Description
The early identification of patients with type 1 diabetes mellitus at risk of developing diabetic retinopathy (DR) is challenging. In this study, we evaluate a non-invasive machine learning (ML) strategy that integrates radiomic features extracted from multimodal retinal imaging with selected clinical variables to predict DR progression at 5 years. Radiomic features were computed from color fundus photography (CFP), optical coherence tomography (OCT), and OCT angiography (OCTA), and were combined across modalities together with demographic, systemic and ocular data. Models were trained and optimized under a double cross-validation scheme, incorporating feature selection and oversampling within training folds to address class imbalance. In a dataset of 199 eyes, corresponding to 133 patients, the best-performing multimodal imaging-driven model achieved an AUC of 0.90 ± 0.03. A clinical-only model comparable to this one was also obtained, based on blood analytics and diabetes duration, which reached an AUC of 0.89 ± 0.02. These results suggest that the non-invasive ML method described has the potential to replace invasive tests such as blood sampling, achieving similar performance for DR progression risk estimation.