Published on 24 September 2026
Cardiorespiratory fitness and later risk of mental disorders: a steady association, an unclear direction
Nature Mental Health · 2026, volume 4, issue 4, pages 653 to 663 · Díaz-Goñi et al.
DOI 10.1038/s44220-026-00599-4
Scientific 68
Editorial 83
The essentials
A systematic review with meta-analysis pools 27 cohort studies and 4,007,638 participants, 30.5% of them women, across nine countries, to examine the link between cardiorespiratory fitness and the later onset of mental disorders or dementia. Twenty-two of these studies enter the pooled analyses. Compared with participants of low fitness, adults of high fitness show a depression risk reported at HR 0.64 (95% CI 0.56 to 0.74, 9 studies, I² 67.9%), an all-cause dementia risk of 0.61 (95% CI 0.55 to 0.68, 7 studies, I² 0%), and a psychotic disorder risk of 0.71 (95% CI 0.65 to 0.77, 2 studies, both all-male). For anxiety disorders, the pooled estimate is 0.90 (95% CI 0.75 to 1.09, 2 studies) and does not cross the threshold of significance. In continuous analysis, each additional MET, that is 3.5 ml/kg/min, is associated with an HR of 0.95 (95% CI 0.92 to 0.98) for depression and 0.81 (95% CI 0.67 to 0.98) for all-cause dementia, the latter estimate carrying a heterogeneity of 97.3% and a prediction interval running from 0.07 to 9.12. The GRADE rating places only one result at the moderate level, all-cause dementia in the high-versus-low comparison; every other result is graded very low. The strength of this work lies in the nature of the exposure: cardiorespiratory fitness is a quantified capacity, not a self-reported behavior, which in principle shields it from the reporting bias that weakens most work on physical activity. The limitation is not a matter of sample size but of study design: these are cohort studies, and reverse causation remains the most serious competing explanation, since a patient sliding into depression moves less and sees their exercise capacity decline well before the diagnosis is made. The authors themselves treat this as the main obstacle to interpreting their own results.
Scope of this analysis. It was prepared from the full text of the publication, consulted on 13 August 2026: the body of the article, Table 1, Figures 1 to 3, the methods, the funding statement, the conflict-of-interest statement, and the Nature Portfolio reporting summary attached to the article. The supplementary material, Supplementary Tables 1 to 10, Supplementary Figures 1 to 4, and the supplementary methods, did not accompany the file and was not consulted. What exists only in that material, notably the GRADE table itself, the study-by-study risk-of-bias detail, the subgroup detail, and the sensitivity-analysis plots, is reported here only as described by the main text, and this is flagged each time it occurs.
Context
The question comes up roughly once a week in consultation. A patient asks whether walking, swimming, or taking up cycling again will help, and the clinician says yes, without knowing precisely what that yes quantifies. The available literature rests heavily on self-reported physical activity, that is, on what participants are willing to say about themselves, with the expected over-reporting and a systematic gap between people who are doing well and those who are not.
Cardiorespiratory fitness is a different variable. It does not measure a behavior but a capacity. It integrates training history, genetics, age, body composition, and comorbidities. It is a physiological marker, less distorted by a patient’s memory, and therefore a cleaner candidate for an etiological question. It is also a variable that an outpatient psychiatrist does not measure, which shifts the interest of this work toward the argument it supplies rather than toward a tool to use.
The topic also touches two major public health problems, depression and dementia, for which modifiable prevention levers remain scarce. The two earlier meta-analyses cited by the authors each pooled fewer than five primary studies and were limited to adults, covering anxiety, depression, and all-cause dementia.
What the study measures, and what it does not
The study links a level of aerobic capacity recorded at baseline to the later onset of a diagnosis. Nothing is measured in between. The discussion devotes several paragraphs to candidate mechanisms, neuroplasticity, cerebral perfusion, systemic inflammation, hippocampal volume preservation, white matter, and the hypothalamic-pituitary-adrenal axis, but these developments draw on outside literature and on none of the data from the cohorts pooled here. No intermediate marker is extracted, and no mediation analysis is conducted. The authors implicitly acknowledge this by noting that few trials have formally examined whether a gain in cardiorespiratory fitness mediates the observed effects of exercise, and by calling for randomized trials to settle the question.
