FedCoRe recovers 50% of AI performance lost to missing medical data
New FedCoRe framework restores half of lost accuracy in healthcare AI when critical scans go missing.
Researchers from NVIDIA (led by Holger R. Roth, Ziyue Xu, and Peter Cnudde) have introduced FedCoRe (Federated Cross-Modal Representation Completion), a novel approach to address missing modalities in federated learning for healthcare AI. The framework tackles a critical real-world problem where hospitals have varying access to medical imaging and data modalities like ECGs or chest radiographs.
FedCoRe operates by learning representation- or logit-space corrections rather than generating synthetic replacements for missing data. In tests using a MIMIC-derived respiratory deterioration task, hiding ECGs reduced model performance by 0.085 AUROC, but FedCoRe's paired-example training restored 0.0415 AUROC (49% of the lost performance). For chest radiographs, the framework recovered 52.8% of lost accuracy. The system uses validation-gated deployment, ensuring corrections are only applied when evidence supports their effectiveness.
The work, accepted at MICCAI's DeCaF 2026 workshop, represents a shift from data generation to targeted corrections in federated learning. By freezing the base multimodal predictor during evaluation, the team isolated the completion module's impact, proving its utility without compromising the core model's integrity.
- FedCoRe recovers 49-52.8% of lost AUROC when critical modalities (ECGs/CXR) are missing in healthcare AI models
- Uses paired-example training and validation-gated corrections instead of synthetic data generation
- Developed by NVIDIA researchers (Holger R. Roth, Ziyue Xu, Peter Cnudde) for federated learning in medical settings
Why It Matters
Enables robust healthcare AI in real-world hospitals with inconsistent data access, improving diagnostic reliability where scans are unavailable.