How to Create AI Programs With Clinical ROI

A recent HealthLeaders AI NOW panel discussed how the technology is being applied to clinical care

Health systems and hospitals are seeing specific benefits from deploying AI technology in clinical care, according to executives taking part in a panel at the recent HealthLeaders AI NOW virtual summit.

While much of the so-called “low-hanging fruit” has so far been tied to back-end and administrative tasks, AI tools have been used with considerable success in radiology, where the technology can pick up details in images that can improve diagnoses. And AI is also being used in places like the Emergency Department, inpatient care, and population health programs.

“Financial ROI is important, but it’s not the only factor that health systems should be focusing on,” said Jared Antczak, chief digital officer at Sanford Health.

The rapid pace of development for AI tools in healthcare is tied to the potential for the technology to solve a wide variety of healthcare’s biggest problems, but without a good enterprise-wide strategy in place, some organizations are launching projects with an uncertain ROI and putting pressure on executives to find value after the fact. Advocates suggest launching small AI programs at first with a defined ROI, especially in areas where the value is clear.

In other words, think before you act.

“AI is not and should not be a strategy in and of itself,” Antczak said. “It’s a potential tool that can be used to solve a problem. But really tools are enablers of strategies, not strategies by themselves. We need to avoid the trap of doing technology for the sake of technology and really leverage technology to create value in people’s lives.”

“Knowledge is expanding faster than our ability to assimilate it and apply it effectively,” he added. AI is “a powerful tool that can sift through the noise and the information overload and really help clinicians by lifting up the things that matter.”

Albert Karam, vice president of data strategy analytics at the Parkland Center for Clinical Innovation, a research institute allied with Dallas-based Parkland Health, said the health system is using a predictive AI tool in Emergency Departments at Parkland Hospital and University of Texas Southwestern Medical Center to assess patients’ mortality over the next 12 to 72 hours, to determine when patients are scheduled for surgery.

“The idea here is that the orthopedic surgeons … use that information to decide whether to take [those patients] into surgery,” he said. “If things are looking a little bit grim … they might … try and get some of those metrics better before taking them in.”

Karam said the score developed by the AI tool is comprised of many data sources and updated hourly.

“It’s one of the life-and-death models [that is] a little bit morbid but incredibly useful,” he said. “They were literally having yelling matches in the hallway between the orthopedic surgeons and some of the other surgeons to decide whether or not to take them into surgery, and that has completely gone away.”

Another AI tool, focused on evaluating a patient for sepsis risk, was introduced in the inpatient setting, Karam said. It worked so well that executives in oncology and OB-GYN asked to have it reconfigured for their departments as well, and just recently the tool was reconfigured again to address whether sepsis is present in a patient on admission in the ED.

Antczak said Sanford Health has several predictive AI tools in use with clinical applications, addressing such issues as risk of colon cancer and chronic kidney disease.

“We’ve developed a number of different algorithms around disease state progression and anticipation to really enable our clinicians and our patients to potentially intervene sooner,” he said.

“Sometimes that word ‘healthcare’ is a bit of a misnomer,” Antczak added. “Really, we’re in the business of sick care. We wait until people are sick, and then we react, and we treat them and try to keep them well. We try to keep disease from progressing. But really if we want to become healthcare providers, we need to get further upstream. We need to look at ways to prevent disease from progressing to begin with, and that’s really where I think … AI can help us.”

Karam noted that programs focused on clinical outcomes often take longer to show ROI, which can be a challenge for a health system looking to contain costs.

“Some of the ROI analysis that we do is in lives impacted and lives saved even though we know that this is going to cost more dollars and cents up front,” he pointed out.

In addition, both he and Antczak noted, it takes a while to properly plan and develop an AI program.

“I don’t think people realize to successfully launch and do appropriate quality assurance on these models, it does take a significant amount of time,” Karam said.

It takes “about a year from ideation to starting a pilot,” he said. “And then we’ll pilot that model in one or two departments for another 4-6 months or so before rolling it out to the whole hospital. So about the minimum amount of time from ideation to implementation, even at a pilot level, is anywhere from a year to a year and a half. Which is not a fast turn-around, but there are so many checks and balances, so much with that data governance.”

And finally, the value of using AI in clinical care has to be measured against the risk. Many healthcare organizations are still trying to figure out how to use AI correctly, with the understanding that bad data or prompts can create bad outcomes—including, potentially, patient harm.

“Everyone is trying to identify where the guardrails are,” said Antczak, who notes Sanford Health used a tiered structure to identify risk in AI programs. Both he and Karam said it’s essential to balance any risky AI programs with human review. In any case where AI impacts a patient, they said, someone other than the technology has to make that final decision.

“It’s absolutely the final decision of the clinician or the nurse,” Karam said.