Computational psychiatry / computational psychosomatics
What the predictive-coding account of interoception looks like when it is pointed at mental illness. Khalsa et al. (2018) fold the wiki’s theoretical machinery — generative models, prediction error, precision, active-inference, allostasis — into a clinical programme: if interoception is inference and control, then interoceptive disorders are specific failures of that inference and control, and the failures are enumerable. This page holds the mechanism; interoceptive-psychopathology holds the clinical nosology it is meant to explain.
The core claim: control needs a model, and disease is where the model fails
The argument the roadmap imports (from Stephan et al. 2016 and Pezzulo, Rigoli & Friston 2015) runs:
- Sensing alone cannot resolve the state of the body, because interosensations are noisy and ambiguous. So the brain runs a generative model and infers.
- The brain’s most fundamental task is not perception but control of bodily states — keeping them in survivable ranges. In information-theoretic terms, choosing actions that minimize the long-term average surprise (entropy) of interosensations.
- Inference and control form a closed loop cast as a hierarchical Bayesian model: predictions, prediction errors, precisions. Reactive homeostatic regulation sits at the bottom (prediction errors trigger reflexes); prospective allostatic regulation modulates the set-points from above.
The mechanistically loaded piece, carried in from Petzschner/Stephan: belief precision determines the force and pace of corrective action — “the tighter the expected range of bodily state, the more vigorous the elicited regulatory action.” A body-state prior held with pathologically high precision therefore drives over-vigorous regulation; this is offered as a novel account of psychosomatic phenomena and placebo effects, and it is where “computational psychosomatics” gets its name (Petzschner et al. 2017 propose it as the basis of a mechanistic differential diagnosis).
Allostatic self-efficacy: a computational route to fatigue and depression
The roadmap’s most concrete clinical hypothesis in this vein is Stephan et al.’s allostatic self-efficacy — a metacognitive belief about one’s own capacity to regulate bodily states. On this account, persistent unresolved interoceptive surprise (dyshomeostasis the system cannot control) is read by higher levels as low self-efficacy over the body, and the downstream psychological expression is fatigue and depression. Depression becomes, in part, a metacognitive verdict about failed bodily control rather than only a mood state. The wiki holds this at one remove (Stephan et al. 2016 is not in raw/), but it connects directly to allostasis, where the precision-weighting account is already recorded, and to the wiki’s own clinical depression material (the Farb & Segal work), which reaches depression from a different, non-computational direction.
The failure mode, in EPIC terms
The roadmap’s psychopathology figure reproduces Barrett & Simmons’s EPIC model and states its illness hypothesis in precision language: interoceptive input (posteriors) becomes decoupled from interoceptive predictions (priors), producing elevated interoceptive prediction error that “may present in the brain as ‘noisy afferent interoceptive inputs’” (Paulus & Stein 2010). Under normal function the agranular visceromotor cortex is relatively insensitive to ascending error, which is why interoceptive predictions stay stable despite bodily fluctuation; the pathological case is that insensitivity breaking down (or over-tightening). See theory-of-constructed-emotion and insular-cortex for the anatomy.
The depression/fatigue lock-in, worked through (Quadt et al. 2018)
Quadt, Critchley & Garfinkel (2018) give the allostatic-self-efficacy and EPIC-failure accounts above their most concrete narrative, and the wiki now holds the mechanism first-hand rather than only through Khalsa’s summary. The fatigue version: aberrant metacognitive beliefs about the brain’s capacity to predictively control the body (low allostatic self-efficacy) can be sustained by chronic stress — raised cortisol disrupts NMDA-receptor function implicated in belief-updating (Corlett et al.), a positive feedback loop grounding the metacognitive verdict that the system cannot regulate itself, so resorting to sickness behaviours and fatigue is an adaptive response to that verdict.
