Ventilator-Derived Mechanical Stress Trajectories Predict Prolonged Mechanical Ventilation in ARDS: A Retrospective Cohort Study Using MIMIC-IV

This registration documents the analytical plan for a retrospective cohort study evaluating ventilator-derived mechanical stress trajectory features as predictors of prolonged mechanical ventilation (14 days or longer) in ARDS patients using the MIMIC-IV v3.1 critical care database. TRANSPARENCY DISCLOSURE: This is a retrospective registration. All analyses were completed prior to registration. This document is filed for transparency and to provide a permanent, timestamped record of the analytical decisions, not to represent a prospective plan. BACKGROUND: Fibroproliferation, the maladaptive repair response to lung injury exacerbated by mechanical ventilation, begins within 24-48 hours of ARDS onset but is not clinically detectable until days later. We hypothesized that trajectory features derived from ventilator parameters over the first 72 hours of mechanical ventilation (capturing cumulative exposure, direction of change, and variability in driving pressure, mechanical power, dynamic compliance, and respiratory rate) would predict prolonged ventilatory dependence independently of baseline illness severity. COHORT: 26,916 adult ICU stays from MIMIC-IV v3.1 meeting modified Berlin ARDS criteria (PaO2/FiO2 ratio 300 or less, mechanical ventilation 24 hours or longer, ventilation onset within 48 hours of ICU admission). Primary outcome: prolonged mechanical ventilation (14 days or longer), n=2,809 events (10.4%). METHODS: 182 trajectory features extracted from 17 ventilator and physiologic parameters using 6-hour time bins over a 72-hour observation window. Primary model: XGBoost with 5-fold stratified cross-validation. Validation: temporal split (train pre-2016, test 2016-2022). Confounding adjustment: residualized trajectories (regressing out baseline severity) and IPTW sensitivity analysis. Secondary analyses: ARDS phenotype stratification (hyperinflammatory vs hypoinflammatory), outcome threshold sensitivity (7/10/14/21 days), trajectory window sensitivity (48/72/96 hours), SHAP explainability. RESULTS SUMMARY: Primary model AUROC 0.837 (95% CI 0.830-0.842) vs severity-only baseline 0.601 (0.591-0.612). Temporal validation AUROC 0.831 (0.819-0.841). Residualized trajectory AUROC 0.825 (0.818-0.832). Signal robust across all sensitivity analyses. DATA: MIMIC-IV v3.1 via PhysioNet. Code: github.com/cadralis/research.