Study at a glance
- Authors: Gidon Levakov, Shira Ganor, Galia Avidan
- Year: 2023
- Journal: Human Brain Mapping
- DOI: 10.1002/hbm.26436
- N: 43
- Design: Concurrent electrogastrography + resting-state fMRI + pulse oximetry; repeated runs (84 runs / 43 subjects), replicated in an open highly-sampled-individual dataset (CERB, 19 sessions, N = 1); methodological validity and test–retest reliability study
Key findings
- The Gastric network survives strict confound control. After regressing global signal, non-grey-matter signal, head motion (6 parameters + derivatives + squares), aCompCor and (in a subsample) cardiac RETROICOR components, occipital, precuneus, bilateral insula and pre/postcentral clusters remain phase-locked to the 0.05 Hz gastric rhythm. Spatial overlap with Rebollo & Tallon-Baudry (2022) exceeds a spin-permutation null (GSR: Dice = .035, p = .03; no GSR: Dice = .114, p = .002).
- Before confound regression the network is enormously overestimated. Significant voxels fall from 25,504 (19.89% of grey matter) to 1,491 (1.16%) once confounds are removed — a 17-fold shrinkage. The gastric network as previously mapped was mostly motion and non-neural signal.
- The confounds are themselves gastric-locked. CSF signal, global signal, x-translation, and x- and z-rotation all show significant phase-locking with the EGG signal (FDR-corrected). Head motion is not independent noise here; it synchronizes with the stomach — plausibly through respiration, which drives both.
- Test–retest reliability of gastric–brain synchrony is absent. Across two runs in the same subjects (n = 23), PLV-delta did not correlate within the significant gastric network (r(21) = −.14, p = .54) or within the grey-matter mask (r(21) = −.08, p = .71). Node-wise PLV across Schaefer-100 parcels also did not differ from zero (mean r = .003, t(1033) = 0.76, p = .446).
- And the null is not an instrument failure. The same participants’ EGG peak frequency correlated r(14) = .74 across lab and scanner sessions months apart, and resting functional connectivity was far more similar within than between subjects (t(1033) = 10.89, p < .01). Both ingredients are reliable; only their coupling is not.
- Every finding except the spatial similarity to prior work replicated in the independent CERB single-subject dataset (Choe et al. 2021). In that dataset the significant clusters (right precuneus, paracentral lobule) showed no spatial similarity to Rebollo & Tallon-Baudry (p = .42).
- Reliability sets a ceiling on validity: the authors invoke Goodwin & Leech (2006) to note that an unreliable measure cannot correlate strongly with anything, which is a direct problem for the interoceptive-biomarker programme (Khalsa & Lapidus 2016).
Limitations
- High exclusion rate. 29 of 122 runs (24%) dropped for poor EGG quality — comparable to Wolpert et al. (2020) but severe; and the reliability analysis required two usable runs, excluding 36 of 59 subjects. The reliability null therefore rests on n = 23, which is small for estimating a correlation, and a null r at that n has wide bounds.
- Restricted sample. BMI 18–25 only (EGG quality degrades with adiposity), mean age 25, healthy. Both the generalizability of the network and the reliability estimate inherit that restriction.
- No causal manipulation. Nothing here establishes that the surviving coupling is driven by the stomach. The causal evidence cited is other people’s — vagotomy and hunger-state modulation in rats (Cao et al. 2022), taVNS in humans (Müller et al. 2022).
- Respiration was not recorded. The paper’s own leading mechanistic explanation for why head motion is gastric-locked is breathing, and it has no respiratory measure to test it with. See Respiratory interoception.
- Not a replication. Different preprocessing pipeline, high-pass-filtered rather than DC EGG recording, repeated measures — the authors say explicitly this is generalization, not replication, so the Dice similarity is a weaker form of corroboration than it first reads as.
- Global signal regression is contested and it changes the answer. Without GSR the map is more widespread and more similar to prior work (Dice .114 vs .035). Which of the two maps is ‘the’ gastric network is a preprocessing choice the field has not settled.
