A Biologically Interpretable Multimodal Model for Prognostic Risk Stratification and Glucocorticoid-Associated Outcome Heterogeneity in Acute Respiratory Distress Syndrome (ARDS)

Abstract The clinical and biological heterogeneity of acute respiratory distress syndrome (ARDS) complicates early prognostic assessment and may contribute to variability in treatment-associated outcomes. In this multicentre study of 2,868 patients from 10 centres, we developed and externally validated a multimodal risk-stratification framework integrating clinical variables, chest CT imaging, and serum proteomic data. Serum proteomic profiles were available in 879 patients and exploratory bronchoalveolar lavage fluid (BALF) single-cell RNA sequencing was performed in 26 patients to characterize pulmonary microenvironment of model-defined risk states. The clinical-imaging model achieved AUROCs of 0.831, 0.805, and 0.781 for 30-day mortality prediction in the training, internal test, and external validation cohorts, respectively, while proteomic integration improved discrimination to 0.892, 0.871, and 0.859 in the proteome-matched subset. Higher predicted risk was associated with impaired early liberation from mechanical ventilation and increased non-invasive ventilation failure. Serum proteomic analyses revealed distinct biological profiles of high-risk group characterized by tissue-injury and inflammatory signatures, such as KRT19, CHIT1, IGFBP2, and IL-6, with enrichment of inflammatory pathways (HMGB1 and TREM1 signalling), metabolic and nuclear receptor pathways (FXR/RXR and PPAR signalling), extracellular matrix remodelling (FAK signalling and collagen degradation), endothelial dysfunction, and coagulation pathways. BALF single-cell RNA further identified neutrophil-enriched pulmonary immune responses in high-risk patients. Glucocorticoid-associated 30-day mortality varied across model-defined risk strata ( P for interaction < 0.001 and 0.033 for the clinical-imaging and multimodal models, respectively), with lower mortality associated with glucocorticoid initiation among high-risk patients. These findings support multimodal risk stratification as a framework for identifying clinically and biologically distinct ARDS risk states and generating hypotheses regarding treatment-associated heterogeneity.