1 / 7

Medical AI Research

MedGraph

RAG

A Knowledge-Graph Grounded
Medical RAG System

From unstructured text to clinically meaningful, graph-grounded answers. Building interpretable, auditable medical QA.

Presented By

Ashwin • Aviral • Dally

Institution

Newton School of Technology

2 / 7

Problem Statement

Unstructured Data Challenges

MedGraph RAG // 02

Medical literature and clinical records contain vast amounts of valuable information, but relying on traditional LLMs and unstructured retrieval creates critical safety and accuracy gaps.

Limited Multi-hop Reasoning

Simple vector retrieval fails to connect disparate facts scattered across different documents.

Entity Aliasing & Ambiguity

Inconsistent terminology breaks retrieval consistency without a unified knowledge schema.

"Black Box" Hallucinations

Standard LLMs generate plausible but unverified answers. Medical decisions require traceable evidence.

3 / 7

Solution Overview

Hybrid KG + RAG Pipeline

MedGraph RAG // 03

Our solution integrates Named Entity Recognition (NER) with Knowledge Graphs to structure medical facts before retrieval, enabling precise, auditable answers grounded in explicit relationships.

1

Input Text

Unstructured clinical notes

2

NER Extraction

spaCy + BERT + Gemini

3

Knowledge Graph

Structured Triplets

4

KG Retrieval

Traverse graph relations

5

Grounded Answer

LLM generates response

4 / 7

Named Entity Recognition

Extracting Medical Entities

MedGraph RAG // 04

Entity Types

MEDICINE

Pharmaceutical names

Examples: Remdesivir, Ibuprofen, Aspirin

PATHOGEN

Bacteria, viruses

Examples: SARS-CoV-2, E. coli, Streptococcus

MEDICAL_CONDITION

Diseases, disorders

Examples: COVID-19, Pneumonia, Diabetes

Model Performance

75.5%

F1 Score

83.3%

Precision

69.0%

Recall

Model Architecture

• Token-to-Vector: MultiHashEmbed

• Transition-Based Parser (Hidden: 64)

• Training: 50 epochs, batch size 100

Annotation Strategy

Hybrid approach: automated NER combined with Gemini API for high-quality relationship extraction.

5 / 7

Knowledge Graph Construction

From Entities to Structured Knowledge

MedGraph RAG // 05

Pipeline Overview

1

Entity Extraction

Use spaCy NER or Gemini API

2

Relationship Extraction

Identify TREATS and CAUSES

3

Triplet Formation

Create (Subject) --[Relationship]--> (Object)

4

Graph Building

Construct graph using NetworkX

5

Visualization

Visualize with Matplotlib

Example Triplets

                                (Remdesivir) --TREATS--> (COVID-19)

                                (SARS-CoV-2) --CAUSES--> (COVID-19)

6 / 7

KG-RAG Architecture

Chain of Exploration (CoE) Framework

MedGraph RAG // 06

Process Flow

1

Exploration Plan

Break queries into sequential steps

2

LLM Evaluation

Refine approach for each step

3

KG Lookup

Cypher queries and vector search

4

RAG Answer

Generate accurate answers

Key Advantages

• Structured reasoning through knowledge graphs

• Multi-hop query support

• Transparent, traceable reasoning paths

• Superior accuracy on complex queries

7 / 7

Results & Future Directions

Achievements and Next Steps

MedGraph RAG // 07

Key Achievements

NER Model

75.5% F1 score for medical entity extraction with 83.3% precision, successfully identifying MEDICINE, MEDICAL_CONDITION, and PATHOGEN entities.

Knowledge Graph

Comprehensive medical KG with TREATS and CAUSES relationships, enabling structured querying and multi-hop reasoning.

KG-RAG System

Chain of Exploration framework for complex multi-hop reasoning, outperforming vanilla LLMs and standard RAG systems.

Future Work

• Expand entity types and relationships

• Improve NER precision and recall

• Scale to larger medical corpora

• Integrate temporal and spatial data

Performance Comparison

Approach	Complex	Multi-hop	Accurate
LLM Only	✗	✗	✗
LLM + RAG	✗	✗	✗
LLM + KG-RAG	✓	✓	✓

Thank You!

MedGraph RAG

Team Members

Ashwin Sathyamoorthy Aviral Singh Dally R

View on GitHub