A Hybrid Deep Learning-Based Approach for Visual Field Test Forecasting.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Informatics

Background:

• Longitudinal visual field (VF) assessment is critical for glaucoma management.
• Traditional VF forecasting requires extensive historical data.
• Deep learning shows promise for more efficient VF prediction.

Purpose of the Study:

• To introduce and evaluate a novel hybrid deep learning framework, Hybrid-VF-Net, for enhanced VF test forecasting.
• To improve the flexibility and accuracy of VF prediction models.
• To assess the impact of data reliability and disease severity on forecasting performance.

Main Methods:

• A retrospective longitudinal study utilizing deep learning models for VF forecasting.
• Trained and compared three models: RNN, CascadeNet-5 (CNN), and Hybrid-VF-Net (RNN + CNN with transformers).
• Evaluated performance using mean absolute error and analyzed factors like data quantity, reliability, and disease severity.

Main Results:

• The proposed Hybrid-VF-Net outperformed existing deep learning methods in VF forecasting accuracy and robustness.
• Hybrid-VF-Net demonstrated resilience to data reliability issues, particularly in severe glaucoma cases.
• Improved performance with fewer prior VF tests was observed, shortening analysis time.

Conclusions:

• Hybrid-VF-Net represents a significant advancement in deep learning-based VF forecasting for glaucoma.
• Forecasting performance is influenced by disease severity, data quality, and temporal factors.
• Future work should focus on refining temporal modeling and utilizing larger datasets to further enhance predictive capabilities.