Probabilistic Feature Imputation and Uncertainty-Aware Multimodal Federated Aggregation
New method improves chest X-ray classification by 5.36% AUC by quantifying reliability of synthesized data.
A team of researchers has introduced a novel framework called the Probabilistic Feature Imputation Network (P-FIN) to solve a major hurdle in privacy-preserving medical AI. Multimodal federated learning allows hospitals to collaboratively train models without sharing sensitive patient data, but many institutions lack a complete set of diagnostic modalities (e.g., specific X-ray views). Existing methods simply guess (impute) the missing data, treating all guesses as equally valid—a dangerous assumption for medical diagnosis. P-FIN changes this by generating both the imputed features and a measure of how uncertain those guesses are.
This uncertainty is leveraged at two critical levels. Locally, a sigmoid gating mechanism uses the uncertainty to reduce the influence of unreliable imputed features before making a classification. Globally, a new aggregation strategy called Fed-UQ-Avg prioritizes model updates from client institutions whose P-FIN produced more reliable, lower-uncertainty imputations. This creates a self-reinforcing system where the overall federated model learns more from trustworthy data.
Tested on federated chest X-ray classification using datasets like CheXpert and PadChest, P-FIN with Fed-UQ-Avg consistently outperformed standard deterministic imputation methods. The most significant improvement was a +5.36% gain in Area Under the Curve (AUC), a key metric for diagnostic accuracy, particularly in challenging scenarios with high rates of missing data. The work, accepted at the Medical Imaging with Deep Learning (MIDL) 2026 conference, represents a crucial step toward safer, more reliable AI for distributed clinical environments.
- P-FIN imputes missing medical data modalities (e.g., X-ray views) and provides calibrated uncertainty estimates for each guess.
- The framework uses uncertainty to down-weight unreliable data locally and prioritize reliable clients during global model aggregation (Fed-UQ-Avg).
- Demonstrated a +5.36% AUC improvement in federated chest X-ray classification on datasets like CheXpert and PadChest.
Why It Matters
Enables safer, more accurate collaborative AI for hospitals without compromising patient privacy, directly addressing a critical barrier to real-world clinical deployment.