Combining structural and microvascular parameters via machine learning for enhanced diagnosis of normal-tension glaucoma among highly myopic eyes: a prospective cross-sectional diagnostic accuracy study Authors: Chang-Zhu He, Qin Qiu, Heng Lai, Chun-Yan Lai, Yu He, Lin Jing Published: 2026 ABSTRACT PurposeThis study aimed to evaluate whether a combination of optical coherence tomography (OCT) and OCT angiography (OCTA) parameters could improve the discrimination between clinically diagnosed normal-tension glaucoma (NTG) and non-glaucomatous high myopia (HM) within a highly myopic population.MethodsIn this prospective cross-sectional diagnostic accuracy study, we consecutively enrolled two groups of participants: patients with high myopia (HM) and those with HM complicated by normal-tension glaucoma (HM-NTG). All clinical and imaging data were collected at a single time point using standardized protocols. Baseline clinical data were collected for all participants. Optic disc structural parameters were acquired using optical coherence tomography (OCT), while optic disc perfusion parameters were obtained via OCT angiography (OCTA). Receiver operating characteristic (ROC) curve analysis was first conducted to evaluate the diagnostic performance of individual parameters. Least absolute shrinkage and selection operator (LASSO) regression was employed for preliminary dimensionality reduction and feature selection, followed by multivariate logistic regression (backward stepwise method) to identify the optimal parameter combination. A diagnostic model was developed based on logistic regression and rigorously validated through bootstrap resampling, calibration assessment, decision curve analysis, and clinical impact curve. A machine learning diagnostic model was constructed using the support vector machine (SVM) algorithm and compared with the conventional regression model. Finally, Shapley additive explanations (SHAP) were applied to interpret the SVM model’s decision-making mechanism and elucidate the individualized contribution of each feature to the prediction outcomes.ResultsA total of 87 patients were enrolled, comprising 42 in the HM group and 45 in the HM-NTG group. Overall, OCT parameters demonstrated superior diagnostic clarity compared with OCTA microvascular indices. The inferior macular ganglion cell complex (GCC) exhibited the strongest discriminatory performance, with an area under the receiver operating characteristic curve (AUC) of 0.85. Among OCTA parameters, the inferior optic disc vessel density (VD) achieved the best diagnostic performance (AUC = 0.72). LASSO regression combined with multivariate logistic regression identified three variables for model construction: inferior GCC thickness (odds ratio [OR] = 0.72), temporal-inferior retinal nerve fiber layer (RNFL) thickness (OR = 0.82), and inferior VD (OR = 0.77). The OCT + OCTA combined model achieved an AUC of 0.909 (95% CI: 0.833–0.984). While this did not significantly exceed the OCT-only model (AUC = 0.888, 95% CI: 0.803–0.973; DeLong test, P = 0.164), the addition of OCTA-derived inferior vessel density yielded significant net reclassification improvement (NRI = 0.679, P = 0.003) and integrated discrimination improvement (IDI = 0.067, P = 0.029), indicating enhanced risk stratification. Bootstrap resampling with 500 iterations yielded an AUC of 0.861 (95% CI: 0.773–0.949). The Hosmer-Lemeshow goodness-of-fit test indicated adequate calibration (χ² = 6.64, P = 0.575), with a Brier score of 0.109. Calibration curves demonstrated close adherence to the ideal diagonal across both low-risk and high-risk probability thresholds. Decision curve analysis indicated favorable net benefit. An SVM model constructed using the same three predictors achieved a comparable AUC of 0.908 (95% CI: 0.833–0.982), providing cross-algorithmic validation of the logistic model’s robustness. SHAP analysis was applied to elucidate individualized feature contributions to patient-specific predictions, confirming that lower values of inferior GCC thickness exerted the greatest directional influence on HM-NTG classification.ConclusionsWithin highly myopic eyes, the combination of inferior GCC thickness, temporal-inferior RNFL thickness, and inferior VD may help distinguish clinically diagnosed HM-NTG from non-glaucomatous HM. The addition of OCTA microvascular parameters did not significantly increase the AUC but improved patient risk reclassification. This multimodal approach, paired with conventional regression and machine learning, offers a promising adjunctive tool for diagnostic evaluation in highly myopic eyes.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12886-026-05065-4. Introduction Glaucoma is a progressive optic neuropathy and a leading cause of irreversible blindness worldwide [1]. Clinically, glaucoma is characterized by structural changes in the optic disc, defects in the retinal nerve fiber layer (RNFL), and functional vision loss [2]. Normal-tension glaucoma (NTG) is a special type of primary open-angle glaucoma (POAG), in patients with NTG, intraocular pressure (IOP) is usually normal but it still shows changes in structure and function [3]. This fact poses challenges to the diagnosis and treatment of ophthalmologists [4]. High myopia (HM) is an independent risk factor for POAG and is commonly defined as a spherical equivalent (SE) of ≤ − 6.00 diopters (D) or an axial length (AL) ≥ 26.0 mm [5]. In HM patients, the stretching of the ocular wall during axial elongation often distorts the optic nerve, resulting in tilted discs and prominent peripapillary atrophy (PPA), this process also precipitates a global, diffuse thinning of the RNFL [6]. These myopia-related changes are morphologically very similar to glaucomatous damage, which increases the difficulty for clinicians to distinguish between high myopic optic neuropathy and glaucomatous optic neuropathy [7]. The mechanism underlying the onset of NTG remains unclear; however, numerous studies have suggested that vascular dysfunction may play a key role. Changes in microvascular density have increasingly been recognized as a potential diagnostic indicator, which could aid in differentiating high myopia-related optic neuropathy from glaucomatous optic neuropathy [2]. Optical coherence tomography angiography (OCTA), as a non-invasive imaging tool, can provide repeatable and objective measurements of retinal microvessels and is increasingly used in the diagnosis of glaucoma [8]. OCTA has become a practical asset for spotting signs of both myopia and glaucoma by giving us a clear view of vessel density (VD) at the optic nerve head. Standard OCT is usually used to measure RNFL thinning for glaucoma, HM often distorts the physical structure of the eye so much that these OCT readings become messy and hard to interpret [9, 10]. Previous studies have shown that combining VD and RNFL measurements can improve diagnostic accuracy, especially in HM patients [11]. In this prospective cross-sectional study, we sought to evaluate whether the integration of OCT structural and OCTA microvascular parameters, analyzed via machine learning, could improve the discrimination between clinically diagnosed HM-NTG and non-glaucomatous HM within a highly myopic population. We acknowledge that the selected parameters may reflect classical glaucomatous changes rather than HM-specific