Sharing Health Data Safely for Better AI Models:
ECDF Project Develops AI for Fall Risk Prevention

Artificial intelligence (AI) can improve healthcare. We all benefit from better diagnoses and treatments. But would you share your health data so that AI models can be trained using it? This is precisely where one of the key hurdles to responsible AI development lies: Patient data is highly sensitive, and its use for research purposes is legally and ethically complex.

A team of researchers from the Einstein Center Digital Future, in a project funded by the Federal Ministry of Research, Technology, and Space (BMFTR), investigated how AI models for fall risk assessment can be developed without compromising the protection of patient data. The project involved data protection officers, nursing staff, computer scientists, and AI developers.

AI for Fall Risk Assessment

In a study conducted in collaboration with one of Germany’s largest hospitals and a specialized geriatric clinic, the team led by ECDF professors Daniel Fürstenau and Felix Bießmann developed and evaluated new methods for fall risk assessment. Ph.D. candidate Ivana Nanevski compared decentralized and centralized learning methods and, in particular, examined the fairness of the models across age and gender groups. Newly developed methods of explainable AI (XAI) also make it possible to understand how individual patient profiles correlate with their specific risk of falling.

Publication: Nanevski et al., The Potential of AI in Nursing Care: Multicenter Evaluation in Fall Risk Assessment, JMIR 2025, doi.org/10.2196/71034.

An Evaluation Protocol for the Safe Handling of Synthetic Health Data

To make AI research on sensitive data possible in the first place, the team has developed an evaluation protocol for synthetic tabular health data. It establishes a framework for systematically assessing the quality of synthetic data, thereby enabling its safe sharing for AI model development.

Publication: Nanevski et al., The Potential of AI in Nursing Care: Multicenter Evaluation in Fall Risk Assessment, JMIR 2025, https://doi.org/10.2196/71034.

Synthius: Open-Source Software for the Evaluation Protocol

With Synthius, the team provides an open-source library that implements a large part of the developed protocol. The software is freely available and generates and evaluates synthetic data created using a variety of generative AI approaches: github.com/calgo-lab/Synthius.

SynTabFall: A New Synthetic Dataset for Fall Risk Research

With SynTabFall, the team has released the first large-scale synthetic tabular dataset for fall risk assessment, which was published in compliance with data protection regulations but can still be used to train AI models. The dataset demonstrates how AI research can advance without compromising patient privacy: models trained on the synthetic dataset achieve a level of predictive accuracy in fall risk assessment that is comparable to models based on real-world data.

Dataset: https://zenodo.org/records/20427694, Nanevski et al., “Anonymized but Useful Synthetic Tabular Health Data for AI-based Fall Risk Assessment,” Nature Scientific Data 2026, https://www.nature.com/articles/s41597-026-07910-z