Cardiology Mortality Atlas LIVING

Cross-class umbrella meta-analysis of all-cause mortality across 24 cardiovascular drug classes

Mahmood Ahmad · RapidMeta Living Evidence Portfolio · Generated 2026-04-10 20:41

Forest plot — drug class effects on all-cause mortality

Per-class details (with freshness and guideline concordance)

Population stratification

Class-level pooled estimates grouped by population category. Pools use DerSimonian-Laird across apps within each stratum.

Bucher indirect comparisons (top 25 most divergent pairs)

For every pair of drug classes, the indirect effect via shared placebo comparator: log(HR_AB) = log(HR_A) − log(HR_B), with variance summed. Statistically significant pairs (CI excludes 1.0) are highlighted in green. Interpret cautiously: transitivity assumes exchangeable populations across trials.

Guideline concordance flags

Classes where the current pooled ACM estimate does not cleanly match the expected strength of the guideline recommendation. Overstrong = guideline is stronger than the pooled evidence supports; understrong = pooled evidence is stronger than the guideline currently acknowledges.

• Bempedoic acid (Class IIa, ASCVD (statin-intolerant)): pooled HR 1.00 (0.87-1.15) — overstrong.
• PCSK9 (Class I, ASCVD on statin): pooled HR 0.94 (0.77-1.15) — overstrong.
• IV iron (Class IIa, HFrEF + iron deficiency): pooled HR 0.94 (0.84-1.04) — overstrong.
• Sotagliflozin (Class IIa, HF + T2DM/CKD): pooled HR 0.82 (0.59-1.14) — overstrong.
• P2Y12 mono (Class IIa, Post-PCI (selected)): pooled HR 0.90 (0.80-1.02) — overstrong.
• Ticagrelor mono (Class IIa, Post-PCI (1-3 mo DAPT)): pooled HR 0.95 (0.85-1.05) — overstrong.
• Icosapent ethyl (Class IIa, Hypertriglyceridemia + ASCVD): pooled HR 0.99 (0.77-1.28) — overstrong.

Advanced analytics (12 supplementary analyses)

Twelve advanced analyses characterize the robustness, interpretability, and clinical relevance of the portfolio-wide mortality pool. Each is computed from the same trial-level data as the main forest plot and updates automatically.

[1] Prediction intervals

The 95% prediction interval gives the expected range of effects in a future trial of the same drug class, accounting for between-study heterogeneity (Higgins 2009 method). A wide PI crossing 1.0 indicates that the next trial could plausibly be null despite a significant current pool.

[2] Leave-one-out fragility

For each class with k>=2 trials, the maximum percent change in pooled HR when any single trial is removed. Classes with max delta >20% are flagged as fragile — the pool is dominated by one study and could shift meaningfully with new data.

[3] Cumulative meta-analysis (stabilization)

For each class, the running pooled HR as trials are added chronologically. Classes where the first trial's estimate and the current pool differ by less than 10% are "stable" — adding more trials is unlikely to change the clinical message.

[4] Low risk-of-bias subanalysis

Pool restricted to trials with overall low risk of bias on all domains. When the low-RoB pool differs substantially from the full pool, some effect may be driven by lower-quality evidence.

[5] Heterogeneity decomposition (Q partitioned)

Total portfolio heterogeneity Q partitioned into within-population-stratum and between-stratum components. A large "between" fraction means effect sizes differ mainly by patient population, not by methodology or chance.

[6] Temporal effect trends

Portfolio-wide weighted meta-regression of log(HR) on trial year across all drug classes. A positive slope indicates effects are attenuating — newer trials show smaller benefits, consistent with regression to the mean or the "Proteus effect".

[7] Number needed to treat (5-year approximation)

Absolute effect translation: NNT = 1 / (baseline_risk × (1 − HR)) using each class's average control-arm mortality rate. Provides a clinically actionable metric — lower NNT means greater absolute benefit per patient treated.

[8] P-score rankings (frequentist SUCRA analog)

P-score (Rücker & Schwarzer 2015) = mean probability that a class is better than a randomly selected comparator class, accounting for uncertainty. Ranges 0-1; higher is better. Unlike SUCRA, does not require Bayesian simulation.

[9] ACM vs primary outcome discordance

For each trial that reports both a primary composite outcome and an ACM secondary, the ratio of HRs quantifies how much the composite inflates apparent benefit beyond mortality alone.

[10] Baseline risk meta-regression

Portfolio-wide weighted regression of log(HR) on logit(control-arm mortality rate). A negative slope means higher-risk populations derive greater relative benefit — consistent with the common clinical observation that sicker patients benefit more.

[11] Cross-population consistency — SGLT2 exemplar

SGLT2 inhibitors are indicated across 4 distinct patient populations in this portfolio. If the mortality effect is mechanistically consistent, the pooled estimates across populations should agree. Divergence implies population-modified effects.

[12] Trial sequential analysis (required information size)

For each class, the required information size (RIS) is the sample size needed for a definitive test with α=0.05 and power=0.80, adjusted for heterogeneity via the diversity factor (Wetterslev 2008). Information fraction >100% means accumulated evidence is already adequate to support the observed effect.

Methods