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

The study shows that in clinical prediction tasks with severe class imbalance, standard uncertainty measures no longer reflect clinical ambiguity but degenerate into a proxy for class membership, causing selective prediction that relies on uncertainty scores to refer ambiguous cases to human experts to systematically penalize minority-class patients and create an illusion of safety. Across 56 synthetic conditions spanning from severe imbalance to near balance and three neurodegenerative disease cohorts—multiple sclerosis, Parkinson's disease, and Alzheimer's disease—the framework proposed by the authors maintained bounded and balanced risk, whereas global metrics either collapsed minority-class sensitivity or protected the minority class at the expense of majority-class patients. CCR with standard entropy at tau=0.5 breaks within-class monotonicity and fails to protect the minority class, whereas CCR combined with a threshold-aware measure restores strict monotonicity, correctly identifies ambiguous patients, and safely mitigates minority-class exclusion. The authors call this bias a correctable structural flaw and recommend that clinical pipelines deploying selective prediction under class imbalance adopt threshold-aligned uncertainty measures and class-stratified evaluation.