Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DoctorRAG: Medical RAG Fusing Knowledge with Patient Analogy through Textual Gradients

Yuxing Lu, Gecheng Fu, Wei Wu, Xukai Zhao, Sin Yee Goi, Jinzhuo Wang

TL;DR

The paper tackles the high-stakes challenge of medical question answering by bridging explicit clinical knowledge with experiential, case-based reasoning. It introduces DoctorRAG, a retrieval-augmented generation framework that uses dual sources (knowledge bases and patient case repositories) together with declarative sentence transformations and ICD-10 tagging to improve retrieval precision. A novel multi-agent Med-TextGrad procedure iteratively refines outputs via textual gradients grounded in context and patient criteria, aiming for higher accuracy, relevance, and faithfulness. Extensive multilingual multitask experiments show DoctorRAG outperforms strong RAG baselines and demonstrates robust doctor-like reasoning across languages and tasks, marking a step toward clinically informed AI systems.

Abstract

Existing medical RAG systems mainly leverage knowledge from medical knowledge bases, neglecting the crucial role of experiential knowledge derived from similar patient cases -- a key component of human clinical reasoning. To bridge this gap, we propose DoctorRAG, a RAG framework that emulates doctor-like reasoning by integrating both explicit clinical knowledge and implicit case-based experience. DoctorRAG enhances retrieval precision by first allocating conceptual tags for queries and knowledge sources, together with a hybrid retrieval mechanism from both relevant knowledge and patient. In addition, a Med-TextGrad module using multi-agent textual gradients is integrated to ensure that the final output adheres to the retrieved knowledge and patient query. Comprehensive experiments on multilingual, multitask datasets demonstrate that DoctorRAG significantly outperforms strong baseline RAG models and gains improvements from iterative refinements. Our approach generates more accurate, relevant, and comprehensive responses, taking a step towards more doctor-like medical reasoning systems.

DoctorRAG: Medical RAG Fusing Knowledge with Patient Analogy through Textual Gradients

TL;DR

The paper tackles the high-stakes challenge of medical question answering by bridging explicit clinical knowledge with experiential, case-based reasoning. It introduces DoctorRAG, a retrieval-augmented generation framework that uses dual sources (knowledge bases and patient case repositories) together with declarative sentence transformations and ICD-10 tagging to improve retrieval precision. A novel multi-agent Med-TextGrad procedure iteratively refines outputs via textual gradients grounded in context and patient criteria, aiming for higher accuracy, relevance, and faithfulness. Extensive multilingual multitask experiments show DoctorRAG outperforms strong RAG baselines and demonstrates robust doctor-like reasoning across languages and tasks, marking a step toward clinically informed AI systems.

Abstract

Existing medical RAG systems mainly leverage knowledge from medical knowledge bases, neglecting the crucial role of experiential knowledge derived from similar patient cases -- a key component of human clinical reasoning. To bridge this gap, we propose DoctorRAG, a RAG framework that emulates doctor-like reasoning by integrating both explicit clinical knowledge and implicit case-based experience. DoctorRAG enhances retrieval precision by first allocating conceptual tags for queries and knowledge sources, together with a hybrid retrieval mechanism from both relevant knowledge and patient. In addition, a Med-TextGrad module using multi-agent textual gradients is integrated to ensure that the final output adheres to the retrieved knowledge and patient query. Comprehensive experiments on multilingual, multitask datasets demonstrate that DoctorRAG significantly outperforms strong baseline RAG models and gains improvements from iterative refinements. Our approach generates more accurate, relevant, and comprehensive responses, taking a step towards more doctor-like medical reasoning systems.

Paper Structure

This paper contains 41 sections, 15 equations, 6 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. Expertise-Experience retrieval and aggregation
  4. Iterative answer optimization with multi-agent Med-TextGrad
  5. Computation graph with gradient backpropagation
  6. Textual gradient computation and backpropagation
  7. Iterative optimization with context and patient optimizers
  8. Experimental Setup
  9. Datasets
  10. LLM backbones
  11. Results
  12. DoctorRAG vs other RAG (RQ.1)
  13. Utility of patient case information (RQ.2)
  14. Impact of Med-TextGrad's iterative refinement (RQ.3)
  15. Different LLM backbone (RQ.4)
  16. ...and 26 more sections

Figures (6)

  • Figure 1: General and Medical RAGs retrieve solely from knowledge bases, whereas DoctorRAG enhance responses by integrating both knowledge-based expertise and case-based experience, mirroring clinical practice.
  • Figure 2: Overview of the proposed DoctorRAG & Med-TextGrad framework. Medical knowledge is transformed into declarative statements and tagged to match the patient's query. Dual-database retrieval generates a response incorporating both clinical expertise and experience. A two-way multi-agent Med-TextGrad optimization process refines the answer through an iterative computation graph involving textual gradient backpropagation.
  • Figure 3: UMAP visualizations of patient embeddings on four distinct medical datasets. Each point corresponds to a patient embedding, and different colors represent different diagnosed diseases.
  • Figure 3: Ablation study of DoctorRAG.
  • Figure 4: Pairwise comparison scores for ground truth (GT), original answer (OA), refined answers of iteration 1-3 (T1-T3) on (a) Comprehensiveness, (b) Relevance, (c) Safety, and (d) Overall. The lower-left triangle of each matrix represents the Y-axis outputs perform better than X-axis output.
  • ...and 1 more figures