Prediction of Sepsis Mortality in ICU Patients Using Machine Learning Methods

Abstract

Problem: Sepsis, a life-threatening condition, accounts for the deaths of millions of people worldwide. Accurate prediction of sepsis outcomes is crucial for effective treatment and management. Previous studies have utilized machine learning for prognosis but have limitations in feature sets and model interpretability. Aim: This study aims to develop a machine learning model that enhances prediction accuracy for sepsis outcomes using a reduced set of features, thereby addressing the limitations of previous studies and enhancing model interpretability. Methods: This study analyzes intensive care patient outcomes using the MIMIC-IV database, focusing on adult sepsis cases. Employing the latest data extraction tools, such as Google BigQuery, and following stringent selection criteria, we selected 38 features in this study. This selection is also informed by a comprehensive literature review and clinical expertise. Data preprocessing included handling missing values, regrouping categorical variables, and using the Synthetic Minority Over-sampling Technique (SMOTE) to balance the data. We evaluated several machine learning models: Decision Trees, Gradient Boosting, XGBoost, LightGBM, Multilayer Perceptrons (MLP), Support Vector Machines (SVM), and Random Forest. The Sequential Halving and Classification (SHAC) algorithm was used for hyperparameter tuning, and both train-test split and cross-validation methodologies were employed for performance and computational efficiency. Results: The Random Forest model was the most effective, achieving an area under the receiver operating characteristic curve (AUROC) of 0.94 with a confidence interval of 0.93 to 0.95. This significantly outperformed other models and set a new benchmark in the literature. The model also provided detailed insights into the importance of various clinical features, with the Sequential Organ Failure Assessment (SOFA) score and average urine output being highly predictive. SHAP (Shapley Additive Explanations) analysis further enhanced the model's interpretability, offering a clearer understanding of feature impacts. Conclusion: This study demonstrates significant improvements in predicting sepsis outcomes using a Random Forest model, supported by advanced machine learning techniques and thorough data preprocessing. Our approach provided detailed insights into the key clinical features impacting sepsis mortality, making the model both highly accurate and interpretable. By enhancing the model's practical utility in clinical settings, we offer a valuable tool for healthcare professionals to make data-driven decisions, ultimately aiming to minimize sepsis-induced fatalities.