Key result
CardioSynth successfully generates a multicenter, multilingual synthetic heart failure dataset for secure clinical research.
Why the study?
CardioSynth was developed to create a multi-centre, multilingual, fully synthetic heart-failure dataset for research and data sharing while preserving patient privacy.
CardioSynth provides a multi-centre, multilingual, fully synthetic heart-failure dataset to facilitate research while preserving patient privacy.
Captured external expert commentary on this paper, strongest first. Original sources are linked where available.
“Having the federated learning backend running against real hospital data is the step that turns an architecture into a usable platform. What we learn from these first tests, particularly on data quality and missingness, will determine how the toolbox performs at the remaining sites.”
“The vision for the toolbox is to provide an all-in-one, user-friendly platform where clinicians and researchers can easily access structured clinical data and perform AI-powered analyses—no programming skills will be required. The initial version is set to launch by the end of 2025, marking the beginning of an iterative process to refine the toolbox and adapt it to meet the specific needs of clinicians.”
“Funded by the European Union's Horizon Europe Framework, DataTools4Heart is dedicated to creating an all-encompassing cardiology data toolbox designed for clinicians, researchers, and data scientists. The toolbox will offer advanced features, including data ingestion and harmonisation, multilingual natural language processing, federated machine learning, and data synthesis.”
Enables privacy-preserving multi-center HF research; extends synthetic methods to notes but leaves open real-world validation.
This milestone document reports the status of CardioSynth, the multi-centre, multilingual and fully synthetic heart-failure dataset produced by DataTools4Heart (DT4H). It briefly reported the methods and results of the full CardioSynth report into three questions: why CardioSynth was built, how it is generated, and what it consists of today (September 2026). CardioSynth is generated inside the secure environment of each participating hospital from adult heart-failure hospitalisations extracted through the DT4H Common Data Model. One pipeline produces structured synthetic records with nine families of generative models, evaluates them against the natural variability between two real samples (the noise floor) and a suite of privacy attacks, and passes every candidate file through an automated release gate. A second pipeline writes one clinical note per synthetic patient, in the language of the centre and in the style of clinician-authored documents about fictional patients, so that no real patient text is ever used. Structured and unstructured generation is complete or ongoing at several centres and planned at the others. The first version of CardioSynth (i.e CardioSynth v1.0) includes data generated with two methods (TVAE and MST) for structured data and derived notes from four clinical centres. Following approval from the local officers, CardioSynth will therefore be released in tiers: open documentation, code and evaluation evidence; gate-cleared synthetic data under a data use agreement on the DataTools4Heart platform; and tailored cohorts generated on request inside the hospitals. The route to a fully open release, and the governance decisions it requires, are described in the final section. Full details of the pipeline to build CardioSynth are published in the Deliverable 7.2 (same date of publication).
No takes yet. Share an insight, caveat, or question.
Capelli et al. (2026) studied heart failure. CardioSynth synthetic dataset generation was evaluated on Generation and evaluation of structured synthetic records and clinical notes. The CardioSynth pipeline successfully generated a multi-centre, multilingual synthetic heart-failure dataset comprising structured records and clinical notes for secure release.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: