Machine Learning-Based Prediction of Albuminuria in Diabetic Patients Using Retinal Microvascular Metrics on Optical Coherence Tomography Angiography.

PURPOSE: To investigate the association between optical coherence tomography angiography (OCTA) parameters and diabetic nephropathy (DN) in adults with diabetes and to evaluate the incremental predictive value of OCTA-derived retinal microvascular metrics using machine learning-based models.

METHODS

This prospective observational study included adults with diabetes who underwent systemic evaluation and OCTA imaging. Vessel density (VD), perfusion density (PD), and foveal avascular zone (FAZ) metrics were quantified from OCTA images. DN severity was classified based on the urine albumin-to-creatinine ratio (UACR). Associations between OCTA parameters and albuminuria were assessed using regression analyses. For prediction of albuminuria, machine learning models were trained using clinical variables alone and with the addition of OCTA parameters, and model performance was evaluated.

RESULTS

A total of 388 eyes were analyzed. VD and PD decreased progressively with increasing albuminuria severity (all P < 0.015), whereas central OCTA measurements and FAZ metrics did not significantly differ across groups (all P > 0.10). In regression analyses, higher UACR was associated with reduced VD and PD in the central, outer, and whole regions (all P < 0.05), while marginal associations were observed in the inner region. In machine learning analyses, inclusion of OCTA parameters improved albuminuria prediction in four of the six algorithms (Δ area under the curve [AUC], 0.002-0.027), with the highest performance observed for the decision tree classifier (AUC, 0.754; 95% confidence interval, 0.620-0.865).

CONCLUSIONS

OCTA parameters are significantly associated with albuminuria in diabetes, and their inclusion in predictive models improves albuminuria identification, supporting OCTA as a noninvasive complementary tool for early detection of DN.

TRANSLATIONAL RELEVANCE: OCTA can be integrated into clinical practice as a noninvasive method for early detection of DN, aiding in timely intervention.