Federated learning allows hospitals to train a shared mortality model without pooling their records. However, published comparisons of aggregation strategies rarely control the evaluation protocol, making it unclear whether performance differences come from the algorithm or from how the experiment is set up.
A new arXiv preprint directly addresses this gap. The study compared seven aggregation strategies for in-hospital mortality prediction, looking at both care-unit and multi-hospital federations. Its central conclusion, stated in the title, is that protocol choices matter more than the aggregation algorithm.
The abstract does not include numerical results, so the size of the effect is not detailed. Still, the finding suggests that researchers should pay closer attention to evaluation design when interpreting federated learning benchmarks.