Second point to hold onto: cardiorespiratory fitness is not exercise. It is partly a consequence of it, but it also depends on determinants that no prescription touches. The authors themselves note that genetics account for close to half of the variation in this capacity, and that part of the observed associations could reflect genetic pleiotropy. Writing that exercise reduces dementia risk by 39% would therefore be a double extrapolation, first from association to causation, then from capacity to behavior. The result does not support it. This 39% figure, incidentally, is not published: it is calculated from the hazard ratio of 0.61.
The study at a glance
| Question, design, criteria | |
|---|---|
| Population | |
| PGeneral population, all ages, with no diagnosed mental or neurocognitive disorder at baseline: 27 cohort studies, 4,007,638 participants, 30.5% women, across nine countries, Finland, Germany, Italy, the Netherlands, Norway, Sweden, Taiwan, the United Kingdom, and the United States. Participants predominantly white, of European ancestry or non-Hispanic in the American cohorts. Mean age at baseline ranging from 10.6 to 72.5 years, with most of the corpus falling between 18 and 64 years. Follow-up of 4 to 29 years. Publications from 2009 to 2025. | |
| Exposure | |
| ICardiorespiratory fitness, assessed by direct measurement of maximal oxygen consumption with gas analysis in three cohorts, and by indirect measurement in all others: maximal or submaximal exercise tests without gas analysis, prediction equations based on performance, or prediction equations without an exercise test. The metrics used are field-test distance or duration, metabolic equivalents, maximal power output, and maximal oxygen consumption. A single measurement at baseline serves as the exposure, with a few cohorts holding repeated later measurements. Continuous analyses are expressed per 1-MET increment, that is 3.5 ml/kg/min. | |
| Comparator | |
| CParticipants with low cardiorespiratory fitness, used as the reference category. Most contributing studies standardize the threshold between high and low fitness by the participant’s age and sex. One study relies solely on an age and sex adjustment within the risk model. | |
| Outcomes | |
| OIncidence of mental and neurocognitive disorders classified under DSM-5-TR, captured either as clinical diagnoses or as symptom-severity measures. Four outcomes are meta-analyzed: anxiety disorders, depression, all-cause dementia, and psychotic disorders. The others, bipolar disorder, obsessive-compulsive disorder, somatic symptom disorders, trauma- and stressor-related disorders, sleep apnea, and attention-deficit/hyperactivity disorder, are each documented by a single study and are not pooled. | |
| Literature search | |
| SourcesFive databases searched from inception through 20 January 2025: PsycINFO, PubMed, Scopus, SPORTDiscus, and Web of Science, supplemented by Google Scholar and citation chasing. 4,385 records identified, 1,832 duplicates removed, 2,553 titles and abstracts screened, 76 full texts assessed, 49 excluded, 27 studies retained in the review and 22 in the meta-analyses. | |
| Study design | |
| DesignSystematic review and meta-analysis of prospective or retrospective cohort studies, with continuous dose-response analyses. Conducted following PRISMA 2020 and MOOSE, registered with PROSPERO. Pooling by random-effects model using the Sidik and Jonkman method, hazard ratios and odds ratios treated as equivalent and combined as HRs, analyses run in R 4.4.0 with the meta and metafor extensions. Methodological quality assessed with the National Institutes of Health tool for observational cohort studies, certainty of evidence with GRADE. Level 2a on the CEBM scale, derived from the published design. |
Quality control
| Point checked | Verdict |
|---|---|
| Overall certainty of evidence | Very low to moderate |
| FindingThis matches how the authors themselves characterize the evidence overall in their abstract. No result in this review can be presented as high-certainty evidence, and only one reaches the intermediate level. | |
| GRADE rating, outcome by outcome | Moderate for dementia, very low elsewhere |
