MERCI: a machine learning approach to identifying hydroxychloroquine retinopathy using mfERG

• Published in: Documenta ophthalmologica 2022-08, Vol.145 (1), p.53-63
• Main Authors: Habib, Faisal, Huang, Huaxiong, Gupta, Arvind, Wright, Tom
• Format: Article
• Language: English
• Subjects: Accuracy; Anti-inflammatory agents; Autoimmune diseases; Eye; Hydroxychloroquine; Immunological diseases; Inflammation; Learning algorithms; Machine learning; Medicine; Medicine & Public Health; Ophthalmology; Original Research Article; Pathology; Patients; Quality of life; Retina; Retinopathy; Support vector machines; Toxicity; Vision
• ISSN: 0012-4486; 1573-2622
• DOI: 10.1007/s10633-022-09879-7

Purpose Hydroxychloroquine (HCQ) is an anti-inflammatory drug in widespread use for the treatment of systemic auto-immune diseases. Vision loss caused by retinal toxicity is a significant risk associated with long term HCQ therapy. Identifying patients at risk of developing retinal toxicity can help prevent vision loss and improve the quality of life for patients. This paper presents updated reference thresholds and examines the diagnostic accuracy of a machine learning approach for identifying retinal toxicity using the multifocal Electroretinogram (mfERG). Methods A retrospective study of patients referred for mfERG testing to detect HCQ retinopathy. A consecutive series of all patients referred to Kensington Vision and Research Centre between August 2017 and July 2020 were considered eligible. Eyes suspect for other ocular pathology including widespread retinal disease and advanced macular pathology unrelated to HCQ or with poor quality mfERG recordings were excluded. All patients received mfERG testing and Ocular Coherence Tomography (OCT) imaging. Presence of HCQ retinopathy was based on ring ratio analysis using clinical reference thresholds established at KVRC coupled with structural features observed on OCT, the clinical reference standard. A Support Vector Machine (SVM) using selected features of the mfERG was trained. Accuracy, sensitivity and specificity are reported. Results 1463 eyes of 748 patients were included in the study. SVM model performance was assessed on 293 eyes from 265 patients. 55 eyes from 54 patients were identified as demonstrating HCQ retinopathy based on the clinical reference standard, 50 eyes from 49 patients were identified by the SVM. Our SVM achieves an accuracy of 85.3% with a sensitivity of 90.9% and specificity of 84.0%. Conclusions Machine learning approaches can be applied to mfERG analysis to identify patients at risk of retinopathy caused by HCQ therapy.