Enhancing central visual field loss representation with a hybrid unsupervised approach.

Area of Science:

• Ophthalmology
• Data Science
• Computer Vision

Background:

• Central visual field (VF) loss is a key indicator of various ocular diseases.
• Accurate representation and prediction of VF loss progression are crucial for patient management.
• Existing methods may lack the granularity to capture individual patient patterns effectively.

Purpose of the Study:

• To develop and validate a hybrid unsupervised approach for representing individual central visual field (VF) loss.
• To compare the efficacy of a hybrid approach combining archetypal analysis (AA) and fuzzy c-means (FCM) against AA alone in analyzing VF data.
• To enhance the prediction of central VF progression using a novel decomposition technique.

Main Methods:

• Utilized a dataset of 7927 10-2 VF tests from 3328 patients across five hospitals.
• Implemented a hybrid approach integrating archetypal analysis (AA) and fuzzy c-means (FCM) for pattern identification and VF decomposition.
• Employed supervised learning for mean deviation (MD) change prediction and linear mixed-effects models to analyze MD slope relationships.

Main Results:

• Identified 10 distinct archetypes representing 10-2 VF test patterns.
• The hybrid AA-FCM approach demonstrated superior performance in predicting MD change over the AA-only method, evidenced by lower mean squared error and higher Pearson correlation.
• Linear mixed-effects models indicated that the hybrid AA-FCM approach provided a better fit for predicting MD slope, with significant reductions in AIC and BIC scores.

Conclusions:

• The hybrid AA-FCM approach effectively visualizes central VF tests into characteristic patterns.
• This method offers enhanced prediction of central VF progression with minimized projection loss compared to the AA single approach.
• Baseline VF characteristics, such as inferior and bilateral hemifield loss, were associated with faster central VF progression.