The depression version (drawing on Barrett, Quigley & Hamilton 2016) is the wiki’s clearest statement of the EPIC lock-in: visceromotor cortical dysfunction overpredicts metabolic demand, driving HPA overactivity and pro-inflammatory cytokines; the resulting elevated interoceptive prediction error is downweighted by precision units so it can no longer inform predictions; to reduce error the network can only maintain the dysfunctional predictions or generate afferents that match them, producing noisier signals that still fail to update the model. The faulty predictions keep demanding energy until the endocrine/immune system is exhausted, at which point “depression, according to EPIC, ensues when the error signals can finally no longer be ignored and must be reduced, enlisting sickness behaviours to conserve energy.” A “locked-in” attractor state — a vicious cycle of faulty predictions and noisy error — biasing the system toward avoidance and withdrawal. This is the same two-factor architecture (granular error vs agranular prediction) the wiki tracks elsewhere, run forward into a self-sustaining pathology.
Computational biomarkers, and the price of clinical relevance (Petzschner et al. 2021)
Petzschner et al. (2021) give this page its author’s own statement of the programme, three years after the roadmap borrowed it — and the emphasis lands differently when it comes first-hand.
The deliverable has a name here: the computational biomarker. The promise is that fitting a generative model to an individual’s behaviour yields quantifiable parameter estimates — precision, learning rate, bias — “linked to illness pathophysiology,” and that these can be used “akin to a blood test” to point at that individual’s specific maladaptive computation, for diagnosis, prognosis or treatment selection. This is what distinguishes computational psychiatry from a re-description of symptoms: the claim is that the parameters are measurable in a person, not merely postulated about a population.
Two conditions are attached, and Petzschner states both plainly:
- The evidence base is nearly empty. “At present there is little empirical work testing such models’ predictions.” The three citations offered for such tests include one of her own (the heartbeat-evoked-potential/attention study) and one placebo-analgesia study — which makes the scarcity point rather than answering it.
- The bar is a drug-development pipeline. To become clinically relevant, computational models “will need to be rigorously translated through a development pipeline not unlike that used during novel drug identification.” An effective biomarker “must demonstrate their utility in improving diagnosis, monitoring, prediction, prognosis, risk susceptibility, or treatment, ideally in individual patients.”
That is a higher standard than this literature usually sets, set by people invested in clearing it. It is also, read against the roadmap’s “indirect so far” (2018) and Quadt et al.’s narrative accounts (2018), a three-year interval in which the framework’s reach grew — anxiety, depression, schizophrenia, somatic symptom disorder, chronic pain, fibromyalgia, functional neurological disorders, addiction — and its tested content did not visibly move.
One structural caution the framework’s clinical use has not absorbed
The same paper notes that models of forecasting — how internal states evolve over time — are the least formalized part of the whole enterprise, “especially in the internal domain.” But the psychopathology story on this page and on interoceptive-psychopathology is largely a story about mis-set predictions of future bodily states (Paulus & Stein’s failure-to-anticipate; the EPIC over-prediction of metabolic demand). A mis-set forecast presupposes a formal model of the forecast. The clinical account is resting its weight on the arc of the sensory-control-loop that nobody has written down.
And a rival that would relocate the pathology
If HRL is the right implementation for a given regulatory behaviour, the pathology is not in precision-weighted prediction at all — it is in valuation, drive, and reward prediction error, with no bodily percept necessarily involved. Petzschner et al. note that HRL has been used for exactly this purpose in addiction (Keramati et al. 2017, cocaine addiction as an HRL disorder). So “which computational disorder is this?” may not have a single vocabulary, and the wiki should not read every clinical claim on this page as forced by the evidence rather than by the framework chosen.
The empirical test finally arrives — and lands on the wrong term
Both concessions above (“little empirical work testing such models’ predictions”; “indirect so far”) can now be answered with exactly one study, and its result is not the one this page’s clinical story predicts.
Harrison et al. (2021), from Stephan’s own group with Petzschner on the author list, fitted an associative learning model to trial-by-trial breathing predictions in 60 people and regressed 7T brain activity on the model’s own per-trial estimates of prediction certainty and prediction error. Both quantities had neural correlates, with the signs predictive coding requires: anterior insula, ACC, MFG and dlPFC deactivated with certainty; aIns, ACC, MFG and the PAG activated with error magnitude.
