Key Takeaways

• Researchers at Duke University developed an AI system designed to determine the optimal treatment timing for patients converting to wet AMD.
• The team found the system’s precision and specificity were particularly strong, reducing false-positive treatment recommendations compared with the baseline models.
• The study was limited by its small single-center dataset and retrospective design, and further studies and additional validation are necessary.

Researchers at Duke University have developed a fully automated AI system designed to determine the optimal treatment timing for patients converting to wet AMD. The study found that the deep-learning framework could predict when anti-VEGF treatment should be initiated while also measuring changes in outer retinal thickness.¹

The study included 122 patients (n = 207 eyes) with intermediate AMD and 2,047 OCT images collected at the Duke Eye Center between January 2006 and November 2020. Researchers divided the data into two datasets to train different components of the AI framework.¹

The modular, multitask, and multimodality deep-learning system consisted of a convolutional neural network to extract the necessary OCT features, and two models to predict treatment initiation and estimate future retinal thickness changes for patients who acutely converted to wet AMD.¹

When evaluated against overall clinical treatment decisions made by retina specialists, the model achieved an area under the receiver operating curve (AUROC) of 0.73, accuracy of 0.87, precision of 0.73, sensitivity of 0.69, and specificity of 0.92. Under a separate single-visit assessment, the framework achieved an AUROC of 0.88, accuracy of 0.89, precision of 0.89, sensitivity of 0.62, and specificity of 0.98.¹

The study found that the AI system’s precision and specificity were particularly strong, reducing false-positive treatment recommendations compared with the baseline models. The authors noted that while some baseline models achieved slightly higher sensitivity, they generated more false positives, potentially leading to unnecessary treatment recommendations.¹

In addition to predicting treatment initiation, the framework estimated future outer retina thickness measurements with a normalized mean absolute error of 11.99%.¹

The authors said the technology may one day be integrated into clinical workflows as a decision-support tool, emphasizing that any recommendations would supplement, not replace, clinician judgment. The study was limited by its small single-center dataset and retrospective design. The researchers also noted that treatment labels were based on real-world clinical decisions rather than standardized image grading protocols. Further studies and additional validation would be necessary before clinical implementation.¹

1. Gao Q, Kuo D, Amason J, et al. Modular multi-task deep learning framework for prediction of treatment initiation in neovascular age-related macular degeneration modular AI for NVAMD treatment initiation study-MANTIS [published online ahead of print September 1, 2026). Br J Ophthalmol. doi.org/10.1136/bjo-2025-328327