Abstract
Purpose :
While hydroxychloroquine (HCQ) retinal toxicity is rare, guidelines recommend routine monitoring, which provides an opportunity to detect other co-existing macular pathology that may compromise vision in this group of patients. We developed and tested a deep learning model to detect abnormal macular OCT scans from a cohort of patients on HCQ therapy, to explore other incidental abnormalities.
Methods :
A 4-class classification model was fine-tuned by applying RETFound, a newly described foundation model, to a large public dataset of labelled OCT images. This dataset has four different classes: Normal, Drusen, Diabetic macular oedema (DMO) and Choroidal neovascularisation (CNV). As part of a virtual HCQ retinal monitoring service provided by expert graders from Liverpool Ophthalmic Reading Centre, 364 patient episodes were graded for presence or absence of HCQ retinal toxicity and other incidental findings between September 2018-August 2020. We applied the newly trained model adapted for binary classification to predict normal vs. abnormal, using spectral domain OCT foveal scans and validated the predictions against graded data.
Results :
Average age of patients in this study was 55 yrs (17-91), with 83.7% being females. Of the 727 scans available, 667 were graded as normal, 60 had incidental abnormalities and no cases of HCQ toxicity were identified. Table 1 gives a breakdown of the different macular pathologies diagnosed. Predictions from our model resulted in an area under the ROC curve value of 0.99 (95% confidence interval (CI) 0.98, 1) and Youden's index of 0.9 (sensitivity 0.92, specificity 0.98) for the entire cohort. (Fig 1) For sensitivity = 1 (95% CI, 1.0, 1.0), specificity was 0.835 (95% CI, 0.804, 0.862). For specificity = 1 (95% CI, 1.0, 1.0), sensitivity was 0.833 (95% CI, 0.719, 0.917). 10 patients out of the 60 (7 with early/ intermediate AMD, 1 with geographic atrophy, 2 with epiretinal membrane) were misclassified as normal.
Conclusions :
Our model has shown promising results and could rule out common co-existing macular pathologies in 83% of OCT scans from this HCQ cohort in which 8.3% of eyes had other incidental abnormalities. The model also showed zero-shot learning capabilities by identifying new unseen conditions. Further refinement may help offset clinical workload by filtering out normal OCT scans in patients on HCQ therapy.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.