Then the clinical part:
- Trait anxiety modulated the prediction term. A valence × group interaction in bilateral anterior insula, on prediction certainty — low-anxiety participants deactivated aIns more when confidently predicting an upcoming resistance; moderate-anxiety participants showed the opposite. That is Paulus & Stein’s insular account of anxiety receiving its first model-based support.
- Trait anxiety did nothing to the error term. No group effect and no interaction anywhere in prediction error.
The second result is a direct constraint on the machinery this page is built from. The EPIC failure mode recorded above locates illness in elevated interoceptive prediction error decoupled from priors; Brewer, Murphy & Bird’s atypical-interoception review and Paulus & Stein (2006, 2010) point the same way. Harrison et al. name the disagreement themselves: the null “contrast[s] with some previously proposed hypotheses.”
How much weight to put on it. Not much yet, and the reasons are worth stating so the page neither ignores nor overreads it. Anxiety is not depression, and the EPIC lock-in above is a depression story. A brief resistive load is not dyshomeostasis. The imaging used a reduced field of view over insula and midbrain, so a null outside that slab means nothing. n = 58, healthy extreme groups, one channel. And the fitted model was Rescorla-Wagner — chosen by pre-registered fallback after no model won selection — which has no precision term at all, so “prediction certainty” here is a transformed point estimate rather than an inferred precision (see interoceptive-precision).
What it does change. Until now every clinical claim on this page was a story about parameters nobody had estimated in a person. One study has now estimated two of them, related one to a psychiatric trait, and reported that the other — the one the framework leans on hardest — did not move. The computational biomarker remains unearned: the fitted learning rate predicted neither group membership nor any questionnaire in this sample. But the drug-development bar Petzschner set has had its first data point, and the honest reading is that the framework’s clinical hypotheses are now falsifiable in practice and one of them has taken a first hit.
One positive result on the same page’s side: FDT metacognitive performance correlated with aIns response to resistance-related prediction errors (r = 0.42) — support for the Stephan et al. (2016) proposal that bodily prediction error is the route by which deviation reaches metacognitive evaluation, which is the mechanism allostatic self-efficacy presupposes.
Where the prediction-error comparator might sit (Nord et al. 2021)
Harrison tested the computation trial-by-trial; Nord, Lawson & Dalgleish (2021) propose an anatomical location for it, from the opposite methodological pole — a transdiagnostic ALE meta-analysis of disrupted interoceptive activation across five psychiatric disorders that converges on the left dorsal mid-insula. Reading the dysgranular precentral insular gyrus’s hybrid connectivity to both the agranular anterior (prediction source) and granular posterior (sensory) insula, they nominate it as the candidate site for encoding interoceptive prediction error, and cast transdiagnostic pathology as a disruption there — “higher weighting of prediction errors” in some conditions, “increased precision of prediction models” in others.
This is a different placement from the EPIC anatomy above (granular mid-and-posterior insula returns error; agranular anterior issues prediction) and from Harrison’s finding that the anterior insula carried both terms. The three do not agree on where in the insula interoceptive prediction error lives — mid (Nord, from convergence + connectivity), anterior (Harrison, from a fitted model), or mid/posterior (EPIC, from cytoarchitecture) — and none is decisive: Nord’s is an anatomical inference from a disruption map with no computational parameters in it, and the paper fitted no learning model. It is the framework’s clinical claim given a specific, falsifiable coordinate rather than a confirmation of it. See insular-cortex, interoceptive-psychopathology, feedforward-vs-predictive-interoception.
The honest caveat, stated by the authors
The roadmap does not oversell this. “The empirical evidence for hierarchical Bayesian principles of interoception and homeostatic/allostatic control is indirect so far. Studies designed to probe hierarchical Bayesian processes under experimentally controlled homeostatic perturbations will be crucial for finessing (or refuting) current computational concepts.” That is the same status the wiki assigns the predictive-coding account generally on feedforward-vs-predictive-interoception: elegant, unifying, anatomically motivated, and not yet the thing that a decisive experiment has confirmed over the feedforward alternative. What computational psychiatry adds is a reason to run those experiments in patients — because if the framework is right, the precision parameters are where the disorders live.