pathophysiology; nevertheless, their combined application may address the diagnostic uncertainty inherent to this anatomically distorted population. The model was constructed by identifying and selecting the most favorable combination of parameters for diagnosis, and machine learning methods were incorporated to assess whether the accuracy of diagnosis can be further improved. Methods Study population Study design and ethics This is a prospectively enrolled cross-sectional diagnostic accuracy study conducted at the Department of Ophthalmology, Chengdu First People’s Hospital, from June 2024 to December 2025. Participants were consecutively recruited based on pre-specified inclusion and exclusion criteria. All clinical assessments, including visual field testing, OCT, and OCTA imaging, were performed at a single time point (baseline visit). The term “prospective” herein refers to the consecutive enrollment and standardized data collection protocol, not to longitudinal follow-up. Ethical approval was originally obtained from the Ethics Committee of Chengdu First People’s Hospital in May 2024 (No. 20240320002759128), prior to the enrollment of the first participant. The approval was subsequently renewed and updated to No. 2026-IIT-0165 when the study period was formally extended. Informed consent was obtained from each participant before any study procedures were performed. Inclusion criteria Each participant completed a standardized clinical workup. This included measuring best-corrected visual acuity (BCVA), which we converted to LogMAR for analysis, alongside non-dilated auto-refraction and non-contact tonometry (CT-1P; Topcon). We also conducted slit-lamp examinations and took fundus photographs using a Visucam 200 (Carl Zeiss Meditec AG). For functional and structural mapping, subjects underwent Zeiss 30 − 2 SITA-Standard visual field testing, while all OCT and OCTA imaging was captured with the RTVue XR Avanti system (Optovue Inc.). The inclusion criteria for this study were as follows: HM Diagnostic Criteria: (1) Spherical equivalent (SE) ≤ -6.00 D or axial length (AL) ≥ 26.0 mm. For data processing consistency, SE was handled as an absolute value (|SE|) in the analytic dataset, with larger values indicating more severe myopia; (2) Clear imaging was achieved with sufficiently transparent refractive media. HM Group: (1) Met the above HM diagnostic criteria; (2) IOP consistently ≤ 21 mmHg without ocular hypotensive medication, confirmed by measurements on at least two separate clinic visits (see IOP Measurement and Verification protocol below); (3) Normal optic disc appearance, with no glaucomatous optic nerve changes. HM-NTG Group: (1) Met the above HM diagnostic criteria; (2) Baseline IOP (without medication) consistently ≤ 21 mmHg, confirmed by measurements on at least three separate clinic visits (see IOP Measurement and Verification protocol below); (3) Clear glaucomatous optic nerve damage confirmed by two independent glaucoma specialists (H.Y and J.L), based on slit-lamp biomicroscopic examination (optic disc cupping, localized rim thinning, or disc hemorrhage) and reproducible glaucomatous visual field defects; (4) Typical glaucomatous visual field defects with anatomical distribution: including paracentral scotomas, arcuate scotomas, or nasal steps, with reproducibility in at least two consecutive tests. Mean deviation (MD) was recorded as an absolute value (|MD|) in the analytic dataset, with larger values indicating greater visual field loss, consistent with the standard Humphrey Field Analyzer output magnitude. The diagnosing physicians did not review quantitative OCT/OCTA parameters (e.g., RNFL or GCC thickness values, vessel density percentages) when establishing the clinical diagnosis. RNFL defects were qualitatively assessed using the color-coded deviation map (green/yellow/red) on the OCT device interface as an auxiliary visual aid only, without reference to specific micrometer values. All visual field tests were required to meet standard reliability criteria: fixation loss rate < 20%, false-positive rate < 15%, and false-negative rate < 20% [12]. Any test failing to meet these criteria was repeated during the same or a subsequent visit until a reliable result was obtained; only reliable visual fields were included in the analysis. Exclusion criteria The exclusion criteria for both groups were as follows: (1) Ocular pathology: Any macular disease other than high myopia (e.g., myopic maculopathy, macular hole), severe non-glaucomatous optic neuropathy, severe dry eye disease, or corneal disease; (2) Surgical and treatment history: History of intraocular surgery (e.g., vitrectomy, anti-glaucoma surgery, excluding simple cataract phacoemulsification), ocular trauma, or laser treatment within the past 3 months; (3) Systemic factors: Presence of uncontrolled systemic diseases, such as hypertension, diabetes (which may interfere with blood flow density measurement), or neurological disorders; (4) Inability to cooperate: Inability to cooperate with the examination or inability to obtain clear imaging. IOP measurement and verification All participants underwent a standardized IOP assessment protocol. At the initial screening visit and a follow-up visit within 2 weeks, IOP was measured using non-contact tonometry (CT-1P; Topcon). At each visit, three consecutive measurements were performed during a single day, and the median value was recorded. The criterion “consistently ≤ 21 mmHg” required that all median values from both visits were ≤ 21 mmHg without ocular hypotensive medication (confirmed by direct patient interview and review of medication history). For participants suspected of glaucoma or NTG, additional IOP assessment was performed using the Corvis ST (Oculus) to measure central corneal thickness (CCT) and corneal biomechanical properties (e.g., stiffness parameter at first applanation). The biomechanically corrected IOP (bIOP) was calculated to exclude the confounding influence of CCT and corneal biomechanics on IOP readings. Only eyes with bIOP consistently ≤ 21 mmHg were eligible. Goldmann applanation tonometry was not used in this study. To rigorously exclude HTG and confirm the NTG phenotype, all suspected NTG patients underwent 24-hour IOP monitoring prior to final enrollment. IOP was measured every 2 h over a 24-hour period (totaling 12 measurements), including nocturnal measurements. Patients with any IOP measurement > 21 mmHg during this 24-hour cycle were excluded from the HM-NTG group. Anterior segment and angle assessment All participants suspected of glaucoma underwent ultrasound biomicroscopy (UBM; SW-3200 L, Suoer) to assess anterior chamber depth and confirm an open angle. If UBM image quality was insufficient due to poor patient cooperation or inadequate resolution, gonioscopy was performed by the glaucoma specialist (H.Y) using a Goldmann 2-mirror goniolens to directly visualize and confirm angle openness. Eyes with angle closure, angle recession, or any evidence of secondary angle abnormality were excluded. Imaging acquisition Peripapillary RNFL imaging was performed using a spectral-domain OCT system (RTVue XR Avanti; Optovue Inc., Fremont, California, USA) equipped with a scanning laser diode with a central wavelength of 840 nm [13, 14]. The RTVue peripapillary scan protocol (4.5-mm diameter) was used to measure RNFL thickness, defined as the distance