Data Challenges of building an AI doctor (5/5): Single Medical Language

This blog is part 5 of a 5 part series. Explore the full series: Part 1  Part 2  Part 3  Part 4  Part 5

Challenge #5: How do we make sure that data generated by any Babylon service can be understood by everyone who needs to understand it?

This is part five of a five part blog series exploring some of the challenges we’ve been facing while building a personal AI doctor.

Many of us know the famous myth of the Tower of Babel in ancient Babylon.

This myth illustrates the power of using a single, common language over many incongruent languages.

The diverse population of services within Babylon can learn from this, as they need to share data with each other, with external services and with us humans.

Problem: How do we make sure that data generated by any Babylon service can be understood by everyone who needs to understand it?

Let’s think about what it takes to create a language (whether it’s a spoken language, programming language or any other!)

We need:

1. Vocabulary - unique concepts that represent the basic building blocks of the language
2. Ability to create more complex meaning from the basic building blocks
3. Ability to be understood by everyone who the language is intended for

1. Vocabulary:

We want to provide accessible and affordable healthcare to everyone in the world - so we must be prepared to deal with every possible disease, symptom or treatment under the sun. That’s why our vocabulary must be broad, rich and comprehensive.

We’ve created our own encyclopedia of unique medical concepts - the Babylon Knowledge Base - that’s continuously evolving towards a complete medical brain.

It includes:

• A vast set of concepts. Includes diseases, drugs, symptoms, risk factors, medical procedures, the human anatomy, and much more
• Synonyms (including layman’s terms)
• Different languages and country-specific codes

Earlier we mentioned the enduring problem of interoperability between disparate medical coding systems. The Babylon Knowledge Base is attempting to solve this for good by being the first in the world to unify pretty much all modern medical standards and vocabularies.

LESSON 13: Build up a versatile vocabulary of unique concepts that is broad and rich enough to support the needs of the mission.

2. Creating Complex Meaning:

Even with such comprehensive vocabulary, it’s still impossible to pre-code every medical entity that could ever be encountered.

A patient may say that they have a “pain around my heart that started gradually”.

None of the existing coding systems has a concept capable of representing this.

However, they CAN represent parts of it.

So by understanding the relationships between the component parts, we can start to connect them in a meaningful way.

Diagram 9: Taking the example phrase “pain around my heart that started gradually”, we use algorithms to map the relationships between the constituent concepts and represent the complex concept in graph format

Using graph format is the best way to do this - concepts are represented by nodes, and relationships between concepts are represented by edges.

Once we determine that a new complex concept is valuable and used frequently enough, we absorb it into our ever-growing Babylon Knowledge Base so that we can reference it more easily and quickly in the future.

LESSON 14: Map out the relationships between basic concepts to create complex concepts.

3. Being Understood:

The Babylon Medical Language needs to be understood by both machines and humans. When Babylon services talk to each other, they represent data elements as unique identifiers which we call Internationalised Resource Identifiers (IRIs).

Diagram 10: When, for example, both Babylon’s symptom checker and Healthcheck are talking about Appendicitis and want to make sure they’re both talking about the same thing, they must use the corresponding unique IRI for Appendicitis

These IRIs don’t make a lot of sense to humans; plus it’s hard for us to comprehend all the relationships between them. That’s why we also made a web interface - the Knowledge Base Explorer - that visualises the Knowledge Base in an interactive way.

Diagram 11: If we humans want to check we’re talking about the same thing as the machines, we can look up Appendicitis or its IRI in the Knowledge Base Explorer and find out more

LESSON 15: We may build AI primarily with machines in mind, but it’s important to adapt what we build for coherence by humans too.

Conclusion

Now you’ve seen the tip of the iceberg of challenges we at Babylon AI are working to overcome on our way to building a personal AI doctor.

Sure, we’ve started figuring some things out (which we’ve shared with you here), but there’s still so, so much for us to learn.

We know some of you reading this have got better ideas than us for solving these tough challenges. And I’m willing to bet that some of you could even picture yourself joining us full-time on this learning journey...

Can you help us learn?

If so, get in touch because we’d love to hear from you! Send us an email on [email protected]

This blog is part 5 of a 5 part series. Explore the full series: Part 1  Part 2  Part 3  Part 4  Part 5

With many thanks to...

AI Engineering: Mohammad Khodadadi, Domenico Corapi, Jonathan Moore AI Product: Maurizio Morriello, Martin Robbins AI Research: Jet Shamdasani, Giorgos Stoilos AI Clinical: Alex Szolnoki Science PR: Edward Sykes, Amy Palin Design: Matt Jakeway Data Trust & Privacy: Cormac West