Abstract
<title>Abstract</title> <p>Background Failure to achieve durable liberation from invasive mechanical ventilation is a clinically important ICU problem. We developed and validated an early prediction model for 7-day ventilator liberation failure using routinely available data from the first 24 hours after initiation of invasive mechanical ventilation. Methods This retrospective multicohort prediction model study used MIMIC-IV for model development, probability calibration, and internal testing; eICU for external validation; and a local ICU cohort from Weifang People's Hospital for additional clinical validation. Among adults who were alive and still receiving invasive mechanical ventilation at the 24-hour landmark, the primary outcome was 7-day ventilator liberation failure, defined as failure to achieve confirmed durable liberation by day 7 after initiation of invasive mechanical ventilation, invasive reventilation within 48 hours after liberation, or death before confirmed durable liberation. Successful liberation required discontinuation of invasive mechanical ventilation by day 7 followed by at least 48 hours without invasive reventilation. Model performance was evaluated using AUROC, AUPRC, Brier score, calibration, decision curve analysis, risk stratification, model comparison, module ablation, and subgroup analyses. Results The analysis included 2,885 patients in the MIMIC-IV test cohort, 12,507 patients in the eICU external validation cohort, and 347 patients in the local validation cohort. Event rates were 36.0%, 30.3%, and 32.6%, respectively. The calibrated random forest model showed similar discrimination across cohorts, with AUROC values of 0.745, 0.744, and 0.749 and AUPRC values of 0.629, 0.590, and 0.592. Brier scores were 0.192, 0.177, and 0.184. Prespecified risk strata separated patients into clinically distinct event-rate groups, and decision curve analysis suggested potential net benefit across plausible threshold probabilities. Conclusions An early calibrated prediction model showed consistent moderate discrimination and clinically interpretable risk stratification for a composite ventilator liberation failure endpoint, although external calibration patterns indicated that local recalibration monitoring would be required before clinical use. The model is best positioned as a risk-stratification aid requiring prospective workflow evaluation, not as a standalone rule for extubation or ventilator discontinuation.</p>