Explainable machine learning to identify patients at risk of developing hospital acquired infections

Hospital-acquired infections (HAIs) contribute to increased mortality rates and extended hospital stays. Patients with complex neurological impairments, secondary to conditions such as acquired brain injury or progressive degenerative conditions are particularly prone to HAIs and often have the worst resulting clinical outcomes and highest associated cost of care. Research indicates that the prompt identification of such infections can significantly mitigate mortality rates and reduce hospitalisation duration. The current standard of care for timely detection of HAIs for inpatient acute and post-acute care settings in the UK is the National Early Warning Score v02 (NEWS2). NEWS2, despite its strengths, has been shown to have poor prognostic accuracy for specific indications, such as infections. This study developed a machine learning (ML) based risk stratification tool, utilising routinely collected patient electronic health record (EHR) data, encompassing over 800+ patients and 400k+ observations collected across 4-years, aimed at predicting the likelihood of infection in patients within an inpatient care setting for patients with complex acquired neurological conditions. Built with a combination of historical patient data, clinical coding, observations, clinician reported outcomes, and textual data, we evaluated our framework to identify individuals with an elevated risk of infection within a 7-day time-frame, retrospectively over a 1-year "silent-mode" evaluation. We investigated several time-to-event model configurations, including manual feature-based and data-driven deep generative techniques, to jointly estimate the timing and risk of infection onset. We observed strong performance of the models developed in this study, achieving high prognostic accuracy and robust calibration from 72-6 hours prior to clinical suspicion of infection, with AUROC values ranging from 0.776-0.889 and well-calibrated risk estimates exhibited across those time intervals (IBS<0.178). Furthermore, by assigning model-generated risk scores into distinct categories (low, moderate, high, severe), we effectively stratified patients with a higher susceptibility to infections from those with lower risk profiles. Post-hoc explainability analysis provided valuable insights into key risk factors, such as vital signs, recent infection history, and patient age, which aligned well with prior clinical knowledge. Our findings highlight our frameworks potential for accurate and explainable insights, facilitating clinician trust and supporting integration into real-world patient care workflows. Given the heterogeneous and complex patient population, and our under-utilisation of the data recorded in routine clinical notes and lab reports, there are considerable opportunities for performance improvement in future research by expanding our models multimodal capabilities, generalisability, and additional model personalisation steps.