Abstract
Purpose :
AMD is one of the major causes of blindness in the developed world. At present, no treatment for AMD. Early intervention can prevent the incidence of late AMD. Considering this, we have developed a prediction model for identifying individual an who will progress from intermediate to late dry or wet AMD.
Methods :
The proposed prediction model considers the individuals who are at an intermediate stage of AMD and predicts if s/he will be progressed to late AMD. The model uses an individual’s color fundus image, socio-demographic and clinical data such as age, sex, race, smoking history, diabetes, sunlight exposure, and fellow eye AMD status. Multiple deep learning based classifiers are built using the color fundus images to classify the images into one of the 12-levels of AMD severity scale, defined by Age-Related Eye Diseases Study (AREDS). Six of the best performing classifiers are then taken to get probability values for every image. Each classifier will output an array of size 12 for each image. A total of 72 probabilities from 6 models are concatenated to obtain a feature array for each image. Further, the socio-demographic parameters and AMD in the fellow eye information are concatenated to augment the feature array. A final machine learning model is built with these features using Logistic Model Tree. We used the patients’ image and socio-demographic parameters from AREDS study. For 1-year AMD incident prediction, we have 135 dry AMD and 182 wet AMD images. For 2-year, we have 148 dry AMD and 195 wet AMD images.
Results :
The prediction model achieved sensitivities and specificities of 90.7% and 84.79% for 1-year (accuracy 87.91%), and 92.2% and 84.3% (accuracy 87.57%) for 2-year late AMD (any type - dry or wet) incident. For late dry AMD, the model achieved 65.94% sensitivity and 95.68% specificity for 1-year incidence and 68.69% sensitivity and 86.2% specificity for 2-year incidence. For wet AMD, the model achieved 68.17% sensitivity and 84.04% specificity for 1-year incidence and 66.23% sensitivity and 93.4% specificity for 2-year incidence.
Conclusions :
We have demonstrated that the prediction of the late dry or wet AMD in short-term is possible and the performance is promising. The model will allow quick analysis of fundus images, help Ophthalmologists with a better decision, a higher confidence level and treat AMD at an early stage.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.