Illustration by TIME; reference image courtesy of Dr. Keith Dreyer

When it comes to putting AI to work in health care, there are plenty of blue-sky promises. But for Keith Dreyer, the practical challenge of figuring out how to fold AI-based strategies into the way doctors diagnose and treat their patients is what consumes his days. With degrees in mathematics and computer science, Dreyer has the ideal background to serve as chief data science officer at Mass General Brigham. It’s his job to oversee the dozens of AI-based algorithms that the health system currently uses in reading images, from strategies that are still being tested to those that have received approval from the U.S. Food and Drug Administration (FDA).

Since models can become outdated quickly, even approved algorithms need constant evaluation to ensure they are still working as intended. “Across the U.S. right now, there is a shortage of radiologists, not because there are fewer radiologists, but because there is so much more imaging data,” he says. “AI is being called upon to solve some of those problems in efficiency and shortages.”

The AI systems that Dreyer uses typically help to triage the mountain of images that inundate doctors—from CT scans, MRIs, and the like—by picking out the ones with the most potential for showing health threats like cancer.

Ultimately, machine learning could lead to more sophisticated AI tools that can detect and diagnose disease, and Dreyer is involved in discussing issues such as reimbursement for AI-based work, the security of patient information used in algorithms, and how much autonomy such technology should, and could, have in medicine.

“I don’t think there is a pathway now for a truly autonomous AI to be approved in health care,” he says, pointing to the imbalance in the way AI-based algorithms and human doctors are certified to practice. For now, AI algorithms only need to pass certain tests to demonstrate their accuracy in order to be approved by the FDA for use in patients, while doctors have to complete the lengthy and continuous process of medical school, state certification and licensing, hospital certification, and continuing medical education. “At some point, as AI becomes a little more autonomous, we may have to reconsider the [AI approval] process,” he says.

More Must-Reads from TIME

Contact us at

Sougwen Chung
Yann LeCun
Abeba Birhane
Audrey Tang