CBEC: a simple retrieval-based framework for population-grounded prediction reliability estimation in chest X-ray diagnosis
Automated chest X-ray diagnosis fails most critically when models produce confident yet incorrect predictions, suppressing clinical oversight at the point of decision-making. Existing uncertainty estimation methods define epistemic uncertainty primarily as a property of model parameters, overlooking whether predictions remain consistent with clinically similar cases. We propose that the divergence between a model’s prediction and the empirical label distribution of neighbouring training cases provides a practical reliability signal—one that correlates with, and can serve as a proxy for, epistemic uncertainty, while acknowledging that it may also reflect additional sources of discrepancy including label noise, representation error, and local population variability. To operationalize this perspective, we introduce a retrieval-based framework that constructs a fixed embedding-space memory of training cases and estimates a non-parametric label distribution over nearest neighbours. This enables direct comparison between model predictions and local case-level structure without introducing additional trainable parameters or modifying the diagnostic backbone. Experiments on ChestX-ray14 demonstrate improved detection of misclassified predictions relative to deep ensembles, with approximately 2.9 percentage-point gains in uncertainty-based error identification and an approximately 23% reduction in calibration error. Under zero-shot transfer to PadChest and CheXpert, the proposed approach exhibits smaller degradation in uncertainty estimation quality, with larger improvements observed for rare pathological conditions. These findings suggest that case-level consistency provides a meaningful practical signal for reliability estimation in medical imaging, whose relationship to the classical epistemic–aleatoric decomposition warrants further theoretical investigation across broader clinical settings.