RetinalVNG-Net, a multimodal deep learning framework integrating fundus photography, OCT, and clinical metadata, achieved macro-averaged AUC-ROC of 0.944 on external validation for simultaneous risk stratification of hypertensive retinopathy, diabetic retinopathy, and neurodegenerative-associated retinal changes, significantly outperforming single-modality models.
RetinalVNG-Net achieved strong internal cross-validation performance across four classification categories.
RetinalVNG-Net demonstrated generalizability on a geographically distinct, device-heterogeneous external test set.
The multimodal RetinalVNG-Net significantly outperformed the best single-modality fundus model.
RetinalVNG-Net employs a multimodal architecture integrating three distinct data streams with cross-modal attention fusion.
The study was a retrospective multi-center design encompassing 2,740 subjects across three independent ophthalmology centers.
The authors note that prospective longitudinal studies are required to establish the value of RetinalVNG-Net for early or predictive detection.
This research describes the development and testing of an artificial intelligence system called RetinalVNG-Net that analyzes eye scans to help identify signs of three different diseases: high blood pressure-related eye damage, diabetic eye disease, and changes associated with neurodegenerative conditions like Alzheimer's disease. The system combines three types of information — color photographs of the retina, detailed 3D scans of retinal layers (OCT), and patient health data — to classify patients into risk categories. Tested on 2,740 patients from multiple medical centers, the AI achieved strong accuracy, correctly distinguishing between disease types about 94% of the time on patients from a new, previously unseen hospital, and performed significantly better than a system using only one type of eye image. This research suggests that combining multiple types of retinal imaging with patient health data in a single AI framework can improve the ability to detect and categorize eye-related signs of serious systemic diseases more accurately than any single imaging approach alone. The system also produces a 'Retinal Biological Age Gap' score, which could potentially serve as a simple number summarizing how much the retina appears older or more damaged than expected for a patient's age. The fact that the system performed well across different hospitals using different equipment suggests it may be adaptable to real-world clinical settings. However, the authors themselves emphasize important limitations: because this was a retrospective study (looking back at existing records), it cannot yet be concluded that the system can predict disease before symptoms appear or improve patient outcomes. The researchers call for prospective studies — following patients forward in time — to determine whether the tool has genuine value for early or predictive detection of these conditions.
Du K, Zheng Y, Wang Z. (2026). Artificial intelligence-integrated multimodal retinal imaging for early detection and risk stratification of systemic vascular and neurodegenerative diseases.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1885303