Shaping Tomorrow's Healthcare: Canada’s AI Apps for 2025

Host 1: Welcome Deep Divers. So, artificial intelligence. It's not some sci-fi dream anymore, is it? It's really here and it's, uh, changing things fast.

Host 2: Oh, absolutely. And maybe nowhere more profoundly or, you know, more critically than in healthcare.

Host 1: Right. So today we want to cut through some of the hype, some of the noise, and get right to what matters. We're diving into a really key document.

Host 2: We are. This is our deep dive mission today. Unpacking the 2025 AI Watch List: Artificial Intelligence in Health Care. It's a horizon scan report from Canada's Drug Agency, the CDA-AMC. Basically, a look at what's coming down the pipe.

Host 1: And this isn't just like one person's opinion, which I think is really important.

Host 2: No, not at all. This watch list came out of a big consensus workshop back in November 2024.

Host 1: So who was there? Like a real mix of people?

Host 2: Yeah, a really diverse group from all across Canada. You had patients, caregivers, policy folks, researchers, people from industry, doctors, nurses, all in one room collaborating.

Host 1: And what was their main goal? Just looking ahead?

Host 2: Pretty much. They wanted to pinpoint the AI technologies and, uh, the big issues around them that are likely to shape Canadian healthcare in the next, say, 5 years.

Host 1: Kind of like a road map then for what's next and what we really need to pay attention to if we want to get it right.

Host 2: Exactly. Responsible integration. That's the key phrase.

Host 1: Okay, so what are we covering for you today? We're hitting two main areas, right?

Host 2: That's right. First we'll look at the top five emerging AI technologies. The stuff with huge potential for making things more efficient, improving patient outcomes, maybe even just making the whole experience better.

Host 1: And then the flip side.

Host 2: And then yeah, the flip side. The top five critical issues. We're talking legal, ethical, environmental, social stuff. The challenges we actually have to grapple with for this to work well.

Host 1: So this is all about giving you the really important nuggets, the key takeaways so you feel properly informed about where things are headed.

Host 2: Let's get into it.

Host 1: Okay, let's dive right in with what's actually, you know, here or really close. The watch list highlights five technologies that could be real game changers.

Host 2: Right. And first up is AI for note-taking.

Host 1: Ah, okay. The admin side of things.

Host 2: Yeah. I mean, think about how much time doctors and other providers spend just writing notes, documenting everything. It's kind of excessive sometimes.

Host 1: It's a massive time sink, isn't it? And a huge factor in burnout. I hear that all the time. Plus maybe less time actually talking to the patient.

Host 2: Exactly, that decreased eye contact, potential communication barriers. It's a real thing.

Host 1: So how does AI actually like step in to help here?

Host 2: Well, these are specific AI applications. They use automatic speech recognition, so listening to the conversation and natural language processing, NLP, which is basically teaching computers to understand language, to transcribe the whole patient-provider chat and then generate draft clinical notes.

Host 1: So like an AI scribe, an assistant?

Host 2: Pretty much. The healthcare provider still has to review it, edit if needed, and sign off. The human is still in control.

Host 1: Gotcha. Are there examples? Is this actually happening now?

Host 2: Oh, yeah. There are tools like Tali AI, PhenoPad, they're out there. And there was a pilot study. Um, it showed these AI scribes cut administrative time by almost 70% in labs.

Host 1: Wow, 70%?

Host 2: Yeah. And in routine practice in Ontario, it saved doctors about 3 hours a week on average. Plus patients actually said they had more meaningful interactions.

Host 1: That sounds like a win-win. Reduced admin, fewer errors maybe, better interactions, happier doctors hopefully.

Host 2: Yeah, the potential upsides are huge. Reduced burnout is a big one.

Host 1: But there's always a but, isn't there? What's the catch?

Host 2: Well, the workshop folks while really enthusiastic, they pointed out these AI scribes are, quote, "imperfect."

Host 1: Meaning they make mistakes?

