Comparing the Use of Measured and Smoothed Data in Forecasting Visual Field Tests Using Deep Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Visual field (VF) tests are crucial for diagnosing and monitoring glaucoma.
• Forecasting VF changes accurately is essential for timely intervention.
• Deep learning (DL) models show promise for VF forecasting but can be affected by test variability.

Purpose of the Study:

• To evaluate the impact of using smoothed visual field (VF) targets versus measured VF targets for training and testing DL models.
• To assess the forecasting accuracy of DL models under different training and testing configurations.

Main Methods:

• Retrospective analysis of 19,437 Humphrey VF tests from 1400 subjects (healthy and glaucoma patients).
• Three DL-based pointwise VF forecasting methods were trained and tested using measured VF targets and smoothed VF targets (constructed via linear regression).
• Models were evaluated using fivefold cross-validation and mean absolute error (MAE).

Main Results:

• Models trained and tested on smoothed VF targets consistently achieved lower mean absolute errors (MAEs) compared to those trained on measured VF targets.
• Performance improvements were most significant in the 0.5- to 1.5-year forecast range.
• Models trained with smoothed targets demonstrated comparable performance when evaluated against measured VF targets.

Conclusions:

• Using smoothed VF targets for training DL models enhances forecasting accuracy by enabling the models to learn long-term trends.
• This method mitigates the impact of noise and short-term variability inherent in VF test data.
• The approach of using smoothed targets is recommended for future DL-based VF modeling efforts to better reflect clinically meaningful functional changes.