NIH Logo
September 7, 2018
GAMING MEDICAL RESEARCH
AI Workshop Surveys Landscape, Assesses Big Ideas

Dr. Ellen Rothenberg discusses T-cell lineage.
AI workshop speaker Craig Mundie (l) casts a bold vision of medical research in the machine-learning era, as NIH director Dr. Francis Collins looks on.

It’s been nearly 160 years since Milton Bradley introduced a “Game of Life” wherein players navigate around a brightly painted board, landing on spaces that dole out good fortune and bad. Although modern editions billed it as a “game of skill and chance,” contemporary Life relied more on the whimsy of a Wheel of Fortune than on personal acumen for success. At the Day of Reckoning you were as likely to wind up a millionaire as a pauper. That’s life, right?

What if players had extraordinary, suprahuman help, though, to foresee dangers, dodge misfortunes and seize advantages? And what if the game ultimately was your real life—or specifically, your health?

That was the fascinating salvo keynote speaker Craig Mundie launched at a recent all-day workshop devoted to “Harnessing Artificial Intelligence and Machine Learning to Advance Biomedical Research.”

We’re seeking prescience about our health, the tech oracle suggested. “How do we structure the problem in such a way that the machine can learn the answer and tell it to us, as opposed to the other way around?” he asked. “This is a big difference between whether you’re just trying to make the machine do what you already know how to do—things it can do more efficiently—or, are you saying, ‘No, the machine can know things I’ll never find out’? That’s a very different way to think.”

Held appropriately in NIH’s on-campus brain center, the Porter Neuroscience Research Center, the meeting gathered thousands of in-person and online attendees—both the AI-experienced and the AI-curious—to explore what NIH director Dr. Francis Collins called “something pretty profound.”

Artificial intelligence, “considered to be the bigger umbrella, can sense, reason, act and adapt,” he said, defining terms at the outset. Machine learning (ML) refers to algorithms “whose performance improves as they are exposed to more data over time, with the learning aspect being the emphasis.”

Collins sought advice on how NIH should engage AI and ML to manage the huge amounts of data the biomedical field is generating and collecting and aim that knowledge toward improving health.

AI, big data, robotics and cloud computing are all included in what some have categorized as “the fourth industrial revolution,” he said, putting the workshop in historical context. “This topic is certainly one that is due for a deep dive by NIH leadership in terms of where we are and where we might need to go.”

The goals of the workshop were to survey the state of AI for biomedical research, dig deeper into specific examples, explore opportunities for NIH to make maximum impact and identify key challenges and obstacles and how can NIH help solve them.

The event also featured numerous innovative ideas presented during two plenary sessions (see sidebar) and discussions afterwards with audience and speakers.

Ultimate Chess Match

Stirring the pot first, though, was former “visionary-in-residence” at Microsoft Mundie with a provocative proposition: AI learns faster and better than humans ever will, so why not give it all of the raw biological data and let the machine learn the pathways and functioning of human biology and use that to predict or find ways to resolve high-dimensional problems such as complex disease?

The longtime tech mega-consultant is known for identifying patterns in information, matching compatible designs and concepts and foreseeing how they could possibly work together to transform whole industries. In the past few years, he’s directed this talent toward the health and biomedical research fields.

Collins invited Mundie to envision NIH’s ideal role regarding the artificial intelligence age; for inspiration, Mundie looked to the computer gaming world. He began by asking a bold question: “Is human biology too complicated for humans to figure out?”

Remarkably in this age of “big data,” he wasn’t endorsing putting more resources into mining gigantic data sets. Mundie’s prescriptions leaned more toward “minimization as opposed to maximization.” So, he posited, we have the human genome and the proteome and maybe that’s as many “omes” as we need to feed into the machines.

“Most of everything we’ve done in data-driven biology and medicine starts with the assumption that we look at a population and try to extract from that population things that we then apply to the individual,” he said. “My prediction is that the future of medicine is all about starting with the individual, and ultimately, as you see enough commonality, you’ll synthesize the population answer—that’s completely upside down from everything the world currently has done.”

Humans, No Monopoly on Knowledge

Mundie (l) suggests the computer gaming world has important lessons for medical research. Collins talks about mining humongous data sets such as the Human Genome Project.
Mundie (l) suggests the computer gaming world has important lessons for medical research. At right, Collins talks about mining humongous data sets such as the Human Genome Project.

PHOTOS: CHIA-CHI CHARLIE CHANG

Mundie’s “penicillin discovery moment” occurred as he encountered recent reports from the computer gaming community. Professional machine-learning gamers studying how AI could be applied to the ancient game “Go” or to solving Rubik’s Cube were coming to terms with their human frailties—computers were increasingly employing the sheer scale of their AI advantage to beat human experts every time. The machines were synthesizing and learning from the different maneuvers and strategies of millions of iterations of the games in a matter of days—more than any human player could hope to experience in a lifetime.

