Work

Multiparametric OCTA Biomarkers for Classifying Mild Diabetic Retinopathy: A Cross-Sectional Evaluation

Year: 2026

Type: article

Abstract: Purpose: To determine whether multiparametric combinations of optical coherence tomography angiography (OCTA) features improve discrimination between mild diabetic retinopathy (DR) versus healthy controls (HCs) and to identify stable features contributing to high-performing models. Methods: Quantitative OCTA parameters, describing perfusion, vessel length, caliber, tortuosity, branching, and fragmentation, were extracted from superficial and deep vascular plexuses across foveal and parafoveal regions. After correlation filtering and principal component analysis (PCA)-guided selection, support vector machine (SVM) models evaluated all feature combinations. Performance was assessed using repeated random partitioning into independent training and validation sets (50 splits), with discrimination quantified by the area under the ROC curve (AUC). Results: Twenty-seven OCTA parameters demonstrated significant group-level differences between mild DR and HCs. Single-parameter models demonstrated moderate discriminative performance (mean AUC range, 0.47-0.88). In contrast, multiparameter combinations yielded consistently higher discrimination, with a subset achieving average AUCs in the ∼0.95 range. Importantly, the most stable contributors to high-performing models were not the strongest individual discriminators but rather features that were weakly correlated with the broader parameter set and recurrently integrated across high-AUC combinations, reflecting complementary vascular information across layers and regions. Conclusions: Multiparametric integration of OCTA features improved discrimination between mild DR and HCs and revealed that feature stability and complementarity, rather than univariate strength alone, underlie robust model performance. Translational Relevance: This work illustrates how structured integration of complementary OCTA features can enhance early DR discrimination and inform future quantitative OCTA analyses without relying on opaque end-to-end models.

Source: Translational Vision Science & Technology

DOI: 10.1167/tvst.15.3.30