Predicting Visual Acuity in Patients Treated for AMD

Abstract

The leading diagnostic tool in modern ophthalmology, Optical Coherence Tomography (OCT), is not yet able to establish the evolution of retinal diseases. Our task is to forecast the progression of retinal diseases by means of machine learning technologies. The aim is to help the ophthalmologist to determine when early treatment is needed in order to prevent severe vision impairment or even blindness. The acquired data are made up of sequences of visits from multiple patients with age-related macular degeneration (AMD), which, if not treated at the appropriate time, may result in irreversible blindness. The dataset contains 94 patients with AMD and there are 161 eyes included with more than one medical examination. We used various techniques from machine learning (linear regression, gradient boosting, random forest and extremely randomised trees, bidirectional recurrent neural network, LSTM network, GRU network) to handle technical challenges such as how to learn from small-sized time series, how to handle different time intervals between visits, and how to learn from different numbers of visits for each patient (1-5 visits). For predicting the visual acuity, we performed several experiments with different features. First, by considering only previous measured visual acuity, the best accuracy of 0.96 was obtained based on a linear regression. Second, by considering numerical OCT features such as previous thickness and volume values in all retinal zones, the LSTM network reached the highest score (R2=0.99). Third, by considering the fundus scan images represented as embeddings obtained from the convolutional autoencoder, the accuracy was increased for all algorithms. The best forecasting results for visual acuity depend on the number of visits and features used for predictions, i.e., 0.99 for LSTM based on three visits (monthly resampled series) based on numerical OCT values, fundus images, and previous visual acuities.

Keywords: OCT; diagnosis of retinal conditions; machine learning; predicting visual acuity.

Figure 1

The third visit, i.e., follow-up (#2), of a patient, right eye only. The right column shows that the values are computed on all of the layers of the retina: between the top layer (i.e., ILM) and the bottom layer (i.e., Bruch’s membrane—BM). The middle column shows nine values for the retinal thickness (black color), and nine values for the retinal volume (red color).

Figure 2

The 9 zones analysed in the retina ($C_0$ is the fovea).

Figure 3

Cumulative explained variance: 99% of the data are represented with 12,000 components.

Figure 4

Cumulative explained variance: 80% of the data are represented with 250 components.

Figure 5

Gradient boosting machine showing how much the numerical OCT features influence the prediction of visual acuity. The case with prediction timestep as a features appears on the left. The case with 1-month resampling appears on the right.

Figure 6

The proposed LSTM architecture.

Figure 7

Autoencoder training and validation plot.

Figure 8

Examples of OCT scan reconstruction and noise filtering using the convolutional autoencoder.

Figure 9

Confusion matrix results for the disease evolution classification using gradient boosting classifier. Label 0 represents a good disease evolution, while 1 stands for a poor evolution. The matrix column represents the actual label.

Figure 10

Average cross-validated $R^2$ scores when predicting future visual acuity only from previous visual acuity data.

Figure 11

Average $R^2$ scores when predicting future visual acuity from all numerical OCT data.

Figure 12

Comparison of feature selection methods’ performance when predicting future visual acuity from all OCT data. Average cross-validated R${}^2$ scores for all neural networks were computed.

Figure 13

Actual vs. predicted VA values for the best model (LSTM with $R^2=0.98$, for three previous visits). The second figure shows that the actual and predicted visual acuities are highly correlated ($r=0.993$).

Figure 14

Actual vs. predicted VA values for the best model using all data (LSTM with $R^2=0.99$, for three previous visits). The second figure shows that the actual and predicted visual acuities are highly correlated ($r=0.994$).

Figure 15

Comparison of feature selection methods’ performance when predicting future visual acuities from all OCT data. Average cross-validated $R^2$ scores for all neural networks were computed.