- A null reliability correlation is a claim about this synchrony measure (PLV-delta averaged over runs), not about gastric–brain coupling as a phenomenon. The authors are careful about this; secondary citation may not be.
Connections
- Methods: Electrogastrography (EGG), Phase-locking value (PLV)
- Concepts: Gastric network, Interoception, Insular cortex, Precuneus, Interoceptive taxonomy (Farb et al. Box 1), computational psychosomatics, Respiratory interoception, Disorders of brain-gut interaction, Body-wandering
- Debates: Can brain–body coupling be measured in a person?, Is interoception domain-general or organ-specific?, Is the heartbeat-counting task a valid measure of cardiac perception?
- Related studies: Haruki & Ogawa (2023), Banellis et al. (2026), Banellis et al. (2026), Berntson & Khalsa (2021), Bonaz et al. (2021), Petzschner et al. (2021), Khalsa et al. (2018), Zeharia et al. (2019)
The wiki’s first gastric evidence of its own, and it arrives as a methods audit rather than a discovery. Three ingests have now cited the stomach at second hand — Banellis et al. put stomach at the top of the brain-constrained CCA loadings (0.58) for Body-wandering; Bonaz et al. built a clinical chapter on brain–gut interaction; the domain-generality debate names gastric as “the notable absence, since it is where coherence was originally claimed.” This paper is the first source here that actually measures a stomach.
What it finds is a two-part verdict that is worth keeping separate, because the paper’s abstract runs them together and its citations will not.
Part one: the network is real, and it was four-fifths noise
Rebollo et al. (2018) reported a resting-state network phase-locked to the stomach’s ~0.05 Hz myoelectrical rhythm — a genuinely novel kind of finding, an intrinsic brain network defined not by its own dynamics but by an organ’s. It has been elaborated since: unimodal rather than transmodal cortex (Rebollo & Tallon-Baudry 2022), modulated by vagus nerve stimulation (Müller et al. 2022).
None of that work regressed head motion. This one does, along with global signal, non-grey-matter signal and cardiac RETROICOR components, and the result is the paper’s single most quotable number:
| significant voxels | % of grey matter | |
|---|---|---|
| minimally preprocessed | 25,504 | 19.89% |
| strict confound regression | 1,491 | 1.16% |
Nearly a fifth of grey matter appeared coupled to the stomach before cleaning. What survives — occipital, precuneus, bilateral insula, pre/postcentral — is spatially concordant with the published network above a spin-permutation null, so the phenomenon is not an artefact. But its published extent largely was, and the claim “the stomach is coupled to a wide swathe of cortex” has to be withdrawn to something much narrower.
The mechanism is the interesting part. These confounds are not incidentally present; they are synchronized with the EGG signal itself — CSF, global signal, and three of six motion parameters. The paper’s proposed route is respiration, which moves the head and modulates gastric signal alike, and whose frequency band can bleed toward the gastric one. It has no respiratory recording to check this with, which is the paper’s sharpest self-inflicted limitation and a standing lesson for Respiratory interoception: breathing is a confound in every other interoceptive channel’s measurement, not merely a channel of its own.
Part two: the coupling has no test–retest reliability, and this is the finding that matters
Two runs, same subjects, same session. The correlation between run 1 and run 2 gastric–brain synchrony:
- within the significant gastric network: r(21) = −.14, p = .54
- within the grey-matter mask: r(21) = −.08, p = .71
- node-wise across Schaefer-100 parcels: mean r = .003, not different from zero
Zero, twice, three ways. And the control analyses close off the easy dismissal: the EGG peak frequency of the same people correlated r = .74 across recordings taken months apart in different rooms on different equipment, and resting functional connectivity was strongly subject-identifiable. The stomach is a reliable measurement. The brain is a reliable measurement. The relationship between them, as currently operationalized, is not a property of the person.
