Every hazard ratio in this calculator traces back to a primary source — a large cohort study or meta-analysis. Below is how we grade evidence, how the model works, then the bibliography grouped by domain.
How we grade evidence
Each chip grades one claim on one endpoint, not a marker as a whole. A marker can carry a B headline on home and A facts on its metric page; those A facts are real and they are not the home number.
Grade ARCT or meta-analysis of RCTs on that endpoint
Someone ran the experiment on that outcome. A trial that finds nothing is still A — the letter is how hard it was tested, not whether the result was flattering.
Grade Bprospective cohort with dose-response, or Mendelian randomization
Large groups followed for years — not trials. Dose-response in a cohort, or a genetic natural experiment.
Grade Ccohort association only
The numbers moved together in a group of people. That is not the same as one causing the other.
Grade Dmechanistic reasoning only
A plausible mechanism. No human outcome study for this claim.
Grade Xno evidence for the claimed endpoint
We looked. This protocol has not been tested against that outcome.
Some cards show METHOD instead of a letter — method notes, not claims.
Every Grade A fact on this site is a randomized trial. Most of them did not cut all-cause deaths, and none of the eight scores is built on one. What the trials found.
Per-metric hazard ratios
Continuous metrics use monotonic PCHIP interpolation between literature-derived anchor points, so risk changes smoothly rather than jumping at thresholds.
Domain aggregation
Metrics are grouped into three domains; each collapses to a single number via a weighted geometric mean, then the three combine the same way. Weights sum to 1 at both levels, so this is a weighted average of log hazard ratios — one metric moves the result by exactly its own weight share, and nothing is double-counted.
Why the composite is an index, not a risk estimate
Because the weights average rather than accumulate, the composite ranks profiles against the population median (1.00) but does not compound risk across markers. Estimating combined risk honestly would need a jointly-fitted multivariable model, and no cohort has measured all 8 markers on the same people. So the composite is published as an index and is never expressed as a percentage change in risk. Per-metric hazard ratios are different: each is a single-marker claim backed by a single-marker study, which is exactly what the cited papers support.
Score & percentile
The composite maps to a 0–100 score. 50 is a typical person on all eight. The scale goes to 100; that is a cap, not a test score, and not a person the studies describe. The percentile looks up that score against a 100,000-profile-per-sex simulation built over the same published marginal distributions used for the curves above, rather than assuming a bell curve. See "Fifty isn’t the middle of the pack" below for why that reference profile lands well above the 50th percentile.
How the citations are checked
Every metric is backed by at least five primary studies, listed in full below. A 2026-08-27 adversarial audit opened each lead paper and required a verbatim source sentence for every quoted statistic. Two live numbers had been hung on the wrong paper in the same bibliography (hearing HR 1.21 is Jia 2025, not Tan 2022; 15% per 10 teeth is Peng 2019, not Romandini 2020) and were relabeled. A number is only as good as the sentence it came from — if we cannot quote that sentence, it does not ship.
What it is not
This is a directional, educational estimate — not a validated clinical risk model or medical advice.
What the eight do not cover
The scored eight are not a lifestyle score. Most of them are readouts — numbers that describe accumulated physiology — not the exposures mortality studies rank when they ask which actions account for most deaths. A person can smoke, drink heavily, eat poorly, and carry high body weight and still look typical or better on all eight. That is a scope choice, not a finding that those exposures do not matter.
Most of the eight are readouts
Grip, VO2 max, resting heart rate, and HbA1c describe what the body is doing now. Steps and fiber are intakes. Sleep is duration, not regularity or clock time. Blood pressure is the closest exception: it is both a reading and close to the pathway the Global Burden of Disease ranks. That ranking is of disease burden (years lost to death and disability), not a deaths table, and it is a population share, not your personal risk.
Smoking, alcohol, diet quality, and body weight are not in the score
They have no input. Joint lifestyle-score cohorts are built from never smoking, weight in a healthy range, activity, moderate alcohol, and diet quality — the usual core in this literature. Physical activity here is scored only as VO2 max and steps, not as hours of exercise. Smoking is the largest single gap against both that core and the GBD ranking; it was considered and declined as a product input. The epidemiology is not in dispute.
Global burden and strength of evidence for 88 risk factors in 204 countries and 811 subnational locations, 1990–2021
Population-level ranking of modifiable risk. Headline list is DALYs, not deaths. High systolic pressure, smoking, and high fasting glucose sit in the top tier. Death-count tables from secondary write-ups are not used here.
Impact of Healthy Lifestyle Factors on Life Expectancies in the US Population
Nurses' Health Study and Health Professionals Follow-up Study. Five-factor lifestyle score: never smoking, BMI 18.5-24.9, activity, moderate alcohol, diet quality. Cited for the factor set, not as a MarkerScore output.
The bibliography below covers the 8 scored metrics. For markers outside the index — sleep regularity, balance, gait speed, hearing, tooth loss, social connection, and ApoB — see More you can do.
Physical Capacity
VO2 Max
Singh et al. (2024) — Journal of Sport and Health Science
Comparison of objectively measured and estimated cardiorespiratory fitness to predict all-cause and cardiovascular disease mortality in adults: A systematic review and meta-analysis of 42 studies representing 35 cohorts and 3.8 million observations
Engine overlay (page meta, not the knot table). 42 studies / 35 cohorts / 3,813,484 observations (81% male), 362,771 ACM deaths. Per 1-MET (3.5 mL/kg/min) ACM RR 0.86 (0.83–0.88), I² 97.66%. Objectively measured 0.90; maximal exercise-estimated 0.85; submaximal 0.94; non-exercise-estimated 0.85 (0.80–0.89) — no statistically significant differences. Non-exercise = prediction equations (age/sex/BMI/PA/RHR), not a consumer watch. Single baseline CRF; authors could not examine whether improved CRF reduced death. The calculator’s PCHIP is constructed; HR=1.0 at 38 M / 31 F is an ACSM midpoint, not Singh. Engine not moved. OA PDF opened 2026-08-29. Hunt: docs/RESEARCH-2026-08-29-vo2-mortality-meta.md.
