Florian A. Wenzl | University of Zurich, Switzerland
Citation: EMJ Int Cardiol. 2026; https://doi.org/10.33590/emjintcardiol/STAZV229
Your work has taken you from the University of Pennsylvania, Philadelphia, USA, and the German Heart Centre, Munich, to the University of Zurich, Switzerland. What experiences along that journey have shaped your interest in AI-driven cardiovascular medicine?
I think this developed over time and is rooted in an inclination towards quantitative science. I was fortunate to work in a very stimulating academic environment and became particularly interested in the idea of harnessing medical data to address clinically meaningful questions.
At the European Society of Cardiology (ESC) Congress 2026, you are presenting on ‘Machine Learning for Diagnosis and Personalised Management of Myocardial Infarction’. What are the key messages you hope delegates will take away from your talk?
The first message is that AI has substantial potential to improve patient care. Training AI models on large real-world datasets allows us to leverage clinical information that would otherwise be difficult to interpret. ECGs, laboratory values, clinical characteristics, and other data contain complementary information, and machine learning can integrate these signals to detect patterns that are difficult to recognise using conventional approaches.
Second, I think we need to move beyond simply predicting whether the diagnostic criteria for myocardial infarction (MI) are met. More important questions are: what disease mechanism is at play? And what treatment may be needed? AI could help distinguish different pathophysiological causes of acute myocardial ischaemia and enable more cause-directed treatment.
Moreover, AI could help diagnose MI earlier, which would be hugely important because time is tissue. Signals recorded by wearable devices could detect evolving changes before hospital presentation, potentially shortening pre-hospital delays. After hospital admission, AI could integrate these signals with ECG, troponin, imaging, and other data to refine diagnosis and management.
Why is MI particularly well suited to machine learning-driven approaches compared with more traditional risk stratification methods?
I think there are several reasons. First, there is scale: we have some of the largest, best-curated multinational registries in medicine, which provide the required data. Second, the underlying biology is inherently complex. And third, personalised treatment is heavily dependent on the assessment of ischaemic and bleeding risk.
What are some of the most promising applications of AI in the diagnosis and management of MI that clinicians may begin to encounter in the near future?
A few areas stand out to me. On the diagnostic side, AI-assisted ECG interpretation is maturing quickly. There are models that can flag subtle ischaemic patterns or detect coronary occlusion before it’s obvious to the human eye. Similarly, integrating high-sensitivity troponin measurements with other clinical data, rather than single cut-offs, can improve rapid rule in and rule out in the emergency department. Along similar lines, we have recently developed the novel biomarker-based Takotsubo syndrome diagnostic score (BioTAK), a disease-specific biomarker signature of Takotsubo syndrome supporting its early differentiation from acute coronary syndromes.
On the management side, I’m excited about models that address complex and previously underrepresented patient populations, such as patients with cancer. Refined risk prediction in these patients could optimise interventional treatment and the intensity and duration of antiplatelet therapy. In this context, we have recently developed a cancer-specific, machine learning-based prediction model, termed the Oncology–Acute Coronary Syndrome (ONCO-ACS) score.
Many risk scores are published each year, but few are adopted. What makes a risk algorithm useful enough to change clinical practice?
Tools that get adopted usually share a set of characteristic features. First, the tool has to address a question that clinicians are actually asking at the bedside. Second, it needs to be usable within existing workflows: a score that requires twenty-five extra variables will not be used in a busy emergency department. Third, external validation in independent populations is key to gain trust. And fourth, it needs to be embedded into hospital information systems for automated calculation.
You developed the AI-enhanced Global Registry of Acute Coronary Events (GRACE) 3.0 How does this differ from traditional risk prediction models, and what advantages does it offer in clinical practice?
GRACE 3.0 was built to address several structural weaknesses of the earlier GRACE versions. These earlier versions of the score did not account for complex relationships and for sex-specific differences. GRACE 3.0 instead uses machine learning to capture non-linear effects and interactions between variables. So, for example, the prognostic relevance of heart rate can shift depending on the patient’s broader clinical picture. Both very high heart and very low heart rates indicate increased mortality risk.
In addition, the previous versions of the GRACE score had been developed and validated in predominantly male cohorts and, therefore, systematically underestimated risk in women. GRACE 3.0 addresses these shortcomings and is now externally validated on data from hundreds of thousands of patients with non-ST-elevation acute coronary syndrome across multiple countries.
The ONCO-ACS score reflects the growing intersection between oncology and cardiology. What are the most important cardiovascular challenges facing cardio-oncology today?
Cardio-oncology has become one of the areas I find most clinically urgent, because patients with cancer are systematically underrepresented in cardiovascular trials. These patients represent a fast-growing population and we need more evidence-based ways to translate existing knowledge from the general cardiovascular patient population to patients with cancer.
For example, patients with cancer who present with acute coronary syndrome face a genuinely different risk landscape, including tumour-related bleeding risk, a higher burden of thrombotic events, and competing mortality risk from the malignancy itself. Existing scores were never designed to capture these influences. To address this issue, ONCO-ACS was specifically developed and validated in this population.
Should future cardiovascular guidelines be based more on personalised risk trajectories than fixed risk thresholds? And do you think AI-driven risk stratification could ultimately reduce healthcare inequalities, or is there a risk that it could unintentionally widen them?
I think this depends heavily on the clinical context. If the clinical setting in which a prediction model is used is clearly defined and the predicted risk is clinically relevant, a single risk estimate can guide treatment. In contrast, constantly changing risk estimates without established clinical meaning offers limited value for patient care. Therefore, proper model design and a clear strategy for clinical deployment are key.
On inequality, I’d say it could go either way, and we should ensure that the right frameworks are in place to ensure that equitable and trust-worthy AI models are delivered to patient care. Which outcome we get is a choice, not an inevitability.
You are involved in several ongoing RCTs investigating novel translational therapies. Which areas of research are you most excited about at present, and why?
I am particularly fascinated by trials investigating new approaches to personalised treatment. For example, one of our trials studies a novel AI-enhanced app aiming to improve guideline implementation. The app features a personalised AI avatar that coaches the patient in how to achieve their individual secondary prevention targets after an MI. In another trial, we use point-of-care testing and a biomarker-based risk algorithm to explore the effect of intensified secondary prevention following an MI in ‘very high-risk’ individuals. In yet another trial, we investigate a targeted biomarker-based therapy in the acute care setting.
Finally, what advice would you give to the next generation of cardiologists and researchers who want to contribute to the rapidly evolving field of AI-driven cardiovascular medicine?
Don’t treat quantitative skills and clinical training as separate tracks that you pick between. The researchers making the most impact in this space are the ones who can move fluidly between the wards and the code. It is critical to understand what questions are clinically worth asking in the first place. If you are clinically trained, invest seriously in statistics and programming. If you come from a data science background, spend time in real clinical environments before you start building models for them.
