Generating More Faithful and Consistent SOAP Notes Using Attribute-Specific Parameters

Abstract

The widespread adoption of SOAP notes for documenting diverse aspects of patient information in healthcare has been prevalent. However, the conventional process of manual note-taking is laborious and can distract healthcare providers from addressing patients’ needs. Prior work by Krishna et al. (2021a) has introduced an end-to-end pipeline for generating SOAP notes, but model-generated notes are susceptible to inaccuracies, irrelevant and missing information. In this work, we assess the performance of large language models (GPT-3.5) for SOAP note generation, compare them with fine-tuned models using automated metrics, and propose a solution to improve the consistency and faithfulness of notes by incorporating attribute-specific information via SOAP section information. To achieve this, we integrate an extra layer of unique section-specific cross-attention parameters to existing encoder-decoder architectures. Our approach is evaluated using a comprehensive suite of automated metrics and expert human evaluators, demonstrating that it leads to more accurate, relevant, and faithful information.