Lang et al. (2024) — British Journal of Sports Medicine
Cardiorespiratory fitness is a strong and consistent predictor of morbidity and mortality among adults: an overview of meta-analyses representing over 20.9 million observations from 199 unique cohort studies
Umbrella of 26 SRs / 199 cohorts / 20.9 million observations — not a new pooling. Headline high vs low ACM HR 0.47 (0.39–0.56) is Han 2022 extracted in Fig 2. Per-MET 11–17% is Laukkanen 0.89 through Qiu eCRF 0.83. GRADE very low-to-moderate (male-heavy). No absolute mL/kg/min ACM target. OA PDF opened 2026-08-29.
Association of Cardiorespiratory Fitness With Long-term Mortality Among Adults Undergoing Exercise Treadmill Testing
Competing number, wrong construct. Cleveland Clinic referred treadmill testers, n=122,007, 13,637 deaths. Age/sex percentile bins of peak estimated METs, not absolute mL/kg/min. Low vs elite adjusted ACM HR 5.04 (4.10–6.20). “No observed upper limit of benefit” — that is why council B dropped ≥45, not a reason to print 5.04 on this curve. Engine low end is 1.6. Eutils abstract opened 2026-08-29; PMC recaptcha.
Cardiorespiratory fitness as a quantitative predictor of all-cause mortality and cardiovascular events in healthy men and women: a meta-analysis
Action / honesty. 33 studies; ACM 102,980 / 6,910. Per 1-MET RR 0.87 (0.84–0.90). Low vs high RR 1.70 (1.51–1.92). Bins: low <7.9 METs, high ≥10.9 — sex-blind MET cutoffs, not this calculator’s 38/31 zero and not a ≥45 mL/kg/min goal. Observational, not a training trial. Eutils abstract opened 2026-08-29; Unpaywall is_oa false.
Objectively Assessed Cardiorespiratory Fitness and All-Cause Mortality Risk: An Updated Meta-analysis of 37 Cohort Studies Involving 2,258,029 Participants
Largest exercise-test ACM meta in Lang Fig 2. 37 cohorts, 2,258,029 people, 108,613 deaths. Top vs bottom tertile RR 0.55 (0.50–0.61). Per 1-MET 0.89 (0.86–0.92). Eutils abstract opened 2026-08-29.
Daily steps and health outcomes in adults: a systematic review and dose-response meta-analysis
Engine paper. Systematic review of 57 studies / 35 cohorts; 31 studies / 24 cohorts entered meta-analyses. The all-cause mortality figure is a 14-study dose-response (not the 57-study review count): 7,000 vs 2,000 steps/day HR 0.53 (95% CI 0.46–0.60), I² 36.3%. Inverse non-linear ACM association with inflection around 5,000–7,000. Authors: 10,000 remains viable for more active people; 7,000 is a more realistic target for some. The 2,000-step reference is the lower bound of the normal range for older adults, not a typical adult. Erratum PMID 40883040.
Paluch et al. (Steps for Health Collaborative) (2022) — The Lancet Public Health
Daily steps and all-cause mortality: a meta-analysis of 15 international cohorts
Harmonised IPD meta of 15 cohorts (47,471 adults, 3,013 deaths). Age-stratified ACM plateaus: ≥60 y at 6,000–8,000 steps/day; <60 y at 8,000–10,000 (p interaction 0.012). Quartile HRs vs lowest quartile: 0.60, 0.55, 0.47. Ding 2025 lists lack of age-specific analysis as a limitation of its own paper — Paluch is still the age-band meta.
Banach et al. (2023) — European Journal of Preventive Cardiology
The association between daily step count and all-cause and cardiovascular mortality: a meta-analysis
17 cohort studies (226,889 people): each additional 1,000 steps/day is associated with 15% lower all-cause mortality (HR 0.85, 95% CI 0.81–0.91). “More the better” over ~3,867 steps/day — a linear-increment pooling, not Ding’s 5,000–7,000 inflection. Same construct (device steps → ACM); different shape from the engine curve.
Del Pozo Cruz et al. (2025) — Annals of Internal Medicine
Step Accumulation Patterns and Risk for Cardiovascular Events and Mortality Among Suboptimally Active Adults
Walking in sustained bouts of ≥15 minutes yielded lower 9.5-year mortality (0.80% vs 4.36%) and CVD incidence (4.39% vs 13.03%) than brief sporadic bursts, independent of total step volume.
Association of Daily Step Patterns With Mortality in US Adults
Action / honesty. NHANES 2005–06 (n=3,101). Not part of Ding’s 14-study pool. Hitting ≥8,000 steps 1–2 days/week vs 0 days: adjusted 10-year ACM risk difference −14.9 percentage points (95% CI −18.8 to −10.9), not an HR. Primary model also adjusts for mean daily step count. Plateau by 3–7 days (aRD −16.5). Pattern is not Ding’s daily-average curve. Do not hang “weekend warrior” on the engine paper.
Del Pozo Cruz et al. (2022) — JAMA Internal Medicine
Prospective Associations of Daily Step Counts and Intensity With Cancer and Cardiovascular Disease Incidence and Mortality and All-Cause Mortality
78,500 UK Biobank adults, median ~7-year follow-up: more daily steps up to ~10,000/day is associated with lower all-cause mortality and lower cancer and cardiovascular incidence and mortality. The paper reports dose-response as change in log-HR per 2,000-step increment rather than as fixed percentage reductions.
Association of Daily Step Count and Step Intensity With Mortality Among US Adults
NHANES cohort of 4,840 adults: higher daily step counts strongly linked to lower all-cause mortality (8,000 vs 4,000 steps/day HR 0.49; 12,000 vs 4,000 HR 0.35), with no independent association for step intensity (cadence) after adjusting for total volume.
