Machine learning for the prediction of urosepsis using electronic health record data

Urosepsis, a medical condition resulting from the progression of urinary tract infection (UTI), is a leading cause of death in hospitals in the United States. Urosepsis commonly occurs due to complicated UTI and constitutes approximately 25% of all sepsis cases. Early prediction of urosepsis is critical in providing personalized care, reducing diagnostic uncertainty, and ultimately lowering mortality rates. While machine learning techniques have the potential to aid healthcare professionals in identifying potential risk factors, and high-risk patients, and recommending treatment options, no existing study has been developed so far to predict the development of urosepsis in patients with a suspected UTI presenting to an outpatient setting. In this research study, we develop and evaluate the utility of multiple machine learning models to predict the likelihood of hospital admission and urosepsis diagnosis for patients with an outpatient UTI encounter, leveraging de-identified electronic health records sourced from a large health care system encompassing a wide range of encounters spanning primary to quaternary care. Inclusion criteria included a positive diagnosis of urinary tract infection indicated by ICD-10 code N30 or N93.0 and positive bacteria result via urinalysis in an ambulatory setting (primary or emergent care settings). For these patients, we extracted demographic information, urinalysis findings, and any antibiotics prescribed for each instance of UTI. Reencounters we defined as all encounters within seven days of the initial UTI encounter. The reencounters were considered urosepsis-related if matching positive blood and urine cultures were found with a sepsis ICD-10 code of A41, R78, or R65. A variety of machine learning models were trained on this rich feature set and were evaluated on two tasks: the prediction of a reencounter leading to hospitalization, and the prediction of Urosepsis. Model performances were stratified by the patient ethnicities. Our models demonstrated high predictive performance with an area under the ROC curve (AUC) of 79.5% AUC and an area under the precision-recall curve (APR) of 13% APR for reencounters, and 90% ROC and 31% APR for Urosepsis. We computed shapley values to interpret our model predictions and found the patient age, sex, and urinary WBC count were the top three predictive features. Our study has the potential to assist clinicians in the identification of high-risk patients, making more informed decisions about antibiotic prescription and providing improved patient care.