Skip to main content

  • Using Machine Learning to Forecast Visual Outcomes in Wet AMD

    By Lynda Seminara
    Selected By: Stephen D. McLeod, MD
    Add to My Bookmarks

    Journal Highlights

    Ophthalmology, July 2018

    Download PDF

    Rohm et al. established a comprehensive data warehouse and applied machine-learning algorithms to predict visual acuity (VA) outcomes of patients who received 3 intravitreal injections for neovascular age-related macular degeneration. They were able to predict VA at 3 months, and results were com­parable to actual measured VA (called ground truth in this study). Moreover, the study showed that the 3-month predictions of VA were more accurate than were the 12-month forecasts.

    Five algorithms were used in the study (AdaBoost.R2, Lasso, Gradient Boosting, Random Forests, and Ex­tremely Randomized Trees). Clinical data obtained from the data warehouse included VA measurements drawn from electronic health records and findings from optical coherence tomography. To provide a quality measure, both mean absolute error (MAE) and root mean square error (RMSE) were calculated for each algorithm. (RMSE penalizes outliers, allowing selection of the most robust algorithm.)

    Three-month forecasts were made for 653 patients (738 eyes). Mean VA before the first injection of an anti–vascular endothelial growth factor (anti-VEGF) drug was 0.54 logMAR (±0.39). Of these patients, 456 (508 eyes) had sufficient follow-up data for the 12-month assessment. Among the 508 eyes, mean VA before the initial injection was 0.56 logMAR (±0.42). The main outcome measure was the difference in predicted versus ground-truth logMAR VA at months 3 and 12 after the start of anti-VEGF therapy.

    Analyses showed that the MAE of predicted VA over ground truth was 0.11 logMAR (5.5 letters) for the 3-month prediction and 0.16 logMAR (8 letters) for the 12-month prediction. The 12-month RMSE was lowest (0.2 logMAR; 10 letters of change) if data from the 4 visits before the third injection were taken into account, but this was not demonstrat­ed for the MAE. The best-performing algorithm was the Lasso L1 regularized linear model. Although 12-month forecasts were not as accurate as their 3-month counterparts, they may be helpful for encouraging patients to stay on their therapy, the authors said.

    The original article can be found here.