Eyeworld

EW RETINA 106 by Liz Hillman EyeWorld Staff Writer Machine learning program developed to predict AMD progression Program predicts drusen regression, which is associated with progression to late AMD R esearchers out of the Medi- cal University of Vienna's Department of Ophthal- mology and Optometry recently published a paper that describes a possible method to predict age-related macular degener- ation (AMD) using optical coherence tomography (OCT) and a machine learning computer model. 1 The study was selected as one of three top abstracts out of 6,000 submitted to the 2017 Association for Research in Vision and Ophthalmology annu- al meeting. While the presence of drusen can indicate early-stage AMD, this research is based on the phenom- enon of drusen regression, which can be indicative of progression from intermediate to late-stage AMD. Being able to predict drusen regression using OCT and machine learning is what Hrvoje Bogunovic, PhD, senior postdoc researcher at the university's Christian Doppler Laboratory for Ophthalmic Image Analysis, and his co-investigators set out to do. "The primary goal of our work was to measure known biomarkers associated with the risk of AMD and see how well their combination can predict the incoming drusen regres- sion," Dr. Bogunovic said. "Because the model was limited to the known biomarkers, it mostly confirms their importance. However, we found that in addition to the size, the internal drusen appearance on OCT was important as well." Bogunovic et al. described previous research that looked at various drusen properties as it related to AMD progression. Ouyang et al., for example, "found the presence of HRF overlying dru- sen and the heterogeneous internal drusen reflectivity to be related with the local onset of atrophy in the ensuing months. Querques et al. reported calcifications inside the regressing drusen," Bogunovic et al. wrote. 2,3 "An important distinction of our approach is that we obtained a personalized predictive model at the level of individual drusen, which enabled us to generate estimates of personalized future regression maps …," Bogunovic et al. explained. "In addition, to the best of our knowl- edge this is the first time quantita- tive properties of [hyperreflective foci] were used and not just the status of their presence." Bogunovic et al. observed patients with AMD every 3 months using OCT for at least 12 months up to 60 months. Several metrics of drusen observed on OCT were recorded, and changes in the drusen were analyzed over time with a ma- chine learning program to create a risk score and predict a timeline for regression of each drusen. The predictive model created by the researchers was then trained and evaluated with 61 eyes (38 patients) over a mean follow-up period of nearly 38 months. At baseline, 944 drusen were identified, and 249 (26%) regressed during the follow-up period. Predictions within the first 2 years were more accurate. Dr. Bogunovic said the model is currently built using only the base- line scan and one follow-up scan 3 months later, allowing it to produce prediction of drusen regression and thus a prediction of disease progres- sion. He added, however, that one would expect a model using longer follow-up data to perform better due to AMD being a slowly progressing disease and some changes not being observable within the 3-month timeframe. Dr. Bogunovic said that once validated, OCT and this machine learning algorithm could be used by ophthalmologists to predict when a patient might progress, allowing them to be most effective in timing their treatment for late-stage AMD. "Ophthalmologists would supply the collected OCTs of their patients to the model in the form of, [for example], a cloud-based service, which would provide them with the expected prognosis. They can then use such provided prediction as a second, computer-generated opin- ion," Dr. Bogunovic said. "Currently, every ophthalmolo- gist builds his or her own predictive model in his or her head based on knowledge and experience," he said. "Machine learning models have the advantage that they can quickly sift through large data sets, going through more OCT scans than any single ophthalmologist can experi- ence during practice, and providing a more objective prognosis." In addition to detecting ear- ly changes from intermediate to late-stage AMD to drive potentially vision-saving therapy, Dr. Bogunovic envisions another benefit of this predictive model. He said it could be used to identify high-risk subjects who could then be recruited for clinical trials, making them shorter in duration and smaller in size, in the hope of bringing down the cost of trials and increasing chances for successful drug discovery. In addition to analyzing OCT scans qualitatively, Dr. Bogunovic said ophthalmologists should ask OCT device suppliers to offer more quantitative analysis. Dr. Bogunovic said the next step is to train and validate the program with a larger cohort beyond this initial pilot study. "Besides that, as opposed to predicting drusen regression, which entails a high risk for developing advanced AMD, we are working on predicting progression to advanced AMD directly and at the same time being able to differentiate and subse- quently understand which eyes will progress to wet and which to the dry form of advanced AMD," he said. EW References 1. Bogunovic H, et al. Machine learning of the progression of intermediate age-related macular degeneration based on OCT imaging. Invest Ophthalmol Vis Sci. 2017;58:BIO141– BIO150. 2. Ouyang Y, et al. Optical coherence tomogra- phy-based observation of the natural history of drusenoid lesion in eyes with dry age-re- lated macular degeneration. Ophthalmology. 2013;120:2656–65. 3. Querques G, et al. Appearance of regressing drusen on optical coherence tomography in age-related macular degeneration. Ophthal- mology. 2014;121:173–9. Editors' note: Dr. Bogunovic has no financial interests related to his comments. Contact information Bogunovic: hrvoje.bogunovic@meduniwien.ac.at September 2017 Researchers created a predictive model of drusen regression using machine learning. Source: Hrvoje Bogunovic, PhD

• Blank Page
• Blank Page (1)