Target-driven Pareto search — set the goal, solve for the minimum set
Cumulative absolute risk reduction by intervention rank
Interventions — select, titrate, observe
Causal diagnostics
Per-intervention contribution
Sensitivity analysis — what-if removal
Pareto frontier — interventions vs risk reduction
Dose-response saturation
Causal DAG — pathway structure
Per-intervention causal metrics
Methodology
| Component | Specification |
|---|---|
| Causal framework | Pearl Structural Causal Model with explicit DAG over pathway-grouped interventions; backdoor criterion applied to remove confounding by shared upstream biology. |
| Effect size pooling | Log-HR aggregation: ln(HRcombined) = Σ wi · ln(HRi) · ai · D(di) · (1 − ρ̄ · Ci), where wi is evidence-shrinkage weight, ai adherence, D(di) the Hill-type dose-response, Ci the pathway-co-occurrence count. |
| Cross-correlation removal | Interventions sharing a pathway node receive a ρ̄-weighted discount on incremental log-HR (default ρ̄ = 0.30 per VanderWeele & Ding 2017 / Greenland 2008). |
| Saturation (κ) | Hill dose-response, D(d) = (d/d*)κ / (1 + (d/d*)κ) with κ = 0.5–1.0 reflecting concavity; binary interventions assigned D = 1 when on. |
| SMD → HR conversion | Where only Cohen's d available, HR ≈ exp(d · π / √3) per Chinn (2000); validated for d ∈ [0.2, 0.8]. |
| Evidence shrinkage | Grade A → 1.00; B → 0.85; C → 0.65; D → 0.40 (multiplier on ln(HR)). Adjustable via slider. |
| E-value | E = HR + √(HR · (HR − 1)) per VanderWeele & Ding (Ann Intern Med 2017); HR' = 1/HR if HR < 1. |
| PNS / PN / PS | Pearl's probabilities of necessity-and-sufficiency derived from RR & P(exposure) per Tian & Pearl (2000); bounds reported. |
| PAF | Population attributable fraction = Pe(HR − 1) / (1 + Pe(HR − 1)); capped at Σ ≤ 100% via Levin overlap correction. |
| Sensitivity | Leave-one-out ΔHR; tornado for ρ̄ ∈ [0, 0.6], κ ∈ [0.5, 1.0]; Monte Carlo 10,000 draws (offline). |
| Antithesis | Every selection challenged: confounding, reverse causation, indication bias, publication bias, model mis-specification (see §below). |
Antithesis & limitations
The case against over-interpreting the combined HR
1 · Heterogeneity of trial populations. The pooled HRs were estimated in populations that rarely overlap with any single patient's profile (age, performance status, comorbidity, molecular subtype). Effect sizes in real-world subgroups frequently shrink by 20–40% relative to RCT estimates.
2 · Cross-correlation underestimation. The default ρ̄ = 0.30 may be too low when interventions share both mechanism and indication-bias (e.g., aerobic exercise + Mediterranean diet + sleep optimization all correlate with health-conscious phenotype). Try ρ̄ = 0.5 and observe the deflation.
3 · Selection & immortal-time bias. Observational HRs (Grade C/D) for lifestyle interventions in Alzheimer's, PD, and CTE are inflated by survivor & immortal-time biases. Mendelian randomization estimates are typically 30–60% smaller.
4 · Indication-by-stage interaction. A combined HR of 0.20 for newly-diagnosed MGMT-methylated GBM does not imply the same for unmethylated recurrent disease, even when the same interventions are toggled. Always re-select stage.
5 · Saturation may be steeper. Exercise and dietary effects appear to saturate well below the inflection point assumed by κ = 0.80; the model may overstate gains from very high doses.
6 · Publication bias. Especially in supplements (vitamin D, omega-3), null trials are under-published. E-values < 1.5 should be treated as no robust signal.
7 · Causal DAG is provisional. The pathway graph encoded here is a reasonable but contestable abstraction. Substantive disagreement among neuro-oncologists, MS-ologists, and dementia researchers about which pathways are upstream of which will shift the backdoor adjustments.
Honest synthesis: the engine is for hypothesis-generation, intervention-ordering, and shared decision-making — not for replacing prospective trials or guideline-concordant care.