MANet: A Multiview Attention Network for Automatic ICD Coding

Hualei Shen, Haizhu Wang, Guoqing Shangguan, Dong Liu

Abstract: Automatically assigning multiple International Classification of Diseases (ICD) codes to a clinical note is challenging due to the unstructured and verbose nature of medical records. Currently, most methods employ deep neural networks to learn the representation of clinical notes from a single perspective. These single-view-based methods overlook the exploitation and fusion of multiview features to enhance the precision of ICD coding. In this paper, we propose a new Multiview Attention Network (MANet) to extract and fuse multiview features for ICD coding. MANet includes a specially designed multiview attention scheme to extract and fuse coarse, global, and local features from clinical notes. Additionally, a novel cascaded multilayer perceptron (MLP) block and a multiscale convolution block are designed to extract global and local features, respectively. Self-attention and cross-attention are integrated to effectively fuse these multiview features, generating more informative and discriminative representations. Extensive experiments conducted on the popular MIMIC-III and MIMIC-IV-ICD9 datasets demonstrate the superiority of our proposed MANet over state-of-the-art methods. On MIMIC-III, MANet achieves a Macro-AUC of 0.953, Micro-AUC of 0.993, Macro-F1 of 0.135, Micro-F1 of 0.596, precision at top 8 (P@8) of 0.773, and precision at top 15 (P@15) of 0.617. On MIMIC-IV-ICD9, MANet achieves a Macro-AUC of 0.968, Micro-AUC of 0.996, Macro-F1 of 0.146, Micro-F1 of 0.614, P@8 of 0.698, and P@15 of 0.526.