Desmedt et al. (2023): Discrepancies in the Definition and Measurement of Human Interoception
A book-length conceptual review (in press, Perspectives on Psychological Science, August 2023) by Olivier Desmedt, the field’s most systematic measurement critic, with Olivier Luminet, Pierre Maurage and Olivier Corneille (Lausanne / UCLouvain). It is the single source in raw/ that states, in full and as its central thesis, two arguments this wiki had been assembling piecemeal across half a dozen pages: that the field’s definition of interoception is internally contradictory, and that its measures are detached from the dimensions they are said to index. Sahib Khalsa is thanked for early feedback — the review takes apart the consensus roadmap he led while its lead author is inside the acknowledgements.
The first discrepancy: the definition contradicts itself
Most interoception papers, the authors observe, endorse both of these in the same article:
- a phenomenon-based definition — interoception is the processing of signals originating within the body;
- a physiological definition — interoception is the activation of specific homeostatic pathways (lamina I spinothalamic afferents and the NTS route), the “homeostatic” machinery Craig identified.
These are treated as interchangeable and are not. The physiological definition equates interoception with slow, high-threshold afferent processing (unmyelinated C and thinly-myelinated Aδ fibres) and dissociates it from fast, low-threshold processing (large-diameter Aβ fibres) assigned to the somatosensory/exteroceptive system. The review’s demolition is empirical: in humans, the processing of internal states runs through the Aβ “non-homeostatic” system too.
| system | homeostatic afferents (the physiological definition’s interoception) | non-homeostatic afferents also involved |
|---|---|---|
| cardiac | baroreceptors; cardiac heart-wall chemo/mechano afferents → vagus/lamina I → NTS → insula | Aβ somatosensory mechanoreceptors (Pacinian). Knapp-Kline et al. (2021): vibrotactile masking of Pacinian/non-Pacinian channels degrades heartbeat detection. The HEP is recordable in somatosensory cortex (Kern et al. 2013); TMS to S1 alters heartbeat perception (Pollatos et al. 2016). Khalsa et al. (2009): a bilateral-insula-damage patient kept normal cardiac awareness that disappeared when chest skin was anaesthetised. |
| respiratory | C and Aδ fibres → “limbic” regions (amygdala, ACC, insula) | Aβ low-threshold mechanoreceptors respond to lung distension; cutaneous thoracic mechanoreceptors track respiratory motion → somatosensory cortex (Davenport & Vovk’s discriminative pathway) |
| gastro-intestinal | C and Aδ fibres → lamina I / medial thalamus → limbic, parietal, frontal | Aβ Pacinian-following fibres → somatosensory cortex via lamina IV–VI and lateral thalamus |
The conclusion is a modal claim and it is the hinge of the paper: no physiological pathway is necessary and sufficient for the processing of internal states. Homeostatic pathways are not necessary (the same tissues are also read by non-homeostatic ones) and not sufficient to mark the boundary (they carry external as well as internal signals). Therefore physiology cannot ground the definition. This is the fullest version of the argument Berntson & Khalsa (2021) gesture at when they catalogue vagal, cranial, sacral, spinothalamic and somatosensory routes; Desmedt et al. draw the definitional consequence they leave implicit. See interoception-exteroception-boundary, where this now sits as the pathway-criterion’s obituary.
The second discrepancy: the dimensions are detached from the measures
The two consensus dimensional schemes — Garfinkel et al.’s (2015) triad (accuracy / sensibility / awareness) and Khalsa et al.’s (2018) eight features — are, the authors argue, “largely detached from current measures.” The evidence is convergence data, and it is worse than the taxonomies assume.
- Accuracy / detection does not travel across organs. Early work found modest cross-channel correlations (13–25% shared variance); more recent work finds essentially none across cardiac, respiratory, gastric, pain, bitterness and balance (Ferentzi et al. 2018 et al.). This is the domain-generality question, reached from the construct-validity side.
- Accuracy does not even converge within a channel. The sharpest single number in the paper: Hickman et al.’s (2020) meta-analysis of 22 studies found the Heartbeat Counting Task and Heartbeat Discrimination Task — both said to measure “interoceptive accuracy” in the same organ — share 4.4% of variance (r = 0.21). See is-the-heartbeat-counting-task-valid.
- Sensibility questionnaires barely converge (r = -0.63 to 0.65). Desmedt et al.’s own (2022) study of the most-cited scales in n = 1003 resolved them into five weakly-related factors, confirming they tap distinct constructs — the same finding Ventura-Bort et al. (2021) reach on the sensibility entry specifically.
