1 of 9

GraphRAG for Exploring

Heterogeneous�Medical Knowledge

​

Giuseppe Futia, PhD

​

1

2 of 9

Building

Ontology Integration, Information Extraction, Data Enrichment, Virtualized Access

Our journey

Retrieving

GraphRAG Agent: from General Medical Knowledge to Virtualized Patient Details

​

Understanding

Medical Domain Challenges, Data Integration Strategies, Technological Solutions

​

https://github.com/giuseppefutia/cdl2025

2

3 of 9

Medical Domain - Data Integration Challenges

Diverse and disconnected data sources, each capturing different aspects of patient health and clinical practice.

​

Their differing formats, structures, and purposes make data integration and analysis both challenging and essential.

​

​

Heterogeneous medical data:

  • Structured information: EHR data capturing diagnoses, lab results, medications, vital signs, and procedures
  • Unstructured contents: Research publications, reports containing diverse medical information
  • Semantic data sources: Biomedical vocabularies and ontologies like ICD-10, SNOMED, and UMLS enabling standardized concepts

3

4 of 9

Medical Domain - Technological Challenges

Strict privacy requirements prevent the use of external services or third-party models, demanding fully self-contained infrastructures.

​

Sensitive medical data must remain on-premise, and systems must ensure secure access without duplicating information across storage layers.

​

Key features :

  • Privacy-centric architecture: Data must be processed locally, while avoiding exposure to external systems
  • Data virtualization: Instead of replicating sensitive data across pipelines, virtualized access layers reduce redundancy and minimizing privacy risks
  • Secure environments: Isolated execution environments and allow analytical workflows while keeping raw data protected and contained

4

5 of 9

Building - Knowledge Graph-based Architecture

Knowledge Graph

Ontology Integration

Information Extraction

Structured

Data Enrichment

Biomedical ontologies

Publications and Reports

Clinical Data

LLM

LLM

LLM

On premises

LLM

Chat

Embedding

Virtualization

5

6 of 9

Ontology Mapping and Information Extraction - A General Framework

Annotated

Data with Entities

Input �Data

Information Background

Multiple

Candidates

for Each Entity

Disambiguated

Entities

for each Entity

Candidate

Selection

Entity

Recognition

(Not always required)

Candidate Disambiguation

6

7 of 9

Retrieving - Knowledge Graph-based Architecture

Knowledge Graph

On premises

LLM

Chat

Embedding

Clinical Data

Direct Access

Virtualized

Access

GraphRAG

Agent

7

8 of 9

Retrieving - Knowledge Graph-based Architecture

Exploring Ontology-based Knowledge

Query Generation

GraphRAG

Agent

EXPLAIN Test

Relation Correction

Query Diagnostics

Query Execution

Runtime Data Materialization

Answer Generation

Exploring Virtualized Patient Data

Virtualized Data Analysis

Ontology Knowledge Analysis

Ontology-

based Answer on Patient Data

Analyzing Patient Date Through Ontologies

8

9 of 9

Thank You!

Discount code (45% off all Manning products) for the Connected Data London 2025 conference: connectdl25

9