AI models using OCT data accurately predicted glaucoma progression
Key takeaways:
- Researchers used machine learning to predict glaucoma progression based on dozens of biomarkers, including those from OCT angiography.
- The approach could enable more personalized, preventive care.
The progression of glaucoma was accurately predicted by machine learning models based on structural, functional and vascular biomarkers, including those from OCT angiography, according to data published in EPMA Journal.
“Glaucoma’s growing prevalence and persistent underdiagnosis produce a profound and multifaceted burden for patients, families, health care systems and societies worldwide, underscoring the critical importance of early detection, individualized monitoring of disease progression and effective interventions,” Natalia I. Kurysheva, of the Ophthalmological Center of the Federal and Medical Biological Agency of the Russian Federation, and colleagues wrote.
To advance that objective, Kurysheva and colleagues tried to develop AI-based glaucoma progression prediction models based on a variety of biomarkers, including those from OCT, OCT angiography, automated perimetry and biomechanical assessments.
“The proposed inclusion of OCT-A was expected to provide unique insights into microvascular alterations that precede and predict structural and functional decline,” the researchers wrote.
Kurysheva and colleagues monitored 114 eyes with different stages of primary open-angle glaucoma for at least 36 months, then used the data to build predictive machine learning models using ranked partial least squares discriminant analysis.
Two models were trained: one for early stage glaucoma trained on 27 variables and one for moderate/advanced stage glaucoma trained on 20 variables. Both showed “high prognostic accuracy” for identifying slow, moderate and rapid glaucoma progression, the researchers wrote.
For early stage primary open-angle glaucoma, the most important variable in the prediction model was retinal nerve fiber layer thickness in the inferotemporal sector.
“This finding emphasizes the vulnerability of the inferotemporal [retinal nerve fiber layer] region in early glaucomatous damage, likely reflecting the anatomical susceptibility of nerve fibers entering the optic disc from this sector,” the researchers wrote.
And for more advanced disease, the most important predictor was ganglion cell complex thickness.
“This reflects the critical importance of macular ganglion cell integrity assessment in advanced disease, where residual functional capacity directly correlates with remaining retinal ganglion cells,” the researchers wrote.
Overall, the findings “emphasize that vascular factors remain significant predictors of glaucoma progression at both early and advanced disease stages,” Kurysheva and colleagues wrote.
The researchers wrote that their machine learning approach could benefit clinical practices by allowing them to fine-tune follow-up intervals and selectively target preventive measures.
“Rather than relying on a few key metrics, our optimized models for early and advanced cohorts incorporate all measured predictors ... to capture the full spectrum of disease heterogeneity,” they wrote. “This comprehensive approach achieves high prognostic accuracy (AUC 0.90) and accommodates the continuous and multifaceted nature of glaucomatous damage.”
Perspective
Back to Top
With the prevalence of glaucoma rising, multimodal AI-based models to personalize glaucoma risk should appeal to clinicians looking to the future of medical optometry.
In this 3-year study utilizing 27 parameters in early primary open angle glaucoma (POAG) and 20 in advanced POAG, researchers were able to predict rates of glaucoma progression translating to around 90% accuracy and establish the most predictive measures of glaucomatous progression for both early POAG and advanced POAG.
This study was limited in its analysis by a small sample size of 114 eyes due to strict exclusion criteria. Despite the strong predictive power of the model, acquiring 20 to 27 variables of data, such as OCT, retinal vasculature, visual fields and cornea hysteresis, may not be feasible. Both the small sample size and large number of variables may impact how this study translates into real-world clinical applications.
Until models are widely available and able to integrate with electronic health records, for now clinicians will have to utilize the piecemeal results of this study — namely, in early POAG, inferotemporal retinal nerve fiber layer thickness is the highest predictor of rate of progression, while it is ganglion cell complex in advanced stages.
A future where doctors work synergistically with AI machine learning to provide more personalized assessments and treatments in glaucoma can only be beneficial to our patients.
Sources/Disclosures
Source:
Kurysheva NI, et al. EPMA. 2025;doi:10.1007/s13167-025-00429-0.
Disclosures: The authors report no relevant financial disclosures. This study was funded by the European Association for Predictive, Preventive and Personalized Medicine.
Ask a clinical question and tap into Healio AI's knowledge base.
- PubMed, enrolling/recruiting trials, guidelines
- Clinical Guidance, Healio CME, FDA news
- Healio's exclusive daily news coverage of clinical data
Read more about
