The approach in one paragraph
We don't predict mortality. We predict physiological age - the age at which someone with median health metrics would carry the same readings you do. Each of 13 domains emits a "years older or younger than chronological" signal, anchored to published research where it exists. Where a weight comes from a published all-cause hazard ratio, it is converted to years with the Gompertz mortality doubling time (about 11.5 × ln HR), so the result reads as a mortality-equivalent effective age; it is not a model trained on death records. Domains compose additively, with safety clamps so no single signal can dominate, and a confidence score so a sparse-data user knows their estimate is provisional. Day-to-day stability is enforced by an exponential moving average (α = 0.35) that's regression-tested to keep variance under 0.5 years on realistic noisy input.
This page is the public version of our internal methodology document. It's published - rather than kept internal - because the alternative is a black-box bio-age number, and health-tech apps that hide their math should not be trusted.
How the weights are sourced. Of roughly 50 weighted coefficients, 13 carry an effect size from a published outcome study. VO₂ max, resting heart rate and former smoking land close to what their studies imply; most of the others are deliberately weighted below it, so that correlated signals (VO₂, steps, resting HR, HRV) don't double-count in an additive model. Most of the rest follow clinical guideline thresholds or reference ranges, and a few are our own product judgement.
1 · Fitness
VO₂ max - the strongest single signal
Predicted VO₂ max for your age, sex, and weight uses the Jackson non-exercise estimate. Years delta is then −1.61 · (measured − expected) / 3.5, clamped to ±6 years.
Mandsager K et al. Association of Cardiorespiratory Fitness with Long-term Mortality Among Adults Undergoing Exercise Treadmill Testing. JAMA Network Open 2018;1(6):e183605.
A 1-MET (3.5 ml/kg/min) advantage over expected mapped to ~12% lower all-cause mortality; converted to ~1.6 years of physiological-age equivalent.
Steps and active minutes
Steps score on a smooth curve against the log of your daily average: up to +3 years at or under 3,000 steps, zero at 8,000, then credit, up to −2 years from 12,000. From 65 the same shape sits lower (zero at 6,000, −2 years from 9,000), and ages 55–65 blend between the two, so neither one extra step nor a birthday moves the number in a jump.
Saint-Maurice PF et al. Association of Daily Step Count and Step Intensity with Mortality Among US Adults. JAMA 2020;323(12):1151-1160.
Cumulative fitness - capacity + behavior
VO₂ measures cardiorespiratory capacity (peak aerobic potential); steps and active minutes measure behavior (habitual movement). They have meaningfully independent contributions.
Stamatakis E et al. Joint associations of objectively-assessed physical activity and cardiorespiratory fitness with all-cause and cardiovascular mortality. BMJ 2019;366:l4570.
Combined high-fitness + high-activity stratum had the largest mortality reduction (HR ≈ 0.45 vs sedentary low-fit). "High-fit / low-active" was substantially worse than "high-fit / high-active".
The engine combines:
| Driver | Range | When it fires |
|---|---|---|
| VO₂ max | ±6y | Always preferred when present |
| Cycling FTP | ±6y | Only when VO₂ absent |
| Steps + active-min | ±2.5y full / ±1.25y discounted | Always when data present |
| Self-reported activity | ±2y | Only when neither capacity nor behavior fires |
| Zone 2 minutes bonus | −0.4y additive | When any primary path fired |
When both capacity and behavior are present, behavior is discounted by min(0.5, |capacityDelta| / 6.0) - Stamatakis 2019's ~50% overlap. The combined sum is clamped at ±7y.
2 · Cardiovascular & autonomic
Blood pressure
Bands match the ACC/AHA 2017 hypertension guideline. The protective athletic-low band (90-114 / 60-74) was added on the back of the SPRINT trial.
| Reading | Years delta |
|---|---|
| 90-114 / 60-74 (athletic-low) | −0.3 |
| <120/80 (normal) | 0 |
| 120-129 / <80 (elevated) | +0.5 |
| ≥130/80 (Stage 1) | +1.5 |
| ≥140/90 (Stage 2) | +3.0 |
| >180/120 (severe) | +4.5 |
Whelton PK et al. 2017 ACC/AHA Guideline for High Blood Pressure in Adults. JACC 2018;71:e127-e248.
Resting heart rate
Bands progress from −1.0y at ≤55 bpm to +3.3y at ~100 bpm.
Saxena A et al. Resting Heart Rate and Mortality. Eur Heart J 2013 - pooled analysis: hazard ratio ~1.09 per 10 bpm above 65.
