Unsupervised Classification of Wide-Field Swept-Source Optical Coherence Tomography Reveals Patterns Associated With Glaucoma Severity and Progression

Abstract

Purpose

We investigated the structural patterns associated with visual field (VF) progression using a novel unsupervised hybrid clustering approach that combined archetypal analysis (AA) and fuzzy c-means (FCM).

Methods

AA was used to identify the representative structural archetypes (ATs) from 1803 wide-field swept-source optical coherence tomography (SS-OCT) scans obtained from 1803 eyes of 1167 patients with glaucoma or suspected glaucoma, resulting in 10 distinct ATs. Subsequently, FCM was used to calculate the decomposition coefficients of each eye. A subset of 372 eyes with at least five reliable 24-2 VF tests was analyzed to evaluate the association between baseline structural patterns and subsequent VF progression. Progression was defined using three criteria: mean deviation slope, Visual Field Index slope, and pointwise linear regression.

Results

We observed stage-dependent differences in archetypal expression. Greater baseline AT9 (bihemiretinal loss) was significantly associated with VF progression in glaucoma suspect eyes; greater baseline AT4 (early superior loss) and AT9 in early glaucoma; and lower baseline AT1 (normal) and greater baseline AT5 (superior hemiretinal loss) in moderate-to-advanced glaucoma. Only 4.3% of eyes met all three VF progression criteria, whereas 81.7% met none, highlighting substantial discordance among definitions.

Conclusions

These findings suggest that the structural patterns associated with VF progression differ according to disease stage and progression criteria. Our hybrid unsupervised method offers a clinically interpretable framework that may facilitate real-time structural assessment and personalized glaucoma monitoring.

Translational Relevance

Hybrid approach combining AA and FCM visualize wide-field SS-OCT images in characteristic patterns. It enables the identification of high-risk eyes in the early phase and allows for individualized patient monitoring.

Introduction

Early detection of structural changes in the optic disc and retinal nerve fiber layer (RNFL) is critical in patients with glaucoma, because structural damage is believed to precede functional visual field loss.^1,2 Optical coherence tomography (OCT) has recently emerged as an essential tool for the early detection of glaucoma and monitoring its progression because of its high reproducibility and objectivity.^3–5 Moreover, OCT is more sensitive than visual field (VF) testing in detecting progression during the early stages of glaucoma.⁶

Swept-source OCT (SS-OCT), which features a faster scanning rate, wider scanning range, and higher output power, has emerged with the advancements in OCT.⁷ The diagnostic performance of wide-field SS-OCT scan is comparable to that of separate scans of the optic disc and macula.^8,9 Furthermore, deep learning models utilizing wide-field SS-OCT have outperformed conventional OCT and VF in distinguishing healthy eyes from those with early glaucoma and have estimated the 24-2 VF more accurately than RNFL and ganglion cell–inner plexiform layer (GCL+) images obtained with spectral-domain OCT.^10,11 A combination of data from both the macular and circumpapillary regions in a single scan can identify glaucomatous patterns because damage to macular retinal ganglion cells (RGCs) and the circumpapillary RNFL (cpRNFL) may follow distinct yet interrelated trajectories.¹²

Conventionally, glaucoma diagnosis depends on the evaluation of morphological changes in the optic nerve head and thinning of the cpRNFL.^2,13 The ganglion cell complex (GCL++), integrating the RNFL and GCL+ is diagnostically comparable to or better than cpRNFL.^6,14–16 Zhang et al.⁶ reported GCL++ as a useful marker for detecting the progression from early to advanced stages of glaucoma, whereas RNFL performance declined as the disease progressed to advanced stage. Based on these findings, we hypothesized that wide-field GCL++ scans might better represent the characteristic patterns of glaucomatous damage across the entire spectrum of disease severity than independent RNFL and GCL+ scans.

Although previous studies have classified patterns in VF tests results using machine learning, no studies have applied machine learning to classify wide-field SS-OCT images.^17–19 Elze et al.¹⁷ conducted archetypal analysis (AA) to classify 24-2 VF test results into 17 distinct patterns, referred to as archetypes (ATs). AA identifies the representative features in the data by estimating the principal convex hull of a dataset.²⁰ However, expressing all data points as convex combinations in AA decomposition within a high-dimensional space can cause projection loss.²¹

We developed an unsupervised hybrid method that combines AA and fuzzy c-means (FCM) to classify both 24-2 and 10-2 VF patterns in glaucomatous eyes more accurately and to address these limitations.^22,23 FCM has been widely used in medical image segmentation and classification owing to its simplicity and accuracy.^24–27 It assigns a membership value to each data point, indicating the degree to which it belongs to each cluster. In addition, it quantifies the relevance of each data point to all the clusters to which it is associated.²⁸

Although the hybrid AA + FCM framework has been applied to VF data, this study is the first to apply it to high-dimensional wide-field SS-OCT pixel data. Unlike 24-2 VF and 10-2 VF data, which consists of only 52 or 68 points, our model processes a 60 × 45 grid (2700 segments) of GCL++ thickness values. This allows for the identification of fine-grained structural patterns that may precede functional loss. Unlike previous supervised learning approaches that depend on labeled OCT or VF datasets, we used an unsupervised hybrid clustering method to analyze wide-field GCL++ thickness maps and identify characteristic patterns across the full spectrum of glaucomatous damage.^29–31 To the best of our knowledge, this is the first study to classify the characteristic patterns of wide-field SS-OCT images using machine learning. We identified distinct wide-field SS-OCT patterns based on the GCL++ thickness value and decomposed individual patient scans using a hybrid unsupervised approach that combined FCM and AA. Next, we investigated the association between the derived patterns and the severity and progression of glaucoma.

Material and Methods

Ethics

This retrospective study was conducted in accordance with the principles of the Declaration of Helsinki. OCT and VF data were collected from patients at Pusan National University Hospital between August 21, 2018 and December 14, 2021. The study protocol was approved by the Institutional Review Board (IRBs) of Pusan National University Hospital (IRB no. 2407-004-140). The requirement for informed consent was waived by the IRB because of the retrospective nature of the study. All data were de-identified, and the authors had no access to information that could identify participants.