| FindingFor the high-versus-low fitness comparison, certainty is moderate for all-cause dementia, and very low for anxiety disorders, depression, and psychotic disorders. For the continuous per-1-MET analysis, it is very low for both depression and all-cause dementia. The authors attribute this downgrading to risk of bias, inconsistency, imprecision, and the inability to assess publication bias. The detailed reasoning appears in two supplementary tables, which we were not able to consult. | |
| Heterogeneity between studies | From zero to 97.3% depending on the outcome |
| FindingThe published values are 56.9% for anxiety disorders, 67.9% for depression, 0% for all-cause dementia, and 0% for psychotic disorders in the high-versus-low comparisons; 17.0% for depression and 97.3% for all-cause dementia in the continuous analyses. That last figure is considerable even by the authors’ own reading grid, which classifies anything above 75% as such. The published prediction intervals confirm it: 0.43 to 0.97 for depression high-versus-low, 0.47 to 0.79 for dementia high-versus-low, 0.72 to 1.25 for depression per MET, and 0.07 to 9.12 for dementia per MET. | |
| Methodological quality of studies | Good for 57.1% of studies |
| FindingThe assessment uses the National Institutes of Health tool for observational cohort studies. 57.1% of studies are rated good quality and 42.9% fair. The three criteria most often failed in the fair-quality studies are a detailed definition of the exposure measure, repetition of that measure, and a follow-up rate above 80%. The study-by-study detail appears in a supplementary table that was not consulted. | |
| Publication bias | Not assessable, fewer than ten studies per outcome |
| FindingThis is not an omission, it is an impossibility the authors declare themselves: with fewer than ten studies per outcome, neither meta-regression nor funnel-plot asymmetry testing was usable. They draw the consequence themselves, stating that the influence of unpublished or selectively reported results cannot be excluded. This impossibility is one of the reasons for GRADE downgrading. | |
| Sensitivity analyses | Three analyses, results unchanged |
| FindingLeave-one-out removal of each study, analyses run separately by type of risk estimator, hazard ratio or odds ratio, and a correction to a standardized effect size using Hedges’ g. All three leave the pooled estimates unchanged. Their plots appear in supplementary figures and a supplementary table that were not consulted. | |
| Composition of the corpus | Close to 70% men, predominantly white corpus |
| Finding30.5% women. Cohorts with measured cardiorespiratory fitness historically draw on occupational, military, or athletic populations with a strong male predominance, and the evidence on psychotic disorders comes exclusively from two cohorts of male conscripts. Participants are also predominantly white, which the authors flag as a limit on generalizability. | |
| Study selection and data extraction | Two independent reviewers, arbitration planned |
| FindingSearching, selection, extraction, quality assessment, and certainty grading were each carried out independently by two researchers, with disagreements resolved by consensus after consulting a third investigator. No inter-rater agreement statistic is published. The meta-analytic dataset and the R code are deposited on the Open Science Framework. | |
| Protocol registration | Number declared, registry not consulted |
| FindingThe publication declares a PROSPERO registration under number CRD42024547081. The registry entry itself could not be opened from our workstation. Whether the deposited protocol matches the published analysis was therefore not checked. | |
| Reverse causation | Not excluded by the design |
| FindingPeople who will go on to develop depression already have, on average, lower fitness. The loop runs both ways, and an observational cohort design cannot settle it. The authors say so explicitly and raise the possibility that the protective effect attributed to cardiorespiratory fitness is overestimated. | |
Results
Hazard ratio for all-cause dementia in adults with high cardiorespiratory fitness, compared with those of low fitness. 95% CI 0.55 to 0.68, across 7 studies and 1,287,561 participants, with zero heterogeneity and a prediction interval of 0.47 to 0.79. It is the only result in this review graded moderate certainty, and it remains an association.
