Companion assets

Asset	Link
Interactive review (HTML)	/evidence-reviews/stress-energy-and-the-capacity-to-function/Stress-Energy-and-Function-Review.html
Evidence table (XLSX, 108 findings)	/evidence-reviews/stress-energy-and-the-capacity-to-function/Stress-Evidence-Table.xlsx
Practical brief	/evidence-reviews/stress-energy-and-the-capacity-to-function/Stress-and-Energy-Practical-Brief.md

Companion review: Anxiety — a critical review of the evidence.

---

Executive summary

Ask why chronic stress leaves people unable to function and you will usually be told a story about cortisol wearing out the adrenal glands and willpower running down like a battery. Both halves of that story are wrong, and the versions that replace them are more interesting.

Nine things the evidence actually supports:

1. Chronic stress is hypermetabolic, not hypometabolic. The best-evidenced version of "stress depletes energy" is not that you run out of fuel, but that maintaining the stressed state is expensive and the bill is paid out of growth, maintenance and repair. Under chronic glucocorticoid exposure, human fibroblasts raise energy expenditure ~60% and age faster on three independent clocks — and the causal experiment shows it is total energy expenditure itself, not any specific dysfunction, that drives the damage (Bobba-Alves et al., 2023).

2. The cortisol story is much weaker than the folk model, and "adrenal fatigue" is refuted. A systematic review of 58 studies concluded flatly that "adrenal fatigue is still a myth" (Cadegiani & Kater, 2016). Real HPA dysregulation exists but predicts health weakly (flatter diurnal slope → poorer health, r = 0.147). Two systematic syntheses nine years apart found no reliable cortisol signature of burnout at all.

3. Willpower is not a depletable resource. Ego depletion failed two enormous preregistered multi-lab tests: d = 0.04 across 23 labs (Hagger et al., 2016) and d = 0.06 across 36 sites (Vohs et al., 2021). Any model of stress-related incapacity built on depletion is built on sand.

4. The better account is that stress raises the price of effort, not that it removes the capacity. Effort is a computed cost entering a value calculation. A day of demanding cognitive work raises glutamate in lateral prefrontal cortex and shifts choices toward low-cost, short-delay options (Wiehler et al., 2022). This distinction — capacity loss versus raised effort price — changes which interventions make sense and removes the moral judgement that usually attaches to "not being able to face it."

5. Acute stress effects on cognition are real but small and fractionated; chronic exhaustion is where the broad damage shows. Acute stress moves working memory only g = −0.20 and does not impair inhibition overall. Clinical burnout, by contrast, impairs attention/processing speed (g = −0.43), executive function, memory and fluency, while leaving crystallised ability untouched.

6. The one large, unambiguous cognitive effect is on memory retrieval, not storage. Stress leaves encoding roughly intact, slightly enhances consolidation, and selectively degrades retrieval — worse for emotionally charged material. That is the mechanism of "blanking" under pressure.

7. Burnout has no stages. The popular twelve-stage model has zero empirical validation in the indexed literature. Longitudinal cohorts recover trajectory classes distinguished by severity and treatment response, never by stage position.

8. The chronic tail is far longer than commonly appreciated. Seven years after clinical exhaustion disorder, only 16% report full recovery and 46% still report extreme fatigue — yet 87% are back at work (Glise et al., 2020). Depression and anxiety resolve in months; exhaustion does not.

9. Individual-level workplace wellness programmes do not work at scale — and the organisational alternative is weaker than its advocates claim. A cluster-randomised trial of 32,974 employees found 2 of 80 prespecified outcomes significant, both self-reported (Song & Baicker, 2019). But the best randomised test of actual work redesign produced eight extra minutes of sleep per night.

What actually helps, ranked by evidence quality × effect size × deliverability: treat the sleep disorder (CBT-I, the only strong guideline recommendation here) → move the body → a structured behavioural protocol (behavioural activation or CBT) → screen for and treat depression → protect recovery structurally rather than exhorting it → change the job. Below that line — mindfulness apps, adaptogens, micro-break gimmicks, corporate resilience training, supplements in people who are not deficient — effects either vanish against active controls or were never there.

And one competing explanation that must be excluded first: chronic sleep restriction produces exactly the fatigue and cognitive impairment people attribute to stress, and — decisively — subjective sleepiness saturates while objective impairment keeps accumulating (Van Dongen et al., 2003). People restricted to six hours are systematically unaware of their own deficit. Any causal model of stress-related fatigue should control for sleep duration before reaching for the HPA axis.

---

How to read this review

Three structural problems that condition every number below

Control-group inflation. Effect sizes in this field collapse as controls get more active. The cleanest demonstration is mindfulness: d = 0.55 versus no treatment, 0.35 versus non-specific active control, 0.23 versus specific active control, and −0.004 versus an evidence-based treatment (Goldberg et al., 2018). Whenever you see an impressive effect size, the first question is what it was compared against.

Self-report and unblinding. Almost every outcome here is a questionnaire filled in by an unblinded participant. The honest counterweight: MetaBLIND found no average difference between blinded and unblinded trial estimates (ROR 0.91, 95% CrI 0.61–1.34; Moustgaard et al., 2020). This is a caution, not a proof of inflation.

Publication bias, sometimes severe. Trim-and-fill cut the app-for-depression effect from g = 0.45 to 0.18 — a ~60% reduction (Kulke et al., 2025). The IPD-Work consortium documented its own: published job-strain→CHD estimates 1.43 versus unpublished 1.16.

Two conventions used throughout

Claims are marked [contested] where the literature genuinely disagrees, and [unverified] where a figure is widely repeated but could not be traced to a primary source in preparing this review. A full register of unverified claims appears in Appendix C — it is deliberately long, because the alternative is quiet laundering of numbers that nobody has checked.

Effect sizes are reported as published. Where a confidence interval could not be verified against the primary text, it is omitted rather than reconstructed.

A note on what this review corrected

Five headline claims were adversarially fact-checked and required revision before inclusion:

• The claim that "cortisol increases deep sleep, so the popular story is backwards" turned out to hold only for acute exogenous cortisol in healthy volunteers. In chronic hypercortisolism (Cushing's syndrome) delta sleep is reduced — 5.8% versus 14.0% in controls — and patients with ACTH-independent Cushing's still show disrupted sleep, which defeats the "it's CRH drive, not cortisol" attribution. Section 1.6 states the narrowed version.
• The frequently quoted line that burnout "does not present the unity expected of a distinct syndrome" is misattributed to Bianchi et al. (2021); the phrase "no syndromal unity" belongs to Verkuilen et al. (2021, Assessment). Corrected in §3.2.
• "Acute stress enhances response inhibition" is a post-hoc moderator result on roughly five degrees of freedom; the overall inhibition effect was null. Corrected in §2.2.
• Fleming (2024) is cross-sectional, not randomised, and the widely repeated per-intervention findings could not be verified from an accessible primary source. Flagged in §4.6.
• Stadje et al.'s (2016) finding that somatic disease is "identical" in tired and untired patients rests on two single studies, one powered at 0.09. Softened in §4.9.

This list is included not as housekeeping but because it is representative: in this literature, roughly one headline claim in three needs narrowing once you read the primary source.

Part I — The biology: what "energy" actually means

1.1 The two response systems, briefly

Two axes carry the stress response on different timescales.

The sympathetic-adrenal-medullary (SAM) axis operates in seconds to minutes: norepinephrine from sympathetic terminals, epinephrine from the adrenal medulla. Its metabolic signature is immediate and large — epinephrine infusion raises blood glucose 50–60% and transiently triples splanchnic glucose output (Saccà et al., 1983).

The hypothalamic-pituitary-adrenal (HPA) axis operates over minutes to hours: CRH and vasopressin from the paraventricular nucleus → ACTH from the anterior pituitary → cortisol from the adrenal cortex. Negative feedback runs at pituitary, hypothalamic and hippocampal/prefrontal levels.

The single most useful conceptual correction to the folk model comes from Sapolsky, Romero & Munck (2000), who dismantled the assumption that glucocorticoids simply "mediate" the stress response. They partition glucocorticoid actions into four categories — permissive, suppressive, stimulatory and preparative — and show that which one applies depends entirely on the endpoint. Cortisol suppresses immune activation while stimulating hepatic gluconeogenesis. It is not a single-valence hormone, and any energy-depletion model treating it as uniformly catabolic is oversimplifying.

The second correction concerns receptors. Cortisol acts through two with roughly tenfold different affinities: the high-affinity mineralocorticoid receptor (MR), largely occupied at basal concentrations and governing tonic appraisal and circadian set-point, and the lower-affinity glucocorticoid receptor (GR), recruited only at stress-range concentrations and driving feedback, energy mobilisation and memory consolidation. De Kloet, Joëls & Holsboer (2005) argue that the pathology lies in the imbalance of this dual system, not the absolute amount of hormone — the ratio and the dynamics, not the level.

1.2 Why the cortisol story is weaker than the folk model needs

The intuitive picture — chronic stress means chronically high cortisol — is wrong, and the field has known it for two decades.

Miller, Chen & Zhou's (2007) meta-analysis, pointedly titled "If it goes up, must it come down?", found that timing dominates. HPA output is elevated at stressor onset and declines as time passes. Stressors that are traumatic, uncontrollable or threaten physical integrity produce a high, flat diurnal profile. HPA activity tracks subjective distress — but is lower in people with PTSD. Chronic stress therefore produces hypercortisolism, hypocortisolism, or neither, depending on stressor type, elapsed time and interpretation.

Fries et al. (2005) formalise the hypocortisolism phenotype: low cortisol across chronic fatigue syndrome, fibromyalgia and PTSD, associated with a triad of enhanced stress sensitivity, pain and fatigue. They propose it follows a prolonged period of hyperactivity — and, importantly, that it may itself be protective, limiting glucocorticoid-mediated damage. A downshift, not a failure.

The most robust cortisol-based health correlate is the flattened diurnal slope. Adam et al. (2017) meta-analysed 179 associations from 80 studies: flatter slopes predicted poorer health at an average r = 0.147, significant in 10 of 12 outcome subtypes and strongest for immune/inflammation outcomes (r = 0.288). Note what that magnitude means. r ≈ 0.15 is about 2% of variance. It is a real signal and a weak predictor simultaneously.

There is one dissociation that maps directly onto the depletion narrative. Chida & Steptoe (2009), across 147 eligible studies from 62 articles, found the cortisol awakening response increase (CARi) positively associated with job stress and general life stress, but negatively associated with fatigue, burnout and exhaustion. Ongoing demand raises the morning burst; exhaustion flattens it.

"Adrenal fatigue" does not exist

This needs stating plainly, because it is the single most commercially exploited claim in the field. Cadegiani & Kater (2016) screened 3,470 articles and included 58 studies (33 in healthy individuals, 25 in symptomatic patients). Their finding was "an almost systematic finding of conflicting results," with limitations including poor-quality fatigue assessment and "the use of an unsubstantiated methodology in terms of cortisol assessment (not endorsed by endocrinologists)." Their conclusion, verbatim: "there is no substantiation that 'adrenal fatigue' is an actual medical condition. Therefore, adrenal fatigue is still a myth." (An erratum exists at BMC Endocr Disord 2016;16:63.)

Keep three things distinct:

Concept	Status
Adrenal insufficiency (Addison's, secondary AI)	Real, dangerous, diagnosable, treatable
HPA dysregulation (flattened slope, blunted CAR, altered feedback)	Real, measurable, weakly predictive
"Adrenal fatigue" (adrenals exhausted by stress; diagnosed by salivary panels; treated with supplements)	Not a recognised diagnosis; refuted

The adrenal cortex does not run out.

Correspondingly, there is no biomarker for burnout. Rothe et al. (2020) reviewed 190 studies and concluded that while major depression shows "a general state of hypercortisolism and glucocorticoid resistance," for burnout there is neither "a sufficient amount of studies… to provide an evidence base" nor consistent construct measurement. Danhof-Pont, van Veen & Zitman (2011) had reached the same conclusion nine years earlier across 31 studies and 38 candidate biomarkers: "No potential biomarkers for burnout were found." Hair cortisol findings even reverse sign between cross-sectional (hyper-) and the best-controlled longitudinal (hypo-) designs.

Methodological warning. The cortisol awakening response literature has a genuine measurement-integrity problem. Expert consensus guidelines were issued in 2016; a formal audit comparing 2013–2015 with 2018–2020 publications found "little improvement in the implementation of central recommendations (especially objective time verification)," prompting Psychoneuroendocrinology to mandate a methodological checklist (Stalder et al., 2022). Treat pre-2016 CAR findings without electronic adherence monitoring as low-confidence.

1.3 Allostatic load: the right framework, imprecisely measured

McEwen & Stellar (1993) introduced allostatic load as "the cost of chronic exposure to fluctuating or heightened neural or neuroendocrine response resulting from repeated or chronic environmental challenge," arguing that homeostasis "has failed to help us understand the hidden toll of chronic stress."

McEwen (1998) elaborated four types, and the taxonomy remains clinically useful because the four have different remedies:

1. Repeated hits — frequent novel stressors, each triggering a full response.
2. Failure to habituate — repeated exposure to the same stressor without attenuation.
3. Failure to shut off — prolonged response after the stressor ends. This is the most directly depleting pattern, and the one that maps onto rumination and impaired recovery.
4. Inadequate response — under-response by one system triggering compensatory overactivity in another (blunted cortisol → unopposed inflammatory cytokines). This is the bridge to §1.8.

Seeman et al. (1997) operationalised the construct as a count of parameters in the highest-risk quartile across ten biomarkers, and Seeman et al. (2001) showed in 1,189 adults aged 70–79 over seven years that baseline allostatic load predicted mortality and functional decline independent of sociodemographics. The structural finding matters: four "primary mediators" (cortisol, norepinephrine, epinephrine, DHEA-S) and six "secondary mediators" (metabolic-syndrome components), with the composite outperforming either subset alone.

The framework's weakness is measurement. Parker et al. (2022) meta-analysed 17 studies: high allostatic load gave HR 1.22 (1.14–1.30) for all-cause mortality and HR 1.31 (1.10–1.57) for cardiovascular mortality — but with I² > 90% in most pooled estimates, reflecting the absence of any standardised operationalisation. "Allostatic load" names a construct measured differently in nearly every study. Notably, the consensus five-biomarker index (McCrory et al., 2023, N = 67,126) contains no HPA marker at all.

1.4 Autonomic load and recovery capacity

Thayer et al.'s (2012) neurovisceral integration model proposes that the default response to uncertainty is the threat response, and that vagally-mediated heart rate variability indexes how effectively top-down cortical appraisal inhibits it. Their meta-analysis identified the amygdala and ventromedial prefrontal cortex as the loci of cross-study association with HRV.

On this reading, low HRV is not merely low parasympathetic tone — it is impaired inhibitory capacity, meaning stress responses initiate more readily and terminate more slowly. That is McEwen's type-3 allostatic load, made measurable.

The prognostic weight is substantial. Jarczok et al. (2022) meta-analysed 37 samples and 38,008 participants: the lowest quartile of five-minute RMSSD against the rest gave HR 1.56 (1.32–1.85) for all-cause mortality. In 3,947 working adults, HRV measures were more strongly associated with global self-rated health than inflammatory markers, blood pressure or lipids (Jarczok et al., 2015).

1.5 The bioenergetic turn — the strongest version of the claim

This is where "stress depletes energy" stops being a metaphor and becomes a measurable quantity, and it is the most important development in the field in the last decade.

The framework. Picard & McEwen (2018) define mitochondrial allostatic load: chronic psychological stress induces mediators that cause structural and functional recalibration of mitochondria, which in turn influence gene expression, epigenetic modification and the rate of cellular ageing. Mitochondria are positioned as both targets and transducers of psychosocial experience.

Direction of causation established. Picard et al. (2015) mutated or deleted four different mitochondrial genes in mice and applied restraint stress. Each defect produced a distinct whole-body stress-response signature across HPA activation, catecholamine output, IL-6, circulating metabolites and hippocampal gene expression. Bioenergetic capacity shapes the stress response; it is not merely downstream of it.

Glucocorticoids act directly on mitochondria, with an inverted-U. Du et al. (2009) showed mitochondrial oxidation, membrane potential and calcium-holding capacity all respond to corticosterone dose-dependently: low doses neuroprotective, high or chronic doses potentiating excitotoxicity. Mechanistically, GR forms a complex with Bcl-2 and translocates into mitochondria; after three days, high — but not low — doses depleted mitochondrial GR and Bcl-2. This is the cellular signature of "the same hormone helps acutely and harms chronically."

The energetic cost, quantified. Bobba-Alves, Juster & Picard (2022) propose the Energetic Model of Allostatic Load: allostasis and stress-induced energy expenditure draw on a finite budget, and when the draw exceeds reserve it is taken from growth, maintenance and repair. Their collated human figures for acute stress-induced energy expenditure: mental arithmetic +9% to +42%; thirty minutes of extended cognitive stress ~+67%; high trait anxiety ~+14% higher expenditure.

The best single experiment. Bobba-Alves et al. (2023) profiled three unrelated primary human fibroblast lines across their entire lifespan under chronic glucocorticoid exposure. Cellular energy expenditure rose ~60%, with a metabolic shift from glycolysis to oxidative phosphorylation. This hypermetabolic state was linked to mtDNA instability, non-linear cytokine changes, and accelerated ageing on three independent measures — DNA methylation clocks, telomere shortening rate, and reduced replicative lifespan.

The decisive control is what makes this experiment matter: pharmacologically normalising OxPhos activity while further raising total energy expenditure worsened the ageing phenotype. It is total energy expenditure itself, not OxPhos dysfunction, that drives the damage. (Scope caveat: this is three fibroblast lines in vitro. "Cellular allostatic load" is the authors' framing, not an in vivo human finding.)

Human brain evidence. Trumpff et al. (2024) combined longitudinal antemortem psychosocial assessment with postmortem dorsolateral prefrontal cortex proteomics. Higher well-being tracked greater abundance of oxidative phosphorylation machinery; higher negative mood tracked lower OxPhos protein content. Combined psychosocial factors explained 18–25% of variance in complex I abundance. Single-nucleus RNA-seq revealed opposite-signed associations in glia versus neurons — meaning bulk-tissue studies systematically underestimate the true effects.

Biomarkers: not yet. Elevated whole-blood mtDNA copy number in depression has replicated, but the authors themselves note that whether it translates to mitochondrial function is unknown. Cell-free mtDNA is weaker: the most rigorous study is negative. Trumpff et al. (2025) ran the crossover design with a resting control condition that prior studies lacked, in 72 volunteers. The stressor reliably raised anxiety, heart rate, blood pressure, cortisol and norepinephrine — but IL-6 and plasma cf-mtDNA rose equally in the stress and control conditions, suggesting these markers reflect non-specific responses to the laboratory protocol (a blood draw) rather than to socio-evaluative stress.

