Developing Interpretable Bayesian Machine Learning Model for Early Sepsis Prediction and Clinical Decision making in U.S. Healthcare Systems

Abstract

Sepsis is one of the major contributors to morbidity and mortality in the health care system and can be diagnosed and treated rapidly enough in intensive care units and emergency departments to enhance patient outcomes. Classical machine learning methods have shown to be effective for predicting sepsis onset, but lack interpretability and difficulty in modelling predictive uncertainty, thereby limiting their clinical uptake. This study suggests building an interpretable Bayesian machine learning model to predict sepsis early and to make clinical decisions in healthcare systems across the United States. The proposed framework uses EHR data such as vital signs, laboratory measurements, demographics and comorbidity profiles to calculate real-time patient-specific sepsis risk. The model makes probabilistic predictions and offers uncertainty estimates to improve transparency and trust by clinicians. To improve understanding of model recommendations, mechanisms of explainability have been incorporated, to help identify the most influential clinical variables involved in the onset of sepsis. The framework is assessed based on the performance benchmarks: accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUROC) and is compared with the existing machine learning and deep learning methods. The results validate that interpretable Bayesian models can provide strong prediction performance, yet provide clinically relevant explanations and uncertainty quantification. The proposed solution can help advance the early detection of sepsis, support evidence-based clinical decisions, and, ultimately, positively impact patient outcomes in different healthcare systems in the United States.