This multicenter retrospective study examined how different algorithm-derived labeling strategies influence machine learning (ML) performance for detecting visual field (VF) progression in glaucoma. Data were aggregated from five tertiary referral hospitals. The authors compared two distinct ground-truth constructions in the absence of an independent clinical reference: an inclusive "Consensus label" and a conservative "Wiggs’ label." The central objective was to isolate the effect of labeling strategy on ML outcomes independent of model architecture.
VF data were collected from five tertiary centers. Two labeling strategies were applied algorithmically. The Consensus label treated a VF series as progressive if any one of five conventional progression algorithms indicated progression. The five algorithms were mean deviation (MD) slope, Visual Field Index (VFI) slope, Advanced Glaucoma Intervention Study (AGIS) criteria, Collaborative Initial Glaucoma Treatment Study (CIGTS) criteria, and pointwise linear regression (PLR). The Wiggs’ label used a region-based event–threshold rule intended to provide spatially consistent confirmation of progression.
Four supervised classifiers were trained separately using each labeling approach: support vector machine (SVM), random forest (RF), logistic regression (LR), and extreme gradient boosting (XGBoost). Model discrimination was assessed using area under the receiver operating characteristic curve (AUC). Sensitivity and specificity were reported, and precision–recall characteristics were summarized by average precision (AP). An ablation analysis evaluated whether Consensus-based performance depended on any single progression criterion or on complementary information across algorithms.
Models trained with the inclusive Consensus label achieved excellent discrimination with AUCs in the range of 0.92–0.95. Sensitivity was high (0.82–0.85) while specificity was near-perfect (0.99–1.00). Precision–recall analysis indicated high reliability of progression detection for the Consensus-trained models, with AP values between 0.93 and 0.94.
By contrast, models trained using the conservative Wiggs’ label demonstrated lower discrimination (AUC 0.88–0.89) and reduced sensitivity (0.63–0.72), while maintaining moderate-to-high specificity (0.87–0.92). AP values for Wiggs’-trained models were lower (0.84–0.85), consistent with the stricter, region-based definition of progression inherent to that labeling strategy.
These results show that the same ML architectures yield different operating characteristics depending on the ground-truth label used during training: the Consensus label favored high sensitivity and very high positive predictive reliability with respect to its own definition, whereas the Wiggs’ label produced a more conservative detector with lower sensitivity and lower AP.
Ablation analysis indicated that the high performance observed with the Consensus label was not attributable to any single progression criterion. Instead, predictive performance reflected complementary information provided by the heterogeneous set of progression algorithms. In other words, the any-positive Consensus rule aggregated signals that individual algorithms might miss when used alone.
The authors emphasize that the observed performance differences primarily reflect model–label compatibility rather than establishing clinical superiority of one labeling system over another. Because no independent clinical reference standard was applied, high performance with respect to a given label demonstrates that the model learned the label definition reliably, but does not validate that labeling definition as the most clinically correct representation of true disease progression.
The study highlights labeling strategy as a major determinant of ML results in VF progression detection and, by extension, a critical design choice for ophthalmic AI studies. The inclusive Consensus label was motivated by clinical reasoning: in an irreversible disease such as glaucoma, the clinical cost of missed progression (false negative) is considered higher than that of a false-positive alert, motivating an any-positive approach to preserve sensitivity in training data.
Conversely, region-based rules like the Wiggs’ label enforce spatial consistency and produce a conservative definition of progression that reduces false positives at the expense of sensitivity. The authors recommend that developers and interpreters of glaucoma ML models explicitly consider how their ground-truth definitions influence model behavior, expected trade-offs between sensitivity and specificity, and the clinical implications of false positives and false negatives.
The study did not use an independent clinical reference standard; both labeling strategies were algorithm-derived. The raw multicenter clinical data are not publicly available due to institutional and ethical restrictions protecting patient privacy. However, processed train/test datasets, sample data, and analysis code are available at the authors’ GitHub repository: https://github.com/kimhb1029/Visual_Field_Progression. The authors note that performance differences indicate compatibility with label definitions rather than definitive clinical validation of either system.