1.6 The scale gap — the field's most important unresolved problem

Here is a tension the literature has not resolved, and which any serious research programme should target.

On one side: chronic glucocorticoid exposure raises cellular energy expenditure ~60% in vitro; acute mental stress raises whole-body VO₂ ~20% and glucose utilisation ~34%, abolished by propranolol (Seematter et al., 2000); chronic low-grade inflammation is estimated to raise systemic energy expenditure up to ~10% (Lacourt et al., 2018).

On the other side: trait stress and anxiety show no association with resting metabolic rate (Wilson et al., 2020), and prior-day stressors predicted lower post-meal resting energy expenditure (Kiecolt-Glaser et al., 2015).

Cellular hypermetabolism and whole-body energetics do not currently reconcile. The honest position is that the mechanistic story is compelling at the cellular level and unproven at the organismal one. Do not present the +60% figure as a human whole-body number — it is cultured fibroblasts.

1.7 Sleep — the competing explanation, not merely a mediator

This section carries more weight than its length suggests, because sleep debt is a sufficient cause of the phenomenology people attribute to stress.

Architecture. Kim & Dimsdale (2007) reviewed 63 polysomnographic studies: experimental stressors consistently decreased slow-wave sleep, REM and sleep efficiency and increased awakenings; naturalistic stressors gave far weaker and less consistent effects. The cleanest causal test (Ackermann et al., 2019) found acute stress nearly doubled sleep-onset latency (17.05 vs 9.03 min, p = 0.005) and reduced early slow-wave activity — but the effect vanished by ~30 minutes and whole-nap efficiency was unchanged. Acute stress disrupts sleep initiation more reliably than whole-night architecture.

A correction, carefully bounded. [contested] It is often claimed that high cortisol destroys deep sleep. The acute pharmacology runs the other way: acute cortisol administration in healthy volunteers increases slow-wave sleep and suppresses REM, probably via feedback inhibition of CRH (Friess et al., 1994: REM 19.9% → 12.2%, SWS 9.4% → 13.9%, both p < 0.05; reviewed in Steiger, 2002).

But this does not generalise to chronic hypercortisolism, and an earlier draft of this review overstated it. In Cushing's syndrome delta sleep is reduced — 5.8 ± 1.4% versus 14.0 ± 2.5% in controls, p < 0.01 (Friedman et al., 1994) — the inverse of the acute pattern. And in ACTH-independent Cushing's, where cortisol is high but ACTH and CRH are suppressed, patients still show poorer sleep continuity and shortened REM latency (Shipley et al., 1992), which defeats the attribution "it is CRH drive, not cortisol." The CRH evidence is itself dose-, route- and age-dependent: Vgontzas et al. (2001) found CRH increased wakefulness and suppressed SWS only in middle-aged men, not young men.

Correct statement: acute exogenous cortisol increases SWS in healthy volunteers; chronic endogenous hypercortisolism reduces it. The popular claim is wrong about acute pharmacology and roughly right about chronic states.

The bidirectional loop. Leproult et al. (1997): partial sleep loss raised evening cortisol +37%, total deprivation +45%, delaying the evening quiescent period by an hour or more. Spiegel, Leproult & Van Cauter (1999): four hours in bed for six nights raised evening cortisol and sympathetic activity and lowered glucose tolerance and thyrotropin. Downstream, the restriction literature shows >40% reduction in glucose tolerance and insulin-sensitivity reductions of 18–24%.

Metabolic recovery depends on early sleep. Roughly 70% of daily growth-hormone output occurs during early sleep, temporally locked to the first slow-wave episode (Van Cauter, Plat & Copinschi, 1998). Selective disruption of slow-wave activity raises CSF amyloid-β in proportion to the degree of suppression, while total sleep time does not correlate (Ju et al., 2017) — the most defensible human clearance finding, given that the broader glymphatic literature is now openly disputed. [contested: Xie et al. (2013) reported faster clearance during sleep; Miao et al. (2024) reported the opposite; Plá et al. (2025) rebutted on methodological grounds.]

The decisive finding. Van Dongen et al. (2003) restricted 48 adults to 4, 6 or 8 hours for 14 days. Chronic restriction to 4–6 hours produced deficits "equivalent to up to 2 nights of total sleep deprivation." And critically: subjective sleepiness saturated early and failed to distinguish the 4-hour from the 6-hour condition, while objective impairment kept accumulating.

This dissociation is the most likely mechanism by which chronically under-slept people misattribute their fatigue to stress. They cannot feel the difference between moderately and severely impaired. Any causal model should control for sleep duration first.

Insomnia is hyperarousal, not sleepiness. Twenty-four-hour VO₂ is elevated in insomnia (Bonnet & Arand, 1995), and over 70% of primary insomnia patients have a mean multiple sleep latency above ten minutes versus 47% of controls — they report exhaustion while denying sleepiness (Singareddy et al., 2010). Note that the HPA account of insomnia is not well supported: the WFSBP Task Force (2024) reports cortisol findings "inconsistent" and dexamethasone/CRH challenge negative.

1.8 Inflammation, sickness behaviour and the dopamine bridge

Glucocorticoid resistance is the mechanism by which cortisol can be high and inflammation uncontrolled. Cohen et al. (2012) supplied the disease endpoint via viral-challenge quarantine: chronic stress predicted developing a cold, OR 1.99 (1.04–3.08), with glucocorticoid receptor resistance predicting the local inflammatory response (IL-6 r = 0.314, TNF-α r = 0.283).

Population-level effects are small. Baumeister et al. (2016) pooled 25 studies on childhood trauma: CRP z = 0.10, IL-6 z = 0.08, TNF-α z = 0.23, with I² = 72.7% for CRP. Acute laboratory stress produces larger effects (IL-1β d = 0.66, IL-6 d = 0.35, TNF-α d = 0.28; Marsland et al., 2017). Elevated inflammation is present only in a subgroup of depressed patients (Miller & Raison, 2016).

Sickness behaviour is the mechanistic account of inflammation-driven fatigue, and it is a better model than "depletion." Dantzer et al. (2008) frame fever, anorexia, withdrawal and fatigue as a coordinated motivational state — not debilitation, but a reorganisation of priorities. Lacourt et al. (2018) give the energetic version: inflammation shifts immune cells to inefficient aerobic glycolysis, raising glucose demand. Human causal evidence exists: 0.8 ng/kg endotoxin produced depressed mood at 3–4 hours (ES = 0.66) with TNF-α correlating r = 0.40–0.75 (Reichenberg et al., 2001).

Inflammation → dopamine → motivational withdrawal. Felger & Treadway (2017) set out the pathway: inflammation-driven oxidative stress depletes tetrahydrobiopterin, the obligate cofactor for tyrosine hydroxylase, impairing dopamine synthesis; inflammation also reduces striatal dopamine packaging and release. In 48 unmedicated depressed patients, higher plasma CRP predicted reduced ventral striatum–vmPFC connectivity (R = −0.56), which correlated with anhedonia (R = −0.47; Felger et al., 2016).

This connects directly to Part II. Draper et al. (2018) found 2 ng/kg LPS selectively reduced acceptance of high-effort options with no effect on reward sensitivity. A placebo-controlled infliximab trial in depressed patients with CRP > 3 mg/L increased willingness to expend effort (Treadway et al., 2024). [contested: Lasselin et al. (2017) found LPS increased high-effort choices when win probability was high; Irwin & Eisenberger's commentary concludes the effect is context-dependent.]

1.9 Metabolic and endocrine consequences

Acute mobilisation is real and large. The Trier Social Stress Test reproduces the hormonal pattern: epinephrine +72%, norepinephrine +148%, ACTH +184%, cortisol +131% (Hitze et al., 2010). Glucocorticoids supply the substrate: methylprednisolone raised leucine oxidation by 62 ± 13% (Oehri et al., 1996) — muscle protein being converted to glucose.

Insulin resistance is dose-dependent. Dexamethasone 3 mg twice daily for 48 hours reduced clamp-measured insulin sensitivity to 45 ± 5% of baseline (Larsson & Ahrén, 1996). But prednisolone 10 mg had minimal metabolic impact while 15–20 mg reduced hepatic and muscle insulin sensitivity (Pofi et al., 2025). The extreme natural experiment: 32% of Cushing's patients have diabetes at diagnosis (ERCUSYN, n = 1,564).

Visceral fat — strong mechanism, weak human effect. [contested] Adipose-selective 11β-HSD1 overexpression produces visceral obesity in mice without elevated circulating glucocorticoids (Masuzaki et al., 2001), but human translation has disappointed. Across 146 cohorts (n = 34,342), the hair cortisol–BMI correlation is r = 0.10 — about 1% of variance (van der Valk et al., 2022). The canonical Epel et al. (2000) finding is a cross-sectional n = 59 study and should not be cited as causal.

Thyroid — weaker than usually claimed. The archetype is non-thyroidal illness syndrome: T3 falls and rT3 rises without compensatory TSH rise, a set-point change rather than a failure (Peeters et al., 2006). But evidence in chronic psychosocial stress specifically is much weaker; the best-controlled human experiment found the Trier test produced a rise in TSH with no change in T3 or T4 (Fischer et al., 2019) — the opposite direction to NTIS.

Gonadal axis — and a decisive result about what is really driving it. In functional hypothalamic amenorrhoea, 24-hour cortisol is elevated and LH pulse frequency reduced by 30%, reversibly. But Loucks, Verdun & Heath's (1998) dissociation experiment is decisive on aetiology: clamping energy availability at 10 versus 45 kcal/kg lean body mass per day reduced LH pulse frequency by 10%, whereas exercise stress at matched energy availability did not. Energy availability, not psychological stress, is the proximate driver.

That finding has a wider moral. Where a metabolic and a psychological explanation compete, the metabolic one has repeatedly won.

1.10 Individual differences: mechanism strong, moderator claims weak

FKBP5 is the best-supported candidate. FKBP51 is a co-chaperone that lowers GR ligand affinity and slows nuclear translocation; because GR activation induces FKBP5 transcription, it forms an ultra-short negative feedback loop — a molecular gain control on cortisol signalling. Klengel et al. (2013) showed childhood trauma produces allele-specific demethylation of FKBP5 intron 7 in risk-allele carriers only.

But calibrate against the base rate. Duncan & Keller (2011) analysed all 103 candidate gene-by-environment studies from 2000–2009: 96% of novel findings were significant versus 27% of replication attempts. Border et al. (2019), with N up to 443,264, found no support for any of 18 historical depression candidate genes. A precise caveat worth preserving: FKBP5 and NR3C1 are not among the 18 that Border et al. tested. That paper establishes the base rate for this class of finding; it does not directly falsify FKBP5. The honest position is that FKBP5 has considerably better molecular support than any classical candidate gene, and has never faced a Border-scale test.

Early-life adversity and HPA programming is genuinely contradictory. [contested] Three meta-analyses of overlapping literatures disagree: Bunea et al. (2017) found blunted reactivity, g = −0.39; Fogelman & Canli (2018) found no significant effect (g = −0.089); Bernard et al. (2017) found null overall but g = 0.24 in agency-referred samples versus g = 0.00 for self-reported maltreatment. Perrone et al. (2023), across 441 effect sizes from 104 studies, found exactly one significant overall effect: bedtime cortisol, r = 0.047. Measurement method and adversity ascertainment moderate results more strongly than adversity itself.

Sex differences are largely a free-fraction artefact. Men show higher salivary cortisol at Trier peak — but Kirschbaum et al. (1999) showed total plasma cortisol shows no group difference at all, consistent with estradiol-driven changes in corticosteroid-binding globulin. The popular claim that men respond to achievement stressors and women to social rejection traces to a single n = 50 study and should be treated as a hypothesis.

Age. Mean cortisol rises 20–50% between ages 20 and 80, with a progressively elevated nocturnal nadir. Otte et al. (2005) meta-analysed 45 studies: older adults showed greater cortisol response to challenge, d = 0.42, nearly threefold stronger in women (0.65) than men (0.24).

Heritability is measure-dependent to the point of uselessness — plasma cortisol 30–60%, hair cortisol 72%, evening cortisol 8%. The decisive molecular result is the CORNET GWAS (Bolton et al., 2014, N = 15,392): under 1% of plasma cortisol variance explained by common variation, and the single genome-wide locus was SERPINA6/SERPINA1 — corticosteroid-binding globulin, not a central HPA gene.

Part II — Function: what actually degrades, and how

2.1 The neural picture

The dominant mechanistic account is Arnsten's: under uncontrollable stress, high catecholamine release weakens dorsolateral prefrontal cortex while strengthening posterior sensory cortices, amygdala and striatum. Noradrenergic α1 and dopaminergic D1 signalling converge on feedforward Ca²⁺–PKC and cAMP–PKA cascades that open potassium channels on dendritic spines, collapsing the recurrent network connectivity that sustains working-memory representations (Arnsten, 2009, 2015). Her phrasing is that this "rapidly flips the brain from reflective to reflexive control of behavior."

The account earned unusual credibility because it produced working drugs: the α1-antagonist prazosin and the α2A-agonist guanfacine both moved into clinical use on this mechanistic basis.

At network scale the same story appears as resource reallocation. Acute stressors shift activity toward a salience network (vigilance, threat detection) at the expense of an executive control network, reversing in the aftermath (Hermans et al., 2014).

Structurally, chronic stress produces opposing remodelling across regions: hippocampal CA3 dendritic atrophy and suppressed neurogenesis, medial prefrontal dendritic retraction, and amygdala changes running the other way (McEwen, Nasca & Gray, 2016). A useful corrective comes from Kim & Kim (2023): glucocorticoid levels alone are not sufficient to explain the effects — amygdala inactivation blocks stress-induced hippocampal LTP deficits, implicating neural activity rather than circulating hormone as the proximate cause.

2.2 Executive function: real, small, and fractionated

The quantitative anchor is Shields, Sazma & Yonelinas (2016). The effects are real and modest, and they do not move together:

Domain	Effect (Hedges' g)	95% CI	Note
Working memory	−0.197	[−0.330, −0.064]	34 studies, N = 1,353
Cognitive flexibility / set-shifting	−0.300	[−0.577, −0.023]	Only 6 studies — thin
Inhibition, overall	−0.076	[−0.243, 0.092]	Not significant
→ Cognitive inhibition (distractor filtering)	−0.208	[−0.379, −0.035]	Impaired
→ Response inhibition (motor stopping)	+0.296	[0.018, 0.573]	Post-hoc moderator, ~5 df

An important correction. It is frequently reported that "acute stress enhances response inhibition." That is not a primary result. The overall inhibition effect was null; the +0.296 emerged only from a post-hoc moderator analysis on roughly five degrees of freedom, with a confidence interval whose lower bound is 0.018 — almost touching zero. What the data support is: no overall effect on inhibition, with divergent subgroup effects. The pattern is still informative — impaired filtering of irrelevant material alongside preserved or enhanced motor stopping is the signature of a system reweighted toward reactive threat monitoring — but it should not be stated as an enhancement finding.

Two moderators matter practically. Impairment scales with load: high-working-memory-load tasks g = −0.303 versus −0.049 for lower-load tasks. And it grows with delay since stressor onset (B = −0.006, p = .044), consistent with a slower glucocorticoid-mediated phase rather than the immediate catecholamine surge. The practical implication is that the worst moment for demanding cognitive work is not during the stressor but twenty to forty minutes afterwards.

Cortisol is a poor proxy for cognitive impact. Cortisol reactivity did not moderate the stress effect, and exogenous cortisol administration produced no comparable impairment (stress g = −0.151, p = .005; cortisol administration g = 0.030, p = .495; difference p = .009). This is a recurring theme: the hormone everyone measures is not the variable that predicts the outcome anyone cares about.

Acute versus chronic. The inverted-U is asserted more often than demonstrated. The cleanest same-conditions demonstration is in rodents (Salehi, Cordero & Sandi, 2010) — and there, the curve's peak shifts with task difficulty and appeared only in high-anxiety animals; other temperament profiles showed monotonic relationships. Treat "optimal stress" as person- and task-specific rather than a universal law.

Chronic stress shows no such upside. In clinical burnout, impairment is broad (Gavelin et al., 2022; 17 studies, 730 patients vs 649 controls):

Domain	Hedges' g	95% CI
Verbal fluency	−0.53	[−1.04, −0.03]
Attention / processing speed	−0.43	[−0.57, −0.29]
Executive function	−0.39	[−0.55, −0.23]
Episodic memory	−0.36	[−0.57, −0.15]
Short-term / working memory	−0.36	[−0.52, −0.20]
Crystallised ability	no difference	—
Visuospatial ability	no difference	—

The fluid/crystallised dissociation has real diagnostic value: vocabulary and general knowledge are preserved while processing speed and executive control degrade. This is why exhausted people can hold a conversation competently and then fail to plan their week.

2.3 Memory: the retrieval dissociation

This is the clearest triple dissociation in the field (Shields, Sazma, McCullough & Yonelinas, 2017):

• Encoding: g+ = −0.109 [−0.282, 0.065] — not significant overall. Direction depends on delay; benefits appear only at near-zero delays with stressor-relevant material.
• Post-encoding consolidation: g+ = +0.206 [0.013, 0.399] — stress enhances consolidation, but only when context is held constant (g+ = 0.380) versus changed context (g+ = −0.196, ns).
• Retrieval: g+ = −0.215 [−0.346, −0.085]. Larger for emotional material: negative g+ = −0.303, positive −0.385, neutral −0.136.

This is the mechanism of "blanking out." The information is encoded and consolidated — it is in there — but retrieval access is selectively degraded, and more so for emotionally charged content, which is precisely the content one is trying to recall in a high-stakes moment. It also explains a common and demoralising experience: recalling the material perfectly twenty minutes after the interview or the exam.

Timing is decisive. Immediate post-stress testing shows no impairment; 10–45-minute delays, capturing the cortisol peak, produce it (Klier & Buratto, 2020). Sex moderates: retrieval impairment was g+ = −0.294 in non-users of hormonal contraception versus +0.101 in users. And once again, cortisol magnitude was unrelated to the memory effects.

2.4 What does not hold up

A rigorous review has to be as clear about the failures as the findings, and this subfield has had several large ones.

Ego depletion — effectively falsified in its strong form.

• Hagger et al. (2016): 23 labs, N = 2,141, d = 0.04 [−0.07, 0.15].
• Vohs et al. (2021): 36 sites, N = 2,463 analysed, preregistered d = 0.06 [−0.02, 0.14], Bayes factor ≈ 4:1 favouring the null.

