Cardiology is rich in data—signals (ECG, intracardiac recordings), imaging (ECHO, CMR, ...), device readouts, genetic sequencing, and unstructured clinical notes, among others. Generative AI now offers a powerful toolbox to synthesize, complete, and reason over these heterogeneous sources. Probabilistic diffusion models, variational and token-based autoencoders, neural fields, and many more increasingly work alongside large language models (LLMs) and multi-agent workflows to connect physiologic signals, images, sequences, and text. This seminar convenes machine-learning researchers, clinical cardiologists, and experts in ethics and regulation to chart a responsible path from algorithms to impact at the bedside.

What makes cardiology special? Cardiac diseases are a major global health burden, and many cardiac tasks are safety-critical and time-sensitive; signals exhibit multi-scale periodic structure; arrhythmia morphologies vary across patients and devices; and imaging and hemodynamics place strict physiological constraints on plausible generations. These properties demand evaluation protocols, interpretability strategies, and fairness safeguards tailored to cardiology–not generic healthcare AI.

Our goals are threefold. First, we will distill where generative models can genuinely aid clinicians: data completion and harmonization across sites; robust simulation for teaching and trial design; synthetic cohorts for privacy-preserving research; and interactive what-if exploration grounded in physiology. Second, we will define rigorous validation and governance: clinically meaningful metrics for signals and images; bias and shift testing across demographics and institutions; and certification pathways aligned with evolving regulations. Third, we will articulate practical human-centered workflows–interfaces that expose rationales and uncertainty; clinician-in-the-loop controls; and deployment patterns that preserve accountability.

Primary deliverable: This Dagstuhl Seminar will produce a jointly authored survey and roadmap on generative AI for cardiology, capturing consensus best practices, open challenges, and translational guidance. Additional outcomes include a living public resource (e.g., benchmark/task repository and protocol checklists) and seeds for cross-disciplinary collaborations.

The program blends concise morning plenaries with structured breakouts and flexible "deep-dive" blocks to pursue promising threads that arise on-site. Short hack sprints will enable hands-on prototyping and metric design. By the end of the week, participants will leave with a shared vocabulary and actionable next steps to responsibly harness generative AI for cardiovascular care—tools that complement, not replace, clinical judgment; methods that respect patients; and pathways that meet the bar for safety, equity, and trust.

By the end of the week, participants will leave with a shared vocabulary and actionable next steps to responsibly harness generative AI for cardiovascular care—tools that complement, not replace, clinical judgment; methods that respect patients; and pathways that meet the bar for safety, equity, and trust.

Creative Commons BY 4.0

Samuel Ruipérez-Campillo, Rajesh Ranganath, Randall Moorman, Stephan Mandt, and Julia E. Vogt