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

Mechanical ventilation is essential for ICU patients with respiratory failure, yet ventilator adjustment and weaning decisions remain experience-dependent and highly variable. Current single-point metrics such as the oxygenation index cannot distinguish true recovery from support-dependent stability-the same oxygenation level may reflect either. This study uses EHR data from 4,232 invasively ventilated ICU patients at Ruijin Hospital (2016-2026) to develop LAVENT, a deep learning-based physiological world model of respiratory dynamics. LAVENT learns the co-evolution of patient physiology and respiratory support, generates 96-hour multivariate trajectories under candidate FiO₂/PEEP settings, and compares alternative ventilation strategies within the same patient state. External validation uses the MIMIC-IV database (n = 29,899). This retrospective observational study uses only existing EHR data, with no prospective intervention or biospecimen collection.

Inclusion Criteria: * Inclusion criteria: * age ≥ 18 years * received invasive mechanical ventilation during ICU stay * had complete 48hour baseline data and 96hour followup data. Exclusion Criteria: * age \< 18 years * duration of mechanical ventilation \< 24 hours * missing rate \> 50% for key variables (FiO₂, PEEP, SpO₂, respiratory rate) * received ECMO therapy.