Deep learning for automatic ICD coding: Review, opportunities and challenges

Background:

The automatic International Classification of Diseases (ICD) coding task assigns unique medical codes to diseases in clinical texts for further data statistics, quality control, billing and other tasks. The efficiency and accuracy of medical code assignment is a significant challenge affecting healthcare. However, in clinical practice, Electronic Health Records (EHRs) data are usually complex, heterogeneous, non-standard and unstructured, and the manual coding process is time-consuming, laborious and error-prone. Traditional machine learning methods struggle to extract significant semantic information from clinical texts accurately, but the latest progress in Deep Learning (DL) has shown promising results to address these issues.

Objective:

This paper comprehensively reviewed recent advancements in utilizing deep learning for automatic ICD coding, which aimed to reveal prominent challenges and emerging development trends by summarizing and analyzing the model’s year, design motivation, deep neural networks, and auxiliary data.

Methods:

This review introduced systematic literature on automatic ICD coding methods based on deep learning. We screened 5 online databases, including Web of Science, SpringerLink, PubMed, ACM, and IEEE digital library, and collected 53 published articles related to deep learning-based ICD coding from 2017 to 2023.

Results:

These deep neural network methods aimed to overcome some challenges, such as lengthy and noisy clinical text, high dimensionality and functional relationships of medical codes, and long-tail label distribution. The Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), attention mechanisms, Transformers, Pre-trained Language Models (PLMs), etc, have become popular to address prominent issues in ICD coding. Meanwhile, introducing medical ontology within the ICD coding system (code description and code hierarchy) and external knowledge (Wikipedia articles, tabular data, Clinical Classification Software (CCS), fine-tuning PLMs based on biomedical corpus, entity recognition and concept extraction) has become an emerging trend for automatic ICD coding.

Conclusion:

This paper provided a comprehensive review of recent literature on applying deep learning technology to improve medical code assignment from a unique perspective. Multiple neural network methods (CNNs, RNNs, Transformers, PLMs, especially attention mechanisms) have been successfully applied in ICD tasks and achieved excellent performance. Various medical auxiliary data has also proven valuable in enhancing model feature representation and classification performance. Our in-depth and systematic analysis suggested that the automatic ICD coding method based on deep learning has a bright future in healthcare. Finally, we discussed some major challenges and outlined future development directions.