Abstract
OBJECTIVE
Despite guidelines for hydroxychloroquine (HCQ) toxicity screening, there are clear challenges to accurate detection and interpretation. In the current report, the feasibility of automated machine learning (ML)-based detection of HCQ retinopathy and prediction of progression to toxicity in eyes without preexisting toxicity has been described.
DESIGN
Retrospective, longitudinal cohort study.
SUBJECTS
Subjects on HCQ therapy.
METHODS
This was an institutional review board-approved, retrospective, longitudinal image analysis of 388 subjects on HCQ. Multilayer, compartmental, retinal segmentation with ellipsoid zone (EZ) mapping was used to harvest quantitative spectral-domain (SD)-OCT biomarkers. Using a combination of clinical features (i.e., cumulative HCQ dose and the duration of therapy) and quantitative imaging biomarkers (e.g., volumetric EZ integrity and compartmental measurements), ML models were created to detect toxicity and predict progression based on ground-truth OCT-based toxicity readings by 2 masked retina specialists. Furthermore, 10-fold cross-validation was performed.
MAIN OUTCOME MEASURES
The model performance was visualized using receiver operator curves and calculating the area under the curve (AUC). The corresponding sensitivity and specificity values were evaluated for the feasibility of HCQ toxicity screening and prediction.
RESULTS
The prevalence of HCQ toxicity in this cohort of 388 patients was 9.8% (n = 38). Twenty-one eyes progressed to toxicity during follow-up. OCT-based features (i.e., partial EZ attenuation, EZ volume, outer nuclear layer volume, and compartmental thicknesses) and clinical features (i.e., HCQ daily dose, HCQ cumulative dose, and duration of therapy) showed significant differences between the toxic and nontoxic groups. Percentage area with partial EZ attenuation (i.e., percentage of the macula with an EZ-retinal pigment epithelium thickness of ≤ 20 μm) was the most discriminating single feature (toxic, 35.7 ± 46.5%; nontoxic, 1.8 ± 4.4%; P < 0.0001). Using a random forest model, high-performance, automated toxicity detection was achieved, with a mean AUC of 0.97, sensitivity of 95%, and specificity of 91%. Furthermore, the toxicity progression prediction model had a mean AUC of 0.89, with a sensitivity and specificity of 90% and 80%, respectively.
CONCLUSIONS
This report described the feasibility of high-performance automated classification models that used a combination of clinical and quantitative SD-OCT biomarkers to detect HCQ retinal toxicity and predict progression to toxicity in cases without toxicity. Future work is needed to validate these findings in an independent dataset.
Collapse