Knowledge Injected Prompt Based Fine-tuning for Multi-label Few-shot ICD Coding.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Automatic International Classification of Diseases (ICD) coding is complex due to high dimensionality and the long-tail distribution of disease codes.
• The long-tail challenge, where rare diseases are infrequently coded, hinders accurate automated medical note classification.

Purpose of the Study:

• To address the long-tail challenge in automatic ICD coding.
• To improve the accuracy of assigning multiple ICD codes to medical notes, particularly for rare diseases.

Main Methods:

• Adapted prompt-based fine-tuning with label semantics for few-shot learning.
• Developed a knowledge-enhanced Longformer model incorporating hierarchy, synonym, and abbreviation knowledge.
• Utilized contrastive learning for additional pretraining on domain-specific knowledge.

Main Results:

• Outperformed state-of-the-art methods on the MIMIC-III-full dataset, achieving a 14.5% increase in macro F1 score (from 10.3 to 11.8).
• On the newly created MIMIC-III-rare50 dataset for rare diseases, the model improved macro F1 from 17.1 to 30.4 and micro F1 from 17.2 to 32.6.

Conclusions:

• The proposed knowledge-enhanced Longformer effectively addresses the long-tail challenge in automatic ICD coding.
• The model demonstrates significant improvements in both general and few-shot rare disease coding tasks.