Comparison of machine learning approaches for structure-function modeling in glaucoma.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Accurate structure-function modeling is crucial for early glaucoma detection and management.
• Current methods for estimating visual field loss have limitations.

Purpose of the Study:

• To evaluate machine learning (ML) approaches for structure-function modeling in glaucoma.
• To estimate visual field (VF) loss using optical coherence tomography (OCT) thickness measurements.
• To compare the performance of different ML models in predicting global and focal VF loss.

Main Methods:

• Trained ML models on OCT thickness data from 832 eyes (537 Asian for training, 148 Asian and 111 Caucasian for testing).
• Estimated global VF mean deviation (VF MD) and focal VF loss using 24-2 standard automated perimetry.
• Compared models using mean absolute errors (MAEs), with baseline MAEs for reference.

Main Results:

• All ML models significantly outperformed baseline estimations.
• Gradient-boosted trees (XGB) achieved the lowest MAE for VF MD estimation in both Asian (3.01 dB) and Caucasian (3.04 dB) test sets.
• XGB and support vector regression (SVR) demonstrated comparable and superior performance in focal VF estimation compared to other models.

Conclusions:

• XGB and SVR models show significant potential for accurate structure-function modeling in glaucoma.
• These ML approaches can effectively estimate both global and focal visual field loss.
• The findings suggest these models could aid in the clinical assessment and monitoring of glaucoma patients.