Longevity Olympics Research Initiative · Hepatology Module v1.0

Hepatic Bayesian Causal Network

A Pearl Structural Causal Model spanning 15 hepatic disease entities, stratified by histologic stage and disease variant, with 180+ literature-anchored interventions. Backdoor-adjusted, cross-correlation-corrected, and reported with E-values, PNS, PAF, and dose-response saturation.

PEARL SCM DO-CALCULUS E-VALUES ρ̄ ADJUSTED

Disease Specification

// Treatment Group X

Intervention Bundle (X = do(·))

// Toggle interventions · Adjust dose intensity
A high (RCT/meta) · B cohort · C case-series/modelled. Lower grades excluded from the combined hazard ratio & all views.
0.30

Pearl-Adjusted Causal Estimate

// E[Y | do(X)] vs baseline
Baseline 5-yr Risk
—
stage + variant adjusted
Pearl-Adj 5-yr Risk
—
post-intervention
Combined HR (Pearl-adj)
—
naive Π HR shown below
NNT (5-year)
—
to prevent 1 event
E-value (VanderWeele)
—
unmeasured confounder bound
PAF / PNS
—
PN, PS reported
→ Risk Trajectory: Baseline vs Adjusted (5-year hazard)

Pareto Frontier — Marginal Intervention Yield

// Per-intervention log-HR contribution · target-RRR minimum-effective-set highlighted
80%
Min interventions to hit target
—
RRR achieved by set
—
Absolute risk after set
—
Marginal cost of last add
—
→ Δ Log-HR per Intervention (larger negative = stronger effect) · amber = inside minimum-effective-set

Slide the target relative-risk-reduction (RRR) bar to set the desired %. The amber-shaded interventions are the minimum subset — added in rank-order of largest log-HR contribution — whose Pearl-adjusted bundle achieves the chosen RRR against the current baseline. Bars beyond the cutoff represent diminishing marginal returns and may already be on the saturation shoulder of the dose-response curve.

Sensitivity Heatmap — ρ̄ × Bundle Size

// Combined HR robustness to cross-correlation assumption
→ How combined HR shifts as ρ̄ and n vary

Selected Intervention Ledger

// Auditable table of every selected do(·) and its causal contribution

Bayesian Causal DAG — Selected Bundle

// X (interventions) → M (mediators) → Y (endpoint); confounders adjusted via backdoor

"The do-operator do(X = x) represents physical intervention rather than passive observation. Backdoor-admissible confounders C must be conditioned on to identify P(Y | do(X)) from P(Y | X, C)." — Pearl, Causality (2nd ed., 2009), Theorem 3.3.2

Antithetical Critique

// Steel-manning the counter-argument (Commandment #7)
Where this model may be wrong
  • Cross-correlation may exceed ρ̄ = 0.30 for highly mechanistically-redundant bundles (e.g., GLP-1 + tirzepatide + bariatric surgery all act on weight/insulin axis; true ρ̄ likely ≥0.55). Slide ρ̄ up to test.
  • Baseline hazards are population averages, not individualized. PNPLA3 GG, TM6SF2, MBOAT7 variants shift MASLD/MASH baselines materially; APOE ε4 modifies cirrhosis HCC risk.
  • Confounding by indication contaminates observational HRs (e.g., statin users are healthier; aspirin users are pre-screened for bleeding). E-value provides only a lower bound on the confounder strength required to nullify.
  • Competing risks dominate at older ages: liver-related death is competing with cardiovascular and cancer mortality. PAF calculations assume independence of competing causes.
  • Some "Tier C" interventions (berberine, vitamin D, omega-3) have HRs derived from surrogate endpoints (LFTs, steatosis %, hepatic fat fraction) rather than hard outcomes. Their inclusion inflates apparent benefit; consider excluding for conservative estimates.
  • Saturation effects beyond dose 1.0 are modeled via exponential approach but real biology may exhibit U-shape (e.g., high-dose UDCA in PSC; selenium in MASLD); flat extrapolation may overshoot.
  • Transplant HRs assume successful listing & graft availability, ignoring waitlist mortality (~10-20% MELD ≥25), recipient comorbidity, and recurrence (HCV cured; HBV well-suppressed; PSC ~20% recur).
Disclaimer: This tool is a research-grade Bayesian causal simulator. It is not medical advice. Hazard ratios are extracted from published literature and combined under explicit Pearl-SCM assumptions (backdoor adjustment, monotonicity, cross-correlation removal). Clinical decisions require individualized assessment by a qualified hepatologist. All HRs cite primary literature; cross-correlation default ρ̄ = 0.30 reflects mid-range mechanistic overlap typical for non-redundant intervention bundles and is the user-established standard for this analytical lineage. Combined HRs > 0.05 floor are enforced; estimates below this likely violate independence assumptions.