| Outcome | Reported estimate |
|---|---|
| Depression, high versus low fitness | HR 0.64 (95% CI 0.56 to 0.74) |
| Reading9 studies, 1,504,915 participants, 18,958 incident cases, follow-up of 4 to 17 years, registry diagnoses or symptoms ranging from mild to severe. Heterogeneity of 67.9%, p=0.005, prediction interval of 0.43 to 0.97. Very low certainty of evidence. The confidence interval clearly excludes unity; the prediction interval comes close to it, which means a future cohort could find almost no difference at all. Expressed as a relative reduction, the central value corresponds to 36%, a figure we calculate from the hazard ratio and that is not published as such. | |
| All-cause dementia, high versus low fitness | HR 0.61 (95% CI 0.55 to 0.68) |
| Reading7 studies, 1,287,561 participants, 4,170 incident cases, follow-up of 8 to 29 years, registry diagnoses. Zero heterogeneity, p=0.625, prediction interval of 0.47 to 0.79. This is the only outcome graded moderate certainty, and it is the combination of zero heterogeneity and a tight prediction interval that justifies it, far more than the width of the confidence interval. That is a 39% relative reduction, again a figure we calculate and that is not published. | |
| Psychotic disorders, high versus low fitness | HR 0.71 (95% CI 0.65 to 0.77) |
| ReadingOnly 2 studies, 1,112,007 participants, 10,420 incident cases, follow-up of 23 to 25 years, zero heterogeneity, p=0.629. Very low certainty of evidence. Both cohorts are exclusively male, one Swedish and one Finnish, and the authors explicitly flag this restriction all the way into their conclusion. One of them contributes 1,109,786 of the 1,112,007 participants and 10,205 of the 10,420 cases: the pooled estimate is, in practice, that of a single conscript cohort. The apparent narrowness of the interval should not be read as agreement between independent studies. | |
| Anxiety disorders, high versus low fitness | HR 0.90 (95% CI 0.75 to 1.09) |
| Reading2 studies, 36,687 participants, 1,009 incident cases, follow-up of 7 to 11 years, symptoms ranging from mild to severe. Heterogeneity of 56.9%, p=0.128. Very low certainty of evidence. This is the only meta-analyzed outcome whose interval spans unity, and the authors state that no significant association was found for anxiety. This result does not demonstrate an absence of association: with two cohorts, about a thousand cases, and an interval running from a 25% reduction to a 9% increase, it mainly says the question remains open. | |
| Depression, per 1-MET increment | HR 0.95 (95% CI 0.92 to 0.98) |
| Reading5 studies, 362,256 participants, 6,158 incident cases, follow-up of 5 to 12 years, heterogeneity of 17.0%, p=0.300, prediction interval of 0.72 to 1.25. Very low certainty of evidence. A modest but steady slope, associated with each MET gained, that is 3.5 ml/kg/min. This is the clinically most useful form of analysis, since it describes a gradient rather than a contrast between extreme categories whose thresholds vary from one cohort to the next. | |
| All-cause dementia, per 1-MET increment | HR 0.81 (95% CI 0.67 to 0.98) |
| Reading5 studies, 123,229 participants, 2,971 incident cases, follow-up of 8 to 24 years. Heterogeneity of 97.3%, p<0.001, prediction interval of 0.07 to 9.12. Very low certainty of evidence. This is by far the most fragile estimate in the review: the upper bound of the confidence interval reaches 0.98, and above all the prediction interval covers nearly the entire space of possible values, from a massive benefit to a massive excess risk. No decision, clinical or argumentative, should rest on this value. | |
| Outcomes documented by a single study | Not pooled, isolated results |
| ReadingIn adults, higher cardiorespiratory fitness is associated with lower risk of bipolar and related disorders, dissociative disorders, obsessive-compulsive disorder, somatic symptom disorders, and trauma- and stressor-related disorders, as well as sleep apnea. In children, a single cohort finds an association with anxiety and attention-deficit/hyperactivity disorder in both boys and girls, and with depression in girls only. A single study per outcome permits no generalization, and the authors do not pool it. | |
What is demonstrated: there is a steady, graded association between higher aerobic capacity and lower subsequent incidence of depression, all-cause dementia, and psychotic disorders in adults, across very large cohorts. Five of the six pooled estimates exclude unity; the one for anxiety disorders does not. What is suggested: that this gradient could serve as a risk-stratification marker, something the authors themselves phrase conditionally, stating that cardiorespiratory fitness may be considered a candidate marker for risk stratification, and that their results should be interpreted with caution given the very low certainty affecting most outcomes. What belongs to expert opinion, including our own: the idea that improving a patient’s fitness will lower their future psychiatric risk. That idea is plausible, consistent with the rest of the literature, and not established by this work.
