Federated Learning (FL) allows institutions to train shared models without exchanging raw data, making it a promising approach for healthcare applications that involve sensitive electronic health records (EHRs). However, despite this distributed design, the gradients exchanged during training can still reveal private information. In this study, we analyze how vulnerable transformer-based language models are to gradient inversion attacks, focusing on the Decepticons method, which can reconstruct original training text from shared gradients. We simulate a cross-silo FL setup with three types of French clinical reports (genetic, anesthesia, and birth records) to evaluate how batch size and sequence length affect reconstruction quality. Our experiments show that a malicious server can recover clinical text with high accuracy: token-level recovery exceeded 95% when training with batch size 1 and remained above 60% for sequences of up to 512 tokens. Reconstructed examples contained identifying elements (names, dates, genetic markers), revealing serious privacy risks for real-world use. These results emphasize that FL alone is insufficient for sensitive clinical text and that privacy-preserving defenses must be integrated before real-world deployment.
Clinical narratives contain rich, detailed information that is essential for medical research but often locked behind privacy constraints due to the presence of personally identifiable information. To enable secondary use while protecting patient confidentiality, these texts must undergo de-identification. Because even a few missed instances can lead to disclosure, systems must achieve high recall. In this work, we present CONORM-DEID, a multilingual model for clinical text de-identification, and investigate how large-scale synthetic pre-training and the availability of human-annotated data impact model robustness in both English and French. We compare models trained from scratch with models incorporating synthetic pre-training across high- and low-data settings. All models trained from scratch achieve high F1-scores (97%+), but robustness, measured as precision at 99% recall, improves only when synthetic pre-training is followed by substantial human fine-tuning. When annotated data are limited, pre-training instead reduces performance, indicating negative transfer. Our code is publicly available at https://github.com/ds4dh/multilingual_deidentification.
High-quality multiple-choice questions (MCQs) are essential for medical education, but costly to design. Large language models (LLMs) can generate MCQs automatically, yet their outputs often fail to meet core quality criteria. This paper presents a self-optimizing multi-agent approach that refines LLM-generated MCQs by aligning them with predefined quality standards. Expert evaluation shows significant improvements for question stems, producing clearer and more answerable items. For answer options, however, improvements remain inconsistent. The findings indicate that automated refinement by self-optimizing agents can reduce manual workload for educators and accelerate the creation of high-quality question stems, while highlighting open challenges for future research.
Poor health data interoperability has been an issue for more than 25 years. With the emergence of data-hungry AI solutions, the need for interoperable and reusable health data is growing. This paper presents a prototype virtual assistant, developed and tested as part of the AIDAVA Horizon Europe project, that creates an interoperable and reusable personal health knowledge graph, compliant with a semantic standard, by maximising automation in the curation of multimodal heterogeneous health data. The paper describes the different transformation steps of a data element from collection to its reusable form, with associated issues. We conclude that the lack of health data interoperability could be resolved by addressing these issues in a structural way through a validated roadmap.
The harmonization of the OMOP Common Data Model (CDM) with HL7 FHIR aims to enhance interoperability in clinical research by harmonizing diverse healthcare datasets. This process, referred to as OMOP-on-FHIR, leverages FHIR Bundles for real-time clinical data exchange and transforms these resources into OMOP CDM format using an ETL process. The ETL pipeline, facilitated by tools like XSLT, enables the extraction, transformation, and loading of data while maintaining semantic consistency. By bridging these two standards, OMOP-on-FHIR promotes the seamless exchange of data across clinical systems and research-oriented databases, supporting global health studies, advanced analytics, and personalized medicine. This methodology advances cross-border research by providing a standardized approach to data management and analysis, thereby improving healthcare outcomes.
