Selecting measures of visual function to classify diabetic retinopathy status: a cross-sectional study

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Diabetic retinopathy (DR) and diabetic macular edema (DMO) are leading causes of vision loss in diabetic patients.
• Accurate staging of DR severity is essential for timely intervention and management.
• Current diagnostic methods may not fully capture the spectrum of visual dysfunction in diabetic eye disease.

Purpose of the Study:

• To determine the most effective combinations of up to three visual function tests for classifying diabetic retinopathy severity.
• To provide detailed measurements from a comprehensive suite of visual function tests.
• To aid researchers and clinical trial planners in selecting optimal visual assessment tools for diabetic eye conditions.

Main Methods:

• 1901 eyes from 1032 participants were evaluated using nine visual function tests.
• DR and DMO severity were graded from fundus, ultra-widefield, and OCT imaging.
• Ensemble machine learning models analyzed combinations of visual function tests for three classification tasks using Area Under the Curve (AUC).

Main Results:

• Top-performing models achieved high accuracy (AUC ≥0.94) for all classification tasks.
• Distance visual acuity was a key component in models for distinguishing diabetes mellitus from healthy eyes and DR with DMO.
• Mesopic microperimetry was prominent in models differentiating DR without DMO from diabetes mellitus alone.

Conclusions:

• Specific combinations of visual function tests, including distance visual acuity, contrast sensitivity, and microperimetry, are highly effective for classifying diabetic retinopathy stages.
• These findings provide a data-driven approach for selecting the best visual function tests in research and clinical settings for diabetic eye disease.
• The study highlights the utility of machine learning in optimizing diagnostic test selection for complex ocular conditions.