Abstract
Purpose :
To develop an interpretable machine learning (ML) model to predict anti-VEGF treatment requirements for patients with neovascular age-related macular degeneration (nAMD).
Methods :
Patients from the ranibizumab 0.5 and 2.0 mg as-needed arms of HARBOR (NCT00891735) who received monthly anti-VEGF injections during a 3-month initiation phase were included. Boundaries of five retinal layers, intra- and subretinal fluid, subretinal hyperreflective material (SHRM), and pigment epithelial detachment (PED) were automatically segmented using ML-based algorithms from spectral-domain optical coherence tomography (SD-OCT) volume scans acquired at each visit. Segmentation results were used to extract quantitative features of layer and fluid features (69 layer and 36 fluid features). BCVA and central subfield thickness (CST) measured at the 3 visits were also included. Low and high treatment groups were defined as requirement of ≤5 or ≥16 injections, respectively, in the 21 months after the initiation phase. Extreme gradient-boosting ML model was used for binary classification (low or high treatment) using stratified 5-fold cross-validation. Feature importance was analyzed using SHapley Additive exPlanations (SHAP).
Results :
Data from 363 patients were analyzed. Low and high treatment groups included 82 and 83 patients, respectively, with mean (±SD) area under the receiver operating characteristic curve scores of 0.81±0.06 and 0.80±0.08 (Fig 1). Low treatment need was most strongly associated with low presence of detected PED at month 2. High treatment need was most strongly associated with low presence of SHRM at month 1, low presence of IRF at day 0 and presence of IRF at month 1 (Fig 2).
Conclusions :
This exploratory study showed the feasibility of identifying low or high treatment needs for patients with nAMD using predefined imaging features from automated fluid and layer SD-OCT segmentations. Further confirmation of model performance will contribute to future development of personalized healthcare algorithms.
This is a 2021 ARVO Annual Meeting abstract.