Free papers detail | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Development and Validation of a Deep-Learning Model to Predict Visual and Anatomical Prognosis of Anti-VEGF Therapy for Neovascular Age-Related Macular Degeneration: A Prospective, Nationwide, Multi-center Study

Authors
Zuyi Yang, Youxin Chen

Presenting
Zuyi Yang

PURPOSE:
The financial burden and uncertain efficacy of anti-vascular endothelial growth factor (anti-VEGF) therapy create critical demand for precise prognosis prediction in neovascular age-related macular degeneration (nAMD). Thus, we aim to develop and validate a deep learning model to predict the short-, medium-, and long-term visual and anatomical prognosis of nAMD patients undergoing anti-VEGF therapy.

PURPOSE:

The financial burden and uncertain efficacy of anti-vascular endothelial growth factor (anti-VEGF) therapy create critical demand for precise prognosis prediction in neovascular age-related macular degeneration (nAMD). Thus, we aim to develop and validate a deep learning model to predict the short-, medium-, and long-term visual and anatomical prognosis of nAMD patients undergoing anti-VEGF therapy.

METHODS:
In this prospective multicenter study across 18 Chinese hospitals, a large dataset was established for nAMD patients receiving Conbercept (0.5 mg/0.05 mL, 3+PRN regimen). A KongMing Model, based on Transformer architecture, was developed to predict short-term (4-6 weeks after a single injection), medium-term (4-6 weeks after the first three injections), and long-term (1-year post-treatment) visual and anatomical outcomes. Model performance was evaluated using metrics including AUC, precision, sensitivity, specificity, human-machine comparisons, mean absolute error (MAE), and Structural Similarity Index Measure (SSIM). Heatmaps and SHAP analysis identified prognosis-related features.

METHODS:

In this prospective multicenter study across 18 Chinese hospitals, a large dataset was established for nAMD patients receiving Conbercept (0.5 mg/0.05 mL, 3+PRN regimen). A KongMing Model, based on Transformer architecture, was developed to predict short-term (4-6 weeks after a single injection), medium-term (4-6 weeks after the first three injections), and long-term (1-year post-treatment) visual and anatomical outcomes. Model performance was evaluated using metrics including AUC, precision, sensitivity, specificity, human-machine comparisons, mean absolute error (MAE), and Structural Similarity Index Measure (SSIM). Heatmaps and SHAP analysis identified prognosis-related features.

RESULTS:
The internal dataset included 34,032 OCT images from 1,240 patients, with 6,828 OCT images from 248 patients in the external set. The model achieved AUCs exceeding 0.94 for all VA change predictions, significantly outperforming ophthalmologists across experience levels (p<0.05). VA value prediction yielded low MAEs (0.048–0.058). Post-treatment OCT predictions showed high structural fidelity (SSIM>0.57). Heatmaps and SHAP images identified critical prognosis-related features.

RESULTS:

The internal dataset included 34,032 OCT images from 1,240 patients, with 6,828 OCT images from 248 patients in the external set. The model achieved AUCs exceeding 0.94 for all VA change predictions, significantly outperforming ophthalmologists across experience levels (p<0.05). VA value prediction yielded low MAEs (0.048–0.058). Post-treatment OCT predictions showed high structural fidelity (SSIM>0.57). Heatmaps and SHAP images identified critical prognosis-related features.

CONCLUSIONS:
The KongMing Model, tested with a nationwide dataset, demonstrated excellent performance in predicting the short-, medium-, and long-term prognosis of nAMD patients undergoing anti-VEGF therapy. It provides a robust and non-invasive method for informed personalized treatment planning, potentially improving treatment adherence and avoid unnecessary interventions.