Dr. Amol Verma, staff physician in General Internal Medicine at St. Michael's Hospital, on Nov.15.Fred Lum/The Globe and Mail
During one of her shifts on the internal medicine unit at St. Michael’s Hospital in Toronto, Yuna Lee received an alert on her phone from CHARTWatch, an AI-powered early-warning system, indicating a patient in the ward was at high risk of dying or needing intensive care.
Dr. Lee, the division head of general internal medicine, checked on the woman and found nothing obviously amiss. She ordered extra blood tests, just to be safe. The results revealed the patient’s liver enzymes were elevated, prompting Dr. Lee to call for an ultrasound of her liver.
As the patient was about to be transferred to the imaging department, she spiked a fever and developed pain in her abdomen – the first overt symptoms of what turned out to be an inflamed gallbladder. CHARTWatch, which was developed by St. Michael’s data science team and analyzes hundreds of points of patient data to produce hourly risk scores, had figured out something was seriously wrong before doctors or nurses did.
“That was very surprising,” Dr. Lee said. “It made me go, ‘Wow, CHARTWatch is amazing.’ ”
CHARTWatch is one of the few examples of machine learning – a branch of artificial intelligence in which computer models, trained on mountains of data, teach themselves to get better at a task over time – that has been integrated into the regular operations of a Canadian hospital long enough to demonstrate results.
In the 20 months after CHARTWatch’s launch in October, 2020, St. Michael’s general internal medicine unit experienced a 26-per-cent reduction in the relative risk of death among non-palliative patients compared with the same period in the four previous years, according to Amol Verma. The general internist and University of Toronto professor of AI in medicine oversaw the development and implementation of the early-warning system.
While public debate about the promise and peril of artificial intelligence is increasing, the use of AI tools on the front line of Canada’s medical system remains “very scarce,” according to Bo Wang, the recently appointed chief AI scientist at Toronto’s University Health Network.
The slow pace of adoption isn’t the fault of Canadian AI scientists, whose research is among the most cutting edge in the world, said Barry Rubin, medical director of the Peter Munk Cardiac Centre and one of the leaders of UHN’s efforts to integrate AI into its hospitals.
There are simply “quite a lot of challenges with implementing AI in the health care setting,” he explained. To name just a few: Provincial governments are reluctant to invest in the necessary computing power at hospitals; privacy must be guaranteed when the anonymized data of thousands of patients are fed into artificially intelligent machines; and if the data excludes some types of patients, the machines could teach themselves to be biased.
Most important, many AI models are not yet accurate enough to supplant the clinical judgment of human doctors. “The stakes are high,” Dr. Rubin said, “because if you get it wrong, you’re influencing a patient’s care.”
That’s why the AI solutions being deployed in Canadian hospitals tend, as with CHARTWatch, to complement physicians rather than replace them. For now, the country’s health leaders are most excited about the prospect of AI helping to alleviate a staffing crisis by taking rote tasks such as writing clinical notes off the plates of overworked nurses and doctors.
“I am absolutely convinced that advanced data analytics and artificial intelligence is going to transform health care as we know it,” said Tim Rutledge, the president and chief executive officer of Unity Health, the network that includes St. Michael’s, St. Joseph’s Health Centre and Providence Healthcare. “If we can automate tasks that are now laborious, it allows our clinicians to spend more quality time interacting with patients.”
St. Michael’s has spent the past couple of years quietly testing about 50 different AI solutions in that vein. A $10-million donation from Hong Kong businessman Li Ka-Shing kick-started the work. About $5-million annually from the hospital’s philanthropic foundation keeps it going.
On a recent afternoon, Muhammad Mamdani, the vice-president of data science and advanced analytics for Unity Health, showed off a wall-sized whiteboard bristling with colourful sticky notes charting the progress that he and his staff of 30 are making on different projects. (At the far end of the board, scrawled in black marker, was a riddle: “What chemical element are data scientists most afraid of?” The answer was sodium. An early version of the CHARTWatch algorithm misinterpreted data about patients’ sodium levels as being “not available” because sodium’s chemical symbol is Na.)
Along with CHARTWatch, other AI solutions developed at St. Michael’s include a tool for assigning emergency department nurses to different posts, such as triaging patients or working in the ER’s resuscitation bay. That assigning task, which used to require hours of manual input on an Excel spreadsheet, is now done by an algorithm in less than 15 minutes.
Another tool analyzes patient information that triage nurses punch into their computers and uses the data to produce wait-time estimates that flash on a screen in the ER, cutting down on the number of times harried staff members are asked, “How long will it be?”
Yet another project synthesizes the electronic medical records of patients with multiple sclerosis into a concise, visual timeline that is particularly helpful for junior doctors who may only have a 10-minute window to prepare for an appointment. The model can summarize seven years’ worth of charts in less than two seconds.
The advantage of having an in-house development process is that it includes regular meetings with end users. The data science team is constantly tweaking models and monitoring them after deployment to be sure they don’t become less accurate over time, Dr. Mamdani said – a risk when intelligent machines train themselves in ways that aren’t always clear to their human creators.
So far, Health Canada, which has issued a growing number of licences for commercial software designed with AI functionality since 2018, has only approved devices that have their algorithms locked. That means that manufacturers have to apply to Health Canada for a licence amendment if their AI-powered devices have the ability to learn or change, according to André Gagnon, a Health Canada spokesman.
If the situation in the United States is any indication, applications for more sophisticated AI-powered devices are on the way here. As of mid-October, the U.S. Food and Drug Administration had approved 171 AI and machine-learning-enabled medical devices, the vast majority of them in radiology, the specialty that interprets medical images such as CT and MRI scans.
Neither the Canadian regulator nor the U.S. FDA have approved any medical devices that incorporate the type of Large Language Models (LLM) that power ChatGPT, the now-famous chatbot developed by OpenAI. But they’re coming. Google has trained its Med-PaLM 2 to provide sophisticated answers to medical questions, while Amazon Web Services and a slew of competitors are rolling out products that use voice recognition and generative AI to turn conversations between doctors and patients into automatic clinical notes.
Dr. Wang’s lab is developing an LLM-powered medical chatbot and automated clinical note generator called Clinical Camel with open-source software, which means anyone can see the code underpinning it. He and Dr. Rubin are among the scientists warning of the risks of leaving LLM-powered medical solutions in the hands of corporate giants who keep their algorithms secret, and who could take services offline if they become unprofitable.
Radiologist Jaron Chong has a different concern about artificial intelligence in health care delivery. Although there are more FDA-approved AI devices in his specialty than any other, he is not worried his profession will go extinct. Rather, he fears that in Canada’s penny-pinching, hide-bound public-health system, human radiologists won’t be able to adopt AI that could make them better diagnosticians.
“There are hospitals today that still chart on paper,” Dr. Chong said. If 30 years down the road, radiologists are doing higher volume but “using the exact same techniques,” he added, “I’ll feel like we have failed.”