Advancing Bloodstream Infection Prediction Using Explanable Artificial Intelligence Framework

Bloodstream infections (BSIs) represent a critical public health concern, primarily due to their rapid progression and severe implications such as sepsis and septic shock. This study introduces an innovative Explanable Artificial Intelligence (XAI) framework, leveraging historical electronic health records (EHRs) to enhance BSI prediction. Unlike traditional models that rely heavily on real-time clinical data, our XAI-based approach utilizes a comprehensive dataset incorporating demographic data, laboratory results, and full medical histories from St. Olavs Hospital, Trondheim, Norway, covering 35,591 patients between 2015 and 2020. We developed models to differentiate between high-risk and low-risk BSI cases effectively, optimizing healthcare resource allocation and potentially reducing healthcare costs. Our results demonstrate superior predictive accuracy, particularly the tree-based models, which significantly outperformed contemporary models in both specificity and sensitivity metrics. Author Summary In this research, we have developed a new tool that uses artificial intelligence to better predict bloodstream infections, which can lead to serious conditions like sepsis if not quickly identified and treated. It is the first of its kind framework that analyzes past health records and helps identify patients who are at high risk of infection more accurately than existing tools. Unlike existing tools our framework can be implemented at any stage of the patient trajectory and is the only framework to achieve good accuracy without the use of intimate patient features such as vital signs. This ability could enable doctors to prioritize care more pre-emptively, effectively, potentially saving lives and reducing unnecessary medical tests. Our approach is designed to be easily understood and used by both medical professionals and those with little technical expertise, making it a valuable addition to hospital systems.