- What if data for machine learning research cannot be shared? -> Federated learning!
- Is this easy when working with images and other data from population-based cohorts? -> No!
- Can it be done? -> Yes! But a great team (and a lot of perseverance) is crucial!
You can check those in 🧠“Multi-Cohort Federated Learning Shows Synergy in Mortality Prediction for MRI-Based and Metabolomics-Based Age Scores” ðŸ§
The paper results from a very successful collaboration involving PhD students, technical experts, data experts, privacy/legal experts, and principal investigators from different institutes participating Netherlands Consortium of Dementia Cohorts (NCDC): Maastricht University Medical Centre, Leiden University Medical Center, Delft University of Technology, Eindhoven University of Technology, Vrije Universiteit Amsterdam, Amsterdam UMC, and University Medical Center Rotterdam.
In particular, Pedro Mateus, Swier Garst, Jing Yu, Inigo Bermejo, Hailiang Mei, and Esther E. Bron contributed equally to this work. From our BIGR, Jing, Alexander, Mahlet, Gennady, and Esther contributed this publication.
In the paper:
- They provide insight into the relation between two types of biological age scores: BrainAge based on MRI scans and MetaboAge based on metabolites from blood.
- They implement a federated learning infrastructure connecting three large-scale population cohorts (the Rotterdam Study, The Maastricht Study and the Leiden Longevity Study) and train a federated deep learning model for BrainAge prediction.
- They compare federated and locally trained deep learning models. The results emphasize the advantage of federated learning in developing models with better generalizability, benefiting cohorts without sufficient data to train a model.
Read the full article here.