from the internal limiting membrane (ILM) to the RNFL boundary. Using the device software, we collected RNFL parameters from different regions of the optic disc. The regions included the superior hemifield (SH), inferior hemifield (IH), nasal-superior (NS), nasal-inferior (NI), temporal-superior (TS), and temporal-inferior (TI) quadrants. For the ganglion cell complex (GCC), we selected Average GCC, Superior GCC, and Inferior GCC. All OCTA measurements were performed using the RTVue-XR Avanti system (AngioVue; Optovue Inc.). The device operates at a central wavelength of 840 nm and a scanning speed of 70,000 A-scans per second. It was used to measure superficial microvascular vessel density within a 4.5 × 4.5 mm peripapillary area [15]. We used the system’s built-in deprojection algorithm to remove artifacts from large retinal vessels. All image processing and quantitative analyses were performed using the AngioVue software (version 2018.1.0.43) [16]. Quality control and segmentation verification: To ensure data reliability, we included only scans with a quality index of ≥ 6 in the analysis. Images with motion artifacts were excluded. All OCT and OCTA scans were systematically reviewed by two experienced technicians (L.H and L.C.Y) for segmentation accuracy. The automated segmentation boundaries (ILM, RNFL boundary, and GCC boundaries) were visually inspected on the B-scan images. If segmentation errors were identified due to peripapillary atrophy, optic disc tilt, or posterior staphyloma-related anatomical distortion, manual correction was performed using the device’s built-in editing tool. Cases with severe segmentation failure that could not be reliably corrected were excluded. The two technicians operated independently and were masked to the participant’s clinical diagnosis and group allocation. Disagreements regarding segmentation quality or the necessity of manual correction were resolved by consultation with the glaucoma specialist (H.Y), who was also masked to the quantitative imaging output. All examinations and image reviews were performed by the same standardized team throughout the study. Statistical analysis Baseline comparison and diagnostic performance evaluation Only one eye per participant was included in the final analysis. For participants in whom both eyes met the inclusion criteria, the eye with the worse visual field mean deviation (or more severe glaucomatous damage for the HM-NTG group) was selected; if severity was equivalent between the two eyes, one eye was randomly chosen. Consequently, all 87 eyes were treated as independent observations. The Shapiro-Wilk test was used to assess normality for continuous variables. Normally distributed variables were expressed as mean ± standard deviation (SD) and compared between groups using independent samples t-tests. Non-normally distributed variables were expressed as median (interquartile range [IQR]) and analyzed using the Mann-Whitney U test. Categorical variables were compared using the chi-square test or Fisher’s exact test, as appropriate. For each OCT and OCTA parameter, receiver operating characteristic (ROC) curves were constructed, and the area under the ROC curve (AUC), sensitivity, specificity, accuracy, and Youden index were calculated to evaluate the diagnostic performance of each individual parameter. Multivariable refinement and model construction All continuous OCT and OCTA parameters were z-score standardized (mean-centered and scaled to unit variance) prior to LASSO regression to ensure that the penalization was not dominated by variables with larger measurement scales. LASSO regression with 10-fold cross-validation was employed for preliminary dimensionality reduction. The optimal regularization parameter (λ) was determined using the λ.1se criterion (the most parsimonious model within one standard error of the minimum mean cross-validation error), which balances predictive performance with model simplicity and enhances generalizability. Candidate variables with non-zero regression coefficients at λ.1se were selected as predictors for subsequent multivariate modeling. Variables identified through LASSO regression were subsequently entered into multivariate logistic regression with backward stepwise selection, using the likelihood ratio test as the elimination criterion (retention threshold: P ≤ 0.05). LASSO was used to filter noise and reduce dimensionality in the presence of correlated OCT/OCTA parameters, yielding a manageable candidate set less prone to overfitting in our small cohort. Backward stepwise regression was then applied to this candidate set to retain only variables with independent statistical significance, ensuring the final model provides conventional, interpretable odds ratios. This two-step approach separates exploratory screening (LASSO) from confirmatory inference (stepwise regression). The final model was constructed using variables selected through this two-step procedure. Model performance was compared against individual parameters, with statistical significance of AUC differences assessed using the DeLong test. Exclusion of specific parameters from model development: The cup-to-disc (C/D) ratio was assessed during clinical examination and was used as a component of the reference diagnostic criteria for the HM-NTG group. To prevent incorporation bias, C/D was excluded from the ROC analysis and predictive modeling. Focal loss volume (FLV) and global loss volume (GLV) are software-derived composite metrics based on deviations from normative GCC thickness databases; they were excluded because they are mathematically redundant with the directly measured GCC thickness parameters (Average, Superior, and Inferior GCC) already included in the analysis, and their inclusion would introduce multicollinearity. Model validation Internal validation was performed using bootstrap resampling (500 iterations) to assess model stability. We acknowledge that this bootstrap procedure validates the final fitted model rather than the entire model-building pipeline including LASSO feature selection. Optimal cutoff derivation: The optimal probability cutoff for binary classification was determined as the point on the ROC curve that maximized the Youden index (sensitivity + specificity − 1). Calibration was evaluated with the Hosmer-Lemeshow goodness-of-fit test, with χ² statistics and corresponding P-values calculated; P > 0.05 indicated adequate fit. The Brier score was employed to quantify probabilistic prediction accuracy, with values < 0.25 considered acceptable. Scatter plots of predicted versus observed probabilities were generated, accompanied by calibration curves. Decision curve and clinical impact analysis: Decision curve analysis (DCA) was conducted to calculate net benefit across different threshold probabilities, comparing the model against “treat all” and “treat none” strategies. The clinical action associated with a positive model result was defined as escalated management for suspected HM-NTG, including: (i) referral to a glaucoma specialist; (ii) addition of 10 − 2 visual field testing to assess central visual field vulnerability; and (iii) shortened follow-up intervals with repeat structural and microvascular imaging [17]. Clinical impact curves were generated using DecisionCurve software, with the y-axis standardized to “number per 1,000 patients” to facilitate clinical