Development and validation of a deep learning model to predict visual and anatomical prognosis of anti-VEGF therapy for neovascular age-related macular degeneration (KongMing Study): a prospective, nationwide, multicentre study

Background The financial burden and the uncertain response of the anti-vascular endothelial growth factor (anti-VEGF) treatment often cause hesitation among patients with neovascular age-related macular degeneration (nAMD), highlighting the need for a reliable method to predict treatment outcomes. We aimed to develop and validate a deep learning model that can predict the visual and anatomical prognosis of patients with nAMD undergoing anti-VEGF therapy. Methods This prospective, nationwide, multicentre study involved 18 tertiary referral hospitals from 12 provinces across China. A large dataset of patients (aged 50–85 years) with nAMD treated with anti-VEGF therapy (Conbercept, 0·5 mg/0·05 mL, Chengdu, China) under a 3+PRN regimen was established. All participants underwent comprehensive ophthalmological examinations, including best-corrected visual acuity (BCVA) assessment and optical coherence tomography (OCT) imaging at baseline, follow-up, and 4–6 weeks after treatment. Patients with other ocular diseases that could confound the diagnosis or prognosis of nAMD were excluded. A Structural-Attention Guided Therapeutic Response Predicting Model (KongMing Model) was developed based on the lesion-aware, transformer-based multitask architecture to facilitate post-single-injection (4–6 weeks after any single injection during the treatment), post-three-loading-injections (4–6 weeks after the first three consecutive injections), and 1-year-post-three doses plus pro re nata (3+PRN; 1-year after treatment) dual predictions of visual and anatomical prognoses. For the prediction of BCVA changes, model performance was evaluated with the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, precision, average precision, and human–machine comparison. The prediction of BCVA values was evaluated with mean absolute error (MAE) and coefficient of determination ( R 2 ). The prediction of post-treatment OCT images was evaluated by the structural similarity index measure (SSIM) compared with the true post-treatment OCT images. Heatmaps and Shapley additive explanation images were used to identify features related to the prognosis of nAMD. Findings From July 1, 2020, to Dec 31, 2023, we collected 29 772 OCT images from 1226 participants after screening to form the internal dataset. From Jan 1, 2023, to May 1, 2024, we collected 3308 OCT images from 172 participants after screening to form the external dataset. In total, we included 603 (43·1%) women and 795 (56·9%) men. For the prediction of post-single-injection, post-three-loading-injections, and 1-year-post-3+PRN BCVA changes, our model achieved AUCs of 0·948 (95% CI 0·942–0·954) in the internal test and 0·941 (0·934–0·948) in the external test, also significantly outperforming ophthalmologists with different levels of experience in the human–machine comparison (all p