A Hybrid Deep Learning-Based Approach for Visual Field Test Forecasting

Ashkan Abbasi; Sowjanya Gowrisankaran; Wei-Chun Lin; Xubo Song; Bhavna Josephine Antony; Gadi Wollstein; Joel S Schuman; Hiroshi Ishikawa

doi:10.1016/j.xops.2025.100803

A Hybrid Deep Learning-Based Approach for Visual Field Test Forecasting

Ophthalmol Sci. 2025 Apr 24;5(5):100803. doi: 10.1016/j.xops.2025.100803. eCollection 2025 Sep-Oct.

Authors

Ashkan Abbasi¹, Sowjanya Gowrisankaran¹, Wei-Chun Lin¹, Xubo Song², Bhavna Josephine Antony³, Gadi Wollstein⁴, Joel S Schuman^{4

5

6}, Hiroshi Ishikawa^{1

2}

Affiliations

¹ Department of Ophthalmology, Casey Eye Institute, Oregon Health & Science University, Portland, Oregon.
² Department of Medical Informatics and Clinical Epidemiology, Oregon Health & Science University, Portland, Oregon.
³ Information Technology, Institute of Innovation, Science & Sustainability, Federation University, Ballarat, Australia.
⁴ Glaucoma Service and Vickie and Jack Farber Vision Research Center, Wills Eye Hospital, Philadelphia, Pennsylvania.
⁵ Department of Ophthalmology, Sidney Kimmel Medical College at Thomas Jefferson University, Philadelphia, Pennsylvania.
⁶ Department of Biomedical Engineering, Drexel School of Biomedical Engineering, Science and Health Systems, Philadelphia, Pennsylvania.

Abstract

Objective: Longitudinal assessment of visual field (VF) testing is essential in glaucoma management. Conventional VF forecasting methods require numerous prior tests, while deep learning techniques have shown promising results with fewer tests. This study introduces a hybrid deep learning framework to enhance flexibility and accuracy in VF test forecasting.

Design: A retrospective longitudinal study using deep learning-based VF forecasting models.

Subjects and controls: A total of 1750 subjects (healthy and glaucoma patients) with 19 437 Humphrey VF (24-2 Swedish Interactive Threshold Algorithm) tests collected from longitudinal glaucoma cohorts at the University of Pittsburgh and New York University.

Methods: Three deep learning models were trained for pointwise forecasting of VF test data: (1) a recurrent neural network (RNN), (2) CascadeNet-5, a convolutional neural network (CNN), and (3) Hybrid-VF-Net, our proposed method that combines an RNN with a CNN equipped with depthwise transformers for both spatial and temporal modeling. The results were analyzed from multiple perspectives, including the impact of varying the amount of prior input data and how data reliability and disease severity influence VF forecasting performance.

Main outcome measures: Mean absolute error between predicted and actual VF test results was evaluated using five-fold cross-validation.

Results: We found that specific VF locations benefited more from either local or temporal modeling, and our proposed methods outperformed the compared approaches using a hybrid strategy. Hybrid-VF-Net exhibited greater resilience to data reliability issues, particularly in managing high false-negative rates often seen in moderate-to-severe glaucoma cases due to increased test-retest variability. Additionally, it demonstrated improved performance with fewer prior VF tests, thus reducing the waiting time needed for progression analysis.

Conclusions: The proposed Hybrid-VF-Net method outperformed the existing deep learning VF methods in terms of performance and robustness. Our findings highlight the influence of disease severity, data quality, and time displacement on forecasting performance, with certain VF locations benefiting more from either local or temporal modeling. Low reliability in data from moderate to advanced glaucoma cases continues to pose a challenge. Therefore, future research could refine temporal modeling and leverage larger datasets to further enhance predictive performance.

Financial disclosures: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

Keywords: Deep learning; Glaucoma progression prediction; Hybrid architecture; Pointwise visual field forecasting; Spatial and temporal modeling.