Study Design and Population

This study included patients clinically diagnosed with glaucoma and those with suspected glaucoma. All participants underwent comprehensive ophthalmological examinations, including best-corrected visual acuity, slit-lamp biomicroscopy, intraocular pressure (IOP) measurement using Goldmann applanation tonometry, gonioscopy, fundus examination, ocular biometry using the IOL Master (Carl Zeiss Meditec, Dublin, CA, USA), and keratometry using an auto-kerato-refractometer (ARK-510A; NIDEK, Hiroshi, Japan).

Eyes were classified as having suspected glaucoma if patients had a history of elevated IOP (> 21 mmHg) and/or a suspicious optic nerve appearance on stereoscopic optic disc photographs, but normal VF test results. Glaucomatous optic neuropathy was defined as the presence of cup-to-disc ratio asymmetry > 0.2 between the two eyes, neuroretinal rim thinning, notching, or characteristic RNFL defects consistent with glaucoma. The eyes were classified as glaucomatous if patients exhibited both glaucomatous optic discs and abnormal VF findings at the baseline visit. An abnormal VF was defined as a pattern standard deviation with P < 0.05 or a Glaucoma Hemifield Test result outside the normal limits. Eyes with established glaucoma were further categorized as early (mean deviation [MD] ≥ −6 dB) or moderate to advanced (MD < −6 dB).³²

Participants with retinal or neurological diseases or severe media opacities possibly affecting the OCT and VF measurements were excluded. The specific exclusion criteria were corneal opacity, cataract, non-glaucomatous optic neuropathy, epiretinal membrane, uveitis, diabetic retinopathy, retinal vein occlusion and age-related macular degeneration.

SS-OCT

Wide-angle scanning was performed using a deep-range imaging Triton SS-OCT device (Topcon, Tokyo, Japan) for all participants. A wide-field 12 × 9 mm² scan centered on the fovea, consisting of 256 B-scans, each comprising 512 A-scans, was acquired to generate 131072 axial scans per volume. Each scan covered the circumpapillary and macular regions and was completed within 1.3 s.^33–35 Images with an image quality score < 40, poorly focused or decentered during foveal scanning, or exhibiting segmentation failure or artifacts caused by eye movements or blinking were excluded. The 12 × 9 mm² area was divided into 200-µm segments, generating a 60 × 45 grid (2700 segments). The GCL++ thickness, including both the RNFL and the GCL+, was extracted for each segment.

24-2 VF Test

Standard automated perimetry was performed within 6 months of the OCT examination using the Humphrey Visual Field Analyzer 750i (Carl Zeiss Meditec) with the Swedish Interactive Thresholding Algorithm standard 24-2 program. Test reliability was considered acceptable if fixation losses were <33%, false-positive rates <20%, and false-negative rates <20%.^36,37

Machine Learning Model

We identified ATs representing the characteristic patterns of wide-field SS-OCT using AA based on 2700 GCL++ thickness value from 60 × 45 grids. For the pattern analysis, GCL++ maps were prioritized over individual RNFL or GCL+ maps. Preliminary comparisons revealed that while RNFL maps lacked central macular detail and GCL+ maps provided limited circumpapillary axonal context, GCL++ offered an intuitive, continuous visualization of the entire GCL++ across the 12 × 9 mm² wide-field area, effectively capturing the interrelated trajectories of macular and circumpapillary damage (Supplementary Fig. S1).

To determine the optimal number of ATs (k), we used a systematic approach balancing statistical rigor and clinical relevance. We did not use a separate test dataset or cross-validation framework for AT selection. We evaluated reconstruction errors across different values of k (1–20) and identified the elbow point in the scree plot (Supplementary Fig. S2). For each value of k, the AA was performed using multiple random initializations (nrep = 3) within the stepArchetypes algorithm, and the solution with the lowest reconstruction error was selected to reduce the likelihood of convergence to suboptimal local minima. To ensure the robustness of our selection, we further conducted a sensitivity analysis by visually inspecting and comparing the archetypal maps for k values ranging from 8 to 12 (Supplementary Fig. S3). This qualitative assessment ensured that the identified patterns were stable and clinically distinct. We ultimately selected k = 10 as the optimal number, as it represented the “elbow” in the scree plot (Supplementary Fig. S2) where the gain in reconstruction accuracy began to plateau while providing the most non-redundant and interpretable structural patterns for clinical use. We applied FCM to obtain the decomposition coefficients (Supplementary Fig. S4).^22,23

The rationale for using a hybrid AA-FCM framework, rather than AA alone, lies in the preservation of data integrity. Although AA excels at identifying characteristic structural topographies (ATs), deriving coefficients solely through AA can incur projection loss when high-dimensional OCT data is forced into a lower-dimensional simplex. By integrating FCM, we assign decomposition weights based on the Euclidean distance between each eye's GCL++ map and the discovered ATs. This distance-based membership minimizes projection artifacts and ensures that the coefficients used for progression analysis are a faithful, lossless representation of each eye's unique structural state.

AA identifies extreme representative patterns located at the boundaries of the data distribution and represents each observation as a mixture of these ATs. Conceptually, this process is analogous to representing colors as combinations of primary colors. Formally, each observation can be approximated as a convex combination of ATs:

$$\begin{array}{c}\hfill X\approx ZA\end{array}$$

where X represents the original data matrix, Z denotes the matrix of ATs, and A represents the decomposition coefficients describing the contribution of each AT to an individual eye. The ATs and coefficients are estimated by minimizing the residual sum of squares between the original and reconstructed data.

After identifying the ATs, FCM clustering was applied to obtain decomposition coefficients for each eye. Unlike hard clustering methods such as k-means, FCM allows each observation to belong to multiple clusters with varying degrees of membership, which is particularly suitable for representing gradual and overlapping structural damage patterns in glaucoma.

FCM estimates cluster memberships by minimizing the following objective function:

$$\begin{array}{c}\hfill min{J}_m\left(U,V\right)=\sum _{i=1}^N\sum _{j=1}^K{u_{ij}^m∥x_i-c_j∥}^2\end{array}$$

where U is the membership matrix, V represents the cluster centers, and m > 1 is the fuzzifier that controls the degree of softness. The membership values U_ki and cluster centers v_i are iteratively updated based on the following update rules.

The algorithm alternated between updating U and v until convergence was achieved. This formulation ensured a robust clustering mechanism, allowing the flexible assignment of data points across multiple clusters, making FCM particularly effective for datasets with overlapping structures.

