Multi-label Few-shot ICD Coding as Autoregressive Generation with Prompt.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Automatic International Classification of Diseases (ICD) coding is crucial for medical record analysis but faces challenges with high-dimensional multi-label assignment and the long-tail problem of infrequent codes.
• Existing methods struggle to effectively assign rare but clinically significant ICD codes, impacting the completeness and accuracy of medical coding.

Purpose of the Study:

• To address the long-tail challenge in automatic ICD coding by reframing it as an autoregressive generation task.
• To develop a novel model that generates text descriptions to infer ICD codes, improving accuracy for both common and rare codes.

Main Methods:

• Introduced a novel pretraining objective to generate diagnoses and procedures using the physician-used SOAP note structure.
• Developed a model that generates lower-dimensional text descriptions instead of directly predicting high-dimensional ICD codes.
• Designed a novel prompt template for multi-label classification and an ensemble learner with a cross-attention reranker.

Main Results:

• The Generation with Prompt (GP_soap) model achieved a macro F1 score of 30.2 on the MIMIC-III-few benchmark, outperforming previous state-of-the-art (SOTA) models.
• The ensemble learner improved macro F1 from 10.4 to 14.6 and micro F1 from 58.2 to 59.1 on the MIMIC-III-full benchmark.

Conclusions:

• Transforming ICD coding into an autoregressive generation task effectively addresses the long-tail challenge.
• The proposed GP_soap model and ensemble learner offer substantial improvements in automatic ICD coding accuracy, particularly for infrequent codes.