Clinical selective prediction misreads imbalance as uncertainty
Original title: When uncertainty quantification fails: structural vulnerabilities and corrections for selective prediction under clinical class imbalance
Across 56 synthetic settings and three disease cohorts, standard entropy failed to protect minority patients.
The paper argues that under severe class imbalance, standard uncertainty scores track class membership rather than clinical ambiguity and can exclude minority patients. Tests covered 56 synthetic conditions and cohorts for multiple sclerosis, Parkinson's disease, and Alzheimer's disease. Combining class-conditional risk with threshold-aware metrics restored within-class monotonicity and reduced minority exclusion. The authors recommend threshold-aligned uncertainty and class-stratified evaluation.
Unconfirmed: Findings are limited to the synthetic conditions and three neurodegenerative cohorts evaluated.
- Standard entropy with CCR broke within-class monotonicity at tau=0.5.
- Cohorts span severe imbalance to near-balanced conditions.
- Global metrics either cut minority sensitivity or harmed the majority.
- The authors call the bias a correctable structural flaw.
From the original article
1 min read at arxiv.orgBackground. Selective prediction systems improve the safety of clinical artificial intelligence by deferring ambiguous cases to human experts while automating high-confidence decisions. However, this safety mechanism relies entirely on the premise that uncertainty scores accurately reflect clinical ambiguity.
By Walter Endrizzi · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.