Research on Respiratory Support Decision-Making for Mechanically Ventilated ICU Patients Based on a Physiological World Model

Mechanical ventilation is the cornerstone of life support for critically ill patients with respiratory failure in the ICU. However, adjustments to ventilator settings and decisions regarding weaning remain highly dependent on clinical experience, resulting in substantial inter-clinician variability. A fundamental limitation of current practice is that commonly used single-point metrics-such as the oxygenation index-cannot reliably reflect a patient's true recovery state across different levels of respiratory support. The same oxygenation level may indicate genuine recovery in a patient on low support, or persistent respiratory failure masked by high ventilatory support. This conflation of "true recovery" with "support-dependent stability" is a major contributor to weaning failure and delayed extubation.

To address this challenge, this study will leverage electronic health record (EHR) data from ICU patients who received invasive mechanical ventilation at Ruijin Hospital, Shanghai Jiao Tong University School of Medicine, between 2016 and 2026 (RuiICU-P1 cohort, n = 4,232), to develop a deep learning-based physiological world model of respiratory dynamics (LAVENT). The model is designed to: (1) learn the co-evolutionary dynamics between a patient's physiological state and respiratory support; (2) generate 96-hour multivariate clinical trajectories given the current patient state and candidate ventilator settings (FiO₂, PEEP); and (3) compare the differential clinical outcomes of alternative ventilator adjustment strategies within the same patient state. External validation will be performed using the public MIMIC-IV database (n = 29,899) to assess model transferability across healthcare systems.

This is a retrospective, observational study using only existing EHR data, without any prospective patient intervention, additional examinations, or biospecimen collection.