Michael G. Semanik

Acute kidney injury (AKI) is an increasingly prevalent problem amongst pediatric inpatients, and is associated with high morbidity and mortality. Unfortunately, current methods of diagnosing AKI rely on “late markers” of injury, making early identification and prevention of AKI difficult. This work describes the development of an “at risk for AKI” clinical phenotype from structured electronic health record data, and its ensuing application in a predictive model. The model performs reasonably well in predicting AKI, with an F1 score of 0.67 and AUC of 0.75. Unstructured data is then added to the model via the inclusion of n-grams derived from ICU clinician notes, which improves performance (the F1 score increases to 0.76 and AUC increases to 0.77). Thus, it is possible to use clinical phenotyping to predict the onset of AKI twenty-four hours before current markers are elevated. This approach may lead to better treatments and preventative strategies for pediatric AKI.