Purchase this article with an account.
Gregory Cyrus Ghahramani, Matthew Brendel, Qingyu Chen, Tiarnan D L Keenan, Kun Chen, Emily Y Chew, Zhiyong Lu, Yifan Peng, Fei Wang; Deep learning survival analysis on the progression to late AMD in the Age-Related Eye Disease Study. Invest. Ophthalmol. Vis. Sci. 2021;62(8):92.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Accurately predicting a patient’s risk of progressing to late age-related macular degeneration (AMD) is crucial for personalized medicine. We conduct deep learning survival analyses to predict AMD progression and to identify which sets of features, derived from Age-Related Eye Disease Study (AREDS) fundus photographs, are key predictors. While existing algorithms consider data from the present visit only, we evaluate how adding data from prior visits improves predictive performance.
The dataset comprised 3,768 AREDS participants without late AMD in either eye by year 3. Four survival models were trained and evaluated (Figure 1) to predict progression to late AMD at year 5 (2-year risk) and year 8 (5-year risk), based on data from year 3 only (Model 1) or years 0, 2, and 3 (Models 2-4). Models 1 and 2 were Cox proportional-hazard models (CoxPH). Models 3 used a deep neural network by concatenating visit information using a multilayer perceptron (MLP). Model 4 used a deep neural network to model the time-dependencies of visits using Long-Short-Term Memory (LSTM). Models were trained and evaluated using (a) drusen size and pigmentary abnormalities (i.e., akin to the 5-step AREDS Simplified Severity Scale, the clinical standard), or (b) drusen size, RPE depigmentation, retinal detachments, hemorrhages, and fibrosis, and geographic atrophy (Figure 2).
Figure 2 conveys the results using the two sets of risk factors. We observe that: 1) using deep learning to model the time-dependencies (LSTM) of feature set b outperformed all other models (AUC@2year: 0.918, AUC@5year: 0.930, c-index: 0.902) (Figure 2b); 2) incorporating data from prior visits improved predictive performance (up to 2% in AUC@2year, 1% in AUC@5year, 1% in c-index); 3) using feature set b improved performance over using drusen/pigment only (up to 6% in AUC@2year, 4% in AUC@5year, 3% in c-index), suggesting that the 5-step AREDS Simplified Severity Scale overlooks useful risk factors.
Existing algorithms for predicting progression to late AMD consider data from one time-point only. We demonstrate that incorporating AREDS grading features from previous years in a deep learning framework provides more accurate predictions than the existing clinical standards and basic survival models.
This is a 2021 ARVO Annual Meeting abstract.
Architectures of survival models
This PDF is available to Subscribers Only