Genetic Risk and Phenotypic Diversity in Primary Open-Angle Glaucoma: Polygenic Risk Scores Association with Laser Trabeculoplasty Outcomes and Data-Driven Phenotyping through Unsupervised Clustering

Abstract

ABSTRACT Objective To examine the relationship between a primary open-angle glaucoma (POAG) polygenic risk score (PRS) and laser trabeculoplasty (LTP) outcomes. Design Retrospective observational follow-up study in a Biobank-linked clinical sample and three case groups identified from population-based cohorts. Participants We included data from patients aged 40 and above with POAG who underwent LTP in the Mass General Biobank, the Nurses’ Health Study, Nurses’ Health Study 2, and the Health Professionals Follow-up Study. Methods A PRS was calculated using prior genome-wide association study (GWAS) summary statistics, and participants with LTP were categorized into either a high- (top 10%) or low-score (remaining 90%) category. Statistical analyses included Kaplan-Meier survival analysis to assess time to LTP failure, Cox proportional hazards models to evaluate the impact of PRS on failure risk, and Area Under the Curve (AUC) to measure predictive accuracy. Main outcome measure LTP failure, defined as less than a 20% reduction in intraocular pressure (IOP) from baseline or the need for additional glaucoma surgical or laser procedures, assessed six weeks after LTP and within two years of follow-up. Results The study included 109 participants, grouped into low PRS (bottom 90th percentile; 72 participants) and high PRS (top 10th percentile; 37 participants). The median time to LTP failure was significantly longer for low PRS patients (12 months, 95%CI: 8.94–16.79) than in those with high PRS (6 months, 95%CI: 5.62–7.36, p.0001). Participants in the high PRS group had a 2.03 times higher risk of LTP failure (95%CI: 1.25–3.32, p=0.004) after adjusting for age, sex, genetic ancestry, LTP type, baseline IOP, IOP-lowering medication number, and visual field mean deviation (MD). A model incorporating PRS in addition to clinical characteristics had higher predictive accuracy over time (AUC = 0.73 vs. 0.66, p=0.03 at 18 months after LTP). Conclusions A higher PRS was significantly associated with increased risk of LTP failure in this sample of POAG patients. These findings suggest PRS may help identify patients at higher risk for LTP failure, but further research with a prospective, population-based sample is needed to confirm this association. ABSTRACT Objective: To compare three unsupervised clustering algorithms: k-means, Fuzzy c-means (FCM), and hierarchical clustering analysis (HCA) for primary open-angle glaucoma (POAG) phenotyping and to identify clinical POAG subtypes using multimodal structural and functional data from Electronic Health Records. Design: Retrospective cohort study. Participants: 4,274 eyes from patients with POAG seen between 2016 and 2023 at a tertiary ophthalmology center. Methods: We extracted 21 continuous clinical features per eye, including baseline and longitudinal visual field indices, intraocular pressure (IOP) parameters, OCT-derived retinal nerve fiber layer (RNFL) metrics, and optic nerve morphology. Dimensionality reduction was performed using Sparse Principal Component Analysis (Sparse PCA), retaining seven components. The optimal number of clusters (K=5) was determined based on silhouette score and HCA dendrogram inspection. We then compared agreement between methods and evaluated cluster performance. Post hoc comparisons across clusters used Kruskal-Wallis and chi-square tests. Main Outcome Measures: Clustering performance was assessed using internal validation metrics (Calinski-Harabasz index, Davies-Bouldin index, Dunn index, and silhouette score) and cluster stability metrics (mean Jaccard similarity for k-means and HCA; fuzzy partition coefficient for FCM). Agreement across methods was quantified using unweighted Cohen’s kappa. Distinct POAG phenotypes were characterized based on demographic, clinical, and treatment features. Results: HCA showed the weakest internal validity (Calinski-Harabasz index = 308.5; Dunn index = 0.02; silhouette score = 0.102) and the lowest cluster stability (mean Jaccard similarity = 0.27). K-means and FCM achieved higher validity (Calinski-Harabasz = 731; Dunn index = 0.013; silhouette scores = 0.133 and 0.132, respectively). FCM yielded the highest overall stability (fuzzy partition coefficient = 0.89). Agreement between FCM and k-means was excellent (Cohen’s kappa = 0.97). Based on FCM-derived clustering, we identified five phenotypes: (1) Small-disc with minimal progression, (2) Early-stage with stable structure-function, (3) RNFL-thinned with preserved function, (4) Rapidly progressing with high IOP variability, and (5) End-stage with apparent functional plateau.

Conclusions: In this large real-world cohort of POAG patients, FCM and k-means outperformed HCA in both internal validity and cluster stability. FCM yielded the highest fuzzy partition coefficient and excellent agreement with k-means, supporting its use for POAG phenotypic segmentation. The five derived POAG subtypes exhibited distinct structural-functional profiles and IOP dynamics. Future work will incorporate archetypal analysis to model spatial patterns of visual field loss and further refine individualized disease trajectories.