MACHINE LEARNING MODELS FOR PREDICTING SEPSIS ONSET IN ICU PATIENTS BASED ON EHR DATA

Abstract

Early detection of sepsis remains a clinical priority in intensive care units due to its association with high morbidity and mortality rates. This study presents a machine learning-based framework for predicting sepsis onset using electronic health record data, with the goal of enabling timely clinical intervention. Multiple algorithms—including ensemble models, support vector machines, and linear classifiers—were trained on a dataset comprising vital signs, lab values, medications, and demographics. Results demonstrated that ensemble models achieved superior predictive performance, with significant gains in accuracy, precision, recall, and F1-score after hyperparameter tuning. Key predictive features identified included lactate concentration, respiratory rate, and white blood cell count. Visualization tools such as SHAP values, temporal risk plots, and feature importance charts enhanced the interpretability of the models. Comparative analysis with conventional clinical scoring systems revealed that the machine learning models offered improved sensitivity and earlier detection. The study also highlights critical challenges such as data heterogeneity, algorithmic transparency, and clinician adoption barriers. Despite these challenges, the findings underscore the transformative potential of machine learning in critical care, particularly in establishing real-time early warning systems for sepsis. The integration of interpretable and accurate models into clinical workflows could significantly reduce diagnostic delays and improve patient outcomes. This research contributes to the advancement of precision medicine by leveraging artificial intelligence for real-world clinical decision support.