I think oncologists need to know a little bit of AI and statistics to be capable of critically judging the tools they are going to use. In fact, many of the oncologists that are now using some techniques, they do it within research projects, so they are probably closer to the field than other potential end users in the future. So they should evaluate all the tools with some perspective and distance, you know, because the end user is going to be probably a very naive AI person who is skilled in medicine...
I think oncologists need to know a little bit of AI and statistics to be capable of critically judging the tools they are going to use. In fact, many of the oncologists that are now using some techniques, they do it within research projects, so they are probably closer to the field than other potential end users in the future. So they should evaluate all the tools with some perspective and distance, you know, because the end user is going to be probably a very naive AI person who is skilled in medicine. That means that they need to understand the biases of the cohorts with which the model was trained, et cetera. I also believe we pay too much attention to model performance in controlled scenarios. And more and more evidence is gathering, indicating that many AI tools that look promising in controlled scenarios are actually performing very poorly in real scenarios. And this has a lot to do with the real use case, where you have patients in the clinics moving around, people are interrupting you, maybe the data you feed to the system is not the best one, etc. So this is one point. Then the interaction with the health system and the objectives with which AI is implemented is probably also unaligned with reality. There are some examples in emergency units where they have implemented AI-assisted tools to manage or support patient care which makes a lot of sense but their expectation was to lower the time in the emergency department and they didn’t achieve that so this means that probably AI is another layer of data management that we have to implement in healthcare, but probably it is not sufficient per se. It is a tool inside a context, inside an environment. And therefore, lots of things that go well beyond the statistics and the AI tool have a big impact on the capacity that these tools will have to improve management of our patients. Which is, by the way, pretty much the same as for standard drugs – you know that we introduce in the clinics – you know, depending on which patient you select, depending on which situation that patient is in, and even in which place that patient is being treated, you may prefer one treatment or the other. So for technology, the differences are probably related to all those factors.
This transcript is AI-generated. While we strive for accuracy, please verify this copy with the video.
