Forecasting Retinal Nerve Fiber Layer Thickness from Multimodal Temporal Data Incorporating OCT Volumes

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Glaucoma diagnosis and monitoring rely on assessing retinal nerve fiber layer (RNFL) thickness.
• Accurate forecasting of RNFL thickness is crucial for timely intervention and personalized patient management.
• Multimodal data, including clinical, structural, and functional parameters, offer a comprehensive view of ocular health.

Purpose of the Study:

• To develop a machine learning model for forecasting future circumpapillary retinal nerve fiber layer (cpRNFL) thickness.
• To utilize multimodal temporal data from healthy, glaucoma suspect, and glaucoma participants for accurate predictions.
• To compare the model's performance against traditional linear regression methods.

Main Methods:

• Retrospective analysis of a longitudinal clinical cohort of 1089 participants.
• Development of four forecasting models using multimodal data (clinical, OCT-derived, visual field) and intervisit intervals.
• Comparison with a linear trend-based estimation (LTBE) model.

Main Results:

• The best model, using 3 visits and deep learning-derived OCT features, achieved high accuracy (mean error ~1.1-1.9 μm) across all groups.
• The model significantly outperformed the LTBE model in glaucoma suspect and glaucoma participants (P < 0.001).
• High correlation (ρ = 0.95-0.96) was observed between forecasted and measured cpRNFL thickness.

Conclusions:

• The developed forecasting model demonstrates consistent and robust performance across different glaucoma disease states.
• Accurate cpRNFL thickness forecasting can aid in personalizing patient care and optimizing intervisit schedules.
• This predictive capability supports timely interventions for glaucoma management.