- The HCT is contaminated by guessing from a believed heart rate (Corneille et al. 2020; Desmedt et al. 2018, 2020) — the mechanism the counting-task debate turns on.
The consequence is stated with unusual statistical bite. Low convergence between measures presumed to assess the same construct makes cross-measure replication failure the expected outcome, not a crisis. Carlson & Herdman (2012) showed that if two measures correlate r = 0.30 and one correlates r = 0.30 with an outcome, the other’s correlation with that outcome can range from -0.95 to 0.95. Since the HCT and HDT correlate only r = 0.21, giving them the same label (“interoceptive accuracy”) manufactures apparent inconsistency out of measures that were never entitled to agree. The recommendation that follows: do not transfer a conclusion from one measure to a broad construct or another measure without evidence of convergence — “we cannot conclude that interoceptive accuracy is associated with depression based on studies showing a correlation between the Heartbeat Counting Task and depressive symptoms.”
The recommended definition
Interoception includes the top-down and bottom-up processes by which an organism senses, interprets, and integrates signals from within itself and below the skin, across conscious and non-conscious levels.
Four deliberate choices, each a response to a reviewed alternative:
- “below the skin” — the skin is the barrier (phenomenon-based), which keeps the definition operational where a pure-physiology definition would be measurable only under laboratory pathway-inhibition.
- classical exteroceptive senses excluded — sight, hearing, smell, taste and touch are ruled out even when trained on the body (hearing one’s heartbeat, seeing one’s chest move) to prevent overlap with exteroception. This narrows Chen et al.’s (2021) “represents rather than originates from” criterion, which would let hearing one’s heartbeat count.
- “top-down and bottom-up” — explicitly builds in predictive processing: perception is priors plus incoming signal, consistent with active inference and computational models. This is what the four inclusive definitions they review (Cameron 2001; Ceunen et al. 2016; Chen et al. 2021; Khalsa/Berntson 2018/2021) each lack in whole or part.
- efferent regulation excluded — unlike Chen et al.’s “senses, interprets, integrates, and regulates,” the descending/regulatory limb is left out, on the ground that it folds fat metabolism and reflex control into the definition.
They add that this is compatible with a continuous physiological difference between interoception and exteroception (Carvalho & Damasio 2021: interoception is proportionally more unmyelinated/lightly-myelinated fibre) — they deny only that physiology draws a categorical boundary.
The hierarchical framework (illustrative)
The proposal for the second discrepancy is structural: replace a flat list of broad dimensions with a hierarchy organised by level of specificity. The top level is interoception; below it, broad factors — interoceptive attention, sensing, interpretation, memory (the first framework to include non-conscious processing and an interoceptive-memory factor); below those, subfactors mapping the old dimensions onto specific constructs (e.g. “interoceptive detection” = Garfinkel’s accuracy; “interoceptive localization” = Khalsa’s discrimination; “somatosensory amplification” and “interoceptive worrying” under interpretation); and at the bottom, measure-related subfactors where one construct has one measure (the original Schandry HCT becomes “the capacity to estimate heart rate via mental counting” and nothing broader). The hierarchy is by specificity, not by processing depth — which is how it differs from the otherwise-similar Suksasilp & Garfinkel (2022) framework it cites. The authors stress the whole scheme is illustrative and needs empirical construction via a formal construct-validity pipeline (Clark & Watson 2019).
Why it matters to this wiki
This is the paper the wiki’s two measurement pages were circling. It converts scattered cautions into a single position with a definition attached:
- It is the canonical citation for the pathway-criterion’s failure on interoception-exteroception-boundary — the “no necessary-and-sufficient pathway” argument is the general form of that page’s nociception and affective-touch case studies.
- It supplies interoceptive-taxonomy with the meta-level diagnosis (broad constructs guarantee low convergence) the three-taxonomies section had reached empirically but not named, plus a constructive alternative.
- It gives is-interoception-domain-general a field position that is a synthesiser rather than a single dataset, and the Carlson & Herdman argument that low convergence makes cross-measure inconsistency expected.
- It gives is-the-heartbeat-counting-task-valid the Hickman et al. (2020) within-domain 4.4%-variance number, the strongest single statistic against treating counting and discrimination as one construct.
The limitation to keep in view: it is a review, and the convergence evidence is largely the authors’ own prior empirical work plus Ferentzi and Hickman. It argues the field should stop over-generalising; it does not itself demonstrate a new interoceptive fact.