Skip to content

Neurobiology and cognition

TL;DR — Reward learning, threat, interoception, cognitive control, set-shifting and habit have all been implicated in AN, but most evidence is cross-sectional and vulnerable to starvation effects. Neuroimaging findings are group averages, not diagnostic markers. The strongest design need is repeated measurement before and after nutritional rehabilitation, with recovered and familial-risk comparators. A mechanistic model must explain both persistent restriction and why it differs across illness stages.

Candidate domains

Domain Recurrent observation Competing interpretation
Reward Altered response to food and monetary outcomes Trait vulnerability versus learned/starvation state
Threat/anxiety Anticipatory anxiety around eating and gain Cause, maintenance loop or consequence
Cognitive control High control during food decisions Adaptive laboratory strategy versus inflexible control
Habit Restriction becomes cue-driven and outcome-insensitive Habit model may fit only a subgroup/stage
Interoception Altered body-signal perception Sensory precision, belief, attention or malnutrition
Set-shifting Reduced flexibility in some meta-analyses Small effects; depression/OCD/weight confounding
Central coherence Detail focus Task and sample dependence

Reviews integrate evidence across neural and genetic levels but emphasize heterogeneity (Bulik 2022, PMID 35524137; Frank 2019, PMID 31443880). Earlier neurobiological accounts proposed altered reward/anxiety circuitry, while later work stresses developmental and metabolic integration (Kaye 2013, PMID 23333342; Bulik 2019, PMID 31056797).

The trait-versus-state argument predates the imaging era. An earlier synthesis observed that people with AN are consistently characterized by perfectionism, obsessive-compulsiveness and dysphoric mood, with high constraint, constricted affect, anhedonia and asceticism, and — the load-bearing claim — that these features often begin in childhood before the onset of an eating disorder and persist after recovery, which is the classic argument for a pre-existing vulnerability rather than a starvation artefact (Kaye 2008, PMID 18164737). Personality-trait continuity is weaker evidence than a prospective high-risk cohort, but it is the observation the state-trait designs below were built to test.

State–trait problem

Design Strength Residual limitation
Acutely ill vs control Detects illness-associated difference Maximum starvation confounding
Pre/post weight restoration Tests reversibility Treatment, time and selection co-vary
Recovered vs control Identifies persistent differences Recovery selection and scars
Unaffected relatives Enriches familial liability Shared environment; low specificity
Prospective high-risk cohort Establishes temporal order Large sample and long follow-up needed

Malnutrition changes endocrine signalling across the hypothalamic-pituitary axis, and the review that catalogues those changes states explicitly that they "might affect neurocognition, anxiety, depression, and the psychopathology of anorexia nervosa" — so almost every cognitive task or scan contrast is potentially downstream of nutritional state (Misra 2014, PMID 24731664). Orexin is one concrete example of a system that regulates the sleep-wake cycle, alertness, vigilance and cognitive functioning and whose central and peripheral levels are reported altered in anorexia nervosa; the wider orexin literature is not AN-specific and offers a mechanism worth testing rather than an established finding (Toor 2021, PMID 34052810). “Persists after recovery” is stronger than acute cross-section but does not prove premorbid causality.

Structural imaging: a large, substantially reversible finding

Brain structure shows a large and replicated neuroimaging effect that is strongly associated with acute nutritional state.

Study Sample Result
King 2015, PMID 25433902 40 acutely ill, 34 long-term weight-restored, 74 age-matched controls Significant cortical thinning across >85% of the cortical surface in acute AN, normalized in long-term recovered patients; mirrored in subcortical volumes; normal age-related thinning trajectories absent during illness; cortical thickness correlated negatively with drive for thinness in extrastriate body-perception regions
Bernardoni 2016, PMID 26876474 47 acutely ill (35 rescanned), 34 recovered, 75 controls Cortical thickness increased at 0.06 mm/month over ~3 months of weight-restoration therapy; the increase was predicted by weight gain alone and not by illness duration, hydration status or symptom improvement; thickness had already normalized at follow-up after >10% BMI increase
Kaufmann 2020, PMID 32366823 26 adult women with severe AN scanned at three BMI stages, 30 controls Reversal confirmed in adults; restoration fastest in the first half of treatment; rate correlated negatively with age but not with duration of illness; residual thinning persisted in superior frontal cortex after weight restoration
Bahnsen 2022, PMID 35390458 Replication (75 acute, 34 recovered, 139 controls) plus mega-analysis (120 acute, 68 recovered, 207 controls) Confirmed widespread thickness and volume reductions and rapid post-refeeding increase; no differences between short- or long-term weight-recovered patients and controls; virtual histology linked the most affected regions to S1 pyramidal-cell and oligodendrocyte expression profiles, and those regions were the most structurally and functionally connected — i.e. the most metabolically demanding

