Machine Learning-based Prediction for Development of New Hydroxychloroquine Retinopathy based on Quantitative Outer Retinal SD-OCT Features and Clinical Variables

Abstract

Purpose : Identifying eyes without current hydroxychloroquine (HCQ) retinal toxicity that are at high-risk to progress to toxicity could facilitate clinician intervention for HCQ dosing modifications. The purpose of the current study is to analyze predictive capability of a machine learning (ML)-based model for identifying eyes that progressed to HCQ toxicity by using a combination of quantitative higher-order SD-OCT features and baseline clinical variables.

Methods : This was an IRB-approved retrospective cohort study of 371 subjects on HCQ without evidence of toxicity where data was collected at baseline and final screening visits. OCTs from all timepoints were analyzed with an automated multi-layer compartmental segmentation system to provide quantitative outer retinal parameters. A selection of baseline clinical features (i.e., cumulative HCQ dose, duration of therapy) and quantitative SD-OCT biomarkers (e.g., volumetric ellipsoid zone (EZ) integrity and compartmental measurements) were compared between eyes that progressed to toxicity (progressors) and eyes that did not progress (nonprogressors). A random forest classifier with 10-fold cross validation was trained using baseline features selected based on univariate analysis.

Results : This analysis includes 371 subjects on HCQ with 21 progressing to hydroxychloroquine retinopathy. Baseline features showed highly statistically significant differences in means between progressors and non-progressors: partial EZ attenuation percentage (9.4±8.7% vs 5.6±3.7 %; p<0.0001, panmacular ONL volume (3.4±0.6 mm3 vs 4.0±0.5 mm3 ; p<0.0001), panmacular EZ volume (1.0±0.2 mm3 vs 1.2±0.2 mm3 ; p<0.0001), ONL-RPE thickness 1 mm nasal to fovea (116.4±21.2 μm vs 128.6±17.8 μm; p<0.0045), and daily HCQ dose (5.3±1.4 mg/kg vs 4.8±1.5 mg/kg; p=0.02). The random forest classifier demonstrated a mean area under curve of 0.89 (0.77-0.94) with sensitivity and specificity of 90% and 80% respectively in identifying progressors.

Conclusions : Combining targeted quantitative SD-OCT biomarkers with HCQ dose enabled the development of a highly discriminating ML-based classification model for the prediction of progression to HCQ toxicity. Future research will be focused on analyzing this model in external datasets and prospective studies.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.