A BAYESIAN FRAMEWORK FOR UNCERTAINTY QUANTIFICATION IN AI-DRIVEN DIAGNOSTIC TOOLS FOR RARE AND UNDERREPRESENTED CLINICAL PHENOTYPES

Abstract

Rare and underrepresented clinical phenotypes pose significant challenges for artificial intelligence (AI)-based diagnostic systems due to limited data and inherent variability. This paper proposes a Bayesian framework for uncertainty quantification (UQ) in diagnostic models, allowing clinicians to assess prediction confidence and reduce diagnostic risk. The Bayesian approach provides a probabilistic perspective that naturally accommodates data scarcity and model ambiguity. We integrate this framework into a diagnostic pipeline using deep neural networks with Bayesian layers and test it on rare disease datasets. The results demonstrate improved interpretability and calibrated confidence estimates. This work underscores the necessity of incorporating UQ in AI diagnostics, especially for rare conditions where traditional models may be unreliable.