Automatic ICD-9 coding via deep transfer learning

Min Zeng, Min Li, Zhihui Fei, Ying Yu, Yi Pan, Jianxin Wang

Abstract: ICD-9 codes have been widely used to describe a patient's diagnosis. Accurate automatic ICD-9 coding is important because manual coding is expensive, time-consuming. Inspired by the recent successes of deep transfer learning, in this study, we propose a deep transfer learning framework for automatic ICD-9 coding. Our proposed method makes use of transferring MeSH domain knowledge to improve automatic ICD-9 coding. We demonstrate its effectiveness by achieving state-of-the-art performance with a value of 0.420 for Micro-average F-measure on MIMIC-III dataset, which indicates that our method outperforms hierarchy-based SVM and flat-SVM. Furthermore, we analyze the deep neural network structure to discover the vital elements in the success of our proposed method. Our experimental results indicate that transfer learning is the key component to improve the performance of automatic ICD-9 coding and deep learning approach is the foundation in the success of our proposed model. In addition, to explore the best network architecture, we also compare the performance of multi-scale and sequential network architectures and find that using multi-scale network is better. Finally, we investigate the effects of transferring different percentage of samples on transfer learning and the results show that the best performance of target domain task can be obtained when 100% number samples are transferred.