Artificial Intelligence Referral for Retinopathy of Prematurity Under Variable Wide-Field Camera Availability: A Cascaded Calibration Evaluation of Healthcare Transfer Timing

Retinopathy of prematurity (ROP) remains a leading cause of preventable childhood blindness, and artificial intelligence (AI) referral systems based on wide-field digital retinal imaging have been proposed to extend screening into districts where ophthalmologists are scarce. Such systems are typically evaluated on image-level discrimination, yet their clinical value depends on a cascaded decision: an image must first be acquired with a camera that is available, then classified with a calibration that transfers across device makes, and only then converted into a transfer decision whose timing is safe. The analysis treats camera availability as a latent regime that governs both the probability of acquiring a gradable image and the calibration of the AI classifier, and it derives the distribution of referral timing error as a function of regime-conditional sensitivity, specificity and acquisition probability. Using a declared variance structure with an intraclass correlation of 0.64, a between-cluster standard deviation of 6.0 and a residual standard deviation of 4.5, the design analysis shows that 54 clusters are required to detect a 5 percentage-point effect at 80% power. The framework clarifies why image-level performance metrics overstate referral safety when camera regimes are heterogeneous, and it identifies the calibration-transfer step as the dominant source of timing error under realistic availability constraints. The contribution is a design-level account of when cascaded AI referral can be trusted for ROP transfer decisions, and a specification of the evidence that a calibration-transfer evaluation must report.