Machine Learning for Evaluation of AA and FCM Decomposition Results

We developed machine learning models to predict MD and Visual Field Index (VFI) using decomposition coefficient derived from wide-field SS-OCT. We used this prediction to compare the performance of AA and FCM decomposition results. According to GIGO (garbage in, garbage out), we expect that a model with input data that includes more relevant information will make a more accurate prediction of MD and VFI, provided that the same machine learning approach with two different input data is applied.^38,39 In this prediction model, the decomposition coefficient from the wide-field SS-OCT is an independent variable, and the MD and VFI are dependent variables. The dataset was randomly split into training, test, and validation sets with a ratio of 0.8:0.1:0.1 using a bootstrap-based sampling strategy. The training set was generated using bootstrap sampling with replacement, while the test and validation sets were preferentially sampled from out-of-bag instances that were not included in the bootstrap training samples.

To build MD and VFI prediction models, we applied elastic net, lasso, linear regression, random forest, and ridge. For each algorithm, two separate models were trained: one using AA decomposition coefficients and the other using FCM decomposition coefficients as input features. All models were implemented using the Python scikit-learn library (version 0.24.2). We used the root mean squared error (RMSE) as the predictive evaluation metric. To ensure robustness of the results, the model training and evaluation procedure was repeated 100 times with different random seeds, while using identical data partitions for AA and FCM in each iteration. To assess whether there were statistically significant differences between the AA and FCM models, paired t-tests were performed on the 100 paired results obtained from the repeated experiments.

$$\begin{array}{c}\hfill RMSE=\sqrt{\frac{1}{N}\sum _{i=1}^N{\left(y_i-x_i\right)}^2}\end{array}$$

where x_i = predicted MD or VFI (by AA and FCM) and y_i = actual MD or VFI

Statistical Analyses

Because both eyes of a single patient were included in the analysis, ocular variables were not independent because of within-subject correlation. This was accounted for using mixed-effects models, incorporating random effects to adjust the inter-eye correlation in addition to fixed effects.^40,41 Differences in the age, IOP, axial length, visual acuity, spherical equivalent, central corneal thickness, circumpapillary total GCL++ thickness, macular total GCL++ thickness, MD, pattern standard deviation, VFI, and decomposition coefficients of the ATs were compared among the different types of glaucoma using a linear mixed model. Differences in sex and eye laterality among different glaucoma types were compared using a binomial logistic mixed model.

The primary purpose of the progression analysis was to determine whether baseline structural patterns identified via the hybrid AA-FCM approach could predict future functional deterioration. To ensure a comprehensive assessment of functional loss, progression was defined using three distinct clinical criteria based on longitudinal 24-2 VF data (minimum of five reliable tests):

• 1.Global Indexes (MD and VFI slopes): The slopes of MD and VFI over time were calculated using linear regression. Progression was defined as a significantly negative slope (P < 0.05).^42–44 To ensure the robustness of the analysis, we additionally performed analyses using stricter criteria of an MD rate of change of −0.5 dB/y or less and a VFI rate of change of −1.8%/y or less.⁴⁵
• 2.Local Sensitivity (Pointwise Linear Regression [PLR]): To capture focal progression, regression slopes of total deviation values at 52 test locations were calculated. VF progression was defined if three or more points displayed a regression slope ≤ −1.0 dB/year with P < 0.01.⁴⁵

To identify which baseline structural ATs were independent predictors of subsequent functional decline, we used generalized estimating equations (GEE) to account for within-subject correlation and the goodness-of-fit of the GEE models was compared using Quasi-likelihood under the independence model criterion (QIC). The final associations were reported as odds ratios with 95% confidence intervals, defining baseline ATs as the predictors and the binary progression status (progressor vs. non-progressor) as the outcome for each of the three criteria.

To account for multiple comparisons among the three severity groups (glaucoma suspect, early glaucoma, and moderate-to-advanced glaucoma), a Bonferroni correction was applied. The significance threshold was set at P < 0.017 (calculated as 0.05 divided by the three pairwise comparisons). This adjusted alpha level was used to evaluate the statistical significance of differences in AT decomposition coefficients between the groups. For all other analyses, a P-value of less than 0.05 was considered statistically significant. Statistical analysis was performed using R version 4.4.2 (R Core Team, 2024, http://cran.r-project.org).

Results

Initially, 2624 sets of wide-field SS-OCT data were collected. After excluding scans with missing values, 1922 sets remained. Among them, 1803 SS-OCT scans from 1803 eyes of 1167 patients with glaucoma or suspected glaucoma were included after applying the exclusion criteria. The mean circumpapillary and macular GCL++ thickness were 123.57 ± 23.63 µm and 93.81 ± 15.12 µm, respectively. MD and VFI were −6.10 ± 7.04 dB and 84.33% ± 21.08%, respectively (Table 1).

Table 1.

Demographic and Clinical Characteristics of the Study Population, Stratified by Diagnosis and Glaucoma Severity

Characteristics	Total	GS	EG	AG	P Valuea
Number of eyes (patients)	1803 (1167)	564 (377)	649 (393)	590 (397)
Age (years)	59.73 ± 14.04	56.76 ± 15.24	59.37 ± 13.41	62.97 ± 12.80	0.318
Sex, Female	968 (54%)	336 (59%)	367 (57%)	265 (45%)	0.411
Eye laterality, right	903 (50%)	294 (52%)	300 (46%)	309 (52%)	0.049
Intraocular pressure (mm Hg)	16.43 ± 4.69	16.45 ± 3.69	16.38 ± 4.83	16.46 ± 5.35	0.090
Visual acuity (logMAR)	0.15 ± 0.19	0.11 ± 0.14	0.12 ± 0.15	0.22 ± 0.24	<0.001
Axial length (mm)	24.17 ± 1.45	24.12 ± 1.36	24.21 ± 1.51	24.18 ± 1.47	0.261
Spherical equivalent (D)	−1.18 ± 2.70	−1.07 ± 2.83	−1.19 ± 2.63	−1.27 ± 2.65	<0.001
Central corneal thickness (µm)	541.56 ± 39.01	542.68 ± 38.85	543.92 ± 39.64	537.94 ± 38.28	0.261
Circumpapillary total GCL++ thickness (µm)	123.57 ± 23.63	141.96 ± 15.01	127.55 ± 16.70	101.62 ± 19.01	<0.001
Macular total GCL++ thickness (µm)	93.81 ± 15.12	105.48 ± 9.07	95.69 ± 11.32	80.58 ± 13.06	<0.001
Mean deviation (dB)	−6.10 ± 7.04	−1.81 ± 3.12	−2.41 ± 1.95	−14.27 ± 6.22	<0.001
Pattern standard deviation (dB)	5.35 ± 4.21	2.41 ± 1.82	3.46 ± 2.24	10.23 ± 3.08	<0.001
Visual Field Index (%)	84.33 ± 21.08	96.66 ± 7.30	95.01 ± 4.48	60.80 ± 21.47	<0.001

AG, moderate to advanced glaucoma; EG, early glaucoma; GS, glaucoma suspect.

