Privacy Noise Damages Calibration in Federated Clinical Models
Hospitals cannot pool patient records, and federated learning is widely proposed as the remedy. Most evidence comes from simulated federations cut from one dataset, which understates the difficulty because every shard shares one measurement process. We instead used a cardiology cohort of 920 patients collected independ...