Machine learning predicts short-term visual acuity of patients with AMD

In this paper, researchers used machine learning to predict the visual acuity of patients with AMD at 3 and 12 months after receiving anti-VEGF injections.

Study design

A data warehouse was used to identify patients with AMD who had received at least 3 injections of an anti-VEGF medication (i.e., aflibercept, bevacizumab or ranibizumab). Before the first injection, baseline visual acuity was 0.56 logMAR. At 3 months, 653 patients (738 eyes) were included in the visual acuity forecast while only 456 of these patients (508 eyes) had sufficient follow-up data to be included in the 12-month predictions.

Researchers used 5 different machine-learning algorithms and input clinical features from electronic medical records and OCT measurement data.

Outcomes

Machine learning algorithms were better for short-term predictions. At 3 months, the difference between predictions and measurements was 0.11 to 0.18 logMAR mean absolute error. For the 12-month forecast, the difference was 0.16 to 0.22 logMAR mean absolute error. The Lasso protocol outperformed the other algorithms.

Limitations

National data registries are still lacking; however, increased availability of data should improve prediction quality. Importantly, the OCT assessment here is a measurement of data (e.g. central retinal thickness), not the OCT images themselves. Developing deep learning of the OCT images remains a work in progress.

Clinical significance

This may be the first step towards designing clinical decision support software that could potentially be tailored to the patient. Maintaining quality electronic medical records is important as it provides reliability at the level of raw data.