Uncertainty-Aware Selective Prediction of Neovascular Age-Related Macular Degeneration Recurrence Using Artificial Intelligence

Neovascular age-related macular degeneration (nAMD) is a progressive, sight-threatening disease predominantly affecting older adults and a leading cause of global vision loss projected to rise with aging populations. Although existing artificial intelligence (AI) studies have investigated nAMD recurrence detection, predictive modeling remains underexplored, with limited performance in a few studies relying on a single type of retinal image. Therefore, we aimed to comprehensively explore AI-based recurrence prediction incorporating multiple types of retinal images and clinical data. We collected retinal photographs and clinical data from 399 nAMD patients. The retinal photographs are six images, comprising fundus photographs and optical coherence tomography (OCT) horizontal/vertical scans taken both before and after treatment. The clinical data includes demographic information, such as age, sex and medication type. Using our collected dataset, we proposed an uncertainty-aware selective-prediction model for nAMD recurrence. The proposed model, built on popular CNN architectures, filters out highly uncertain predictions to enhance reliability of the recurrence prediction. Results showed that OCT horizontal post-treatment images achieved an area under the curve (AUC) of 0.662, which increased to 0.836 with Monte Carlo (MC) dropout by excluding uncertain predictions, while the double-threshold method yielded an AUC of 0.744. Incorporating clinical parameters further improved performance, raising the AUC to 0.855, and drug-specific modeling produced particularly high predictive performance, with AUCs of 0.934 for ranibizumab and 0.978 for aflibercept. Our experimental results highlight the potential of AI-based multimodal modeling, selective prediction, and treatment-specific analysis in improving prognostic prediction and guiding therapeutic decision-making for patients with nAMD.