Machine learning prediction models for mechanical ventilation requirement in patients with embolic stroke: a retrospective cohort study based on MIMIC-IV.

OBJECTIVE: This study seeks to develop and compare four machine learning models using the MIMIC-IV database to identify predictive variables associated with invasive mechanical ventilation (MV) requirement in patients with embolic cerebral infarction. METHODS: This study included 568 patients with embolic cerebral infarction from the MIMIC-IV database. Data collected encompassed demographic information, vital signs, laboratory results, and other relevant variables. The dataset was randomly split into a development set and a validation set in a 7:3 ratio. Feature selection was performed using univariable and multivariable analyses. Four prediction models, namely decision tree (DT), logistic regression (LR), eXtreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM), were developed. Model performance was assessed using receiver operating characteristic (ROC) curve analysis. Calibration was evaluated through calibration curves and Brier scores. Clinical utility was examined using decision curve analysis (DCA), and Shapley additive explanations (SHAP) were used to interpret model predictions. ICU admission was t0 and predictors were summarized over the first 24 h. Because exact MV-initiation times were unavailable, robustness analyses addressed optimism, class imbalance, and temporal ambiguity using repeated nested cross-validation, an all-MV benchmark, inverse-frequency weighting, and exclusion of pneumonia and log input amount. RESULTS: Among 568 patients, 438 received invasive MV. XGB achieved an AUROC of 0.839 (95% CI, 0.796-0.883) in the development set and 0.720 (95% CI, 0.637-0.801) in the validation set; repeated nested cross-validation yielded 0.713 (95% CI, 0.665-0.757). The all-MV benchmark had a specificity of 0 and 0.500 balanced accuracy, whereas XGB achieved 0.700 and 0.659, respectively. Excluding pneumonia and log input amount reduced the nested-CV AUROC to 0.654 (95% CI, 0.595-0.703). SHAP analysis identified pneumonia, hematocrit, INR, weight, heart failure, log input amount, and glucose as the highest-importance features. CONCLUSION: This study developed four machine learning models to assess the necessity of mechanical ventilation support in patients with embolic cerebral infarction and compared their performance. XGB showed the most favorable overall validation profile but requires time-stamped external validation before clinical use.