The authors' own conclusion: depletion "is not as reliable or robust as previously assumed." Do not build a model of stress-related incapacity on willpower depletion. The phenomena it was invented to explain are better handled by the effort-cost framework in §2.5.

Stress → temporal discounting — null. The frequently asserted "stress makes you impulsive" claim does not survive current meta-analysis: Forbes et al. (2024) pooled 11 studies, N = 1,097, and found SMD = −0.18 [−0.57, 0.20], p = .32 — null across sex, stressor type and incentivisation.

Stress → habit shift — failing exact replication. Schwabe & Wolf's influential claim that stress biases behaviour from goal-directed to habitual control is in serious trouble. Smeets et al. (2023) ran two preregistered exact replications in which the control group failed to show goal-directed control at all, invalidating the critical test. Zwosta et al. (2025) tested 129 participants across two stressors and found no effect, with Bayes factors around 5:1 favouring the null, noting the original effects were "restricted to the very first trials."

Scarcity and cognitive bandwidth — partially replicated, magnitude unresolved. [contested] Mani et al. (2013) found Indian sugarcane farmers scored worse on Raven's matrices pre- versus post-harvest (4.35 vs 5.45 items, p < .001), a result popularised as costing "13 IQ points." The strongest challenge is Carvalho, Meier & Wang (2016): N = 3,821 low-income participants randomised to before- versus after-payday testing, null across Flanker, Stroop, working memory and cognitive reflection. To reconcile, the harvest shock would have to be 11–20× larger than the payday variation. O'Donnell et al. (2021) replicated 20 scarcity studies and found "considerable variability… some strong successes and other undeniable failures." That scarcity imposes some cognitive load is plausible; the "13 IQ points" figure should not be used as an established quantity.

Stress → risky decision-making — unresolved, with multiple well-powered nulls.

Stress → emotion-regulation failure — inconsistent. [contested] The canonical citation is Raio et al. (2013), "Cognitive emotion regulation fails the stress test." But Shermohammed et al. (2017) found reappraisal intact under a verified stress induction (self-reported stress, cortisol and heart rate all rose; regulation did not budge), and at least one study reports acute stress improving regulation in men. The neural mechanism for a vicious cycle is well specified; the human behavioural demonstration is inconsistent. Regulation may be more robust than the theory predicts.

2.5 The reframe that matters: effort as a price, not a fuel gauge

If depletion is dead, what explains "I just can't face it"?

The productive move is to treat cognitive effort as a cost entering a value computation rather than a substance being consumed. People reliably discount reward as a function of cognitive demand, choosing smaller-reward/lower-demand options over larger-reward/higher-demand ones (Westbrook & Braver, 2016). Their framing is the crucial one: "allocation of working memory is a motivated process." Capacity and the decision to deploy it are separate variables.

Dopamine's role is activational, not hedonic. Salamone & Correa's programme reframes mesolimbic dopamine as regulating the willingness to exert effort to overcome obstacles, not the pleasure of the goal. Depression and schizophrenia show "reduced selection of high-effort activities" — motivational, not capacity, impairment.

The phenomenology of effort may be a computed signal, not a readout of a store. Kurzban et al. (2013) argue that the feeling of effort is the computed opportunity cost of continuing the current task — an adaptive prompt to reallocate, not evidence of depletion. Müller & Apps (2019) formalise "motivational fatigue": exertion inflates subsequent effort costs, biasing choice away from demanding options.

And it now has physical grounding. Wiehler et al. (2022) is the study that turned this from a modelling framework into a biological claim. Across a simulated workday, high-demand cognitive work (versus low-demand) produced elevated glutamate concentration and glutamate/glutamine diffusion in lateral prefrontal cortex — but not in a visual-cortex control region. Behaviourally, the high-demand group showed reduced pupil dilation during decisions and a modelled low-cost bias toward short-delay, low-effort options. The interpretation: glutamate accumulation triggers a regulatory mechanism that makes lateral PFC activation more costly, not impossible. This is synthesised in Pessiglione et al. (2025) as the MetaMotiF framework — fatigue originating in metabolic changes in control regions, which raise the price of controlled thought.

Why the distinction is not academic. A person who is "too tired to work" may have near-intact capacity but a raised effort price. The two states call for different responses:

	Capacity loss	Raised effort price
What's wrong	The machinery is degraded	The machinery works but costs more to run
What helps	Reduce load; offload cognition; simplify tasks; protect the hardest work for the best hours	Change the cost–benefit structure; restore genuine recovery periods; reduce competing demands; in clinical cases, target motivational systems
What doesn't	Exhortation, incentives	Exhortation, incentives — and simply resting without changing what made effort expensive

Conflating them produces both bad advice and unwarranted moral judgement. "You just need to push through" assumes capacity loss and treats a price signal as weakness.

2.6 Subjective and objective impairment come apart

This dissociation is routine, consequential, and almost always ignored in practice.

Wiehler et al. (2022) found objective fatigue markers only in the high-demand group, while subjective fatigue ratings rose in both — self-report tracked time-on-task, not the underlying signature. Shermohammed et al. (2017) produced a textbook instance: a successful stress induction on every physiological measure, with no measurable regulation deficit. People felt stressed; their performance held.

The reverse also occurs. Österberg et al. (2014) found persistent attention deficits in recovered exhaustion-disorder patients with no relationship between test performance and extent of work resumption, and little correspondence between subjective and objective cognitive impairment. Van Dongen's sleep-restriction dissociation (§1.7) is the same phenomenon in a different domain.

Practical implication: self-reported impairment and tested impairment are separate constructs and should be measured separately. Neither is a valid proxy for the other. For research design this means never using a fatigue questionnaire as a stand-in for a cognitive outcome. For clinical and personal purposes it means the felt sense of "I'm fine" and "I'm useless" are both unreliable, in opposite directions and at different times.

2.7 What this costs in the world

Clinical safety. Hodkinson et al. (2022) pooled 170 studies and 239,246 physicians. Burnout roughly doubled patient safety incidents (OR 2.04 [1.69–2.45]), low professionalism (OR 2.33) and patient dissatisfaction (OR 2.22). Career effects were larger still: job dissatisfaction OR 3.79, career regret OR 3.49, turnover intention OR 3.10. Highest risk in emergency and intensive care, ages 31–50, and trainees.

Industrial safety. Nahrgang, Morgeson & Hofmann (2011): burnout → adverse events ρc = .29; → unsafe behaviour ρc = .32; → accidents/injuries ρc = .13. Engagement was protective (→ adverse events ρc = −.32).

Driving. The most striking real-world figure in this review comes from naturalistic data: 3,542 drivers, 35+ million miles, 905 crashes (Dingus et al., 2016). Driving while observably emotional — anger, sadness, crying, agitation — carried OR 9.8 (5.0–19.0), higher than drowsiness (3.4), handheld phone use (3.6) or distraction generally (2.0). Its prevalence was low (0.22% of baseline driving), so the population attributable risk is modest, but the per-episode multiplier is extreme. This is the least-discussed acute risk of emotional distress.

Job performance is affected less than you would guess. Corbeanu et al. (2023): exhaustion → performance r = −.17 (N = 18,019); depersonalisation r = −.16; inefficacy r = −.23. Small-to-moderate — and note the gap between these modest correlations and the much larger safety-incident odds ratios. The plausible reading is that stress degrades peak reliability and error rates more than average throughput. People keep producing; they just produce worse tails.

Academic. Across 177 studies and 906,311 participants, math anxiety ↔ performance r = −.30, test anxiety ↔ performance r = −.23 (Caviola et al., 2022). Notably, working-memory mediation was "negligible" — which cuts against the simple cognitive-load story.

Relationships. Neff & Karney (2009) found that external stress increases reactivity: daily marital satisfaction becomes more tightly coupled to specific daily events. Stress does not merely lower mood; it removes the buffering that normally stops a bad day becoming a verdict on the relationship. Individual differences in self-esteem and attachment did not moderate this — it is not a resilience trait.

Economic. Martinez et al. (2025) modelled annual burnout cost per employee at $3,999 (nonmanagerial hourly) to $20,683 (executives), or $5.04 million per year for a 1,000-person organisation, with 801.7 QALYs lost annually. Presenteeism dominates absenteeism by 5–10× (Evans-Lacko & Knapp, 2016).

Part III — The chronic trajectory: how acute stress becomes chronic dysfunction

3.1 What burnout is, officially and unofficially

The construct. Freudenberger coined the term in 1974; Maslach & Jackson (1981) operationalised it as three dimensions — emotional exhaustion, depersonalisation/cynicism, and reduced personal accomplishment. The Maslach Burnout Inventory dominates the literature: 85.7% of studies in the largest physician review used it.

Two alternatives matter. The OLBI reduces it to exhaustion and disengagement. The Burnout Assessment Tool (Schaufeli, Desart & De Witte, 2020) adds two dimensions the MBI omits — impaired emotional control and impaired cognitive control — which is significant, because those are precisely the deficits Part II documents. The BAT was also designed to permit a single composite score, which the MBI forbids and which European welfare systems need for individual-level assessment.

ICD-11. The WHO position, verbatim:

"Burn-out is included in the 11th Revision of the International Classification of Diseases (ICD-11) as an occupational phenomenon. It is not classified as a medical condition."

It sits in the chapter "Factors influencing health status or contact with health services," and is defined as "a syndrome conceptualized as resulting from chronic workplace stress that has not been successfully managed," characterised by "feelings of energy depletion or exhaustion; increased mental distance from one's job…; and reduced professional efficacy." And critically: "Burn-out refers specifically to phenomena in the occupational context and should not be applied to describe experiences in other areas of life."

Two things routinely dropped from summaries. First, the WHO news release does not itself state the code QD85 — that comes from the ICD-11 browser. Second, and more substantively: "Burn-out was also included in ICD-10, in the same category as in ICD-11, but the definition is now more detailed." The 2019 announcement was widely reported as burnout being newly recognised as a condition. It was neither new nor recognised as a condition.

Where it is a clinical diagnosis. Sweden's exhaustion disorder (utmattningssyndrom, ICD-10 F43.8A; criteria published 2003, code added 2005) requires physical and mental exhaustion for ≥2 weeks following identifiable stressors present ≥6 months, markedly reduced mental energy, and ≥4 of six symptoms including memory complaints, reduced tolerance of demands, emotional instability and sleep disturbance. Its sharpest structural difference from ICD-11 is criterion F, which makes exhaustion disorder explicitly co-morbid with depression, dysthymia or GAD rather than excluded by them. Where WHO treats burnout and mood disorders as mutually exclusive, Sweden treats exhaustion as stackable. Exhaustion disorder also admits non-work stressors; ICD-11 burnout does not.

The Dutch approach is architecturally opposite again: burnout (ICPC Z29.01) is a severity/chronicity subtype of overspanning — present for over six months with fatigue and exhaustion strongly prominent.

Why this bureaucratic detail matters: it determines who gets paid. Sweden's F43.8A carries sick-leave entitlement and clinical guidance (full or partial leave up to six months acutely, longer with persisting cognitive difficulty; direct return to full-time work is described as "oftast kontraproduktivt" — usually counterproductive). ICD-11 QD85 confers none of that. The clearest empirical consequence is Dutch: in the same national survey in the same year, 20.1% of employees reported burnout complaints while only 1.6% had a physician-diagnosed case — a twelvefold gap generated entirely by case definition.

3.2 The depression problem

This is the most serious unresolved challenge to burnout's status as a distinct construct, and it is largely unanswered on its own terms.

Bianchi, Verkuilen, Schonfeld and colleagues (2021) pooled 14 samples, N = 12,417, across six countries. The disattenuated exhaustion–depression correlation was r = .80, with bifactor explained-common-variance indices of .67–.87 (median .82) indicating essential unidimensionality, and exhaustion and depression items loading similarly on the general factor. The published conclusion is that burnout "problematically overlaps with depression."

(A note on a widely circulated quotation: the line "does not present the unity expected of a distinct syndrome" is frequently attributed to this paper. It is not in it. The Bianchi group's phrase "burnout showed no syndromal unity" comes from Verkuilen et al., 2021, in Assessment 28(6). The confidence interval [.75, .84] is also not in the CPS abstract and could not be verified here.)

Longitudinally, Bianchi et al. (2015) found in 627 French teachers over 21 months that burnout at T1 did not predict depressive symptoms at T2 once baseline depression was controlled — baseline depression accounted for roughly 88% of the association.

The counter-position exists but is weaker. Koutsimani et al. (2019) report r = .520 and argue for distinct constructs. Much of the gap is a psychometrics argument rather than a data dispute: Bianchi disattenuates for measurement error and Koutsimani does not. Ahola & Hakanen (2007) found a reciprocal relationship over three years in 2,555 dentists (burnout→depression OR 2.6; depression→burnout OR 2.2).

Exhaustion disorder fares no better on discriminant validity. Sennerstam et al. (2025) compared exhaustion disorder (n = 352), major depression (n = 99) and adjustment disorder (n = 302) in primary care. Exhaustion disorder and depression differed on only two of nine self-report symptom scales. It did differ systematically from adjustment disorder.

The ME/CFS boundary. Among 151 fatigued employees on sick leave, 43.7% met CFS research criteria and 50.3% met burnout criteria — heavily overlapping populations, with the discriminating variable being causal attribution: somatic for the CFS-like, psychological for the burnout cases (Huibers et al., 2003). In clinical practice the operative discriminator is post-exertional malaise — delayed, disproportionate worsening after exertion that does not remit with rest — which is a required feature of ME/CFS and appears in no burnout or exhaustion-disorder criterion set. [unverified: no peer-reviewed head-to-head study demonstrating PEM's discriminative performance between burnout and ME/CFS was located. This reflects criteria structure, not a validated differential-diagnosis algorithm.]

What this means practically. If you are exhausted and not functioning, the probability that a competent assessment finds treatable depression is high, and the probability that "burnout" as a label adds anything actionable beyond that is low. The label's main function in most systems is social — it explains the state without pathologising the person. That is a real benefit. It is not a diagnosis.

3.3 There are no stages

The twelve-stage model of burnout (Freudenberger & North, 1985) circulates widely in workplace training, HR materials and popular articles. It has no empirical validation whatsoever.

Systematic phrase searches of Europe PMC (~40M records) and PubMed, with validated positive controls, return:

Search phrase	Hits
"Maslach Burnout Inventory" (positive control)	10,079
"Freudenberger" (positive control)	1,354
"stages of burnout"	112
"12 stages of burnout"	0
"twelve stages of burnout"	0
"12 phases of burnout"	0
"Freudenberger and North"	5 — none an empirical test

The model has no instrument, no staging criteria, no inter-rater reliability data and no prospective test. Even sympathetic reviews concede that "the understanding of the precise development of burnout and how many stages are included are still not unified… There are also different versions of stage order." Golembiewski's competing eight-phase model assigns phases by median-splitting cross-sectional MBI scores rather than observing progression.

What the longitudinal evidence actually shows is stability at heterogeneous levels, not sequence:

• Dunford et al. (2012), N = 2,089, five measurements over two years: burnout "relatively stable for organizational insiders but slightly dynamic for newcomers and internal job changers."
• Mäkikangas et al. (2012), growth mixture modelling, N = 433: four exhaustion classes and four cynicism classes, with exhaustion and vigour behaving as independent constructs — directly contrary to a sequential energy-depletion model.
• Hätinen et al. (2009): three trajectories — low burnout, high burnout/benefited, high burnout/not benefited. Classes differ by level and treatment response, never by stage position.
• Mäkikangas & Kinnunen (2016), systematic review of 24 person-oriented studies: heterogeneity of levels, no evidence of a shared sequence.

Why it matters that this is folk psychology. Stage models create a false prognostic promise ("you're at stage 7, here's what's next") and a false reassurance ("I'm only at stage 3"). The real picture — high stability, heterogeneous levels, differential treatment response — implies something less tidy and more useful: your current level is a poor guide to your trajectory, and whether you respond to intervention is a largely separate question from how bad things currently are.

3.4 Mechanisms of the transition — which models earn their keep

Model	Best current evidence	Verdict
Effort–Recovery (Meijman & Mulder, 1998)	Detachment↔exhaustion r̄ = −0.36 [−0.42, −0.30], k = 23, N = 7,007; detachment↔fatigue ρ = −.39	Strongest and most consistent effect sizes of any model here
Allostatic load (McEwen)	AL index → all-cause mortality HR 1.22 [1.14, 1.30], but I² > 90%	Predicts — but not via the mechanism it names
Effort–Reward Imbalance (Siegrist, 1996)	ERI → CHD HR 1.16 [1.00, 1.35], N = 90,164; ERI → depression RR 1.49 [1.23, 1.80]	Effect carried entirely by the reward limb (reward 1.18, effort 0.99)
Demand–Control(–Support) (Karasek, 1979)	Job strain → CHD HR 1.23 [1.10, 1.37]	Main effects yes; the interaction term does not replicate
Job Demands–Resources (Demerouti et al., 2001)	Longitudinal demands(T1)→burnout(T2) β ≈ .10	Founding cross-sectional β = .91 versus longitudinal β ≈ .10
Recovery paradox (Sonnentag, 2018)	Job demands → lower detachment r̄ = −0.25, k = 60, N = 28,507	First half supported; the paradoxical half untested meta-analytically

Three findings deserve emphasis.

The biological mechanism is missing. Two systematic syntheses nine years apart found no reliable HPA signature of burnout (§1.2). The effort–reward imbalance↔HPA activity correlation is r = 0.05 across 14 studies. The models that predict health outcomes do not do so through the physiological pathway they were built to describe.

The reverse causal path is larger than the forward one. Guthier, Dormann & Voelkle (2020), k = 48, N = 26,319: "The stressor-effect is small, whereas the strain-effect is larger." Burnout predicts subsequently reported stressors better than stressors predict burnout. There are two readings — burnout genuinely degrades the work environment (through withdrawal, conflict, reduced performance), or exhausted people report their environment more negatively. Both are probably true, and both undermine the simple demands-cause-burnout picture.

The interaction terms fail. Madsen et al. (2017), N = 120,221: "There was no statistical interaction between demands and control." Burr et al. (2021): "None of the studies found statistical interaction." Demerouti et al. (2001) explicitly declined to test their own buffer hypothesis, citing exactly this weakness — it was grafted on in 2005 against the founding paper's own reading. Schaufeli & Taris (2014) concede the JD-R is "a descriptive model… rather than an explanatory model" whose core claims follow "by definition."

Which model predicts best? A meta-review of 72 reviews (Niedhammer et al., 2021) found that for coronary heart disease "the magnitude of the association was similar for job strain, long working hours, and effort-reward imbalance." There is no evidential basis for ranking them. For burnout specifically, the same meta-review found only "two literature reviews, of low or moderate quality… only one was based on prospective design" — burnout is the thinnest-evidenced outcome in the entire meta-review.

