Prediction models show moderate discrimination for ARDS occurrence and mortality in sepsis patients

BackgroundSepsis is a life-threatening syndrome driven by dysregulated inflammation and immunity, often leading to multiple organ dysfunction. The lung is commonly affected early, and sepsis-related lung injury, including acute respiratory distress syndrome (ARDS), is associated with poor survival. Although models have been proposed to predict lung injury and short-term mortality once injury occurs, their performance, methods, and certainty of evidence remain insufficiently assessed.MethodsWe searched PubMed, Embase, and the Cochrane Library for studies published up to December 11, 2025. Risk of bias and applicability were assessed with PROBAST, and certainty of evidence was appraised using an AUC-based adaptation of GRADE. The protocol was registered in PROSPERO. The individual prediction model was the unit of analysis. We extracted AUC, sensitivity, and specificity. Pooled estimates were calculated separately for ARDS occurrence and short-term mortality in sepsis-associated lung injury, and separately for training, validation, and test phases, with no pooling across phases.ResultsNine studies were included, eight of them from China. Together they reported 68 model phase units: 24 training, 21 validation, and 23 test AUCs. PROBAST classified 4 studies as having high overall risk of bias and 6 as having unclear risk; only three studies had low concern for applicability. Certainty of evidence was low for all outcome families and modeling phases. For ARDS occurrence, the pooled test-phase AUC was 0.749 (95% CI, 0.648–0.849; I2 = 98.9%). For short-term mortality, pooled AUCs were 0.800 in training (95% CI, 0.761–0.838; I2 = 97.9%), 0.778 in validation (95% CI, 0.751–0.804; I2 = 63.5%), and 0.815 in testing (95% CI, 0.780–0.850; I2 = 75.7%). Leave-one-out analyses indicated that no single study substantially changed the pooled estimates. Within each outcome family, machine learning models did not consistently outperform logistic regression. Heterogeneity remained high, particularly for ARDS occurrence and mortality training models.ConclusionCurrent models showed moderate discrimination, but their clinical use is limited by bias, weak methods, low certainty, and heterogeneity. ARDS occurrence and mortality should be developed, validated, and reported separately. Future work needs transparent designs and external validation before implementation.Systematic review registrationhttps://www.crd.york.ac.uk/PROSPERO/view/CRD420251275870, identifier CRD42025127587.