Machine learning based diagnosis and staging of primary open angle glaucoma using lamina cribrosa parameters

Purpose The aim of this study was to build and internally validate a machine learning (ML) model that can discriminate between normal individuals and patients with different glaucoma stages using the lamina cribrosa (LC) structural parameters. Patients and methods This retrospective analysis used a dataset of 92 eyes (52 glaucomatous eyes of 28 primary open angle glaucoma patients and 40 normal eyes of 20 age-matched normal participants). Five LC parameters, specifically prelaminar neural tissue (PLNT) area, PLNT thickness, LC thickness, LC area, and anterior LC surface depth, were derived from enhanced depth imaging optical coherence tomography of the LC. These parameters were used to train a Random Forest Classifier for classifying the eyes into three groups: normal, mild to moderate glaucoma, and severe glaucoma. Internal validation was performed using a five-fold stratified cross-validation, and SHapley Additive exPlanations (SHAP) values applied to determine the model interpretability. Results The ML model average accuracy was 0.6392±0.1270, with F1-scores (the harmonic mean of precision and recall) of 0.7124 for “normal,” 0.5514 for “mild to moderate glaucoma,” and 0.5921 for “severe glaucoma.” The SHAP analysis identified PLNT area, PLNT thickness, and anterior LC surface depth as the most influential parameters for classification. Conclusion The ML model created using enhanced depth imaging optical coherence tomography LC parameters provided an objective approach in glaucoma diagnosis and staging, while offering clear explanations for its classifications. By identifying parameters such as PLNT area as highly predictive features, the model enhanced the clinical understanding of the structural changes in the LC. External validation is required for broad clinical application.