Abstract
Purpose :
Models that can quantify and predict geographic atrophy (GA) growth would provide valuable insights to the design and efficacy of clinical trials. This work aims to develop learning-based prediction models for accurate GA estimation at variable timepoints using patient demographic and image-based variables.
Methods :
As part of the Age-Related Eye Disease Study 2 (AREDS2) clinical trial, data was collected from 890 participants (comprising 1215 eyes and 3411 study visits, spanning 1-4 yrs) who exhibited prevalent or incident GA. Color fundus photographs were graded for GA area, involvement and proximity to center, and GA configuration. Patient demographic information (i.e., age, race, gender, smoking status) was also collected. Random forest (RF) regression models were developed to predict GA growth rates using demographic and image-based predictor variables. Predictions were made for growth rates computed at 1 year (3296 pairs), 2 year (2215 pairs), 3 year (1361 pairs), and 4 year time intervals (719 pairs) and after combining data from all time spans. The performance was quantified using RMSE and correlation coefficient (R2) in a ten-fold cross-validation setting.
Results :
Growth rates measured using GA areas at different timespans were (mean ± stdev) 0.31 ± 0.43 mm/yr over 1 year, 0.29 ± 0.28 mm/yr over 2 years, 0.28 ± 0.22 mm/yr over 3 years, and 0.27 ± 0.19 mm/yr over 4 years. The average growth rate measured using all time intervals was 0.29 ± 0.33 mm/yr. The RF prediction model estimated growth rates with RMSE = 0.43 mm/yr (R2=-0.01) over 1 year, 0.26 mm/yr (R2=0.09) over 2 years, 0.20 mm/yr (R2=0.15) over 3 years, and 0.18 mm/yr (R2=0.10) over 4 years. After combining growth rates measured across different time spans, the RF model exhibited RMSE = 0.30 mm/yr (R2=0.05). Both measured growth rates and errors of the prediction models decreased with increasing measurement time-intervals.
Conclusions :
At longer time-spans, both GA growth rate variability and learning-based prediction accuracy worsened gradually. At shorter time-spans, measurement error could constitute a larger portion of the growth rate and could result in higher variability and more challenging estimation by prediction models. Future work investigates additional automated image-derived metrics to improve the prediction accuracy.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.