Machine Learning Predicts Anti-VEGF Treatment Demand
By Jean Shaw
Selected By: Andrew P. Schachat, MD
Journal Highlights
Ophthalmology Retina, July 2021
Download PDF
Gallardo et al. assessed the potential of machine learning to predict low and high treatment demand in patients with neovascular age-related macular degeneration (AMD) or retinal vascular disease who received treat-and-extend injections in a routine clinical setting. They found that machine learning can predict treatment demand and potentially could be used to establish patient-specific treatment plans.
For this retrospective cohort study, the researchers evaluated 340 patients (377 eyes) with neovascular AMD and 285 patients (333 eyes) with retinal vascular disease. The latter group comprised 150 patients with retinal vascular occlusion (RVO) and 135 patients with diabetic macular edema (DME). All eyes were treated with either aflibercept or ranibizumab according to a predefined treat-and-extend protocol at the University of Bern, Switzerland. The study period ran from 2014 to 2018, and patients received anti-VEGF injections for at least one year.
The researchers defined eyes as low-, moderate-, or high-treatment demanders, using the average interval between treatments (low: ≥10 weeks; high: ≤5 weeks; moderate: all other eyes). They then trained two random forest models to predict the long-term treatment demand of a new patient. Both models used patient demographic information and morphological features automatically extracted from OCT volumes at baseline and after two consecutive visits. Mean area under the curve (AUC) of both models was measured.
In the cohort of patients with AMD, the researchers identified 127 low-, 42 high-, and 208 moderate-treatment demanders. Of those with RVO or DME, 61 patients were low-, 50 were high-, and 222 were moderate-treatment demanders. For patients with AMD, the mean AUC was 0.79 for both low and high demanders. For those with retinal vascular disease, the mean AUC was 0.76 and 0.78 for low and high demanders, respectively. To predict low treatment demand, only the information derived at baseline was necessary. In contrast, accurate prediction of high treatment demand required additional information from follow-up visits.
The original article can be found here.