Abstract
Patients with intracerebral hemorrhage (ICH) are highly susceptible to sepsis. This study evaluates the efficacy of machine learning (ML) models in predicting sepsis risk in intensive care units (ICUs) patients with ICH. We conducted a retrospective analysis on ICH patients using the MIMIC-IV database, randomly dividing them into training and validation cohorts. We identified sepsis prognostic factors using Least Absolute Shrinkage and Selection Operator (LASSO) and backward stepwise logistic regression. Several machine learning algorithms were developed and assessed for predictive accuracy, with external validation performed using the eICU Collaborative Research Database (eICU-CRD). We analyzed 2,214 patients, including 1,550 in the training set, 664 in the validation set, and 513 for external validation using the eICU-CRD. The Random Forest (RF) model outperformed others, achieving Area Under the Curves (AUCs) of 0.912 in training, 0.832 in internal validation, and 0.798 in external validation. Neural Network and Logistic Regression models recorded training AUCs of 0.840 and 0.804, respectively. ML models, especially the RF model, effectively predict sepsis in ICU patients with ICH, enabling early identification and management of high-risk cases.