extrapolation and cross-study comparability, following standard methodology [18]. SVM model development and hyperparameter optimization Given the relatively small sample size, a support vector machine (SVM) with a radial basis function (RBF) kernel was employed for binary classification. To prevent data leakage and overfitting, hyperparameter tuning was performed using grid search coupled with 5-fold cross-validation [19, 20]. The final SVM model was then trained using these optimal hyperparameters, and the pooled predicted probabilities from all validation folds were used to construct the ROC curve and calculate the AUC. The SVM served primarily as a cross-algorithmic sensitivity check rather than a performance competitor to logistic regression. Model performance and interpretability DeLong test was used to compare the models and individual parameters. To quantify the incremental clinical value of model combinations, we calculated the continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI). The continuous NRI (category-free) was used to avoid dependence on predefined risk thresholds [21]. The 95% confidence intervals were derived using bootstrap resampling (500 iterations) via the nricens package in R. SHAP (SHapley Additive exPlanations) analysis was subsequently applied to the SVM to provide individualized, patient-level quantification of each feature’s directional contribution to prediction outcomes, enhancing clinical interpretability and model transparency [22]. All statistical analyses were performed using R software (version 4.4.2). The packages ‘tidyverse’, ‘glmnet’, ‘geepack’, ‘ggROC’, ‘caret’, ‘kernlab’, ‘pROC’, ‘nricens’, ‘PredictABEL’, ‘fastshap’, and ‘shapr’ were used for data analysis and visualization. Sensitivity analyses To evaluate whether the diagnostic performance of the identified predictors was confounded by baseline axial length, age, or sex, we performed sensitivity analyses by reconstructing the multivariable logistic regression model with additional covariates. In Model 2, AL was added as a continuous variable to the original three-predictor model. In Model 3, age and sex were further added to create a fully adjusted model. Model discrimination was compared using AUC, and the stability of the core imaging predictors was assessed by examining changes in ORs and their 95% CIs. Results Baseline characteristics A total of 87 patients (87 eyes) were enrolled, comprising 42 patients (42 eyes) in the HM group and 45 patients (45 eyes) in the HM-NTG group (Table 1). No significant differences were observed between the two groups in sex, age, BCVA, AL, SE, or IOP. Baseline axial length did not differ significantly between the HM and HM-NTG groups (26.92 ± 1.17 mm vs. 27.25 ± 1.60 mm, P = 0.26). Overall, 11 eyes (12.6%) had an AL ≥ 28 mm, comprising 4/42 (9.5%) in the HM group and 7/45 (15.6%) in the HM-NTG group. The between-group difference in the proportion of eyes with AL ≥ 28 mm was not statistically significant (P = 0.398). Regarding visual field parameters, the visual field index (VFI), mean deviation (MD), and pattern standard deviation (PSD) all demonstrated significant between-group differences (all P < 0.001). For OCT-derived parameters (Table 2), focal loss volume (FLV) and global loss volume (GLV) differed significantly between groups (both P < 0.001). In sectoral comparisons of peripapillary RNFL thickness, no significant differences were detected in the nasal superior and nasal inferior sectors (P = 0.16 and P = 0.983, respectively); all remaining sectors showed significant between group differences. For OCTA parameters, whole vessel density, inferior hemifield vessel density, and inferior vessel density exhibited significant intergroup differences (P = 0.005, P = 0.02, and P < 0.001, respectively). No significant differences were observed for other sectors. Detailed results are provided in Table 3. Scan quality assessment: The mean OCT quality index did not differ significantly between the HM and HM-NTG groups (7.2 ± 0.8 vs. 7.0 ± 0.9, P = 0.31), indicating comparable image acquisition quality between groups. Table 1Demographics and baseline ocular characteristics of highly myopic eyes with and without normal-tension glaucomaCharacteristicsALL (n = 87)HM (n = 42)HM-NTG (n = 45) P Sex0.929 Female46(52.9%)22(52.4%)24(53.3%) Male41(47.1%)20(47.6%)21(46.7%)Age, years42.58 ± 12.5740.42 ± 13.1045.10 ± 11.630.1BCVA, logMAR0.21 ± 0.410.18 ± 0.410.24 ± 0.430.13IOP, mmHg17.28 ± 3.2417.36 ± 3.2917.18 ± 3.240.96CCT, µm548.00 [530.00;566.50]548.00 [545.50;571.00]545.00 [522.00;562.50]0.116AL, mm27.06 ± 1.3726.92 ± 1.1727.25 ± 1.600.26AL≥28 mm11(12.6%)4(9.5%)7(15.6%)0.398SE, d7.50 [7.00;8.75]7.50 [7.00;9.34]7.50 [6.50;8.38]0.175VFI (%)89.94 ± 14.4096.50 ± 4.0182.32 ± 18.06< 0.001MD, dB5.20 ± 4.572.86 ± 2.107.91 ± 5.15< 0.001PSD, dB5.45 ± 4.803.19 ± 2.408.07 ± 5.54< 0.001Abbreviations: BCVA, best-corrected visual acuity; IOP, intraocular pressure; CCT, central corneal thickness; SE, spherical equivalent; Axial length, AL; VFI, visual field index; MD, mean deviation; PSD, pattern standard deviation; dB, decibel; HM, high myopia; NTG, normal-tension glaucoma. P-values < 0.05 were considered statistically significant. For consistency with the internal data processing and machine learning pipeline used in this study, spherical equivalent (SE) and mean deviation (MD) are reported as absolute values (diopters and decibels, respectively). Larger absolute values indicate more severe myopic refractive error (SE) or more extensive visual field loss (MD). In standard clinical convention, these values are expressed as negative numbers; the underlying clinical interpretation remains identical Table 1Demographics and baseline ocular characteristics of highly myopic eyes with and without normal-tension glaucomaCharacteristicsALL (n = 87)HM (n = 42)HM-NTG (n = 45) P Sex0.929 Female46(52.9%)22(52.4%)24(53.3%) Male41(47.1%)20(47.6%)21(46.7%)Age, years42.58 ± 12.5740.42 ± 13.1045.10 ± 11.630.1BCVA, logMAR0.21 ± 0.410.18 ± 0.410.24 ± 0.430.13IOP, mmHg17.28 ± 3.2417.36 ± 3.2917.18 ± 3.240.96CCT, µm548.00 [530.00;566.50]548.00 [545.50;571.00]545.00 [522.00;562.50]0.116AL, mm27.06 ± 1.3726.92 ± 1.1727.25 ± 1.600.26AL≥28 mm11(12.6%)4(9.5%)7(15.6%)0.398SE, d7.50 [7.00;8.75]7.50 [7.00;9.34]7.50 [6.50;8.38]0.175VFI (%)89.94 ± 14.4096.50 ± 4.0182.32 ± 18.06< 0.001MD, dB5.20 ± 4.572.86 ± 2.107.91 ± 5.15< 0.001PSD, dB5.45 ± 4.803.19 ± 2.408.07 ± 5.54< 0.001Abbreviations: BCVA, best-corrected visual acuity; IOP, intraocular pressure; CCT, central corneal thickness; SE, spherical equivalent; Axial length, AL; VFI, visual field index; MD, mean deviation; PSD, pattern standard deviation; dB, decibel; HM, high myopia; NTG, normal-tension glaucoma. P-values < 0.05 were considered statistically significant. For consistency with the internal data processing and machine learning pipeline used in this study, spherical equivalent (SE) and mean deviation (MD) are reported as absolute values (diopters and decibels, respectively). Larger absolute values indicate more severe myopic refractive error (SE) or more extensive visual field loss (MD). In standard clinical convention, these values are expressed as negative numbers; the underlying clinical interpretation remains identical Demographics and baseline ocular characteristics of highly myopic eyes with and without normal-tension glaucoma Abbreviations: