DLGNet: Adversarial Reinforcement Learning for Disease Label Generation

Abstract

International Classification of Diseases (ICD) coding has been considered as a multi-label prediction problem, requiring the assignment of one or more codes to a detailed discharge summary. Existing automatic ICD coding algorithms struggle to effectively classify medical diagnosis texts representing deep sparse categories. We propose Disease Label Generation Network (DLGNet), a novel adversarial network that transforms ICD codes into a label generation challenge. This strategy faces three major challenges: (1) How to extract the relationship between clinical text and ICD codes? (2) What training methods should be used to improve the generalization and effectiveness of network? (3) How to evaluate the quality of disease labels generated by the DLGNet? For (1), we develop an information integration module (MIM) to encode the relationship between clinical text and ICD codes. For (2), we present adversarial training algorithms, such as reinforcement causal learning and adversarial perturbation regularization. For (3), we present a Label Discriminator (LD) that calculates the reward for each ICD code in the Label Generator (LG). In conclusion, DLGNet outperforms existing state-of-the-art approaches on evaluation measures such as micro-F1, leading to the formation of a new SOTA. The code is available at anonymous github link: https://anonymous.4open.science/r/DLGNet-787D.