Values are reported as mean ± standard deviation.

^a P value of continuous variables from a linear mixed model and that of categorical variables from the logistic mixed model. Significance level α = 0.05.

Representation of Characteristic Wide-Field GCL++ Defect Patterns Using 10 ATs

AA was performed using 1803 GCL++ thickness data from 1803 eyes of 1167 patients. For each eye, 60 × 45 GCL++ thickness values were extracted from the wide-field superpixel grid of the SS-OCT images. FCM decomposition was applied to characterize the damage patterns in individual eyes.

Figure 1 represents 10 characteristic wide-field GCL++ defect patterns identified using AA, with the average FCM decomposition weights. AT1 represented a normal pattern, and AT2 corresponded to early macular loss. AT3 exhibited inferior loss, whereas AT4 and AT5 exhibited superior defects. AT6 showed a nasal loss. AT7, AT8, and AT9 showed bihemiretinal loss patterns. AT10 denoted the total loss.

Figure 1.

Characteristic wide-field swept-source OCT pattern identified using archetypal analysis based on GCL++ thickness. Blue indicates areas of low thickness, whereas red indicates areas of high thickness. (A) Normal pattern with 24-2 visual field test points. Orange circle indicates fovea center, and navy circle indicates optic disc center. (B) Ten characteristic patterns. The mean fuzzy c-means decomposition ratios for each archetype are shown on the right. All OCT scans are provided in right-eye format.

Comparison of the Predictive Performance of AA and FCM Decomposition

We used 1803 eyes of 1167 patients as MD and VFI prediction models. The RMSE values of the FCM decomposition were significantly lower than those of the AA decomposition in all five prediction models (all P ≤ 0.002). These results represent that VF prediction using FCM decomposition derived from SS-OCT is superior to AA decomposition (Table 2).

Table 2.

Comparison of Prediction Error (Root Mean Squared Error, dB) Between Archetypal Analysis Decomposition and Fuzzy C-Means Decomposition in Predicting Mean Deviation and Visual Field Index Using Machine Learning Models

	Mean deviation Prediction (dB)			Visual Field Index Prediction (%)
	AA	FCM	P Valuea	AA	FCM	P Valuea
ElasticNet	4.940 ± 0.299	4.473 ± 0.340	<0.001	14.701 ± 0.992	13.086 ± 1.159	<0.001
Lasso	4.939 ± 0.298	4.471 ± 0.340	<0.001	14.692 ± 0.990	13.087 ± 1.161	<0.001
Linear regression	4.869 ± 0.295	4.479 ± 0.342	<0.001	14.511 ± 0.981	13.108 ± 1.165	<0.001
Random forest	4.576 ± 0.323	4.510 ± 0.325	0.002	13.469 ± 1.124	13.183 ± 1.156	<0.001
Ridge	4.933 ± 0.298	4.474 ± 0.341	<0.001	14.683 ± 0.989	13.089 ± 1.161	<0.001

AA, archetypal analysis; FCM, fuzzy c-means.

^aSignificance level α = 0.05.

Characteristic Wide-Field GCL++ Defect Pattern Stratified by Glaucoma Severity

The most frequently observed pattern in the glaucoma suspect group was early macular loss (AT2), whereas bihemiretinal loss (AT9) was the most prevalent pattern in both the early and moderate-to-advanced glaucoma groups. Comparisons among the three groups demonstrated statistically significant differences in the distribution of decomposition coefficients for all ATs (all P ≤ 0.016 for multiple comparisons) except for AT5 (Table 3).

Table 3.

Comparison of the Mean Decomposition Coefficient for Each Archetype Among Glaucoma Suspect, Early Glaucoma, and Moderate to Advanced Glaucoma Groups

					P Valueb
	GS	EG	AG	P Valuea	GS vs. EG	EG vs. AG	GS vs. AG
AT1	0.078 ± 0.039	0.052 ± 0.037	0.029 ± 0.014	<0.001	<0.001	<0.001	<0.001
AT2	0.151 ± 0.079	0.093 ± 0.053	0.047 ± 0.029	<0.001	<0.001	<0.001	<0.001
AT3	0.097 ± 0.021	0.109 ± 0.041	0.087 ± 0.043	<0.001	<0.001	<0.001	<0.001
AT4	0.128 ± 0.050	0.100 ± 0.041	0.062 ± 0.031	<0.001	<0.001	<0.001	<0.001
AT5	0.151 ± 0.047	0.144 ± 0.057	0.108 ± 0.056	<0.001	0.065	<0.001	<0.001
AT6	0.087 ± 0.033	0.080 ± 0.024	0.065 ± 0.018	<0.001	0.016	<0.001	<0.001
AT7	0.070 ± 0.020	0.074 ± 0.016	0.066 ± 0.019	<0.001	<0.001	0.002	<0.001
AT8	0.066 ± 0.013	0.077 ± 0.025	0.072 ± 0.018	<0.001	<0.001	<0.001	<0.001
AT9	0.112 ± 0.041	0.178 ± 0.071	0.233 ± 0.078	<0.001	<0.001	<0.001	<0.001
AT10	0.061 ± 0.028	0.093 ± 0.054	0.232 ± 0.152	<0.001	<0.001	0.002	<0.001

AG, moderate to advanced glaucoma; EG, early glaucoma; GS, glaucoma suspect.

P value of continuous variables from a linear mixed model.

^aSignificance level α = 0.05.

^bSignificance level α = 0.017 for multiple comparisons.

The normal pattern (AT1), focal loss patterns (AT4, AT5, and AT6), and early macular loss pattern (AT2) demonstrated decreasing trends as glaucoma progressed from the suspected to the advanced stage, whereas the diffuse loss patterns (AT9 and AT10) displayed increasing trends. Inferotemporal dominant loss patterns (AT3, AT7, and AT8) initially increased from the suspected glaucoma stage to early glaucoma; however, these subsequently decreased with the progression of the disease to the moderate-to-advanced stage (Fig. 2).