BCVA, best-corrected visual acuity; IOP, intraocular pressure; CCT, central corneal thickness; SE, spherical equivalent; Axial length, AL; VFI, visual field index; MD, mean deviation; PSD, pattern standard deviation; dB, decibel; HM, high myopia; NTG, normal-tension glaucoma. P-values < 0.05 were considered statistically significant. For consistency with the internal data processing and machine learning pipeline used in this study, spherical equivalent (SE) and mean deviation (MD) are reported as absolute values (diopters and decibels, respectively). Larger absolute values indicate more severe myopic refractive error (SE) or more extensive visual field loss (MD). In standard clinical convention, these values are expressed as negative numbers; the underlying clinical interpretation remains identical Table 2Comparison of circumpapillary and macular OCT parameters between the HM and HM-NTG groupsCharacteristicsAllHMHM-NTG P FLV (%)5.57 ± 6.451.76 ± 3.2210.00 ± 6.47< 0.001GLV (%)12.17 ± 1.794.89 ± 5.0020.61 ± 11.82< 0.001RNFL, µmSH90.00 ± 17.6896.19 ± 14.2882.81 ± 18.71< 0.001IH80.79 ± 17.6689.81 ± 13.5170.32 ± 16.21< 0.001Average85.36 ± 15.9792.50 ± 12.3477.06 ± 15.86< 0.001Superior103.30 ± 24.90111.00 ± 24.5294.35 ± 22.540.008Inferior97.67 ± 30.12111.42 ± 28.1581.71 ± 24.12< 0.001NS68.79 ± 18.2870.44 ± 17.4366.87 ± 19.320.16NI63.34 ± 23.1961.28 ± 14.0365.74 ± 30.700.93TS80.60 ± 22.6589.42 ± 19.7370.35 ± 21.74< 0.001TI63.39 ± 19.9173.42 ± 15.6051.74 ± 18.09< 0.001GCC, µmSuperior89.04 ± 14.4895.25 ± 7.1881.84 ± 17.36< 0.001Inferior83.70 ± 13.8292.42 ± 6.7473.58 ± 13.06< 0.001 C/D V. Ratio0.61 ± 0.240.47 ± 0.230.77 ± 0.13< 0.001H. Ratio0.69 ± 0.210.57 ± 0.210.83 ± 0.10< 0.001Abbreviations: OCT, optical coherence tomography; HM, high myopia; NTG, normal-tension glaucoma; FLV, focal loss volume; GLV, global loss volume; RNFL, retinal nerve fiber layer; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior; GCC, ganglion cell complex; C/D, cup-to-disc; V. Ratio, vertical cup-to-disc ratio; H. Ratio, horizontal cup-to-disc ratio; Area Ratio, cup-to-disc area ratio. P-values < 0.05 were considered statistically significant Table 2Comparison of circumpapillary and macular OCT parameters between the HM and HM-NTG groupsCharacteristicsAllHMHM-NTG P FLV (%)5.57 ± 6.451.76 ± 3.2210.00 ± 6.47< 0.001GLV (%)12.17 ± 1.794.89 ± 5.0020.61 ± 11.82< 0.001RNFL, µmSH90.00 ± 17.6896.19 ± 14.2882.81 ± 18.71< 0.001IH80.79 ± 17.6689.81 ± 13.5170.32 ± 16.21< 0.001Average85.36 ± 15.9792.50 ± 12.3477.06 ± 15.86< 0.001Superior103.30 ± 24.90111.00 ± 24.5294.35 ± 22.540.008Inferior97.67 ± 30.12111.42 ± 28.1581.71 ± 24.12< 0.001NS68.79 ± 18.2870.44 ± 17.4366.87 ± 19.320.16NI63.34 ± 23.1961.28 ± 14.0365.74 ± 30.700.93TS80.60 ± 22.6589.42 ± 19.7370.35 ± 21.74< 0.001TI63.39 ± 19.9173.42 ± 15.6051.74 ± 18.09< 0.001GCC, µmSuperior89.04 ± 14.4895.25 ± 7.1881.84 ± 17.36< 0.001Inferior83.70 ± 13.8292.42 ± 6.7473.58 ± 13.06< 0.001 C/D V. Ratio0.61 ± 0.240.47 ± 0.230.77 ± 0.13< 0.001H. Ratio0.69 ± 0.210.57 ± 0.210.83 ± 0.10< 0.001Abbreviations: OCT, optical coherence tomography; HM, high myopia; NTG, normal-tension glaucoma; FLV, focal loss volume; GLV, global loss volume; RNFL, retinal nerve fiber layer; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior; GCC, ganglion cell complex; C/D, cup-to-disc; V. Ratio, vertical cup-to-disc ratio; H. Ratio, horizontal cup-to-disc ratio; Area Ratio, cup-to-disc area ratio. P-values < 0.05 were considered statistically significant Comparison of circumpapillary and macular OCT parameters between the HM and HM-NTG groups Abbreviations: OCT, optical coherence tomography; HM, high myopia; NTG, normal-tension glaucoma; FLV, focal loss volume; GLV, global loss volume; RNFL, retinal nerve fiber layer; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior; GCC, ganglion cell complex; C/D, cup-to-disc; V. Ratio, vertical cup-to-disc ratio; H. Ratio, horizontal cup-to-disc ratio; Area Ratio, cup-to-disc area ratio. P-values < 0.05 were considered statistically significant Table 3Comparison of peripapillary and intrapapillary vessel density between the HM and HM-NTG groupsCharacteristicsAllHMHM-NTG P VD (%) Whole43.67 ± 5.3545.13 ± 4.3841.98 ± 5.920.005Inside Disc49.85 ± 6.4651.59 ± 4.7947.82 ± 7.560.056Peripapillary44.39 ± 7.0845.71 ± 6.5442.86 ± 7.460.072SH45.43 ± 7.6646.23 ± 7.4244.52 ± 7.960.31IH43.07 ± 7.5145.14 ± 6.1740.67 ± 8.290.020Superior46.37 ± 9.7448.31 ± 8.4944.13 ± 10.710.078Inferior40.48 ± 11.8447.11 ± 7.3932.77 ± 11.42< 0.001NS41.70 ± 8.2741.42 ± 8.7542.03 ± 7.810.97NI42.07 ± 9.5541.78 ± 7.7542.42 ± 11.410.57TS48.30 ± 10.1448.64 ± 9.1947.90 ± 11.280.95TI43.76 ± 11.3845.22 ± 11.2842.06 ± 11.430.19Abbreviations: OCTA, optical coherence tomography angiography; HM, high myopia; NTG, normal-tension glaucoma; VD, vessel density; Inside Disc, intrapapillary vessel density; Peripapillary, peripapillary vessel density; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior. P-values < 0.05 were considered statistically significant Table 3Comparison of peripapillary and intrapapillary vessel density between the HM and HM-NTG groupsCharacteristicsAllHMHM-NTG P VD (%) Whole43.67 ± 5.3545.13 ± 4.3841.98 ± 5.920.005Inside Disc49.85 ± 6.4651.59 ± 4.7947.82 ± 7.560.056Peripapillary44.39 ± 7.0845.71 ± 6.5442.86 ± 7.460.072SH45.43 ± 7.6646.23 ± 7.4244.52 ± 7.960.31IH43.07 ± 7.5145.14 ± 6.1740.67 ± 8.290.020Superior46.37 ± 9.7448.31 ± 8.4944.13 ± 10.710.078Inferior40.48 ± 11.8447.11 ± 7.3932.77 ± 11.42< 0.001NS41.70 ± 8.2741.42 ± 8.7542.03 ± 7.810.97NI42.07 ± 9.5541.78 ± 7.7542.42 ± 11.410.57TS48.30 ± 10.1448.64 ± 9.1947.90 ± 11.280.95TI43.76 ± 11.3845.22 ± 11.2842.06 ± 11.430.19Abbreviations: OCTA, optical coherence tomography angiography; HM, high myopia; NTG, normal-tension glaucoma; VD, vessel density; Inside Disc, intrapapillary vessel density; Peripapillary, peripapillary vessel density; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior. P-values < 0.05 were considered statistically significant Comparison of peripapillary and intrapapillary vessel density between the HM and HM-NTG groups Abbreviations: OCTA, optical coherence tomography angiography; HM, high myopia; NTG, normal-tension glaucoma; VD, vessel density; Inside Disc, intrapapillary vessel density; Peripapillary, peripapillary vessel density; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior. P-values < 0.05 were considered statistically significant Diagnostic performance of individual parameters Tables 4 and 5 present the AUC, sensitivity, specificity, Youden index and accuracy of the OCT and OCTA parameters, respectively. Diagnostic performance was assessed for individual OCT variables. ROC analysis revealed that inferior GCC thickness exhibited the optimal discriminatory ability, with an AUC of 0.85 (95% CI: 0.75–0.95). Among peripapillary RNFL parameters, inferior hemifield thickness demonstrated comparable diagnostic accuracy, achieving an AUC of 0.85 (95% CI: 0.73–0.92). For OCTA parameters, ROC analysis indicated that inferior optic disc VD displayed the strongest diagnostic performance, with an AUC of 0.72 (95% CI: 0.64–0.79). Vessel densities in remaining sectors did not exhibit superior diagnostic performance. Table 4Diagnostic performance of OCT parameters in discriminating eyes with HM-NTG from HMCharacteristicsAUC (95% CI)SensitivitySpecificityYouden IndexAccuracy RNFL