Three consequences follow. First, cross-sectional "brain differences in AN" are substantially confounded by malnutrition. Second, the speed of reversal (months, not years) argues against neuronal loss and against simple dehydration, and points toward glial and macromolecular remodelling — Bernardoni and colleagues state this explicitly as drawing dehydration and neurogenesis explanations into question (PMID 26876474). Third, the absence of normal age-related cortical trajectories during illness raises a developmental question, but it too may be state-dependent and has not been tested prospectively.

Reward: prediction error and its direction of causation

The most systematically pursued functional finding in AN is an elevated dopamine-related reward prediction error (PE) response during sucrose taste conditioning.

  • In 21 female adolescents with AN scanned before and after treatment against 21 controls, PE response was elevated in caudate, ventral caudate/nucleus accumbens and anterior and posterior insula; PE and unexpected-reward-omission responses tended to normalize with treatment while unexpected-reward-receipt response remained elevated; and greater caudate PE response when underweight predicted less weight gain during treatment (DeGuzman 2017, PMID 28231717).
  • A larger multimodal replication (56 AN, mean age 16.6, mean BMI 15.9; 52 controls) confirmed elevated PE in caudate head, nucleus accumbens and insula (Wilks λ 0.707, p = .02, partial η² = 0.296). Orbitofrontal PE correlated positively with harm avoidance (ρ ≈ 0.32) and negatively with treatment BMI change (ρ ≈ −0.28), and ventral striatal→hypothalamic effective connectivity strength correlated with insula and orbitofrontal PE (Frank 2018, PMID 30027213).
  • Extending the paradigm across 317 young women (197 with eating disorders, 120 controls) reframed it: PE response was elevated in AN but, across the eating-disorder spectrum, inversely correlated with BMI (left nucleus accumbens r = −0.291, 95% CI −0.413 to −0.167, p < .001) and with binge-eating tendency and trait anxiety (Frank 2021, PMID 34190963).

That last result is the crux. If PE magnitude tracks BMI continuously across diagnoses, the elevation in AN may be a readout of energy status rather than a disorder-specific trait — the authors' own conclusion is that BMI modulates the circuit. The predictive finding (higher PE → less weight gain) is nevertheless clinically interesting whichever way causation runs, and has been replicated within the same group but not independently.

Habit: a model with genuinely mixed support

The habit account holds that restriction, initially goal-directed, becomes automatic, outcome-insensitive and dorsal-frontostriatally mediated (Steinglass 2006, PMID 16523472; Uniacke 2018, PMID 30039342). Its strongest evidence is disorder-specific: during a food-choice task, individuals with AN engaged the dorsal striatum more than controls, and fronto-striatal activity correlated with what they actually ate at a laboratory meal the following day (Foerde 2015, PMID 26457555).

Its weakest evidence is the general case. Two parallel outcome-devaluation experiments — the slips-of-action paradigm in 23 women with AN vs 18 controls, and a second study adding 14 recovered participants and an aversive-habit paradigm — found no bias toward habitual over goal-directed control in AN, acute or recovered (Godier 2016, PMID 27497292). The reconciliation offered in the literature is that AN habits are domain-specific to eating rather than a general learning-style deficit (Uniacke 2018, PMID 30039342). Current food-choice tasks do not cleanly distinguish that account from one in which food-specific valuation, rather than habit, is altered. This remains unresolved.

