Anomaly detection in OCT images in patients on hydroxychloroquine: A data science approach.

Abstract

Purpose : The research aims to use OCT to detect central retinal abnormalities in patients on hydroxychloroquine therapy.
This is a tool that would allow clinicians in community clinics (e.g., rheumatology) to triage these patients and refer only abnormal patients to ophthalmologists, thereby relieving the burden on the ophthalmology clinic.

Methods : This is an institutional review board-approved retrospective and longitudinal image analysis of subjects on HCQ to find features to detect abnormalities in OCT and autofluorescence. After having the ethics approval, we collected OCT images for the patient throughout several years of taking HCQ treatment. The automated segmentation software from Heidelberg's Spectralis (SD)-OCT system was used to harvest quantitative biomarkers in the spectral domain (SD)-OCT. We developed a model using these extracted OCT biomarkers to detect HCQ toxicity and other central retinal abnormalities. In the abnormal group, we include retinal toxicity, wet and dry AMD, and any other internal or external changes.

Results : Our study involves 158 eyes of 79 patients. We chose to study each eye independently to augment our data set. The prevalence of abnormalities in this sample is 17,1% (27 eyes), of which 6 eyes show changes in the inner layers and 8 in the outer layers. Firstly, we verified that quantitative biomarkers from normal and abnormal distributions are statistically significant (p-value= 0.0431). We extracted features from retinal layer detections using principal component analysis and then trained an ensemble classifier to detect abnormal cases. We validated the system with 10-fold cross-validation. It displayed a sensitivity of 99,3% and a specificity of 62,5%.

Conclusions : Recent studies have acknowledged the feasibility of automated machine learning (ML)-based detection of HCQ retinopathy. Our study aims to validate these findings in an independent dataset and create an automated tool to triage patients under HCQ treatment in order to detect HCQ toxicity along with other abnormalities. The current results show the presence of discriminative features in abnormal cases. While our data collection continues to increase, it will enable to extract more discriminative features with deep-learning models.

This abstract was presented at the 2023 ARVO Imaging in the Eye Conference, held in New Orleans, LA, April 21-22, 2023.