Purchase this article with an account.
Filippo Arcadu, Fethallah Benmansour, Andreas Maunz, Jeffrey Ryuta Willis, Marco Prunotto, Zdenka Haskova; Deep learning algorithm for patient-level prediction of diabetic retinopathy (DR) response to vascular endothelial growth factor (VEGF) inhibition. Invest. Ophthalmol. Vis. Sci. 2019;60(9):2806. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
DR is the leading cause of blindness among working-age adults due to vision threatening complications such as diabetic macular edema (DME), and the risk of vision loss increases with DR worsening. Several large studies found that VEGF inhibition with ranibizumab regressed DR in a significant proportion of patients with DR with or without DME, and that baseline DR severity was predictive of DR response to this therapy. However, there is a lack of tools to predict patient-level DR treatment response. As part of Roche’s comprehensive initiative in Ophthalmology Personalized Healthcare, we have developed a predictive algorithm of individual treatment response to help inform the optimal timing for start of ranibizumab therapy in DR.
RIDE (NCT00473382) and RISE (NCT00473330) were identical double-masked phase 3 trials of ranibizumab in patients with DR and vision loss due to DME (N = 759). DR severity was prospectively graded on stereoscopic 7-field color fundus photographs (CFPs) by masked reading center evaluators using the Early Treatment Diabetic Retinopathy Study (ETDRS) Diabetic Retinopathy Severity Scale (DRSS). The 7-field CFPs from patients treated with monthly ranibizumab 0.3 mg or 0.5 mg were used to train and validate deep learning algorithms predictive of ≥2-step DRS improvement after 6, 12, and 24 months from baseline. The performance of the deep learning (DL) models was assessed through calculation of the area under the curve (AUC).
Manually detected rates of ≥2-step improvement in treated eyes at month 6, 12, and 24 from baseline were 30.7%, 33.7%, and 40.1%, respectively. At 6 months, the best DL algorithm was able to predict ≥2-step DRSS improvement with an AUC of 0.82 ± 0.03, sensitivity 87% ± 8%, and specificity 72% ± 6%. At 12 months, the best DL algorithm had an AUC of 0.85 ± 0.05, sensitivity 81% ± 10%, and specificity 82% ± 13%. At 24 months, the best DL algorithm had an AUC of 0.87 ± 0.04, sensitivity 84% ± 9%, and specificity 82% ± 7%.
Our work demonstrated feasibility of a deep learning algorithm that can predict response to anti-VEGF therapy from the baseline color fundus photographs (CFPs) at an individual patient level. CFPs from DR patients treated with anti-VEGFs in the real world are needed for further validation and to expand the applicability of the current pilot algorithm.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only