Background and Objectives: Glaucoma (GL) classification is crucial for early diagnosis and treatment, yet relying solely on stand-alone models or International Classification of Diseases (ICD) codes is insufficient due to limited predictive power and inconsistencies in clinical labeling. This study aims to improve GL classification using stacked weight-based machine learning models. Materials and Methods: We analyzed a subset of 33,636 participants (58% female) with 340,444 visual fields (VFs) from the Mass Eye and Ear (MEE) dataset. Five clinically relevant GL detection models (LoGTS, UKGTS, Kang, HAP2_part1, and Foster) were selected to serve as base models. Two multi-layer perceptron (MLP) models were trained using 52 total deviation (TD) and pattern deviation (PD) values from Humphrey field analyzer (HFA) 24-2 VF tests, along with four clinical variables (age, gender, follow-up time, and race) to extract model weights. These weights were then utilized to train three meta-learners, including logistic regression (LR), extreme gradient boosting (XGB), and MLP, to classify cases as GL or non-GL. Results: The MLP meta-learner achieved the highest performance, with an accuracy of 96.43%, an F-score of 96.01%, and an AUC of 97.96%, while also demonstrating the lowest prediction uncertainty (0.08 ± 0.13). XGB followed with 92.86% accuracy, a 92.31% F-score, and a 96.10% AUC. LR had the lowest performance, with 89.29% accuracy, an 86.96% F-score, and a 94.81% AUC, as well as the highest uncertainty (0.58 ± 0.07). Permutation importance analysis revealed that the superior temporal sector was the most influential VF feature, with importance scores of 0.08 in Kang’s and 0.04 in HAP2_part1 models. Among clinical variables, age was the strongest contributor (score = 0.3). Conclusions: The meta-learner outperformed stand-alone models in GL classification, achieving an accuracy improvement of 8.92% over the best-performing stand-alone model (LoGTS with 87.51%), offering a valuable tool for automated glaucoma detection.

Glaucoma is a leading cause of irreversible blindness worldwide, with early intervention often being crucial. Research into the underpinnings of glaucoma often relies on electronic health records (EHRs) to identify patients with glaucoma and their subtypes. However, current methods for identifying glaucoma patients from EHRs are often inaccurate or infeasible at scale, relying on International Classification of Diseases (ICD) codes or manual chart reviews. To address this limitation, we introduce (1) OphthaBERT, a powerful general clinical ophthalmology language model trained on over 2 million diverse clinical notes, and (2) a fine-tuned variant of OphthaBERT that automatically extracts binary and subtype glaucoma diagnoses from clinical notes. The base OphthaBERT model is a robust encoder, outperforming state-of-the-art clinical encoders in masked token prediction on out-of-distribution ophthalmology clinical notes and binary glaucoma classification with limited data. We report significant binary classification performance improvements in low-data regimes (p < 0.001, Bonferroni corrected). OphthaBERT is also able to achieve superior classification performance for both binary and subtype diagnosis, outperforming even fine-tuned large decoder-only language models at a fraction of the computational cost. We demonstrate a 0.23-point increase in macro-F1 for subtype diagnosis over ICD codes and strong binary classification performance when externally validated at Wilmer Eye Institute. OphthaBERT provides an interpretable, equitable framework for general ophthalmology language modeling and automated glaucoma diagnosis.