Feasibility of Machine Learning-based Assessment for Hydroxychloroquine Toxicity Utilizing Multi-Layer SD-OCT Feature Interrogation

Abstract

Purpose : To evaluate feasibility of a machine learning (ML)-based classifier for identifying eyes with underlying hydroxychloroquine (HCQ) toxicity utilizing quantitative higher-order SD-OCT features and clinical variables.

Methods : In this IRB-approved retrospective study, clinical characteristics and SD-OCT scans for eyes that were identified to be on HCQ and had undergone concurrent screening SD-OCT scans were collected. SD-OCT scans were analyzed with an automated multi-layer feature extraction platform that enabled assessment of ellipsoid zone (EZ) and outer nuclear layer (ONL) characteristics (e.g., thickness, volume, en face integrity mapping). Segmentation was reviewed and corrected by an expert image analyst. Masked qualitative SD-OCT review was performed by a retina specialist to identify eyes with HCQ toxicity. Univariate assessment was performed for feature comparison between the toxic (T) and non-toxic (NT) groups. A random forest classifier (RFC) with 5-fold random validation was trained using a 50% split between training and test data. Binary classification performance into Toxic (T) or Non Toxic (NT) groups was analyzed using a set of 8 biomarkers (clinical: 2, SD-OCT biomarkers: 6).

Results : This analysis included 385 eyes, with mean patient age of 56.8 years, mean duration on HCQ of 5.8 years, and mean HCQ dose of 4.84 mg/kg that were screened for HCQ toxicity. Utilizing the RFA classifier model had an AUC of 1.00 with sensitivity of 100% and specificity of 98.4% for our training set and an AUC of 0.90 with a sensitivity of 88.0% and specificity of 92.4% for detecting toxicity in the test set. The features noted to be of greatest impact were temporal ONL-RPE (Retinal Pigment Epithelium) thickness, mean central macular (i.e., 2-mm) ONL-RPE thickness, mean central macular EZ-RPE thickness, and cumulative dose of HCQ.

Conclusions : Utilizing a combined approach of outer retinal quantitative feature extraction and ML-based classification, a high performance model was created for detection of HCQ toxicity. Further research will be focused on validating these findings in prospective and external datasets. An automated risk assessment tool for HCQ toxicity could provide important feedback to clinicians for evaluating likelihood of underlying retinal damage and facilitate early detection.

This is a 2021 Imaging in the Eye Conference abstract.