Modelled ceiling — not a personal forecast
The headline figure on this page is an upper bound under best-case assumptions; a conservative floor (the lower confidence bound, or the same model run at a higher cross-correlation) is materially lower. Any individual’s result is unknown and may fall anywhere in — or below — this range. This is a population model, not a prediction about you. Discuss with your physician before acting.
What this number assumes
  • Combines interventions each studied separately, added as if independent (log-additivity); real combinations may overlap or compete, which lowers the total.
  • The cross-correlation discount (ρ̄ = 0.30) is an assumption, not a measured per-pair value; a higher value lowers the estimate.
  • Some inputs are observational hazard ratios; for those, residual confounding can inflate the apparent effect.
  • The life-table step assumes the modelled reductions persist across the population age/sex structure.
  • It is a population-model ceiling, not a forecast for any individual.

Pearl Bayesian ACM v4.4 — Age & Sex Adaptive v10

Stuart J. Berkowitz · Pearl causal-DAG reasoning · log-additive HR aggregation (ρ̄-discounted ceiling) · Statistics Canada life tables · GBD 2019 pathway weights
Demographics
72

ACM reduction by pathway (age/sex-weighted)

Intervention contributions (top 18)

Model comparison

Causal DAG — all-cause mortality. The structural model behind every figure in this atlas: modifiable intervention groups act on eight mechanistic pathways (mediators), and each pathway carries its share of all-cause mortality — the edge width into the endpoint is proportional to that GBD-2019 share (representative, ~age 65). Interventions loading on the same pathway overlap, which is exactly what the pathway-ρ correction discounts. Left-side edge width ∝ each group’s aggregate loading on that pathway.
INTERVENTION GROUPS MECHANISTIC PATHWAYS ENDPOINT Exercise (3) Thermal (1) Diet (3) Sleep (1) Diagnostics (1) Cardiovascular (2) Senolytics (4) Anti-infective (1) Supplements (16) Medications (3) Lifestyle (3) Antioxidants (7) Cardiovascular 32% of mortality Oncological 24% of mortality Neurological 11% of mortality Metabolic 11% of mortality Immune / Inflam. 8% of mortality Cell Senescence 6% of mortality Frailty / Musc. 4% of mortality Oxidative / Detox 4% of mortality All-Cause Mortality
Cardiovascular (32%) and oncological (24%) pathways dominate all-cause mortality, which is why interventions that hit them — cardiorespiratory fitness, non-smoking, Mediterranean diet — sit atop the Pareto frontier. This is a schematic of the causal DAG (top-two pathway edges per group shown for legibility), not an exhaustive edge list.

Cause-specific ACM pathway breakdown (GBD 2019, age/sex-adjusted)

Emax saturation: Hill-equation dose-response for each pathway. Vertical bar = EC50. Stuart's protocol saturation levels shown relative to pathway Emax.
Cross-correlation ρ_ij derived from dose-response saturation studies. Red = high redundancy. Green = independent.

Pairwise ρ_ij — source pairs (with CI & evidence grade)

Derived pathway ρ — what the model actually uses

Combined-benefit aggregation mode

Effective dimensionality of the active stack (eigenvalue)

Pathway loading heatmap

Statistics Canada life table q(x) — annual ACM probability by age and sex

GBD 2019 cause-specific pathway weights by age and sex

Sex-adjusted HR modifiers applied to specific interventions

Pearl counterfactual: ACM increase if X removed, all others held constant. Uses age/sex-adjusted baseline and pathway weights.

If not for X — % ACM increase

Marginal benefit of inactive interventions, corrected for pathway saturation at current age/sex-weighted pathway loads.

Marginal benefit if added

Combined HR vs global ρ scale

E-values (VanderWeele & Ding 2017)

Pathway ρ ±0.15 tornado

Search papers:
Intervention · Academic papers (expand ▸) Base HRSex HRnEvE-valρ disceff.logRR
Methodology summary. Interventions are organised on a Pearl-style causal directed acyclic graph (DAG) used to reason about pathway structure and shared confounding. The combined all-cause-mortality (ACM) figure is then produced by log-additive aggregation of each intervention's −log(hazard ratio), shrunk by per-pathway redundancy discounts to avoid double-counting overlapping mechanisms. Each pathway's ρ is derived as the load-weighted mean of the pairwise ρ_ij shown in the Cross-Corr tab (so the displayed matrix is the source of the number the model uses), with a 95% band set by each pair's A–D evidence grade; where a pathway's ρ has been estimated from joint cohort data (NHANES §07) that measured value overrides the literature-derived one. This is an effect-aggregation heuristic — not a backdoor/front-door–identified estimate: no joint individual-level dataset is used, so the headline is a modelled ceiling, not a do-calculus–identified causal quantity. Except where marked measured, the pairwise ρ_ij are literature/mechanism-derived, not estimated from a joint individual-level dataset — the ρ-sensitivity band sweeps each pathway across its own derived 95% CI to show how the headline moves. Age- and sex-stratified baselines: Statistics Canada Life Tables 2020–2022; cause-specific pathway weights: Global Burden of Disease (GBD) 2019 Canadian decomposition.

