Medical Knowledge Graph Aggregation with Privacy-Preserving Federated Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The challenge in creating medical knowledge graphs lies in the acquisition of data sources due to privacy concerns and commercial competition, leading to isolated data silos that hinder centralized model training.

Innovation Solution

A method involving federated learning is employed to create and aggregate base knowledge graph models from different sites, sharing only model parameters while keeping raw data secure within each site, using a central aggregator to form a comprehensive central knowledge graph model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If data sources are centralized for knowledge graph training, then model robustness and granularity are improved, but data privacy and security are compromised

Engineering Contradiction:
Improvemodel robustnessVSAvoiddata privacy risk
Core Design Contradiction:
ReliabilityVSObject-affected harmful factors

Solution Approach 1:

The patent segments the centralized training process into distributed local training at multiple sites. Each site trains its own base knowledge graph model locally using its private data, preventing data from leaving the premises while still contributing to the overall model development. This segmentation resolves the contradiction by maintaining data privacy at each location while achieving model robustness through aggregation of multiple trained models.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a central aggregator as an intermediary that receives and combines trained base knowledge graph models from multiple sites without directly accessing the raw data. The aggregator acts as a mediator that facilitates knowledge graph generation from distributed models while preserving data boundaries, thus enabling model robustness without compromising data privacy.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Object-affected harmful factors

If data is kept isolated at different sites, then data security is maintained, but model training quality deteriorates due to limited data diversity

Engineering Contradiction:
Improvedata securityVSAvoidmodel training quality
Core Design Contradiction:
Object-affected harmful factorsVSReliability

Solution Approach 1:

The patent creates copies of the base knowledge graph model structure at each site, which are then trained locally with site-specific data. These model copies serve as vehicles for transferring learning capabilities across sites without transferring the actual sensitive data. The aggregation of these model copies produces a comprehensive knowledge graph that benefits from data diversity while maintaining security.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The central aggregator serves as an intermediary that synthesizes knowledge from multiple isolated trained models. It combines the learned patterns and relationships from each site's private data without requiring direct access to the raw data, thus maintaining data security while achieving high model training quality through diverse knowledge integration.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Loss of information

If base knowledge graph models are aggregated from multiple sites, then knowledge graph comprehensiveness is improved, but system complexity increases

Engineering Contradiction:
Improveknowledge completenessVSAvoidsystem complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The patent segments the knowledge graph construction into independent base model training phases at each site followed by aggregation. This segmentation allows each site to independently develop its model without coordination complexity, while the aggregator handles the integration. The segmented approach achieves comprehensive knowledge aggregation while minimizing system complexity through clear separation of concerns.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The central aggregator acts as a specialized intermediary that handles the complexity of model aggregation. It receives standardized base knowledge graph models from multiple sites and combines them into a comprehensive knowledge graph, isolating the aggregation complexity within the aggregator module while keeping individual site systems simple and uniform.

Inventive Principle:
Principle #24Intermediary (Mediator)

4Object-affected harmful factors

If federated learning is used to share only model parameters, then data privacy is maintained, but training efficiency decreases due to iterative communication

Engineering Contradiction:
Improvedata privacy protectionVSAvoidtraining efficiency
Core Design Contradiction:
Object-affected harmful factorsVSProductivity

Solution Approach 1:

The patent performs preliminary local training of base knowledge graph models at each site before aggregation, so that the majority of learning occurs locally without iterative communication. This preliminary action reduces the need for repeated parameter exchanges, maintaining data privacy while improving training efficiency by minimizing communication overhead.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent creates and shares copies of the base knowledge graph model structures rather than iterating on parameters. Each site receives a model copy, trains it locally with its data, and returns the trained copy. This copy-based approach maintains privacy while improving efficiency by eliminating iterative parameter tuning and reducing communication rounds.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP4583009A1Computer implemented method for generating a knowledge graph model in a medical environment
Publication Date: 2025.07.09 KONINKLIJKE PHILIPS NV
  • EP4583009A1 patent drawingFigure 1
  • EP4583009A1 patent drawingFigure 2
  • EP4583009A1 patent drawingFigure 3

AI summary

Computer implemented method for generating a central knowledge graph model (10) in an medical environment, comprising the steps of: providing a first base knowledge graph model and a second base knowledge graph model (S10); training the first base knowledge graph model with a first data set in a first site (15) of the medical environment (S20); training the second base knowledge graph model with a second data set in a second site (16) of the medical environment (S30); transferring the trained first base knowledge graph model (13) and the trained second base knowledge graph model (14) to a central aggregator (11) (S40); aggregating by the central aggregator (11) the trained first base knowledge graph model (13) and the trained second knowledge graph model (14) to a central knowledge graph model (10) (S50); providing the central knowledge graph model (10) for further use (S60).