Why medicine is creating exciting new frontiers for machine learning

This week, I gave a keynote presentation at ICLR 2020, which—as many will know—is a huge deep learning conference that draws an impressively diverse crowd of participants and speakers. 

Like many, I had been looking forward to meeting colleagues and sharing views against the inspiring backdrop of Addis Ababa, but sadly that was not to be. For obvious reasons, this year’s conference was entirely online, and I’m grateful to its organizers for making the absolute best of a really difficult situation. 

To kick off my keynote, I explained why I think medicine is such an exciting new frontier for machine learning. Some of the reasons are shared below, as are a few clips from the presentation itself. 

Complexities, challenges, and opportunities to solve new problems 

Medicine stands apart from other areas where machine learning can be applied. Where we have seen advances in other fields driven by lots of data, it is the complexity of medicine, not the volume of data, that makes the challenge so hard. But at the same time this makes medicine the most exciting area for anyone who is really interested in exploring the boundaries of machine learning, since we are given real-world problems to formalize and solve. And not only that—the solutions are ones that are societally important, and they potentially impact us all (just think COVID-19!).

Machine learning has of course already achieved very impressive results in numerous areas. Some standout examples include computer vision and image recognition (e.g. recognizing pictures of cats vs. dogs using convolutional neural networks), playing games (e.g. AlphaGo using deep reinforcement learning) or in teaching robots how to act (e.g. using imitation learning). AI empowered by machine learning is so good at mastering these things because they’re easily-stated problems where the solutions are well-defined and easily verifiable. 

“Easily-stated problems” have a clear challenge to solve and clear rules to play by; “well-defined solutions,” fall into a easily recognisable class of answers; while a “verifiable solution” is one that we as humans can understand in terms of judging whether the model has succeeded or not. Unfortunately, in medicine the problems are not well-posed, the solutions are often not well-defined, and they aren’t easy to verify.

Extracting intelligence from data is necessary to advance to new and improved levels of prognosis, diagnosis and treatment. To do this we have more and more data available from a wider and wider variety of sources (for example, imaging, genomics, vital signs and lab test data) all collected over time. We need to learn from this complex data to understand complex diseases such as Alzheimer’s, cancer, cardiovascular disease, cystic fibrosis, and diabetes—all of which exhibit complex phenotypes that vary between individuals and over time, and which will require sophisticated data analysis of data.

Progress toward understanding, preventing and treating such diseases therefore requires new methods. These methods have to break the barriers between machine learning, statistics and mathematics. They should allow us to discover new links, causes and causal relationships among clinical data, genetic data, metabolic data, environmental data, social data, letting us assess the impact of each on the risk, incidence, treatment and trajectory of disease.

Being part of a revolution 

AI and machine learning have certainly not moved slowly in bringing seismic change to countless areas including retail, logistics, advertising, and software development. But I truly feel that I’m working in the area where there is the most unexploited potential for complete revolution. To be clear, I am not talking about doctors becoming obsolete—quite the opposite. We can use AI and machine learning to empower medical professionals by enhancing the guidance and information available to them. We can offer clinicians new approaches to diagnosing, staging and treating diseases or allocating resources, and we can open up new pathways of care. Researchers will benefit from new ways of understanding diseases (and the relationships between diseases), new insight into the effects of treatments, and new methods pf preventing diseases—including drug discovery and clinical trials. 

One small example from a recent post: by using INVASE and subsequently symbolic metamodeling to turn machine learning black boxes into white boxes, we learned something new about mortality in breast cancer patients. Specifically, we learned that i) subtype definition can be refined through tumour grades, and ii) mortality risk increases quadratically with the number of affected lymph nodes in the patient. This is valuable knowledge that allowed us to provide actionable recommendations about detecting and treating a disease that affects millions of women every year. Being able to make discoveries like this on a regular basis is incredible. 

The chance to collaborate on real-world problems 

For me, collaboration is one of the most appealing aspects of straddling the divide between machine learning and medicine. My team works with doctors and clinical researchers because they are invaluable contributors in all our research. True collaboration is a challenge, of course: we are all highly specialized in our respective areas, with different ways of thinking and different professional languages and approaches, so we must each make extra effort to reach the middle ground. But it’s a fascinating and invigorating way to work. Listening to clinicians or researchers can guide us to where problems and challenges might lie, and then we can start being creative in trying to solve them. I also love being able to see the real-world impact of machine learning around me. The problems we solve (though difficult to initially formulate) are far from theoretical: they arise from issues facing doctors and patients every day. They are grounded in the very tangible reality of health and disease. The revolution that machine learning in medicine brings will change how we all live for the better.

If you’d like to see what I said on this topic at ICLR, please check out the video embedded directly below.

Over the next couple weeks, I’m hoping to put together a few more posts on the other topics I touched on in my keynote, as well as highlighting some wonderful work by my Ph.D. students, who had two papers selected for the conference.