Abstract
Purpose :
To identify imaging biomarkers and patient characteristics that predict future need for anti-VEGF therapy in diabetic retinopathy (DR) using an integrative machine learning predictive model.
Methods :
Retrospective image analysis study of anti-VEGF naive eyes with DR who underwent ultra-widefield fluorescein angiography (UWFA) and optical coherence tomography (OCT). Quantitative UWFA analysis was performed using a custom semi-automated platform for leakage index (LI), ischemic index, microaneurysm (MA) count and vascular density (VD) parameters (skewness. kurtosis, variance, and zero VD index). OCT scans were reviewed for mean central subfield thickness (CST) and the presence of diabetic macular edema (DME). Two random forest classifiers grown with 1000 trees using 3 randomly sampled features as candidates at each split were created to differentiate eyes that required anti-VEGF within 3 months (including on the day of UWFA) and after 3 months of the imaging visit from the eyes that did not require any anti-VEGF during the follow-up period. Features including age, gender, follow-up time, visual acuity (VA), panretinal LI, macular LI, ischemic index MA count, CST, presence of DME and VD parameters were selected for model testing.
Results :
Eyes requiring anti-VEGF within 3 months (n=38) of the UWFA imaging session and eyes requiring anti-VEGF after 3 months (n=34) following the imaging visit were compared to the eyes that did not require any anti-VEGF treatment (n=101), independently. The mean follow-up time was 22 (11- 43) months. The area under the curve (AUC) for the anti-VEGF initiation within 3 months model was 0.93 ± 0.03 with the most important features being VA, CST, macular LI, and total LI (importance: 0.211, 0.197, 0.100 and 0.069 respectively). The AUC for differentiating eyes that required anti-VEGF treatment after 3 months vs eyes that did not require any anti-VEGF treatment was 0.77 ± 0.04 with the four most important features being VA, macular leakage index, total leakage index, and follow-up time (importance: 0.112, 0.112, 0.102 and 0.100 respectively).
Conclusions :
A ML predictive model was successful identifying eyes requiring immediate and future anti-VEGF therapy. The most important predictive features in both models included VA and macular LI.
This is a 2021 ARVO Annual Meeting abstract.