Bayesian Uncertainty–Aware Metric Learning for Few-Shot Rare Disease Diagnosis in Clinical AI Systems

Abstract

Rare disease diagnosis in clinical AI systems is challenged by extreme data scarcity and the need for reliable uncertainty-aware decision-making. Few-shot learning is a promising solution in this context, but existing metric learning approaches typically rely on deterministic representations and post-hoc uncertainty estimation. This paper presents a Bayesian uncertainty-calibrated metric learning framework for few-shot rare disease diagnosis. The approach introduces a Bayesian Metric Learner that models class prototypes as multivariate Gaussian distributions inferred via variational inference, enabling uncertainty-aware similarity measurement through a reformulated Mahalanobis distance. A confidence-aware prediction head jointly outputs class probabilities and predictive uncertainty, flagging low-confidence cases for clinician review. The framework supports multimodal clinical data using a hierarchical Vision Transformer backbone and provides interpretability via gradient-based saliency maps. Experiments on real-world clinical datasets demonstrate improved diagnostic accuracy and uncertainty calibration over conventional baselines, while maintaining computational efficiency and scalability for clinical deployment.