Unsupervised Gaussian Mixture-Model With Expectation Maximization for Detecting Glaucomatous Progression in Standard...

Area of Science:

• Ophthalmology
• Machine Learning
• Medical Data Analysis

Background:

• Glaucoma is a progressive optic neuropathy leading to visual field (VF) defects.
• Accurate detection of glaucomatous progression is crucial for timely intervention.
• Current methods for detecting progression may lack sensitivity or localized information.

Purpose of the Study:

• To validate Gaussian mixture-model with expectation maximization (GEM) and variational Bayesian independent component analysis mixture-models (VIM) for detecting glaucomatous progression.
• To assess the performance of GEM-progression of patterns (POP) and VIM-POP in identifying VF defect patterns.
• To compare GEM-POP and VIM-POP against established methods like point-wise linear regression (PLR).

Main Methods:

• GEM and VIM models were used to cluster cross-sectional abnormal and normal VFs.
• Clusters were decomposed into independent axes, and a confidence limit (CL) of stability was established.
• Sensitivity for detecting progression was assessed in eyes with known progressive glaucomatous optic neuropathy (PGON), comparing GEM-POP and VIM-POP to PLR, PoPLR, LR of MD, and VFI.

Main Results:

• GEM and VIM models demonstrated high sensitivity and specificity in detecting glaucomatous VFs (GEM: 89.9% sensitivity, 93.8% specificity; VIM: 93.0% sensitivity, 97.0% specificity).
• Receiver operating characteristic (ROC) curve areas showed strong performance for VIM-POP (0.82) and GEM-POP (0.86).
• GEM-POP exhibited significantly higher sensitivity to PGON compared to PoPLR, LR of MD, and VFI.

Conclusions:

• Gaussian mixture-model with expectation maximization (GEM-POP) demonstrated superior sensitivity in detecting progressive glaucomatous optic neuropathy (PGON).
• These unsupervised machine learning models provide localized progression information, enhancing detection capabilities.
• Assessing longitudinal changes in localized glaucomatous defect patterns identified by machine learning can improve glaucoma progression detection.