Heart rate variability (RMSSD)
Hillebrand S et al. Heart Rate Variability and First Cardiovascular Event in Populations Without Known Cardiovascular Disease: meta-analysis and dose-response. Europace 2013;15:742-749.
3 · Sleep & recovery
Sleep duration uses an inverted-U curve with its optimum at 7-8.5h, with bands grounded in:
Yin J et al. Relationship of Sleep Duration With All-Cause Mortality and Cardiovascular Events: A Systematic Review and Dose-Response Meta-Analysis of Prospective Cohort Studies. J Am Heart Assoc 2017;6:e005947.
Sleep efficiency (% of time-in-bed actually asleep), bedtime regularity, and time-in-deep-sleep + REM each contribute additional small deltas. The sleep domain as a whole is bounded at −1 / +5 years.
4 · Body composition
Body fat % bands are sex-specific and preferred when available. BMI is the fallback, following the J-curve: a small protective band at 22-24, penalties below 18.5 and above 25. Waist circumference is a separate visceral-fat signal using the WHO absolute thresholds: 94 / 102 cm for men and 80 / 88 cm for women (above the first is a small penalty, above the second a larger one).
Flegal KM et al. Association of all-cause mortality with overweight and obesity using standard BMI categories: a systematic review and meta-analysis. JAMA 2013;309(1):71-82.
5 · Cognitive
The in-app cognitive battery (reaction time, word recall, spatial memory, symbol match) emits a small (±2y) bonus or penalty, weighted by how much the user's score deviates from age-matched norms. This is a deliberately modest contribution - single-session cognitive scores have high variance.
6 · Physical function
Grip strength, sit-to-stand reps, and balance time. Each is a strong independent predictor of mortality in older adults; in younger users they're smaller signals.
Leong DP et al. Prognostic value of grip strength: findings from the Prospective Urban Rural Epidemiology (PURE) study. Lancet 2015;386(9990):266-273.
Each 5kg decrement in grip strength was associated with a 16% increase in all-cause mortality.
7 · Lab biomarkers (AI-aggregated)
When you upload a lab report (PDF or photo), an AI pass reads it. Four markers are pulled out and scored individually with published bands (HbA1c, ApoB, Lp(a) and fasting insulin; see §9). Separately, the AI returns one aggregate years estimate for the report as a whole, which is how an epigenetic-clock or telomere result enters the model. That aggregate is weighted by how recent the test is (full weight under 6 months, dropped after 18) and capped at ±15 years.
8 · Inflammation (hsCRP)
High-sensitivity CRP is the single inflammation marker in the model.
Ridker PM. A Test in Context: High-Sensitivity C-Reactive Protein. JACC 2016;67(6):712-723.
hsCRP <1 mg/L low risk; 1-3 mg/L average; >3 mg/L elevated; >10 mg/L acute-process flag (excluded from chronic-disease estimation).
9 · Lab markers - structured per-marker
Four markers are scored individually, each with its own band table and the same 18-month freshness weight. Together they are capped at −1.5 / +5 years.
ApoB
Sniderman AD et al. Apolipoprotein B Particles and Cardiovascular Disease: A Narrative Review. JAMA Cardiol 2019;4(12):1287-1295.
ApoB outperforms LDL-C as a CV-risk predictor; targets <80 mg/dL standard, <60 mg/dL aggressive.
Bands (mg/dL): <70 −0.3y · 70-99 0 · 100-129 +0.7y · ≥130 +1.5y.
HbA1c
Selvin E et al. Glycated Hemoglobin, Diabetes, and Cardiovascular Risk in Nondiabetic Adults. NEJM 2010;362:800-811.
Bands (%): <5.7 0 · 5.7-5.9 +0.4y · 6.0-6.4 +0.8y · 6.5-7.4 +1.5y · 7.5-8.9 +2.5y · ≥9.0 +3.5y.
Lp(a)
Genetic, single-test. Bands (mg/dL): <30 0 · 30-49 +0.3y · 50-99 +0.8y · ≥100 +1.5y. Kronenberg F et al. Lipoprotein(a) in atherosclerotic cardiovascular disease and aortic stenosis: a European Atherosclerosis Society consensus statement. Eur Heart J 2022;43(39):3925-3946.
