Artificial intelligence and machine learning have crossed the threshold from theory and pilots to real implementations in hospitals. But the reality of machine learning’s role is a bit different from the idyllic vision of IBM Watson reading the New England Journal of Medicine and setting a care plan, Beth Israel Deaconess Medical Center CIO Dr. John Halamka said in a talk today at the Innovation Live Pavilion at HIMSS19.

“We’re focused on use cases that are high value to increase efficiencies,” he said.

For example, the hospital has always given surgical patients one hour in the operating room. But many patients — including younger patients or patients having more routine operations — don’t need the full hour, which means a lot of valuable operating room time is being wasted.

So BIDMC is feeding a lot of historical data into a machine learning algorithm run on the Amazon cloud, and letting that algorithm set times for surgeries.

“In doing this, we have opened up 30 percent of the operating capacity of our hospital by simply scheduling the OR more logically,” Halamka said.

That’s just one example. The medical center is also working with Google to use machine learning to keep track of surgical consent forms. They’re also using the technology to predict no-shows and reach out proactively.

“There’s a lot of reasons patients might not show up to an appointment,” Halamka said. “Maybe they don’t have a car. Maybe English isn’t their first language. So we use machine learning to figure out who is unlikely to show up, and then we figure out what we can do to help them.”

They’re also working with Google to implement semantic search in the EHR. Semantic search is capable of understanding confusing or imperfect human sentences that can perplex a normal search function.

Ultimately, Halamka is excited by the prospect of machine learning being incorporated into the EHR in even more ways, using data-based insights to compensate for gaps in physician knowledge or experience.

For instance, he told a story about visiting India and treating a patient with abdominal tuberculosis, a condition virtually unheard of in the United States where he normally practices but fairly common there. A machine learning algorithm might have been able to point him in the direction of that diagnosis sooner.

In another example he gave, the presence of cannabis in a 24-year-old patient’s lab results caused an emergency physician to dismiss odd behavioral symptoms — which turned out to be symptoms of meningitis. An algorithm offering diagnostic suggestions might also have helped to mitigate that doctor’s dangerous human bias.

“I’m really looking forward to the next couple of years when our EHRs are surrounded by app and cloud services that make them much more useful,” Halamka said.