Host 2: They can. Errors or sometimes these weird AI hallucinations where they just make stuff up or distort information. And they might struggle with different languages or accents.

Host 1: Okay, so that human review step is absolutely critical.

Host 2: Non-negotiable. Plus, getting these tools to work smoothly with existing electronic health records, the EHRs, that's still a, quote, "challenging endeavor." It's not plug-and-play.

Host 1: Right. But maybe seen as a safer starting point for AI compared to say, diagnosis.

Host 2: Exactly. That was the feeling. A good place for sort of safer, early adoption.

Host 1: Okay, what's next on the technology list?

Host 2: Number two is AI tools to accelerate and optimize clinical training and education.

Host 1: Ah, interesting. Moving beyond just memorizing textbooks.

Host 2: Totally. Traditional medical education, yeah, lots of memorization. AI could potentially revolutionize that with more personalized learning, real-time feedback.

Host 1: So what do these tools actually do?

Host 2: They can summarize massive amounts of medical evidence, give background info, pull up the latest research, basically helping doctors, students, even patients stay up to date. It supports, uh, upskilling and reskilling.

Host 1: I saw some examples like Open Evidence, a language model built just for medicine.

Host 2: That's right. It gathers and synthesizes clinical evidence, and crucially, gives you the citations so you can check its work. Apparently it did well on US medical licensing exams.

Host 1: Impressive. And things like ChatGPT, can they play a role?

Host 2: They can. You can use them to create virtual patient scenarios for practice, make quizzes, even get feedback on simulated doctor-patient chats. But, big warning again about those AI hallucinations, they can sometimes just fabricate references.

Host 1: So gotta be careful. Verify, verify, verify.

Host 2: Always. There's also something called AI VSP, artificial intelligence virtual simulated patients. Sounds pretty cool. Creates more immersive interactive learning.

Host 1: The benefits seem pretty clear here too. Better learning, faster training, less research grunt work, hopefully leading to better patient care overall.

Host 2: Definitely. And it could help with the healthcare resource crisis too, finding innovative ways to train people.

Host 1: What really struck me here was the dual role of AI you mentioned.

Host 2: Yeah, that's key. AI is both a subject to learn about, providers need to understand how it works, its limits, how to critique it, and it's a tool for learning.

Host 1: And they also mentioned educating patients.

Host 2: Right. Helping patients understand these tools, too, so they can be more proactive in managing their own health. Makes sense.

Host 1: Okay, shifting gears now. This one feels a bit more futuristic, maybe: AI for disease detection and diagnosis.

Host 2: Yeah, this is a big one. Getting that early, accurate diagnosis is obviously critical for treatment. But it's hard work for doctors, really cognitively challenging.

Host 1: And the US FDA has already approved quite a few AI medical devices for this, right? Like hundreds.

Host 2: Almost a thousand, yeah. Mostly in this detection and diagnosis space. So this is about using AI, particularly machine learning, to help doctors spot diseases.

Host 1: Uh, what kind of data does it use?

Host 2: All sorts. Medical images are huge, especially in radiology, which is really leading the way here. But also things like physical exam results, family history. AI models learn patterns from this data.

Host 1: I remember reading about assist TBI, that helps find traumatic brain injuries on CT scans really fast.

Host 2: Right. With high accuracy, over 80%. The idea is it could potentially speed things up, maybe even let doctors bypass waiting for a radiologist review in urgent cases.

Host 1: And Luminaero, detecting things like Alzheimer's early using eye scans. That sounds incredible.

Host 2: Yeah, using low-cost, non-invasive imaging to look for protein biomarkers in the retina. Truly cutting-edge stuff.

Host 1: So the potential positives: better accuracy, maybe more accessible diagnostics, finding diseases earlier, maybe cutting wait times.

Host 2: Huge potential, definitely.

Host 1: But the potential downsides or maybe unintended consequences?

Host 2: Well, the workshop participants flagged a concern. AI might actually increase the demand for diagnostic tests.

Host 1: Oh, interesting. So it finds more things leading to more tests, potentially straining the system.

