How AI is transforming glaucoma care: from research to real-world application

Introduction

Artificial intelligence is not the future of ophthalmology, it’s already reshaping it. Glaucoma, the leading cause of irreversible blindness, affects more than 80 million people worldwide. Among them, primary open-angle glaucoma (POAG) alone impacts 57.5 million individuals, with prevalence strongly correlated with age. 

In Europe, over 7.8 million people were affected in 2020, and in Switzerland, it's estimated that up to 100,000 people live with glaucoma, often undiagnosed. Left untreated, the disease progressively damages vision, reducing quality of life and increasing the risk of social isolation and depression.

At PeriVision, we are building the next generation of vision testing systems powered by AI, virtual reality, and cloud computing to fundamentally change how we test and understand visual function.

Why AI matters in ophthalmology

Eye diseases like glaucoma and age-related macular degeneration (AMD) are complex, progressive, and difficult to monitor. Traditional tools often fail to detect early changes, leading to delays in treatment and irreversible vision loss. AI brings a new level of precision and personalization to this space by learning patterns from vast datasets and making accurate predictions about disease progression.

At PeriVision, we apply AI across these core areas:

• Eye test optimization: AI adapts visual field tests to reduce duration, improve accuracy and automate assistance.

• Diagnostics & prognostics: Algorithms analyze results and flag early signs of deterioration, as well as predict future progression.

In this blog post, we will mainly focus on the latter, i.e. diagnostics and prognostics. 

Predicting glaucoma progression with machine learning

One of the most promising applications of AI in ophthalmology is predicting progression more accurately in individual patients. Glaucoma is a leading cause of irreversible blindness, and early detection of disease advancement is critical to prevent vision loss. However, traditional linear models often fall short in identifying fast-progressing cases and fail to capture the complex patterns inherent in visual field degradation.

That’s why PeriVision’s research and work together with the University of Bern and Inselspital is so important. Their research, centered on Advancing Glaucoma Progression Prediction with Machine Learning Models, investigates how AI-based approaches can outperform conventional methods in predicting glaucoma progression using visual field test data.

Their research covers three core predictive tasks:

1. Binary progression classification (yes/no)

2. Glaucoma stage classification

3. Visual field forecasting

   
   

Importantly, the team’s work on glaucoma stage classification “Visual field-based machine learning model for predicting disease stage in glaucoma“ (Colmenar Herrera et al, 2025) has been selected for publication in the 2025 edition of Current Directions in Biomedical Engineering, the joint journal of the German, Swiss and Austrian Societies for Biomedical Engineering.

Meet the researcher: Marta Colmenar Herrera

Marta is a PhD candidate at the ARTORG Center for Biomedical Engineering Research and a data scientist at PeriVision. Her background in biomedical engineering and AI in medicine spans multiple countries and fields, including neurophysiology and bioinformatics.

“One of the most valuable lessons I’ve learned from working with ophthalmologists is the importance of clinical context and usability in AI development,” she explains. “Even the best-performing model must be easy to use and trusted by clinicians.”

Study design

The models developed for the three predictive tasks were trained and validated using visual field data from three independent clinical datasets from Bern (Inselspital Eye Clinic, Switzerland), Rotterdam (public), and Washington (public).

The study employs visual field triplets (VF₁ , VF₂ → predict VF₃) to capture temporal disease progression. 

1. For binary progression classification, the team developed a neural network model to address the fundamental clinical task of determining whether a patient is progressing or not progressing, enabling early identification of at-risk individuals.

   
   

2. For future glaucoma stage classification, a non-linear ensemble learning model, such as Adaboost, is benchmarked against traditional logistic regression.

   
   

3. For visual field forecasting, the team developed an unsupervised neural network (NN) model capable of reconstructing future visual fields from prior examinations—independent of the number of available previous tests or the prediction time horizon.

Pictures: Architecture of the three proposed models.

