Multimodal Deep Learning for Longitudinal Prediction of Glaucoma Progression Using Sequential RNFL, Visual Field, and...

Area of Science:

• Ophthalmology and Artificial Intelligence
• Medical Imaging and Machine Learning

Background:

• Glaucoma progression forecasting is essential for preventing irreversible vision loss.
• Existing methods struggle with long-term prediction accuracy and multimodal data integration.

Purpose of the Study:

• To develop and validate a multimodal, longitudinal deep learning framework for predicting glaucoma progression.
• To forecast disease advancement over a two-to four-year horizon using integrated patient data.

Main Methods:

• Utilized a retrospective cohort of 10,864 glaucoma patients.
• Integrated sequential structural (OCT RNFL scans), functional (visual-field maps), and clinical data.
• Employed four deep learning backbone architectures (ConvNeXt-V2, ViT, MobileNet-V2, EfficientNet-B0) with a bidirectional LSTM for temporal analysis.

Main Results:

• The ConvNeXt-V2-based model achieved high performance (0.97 AUC, 0.94-0.96 accuracy).
• Demonstrated robust performance across diverse demographic subgroups (sex, race) with minimal impact in older adults (>70 years).
• Saliency maps indicated clinical relevance, localizing predictions to relevant retinal nerve fiber layer regions.

Conclusions:

• The developed deep learning framework effectively fuses multimodal data for accurate, long-horizon glaucoma progression risk stratification.
• The model offers interpretable and equitable predictions, advancing personalized glaucoma management strategies.
• This approach holds significant potential for proactive intervention and vision preservation in glaucoma patients.