Utilizing Machine Learning Predictive Analytics to Enhance Early Sepsis Diagnosis in Critical Care Setting

Abstract

Sepsis remains a critical medical challenge worldwide, with millions of cases and fatalities annually, particularly affecting vulnerable populations like children. The evolving understanding of sepsis has led to refinements in diagnostic criteria, including the introduction of novel metrics like qSOFA. Despite advancements, early diagnosis remains pivotal for improved outcomes. This study leverages electronic medical records (EMR) and machine learning to propose a predictive model for early sepsis diagnosis in critical care patients. Using the eICU Collaborative Research Database, the dataset comprising vital signs was preprocessed and balanced. Various machine learning algorithms, including deep learning models, were evaluated using 10-fold cross-validation. Performance metrics are AUROC, accuracy, F1 score, and Brier score were computed to assess model effectiveness. XGBoost emerged as the top-performing algorithm, exhibiting superior performance across metrics compared to other models with AUROC = 0.78, Accuracy = 80, F1-Score = 72, Brier-Score = 0.20. While demonstrating promising results, XGBoost exhibited higher false-positive rates, indicating the need for further refinement. Overall, the study underscores the potential of machine learning in early sepsis diagnosis, providing insights for future optimization and clinical application.