Then, Mundie observed a chance situation in biology where the learning system compensated for a human error in providing requisite data and got a better answer without it.

“These things learn at a scale that humans cannot do and never will be able to do,” he explained.

Showing an article from the most recent Harvard Business Review, Mundie said, “They like many others are wrestling with the question of ‘What does it mean for mankind when the computer does all of the things that we do?’” They suggest a collaboration between human and machine where the humans provide three key capabilities— training, explaining and sustaining. He questioned whether the first two were really required.

Clue to Health?

Perhaps the biomedical research community—and NIH in particular—should get into the game-design business, Mundie said, sketching an outline of the potential competition.

“Life is just a tournament of multi-player games,” he said. “To win this game, teams have to get the most people to have the longest, highest quality existence. Every game is unique.”

Who are the players? “Mother Nature is a player and she has two tools—randomness and evolution,” Mundie said.

Other contenders are the environment, which includes the physical world as well as political, financial, policy, governance and computing infrastructure factors. Of course, AI and ML—“the bots”—are playing the game too. Their role is to help humans, who include individuals and the people employed in the medical research, health care delivery and wellness and prevention ecosystems.

“This is really an interesting way to think about applying all the different elements of AI to this problem,” Mundie concluded. “This approach discards existing history about biology and medicine and builds an understanding of pathways and interventions up from the individual rather than down from the population. Ultimately, humans will trust—rather than understand— what the machines know and do, just as they do with most of the computerized infrastructure in their lives.”

And how would success be measured? Many in the audience wanted to know.

“What I consider to be winning for me personally, I’d like to live my life where my physical body and my mental capacity stay in sync and in the end they kind of get extinguished together,” Mundie concluded.

The entire workshop is archived at https://videocast.nih.gov/summary.asp?Live=28053&bhcp=1.

AI Applied in Clinical Settings

In addition to a disruptive keynote talk by tech thought leader Craig Mundie, the recent all-day artificial intelligence (AI) workshop heard a number of practical considerations and applications of machine learning in clinical settings that provided NIH with plenty of food for thought.

Dr. Dina Katabi of MIT discusses AI innovations in clinical research and clinical care.
Dr. Dina Katabi of MIT discusses AI innovations in clinical research and clinical care.

Dr. Dina Katabi of MIT’s Center for Wireless Networks and Mobile Computing introduced the “health-aware home,” her lab’s AI clinical application prototype.

Using the electromagnetic field surrounding everything, the group’s wi-fi-like technology can track and measure people’s breathing, sleep, heart rate, falls and gait speed—all without attaching sensors to the subject. The device transmits low-level wireless signals and uses AI to analyze the signals. It’s already been tested in assisted-living facilities with people who are elderly or who have sleep and breathing disorders, Parkinson’s disease or other chronic health problems.

“The future in health care is terrific,” Katabi said, “and computer science with additional technology, AI and machine learning will come together to provide much stronger types of data and also provide information for the health care system to improve the quality of care.

“Imagine what would happen if we had a device like this in the home of every chronic disease patient—how much more we would learn about diseases, how much more we would be able to intervene before exacerbation. I think we can do that if we start thinking together, bringing the computer scientists, people who work on AI, with people who are in the health care system—the doctors, the biologists, the bioengineers, the chemists—and creating that future.”

NIH’s own Dr. Ronald Summers of the Clinical Center offered proof that machine learning can significantly improve precision and accuracy in biomarker imaging diagnostics.

Dr. Judith Dexheimer of Cincinnati Children’s Hospital talked about her group’s research on making the wealth of data contained in electronic health records, or EHRs, more universally useful and accessible.

The event’s line-up also included intriguing talks by Eileen Koski of Microsoft, Dr. David Heckerman of Amazon and Dr. Anshul Kundaje of Stanford.

It’s clear that “we need to rather quickly escalate our involvement and our investment,” said NIH director Dr. Francis Collins, offering closing thoughts at day’s end.

Other action items he noted include the need to prioritize, harmonize and clean up data sets and better enforce provisions giving researchers access to such data; plan for quantum computing; and build a broader community for people in AI-related professions exploring health research applications.

Collins also said he intends to bulk up NIH’s brain trust, “deepening our own bench.” He’ll follow up the AI conference by convening a working group within his advisory committee to the director.

“We have a critical opportunity here now and we don’t want this moment to come up and go down without having some legs,” he concluded. “What exactly those legs should look like will have to involve some folks who can think big and give us big advice…

“We want our approach to AI to be more than just ‘A,’” Collins concluded. “The ‘artificial’ part comes really easily, but we need lots of ‘I,’ the intelligence part—and better yet, we need both of those together. And we’re only going to get there if you help us.”

back to top of page