Set-shifting, coherence and the autism overlap

The neuropsychological profile is real but small, and — importantly — not specific to being underweight. A meta-analysis of 16 studies (1,112 participants) for central coherence and 38 studies (3,505 participants) for set-shifting found AN significantly poorer than controls on central coherence (Hedge's g = −0.53, 95% CI −0.80 to −0.27) and set-shifting (g = −0.38, 95% CI −0.50 to −0.26) — but bulimia nervosa showed comparable inefficiencies (g = −0.70 and −0.55) that did not significantly differ from AN (Keegan 2021, PMID 33305366). A profile shared with a non-underweight eating disorder cannot be attributed to low weight, and cannot be used as an AN-specific endophenotype.

Duration matters: across 53 studies, deficits in central coherence, cognitive flexibility and emotion recognition were more pronounced in longer-duration AN, which the authors read as bidirectional — impairments may prolong illness and prolonged illness may deepen impairments ("neurological scar effect") (Saure 2020, PMID 32181530). The same profile is the reason autism and AN are repeatedly linked. A meta-analysis of 22 studies (1,172 AN, 2,747 controls) found substantially elevated autistic traits (g = 0.88, 95% CI 0.65–1.12), a modest correlation with eating-disorder symptom severity (r = 0.28, 95% CI 0.11–0.44), and 29% (95% CI 19–38%) of AN participants scoring above an ADOS cut-off (Inal-Kaleli 2025, PMID 39530423). Whether that represents genuine co-occurrence, starvation-induced phenocopy, or measurement overlap is unsettled; exceeding an ADOS cut-off is not equivalent to a confirmed clinical diagnosis.

Interoception

Insula findings persist beyond acute starvation, which makes them one of the few candidate trait markers. In 15 weight-restored, medication-free women with restricting-type AN and 15 controls performing an interoceptive attention task, a group × modality interaction was significant in two distinct insular regions: dorsal mid-insula, driven by stomach interoception (p = 0.002, Bonferroni corrected), and anterior insula, driven by heart interoception (p = 0.03). Dorsal mid-insula activation during stomach interoception correlated with anxiety and psychopathology measures (Kerr 2016, PMID 26084229). Altered insula response to sweet taste has likewise been reported after recovery from both AN and bulimia nervosa (Oberndorfer 2013, PMID 23732817). Sample sizes are 15–30 per group; these are hypothesis-generating.

Leptin, hyperactivity and the animal model

Hyperactivity in AN is not simply volitional exercise. Semi-starvation reliably induces wheel-running in rodents — the activity-based anorexia (ABA) paradigm, which reproduces food restriction, hyperactivity, weight loss and endocrine change and is the field's principal mechanistic model (Spadini 2021, PMID 34600568; Foldi 2024, PMID 38103992). The leptin hypothesis proposed that hypoleptinaemia drives this: leptin administration suppressed semi-starvation-induced hyperactivity in rats (Exner 2000, PMID 11032380), and the clinical implications were argued in detail (Hebebrand 2007, PMID 17060920).

It has since been directly contradicted. In 74 Sprague-Dawley rats compared under matched food restriction, housing at 32 °C significantly reduced wheel-running and weight loss whereas leptin infusion did not; leptin-treated animals at 21 °C had significantly reduced body temperature during restriction, and those without a stable prior activity pattern showed circadian collapse. The authors' conclusion is that ambient temperature plays a more critical role than leptin (Fraga 2020, PMID 32210308). Both results are in the literature; neither has been overturned.

On the human side the evidence is a three-patient off-label case series: metreleptin for up to 14 days reduced drive for activity, food preoccupation, inner restlessness and weight phobia in two of three severely ill women, with rapid improvement in depression in all three and no serious adverse events — reported by its authors as requiring placebo-controlled confirmation (Milos 2020, PMID 32855384). That is a signal, not evidence of efficacy; see pharmacotherapy.

