Development and validation of a machine learning model to predict 30-day mortality in ischemic stroke patients with consciousness impairment: Insights from MIMIC-IV database and multicenter ICU data in China.

Abstract

Patients with ischemic stroke complicated by consciousness disorders remain associated with high mortality risks. This study aims to develop and validate an interpretable machine learning model using multicenter ICU data to predict 30-day mortality in this population. The MIMIC-IV database was utilized for model training, with data from three Chinese top-tier hospitals employed for external validation. Clinical variables extracted from the first 24 h following ICU admission were screened for key predictors using XGBoost and Random Forest algorithms. Ten-fold cross-validation and SMOTENC were implemented to balance datasets. Seven maximum likelihood estimation algorithms were implemented to develop predictive models. Model performance was comprehensively evaluated using the AUC-ROC, calibration curves, PR curves, and DCA et al. Additionally, the SHAP method was applied to interpret the model's decision-making mechanism. Among 529 MIMIC-IV patients, 158 (29.87 %) died within 30 days of ICU admission. Nine feature variables were selected for model development. The CatBoost model demonstrated the best predictive performance, with an internal validation AUC of 0.873 and an external validation AUC of 0.848. SHAP interpretability analysis revealed that the anion gap was the strongest driving factor for mortality risk; advanced age, hyperglycemia, and increased respiratory rate were significantly associated with higher risk, while systolic blood pressure, diastolic blood pressure, Na⁺, Ca²⁺, and AST levels further aided in risk stratification. We developed a CatBoost-based machine learning predictive model to estimate the 30-day mortality risk in ICU patients with ischemic stroke and impaired consciousness. The model's reliability and potential advantages were validated. SHAP interpretability analysis, along with the development of an interactive web-based prediction tool, provides clinicians with valuable support for the early identification of high-risk patients.