The one model that consistently earns its keep is the least theoretically ambitious: effort–recovery. Whether you detach from work predicts exhaustion and fatigue more reliably than any structural model of the job.

3.5 Prevalence: a measurement catastrophe

Rotenstein et al. (2018) is the definitive demonstration. 182 studies, 109,628 individuals, 45 countries. Findings:

• Overall burnout prevalence: 0% to 80.5%
• Emotional exhaustion: 0–86.2%; depersonalisation: 0–89.9%
• At least 142 unique definitions of meeting burnout criteria, including ≥47 distinct definitions among MBI-based studies alone
• Meta-analytic pooling was planned and abandoned: "variation in study designs and burnout ascertainment methods, as well as statistical heterogeneity, made quantitative pooling inappropriate"

Any single pooled burnout prevalence figure is making a choice that Rotenstein explicitly refused to make.

Instrument choice alone moves estimates twofold. Psychiatrists: MBI 25.9% versus CBI 50.3% in the same meta-analysis. Teachers: CBI 64%, MBI 57%, BAT 29%. Switzerland, one country, one review: clinical/severe burnout 4% versus overall burnout 18%. Nurses: three peer-reviewed meta-analyses span 11.23% to 59.5%.

Physician burnout is non-monotonic, not steadily worsening — a fact obscured by almost all commentary. Serial cross-sectional surveys, percentage with ≥1 burnout symptom: 2011 45.5% → 2014 54.4% → 2017 43.9% → 2020 38.2% → 2021 62.8% → 2023 45.2% (Shanafelt et al., 2025). The 2023 level is statistically indistinguishable from 2011. The pandemic spike was real, large and largely resolved.

European working-population trends are gently upward but confounded by methodological breaks: Netherlands 14.4% (2014) → 20.7% (2025), with a survey break in 2022; Finland 6% (2019) → 8–10% (2023–25); Sweden's administrative F43 caseload +25% from 2019 to 2024, 79% women. Eurofound's (2018) summary remains the honest one: severe burnout around 2–3%, moderate forms 15–25%, and "it is difficult to draw a general picture, as the findings of the different studies are not comparable."

On the famous Gallup figures. The widely cited "76% of employees experience burnout" is from a 2020 Gallup report, is US-only, from 2019 fieldwork, uses a single unvalidated item, and the 76% comprises 28% "very often/always" plus 48% "sometimes." It is not a prevalence estimate in any clinical sense. Gallup's separate global engagement series (23% → 21% → 20% engaged, 2022–2025; daily stress 40%) revises its own back-series between editions and reports no confidence intervals.

3.6 Long-term health consequences: real, but smaller than advertised

Exposure → outcome	Estimate	Note
Job strain → CHD	HR 1.23 [1.10, 1.37] → 1.17 adjusted for SES	N = 197,473; PAR only 3.4%
≥55 h/week → stroke	RR 1.33 [1.11, 1.61]	N = 603,838; p-trend < 0.0001
≥55 h/week → CHD	RR 1.13 [1.02, 1.26]	Same
Burnout → CVD (pooled)	OR 1.21 [1.03, 1.39]	CHD subgroup 1.79 [0.79, 2.79] — not significant
Job strain → type 2 diabetes	1.26 [1.16, 1.37] → 1.12 [0.99, 1.26] fully adjusted	Null on full adjustment
Incident metabolic syndrome	OR 1.53 [0.82, 2.87]	Null; only hypertension (1.63) significant
Burnout → all-cause mortality	+35% per unit [1.07, 1.71] — but only under age 45	"Not related to mortality among the older employees"
Burnout → chronic work disability (severe)	HR 3.8 → 1.57 [1.09, 2.26] adjusted; mild burnout null
Long working hours, global burden (2016)	745,194 deaths, 23.3M DALYs	488M people worked ≥55h/wk

Read the third column. The IPD-Work consortium's own conclusion on job strain is that tackling it "would have a much smaller effect than would tackling of standard risk factors, such as smoking" — population attributable risk 3.4%. The consortium also documented its own publication bias: published estimates 1.43 versus unpublished 1.16.

Dementia: the evidence does not support a work-stress pathway. [contested, but the contest is one-sided] Sindi et al. (2017) found an association at first follow-up that disappeared over extended follow-up. Crowe et al. (2007) found job dissatisfaction and high demands "not associated with dementia risk." Nabe-Nielsen et al. (2019) found "no indications of a higher risk" at ≥45 h/week. Structurally, the "passive job" (low demand, low control) repeatedly rivals or exceeds "high strain" as a predictor, and high cognitive stimulation at work is protective (HR 0.82 [0.68, 0.98]; Kivimäki et al., 2021). The signal points to cognitive under-stimulation, not stress.

3.7 Recovery: the long tail nobody warns you about

This is, for anyone personally affected, the most important section in the review.

Symptom domains diverge sharply. In 232 Swedish clinical exhaustion-disorder patients (Glise et al., 2012), clinically significant burnout fell 93% → 66% (3 mo) → 52% (6 mo) → 38% (12 mo) → 33% at 18 months, while anxiety fell 65% → 11% and depression 34% → 6%.

Depression and anxiety resolve on a months timescale. Exhaustion does not.

The seven-year data are the headline (Glise, Wiegner & Jonsdottir, 2020; N = 217):

• Only 16% reported full recovery
• 46% still reported extreme fatigue
• 73% reported decreased stress tolerance
• 43% reported memory problems
• 31% were still clinically judged to have stress-related exhaustion
• But 87% were not on sick leave (3% full-time, 6% part-time, 4% disability pension)

At ten years: 83% working or studying, 31.5% still self-reporting exhaustion disorder, and 73% had changed workplace — against 25.5% over five years in the general Swedish population. "Symptoms of burnout, anxiety, and depression remained stable from the 1- to the 10-year follow-up."

The pattern that emerges is functional recovery without symptomatic recovery. People go back to work, and often to different work, while continuing to carry reduced stress tolerance and fatigue. That is a substantially less optimistic picture than "burnout, then rest, then recovery," and it is worth knowing in advance — both because it sets realistic expectations and because it argues strongly for not reaching the clinical threshold in the first place.

Residual cognitive deficits are real but modest (see §2.2 for the effect sizes) and persist after clinical recovery. Österberg et al. (2014) found "persistent signs of a minor attention deficit, despite considerable general recovery and return to work" — with, importantly, no relationship between test performance and extent of work resumption.

Relapse. 27.9% recurrent sickness absence within 12 months of return to work, with anxiety/depression severity at return predicting recurrence (Deprez et al., 2026).

And the interventions largely do not work. This is uncomfortable but well documented. Perski et al. (2017), 8 trials: full return-to-work OR 1.33 [0.59, 2.98] — not significant, with no significant effects on exhaustion, depression or anxiety. Salomonsson et al. (2018), 45 RCTs: sick leave g = 0.15, symptoms g = 0.21. Two well-conducted trials found interventions increased sickness absence: Noordik et al. (2013) found exposure-based return-to-work prolonged time to full return (209 vs 153 days; "We recommend that occupational physicians do not apply RTW-E"), and Finnes et al. (2019) found an ACT-plus-workplace intervention generated more sickness absence than treatment as usual. Franke Föyen et al. (2025), n = 300: "no specific effect of CBT on sick leave or cognitive functioning" — cognition improved d = −0.72 regardless of arm.

3.8 Who tips — and the single most striking moderator finding

Perfectionism is the best-verified disposition: perfectionistic concerns → total burnout r+ = .41 [.36, .45] across 43 studies and N = 9,838. Interestingly, perfectionistic strivings are protective in sport and education but maladaptive specifically at work (+.11 in work versus −.15 in education). Self-efficacy → burnout r̄ = −.33.

Work conditions carry the strongest prospective evidence. Aronsson et al. (2017), restricted to prospective and case-control designs and GRADE-rated, found moderately strong evidence that job control reduces and low workplace support increases emotional exhaustion, with limited evidence for workplace justice, demands, workload, low reward and job insecurity.

Job insecurity → CHD RR 1.32 age-adjusted, attenuating to 1.19 fully adjusted; → depressive symptoms OR 1.29, modestly higher than unemployment itself (1.19). There is no meta-analysis of financial strain or precarious employment with burnout as an outcome — a genuine gap.

Work–family conflict → burnout β = 0.11 across 52 longitudinal studies and 112,714 participants, adjusted for baseline mental health, with no significant moderation by gender. And contrary to lay assumption, across >350 samples and N > 250,000, "men and women generally do not differ on their reports of work–family conflict."

Discrimination and mistreatment produce the standout finding in this entire section. Hu et al. (2019) surveyed 7,409 US surgical residents — 99.3% of those eligible. Gender discrimination affected 65.1% of women. Mistreatment at least a few times a month gave burnout OR 2.94 [2.58, 3.36] and suicidal thoughts OR 3.07 [2.25, 4.19].

The critical result: the gender gap in burnout (OR 1.33 [1.20, 1.48]) disappeared entirely after adjusting for mistreatment (OR 0.90 [0.80, 1.00]). Independently replicated by Dyrbye et al. (2022) in 6,512 physicians, with a clean dose-response (OR 1.27 → 1.70 → 2.20) and "no difference in the odds of burnout by gender after controlling for experiencing mistreatment and discrimination."

The women were not less resilient. They were being treated worse. This is the clearest example in the literature of an apparent individual-difference finding dissolving into an environmental one — and a warning about how many other "vulnerability" findings might do the same. (Caveat: cross-sectional, so associational. The intersectional gradient is also documented: female + non-white + LGB versus male + white + heterosexual medical students, adjusted mean difference in exhaustion 1.96 [1.47, 2.44].)

Sleep may matter more for failure to recover than for onset. Jansson-Fröjmark & Lindblom (2010), N = 1,258, found neither insomnia nor burnout predicted the other's incidence — while insomnia strongly predicted the persistence of emotional exhaustion (OR 3.02).

The individual-versus-organisational question is design-driven, not settled. [contested] Bianchi et al. (2021), across three teacher samples, found work factors explained 26–32% of variance and individual/non-work factors 68–74%, with neuroticism alone accounting for 28–34% — roughly equal to all work factors combined. But this is cross-sectional, teachers only, with a single dispositional trait competing against many work variables and shared method variance with self-reported burnout. Aronsson et al. (2017) — prospective and GRADE-rated — reaches the opposite conclusion. The two literatures differ in design, not merely in conclusion, and a live exchange runs in Work & Stress 39(2) (2025).

Part IV — What actually helps

4.1 The evidence tier table

Tier	Intervention	Best effect size (95% CI)	Source
STRONG	CBT-I for insomnia	ISI g = 0.98; sleep efficiency 0.71; total sleep time only 0.16	van Straten 2018; AASM strong recommendation (Edinger 2021)
STRONG	Exercise for depression	SMD −0.43 (IQR −0.66 to −0.27); vs active controls g −0.42 to −0.62	Singh 2023; Noetel 2024
STRONG	Behavioural activation for depression	SMD 0.67 (0.54–0.80); prediction interval −0.11 to 1.46	Cuijpers 2026
STRONG	Social connection (epidemiological association)	Survival OR 1.50 (1.42–1.59)	Holt-Lunstad 2010
MODERATE	Exercise for anxiety	SMD −0.42; resistance training Δ 0.31 (0.17–0.44)	Singh 2023; Gordon 2017
MODERATE	CBT-based occupational stress management	d = 1.164 (0.456–1.871) — but k = 7 only	Richardson & Rothstein 2008
MODERATE	Mindfulness for anxiety/depression vs non-specific controls	0.38 (0.12–0.64); 0.30 (0.00–0.59)	Goyal 2014
MODERATE	Caffeine restriction before bed	Recovers 45.3 min (29.0–61.5) total sleep time	Gardiner 2023
MODERATE	Alcohol reduction	Recovers 2.8% (1.7–3.9) REM proportion	Gardiner 2025
MODERATE	Psychological detachment (correlational)	Fatigue r = −0.42, prediction interval crosses zero	Wendsche 2017
LIMITED	Micro-breaks	Vigour d 0.36 (0.16–0.55); performance 0.16 (−0.04 to 0.37) NS	Albulescu 2022
LIMITED	Burnout interventions (all types)	Overall SMD −0.29 (−0.42 to −0.16) ≈ 3 MBI points	Panagioti 2017
LIMITED	Organisation-directed burnout interventions	−0.45 (−0.62 to −0.28), but very low GRADE elsewhere	Panagioti 2017; Bes 2023
LIMITED	Professional coaching (physicians)	Depersonalisation −0.30 (−0.42 to −0.19), moderate certainty	Collett 2026
LIMITED	HRV biofeedback / slow breathing	Breathwork on stress g −0.35 (−0.55 to −0.14)	Fincham 2023
LIMITED	Digital / app interventions	Stress g 0.35 — but 0.09 (NS) vs a placebo app	Linardon 2019
LIMITED	Vacations	d = +0.43, fade-out −0.38, only 7 studies	de Bloom 2009
LIMITED	Loneliness interventions	RCT-only effect −0.198 (−0.32 to −0.08)	Masi 2011
NULL / INSUFFICIENT	Sleep hygiene alone	AASM recommends against as a single component	Edinger 2021
NULL / INSUFFICIENT	Individual workplace wellness programmes	2 of 80 outcomes, both self-reported	Song & Baicker 2019/2021; Jones 2019
NULL / INSUFFICIENT	Mindfulness vs an evidence-based treatment	d = −0.004 (−0.15 to 0.14)	Goldberg 2018
NULL / INSUFFICIENT	Omega-3 for depression prevention	RR 1.01 (0.92–1.10), I² = 0%, N = 41,470	Deane 2021
NULL / INSUFFICIENT	Vitamin D for depression (non-deficient)	HR 0.97 (0.87–1.09), N = 18,353, 5.3 y	Okereke 2020
NULL / INSUFFICIENT	B vitamins for fatigue (non-deficient)	No effect; fatigue not analysable for lack of data	Markun 2021
NULL / INSUFFICIENT	Rhodiola for fatigue	RCT: placebo better, MD −17.3 (−30.6 to −3.9)	Punja 2014
NULL / INSUFFICIENT	Cross-stressor adaptation (blunted reactivity)	Not replicated at N = 832 or N = 116	Ensari 2020; van der Mee 2023

4.2 Sleep — the strongest single lever

CBT-I is the only intervention in this review carrying a strong recommendation from a major guideline body (AASM; Edinger et al., 2021). The largest synthesis (van Straten et al., 2018; 87 RCTs, 3,724 treated) gives Hedges' g of 0.98 for insomnia severity, 0.71 for sleep efficiency, 0.63 for wake-after-sleep-onset, 0.57 for sleep-onset latency — and 0.16 for total sleep time, the smallest effect in the paper.

Read that pattern honestly: CBT-I substantially improves sleep quality and continuity but barely increases sleep duration. If the problem is that you are not getting enough sleep because of hours, obligations or shift work, CBT-I is not the answer to that problem.

Sleep hygiene is the weakest component and the most widely disseminated. The AASM guideline explicitly recommends against sleep hygiene as single-component therapy. The active ingredients are stimulus control and sleep restriction — which are behaviourally demanding and initially worsen daytime sleepiness, which is exactly why the easy version gets promoted instead.

Dose: at least four face-to-face sessions outperformed self-help or shorter formats. Digital CBT-I is a genuine partial answer — apps for sleep problems showed g = 0.71 (0.51–0.92) with no detectable publication bias, unlike depression and anxiety apps.

Sleep extension is weaker than assumed. A systematic review of nine studies found extension improved sustained attention and reaction time but had no beneficial effect on executive function, with benefits "evident but short-lived," and concluded it does not support sleep extension as a sole intervention for cognitive deficits following sleep debt (Yu et al., 2025). You cannot simply bank sleep back.

4.3 Physical activity — strong for mood, more complicated for fatigue

The umbrella review (Singh et al., 2023; 97 reviews, 1,039 RCTs, >128,119 participants) gives median SMD −0.43 for depression and −0.42 for anxiety. Note the quality warning: AMSTAR-2 rated 77 of 97 included reviews as critically low quality, and the frequently quoted psychological-distress figure (−0.60) rests on a single review of 6 RCTs.

The best-designed synthesis is Noetel et al. (2024), a network meta-analysis of 218 studies against active controls: walking/jogging g = −0.62, yoga −0.55, strength training −0.49, mixed aerobic −0.43, tai chi/qigong −0.42. Their own caveat: "only one study met the Cochrane criteria for low risk of bias."

Two counterintuitive findings deserve care.

Dose-response runs backwards in the meta-analysis. Shorter interventions outperformed longer (≤12 weeks −0.84; ≥24 weeks −0.28), and ≤150 min/week outperformed >150 min/week. This is far more likely an artefact of adherence decay and small-study effects than a biological ceiling. Treat it as a warning about the literature, not a prescription to exercise less.

Exercise reduces fatigue — but the placebo-controlled evidence is null. Puetz, O'Connor & Dishman (2006) found chronic exercise increased energy and reduced fatigue by mean Δ = 0.37 — with the decisive qualification in their own words: "Investigations that used a placebo control and examined chronic exercise alone found no effect of chronic exercise on feelings of energy and fatigue." In clinical fatigue the signal is cleaner (cancer-related fatigue Δ = 0.32–0.38). So: exercise for mood on strong evidence, and for fatigue on weaker evidence, while recognising that expectancy is doing some of the work.

Cross-stressor adaptation is largely unsupported by modern tests. The hypothesis that fitness blunts psychological stress reactivity has older meta-analytic support but fails two well-powered modern tests: Ensari et al. (2020), N = 832, ambulatory monitoring — "did not support the hypothesis"; van der Mee et al. (2023), N = 116, three exposure definitions — "We did not find evidence for the cross-stressor adaptation hypothesis, irrespective of ANS or affective outcome measure." Exercise helps mood; the specific claim that it makes you physiologically less reactive to stress does not survive.

Practical dose: 3–5 sessions/week, moderate-to-vigorous, either resistance or aerobic. Benefit appears at volumes below public-health guidelines, with the steepest gradients at low activity volumes (Pearce et al., 2022; 191,130 participants). The largest return is on the first increment from doing nothing.

4.4 Psychological interventions

Behavioural activation has the best recent evidence of any psychotherapy here and is the most scalable. Cuijpers et al. (2026), 105 trials, 13,933 patients: SMD 0.67 (0.54–0.80) versus controls, with no significant difference from other therapies. Self-guided BA still worked (0.36). (Note the 95% prediction interval −0.11 to 1.46 crosses zero — the pooled effect is solid, the expected effect in a new setting is not guaranteed.)

