We’ve made great progress
We have a long way to go to reach our mission, but we’re incredibly proud of our achievements in AI so far.
We’re proving the credibility of AI in primary care
In early pilot studies, in association with Stanford and Yale, our AI was compared against seven highly-experienced primary care doctors using 100 independently-devised symptom sets (or vignettes). In these specific tests, our AI scored 80% for accuracy, while the seven doctors achieved an accuracy range of 64-94%. We’re currently preparing a number of studies to evaluate the impact of our Al in real-world settings.
We’ve built an AI for medicine that is not just a ‘black box’
Unlike a 'black box' that makes data difficult to interpret for scientists, programmers and users, our AI model is fully explainable. So humans can understand how we've reached our conclusions. This is known as being ‘interpretable’ or ‘transparent’ in data science speak. In this way we ensure that while AI continues to develop behind the scenes, all of our app updates have been approved by clinicians. And we're constantly improving our privacy policies.
We’re deploying AI in healthcare at scale
Our models reach millions of people in multiple countries across the globe through our app.
We’ve pushed what natural language processing can do.
We have published numerous pieces of peer-reviewed research on Natural Language Processing. Our AI has been created to understand medical terms and data so it can gather information from medical datasets - but it can also read and learn from patient health records, including the consultation notes made by our clinicians in the different countries where we work.
There’s much more to come
It’s fantastic to see progress but it’s also important to recognize that we’ve got so much more to do. One of the things we’re most excited about is that we’re pushing a whole new area of AI:
Machines have been programmed to reason by associating a potential cause to a set of conditions. As Judea Pearl, Turing Award-winner and professor of computer science at UCLA, puts it, the current common practice in AI “amounts to curve fitting.”
Machines don’t actually know how much about the causal relationship between variables. Put another way, they have the ability to associate fever and malaria, but not the capacity to reason that malaria causes fever. Once this kind of causal framework is in place, it becomes possible for machines to ask counterfactual questions how causal relationships would change given some kind of intervention. This is a whole new area of AI, and its implications are colossal. Read more on Deep Learning Models with Constrained Adversarial Examples..
We’re part of the community
We contribute to the AI community by publishing papers, speaking at conferences and open-sourcing some of our work for the benefit of all.
Estimating Mutual Information Between Dense Word Embeddings
Vitalii Zhelezniak, Aleksandar Savkov, April Shen, Nils Hammerla
Hybrid Reasoning Over Large Knowledge Bases Using On-The-Fly Knowledge Extraction
Stoilos, Giorgos and Juric, Damir and Wartak, Szymon and Schulz, Claudia and Khodadadi, Mohammad
Can Embeddings Adequately Represent Medical Terminology? New Large-Scale Medical Term Similarity Datasets
Claudia Schulz, Damir Juric