Resting heart rate and all-cause and cardiovascular mortality in the general population: a meta-analysis
Engine paper (page meta). General-population only; search to 1 Jan 2015. 46 studies. ACM: 1,246,203 people / 78,349 deaths. Multivariate RR 1.09 (1.07–1.12) per 10 bpm, I² 92.3%, n=35 — RR, not HR. Trim-and-fill 1.04 (1.02–1.06). Categorical vs lowest: 60–80 bpm RR 1.12; >80 bpm RR 1.45. ACM spline vs 45 bpm is linear (p nonlinearity = 0.1); 90 bpm is the CV-mortality threshold, not an ACM cliff. ECG subgroup 1.06 vs non-ECG 1.15 per 10 bpm. The calculator’s PCHIP (70=1.0; 40–50 floor 0.85; steepens above 80) is constructed. Engine not moved. Hunt: docs/RESEARCH-2026-08-29-rhr-mortality-meta.md.
Aune et al. (2017) — Nutrition, Metabolism and Cardiovascular Diseases
Resting heart rate and the risk of cardiovascular disease, total cancer, and all-cause mortality – A systematic review and dose-response meta-analysis of prospective studies
Larger ACM meta (search to 29 Mar 2017). 87 studies overall; ACM n=48, RR 1.17 (1.14–1.19) per 10 bpm, I² 94.0%. Broader prospective mix than Zhang’s general-population-only pool. Webappendix high-vs-low ACM 1.69 (1.52–1.87). Trim-and-fill not quoted this pass (accepted manuscript login-walled). Engine not moved onto 1.17.
Age, sex and disease-specific associations between resting heart rate and cardiovascular mortality in the UK BIOBANK
Competing number, single cohort. UK Biobank n=502,534 (analysis 228,594 men / 272,737 women after 0.2% missing RHR), 7–12 y. Fully adjusted ACM per 10 bpm: men HR 1.22 (1.20–1.24), women 1.19 (1.16–1.22). Sitting Omron pulse, mean of two readings — closer to this slider than Zhang’s mixed ECG/non-ECG pool, and higher than Zhang’s 1.09. CVD mortality SHR 1.17 men / 1.14 women. PLOS HTML opened 2026-08-30. Engine not moved.
Relationship of resting heart rate and blood pressure with all-cause and cardiovascular disease mortality
Retrospective cohort of 67,028 Chinese adults aged 60 and older, 361,975 person-years of follow-up, 9,326 deaths. Adults in the highest resting-heart-rate quartile had 25% higher all-cause mortality risk than those in the lowest quartile (HR 1.25, 95% CI 1.17-1.33).
Resting heart rate and cause-specific death in a 16.5-year cohort study of the Japanese general population
Japanese general-population cohort of 8,800 adults (National Survey on Circulatory Disorders), followed 16.5 years. Middle-aged men (30-59y) in the highest resting-heart-rate quartile had 45% higher all-cause mortality (RR 1.45, 95% CI 1.06-2.00) than those in lower quartiles; associations were not significant in adults 60 and older.
Custodis et al. (2016) — Clinical Research in Cardiology
Resting heart rate is an independent predictor of all-cause mortality in the middle aged general population
Action / honesty. Heinz Nixdorf RECALL, n=4,091 without known CAD (3,348 free of rate-lowering meds), median 10.5 y, 398 deaths. Off meds, RHR ≥70 vs <70: ACM HR 1.68 (1.30–2.18); coronary events 1.20 (0.82–1.77), NS. Per 5 bpm ACM 1.13 (1.07–1.20); coronary 1.02 (0.94–1.11), NS. Authors: independent risk marker for ACM but not for coronary events. Marker, not a rate-lowering trial. Eutils abstract opened 2026-08-30.
Prognostic value of grip strength: findings from the Prospective Urban Rural Epidemiology (PURE) study
Engine paper (slope). PURE, 17 countries, ages 35–70. 139,691 with known vital status; 3,379 deaths; median 4.0 y. Jamar; mean of each hand’s maximum (protocol amended mid-study from non-dominant only). ACM HR per 5 kg reduction 1.16 (1.13–1.20); CVD mortality 1.17 (1.11–1.24). Per-SD ACM 1.37 vs SBP 1.15. Abstract/PDF report only this slope, not 40/25 knots. The calculator’s curve is constructed: 40 kg male / 25 kg female = HR 0.8 (not 1.0); below that, additive +0.16 per 5 kg; above, −0.005/kg to floor 0.7. UI is one hand; Leong is mean of both. NATAP PDF opened 2026-08-29; Lancet HTML login-walled. Hunt: docs/RESEARCH-2026-08-29-grip-mortality-meta.md.
López-Bueno et al. (2022) — Ageing Research Reviews
Thresholds of handgrip strength for all-cause, cancer, and cardiovascular mortality: A systematic review with dose-response meta-analysis
Page meta (kg dose-response, not the engine knots). 48 prospective cohorts of healthy adults, 3,135,473 people, search to 8 Feb 2022. ACM inverse 26–50 kg, close-to-linear, I² 45.7%. Does not report 16% per 5 kg. Tertile forest (12 studies, vs strongest third): weakest ACM HR 1.58 (1.40–1.78), middle 1.30 (1.17–1.44). OA PDF opened 2026-08-29. Engine not moved.
Wu et al. (2017) — Journal of the American Medical Directors Association
Association of Grip Strength With Risk of All-Cause Mortality, Cardiovascular Diseases, and Cancer in Community-Dwelling Populations: A Meta-analysis of Prospective Cohort Studies
Competing per-5 kg ACM meta. 42 studies / 3,002,203 people. Per 5 kg decrease ACM HR 1.16 (1.12–1.20) — same centre as Leong, pooled. Lowest vs highest ACM 1.41 (1.30–1.52). Linear within 56 kg. Abstract opened 2026-08-29; PDF not. Engine not moved onto this pool.