Its practical appeal is that it is the lowest-complexity protocol with a top-tier evidence base: schedule and re-engage with valued and rewarding activity, on a plan rather than on how you feel. It also happens to map neatly onto the effort-cost model in §2.5 — it changes the cost–benefit structure rather than trying to increase capacity.

Occupational stress management. Richardson & Rothstein (2008), 55 interventions: overall d = 0.526; cognitive-behavioural d = 1.164 (0.456–1.871); relaxation 0.497; multimodal 0.239; organisational 0.144 (−0.123 to 0.411) — null.

Burnout interventions — report these honestly, they are small. Panagioti et al. (2017), 20 comparisons, n = 1,550 physicians, controlled designs only: overall SMD −0.29 (−0.42 to −0.16), described by the authors as "a drop of 3 points on the MBI emotional exhaustion domain above change in controls." Organisation-directed −0.45; physician-directed −0.18; subgroup difference P = .04 — a single-degree-of-freedom test across 20 comparisons, and fragile.

West et al. (2016) is widely cited for burnout falling 54% → 44%, but these are pooled within-group pre-post changes including 37 uncontrolled cohorts — they are not efficacy estimates, and the paper publishes no separate subgroup point estimates. Any source quoting subgroup effect sizes for West 2016 is fabricating them.

Cochrane (Ruotsalainen et al., 58 studies, 7,188 participants): CBT ± relaxation SMD −0.38 (−0.59 to −0.16); physical relaxation −0.47; mental relaxation −0.50 (−1.15 to 0.15, null); changing work schedules −0.55 but from 2 trials and 180 people. "Other organisational interventions were not more effective than no intervention." GRADE: low for all but one comparison.

The most recent update (Collett et al., 2026; 99 trials, 9,330 participants) found professional coaching reduced emotional exhaustion −0.37 and depersonalisation −0.30, the latter at moderate certainty — the strongest certainty rating anywhere in this literature. Mindfulness for physicians was null on both.

ACT is thinner than commonly claimed for this indication — available meta-analyses are small and population-specific. Treat as promising-but-underpowered.

4.5 Mindfulness: moderate for some things, null against real comparators

The anchor is Goyal et al. (2014), notable because every included trial had an active control. Against non-specific active controls: anxiety 0.38 (0.12–0.64), depression 0.30 (0.00–0.59), pain 0.33 (0.03–0.62) — all rated moderate strength of evidence. For stress/distress, positive mood, attention, substance use, sleep and weight the rating was low or insufficient, and no pooled stress effect size exists because the authors declined to pool. Their own conclusion: "We found no evidence that meditation programs were better than any active treatment (ie, drugs, exercise, and other behavioral therapies)."

Nothing since has overturned it. Galante et al. (2021), 136 RCTs, N = 11,605: distress SMD −0.45 versus no intervention, but −0.14 (−0.51 to 0.23), non-significant, versus non-specific active control. Workplace mindfulness (Vainre et al., 2025; 99 studies, N = 16,054): task performance g = 0.25 versus passive controls but 0.12 (NS) versus active controls, with GRADE confidence "very low."

Van Dam et al.'s (2018) critique remains unanswered: no agreed technical definition of mindfulness; questionnaires that rate binge drinkers as more mindful than experienced meditators; "less than 25% of meditation trials actively assess adverse events," potentially underestimating adverse-event frequency more than twentyfold; and only 9% of MBI research conducted at the active-control efficacy stage. Adverse events are real: pooled prevalence 8.3%, rising to 33.2% in observational studies (Farias et al., 2020).

Reporting bias is measurable: 88% of 124 published mindfulness trials concluded the intervention was effective — 1.6× the number expected given d = 0.55 — and 62% of registered trials remained unpublished at 30 months.

One genuinely strong result deserves credit. Hoge et al. (2023): in a 276-participant randomised noninferiority trial, MBSR was noninferior to escitalopram for anxiety disorders, with study-related adverse events in 78.6% of the escitalopram arm versus 15.4% of the MBSR arm. That is a real and useful finding for people who want to avoid medication.

4.6 Recovery, detachment, and the paradox at the centre of everything

Psychological detachment — mentally disengaging from work during non-work time — is the best-supported single construct in the recovery literature. Wendsche & Lohmann-Haislah (2017), 91 samples, N = 38,124: detachment correlates with fatigue r = −0.42, exhaustion −0.36, wellbeing 0.32, sleep 0.30.

Two caveats a rigorous review must state. I² is 80–98% for nearly every outcome, and the 95% prediction intervals cross zero for fatigue, wellbeing, sleep and life satisfaction — a new study could plausibly find nothing. And detachment is not uniformly good: task performance r = 0.09, but contextual performance −0.13 and creativity −0.11.

A refinement that matters. Bennett, Bakker & Field (2018), 54 samples, N = 26,592: detachment→fatigue ρ = −0.39, but detachment→vigour only ρ = 0.14. It is control (ρ = 0.31) and mastery (ρ = 0.29) that predict vigour.

Recovery that removes exhaustion is not the same as recovery that restores energy. Switching off reduces fatigue; doing something you choose and get better at is what generates vigour. Collapsing on the sofa and learning the guitar are not interchangeable, and the literature can tell them apart.

The recovery paradox (Sonnentag, 2018) is the single most policy-relevant idea in this field: "job stressors are not associated with a higher — but a lower — likelihood of recovery-enhancing processes." Job demands → lower detachment, r̄ = −0.25 across 60 studies and N = 28,507. The people who most need recovery are least able to initiate it — because negative activation keeps work cognitions accessible, because resource depletion makes effortful recovery harder to start, and because connectivity removes the boundary.

This predicts that advising depleted people to "recover better" will fail, and it is the strongest argument in the literature for making recovery structural rather than exhortative. If recovery depends on the depleted person's initiative, it will not happen.

Micro-breaks (Albulescu et al., 2022; 22 samples): vigour d = 0.36 (0.16–0.55), fatigue d = 0.35 (0.19–0.50), but performance d = 0.16 (−0.04 to 0.37), non-significant. And break duration positively moderated performance — the popular "even a tiny break boosts productivity" claim is the opposite of what the paper found.

Vacations fade fast. Effect d = +0.43, fade-out −0.38, from only 7 studies (de Bloom et al., 2009). Kühnel & Sonnentag (2011): burnout decreased and engagement rose after vacation, but "these beneficial effects faded out within one month" — accelerated by post-vacation job demands, delayed by leisure relaxation. A holiday is a genuine intervention with a one-month half-life. It is not a treatment for a structural problem.

4.7 Organisational interventions: everyone's evidence is weaker than they claim

Individual wellness programmes: null. Song & Baicker (2019) cluster-randomised 160 worksites and 32,974 employees for 18 months. Of 80 prespecified outcomes, 2 were significant — both self-reported (regular exercise +8.3pp, actively managing weight +13.6pp). All 10 clinical measures null. All 38 spending/utilisation outcomes null. All 3 employment outcomes null. At three years the pattern was unchanged. Mean uptake: 1.3 of 8 modules — so this is as much a test of weak implementation as of the concept.

Jones, Molitor & Reif (2019), ~4,834 individuals randomised, found "strong patterns of selection" (participants were already healthier and cheaper before the intervention) and no causal effect on spending, behaviours, productivity or self-reported health after two years. Their confidence intervals "rule out 84% of previous estimates." The selection finding is the substantive one: apparent wellness ROI in observational studies can be entirely a sorting artefact.

Fleming (2024) studied 46,336 employees across 233 organisations and found participants in resilience training, stress management and coaching were no better off. Important design caveat: this is cross-sectional — the published subtitle is literally "Cross-sectional evidence from the United Kingdom" — comparing participants with non-participants, not a randomised trial. [unverified: whether propensity-score matching was used, and the widely repeated per-intervention findings ("volunteering the only positive signal," "mindfulness negative on some outcomes"), could not be confirmed from an accessible primary source. Cite the headline null; do not cite the details.]

But the organisational case is also weaker than advertised. It rests on two fragile subgroup tests (Panagioti's Q = 4.15, P = .04; West's interaction p = 0.03 with I² = 79%) and is contradicted by Richardson & Rothstein (organisational d = 0.144, null) and by Cochrane ("other organisational interventions were not more effective than no intervention"). Where organisational effects are pooled directly (Bes et al., 2023; 13 studies): exhaustion ES −0.30 (−0.42 to −0.18) — at very low GRADE certainty.

The best randomised test of actual work redesign is sobering. The Work, Family and Health Network STAR trial increased schedule control and supervisor support. Results: +8 minutes per day of sleep in an IT firm; +9 minutes at 12 months; no significant sleep effect at all in nursing homes; no main effect on cardiometabolic risk in either industry. Egan et al. (2007) reviewed 18 employee-control interventions and found no RCTs at all — and two participatory interventions running alongside redundancies worsened employee health.

On the four-day week. The widely cited UK pilot had 61 self-selected companies, no control group, no randomisation, all self-report. The famous "71% reduced burnout" figure is the proportion of individuals whose score moved down at all — not a 71% reduction. Fan et al. (2025, Nature Human Behaviour) reports improvements across 2,896 employees with 12 non-randomised control companies, but [unverified: the abstract gives direction only; no coefficient or CI could be obtained].

Honest summary of this section: the strongest finding in the organisational literature is a negative one — individual-level wellness programmes do not work at scale. The organisational alternative is directionally better supported but rests on very low GRADE certainty and on randomised trials that produced eight extra minutes of sleep. Anyone claiming this question is settled, in either direction, is ahead of the evidence.

4.8 Social connection

Epidemiologically strong, interventionally weak. Holt-Lunstad et al. (2010), 148 studies, N = 308,849: OR 1.50 (1.42–1.59) for survival with stronger social relationships — though note the spread by operationalisation, from complex social integration 1.91 down to living alone 1.19 (non-significant). Fully adjusted in the 2015 update (N = 3,407,134): social isolation 1.29, living alone 1.32, loneliness 1.26 — and the three did not differ significantly from one another, so claims that loneliness specifically is the operative factor are unsupported. Effects were stronger in younger adults.

(The popular "equivalent to 15 cigarettes a day" gloss appears nowhere in Holt-Lunstad's papers.)

Interventions are small. Masi et al. (2011), 50 studies: randomised comparisons gave mean effect −0.198 (−0.32 to −0.08) — less than half the uncontrolled pre-post estimate. Within those 20 RCTs, only 6 showed efficacy. The subtype that worked best (addressing maladaptive social cognition, −0.598) rests on 4 studies with a borderline moderator test.

4.9 Caffeine, alcohol, diet, supplements

Caffeine — moderate evidence, clear numbers. Gardiner et al. (2023), 24 controlled crossover studies: caffeine reduces total sleep time by 45.3 min (29.0–61.5), increases sleep-onset latency 9.1 min, increases wake-after-sleep-onset 11.8 min, reduces sleep efficiency 7.0% and slow-wave sleep 11.4 min; REM effects null. The timing gradient is modest: the total-sleep-time decrement shrinks 2.8 min per additional hour before bed. Derived cut-offs: a coffee (107 mg) ≥8.8 h before bed; a pre-workout dose (217.5 mg) ≥13.2 h.

Correct a widespread mis-citation: the "caffeine 6 hours before bed costs you an hour of sleep" claim traces to a study with n = 12, no polysomnography, funded by the device manufacturer, whose diary effect at 6 h was −41 min, p = 0.08, non-significant, and which explicitly states the three timings "did not produce differential sleep disruption."

The caffeine–fatigue loop. Withdrawal is well characterised: 10 validated symptoms, headache incidence 50%, onset 12–24 h, peak 20–51 h, duration 2–9 days, from doses as low as 100 mg/day. [contested] Whether habitual caffeine confers net benefit is genuinely disputed — Rogers et al. (2010, n = 379) found "no net benefit for alertness." A defensible framing: tolerance appears near-complete for subjective alertness in habitual consumers, while psychomotor benefits persist. The blanket claim "caffeine only reverses withdrawal" is an overreach.

Alcohol — and a correction. Gardiner et al. (2025), 27 studies: REM onset latency +18.0 min; REM duration −11.3 min; REM proportion −2.8%; total sleep time null; sleep efficiency null; WASO null. Dose-response per 1 g/kg: REM onset +30.1 min, REM duration −40.4 min, with effects emerging from ~0.35 g/kg — roughly two to three drinks. Note this contradicts the "REM rebound" story: whole-night REM is reduced, not rebounded.

Diet — limited, and weaker than the headlines. The corrected figure for dietary interventions on depression is g = 0.162 (0.055–0.269), about 40% smaller than the widely circulated 0.275, with anxiety null. Only 1 of 16 trials studied clinically depressed people (SMILES), which randomised 67 of a targeted 176, used a befriending control against a dietician-plus-free-food-hampers arm, and carries a published correction. Restricted to actual depressive disorders (Tavakoly et al., 2025; 5 RCTs, n = 952), both comparisons were null.

Adaptogens and supplements — mostly overhyped

Ashwagandha. Pooled anxiety effects look spectacular: SMD −1.55 and −1.62 across two meta-analyses, falling to −1.13 after outlier correction. Treat with deep scepticism, for five reasons:

1. The literature contains no negative trials — "one showed a null effect and none reported negative results… confirmed by Egger's test p < 0.001."
2. Mean trial size ~51 participants (~25 per arm), median duration 8 weeks, no long-term data.
3. 57% of trials are from a single country, with limited pre-registration.
4. GRADE low to very low throughout.
5. Most tellingly: the one review reporting both outcomes in explicit units found cortisol fell (−1.16 µg/dL) while perceived stress was flatly null (SMD −0.355, p = 0.40). The biomarker moves; the symptom does not.

Safety is not trivial. LiverTox assigns ashwagandha likelihood score B — "likely cause of clinically apparent liver injury" — with latency 2–12 weeks. One Indian case series reports 23 patients, of whom 3 developed acute-on-chronic liver failure and all 3 died; chemical analysis found no adulterants. (On regulation: Denmark issued a 2020 risk assessment concluding no safe intake level could be established. The commonly repeated "Denmark banned ashwagandha" is not accurate as stated.)

Rhodiola rosea. No meta-analysis is possible — "no two studies reported the same outcomes"; "all of the included studies exhibit either a high risk of bias or have reporting flaws." The confirmatory trial went the wrong way: Punja et al. (2014), n = 48, double-blind, found vitality favoured placebo, MD −17.3 (−30.6 to −3.9), p = 0.011 — rhodiola worsened fatigue.

Magnesium. The flagship review states plainly: "No study administered a validated measure of subjective stress as an outcome," none recruited magnesium-depleted samples, and "all of the studies which included a placebo demonstrated significant placebo effects."

Vitamin D, omega-3, B vitamins in non-deficient people: null. VITAL-DEP (N = 18,353, median 5.3 years): depression HR 0.97 (0.87–1.09). Omega-3 across 31 trials and 41,470 participants: RR 1.01 (0.92–1.10), I² = 0%, moderate quality — and the VITAL omega-3 arm found a significant increase in depression risk (HR 1.13, 1.01–1.26). B vitamins: no effect on cognition or depression, with fatigue not analysable for lack of data.

4.10 When to involve a clinician — information, not medical advice

This section describes what clinicians investigate. It is not medical advice and is not a substitute for individual assessment. Anyone with persistent unexplained fatigue should see a doctor.

The single most useful number in this domain comes from Stadje et al. (2016), a systematic review of 26 primary-care studies of tiredness as a presenting complaint:

Cause	Pooled prevalence
Anaemia	2.8% (1.6–4.8) — pooled from 3 studies
Malignancy	0.6% (0.3–1.3) — 3 studies
Serious somatic disease	4.3% (2.7–6.7) — 3 studies
Depression	18.5% (16.2–21.0) — 6 studies

And the finding that reframes the popular "get your bloods checked" advice: in studies with control groups, the prevalence of somatic disease did not differ significantly between patients with and without tiredness, while depression was substantially more frequent among the tired. The authors' conclusion: "Extensive investigations are only warranted in case of specific findings from the history or clinical examination. Instead, attention should focus on depression and psychosocial problems."

Two honest caveats on this study: the pooled figures rest on very few studies each; only 6 of 26 had control groups; the somatic-disease "no difference" rests substantially on two single studies, one of which was powered at 0.09 to detect the difference it failed to find; and the paper's own discussion gives 3.1% where the abstract gives 4.3%. Soften "identical" to "did not differ significantly in the few studies with controls."

A basic screen is still cheap and reasonable — a competent clinician will typically consider full blood count, thyroid function, HbA1c or glucose, ferritin, renal and liver function, and coeliac serology where indicated, plus screening for obstructive sleep apnoea (snoring, witnessed apnoeas, daytime sleepiness, BMI, neck circumference) and structured screening for depression and anxiety. The yield of extensive investigation without specific findings is low.

Antidepressants where major depression is diagnosed: all 21 studied drugs beat placebo, ORs from 2.13 (amitriptyline) to 1.37 (reboxetine), across 522 trials and 116,477 participants (Cipriani et al., 2018) — with differences between active drugs small and certainty of evidence moderate to very low. Antidepressants treat depression. They are not an intervention for non-depressive fatigue.

4.11 Breathing, biofeedback and apps

HRV biofeedback / slow-paced breathing — limited: real signal, poor methods. Goessl et al. (2017), 24 studies: pre-post g = 0.81, between-group g = 0.83 (0.34–1.33) — but only 2 of 24 studies had independent allocation concealment and only 2 of 13 comparators were sham. The between-group estimate is largely biofeedback versus doing nothing, with a CI spanning "small" to "very large."

Breathwork generally (Fincham et al., 2023): stress g = −0.35 (−0.55 to −0.14), anxiety −0.32, depression −0.40, with no funnel asymmetry. The authors' own candour is worth quoting — significance against active controls "could be due to poor quality of the active controls," and they "urge caution… to avoid a miscalibration between hype and evidence."

Practical protocol as studied: ~6 breaths/min, 10–20 min/day, 4–10 weeks. Cheap, low-risk, expect a small effect.

Apps — small effects, catastrophic real-world attrition. Linardon et al. (2019), 66 RCTs: stress g = 0.35 overall, but 0.47 versus waitlist and 0.09 (−0.05 to 0.24), non-significant, versus an attention/placebo app. Against active interventions, depression g = 0.13 (NS).

The attrition figures are the real story. Pooled trial dropout 26.2%, rising to 47.8% after trim-and-fill. In the wild, across 93 Android mental-health apps with ≥10,000 installs: median 30-day retention 3.3%, median daily active users 4.0%. Breathing-exercise apps had a median day-30 retention of 0.0%.

