A Judea Pearl Structural Causal Model for the highest-mortality condition in the atlas — and an anti-hype one. The dominant causal lever is time to effective treatment, not a proprietary drug: early antibiotics, source control, and bundle completion move mortality, while a generation of celebrated adjuncts (protocolised EGDT, high-dose vitamin C) failed in randomised trials. All effect sizes are real published values; the engine credits only what trials support and shows the failed levers at or below zero.
Causal question. Across interventions with published mortality effects, what is the identifiable causal contribution of each to a reduction in sepsis mortality, after backdoor confounding control and after removing shared-pathway double-counting (the common "good care" mediator) via cross-correlation shrinkage?
| Element | Specification |
|---|---|
| Population | Adults with sepsis / septic shock (Sepsis-3); ~49 million cases and ~11 million deaths per year globally |
| Primary endpoint | In-hospital / 28–90-day all-cause mortality |
| Dominant lever | Time to effective antibiotics & bundle completion (a time-dependent hazard, not a fixed dose) |
| Anchor | Expected mortality under prompt, complete, guideline care for this severity (user-set; septic shock ≈ 30–45%) |
| Severity strata | Septic shock (vasopressor-dependent) vs sepsis without shock — moderates the timing slope and steroid benefit |
Sepsis is unusual: the strongest do(·) is not a molecule but a clock. Recognition speed drives time-to-antibiotics and bundle completion, which drive mortality; severity (shock) steepens every slope. Confounders (age, comorbidity, source) are closed by trial randomisation where it exists and by risk-adjustment in the observational timing studies.
| Operation | Method | Implementation here |
|---|---|---|
| Confounding control | Backdoor criterion (Pearl) | RCTs for steroids/fluids/EGDT/vitamin C; timing studies risk-adjusted for severity |
| Time dependence | Per-hour mortality hazard | Delay beyond the 1-hour target ×1.04/hr (sepsis) or ×1.07/hr (septic shock) |
| Effect aggregation | Multiplicative on risk, shrunk for overlap | Timing primary; bundle/steroids/fluids deflated by (1−ρ̄) — they share the "good care" pathway |
| Failed-lever honesty | Credit only RCT-supported effects | Protocolised EGDT ×1.00; high-dose vitamin C ×1.05 (LOVIT harm signal) |
| Robustness | E-value (VanderWeele & Ding) | Computed for the dominant lever (early antibiotics, OR 0.67) |
| Minimum-effective set | Pareto frontier (threshold slider) | Levers ranked by mortality reduction in Section IV |
| Intervention | Mortality effect | Source / grade |
|---|---|---|
| Early antibiotics (≤1 h vs >1 h) | OR 0.67 (0.59–0.75) | door-to-antibiotic meta-analysis 2024 B |
| Each hour to 3-h bundle completion | OR 1.04/h (1.02–1.05) | Seymour 2017, NEJM (n=49,331) B |
| Each hour to antibiotics, septic shock | ~+7%/h (OR 1.07) | JID time-to-antibiotic analysis B |
| Complete SSC Hour-1 / 3-h bundle | 28-d HR 0.75 (0.59–0.95) | SSC bundle-compliance cohort B |
| Hydrocortisone + fludrocortisone (severe shock) | 90-d 43.0% vs 49.1% (RR ≈ 0.88) | APROCCHSS, NEJM 2018 A |
| Hydrocortisone alone | 90-d 27.9% vs 28.8% (NULL) | ADRENAL, NEJM 2018 — faster shock resolution only A |
| Balanced crystalloids vs saline | ≈ RR 0.90 (sepsis subgroup) | SMART 2018; BaSICS/PLUS neutral B |
| Protocolised early goal-directed therapy | RR ≈ 1.0 (NO benefit) | ProCESS / ARISE / ProMISe; PRISM meta A |
| High-dose vitamin C | RR ≥ 1.0 (harm signal) | LOVIT 2022 (NEJM); VITAMINS null A |
The pattern: the levers that move mortality are timing and completeness of basic care, plus steroids in the sickest. The branded adjuncts that promised a step-change did not deliver one — and the engine refuses to pretend otherwise.
Levers ranked by relative mortality reduction (1 − RR) at the current severity. Drag the threshold to set a minimum benefit: levers at or above the line form the minimum-effective set; the dimmed tail includes the failed adjuncts, which sit at or below zero benefit — a visual statement of where the evidence actually is.
Most diseases offer a dose; sepsis offers a deadline. The same antibiotic given at hour 1 versus hour 6 is a different intervention in causal terms — the do(·) is timing. Systems that compress recognition-to-treatment time act on the dominant node directly.
EGDT, vitamin C, and high-dose vasopressin adjuncts each promised to outflank the basics. Randomised against good usual care, they didn't — because the usual-care comparator had already captured most of the available benefit through timing and source control.
ADRENAL (hydrocortisone alone) was null; APROCCHSS (hydrocortisone + fludrocortisone, sicker patients) cut mortality. The benefit concentrates in vasopressor-dependent shock — so the engine applies it only in that stratum.