Associations of grip strength with cardiovascular, respiratory, and cancer outcomes and all cause mortality: prospective cohort study of half a million UK Biobank participants
Action / honesty. UK Biobank, 502,293 adults aged 40–69, mean follow-up 7.1 years, 13,322 deaths. Mean of right and left hands, not this UI’s one hand. Per 5 kg lower: women ACM HR 1.20 (1.17–1.23), men 1.16 (1.15–1.17). Splines: no deviation from linearity. FNIH weakness <26 kg men / <16 kg women is not this calculator’s 40/25. Observational. Leong: further research needed to test whether improvement in strength reduces death. Marker, not a training effect. BMJ HTML opened 2026-08-29.
Associations of handgrip strength with all-cause and cancer mortality in older adults: a prospective cohort study in 28 countries
SHARE, 121,383 adults (mean ~64) in 27 European countries and Israel, 896,836 person-years. Top vs bottom tertile ACM HR 0.41 (0.34–0.50) men / 0.38 (0.30–0.49) women. Maximal ACM threshold 42 kg men / 25 kg women — close to, not equal to, the engine’s 40/25. Not the engine knots.
Wang et al. (2022) — Journal of Science and Medicine in Sport
Association of handgrip strength with all-cause mortality: a nationally longitudinal cohort study in China
China Health and Retirement Longitudinal Study (CHARLS) cohort of 11,618 adults aged 45 and older, roughly 8-year follow-up, 1,290 deaths. Comparing the highest to lowest handgrip-strength tertile, all-cause mortality odds were 53% lower in men (OR 0.47, 95% CI 0.35-0.64) and 49% lower, though not statistically significant, in women (OR 0.51, 95% CI 0.24-1.08).
Lewington et al. (Prospective Studies Collaboration) (2002) — The Lancet
Age-specific relevance of usual blood pressure to vascular mortality: a meta-analysis of individual data for one million adults in 61 prospective studies
Engine overlay (page meta, not the joint rule). IPD of 1 million adults, 61 prospective studies, no prior vascular disease. Findings quantify vascular death: ages 40–69, each 20 mmHg usual SBP (≈10 mmHg usual DBP) is associated with more than a twofold difference in stroke death and twofold differences in IHD and other vascular death. No threshold down to 115/75. Interpretation says “vascular (and overall) mortality”; abstract does not report an ACM relative-risk table. The calculator’s joint min-reward / max-penalty rule is constructed (LOW confidence); high-side 1.08 per 10 mmHg SBP is not Lewington’s ~2.0 per 20. Unpaywall is_oa false 2026-08-29; eutils abstract opened. Hunt: docs/RESEARCH-2026-08-29-bp-mortality-meta.md.
Systolic Blood Pressure Reduction and Risk of Cardiovascular Disease and Mortality: A Systematic Review and Network Meta-analysis
Nearest ACM pool — treatment achieved SBP, not usual-BP observational. 42 trials, 144,220 treated hypertensives. Lowest ACM at mean achieved 120–124 mmHg. ACM HR vs 130–134: 0.73 (0.58–0.93); vs 140–144: 0.59 (0.45–0.77); vs ≥160: 0.47 (0.32–0.67). No matching-construct usual office/home BP → ACM continuous general-pop meta opened this pass. OA PDF opened 2026-08-29. Engine not moved.
Longer-Term All-Cause and Cardiovascular Mortality With Intensive Blood Pressure Control: A Secondary Analysis of a Randomized Clinical Trial
Action / honesty. SPRINT extended: intervention-period ACM HR 0.83 (0.68–1.01); at median 8.8 y total follow-up, after intensity lapsed, ACM HR 1.08 (0.94–1.23). Outpatient SBP in the intensive arm rose 132.8 → 140.4 mmHg by year 10. Marker, not a one-time treatment effect. Eutils abstract opened 2026-08-29; PMC PDF not this pass.
SPRINT Research Group (Wright et al.) (2015) — New England Journal of Medicine
A Randomized Trial of Intensive versus Standard Blood-Pressure Control
One trial, not the engine. 9,361 high-risk adults without diabetes; SBP target <120 vs <140. Median 3.26 y, stopped early. ACM HR 0.73 (0.60–0.90). Jaeger is the same trial after intensity lapsed. PMC HTML opened 2026-08-29.
Blood Pressure Lowering Treatment Trialists’ Collaboration (Rahimi et al.) (2021) — The Lancet
Pharmacological blood pressure lowering for primary and secondary prevention of cardiovascular disease across different levels of blood pressure: an individual participant-level data meta-analysis
Treatment IPD, primary endpoint MACE not ACM. 48 trials, 344,716 people. Per 5 mmHg SBP: MACE overall HR 0.90 (0.88–0.92); ACM 0.98 (0.96–1.01); CV death 0.95 (0.92–0.99). Context line is stroke/HF/IHD/CV-death −13/−13/−8/−5%, not ACM. OA PDF opened 2026-08-29.
Blood pressure and incidence of twelve cardiovascular diseases: lifetime risks, healthy life-years lost, and age-specific associations in 1.25 million people
CVD incidence, not ACM. CALIBER 1.25 million, 83,098 first CVD presentations. Lowest CVD risk at SBP 90–114 / DBP 60–74; hypertension lifetime CVD risk 63.3% vs 46.1%. PMC HTML opened 2026-08-29. Not the UI endpoint.
HbA1c and Risks of All-Cause and Cause-Specific Death in Subjects without Known Diabetes: A Dose-Response Meta-Analysis of Prospective Cohort Studies
Engine overlay (page meta, not the knot table). 12 studies; ACM dose-response 113,526 people / 11,301 deaths. HR per 1% 1.03 (1.01–1.04), I² 28.9%. Non-linear: curves “relatively flat for HbA1c less than around 5.7%, and rose steeply thereafter.” After excluding undiagnosed diabetes (HbA1c ≥6.5%): ACM HR 1.01 (0.99–1.03). After also excluding prediabetes: ACM HR 1.01 (0.98–1.03), NS. Authors: the association “is driven by those with undiagnosed diabetes or prediabetes.” The calculator’s PCHIP is constructed (flat 1.00 to 5.6%; 6.0=1.15); 1.15 at 6.0 is not Zhong’s 1.03/1%. EuropePMC XML opened 2026-08-29. Hunt: docs/RESEARCH-2026-08-29-hba1c-mortality-meta.md.