SH0.76(0.62–0.90)0.78(0.66–0.97)0.71(0.48–0.88)0.49(0.35–0.73)0.77(0.69–0.87)IH0.85(0.73–0.92)0.86(0.67–0.97)0.77(0.63-1.00)0.64(0.53–0.82)0.84(0.76–0.92)Average0.82(0.72–0.92)0.72(0.60-1.00)0.87(0.58–0.97)0.59(0.44–0.76)0.81(0.72–0.89)Superior0.69(0.55–0.84)0.72(0.34–0.96)0.61(0.37-1.00)0.34(0.23–0.57)0.70(0.60–0.78)Inferior0.81(0.68–0.91)0.83(0.55-1.00)0.65(0.43–0.97)0.48(0.37–0.70)0.77(0.68–0.86)NS0.60(0.48–0.73)0.42(0.38–0.88)0.81(0.33–0.88)0.22(0.12–0.48)0.64(0.54–0.75)NI0.51(0.39–0.65)0.77(0.05–0.96)0.31(0.16-1.00)0.08(0.04–0.32)0.57(0.47–0.66)TS0.77(0.64–0.88)0.89(0.67–0.97)0.65(0.46–0.84)0.53(0.32–0.73)0.78(0.66–0.87)TI0.84(0.74–0.93)0.92(0.79-1.00)0.71(0.57–0.88)0.63(0.48–0.79)0.83(0.76–0.90) GCC Superior0.83(0.74–0.92)0.69(0.60-1.00)0.87(0.52–0.97)0.57(0.46–0.76)0.80(0.72–0.87)Inferior0.85(0.75–0.95)0.92(0.88-1.00)0.77(0.56–0.89)0.69(0.55–0.87)0.86(0.78–0.94)Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; HM, high myopia; NTG, normal-tension glaucoma; OCT, optical coherence tomography; FLV, focal loss volume; GLV, global loss volume; RNFL, retinal nerve fiber layer; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior; GCC, ganglion cell complex Table 4Diagnostic performance of OCT parameters in discriminating eyes with HM-NTG from HMCharacteristicsAUC (95% CI)SensitivitySpecificityYouden IndexAccuracy RNFL SH0.76(0.62–0.90)0.78(0.66–0.97)0.71(0.48–0.88)0.49(0.35–0.73)0.77(0.69–0.87)IH0.85(0.73–0.92)0.86(0.67–0.97)0.77(0.63-1.00)0.64(0.53–0.82)0.84(0.76–0.92)Average0.82(0.72–0.92)0.72(0.60-1.00)0.87(0.58–0.97)0.59(0.44–0.76)0.81(0.72–0.89)Superior0.69(0.55–0.84)0.72(0.34–0.96)0.61(0.37-1.00)0.34(0.23–0.57)0.70(0.60–0.78)Inferior0.81(0.68–0.91)0.83(0.55-1.00)0.65(0.43–0.97)0.48(0.37–0.70)0.77(0.68–0.86)NS0.60(0.48–0.73)0.42(0.38–0.88)0.81(0.33–0.88)0.22(0.12–0.48)0.64(0.54–0.75)NI0.51(0.39–0.65)0.77(0.05–0.96)0.31(0.16-1.00)0.08(0.04–0.32)0.57(0.47–0.66)TS0.77(0.64–0.88)0.89(0.67–0.97)0.65(0.46–0.84)0.53(0.32–0.73)0.78(0.66–0.87)TI0.84(0.74–0.93)0.92(0.79-1.00)0.71(0.57–0.88)0.63(0.48–0.79)0.83(0.76–0.90) GCC Superior0.83(0.74–0.92)0.69(0.60-1.00)0.87(0.52–0.97)0.57(0.46–0.76)0.80(0.72–0.87)Inferior0.85(0.75–0.95)0.92(0.88-1.00)0.77(0.56–0.89)0.69(0.55–0.87)0.86(0.78–0.94)Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; HM, high myopia; NTG, normal-tension glaucoma; OCT, optical coherence tomography; FLV, focal loss volume; GLV, global loss volume; RNFL, retinal nerve fiber layer; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior; GCC, ganglion cell complex Diagnostic performance of OCT parameters in discriminating eyes with HM-NTG from HM Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; HM, high myopia; NTG, normal-tension glaucoma; OCT, optical coherence tomography; FLV, focal loss volume; GLV, global loss volume; RNFL, retinal nerve fiber layer; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior; GCC, ganglion cell complex Table 5Diagnostic performance of OCTA parameters in discriminating eyes with HM-NTG from HMCharacteristicsAUC (95% CI)SensitivitySpecificityYouden IndexAccuracy VD Whole0.70(0.57–0.82)0.67(0.57–0.93)0.81(0.49–0.91)0.47(0.31–0.65)0.74(0.64–0.82)ID0.64(0.48–0.77)0.50(0.35-1.00)0.74(0.23–0.94)0.24(0.11–0.49)0.65(0.54–0.78)Peripapillary0.63(0.49–0.76)0.67(0.31–0.97)0.61(0.34–0.95)0.28(0.17–0.53)0.67(0.57–0.78)SH0.57(0.46–0.71)0.89(0.17–0.97)0.29(0.21–0.99)0.18(0.09–0.43)0.62(0.48–0.73)IH0.67(0.56–0.80)0.67(0.41–0.97)0.65(0.29–0.95)0.31(0.20–0.61)0.69(0.60–0.81)Superior0.63(0.50–0.75)0.67(0.31–0.97)0.58(0.20–0.95)0.25(0.09–0.50)0.65(0.54–0.75)Inferior0.72(0.64–0.79)0.53(0.48–0.84)0.86(0.55–0.92)0.38(0.28–0.53)0.70(0.63–0.77)NS0.50(0.38–0.64)1.00(0.08-1.00)0.08(0.07-1.00)0.08(0.04–0.31)0.56(0.45–0.66)NI0.54(0.39–0.66)0.55(0.07–0.75)0.61(0.48-1.00)0.16(0.03–0.40)0.61(0.51–0.71)TS0.50(0.38–0.63)0.19(0.06–0.84)0.92(0.34-1.00)0.11(0.04–0.30)0.60(0.52–0.69)TI0.59(0.46–0.73)0.50(0.35–0.91)0.74(0.32–0.89)0.24(0.09–0.45)0.63(0.54–0.73)Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; HM, high myopia; NTG, normal-tension glaucoma; OCTA, optical coherence tomography angiography; VD, vessel density; ID, intrapapillary vessel density (inside disc); Peripapillary, peripapillary vessel density; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior Table 5Diagnostic performance of OCTA parameters in discriminating eyes with HM-NTG from HMCharacteristicsAUC (95% CI)SensitivitySpecificityYouden IndexAccuracy VD Whole0.70(0.57–0.82)0.67(0.57–0.93)0.81(0.49–0.91)0.47(0.31–0.65)0.74(0.64–0.82)ID0.64(0.48–0.77)0.50(0.35-1.00)0.74(0.23–0.94)0.24(0.11–0.49)0.65(0.54–0.78)Peripapillary0.63(0.49–0.76)0.67(0.31–0.97)0.61(0.34–0.95)0.28(0.17–0.53)0.67(0.57–0.78)SH0.57(0.46–0.71)0.89(0.17–0.97)0.29(0.21–0.99)0.18(0.09–0.43)0.62(0.48–0.73)IH0.67(0.56–0.80)0.67(0.41–0.97)0.65(0.29–0.95)0.31(0.20–0.61)0.69(0.60–0.81)Superior0.63(0.50–0.75)0.67(0.31–0.97)0.58(0.20–0.95)0.25(0.09–0.50)0.65(0.54–0.75)Inferior0.72(0.64–0.79)0.53(0.48–0.84)0.86(0.55–0.92)0.38(0.28–0.53)0.70(0.63–0.77)NS0.50(0.38–0.64)1.00(0.08-1.00)0.08(0.07-1.00)0.08(0.04–0.31)0.56(0.45–0.66)NI0.54(0.39–0.66)0.55(0.07–0.75)0.61(0.48-1.00)0.16(0.03–0.40)0.61(0.51–0.71)TS0.50(0.38–0.63)0.19(0.06–0.84)0.92(0.34-1.00)0.11(0.04–0.30)0.60(0.52–0.69)TI0.59(0.46–0.73)0.50(0.35–0.91)0.74(0.32–0.89)0.24(0.09–0.45)0.63(0.54–0.73)Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; HM, high myopia; NTG, normal-tension glaucoma; OCTA, optical coherence tomography angiography; VD, vessel density; ID, intrapapillary vessel density (inside disc); Peripapillary, peripapillary vessel density; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior Diagnostic performance of OCTA parameters in discriminating eyes with HM-NTG from HM Abbreviations: ROC, receiver operating characteristic; AUC, area under the curve; CI, confidence interval; HM, high myopia; NTG, normal-tension glaucoma; OCTA, optical coherence tomography angiography; VD, vessel density; ID, intrapapillary vessel density (inside disc); Peripapillary, peripapillary vessel density; SH, superior-hemifield; IH, inferior-hemifield; NS, nasal-superior; NI, nasal-inferior; TS, temporal-superior; TI, temporal-inferior Feature selection via LASSO & multivariable logistic regression All OCT and OCTA parameters were subjected to LASSO regression analysis. Ten-fold cross-validation was employed to determine the optimal regularization parameter (λ.1se = 0.033). As illustrated in Fig. 1, this λ value achieved an optimal balance between minimizing prediction error and maintaining model parsimony, yielding 9 variables with non-zero coefficients that were selected as candidate predictors for subsequent multivariate analysis. These 9 candidates were entered into multivariate logistic regression analysis (Table 6) using backward stepwise selection. This procedure identified 4 variables; after excluding those with P > 0.05, the final model retained 3 independent predictors: inferior GCC thickness (OR = 0.722, 95% CI: 0.56–0.94, P < 0.05), temporal-inferior RNFL thickness (OR = 0.82, 95% CI: 0.70–0.96, P < 0.05), and inferior VD (OR = 0.78, 95% CI: 0.62–0.97, P < 0.05). Fig. 1Feature selection using LASSO regression. (A) LASSO coefficient profiles of the candidate OCT and OCTA parameters. Each curve represents the coefficient of a specific variable as a function of the L1-norm. (B) Selection of the optimal tuning parameter (λ) in the LASSO model via 10-fold cross-validation. The