Host 2: Possibly. So you've got to balance finding things early with the risk of overdiagnosis, finding things that might never have caused a problem, and misdiagnosis, like false positives or negatives.

Host 1: Tricky balance. And what about using things like ChatGPT for diagnosis?

Host 2: Yeah, the report was clear on that. While these large language models can be surprisingly good at diagnostic reasoning, they absolutely should not be used for autonomous diagnosis without physician oversight. Human judgment is still essential.

Host 1: Makes sense. They also mentioned something called opportunistic screening.

Host 2: Right, using data that's already being collected for another reason. Like if you get a routine CT scan, maybe AI could also screen that scan for early signs of pancreatic cancer incidentally, without needing a separate test. Could be a way to reduce the screening burden.

Host 1: Okay, let's move on to number four: AI for disease treatment. Once you have the diagnosis.

Host 2: Right, then it's about choosing the best treatment. That's core clinical decision-making, right? Weighing the evidence, the patient's own needs and values, the costs.

Host 1: And AI can help with that selection process?

Host 2: Yeah, it offers new ways to get patients more optimal, personalized treatment working alongside traditional care.

Host 1: How so? Like suggesting medications?

Host 2: It could involve identifying the best treatment plan medication, dosage, checking for drug interactions, maybe even considering someone's genetic profile, and potentially updating that plan as new info comes in.

Host 1: Okay.

Host 2: It can also help with things like triage in the ER, sorting patients faster, assessing risk better for earlier intervention.

Host 1: It sounds like that shift from one-size-fits-all medicine to something much more tailored.

Host 2: Exactly. That personalization is a key theme.

Host 1: I saw examples like Kaia Health, using AI for managing pain or COPD at home. And Wysa, that AI mental health chatbot.

Host 2: Yep, digital therapeutics. Kaia uses machine learning for self-management programs. Wysa offers 24/7 mental health support, guiding people through things like cognitive behavioral therapy, CBT.

Host 1: And even in drug discovery, like Valence Labs.

Host 2: Right. Using large language models to help come up with hypotheses for new drugs and design experiments, so it spans from patient care right back to the lab.

Host 1: The potential impact seems huge again. Efficiency, sustainability, better outcomes, maybe helping with the shortage of healthcare workers.

Host 2: All of that. Enhanced satisfaction for patients and providers too, potentially.

Host 1: But the workshop folks mentioned this is hard to implement. Requires significant resources.

Host 2: It does. You need money obviously, big financial investment. You need training for all the healthcare professionals, ongoing tech support.

Host 1: And just whether the whole system is ready for it.

Host 2: Yeah, system readiness was a big concern. Funding differences between regions, and how this ties into those other big issues we're going to talk about like privacy and data. It's complicated to integrate properly.

Host 1: Okay. Final technology: AI for remote monitoring. This uses sensors outside the hospital, right? Like wearables.

Host 2: Exactly. Biomedical sensors collecting health data while you're at home or just going about your day. Then AI steps in to make sense of all that data looking for actionable insights.

Host 1: So AI analyzes data from my smartwatch or maybe sensors in my home?

Host 2: Could be wearables, smart home devices, even just your smartphone. It collects, analyzes, interprets this data in real time.

Host 1: And what does it do with it? Send alerts?

Host 2: Yeah, it can generate alerts if something looks off, prompting timely interventions. It can also try to predict potential health issues before they become serious. And the algorithms learn and improve over time.

Host 1: That sounds incredibly useful, especially for managing chronic conditions.

Host 2: Hugely. Think about AlayaCare. Their software apparently improved event predictions by 11% and cut overdiagnoses by 54%.

Host 1: Wow. And there was a Canadian study, too?

Host 2: Yeah, using AlayaCare for patients with COPD or heart failure. Over 3 months, they saw a 68% drop in ER visits and a 35% decrease in hospital stays.

Host 1: Those are significant numbers.

Host 2: Very significant. And even simpler things like Hyfe, using AI sound analysis via a smartphone to monitor chronic coughs remotely.

