Knowledge Graphs

What are the core components of a knowledge graph?

A knowledge graph is a directed label graph representing relationships between different entities. It has three key components:

• Nodes — the entities
• Edges — the relationships between entities
• Labels — the meaning of those relationships

Combined, these three components form what's known as a triple, typically represented as:

• A (the original entity, the first node) = the subject
• B (the link, or edge) = the predicate
• C (the target entity, the second node) = the object

This relatively simple structure can link together vast amounts of heterogeneous information — such as test results, genomic data, air quality data, and previous prescriptions.

Why are knowledge graphs useful?

This way of structuring data allows researchers to infer, or make an informed guess about, new and previously unknown information. For example, if the fact "Prince Harry is brother of Prince William" was added to the family tree above, it would be possible to infer that Princess Diana was also the mother of Prince Harry — even if that specific fact hadn't previously been recorded.

These links can simply make known information more discoverable (e.g., the connection between a set of symptoms and a specific disease), or they can enable inferences about entirely new information (e.g., whether a drug used to treat Disease A might also treat Disease B). In this way, knowledge graphs enable a deeper understanding of a particular domain.

This inference capability is what makes knowledge graphs popular in widely used technologies:

• Google uses a knowledge graph to power its search functionality
• Facebook uses one to identify connections between people
• Netflix uses one to recommend films or TV shows to viewers

How are knowledge graphs used in medicine?

Knowledge graphs are used extensively in healthcare, including for:

Diagnostics — representing links between symptoms and diseases, helping technologies like chatbots provide decision support to clinicians.

Drug repurposing — showing similarities between diseases, so researchers can assess whether a drug designed to treat Disease A might also treat Disease B.

Personalised treatment planning — showing which drugs are most likely to be effective for a specific patient, based on how similar they are to other patients who've been treated with that drug.

Treatment recommendation decision support — a knowledge graph where nodes represent medicines, diseases, and patients, and edges represent the interactions between them, can recommend the best treatment for a specific patient.

Identifying misinformation — comparing information found online (e.g., claims about a drug's effectiveness) against the authoritative information contained in the knowledge graph, to help identify and suppress misinformation.

Drug discovery and drug repurposing — a knowledge graph made up of entities (e.g., a drug's ingredients) and their relationships (e.g., the interaction between Drug A and Drug B) can predict dangerous drug-to-drug interactions, flag potential repurposing opportunities, or help identify entirely new treatments.

How are knowledge graphs created?

Knowledge graphs can be built manually, following a series of steps:

1. Identify the scope and objectives of the knowledge graph and develop the data schema
2. Gather data from different sources — medical literature, clinical trial results, and real-world data such as electronic health records (EHRs)
3. Extract and transform the data into a structured format by identifying entities and defining relationships between them
4. Map the entities and relationships to the chosen data schema
5. Use prediction models to infer missing information, such as missing relationships
6. Validate the knowledge graph to ensure accuracy
7. Regularly update the knowledge graph — incorporating new data and knowledge, refining the schema, and evaluating quality over time

This manual process is time-consuming and not well suited to the exponential growth of medical information seen today.

Is AI changing how knowledge graphs are built and used?

Yes. Increasingly, AI techniques — including natural language processing (NLP) — are used to automate the creation of knowledge graphs, independently "learning" (or figuring out) connections between pieces of information within a given dataset. Historically, this was a manual process carried out by human experts such as doctors.

AI's relationship with knowledge graphs also runs the other way: existing knowledge graphs can act as input for AI models, serving as training material and reducing the need to create large labelled datasets from scratch. They can also help make machine learning models more explainable, by representing the underlying knowledge — and therefore the model's decision-making process — in a structured, relatively transparent format.

‍