Artificial intelligence applied to medicine faces the challenge of patient data privacy and the heterogeneity of clinical datasets, especially for rare diseases. To overcome these barriers, the UPM has introduced FedSDS, an innovative federated learning strategy designed for survival analysis, which predicts the time until a relevant clinical event occurs.
This new approach avoids transferring medical records between centers. Instead, each participating hospital works with synthetic data generated locally, replicating statistical patterns without exposing real patient information. The system also selects synthetic data most similar to each hospital's reality, adapting to contexts with scarce or unbalanced data.
The research results, validated with well-known oncological datasets and real breast cancer clinical data, show that this approach improves performance compared to reference federated strategies, particularly in complex scenarios with limited data, significant differences between centers, or missing key clinical variables.
“"Beyond the technical advancement, the work addresses a very specific need of the healthcare system: collaboration without sacrificing privacy."
Tools like FedSDS could enable hospitals of various sizes to train more robust predictive models without centralizing sensitive patient information, which is particularly valuable for rare diseases or settings with few available cases. The research, published in the journal Computers in Biology and Medicine, was conducted by Patricia A. Apellániz, Juan Parras, and Santiago Zazo from the Information Processing and Telecommunications Center (IPTC) and the Higher Technical School of Telecommunication Engineering (ETSI Telecomunicación) at UPM, with support from European projects like GenoMed4All and SYNTHEMA.
