Conference proceeding
Dynamic Retrieval Strategy for Summarizing Doctor-Patient Dialogues with RAG
IEEE International Conference on Big Data, pp 8070-8079
08 Dec 2025
Abstract
Retrieval-augmented generation (RAG) has emerged as a promising solution to overcome the limitations of large language models (LLMs), particularly their reliance on training data with a fixed knowledge cut-off and the possibility of generating hallucinated or outdated content. By integrating real-time information retrieval with generative capabilities, RAG enables more accurate and contextually relevant outputs. While LLMs have attracted growing interest for various biomedical summarization tasks, the potential advantages of RAG for this domain remain underexplored. In this study, we investigate the effectiveness of RAG-based methods for generating "assessment and plan" summaries from doctor-patient dialogues. Specifically, we compare three RAG strategies: naive RAG, advanced RAG with post-retrieval filtering, and a novel approach, dynamic RAG, which adaptively determines the optimal number of retrieved chunks based on content novelty. Our results show that dynamic RAG can generate high-quality summaries using only a subset of the most relevant content, effectively balancing efficiency and completeness and offering a promising direction for scalable summarization systems.
Metrics
1 Record Views
Details
- Title
- Dynamic Retrieval Strategy for Summarizing Doctor-Patient Dialogues with RAG
- Creators
- Yiwen Shi - Drexel UniversityXiaohua Hu - Drexel University
- Publication Details
- IEEE International Conference on Big Data, pp 8070-8079
- Publisher
- IEEE
- Resource Type
- Conference proceeding
- Language
- English
- Academic Unit
- Information Science
- Other Identifier
- 991022166454104721