ObjectiveTo develop machine learning classification models for primary open-angle glaucoma (POAG) using optical coherence tomography (OCT) optic disc parameters and interocular asymmetry features. MethodsA prospective case-control study. From January 2023 to June 2025, 68 patients (136 eyes) diagnosed with POAG (POAG group) and 83 individuals (166 eyes) undergoing routine eye health examinations (healthy controls group) at the Department of Glaucoma of Nanning Aier Eye Hospital were enrolled. The optic disc was scanned using the Glaucoma Module Premium Edition mode of the Spectralis OCT to obtain peripapillary retinal nerve fiber layer (pRNFL) thickness and Bruch's membrane opening-minimum rim width (BMO-MRW) measurements for the global, temporal, superotemporal, superonasal, nasal, inferonasal, and inferotemporal quadrants. The normalized absolute difference in corresponding quadrant parameters between eyes was calculated as interocular asymmetry indix. Elastic Net feature selection was performed on 47 feature variables, and 11 machine learning classification models for POAG were constructed based on the selection results. The overall classification performance was evaluated using achieving an area under the curve (AUC), accuracy, and F1-score. The top five models with the best performance were selected based on AUC. The SHAP (SHapley Additive exPlanations) values were used to analyze the contribution of each feature to prediction outcome in the best-performing model. ResultsCompared with the healthy control group, both the BMO-MRW and pRNFL thickness in all quadrants were significantly reduced, and interocular asymmetry was significantly greater in POAG group, the differences were all statistically significant (P＜0.05). After feature selection and model comparison, the support vector machine model constructed based on 5 features performed the best. Its AUC was 0.981 (95% confidence interval 0.950-1.000), with an accuracy of 0.957 and an F1 score of 0.952. SHAP analysis indicated that interocular asymmetry in temporal and superotemporal BMO-MRW contributed highly to the model's classification decisions, second only to the inferonasal BMO-MRW thickness of the right eye. ConclusionsMachine learning classification models incorporating OCT optic disc parameters and their interocular asymmetry features effectively distinguish POAG patients from healthy individuals. The asymmetry parameters of BMO-MRW between the two eyes play an important role in the model.