vertical dashed lines indicate the minimum mean squared error (λ min) and the one-standard-error rule (λ 1se) Feature selection using LASSO regression. (A) LASSO coefficient profiles of the candidate OCT and OCTA parameters. Each curve represents the coefficient of a specific variable as a function of the L1-norm. (B) Selection of the optimal tuning parameter (λ) in the LASSO model via 10-fold cross-validation. The vertical dashed lines indicate the minimum mean squared error (λ min) and the one-standard-error rule (λ 1se) Table 6Multivariable logistic regression analysis for independent predictors of NTG in highly myopic eyesCharacteristicsβORCIUnit P Intercept41.874----TI RNFL-0.1980.820.702–0.958per 1-µm increase< 0.05Inferior GCC-0.3250.7220.555–0.940per 1-µm increase< 0.05Inferior VD-0.2590.7720.616–0.967per 1-percentage-point increase< 0.05Abbreviations: OR, odds ratio; CI, confidence interval; TI, temporal-inferior; RNFL, retinal nerve fiber layer; GCC, ganglion cell complex; VD, vessel density. P-values < 0.05 were considered statistically significant Table 6Multivariable logistic regression analysis for independent predictors of NTG in highly myopic eyesCharacteristicsβORCIUnit P Intercept41.874----TI RNFL-0.1980.820.702–0.958per 1-µm increase< 0.05Inferior GCC-0.3250.7220.555–0.940per 1-µm increase< 0.05Inferior VD-0.2590.7720.616–0.967per 1-percentage-point increase< 0.05Abbreviations: OR, odds ratio; CI, confidence interval; TI, temporal-inferior; RNFL, retinal nerve fiber layer; GCC, ganglion cell complex; VD, vessel density. P-values < 0.05 were considered statistically significant Multivariable logistic regression analysis for independent predictors of NTG in highly myopic eyes Abbreviations: OR, odds ratio; CI, confidence interval; TI, temporal-inferior; RNFL, retinal nerve fiber layer; GCC, ganglion cell complex; VD, vessel density. P-values < 0.05 were considered statistically significant Construction of the combined diagnostic model A diagnostic model incorporating inferior GCC thickness, temporal-inferior RNFL thickness, and inferior VD was developed based on multivariate logistic regression analysis. The model demonstrated favorable discriminative performance, achieving an AUC of 0.909 (95% CI: 0.833–0.984) (Fig. 2), which was significantly superior to any single-parameter model (DeLong test vs. inferior GCC: P < 0.05). At the optimal cutoff (determined by the maximum Youden index, probability = 0.605), the model exhibited a sensitivity of 74.2%, specificity of 97.2%, and overall accuracy of 85.7%. The final 3-variable model was compared against the best single predictor (inferior GCC thickness, AUC = 0.85). The DeLong test confirmed that the combined model achieved significantly superior discrimination (P < 0.05). The NRI was 0.679 (95% CI: 0.228–1.131, P = 0.003) and the IDI was 0.067 (95% CI: 0.007–0.128, P = 0.029), indicating substantial improvement in risk classification and discrimination slope. Fig. 2ROC curve of the predictive model built using multivariable logistic regression ROC curve of the predictive model built using multivariable logistic regression Final logistic regression model specification: The three selected predictors were entered into the final multivariable logistic regression model as raw (unstandardized) continuous variables to preserve clinical interpretability. The unit for RNFL and GCC thickness was micrometers (µm), and the unit for vessel density was percentage points (%). The full prediction equation is:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned}\mathrm{logit}\left(\mathrm{P}\right)&=41.874-0.198\times\left(\mathrm{TI}\:\mathrm{RNFL}\:\upmu\mathrm{m}\right)\cr&\quad-0.325\times\left(\mathrm{Inferior}\:\mathrm{GCC}\:\upmu\mathrm{m}\right)\cr&\quad-0.259\times\left(\mathrm{Inferior}\:\mathrm{VD}\:\%\right)\end{aligned}$$\end{document} Model validation Internal validation was performed using bootstrap resampling (500 iterations) to assess the stability of model discrimination (Fig. 3A), yielding an optimism-corrected AUC of 0.861 (95% CI: 0.773–0.949). Overall model calibration was evaluated with the Hosmer-Lemeshow goodness-of-fit test (χ² = 6.64, P = 0.575). The Brier score was 0.109, indicating satisfactory accuracy in individual probability prediction. Calibration plots (Fig. 3B) demonstrated close agreement with the ideal diagonal line in both low-risk (predicted probability < 0.3) and high-risk (predicted probability > 0.8) regions. Decision curve analysis (Fig. 3C) revealed that net benefit from applying this diagnostic model exceeded that of both “treat all” and “treat none” strategies across clinically relevant threshold probabilities ranging from 10% to 80%. Clinical impact curves (Fig. 3D) displayed the estimated number of patients classified as high-risk and the corresponding number of true positives per 1,000 hypothetical patients at various threshold levels, following standard DecisionCurve software convention. Fig. 3Model Validation (A) ROC curve for bootstrap internal validation. The red solid line denotes the original ROC curve, and the gray solid line represents the ROC curve from 500 bootstrap resamples. (B) Calibration curve. The X-axis indicates predicted probability, and the Y-axis indicates observed probability. The black solid line represents the model calibration line, and the gray dashed line represents the ideal calibration line (45° diagonal). (C) Decision curve analysis. The X-axis indicates threshold probability, and the Y-axis indicates net benefit. The red line (Treat All) represents the “treat all” strategy, the green line (Treat None) represents the “treat none” strategy, and the blue line represents the net benefit curve of the present model. (D) Clinical impact curve. The red solid line indicates the total number of patients classified as high-risk by the model, and the blue dashed line indicates the number of actual HM-NTG patients among them (true positives) Model Validation (A) ROC curve for bootstrap internal validation. The red solid line denotes the original ROC curve, and the gray solid line represents the ROC curve from 500 bootstrap resamples. (B) Calibration curve. The X-axis indicates predicted probability, and the Y-axis indicates observed probability. The black solid line represents the model calibration line, and the gray dashed line represents the ideal calibration line (45° diagonal). (C) Decision curve analysis. The X-axis indicates threshold probability, and the Y-axis indicates net benefit. The red line (Treat All) represents the “treat all” strategy, the green line (Treat None) represents the “treat none” strategy, and the blue line represents the net benefit curve of the present model. (D) Clinical impact curve. The red solid line indicates the total number of patients classified as high-risk by the model, and the blue dashed line indicates the number of actual HM-NTG patients among them (true positives) Comparative value of OCTA beyond structural OCT To evaluate the incremental contribution of OCTA beyond structural OCT, we constructed an OCT-only diagnostic model incorporating all OCT parameters that entered the LASSO screening. This model achieved an AUC of 0.888 (95% CI: 0.803–0.973) (Fig. 4). The subsequent addition of OCTA-derived inferior VD to create the final OCT + OCTA combined model yielded an AUC of 0.909 (95% CI: 0.833–0.984). The incremental