Labeling matters: A multicenter machine learning study on visual field progression in Glaucoma

To compare machine learning (ML) performance for detecting visual field (VF) progression across different labeling strategies using a large multicenter dataset.

In this multicenter retrospective study, VF data were collected from five tertiary referral hospitals. Two algorithm-derived labeling approaches were evaluated without an independent clinical reference standard: an inclusive Consensus label, defined as progression detected by at least one of five conventional algorithms (mean deviation slope, Visual Field Index slope, Advanced Glaucoma Intervention Study, Collaborative Initial Glaucoma Treatment Study, and pointwise linear regression), and a conservative Wiggs' label, based on a region-based event-threshold rule. Four ML classifiers, support vector machine, random forest, logistic regression, and extreme gradient boosting, were trained using each labeling strategy. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and precision-recall analysis summarized by average precision (AP).

Using the Consensus label, all models demonstrated excellent discrimination (AUC, 0.92-0.95), with high sensitivity (0.82-0.85) and near-perfect specificity (0.99-1.00). Precision-recall analysis showed consistently high reliability of progression detection, with AP values ranging from 0.93 to 0.94. In contrast, models trained with the Wiggs' label exhibited lower AUCs (0.88-0.89) and reduced sensitivity (0.63-0.72), while maintaining moderate-to-high specificity (0.87-0.92) and lower AP values (0.84-0.85), reflecting a stricter, region-based progression definition. Ablation analysis showed that Consensus-based performance was not driven by any single criterion, but rather by complementary information across heterogeneous progression algorithms.

In this multicenter study, the labeling strategy was a major determinant of ML performance in VF progression detection. The Consensus label enabled sensitive and reliable identification of progression with high specificity with respect to the Consensus label definition across heterogeneous clinical settings, whereas the Wiggs' label provided conservative, spatially consistent confirmation. The observed performance differences primarily reflect model-label compatibility rather than the clinical validity of either detection system, underscoring that careful definition of ground truth is critical for interpreting ML-based glaucoma progression research.