Abstract
Purpose: To develop an equitable deep learning model with knowledge distillation to enhance demographic equity in glaucoma progression prediction. Methods: We developed a novel deep learning model called FairDist which uses baseline optical coherence tomography (OCT) scans to predict glaucoma progression. First, an equity-aware EfficientNet termed EqEffNet was trained for glaucoma detection. Next, the pretrained detection model was adapted for progression prediction using knowledge distillation, which minimizes image and identity feature differences between the detection and progression models. Progression was defined based on longitudinal visual field maps from at least five visits up to six years. Model performance was measured by the area under the receiver operating characteristic curve (AUC) and equity-scaled AUC (ES-AUC). Results: For MD Fast Progression, FairDist achieved the highest AUC and ES-AUC for gender (AUC: 0.738, ES-AUC: 0.693) and race (AUC: 0.778, ES-AUC: 0.677). For TD Pointwise progression, FairDist achieved the best AUC and ES-AUC for gender (AUC: 0.743, ES-AUC: 0.719) and race (AUC: 0.746, ES-AUC: 0.645). Conclusions: FairDist enhances both model performance and equity in glaucoma progression prediction after integrating equity-aware learning and knowledge distillation components.
Introduction
Glaucoma is a leading cause of irreversible blindness worldwide. Predicting its progression is a critical clinical challenge, traditionally relying on longitudinal visual field (VF) tests. However, VF tests suffer from high "test-retest" variability and require years of data to establish reliable trends. While Optical Coherence Tomography (OCT) provides objective structural data, predicting future decline from a single baseline scan is difficult due to data scarcity in longitudinal datasets and inherent algorithmic bias. Previous medical AI models have shown performance disparities across different racial and gender groups, often due to skewed training data or unaddressed anatomical variations. This study introduces FairDist, a framework designed to leverage large-scale detection data to improve the accuracy and fairness of progression prediction.
Methodology
The FairDist framework employs a two-stage approach to address data scarcity and bias:
1. EqEffNet: Equity-aware EfficientNet
The core architecture, EqEffNet, modifies the standard EfficientNet by integrating a fairness-aware attention mechanism. This layer processes demographic attributes (race and gender) to weight OCT image features, ensuring the model learns representations that are robust across diverse patient identities.
2. Knowledge Distillation and Multi-Task Co-Training
To overcome the limited size of longitudinal progression datasets, FairDist utilizes knowledge distillation. A "teacher" model (EqEffNet trained on a large-scale glaucoma detection dataset) transfers its learned equity-aware features to a "student" model focused on progression prediction. This process minimizes the feature space distance between the two tasks, allowing the progression model to inherit the teacher's robustness and clinical knowledge.
Progression Definition and Datasets
Progression was defined using two clinical standards derived from longitudinal VF data: Mean Deviation (MD) Fast Progression and Total Deviation (TD) Pointwise Progression. These labels were established using at least five clinical visits over a six-year period.
The study utilized two datasets from the Massachusetts Eye and Ear service: a detection dataset of 10,000 OCT samples and a progression dataset of 500 OCT samples with associated longitudinal VF data.
Results
FairDist consistently outperformed baseline models and alternative bias-mitigation strategies. Performance was evaluated using standard AUC and the Equity-Scaled AUC (ES-AUC), which adjusts the score by accounting for subgroup disparities.
- MD Fast Progression: FairDist achieved an AUC of 0.738 (ES-AUC: 0.693) for gender and an AUC of 0.778 (ES-AUC: 0.677) for race.
- TD Pointwise Progression: The model reached an AUC of 0.743 (gender) and 0.746 (race), maintaining superior equity scores.
Qualitative analysis via class activation maps confirmed that FairDist focused more accurately on relevant retinal layers across all demographic groups compared to standard models, which often attended to irrelevant areas in minority scans.
Conclusion
The FairDist framework demonstrates that integrating equity-aware attention with knowledge distillation can significantly improve both the accuracy and fairness of glaucoma progression prediction. By treating fairness as a feature-weighting problem and leveraging large-scale detection data, FairDist provides a promising path toward reducing healthcare disparities in ophthalmic AI.
Summary updated