Why It Matters
IN PATIENTS with cerebral infarction, machine learning models using neurophysiological and clinical data predicted ICU readmission, with logistic regression delivering the highest discrimination and clinically interpretable predictors that can inform bedside risk stratification.
Study Design and Analytical Approach
Investigators conducted a retrospective cohort analysis of 3,348 adults with cerebral infarction from the MIMIC-IV database. Feature selection used LASSO to reduce dimensionality before multivariable modeling. Seven machine learning algorithms were trained and evaluated, including Decision Tree, K-Nearest Neighbors, LightGBM, Naïve Bayes, Random Forest, Support Vector Machine, and XGBoost. The logistic regression model achieved the top area under the receiver operating characteristic curve at 0.682 with a 95% confidence interval of 0.630 to 0.733. This performance advantage, together with transparent coefficients, supports clinical adoption where interpretability is essential.
Key Predictors and Clinical Signals
Significant predictors of ICU readmission included peptic ulcer disease, inpatient glucocorticoid use, serum potassium levels, and red blood cell count. These variables capture comorbidity burden, treatment exposure, electrolyte balance, and hematologic status. The profile indicates that gastrointestinal vulnerability, steroid-related risk, dyskalemias, and anemia may signal instability that warrants closer monitoring and proactive discharge planning in cerebral infarction.
Machine Learning for ICU Readmission Risk
The machine learning framework identified patients at elevated risk of returning to the ICU after stroke. While multiple algorithms were explored, logistic regression offered a balance of accuracy and interpretability that aligns with current clinical workflows. The authors conclude that model-based risk stratification may enable targeted surveillance, optimization of electrolytes and hematologic parameters, and earlier mitigation of complications in this population.
Reference: Su H et al. Predicting ICU Readmission in Patients With Cerebral Infarction: A Machine Learning Approach Using Neurophysiological and Clinical Data. Brain and Behavior. 2025;15(10):e70958.
