Early prediction of prolonged mechanical ventilation in invasively ventilated patients with subarachnoid hemorrhage: A MIMIC-IV prediction model study.

OBJECTIVE: To develop and internally validate an early prediction model for prolonged mechanical ventilation (PMV) in invasively ventilated patients with subarachnoid hemorrhage (SAH) using MIMIC-IV. METHODS: This retrospective cohort study used MIMIC-IV (version 3.1). Adult patients with coded SAH who received invasive mechanical ventilation during the indexed ICU stay were included. PMV was defined as a cumulative invasive ventilation duration of more than 7 days during the indexed ICU stay. The prespecified primary manuscript model was a multivariable logistic regression model with grouped Glasgow Coma Scale (GCS) categories. Apparent performance was assessed using the area under the receiver operating characteristic curve (AUROC), the area under the precision-recall curve (AUPRC), Brier score, and calibration. Clinical utility was evaluated using decision curve analysis. Internal validation was performed with bootstrap optimism correction using 1000 resamples. Missing-data robustness was examined in a multiple-imputation sensitivity analysis. Random forest, XGBoost, and LightGBM were evaluated in exploratory comparative analyses. RESULTS: The final main cohort included 623 invasively ventilated patients with SAH, of whom 151 (24.2%) developed PMV. The complete-case derivation cohort comprised 557 patients, including 144 PMV events (25.9%). In the primary grouped-GCS logistic regression model, SOFA score, respiratory rate, and temperature were the most consistent early predictors of higher PMV risk. Apparent model performance was modest, with an AUROC of 0.677 (95% CI: 0.624-0.727), an AUPRC of 0.400 (95% CI: 0.333-0.488), and a Brier score of 0.1794. After bootstrap internal validation, the optimism-corrected AUROC was 0.608 and the optimism-corrected calibration slope was 0.600, indicating substantial optimism and overfitting. In the multiple-imputation sensitivity analysis, mean AUROC was 0.671. Decision curve analysis demonstrated that the model provided net benefit exceeding the treat-all strategy across threshold probabilities of approximately 16-60%. Comparative machine-learning models showed similarly modest performance and did not demonstrate a clear advantage over the primary logistic model; mean AUROCs were 0.609 for random forest, 0.612 for XGBoost, and 0.609 for LightGBM. CONCLUSIONS: In invasively ventilated patients with SAH, a grouped-GCS logistic regression model provided an interpretable but modest-performing approach to early PMV risk stratification. SOFA score, respiratory rate, and temperature were the most stable early predictors across analyses. Decision curve analysis suggested potential clinical utility at threshold probabilities approximating the observed PMV prevalence, though these findings are exploratory. Exploratory comparative machine-learning models did not demonstrate a clear advantage over the primary logistic model. Given the modest discrimination, substantial internal optimism, and single-centre derivation setting, external validation and further refinement are required before any clinical implementation.