Machine Learning Can Predict Anti–VEGF Treatment Demand in a Treat-and-Extend Regimen for Patients with Neovascular AMD, DME, and RVO Associated Macular Edema

Machine learning artificial intelligence models have the potential to improve our ability to predict a patient's response to various treatments, and can play an important role in our move towards more individualized patient care.  This study endeavored to predict the anti-VEGF injection treatment demand in patients with neovascular age-related macular degeneration, retinal vein occlusion, and diabetic macular edema.

This was a retrospective study involving 710 eyes of 625 patients (377 eyes with AMD, 155 eyes with RVO, and 178 eyes with DME) who were treated with anti-VEGF injections (ranibizumab or aflibercept) for at least 1 year according to a predefined treat and extend regimen. The RVO and DME eyes were then lumped together into the “retinal vascular diseases” group.

The treatment burden was defined as being low demand, moderate demand, or high demand. Low treatment demand was defined as having an average treatment interval of 10 weeks or more, and high treatment demand was defined as having a treatment interval of 5 weeks or less, and moderate treatment demand applied to the remaining eyes.

The machine learning model used random forest plots to predict the likelihood of requiring low, moderate, or high treatment demand.  The machine learning model employed demographic information (age and gender) and morphological features from OCT volume scans at the baseline visit and the next 2 visits.  Some of the biomarkers studied included subretinal fluid, intraretinal fluid, hyperreflective foci, drusen, reticular pseudodrusen, epiretinal membrane, geographic atrophy, outer retinal atrophy, and fibrovascular pigment epithelial detachment. Additionally, segmentation data of the retinal and choroidal layers, intraretinal fluid, subretinal fluid, and pigment epithelial detachment were also used.

Ultimately, the machine learning technique utilized a very large number of features (153 features for the baseline visit, 607 features for the baseline and subsequent visit, and 1061 features for the baseline and 2 subsequent visits).  A tenfold cross validation was performed with 90% of the data used for training and 10% of the data used for validation.  No patient in the training set was used in the validation set.  The accuracy of the predictive model was evaluated using the area under the curve (AUC).  Additionally, the importance of each feature in contributing to the performance of the predictive model was computed.

In the AMD group, 127 eyes, 42 eyes, and 208 eyes were identified as having low, high, and moderate treatment demand, respectively.  In the retinal vascular diseases group, 61 eyes, 50 eyes, and 222 eyes were identified as having low, high, and moderate treatment demand, respectively.

The machine learning model was able to predict the 1-year treatment with reasonable accuracy (AUC of 0.76-0.79) for patients with age-related macular degeneration and retinal vascular diseases (RVO and DME).  Prediction for low treatment demand was possible with data from just the baseline visit prior to receiving an intravitreal injection, whereas prediction for a high treatment demand required data from the baseline and subsequent two visits.

The most informative features for patients with AMD were subretinal fluid, intraretinal fluid, and total retinal thickness.  The most informative features for retinal vascular diseases were intraretinal fluid and total retinal thickness.

This study has several limitations.  First, the cohort size precluded separate assessment of DME and RVO, which instead had to be considered as a single entity (retinal vascular disease). Second, the image quality of some of the B-scan images included was poor.  Third, there was some variability between the prescribed treatment intervals and the true treatment intervals in this real-world study.  Fourth, BCVA, additional demographic data (beyond age and gender), and other imaging modalities (such as fluorescein angiography) were not utilized.  Fifth, the age of patients in this study was higher than in some other studies which may limit the generalizability to younger patients.  Sixth, both eyes of patients were allowed to be included in the study (and these eyes may not be truly independent).

In conclusion, this study identified that, using a machine learning model, a low treatment demand can be predicted with reasonable accuracy using just data from a baseline visit (prior to receiving any intravitreal injections) for patients with AMD and retinal vascular diseases (RVO and DME), and high treatment demand can be predicted using data from the baseline and 2 subsequent visits.  In addition, it identified several important predictive features as OCT biomarkers for training machine learning models.  This work serves as an important stepping stone to further artificial intelligence analyses that may enhance our individualized care of patients with retinal diseases.