive coding requires (Feldman & Friston 2010).
• Prediction error magnitude → activation of aIns, ACC, MFG and the periaqueductal gray.
• Valence mattered only for errors: unexpectedly receiving a resistance activated left posterior insula more than unexpectedly escaping one.

Two things worth pinning down for Insular cortex. First, the anterior insula deactivates for certainty and activates for error — one region holding both sides of the update equation, which is what a region “representing and updating models of the body” should look like, and which the wiki’s usual posterior-error/anterior-prediction split does not straightforwardly predict. Second, and the authors say so, the anterior/posterior dissociation of predictions from prediction errors that Seth, Barrett & Simmons and Stephan et al. all propose did not appear. The single crumb for it is that valence effect in posterior insula, which the authors read as homeostatically-relevant inputs being enhanced when they threaten homeostasis.

The PAG result is the other novelty: it tracked prediction error magnitude but not prediction certainty, which fits its assignment as a homeostatic-control structure receiving regulatory error rather than holding beliefs. See Periaqueductal gray (PAG).

The methodological standard, which is the highest in this wiki

Worth recording because most of what the wiki ingests does not do this: a time-stamped pre-registered analysis plan on GitLab (including the fallback rule that produced the Rescorla-Wagner choice); simulated parameter recovery and model identifiability before fitting; a held-out validation sample; a null-model likelihood-ratio check per participant, with the two failures excluded; an independent code review; summary statistical maps deposited publicly and code archived on Zenodo. Raw data are withheld for GDPR reasons with a stated request route.

The pre-registration did real work rather than decorative work: no model reached the pre-specified 90% PXP threshold, and the plan’s fallback — take the simplest — is what licensed using RW rather than a post-hoc justification. That is also the paper’s largest weakness, and the two facts are the same fact.

Placement

Read as the wiki’s first empirical, model-based test of interoceptive predictive coding in humans, and as its most complete single-channel measurement of the interoceptive hierarchy in one sample. It contradicts nothing here. What it does is supply three things the wiki has been asking for by name: an anxiety measure against a non-cardiac interoceptive task (Anxiety sensitivity’s standing gap), a first-hand source for the FDT, and evidence about where in the hierarchy an affective trait attaches — which turns out to be the level furthest from the body.