Abstract
Acute Kidney Injury (AKI) is a quickly growing complication in critical care patients. AKI is associated with increased period of hospital stay, increased risk of long-term kidney injury, chronic kidney disease and mortality. Hence it is crucial to develop methods for early detection and prediction of AKI to improve patient outcomes. The existing solutions to AKI prediction suffer from data sparsity issue. In other words, they rely on a well-balanced dataset to achieve better prediction results. However, class imbalance usually abounds in medical diagnosis due to rareness of the studied diseases. To solve this issue, we develop a new framework for early AKI detection and prediction using Conditional Variational Autoencoder (CVAE). Extensive experiments show that our model outperformed the state-of-the-art models for all the six continuous 8-hour windows. Our framework can be effectively integrated into hospitals’ Electronic Health Records (EHR) systems to assist healthcare professionals and providers in predicting AKI occurrences up to 48-hours in-advance for real-time patient monitoring.