"The AI revolution may already have arrived, but evidence, workflow integration, trust, and organizational transformation are still missing," says Prof. Alexander Meyer, Director of the Institute for Artificial Intelligence in Medicine at Charité – Universitätsmedizin Berlin. In this interview with ICT&health Global, he explains how healthcare organizations can become AI-ready.
You began your career as a cardiac surgeon. Now you lead an institute shaping AI in medicine.
How did it happen?
I don't think I actually switched roles. I've always been passionate about data: how to represent it, process it, and extract meaningful information from it. Computer science was my first field, and medicine became the ideal playground to apply those techniques to real-world problems.
Working as a physician allowed me to experience medicine from the inside. I came to understand its challenges, its language, and, above all, the enormous potential of using data more effectively. Today, that combination of backgrounds allows me to apply computer science to medicine in a much more meaningful and impactful way.
For someone outside cardiology, what has changed in this field since you started your career? How do we practice cardiology today compared with when you began?
Throughout my career, people have repeatedly told me, "In a couple of years, medicine will be completely different." It hasn't happened.
I remember hearing the same prediction during medical school, when genomics and technologies like polymerase chain reaction were emerging. Our professors were convinced that by the time we graduated, medicine would be fundamentally transformed. It wasn't.
So the honest – and perhaps disappointing – answer is that clinical practice hasn't changed as much as many people expected. Patients still undergo largely the same procedures, and everyday medicine looks remarkably similar to what it did just a few years ago.
What has changed is the ChatGPT moment. For the first time, the general public and many clinicians began to appreciate the capabilities of artificial intelligence. It shifted the conversation from whether AI could play a role in medicine to how it could be applied meaningfully in everyday clinical practice.
Let's go back to another major turning point: smartphones and smartwatches. They have also become important in cardiology, as many people now wear devices that continuously measure health data and can even detect conditions such as atrial fibrillation.
Has this changed predictive cardiology or cardiovascular prevention?
Of course, these technologies have changed parts of cardiology. Detecting atrial fibrillation, for example, is an important advance with significant preventive potential. Identifying patients who appear to be in normal sinus rhythm but actually experience intermittent atrial fibrillation can make a real clinical difference.
At the same time, I wouldn't say wearables have fundamentally reshaped cardiology. That raises a broader question: what do we actually expect from technology? We often describe innovations as disruptive or transformative, but most technologies don't change an entire field overnight. They become powerful tools that gradually improve the way we work.
The same applies to smartphones. They have become platforms for entirely new ways of communicating and delivering healthcare, but they haven't transformed medicine on their own. I think the media often creates unrealistic expectations. Every new technology is presented as revolutionary. In reality, progress comes from many individual innovations that, over time, accumulate into meaningful change.
Some of the early visions of digital health depicted patients using wearables that continuously transmitted data to physicians, who would monitor dashboards and predict events such as strokes before they occurred.
Were those expectations unrealistic?
I think so. Sometimes those visions even seem a bit anachronistic to me. When I see concepts based on command centers with physicians watching endless streams of patient data, I think they're actually the opposite of what we should aim for. An intelligent system shouldn't overwhelm clinicians with every available data point. It should surface only the information that matters for a particular patient at a particular moment. That's what intelligence should mean.
The problem is that we've often created unrealistic expectations around AI because of its enormous potential. AI is a platform technology. The same underlying technology can power applications in cardiology, internal medicine, neurology, hospital administration, public health, and many other areas. That's what makes AI different. It is one technology with the potential to solve many different problems, but that doesn't mean it will transform healthcare overnight.
One of the biggest challenges is that many promising AI technologies never make it into routine clinical practice. You often argue that the barriers include a lack of trust, insufficient evidence, and a lack of the right structures.
What are the key obstacles preventing AI from moving from innovation to everyday clinical use?
There isn't just a single barrier, so let me list the most important ones. The first is financial. AI isn't free, and implementing it can be expensive. Like any digital transformation, it requires upfront investment. If implemented poorly, it can even increase costs rather than reduce them. The second barrier is evidence, and here we need to distinguish between different types of AI.
For classical machine learning applications, such as image or ECG analysis, we already have strong evidence in many cases. Mammography is a good example. The question is no longer whether these technologies work, but why they are not yet used more widely. Is it a matter of funding? Education? Or has clinical practice simply not caught up with the evidence?
