Early Versus Deferred De-escalation From Controlled Ventilation After a 48-Hour Landmark of Physiological Stability

Overview

This retrospective target trial emulation evaluates the optimal timing of de-escalation from controlled to assisted mechanical ventilation in critically ill adults. Using the clone-censor-weight (CCW) estimator with stabilised inverse-probability-of-censoring weights, we compare two strategies over a 24-hour grace window following a 48-hour landmark: early de-escalation (step-down in ventilatory support mode or successful extubation) versus deferred de-escalation (support maintained or increased). The primary outcome is 28-day in-hospital mortality. Secondary outcomes are 28-day restricted mean survival time (RMST) and ventilator-free days (VFD28). The study uses MIMIC-IV v2.2 (derivation cohort, n=1,902 analysable) with external validation in eICU-CRD v2.0 (validation cohort, n=10,957 analysable). All analyses employ stabilised IPCW with bootstrap confidence intervals, including six sensitivity and robustness analyses (extended covariate adjustment, alternative weight truncation, doubly robust AIPW estimator) and an exploratory FiO₂×PEEP threshold grid.

Background: Mechanical ventilation is life-saving, yet optimal timing of transition from controlled to assisted modes remains uncertain. Prolonged controlled ventilation exposes patients to sedation, diaphragm dysfunction, and lung injury, while premature de-escalation risks asynchrony and reintubation. Design: Landmark target trial emulation with 48-hour eligibility ascertainment. Eligibility: Adults ≥18 years, invasive mechanical ventilation \>24 hours, alive and ventilated at 48 hours with classifiable Day-2 mode (controlled or assisted), Day-2 FiO₂ ≤50%, Day-2 PEEP ≤10 cmH₂O. Strategies: Early-support level decreased by Day 3 (mode step-down from controlled to assisted, or successful extubation while alive); Deferred-support maintained or increased. Analysis: Single-interval CCW estimator with stabilised IPCW (truncated 1st-99th percentile), propensity model with 21 baseline covariates, bootstrap 95% CIs (2,000 resamples primary; 800 sensitivity). Sensitivity analyses: no truncation, ICU mortality, unweighted per-protocol, extended 40-covariate adjustment, 5th-95th truncation, AIPW doubly robust estimator. E-values computed for unmeasured confounding. External validation: Harmonised protocol in eICU-CRD v2.0 (208 hospitals, n=10,957), outcome: in-hospital mortality.