>> Mr. Nagel, in the 05/2018 issue of “Monitor Healthcare Research” we had a debate about Big Data with Prof. Antes and Prof. Häussler, which was primarily about the potential and obstacles of Big Data. More than two years later, we would like to revisit the topic with you. From today's perspective, what do you think of the statements made back then?
There was a lot of “theoretical explosive” in this interview. However, in practice we approach the topics much less dogmatically - but perhaps we will try to sort things out a little.

You're welcome, then let's start with the core question of the pro/con interview at the time: How do big data and causality fit together?
Big data as a term is not very well defined, it is more of a conglomerate of trends and fictions - I agree with Prof. Antes. What seems more important to me is the fact that the increased performance of computer science, combined with the convergence of technologies, allows us to achieve results more and more quickly. I think they tried to represent this with the so-called four or five V properties of big data. But otherwise there is quite a confusion of terms (digitalization, machine learning, etc.). The term used by Prof. Häussler describes it in a more down-to-earth way. He speaks of “exploratory analysis”. This is certainly always the first step in using data, but it should not be the only one.

You just mentioned the four or five V properties: Volume, Variety, Velocity, Validity, Value. Where do you see problems here?
By the way, the properties of validity and value were also addressed in your interview with Prof. Häussler and Prof. Antes. The V, which stands for variety and the increasing breadth of data, seems to be more and more important to me. This is precisely where new opportunities lie. Of course, big data does not directly lead to an understanding of causality. It is true: Large amounts of data lead to nothing without suitable methods and verification processes. But neither do randomized controlled trials, or RCTs for short. As far as I can observe, the data science community is currently increasingly researching ways to at least approximate causality with models. The diversity or breadth of the data plays a much larger role than its quantity or the other Vs. The awareness, which is increasing through current discussions, that correlations in observational data are of little significance and in their place the questions about the mechanisms behind the correlations has to ask is helpful. I agree with Yoshua Bengio, who said in the specialist magazine “WIRED” in 2019: “The central question is to use algorithms to understand the 'why' and not 'just' recognize patterns.” So it's more possible at this point the question: “Why is something happening?” rather than the question “What is happening?” Polemizing examples of the children and the storks or the cheese and the blanket that were given in the interview with the two professors, on the other hand, attempt to discredit big data and contribute to the debate about this relatively young field of research in which there is still a lot to be done there is little to discover.

Since you are a practitioner, do you have an example?
Of course. We have been dealing with the analysis of structured data using new methods and technologies for a long time. For example, we have been cooperating for some time with the Munich start-up Xplain Data, whose technology has helped us, among other things. helps to go beyond banal correlations to identify possible cause-effect relationships in data or at least to better understand connections.

Like that?
This explains Dr. Michael Haft from what?” is the basis for intelligent
Interventions and thus also for the vision of precision medicine.

Can you be a little more specific?
One of many examples: The Federal Institute for Drugs and Medical Devices (BfArM) warned some time ago that the administration of fluoroquinol antibiotics can lead to serious and long-lasting side effects in the muscles and joints as well as the nervous system, including increased suicide rates. In my opinion, at least two central questions immediately arise in practice:
1. If the BfArM issues such a warning: How does the patient actually find out about it?
2. Can we find the cause-and-effect relationship for fluoroquinol antibiotics that has now been proven by medicine and pharmacology in the health data - in our context with big data methods?

That means?
The question that automatically arises for me is whether we won't be able to detect such side effects systematically and earlier using suitable methods from routine data in the future?

The answer to that is?
To the first question: Not at all or through the doctor, at least not specifically and promptly. In my opinion, it would be necessary to inform patients proactively and systematically via various channels. This is where we should start, because the data is available. This means: We know, or at least could know, who has received the medication, who is at risk of a side effect and in which cases we need to be particularly careful. To put it somewhat polemically: treating a person who is prone to depression with quinolones provokes suicide. And we allow that too!

And to the second question?
This is where you can start if you have the necessary data. Then you find relationships in the data with a “degree of certainty” that goes well beyond a normal correlation.

But isn't this a causal relationship?
No not that. But there are indications of cause-and-effect relationships in the data, which is crucially pre-qualified. The term “prequalification” is very appropriate here, because you need methods that quickly limit the “forest of correlations” to a few relevant hypotheses. Otherwise you can't see the forest for the trees. These prequalified hypotheses must then still be evaluated in discussions with experts from other disciplines (medicine, pharmacology, etc.).

At this point, the objection about unknown confounders certainly often comes up.
Of course, these can and will exist - for this reason the expert must also interpret the results. Nevertheless, the forest of data should first be cleared as best as possible based on the existing data. Big data goes a long way in this respect: more and as diverse data as possible and strong search methods mean that fewer confounders remain undetected. Ultimately, I agree with George Box, who said back in the 1970s: “All models are wrong, but some are useful.” The central question is always: Are new clues or insights emerging that we would otherwise not have or would only have with much greater effort?