HbA1c levels in non-diabetic older adults - No J-shaped associations with primary cardiovascular events, cardiovascular and all-cause mortality after adjustment for confounders in a meta-analysis of individual participant data from six cohort studies
Nearest IPD ACM pool — non-diabetic ≥50, not the knot table. 6 European/US cohorts, 28,681 people, 6,769 deaths, mean 10.7 y. Increased (6.0–<6.5%) vs low (5.0–<5.5%) ACM HR 1.14 (1.03–1.27) after CV-risk adjustment (~50% of excess explained). Very-low (<5.0%) pooled NS. NHANES J lost significance after race/ethnicity, alcohol, BMI, iron-deficiency anemia and liver markers. Authors: small effects “do not support the notion of a J-shaped association.” Engine 6.0=1.15 is nearby; engine 6.5=1.35 is outside this paper. EuropePMC XML opened 2026-08-29. Engine not moved.
Glycated haemoglobin A1c as a risk factor of cardiovascular outcomes and all-cause mortality in diabetic and non-diabetic populations: a systematic review and meta-analysis
Competing number, categorical U, not this curve. 74 studies / 46 in the MA. Non-diabetic ACM >6.0% HR 1.74 (1.38–2.20); <5.0% HR 1.19 (1.04–1.36). Diabetic >9.0% 1.69 (1.09–2.66); <6.0% 1.57 (1.14–2.17). Authors: optimal 5.0–6.0% without diabetes, 6.0–8.0% with diabetes. Do not paste 1.74 onto the engine’s 1.15 at 6.0%, or 1.19 onto the flat 1.00 below 5.6% (council 2–1 against a low-end uptick). EuropePMC XML opened 2026-08-29. Engine not moved.
ACCORD Study Group (Gerstein et al.) (2008) — New England Journal of Medicine
Effects of Intensive Glucose Lowering in Type 2 Diabetes
Action / honesty. High-risk T2D, intensive target <6.0% (median achieved 6.4%) vs standard 7.0–7.9% (median 7.5%). Primary MACE HR 0.90 (0.78–1.04), NS. ACM 257 vs 203, HR 1.22 (1.01–1.46), P=0.04. Targeting a number this calculator scores as 1.00–1.15 raised death in that trial. Marker, not a DIY treatment target. 2011 follow-up (PMID 21366473) 5-year ACM 1.19 (1.03–1.38). Eutils abstract opened 2026-08-29; NEJM wall. Engine not moved.
Low hemoglobin A1c in nondiabetic adults: an elevated risk state?
ARIC, n=13,288. HbA1c <5.0% vs 5.0–<5.7%: ACM HR 1.32 (1.13–1.55); cancer death 1.47 (1.16–1.84). Authors: “generalized marker of mortality risk,” a mix of healthy people and people in whom low HbA1c signals underlying illness. Council 2026-08-29 minority (Kelly 1.05–1.10 below 5.0%); not implemented. EuropePMC XML opened 2026-08-29.
Glycated haemoglobin, diabetes, and mortality in men in Norfolk cohort of european prospective investigation of cancer and nutrition (EPIC-Norfolk)
EPIC-Norfolk, 4,662 men 45–79. Each 1% higher HbA1c: +28% ACM (P<0.002). After excluding known diabetes, HbA1c ≥7%, or prior MI/stroke: RR 1.46, P=0.05. 82% of excess population mortality at HbA1c 5.0–6.9%. Lowest rates below 5% — do not use Khaw backwards as a low-end-uptick paper. Eutils abstract opened 2026-08-29; PMC XML empty this pass.
Association between prediabetes and risk of all cause mortality and cardiovascular disease: updated meta-analysis
129 studies, 10.07 million. General-population prediabetes ACM RR 1.13 (1.10–1.17). ACM was raised for HbA1c-IEC 6.0–6.4% (1.21, 1.06–1.38) and for IFG/IGT; not for “other definitions,” which includes HbA1c-ADA 5.7–6.4%. CVD was raised for HbA1c-ADA (1.17, 1.03–1.33). Binary prediabetes ≠ this continuous curve. EuropePMC XML opened 2026-08-29.
Nonlinear association between glycated hemoglobin levels and mortality in elderly patients with non-diabetic chronic kidney disease: a national health and nutrition examination survey analysis
NHANES elderly non-diabetic CKD, n=1,931, mean age 73. Q1 3.7–5.3% ACM HR 1.48 (1.18–1.87) vs 5.7–5.8%. Wrong population for this calculator. Eutils abstract opened 2026-08-29.
Yin et al. (2017) — Journal of the American Heart Association
Relationship of Sleep Duration With All-Cause Mortality and Cardiovascular Events: A Systematic Review and Dose-Response Meta-Analysis of Prospective Cohort Studies
Engine paper. Prospective cohorts in generally healthy adults, search to December 2016. Nadir ≈7 h. Table 2 (nonlinear, 40 ACM risk estimates, reference 7 h): 8 h RR 1.04 (1.04–1.05), 9 h 1.15 (1.14–1.16), 10 h 1.32 (1.29–1.35). Linear short-sleep slope RR 1.06 (1.04–1.07) per hour below 7 h (P for nonlinearity on the short side = 0.12); linear long-sleep slope 1.13 (1.11–1.15) per extra hour. Extreme categorical pooling is 57 reports (shortest vs ref RR 1.13, longest 1.35) — not Table 2. The engine uses Table 2 on the long side and the linear 1.06/h on the short side (4/5/6 h = 1.19/1.12/1.06), not Table 2’s nonlinear 1.08/1.04/1.01.
