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

Area of Science:

• Computer Science
• Medical Informatics
• Artificial Intelligence

Background:

• Automatic International Classification of Diseases (ICD) coding is crucial for healthcare data management, but faces challenges with complex Electronic Health Records (EHRs).
• Manual coding is inefficient and error-prone, while traditional machine learning methods struggle with clinical text semantics.
• Deep Learning (DL) shows promise in addressing these limitations for accurate medical code assignment.

Purpose of the Study:

• To comprehensively review recent advancements in deep learning for automatic ICD coding.
• To identify prominent challenges and emerging development trends in DL-based ICD coding.
• To analyze models based on their year, design, neural networks, and auxiliary data.

Main Methods:

• Systematic literature review of deep learning-based automatic ICD coding methods.
• Screening of 5 major online databases (Web of Science, SpringerLink, PubMed, ACM, IEEE).
• Collection and analysis of 53 relevant articles published between 2017 and 2023.

Main Results:

• Deep neural networks like CNNs, RNNs, attention mechanisms, Transformers, and PLMs effectively handle lengthy, noisy clinical text and complex code relationships.
• Emerging trends include integrating medical ontologies (code descriptions, hierarchies) and external knowledge (Wikipedia, CCS) for improved coding.
• These methods address challenges like high dimensionality and long-tail label distribution in medical coding.

Conclusions:

• Deep learning, particularly neural networks and attention mechanisms, has demonstrated success in automatic ICD coding tasks.
• Incorporating auxiliary medical data enhances model performance and feature representation.
• Deep learning-based automatic ICD coding holds significant future potential in healthcare, with ongoing challenges and future directions identified.