Deep Learning Algorithm for the Diagnosis and Prediction of Hydroxychloroquine Retinopathy: An International, Multi-institutional Study.
PURPOSE: We present a deep learning algorithm-HCQuery-that detects the presence of hydroxychloroquine retinopathy and predicts its future occurrence from spectral-domain OCT (SD-OCT) images. DESIGN: We trained and validated a deep learning algorithm using retrospective SD-OCT images from patients taking hydroxychloroquine. SUBJECTS: The study involved a retrospective, nonconsecutive collection of 409 patients (171 positive for hydroxychloroquine retinopathy and 238 negative) and 8251 SD-OCT b-scans (1988 volumes) from 5 independent international clinical locations. METHODS: Imaging macular volumes from 2 different SD-OCT devices (Heidelberg Spectralis and Zeiss Cirrus) at 2 clinical sites were used to train and validate a convolutional neural network (EfficientNet-b4) to produce a likelihood of retinopathy score for each SD-OCT b-scan. Likelihood of retinopathy score were processed across SD-OCT volumes for an eye-level and patient-level binary decision output for the presence or absence of retinopathy. The adjudicated consensus of ≤3 independent retina specialists using patient clinical data and multimodal testing served as the reference standard for hydroxychloroquine retinopathy. The algorithm was tested on 4 withheld test sets, 1 internal (data set 1), and 3 external (data sets 3-5). The test sets were obtained in 2 countries (United States and United Kingdom) and represented 2 SD-OCT devices each with diverse acquisition parameters. MAIN OUTCOME MEASURES: Sensitivity, specificity, accuracy, negative predictive value, positive predictive value, area under the receiver-operator characteristic curve, and area under the precision-recall curve for the detection of hydroxychloroquine retinopathy either at the time of clinical diagnosis or ≤18 months in advance of clinical diagnosis. RESULTS: The algorithm discriminated hydroxychloroquine retinopathy at the time of clinical diagnosis as well as in advance of clinical diagnosis (mean: 220.8 days before clinical diagnosis; accuracy: 0.987 [95% CI: 0.962-1.00]; sensitivity: 1.00 [95% CI: 0.833-1.00]; specificity: 0.983 [95% CI: 0.952-1.00]; positive predictive value: 0.944 [95% CI: 0.836-1.00]; negative predictive value: 1.00 [95% CI: 0.937-1.00]). For eyes that developed retinopathy, it was identified as positive 2.74 years in advance of the clinical diagnosis on average. CONCLUSIONS: Our algorithm can detect retinopathy at all stages of disease, as well as predict retinopathy years in advance of clinical diagnosis. FINANCIAL DISCLOSURE(S): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.
Duke Scholars
Published In
DOI
EISSN
Publication Date
Location
Citation