SickKids in Toronto has been building AI infrastructure for more than a decade and is now preparing to trial a system that could order diagnostic tests before a physician has seen the patient.THE CANADIAN PRESS/Frank Gunn
It’s lunchtime at The Hospital for Sick Children in Toronto. A child arrives in the emergency department, doubled over in abdominal pain. The standard intake information is collected: vital signs, symptoms, history. An artificial intelligence model assesses the data, flags a high likelihood of appendicitis, and orders an ultrasound.
No waiting for a physician to work through a queue of patients. By the time the child is seen by a doctor, the imaging results are in hand. A same-day surgery is booked, and the child is soon recovering at home.
That future is not far off. This summer, SickKids will launch a trial that brings this scenario closer to reality.
“It represents a massive leap forward in how AI will accelerate patients towards a diagnosis faster,” says Dr. Devin Singh, co-lead of the SickKids’s artificial intelligence service, known as SKAI.
During the first phase of the trial, a trained clinician will review and approve each AI-generated recommendation. The model has been running silently in the background at SickKids for more than 18 months, and the results, Dr. Singh says, are impressive. It’s the culmination of more than five years of development, and a glimpse of what hospital medicine will look like in the decade ahead.
The goal with the trial is to use data captured at triage to anticipate which patients are likely to need particular tests before a physician has formally assessed them.
“Why have a patient sit for four to six hours waiting if we can make this prediction with such great precision?” Dr. Singh says. “Let’s just order the test right away.”
The longer-term vision, once the clinician-review phase is proven safe, is for AI to communicate directly with patients and families to initiate testing. Getting there, however, will require additional research into consent and patient experience. How far and how fast that goes depends, Dr. Singh says, on having the institutional foundations to match the ambition.
SickKids has been investing in AI infrastructure far longer than most global healthcare institutions. The hospital began building machine learning projects more than a decade ago, well before artificial intelligence became a mainstream conversation. Because they were so far ahead of the curve, they not only had to build the tool, but they also had to create the regulatory and ethical safeguards to use them safely.
“We discovered that governance was a huge obstacle and barrier to getting these technologies translated,” Dr. Singh says. “Eight to 10 years ago when we were thinking: ‘Okay, we’ve got some models, now how do we deploy them?’ There was no runway in place.”
Dr. Devin Singh, co-lead of SickKids' AI service, says the goal is to use predictive models to accelerate patients toward a diagnosis.Supplied
So they developed one. From the earliest stages of any project, SKAI now convenes a team that would look unusual in most technology development environments: privacy officers, risk managers, ethics specialists, quality improvement experts, cybersecurity teams and hospital executives alongside the engineers and clinicians. Patient and family voices are recruited into the process. That work, Dr. Singh says, is helping to inform the rules Canada is only beginning to write for AI in healthcare.
With an ethical framework in place, SKAI has been able to push into territory few institutions have attempted. For example, Dr. Singh and his team are also using an AI-analyzed 3D photography model to screen newborns for craniosynostosis, a condition in which the skull’s sutures fuse prematurely.
For this disorder, the timing of diagnosis is everything. Babies diagnosed before four months of age can undergo a minimally invasive endoscopic procedure and recover quickly. Miss that window, and the surgery required becomes dramatically more complex: the skull is removed, broken apart, reformed and replaced. Detecting the condition early is difficult, however, and requires expert clinical assessment that falls outside the scope of most primary care physicians.
The long-term goal is to encode the craniosynostosis model so it runs on a cell phone.
“Your community pediatrician or your family doctor would just use their mobile phone, capture the image and instantly you’re getting the expertise of a craniofacial surgeon at SickKids,” Dr. Singh says. “Think of how just insanely impactful that’s going to be.”
That vision is grounded in a clear-eyed understanding of what AI can and cannot do. No model is perfect, he says. The mistake is assuming one could be. Instead, his team invests in understanding exactly where a model will perform well and where it will struggle, then designs clinical workflows to compensate.
The predictive diagnostics models are a case in point. They are optimized to be right, not to catch everything. The objective was to create a system that, when it flags a patient as needing a test, is almost never wrong. A negative prediction does not send a patient home. It simply means they continue through the normal physician assessment pathway.
“We would only ever use the model to accelerate care forward,” Dr. Singh says. “It’s not about blindly embracing technology, but recognizing that here’s a tool and an opportunity to dramatically expand our capacity to provide more care to patients, to reduce wait times and to get to diagnoses faster, and quite literally, save lives.”
While institutions such as SickKids are setting the pace for AI integration, national medical leaders are weighing the broader implications for the profession. Dr. Margot Burnell, president of the Canadian Medical Association, says the sector is only beginning to grapple with what responsible AI adoption actually requires of physicians.
“It’s about trying to find that sweet spot of making sure that patient safety is number one balanced against innovation,” she says.
Part of finding that balance, Ms. Burnell argues, is ensuring clinicians understand the tools they are working with. In partnership with SickKids’ Learning Institute, SKAI runs an enterprise-wide AI literacy program that encourages staff to think critically about the tools they use. She says that this kind of education will be essential across the sector.
“Physicians and their teams are going to have to be educated about the algorithms that are used in these AI tools, what the training platforms were, what the validation platforms were, and do they really apply to the patients in front of them?” Ms. Burnell says. “We need to make sure that we always use our first principles of analyzing a situation and not totally rely on AI.”
At the centre of all of it, Dr. Singh says, are the patients and families the technology is built to serve.
“We’ve just scratched the surface of how AI is going to impact patient care at the hospital,” he says. “Our patients and families are going to grow it alongside us as partners and collaborators.”