Purpose
Current models analyzing risks factors for choroidal neovascularization (CNV) in early or intermediate age-related macular degeneration (AMD) patients have limited performance and are aimed for making predictions over long time intervals (> 5 years), limiting their utility in guiding real-world screening intervals or conducting preventive drug trials. We modified an experimental predictive model to improve short-time predictions and evaluated its accuracy in predicting imminent (< two months) CNV in an independent database acquired at a separate institution.
Methods
Using machine learning techniques we applied a predictive model that produces a score related to the chances of a CNV event at any desired short-term interval based on automated processing of SD-OCT images and patient historical and demographic information. The model was trained on 790 longitudinal clinic visits acquired in our institution over 6 years from 186 eyes presenting with early or intermediate AMD (mean age 77.8 years, 58.6% female), out of which 21 progressed. We tested the performance in an independent longitudinal dataset acquired at a separate institution consisting of 258 clinic visits from 29 eyes observed over 2.7 years presenting with early or intermediate AMD (mean age 72.9 years, 65.5% female), out of which 8 eyes developed CNV event during the course of the study. Eyes had an average of 6.33 clinic visits, with a mean follow-up time of 1.32 months between visits. We evaluated the accuracy of the predictions at 258 transitions (prediction at follow-up).
Results
The computational model was able to predict with 87.5% sensitivity and 76.52% specificity the probability of progression at follow up within a mean of 1.32 months (SD 1.73 months). Choosing an operating criterion of 100% sensitivity, specificity was 43.2%. The mean area under a receiver operating characteristic curve was 0.8 (95% confidence interval of (0.61, 0.88)).
Conclusions
Our improved predictive model demonstrated high potential for discerning patients which have higher risk of a CNV event within a mean of less than two months when evaluated in an independent dataset. A model of these characteristics may be used to assess subtle changes in an early and intermediate AMD phenotype that may be related to CNV risk with higher accuracy and time resolution than in current practice.