AI in healthcare: It all depends on the right model

Ask complex questions, get quick answers – that's how many people use AI-based chatbots today. But these bots have weaknesses; they are known to invent or distort information. Can programs like these still be used in healthcare at some point?

In an interview with MEDICA.de, Prof. Paul Lukowicz explains the impact that the AI models behind the bots (large language models or foundation models) have on healthcare and why they are not a miracle cure but can still be useful.

Prof. Lukowicz, what trends are you currently observing in the areas of AI and big data in healthcare?

Prof. Paul Lukowicz: I think, as in all areas of AI, the hot topic right now is foundation models. These were originally called large language models, i.e., models such as ChatGPT. They have the effect that you no longer have to train a system extensively for each specific application.

These are pre-trained systems that you feed some additional information into, and they are suddenly able to deliver solutions. But there is a problem: how do I ensure that these solutions actually make sense and don't just look like they do?

Why is this particularly exciting right now?

Lukowicz: These models have given a large number of companies and applications access to high-performance AI. The effort required to train such a system, i.e., access to data, computing, infrastructure, and expertise, previously required dedicated AI departments. Now, even small businesses can gain access to such models.

And that's the problem: it sometimes creates the illusion that you can use these models without in-depth AI expertise. This can have negative consequences because the product may not work as you imagine it will in the end.

Another trend is that these models will change how we use wearables. Wearables don't understand our lives today; they don't tell us, "This is what you need to do to stay healthy." That works in the lab, but not in reality. Language models could change that in the future. However, this will play a greater role in wellness and personal lifestyle than in healthcare.

In your opinion, what do research and industry need to do to ensure that patients and medical professionals trust AI and accept its use?

Lukowicz: I think you have to communicate clearly what such a model can and cannot do. And users must, of course, also question what kind of model they are using and where the company that operates this model is based, because this is about the protection of health data.

In Germany, on the one hand, we have a tendency to emphasize data protection in many areas to such an extent that it also restricts us. On the other hand, there is undeniably the idea that someone is sitting there collecting my data, and I don't know who that is or what for. And when it comes to the most intimate questions about health, this is highly problematic for users. So, this trust has to be established, and that also means that I have to say what is possible and what is not.

How do you think we will be using AI in healthcare in five years?

Lukowicz: I believe that AI will change almost all areas of healthcare. It is always difficult to say whether and to what extent this will be implemented in five years. Five years ago, ChatGPT didn't even exist. This shows that a lot is happening.

Overall, we will be able to work more efficiently in healthcare. Take, for example, the issue of healthcare in rural areas, where many places do not have the care they should have. Of course, AI will not replace doctors here, nor should it. But if, for example, there is a local general practitioner, then that person does not have specialist knowledge in many areas. Here, AI will offer the possibility of accessing much more information. The doctor will then be able to evaluate this information and do much more with it than before, without having to send people directly to a specialist practice.

The same is conceivable for smaller hospitals, which can use AI to make better diagnoses, comparable to the top tier of university hospitals. AI tools are also ideal for administration and documentation, areas in which hospitals consume a great deal of resources.