Inductive Representation Learning of Multiple ICD Codes for Healthcare
Sheng Jie Lui, Xiang Cheng, Shonali Priyadarsini Krishnaswamy · 2022
The International Classification of Disease (ICD) coding scheme authorized by the World Health Organization, is a standard used to classify medical diagnosis in health insurance claims. One of the key challenges in adopting machine learning techniques for health insurance claims processing lies in the sparse and complex nature of ICD codes. There are over 69,000 unique diagnosis codes in the 10th version of the ICD coding scheme where multiple codes are used to represent complex medical diagnosis. In this work, we address the challenge of embedding both single and multiple ICD codes into a vector space. This representation addresses the sparse representation of ICD codes by providing a more compact representation that captures their diagnosis relationship. Our proposed inductive multi-code representation introduces a novel application of inductive node embedding by representing multiple ICD codes as a multi-code node. This enables both single and multiple ICD codes to be represented in the same vector space. For multi-code relevance, our proposed approached surpassed the benchmark node2vec implementation by 34%, highlighting its effectiveness.