Prediction models for organ transplants

At the end of 2024, the number of people in Germany on the waiting list for an organ transplant exceeded 8,500, of whom 6,300 were hoping for a donor kidney. There is an acute shortage – every heart and every kidney is a precious and potentially life-saving good. At Darmstadt University of Applied Sciences (h_da), an interdisciplinary team of researchers from mathematics and computer science is working on AI models to gauge the chances of success of such transplants. The aim is to use these models to assist doctors. The Federal Ministry of Research, Technology and Space has allocated project funds of €265,000 for three years.

By Christina Janssen, 8.12.2025

Let’s take University Medical Centre Mainz as an example: in March 2025, there were 200 patients on its waiting list, as was reported in the news, but each year only around 50 kidney transplants are possible. And the waiting lists are growing longer and longer. But at least the success rate is high, especially for kidney transplants. Of 100 transplanted organs, around 90 percent still function one year after the operation. After five years, at 75 percent the figure is still high.

A team of researchers at h_da is analysing the complex relationships between patient, organ and the chances of a successful transplant. In a research project with the German Organ Procurement Organisation (DSO), the pharmaceutical company Merck and the software company Accso, mathematician Antje Jahn and computer scientist Gunter Grieser are taking a closer look at data from the National Transplant Register, which was launched just a few years ago. They want to use this data to develop an AI tool that visualises these relationships in a transparent and traceable way.

Jahn and Grieser, who both teach data science students at h_da, have been waiting for this opportunity for a long time. “The National Transplant Register was only introduced in 2021,” explains Professor Antje Jahn, which is why the team’s precursor project had to rely on support from the USA (impact article “Well-ordered chaos”). “Now, all the relevant data is being stored for the first time in one place in Germany too. We want to know exactly what information these datasets contain and what it can be used for.”

When data science experts like Jahn and Grieser talk about “data”, they mean more than just a patient file with details of age, gender, weight, blood group or pre-existing conditions. “For every transplant patient, there are tables with hundreds of columns,” explains Jahn. “Countless values are documented – creatinine, urea, tissue type, results of liver tests, and so on.” The first step is to cleanse the data. “These parameters are often recorded manually in hospitals and doctors’ surgeries, which makes them susceptible to error,” explains Professor Gunter Grieser. If the columns for height and weight are switched, for example, the result could be a person who is 90 centimetres tall and weighs 160 kilograms. This sounds funny, but it is a genuine problem. Researchers use automated plausibility checks to filter out such outliers: a Sisyphean task – before the real work can even start.

The h_da team was already able to solve another problem with the data from the transplant register: “Patients usually go for a follow-up once a year after their transplant,” says Professor Jahn, describing the procedure. “This means that certain parameters are only monitored on an annual basis, although they change continuously. Or the data from a patient’s follow-up appointments is missing altogether. “Perhaps they have moved abroad and are continuing their treatment there?” suggests computer scientist Grieser. “This leads to a distorted picture.” For such cases, the team has developed a statistical method that takes the uncertainty factor into account. “This is one of the scientific results we have achieved so far,” says Jahn, summing up.

The second interim result from the project is an “explorer”: “I can use it to zoom in on the data and filter it according to various parameters,” explains Grieser, who specialises in theoretical computer science and artificial intelligence. “For example, you can examine the progression in women between 20 and 40 years of age with creatinine levels between x and y. I can then compare all this data with that of men of the same age and produce a graph.” In addition, the explorer can predict survival probability and transplant tolerance – and it delivers the reasons for its prediction at the same time. A prototype for the planned support tool.

Thanks to the team’s close collaboration with the German Organ Procurement Organisation (DSO), it has its finger firmly on the pulse vis-à-vis current questions in medical practice, as Grieser continues. This is also how the team encountered a problem that data science students at h_da had worked on extensively in the previous semester: “Not all donated kidneys are actually transplanted,” explains biostatistician Jahn. This phenomenon is known as “organ discard” and describes the situation where a suitable patient cannot always be found quickly enough within the crucial time window after an organ has been donated.

“There are cases where patients turn down an organ because, for example, they themselves are still very young, but the kidney comes from an older donor.” This means that the first hospital offered the organ has only a few hours to decide, after which it is the next hospital’s turn. In a spin-off from the main project, students supervised by Jahn and Grieser have developed interactive diagrams on this discard problem: “You can enter organ data into them and gauge the probability of whether a kidney will be accepted or declined,” says Grieser.

A first test case for the students’ model: the director of a transplant centre reported during a project meeting on a patient in intensive care. The medical team was trying, he said, to lower the creatinine level because experience shows that this improves the chances of a successful transplant. To test the students’ model, the doctor entered the patient’s data into it and adjusted the creatinine value: the probability of organ acceptance that the model had predicted promptly changed. The conclusion: “This approach is worth pursuing.”

Jahn and Grieser found one of the findings produced in the project with their students particularly astonishing: it emerged that predictions are not necessarily better the more complex an AI system is. “The deep learning methods on everyone’s lips right now did not perform any better in our student project than a relatively simple one,” reports Antje Jahn. She says this is due to the comparatively small amount of data. “We only have a few thousand datasets here, and not billions.”

The advantage of less complex systems is obvious: the simpler the model, the easier it is to explain the results. “In a critical health situation, you want not only a prediction but also to know how it came about and whether there are factors that you can influence yourself,” says Antje Jahn. That is why explainability is a key factor, she says, when it comes to the acceptance of new methods. “Tremendous progress has been made in this area in recent years, which we are leveraging for our model.”

Both researchers emphasise that the support tool is still at the development stage and that there is a long way to go before it can be applied in practice. “No medical professionals have examined it in detail yet,” Grieser makes clear. “The system probably still contains some errors, but it works in principle.” To progress further in the direction of practical application, the researchers plan to bring medical experts on board in a follow-up project. “This is something our project typifies: interdisciplinary work is indispensable when tackling such questions. I am convinced that machine learning is impossible if you do not integrate expert know-how,” concludes Professor Jahn. Even if many perhaps assume otherwise in times of ChatGPT & Co.