4.12 The overhyped list

1. Sleep hygiene as a standalone intervention — the AASM recommends against it.
2. Individual corporate wellbeing programmes — resilience training, stress-management workshops, mindfulness apps, EAP-as-population-intervention. Two randomised trials and one large cross-sectional study converge on null.
3. Mindfulness as a general-purpose intervention — moderate for anxiety, depression and pain; low or insufficient for stress, sleep, attention and mood; no advantage over any active treatment; adverse events unassessed in over 75% of trials.
4. Ashwagandha — huge pooled effects from tiny trials with I² ≈ 94%, Egger's p < 0.001, no negative trials in the entire literature, and a real hepatotoxicity signal.
5. Rhodiola — the confirmatory RCT found it worsened vitality relative to placebo.
6. Magnesium for stress — no study used a validated subjective stress measure.
7. Omega-3, vitamin D and B vitamins in non-deficient people — consistently null in megatrials; VITAL found omega-3 increased depression risk.
8. "Caffeine 6 hours before bed costs an hour of sleep" — from an n = 12, manufacturer-funded study whose result was non-significant and which found no timing gradient.
9. "Alcohol causes REM rebound" — whole-night REM is reduced, not rebounded.
10. "The four-day week cut burnout by 71%" — the 71% is the proportion of individuals whose score moved down at all, in an uncontrolled self-selected self-report pilot.
11. "Micro-breaks boost productivity" — pooled performance effect non-significant, with small-study bias, and longer breaks did better.
12. Wellness apps — g = 0.09 versus a placebo app; ~3% real-world 30-day retention.
13. "Loneliness kills as much as 15 cigarettes a day" — not a figure in Holt-Lunstad's papers.
14. "Adrenal fatigue" — refuted by systematic review.
15. "The 12 stages of burnout" — zero empirical validation in the indexed literature.

---

Part V — Synthesis

5.1 A defensible causal model

Putting the parts together, here is what the evidence supports as a working model. Arrows marked (w) are weak or contested; arrows marked (s) are well supported.

                         CHRONIC DEMAND
                    (workload, low control, low reward,
                     mistreatment, insecurity, caregiving)
                                 │
              ┌──────────────────┼──────────────────┐
              │                  │                  │
              ▼                  ▼                  ▼
      IMPAIRED RECOVERY    SUSTAINED             SLEEP LOSS
      (low detachment)     ACTIVATION            (initiation,
        (s) r ≈ −.36       (SAM + HPA)           continuity)
              │                  │                  │
              │                  ▼                  │
              │         ┌────────────────┐          │
              │         │  METABOLIC     │          │
              └────────▶│  COST OF       │◀─────────┘
                        │  ALLOSTASIS    │
                        │  (s, cellular; │
                        │   w, whole-body│
                        │   — the gap)   │
                        └───────┬────────┘
                                │
          ┌─────────────────────┼─────────────────────┐
          ▼                     ▼                     ▼
   LOW-GRADE              RAISED EFFORT          IMPAIRED
   INFLAMMATION           PRICE                  PFC FUNCTION
   (w at population       (s, mechanism;         (s, acute small;
    level, z ≈ .10)       Wiehler/MetaMotiF)      s, chronic broad)
          │                     │                     │
          └──────────┬──────────┴──────────┬──────────┘
                     ▼                     ▼
              MOTIVATIONAL           MEASURED DEFICITS
              WITHDRAWAL             (retrieval −0.22;
              ("can't face it")      burnout attention −0.43)
                     │                     │
                     └──────────┬──────────┘
                                ▼
                    FUNCTIONAL CONSEQUENCES
              (errors, safety incidents OR ≈ 2,
               presenteeism, relational reactivity)
                                │
                                ▼
                    FEEDBACK: burnout → stressors
                    (s, and LARGER than the forward path)

Four features of this diagram are worth stating explicitly.

The cortisol box is missing on purpose. Cortisol is a mediator in the classical account and appears nowhere as a load-bearing arrow here, because the measured associations are r ≈ 0.10–0.15, direction is unpredictable, and two decades of searching found no burnout biomarker. Cortisol is a marker of activation, not a good predictor of outcome.

The main pathway is metabolic, not endocrine. The strongest mechanistic claim available is that maintaining the stressed state is expensive and the cost is paid from maintenance and repair. That claim is well evidenced at the cellular level and unproven at the whole-body level (§1.6). It is the field's central open question.

"Effort price" replaces "energy reserve." This is not a semantic preference. Depletion predicts that rest restores capacity; effort-cost predicts that rest helps only insofar as it lowers the price, which is why people return from holidays feeling fine and collapse again within a month (§4.6).

The feedback loop is larger than the forward path. Guthier et al. (2020) found the strain→stressor effect exceeds the stressor→strain effect. Whatever starts the process, once it is running, the state maintains its own conditions.

5.2 The five open questions worth a research programme

1. The scale gap: does cellular hypermetabolism show up in whole-body energetics? Chronic glucocorticoid exposure raises fibroblast energy expenditure ~60% and psychosocial factors explain 18–25% of variance in brain complex I abundance — yet trait stress shows no association with resting metabolic rate, and recent stressors predict lower postprandial thermogenesis. Either the cellular effect is too small a fraction of total expenditure to detect, or it is compartment-specific, or the in vitro model does not translate. Resolving this would settle whether "stress depletes energy" is literally true at the organismal level. Design consideration: the discriminating study is probably not another RMR correlation but a within-person longitudinal design with doubly-labelled water plus tissue-level markers across a documented stress transition.

2. Can effort price be measured cheaply enough to be clinically useful? Wiehler's glutamate finding requires 7T MRS. Effort-discounting tasks and pupillometry are cheaper proxies but under-validated for individual-level use. A validated bedside measure of effort cost — distinct from both capacity testing and fatigue questionnaires — would be genuinely new, and §2.6 shows why neither existing measure substitutes for it.

3. Why is the subjective–objective dissociation so large, and is it prognostic? Recovered exhaustion patients show attention deficits unrelated to work resumption; sleep-restricted people cannot feel their own impairment; a verified stress induction can leave regulation intact. Is the size of an individual's subjective–objective gap itself a predictor — of recovery, relapse, or of who pushes through into clinical exhaustion? Nobody appears to have asked.

4. Does burnout survive as a construct, and if not, what replaces it? Disattenuated exhaustion–depression r = .80 with essential unidimensionality is a serious result, and exhaustion disorder differs from depression on two of nine symptom scales. But the Swedish long-term data (exhaustion persisting at seven years while depression resolves in months) suggest something distinguishes the trajectories even if the cross-sectional symptom profiles overlap. The discriminating test is longitudinal, not cross-sectional — and it has not been run properly.

5. Why do burnout interventions fail, and does the recovery paradox explain it? Return-to-work intervention effects are null (OR 1.33, ns), two trials increased sickness absence, and one large trial found cognition improved equally in both arms. Sonnentag's paradox predicts exactly this: interventions requiring initiative from a depleted person will fail. The test is a trial comparing an intervention that requires initiative against one that removes demand structurally without requiring anything of the participant. As far as this review can establish, that trial has not been done.

5.3 Practical translation

Everything below follows from the evidence above. Nothing here is medical advice, and anyone whose functioning is substantially impaired should see a clinician — see §4.10 for what a competent assessment covers.

Establish the base rate before reaching for stress

Three things are more common than the stress explanation and are routinely missed:

• Chronic sleep restriction. You cannot feel the difference between six hours and four (Van Dongen). Before attributing fatigue to stress, count actual sleep for two weeks. If it is under seven hours on most nights, that is a sufficient explanation and should be dealt with first.
• Depression. 18.5% of fatigue presentations in primary care. Exhaustion and depression correlate at r ≈ .80 once you correct for measurement error. "Burnout" is a more comfortable label and a less useful one.
• A sleep disorder, most commonly apnoea. Cheap to screen, transformative to treat, and invisible to the person who has it.

Rank interventions honestly

Do this first	Why
Fix the sleep disorder — CBT-I if insomnia, investigate if apnoea is plausible	The only strong guideline recommendation in this review (g ≈ 0.98 for insomnia severity)
Move, three to five times a week	Strong evidence for mood (SMD ≈ −0.43 to −0.62 against active controls); the largest returns are on the first increment from nothing
Screen for and treat depression	Highest base rate of any treatable cause; strong treatment evidence
A structured behavioural protocol (behavioural activation)	SMD 0.67, and it works self-guided (0.36) — the highest evidence-to-complexity ratio available
Protect detachment structurally	The single best-supported recovery construct (r ≈ −.42 with fatigue)

Worth doing, modest returns	Caveat
Caffeine cut-off ~9 h before bed	Recovers ~45 min sleep; the 6-hour rule is a mis-citation
Reduce alcohol	Costs ~2.8% of REM and delays REM onset ~18 min; total sleep time unaffected, which is why people don't notice
Slow-paced breathing, ~6/min, 10–20 min	g ≈ −0.35; cheap, low-risk, small
Real breaks, and longer ones	Vigour d = 0.36 — but the performance effect is null, so do it for recovery, not output
Mastery activity, not just rest	Control and mastery predict vigour (ρ ≈ 0.30); switching off only reduces fatigue

Don't bother	Why
Sleep hygiene alone	Recommended against by the AASM
Adaptogens (ashwagandha, rhodiola)	No negative trials in the ashwagandha literature at all; rhodiola RCT favoured placebo; real hepatotoxicity signal
Supplements if you're not deficient	Null in megatrials; omega-3 arm of VITAL increased depression risk
Resilience training and wellness apps	2 of 80 outcomes in a 32,974-person RCT; g = 0.09 versus a placebo app
Waiting for a holiday to fix it	Effect fades within one month

The two reframes worth internalising

"I have no energy" is usually "effort has become expensive," not "the tank is empty." This matters because the two call for different responses. An empty tank suggests rest. An expensive effort price suggests changing what makes effort expensive — reducing competing demands, restoring genuine detachment, removing the low-control or low-reward features of the situation, and protecting the hardest cognitive work for the hours when the price is lowest. Rest alone changes the price only temporarily, which is precisely the vacation fade-out result.

Recovery does not happen by intention when you are depleted. That is the recovery paradox, and it is the best-supported reason that "you just need to look after yourself better" fails as advice. The people most in need of recovery are least able to initiate it. The workable response is to make recovery structural — boundaries that do not depend on your discipline on a bad day — rather than aspirational.

A note on the long tail

If you are reading this because you are somewhere in this territory personally, the seven-year data (§3.7) deserve to be stated plainly rather than softened: among people who reached the clinical threshold, only 16% reported full recovery, 46% still reported extreme fatigue, and 73% reported reduced stress tolerance — while 87% were nevertheless back at work.

Two things follow. The pessimistic one is that reaching that threshold has consequences that persist far longer than the sick note. The more useful one is that functional recovery arrives long before symptomatic recovery, and reduced stress tolerance is compatible with a working life — 73% of that cohort had changed workplace, which reads less like defeat than like accurate adjustment. The strongest argument in the data is for acting well before the threshold, and the second strongest is that the trajectory afterwards is not the cliff it feels like from inside it.

Appendix A — Key numbers

Finding	Value	Source
BIOENERGETICS
Chronic glucocorticoid → cellular energy expenditure (human fibroblasts)	~+60%, causally linked to accelerated ageing on 3 clocks	Bobba-Alves 2023, Psychoneuroendocrinology 155:106322
Acute mental arithmetic → energy expenditure	+9% to +42%	Bobba-Alves 2022, Psychoneuroendocrinology
30-min extended cognitive stress → energy expenditure	~+67%	Bobba-Alves 2022
Acute mental stress → VO₂ / glucose utilisation (β-blockade abolished)	+20% / +34%	Seematter 2000, Am J Physiol Endocrinol Metab
Psychosocial factors → variance in brain OxPhos complex I	18–25%	Trumpff 2024, PNAS 121:e2317673121
Trait stress/anxiety → resting metabolic rate	No association — the scale gap	Wilson 2020, Appl Physiol Nutr Metab
CORTISOL & ALLOSTATIC LOAD
Flatter diurnal cortisol slope → poorer health	r = 0.147 (immune/inflammation 0.288)	Adam 2017, Psychoneuroendocrinology 83:25–41
"Adrenal fatigue" systematic review	58 studies; "adrenal fatigue is still a myth"	Cadegiani & Kater 2016, BMC Endocr Disord 16:48
Burnout biomarkers identified in 31 studies / 38 candidates	None	Danhof-Pont 2011, J Psychosom Res 70:505–524
High allostatic load → all-cause mortality	HR 1.22 (1.14–1.30); CVD 1.31 (1.10–1.57); I² > 90%	Parker 2022, Am J Prev Med 63:131–140
Lowest-quartile 5-min RMSSD → all-cause mortality	HR 1.56 (1.32–1.85), N = 38,008	Jarczok 2022, Neurosci Biobehav Rev 143:104907
Common genetic variation explaining plasma cortisol variance	<1% (top locus = a binding-protein gene)	Bolton 2014, PLoS Genet 10:e1004474
SLEEP
6 h/night × 14 nights	≈ 2 nights of total sleep deprivation — unnoticed by the sleeper	Van Dongen 2003, Sleep 26:117–126
Partial / total sleep loss → evening cortisol	+37% / +45%	Leproult 1997, Sleep 20:865–870
Sleep restriction → glucose tolerance / insulin sensitivity	−40% / −18–24%	Spiegel 1999, Lancet 354:1435–1439
Daily GH output occurring during early sleep	~70%	Van Cauter 1998, Sleep 21:553–566
Acute cortisol administration → REM / SWS (healthy volunteers)	REM 19.9%→12.2%; SWS 9.4%→13.9%	Friess 1994, J Sleep Res 3:73–79
Chronic hypercortisolism (Cushing's) → delta sleep	5.8% vs 14.0% in controls — inverse of the acute pattern	Friedman 1994, Neuroendocrinology 60:626–634
INFLAMMATION
Chronic stress → developing a cold on viral challenge	OR 1.99 (1.04–3.08)	Cohen 2012, PNAS 109:5995–5999
Childhood trauma → CRP / IL-6 / TNF-α	z = 0.10 / 0.08 / 0.23	Baumeister 2016, Mol Psychiatry 21:642–649
Acute lab stress → IL-1β / IL-6 / TNF-α	d = 0.66 / 0.35 / 0.28	Marsland 2017, Brain Behav Immun
Chronic low-grade inflammation → systemic energy expenditure	up to ~+10% (estimate)	Lacourt 2018, Front Behav Neurosci 12:78
COGNITION
Acute stress → working memory	g = −0.197 [−0.330, −0.064]	Shields 2016, Neurosci Biobehav Rev
Acute stress → inhibition, overall	g = −0.076 [−0.243, 0.092] — null	Shields 2016
High- vs low-load working-memory tasks	g = −0.303 vs −0.049	Shields 2016
Exogenous cortisol → executive function	g = 0.030, p = .495 — null	Shields 2016
Stress at encoding / consolidation / retrieval	−0.109 (ns) / +0.206 / −0.215	Shields 2017, Psychol Bull 143:636–675
Retrieval, by valence	negative −0.303; positive −0.385; neutral −0.136	Shields 2017
Clinical burnout → fluency / attention-speed / EF	g = −0.53 / −0.43 / −0.39; crystallised unaffected	Gavelin 2022, Work & Stress 36:86–104
Ego depletion (23 labs, N = 2,141)	d = 0.04 [−0.07, 0.15]	Hagger 2016, Perspect Psychol Sci
Ego depletion (36 sites, N = 2,463)	d = 0.06 [−0.02, 0.14], BF ≈ 4:1 null	Vohs 2021, Psychol Sci 32:1566–1581
Acute stress → delay discounting	SMD −0.18 [−0.57, 0.20], p = .32 — null	Forbes 2024, Neurobiol Stress 31:100653
Stress → habit shift (129 pp., 2 stressors)	BF₀₁ = 4.67–4.98 favouring null	Zwosta 2025, PLoS ONE 20:e0327807
Payday replication of scarcity effects (N = 3,821)	Null; needs 11–20× larger shock to reconcile	Carvalho 2016, Am Econ Rev 106:260–284
REAL-WORLD FUNCTION
Physician burnout → patient safety incidents	OR 2.04 [1.69–2.45], 239,246 physicians	Hodkinson 2022, BMJ 378:e070442
Physician burnout → turnover intention	OR 3.10 [2.30–4.17]	Hodkinson 2022
Driving while observably emotional	OR 9.8 [5.0–19.0] (vs drowsy 3.4, phone 3.6)	Dingus 2016, PNAS 113:2636–2641
Burnout → workplace adverse events	ρc = .29 (N = 12,144)	Nahrgang 2011, J Appl Psychol 96:71–94
Exhaustion → job performance	r = −.17 (N = 18,019)	Corbeanu 2023, EJWOP 32
Burnout cost per executive per year	$20,683; $5.04M per 1,000 employees	Martinez 2025, Am J Prev Med
Presenteeism vs absenteeism cost	5–10× higher	Evans-Lacko & Knapp 2016
BURNOUT & TRAJECTORY
Physician burnout prevalence range	0% – 80.5%, ≥142 unique definitions	Rotenstein 2018, JAMA 320:1131–1150
Exhaustion–depression disattenuated correlation	r = .80, N = 12,417	Bianchi 2021, Clin Psychol Sci 9(4)
Exhaustion disorder vs MDD: symptom scales differing	2 of 9	Sennerstam 2025, Scand J Psychol
Empirical validations of "the 12 stages of burnout"	0 (Europe PMC + PubMed, positive controls used)	This review
Detachment ↔ exhaustion	r̄ = −0.36 [−0.42, −0.30], N = 7,007	Wendsche 2017, Front Psychol 7:2072
Detachment ↔ fatigue / ↔ vigour	ρ = −0.39 / only +0.14 (control 0.31, mastery 0.29)	Bennett 2018, J Organ Behav 39:262–275
Job demands → lower detachment (recovery paradox)	r̄ = −0.25, k = 60, N = 28,507	Sonnentag 2018
Strain → stressor path vs stressor → strain path	Reverse path larger	Guthier 2020, Psychol Bull 146:1146–1173
Physician burnout, ≥1 symptom, 2011 → 2023	45.5% → 62.8% (2021) → 45.2% — non-monotonic	Shanafelt 2025, Mayo Clin Proc 100:1142–1158
NL: burnout complaints vs diagnosed burnout	20.1% vs 1.6% (same survey, same year)	CBS/TNO NEA
HEALTH OUTCOMES
Job strain → CHD	HR 1.23 [1.10, 1.37]; PAR 3.4%; published 1.43 vs unpublished 1.16	Kivimäki 2012, Lancet 380:1491–1497
≥55 h/week → stroke / CHD	RR 1.33 [1.11, 1.61] / 1.13 [1.02, 1.26]	Kivimäki 2015, Lancet 386:1739–1746
Burnout → CVD (pooled)	OR 1.21 [1.03, 1.39]; CHD subgroup 1.79 [0.79, 2.79] ns	John 2024, Front Psychiatry 15:1326745
Job strain → type 2 diabetes	1.26 → 1.12 [0.99, 1.26] fully adjusted — null	Nyberg 2014, Diabetes Care 37:2268–2275
Burnout → all-cause mortality	+35% per unit, under age 45 only; null ≥45	Ahola 2010, J Psychosom Res 69:51–57
Deaths from long working hours (2016)	745,194 [705,786–784,601]; 23.3M DALYs	Pega 2021, Environ Int 154:106595
Mistreatment → burnout (surgical residents)	OR 2.94 [2.58, 3.36], N = 7,409	Hu 2019, NEJM 381:1741–1752
Gender gap in burnout after adjusting for mistreatment	1.33 → 0.90 [0.80, 1.00]	Hu 2019
RECOVERY
Still clinically exhausted at 18 months	33%	Glise 2012, BMC Psychiatry 12:18
Reporting full recovery at 7 years	16%; 46% extreme fatigue; 73% reduced stress tolerance	Glise 2020, BMC Psychology 8:26
Not on sick leave at 7 years	87%	Glise 2020
Changed workplace at 10 years	73% (vs 25.5% population base rate over 5 y)	Eskilsson 2024, BMC Psychiatry 24:525
Recurrent sickness absence within 12 months of RTW	27.9%	Deprez 2026, J Occup Rehabil
RTW intervention → full return to work	OR 1.33 [0.59, 2.98] — ns	Perski 2017, Scand J Psychol 58:551–561
INTERVENTIONS
CBT-I on insomnia severity / on total sleep time	g = 0.98 / g = 0.16	van Straten 2018, Sleep Med Rev 38:3–16
Exercise on depression (umbrella / vs active controls)	SMD −0.43 / g −0.42 to −0.62	Singh 2023 BJSM; Noetel 2024 BMJ 384:e075847
Resistance training on depression	Δ 0.66 [0.48, 0.83]	Gordon 2018, JAMA Psychiatry 75:566–576
Exercise on energy/fatigue with placebo control	No effect	Puetz 2006, Psychol Bull 132:866–876
Behavioural activation	SMD 0.67 [0.54, 0.80]; self-guided 0.36	Cuijpers 2026, Clin Psychol Rev 128:102783
Mindfulness vs evidence-based treatment	d = −0.004 [−0.15, 0.14]	Goldberg 2018, Clin Psychol Rev
Mindfulness adverse events (observational studies)	33.2% [25–41%]	Farias 2020, Acta Psychiatr Scand
MBSR vs escitalopram for anxiety	Noninferior (CGI-S diff −0.07 [−0.38, 0.23])	Hoge 2023, JAMA Psychiatry
Burnout interventions, overall	SMD −0.29 [−0.42, −0.16] ≈ 3 MBI points	Panagioti 2017, JAMA Intern Med 177:195–205
Coaching → depersonalisation (moderate certainty)	−0.30 [−0.42, −0.19]	Collett 2026, Ann Intern Med 179:51–66
Workplace wellness RCT	2 of 80 outcomes, both self-reported, N = 32,974	Song & Baicker 2019, JAMA 321:1491–1501
STAR work-redesign RCT	+8 min sleep/night; null in nursing homes; null cardiometabolic	Olson 2015; Marino 2016; Berkman 2023
Micro-breaks on vigour / on performance	d = 0.36 [0.16, 0.55] / 0.16 [−0.04, 0.37] ns	Albulescu 2022, PLoS ONE 17:e0272460
Vacation effect / fade-out	+0.43 / −0.38 (7 studies); faded within 1 month	de Bloom 2009; Kühnel & Sonnentag 2011
Caffeine on total sleep time	−45.3 min [29.0, 61.5]; 2.8 min recovered per hour earlier	Gardiner 2023, Sleep Med Rev 69:101764
Alcohol on REM proportion / total sleep time	−2.8% [−3.9, −1.7] / null	Gardiner 2025, Sleep Med Rev 80:102030
Ashwagandha on perceived stress (vs on cortisol)	SMD −0.355, p = 0.40 (null) vs cortisol −1.16 µg/dL	Albalawi 2025, Nutr Health 31:1395
Rhodiola on vitality	Placebo better, MD −17.3 [−30.6, −3.9]	Punja 2014, PLoS ONE 9:e108416
Breathwork on stress	g = −0.35 [−0.55, −0.14]	Fincham 2023, Sci Rep 13:432
App for stress vs placebo app	g = 0.09 [−0.05, 0.24] ns	Linardon 2019, World Psychiatry 18:325–336
App real-world 30-day retention	3.3% (breathing apps: 0.0%)	Baumel 2019, JMIR 21:e14567
Serious somatic disease in fatigue presentations	4.3% [2.7–6.7] — no different from non-fatigued controls	Stadje 2016, BMC Fam Pract 17:147
Depression in fatigue presentations	18.5% [16.2–21.0]	Stadje 2016
Social relationships → survival	OR 1.50 [1.42, 1.59], N = 308,849	Holt-Lunstad 2010, PLoS Med 7:e1000316
Loneliness interventions (RCTs only)	−0.198 [−0.32, −0.08]; 6 of 20 RCTs efficacious	Masi 2011, Pers Soc Psychol Rev 15:219–266