Core identities

naive_logRR(stack) = Σᵢ −log(HRᵢ) Pearl_logRR(stack) = Σᵢ −log(HRᵢ) · (1 − ovlpᵢ) where ovlpᵢ = Σ_pathway ρ_pathway · min(loadᵢ_pathway, accumulated_pathway) HR_combined = exp(−Pearl_logRR) ACM_posterior = HR_combined × HR_risk × q_baseline(age,sex)
Sex-adjusted HR (protective interventions, female): HR_F = HR_base + (1 − HR_base) · (1 − adj_factor_F) → HR drifts toward 1.0 by adj_factor_F proportion of the protective gap.
Proportional-hazards life expectancy: S_adj(t) = S_base(t)^HR → q_adj(x) = 1 − (1 − q_base(x))^HR e(x) = Σ_t S(t) − 0.5 (curtate→complete actuarial correction; Bowers §5.3)
Bio-age forward life expectancy: bioAge = q⁻¹(HR · q_base(chronAge), sex) e_bioFwd = e(bioAge, sex, HR=1.0) → projected death age = chronAge + e_bioFwd

Data sources & tools

DomainSourceApplication
Baseline ACM q(x)Statistics Canada Life Tables 2020-2022 (Table 13-10-0114-01)Age- and sex-specific annual death probability
Pathway weightsGBD 2019 — Institute for Health Metrics and Evaluation (IHME)Canadian cause-of-death fractions by age and sex across 8 pathways
Intervention HRsPeer-reviewed meta-analyses and RCTs (122 papers indexed in Evidence tab)Base hazard ratio with 95% CI for each intervention
Cross-correlation ρEmax dose-response saturation studies (Riedl 2017; Xu 2018; Ravussin 2015 CALERIE)Pathway-level redundancy discount
VO₂max referenceFRIEND Registry (Kaminsky 2022, n=750,302)Age/sex-stratified 50th-percentile VO₂max for ΔMETs; Mandsager 2018 HR=0.855^ΔMETs
Sex modifiersUSPSTF 2022 (aspirin); CTT Collaboration sex-stratified (statin); Laukkanen 2018 KIHD (sauna)Female-specific HR adjustment for 3 interventions
E-valuesVanderWeele & Ding 2017 Ann Intern MedRobustness to unmeasured confounding: E = RR + √(RR(RR−1))

Key assumptions and their sensitivity

AssumptionJustificationSensitivity
Multiplicative HR composition on log scaleStandard Cox PH composition; assumes log-additivity of independent effectsρ-discount removes overcounting from non-independence
Pathway loads sum to 1.0 per interventionMechanism-of-action allocation across 8 GBD pathwaysAllocation uncertainty propagated through ±0.15 ρ tornado in Sensitivity tab
Independence on logRR scale for stack SEConservative — correlation would narrow CI±1 SD bands shown in CI / SD tab stack uncertainty chart
q(x) ≥ 95 extrapolated by Gompertzβ = ln(q₉₅/q₉₀)/5 fitted to top life-table anchorNegligible impact for users below age 90
Curtate→complete correction = −0.5Trapezoidal rule, Bowers et al. Actuarial Mathematics §5.3±0.5 yr precision on life expectancy
Sex modifiers shift HR toward 1.0Most sex-stratified data shows attenuation, not reversalOnly 3 interventions affected (aspirin, statin, sauna)
Causal DAG for VO₂max decline. Click any node below to see its causal mechanism, evidence, and role in the causal DAG. Key structural finding: physical inactivity is the central modifiable hub through which most "age effects" operate, not a peripheral risk factor.

