How to Balance AI Adoption with Governance

Duke Health CHIO Eric Poon says healthcare leaders can move quickly and fail fast, but they had better learn from those mistakes and adjust accordingly

The Duke University Health System is investing in roughly 80 AI products. And they're fine with the fact that not all of these tools are going to work.

"We know that not everything that we put our hands on or that comes our way is going to pan out," says Eric Poon, MD, MPH, FACMI, the health system's Chief Health Information Officer. "We don't ask for a mountain of evidence up front. Our governance philosophy is that we need to fail often and fail fast, so we provide just-in-time advice."

Welcome to what some have called the Wild, Wild West of AI in healthcare. It's a heady time to be an innovation executive these days, working with a technology that has strong potential to transform a troubled industry. But Poon, like so many of his colleagues in health systems and hospitals across the country, knows the path forward has to be managed just right. Mistakes are OK, but they had better be learning lessons.

"We openly tell our clinicians that AI is not perfect and AI is capable of making mistakes, so we as clinicians need to be responsible for everything that we take advantage of out of these products," he says. "In this case anything that we put in the chart, we need to take direct responsibility for, which means that we need to take very good care in reviewing the notes."

One example is ambient AI, in which Duke Health has partnered with Abridge to deploy a tool for physicians to capture the doctor-patient encounter. More than 1,200 clinicians in the health system are now using the tool.

Poon says clinicians have been using voice-to-text technology for some time, so the idea of capturing the doctor-patient conversation isn't entirely new. But with AI layered onto the process, the notes have more context and are more easily integrated into the medical record, potentially reducing the time spent by the doctor translating notes while enhancing opportunities to identify care pathways and coding.

Clinicians were given some training on how to use the technology, Poon says, but they were also given some freedom to integrate the tool into their own workflows.

As they become comfortable with the technology, Duke Health will look at how it can be integrated into other clinical pathways. Poon says this will include "workflows upstream and downstream from the generation of the note."

Therein lies the true benefit of working with AI. Accessing and analyzing more data gives clinicians the opportunity to better manage the patient's care journey—before, during and after the patient-doctor encounter. And while this data can unlock those opportunities to improve clinical care, clinicians have to be careful that the data being used is accurate and reliable.

The key to keeping track of the successes—and failures—is governance. Poon says AI governance was launched initially as a separate entity to technology governance, then quickly integrated when it became clear both were looking for the same things.

"We've figured out that you cannot have one single group squirreled away trying to perfect data," he says. "It needs to be a team sport, so we have a distributed data governance process and I think we in some ways take a risk-based approach."

Duke Health's approach isn't unique. While some healthcare organizations separate their AI governance, more health systems and hospitals are working those responsibilities into existing committees and workflows. Some simply don't have the time or resources to create more layers of governance, while others are recognizing that AI technology is fast becoming a part of many workflows.

Poon says healthcare executives need to understand that AI is advancing faster than many other technologies in healthcare, in part because the technology is easy to use and is being embraced just as quickly by consumers. The temptation for healthcare leaders is to move that quickly as well, which is why governance is so important.

"We've put a lot of time and effort" into responsible AI development, he says.