With ~11 million sepsis deaths a year, even a small shift in median time-to-antibiotics — a systems intervention, not a drug — would avert more deaths than any single pharmacologic adjunct yet tested.
| Challenge | Why it threatens the conclusion | Mitigation here |
|---|---|---|
| Timing evidence is observational | Seymour/Kumar are risk-adjusted cohorts, not RCTs; sicker patients may be treated faster or slower | Graded B; the per-hour slope is presented as an association, with severity-specific slopes |
| Risk-multiplier approximation | Mortality is not rare (~40%), so multiplying ORs/HRs/RRs as risk ratios is an approximation | Framed as a ceiling; ρ̄ shrinkage and a cap prevent runaway products; flagged here |
| Bundle ↔ timing overlap | "Complete bundle" already includes early antibiotics — counting both double-counts | ρ̄ deflates the bundle and adjunct terms; timing carries full weight as the primary lever |
| Steroid heterogeneity | ADRENAL and APROCCHSS disagree; drug, dose, and severity differ | Benefit applied only in septic shock and labelled as the APROCCHSS (hydrocortisone+fludrocortisone) result |
| Balanced-fluid uncertainty | SMART favoured balanced; BaSICS and PLUS were neutral | Modelled as a small benefit (RR ≈ 0.90), graded B and flagged as uncertain |
timeFactor = perHr ^ max(t − 1, 0) ; perHr = shock ? 1.07 : 1.04
riskMult = timeFactor × bundlePenalty^(1−ρ̄) × steroid^(1−ρ̄) × balanced^(1−ρ̄) × egdt × vitC
bundlePenalty = complete ? 1.00 : 1.33 (=1/0.75) ; steroid = (shock & on) ? 0.88 : 1.00
balanced = on ? 0.90 : 1.00 ; egdt = on ? 1.00 : 1.00 ; vitC = on ? 1.05 : 1.00
modelled = min(anchor × riskMult, 0.95)
best-achievable = anchor × (shock ? 0.88 : 1)^(1−ρ̄) × 0.90^(1−ρ̄) (prompt, complete, no harmful adjunct)
avoidable excess = max(modelled − best-achievable, 0) ; E-value(HR) = RR* + √(RR*·(RR*−1)), RR* = 1/HR
Worked example: septic shock, anchor 40%, antibiotics at 5 h (delay 4 h), bundle incomplete, steroids on, ρ̄ 0.30 → timeFactor 1.07^4 = 1.31; bundle 1.33^0.7 = 1.22; steroid 0.88^0.7 = 0.91 → riskMult 1.46 → modelled 58.3%. Best-achievable = 40%×0.91×0.93 = 33.9%. Avoidable excess ≈ 24 pp — almost all of it timing and bundle.
1. Seymour CW, Gesten F, Prescott HC, et al. Time to Treatment and Mortality during Mandated Emergency Care for Sepsis. N Engl J Med 2017;376:2235–2244. Each hour to 3-h bundle completion OR 1.04 (1.02–1.05).
2. Kumar A, et al. Duration of hypotension before initiation of effective antimicrobial therapy. Crit Care Med 2006;34:1589–1596. Each hour of delay over the first 6 h ≈ 7.6% decrease in survival.
3. Door-to-antibiotic time and mortality in sepsis: systematic review and meta-analysis. Eur J Intern Med 2024. Early (≤1 h) vs later antibiotics pooled OR 0.67 (0.59–0.75).
4. Venkatesh B, et al. (ADRENAL). Adjunctive Glucocorticoid Therapy in Patients with Septic Shock. N Engl J Med 2018;378:797–808. 90-day mortality 27.9% vs 28.8% (NS); faster shock resolution.
5. Annane D, et al. (APROCCHSS). Hydrocortisone plus Fludrocortisone for Adults with Septic Shock. N Engl J Med 2018;378:809–818. 90-day mortality 43.0% vs 49.1% (P=0.03).
6. Semler MW, et al. (SMART). Balanced Crystalloids versus Saline in Critically Ill Adults. N Engl J Med 2018;378:829–839. Lower MAKE30; sepsis-subgroup mortality benefit. BaSICS (2021) and PLUS (2022) neutral overall.
7. ProCESS (NEJM 2014), ARISE (NEJM 2014), ProMISe (NEJM 2015), PRISM meta-analysis (NEJM 2017): protocolised early goal-directed therapy showed no mortality benefit vs usual care.
8. Lamontagne F, et al. (LOVIT). Intravenous Vitamin C in Adults with Sepsis in the ICU. N Engl J Med 2022;386:2387–2398. Higher risk of death or persistent organ dysfunction. VITAMINS (JAMA 2020) null.
Longevity Research Initiative — Bayesian Causal Atlas. SCM methodology after Judea Pearl; E-values after VanderWeele & Ding. All effect sizes are published values cited above; none were fabricated. Outputs are upper-bound ceilings for research, not clinical guidance. Report generated .