Biomarker status

Candidate imaging, electrophysiological and peripheral markers are not validated diagnostic tests. An opinion review surveying genetic, metabolomic, microbiomic, endocrine, immunological, haematological, electrophysiological and neuroimaging parameters positions all of them as adjuncts across the treatment cycle — diagnosis, diagnostic specification, risk management, choice of therapy, therapy monitoring and treatment review — and concludes that history-taking, physical and neuropsychological examination, clinical observation and the judgements of the patient, carers and multidisciplinary team remain essential to interpret them (Himmerich 2024, PMID 38331700). The criteria a marker would have to meet to displace any of that — a locked measurement protocol, a prespecified threshold, external replication, incremental value beyond clinical assessment and a demonstrated clinical use — are this wiki's framing, not the review's claim, and as of September 2026 no biomarker in anorexia nervosa satisfies them. Classification accuracy from resampled small datasets is not clinical validity.

Treatment implications

Cognitive remediation, exposure, habit-focused interventions and neuromodulation derive rationales from these models. The adult network meta-analysis found low-confidence comparative evidence across psychological interventions (Solmi 2021, PMID 33600749), while most mechanism-focused studies do not test whether target engagement mediates durable clinical outcome. The required chain is model → measurable target → target engagement → clinical mediation.

Open questions

  • Which neural/cognitive findings precede weight loss and predict onset (Bulik 2022, PMID 35524137)?
  • Does change in habit or reward learning mediate sustained nutritional recovery?
  • Which abnormalities normalize with weight restoration, and on what time scale (Frank 2019, PMID 31443880)?
  • Is elevated reward prediction error a trait of AN or a readout of energy status, given that it scales inversely with BMI across the whole eating-disorder spectrum (Frank 2021, PMID 34190963; DeGuzman 2017, PMID 28231717)?
  • Are AN "habits" domain-specific to eating, or is the habit model unfalsifiable as currently specified, given intact general outcome-devaluation performance (Foerde 2015, PMID 26457555; Godier 2016, PMID 27497292)?
  • Does the absence of normal age-related cortical thinning during illness have developmental consequences, and is it itself a nutritional-state effect (King 2015, PMID 25433902; Bahnsen 2022, PMID 35390458)?
  • Do the 29% of people with AN who score above an ADOS cut-off represent co-occurring autism, a starvation phenocopy, or instrument overlap (Inal-Kaleli 2025, PMID 39530423; Saure 2020, PMID 32181530)?
  • Is hypoleptinaemia or thermoregulation the dominant driver of starvation-induced hyperactivity, and does the answer transfer to humans (Exner 2000, PMID 11032380; Fraga 2020, PMID 32210308; Milos 2020, PMID 32855384)?