Critical appraisal
| Domain | Judgment |
|---|---|
| Reverse causation | Core limitation |
| FindingThis is the main weakness, and it is not reducible by sample size. Social withdrawal, psychomotor slowing, and loss of drive appear months to years before a diagnosis is made, and they lower exercise capacity. The same reasoning applies to dementia, whose prodromal phase is long. The authors devote an entire discussion to this: every included study excluded participants with an established diagnosis at baseline, but undetected early symptoms could still have bent the trajectory of fitness. As a counterpoint they cite bidirectional Mendelian randomization studies arguing for a causal effect of physical activity on depression with no effect in the reverse direction, and a genetically informed analysis of Alzheimer’s disease that found no significant result. | |
| Exclusion of the early follow-up years | No latency analysis reported |
| FindingThis is the sensitivity analysis that would help address the previous objection: redoing the estimate after excluding cases occurring in the first years after the measurement. The three sensitivity analyses reported in the main text do not do this: they remove studies one at a time, separate risk estimator types, and apply a Hedges’ g correction. None addresses the latency between exposure and diagnosis. The supplementary methods, which we did not consult, might describe others, but the main text mentions none. | |
| Nature of the exposure | Measured, not self-reported, capacity |
| FindingThis is what sets this work apart from meta-analyses built on activity questionnaires, and what justifies reading it. The advantage is graded, though, and the full text finally lets us measure it: maximal oxygen consumption is directly measured, with gas analysis, in only three cohorts. All the others rely on maximal or submaximal tests without gas analysis, on prediction equations based on performance, or on prediction equations without an exercise test at all. The very title of one of the two studies contributing to the anxiety outcome refers to cardiorespiratory fitness estimated without an exercise test. The exposure remains a capacity rather than a self-reported behavior, but it is not measured with the same rigor everywhere, which the authors themselves cite as a reason for downgrading quality. | |
| Consistency of the two dementia scales | Two different sets of studies |
| FindingAn observation of our own, obtained by cross-referencing published values, and one the authors do not state. The extreme-category contrasts are nearly identical for dementia and depression, 0.61 and 0.64. The continuous slopes, however, differ by a considerable factor, 0.81 and 0.95 per MET. The explanation becomes legible in the reference lists of the two figures: the high-versus-low comparison for dementia rests on 7 studies and 1,287,561 participants, the continuous analysis on 5 studies and 123,229 participants, and only three studies are common to both. The first has zero heterogeneity, the second a heterogeneity of 97.3%. These are not two readings of the same signal, they are two analyses on largely distinct sets of studies, one of which is highly unstable. | |
| Weight of the very large cohorts | One cohort carries most of an outcome |
| FindingFor psychotic disorders, a single cohort of Swedish conscripts contributes 1,109,786 of the 1,112,007 participants and 10,205 of the 10,420 incident cases. For depression high-versus-low, the same family of registries contributes 1,117,292 participants and 12,427 cases out of a total of 1,504,915 and 18,958. The correspondence between sample size and statistical weight is not mechanical, though: in a random-effects model, weight depends on the precision of each estimate, not the number of participants, as illustrated by dementia, where the largest cohort, 1,174,483 participants, contributes only 662 incident cases, while a cohort of 19,458 participants contributes 1,659. The reader should therefore keep in mind that these pooled estimates are not the convergence of numerous independent studies of comparable size. | |
| Composition of the corpus | Close to 70% men |
| Finding30.5% women. On a question where the prevalence of depression is markedly higher in women, this imbalance weighs on external validity. Sex-specific analyses exist for depression and are reassuring in direction: 0.51 (95% CI 0.39 to 0.66) in women, 0.55 (95% CI 0.51 to 0.60) in men, with no significant difference between subgroups. By contrast, the entirety of the evidence on psychotic disorders comes from male conscript cohorts and does not transpose to women, as the authors themselves state. | |
| Residual confounding | Inherent to the design |
| FindingPeople in good physical condition differ from others along many axes. No adjustment model neutralizes that gap, and the authors list it out themselves: genetic predisposition, chronic pain, multimorbidity, social support, and lifestyle behaviors including physical activity itself, diet, and smoking. The adjustment variables used cohort by cohort were extracted but relegated to a supplementary table we did not consult: we therefore cannot judge how fine-grained the study-by-study adjustments were. | |
| External validity | General population, not a clinical caseload |
| FindingEnrollment draws on the general population, with no established disorder at baseline. Transposing this to an already-constituted psychiatric caseload remains an extrapolation: treated patients, often on psychotropic medication and already deconditioned, are not the population studied. The authors add two geographic and demographic restrictions: only nine countries, all European, North American, or East Asian; a predominantly white corpus; and an underrepresentation of those under 18 and over 65. | |
| What this reading could not examine | Supplementary material not consulted |
| FindingThe full text was read, but not the supplementary material, which contains the detailed GRADE table, the study-by-study risk-of-bias assessment, each cohort’s characteristics and adjustments, the subgroup-analysis detail, the sensitivity-analysis plots, and the search strategy. Whatever is reported here from that material comes from what the main text says about it, never from the material itself. | |
Level of evidence
Confidence is high regarding the existence of an association and its direction for depression, all-cause dementia, and psychotic disorders. It is also high regarding the nature of the exposure, which is this work’s real contribution. From there it follows the authors’ own grading, and that grading is strict: only one result out of six reaches moderate certainty, all-cause dementia in the high-versus-low comparison; every other one is graded very low. It is low for psychotic disorders, whose estimate rests in practice on a single cohort of male conscripts. It is frankly low for the continuous estimate for dementia, whose heterogeneity reaches 97.3% and whose prediction interval runs from 0.07 to 9.12. It is nil regarding causation: an observational design cannot establish it, whatever the number of participants. Four million people do not turn a correlation into a mechanism, they only make the correlation more precise.
