Diabetes mellitus is commonly diagnosed and monitored using invasive blood-based assays, which can reduce screening adherence in low-resource and high-throughput settings. This study investigates a non-invasive biometric alternative based on iris topography and deep learning. The central technical challenge is geometric mismatch: conventional convolutional neural networks (CNNs) are optimized for Cartesian image grids and can underperform on circular iris structures, while also overfitting peri-ocular confounders such as eyelashes and scleral regions. A second challenge is methodological validity; many small-cohort studies report inflated performance due to image-level splitting that leaks patient identity across training and testing subsets. To address both issues, we implement a dual-geometry ensemble on AWS SageMaker that (i) isolates stromal tissue, (ii) processes both spatial and polar representations, and (iii) enforces strict patient-level partitioning. The dataset contains 325 iris images from 196 unique patients, with an independent zero-leakage validation vault. In parallel, a SHA-256-based privacy layer masks personally identifiable information (PII) to support compliant handling of biometric data. The present study addresses binary diabetic-versus-control discrimination and does not attempt Type 1/Type 2 subclassification
Almonzer et al. (2026) studied this question.