Deferral-Rule Geometry, Not Multidimensional Reliability, Explains Selective-Prediction Gains in Clinical AI: A Three-Database Intensive Care Study

Multidimensional reliability scores have been proposed as better criteria than confidence for deferring clinical predictions to clinicians. We evaluated reliability-gated deferral for sepsis, mortality, and vasopressor prediction in intensive care patients from MIMIC-IV (68,546 stays), MIMIC-III (42,242), and eICU (119,757 stays; 206 hospitals), with patient-level splits, held-out hospitals, and MIMIC-IV-to-eICU transfer. Against a one-sided gate that defers the lowest predicted risks, a reliability composite reduced the Brier score of retained predictions 1.5- to 2.1-fold at 80% coverage. Against a symmetric gate, maxp~,1−p~, which minimizes this error for a calibrated model, the composite offered no advantage in any of 12 database–task pairs; the symmetric gate was better in five. At matched predicted risk, missingness-aware reliability dimensions and an error model given the entire patient record showed no statistically detectable improvement in deferral beyond calibrated confidence in any of 75 planned tests (54 confirmatory), although injected error signals worth 0.004–0.007 in retained Brier score were detectable. After transfer between databases, calibration error reached 19.7% and some scores carried outcome information beyond predicted risk, but recalibration at the target site improved retained predictions far more than any score. Deferral signals should be evaluated against symmetric confidence at matched risk; shift calls for recalibration.