Prediction Models for Mortality in Patients With Moderate to Severe ARDS Treated in the ICU

Background

Mortality prediction models have been developed for patients in the ICU, but infrequently are targeted for specific conditions. Because ARDS is characterized by high morbidity and mortality, ARDS-specific models for outcome prediction could be valuable for informing patients and relatives, for clinical decision-making, for targeted interventions, and for research.

Research Question

What are the available prediction models for moderate to severe ARDS and what is their capacity to predict mortality?

Study Design and Methods

In this systematic review and meta-analysis, we searched for eligible studies in PubMed MEDLINE, Embase, PsycINFO, Web of Science, Scopus, CINAHL, Cochrane Library, and Google Scholar databases up to March 11, 2024. We included studies that developed or validated multivariable prediction models for mortality in moderate to severe ARDS, applied within 24 hours after ICU admission. Calibration, discrimination, and clinical usefulness were summarized across models. The pooled area under the receiving operating characteristic curve (AUC) was calculated with random effects models both overall and in subgroups of models and study type (development or validation). Heterogeneity was evaluated using the I² statistic.

Results

Of the 7455 screened articles, 14 were included, evaluating 20 unique models. Discrimination was reported for all models, whereas calibration was reported in 16 models. The pooled AUC was 0.782 (95% CI, 0.748-0.817) with an I² of 99.5% (P < .0001). In subgroup analysis, the pooled AUC for the Sequential Organ Failure Assessment (SOFA) score was 0.802 (95% CI, 0.719-0.885), the age, plateau, and Pao₂ to Fio₂ ratio score was 0.724 (95% CI, 0.643-0.805), the Acute Physiology and Chronic Health Evaluation (APACHE) II score was 0.667 (95% CI, 0.613-0.721), and all other scores were 0.813 (95% CI, 0.774-0.852; P = .0001 for subgroup differences). The pooled AUC was higher for derivation vs validation studies (0.816 [95% CI, 0.760-0.872] vs 0.767 [95% CI, 0.725-0.809]; P = .17 for subgroup differences).

Interpretation

Substantial variability in discrimination exists across the included models, with calibration frequently unreported. Although models developed specifically for this patient population demonstrate superior performance, general disease severity models like APACHE and SOFA are validated more extensively. Presently, no extensively validated prediction model exists showing good discrimination and calibration for moderate to severe ARDS.

Clinical Trial Registry

International Prospective Register of Systematic Reviews; No.: CRD42022342893; URL: https://www.crd.york.ac.uk/prospero/