Unsupervised Clustering for POAG Phenotyping

To identify and characterize clinically meaningful phenotypic subtypes of POAG using unsupervised clustering of multimodal clinical data.

This retrospective cohort study included 4274 eyes from 4274 patients aged ≥40 years with POAG. Twenty-one clinical features encompassing visual field indices and progression slopes, retinal nerve fiber layer (RNFL) thickness, optic nerve head parameters, and IOP level and variability were analyzed. Hierarchical clustering, K-means, and fuzzy C-means algorithms were applied, with stability assessed via internal validation metrics and visual field archetypal analysis.

Five reproducible POAG phenotypes were identified The. k-means and fuzzy C-means performed best (Calinski-Harabasz 731.76 and 731.43; Davies-Bouldin 1.59 and 1.60), with excellent pairwise agreement between them (κ = 0.97) compared with moderate agreement with hierarchical clustering (κ ≈ 0.55), and high overall stability (mean Adjusted Rand Index, 0.98 ± 0.02). Clusters 1 and 2 represented mild, stable phenotypes (mean deviation [MD] slopes of +0.14 and +0.10 dB/year), differing primarily in IOP burden. Cluster 3 showed structural-functional dissociation-significant RNFL thinning (71.40 ± 8.14 µm) despite limited functional progression (MD slope +0.08 dB/year). Cluster 4 exhibited the most aggressive course, with rapid functional decline (MD slope of -0.98 ± 0.58 dB/year), greatest IOP variability (3.88 ± 1.82 mm Hg), and advanced structural loss (RNFL 64.44 ± 10.05 µm). Cluster 5 demonstrated advanced baseline damage (MD of -15.28 ± 4.72 dB; RNFL, 62.87 ± 9.53 µm) with relative longitudinal stability. Visual field archetypes aligned with cluster-specific severity.

Multimodal clustering identifies distinct POAG phenotypes, supporting improved risk stratification and targeted management.