Validation of a Machine Learning Approach to the Analysis of Multifocal Electroretinograms for Hydroxychloroquine...

Hydroxychloroquine (HCQ) retinopathy is detectable through multimodal ophthalmic screening, yet individual diagnostic tests each have inherent limitations. Machine learning may simplify monitoring, but few HCQ screening algorithms have undergone rigorous external validation. This study evaluated the clinical utility of the Multifocal Electroretinogram Classification Interface (MERCI) algorithm by assessing its ability to predict HCQ retinopathy compared with diagnoses derived from American Academy of Ophthalmology (AAO) guidelines.

Patients referred for HCQ toxicity screening underwent perimetry, spectral-domain optical coherence tomography, fundus autofluorescence photography, and multifocal electroretinogram (mfERG) testing. Performance was assessed in a temporal dataset (participants overlapping with the development dataset) and a novel dataset (newly enrolled participants). Reference diagnoses were assigned using AAO guideline-based decision trees informed by ancillary tests results. MERCI generated corresponding mfERG-based predictions, which were compared with reference diagnoses to compute performance metrics for each dataset.

Using the decision trees, the temporal (n = 145) and novel (n = 300) datasets showed retinopathy prevalences of 12.4% and 9.7%, respectively. MERCI yielded 63/145 (43.4%) and 115/300 (38.3%) false-positive cases, respectively. MERCI achieved strong sensitivity (1.000 temporal; 0.931 novel) and negative predictive value (NPV) (1.000 temporal; 0.987 novel), with comparable area under the curve between datasets (0.759 temporal; 0.805 novel).

MERCI retained high sensitivity and NPV when validated against temporally distinct cohorts using a clinical definition of HCQ toxicity. This suggests MERCI's potential as a screening tool for HCQ retinopathy.

The mfERG-based machine learning models may improve screening efficiency and clinical decision-making for HCQ users.