Nature Biomedical Engineering, Published online: 26 August 2026; doi:10.1038/s41551-026-01766-9 We present a multiple-instance learning framework that aggregates patch-level features into reliable slide-level predictions through the use of random sampling at both patch and feature levels. Our approach enables large-batch optimization and consistent performance gains across 35 clinical tasks and 4 foundation models, and supports principled uncertainty estimation for every slide-level prediction.