Sparse-Observation Multi-Horizon Glaucoma Progression Forecasting with Biologically Constrained Temporal Consistency:...

Area of Science:

• Ophthalmology
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Glaucoma progression forecasting is challenging with sparse longitudinal data due to model limitations.
• Existing models often require dense, multi-year data and lack biological constraint integration.
• Irreversible disease progression necessitates models that respect biological constraints.

Purpose of the Study:

• To develop a novel sparse-observation framework for predicting multi-horizon glaucoma progression outcomes.
• To incorporate biological constraints, specifically monotonic risk ordering, into a forecasting model.
• To enable accurate and calibrated long-horizon risk prediction using limited patient visit data.

Main Methods:

• Introduced a sparse-observation framework utilizing only two patient visits.
• Developed a Temporally Consistent Multi-Horizon (TCMH) loss function to enforce monotonic risk ordering.
• Integrated circumpapillary retinal nerve fiber layer (cpRNFL) images, visual field total deviation (VFTD) maps, and clinical covariates using a ConvNeXt architecture.

Main Results:

• Achieved high performance in two-, three-, and four-year glaucoma progression forecasting (AUROC 0.968, accuracy 0.947).
• Demonstrated improved calibration (10.2% better than baseline) and minimal demographic disparity (0.0163 max).
• Outperformed specialist graders in specificity (0.98 vs. 0.57-0.73) on held-out data, enabling efficient risk stratification.

Conclusions:

• Sparse-observation temporally consistent forecasting is a viable paradigm for calibrated long-horizon risk prediction in progressive diseases.
• The developed framework accurately predicts glaucoma progression using limited data, respecting disease biology.
• This approach facilitates automated risk stratification, optimizing clinical review for uncertain cases.