Universitetsavisen
Nørregade 10
1165 København K
Tlf: 35 32 28 98 (mon-thurs)
E-mail: uni-avis@adm.ku.dk
—
Science
Artificial intelligence — A new AI model with contributions from the University of Copenhagen can predict the onset of illness across entire populations. But it also raises ethical questions about data security — and about people’s right not to know which diseases might affect them in the future.
Doctors traditionally rely on textbooks, experience, and maybe a bit of intuition to make a diagnosis.
But now they may have a new additional tool for making clinical decisions when facing a patient: A kind of high-tech prediction model.
More precisely, it’s an AI model that is capable of analysing millions of patient cases and pointing to the most likely diseases in their future. The model has been developed by a team of international researchers with participation from the University of Copenhagen (UCPH), and the study describing it has just been published in the scientific journal Nature.
One of the driving forces behind the project is Søren Brunak, a professor at the Department of Public Health, who has spent decades combining biomedicine with vast amounts of health data. For him, it’s not just about technology — it’s about changing how we understand the development of disease through the course of a life. The model can’t predict everything, but it can outline probable patterns in the diseases that we are likely to develop — and in what order.
Søren Brunak explains it with reference to the words of Danish 19th century philosopher Søren Kierkegaard. Life can only be understood backwards, but it must be lived forwards.
»What we’re trying to do here is challenge that idea. To predict something before we’ve lived it,« he says.
In its current form, the model is best at predicting disease trends across a population as a whole. It has been trained on data from 400,000 people from the UK in the UK Biobank, where researchers have tracked participants’ health and lifestyle over several decades.
For most research projects, 400,000 patients is a massive dataset, but in Søren Brunak’s world it’s relatively small — he’s used to working with populations of millions.
Even so, the results proved surprisingly robust. When the model was tested on the health histories of 1.9 million Danes, it was able to identify patterns and make projections that gave an accurate picture of public health.
Uncertainties emerge when it comes to individuals however.
»It’s not yet good enough to predict the next illness in a single person, but it is good enough to forecast how diseases will develop across an entire country,« says Søren Brunak.
If it were trained on data from 50 million people, I could easily imagine it being, on average, better than a doctor in five years’ time
Søren Brunak, professor at the Department of Public Health
The model can, in other words, estimate how many Danes will have diabetes or heart disease in ten years’ time — but not estimate with any degree of certainty whether a particular, individual, Dane will be among them. That needs inputs that are not yet available to the algorithm — all the things the doctor sees, and senses, in the doctor’s office, but which are not recorded in the medical record.
Some of the limitations are specific: The order in which diseases occur matters, but this is hard to capture in individual cases, and certain conditions — like pregnancy complications — are underrepresented, as the UK Biobank primarily contains data from older people.
A model, on the other hand, can learn from millions of patterns. Even the most experienced doctors would never encounter this number in a single career. Where a family doctor might only see a rare type of cancer once in a lifetime, an algorithm can identify it across thousands of patient cases and link it to factors that would otherwise go unnoticed. In this way, the machine can act as a supplement — expanding the doctor’s perspective and reducing the chance of mistakes.
And this is exactly why Søren Brunak believes that once the model is trained on much larger datasets, it could surpass humans in making diagnoses.
»If it were trained on data from 50 million people, I could easily imagine it being, on average, better than a doctor in five years’ time — even at diagnosing individual patients,« he says.
The reason a model like this can even function is the vast amounts of health data that is available to researchers. In Denmark, the National Patient Register is particularly valuable, as it covers the entire population and goes back nearly 50 years. This gives researchers a unique view of disease development over time — something that few other countries can match. But the data isn’t always complete, and important details from general practitioners or specialists are often missing.
This is why Søren Brunak sees great potential in the EU’s new initiative, the European Health Data Space. Here, member states will gather and standardise their healthcare data, so it can be used in both research and clinical settings.
This would make it possible to study rare diseases — of which only a few cases exist in a small population like Denmark — throughout Europe. Or to provide better treatment for immigrants who represent too small a group in one country, because the data from their countries of origin can now be included in the training: Our genes differ, and some population groups therefore show different disease patterns.
It’s a kind of decision support for the doctor — rather than a worry machine designed to ruin your life
Søren Brunak, professor at the Department of Public Health
The idea of the European Health Data Space is also practical: If a Dane, for example, breaks a leg in Barcelona, the doctors there should be able to access vital information — from chronic conditions to ongoing treatments.
Data security is an absolute necessity, however. Søren Brunak emphasises that researchers never move raw data between countries.
Instead, the models are trained locally and then transferred in a form where patient information cannot be extracted.
»We’ve locked up the data, but we’ve also locked up the model itself,« as he puts it.
He points out that trust is key. If citizens aren’t confident that their health data is being handled responsibly, the whole foundation collapses. That’s why the work is carried out in secure supercomputer environments without internet access — so that neither researchers nor hackers can gain access to the data.
The ability to predict diseases decades into the future inevitably raises a series of ethical questions. For while the technology can be a powerful tool in healthcare, it’s far from certain that patients themselves benefit from knowing their risk of future diagnoses.
Søren Brunak emphasises that the model is first and foremost a tool for healthcare professionals who need to make fast, well-informed decisions — not a system designed to hand out grim predictions directly to the public.
»It’s a kind of decision support for the doctor — rather than a worry machine designed to ruin your life,« he says.
But there is a clear dilemma: The earlier a disease is detected, the greater the chance of effective treatment. At the same time, numbers on probabilities and risk are a burden if otherwise healthy people are left to deal with them.
Should a 30-year-old woman, for example, be told she has an elevated risk of breast cancer in 20 years time — even if the disease never develops? Or should a young man be told that his genes and lifestyle make it likely he’ll get type 2 diabetes in his 50s?
According to Søren Brunak, studies show that many people actually want to know their risk — no matter how uncomfortable it might be. For some, it provides the chance to act in time and change their lifestyle. For others, it’s a way to prepare or protect their family. So the real question isn’t just whether the technology can predict disease, but also who should have access to the results — and how they should be communicated.
Søren Brunak himself stresses that choice is essential. Patients should be able to say no to having their future laid out in probabilities, but doctors should be able to use the tool when it can clearly make the difference between overtreatment, undertreatment — or the right intervention at the right time.
With these reservations, Søren Brunak is convinced that changes are swiftly unfolding. Today, the model is primarily a research tool, but as it’s fed datasets that are much larger, he sees a future where artificial intelligence becomes an indispensable part of daily healthcare.
Already now, projects are being planned to train models on 40 to 50 million patient cases. And with the upcoming European Health Data Space, the number could reach several hundred million. According to Søren Brunak, this will mean that many of the blind spots currently limiting the model’s accuracy will gradually disappear.
The high-tech prediction model for healthcare is still only a prototype. But in spite of multiple ethical dilemmas, Søren Brunak believes that AI models like this will, in a few years time, become widespread in the interaction between doctor and patient.
They will challenge Kierkegaard’s dictum: We will not only understand life backwards — we will also be able to interpret what’s ahead, looking forwards.
This article was first written in Danish and published on 2 October 2025. It has been translated into English and post-edited by Mike Young.