A European research project designed to make hospital cardiology data more accessible for research has entered a new phase. The DataTools4Heart (DT4H) platform is now being tested with real clinical data at seven hospitals across Europe. The aim is to enable researchers to analyse data from multiple institutions without transferring sensitive patient information between them.
The Horizon Europe Research and Innovation Action addresses a persistent challenge in cardiovascular research. Hospitals generate large amounts of data during routine care, but differences in formats, languages and IT systems often make those data difficult to reuse. Relevant information can also be embedded in unstructured clinical notes. According to the project, this represents a missed opportunity in a field with a substantial disease burden. Cardiovascular disease causes around one in three deaths in the European Union and is estimated to cost the EU €282 billion annually.
Data stays inside the hospital
DataTools4Heart uses a federated approach. Rather than collecting patient information in a central database, data remain within the hospitals where they were generated. Analytical tools and machine learning models are brought to the data instead. The platform combines several technologies. These include tools for standardising and harmonising data, multilingual natural language processing to extract information from clinical text and federated machine learning for training models across different institutions.
Other components include CardioSynth, which generates differentially private synthetic data, language models developed for cardiology and virtual assistants intended to help clinicians and researchers navigate information from multiple sources. According to the consortium, the tools are being developed according to privacy-by-design principles and European regulations and data standards. Testing is taking place at clinical sites in the Netherlands, Sweden, the United Kingdom, Italy, Spain, the Czech Republic and Romania.
Real-world data put platform to the test
Moving from technical development to real hospital records is an important test for the project. Clinical datasets are rarely complete or uniform. Missing information, inconsistent coding and differences between hospitals can affect the performance of analytical models.
The project team has therefore adapted its data ingestion tools to identify inconsistencies and coding errors earlier. It is also evaluating several methods for dealing with missing data and has simplified the process for connecting participating hospitals to the platform. “Having the federated learning backend running against real hospital data is the step that turns an architecture into a usable platform,” said DataTools4Heart coordinator Karim Lekadir, ICREA Research Professor at the University of Barcelona and director of the BCN-AIM lab. According to Lekadir, lessons about data quality and missing information during these initial tests will be important for the platform’s performance at the other sites.
The European Society of Cardiology is involved in defining clinical requirements and assessing whether the tools fit everyday cardiology practice. Its volunteers and network of national cardiac societies also support dissemination of the project results. The current testing phase should provide a clearer picture of whether the federated approach can overcome practical differences between European hospital datasets while maintaining patient privacy.
AI as supportive diagnostics tool
Earlier this year we wrote about how AI could help clinicians detect occlusive myocardial infarction (MI) earlier in patients who do not show ST elevation on an ECG. A prospective study presented at ESC Acute CardioVascular Care 2026 evaluated a CE-certified, smartphone-based AI algorithm in 1,490 patients with suspected acute coronary syndrome.
The AI analysed patients’ initial ECGs alongside the standard diagnostic pathway, which included clinician ECG interpretation, troponin testing and, when necessary, coronary angiography. It identified occlusive MI in 108 patients, representing 7% of the study population. The algorithm achieved 84% accuracy, with 77% sensitivity, 99% specificity and a 98% negative predictive value. Conventional ECG interpretation correctly identified occlusive MI in 42% of cases.
The findings suggest that AI-supported ECG analysis could complement existing diagnostic methods and potentially accelerate treatment decisions for patients whose heart attacks are difficult to recognise using conventional ECG interpretation. Further validation is needed before broader clinical implementation.
References
Add ICT&health on Google
Show more content from ICT&health in Google Search.