Interactive causal DAG — click any node

VO₂max — Pearl-style Causal DAG Roots → exposures → mediators → outcome · dashed = feedback loop EXOGENOUS ROOTS non-modifiable LIFESTYLE EXPOSURES CENTRAL MODIFIABLE HUB MEDIATORS (downstream of inactivity) DIRECT CAPS OUTCOME 🧬 Genetics ≈50% heritable ⏱ Age indirect via mediators ⚥ Biological sex 18% (Hb, SV) Smoking → CVD, COPD Diet quality → obesity, T2DM Air pollution PM₂.₅ → COPD ⭐ Physical inactivity CENTRAL MODIFIABLE HUB ~60% of "age effect" routes through here Obesity 18% · 85% via PA Sarcopenia 16% · 80% via PA ⭐ Osteoarthritis 14% · 88% via PA Type 2 diabetes 12% · 75% via PA Depression 7% · 92% via PA Sleep apnea 8% · 75% via obesity CVD / heart failure 22% · 60% intrinsic COPD 20% · 65% intrinsic Anemia / low Hb 10% · direct cap 🎯 VO₂max cardiorespiratory fitness hub causation mediator path exposure path feedback loop indirect (via mediators)
How to read this DAG: Roots are exogenous (genetics, age, sex). Lifestyle exposures act primarily via mediators. The central red node is the modifiable hub — most of what we call "age effects" routes through it. Dashed coral arrows show feedback loops where mediators feed BACK to the hub, accelerating decline. All paths terminate at the VO₂max outcome.
← Click any node in the diagram above to see its causal mechanism, evidence, and role in the VO₂max causal DAG
The key geroscience question: How much of what we call "aging" is actually accumulated physical inactivity? On average across all components, ~60% of the apparent "age effect" on VO₂max is inactivity-mediated and theoretically reversible. The remaining ~40% is truly intrinsic to biological time — irreversible senescence driven by somatic mutations, telomere attrition, AGE cross-linking, and SA-node decline.

Decomposition of each "age effect" on VO₂max — what fraction is inactivity vs intrinsic?

Total VO₂max variance — true root attribution

After correctly routing inactivity-mediated paths through the central hub node.

Inactivity-mediated fraction by component

Red = inactivity-driven (modifiable). Blue = intrinsic aging (not reversible).

Three self-amplifying feedback loops: Physical inactivity doesn't just cause downstream conditions — those conditions feed back to cause MORE inactivity, creating vicious cycles that accelerate VO₂max decline exponentially with age. This is why the decline accelerates after 65 (multiple loops all active simultaneously) and why early intervention is disproportionately valuable.

Loop 1: Inactivity → Obesity → OA → more Inactivity

Inactivity → weight gain → mechanical OA loading → joint pain → exercise avoidance → more inactivity. DPPRG 2002: breaking this loop with exercise reduced obesity-related OA risk 58%.

Loop 2: Inactivity → Sarcopenia → more Inactivity

Reduced muscle mass limits exercise capacity, which reduces motivation and ability to exercise, accelerating further sarcopenia. Goodpaster 2006: this loop drives 80% of age-related sarcopenia.

Cumulative VO₂max loss acceleration: single active vs all three feedback loops

Simulation: starting at VO₂max 40 mL/kg/min at age 45, projecting forward with 0, 1, 2, or 3 feedback loops active. Each loop adds ~0.4% additional annual decline.

Loop 3: Inactivity → Depression → more Inactivity (anhedonia cycle)

StepCausal linkEffect sizeEvidence
1Physical inactivity → depression onsetHR 1.44 for new MDD onsetMammen 2013 Am J Prev Med meta n=30,000
2Depression → reduced voluntary PA−40 min/week moderate PASchuch 2017 Psychol Med (MANOVA)
3Reduced PA → VO₂max loss−2.5 mL/kg/min per 50 min/wk lossSloth 2013 Scand J Med Sci Sports
4 (loop)VO₂max loss → worsening depressionBidirectional: low fitness predicts depression HR 1.32Blumenthal 2012 Psychosom Med
Breaking the loop: Exercise therapy produces effect equivalent to antidepressants (d=0.68) AND simultaneously restores VO₂max — treating the bottleneck node treats both conditions simultaneously.
Key structural finding: Truly exogenous (non-modifiable) root causes account for only ~40% of the total variance in VO₂max decline. The remaining ~60% flows through physical inactivity — either directly (inactivity causes deconditioning) or through the downstream mediators that inactivity causes (obesity, sarcopenia, T2DM, depression). The single highest-ROI intervention is to protect the physical activity node — this simultaneously interrupts all three feedback loops and all downstream mediator pathways.

Complete evidence table — causal pathway attribution and modifiability

FactorDAG roleTrue causal
weight
% via
inactivity
Intrinsic
component
Key evidenceModifiable?
Truly exogenous roots — set by nature, cannot be changed
Genetic endowmentRoot47%15%85% (cardiac morphology, Hb affinity)Baseline VO₂max ≈50% heritable (Bouchard 1998, sedentary state); the 47% figure is training-response heritability (Bouchard 1999, n=481, 98 families) — not baselineNo
Intrinsic SA node agingRoot (via Age)~15%30%70% (~1 bpm/yr; athletes vs sedentary same rate)Tanaka 1997; Hawkins 2001 (lifetime athletes still decline)Partial
Vascular AGE cross-linkingRoot (via Age)~8%40%60% (irreversible collagen/elastin glycation)Laurent 2006 Eur Heart J; Tanaka 2000 JAMAPartial
Biological sexRoot18%20%80% (Hb ~10-15% higher males, cardiac SV)Kaminsky 2022 FRIEND n=750,302No
Physical inactivity — the central modifiable hub
Physical inactivity (direct deconditioning)Central hub~35%100%0%Saltin 1968 bed rest; Mujika 2001 Eur J Appl PhysiolYes
Mediators downstream of inactivity — caused by inactivity, causing more inactivity
Mitochondrial dysfunctionMediator ← inactivity~12%60%40% (somatic mtDNA mutations)Lanza 2008 PNAS: trained elderly = young sedentary mitochondriaMostly yes
SarcopeniaMediator ← inactivity ↔ feeds back16%80%20% (satellite cell senescence)Goodpaster 2006; Mitchell 2012 — resistance training reverses mostMostly yes
ObesityMediator ← inactivity + diet ↔ feeds back18%85%15% (hormonal, GH decline)Fogelholm 2010 Obes Rev; DPPRG 2002Mostly yes
OsteoarthritisMediator ← obesity+age ↔ inactivity14%88% via PA12% (systemic IL-6)Bartels 2016 Cochrane; Wallis 2013; van Wessel 2010Via PA modality
Type 2 DiabetesMediator ← inactivity + obesity12%75%25% (beta-cell intrinsic)DPPRG 2002: exercise −58% T2DM risk; Warburton 2006Mostly yes
DepressionMediator ← inactivity ↔ feeds back7%92%8% (HPA catabolism)Blumenthal 2012; Mammen 2013 meta n=30,000Yes
Sleep apneaMediator ← obesity+anatomy8%75%25% (nocturnal hypoxia cardiac)Alahmari 2023; Brutinel 2016Yes (CPAP)
Conditions with large DIRECT physiological components (partially bypass activity)
CVD / Heart failureMediator + direct cap22%40%60% (cardiac output ceiling)Guazzi 2016 JACC; Lakatta 2003 CirculationPartially
COPDMediator + direct cap20%35%65% (ventilatory ceiling)Oga 2003 Eur Resp J; Casaburi 2009Partially
Anemia / Low HbDirect physiological cap10%10%90% (O₂ carrying capacity)Calbet 2006 Acta Physiol; Ekblom 1972Yes (treat)
Pareto-optimal intervention portfolio. Greedy selection: at each step, add the intervention that provides the largest marginal gain in Pearl logRR not yet captured by already-selected interventions. The Pareto frontier is steep — VO₂max alone gives 81% of the maximum, and just 6 interventions deliver 95%.
Selection driver:
Target coverage:
97% of maximum 97.7% ACM reduction
Pathway coverage
Pareto-optimal portfolio — greedy order (each row adds the most new non-overlapping benefit)
Pareto curve: cumulative effect vs number of interventions
Uncertainty propagation methodology: SE(logRR) = [ln(CI_hi) − ln(CI_lo)] / (2 × 1.96) from each study's 95% CI. Stack SE propagated as √(Σ SE_i²) assuming independence on logRR scale (conservative — correlation would narrow the CI). Active protocol (✓ rows) reflects current sidebar selections. Green-highlighted rows = currently active.
Stack uncertainty table
StacknPearl HRACM Red% 95% CI lo95% CI hi68% CI lo68% CI hi CI widthSE(logRR)
Individual intervention CI table (✓ = currently active)
InterventionEvACM Red% CI lo%CI hi% SD×100SE(logRR) HR 95% CInBar
Forest plot — individual intervention HRs with 95% CI whiskers (top 20 by effect size)
Stack uncertainty — ACM reduction 95% CI and ±1 SD by intervention portfolio size
Cause-of-death distribution
Age/sex-specific. GBD 2019 + Statistics Canada 2020-2022.
Before vs after protocol — top 12 causes
Disease-specific HRs from cause-specific meta-analyses, combined multiplicatively for active interventions.
Disease-by-disease breakdown — all causes with ≥1% mortality weight