Prediction of Function in ABCA4-Related Retinopathy Using Ensemble Machine Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Recessive Stargardt disease, or ABCA4-related retinopathy, poses challenges for clinical trials due to variability in functional tests.
• Traditional measures like full-field electroretinogram (ERG) and best-corrected visual acuity (BCVA) have prognostic value but are limited by patient burden and variability.

Purpose of the Study:

• To develop and evaluate an ensemble machine-learning model to predict functional outcomes in Stargardt disease.
• To differentiate Stargardt patients from controls and predict disease categories using microstructural imaging and patient data.

Main Methods:

• Utilized spectral-domain optical coherence tomography (SD-OCT) and patient data to train an ensemble machine-learning model.
• Developed 'inferred ERG', 'inferred visual impairment', and 'inferred BCVA' as novel outcome measures.
• Analyzed feature importance, identifying foveal status, outer retinal layer thickness, and age of onset as key predictors.

Main Results:

• Achieved high accuracy (up to 99.53 ± 1.02%) for 'inferred ERG' and 'inferred visual impairment'.
• Predicted BCVA values ('inferred BCVA') with a precision of ±0.3 LogMAR in 85.31% of eyes.
• Demonstrated that microstructural imaging can accurately estimate functional deficits.

Conclusions:

• 'Inferred ERG', 'inferred visual impairment', and 'inferred BCVA' provide accurate, quasi-functional parameters for Stargardt disease assessment.
• These novel measures offer potential for refined patient stratification and monitoring of treatment effects or disease progression in clinical trials.
• The machine-learning approach addresses limitations of traditional functional tests, reducing patient burden and variability.