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 the US. Urosepsis occurs due to complicated UTI and constitutes ~25% of sepsis cases. Early prediction of urosepsis is critical in providing personalized care, reducing diagnostic uncertainty, lowering mortality rates. While machine learning (ML) techniques have the potential to aid healthcare professionals in identifying risk factors and recommending treatment options, no study has been developed to predict the development of urosepsis in patients with a suspected UTI in outpatient settings. We develop and evaluate ML models in predicting hospital admission and urosepsis diagnosis for patients with an outpatient UTI encounter, leveraging de-identified electronic health records sourced from UCLA. Inclusion criteria included a positive diagnosis of urinary tract infection indicated by ICD-10 code N30/N93.0 and positive bacteria result via urinalysis in an ambulatory setting. W extracted demographic information, urinalysis findings, and antibiotics prescribed for each instance of UTI. Reencounters we defined as encounters within seven days of the initial UTI. Reencounters were considered urosepsis-related if matching positive blood and urine cultures were found with a sepsis ICD-10 code of A41/R78/R65. ML models were trained and evaluated on two tasks: prediction of a reencounter leading to hospitalization, prediction of Urosepsis. Model performances were stratified by ethnicities. Random forest models achieved significant improvement over baseline performance (APR = 0.004), with APR = 0.15 for reencounters and 0.31 for urosepsis prediction. While these APR values reflect the challenge of predicting rare events (0.4% prevalence), they represent meaningful predictive power for clinical risk stratification. We computed Shapley values to interpret model predictions and found 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, informed decisions about antibiotic prescription and improving patient care.