Imbalanced sleep increases mortality risk by 14–34%: a meta-analysis
Newest categorical ACM pooling (search to October 2024): 79 cohort studies. Short sleep (<7 h vs 7–8 h) HR 1.14 (1.10–1.18); long sleep (≥9 h) HR 1.34 (1.26–1.42). Not a dose–response and not the engine curve. First author is editor-in-chief of the publishing journal. Engine not moved.
Kwok et al. (2018) — Journal of the American Heart Association
Self-Reported Sleep Duration and Quality and Cardiovascular Disease and Mortality: A Dose-Response Meta-Analysis
74-study review; the reported dose–response is 30 studies (>1 million people). Linear model: 9 h RR 1.14 (1.05–1.25), 10 h 1.30 (1.19–1.42), 11 h 1.47 (1.33–1.64). No significant increase below 7 h — that is the short-side disagreement with the engine’s 1.06/h tax. Long side sits next to Yin Table 2. Engine not moved.
Nighttime sleep duration, 24-hour sleep duration and risk of all-cause mortality among adults: a meta-analysis of prospective cohort studies
35 articles. Nighttime Table 1 vs 7 h (36 results, 146,830 deaths / 1,526,609 people): 4 h RR 1.07 (1.03–1.13), 6 h 1.01, 11 h 1.55 (1.47–1.63). 24-hour sleep is a different column (11 h RR 1.84). Short side is closer to Yin Table 2 than to the engine’s 1.19 at 4 h.
Sleep duration and risk of all-cause mortality: A flexible, non-linear, meta-regression of 40 prospective cohort studies
Closed full text (Unpaywall is_oa false as of 2026-08-29). PubMed abstract: 40 cohorts, 2,200,425 people, 271,507 deaths. J-shaped vs 7 h (24-hour sleep, n=29): 4 h RR 1.05 (1.02–1.07), 11 h 1.38 (1.33–1.44). Female short-sleep signal; male short sleep not significant in the subgroup. Not the engine paper.
Sleep duration and all-cause mortality: a systematic review and meta-analysis of prospective studies
Action / honesty. Foundational categorical pairing, not a dose–response: 16 studies / 27 cohorts, 1,382,999 people, 112,566 deaths. Short sleep RR 1.12 (1.06–1.18); long sleep RR 1.30 (1.22–1.38). This is the 1.12 / 1.30 pair the old page led with. Not the engine paper and not a sleep-extension trial. Marker, not a treatment effect.
Carbohydrate quality and human health: a series of systematic reviews and meta-analyses
Engine paper. WHO-commissioned series (185 prospective publications / 58 trials). All-cause mortality highest-vs-lowest is a 10-study observational pooling (not the 185-publication count): RR 0.85 (0.79–0.91), GRADE moderate, people without chronic disease. Fig 1A: assuming linearity, RR 0.93 (0.90–0.95) per 8 g more fibre/day (68,183 deaths). Greatest categorical benefit at 25–29 g/day (6 of 7 critical outcomes); dose-response linear with no plateau in the available data. The engine uses the 0.85 floor at/above the age/sex target; that target itself is the IOM/USDA AI (34/30/26/21 g), not Reynolds’s 25–29 g band. Erratum PMID 30712898.
Dietary intake of total vegetable, fruit, cereal, soluble and insoluble fiber and risk of all-cause, cardiovascular, and cancer mortality: systematic review and dose-response meta-analysis of prospective cohort studies
Newer matching-construct ACM meta (search to August 2023). Highest vs lowest: 16 studies, RR 0.81 (0.77–0.86), I² 71.9%. Per 10 g/day: 14 studies, RR 0.90 (0.86–0.93), I² 86.1% — the 0.90 in the abstract is this increment, not highest-vs-lowest. Non-linear (P=0.0096): steeper below 15 g/day, more gradual after. Same 15 g place as the engine’s kink; Yao does not set HR=1.0 there. Engine not moved onto 0.81 / 0.90.
Total and different dietary fiber subtypes and the risk of all-cause, cardiovascular, and cancer mortality: a dose-response meta-analysis of prospective cohort studies
Action / honesty. 28 prospective studies, 1,613,885 people, general population. Higher total and most subtypes (cereal, vegetable, legume, soluble, insoluble) associated with lower ACM; fruit fibre was not. Subtype HRs 0.77 (insoluble) to 0.93 (legume). Inverse dose-response for total fibre vs ACM / CVD / cancer death; GRADE moderate for those total-fibre associations. Abstract does not report a single total-fibre highest-vs-lowest HR. Calculator collects total grams, not subtype — moving the slider is not “eat fruit fibre.”
Yang et al. (2015) — American Journal of Epidemiology
Association between dietary fiber and lower risk of all-cause mortality: a meta-analysis of cohort studies
Pre-Reynolds tertile pooling: 17 cohorts (982,411 people, 67,260 deaths). Top vs bottom tertile ACM RR 0.84 (0.80–0.87), I² 41.2%. Per 10 g/day RR 0.90 (0.86–0.94), I² 77.2%. Same per-10 g point estimate as Yao, older and smaller. Not the engine paper.
We correct calibration issues when we find them, and we disclose them here rather than waiting for someone else to find them first. Corrected 2026-07-25.
Calibration was corrected
Earlier versions anchored several curves’ "no added risk" point (HR = 1.0) at values well below what a typical adult actually does — the daily-steps curve, for example, treated 2,000 steps/day as the neutral baseline, when that is closer to the bottom of the range than the middle. The effect was that ordinary results looked better than they were. Scores are now expressed relative to a defined median reference profile for a 45-year-old US adult, so matching that profile on every marker scores exactly 50 — typical, not a test out of 100. See "Fifty isn’t the middle of the pack" below for why that is a higher bar than it sounds.
What the per-metric hazard ratios mean
Where a study reports a usable number at a named exposure, that anchor is the paper’s figure against that paper’s own comparator. Interiors of the PCHIP curves, and the constructions listed in the next card, are ours — direction from the papers, specific shape not. Recalibration of the composite left those per-metric numbers in place so the published anchors stay traceable. Only the composite score and percentile are expressed relative to the median reference.
