Predicting Imminent Conversion to Exudative Age-Related Macular Degeneration Using Multimodal Data and Ensemble...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Exudative age-related macular degeneration (eAMD) is a leading cause of irreversible central vision loss.
• Early identification of patients at high risk for eAMD conversion is crucial for timely treatment and improved patient outcomes.

Purpose of the Study:

• To develop and compare classical machine learning (ML) and deep learning (DL) models for predicting imminent eAMD conversion within six months.
• To integrate optical coherence tomography (OCT) imaging with clinical data into a single predictive model.

Main Methods:

• A retrospective cohort study utilizing spectral domain OCT volumes and clinical data (age, visual acuity, sex, fellow-eye status) from patients with eAMD.
• Development and comparison of ResNet-50, Random Forest, and XGBoost models for predicting eAMD conversion.
• Creation of a multimodal deep learning model (MLP) integrating OCT features and clinical data.

Main Results:

• The best-performing models, both based on DL (ResNet-50) architecture, achieved an area under the operating characteristic curve (AUC) of 0.76 (MLP multimodal) and 0.75 (CNN OCT).
• The multimodal model incorporating both OCT and clinical data demonstrated superior performance compared to the OCT-only model for predicting first-eye and all-eye conversion.
• The MLP multimodal model achieved an AUC of 0.76 (95% CI: 0.71-0.80), outperforming the CNN OCT model (AUC: 0.75, 95% CI: 0.70-0.79).

Conclusions:

• 3D deep learning models trained on OCT volumes can effectively predict imminent eAMD conversion.
• The integration of clinical data with OCT imaging further enhances the predictive performance of these deep learning models.
• These predictive models hold potential as screening tools to prioritize patients requiring urgent retinal care, pending prospective validation.