Abstract
Purpose :
Retinal vein occlusion (RVO) is often complicated by sight limiting macular oedema (MO). Intravitreal injections (IVIs) of anti-vascular endothelial growth factor (anti-VEGF) carry some clinical risk to patients but improve MO and vision for many. A significant minority show partial or no response despite many IVIs. If treatment outcomes could be forecast at RVO presentation, such patients may decline anti-VEGF treatment. This prognostic accuracy study aimed to test the feasibility of such forecasts using Random Forest technique (RF).
Methods :
A retrospective dataset of RVO treated with anti-VEGF IVIs for MO was assembled from the electronic medical record (EMR) of a large provincial UK ophthalmology centre. 412 eligible eyes were identified (212 left eye, 194 CRVO, 205 male, mean age 72.6, mean delay between diagnosis and treatment 110.9 days) with a mean visual acuity (VA) at treatment initiation of 50.3 early treatment diabetic retinopathy study (ETDRS) letters and a mean VA of 58.8 letters following 1 year of treatment. Fovea centred optical coherence tomography slices, taken at treatment initiation, were included. The dataset was divided 80:20 for training and testing. Features of the images were extracted using Histogram Oriented Gradient feature descriptor. To evaluate the performance of the RF model, 11 different ophthalmology doctors provided similar VA forecasts for two subsets (n=41) of the RF test set (n=82).
Results :
The mean absolute error (MAE) of the RF model is 14.3 (SD=16.2), which was not significantly different to the pooled MAE of 16.7 (SD=22.4) from all clinicians’ forecasts (p=0.42). The root mean squared errors were 19.55 and 23.53 for the RF model and pooled clinician forecasts respectively. The best performing clinician achieved a MAE of 14.8, which was not significantly different to the RF model’s MAE of 12.9 on the same subset (p=0.35).
Conclusions :
Such a RF model can match ophthalmologists in their ability to forecast VA outcomes for RVO complicated by MO undergoing 1 year of anti-VEGF treatment. The error reported here is too great to be of clinical use, but further work is required to establish if larger datasets and more powerful techniques could help to support patients’ treatment decisions.
This is a 2021 ARVO Annual Meeting abstract.