More curve shapes are our own construction than three
Grip strength, VO2 max, and HRV are the three we previously named: Leong reports only a per-5kg slope, Mandsager uses age/sex percentile bins, Jarczok compares quartiles (its HRV curve was rescaled 2026-08-29 to Jarczok’s covariate-adjusted estimate rather than the unadjusted one). Also constructed, and now named: the daily-steps PCHIP interiors (0 / 3k / 4k / 5k / 10k / 12k / 16k — only 2,000 = 1.0 and 7,000 = 0.53 are Ding 2025); fibre’s 15 g kink, the slope below 15 g, and the extra −0.005/g past the age/sex target (only the 0.85 floor is Reynolds 2019 highest-vs-lowest; the 34/30/26/21 g targets are IOM/USDA Adequate Intakes, not Reynolds’s 25–29 g greatest-benefit band); sleep duration’s short-side PCHIP anchors (4 h = 1.19, 5 h = 1.12, 6 h = 1.06) apply Yin 2017’s linear 1.06 per hour below 7 h, not Table 2’s nonlinear 1.08 / 1.04 / 1.01 (long-side 8 / 9 / 10 h are Table 2); gait speed’s 1.2 m/s recentre for a ~45-year-old (Studenski measured adults 65+); sleep-regularity’s log-linear fill between Windred’s 1.0 and 0.70 quintile endpoints (0.70 is the fully adjusted most-regular HR; 0.52 is the same contrast in the minimal model, not used; interiors 0.91 / 0.84 / 0.76 are ours, not Table 2); social’s three-level 0.85 / 1.0 / 1.25 (Strong / Average / Weak) is constructed around typical = 1.0 — not Wang 2023’s isolated-vs-not 1.32, and 0.85 is not 1/1.32 (isolation, loneliness, and living-alone are three separately pooled constructs); and the blood-pressure combination rule when systolic and diastolic disagree. Dental (poor HR 1.15, Wu 2021) and hearing (loss HR 1.13, Tan 2022) are directly-sourced figures, not constructed. Direction comes from the papers; those specific shapes are ours.
Balance is one clinic cohort, not the 2024 meta
Fail HR 1.84 is Araujo 2022 (CLINIMEX clinic volunteers, 10-second one-legged stance — the same test the calculator asks). Das et al. 2024 pooled 15 cohorts of mixed static-balance tests at HR 1.14 (1.07–1.21), I² 88%; that pool is not one-leg-only. Xie 2023 (CHARLS) is a 10-second semi-tandem stand, a different and easier test (1% failed vs Araujo’s 20%). The engine has not been moved off Araujo; a council brief is written, not voted.
The percentile comes from a simulation, not measured people
The score-to-percentile conversion used to assume population scores follow a bell curve with a fixed spread. It now looks up a 100,000-profile-per-sex simulation built over the same published marginal distributions as the curves above — a real improvement, but still not measured population data. Each simulated profile samples its 8 markers independently, so it misses the real correlation between them (someone with a high VO2max also tends to have a lower resting heart rate and better blood pressure), which means the simulation understates how spread out the real population actually is. Treat the percentile as a rough band, not a precise rank.
Fifty isn’t the middle of the pack
The reference profile behind score 50 is median on every one of the 8 markers — but almost nobody actually is median on all 8 at once. Because the hazard-ratio curves punish a bad value harder than they reward an equally-sized good one, matching that all-median profile beats a large share of the independently-sampled simulated population, not 50% of it. A copula check that lets markers correlate widens the spread; treat the percentile as a band, not a precise rank.
The roadmap projects, it does not promise
Roadmap point gains are computed by re-running this same model with one input changed — so they can never drift away from your score — but they are arithmetic, not outcomes. They assume you hold every other metric where it is, that a change you make is genuine rather than a measurement artefact, and that the association the underlying study observed would apply to you. Observational cohorts show that people with better numbers die less often; they cannot prove that moving your number moves your risk by the amount shown.
Self-reported inputs carry real error
Grip strength, VO2 max, HRV, and sleep regularity in particular depend on the user’s equipment and technique, and are only as accurate as what the user measures and enters.
This is not a medical device
This is a general-wellness and educational tool. It is not a diagnostic device, it does not provide medical advice, and it must not be used to make treatment decisions. Discuss your results with a clinician.
Reference-profile changelog
2026-08-29.A three-person science council voted 3-0 on four changes. Hearing-loss HR moved from an uncited 1.25 to Tan et al. 2022's pooled 1.13. Dental was redefined from a borrowed clinical-periodontitis HR (1.30) to a self-reported-tooth-loss HR of 1.15 (Wu et al. 2021, the only cited cohort using this app's own self-report modality) — the question itself changed from gum-disease status to natural tooth loss; the underlying 'good'/'poor' values and shareable-link encoding did not. HbA1c's low-end reward (HR 0.88 below 5.4%) was removed — Schöttker et al. 2016 and Zhong et al. 2016 found that association was inconsistent, non-significant, or driven by undiagnosed diabetes — so the curve is now flat (HR 1.0) up to 5.6% rather than U-shaped. HRV's 10ms endpoint moved from a constructed 1.40 (the engine never sat at Jarczok's unadjusted 1.56) to Jarczok et al. 2022's covariate-adjusted lowest-quartile estimate (~1.23); the other HRV anchors were pulled toward 1.0 at 33ms with per-point log-space factors of 0.587–0.652, not one uniform scale. HRV's weight within the Recovery domain fell from 0.5 to 0.3 (hearing and dental each rose from 0.25 to 0.35 to renormalize). Re-deriving the distribution after these changes moved the score-band cutoffs from 38/46/51/57/63 to 39/47/51/57/63, and the median reference profile's percentile from ~p75.4-75.8 down to ~p74.6-75.1 (male 74.636, female 75.058). Any score computed before this change is not directly comparable to one computed after it.
