Abstract
Acute kidney injury (AKI) is associated with mortality greater than 50% in critical care due to the fact that the overall incidence of AKI in ICU patients ranges from 20% to 50%. The past decade has witnessed the rapid development of machine learning-driven techniques for AKI monitoring. In this paper, we present a novel framework for AKI prediction using Graph Neural Network and gradient boosting. We cast the real-time AKI prediction into a graph classification problem by treating each patient as an attributed graph where the attribute of each node (a specific type of laboratory tests) evolves continuously over time. We use Adaptive Graph Convolutional Recurrent Network to learn the adjacency matrix by capturing hidden dependencies between AKI-related laboratory tests and make numerical predictions on various laboratory tests. Extensive experiments using MIMIC-IV show that our model outperforms the state-of-the-art approaches in AKI prediction for up to 48 hours in terms of all effectiveness metrics.