For many other machine learning applications, however, the evidence is still emerging. Over the past few years, we've seen a growing number of high-quality clinical trials, which is encouraging because they are building a solid evidence base. Once an AI application has demonstrated sufficient clinical benefit, it should be treated like any other medical device: adopted by hospitals, integrated into clinical practice, and eventually reflected in clinical guidelines.
Generative AI presents a different challenge.
If you follow the scientific literature, you'll notice a striking pattern. One week, a high-profile paper in Nature Medicine suggests that foundation models are ready for medicine. The next week, another paper – also in Nature Medicine – argues that they are not. This isn't a contradiction so much as a reflection of the technology itself.
Foundation models are designed to perform many different tasks. Their strength is their versatility, but that's also what makes them difficult to validate. There are simply too many possible medical applications. That creates a regulatory challenge. Naturally, people ask whether we could certify the overall medical capability of these models and then allow them to be used across multiple clinical scenarios. But that's not how today's Medical Device Regulation works. Every intended use has to be validated separately.
As a result, an extremely powerful general-purpose technology may be approved only for one narrowly defined task. It may seem inefficient, but that's the current regulatory reality. One of the most important questions for the coming years is how we can regulate these general-purpose models in a way that preserves safety while allowing innovation to reach patients more quickly.
Speaking about science that can’t keep up with the speed of AI evolution: A few weeks ago, another paper in Nature Medicine suggested that agentic AI is coming to healthcare.
Do you think cardiology is ready to give AI agents some degree of autonomy?
Agentic AI builds on generative AI. Foundation models learn not only to generate content but also to use tools and carry out workflows autonomously. Personally, I wasn't surprised by the recent results. I never really doubted that this approach could work in medicine. But conviction is no substitute for evidence – that still has to be established through real-world validation.
The real challenge now is implementation. Most studies have been conducted in simulated or experimental environments. The next step is to integrate these systems into real hospitals and evaluate them under real-world conditions. That also means making hospitals themselves ready for agentic AI by building the necessary infrastructure and connecting existing clinical systems.
As for where agentic AI can be applied, everything depends on the level of risk and the available evidence. If a system has been rigorously validated for a particular use case, I see no fundamental reason why it couldn't perform certain tasks autonomously.
We already accept autonomous systems in medicine. A pacemaker, for example, continuously makes decisions without human intervention, even though it relies on deterministic rather than generative algorithms. The principle is the same: if a system is validated, performs reliably, and meets the appropriate safety standards, autonomous operation can be highly beneficial.
Giving autonomy to the machines means stopping human oversight. Do we always have to keep a human in the loop – a cardiologist in the loop – even if AI performs better than a human cardiologist?
It depends entirely on the use case. I don't think there can be a universal rule requiring a human to remain in the loop. Take an insulin pump, for example. It already operates as a closed-loop system, and in this case, continuous human intervention would actually reduce performance rather than improve it.
The discussion is different for probabilistic systems such as generative AI. That's also why I prefer to talk about AI systems rather than AI models. A well-designed system includes guardrails, predefined behaviors, and multiple layers of safeguards. The level of human oversight should depend on how the entire system is designed, not just on the underlying model. That's why the question has to be answered on a case-by-case basis. Personally, I don't believe a human always needs to remain in the loop.
Let's take another example: a smartwatch that detects or predicts atrial fibrillation. If such a system were validated to be highly accurate, should it automatically recommend that a patient see a doctor, or should a physician always be involved before that recommendation is made?
I think that's already how these systems work today. Their role is to alert the patient, not to make decisions on the patient's behalf. Take the Apple Watch, for example. If it detects something unusual, it generates a warning and recommends seeing a doctor. I think that's exactly the right approach.
Ultimately, the patient makes the final decision. Even if a device warns me that something may be life-threatening and advises me to seek medical attention immediately, I can still choose not to. AI should provide recommendations, not commands.
So it should always remain a recommendation because patients have free will?
Of course. Absolutely.
You lead the Institute for Artificial Intelligence in Medicine at the Charité. What are you currently working on? What are the institute's main projects and goals?