Fasting insulin
Bands (μIU/mL): <10 0 · 10-14 +0.4y · 15-19 +0.8y · ≥20 +1.2y. Fasting hyperinsulinaemia tends to precede a rise in glucose. Reaven GM. The metabolic syndrome or the insulin resistance syndrome? Endocrinol Metab Clin North Am 2004;33(2):283-303.
Other markers on a report (vitamin D, ferritin, liver and kidney panels) are not scored individually by the engine.
10 · Genetic (APOE ε4)
When a user uploads 23andMe / AncestryDNA raw data, the engine reads APOE ε4 status. e3/e3 baseline; one ε4 copy +0.5y; two copies (e4/e4) +2.0y. This is a deliberately small contribution - APOE is correlative, not deterministic.
Liu C-C et al. Apolipoprotein E and Alzheimer disease: risk, mechanisms and therapy. Nat Rev Neurol 2013;9:106-118.
11 · Family history (parental longevity)
Self-reported parental longevity, scored as one whole-family category: both parents lived or are living past 90, −1.0y; both past 80, −0.5y; typical, 0; one parent died before 70, +0.5y; both parents died before 70, +1.5y.
Sebastiani P et al. Families enriched for exceptional longevity also have increased health-span. Front Public Health 2013;1:38.
12 · Mental health
PHQ-9 (depression), GAD-7 (anxiety), and PSS-10 (perceived stress) - each user-administered in-app. Major depression is associated with ~+1.7y physiological age; severe anxiety adds another small layer; high perceived stress is a separate signal.
Han LKM et al. Contributing factors to advanced brain aging in depression and anxiety disorders. Translational Psychiatry 2021;11:402.
13 · Clinical flags
Smoking: current +8y, former +2y (no quit-year tiers). Chronic conditions: +1 / +2 / +3y for one, two, or three or more reported, plus up to +2y more for the most serious cardiometabolic ones (diabetes or pre-diabetes, heart disease, and high blood pressure when no measured BP is already scored). The clinical domain is capped at +10y. Severe BP (>180/120) also surfaces with a "seek prompt review" message - at that level, the algorithm should not silently average the data into a score.
Composite, caps, and smoothing
Domains compose additively, with two safety mechanisms:
- Per-domain caps. Each domain has its own bound: fitness ±7 years, cardiovascular −2 / +6, sleep −1 / +5, cognitive ±2, physical function −2.5 / +3.5, mental health 0 / +2.5, the four structured lab markers −1.5 / +5, clinical −0.5 / +10. The AI lab aggregate is the widest at ±15, because a measured clock can genuinely sit that far out. Prevents one outlier signal from dominating.
- Two buckets, then a total clamp. The lifestyle domains (fitness, cardiovascular, sleep, body composition, mental health, clinical) are summed and capped by age: −3 / +8 years at 22 and under, widening in a straight line to −18 / +20 at 60 and over, because a 25-year-old can't plausibly be a decade younger on lifestyle alone. Measured biology (labs, inflammation, cognitive, physical function, genetic, family history) is summed and clamped at ±25. The combined result is bounded at −30 / +35 years, and bio age never goes below 18.
Day-to-day stability is enforced by an exponential moving average (α = 0.35) that's regression-tested on noisy realistic input. The smoother lives in BioAgeSmoother.swift and is mirrored on the server as a Postgres RPC for cross-platform parity.
Confidence & uncertainty
Every reading carries a confidence score (0-1) computed from how many domains had data. Sparse-data users see a "provisional" badge on their bio age until enough domains light up. The confidence breakdown is visible in the in-app Bio Age Breakdown sheet, alongside the per-domain contribution chart.
What this isn't
- Not a mortality prediction. Mortality models (e.g. Klemera-Doubal, PhenoAge) train on death-event data; we don't. Where our years come from published hazard ratios, they are translated into a mortality-equivalent effective age, not a prognosis.
- Not yet validated against outcomes or an epigenetic clock. Since September 2026 the app logs, for members who upload a DNA-methylation result, how far the engine (with that lab result left out) sits from their clock age. We'll publish those numbers once the sample is big enough to mean something.
- Not a methylation clock. Methylation clocks (Horvath, GrimAge) measure DNA methylation directly. This model uses behavioral, physiological, and biomarker proxies.
- Not a disease diagnosis. A high bio age doesn't mean you have a condition; it means signals are pointing in a direction worth understanding.
- Not personalised therapy. Recommendations are population-level. For anything clinically significant, see a doctor.
Methodology v1.4, last revised 2026-10-03. The full coefficient-level document, with every band and citation, is available on request: hi@thier.app.