Clinical Phenotyping in the Prediction of Pediatric Acute Kidney Injury

Abstract

Predicting pediatric acute kidney injury is a difficult but important task. Accurate prediction would allow preventative measures to be taken before kidney injury occurs, decreasing the morbidity and mortality associated with this disease. This work describes the process of creating an “at risk for AKI” clinical phenotype from electronic health record data, which is then used to predict AKI in a retrospective data set. This predictive model has reasonable performance, with an F1 score of 0.67 and AUC of 0.75. In a subset of intensive care unit patients, the addition of unstructured data from clinician notes improves the model’s F1 score to 0.72 and AUC to 0.77, suggesting a possible role for natural language processing in refining clinical phenotypes. Interpreting these models requires careful consideration of the information contained within each variable – specifically, the extent to which that information describes biologic processes within a patient or systemic processes within a hospital. Further evaluation of the use of clinical phenotyping in predicting pediatric AKI is necessary to confirm the utility of these models.