Host 1: The positives seem obvious. More monitoring, better access, especially for remote areas, keeping people at home instead of the hospital, saving costs, freeing up staff time.

Host 2: Right. Better quality of life, too, maybe. More dynamic, responsive care.

Host 1: Especially crucial with, you know, our aging population and more chronic disease. But what are the hurdles here?

Host 2: Well, first, practical stuff. Limited internet access is still a reality in many remote or rural areas.

Host 1: Good point.

Host 2: And then there's technical literacy. Not everyone, especially maybe older adults, is comfortable using these devices or apps.

Host 1: True. And what about the accuracy of the data itself?

Host 2: That's a critical one. The report highlights studies showing, for example, that pulse oximeters, which measure blood oxygen, can be less accurate for people with darker skin tones.

Host 1: Oh, wow. That's a serious bias issue.

Host 2: It is, and it underlines why AI tools need rigorous testing across diverse populations to make sure they don't actually worsen health inequities. We need to be really careful there.

Host 1: Okay, so that's the tech. Amazing potential, but clearly lots to think about. Which brings us perfectly to the second half: the critical issues.

Host 2: Exactly. With great power comes great responsibility, right? So let's get into the top five issues the watch list flagged, starting with a really fundamental one.

Host 1: Which is?

Host 2: Privacy and data security.

Host 1: Ah, yeah. The big one. AI needs tons of data, and in healthcare, that data is incredibly sensitive personal health information.

Host 2: Precisely. So how do we keep it confidential? How do we keep it secure? These are massive questions.

Host 1: I saw a stat that really jumped out: a 2024 survey found 79% of Canadian doctors were not confident or unsure about AI and patient confidentiality.

Host 2: That says a lot, doesn't it? And Canada's track record isn't perfect, either, ranked 10th globally for cyber breaches with quite a few major attacks on health networks recently: ransomware, data theft.

Host 1: Scary stuff. And the rules around this, are they clear enough for AI?

Host 2: Well, Canada has laws like PIPEDA federally, and provincial rules, too. But here's a key point from the report: many publicly available AI technologies, think ChatGPT or maybe some of those AI scribes, if used directly by patients or clinicians, they're often not regulated in Canada.

Host 1: Meaning they might not follow Canadian privacy rules or even US rules like HIPAA?

Host 2: Potentially, yeah, especially concerning where the data actually gets stored. Patients are understandably worried about consent, whether their data is truly anonymized, who they even go to if something goes wrong.

Host 1: So what's the path forward? How do we build trust here?

Host 2: It involves a few things: proactively adopting strong privacy and security measures, encryption, anonymization techniques, training for everyone involved: doctors, staff, patients on how to use AI securely.

Host 1: And making sure tools actually comply with local laws.

Host 2: Definitely. And creating policies that empower patients, giving them control over their data, the ability to consent, to see their data, correct errors. Quebec's Bill 64 was mentioned as an example, giving patients rights like knowing how an AI decision was made and asking for a human review.

Host 1: Okay, crucial groundwork. What's the second major issue?

Host 2: Liability and accountability.

Host 1: Right. If an AI makes a mistake or contributes to harm, who's responsible? Who gets sued?

Host 2: Exactly. It's super complex, especially because the doctor using the AI often didn't build it, might not fully understand how it works, that black box problem, the opacity, the lack of explainability sometimes.

Host 1: Is the legal system ready for this? Does Canada have clear answers?

Host 2: Not really, no. The report says the legal system does not have clear-cut answers yet. There's Bill C-27, the proposed AI and Data Act, but it's still being debated.

Host 1: I saw that Air Canada example. They were held liable because their chatbot gave bad info.

Host 2: Yeah, a really interesting case outside healthcare, but it shows the direction things might go: who is ultimately accountable when an automated system messes up?

Host 1: And the workshop participants felt like putting all the responsibility on the doctor using the tool might not be fair.

Host 2: Right. They felt that burden might be inappropriate for end users who aren't AI experts themselves. It ties into that idea of needing a human in the loop, but even designing that interaction is complex. Who's really in charge?

