Development and Comparison of Machine Learning Algorithms to Determine Visual Field Progression.

Area of Science:

• Ophthalmology
• Medical Technology
• Data Science

Background:

• Visual field progression is a key indicator of glaucoma severity and disease advancement.
• Accurate detection of visual field progression is crucial for timely treatment adjustments and patient management.
• Conventional algorithms for visual field progression analysis can exhibit bias in classifying uncertain cases.

Purpose of the Study:

• To develop and evaluate machine learning classifiers (MLCs) for the accurate determination of visual field progression.
• To compare the performance of MLCs against established progression algorithms using a large dataset.

Main Methods:

• Utilized 90,713 visual fields from 13,156 eyes for training and testing MLCs.
• Applied six conventional algorithms and six MLCs (logistic regression, random forest, XGBoost, SVC, CNN, FCNN).
• Input data for MLCs involved averaging visual field locations over time in the first and second halves of the study period.

Main Results:

• MLCs demonstrated performance metrics comparable to conventional algorithms, with accuracy ranging from 87% to 91%.
• MLCs achieved sensitivity between 0.83-0.88 and specificity between 0.92-0.96.
• Unlike conventional algorithms which showed significant class bias, MLCs were balanced in classifying uncertain cases (P ≥ 0.08).

Conclusions:

• Machine learning classifiers exhibit moderate to high accuracy, sensitivity, and specificity in determining visual field progression.
• MLCs provide a more balanced approach to classifying uncertain visual field changes compared to traditional methods.
• MLCs hold promise as a valuable tool for improving the assessment of visual field progression in clinical practice.