Figure 2.

Distribution of decomposition coefficient for archetypes according to glaucoma severity. Five archetypes (AT1, AT2, AT4, AT5, and AT6) exhibited a decreasing trend, whereas two archetypes (AT9 and AT10) displayed an increasing trend as glaucoma progressed from early to advanced stages. Inferior-dominant loss patterns (AT3, AT7, and AT8) increased from the glaucoma suspect stage to early glaucoma, followed by a decrease from early to advanced stages. The most common pattern in glaucoma suspect eyes was early macular loss (AT2), whereas bihemiretinal loss (AT9) was the most prevalent pattern in both early and moderate to advanced glaucoma groups. *P < 0.001, **P < 0.017. Significance level α = 0.017 for multiple comparisons.

In the correlation analysis with MD, AT1 (r = 0.590), AT2 (r = 0.519), and AT4 (r = 0.547) showed a positive correlation. By contrast, AT9 (r = −0.483) and AT10 (r = −0.778) showed a negative correlation.

AT6 appears to be less characteristic of primary glaucomatous damage. Correlation analysis between AT6 and baseline characteristics revealed significant associations with age (r = −0.246), axial length (r = 0.314), and refractive error (r = −0.253) (all P < 0.001). Of the 1803 eyes, 34 demonstrated AT6 as the highest value among the 10 ATs. The mean axial length of the 34 eyes was 26.39 ± 1.63 mm, the mean refractive errors was – 3.83 ± 4.68 D, and the mean age was 42.94 ± 16.05 years.

Baseline ATs Associated With Glaucoma Progression

To investigate whether the baseline archetypal patterns possess independent prognostic value beyond existing clinical classifications, we performed a within-stage progression analysis with 372 eyes from 279 patients with at least five reliable 24-2 VFs. By stratifying eyes into glaucoma suspect, early glaucoma, and moderate-to-advanced glaucoma groups, we identified specific structural ATs that predict future functional decline independently of the baseline severity. No significant differences were reported in age, IOP, axial length, refractive errors, or central corneal thickness among groups (all P ≥ 0.283) (Table 4). Among the 372 eyes with at least five reliable VF tests, 16 eyes (4.3%) met all three progression criteria, whereas 304 eyes (81.7%) did not meet any. This highlights the limited overlap between the criteria, suggesting that each may capture distinct aspects of functional progression (Supplementary Fig. S5).

Table 4.

Demographic Characteristics of Patients Who Underwent Five or More 24-2 Visual Field Tests According to Glaucoma Severity

Characteristics	Total	GS	EG	AG	P Valuea
Number of eyes (patients)	372 (279)	64 (50)	156 (112)	152 (117)
Age (years)	57.17 ± 13.19	55.17 ± 13.48	56.20 ± 13.27	59.01 ± 12.82	0.623
Sex, Female	184 (49.5%)	41 (64.1%)	86 (55.1%)	57 (37.5%)	<0.001
Eye laterality, right	194 (52.2%)	47 (71.9%)	65 (41.7%)	82 (54.6%)	<0.001
Intraocular pressure (mm Hg)	16.67 ± 4.57	16.66 ± 3.53	16.40 ± 4.43	16.96 ± 5.07	0.309
Visual acuity (logMAR)	0.15 ± 0.18	0.14 ± 0.18	0.10 ± 0.14	0.20 ± 0.21	<0.001
Axial length (mm)	24.32 ± 1.45	24.10 ± 1.18	24.40 ± 1.48	24.34 ± 1.52	0.317
Spherical equivalent (D)	−1.75 ± 2.90	−1.81 ± 3.19	−1.75 ± 2.79	−1.72 ± 2.89	0.968
Central corneal thickness (µm)	541.96 ± 38.61	548.09 ± 40.96	541.63 ± 39.72	539.69 ± 36.37	0.283
Circumpapillary total GCL++ thickness (µm)	118.46 ± 24.15	141.04 ± 16.23	127.11 ± 15.56	100.07 ± 21.00	<0.001
Macular total GCL++ thickness (µm)	90.07 ± 15.23	103.64 ± 10.23	95.53 ± 10.36	78.74 ± 13.36	<0.001
Mean deviation (dB)	−7.32 ± 7.45	−1.53 ± 1.78	−2.60 ± 1.96	−14.61 ± 6.37	<0.001
Pattern standard deviation (dB)	6.13 ± 4.46	1.96 ± 0.82	3.58 ± 2.30	10.50 ± 3.15	<0.001
Visual Field Index (%)	81.24 ± 22.81	97.98 ± 2.57	95.17 ± 4.13	59.89 ± 21.90	<0.001

P value of continuous variables from a linear mixed model, and that of categorical variables from the logistic mixed model.

^aSignificance level α = 0.05.

Table 5 provides the results of the GEE models analyzing the associations of circumpapillary total GCL++ thickness, macular total GCL++ thickness, and baseline ATs as predictors of VF progression in 372 eyes of 279 patients with varying degrees of glaucoma severity. Supplementary Table S1 presents the results obtained from the same subjects using stricter VF progression criteria than those applied in Table 5.

Table 5.

Results of Generalized Estimating Equations Analysis Assessing the Association Between the GCL++ Thickness and Baseline Archetypes as a Predictor of Visual Field Progression in 372 Eyes From 279 Patients With Varying Glaucoma Severity

