AnEMIC: A Framework for Benchmarking ICD Coding Models.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Computational Health

Background:

• Manual diagnostic coding (ICD coding) using clinical notes is laborious and prone to errors.
• Existing automatic ICD coding research lacks standardized benchmarking frameworks.
• The International Classification of Diseases (ICD) standard is crucial for healthcare data management.

Purpose of the Study:

• To introduce AnEMIC, an open-source, user-friendly tool for automatic ICD coding.
• To establish a standardized framework for benchmarking ICD coding models.
• To improve the accuracy and efficiency of assigning diagnosis codes from clinical notes.

Main Methods:

• Developed a streamlined pipeline for data preprocessing, model training, and evaluation in automatic ICD coding.
• Corrected preprocessing errors present in existing works.
• Integrated explainable AI for model analysis and an interactive demo for real-time inference.

Main Results:

• Provided corrected datasets and pre-trained models with weights for automatic ICD coding.
• Demonstrated a streamlined workflow for preprocessing, training, and evaluation.
• Enabled real-time inference and model interpretability through visualizations.

Conclusions:

• AnEMIC offers a standardized framework to advance research in automatic ICD coding.
• The tool aims to assist healthcare professionals in leveraging automated coding solutions.
• Open-sourcing AnEMIC facilitates reproducible research and exploration of ICD coding technologies.