The research team used 301 eye-visit records from 176 patients treated in the ophthalmology department of the Second Affiliated Hospital of Harbin Medical University in Harbin, Heilongjiang. The dataset included 99 records from eyes with early primary open-angle glaucoma, 85 from eyes classified as suspected primary open-angle glaucoma and 117 from healthy eyes. These records were divided among groups of 45 patients with early glaucoma, 69 with suspected glaucoma and 62 healthy participants. The retrospective design drew on examinations that clinicians routinely use in glaucoma assessment: nonmydriatic fundus photography, circumpapillary retinal nerve fiber layer thickness measured by optical coherence tomography, intraocular pressure, standard automated perimetry and age. All examinations for an individual patient were conducted within one week.

The study’s central technical step was a deep-learning system that segmented the optic nerve head in each fundus photograph. Three glaucoma specialists manually marked the optic disc and cup in the images, and a senior glaucoma specialist with more than 10 years of experience reviewed the annotations. The segmentation architecture combined a DeepLabV3+ network with a MobileNetV2 backbone, processing cropped optic-disc regions at a resolution of 512 by 512 pixels. From the resulting masks, the researchers calculated vertical and horizontal cup-to-disc ratios and measured the widths of the superior, inferior, temporal and nasal neuroretinal rims. The fundus photographs served as the input to this segmentation process; high-dimensional image features learned directly from the photographs were not passed to the final classifier.

The machine-learning stage used 13 clinical features as its input: age, mean intraocular pressure, six sector-specific retinal nerve fiber layer thickness measurements, the visual field index, mean deviation, pattern standard deviation, and the horizontal and vertical cup-to-disc ratios derived from fundus-image segmentation. The six nerve fiber layer measurements covered the temporal, nasal, superior temporal, superior nasal, inferior temporal and inferior nasal sectors around the optic disc. The researchers tested support vector machines, random forests, logistic regression and an artificial neural network with one hidden layer of 512 units. Each classifier underwent five-fold cross-validation, with four folds used for training and the remaining fold used for validation in each iteration. Reported accuracy was the average of the five validation accuracies, while area under the curve was based on averaged cross-validation performance.

Support vector machines produced the strongest overall results in the study. The support vector machine’s accuracy was 0.9478, with a 95% confidence interval of 0.9114 to 0.9644; sensitivity was 0.9577, with a 95% confidence interval of 0.9236 to 0.9771, and specificity was 0.9710, with a 95% confidence interval of 0.9584 to 0.9844. Its F1 score was 0.9513 and its Matthews correlation coefficient was 0.9255. The area under the curve was 0.9858, with a 95% confidence interval of 0.9724 to 0.9961. Sensitivity was 0.9714 at a specificity of 80% and 0.9570 at a specificity of 90%. Random forest achieved an accuracy of 0.9217 and an area under the curve of 0.9851. Logistic regression and the artificial neural network each achieved an accuracy of 0.9130.

Feature analysis indicated that structural measurements carried particular weight in the random-forest model. Vertical cup-to-disc ratio had the highest importance score at 0.15. Inferotemporal and superior temporal retinal nerve fiber layer thickness followed, each with a score of 0.12, followed by horizontal cup-to-disc ratio at 0.10 and pattern standard deviation at 0.09. Models using only optical coherence tomography data reached a mean area under the curve of 0.9756. Models based on cup-to-disc ratio measurements reached 0.8569, those using visual-field parameters reached 0.7761, and those combining age with mean intraocular pressure reached 0.7817. Removing optical coherence tomography parameters from the full support vector machine reduced the area under the curve from 0.9858 to 0.9510. These comparisons show the contribution of the selected feature groups within this dataset; they do not isolate their effects on glaucoma progression or clinical outcomes.