	Mean Deviation Slope				Visual Field Index Slope				Pointwise Linear Regression
Variables	Odds Ratio	95% CI	P Valuea	QIC	Odds Ratio	95% CI	P Valuea	QIC	Odds Ratio	95% CI	P Valuea	QIC
Glaucoma suspect
Circumpapillary total GCL++ thickness	0.956	0.901 – 1.014	0.135	27.633	0.936	0.879 – 0.996	0.038*	19.531	0.934	0.872 – 1.000	0.050	33.050
Macular total GCL++ thickness	0.963	0.872 – 1.065	0.465	29.255	0.934	0.845 – 1.033	0.183	20.811	0.927	0.845 – 1.017	0.109	34.012
AT9	1.191	0.977 – 1.450	0.083	25.643					1.297	1.048 – 1.606	0.017*	27.626
Early glaucoma
Circumpapillary total GCL++ thickness	0.985	0.954 – 1.017	0.347	112.302	0.985	0.954 – 1.018	0.371	143.034	0.990	0.909 – 1.054	0.746	76.158
Macular total GCL++ thickness	0.958	0.914 – 1.004	0.075	110.489	0.941	0.873 – 1.013	0.108	148.003	0.994	0.928 – 1.065	0.865	74.421
AT2	1.213	0.995 – 1.480	0.057	102.511
AT4									1.400	1.015 – 1.932	0.041*	64.796
AT6									1.527	0.779 – 2.993	0.218
AT9	1.195	1.050 – 1.359	0.007*						1.307	0.971 – 1.760	0.077
Moderate to advanced glaucoma
Circumpapillary total GCL++ thickness	0.980	0.961 – 0.999	0.041*	135.558	0.980	0.962 – 0.997	0.025*	148.714	0.993	0.969 – 1.017	0.542	78.524
Macular total GCL++ thickness	0.975	0.945 – 1.007	0.121	137.050	0.972	0.943 – 1.001	0.056	149.448	1.000	0.964 – 1.037	0.997	78.825
AT1	0.765	0.578 – 1.012	0.061	136.144	0.720	0.538 – 0.963	0.027*	148.465
AT5									1.123	1.014 – 1.243	0.025*	75.541

AT, archetype; CI, confidence interval; QIC, Quasi-likelihood under the independence model criterion.

^aP value from a generalized estimating equation.

^*Significance level α = 0.05.

In the glaucoma suspect group, higher bihemiretinal loss pattern (AT9) was associated with an increased risk of PLR progression (P = 0.017). In the PLR model, the QIC value for the hybrid model was 27.626, which was lower than those of the OCT measures. In the early glaucoma group, higher bihemiretinal loss pattern (AT9) was associated with an increased risk of MD progression (P = 0.007) and higher early superior loss pattern (AT4) was associated with an increased risk of PLR progression (P = 0.041). In the MD slope and PLR models, the QIC values for hybrid models were 102.511 and 64.796, respectively, which were lower than those of the OCT measures. In the moderate to advanced glaucoma group, lower normal pattern (AT1) was associated with VFI progression (P = 0.027), and higher superior hemiretinal loss pattern (AT5) was associated with PLR progression (P = 0.025). In the VFI slope and PLR models, the QIC values for the hybrid models were 148.465 and 75.541, respectively, which were also lower than those of the OCT measures.

When progression was defined as MD rate of change of −0.5 dB/y or less, the numbers of progressed eyes in the glaucoma suspect, early glaucoma, and moderate-to-advanced glaucoma groups were 2 (3.13%), 15 (9.62%), and 24 (15.79%), respectively. When VFI rate of change of −1.8%/y or less was applied, the numbers were 2 (3.13%), 18 (11.54%), and 27 (17.76%), respectively.

Even when stricter VF progression criteria were applied, higher bihemiretinal loss pattern (AT9) was associated with an increased risk of MD progression in the glaucoma suspect and early glaucoma groups (P = 0.019 and P = 0.016, respectively). Lower normal pattern (AT1) was associated with VFI progression in the moderate to advanced glaucoma group (P = 0.016) (Supplementary Table S1).

Supplementary Table S2 presents the area under curve (AUC) values and cutoff points of the ATs that were significant in the GEE models shown in Table 5. In the glaucoma suspect group, the AUC values for predicting PLR progression were 0.792 for AT9, indicating good diagnostic performance.

Representative Cases

Supplementary Fig. S6 illustrates two representative cases decomposed by FCM. Supplementary Fig. S6A shows early glaucoma with superior RNFL defect. Superior hemiretinal loss pattern (AT5, 51.1%) is the representative pattern. Supplementary Fig. S6B shows advanced glaucoma with inferior-dominant bihemiretinal RNFL defect. Bihemiretinal loss pattern (AT9, 51.3%) accounts for the largest proportion.

Supplementary Fig. S7 illustrates the longitudinal changes in VF along with representative OCT patterns. Supplementary Fig. S7A shows early glaucoma at the initial visit. Baseline wide-field SS-OCT shows inferior RNFL defect. Inferior hemiretinal loss (AT3) is the most prominent pattern. VF showed superior arcuate loss with central involvement, which remained stable without progression over seven years. All three VF progression criteria—MD slope, VFI slope, and PLR analysis—demonstrated no progression. Supplementary Fig. S7B shows moderate stage glaucoma at the initial visit. Baseline wide-field SS-OCT shows superior RNFL defect. Superior hemiretinal loss pattern (AT5) is the representative SS-OCT damage pattern. On the initial VF test, there was an inferior arcuate loss with central involvement, and the VF defect further extended inferiorly over six years. All three VF progression criteria—MD slope, VFI slope, and PLR analysis—demonstrated progression. Supplementary Fig. S7C shows early glaucoma at the initial visit. Baseline wide-field SS-OCT shows inferior-dominant bihemiretinal RNFL defect. Bihemiretinal damage pattern (AT9) is the most predominant. At the initial visit, the VF defect showed superior arcuate loss with central involvement, which progressed to a double arcuate pattern over 6.5 years. All three VF progression criteria—MD slope, VFI slope, and PLR analysis—demonstrated progression.

Discussion

To the best of our knowledge, this is the first study to apply unsupervised machine learning to classify and analyze glaucomatous defect patterns in wide-field SS-OCT images of patients with glaucoma. Using AA, we identified 10 characteristic wide-field patterns of glaucomatous damage and applied FCM to decompose the patterns of individual patients. Eyes with baseline wide-field SS-OCT results with higher bihemiretinal loss were more likely to demonstrate VF progression in the glaucoma suspect group. Eyes with baseline wide-field SS-OCT results exhibiting increased bihemiretinal and early superior losses were more likely to show VF progression in the early glaucoma group. Eyes with baseline wide-field SS-OCT results showing less normal pattern and more superior hemiretinal losses were more likely to show VF progression in the moderate to advanced glaucoma group. Even when more stringent VF progression criteria were applied, higher bihemiretinal loss pattern was associated with an increased risk of VF progression in the glaucoma suspect and early glaucoma group. Lower normal pattern was associated with VF progression in the moderate to advanced glaucoma group. Furthermore, we quantitatively compared the performance of AA and FCM decomposition in predicting MD and VFI using machine learning models, demonstrating the superiority of the hybrid approach.