The colleague test
What an experienced colleague might say about this study in two minutes, between two consultations.
“It doesn’t change my consultation tomorrow morning, but it finally gives me a number to counter with when a patient asks if it actually helps. On dementia, 0.61 with zero heterogeneity and certainty rated moderate, that holds up. On depression I stay cautious: a patient who’s depressed moves less, so which way the arrow points, nobody’s shown me that. And on psychosis, two conscript cohorts, all men, one of which carries the result on its own, I’m not taking that to a female patient.”
Translated for practice: the work legitimizes and quantifies a recommendation most clinicians already give, without granting it the status of a demonstrated effect. It strengthens the argument, it does not change the therapeutic strategy.
What you can do with this
- Have a usable order of magnitude for consultation, provided it is phrased correctly. Not “exercising cuts your dementia risk by 39%,” but “people with better exercise capacity develop dementia less often, across follow-ups covering several million people.” The nuance is not cosmetic, it is what protects you from making a promise the data does not cover.
- Favor the gradient over the extreme-category contrast. One MET gained, that is 3.5 ml/kg/min, corresponds to an HR of 0.95 for depression. It is modest, cumulative, and it describes what is happening in a given patient better than a comparison between the fittest and least fit groups, whose thresholds vary from cohort to cohort.
- Do not transpose the result on psychotic disorders to women. It comes from two exclusively male conscript cohorts, one of which carries nearly the entire sample, and the authors themselves flag this restriction.
- Ask about exercise capacity the way you ask about sleep or appetite. Climbing a flight of stairs, sustaining a brisk walk, these conversational benchmarks are better than nothing, without pretending to estimate aerobic capacity. No standardized tool for measuring cardiorespiratory fitness is available in outpatient practice, and the study offers none; the authors themselves call for standardized thresholds to finally be defined.
- Account for the composition of the corpus when you transpose these findings: close to 70% men, participants predominantly white, general population with no established disorder at baseline, no cohort from Africa, Latin America, or Oceania.
- Do not substitute any of this evidence for an ongoing treatment. Nothing in this work concerns treated patients, or the course of an established disorder.
- The course of action is set out in the NICE decision tree for generalised anxiety and panic disorder.
Prescribing, delivery and reimbursement: three separate questions
This framework sits outside the study itself, which evaluates no prescribing pathway. It is worth setting out here because it is the question a reader often asks next: if fitness tracks with lower risk, can physical activity be prescribed, and will it be paid for? In most health systems these are three separate questions: whether an intervention can be formally authorized or prescribed, who is qualified to deliver it, and whether it is collectively reimbursed. A decision not to reimburse something, or to reimburse it only under specific conditions, is a financing and administrative choice. It says nothing about whether the intervention works.
France illustrates the point cleanly. Under a law passed in 2022, article L.1172-1 of the French Public Health Code allows a physician managing a patient with a long-term or chronic condition, a risk factor, or a loss of autonomy to prescribe what French law calls activité physique adaptée, physical activity adapted to the patient’s condition, physical capacity, and medical risk. Delivery is restricted to specifically qualified professionals, among them physiotherapists, occupational therapists, and psychomotor therapists. Yet the same law sets up no funding mechanism, and as of a 23 March 2026 update on the French national health insurance fund’s page for physicians, adapted physical activity is not covered by statutory health insurance as a matter of common right. Targeted funding exists through regional health agencies, local authorities, or some supplementary insurers, and a derogatory scheme opened in March 2026 for patients with complex obesity. This regulatory framework was verified against Légifrance and the health insurance fund’s site on 13 August 2026; these texts change and should be checked again before any decision is based on them.
