ClinicSum: Utilizing Language Models for Generating Clinical Summaries from Patient-Doctor Conversations

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Clinical summarization of patient-doctor conversations is crucial for efficient healthcare documentation.
• Existing methods often struggle with accuracy and completeness in capturing essential clinical information.
• Automating this process can alleviate clinician burden and improve data quality.

Purpose of the Study:

• To introduce ClinicSum, a novel framework for automated clinical summary generation.
• To leverage a two-module architecture combining retrieval and Pre-trained Language Models (PLMs) for enhanced summarization.
• To evaluate ClinicSum's performance against state-of-the-art methods using both automated metrics and expert assessments.

Main Methods:

• Developed a retrieval-based module to extract Subjective, Objective, Assessment, and Plan (SOAP) information.
• Employed fine-tuned PLMs in an inference module to generate abstracted clinical summaries from SOAP data.
• Created a training dataset of 1,473 conversation-summary pairs from public datasets (FigShare, MTS-Dialog) with Subject Matter Expert (SME) validation.

Main Results:

• ClinicSum demonstrated superior performance compared to state-of-the-art PLMs.
• Achieved higher precision, recall, and F-1 scores in automatic evaluations (ROUGE, BERTScore).
• Received high preference ratings from SMEs in human assessments, indicating clinical utility.

Conclusions:

• ClinicSum offers a robust and effective solution for automated clinical summarization.
• The framework shows significant potential to improve the efficiency and accuracy of clinical documentation.
• Further development could integrate ClinicSum into clinical workflows to support healthcare professionals.