Multitask gated interactive network for automatic international classification of diseases coding with dual denoising mechanism

Xiaobo Li, Yi-Jia Zhang, Xiaodi Hou, Shilong Wang, Wen Qu

Abstract: Recent advancements in deep learning have significantly impacted the field of medicine and healthcare, offering promising solutions for automatic International Classification of Diseases (ICD) coding. ICD coding plays a vital role in mapping clinical records to the corresponding ICD codes, providing doctors with faster and more accurate diagnostic support. However, traditional manual coding is inefficient and error-prone, while deep learning models face challenges such as noisy and lengthy clinical text, complex connections among labels, and imbalanced label distributions. Therefore, we propose a novel Multitask Gated Interactive Network (MGIN) with a dual denoising mechanism. Specifically, the gated recalibration module is designed to combine cascaded convolutions and gated units to filter irrelevant text and extract the most ICD-relevant fragments from lengthy clinical notes. The multitask interactive sharing module aims to jointly train ICD and Clinical Classification Software (CCS) codes for sharing parameter information and incorporates a co-occurrence frequency matrix to capture the complex associations between medical codes. Additionally, a dual denoising module simultaneously combats class imbalance by emphasizing rare samples through adaptive loss weighting while further suppressing noisy samples by dynamically discard high-loss training instances. Finally, comprehensive experiments on real-world datasets have demonstrated that our proposed model MGIN outperforms multiple competitive baselines.

External IDs:dblp:journals/eaai/LiZHWQ25