The consequence is stated in the paper via Goodwin & Leech (2006) and it is the general principle, not a gastric one: reliability bounds validity. A measure that does not correlate with itself cannot correlate with hunger, with symptom severity, with eating pathology, or with anything else. Every individual-differences use of gastric–brain coupling — including the first author’s own 2021 paper predicting weight loss from brain–gastric interaction (Levakov et al., NeuroImage 224:117403) — is under this ceiling.
That self-implication is why the paper is worth trusting. The authors published a result that undercuts their own prior finding, and say so obliquely by citing it in the discussion as the kind of work that “will allow future work to explore the functional significance.”
Why this matters beyond the stomach
The wiki now holds three independent instances of one pattern, and they should be read together:
| channel | what is reliable | what is not |
|---|---|---|
| cardiac | — | heartbeat counting as a perceptual measure (Is the heartbeat-counting task a valid measure of cardiac perception?; <10% can do it, Does et al. (2000)) |
| cardiac ↔ respiratory | confidence (r ≈ .51–.64) | cross-channel performance (Banellis et al. (2026)) |
| gastric ↔ brain | the EGG signal (r = .74), the fMRI signal | their coupling (r ≈ 0) |
Three different labs, three different instruments, three different levels of analysis, and the same shape each time: the components are measurable and the interoceptive quantity built from them is not. This is now sufficiently repeated that it should be treated as a property of the field’s measurement programme rather than a series of local failures — see Can brain–body coupling be measured in a person?, created with this ingest.
It also sharpens what Banellis et al. (2026) can be taken to have shown. That study reported “convergent patterns in respiration and electrogastrography” alongside its body-wandering dimension. Nothing in Levakov et al. touches EGG measurement itself — which is reliable — so the signal-level claim stands. What would not stand is a follow-up correlating body-wandering with gastric–brain coupling, and that is the obvious next study for that group.
What it does not say
It does not say the gastric network is spurious. The authors are explicit that it generalizes, that rodent vagotomy work (Cao et al. 2022) and human taVNS (Müller et al. 2022) supply causal support neither of which can be explained by motion, and that group-level gastric–brain coupling is a robust phenomenon. The distinction the paper insists on and this wiki should preserve:
- Group-level existence: established.
- Individual-level measurement: not yet a thing that can be measured.
The second is what a biomarker requires. Petzschner et al. already conceded that the computational-biomarker programme “has almost no empirical test behind it”; this is what one of those tests looks like when it is run.
The same map, reached without recording a stomach (Haruki & Ogawa 2023)
Ingested the run after this one, Haruki & Ogawa (2023) report that the visual-cortex territory recruited when people simply attend to their stomachs largely overlaps the Rebollo & Tallon-Baudry (2022) gastric map — the same NeuroVault collection this paper audits.
Two ways to read the pair, and the wiki holds both:
- Encouraging. A network defined by phase-locking to an unfelt 20-second rhythm and a map produced by a one-word instruction land in overlapping cortex, by methods sharing no assumptions. Convergence across that gap is worth something.
- Cautionary, and this paper is why. Haruki & Ogawa assert the overlap from a figure — no Dice coefficient, no spin-permutation null, no confound audit of the sort performed here. This paper’s whole contribution is that spatial claims about this map had been made at a standard that did not survive testing, and the very next study to make one used a looser standard still.
The comparison the field needs is quantitative overlap between the confound-controlled gastric map (1.16% of grey matter, not 19.89%) and the attention map. Neither paper ran it, and both datasets are public.
Provenance notes
Open access (CC-BY), Israel Science Foundation funded, data and analysis code public (github.com/GidLev/brain_gastric_synchronization_2023). The authors thank Rebollo and Tallon-Baudry for sharing code and maps — i.e. the group whose finding this paper trims did the trimming with them, which is a better sign about this literature than most. Nicolas Legrand and Micah Allen’s Systole package supplies the cardiac peak detection, connecting this paper’s toolchain to the Aarhus group whose work occupies the wiki’s other recent ingests. No researcher page created for Levakov, Ganor or Avidan: single appearance in raw/, per the wiki’s convention.