---

Appendix B — Master reference list

Frameworks and stress biology

1. Sapolsky RM, Romero LM, Munck AU (2000). How do glucocorticoids influence stress responses? Endocr Rev 21:55–89. https://pubmed.ncbi.nlm.nih.gov/10696570/
2. de Kloet ER, Joëls M, Holsboer F (2005). Stress and the brain: from adaptation to disease. Nat Rev Neurosci 6:463–475. https://doi.org/10.1038/nrn1683
3. McEwen BS, Stellar E (1993). Stress and the individual. Arch Intern Med 153:2093–2101. https://pubmed.ncbi.nlm.nih.gov/8379800/
4. McEwen BS (1998). Protective and damaging effects of stress mediators. NEJM 338:171–179. https://pubmed.ncbi.nlm.nih.gov/9428819/
5. Miller GE, Chen E, Zhou ES (2007). If it goes up, must it come down? Psychol Bull 133:25–45. https://pubmed.ncbi.nlm.nih.gov/17201569/
6. Cadegiani FA, Kater CE (2016). Adrenal fatigue does not exist: a systematic review. BMC Endocr Disord 16:48. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4997656/
7. Adam EK et al. (2017). Diurnal cortisol slopes and mental and physical health outcomes. Psychoneuroendocrinology 83:25–41. https://pubmed.ncbi.nlm.nih.gov/28578301/
8. Chida Y, Steptoe A (2009). Cortisol awakening response and psychosocial factors: a systematic review and meta-analysis. Biol Psychol 80:265–278. https://pubmed.ncbi.nlm.nih.gov/19022335/
9. Stalder T et al. (2025). The cortisol awakening response: regulation and functional significance. Endocr Rev 46:43–59. https://doi.org/10.1210/endrev/bnae024
10. Seeman TE et al. (2001). Allostatic load as a marker of cumulative biological risk. PNAS 98:4770–4775. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC31909/
11. Parker HW et al. (2022). Allostatic load and mortality: systematic review and meta-analysis. Am J Prev Med 63:131–140. https://pubmed.ncbi.nlm.nih.gov/35393143/
12. Thayer JF et al. (2012). A meta-analysis of heart rate variability and neuroimaging studies. Neurosci Biobehav Rev 36:747–756. https://pubmed.ncbi.nlm.nih.gov/22178086/
13. Jarczok MN et al. (2022). Heart rate variability in the prediction of mortality. Neurosci Biobehav Rev 143:104907. https://pubmed.ncbi.nlm.nih.gov/36243195/
14. Rothe N et al. (2020). Examination of peripheral basal and reactive cortisol levels in MDD and burnout. Neurosci Biobehav Rev 114:232–270. https://pubmed.ncbi.nlm.nih.gov/32088345/

Bioenergetics and mitochondria

15. Picard M, McEwen BS (2018). Psychological stress and mitochondria: a conceptual framework. Psychosom Med 80:126–140. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5901651/
16. Picard M, McEwen BS (2018). Psychological stress and mitochondria: a systematic review. Psychosom Med 80:141–153. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5901654/
17. Picard M et al. (2015). Mitochondrial functions modulate neuroendocrine, metabolic, inflammatory and transcriptional responses to acute psychological stress. PNAS 112:E6614–23. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4672794/
18. Du J et al. (2009). Dynamic regulation of mitochondrial function by glucocorticoids. PNAS 106:3543–3548. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2637276/
19. Bobba-Alves N, Juster R-P, Picard M (2022). The energetic cost of allostasis and allostatic load. Psychoneuroendocrinology 105951. https://doi.org/10.1016/j.psyneuen.2022.105951
20. Bobba-Alves N et al. (2023). Cellular allostatic load is linked to increased energy expenditure and accelerated biological aging. Psychoneuroendocrinology 155:106322. https://europepmc.org/article/MED/37423094
21. Trumpff C et al. (2024). Psychosocial experiences are associated with human brain mitochondrial biology. PNAS 121:e2317673121. https://europepmc.org/article/MED/38889126
22. Trumpff C et al. (2025). Effects of acute psychological stress on blood cf-mtDNA: a crossover experimental study. Psychoneuroendocrinology 182:107644. https://pubmed.ncbi.nlm.nih.gov/41129907/

Sleep

23. Van Dongen HPA et al. (2003). The cumulative cost of additional wakefulness. Sleep 26:117–126. https://pubmed.ncbi.nlm.nih.gov/12683469/
24. Spiegel K, Leproult R, Van Cauter E (1999). Impact of sleep debt on metabolic and endocrine function. Lancet 354:1435–1439. https://pubmed.ncbi.nlm.nih.gov/10543671/
25. Kim E-J, Dimsdale JE (2007). The effect of psychosocial stress on sleep: a review of polysomnographic evidence. Behav Sleep Med 5:256–278. https://pubmed.ncbi.nlm.nih.gov/17937582/
26. Friess E et al. (1994). Effects of pulsatile cortisol infusion on sleep-EEG and nocturnal growth hormone release. J Sleep Res 3:73–79. https://europepmc.org/article/MED/10607111
27. Friedman TC et al. (1994). Decreased delta-sleep and plasma delta-sleep-inducing peptide in patients with Cushing syndrome. Neuroendocrinology 60:626–634. https://europepmc.org/article/MED/7700506
28. Shipley JE et al. (1992). Sleep architecture and sleep apnea in patients with Cushing's disease. Biol Psychiatry 32:146–155. https://europepmc.org/article/MED/1330006
29. Ju Y-ES et al. (2017). Slow wave sleep disruption increases cerebrospinal fluid amyloid-β levels. Brain 140:2104–2111. https://academic.oup.com/brain/article/140/8/2104/3933862
30. Nicolaides NC, Vgontzas AN, Kritikou I, Chrousos G (2020). HPA axis and sleep. Endotext. https://www.ncbi.nlm.nih.gov/books/NBK279071/

Inflammation

31. Cohen S et al. (2012). Chronic stress, glucocorticoid receptor resistance, inflammation, and disease risk. PNAS 109:5995–5999. https://www.pnas.org/doi/10.1073/pnas.1118355109
32. Dantzer R et al. (2008). From inflammation to sickness and depression. Nat Rev Neurosci 9:46–56. https://www.nature.com/articles/nrn2297
33. Felger JC, Treadway MT (2017). Inflammation effects on motivation and motor activity. Neuropsychopharmacology 42:216–241. https://www.nature.com/articles/npp2016143
34. Lacourt TE et al. (2018). The high costs of low-grade inflammation. Front Behav Neurosci 12:78. https://www.frontiersin.org/articles/10.3389/fnbeh.2018.00078/full
35. Baumeister D et al. (2016). Childhood trauma and adulthood inflammation: a meta-analysis. Mol Psychiatry 21:642–649. https://www.nature.com/articles/mp201567

Cognition and effort

36. Arnsten AFT (2009). Stress signalling pathways that impair prefrontal cortex structure and function. Nat Rev Neurosci 10:410–422. https://www.nature.com/articles/nrn2648
37. Arnsten AFT (2015). Stress weakens prefrontal networks. Nat Neurosci 18:1376–1385. https://www.nature.com/articles/nn.4087
38. Hermans EJ et al. (2014). Dynamic adaptation of large-scale brain networks in response to acute stressors. Trends Neurosci 37:304–314. https://pubmed.ncbi.nlm.nih.gov/24766931/
39. Shields GS, Sazma MA, Yonelinas AP (2016). The effects of acute stress on core executive functions: a meta-analysis. Neurosci Biobehav Rev. https://pubmed.ncbi.nlm.nih.gov/27371161/
40. Shields GS et al. (2017). The effects of acute stress on episodic memory: a meta-analysis. Psychol Bull 143:636–675. https://escholarship.org/uc/item/7ds6x4h2
41. Gavelin HM et al. (2022). Cognitive function in clinical burnout: systematic review and meta-analysis. Work & Stress 36:86–104. https://doi.org/10.31234/osf.io/n2htg
42. Kim EJ, Kim JJ (2023). Neurocognitive effects of stress: a metaparadigm perspective. Mol Psychiatry 28:2750–2763. https://www.nature.com/articles/s41380-023-01986-4
43. Hagger MS et al. (2016). A multilab preregistered replication of the ego-depletion effect. Perspect Psychol Sci 11:546–573.
44. Vohs KD et al. (2021). A multisite preregistered paradigmatic test of the ego-depletion effect. Psychol Sci 32:1566–1581. https://www.asc.upenn.edu/sites/default/files/2022-06/Vohs%20et%20al%202021.pdf
45. Wiehler A et al. (2022). A neuro-metabolic account of why daylong cognitive work alters the control of economic decisions. Curr Biol 32:3564–3575.e5. https://pubmed.ncbi.nlm.nih.gov/35961314/
46. Pessiglione M, Blain B, Wiehler A, Naik S (2025). Origins and consequences of cognitive fatigue. Trends Cogn Sci 29:730–749. https://pubmed.ncbi.nlm.nih.gov/40169294/
47. Westbrook A, Braver TS (2016). Dopamine does double duty in motivating cognitive effort. Neuron 89:695–710. https://www.cell.com/neuron/fulltext/S0896-6273(15)01131-9
48. Salamone JD, Correa M (2024). The neurobiology of activational aspects of motivation. Annu Rev Psychol 75:1–32. https://www.annualreviews.org/content/journals/10.1146/annurev-psych-020223-012208
49. Kurzban R, Duckworth A, Kable JW, Myers J (2013). An opportunity cost model of subjective effort and task performance. Behav Brain Sci 36:661–679. https://gwern.net/doc/psychology/willpower/2013-kurzban.pdf
50. Forbes PAG et al. (2024). No effects of acute stress on monetary delay discounting. Neurobiol Stress 31:100653. https://pmc.ncbi.nlm.nih.gov/articles/PMC11201353/
51. Zwosta K et al. (2025). No evidence for increased habitual or decreased goal-directed action control after acute stress. PLoS ONE 20:e0327807. https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0327807
52. Carvalho LS, Meier S, Wang SW (2016). Poverty and economic decision-making. Am Econ Rev 106:260–284. https://sites.pitt.edu/~swwang/papers/Payday.pdf
53. O'Donnell M et al. (2021). Empirical audit and review… psychological consequences of scarcity. PNAS 118:e2103313118. https://pubmed.ncbi.nlm.nih.gov/34711679/

Real-world function

54. Hodkinson A et al. (2022). Associations of physician burnout with career engagement and quality of patient care. BMJ 378:e070442. https://pubmed.ncbi.nlm.nih.gov/36104064/
55. Dingus TA et al. (2016). Driver crash risk factors and prevalence evaluation using naturalistic driving data. PNAS 113:2636–2641. https://www.pnas.org/doi/10.1073/pnas.1513271113
56. Nahrgang JD, Morgeson FP, Hofmann DA (2011). Safety at work: a meta-analytic investigation. J Appl Psychol 96:71–94.
57. Neff LA, Karney BR (2009). Stress and reactivity to daily relationship experiences. J Pers Soc Psychol 97:435–450.
58. Caviola S et al. (2022). Math performance and academic anxiety forms. Educ Psychol Rev 34:363–399. https://link.springer.com/article/10.1007/s10648-021-09618-5

