Breaking barriers in ICD classification with a robust graph neural network for hierarchical coding

Abstract

Abstract The accurate classification of International Classification of Diseases (ICD) codes is a complex and critical multi-label task in clinical documentation, involving the assignment of diagnostic codes to medical discharge summaries. Existing automated methods face challenges due to the sparsity and nuanced nature of medical text, while traditional backpropagation-based models often lack flexibility and robustness. To address these issues, we propose Labeled Graph Generation with Node Representation Grasp (LGG-NRGrasp), an advanced adversarial learning framework that models ICD coding as a labeled graph generation problem. By leveraging a hierarchical structure to refine feature learning, our approach addresses the issue of over-smoothing in deep graph neural networks. A key innovation of LGG-NRGrasp is the integration of adversarial reinforcement learning and domain adaptation techniques, which enhance its ability to generalize across heterogeneous datasets. Extensive evaluations on benchmark datasets indicate that LGG-NRGrasp markedly surpasses leading models, exhibiting enhanced performance and dependability in automated ICD coding.