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

Glaucoma progression forecasting from sparse longitudinal data remains unsolved, as existing models require dense multi-year inputs and lack the biological constraints. We introduce a sparse-observation framework that predicts multi-horizon outcomes from only two visits, incorporating a novel TCMH (Temporally Consistent Multi-Horizon) loss that enforces monotonic risk ordering to reflect irreversible disease biology. Applied to glaucoma progression, we integrate circumpapillary retinal nerve fiber layer (cpRNFL) images, visual field total deviation (VFTD) maps, and clinical covariates through ConvNeXt architecture trained with TCMH loss. In 3,593 patients (13,087 sequences), our model achieved AUROC 0.968 and accuracy 0.947 for two-, three-, and four-year progression forecasting, with 10.2% better calibration than baseline and 0.0163 maximum demographic disparity. At 90% coverage, classification error remained 2.5%, enabling automated risk stratification with expert review for only uncertain cases. The model exceeded three independent specialist graders on specificity (0.98 vs. 0.57-0.73) on 108 held-out eyes. These results establish sparse-observation temporally consistent forecasting as a generalizable paradigm for calibrated long-horizon risk prediction in irreversible progressive diseases.