Predicting future visual acuity outcomes from early morphologic and functional response in anti-VEGF treated retinal vein occlusion.

Abstract

Purpose : To determine predictive factors in longitudinal OCT data for visual acuity outcomes after 6 months in patients with branch retinal vein occlusion (BRVO) and central retinal vein occlusion (CRVO).

Methods : Baseline OCT scans and 3 monthly follow-up scans of patients with macular edema secondary to CRVO (n=144) and BRVO (n=96) were included. All patients received ranibizumab injections for three months followed by a PRN regimen.
All OCT scans were automatically spatially aligned using the fovea and optic disk center. Automatic segmentations of intraretinal cystoid fluid and subretinal fluid (IRF/SRF) were created using convolutional neural networks. From the segmentations, IRF and SRF volume was computed for each section of the fovea centered ETDRS grid. Total retinal thickness maps were automatically computed using the Iowa layer segmentation reference algorithm.
To predict visual acuity (VA) and VA change to baseline for month 6, elastic-net regularized linear regression models were trained on (1) BCVA letter scores at baseline, (2) BCVA values of the loading phase with their relative change from baseline, and (3) with additional segmented imaging information from the initial 3 months.
To assess the predictive power of the models, a 5-fold cross-validation setting was used. As error metrics, the mean absolute error (MAE) in BCVA letters, and the predicted R² were computed.

Results : Performance results of the prediction models are provided in Table 1. Adding BCVA response during the loading phase increased the predicted R² of absolute/relative BCVA from 0.51/-0.01 to 0.78/0.54 for CRVO, and from 0.19/0.24 to 0.69/0.71 for BRVO. Adding imaging features did neither reduce the MAE nor increase the predicted R².

Conclusions : The early response in BCVA score is a good predictor of visual acuity outcomes in ranibizumab-treated CRVO and BRVO eyes. Imaging information such as total retinal thickness and IRF/SRF did not add an additional predictive value to the model.

This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.