We started with a simple observation: medicine has an implementation gap. There is no shortage of AI innovation or promising technologies, yet very few have become part of routine clinical practice. Our goal is to close that gap by making hospitals ready for AI. Together with different organizational units and partners across Charité, we developed an AI strategy. But we didn't wait for the strategy to be finalized before taking action. It was officially in effect in March, and since then we've been fully focused on putting it into practice.
A key part of that effort is building an AI platform that provides secure access to foundation models across the organization. We want clinicians, researchers, and other staff to be able to use generative AI responsibly, while also creating the capability to develop and deploy our own agentic AI solutions.
At the same time, we're laying the technical and organizational foundations for the future. Alongside several research projects on foundation models and agentic AI, we're creating an environment where these technologies can be evaluated using real-world clinical data rather than isolated research settings. Ultimately, that's our mission: to build a hospital where AI can be introduced safely, responsibly, and at scale.
So it's better to start with an AI strategy and build the right infrastructure before introducing individual technologies into an organization?
Absolutely. Otherwise, AI adoption quickly becomes fragmented and difficult to manage. An organization first needs to define what it wants to achieve. Is the goal to remain at the forefront of research? To improve the quality of care? To increase efficiency? Or all of the above? Only then should it decide which AI technologies best support those objectives. AI should be implemented strategically, not opportunistically.
When implementing AI, organizations also have to decide which models to use. Is there already enough evidence to convince decision-makers that AI improves quality of care or patient outcomes?
It depends on what you mean by a clinical application. The first distinction is whether we're talking about an MDR-regulated medical device or a non-MDR use case. For non-MDR applications, convincing people is no longer the biggest challenge. In fact, the opposite can be true. Many people now see AI as a technology that can solve almost any problem, and that's a risk in itself.
That's why having an AI strategy is so important. It provides clear priorities, defines boundaries, and ensures that AI is introduced in a controlled and responsible way. Even outside the scope of the Medical Device Regulation, we still need to make sure that the systems we develop and deploy are safe, reliable, and appropriate for a high-stakes environment like healthcare. MDR-regulated applications, of course, are a different matter altogether because they must meet a much stricter regulatory framework before they can be used in clinical care.
With so many AI models evolving so quickly, how do you personally decide whether a model is ready for clinical use?
Different clinical applications require different model capabilities. In most cases, the available foundation models weren't developed for your specific use case, so you can't simply assume they'll perform well in your environment. That's why local evaluation is essential. Currently, every model has to be tested against the clinical task it is intended to support before it can be used in practice.
As a physician, what does it mean for you to trust medical AI? When do you personally trust AI?
As a physician, I judge AI the same way I would judge a new drug, a biomarker, or a diagnostic test. I look for robust clinical evidence demonstrating safety and improved patient outcomes. The standard shouldn't be any different for AI. If the evidence is there, I trust the technology.
So it’s about evidence, not a generic, often abstract concept of trust.
Yes. If there is solid clinical evidence, then I trust AI. Medicine is full of examples where we don't fully understand the underlying mechanism of action. Lithium is one of them. We prescribe it because decades of clinical evidence show it works, even though we still don't fully understand why.
Ultimately, what matters is the balance between benefits and risks, supported by robust evidence. The same principle should apply to AI. We don't need to understand every aspect of how a model reaches its conclusions, as long as there is convincing evidence that it is safe, effective, and improves patient outcomes.
What should a hospital like Charité look like by 2030?
This is, of course, my personal vision, and it may be influenced by the fact that I'm deeply involved in the field of AI. I envision an agentic AI-ready hospital where physicians, nurses, administrators, and researchers can create AI agents to solve problems in their daily work. Some agents would support standardized workflows, while others could be tailored to individual patients, specific patient cohorts, hospital wards, departments, or even the entire organization. Imagine an intensive care physician spinning up a monitoring agent tailored to a single critically ill patient, or a ward physician asking an agent to check guideline adherence across all patients on the ward.
The range of applications is enormous: from patient monitoring and personalized alerts to guideline checking, scheduling, billing, and many other administrative and clinical tasks. If we succeed in building the right infrastructure, the opportunities to improve healthcare will be almost limitless.
But greater capability also brings greater responsibility. The more powerful these systems become, the greater the potential to create new problems. That's why hospitals need the right safeguards, governance, and oversight from the very beginning.
