Hybrid Attention Knowledge Fusion Network for Automated Medical Code Assignment

Shilong Wang, Xiaobo Li, Wen Qu, Hongfei Lin, Yi-Jia Zhang

Abstract: Automated medical code assignment aims to allocate disease and procedure codes to patient's discharge summaries, which is crucial for health statistics, medical decision-making, and reimbursement. To alleviate the burden on manual coders and improve coding efficiency, several methods for automated code assignment have been proposed. However, challenges persist in automated medical code assignment due to knowledge gaps, the complexity and verbosity of clinical text, and class imbalance issue. In this study, we present a Hybrid Attention Knowledge Fusion Network (HAKFN) with the Transformer-based architecture to overcome the aforementioned challenges. We introduce structured tabular data from patient's electronic health records (EHRs) as additional auxiliary knowledge for the model and design a hybrid attention module to integrate clinical text with tabular data to enhance text representations. To mitigate class imbalance issue, we employ focal loss to rebalance the model's attention toward low-frequency and high-frequency codes. Results of the experiments demonstrate that our method achieves competitive performance.

External IDs:dblp:conf/isbra/WangLQLZ24