An Effective Fusion Feature Extraction Method and Random Forest for Predicting Anti-Vegf Treatment Demand for RVO Patients

Retinal Vein Occlusion (RVO) is the most common retinal vascular occlusive disorder and is usually associated with visual loss to variable degrees. The most common cause of vision loss in the eyes related to RVO is macular oedema (MO). In many cases, MO can be successfully treated or managed with intravitreal injections of anti-vascular endothelial growth factor (anti-VEGF) agents. VEGF agents represent one of the most substantial advances in contemporary medicine. It can be considered as the current treatment for many diseases, including RVO disease. However, while many patients with RVO show an excellent response to treatment, others show no or only a partial response. At present, ophthalmologists have no way of predicting who will improve and who will not. This has profound implications, as treatment is costly, and an injection into the eyeball risks blinding. Knowing at the start of the treatment whether or not a patient will respond well, partially respond, or not respond would help to plan, set patient expectations, and improve healthcare efficiency. In this work, Random Forest (RF) was trained to predict high and low anti-VEGF demand at the early stage of treatment, using a Histogram of oriented gradients (HOG) and Gray-level co-occurrence matrix (GLCM) feature extraction methods for the analysis of Optical coherence tomography (OCT) images and fusing these features with electronic medical records (EMRs) to improve the system's performance. In addition, regression of the number of anti-VEGF injections given after 12 months was predicted using the Random Forest regressor. Using the proposed method, the area under the curve (AUC) suggested equivocal performance with a trend towards model superiority.