Abstract
Purpose :
The purpose of this study was to identify baseline spectral-domain optical coherence tomography (SD-OCT) features for predicting response to anti-VEGF therapy in patients with neovascular AMD, and to build a model capable of providing individual patient predictions.
Methods :
Clinical trial data of patients (N=793) with neovascular AMD treated with ranibizumab from all arms of the HARBOR study (NCT00891735) was used to train and validate the models to predict response. Response to anti-VEGF therapy was defined as Best Corrected Visual Acuity (BCVA) ≥20/40 at month 12 in those subjects with BCVA ≤ 20/40 at baseline. Baseline age, sex, SD-OCT, and BCVA data were included. Automated segmentation of retinal layers and fluid-filled regions over a 6 X 6mm cube of SD-OCT images centered on the fovea was used to extract 62 SD-OCT features: 44 layer related features in 2-D space, 9 layer related in a 3D space and 9 fluidic area related features in 3-D space. To gain insight into the CatBoost model trained from the data, SHAP (SHapley Additive exPlanations) method was used to interpret patient-level model predictions. Model performance was assessed in terms of area under the receiver operating curve (AUROC) using 5-fold cross validation.
Results :
AUROC to predict response at month 12 using only baseline data was 0.77 (95% CI 0.73 – 0.82). Baseline BCVA, central subfield thickness, central subfield volume, and intra-retinal fluidic volume were among the most impactful measurements to predict response according to SHAP analysis, driving predictions, up, down, down, and up, respectively, in favor of reaching BCVA ≥20/40.
Conclusions :
We proposed and evaluated a machine learning methodology to predict probability of achieving functional BCVA from SD-OCT scans taken at treatment initiation. In the new era of therapies targeting multiple pathways in the management of neovascular AMD, the results of this retrospective analysis allow identification of patients who are likely to be good responders to anti-VEGF, thus enabling selection of appropriate patient population for the novel therapies and a precision medicine approach.
This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.