BVFLMSP : Bayesian Vertical Federated Learning for Multimodal Survival with Privacy
New Bayesian federated learning model improves C-index by 0.02 while protecting sensitive medical data across institutions.
A research team led by Abhilash Kar has introduced BVFLMSP (Bayesian Vertical Federated Learning for Multimodal Survival with Privacy), a novel framework that addresses two critical challenges in medical AI: data privacy and prediction reliability. The system enables multiple healthcare institutions to collaboratively train survival prediction models without sharing sensitive patient data, using a vertical federated learning approach where each client processes different data modalities (like imaging, genomics, and clinical records) locally. The framework incorporates Bayesian neural networks to provide uncertainty estimates alongside predictions, addressing the limitation of traditional models that offer single deterministic outputs without confidence measures.
BVFLMSP employs a Split Neural Network architecture where clients independently process their data modalities using Bayesian neural networks, then transmit perturbed intermediate representations to a central server that aggregates information for survival risk prediction. The system integrates differential privacy mechanisms by adding carefully calibrated noise to client-side representations before transmission, providing formal privacy guarantees against information leakage during federated training. This approach allows institutions to maintain control over their sensitive data while contributing to more accurate multimodal models.
In experimental evaluations, BVFLMSP demonstrated consistent improvements over existing approaches, achieving up to 0.02 higher C-index compared to the centralized multimodal baseline MultiSurv across various modality combinations. The research team systematically evaluated the tradeoff between predictive performance and privacy protection under different privacy budgets, showing that BVFLMSP remains robust even under strict privacy constraints. The framework's ability to incorporate uncertainty estimates makes it particularly valuable for high-stakes medical decision-making, where understanding model confidence is as important as the predictions themselves.
- Uses Bayesian neural networks to provide uncertainty estimates alongside survival predictions, addressing reliability concerns in medical AI
- Achieves up to 0.02 higher C-index than MultiSurv baseline while maintaining formal privacy guarantees through differential privacy
- Enables collaborative training across healthcare institutions without sharing sensitive patient data using vertical federated learning architecture
Why It Matters
Enables hospitals to build better predictive models for diseases like cancer while protecting patient privacy and providing confidence estimates for clinical decisions.