Artificial Intelligence Computer Vision for Sickle Cell Retinopathy Screening Under Variable Fundus Camera Flash Intensity: A Test-Time Calibration Evaluation of Healthcare Referral Sensitivity

Sickle cell retinopathy screening programmes that deploy artificial intelligence (AI) computer vision on fundus photographs face a calibration problem that is rarely modelled explicitly: flash intensity varies across cameras, operators and clinic conditions, and a classifier trained at one illumination level may shift its referral threshold when deployed at another. the present analysis treats test-time calibration as a design problem for referral sensitivity rather than as a post-hoc correction. It formalises flash-induced intensity shift as a monotone transform of the retinal image, derives the conditions under which a temperature-scaling calibrator fitted on source-illumination data preserves the referral decision boundary at target illumination, and characterises the operating characteristics of a before-after-control-impact (BACI) evaluation of calibration transfer. The design uses 8, 16, 24 and 32 clusters, effect sizes of 3, 5 and 8 percentage points, and 2000 replications per cell under a declared variance structure with between-cluster SD 6.0, residual SD 4.5 and intraclass correlation 0.64.56 clusters are required to detect a 5-percentage-point referral-sensitivity gain at 80% power, and that the intraclass correlation of 0.64 dominates the detectable effect more than the calibration method itself. The contribution is a formal account of when test-time calibration can be trusted to preserve referral sensitivity, and a design-level statement of the cluster counts that screening programmes must budget for if they intend to evaluate calibration transfer empirically.