References

  1. Bulik CM, et al. Genetics and neurobiology of eating disorders. Nat Neurosci. 2022. PMID 35524137.
  2. Frank GKW. The neurobiology of eating disorders. Child Adolesc Psychiatr Clin N Am. 2019. PMID 31443880.
  3. Kaye WH, et al. Nothing tastes as good as skinny feels: the neurobiology of anorexia nervosa. Trends Neurosci. 2013. PMID 23333342.
  4. Bulik CM, et al. Reconceptualizing anorexia nervosa. Psychiatry Clin Neurosci. 2019. PMID 31056797.
  5. Kaye W. Neurobiology of anorexia and bulimia nervosa. Physiol Behav. 2008. PMID 18164737.
  6. Misra M, Klibanski A. Endocrine consequences of anorexia nervosa. Lancet Diabetes Endocrinol. 2014. PMID 24731664.
  7. Toor B, et al. Sleep, orexin and cognition. Front Neurol Neurosci. 2021. PMID 34052810.
  8. Himmerich H, et al. Anorexia nervosa: diagnostic, therapeutic, and risk biomarkers in clinical practice. Trends Mol Med. 2024. Opinion article. PMID 38331700.
  9. Solmi M, et al. Comparative efficacy and acceptability of psychological interventions for adult outpatients with anorexia nervosa. Lancet Psychiatry. 2021. PMID 33600749.
  10. King JA, et al. Global cortical thinning in acute anorexia nervosa normalizes following long-term weight restoration. Biol Psychiatry. 2015;77:624-632. PMID 25433902.
  11. Bernardoni F, et al. Weight restoration therapy rapidly reverses cortical thinning in anorexia nervosa: a longitudinal study. Neuroimage. 2016;130:214-222. PMID 26876474.
  12. Kaufmann LK, et al. Age influences structural brain restoration during weight gain therapy in anorexia nervosa. Transl Psychiatry. 2020;10:126. PMID 32366823.
  13. Bahnsen K, et al. Dynamic structural brain changes in anorexia nervosa: a replication study, mega-analysis, and virtual histology approach. J Am Acad Child Adolesc Psychiatry. 2022;61:1168-1181. PMID 35390458.
  14. DeGuzman M, et al. Association of elevated reward prediction error response with weight gain in adolescent anorexia nervosa. Am J Psychiatry. 2017;174:557-565. PMID 28231717.
  15. Frank GKW, et al. Association of brain reward learning response with harm avoidance, weight gain, and hypothalamic effective connectivity in adolescent anorexia nervosa. JAMA Psychiatry. 2018;75:1071-1080. PMID 30027213.
  16. Frank GKW, et al. Association of brain reward response with body mass index and ventral striatal-hypothalamic circuitry among young women with eating disorders. JAMA Psychiatry. 2021;78:1123-1133. PMID 34190963.
  17. Steinglass J, Walsh BT. Habit learning and anorexia nervosa: a cognitive neuroscience hypothesis. Int J Eat Disord. 2006;39:267-275. PMID 16523472.
  18. Uniacke B, et al. The role of habits in anorexia nervosa: where we are and where to go from here? Curr Psychiatry Rep. 2018;20:61. PMID 30039342.
  19. Foerde K, et al. Neural mechanisms supporting maladaptive food choices in anorexia nervosa. Nat Neurosci. 2015;18:1571-1573. PMID 26457555.
  20. Godier LR, et al. An investigation of habit learning in anorexia nervosa. Psychiatry Res. 2016;244:214-222. PMID 27497292.
  21. Keegan E, et al. Central coherence and set-shifting between nonunderweight eating disorders and anorexia nervosa: a systematic review and meta-analysis. Int J Eat Disord. 2021;54:229-243. PMID 33305366.
  22. Saure E, et al. Characteristics of autism spectrum disorders are associated with longer duration of anorexia nervosa: a systematic review and meta-analysis. Int J Eat Disord. 2020;53:1056-1079. PMID 32181530.
  23. Inal-Kaleli I, et al. Investigating the presence of autistic traits and prevalence of autism spectrum disorder symptoms in anorexia nervosa: a systematic review and meta-analysis. Int J Eat Disord. 2025;58:66-90. PMID 39530423.
  24. Kerr KL, et al. Altered insula activity during visceral interoception in weight-restored patients with anorexia nervosa. Neuropsychopharmacology. 2016;41:521-528. PMID 26084229.
  25. Oberndorfer TA, et al. Altered insula response to sweet taste processing after recovery from anorexia and bulimia nervosa. Am J Psychiatry. 2013;170:1143-1151. PMID 23732817.
  26. Spadini S, et al. Activity-based anorexia animal model: a review of the main neurobiological findings. J Eat Disord. 2021;9:123. PMID 34600568.
  27. Foldi CJ. Taking better advantage of the activity-based anorexia model. Trends Mol Med. 2024;30:330-338. PMID 38103992.
  28. Exner C, et al. Leptin suppresses semi-starvation induced hyperactivity in rats: implications for anorexia nervosa. Mol Psychiatry. 2000;5:476-481. PMID 11032380.
  29. Fraga A, et al. Temperature but not leptin prevents semi-starvation induced hyperactivity in rats: implications for anorexia nervosa treatment. Sci Rep. 2020;10:5300. PMID 32210308.
  30. Hebebrand J, et al. The role of leptin in anorexia nervosa: clinical implications. Mol Psychiatry. 2007;12:23-35. PMID 17060920.
  31. Milos G, et al. Short-term metreleptin treatment of patients with anorexia nervosa: rapid on-set of beneficial cognitive, emotional, and behavioral effects. Transl Psychiatry. 2020;10:303. PMID 32855384.