Predicting Visual Acuity in Patients Treated for AMD.

Area of Science:

• Ophthalmology
• Medical Imaging
• Machine Learning

Background:

• Optical Coherence Tomography (OCT) is a key diagnostic tool but cannot predict retinal disease progression.
• Age-related macular degeneration (AMD) can lead to irreversible blindness if not treated promptly.
• Accurate forecasting of disease progression is crucial for timely intervention.

Purpose of the Study:

• To develop machine learning models for forecasting retinal disease progression.
• To assist ophthalmologists in identifying optimal timing for early treatment.
• To prevent severe vision impairment and blindness caused by AMD.

Main Methods:

• Utilized machine learning techniques including linear regression, gradient boosting, random forest, extremely randomized trees, bidirectional recurrent neural networks, LSTM, and GRU networks.
• Addressed challenges of small time-series data, variable visit intervals, and varying numbers of patient visits (1-5).
• Experimented with predicting visual acuity using historical visual acuity, numerical OCT features (thickness, volume), and fundus image embeddings from a convolutional autoencoder.

Main Results:

• Linear regression achieved 0.96 accuracy using only previous visual acuity.
• LSTM networks reached an R2 score of 0.99 using numerical OCT features.
• Incorporating fundus image embeddings further improved accuracy across all algorithms.
• The best forecasting (0.99 accuracy) was achieved by an LSTM model using three monthly resampled visits, incorporating OCT values, fundus images, and prior visual acuity.

Conclusions:

• Machine learning models can effectively forecast retinal disease progression.
• Combining OCT data, fundus images, and visual acuity history enhances prediction accuracy.
• This approach supports early treatment decisions to prevent vision loss in AMD patients.