This Tech Pro Says AI Is Ready to Revolutionize Our Health Care in Canada

With Canada’s health care system in such disarray, it might be time to call in the machines. Currently, there are more than 300 homegrown startups working on health innovations fuelled by artificial intelligence, from smartphone tools that judge the severity of a wound to a handheld digital device that ​detects cardiac disease to a platform that predicts the global spread of viruses like COVID and monkeypox. While most of these systems aren’t ready for rollout quite yet—though in some Toronto hospitals, AI is already flagging at-risk patients who may require a transfer to the ICU—they’re poised to make a big impact in health care soon.

To better understand what this future could look like, we spoke with Azra Dhalla, the Director of Health AI Implementation at Toronto’s Vector Institute for Artificial Intelligence. “AI has tremendous capabilities, but its transformative powers have yet to be fully realized in health, and that’s what we’re trying to change,” she says. “So we’re working with hospitals, health care agencies and academia to take this world-class research and translate it into something that’s really tangible.” Here, Dhalla discusses AI’s potential to bring hospital wait times way down, the need for more diverse health data and why a robot probably won’t be the one tinkering with your wonky knee.

Before we get to the future of health care, are there places where AI is already being used in medicine?

Yes. There’s ChartWatch, which focuses on predictive analytics. It’s an early-warning system that pulls vitals from patients in the internal-medicine ward to predict whether a transfer to the ICU or a death will occur. So the predictive power of these solutions can really lead to improved decision-making and the ability to intervene early.

What about other AI models working away in the background?

There are some, but I would say that in health care, very few models have been deployed in a clinical setting. Health care has a number of challenges, especially when we’re dealing with data—we have to be very stringent about security, about privacy, about confidentiality. The thing with AI algorithms is that, similar to how you and I learn, AI algorithms get better as more data is provided. But that can only happen if we can actually get access to it, which is very difficult in health care. However, we’ve partnered with Gemini, a data collaborative of more than 30 hospitals’ data in Ontario, the largest of its kind in Canada, and that’s allowed Vector researchers to develop cutting-edge AI models and solutions, including studies related to COVID-19.

What are some of those projects in development right now?

I’d say there are three areas worth highlighting: personalized medicine, drug discovery and creating a more efficient health system. With personalized medicine, algorithms can help us predict illness and support patients long-term. So, for example, you can use AI to predict Alzheimer’s disease based on changes to speech patterns. Or you could use it to discover insights within imaging data that can guide treatment and therapy decisions for patients with breast cancer. With drug discovery, AI can analyze pharmacological and health data to find different combinations of drugs that can be used to target existing and emerging viruses, or treat conditions that the drug might not have originally been prescribed for. And with health systems, it could help alleviate wait times faced by patients in hospitals, which is a big issue in Canada right now. When you bring in AI, the potential for us to better allocate resources, both in terms of staffing and funding, is fantastic and leads to better patient outcomes.

On the other hand, what isn’t going to happen with AI and health care? You must hear some pretty wild theories when people find out what you do.

One thing is that AI is not going to replace a physician—it will augment clinical decision-making, but it won’t replace it. It’s more like a virtual second opinion, not meant to override human judgement or expertise.

So I’m not going to roll up for knee surgery and find a robot about to perform it on me?

Well—I can’t predict the future. But I don’t think that’s going to happen.

What do you hear from doctors and health care practitioners when you talk to them about AI?

They really do want to know how we’re using AI to revolutionize health care, and they want to know not just on a theoretical level but a practical level. How can they use these solutions in a clinical setting? What does it mean for patient care overall? That’s always their number-one question—well, actually, I’ll say there are two questions. Number one, will it be disruptive to my workflow? And the second is, what are the outcomes that can be produced for a patient?

What worries them about their workflow?

What they say is: We don’t want another button to press. We want it to be very seamless. And also they worry whether this all happens in a black box. Explainability in AI is very important—we don’t want to just use this blindly. So if an algorithm makes some kind of decision, we need to know how it has actually come up with that decision.

We hear a lot about bias in AI. How can bias skew an algorithm’s results?

You hear the expression “garbage in, garbage out.” AI algorithms will always reinforce bias if the data they’re trained on is biased. If we’re looking at a pool of health care data that is only representative of a certain segment of the population—

Say, white men of a certain age?

That’s right. Then when you try to apply the AI model to a different or a more diverse population segment, it doesn’t work, or it won’t work in the same way. A good example is an image-recognition model that wasn’t able to recognize melanoma in patients with different skin types, because the model wasn’t trained on data that was representative of the whole population. I will say that there’s much work being done on responsible AI, making sure that we correct for inherent biases.

And how do we do that?

By ensuring that there’s access to very diverse data. And then by looking at that data to really say when it isn’t representative of an entire population, so that if there are inherent biases, we can correct that at the forefront. We also want to make sure our models work for everyone. So in AI implementation, we do these silent trials, where we test out the solution in, say, a hospital, before it goes into practice. Because we don’t want to just say, hey, this tool works fantastically, we’re gonna implement it now. Being able to pilot it is extremely important.

People are understandably quite anxious about the state of Canadian health care. What do you see as the potential for these AI programs, whenever they do get rolled out?

I truly believe that AI has transformative benefits for patients. There is a machine learning model that can create radiation therapy treatment plans for patients with prostate cancer. That can take a clinician more than a day to develop, and the model produces plans within hours that are deemed to be as good as or even better, nine times out of 10. If I were a patient, this is exactly what I’d want: something that creates efficiencies and frees up resources so that I not only have a personalized treatment plan sooner—but I get to spend more time with my physician. That’s extremely beneficial to a patient’s quality of life and the quality of care they receive.

This interview has been edited and condensed.

Next: The Forces That Shape Health Care for Black Women