A patient safety knowledge graph supporting vaccine product development.

Area of Science:

• Pharmacovigilance and Drug Safety
• Data Science and Knowledge Representation
• Public Health Surveillance

Background:

• COVID-19 vaccine adverse event data presented challenges due to varied representations and unprecedented volume.
• A robust data store was needed to accommodate and harmonize this large-scale surveillance data.
• The Patient Safety Knowledge Graph (PSKG) was developed to address these challenges.

Purpose of the Study:

• To detail the development and implementation of a post-marketing safety system using public data sources for COVID-19 vaccine surveillance.
• To demonstrate the utility of a knowledge graph in managing and analyzing complex, high-volume adverse event data.
• To support timely responses to health authority inquiries regarding vaccine safety.

Main Methods:

• Designed a semantic model to represent key safety concepts and integrated public data sources.
• Developed an extract-transform-load (ETL) pipeline for data parsing, alignment (e.g., vaccine names), and quality control.
• Utilized the Medical Dictionary for Regulatory Activities (MedDRA) and deployed the PSKG in a Neo4J graph database with API access.

Main Results:

• Ingested and aligned adverse event and vaccine exposure data from 250 countries, creating a graph with over 4.3 million nodes and 30.5 million edges.
• Enabled ad-hoc analyses and periodic reporting for widely used COVID-19 vaccines.
• Analysis code utilizing the PSKG was 80% shorter and over 200 times faster than traditional Python implementations.

Conclusions:

• Organizing safety data in a knowledge graph simplifies analysis by centralizing data parsing and transformation.
• The PSKG transformed the approach to answering scientific and medical questions, allowing for rapid iteration and response to evolving requests.
• This system enhances the ability to quickly assess vaccine safety profiles based on specific adverse events and geographical data.