Host 1: So how do we manage this liability risk?

Host 2: Solutions suggested include things like clear agreements upfront about responsibilities, strong disclosure requirements about AI use, better hospital policies for safe AI implementation, clear guidance for doctors on how to weigh AI recommendations, and pushing for more algorithm transparency so the systems are less of a black box.

Host 1: And Canada launched that AI Safety Institute.

Host 2: Yes, in November 2024. A $50 million initiative to support research and help manage these kinds of risks. It's a start.

Host 1: Okay. Issue number three: this feels connected to the first two.

Host 2: Very connected. It's data availability, quality, and bias.

Host 1: Right. The whole garbage in, garbage out problem. AI is only as good as the data it learns from.

Host 2: Precisely. You need lots of data, and it has to be accurate, reliable, complete, relevant. If the data is poor quality, or worse, biased, the AI can end up making unfair, discriminatory, or just plain wrong decisions.

Host 1: And bias can creep in in different ways.

Host 2: Yeah, they talked about data bias: maybe the data doesn't represent everyone fairly; algorithmic bias: errors in the AI's code or training; or even user bias: how people interpret the AI's output based on their own biases.

Host 1: The examples are worrying: chatbots repeating harmful, race-based medical myths?

Host 2: Yes. And systemic biases that already exist in healthcare, where certain groups, equity-deserving groups, have historically been left out or misrepresented in data can get baked right into the AI.

Host 1: Like the examples you mentioned earlier, algorithms making wrong assumptions about muscle mass based on race or pulse oximeters being less accurate for darker skin.

Host 2: Exactly. Or AI trained mostly on male heart attack symptoms potentially missing them in women. It highlights how crucial diverse, representative data is.

Host 1: And in Canada, just getting enough good data is a challenge in itself, because it's all fragmented.

Host 2: That's a major hurdle. Health data is scattered across different provinces, territories, health systems, hard to pull it together to train robust AI models. There are initiatives trying to fix this, like a Pan-Canadian Interoperability Roadmap and Bill C-72 aiming to stop data blocking and help information flow better.

Host 1: The workshop called this foundational, right? Something we need to get right from the start.

Host 2: Absolutely. They pointed out gaps in the data we currently collect: a lack of info on social factors, patient experiences, economic situations. And there's a debate: do we wait until the data is perfect or use AI carefully with the imperfect data we have now?

Host 1: And that black box issue makes it hard to even find the bias sometimes.

Host 2: Right. And clinical practice changes. The example given was the Ontario Renal Network saying we should stop using race in kidney function calculations, eGFR. If AI developers aren't kept up to date with these clinical shifts, their models could become outdated or even harmful.

Host 1: So tackling this requires better data sharing, better quality control.

Host 2: Yes, more pan-Canadian collaboration on data, legislation to improve quality, letting patients access and correct their own data, and specific steps to mitigate bias: diverse development teams, combining different data sets, standardizing data collection, testing AI in the real world, and monitoring it after deployment using fairness guidelines. It's a multipronged approach.

Host 1: Okay, that leads into issue number four: data sovereignty and governance.

Host 2: Yeah, this builds on the data discussion. It's about who actually owns and controls all this health data that AI uses.

Host 1: Sovereignty versus governance, what's the difference?

Host 2: Data sovereignty is basically the right of a specific group, like Indigenous Peoples, to control their own data according to their own principles. Data governance is the broader framework defining who has the authority and responsibility to manage data generally.

Host 1: And this is especially critical for Indigenous communities in Canada because of historical issues.

Host 2: Absolutely. Due to colonialism, their health data was often collected, controlled, sometimes even destroyed by the government without their consent or benefit. So now there are frameworks like the UN Declaration on the Rights of Indigenous Peoples, the CARE Principles, OCAP for First Nations—

Host 1: OCAP meaning ownership, control, access, and possession.

