A research team has proposed a way to train medical AI across hospitals even when each site is missing some imaging data.
The framework, called ProMoE-FL, tackles a specific bottleneck in federated learning: what happens when one hospital has chest X-rays but not the paired reports, or vice versa. Rather than requiring a shared public dataset as a crutch — a common workaround — or just ignoring the gap, ProMoE-FL builds a shared "prototype bank" that captures statistical patterns about each modality across all participating institutions. A Mixture-of-Experts layer then uses those prototypes to synthesize plausible stand-ins for whatever data is absent at inference time. The researchers tested the system on four public chest X-ray datasets — MIMIC-CXR, NIH Open-I, PadChest, and CheXpert — and reported consistent improvements over existing approaches in both uniform and mixed data environments.
Federated learning was supposed to solve the problem of training AI on sensitive medical data without centralizing it. In practice, the missing-modality problem has quietly undermined that promise: real hospital networks rarely have the same sensors, workflows, or data-collection histories, so models trained assuming complete data often degrade badly in the field. ProMoE-FL's prototype bank approach keeps sensitive data local while still letting the model learn what a missing input should look like.
The work is a preprint and has not yet been peer-reviewed, so the benchmark gains should be treated as promising rather than proven — a distinction that tends to get lost when AI research moves from arXiv to press release.