2026-07-29. The gait-speed addition below was revised after an internal review flagged two problems with the first version. First: the 1.2 m/s reference point is an unsourced judgment call, not a measured figure— Studenski et al. 2011 only measured adults 65+ (pooled mean 0.92 m/s); 1.2 m/s is this app's own recentring for a younger adult, now called out explicitly here, next to the gait-speed slider in the calculator, and in the source note below, rather than being folded silently into the model. Second, the weighting was fixed:the first version funded gaitSpeed's 0.15 weight in the Physical Capacity domain partly by cutting grip strength from 0.15 to 0.10 — grip strength has comparably strong, arguably more broadly-replicated all-cause-mortality evidence, so this read as favoring a newer, more speculative metric over an established one. Grip strength is restored to 0.15; the room for gaitSpeed now comes only from gaitSpeed's own initial allocation (0.15 → 0.10), leaving vo2max, steps, resting heart rate, and balance unchanged throughout. Re-deriving the distribution after this weight change moved the score-band cutoffs from 38/47/52/57/65 to 38/46/51/57/63 (close to the original pre-gait-speed 38/46/52/57/64), and the median reference profile's percentile from ~p73.7-74.1 back up to ~p75.4-75.8. Any score computed before this change is not directly comparable to one computed after it.
2026-07-29. Added usual gait speed (walked over a measured ~4-meter distance) as a 17th metric, in the Physical Capacity domain. Source: Studenski et al. 2011 (JAMA305(1):50-58, PMID 21205966), a pooled analysis of 9 cohorts covering 34,485 community-dwelling adults 65 and older, one of the most validated all-cause-mortality predictors in geriatrics — age-adjusted HR for death of 0.88 (95% CI 0.87-0.90) per each 0.1 m/s higher gait speed. Because that cohort is older adults (mean age 73.5) and this app models a ~45-year-old reference population, the population median used for scoring (1.2 m/s) is a recentred judgment call rather than the source's own 0.92 m/s pooled mean — documented in full in lib/model.ts and scripts/simulate-population.ts. Adding a 17th metric shifted the simulated distribution slightly: score-band cutoffs moved from 38/46/52/57/64 to 38/47/52/57/65, and the median reference profile's percentile moved from ~p74.4-75.7 to ~p73.7-74.1. Any score computed before this change is not directly comparable to one computed after it.
2026-07-29.A literature-sourcing pass checked every ASSUMED spread parameter behind the simulated population used for percentiles and score bands. Most held up as the best defensible estimate and were left unchanged with stronger citations. One did not: VO2max's population SD was based on an inferred ACSM percentile-band width (6 mL/kg/min for both sexes); the FRIEND registry (Kaminsky et al. 2015, a large directly-measured cardiopulmonary exercise test cohort) reports the real figure as 9.6 (male) / 7.7 (female) — meaningfully wider. That widened the whole simulated distribution: the score-band cutoffs shifted (p80, previously 50, is now 52), and the median reference profile's percentile moved from ~p78.5-80.1 down to ~p74.4-75.7. Any score computed before 2026-07-29 is not directly comparable to one computed after it.
2026-07-27.The median reference profile behind every score was re-anchored on two markers. Daily steps moved from 5,500 steps/day (the same value for both sexes, which had no supporting source) to 8,500 (male) / 7,000 (female), based on NHANES accelerometer data and the All of Us Research Program. Grip strength moved from 42 kg (male) / 26 kg (female) to 41 kg / 29 kg, based on dominant-hand, best-of-trials norms from a NIH Toolbox study — and the model now states explicitly, for the first time, that it uses the dominant-hand convention throughout. VO2 max was reviewed against two disagreeing primary sources and deliberately left unchanged at 38 (male) / 31 (female), with the reasoning for that choice now written up in the model's source comments. Because the reference profile moved, the score-band cutoffs and the percentile shown next to a score also shifted slightly. Any score computed before 2026-07-27 is not directly comparable to one computed after it — re-run your numbers if you want an apples-to-apples comparison.
How much would the score move under different weights?
The weights above are one defensible allocation, not the uniquely correct one. To make that tradeoff inspectable rather than a single judgment call, we recompute the composite for a panel of 26 representative profiles (median-profile and single-metric deviations, both sexes) under four alternative weight sets: equal-weight across domains and members, cardiometabolic-heavy (50/30/20), physical-heavy (55/30/15), and one that swaps the sleep/fiber emphasis inside Lifestyle. Full method and numbers in scripts/weight-sensitivity.ts (run with npm run weight-sensitivity). The live model uses three domains and eight scored metrics.
The largest single swing observed, across every profile and every alternative weighting, is 11.0 points — a very-high-HbA1c profile scores 11 points lower under cardiometabolic-heavy weighting than under the current weights, because that alternative lifts the Cardiometabolic domain from its current 35% to 50% and HbA1c is the metric carrying that profile down. The next largest is 9.0 points: a very-high-VO2max profile scores 9 points lower under equal-weighting, which cuts VO2 max from its current 0.375 share of Physical Capacity to a uniform 0.25 and the domain itself from 41% to 33%. Averaged across all 26 profiles, the mean absolute swing is 2.1–2.7 points for the three domain-level alternatives. Swapping the sleep/fiber emphasis inside Lifestyle is by far the smallest — at most 1 point, 0.2 on average — since it reorders two members inside the lightest domain rather than touching the domain weights themselves.
This is presented as a bound on the current weights' influence, not a defense that they are uniquely correct: a handful of reasonable people could disagree on the weighting and move any individual score by about half a score band (5 points) on average, or just over two bands in the worst case we tested. Worth holding next to a fact from the roadmap — one realistic lifestyle change is typically worth 1–4 points. For the profiles at the extremes, the choice of weighting can matter more than the change you make.
Disclaimer. This tool is for educational purposes only and is not medical advice, diagnosis, or treatment. Consult a qualified healthcare professional about your own health. Back to the calculator →