Seasonal dynamics and prognostic determinants of respiratory failure in critically ill adults: a multicenter retrospective observational study in Northern China

Abstract Background Respiratory failure is a common and severe syndrome in critical care, associated with poor prognosis, and its incidence and healthcare burden among critically ill patients are increasing. This study aimed to investigate seasonal patterns of respiratory failure admissions and their associations with meteorological factors and to evaluate the contributions of clinical and meteorological features to model-predicted in-hospital mortality. Methods This multicenter retrospective observational study included adults with respiratory failure admitted to emergency intensive care units (EICUs) between 2022 and 2024. Clinical and meteorological datasets were linked based on admission time and regional location. Seasonal trends were analyzed using seasonal-trend decomposition with Loess (STL), while quasi-poisson generalized additive models (QP-GAMs) and distributed lag non-linear models (DLNMs) were used to examine meteorological effects. A random forest model combined with SHapley Additive exPlanations (SHAP) was used to quantify global and individual feature contributions to model-predicted in-hospital mortality. Results Among 1,837 critically ill patients with respiratory failure, 57.8% were male, with a median age of 66 years; 34.2% had hypoxemic type and 64.1% hypercapnic type, with overall in-hospital mortality of 54.3%. STL analysis revealed marked seasonal variation, peaking in winter and reaching a minimum in summer. QP-GAMs and DLNMs suggested nonlinear and lagged associations of temperature and relative humidity with monthly respiratory failure admission counts. In the random forest model, age and mechanical ventilation had the largest mean absolute SHAP values, followed by other clinical features. Conclusion This study revealed nonlinear and lagged associations between meteorological factors and monthly respiratory failure admissions, with the highest admission counts observed in winter, followed by spring. In the fitted random forest model, clinical and physiological features contributed more to mortality predictions than meteorological variables.