Client Participation per Round in Federated Learning for Multiple Sclerosis with Real-World Data

We quantify how the per round client participation rate k affects performance and runtime in a federated learning study predicting two year confirmed disability progression in multiple sclerosis using routine clinical data. Using the original study's data, preprocessing, model, and evaluation protocol, we vary k at 1.0, 0.6, and 0.4. Lower participation shortens wall time but can slightly reduce ROC - AUC and AUC - PR; k ≈ 0.6 preserved nearly all performance while cutting runtime by about one third.