For the clinician comparison, three glaucoma specialists and three general ophthalmologists, all with at least five years of clinical experience, independently assessed the records. Diagnostic labels were hidden from the clinicians. They reviewed optical coherence tomography reports, visual-field reports, fundus photographs and 11 features consisting of age, intraocular pressure, six-sector retinal nerve fiber layer thickness, visual field index, mean deviation and pattern standard deviation. Each clinician assigned cases to one of three categories: healthy eye, suspected glaucoma or early primary open-angle glaucoma. The glaucoma specialists had a mean accuracy of 0.9493, sensitivity of 0.9693 and specificity of 0.9475. The general ophthalmologists had a mean accuracy of 0.8725, sensitivity of 0.9194 and specificity of 0.8632. Interobserver agreement was higher among the glaucoma specialists, with Fleiss’ kappa of 0.7708 and a 95% confidence interval of 0.7220 to 0.8169, than among the general ophthalmologists, whose kappa was 0.5633 with a 95% confidence interval of 0.5054 to 0.6186.

The nearly identical overall accuracy values of the support vector machine and glaucoma specialist group do not demonstrate that the two approaches were statistically equivalent. The algorithm and clinicians also did not work under identical conditions: the model incorporated image-derived cup-to-disc measurements, while the clinicians received clinical reports and interpreted the same images and measurements. The study did not randomize cases between decision makers, measure diagnostic benefit in routine care, or assess whether use of the tool improved testing, referrals or treatment decisions. Its comparison therefore concerns classification performance on this selected retrospective dataset rather than superiority or interchangeability in clinical practice.

Several features of the dataset also require careful interpretation. Participants had to have best-corrected visual acuity of at least 0.5 logMAR, refractive error between +3 and −6 diopters, open angles on gonioscopy and reliable visual-field examinations with false-positive errors, false-negative errors and fixation losses below 15%. Patients with other retinal diseases, secondary glaucoma, other causes of optic neuropathy, previous intraocular surgery, diabetes or hypertension were excluded. Mean age was 40.16 years in the early-glaucoma group, 39.19 years in the suspected-glaucoma group and 39.39 years in the healthy group. Apart from age, all measured features differed significantly among the three groups at P < 0.05 by Welch analysis of variance. Visual field index, mean deviation and pattern standard deviation did not differ significantly between the suspected-glaucoma and healthy groups.

The principal limitation was the study’s size and setting: it included Chinese patients from one center, and the authors said the training sample was relatively small. Cross-validation was conducted at the eye-record level without enforced patient-level separation, so records from the same patient could appear in both training and evaluation folds. This correlated-data design may have produced optimistic performance estimates. The research also faced incorporation bias because retinal nerve fiber layer thickness, visual-field measurements, intraocular pressure and cup-to-disc ratios helped define the diagnostic groups while also serving as model predictors. Cross-validation cannot eliminate that problem. In addition, the segmentation system sometimes performed poorly on small optic cups and disc notches, although the researchers considered most resulting cup-to-disc measurements credible.

Taken together, the results show that a hybrid deep-learning and machine-learning pipeline can combine automated fundus-image measurements with conventional structural and functional examinations to distinguish three clinically relevant categories in a controlled study cohort. The most useful next step is not a claim of population-wide screening performance but independent, multicenter external validation in previously unseen patients, using independently adjudicated diagnoses and longitudinal clinical outcomes. Such studies would need to determine whether performance persists across different populations, imaging equipment and clinical workflows, and whether recalibration is needed before use. Until then, the evidence supports the system as a potential decision-support and triage tool that has not yet been validated for independent diagnosis or demonstrated to improve real-world glaucoma care.

Subject of Research: Evaluation of a multimodal deep-learning and machine-learning system for early primary open-angle glaucoma diagnosis.

Article Title: Early diagnosis program for multimodal primary open-angle glaucoma based on clinical process

Article References: Ma, J., Lv, B., Li, Y., Fan, P., Zhao, X., Zhang, Y., & Yuan, H. (2026). Early diagnosis program for multimodal primary open-angle glaucoma based on clinical process. Frontiers in Ophthalmology, 6, 1975341. https://doi.org/10.3389/fopht.2026.1975341

Image Credits: AI Generated

DOI: 10.3389/fopht.2026.1975341

Keywords: primary open-angle glaucoma, early glaucoma diagnosis, multimodal artificial intelligence, deep learning, machine learning, fundus photography, optical coherence tomography, visual-field testing, clinical decision support, retrospective diagnostic study, Early, diagnosis