Automatic ICD Coding Based on Bias Removal
Xun Peng, Tengkai Tan, Teng Fan · 2024
The automated coding task aims to match medical text records with the corresponding International Classification of Diseases (ICD) codes to improve the efficiency and accuracy of medical record management. In the automatic coding task, existing methods often face the challenge of label bias, a problem that affects the overall performance of the model. To address this challenge, we propose a new bias removal method that aims to optimize model performance. Furthermore, we note that some samples are difficult to be recognized during the encoding process due to their complexity. Given the huge amount of data contained in the large language models, we attempted to use these models to recognize these hard samples and compared their effectiveness with our debiasing method. Test results on the MIMIC-III dataset show that we find our proposed debiasing method significantly outperforms the method that relies only on large language models in dealing with hard samples, thus confirming the effectiveness and superiority of our method.