This paper presents a novel framework for automatic International Classification of Diseases (ICD) coding based on Chinese electronic medical records (EMRs), MKE-Coder. Due to the concise description and special internal structure of Chinese EMRs, it is difficult to extract disease code-related information, and existing methods fail to utilize disease-based multi-axis knowledge and lack correlation with clinical evidence. To solve this problem, MKE-Coder first identifies candidate codes for diagnoses and classifies them into four coding axes. Then, it searches for corresponding clinical evidence from the comprehensive content of EMRs and filters reliable evidence through a scoring model. Finally, it verifies the validity of candidate codes through an inference module based on masked language modeling strategy, checks whether all axis knowledge related to the candidate codes are supported by evidence, and provides recommendations accordingly. Experiments are conducted using a large Chinese EMR dataset collected from various hospitals, and the results show that MKE-Coder performs well in automatic ICD coding tasks based on Chinese EMRs. Practical evaluations in simulated real-world coding scenarios demonstrate that it greatly contributes to improving the coder’s coding accuracy and speed.
