When uncertainty quantification fails: structural vulnerabilities and corrections for selective prediction under clinical class imbalance

Walter Endrizzi, Flavio Ragni, Raffaele Marchesi, Stefano Bovo, Monica Moroni et al.|October 27, 2026arXiv

Key Takeaway

When deploying selective prediction in clinical settings with class imbalance, standard uncertainty scores become proxies for class membership rather than true ambiguity—use class-stratified, threshold-aligned uncertainty measures instead to avoid systematically excluding minority patients.

Summary

This paper reveals a critical flaw in uncertainty quantification for clinical AI systems with imbalanced data: standard uncertainty measures fail to identify truly ambiguous cases in minority classes, instead just flagging minority patients as uncertain.

safety evaluation

Key Terms

selective-prediction uncertainty-quantification class-imbalance threshold-aware-metrics