Longitudinal Subphenotyping of Acute Respiratory Distress Syndrome: Implications in Prognostic Prediction and Clinical Intervention

BACKGROUND AND OBJECTIVE: To identify the longitudinal subphenotypes (LSPs) in patients with acute respiratory distress syndrome (ARDS) and their transitions, and to evaluate their potential for prognostic prediction and guiding interventions. METHODS: This retrospective multicohort study derived its cohort from Zhongshan Hospital, Fudan University, China. Feature selection was performed using univariate analysis, recursive feature elimination, and correlation analysis, followed by longitudinal latent profile analysis. Cox regression model was used to compare differences in mortality and responses to interventions. A predictive model was developed through selection from nine candidate machine learning algorithms followed by grid search optimization and subsequently applied to an independent validation cohort. RESULTS: Nine hundred ninety seven patients were included in the derivation dataset. Utilizing the 20 most prognostically relevant variables, three distinct LSPs were identified. Based on Day 1, the LSPs accounted for 36.41%, 36.71%, and 26.88%, respectively. LSP 1 (HR 5.119; 95% CI: 3.657-7.165) and LSP 2 (HR 2.922; 95% CI: 2.063-4.139) were associated with higher mortality. Both high-dose corticosteroids and invasive mechanical ventilation failed to elicit a treatment response in LSP 2 and LSP 3. Conversely, prone positioning proved to be an effective intervention for LSP 2. An early shift from LSP 1 to LSP 2 was associated with increased mortality (HR 1.679; 95% CI: 1.051-2.683). Furthermore, the optimized random forest model demonstrated superior performance in differentiating the three LSPs and could identify consistent subphenotypes in validation cohort. CONCLUSIONS: Our findings underscore the importance of incorporating temporal subphenotype evolution into prognostic stratification and personalized treatment.