Burnout, taxonomy and trajectory

59. WHO (2019). Burn-out an "occupational phenomenon": ICD-11. https://www.who.int/news/item/28-05-2019-burn-out-an-occupational-phenomenon-international-classification-of-diseases
60. Maslach C, Schaufeli WB, Leiter MP (2001). Job burnout. Annu Rev Psychol 52:397–422.
61. Schaufeli WB, Desart S, De Witte H (2020). Burnout Assessment Tool (BAT). IJERPH 17:9495. https://www.mdpi.com/1660-4601/17/24/9495
62. Rotenstein LS et al. (2018). Prevalence of burnout among physicians: a systematic review. JAMA 320:1131–1150. https://pubmed.ncbi.nlm.nih.gov/30326495/
63. Bianchi R et al. (2021). Is burnout a depressive condition? Clin Psychol Sci 9(4). https://journals.sagepub.com/doi/10.1177/2167702620979597
64. Verkuilen J, Bianchi R, Schonfeld IS, Laurent E (2021). Burnout–depression overlap: exploratory structural equation modeling bifactor analysis. Assessment 28:1583–1600. https://pubmed.ncbi.nlm.nih.gov/32153199/
65. Lindsäter E et al. (2022). Exhaustion disorder: scoping review. BJPsych Open 8(6). https://doi.org/10.1192/bjo.2022.559
66. Sennerstam V et al. (2025). Exhaustion disorder in primary care. Scand J Psychol. https://doi.org/10.1111/sjop.13087
67. Huibers MJH et al. (2003). Fatigue, burnout and chronic fatigue syndrome among employees on sick leave. Occup Environ Med 60(Suppl 1):i26–i31. https://pubmed.ncbi.nlm.nih.gov/12782744/
68. Mäkikangas A, Kinnunen U (2016). The person-oriented approach to burnout: a systematic review. Burnout Research 3:11–23.
69. Guthier C, Dormann C, Voelkle MC (2020). Reciprocal effects between job stressors and burnout. Psychol Bull 146:1146–1173.
70. Sonnentag S (2018). The recovery paradox. Res Organ Behav 38:169–185. https://www.helsinki.fi/assets/drupal/2024-04/Sonnentag.pdf
71. Wendsche J, Lohmann-Haislah A (2017). A meta-analysis on antecedents and outcomes of detachment from work. Front Psychol 7:2072. https://www.frontiersin.org/articles/10.3389/fpsyg.2016.02072/full
72. Bennett AA, Bakker AB, Field JG (2018). Recovery from work-related effort: a meta-analysis. J Organ Behav 39:262–275.
73. Danhof-Pont MB, van Veen T, Zitman FG (2011). Biomarkers in burnout: a systematic review. J Psychosom Res 70:505–524. https://pubmed.ncbi.nlm.nih.gov/21624574/
74. Niedhammer I, Bertrais S, Witt K (2021). Psychosocial work exposures and health outcomes: a meta-review of 72 reviews. Scand J Work Environ Health 47:489–508. https://doi.org/10.5271/sjweh.3968
75. Shanafelt TD et al. (2025). Changes in burnout and satisfaction with work-life integration, 2011–2023. Mayo Clin Proc 100:1142–1158. https://doi.org/10.1016/j.mayocp.2024.11.031
76. Glise K, Ahlborg G, Jonsdottir IH (2012). Course of mental symptoms in patients with stress-related exhaustion. BMC Psychiatry 12:18.
77. Glise K, Wiegner L, Jonsdottir IH (2020). Long-term follow-up of residual symptoms in patients treated for stress-related exhaustion. BMC Psychology 8:26. https://bmcpsychology.biomedcentral.com/articles/10.1186/s40359-020-0395-8
78. Eskilsson T et al. (2024). Ten-year follow-up of patients with stress-related exhaustion disorder. BMC Psychiatry 24:525.
79. Österberg K, Skogsliden S, Karlson B (2014). Neuropsychological sequelae of work-stress-related exhaustion. Stress 17:59–69.
80. Aronsson G et al. (2017). A systematic review including meta-analysis of work environment and burnout symptoms. BMC Public Health 17:264. https://pmc.ncbi.nlm.nih.gov/articles/PMC5356239/
81. Hu Y-Y et al. (2019). Discrimination, abuse, harassment, and burnout in surgical residency training. NEJM 381:1741–1752. https://pubmed.ncbi.nlm.nih.gov/31657887/
82. Hill AP, Curran T (2016). Multidimensional perfectionism and burnout: a meta-analysis. Pers Soc Psychol Rev 20:269–288.
83. Kivimäki M et al. (2012). Job strain as a risk factor for coronary heart disease. Lancet 380:1491–1497. https://pubmed.ncbi.nlm.nih.gov/22981903/
84. Kivimäki M et al. (2015). Long working hours and risk of coronary heart disease and stroke. Lancet 386:1739–1746. https://pubmed.ncbi.nlm.nih.gov/26298822/
85. Pega F et al. (2021). Global, regional and national burdens of ischaemic heart disease and stroke attributable to long working hours. Environ Int 154:106595. https://pmc.ncbi.nlm.nih.gov/articles/PMC8204267/
86. Perski O et al. (2017). Interventions and return to work after stress-related mental disorders. Scand J Psychol 58:551–561.

Interventions

87. Edinger JD et al. (2021). Behavioral and psychological treatments for chronic insomnia disorder in adults: AASM clinical practice guideline. J Clin Sleep Med 17:255–262. https://doi.org/10.5664/jcsm.8986
88. van Straten A et al. (2018). Cognitive and behavioral therapies in the treatment of insomnia: a meta-analysis. Sleep Med Rev 38:3–16. https://doi.org/10.1016/j.smrv.2017.02.001
89. Singh B et al. (2023). Effectiveness of physical activity interventions for improving depression, anxiety and distress. Br J Sports Med. https://pubmed.ncbi.nlm.nih.gov/36796860/
90. Noetel M et al. (2024). Effect of exercise for depression: systematic review and network meta-analysis. BMJ 384:e075847. https://pmc.ncbi.nlm.nih.gov/articles/PMC10870815/
91. Puetz TW, O'Connor PJ, Dishman RK (2006). Effects of chronic exercise on feelings of energy and fatigue. Psychol Bull 132:866–876. https://pubmed.ncbi.nlm.nih.gov/17073524/
92. Cuijpers P et al. (2026). Behavioral activation for depression. Clin Psychol Rev 128:102783. https://pubmed.ncbi.nlm.nih.gov/42492146/
93. Richardson KM, Rothstein HR (2008). Effects of occupational stress management intervention programs: a meta-analysis. J Occup Health Psychol 13:69–93.
94. Panagioti M et al. (2017). Controlled interventions to reduce burnout in physicians. JAMA Intern Med 177:195–205. https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2588814
95. West CP et al. (2016). Interventions to prevent and reduce physician burnout. Lancet 388:2272–2281. https://pubmed.ncbi.nlm.nih.gov/27692469/
96. Collett G et al. (2026). Interventions to reduce burnout in physicians. Ann Intern Med 179:51–66. https://pubmed.ncbi.nlm.nih.gov/41248499/
97. Ruotsalainen JH et al. Preventing occupational stress in healthcare workers. Cochrane Database Syst Rev CD002892. https://www.cochranelibrary.com/cdsr/doi/10.1002/14651858.CD002892.pub5/references
98. Goyal M et al. (2014). Meditation programs for psychological stress and well-being. JAMA Intern Med 174:357–368. https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/1809754
99. Goldberg SB et al. (2018). Mindfulness-based interventions for psychiatric disorders: a systematic review and meta-analysis. Clin Psychol Rev 59:52–60. https://pmc.ncbi.nlm.nih.gov/articles/PMC5741505/
100. Van Dam NT et al. (2018). Mind the hype. Perspect Psychol Sci 13:36–61. https://pure.rug.nl/ws/files/56912114/Mind_the_Hype.pdf
101. Galante J et al. (2021). Mindfulness-based programmes in non-clinical settings. PLoS Med 18:e1003481. https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1003481
102. Hoge EA et al. (2023). Mindfulness-based stress reduction vs escitalopram for anxiety disorders. JAMA Psychiatry. https://jamanetwork.com/journals/jamapsychiatry/fullarticle/2798510
103. Albulescu P et al. (2022). "Give me a break!" A meta-analysis on micro-breaks. PLoS ONE 17:e0272460. https://journals.plos.org/plosone/article?id=10.1371%2Fjournal.pone.0272460
104. de Bloom J et al. (2009). Do we recover from vacation? J Occup Health 51:13–25. https://pubmed.ncbi.nlm.nih.gov/19096200/
105. Song Z, Baicker K (2019). Effect of a workplace wellness program on employee health and economic outcomes: a randomized clinical trial. JAMA 321:1491–1501. https://pubmed.ncbi.nlm.nih.gov/30990549/
106. Jones D, Molitor D, Reif J (2019). What do workplace wellness programs do? Q J Econ 134:1747–1791. https://academic.oup.com/qje/article/134/4/1747/5550759
107. Fleming WJ (2024). Employee well-being outcomes from individual-level mental health interventions: cross-sectional evidence from the United Kingdom. Ind Relat J 55:162–182. https://doi.org/10.1111/irj.12418
108. Egan M et al. (2007). The psychosocial and health effects of workplace reorganisation. J Epidemiol Community Health 61:945–954. https://pmc.ncbi.nlm.nih.gov/articles/PMC2465601/
109. Holt-Lunstad J, Smith TB, Layton JB (2010). Social relationships and mortality risk. PLoS Med 7:e1000316. https://journals.plos.org/plosmedicine/article?id=10.1371%2Fjournal.pmed.1000316
110. Masi CM et al. (2011). A meta-analysis of interventions to reduce loneliness. Pers Soc Psychol Rev 15:219–266.
111. Gardiner C et al. (2023). The effect of caffeine on subsequent sleep: a meta-analysis. Sleep Med Rev 69:101764. https://doi.org/10.1016/j.smrv.2023.101764
112. Gardiner C et al. (2025). The effect of alcohol on subsequent sleep: a meta-analysis. Sleep Med Rev 80:102030. https://doi.org/10.1016/j.smrv.2024.102030
113. Marchi M et al. (2025). Withania somnifera for stress and anxiety: systematic review and meta-analysis. BJPsych Open. https://pmc.ncbi.nlm.nih.gov/articles/PMC12569615/
114. Punja S et al. (2014). Rhodiola rosea for physical and mental fatigue: a randomized controlled trial. PLoS ONE 9:e108416.
115. Fincham GW et al. (2023). Effect of breathwork on stress and mental health. Sci Rep 13:432. https://www.nature.com/articles/s41598-022-27247-y
116. Linardon J et al. (2019). The efficacy of app-supported smartphone interventions for mental health problems. World Psychiatry 18:325–336. https://onlinelibrary.wiley.com/doi/full/10.1002/wps.20673
117. Baumel A et al. (2019). Objective user engagement with mental health apps. J Med Internet Res 21:e14567. https://www.jmir.org/2019/9/e14567/
118. Stadje R et al. (2016). The differential diagnosis of tiredness: a systematic review. BMC Fam Pract 17:147. https://pmc.ncbi.nlm.nih.gov/articles/PMC5072300/
119. Cipriani A et al. (2018). Comparative efficacy and acceptability of 21 antidepressant drugs. Lancet 391:1357–1366. https://pmc.ncbi.nlm.nih.gov/articles/PMC5889788/

Additional inspectable records

These close the Appendix B gaps named in review and the remaining findings-table citations that could be matched to a stable record without guessing.

120. Maslach C, Jackson SE (1981). The measurement of experienced burnout. J Organ Behav 2:99–113. https://doi.org/10.1002/job.4030020205
121. Leproult R, Copinschi G, Buxton O, Van Cauter E (1997). Sleep loss results in an elevation of cortisol levels the next evening. Sleep 20:865–870. https://doi.org/10.1093/sleep/20.10.865
122. Van Cauter E, Plat L, Copinschi G (1998). Interrelations between sleep and the somatotropic axis. Sleep 21:553–566. https://doi.org/10.1093/sleep/21.6.553
123. Seematter G et al. (2000). Effects of mental stress on insulin-mediated glucose metabolism and energy expenditure in lean and obese women. Am J Physiol Endocrinol Metab 279:E799–E805. https://doi.org/10.1152/ajpendo.2000.279.4.E799
124. Reichenberg A et al. (2001). Cytokine-associated emotional and cognitive disturbances in humans. Arch Gen Psychiatry 58:445–452. https://doi.org/10.1001/archpsyc.58.5.445
125. Ahola K et al. (2010). Burnout as a predictor of all-cause mortality among industrial employees. J Psychosom Res 69:51–57. https://doi.org/10.1016/j.jpsychores.2010.01.002
126. Nyberg ST et al. (2014). Job strain as a risk factor for type 2 diabetes. Diabetes Care 37:2268–2275. https://doi.org/10.2337/dc13-2936
127. Bolton JL et al. (2014). Genome wide association identifies common variants at the SERPINA6/SERPINA1 locus influencing plasma cortisol. PLoS Genet 10:e1004474. https://doi.org/10.1371/journal.pgen.1004474
128. Olson R et al. (2015). A workplace intervention improves sleep: results from the randomized controlled Work, Family, and Health Study. Sleep Health 1:55–65. https://doi.org/10.1016/j.sleh.2014.11.003
129. Miller AH, Raison CL (2016). The role of inflammation in depression: from evolutionary imperative to modern treatment target. Nat Rev Immunol 16:22–34. https://doi.org/10.1038/nri.2015.5
130. Evans-Lacko S, Knapp M (2016). Global patterns of workplace productivity for people with depression: absenteeism and presenteeism costs across eight diverse countries. Soc Psychiatry Psychiatr Epidemiol 51:1525–1537. https://doi.org/10.1007/s00127-016-1278-4
131. Marsland AL, Walsh C, Lockwood K, John-Henderson NA (2017). The effects of acute psychological stress on circulating and stimulated inflammatory markers: a systematic review and meta-analysis. Brain Behav Immun 64:208–219. https://doi.org/10.1016/j.bbi.2017.01.011
132. Gordon BR et al. (2018). Association of efficacy of resistance exercise training with depressive symptoms. JAMA Psychiatry 75:566–576. https://doi.org/10.1001/jamapsychiatry.2018.0572
133. Ackermann S et al. (2019). Psychosocial stress before a nap increases sleep latency and decreases early slow-wave activity. Front Psychol 10:20. https://doi.org/10.3389/fpsyg.2019.00020
134. Farias M, Maraldi E, Wallenkampf KC, Lucchetti G (2020). Adverse events in meditation practices and meditation-based therapies: a systematic review. Acta Psychiatr Scand 142:374–393. https://doi.org/10.1111/acps.13225
135. Ensari I et al. (2020). Testing the cross-stressor hypothesis under real-world conditions. J Behav Med 43:989–1001. https://doi.org/10.1007/s10865-020-00155-0
136. Wilson PB et al. (2020). Life stress and background anxiety are not associated with resting metabolic rate in healthy adults. Appl Physiol Nutr Metab 45:812–816. https://doi.org/10.1139/apnm-2019-0875
137. Okereke OI et al. (2020). Effect of long-term vitamin D3 supplementation vs placebo on risk of depression or clinically relevant depressive symptoms. JAMA 324:471–480. https://doi.org/10.1001/jama.2020.10224
138. Markun S et al. (2021). Effects of vitamin B12 supplementation on cognitive function, depressive symptoms, and fatigue: a systematic review, meta-analysis, and meta-regression. Nutrients 13:923. https://doi.org/10.3390/nu13030923
139. Corbeanu A et al. (2023). The link between burnout and job performance: a meta-analysis. Eur J Work Organ Psychol 32:599–616. https://doi.org/10.1080/1359432X.2023.2209320
140. Berkman LF et al. (2023). Employee cardiometabolic risk following a cluster-randomized workplace intervention from the Work, Family and Health Network, 2009–2015. Am J Public Health 113:1322–1331. https://doi.org/10.2105/AJPH.2023.307413
141. John A, Bouillon-Minois JB, Bagheri R, Pélissier C, Charbotel B, Llorca PM (2024). The influence of burnout on cardiovascular disease: a systematic review and meta-analysis. Front Psychiatry 15:1326745. https://doi.org/10.3389/fpsyt.2024.1326745
142. Martinez B, Rutledge T, Conley CS, Mahrer NE, Guinto R (2025). The health and economic burden of employee burnout to U.S. employers. Am J Prev Med 68:645–655. https://doi.org/10.1016/j.amepre.2025.01.011
143. Albalawi KS et al. (2025). Dual impact of ashwagandha: significant cortisol reduction but no effects on perceived stress. Nutr Health 31:1395–1408. https://doi.org/10.1177/02601060251363647
144. Deprez A et al. (2026). Determinants of sustainable return to work after burnout or depression: a longitudinal cohort study. J Occup Rehabil. https://doi.org/10.1007/s10926-026-10407-6

---

Appendix C — Register of unverified and corrected claims

Claims that could not be traced to an accessible primary source during preparation. None is used as a load-bearing figure in this review; they are listed so that nobody launders them onward.

Claim	Status
"Burnout does not present the unity expected of a distinct syndrome" attributed to Bianchi et al. (2021)	Misattributed. Zero full-text hits in Europe PMC. The Bianchi group's phrase "no syndromal unity" is from Verkuilen et al. (2021, Assessment). The CPS 2021 conclusion is the weaker "burnout problematically overlaps with depression."
Bianchi et al. (2021) CI [.75, .84]	Not in the abstract; body paywalled. Point estimate r = .80 verified.
"Acute stress enhances response inhibition"	Overstated. Overall inhibition effect null (g = −0.076); +0.296 is a post-hoc moderator on ~5 df.
"Cortisol destroys deep sleep is backwards"	Bounded. True for acute exogenous cortisol in healthy volunteers only; chronic hypercortisolism reduces delta sleep, and ACTH-independent Cushing's still disrupts sleep.
Stadje et al. (2016) "identical" somatic disease prevalence	Softened. Only 6 of 26 studies had controls; support rests on two single studies, one powered at 0.09. Abstract gives 4.3%, discussion gives 3.1%.
Fleming (2024) — propensity-score matching; "volunteering the only positive signal"; "mindfulness negative"	Unverified. Design confirmed cross-sectional; per-intervention details not in the abstract and full text inaccessible.
ICD-11 code "QD85"	Not stated on the WHO news release; sourced from the ICD-11 browser. Note also that burnout appeared in ICD-10 in the same category.
Fan et al. (2025, Nat Hum Behav) four-day-week effect sizes	Abstract gives direction only; no coefficient or CI obtainable.
West et al. (2016) subgroup point estimates	The paper publishes only an interaction p-value. Any source quoting subgroup effect sizes is fabricating them.
Lehrer et al. (2020) HRV-biofeedback pooled effect sizes	Closed access; no OA copy located.
van Straten (2018), Linardon (2024), Moshe (2021), Bennett (2018) confidence intervals	Not in abstracts; not verified.
Social-support buffering meta-analyses (Viswesvaran 1999; Häusser 2010)	Inaccessible.
Starcke & Brand (2016) stress–risk pooled effect sizes	Full text inaccessible; a published correction exists.
Pooled effect size for stress → vigilance decrement or mind-wandering	No verified meta-analysis located.
WHO "US$1 trillion in lost productivity"	WHO fact sheet provides no source citation or methodology. Cite as a fact sheet, not a peer-reviewed estimate.
WHO "871,000 loneliness deaths per year"	Figure verified as WHO's claim; derivation not traced to a methods annex.
€617bn "cost of work-related stress"	Mis-citation. It is an estimate of work-related depression (Matrix 2013 for EAHC), and the report states verbatim that "Stress was not used." Better-founded work-attributable estimate: €44.7–103.1bn.
"Loneliness equals 15 cigarettes a day"	Not a figure in Holt-Lunstad's papers.
Post-exertional malaise as a validated discriminator between burnout and ME/CFS	Reflects criteria structure; no head-to-head validation study located.
Alsanie et al. (2026) ashwagandha SMD −5.9 to −6.9	Biologically implausible; almost certainly raw mean differences mislabelled. Do not cite.
Goode, Sturm & Picard (2026) norepinephrine/glycolysis figures	Preprint, not peer-reviewed.
Deane 2021 (omega-3 depression-prevention RR 1.01, N = 41,470); van der Mee 2023 (cross-stressor N = 116); Marino 2016 (STAR nursing-home arm)	Bibliographic line not attached. Numbers stay as reported in the findings table; a stable full-text record was not matched at ship without guessing the paper.

---

Prepared 26 July 2026. Four domain dossiers (biology, cognition, burnout trajectory, interventions) and two adversarial verification passes underlie this synthesis; the source dossiers are available separately.