Discrimination transfers but calibration does not:1prevalence-dominated miscalibration of diabetic2retinopathy screening across fundus cameras

<title>Abstract</title> <p>Deep learning can screen for referable diabetic retinopathy (DR) from fundus photographs, but a model trained on tabletopcameras must stay trustworthy when deployed on the handheld and portable devices used in low-resource screening. Weask whether a referable-DR classifier trained on tabletop images preserves not only its discrimination but its calibration — theagreement between predicted confidence and observed accuracy — when applied without adaptation to handheld images andto an independent external dataset. Across two convolutional backbones and five random seeds, discrimination transferredwell, reaching within about 0.04 area-under-the-ROC-curve of a target-trained ceiling, whereas calibration degraded five- totwelve-fold. Decomposing the degradation showed it was driven predominantly by the change in disease prevalence betweensettings (about 80%) rather than by the change of camera (about 20%). The miscalibration was inexpensively reversible:recalibrating on 100–200 labelled target images with Platt or isotonic scaling restored calibration to near in-domain levels,whereas temperature scaling could not, because the dominant driver is a base-rate shift. Finally, 8-bit integer quantizationre-broke calibration and reduced discrimination in a way post-hoc recalibration cannot recover, whereas half-precision preservedboth. Reliable cross-device DR screening therefore requires explicit, inexpensive recalibration, not accuracy alone.</p>