Introduction: Tracheostomy decision and timing in patients with sepsis-associated acute respiratory failure (ARF) requiring invasive mechanical ventilation (IMV) remains challenging. Mechanical power (MP), a marker of ventilatory load, has been associated with adverse outcomes, including mortality and weaning failure. We hypothesized that MP, calculated at the time of intubation and at 72 hours, along with demographics and cumulative ventilator days, could improve prediction of tracheostomy need in this population. Methods: We conducted a retrospective cohort study using adult ICU data from the MIMIC-IV and K-MIMIC databases. Adults with ARF and sepsis who required IMV were included. Exclusion criteria included age < 18 years, tracheostomy on admission, death within 48 hours, reintubation, extubation within 72 hours, and admissions during 2020–2022. MP was calculated at intubation and 72 hours using the equation: MP = 0.098 × respiratory rate × tidal volume × (peak pressure – PEEP) + PEEP. Gradient Boosting (GB) and Random Forest (RF) models were trained on MIMIC-IV using a train-test split, with class imbalance addressed via SMOTE. External validation was performed on K-MIMIC. Model performance was assessed using ROC AUC, accuracy, and precision. Results: For models incorporating MP, age, sex, and cumulative ventilator days achieved strong internal performance on MIMIC-IV (RF: AUC 0.87, accuracy 0.82, precision 0.72; GB: AUC 0.85, accuracy 0.80 GB, precision, 0.67) but declined on external validation with K-MIMIC (RF AUC 0.61, GB AUC 0.59). MP at 72 hours outperformed MP at intubation in predictive contribution. Including MP improved performance beyond demographic and clinical features alone. Conclusions: Mechanical power at 72 hours provides measurable prognostication in anticipating tracheostomy in septic ARF patients. While not an independent predictor, MP may serve as an adjunct variable in early tracheostomy decision. External validation limitations demonstrate the need for improving generalizability between databases. Future work should explore improved regularization and robust modeling across diverse ICU populations.
Vijay et al. (Sun,) studied this question.