ISC 2026 | AI in stroke care: risks and strategies for implementation

At the moment, there aren’t a lot of areas where they are integrated, so we have to be careful. The obvious risks are abnormal results, so either false positive, false negative. So you either say something that isn’t true, that’s very common for a general large language model, hallucination, for instance, or false negative, you miss something. So at the moment, there are specialized tools for certain areas I’ll come to, in particular for image analysis, as we just discussed. The next thing would be the risk of privacy disclosure. And that’s a very big risk. There are tools that can prevent that. But unfortunately, they’re not really widely implemented, because in many ways, the regulatory bodies are still catching up with many of the tools that are being introduced into the marketplace. For instance, I think people should know both providers and patients should know, that small startup companies, direct-to-consumer products, are often not covered by medical privacy regulations. That’s certainly true in the US. It’s less true in Europe because of GDPR or UK GDPR in Britain. But generally, we have to be careful that privacy may not be maintained. That is really true for large language models. So then what kind of, so that covers the two main risks, bad results and privacy. What about integration? So there are some areas where that’s been done really, really well, particularly in enclosed healthcare systems. So I think the NHS in the UK is working on systems. I don’t think they’re fully integrated yet, but I think that’s a good example of where you have essentially control of the data. It’s to some extent controlled in a separate ecosystem. And you can limit access by external companies or agents and maintain quality. So the results are right and privacy. Another good example of that in the US would be the Mayo Clinic Health System, who pioneered a lot of the AI work, starting with cardiology atrial fibrillation detection, but extending more broadly, including work on stroke. I think it might be difficult in the general marketplace. And I would say that if people are considering using these tools, which of course we all are, then we really have to be careful that this data is not widely available and that it’s accurate. So again, adjunctive use, be wary of privacy.

This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.