Ottawa surgeon is on a mission to build safer health-information AI

Many patients have common questions about their reactions and recovery from dental surgery. To help clinics answer them, Dr. Hassan Moghadam, an oral and maxillofacial surgeon in Ottawa, co-founded an artificial intelligence (AI) platform called Lillian.

He knew the most useful way to test it: “I would try to make it fail.”

Dr. Moghadam had his team upload approximately 450 textbooks and 350,000 journal articles into an early version of Lillian. He then posed difficult clinical questions and watched for answers that were distorted or wrong.

One result was particularly concerning: a company had published so extensively about a dental material that the AI interpreted the volume as evidence the product was sound. “We know [the product] is not great. If you just use AI, it looks at, ‘Oh, there’s 10,000 articles about this – it must be good,’” says Dr. Moghadam, a partner at the Argyle Associates clinic.

The team changed how Lillian weighed its sources, screening for products backed by corporations, and favouring university-cited research. It also added material presented at conferences, to keep on the cutting-edge of new techniques.

Large language models (LLMs) are designed to predict likely text, not to establish medical fact, cautions Matthew Guzdial, an associate professor who researches AI at the University of Alberta’s Faculty of Science. “They don’t have any notion of ground truth. In some respects, they’re always hallucinating.”

According to a recent survey by the Canadian Medical Association, Canadians who followed health advice from AI were five times more likely to experience harm than those who did not. The survey also found that 69 per cent of Canadians had become skeptical of health information online, including material from sources they believed they should trust. Only 27 per cent trust AI to provide accurate information about health care or the health system.

AI app developers can add filters and system instructions, asking the tool to cite where it’s getting its sources. But even citations can be unreliable, says Mr. Guzdial, as a model may attach a made-up source. “There’s no actual understanding here. It’s just text in, text out.”

Lillian’s developers are also applying AI to analyze rejected dental insurance claims, identify submission problems and resubmit claims with revised information.

The project is not Dr. Moghadam’s first effort to build dental and medical technology. Earlier attempts included using AI to improve optical imaging and capture detailed images of tissue, and creating a “GPS” for surgery to help doctors navigate to exact locations inside the body.

Dr. Moghadam was unable to secure funding for those projects, “but the key here was pivot, pivot, pivot.”

His experimentation with AI extends beyond health care. Nine months ago he opened the Sofia Vintage Motor Café in Ottawa. From winter to summer, sales plummeted by 30 per cent and Dr. Moghadam was eager to diagnose why.

He hired a Carleton University business student to analyze the situation and to feed sales data into ChatGPT. They realized that while customer traffic had held steady, sandwich sales had fallen, most likely because another sandwich shop had opened nearby.

In response, Dr. Moghadam cut sandwich prices by $1, changed the bread, introduced a sandwich-and-drink deal, and moved prepared sandwiches to the counter “so people are buying with their eyes,” he says. All were sales tactics suggested by AI.

The café still hasn’t turned a profit, but it has become a community hub, raising tens of thousands of dollars for local hospices.

As Lillian remains in development, Dr. Moghadam remains convinced that AI will have a positive impact. Still, human intervention is critical. While a patient might ask the AI whether bleeding is normal in a certain instance, and receive initial medical guidance, the user can always opt to speak to a doctor by activating a button on the interface.

“Anything that’s very creative still requires a human mind,” Dr. Moghadam says.