Prediction for Respiratory Failure in Ischemic Stroke Patients Admitted to ICU: A Retrospective Analysis Based on MIMIC‐IV Database

ABSTRACT Introduction There remains a lack of studies evaluating the risk of respiratory failure in intensive care unit (ICU)‐admitted ischemic stroke (IS) patients. We aim to develop a nomogram for the prediction of respiratory failure in those patients and the identification of the patients with high risk of respiratory failure, to facilitate early intervention. Methods The medical data of IS patients in the Medical Information Mart for Intensive Care (MIMIC)‐IV database were extracted. Variables were selected using Cox stepwise regression, and variables with statistical significance were finally included in the nomogram. The marginal structural Cox model (MSCM) was to adjust for baseline and time varying confounding factors. The calibration curve and Receiver operating characteristic curve (ROC) were applied to assess the performance of the model. Results External validation using IS patient data from the eICU collaborative ersearch database (eICU‐CRD). A total of 3462 eligible patients (2424 in the training set and 1038 in the validation set) were included. The following variables were finally included in the model: infarction location, atrial fibrillation, A alkaline phosphatase (ALP), anion gap (AG), lactic dehydrogenase (LDH), and Na 2+ concentration. The direction of the hazard ratios (HR) of the variables in the model is consistent with the MSCM results. The area under the ROC curve (AUC‐ROC) of respiratory failure occurring between 1 and 7 days after ICU admission was 0.839 and 0.760 in the training set, 0.839 and 0.769 in the validation set, and 0.687 and 0.733 in the eICU set, respectively. The calibration curve showed acceptable consistency, indicating the model was of satisfactory performance. Conclusion We have developed a nomogram model for the prediction of respiratory failure in IS patients admitted to the ICU, validated using external data. The model could perform effective prediction and thus provide more information for clinicians.