High interpretable machine learning classifier for early glaucoma diagnosis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning in Healthcare

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Early detection of glaucoma is crucial for preserving vision.
• Peripapillary retinal nerve fiber layer (RNFL) thickness is a key indicator in glaucoma diagnosis.

Purpose of the Study:

• To develop an interpretable machine learning classifier for differentiating healthy eyes from early glaucoma.
• To identify the most significant RNFL thickness parameters for glaucoma classification.
• To utilize optical coherence tomography (OCT) data for enhanced diagnostic accuracy.

Main Methods:

• Utilized RNFL thickness data (quadrant, clock-hour, average) from 90 early glaucoma patients and 85 healthy controls.
• Employed tree gradient boosting algorithms to determine parameter importance and build an interpretable classifier.
• Applied cluster analysis for outlier detection and cross-validation for model evaluation.

Main Results:

• Average and 7 clock-hour RNFL thicknesses were the most important predictors.
• A decision tree model identified average RNFL thickness < 82 µm as a strong predictor of early glaucoma.
• The classifier achieved an Area Under the ROC Curve (AUC) of 0.953 and 89% accuracy.

Conclusions:

• Gradient boosting models offer accurate and interpretable classification for early glaucoma detection.
• Average and 7-hour RNFL thicknesses demonstrate significant discriminant power.
• This approach enhances the ability to differentiate between healthy and early glaucoma eyes.