Retrieve and rerank for automated ICD coding via Contrastive Learning.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Computational Linguistics

Background:

• Automated ICD coding is crucial for healthcare but faces challenges with large, imbalanced label sets in deep learning.
• Existing methods struggle with the complexity of multi-label prediction for disease codes.

Purpose of the Study:

• To develop a novel retrieve and rerank framework to enhance automated ICD coding accuracy.
• To address the limitations of large label sets and data imbalance in deep learning models for this task.

Main Methods:

• Proposed a retrieve and rerank framework incorporating Contrastive Learning (CL) for label retrieval.
• Utilized CL as a training strategy to simplify the label space and capture code co-occurrence.
• Employed a Transformer-based model for refining and reranking candidate ICD codes from clinical notes.

Main Results:

• The framework demonstrated improved accuracy by preselecting a smaller subset of relevant ICD codes.
• Achieved high performance metrics: 0.590 Micro-F1 and 0.990 Micro-AUC on the MIMIC-III benchmark.
• The CL-based retriever implicitly captured code co-occurrence, overcoming cross-entropy's independent label assignment limitations.

Conclusions:

• The retrieve and rerank framework with Contrastive Learning significantly enhances automated ICD coding.
• This approach offers a more accurate and efficient method for assigning disease codes from clinical text.
• The model shows strong potential for real-world clinical applications requiring precise diagnostic coding.