An Efficient and Credible Prediction Network for High-Dimensional Critical Care Data

Medical data prediction represents a key area of AI in medicine, where it plays a significant role in early disease detection and supports doctors' decision-making. With the continuous development of medical information systems, revealing more complex heterogeneous features and interrelationships among them. Existing methods that rely on manual feature engineering depend heavily on doctors' prior knowledge, which impedes the efficient prediction of patient diagnosis and prognosis. To address this issue, We propose an efficient and credible prediction network (FRGN) based on graph networks. This network extracts feature relationships from data and provides explanations for prediction results. Specifically, we first embed the heterogeneous features into the same feature space. Then, we propose a graph structure initialization method tailored to the characteristics of critical care data, which applies prior constraints to the graph to enhance the model's training speed. Finally, the graph structure learner and feature relationship encoder jointly encode the relationships among features, generating high-quality node feature representations and an accurate feature relationship model for downstream tasks. We conduct experiments on four large public critical care datasets, and the results demonstrate that compared to previous methods, the proposed FRGN network achieves greater efficiency and accuracy in predictions, while elucidating a credible feature relationship structure.