What distinguishes a “normal” from a causal correlation?
Causality cannot be proven at all based on pure observational data. However, many hypotheses can be ruled out because direct causality must meet certain conditions.

An example?
For example, we would only assume a causal connection between two variables X and Y if Y follows X in time, because the effect (Y) must always follow the cause (X). In addition, there must be a clear connection between X and Y, which is then called a “normal” correlation. The problem is that there are almost any number of such “correlations” and most of them are completely meaningless - Prof. Antes is absolutely right. Therefore, the third condition is the most important: the observed relationship cannot be explained (or only partially) by other variables.

What do you get out of it then?
Only with the third condition can we rule out many hypotheses - and in doing so, at least bring some light into the forest, so to speak. To put it very simply, we could say: The “direct causal part” of a normal correlation is the significant part that cannot be explained by other variables. However, what is important in practice is the second part of this condition statement, which reads: “cannot be explained using other information”. In practice, of course, this means that this causal part cannot be explained by the other available information - with the emphasis on "available". In this specific case, the information is often not present in the data or is not available in the form of variables.

So a zero-sum game?
But no. This is where the other Vs come into play: Volume, Variety and Velocity. The clear advantage of Big Data is that as contextual information grows – Variety! – an initially weak assumption of causality can become an increasingly interesting hypothesis. Therefore, my clear request is that the existing data must be made much more accessible to research than has previously been the case.

Another sticking point in the interview with Professors Antes and Häussler about Big Data was the statement that more data does not automatically lead to more or proportionally more insights.
That's correct. But with regard to the healthcare system, one simply has to say: In no other market do the players sit so heavily on their data and - here I use the keywords "data protection" and the "transparent patient" as examples - are not prepared to share their data. Unfortunately, many past initiatives have shown this. Health Minister Spahn has long since brought dynamism back into the process. One can only hope that it isn't a flash in the pan.

One could almost get the impression that you are a bit at loggerheads with data protection?
Not at all. In my opinion, the requirements posed by data protection are largely met and certainly achievable. Just don't mix things up. When I read articles - put here as an example - that "many citizens are concerned that their sensitive patient data will be made available to research without being given the right to object", then I have maximum doubts. What does many mean? Where is the evidence? In my view, a problem of the actors is being projected onto the patients.

What do patients say about your projects?
In all the projects in which we work with patients or their representatives, the patients have the fewest problems with data protection - and not because they handle their data carelessly, but because they hope to gain concrete insights for their care help to improve their everyday life and better understand their individual situation.

Could you also give an example of this?
In the national transplant register, which we are allowed to set up and operate on behalf of the German Medical Association, the National Association of Statutory Health Insurance Funds and the German Hospital Association, the data protection standards were very high right from the start and are of course fully coordinated with the Federal Commissioner for Data Protection. The patient representatives were also involved from the start. It will also be possible by law to export this data for quality assurance and research.

What happens to the data itself?
The problem addressed is of a theoretical nature, at least in healthcare. Simply because we find it extremely difficult to build a really large data pool. From a competitive perspective, this is a maximum disadvantage for the German big data and AI research fields compared to the heavyweights USA and China - even if I don't want to import either model to Europe. You can also assume that we are monitoring national initiatives very closely, such as the National Cohort Health Study (NAKO) or the “Artificial Intelligence” Study Commission that has just been launched.

Is there a reason for this?
Let's take the corona pandemic: We are confronted with a wealth of speculation about which comorbidities cause a difficult course or which active ingredients prevent it - keyword: hydroxychloroquine. However, correlations are mostly reported. We are all currently experiencing how difficult it is to find answers to such a new challenge. A large part of the time is wasted laboriously collecting the data we have - be it the data from health insurance companies, clinics, health authorities or the data that we have yet to generate via apps and studies.

What specific contribution could you make?
At the health forums, we are in the fortunate position that we are able to keep health insurance data in a research database as well as operate various registers and national data pools, such as the hospital quality reports of the Federal Joint Committee (G-BA). Nevertheless, general access by research institutions to the available data is extremely poor. The Federal Ministry of Health wants to address this problem with the new research data center, but looking at the last 20 years, the hurdles have not really been reduced.

And what do you see as the reason for this?
System worlds collide here. The question that still needs to be answered is: Should we organize our healthcare system more in a market-based manner? Then the data in the sense of “assets” is the new oil for the players, which are documented by all the “real world evidence” and data science initiatives and organizational units.

The antipode would be?
Do we afford at least enough transparency for society's public health services in the area of health so that we can answer central control questions?

