DRCNNTLe: A deep recurrent convolutional neural network with transfer learning through pre-trained embeddings for automated ICD coding

The International Classification of Diseases (ICD), which is endorsed by the World Health Organization, is a diagnostic classification standard. ICD codes store, retrieve, and analyze health information to make clinical decisions. Currently, ICD coding has been adopted by more than 137 countries. However, in Pakistan, very few hospitals have implemented ICD coding and conducted different epidemiological studies. Moreover, none of them have reported the spectrum of liver disease burden based on ICD coding, nor implemented automated ICD coding. In this study, we annotated ICD codes for the database of the liver transplant unit of the Pir Abdul Qadir Shah Jeelani Institute of Medical Sciences. We named this database Medical Information Mart for Liver Transplantation (MIMLT). The results revealed that the database contains 34 ICD codes, of which V70.8 is the most frequent code. Furthermore, we determined the spectrum of liver disease burden in liver recipients based on ICD coding. We found that chronic hepatitis C (070.54) is the most frequent indication for liver transplantation. Additionally, we implemented automated ICD coding utilizing the MIMLT database and proposed a novel Deep Recurrent Convolutional Neural Network with Transfer Learning through pre-trained Embeddings (DRCNNTLe) model, which is an extended version of our DRCNN-HP model. DRCNNTLe extracts robust text representations from its pre-trained embedding layer, which is trained on a large domain-specific MIMIC III database corpus. The results indicate that utilizing pre-trained word embeddings, which are trained on large domain-specific corpora can significantly improve the performance of the DRCNNTLe model and provide state-of-the-art results when the target database is small.