Predicting Longitudinal Visual Field Progression With Class Imbalanced Data.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Informatics

Background:

• Glaucoma is a primary cause of irreversible blindness globally.
• Standard automated perimetry for visual field (VF) testing is the current diagnostic standard.
• Class imbalance in ophthalmic prediction tasks, particularly with longitudinal VF data, remains a significant challenge.

Purpose of the Study:

• To propose VF-Transformer, a transformer-based framework for predicting visual field progression in glaucoma.
• To address the challenge of class imbalance in longitudinal VF data prediction.
• To evaluate the framework's performance against existing methods and imbalance handling strategies.

Main Methods:

• Development of VF-Transformer, a deep learning framework utilizing transformer architecture.
• Incorporation of a novel inverted class-dependent temperature (ICDT) loss function.
• Integration of weight normalization to mitigate class imbalance effects.
• Validation on public and external hospital datasets using standard performance metrics (accuracy, sensitivity, specificity, AUC).

Main Results:

• The VF-Transformer framework demonstrated effectiveness in predicting VF progression.
• The proposed ICDT loss and weight normalization strategies successfully addressed class imbalance.
• The framework achieved strong performance metrics, outperforming existing methods in the presence of imbalanced data.

Conclusions:

• VF-Transformer offers a robust solution for predicting glaucoma progression from longitudinal VF data.
• The framework's ability to handle class imbalance enhances its clinical applicability.
• This approach holds promise for improving glaucoma management and preventing irreversible blindness.