OCT evaluates the macular and circumpapillary regions to detect changes in the RGCs and their axons, as the density of RGCs is the highest in the macula, and their axon bundles are concentrated in the circumpapillary region where the RNFL is the thickest.⁴⁶ A 12 × 9 mm wide-field scan mode has become available with the recent introduction of SS-OCT, allowing simultaneous imaging of the optic nerve head and macula. Wide-field imaging offers superior visualization of the temporal margin and the peripheral areas that extend beyond the optic disc.⁴⁷ Lee et al.⁴⁷ reported that the wide-field RNFL thickness map obtained using SS-OCT demonstrated the highest sensitivity for diagnosing preperimetric and early glaucoma compared to cpRNFL, macular GCL+, and GCL++ thickness maps based on abnormality criteria. The GCL++ thickness exhibited a higher AUC value than the individual RNFL and GCL+ thicknesses, although the difference was not statistically significant, suggesting that the diagnostic performance of GCL++ thickness is comparable to that of the individual RNFL and GCL+ thicknesses.^14,15 Yang et al.¹⁴ suggested that continuous maps of abnormalities may be more useful clinically. Based on these findings, we selected to use wide-field GCL++ scans for the analysis.

The number of ATs was set to 10, which was clinically interpretable and had a low reconstruction error. AT1 represented an intact pattern, whereas AT2–AT10 represented different patterns of glaucomatous damage in the GCL++ map. Normal (AT1) and early damage patterns (AT2, AT4) showed a positive correlation with MD, whereas advanced patterns (AT9, AT10) showed a negative correlation. Early macular loss (AT2) was the most common pattern observed in the glaucoma suspect group. This finding is in agreement with the results of previous studies, which reported that approximately 50% of the RGCs were located within 4.5 mm of the foveal center, and 16% of the eyes with normal 24-2 VF were classified as abnormal on 10-2 VF.^48,49

In the glaucoma group, normal and focal loss patterns showed a decreasing tendency, whereas diffuse loss patterns showed an increasing tendency as glaucoma worsened. Subsequently, bihemiretinal loss (AT9), followed by total loss (AT10) was the most common pattern in the moderate to advanced glaucoma group. Among the 10 ATs, AT6 (nasal loss), which is considered less characteristic of primary glaucomatous damage, was also included. In a subgroup analysis of the 34 eyes in which AT6 was the predominant pattern, a myopic tendency was observed, which may be attributable to the nasal RNFL thinning associated with myopia.^50,51

Overall, three wide-field SS-OCT patterns showed inferotemporal dominant losses (AT3, AT7, and AT8), whereas two patterns displayed superotemporal dominant losses (AT4 and AT5). In addition, inferotemporal dominant losses (AT3, AT7, and AT8), which indicated glaucomatous damage in the macular vulnerability zone (MVZ), tended to increase from the glaucoma suspect stage to early glaucoma, followed by a decrease as glaucoma progressed to advanced glaucoma. This pattern likely reflects the unique pathophysiology of glaucomatous progression in the macula. The initial rise in these coefficients aligns with the concept of the MVZ in the inferior-temporal retina, which is often the first site of measurable structural loss in early glaucoma.¹² Hood et al.¹² proposed a schematic model of the MVZ, highlighting its susceptibility to glaucomatous damage. Although their study reported damage patterns based on a subjective visual inspection, our study objectively identified glaucoma patterns using unsupervised machine learning. By objectively presenting wide-field SS-OCT patterns, we achieved an unbiased classification of glaucomatous damage. However, as the disease reaches more advanced stages, these focal coefficients tend to decrease. This decrease is accompanied by a concomitant rise in coefficients for AT9 (bihemiretinal loss) and AT10 (total loss). This suggests that as localized inferior defects expand to encompass both hemispheres, the hybrid model mathematically re-allocates the weights from focal ATs toward more diffuse, higher-severity patterns. Therefore, the decrease in focal inferior ATs in advanced stages should be interpreted as a transition toward comprehensive structural collapse rather than a regression of damage.

The decomposition coefficients of all ATs showed statistically significant changes in proportion from the early to advanced stages. These results suggest that glaucoma progression can be detected on wide-field SS-OCT using our hybrid unsupervised approach. AT3, AT7, and AT8 exhibited statistically significant differences between the early and moderate to advanced stages, although the changes were minimal, indicating that FCM decomposition was capable of detecting even minor changes. These results can be attributed to the inherent advantage of the FCM algorithm, which enables lossless decomposition.^22,23 Unlike AA, FCM decomposition reflects the original distances in the Euclidean space, thereby overcoming the projection loss associated with AA.^52,53

In the subgroup analysis of eyes with at least five reliable VFs, GEE analysis identified AT9 (bihemiretinal loss) as a significant predictor of VF progression in glaucoma suspect and early glaucoma. Importantly, although AT9 (bihemiretinal loss) was associated with VF progression, its biological and clinical implications likely differ according to glaucoma severity. In glaucoma suspect eyes, AT9 may reflect early, spatially diffuse structural vulnerability that precedes overt functional loss. In contrast, in early glaucoma, AT9 likely represents cumulative or advanced bihemiretinal damage indicative of established structural reserve depletion. Thus the same AT may capture different pathophysiological states depending on disease stage, rather than implying a uniform mechanism of progression across the glaucoma spectrum.

AT4 (early superior loss) and AT5 (superior hemiretinal loss) were significant predictors of PLR progression in the early glaucoma and moderate to advanced glaucoma, respectively. These findings are consistent with those of previous studies that reported reduced OCT-derived superior quadrant RNFL thickness is a significant predictor of VF progression.⁵⁴ Nouri-Mahdavi et al.⁵⁵ demonstrated that the baseline superior hemimacular GCL+ thickness predicted VF progression according to the PLR criteria in moderate to advanced glaucoma. Although previous studies relied solely on OCT structural parameters to predict VF progression,^54,55 we used ATs derived using a novel hybrid approach that combines AA and FCM. In addition, Wang et al.³⁷ reported that a higher coefficient of inferonasal 10-2 VF loss, corresponding to superotemporal GCL++ damage, is associated with faster VF progression.

