Recent Advances in AI for Automated ICD Coding: A Systematic Literature Review

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Documentation Improvement

Background:

• Manual International Classification of Diseases (ICD) coding is inefficient, costly, and prone to errors, hindering healthcare analytics and workflow.
• Automated ICD code assignment using Artificial Intelligence (AI) presents a potential solution to these challenges.
• This systematic review examines AI models for automated ICD coding from various clinical documents.

Purpose of the Study:

• To systematically review the current research landscape of AI-driven automated ICD coding.
• To analyze the methodologies, datasets, and performance of existing AI models.
• To identify critical gaps and propose a future research agenda for AI in ICD coding.

Main Methods:

• Systematic literature search following PRISMA guidelines across six databases (2019-2024).
• Selection of 54 relevant studies from 4,280 initial citations.
• Analysis of datasets, preprocessing, feature extraction, and AI model architectures (ML, DL, CNN, RNN, Transformers, Hybrid).

Main Results:

• Diverse datasets and preprocessing techniques were employed, with a shift towards deep learning models.
• Model performance varied, with higher accuracy on frequent codes; challenges remain for rare codes and generalizability.
• Significant gaps identified include dataset limitations, lack of interpretability, and inconsistent evaluation protocols.

Conclusions:

• AI shows significant potential to revolutionize ICD coding, enhancing efficiency and accuracy.
• Addressing issues of generalizability, rare code prediction, model interpretability, and standardization is crucial for clinical adoption.
• A proposed 5P research agenda (Population Diversity, Performance Robustness, Prediction of Rare Codes, Provenance Transparency, Practical Integration) guides future development.