Automated ICD Coding via Contrastive Learning With Back-Reference and Synonym Knowledge for Smart Self-Diagnosis Applications

Zhizhuo Zhao, Wenpeng Lu, Xueping Peng, Lumin Xing, Weiyu Zhang, Chaoqun Zheng

Abstract: Smart applications are essential in intelligent consumer electronics. With the rising focus on health concerns, self-diagnosis applications on smart devices have gained widespread popularity, attracting significant attention from both consumers and researchers. These applications play a crucial role in accurately predicting potential diseases based on user-provided symptom descriptions, thereby facilitating effective treatment. Distinguishing between diseases with similar symptoms yet requiring different treatments is essential. However, automated ICD coding faces challenges due to subtle differences between some ICD codes. Although existing methods has utilized contrastive learning to differentiate between codes, precise selection of high-quality negative samples and accurate matching of code descriptions with patient conditions remain critical factors influencing their performance. Drawing inspiration from medical practices, doctors tend to utilize back-reference knowledge to differentiate between similar diseases. We propose a novel framework for automated ICD coding via contrastive learning with b ack- r eference and s ynonym k nowledge (BRSK) for smart self-diagnosis applications. Specifically, by incorporating back-reference knowledge, we generate challenging negative samples, enhancing the effectiveness of contrastive learning. Additionally, multi-view synonym matching networks are used to fully leverage synonym knowledge, resulting in more precise modeling of semantic matching between codes and patient descriptions. Extensive experiments on MIMIC-III datasets demonstrate that BRSK significantly outperformsstate-of-the-art baselines The source code and data are publicly available at https://github.com/zhaozhizhuo/ICD-BRSK .