Like many bench scientists, neuroscientist Dwight Bergles says advances in AI are transforming his research by rapidly accelerating the speed of cellular analysis.

His lab focuses on the repair of myelin, the insulation around nerve cells that is destroyed in multiple sclerosis, by using microscopy to look into the living brain of a mouse that has been engineered to allow certain cells to fluoresce. “Traditionally,” he says, “we would have to do all of that work by hand, which greatly exceeds what a trainee can do over the period of a Ph.D. or postdoctoral fellowship. It takes months and months and months — sort of a brute force problem — before you can start to get to the interesting parts of the research.”

Using AI not only speeds the analysis, he says, but “it dramatically allows us to expand our ability to analyze the entire brain, looking at these more regional differences,” rather than focusing on just one small part of the brain in great detail.

Cardiologist Natalia Trayanova and biomedical engineer Sridevi Sarma have both been early leaders in creating and using AI algorithmic tools to improve clinical care. Both cite current concern around the “black box” of AI (see glossary) as a challenge that must be addressed.

“The algorithms are there to help physicians make clinical decisions,” says Trayanova. “We want them to be as transparent as possible. The clinicians need to see why the algorithm is making that decision. It’s important for clinical acceptance.”

In Sarma’s current work to improve treatment for epilepsy, she has developed a novel computational tool, EZTrack, that analyzes hundreds of electroencephalogram readings to pinpoint the brain’s epileptogenic zone, where seizures originate, to improve the accuracy of surgical removal and surgical outcomes. Crucially, her “systems approach” offers an opportunity for clinicians to peer into the black box to understand how AI reaches its conclusions.

She uses imaging data to build a digital brain and then examines connectivity properties of the virtual brain to identify pathological regions. “This is patient-specific, with mathematical equations you can visualize and interpret,” she says. “It only takes a few minutes, and I can hand my deck of slides to the neurosurgeon.” Sarma is currently taking a similar approach to develop AI tools for autism, major depressive disorder, ADHD and frontal temporal dementia.

Real-World Deployment

In her research, computer scientist Suchi Saria has collaborated with federal agencies like the FDA to develop new rubrics and tools for improving real-world deployment and evaluation of AI. She is sharing her insights with the National Academy of Medicine, helping to create guidelines for AI use.

“When you’re developing a new tool, you take a data set in a lab and test it. That is a fixed, limited data set. It’s controlled. As you move from the lab to the real world, lots of changes can occur. Maybe you’ve trained it one type of population and now you’re deploying it elsewhere. There could be differences in how data is collected or in demographics.”

AI needs to account for these changes, she says, while recognizing differences between global, local and even individual patterns. Further, it needs to be transparent, easy to use and responsive to the care team’s input.