Machine Learning for Classification of Functional Phenotypes in Stargardt Disease from Full-Field Electroretinography

Abstract

Purpose : This study aims to test the applicability of machine learning (ML) to the analysis of full-field electroretinography (ERG) for the identification of clinically relevant phenotypic subtypes in patients diagnosed with ABCA4-retinopathy (Stargardt disease).

Methods : Patients with ABCA4-retinopathy who had undergone full-field were ascertained. All patients had a molecularly confirmed, likely-disease causing, genotype in the ABCA4 gene and a clinical presentation consistent with ABCA4-retinopathy. All ERGs adhered to the International Society for Clinical Electrophysiology of Vision (ISCEV) standard. Based on interpretation of the dark-adapted strong flash (DA 10.0) ERG, light-adapted single flash cone (LA 3.0) ERG and 30Hz flicker (LA 30Hz) ERG, data were subdivided into three groups by two experienced electrophysiologists: normal (group 1 n=344), consistent with cone dystrophy (group 2 n=45) or cone-rod dystrophy (group 3 n=210). For model development, ERG data were divided 80/20 at patient level into training and test datasets. Individual ERG traces, their derivatives, patient age, and pupil size were input into three logistic regression models corresponding to each of the above stimuli. The collated probabilities from an individual patient's ERG traces were input into a further model generating one phenotype prediction.

Results : Expert analysis and ML methods predicted phenotypic subtypes with 88.2% to 94.2% concordance in the unseen ERG data. In a 5-fold cross validation average test group accuracy was 91.1% with a kappa value of 0.83. Combining phenotypes 2 and 3 in a binary classification gave an average accuracy of 93.8%, sensitivity of 0.93, and specificity of 0.94.

Conclusions : Machine learning classification and human analysis of ERG phenotypes in ABCA4-retinopathy show a high degree of concordance. Machine learning methods have the potential to deepen understanding of retinal dysfunction and disease, previously dependent on specialist expertise and extensive training.

This is a 2021 ARVO Annual Meeting abstract.