A pilot study using machine learning (ML) to predict treatment outcomes in patients with neovascular age-related macular degeneration (nAMD) using phase 2 trial data

Abstract

Purpose : To develop ML models to predict treatment outcomes in patients with nAMD using baseline (BL) characteristics and optical coherence tomography (OCT) imaging data from patients treated with faricimab in the phase 2 AVENUE (NCT02484690) trial.

Methods : 185 faricimab-treated eyes (80% training; 20% test) were included. 5-fold cross-validation was performed on the training set. Age, gender, best-corrected visual acuity (BCVA), central subfield thickness (CST), low-luminance visual acuity, and OCT images on study day 1, together with treatment assignment, were included in the model. Regression and binary classification models were developed to predict BCVA at month 9 and CST reduction of >35% at month 9, respectively. Symbolic models (linear model and extreme gradient boost tree) were trained on BL characteristics, and deep neural networks (DNNs; based on Inception-v3 with ImageNet weights) were trained on B-scans. Image data and BL characteristics were merged using: (1) a model stacking approach, which uses the prediction from the DNN as one of the input features for the symbolic model, and (2) a model averaging approach, which averages predictions from the DNN using OCT volume and from the symbolic model using BL characteristics.

Results : The BL linear model had an R² score of 0.30 and area under the receiver operating characteristic (AUROC) of 0.87. The BL DNN showed an R² score of 0.079 for BCVA regression and AUC of 0.70 for CST reduction binary classification. After model stacking with linear model, R² score and AUC were 0.32 and 0.87, respectively. The model averaging approach with linear model showed an R² score of 0.27 and AUC of 0.85.

Conclusions : The results suggest that the most predictive features for both BCVA and CST at month 9 were captured in BL measurements, and adding image data did not show significant improvements after stacking or averaging given the sample size. This pilot study highlights the potential for ML to support clinicians to make treatment decisions for optimal patient outcomes. To fully explore the predictive capacity of models using image data and ascertain benefits of combining image data with BL characteristics, the methodology needs to be validated on a larger data set.

This is a 2021 ARVO Annual Meeting abstract.

 

Fig 1. Model stacking vs model averaging

 

Table 1. Model performance in 5-fold cross-validation and on the holdout set