The Evidence Layer for Scientific AI

 Quick Analysis Plan Copied

 Long Analysis Plan

Scope check: the premise is not supported by the supplied records

No supplied paper is a meta-analysis authored by Marta Arrieta; the evidence set contains seven primary studies — an autopsy ECM study Autopsy ECM remodeling study, two sepsis/etiology outcome cohorts Sepsis vs non-sepsis ARDS outcomesDirect vs indirect ARDS predictors, a prone-positioning oxygenation study Prone positioning oxygenation study, a prognostic tertile model Risk tertile prognostic model, an NIV failure cohort NIV failure risk factors, a 1989 diagnostic-criteria reappraisal ARDS criteria reappraisal, and an FGF21 biomarker study FGF21 prognostic biomarker study. Treating this set as a "meta-analysis" would be a framing error.

The clearest cross-study signal concerns whether etiology changes prognosis: mortality was similar between direct and indirect ARDS (28% vs 31%) despite different severity profiles (SAPS II 51 vs 62) and different mortality predictors Direct vs indirect ARDS predictors, while sepsis-related ARDS carried higher 60-day mortality than non-sepsis ARDS Sepsis-related ARDS outcomes. These are partially divergent framings of the same underlying question — a known tension, not a contradiction.

Battle-testing: key blind spots

• Small validation samples: the tertile model's external validation used only 50 patients with zero events in the lowest tertile (0% mortality), making the AUC 0.810 fragile and prone to optimism Small external validation of tertile model.
• Single-center and observational designs dominate the NIV, prone, and FGF21 evidence, limiting causal inference and generalizability Single-center NIV cohort limitationsSingle-center FGF21 limitations.
• Autopsy-only ECM data cannot inform prognostic metrics for survivors, and the 1989 spectrum hypothesis was tested in 50 patients with incomplete biomarker capture Autopsy-only limitation.
• Missing information: no publication-bias assessment, no heterogeneity statistics, no GRADE-style quality appraisal — the components a true meta-analysis would require — are absent from these records.

What would change these conclusions

An independent, multi-center cohort validating the age/Pplat/PaO2-FiO2 tertile cutoffs in non-lung-protective ventilation settings, or replication of FGF21 dynamics in a second center, would either strengthen or falsify the two most quantitative prognostic claims. A large direct-vs-indirect cohort showing divergent mortality would overturn the "similar mortality, different predictors" pattern. Until then, confidence is moderate for the tertile and severity-gradient findings and low for biomarker-based prognostication.