“We have reached a point where AI is teaching us”

On Friday, 18 October, the annual Olof Rudbeck Day will take place, with the aim of disseminating information about current medical research on a specific theme. This year’s topic is ‘AI for health and care – opportunities and risks.’

“This subject engages me deeply. There are significant challenges in healthcare today, and AI comes with a promise to solve many of them. That said, it is not a solution to be taken lightly; we must collectively understand the most likely paths for AI in healthcare in the short and long term,” says Johan Sundström, cardiologist and Professor of Epidemiology, and one of the organisers of the programme.

Johan and his research group are running several projects involving AI as a tool in healthcare, and he is part of the interdisciplinary initiative AI4Research. He looks forward to a busy event in October.

“I think it’s fantastic to enjoy all of this in one day. There are so many aspects of AI, both good and bad, that need to be discussed, and this is a perfect forum.”

Finding patterns in ECG

AI models and algorithms can be essential tools in many areas of healthcare, but perhaps less suitable in others. In image interpretation, which is the focus of one of Johan Sundström’s research projects, significant progress has already been made.

“There is a lot of research going on in AI for image interpretation and pattern recognition, and it is primarily in this area that I see the greatest potential right now,” says Johan.

When an AI model is trained on a large number of human interpretations of ECGs, the AI will eventually become as skilled as a human in interpreting them, reaching a ‘human’ level of expertise. Johan Sundström’s research group has taken this a step further.

“In addition, we have trained the model to recognise other patterns by allowing the AI to learn from patient outcomes, whose ECGs were previously interpreted. For instance, one patient needed a coronary angioplasty after their acute myocardial infarction, while another did not. Ultimately, the AI model learns to recognise patterns regarding which interventions a patient might need. This means it can do things that we humans cannot. Suddenly, it seems that the AI is ‘superhuman’, and it starts teaching us how it detected the patterns by highlighting where on the ECG curve it found something.”

AI takes shortcuts

Does that sound astounding? It is. But it does not mean it is a completely reliable system. When AI models predict patient outcomes, they sometimes take shortcuts.

“In other research projects, models have interpreted factors other than the disease itself. For example, a model for chest X-rays interpreted that an image came from the intensive care unit’s portable X-ray machine. Statistically, this means a higher likelihood of certain severe diseases than if an image comes from the regular radiology department. We need to be aware of such things,” notes Johan Sundström.

How can a patient trust an interpretation by a machine when this is introduced into healthcare for real?

“That is a very good question, and I almost want to say ‘you’ll have to wait and hear the answer during Olof Rudbeck Day,’ Johan laughs and continues:

“But to give a little teaser, AI algorithms need to be tested as rigorously as anything else being introduced into healthcare, including randomised clinical trials. If it withstands that, it can be integrated into healthcare for real. Another challenge is how it should be used and the extent of human involvement. I believe it is important to have a ‘human in the loop’, as we call it. Healthcare personnel should always be present and make active decisions. Just because the AI always seems to be right, does not mean we should step back and leave decision making to it.”

An important question about discrimination

The issue of responsibility for AI and decisions in healthcare is crucial. Another ethical issue, which will also be raised during Olof Rudbeck Day, concerns discrimination. The patients for whom there is the most data will also be the patients for whom the AI model makes the most accurate diagnoses.

“As part of this, we have a project in the emergency department where patients are asked to draw their pain – where it is located and how it feels. This creates an image that an AI then interprets, as AIs are good at image interpretation. So, even though abdominal pain can mean several different diagnoses, the way it is drawn might help the AI determine that it is, for example, kidney stones.”

This, along with having a generative AI model that reads patient histories and a model that interprets ECGs, creates a whole that reduces the risk of discrimination by providing the model with various types of data that can complement each other.

“At least that is our hope and our attempt to ensure that more people can benefit from AI in healthcare, regardless of who you are and where you come from. Drawing your pain works well even if Swedish is not your first language, and if you cannot draw for some reason, there are other ways to cover that,” says Johan.

More about the opportunities, but also the challenges, of AI in healthcare will be discussed during Olof Rudbeck Day on 18 October in Grönwallsalen at Uppsala University Hospital. During the event, this year’s Olof Rudbeck Prize will also be awarded by Upsala Läkareförening, which co-organises Olof Rudbeck Day with the Disciplinary Domain of Medicine and Pharmacy at Uppsala University and Uppsala University Hospital.

Robin Widing