Mr. Wright, you have been a professor at the University of Bremen and BIPS for two and a half years now. First of all, how did you experience your first time as a professor and what does your day-to-day work look like?

It was very stressful at first. I think everyone who starts a new professorship says that, if only because there are so many new tasks and new responsibilities: Many things, such as preparing courses, are particularly hard work at the beginning. You also have to grow into many other tasks first. What's more, I started during the coronavirus period. I have three small children and childcare wasn't always reliable. Every week you had to look at: how do you actually plan the week? When both parents work, it's challenging to combine childcare and work.

But of course it is also exciting to do many things for the first time and to be given new tasks. The Emmy Noether funding gives me the opportunity to conduct research very independently and autonomously. Doing my own research is important to me. And by that I really mean programming myself, coming up with concepts and developing solutions.
In my day-to-day work, my main tasks are research, supervising doctoral students and teaching, even though I have a reduced teaching load due to my cooperative professorship between the University of Bremen and BIPS.

Keyword: cooperative professorship. How can we imagine this?

The special thing about a cooperative professorship is that you are jointly appointed by a university and a non-university research institute and then work at both institutions: My main place of work is the non-university institute, in my case the Leibniz Institute BIPS. This is where I mainly do my research. I also work at the university. There I have all the rights and obligations of a normal university professor. That means I have the right to participate in teaching, commissions and examination boards at the university.

To what extent was the scientific location of Bremen and the connection to the research focus Minds, Media, Machines (MMM) decisive for your decision to stay in Bremen after your Emmy Noether junior research group was funded?

On the one hand, I feel very comfortable here at BIPS. I like the fact that we are so close to the application in the health sciences, but can still do methodical research. Beyond BIPS, and particularly in relation to Minds, Media, Machines, I like the fact that there is so much collaboration in Bremen. For example, I am involved in the "AI in Health" initiative and the "AI Center in Healthcare". Here, various non-university and university partners work together on a topic. The good cooperation is not only due to the size of the location, but above all to its culture. Thanks to the successful collaboration between BIPS and the university and its WSPs MMM (and Health), we were able to acquire the DFG AI research group Lifespan AI. Through this research group, we were able to establish my bridge professorship.

As head of the Emmy Noether Junior Research Group "Beyond Prediction - Statistical Inference with Machine Learning" and co-spokesperson of the DFG Research Unit "Lifespan AI", you develop statistical methods for machine learning processes. What do you see as the particular challenges in modeling health data over the lifespan?

In principle, data protection is of course a major challenge when dealing with health data. We are talking about personal data that belongs to the people themselves. This should not simply be made completely public, but should be protected. However, there is also a trade-off: if data protection requirements are too strict, this can slow down research. You may not gain insights as quickly or you may not be able to combine data from different sources. In Germany, for example, it is typical that different data sources do not have a unique identifier. Such an identifier would make it possible to link anonymized health insurance data with equally anonymized vaccination data. However, as it is not known which vaccination data belongs to which health insurance data, any statistical correlations with this data cannot be investigated.

Another major challenge is time. In Lifespan AI, we want to look at very long periods of time. However, most standard methods are made for a snapshot, in which there is no "time" factor, but with which you want to make a prediction for the future at a certain point in time. In addition, most methods that take the "time" factor into account are often optimized for very small, short time intervals over a comparatively short period of time. Share prices would be a typical example of such a time series.

If we now conduct a study that runs for 20 years, it looks completely different. For example, we receive data on individuals from such studies every two years. In Lifespan AI, however, we want to go one step further and create models covering the entire lifespan. There are very few studies covering the entire lifespan. In order to be able to use a broad database, we are combining these studies. However, this also gives us different data sources that measure very different things. Even if the same things are measured, this is sometimes done in a different way or with different devices. Then new measurement methods are added, others are dropped over time. Taking all of this into account requires many methodological developments.

The problem has long been that we don't have that much health data, let alone long-term data. How do you see the current influx of such health data? How will this continue to develop? And how can we perhaps motivate people to donate their data from the perspective of Bremen research in this area?

I believe that, in principle, there will always be more data. One problem is that a lot of data was originally collected for other purposes. But now we want to use it for health research. If my mobile device records my movements, it is initially unclear whether this recording of movement is at all as precise as the measurement in the laboratory, for example to estimate the body's energy consumption. Another example is health insurance data. This data is often collected for billing purposes and not for research purposes, meaning that much of the information relevant to research is either not included or is at least difficult to derive. In projects such as Lifespan AI, we must therefore be prepared for the fact that much of the data does not come from controlled studies. The current situation is that, on the one hand, the willingness to participate in health science studies is declining, but on the other hand, many people in everyday life carelessly hand over their health data to Google, Apple and co. by simply clicking on "Yes, accept". As we have seen with the Corona Warn app, for example, many people are quite willing to donate their data if they (a) recognize that the donations will benefit themselves and others, (b) the use of the data is transparent and understandable, and (c) they have access to all components (such as the code). I think a helpful component for research could be to accommodate people here and also digitize the research processes in the field of health science as much as possible. Here we should help to create an awareness of who you can donate data to as a user or patient, what happens to this data and by whom it is actually used responsibly.

If you have health data that you want to apply machine learning to, how do you go about making your models explainable and interpretable?

That is our core topic. We look at predictive models with the aim of understanding how they actually arrive at their predictions. Then we try to go one step further and clarify how a disease develops and what the risk factors are that lead to its development. There are many different methods for this.

But perhaps the most important question, which is often underestimated, is: What do I actually want to explain to whom and what do I expect from the explanation? Explainability is not an end in itself. It is completely different if a bank wants or needs to explain to its customer why a loan has been refused than if a patient needs to be told how a genetic marker is related to a disease. It may be that the same methods are used for this. But you may also need completely different methods. It is therefore very important to start by asking: "What question do I actually want to answer?" and then either select suitable methods or develop new ones.

What are typical misconceptions or misunderstandings that you encounter when talking to people outside your field?

My professorship is called 'Machine Learning in Statistics'. And funnily enough, these are two terms that are received very differently. When I say I do statistics or biostatistics, many people think of sports statistics or creating tables, which is very dry and boring. So if you answer the question "And what do you do?" with "statistics", then that's very suitable if you don't want to talk about work (laughs). But the opposite is currently the case with machine learning. Especially recently, it's a typical misconception that 'machine learning' is often equated with large language models like ChatGPT - so the questions now often go in that direction. Others first think of robotics when they think of machine learning, which also plays an important role in MMM. However, very few people think that it can be used to do things like investigate the causes of illness based on health data

There are currently new initiatives in the European Parliament to define EU directives for machine learning procedures. From your perspective, what do you think about such directives, especially for the healthcare sector? Would you like to see this?

For the healthcare sector in particular, I would like to see the technologies used primarily for research purposes and for this type of use to be supported in principle and not restricted or even prevented. At the same time, however, you have to be particularly careful here with corporate interests. On the other hand, companies often develop very helpful procedures, such as AI procedures, which can improve the chances of curing cancer. In my opinion, this should not be regulated away, but on the contrary, research and application should be supported.
In other areas, there are certain AI applications that I believe should definitely be regulated. Take, for example, news that is tailored to personal interests. This means that the messages you receive are so personalized that they either only support your own existing opinion or, even worse, can be used to suggest an opinion. In my opinion, however, the solution does not lie in limiting the models to a certain number of parameters, for example. Instead, we should again consider the application, in principle similar to explainability: What do I actually want to prevent? We should therefore try to regulate in a targeted manner. Regulating AI in general will not work.