Posted by Rob Knies

Microsoft Research Machine Learning Summit logo

The Microsoft Research Machine Learning Summit 2013 concluded with a plenary panel discussion titled Data Challenges and Opportunities in the Next Decade. Chaired by Jeannette Wing, Microsoft vice president and head of Microsoft Research International, the discussion included Eric Horvitz, Microsoft distinguished scientist and managing co-director of Microsoft Research Redmond; Michel Cosnard, president of InriaIain Buchan of the University of Manchester; and Lionel Tarassenko of the University of Oxford.

My previous post ended with Hermann Hauser, co-founder of Amadeus Capital Partners, stating that machine learning would have a profound effect on the future of health care. That was interesting, because I had planned for the final post from the summit to focus on that very subject.

Buchan is quite aware of that potential. A clinical professor of Public Health Informatics at the University of Manchester and director of the MRC Health eResearch Centre, his research interests lie in building effective models of health and in connecting citizens, patients, and health professionals with more potent health information.

He graciously took a few minutes during a break in the event to discuss the possibilities machine learning can offer for health care.

“There are two areas of transformative potential,” Buchan said. “First is the provision of more usefully complex models of health, relevant to patients with multiple conditions as they get older, patients on more than one medication, women of child-bearing age—the people left out of conventional clinical trials—harnessing that useful complexity by having more brains involved in the modeling, linked by machines aware that Investigator A is touching common data in ways that are relevant to Investigator B, building something that supports health care for people where there are gaps in medical knowledge at the moment.

“Bear in mind that current medical knowledge predicts less than a third of the outcome of the average patient to the average treatment. The natural experiments of health care that happen every day—capturing that knowledge into a bigger funnel of health science requires machines to help people think and work together to fill that two-thirds gap in the knowledge base.”

The second area that Buchan believes holds promise for improved health care involves us all.

“The other big area with potential,” he said, “is having a richer, longitudinal signal from individuals about their own health, machines learning what is engaging for healthier behaviors, not just crunching numbers to relate patients' genes to doctors' decisions. That’s an easy challenge. But machines inferring when you’re likely to respond to a question in your daily life, versus when it would irritate you—getting a really rich signal over time—we don’t have it. It’s not done. Most of the current health-care apps on smartphones, for example, are a best guess at what might work, rather than engineering founded on best evidence.

“But theoretically, there is huge potential to engage with the individual through machine-learned methods that sit in a triangle between psychology, sociology, and technology—the health transactions of daily life.”