ICD coding of death certificates with generative language models.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Medical Informatics

Background:

• Large language models (LLMs) excel at text generation but are underutilized for text classification tasks.
• International Classification of Diseases (ICD) coding of clinical documents is a critical but challenging classification task.

Purpose of the Study:

• To adapt a LLaMA generative language model for the ICD coding task.
• To evaluate the performance of generative LLMs against traditional encoder models for ICD coding.

Main Methods:

• Utilized a LLaMA generative model trained with a language modeling objective.
• Employed constrained decoding at inference time instead of discriminative fine-tuning.
• Applied the model to Portuguese death certificates for underlying cause of death coding.
• Compared performance against a BERT encoder model.

Main Results:

• Generative language models achieved strong results in ICD coding of death certificates.
• Classification accuracy was comparable to that of established encoder models.
• Demonstrated the suitability of language generation for ICD coding.

Conclusions:

• Language generation models offer a viable and effective approach for ICD coding.
• A single unified generative model can handle multiple related coding tasks, such as underlying and multiple cause coding.