Drug-Induced Acute Kidney Injury Risk Prediction Models

Abstract

Background: Acute kidney injury (AKI) risk prediction models can predict AKI with short lead times and excellent model performance. However, these prediction models have not ascertained the etiology of the AKI. Drugs are an important contributor to AKI, and it is difficult to distinguish drug causes from other etiologies.

Summary: Clinical adjudication of AKI etiology can reduce misclassification associated with temporal relationships, since expert adjudicators are trained to assess an outcome in a consistent manner using standardized definitions. Drug-induced acute kidney injury (DI-AKI) varies by drug and is heterogeneous in onset and mechanisms, limiting a universal approach to risk prediction and necessitating expert review. DI-AKI models should be constructed using a high-quality prospective dataset to maximize model performance, internal and external validity. Predictor selection and engineering requires careful attention to unique issues arising from interactions such as drug dose and concentrations. Various statistical methods, such as least absolute shrinkage and selection operator regression or advanced machine learning techniques, may be applied to perform feature selection and capture trends and interactions between predictors. Finally, the model should be carefully evaluated by internal and external validation.

Key messages: The development of DI-AKI risk prediction models is needed to identify high-risk patients, identify new risk factors, formulate, and apply preventative strategies. DI-AKI risk prediction models require a well-defined dataset of clinically adjudicated cases to improve model performance, validity, and reduce the risk of misclassification.

Trial registration: ClinicalTrials.gov NCT02159209.