Aren't we doing that today?
This can best be explained with an example: With the introduction of the health fund, diagnoses have become a central parameter of distribution. One can therefore assume that the diagnoses are well documented - with all the distortions that this entailed and which I do not want to discuss again here. As early as 2012, we were able to show in a private study based on routine data that the prevalence of multiple sclerosis in Germany is far higher than all prevalence figures previously published in medical publications. After a publication we carried out together with partners in 2015, the medical societies also adopted the numbers, although only in 2018/2019, but at least. But that also means that there was a six to seven year delay in adopting even the simplest, descriptive results based on data that is actually available. And we're not even talking about correlations. Another current example: In the context of the Covid-19 pandemic, there were naturally many questions at the beginning. The topic was new for all of us. Therefore, it should also be scientifically accepted that one cannot provide all the answers immediately and that one has to change one's opinion if the data situation changes. Nevertheless, some theses could have been verified or falsified very quickly on the basis of routine data, provided that this routine data had been allowed to be used in a timely manner. In reality, we only started collecting the data with the pandemic. Not that I'm being misunderstood: the current data will usually not be sufficient for new topics, but not having it available and being able to use it simply uses up valuable time.

And your company could do that?
We are certainly not the only ones in the Republic, but we have been able to monitor the pandemic since its outbreak based on routine data from our statutory health insurance partners. I would have liked a national, open research pool. We would have been significantly faster that way. The availability of such data ultimately determines how leading or non-leading Germany is in the development of new big data and AI methods.

Is that desired?
That is the question of all questions. In this case, it is helpful to look at the arguments and ask yourself “What is the motivation?” or “Cui bono?”, i.e. “Who benefits from the argument?”

Let's look at the key players and their arguments, such as politics.
As just briefly explained, the BMG is currently trying to set up a research data center that is located at the German Institute for Medical Documentation and Information (DIMDI), which has now been integrated into the BfArM. This option has actually been available for years - but it is cumbersome and the data is not up-to-date enough. Nevertheless, the minister is criticized for being too “industry-friendly” and not “data protection compliant”.

How do the health insurance companies and doctors act?
In my view, the central problem of health insurance companies lies in competition with one another. Here the data is now understood as assets. When it comes to doctors, their lobby organizations have been torpedoing all efforts that lead to more transparency and curtail their own sovereignty for decades. This can be clearly observed with the health card and telematics infrastructure. Not that I'm being misunderstood: If many parties are fighting to get a piece of the pie, it just shows that it is now understood how much potential there is in data and digitalization. Therefore, everyone claims as large a piece as possible for themselves. Nobody wants to give up or lose the sovereignty of interpretation.

That can be worrying, right?
Secure. My main concern is that we will continue to lose touch in the area of AI and big data in Germany and Europe if we do not solve the problem of data availability.

And almost as always, we come to the citizens last.
Most people are generally willing to participate with their data - we have seen this in many ways, especially during the pandemic. Expressions such as the “transparent patient” predominantly attempt to exploit citizens or patients in the debate.

Another central point of contention in the debate about the pros and cons of big data concerned the need for the strictest scientific approach versus aimless digging around in data. Which data is suitable?
Prof. Wascher took up this in Rapid Report A19-43 in May of this year and thereby disqualified the routine and patient record data as a well-founded research source. Only registers and studies are the method of choice, says the Institute for Quality and Efficiency in Health Care (IQWiG).
What do you think?
I think we are making a mistake by excluding routine data.

What does that mean?
A simple example: A patient goes to the doctor and receives a prescription. The process is noted in the doctor's information system, so that the data indirectly becomes the doctor's data. In 2020, the patient still receives a piece of paper called a prescription. With that he goes to the pharmacy. The paper is then scanned there and recorded in the pharmacy information system, whereby its contents indirectly become pharmacy data. The pharmacy then sends the prescription to a pharmacy data center for billing, which then sells this data to various brokers and analysts. The data record is then sent to the health insurance company's billing center and from there to the relevant health insurance company, which triggers the payment process and thus becomes the proud owner of routine data. The following applies to all of these intermediate steps: The digitalization of this process does not significantly change the ownership and protection relationships. But here too, all of this data is still not systematically available for research. So the point is: There is data in many places, in many organizations - just unfortunately not nationally.

A final aspect from the discussion mentioned at the beginning referred to the “data scientist”, who did not fare particularly well in terms of his skills. What do you think?
When I studied computer science, there was the term “data science”
not yet. This only emerged systematically in the mid-1990s. Today, computer science should be allowed to stand on an equal footing with the disciplines of biometrics, medicine, epidemiology, genetics, biology and mathematics. It doesn't help to defend ourselves against what will inevitably come our way: ever-increasing complexity combined with ever-increasing speed. In order to solve the problems at hand, different disciplines are inevitably needed.

Mr. Nagel, thank you very much for the interview. <<

The interview was conducted by MVF editor-in-chief Peter Stegmaier.