None of this bears on the evidence reviewed above. Authorization to prescribe, who may deliver an intervention, and whether it is reimbursed are three distinct decisions, made on different grounds and by different bodies, and a funding gap is not a verdict on efficacy. Readers practicing outside France should check the prescribing, delivery, and reimbursement rules that apply in their own jurisdiction before advising a patient, since these arrangements for adapted or supervised physical activity vary widely and change over time.
Frequently asked questions
Can I tell a patient that exercise cuts their dementia risk by 39%?
No, for three reasons. The 39% figure is not published: it is calculated from the hazard ratio of 0.61. That ratio describes a risk difference between two fitness groups, not the effect of an intervention. It is also relative, so it says nothing about the absolute benefit expected for a given individual. And cardiorespiratory fitness cannot be reduced to exercise performed, since the authors themselves note that close to half of its variation is genetic in origin.
Is the dementia finding stronger than the depression finding?
Yes, and the authors say so explicitly. The GRADE rating places all-cause dementia at moderate certainty for the high-versus-low fitness comparison, and depression at very low certainty. The gap comes down to heterogeneity, zero for dementia and 67.9% for depression, and to the prediction interval, 0.47 to 0.79 for dementia against 0.43 to 0.97 for depression. One caveat though: this hierarchy holds for the extreme-category comparison. In the continuous per-MET analysis, both outcomes drop back to very low certainty, and dementia even becomes the most unstable estimate in the entire review.
Did the analysis exclude cases occurring in the first years of follow-up?
No, at least not in what the main text reports. The three published sensitivity analyses remove each study one at a time, separate hazard ratios from odds ratios, and apply a Hedges’ g correction. None addresses the latency between the fitness measurement and the diagnosis. Yet this is exactly the analysis that would help address the reverse-causation objection, and its absence is the main reason that objection stands unresolved.
What are the anxiety disorder results worth?
The pooled estimate is 0.90, with a 95% confidence interval of 0.75 to 1.09, across two studies, 36,687 participants, and 1,009 incident cases, with a heterogeneity of 56.9% and very low certainty of evidence. The authors conclude that no significant association was found. That does not mean the association is null: two cohorts and an interval running from a one-quarter reduction to a 9% increase leave the question open, and an absence of statistical significance in such a sparsely powered analysis is not a demonstration of no effect.
Does the result hold in women and across age?
For depression, yes, as far as the subgroup analyses allow: 0.51 (95% CI 0.39 to 0.66) in women and 0.55 (95% CI 0.51 to 0.60) in men. Before age 50, depression sits at 0.64 (95% CI 0.52 to 0.79) and dementia at 0.62 (95% CI 0.54 to 0.70); after 50, at 0.65 (95% CI 0.52 to 0.80) and 0.60 (95% CI 0.47 to 0.72). No significant difference appears between the compared subgroups. Two caveats: these analyses rest on study-level aggregated data rather than individual-participant data, which the authors themselves list as a limitation, and the evidence on psychotic disorders includes no women at all.
Should I measure my patients’ cardiorespiratory fitness?
That is neither realistic nor useful in outpatient psychiatry, and the study does not propose it. The objective measurement here is a research tool, chosen for the validity of the exposure. The authors themselves note that no standardized categories of cardiorespiratory fitness with defined thresholds currently exist, and call for building them. Its clinical transposition therefore plays out in what you say to the patient, not in an act you perform.
Who funded this work?
Public bodies only, according to the printed declaration: Uruguay’s national research and innovation agency, the University of Castilla-La Mancha with co-financing from the European Social Fund, the Spanish national research agency and the European Regional Development Fund, the Spanish Ministry of Science through a mobility grant, and the Swedish Research Council for Health, Working Life and Welfare. Open-access publication fees were covered by the Karolinska Institutet. The authors state that funders had no role in the study’s design, data collection and analysis, the decision to publish, or manuscript preparation, and they declare no conflicts of interest.
