Case study for the development of an acute kidney injury prediction model for clinical use

Abstract

To apply an artificial intelligence disease prediction model in actual clinical settings, it is essential to ensure interpretability consistent with medical research and provide a multidimensional analysis of the model’s operation. Particularly in the case of Acute Kidney Injury (AKI), since the etiology is diverse, it is necessary to determine whether the model accurately predicts the reasons for the occurrence of acute kidney injury to enable appropriate preventive measures. In this study, we implemented an RNN-based AKI prediction model and achieved performance at a level suitable for clinical use through feature engineering that incorporates medical knowledge. We also applied Layer-wise Relevance Propagation methods to the model to provide a foundation for its clinical application. Additionally, we present the results of individual case analyses of the dataset predicted by the model. This includes cases that were classified as false negatives in early prediction but eventually indicated kidney injury, cases not classified as acute kidney injury but with a poor prognosis, and cases where alerts are practically meaningful. These analyses offer examples of patient types that can occur in clinical applications.