Multi-Label Classification of ICD Coding Using Deep Learning

This study uses deep learning approach to tackle the multi-label classification problem in ICD coding. The discharge summaries on MIMIC-III dataset are adopted to explore the training methods of text preprocessing, label preprocessing, and model training. Specifically, the methods of label-to-chapter and common label classification are experimented, hoping to find the best recommendations of ICD codes for each diagnosis by the physicians. The result shows CNN has the best performance 76% by micro F1measure in label-to-chapter. In addition, it also shows CNN outperforms other methods within top-50 and top-100 common labels.