Introduction: Prolonged respiratory failure (PRF) following initiation of invasive mechanical ventilation is associated with increased morbidity, mortality, and resource utilization. While predictive scoring systems such as iTRACH have been developed, the duration and determinants of PRF remain inadequately characterized across diverse hospital settings. Methods: We performed a retrospective cohort study using data from six CLIF Consortium hospital systems and the MIMIC-IV database, including adult MICU patients who received invasive mechanical ventilation ≥24 hours between January 2018 and December 2024. Patients with tracheostomies on admission were excluded. PRF was defined as ventilation >14 days, and early deaths as those occurring 110), renal dysfunction (BUN >25, creatinine >2.0), and acidemia (pH < 7.25, HCO₃ < 20), was used to predict PRF, with a cutoff of ≥4. Demographic, clinical, and hospital-level data were extracted from each site’s EHR, and aggregated using a federated approach. Large language models assisted with coding and drafting; all outputs were author-reviewed. Results: The cohort consisted of 19,444 admissions (median age range across the 7 systems: 61–65 years, 43% Female, 28% Black, 7% Hispanic), with the median intubation duration ranging from 3.0-4.4 days and ICU stays ranging from 5.7-10 days. 12.2% of patients experienced PRF. 25% of patients had early death. Patients with COVID-19 had a longer median intubation duration (4-14 days), consistent with what has previously been described. The overall median ITRACH score was 1 (IQR: 0-2), and ranged from 1-2 across sites for those with PRF. Using an iTRACH cutoff of ≥4, specificity was high (0.96-0.98), but sensitivity was low (0.02–0.07). Conclusions: PRF is common among mechanically ventilated patients across seven hospital systems, with notable variation in duration by site. Existing tools like iTRACH showed poor predictive performance. Early identification of at-risk patients could improve prognostication and guide targeted interventions. More advanced approaches, such as machine learning, may offer greater promise in predicting PRF.
Liao et al. (Sun,) studied this question.