Key results

The ML models demonstrated consistent and substantial improvement over traditional methods:

• Stage classification (Cohen’s Kappa):

  • AdaBoost: up to 0.86 across datasets

  • Logistic Regression: typically lower, with values as low as 0.66
    

• Binary progression classification (ROC AUC):

  • All datasets showed AUC values above 0.80, confirming strong discriminative power using a deep learning approach
    

• Visual field forecasting (Mean Absolute Error - MAE):

  • Neural Network: MAE as low as 2.41 ± 2.87

  • Outperformed linear regression baselines, with gains especially pronounced in advanced glaucoma
    

• Visual field forecasting  (Cohen’s Kappa):

  • Stage classification using NN-generated visual fields Kappa values reached up to 0.95, indicating high fidelity of the model’s predictions in capturing disease stage

Picture: Example from the Bern Dataset showing VFT forecasting on the test dataset, with real and predicted VF images alongside patient age, Mean Defect (MD), and Loss Variance square root (sLV).

These results highlight the superior performance and robustness of non-linear ML models in capturing the intricacies of glaucoma progression, especially when compared to conventional statistical methods.

Scientific and clinical impact

The study's findings shows the potential of AI in glaucoma care by enabling:

• Earlier detection of fast-progressing patients

• More personalized treatment plans, including optimized timing for therapeutic interventions (e.g. medications, surgery)

• Reduced clinical burden via more targeted follow-ups for stable patients

Next step: building the AI assistant for ophthalmologists

To truly unlock AI’s potential in clinical settings, we must also address key challenges. That’s why we are building an AI portfolio together with OphthaPro to tackle core obstacles in today’s glaucoma care workflows:

a. Isolated data silosExisting perimetry devices and glaucoma software often operate independently, making it difficult to organize patient data and extract longitudinal insights.

b. Crude statistical approachesMost current tools rely on static, linear models and sometimes fail to account for important contextual factors such as treatment changes or coexisting ocular conditions.

c. Time-consuming analysis and specialist dependencyReviewing and interpreting visual field test (VFT) data is labor-intensive and often requires highly specialized expertise to avoid misinterpretation or missed trends.

By addressing these issues, we aim to streamline glaucoma management from both a technological and clinical standpoint bringing AI from research to routine use.

At PeriVision, we are integrating these insights into our VisionOne™ platform. By combining AI-powered test results with real-world patient data, we aim to:

• Identify fast-progressing patients earlier

• Support timely referrals for surgical intervention or adjustments of medications

• Reduce unnecessary follow-ups for stable patients
  

Ultimately, this enables more proactive, personalized glaucoma care and better long-term outcomes.

Our vision: AI as a true clinical co-pilot

Looking ahead, we envision AI not just as a diagnostic tool, but as a true clinical co-pilot. The upcoming OphthaPro AI suite will integrate data from visual field tests (VFT), intraocular pressure (IOP), imaging or treatment history to provide a holistic, longitudinal view of each patient. 

By combining a rich dataset with advanced ML models, we aim to deliver both diagnostic and prognostic insights helping doctors detect deterioration earlier and plan treatments more precisely. Ultimately, our goal is to support ophthalmologists in making faster, smarter decisions without increasing their workload.

AI in action: VisionOne by PeriVision

VisionOne is PeriVision’s flagship product: a portable VR platform for visual field testing designed for flexibility, patient comfort, and integration into modern clinical workflows. While our research is actively advancing AI-powered diagnostics and prognostics, VisionOne today focuses on delivering reliable, high-quality tests with minimal setup.

Features that make the difference

• Autonomous visual field testing guided by a multilingual virtual assistant

• Industry-leading test-retest variability (1.30 dBs)

• Comfortable, ergonomic design optimized for patient experience

• CE, FDA, and MHRA certified for international clinical use

Conclusion

From groundbreaking academic research to real-world diagnostic tools, AI is rapidly changing the landscape of ophthalmology. At PeriVision, we are proud to be at the forefront of this transformation, combining Swiss engineering, international clinical research, and next-generation software to make eye care more intelligent, accessible, and patient-centered.

Want to learn more about how AI is shaping the future of vision care? 

Book a demo