Machine Learning Applied to Visual Fields of Dominant Optic Atrophy Patients

Area of Science:

• Ophthalmology
• Machine Learning
• Genetics

Background:

• Dominant optic atrophy (DOA) is a genetic optic neuropathy characterized by progressive vision loss.
• Understanding the visual field (VF) patterns associated with DOA is crucial for diagnosis and management.
• Existing methods may not fully capture the spectrum of VF defects or their relationship to genetic subtypes.

Purpose of the Study:

• To identify and quantify characteristic visual field (VF) patterns in patients with dominant optic atrophy (DOA).
• To utilize the archetypal analysis (AA) machine learning algorithm for this purpose.
• To explore genotype-phenotype correlations by linking VF patterns to OPA1 mutation subtypes.

Main Methods:

• Retrospective analysis of 30-2 or 24-2 Humphrey Visual Field tests from 144 patients with molecularly confirmed DOA (OPA1 heterozygous mutation).
• Development of an AA model using a training set (80% of VFs) to decompose VFs into archetypes (ATs).
• Correlation analysis between AT weights and visual acuity (VA)/mean deviation (MD), and statistical comparison of AT weights across mutation subtypes.

Main Results:

• The developed DOA-AA model, comprising eight ATs, demonstrated high performance (R2 = 0.88) in the test set.
• Archetypes representing central/ceco-central scotoma and superior defects showed the highest weights.
• Defect severity correlated with MD, while total loss ATs correlated with VA; mutation subtypes associated with worse outcomes had significantly higher weights for more severe ATs.

Conclusions:

• The AA model successfully identified and quantified VF patterns specific to DOA.
• The study supports a clinical genotype-phenotype association, with VF AA decomposition reflecting disease severity.
• AA provides an objective method for quantifying VF defects in DOA, offering functional insights based on genetic information.