Rao, Anil; Chandra, Shruti; Sivaprasad, Sobha; (2022) Prediction of treatment response in patients with neovascular age-related macular degeneration

Abstract

Neovascular age-related macular degeneration (nAMD) is a common cause of visual impairment, and is currently treated with intravitreal anti-vascular endothelial growth factor agents such as aflibercept. While these treatments may improve visual acuity (VA) in some patients, clinicians cannot currently predict who is likely to benefit before treatment starts. The aim of this study is to explore the effectiveness of using Deep Learning approaches to train models for predicting whether a patient's VA will respond favourably to three months of aflibercept therapy, using pre-treatment OCT images and clinical/demographic variables. We train a number of models using standard machine learning, Deep Learning transfer learning, and fully trained Deep Learning approaches in two experiments using outcomes based on the VA at 4-10 weeks after the final dose. In experiment one, we trained models to predict whether the VA will be at least 54 Early Treatment Diabetic Retinopathy Study (ETDRS) letters, while in experiment two we trained them to predict whether the VA will have increased by 10 or more letters. Model prediction quality was assessed using the Area Under the Curve (AUC) of the Receiver-Operating-Characteristic (ROC) curves. We found that all models performed significantly better than chance in both experiments, except for the fully trained Deep Learning model using just images in experiment two. The best performing model for experiment one was the Deep Learning transfer model using images and clinical/demographic variables (AUC=0.901), while in experiment two, none of the Deep Learning approaches performed better than a random forest using only clinical/demographic variables (AUC=0.751). Our experiments suggest that different Deep Learning approaches are required for predicting the second outcome if we want the models to perform better than those that use clinical/demographic variables alone.