Dual-Level Pattern Tree for Visual Field Improves Glaucoma Progression and Polygenic Risk Prediction

Purpose: To develop a dual-level pattern tree to characterize visual field (VF) loss subtypes that can be used to better predict glaucoma progression and glaucoma polygenic risk scores (PRSs). Methods: This study included 113,030 patients from three datasets, each used for a specific purpose: (1) model training, (2) progression forecasting, and (3) PRS correlations. We applied archetypal analysis to cluster 24-2 VFs into trunk patterns and their branch patterns. The Cox regression model was used to forecast VF progression using mean deviation (MD) slope, MD-fast slope, total deviation (TD) pointwise slope, and visual field index (VFI) slope. Multivariable regression analyses were used to link VF patterns with glaucoma PRSs. The Akaike information criterion (AIC) was used for model comparisons. Results: We identified 17 trunk patterns and 169 branch patterns, with a mean of 9.9 ± 1.6 branches per trunk. Trunk–branch (T-B) patterns were consistently superior to trunk patterns (all contrast P < 0.05) for forecasting 5-year progression using the area under the receiver operating characteristic curve: MD, 0.60 vs. 0.58; MD-fast, 0.84 vs. 0.78; TD pointwise, 0.68 vs. 0.65; and VFI, 0.64 vs. 0.63. The trunk–branch patterns were superior in predicting PRSs (linear regression showed AIC improvement of 26). Conclusions: Trunk–branch VF classifiers were superior to trunk-only characterizations for predicting functional progression and glaucoma PRS. Translational Relevance: High-quality clustering of patient VF characteristics may allow physicians to better manage glaucoma patients by aligning with their goal of care and provide researchers with insights into glaucoma subtypes.