Federated learning is an emerging paradigm in artificial intelligence that enables the training of robust models across decentralized datasets without requiring the physical transfer of sensitive data. This privacy-preserving approach has gained increasing traction in medical research, where data fragmentation and legal barriers often hinder the development of multicentric trials and AI applications.

In this review, we first provide an explanation of federated learning's process and functioning. We then provide a structured overview of its implementation in clinical research, highlighting key multicentric studies in several fields of medicine. These studies consistently demonstrate that FL achieves comparable, and in some cases superior, diagnostic and prognostic performance in comparison to centralized learning approaches, with area under the curve values often exceeding 0.80. We then consider the potential of federated learning in the context of inflammatory bowel diseases, where data heterogeneity, geographic dispersion, and patient privacy concerns currently limit the development of large-scale predictive models. In doing so, we will provide specific focus on its application in multicentric trials and basic research. Finally, aspects like semantic interoperability in federated learning and privacy issues will also be discussed.

We believe that federated learning could transform the way inflammatory bowel diseases datasets are utilized across institutions, facilitating collaborative algorithm development in areas such as treatment response prediction, endoscopic image analysis, and disease phenotyping—without compromising patient confidentiality.