An unreviewed preprint examines how clinical prediction systems select uncertain cases for human assessment when outcomes are unevenly represented. Global rankings can disproportionately remove the rarer outcome class from automated decisions, making average performance misleading. The authors compare class-conditioned selection with global methods using synthetic experiments and retrospective neurological disease cohorts.
A clinical AI preprint examines what happens when uncertain predictions are sent for human review while more confident ones remain automated. If a rare outcome attracts higher uncertainty scores, a global ranking can remove it disproportionately from automated decisions. The average error can then improve while obscuring which cases remain. Here the minority means a less frequent outcome class, rather than a demographic population.[1]
The researchers combine a threshold-aware uncertainty score with ranking inside each predicted class and separate risk evaluation for each outcome. The same share is deferred from each predicted class without consulting the true label. This preserves predicted-class proportions, but cannot itself repair a confident misclassification. The authors also note that threshold-adjusted margin ranking gives the same ordering at a fixed threshold, limiting what can be credited to a new score alone.[1]
Tests use synthetic data and retrospective cohorts concerning multiple sclerosis, Parkinson’s disease and Alzheimer’s disease. Comparisons include global selection, class-conditioned selection and random selection within classes. The study concerns binary decisions with a known threshold. It is an unreviewed preprint, and its methodological results do not demonstrate prospective improvements in patient outcomes.[1]
