Towards Explainable Medical Risk Prediction under Epistemic Uncertainty via Dempster-Shafer Theory Within a Meta-Fairness Framework

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Explainability and fairness are widely regarded as essential for trustworthy healthcare prediction systems, yet there is no consensus on their precise definition or implementation. In medical risk prediction, both concepts are closely linked to uncertainty. While such tools estimate outcome probabilities and thus primarily model aleatory variability, epistemic uncertainty arising from limited knowledge about model structure, parameters, or data is often insufficiently addressed. In this work, we investigate how different representations of epistemic uncertainty help improve modeling and affect explainability in medical risk prediction, and how this may support decision-making from a meta-fairness perspective. As a use case, we consider the prediction of BRCA1/2 mutation probabilities for hereditary breast and ovarian cancer syndrome. We compare our earlier framework based on Dempster-Shafer theory (DST) with logistic regression and its DST-based interpretation using a synthetic dataset. We extend this two-stage DST model by incorporating interval arithmetic with Yager functions and thick intervals. Our functionally grounded analysis shows that explicit representations of epistemic uncertainty can provide additional information beyond point-valued predictions and make the propagation of uncertainty more algorithmically transparent, contributing to meta-fairness. Finally, we describe the results of a light user study examining how different uncertainty representations are perceived by non-expert users in order to assess their implications for explainability. DST with simple intervals achieved the highest self-reported comprehension, whereas more complex representations were perceived as harder to interpret but more useful for representing uncertainty by one participant group.