Host 2: Exactly. And similar principles like Inuit Qaujimajatuqangit, OCS for Métis, and EGAP for Black communities in Ontario. They all emphasize the right of these communities to govern their own data.

Host 1: So when developing AI, you can't just take the data. You need partnership.

Host 2: Genuine partnership. The report stresses working together with equity-deserving groups from the start so they control how their data is used and interpreted. The goal is to ensure AI actually helps reduce health inequities, not make them worse by repeating past mistakes.

Host 1: And respecting evolving language and ways of knowing, too.

Host 2: Right. Health understanding isn't static. AI needs to keep up with how communities describe their own health needs and experiences. The workshop saw this as absolutely vital for building trust and making sure AI aligns with our societal values.

Host 1: So the solution boils down to empowering patients and communities.

Host 2: Essentially, yes. Policies giving patients control, access, and interpretation rights, and participatory approaches where communities are co-designers, establishing control right from the beginning.

Host 1: All right. The final issue, this one might catch some people by surprise: environmental costs.

Host 2: Yeah, it's not always the first thing people think about with AI and healthcare, but it's significant. AI has an environmental footprint.

Host 1: How so? Energy use?

Host 2: That's a big part of it. Training complex AI models and running the data centers they live in consumes a lot of energy. And those data centers also need huge amounts of water for cooling.

Host 1: Plus the hardware itself, right? Mining rare earth metals, dealing with e-waste.

Host 2: Exactly. Making the ships, the servers, it it all has an impact. And then disposal of outdated electronic devices creates e-waste, often ending up in landfills in poorer countries.

Host 1: The stats were pretty eye-opening: over 239 data centers in Canada, Ontario's electricity demand potentially jumping 75% by 2050, partly due to AI.

Host 2: Yeah, and big tech companies like Microsoft, Google, Amazon are looking at nuclear power just to meet AI's energy needs. That tells you something about the scale.

Host 1: The workshop participants felt this was often overlooked, an afterthought.

Host 2: They did. They felt there needs to be more advocacy around this, mentioning the recent devastating forest fires in Canada, flooding in Europe. It brings the urgency of climate change home. AI can be used to optimize energy use in some ways, but its overall net impact is a major concern.

Host 1: It really forces a broader view of health, doesn't it? Including planetary health.

Host 2: Absolutely. The core question is how we balance the potential patient benefits of AI against these potential environmental harms. We need sustainable AI.

Host 1: So what can be done? Greener AI?

Host 2: Things like building more energy-efficient AI models, creating green computing systems, doing regular life cycle assessments to design hardware more sustainably, eco-design, managing data better, reducing waste, using renewable energy sources, handling hardware disposal responsibly, and just raising awareness among everyone involved.

Host 1: Wow. Okay, that's a lot to take in. We've covered these five really promising technologies: note-taking, training, detection, treatment, remote monitoring, and then these five really critical, interconnected issues: privacy, liability, data quality and bias, sovereignty, and the environment.

Host 2: It really paints a picture, doesn't it? The watch list makes it clear these aren't separate silos. You need systems-level thinking to navigate all this. It's all linked.

Host 1: And the overall feeling from that workshop was that AI in healthcare, at least some forms of it, is basically inevitable.

Host 2: Yeah, that was the strong sense, especially the consumer-facing stuff. Whether the healthcare system is fully ready or not, these technologies are coming. They're arriving.

Host 1: So the call to action is really about getting ready, preparing the system.

Host 2: Exactly. The report calls for system readiness. That means getting those foundational pieces in place: strong governance, clear privacy rules, sorting out liability to build public trust, and allow for responsible adoption down the line. It's about building a solid foundation.

Host 1: It's a landscape that's changing incredibly fast. And it really makes you think, doesn't it? How do you, listening right now, see yourself contributing to this future or preparing for it? That future where AI and human judgment have to work together for the best health outcomes.

Host 2: Yeah, what responsibility do we all share in making sure this powerful technology leads to a healthcare system that's not just efficient, but also ethical and equitable for everyone?

Host 1: Definitely a lot to mull over. This is really just the start of the conversation.