Predicting geographic atrophy growth rate using optical coherence tomography extracted tabular features

Abstract

Purpose : To predict annualized growth rate of geographic atrophy (GA) using tabular features derived from the layer segmentation of optical coherence tomography (OCT) volumes from baseline visits and to identify the top risk factors.

Methods : A OCT segmentation model (EyeNotate) was utilized to segment OCT volume scans into retinal layers and drusen. Subsequently, eighteen numerical features including area and number of layer disruptions, average layer thickness, and length of area disruption reaching scan boundaries were derived. Ten clinical tabular features such as demographics (age, sex, smoking history), functional measurements (visual acuity, low luminescence deficits), manual readings from non-OCT imaging modalities (presence of RPD, fovea involvement of GA, the distance of GA to fovea on FAF, the connectivity of GA lesion of FAF, GA area) were also included. Baseline study eyes of patients with bilateral GA (NCT02247479; NCT02247531; and NCT02479386, n=1012) were split into training and validation sets. Lasso and XGBoost models were trained to predict the annualized GA growth rate measured on FAF using 5x3 nested cross-validation (3 inner folds for hyperparameter tuning and 5 predefined outer folds for performance reporting). Feature importance was determined via partial dependence plots (PDP), defined as the ratio of variance of PDP per feature versus the variance of sum of all PDPs.

Results : OCT derived features outperformed the clinical feature set and added performance on top of the clinical feature set (Table 1). The validation R2 (SE) for GA growth rate prediction for the model with both sets of features was 0.334 (0.022) for Lasso and 0.342 (0.012) for XGBoost. The two three factors identified by both models include baseline ellipsoid zone (EZ) area loss and ratio of retinal pigment epithelium (RPE) area loss to EZ area loss (Fig. 1).

Conclusions :
Baseline EZ loss area and ratio of RPE to EZ loss are top risk factors for future GA growth rate.

This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.

 

Fig. 1. Top 14 risk factors identified by the two models (Left: Lasso; Right: XGBoost), with demographic and functional features in blue, non-OCT imaging features in red, OCT features in black.

 

Table 1. The validation R2 of different models using different covariates.