Unsupervised Clustering for POAG Phenotyping

Area of Science:

• Ophthalmology
• Glaucoma Research
• Data Science in Medicine

Background:

• Primary open-angle glaucoma (POAG) is a leading cause of irreversible blindness worldwide.
• Understanding POAG heterogeneity is crucial for effective management and treatment.
• Current classification may not fully capture the diverse clinical trajectories of POAG.

Purpose of the Study:

• To identify and characterize clinically meaningful phenotypic subtypes of POAG.
• To apply unsupervised clustering algorithms to multimodal clinical data for POAG subtyping.
• To explore the potential for improved risk stratification and targeted management based on identified phenotypes.

Main Methods:

• Retrospective cohort study of 4274 eyes from patients with POAG (age ≥40 years).
• Analysis of 21 clinical features including visual field indices, RNFL thickness, optic nerve head parameters, and IOP.
• Application of hierarchical clustering, K-means, and fuzzy C-means algorithms, with stability assessed by internal validation and visual field archetypal analysis.

Main Results:

• Five reproducible POAG phenotypes were identified using k-means and fuzzy C-means clustering.
• Phenotypes varied in disease severity, progression rates, structural damage (RNFL thickness), and intraocular pressure (IOP) variability.
• Distinct clusters represented mild stable disease, structural-functional dissociation, aggressive progression, and advanced baseline damage with stability.

Conclusions:

• Multimodal data clustering effectively identifies distinct POAG phenotypes.
• These identified phenotypes offer a basis for improved risk stratification in POAG.
• Targeted management strategies can be developed based on these specific POAG subtypes.