External Validation And Artificial Intelligence Assisted Population- Specific Recalibration Of The Ohts- Egps 5-Year Glaucoma Risk Model In A Western Black Sea Cohort

Glaucoma risk prediction models derived from clinical trials may preserve discriminative ability across populations; however, absolute risk estimation can vary due to geographic and demographic differences, leading to reduced calibration accuracy. This study aimed to externally validate the OHTS-EGPS 5-year primary open-angle glaucoma (POAG) risk model in a Western Black Sea ocular hypertension cohort and to optimize absolute risk estimation through AI-assisted population-specific recalibration, thereby developing a regionally adapted risk calculator.

This retrospective cohort study was conducted at the Department of Ophthalmology, Zonguldak Bülent Ecevit University , a tertiary referral center in the Western Black Sea region of Türkiye. Clinical records of patients diagnosed with ocular hypertension and followed regularly for at least five years were reviewed, reflecting a real-world clinical practice setting.

This retrospective cohort study included 212 eyes from 106 patients with ocular hypertension diagnosed between 2010 and 2022 and followed for at least five years. Each eye was analyzed separately. The OHTS-EGPS 5-year POAG risk calculator was applied to generate predicted risk probabilities using age, intraocular pressure, central corneal thickness, vertical cup-to-disc ratio, and pattern standard deviation. The primary outcome was POAG development within five years. Model performance was evaluated using AUC for discrimination and calibration slope, intercept, and Brier score for calibration. Predicted probabilities were subsequently updated using logit-based machine learning logistic regression for population-specific recalibration.

A total of 106 patients (212 eyes) were included, and POAG developed in 30 patients (28.3%) within five years. The original OHTS-EGPS model demonstrated good discriminative performance (AUC=0.770; 95% CI: 0.660–0.872), confirming preserved risk stratification in this independent cohort. Calibration analysis revealed moderate overestimation of absolute risk (calibration slope=0.845; intercept=0.409; Brier score=0.165). After AI-assisted population-specific recalibration, discrimination remained unchanged (AUC=0.770), while calibration improved markedly (slope≈0.98; intercept≈−0.02), and overall prediction error decreased (Brier=0.157), reflecting closer agreement between predicted probabilities and observed clinical outcomes.

The OHTS-EGPS 5-year POAG risk model demonstrated good discriminative performance in a Western Black Sea ocular hypertension cohort. AI-assisted population-specific recalibration improved absolute risk estimation while preserving discrimination, thereby enhancing the model’s clinical applicability. This study demonstrates that established glaucoma risk calculators can be regionally optimized using local data and machine learning–based calibration approaches, enabling more reliable, precise, and personalized clinical decision support across diverse populations. These findings support a scalable framework for adapting validated risk models to different geographic and demographic settings without redeveloping the original predictive structure.