Machine learning techniques increase the accuracy of visual outcome prediction in retinal vein occlusion complicated by macular oedema

Introduction: Retinal vein occlusion (RVO) is the second most common retinal vascular disease which can cause sight loss
through macular oedema. With anti-VEGF treatment many patients regain vision, but to what extent is often unclear at
presentation. We aim to increase the accuracy of prognostication to improve treatment discussions with patients.
Methods: The electronic medical record at the Royal Victoria Infirmary was interrogated for cases that received anti-VEGF
intravitreal injections, over 1 year or more, for macular oedema secondary to RVO. 428 eligible eyes were identified
(222 left eyes, 200 central RVO, 213 male, mean age 72.6 and mean interval between diagnosis and treatment 118 days).
Mean baseline visual acuity was 50.6 ETDRS letters improving to 59.0 letters following 1 year of anti-VEGF treatment.
Linear regression and random forest regression analyses were performed to predict 1-year visual acuity from baseline data.
Results: Linear regression produced a model accounting for 57% of the variability seen in 1 year visual acuity (R2=0.57).
Using the same data, random forest regression surpassed this with the model accounting for 62% of variability (R2=0.62).
Discussion: With basic demographic data machine learning techniques demonstrate a superior ability to predict visual
prognosis for patients with RVO. Unlike traditional statistical techniques this approach also has the potential to draw on retinal
imaging to further improve its performance. Such an enhanced tool could help clinicians personalise patient information and
commissioners to tailor cost-utility analyses to each case.