A potential concern in unsupervised learning is whether the discovered patterns merely reflect pre-existing clinical labels, a phenomenon often referred to as 'label leakage'. However, our ATs were derived in a completely label-agnostic manner, using only raw GCL++ thickness maps. The validity of these patterns as independent structural phenotypes is supported by two findings: first, the non-monotonic trend of certain focal loss patterns (AT3, AT7, and AT8) across severity stages; and second, their ability to predict future functional loss within patients of the same baseline severity. These results suggest that our hybrid AA-FCM approach identifies “sub-phenotypes” of glaucomatous damage that provide incremental prognostic information beyond conventional MD-based or thickness-based classification.

A key strength of the proposed hybrid AA-FCM framework is its potential for longitudinal monitoring. Although the current study evaluated structural ATs only at baseline to predict subsequent functional loss, tracking longitudinal shifts in AT coefficients could provide a more nuanced understanding of disease trajectories. Specifically, measuring the rate of change (slope) in decomposition coefficients for each AT may allow clinicians to identify eyes where structural patterns are actively transitioning from normal (AT1) to pathological states.

Such longitudinal AT analysis might detect progression earlier than conventional global or sectoral thickness metrics, as it captures localized spatial redistributions of the GCL++. For instance, a rapid increase in the coefficient for AT9 (bihemiretinal loss) could serve as a quantitative biomarker for aggressive disease progression, signaling the need for more intensive treatment before significant VF loss occurs. As shown in Supplementary Figure S7, even when initial findings consistent with early glaucoma, higher baseline bihemiretinal loss pattern (AT9) is associated with a faster rate of progression. Future studies using longitudinal wide-field OCT data will be essential to validate these AT-based structural slopes as markers for glaucoma progression.

The limited overlap among the three progression criteria underscores the need for multi-faceted approaches to detect functional deterioration. Furthermore, it suggests that the structural patterns associated with disease progression may differ depending on the selected reference standard. As our study stratified patients by disease stage (suspected glaucoma, early glaucoma, and moderate-to-advanced glaucoma), the observed structural-functional associations may reflect the definition of progression and the stage-specific characteristics of glaucomatous damage. These results suggest that multiple structural ATs may contribute to VF progression in different ways and that their clinical utility may vary depending on both the progression criterion and disease stage.

The substantial discordance among the three functional progression criteria, where only 4.3% of eyes met all definitions, highlights the different spatial scales at which these algorithms operate, as noted by Saeedi et al.⁵⁶ Our findings suggest that specific structural ATs at baseline can help predict which functional criteria an eye is most likely to fail. Specifically, focal ATs such as AT4 (early superior loss) and AT5 (superior hemiretinal loss) were more closely associated with PLR progression, identifying localized functional decay that global averages might initially overlook. Conversely, AT9 (bihemiretinal loss) showed significant associations with both global and local progression criteria. This suggests that extensive bihemiretinal thinning represents a high-risk phenotype for overall functional collapse.

Our study had certain limitations. First, more ATs with inferotemporal damage were present than with superotemporal damage. As the AA classified inferior damage into four distinct patterns across the entire cohort, the mean decomposition coefficient for inferior damage patterns may have been underestimated compared with that for superior loss patterns. However, when the ATs associated with glaucomatous damage in the MVZ were combined, the mean decomposition coefficient for inferior damage patterns was greater than that for superior damage patterns. Second, a longitudinal analysis of the wide-field SS-OCT data was not performed because of the retrospective nature of the study. Pattern analysis may be useful for tracking the longitudinal changes in glaucomatous damage patterns over time. In future studies, we plan to investigate longitudinal changes in ATs to better detect glaucoma progression. Third, the study population consisted entirely of Asian participants, which may limit the generalizability of our identified ATs to other ethnic groups. Anatomical variations, such as optic disc size and axial length, differ significantly across races. Previous studies have reported that Asians and individuals of African descent often have larger optic disc areas compared to Caucasians, which can influence the physiologic cup-to-disc ratio and the spatial distribution of the RNFL.⁵⁷ Furthermore, the high prevalence of myopia and longer axial lengths in East Asian populations may lead to different structural vulnerability patterns, such as more pronounced temporal or inferotemporal thinning.⁵⁸ OCT parameters also exhibit racial variations; for example, Asians have been shown to have thicker global and superior RNFL but thinner GCL++ and smaller nasal neuroretinal rim values compared to non-Hispanic Whites.^59,60 These ethnic-specific anatomical characteristics likely influenced the formation of our 10 ATs. Therefore, although our hybrid AA-FCM approach is a robust methodology, the specific topography of the ATs might shift when applied to Caucasians or other ethnicities. Future studies incorporating multi-ethnic cohorts are essential to validate these patterns globally. Fourth, one important limitation of this study is the absence of external validation using independent datasets. Although the hybrid clustering approach showed consistent association between certain ATs and VF progression within our cohort, the generalizability of these findings remains to be confirmed. External validation using independent cohorts would strengthen the robustness and clinical applicability of the proposed method and should be addressed in future studies. Furthermore, our study only included eyes that met both structural and functional criteria for glaucoma. Consequently, eyes exhibiting isolated visual VF loss without clinically evident optic neuropathy were excluded. Although this strict definition ensured the specificity of our pattern analysis, it may have limited our ability to capture the full spectrum of very early or atypical glaucoma cases. Future longitudinal studies including a wider variety of glaucoma phenotypes are warranted to further validate our findings.

Another limitation of this study is the absence of a cross-modality validation between OCT-derived and VF–derived ATs. Although the same hybrid AA–FCM framework has previously been applied to VF data, the present study focused exclusively on identifying and validating structural ATs from wide-field SS-OCT GCL++ maps and their associations with subsequent VF progression. A joint analysis directly assessing the alignment between OCT-based ATs and VF–based ATs, ideally in an independent cohort with both modalities available, would provide valuable cross-modality validation and further strengthen the biological and clinical interpretation of the proposed framework. Such an integrated structure–function archetypal analysis represents an important direction for future research.

Conclusions

A hybrid approach combining AA and FCM effectively demonstrated glaucomatous defect patterns in wide-field SS-OCT images. Despite certain limitations, this study represents a significant initial step toward the classification and application of characteristic OCT patterns using unsupervised machine learning. These findings may inform the future development of AI-based diagnostic tools that visualize wide-field OCT patterns in real-time, aid clinicians in the early identification of high-risk eyes, and facilitate personalized monitoring strategies.