Annotated bibliography
Source study. Díaz-Goñi V, López-Gil JF, Rodríguez-Gutiérrez E, Visier-Alfonso ME, Jiménez-López E, Sequí-Domínguez I, Ortega FB, Castro-Piñero J, Mesas AE, Sánchez-López M, Martínez-Vizcaíno V, Bizzozero-Peroni B. Cardiorespiratory fitness and risk of mental disorders and dementia: a systematic review and meta-analysis. Nature Mental Health, 2026, volume 4, issue 4, pages 653 to 663. DOI 10.1038/s44220-026-00599-4. Received 24 June 2025, accepted 26 January 2026, published online 20 March 2026, open access under a Creative Commons Attribution 4.0 license. No PubMed identifier is available for this reference as of the verification date: the publication is not indexed in PubMed, and a DOI search there returns zero results as of 13 August 2026. PROSPERO registration CRD42024547081. The meta-analytic dataset and R code are deposited on the Open Science Framework, DOI 10.17605/OSF.IO/BJSGN.
Systematic review and meta-analysis of cohort studies, 27 studies included and 22 pooled, 4,007,638 participants, 30.5% women, nine countries, follow-up of 4 to 29 years, with continuous dose-response analyses per 1-MET increment. Public funding exclusively, publication fees covered by the Karolinska Institutet; the authors declare no conflicts of interest. Named peer review, the journal crediting Riccardo De Giorgi, David Stacey, and Jesse Stewart. Contribution: a quantified, non-self-reported exposure on a question usually handled by questionnaire, and sample sizes that give the estimates unusual precision. Limitation: an observational design that leaves reverse causation open, with no latency analysis to address it; a corpus close to 70% male and predominantly white; evidence on psychotic disorders drawn from two male conscript cohorts; and certainty of evidence graded moderate for only the all-cause dementia high-versus-low outcome, very low for every other one.
What was consulted. Figures, method, and references verified by independent double reading against the full text of the published version, consulted on 13 August 2026: the article’s eleven pages, pages 653 to 663, comprising the abstract, introduction, results, discussion, conclusion, methods, Table 1, the flow diagram in Figure 1, the two forest plots in Figures 2 and 3, the acknowledgments, the funding statement, the conflict-of-interest statement, and the affiliations, together with the Nature Portfolio reporting summary attached to the article. Every sample size, every hazard ratio, every confidence interval, every prediction interval, every heterogeneity statistic, and every GRADE level cited here was recorded from its location in this source. The supplementary material, Supplementary Tables 1 to 10, Supplementary Figures 1 to 4, and the supplementary methods, did not accompany the file and was not consulted: what exists only in that material, notably the detailed GRADE table, the study-by-study risk of bias, each cohort’s characteristics and adjustments, the subgroup detail, and the sensitivity-analysis plots, is reported here only as described by the main text, and this is flagged each time it occurs. The PROSPERO registry entry CRD42024547081 could not be opened from our workstation, its page requiring script execution: the number is reported as declared in the publication and was not checked against the registry. The meta-analytic dataset deposited on the Open Science Framework was not downloaded. The title, the list of twelve authors, the journal, the year, the volume, the pagination, and the DOI were recorded from the publication itself; the issue number comes from the publisher’s record, the publication itself carrying only the mention of April 2026. No PMID chip appears under the title because no PMID is available for this reference: the publication is not indexed in PubMed, and a DOI search there returns zero results as of 13 August 2026. Values reported as calculated by us, the 36% and 39% relative reductions, one cohort’s share of a total, and the number of studies common to two analyses, are flagged as such where they appear and do not appear in the publication. The French regulatory framework discussed above was verified on 13 August 2026 against Légifrance, for articles L.1172-1 and D.1172-1 to D.1172-5 of the Public Health Code, and against the national health insurance fund’s site for the reimbursement question; these texts change and should be checked again before any decisional use.
Editorial collections
Tags
Verified on August 13, 2026 against the full text of the publication and its supplementary material where available. This analysis underwent an independent double reading. The English version was checked for conformity on September 23, 2026, against the figures of the French version and against the source. How we verify what we publish
This analysis is intended for healthcare professionals. It does not constitute a prescribing recommendation and does not replace individual clinical judgment.
Analysis from Psychiatry Evidence Base, evidence-based psychiatry, explained with rigor.