improvement in AUC (AUC = 0.021) did not reach statistical significance on the DeLong test (P = 0.164). However, reclassification metrics indicated that the inclusion of OCTA provided meaningful clinical utility. The continuous NRI was 0.504 (95% CI: 0.039–0.968, P = 0.034), and the IDI was 0.057 (95% CI: 0.001–0.114, P = 0.046). These findings indicate that while OCTA did not substantially expand the overall discrimination envelope (AUC), it improved the model’s ability to correctly shift patients across risk categories and sharpened the separation of predicted probabilities between groups. Fig. 4Comparison of diagnostic performance between the OCT-only model and the OCT + OCTA combined model Comparison of diagnostic performance between the OCT-only model and the OCT + OCTA combined model Machine learning validation and SHAP interpretability The SVM model achieved an AUC of 0.908 (95% CI: 0.833–0.982) (Fig. 5), nearly identical to the logistic regression model (AUC = 0.909). Cross-validation stability analysis demonstrated consistent discriminative performance across all 5 folds, with individual fold AUCs ranging from 0.887 to 0.916 (mean 0.905, SD 0.014) (Fig. 6). Detailed hyperparameter search ranges and final selected values are provided in Supplementary Table S1. This convergence across algorithmically distinct approaches suggests that the discriminatory signal of the three selected predictors is robust and not dependent on a specific modeling framework. SHAP analysis was employed to elucidate individualized feature contributions to patient-specific predictions (Fig. 7). Each dot in the beeswarm plot represents one patient’s SHAP value for a given feature, revealing how that feature influenced the prediction for that individual. Lower values of inferior GCC thickness, temporal-inferior RNFL thickness, and inferior VD were consistently associated with increased predicted probability of HM-NTG. Inferior GCC thickness displayed the widest SHAP value distribution range, indicating that this parameter exerted the largest individualized influence on model decisions. This patient-level granularity provides clinically actionable insight beyond population-average odds ratios, enabling physicians to understand the specific drivers of risk for individual patients. Fig. 5ROC curve of the predictive model built using SVM model ROC curve of the predictive model built using SVM model Fig. 6Five-fold cross-validation ROC curves of the support vector machine (SVM) diagnostic model Five-fold cross-validation ROC curves of the support vector machine (SVM) diagnostic model Fig. 7SHAP Beeswarm Plot of the SVM Model; The SHAP beeswarm plot illustrates the individualized contribution of each feature to SVM-based HM-NTG diagnosis. Each dot represents the SHAP value for an individual patient. The X-axis denotes SHAP value magnitude (positive values indicate increased probability of HM-NTG prediction), and the Y-axis displays the three predictive features. The color scale on the right indicates actual feature values: yellow represents higher values, and purple represents lower values SHAP Beeswarm Plot of the SVM Model; The SHAP beeswarm plot illustrates the individualized contribution of each feature to SVM-based HM-NTG diagnosis. Each dot represents the SHAP value for an individual patient. The X-axis denotes SHAP value magnitude (positive values indicate increased probability of HM-NTG prediction), and the Y-axis displays the three predictive features. The color scale on the right indicates actual feature values: yellow represents higher values, and purple represents lower values Sensitivity analyses The addition of AL as a continuous covariate did not improve model discrimination (AUC = 0.910 vs. 0.909 for the original model), and AL was not independently associated with HM-NTG status (OR = 1.08, 95% CI: 0.82–1.42). In the fully adjusted model incorporating AL, age, and sex, the AUC was 0.918, representing a negligible incremental gain. Neither age (OR = 1.02, 95% CI: 0.98–1.06) nor sex (OR = 0.78, 95% CI: 0.41–1.48) was a significant independent predictor. The three core imaging predictors (inferior GCC, temporal-inferior RNFL, and inferior VD) retained their statistical significance, effect direction, and comparable magnitude across all sensitivity models (Supplementary Table S2). Discussion Axial elongation secondary to high myopia induces diffuse thinning of the entire RNFL, which may confound glaucoma diagnosis to some extent. This study developed and validated a diagnostic model integrating multimodal imaging parameters (including OCT and OCTA metrics) to identify NTG among patients with HM. We demonstrated that inferior GCC thickness, temporal-inferior RNFL thickness, and inferior VD constitute the optimal combination of independent predictors for distinguishing between the two groups. The logistic regression model constructed from these parameters exhibited excellent discriminative performance, significantly outperforming any single metric. This is consistent with prior glaucoma research emphasizing the superiority of multiparameter approaches over isolated indices. Furthermore, internal validation confirmed the model’s robust stability and clinical utility. Incorporation of the SVM algorithm maintained favorable diagnostic performance, with SHAP analysis identifying inferior GCC thickness as the predominant driver of model decisions. The SVM model achieved an AUC nearly identical to that of the logistic regression model (0.908 vs. 0.909). We do not interpret this as a failure of machine learning to improve performance, but rather as cross-algorithmic validation that the three selected predictors carry genuine discriminatory signal. The comparable performance suggests that any nonlinear or threshold effects present in this dataset are modest and adequately captured by the linear logistic model, though this cannot be assumed without empirical testing of a nonlinear learner. The primary contribution of the machine learning component lies not in superior AUC, but in the application of SHAP for individualized prediction explanation. While logistic regression provides population-average odds ratios, SHAP values reveal how each feature contributes to the prediction for every individual patient. In clinical practice, this granularity is essential: a physician can examine a specific patient’s SHAP profile and understand that their high-risk classification was driven primarily by a markedly reduced inferior GCC thickness, rather than by a generic coefficient. This transparency addresses growing concerns about the “black box” nature of machine learning in medicine and supports the regulatory and ethical imperative for explainable AI in clinical decision support systems. The practical utility of this diagnostic model lies in its ability to stratify highly myopic patients into distinct management pathways. For patients with a predicted probability exceeding the optimal threshold (0.605), the model recommends classification as HM-NTG suspect and triggers a tiered management protocol: referral to a glaucoma specialist, addition of 10 − 2 visual field testing to evaluate central visual field vulnerability, and shortened follow-up intervals with repeat OCT/OCTA monitoring. This approach is supported by Nakajima et al. [17], who demonstrated that central visual field progression is particularly relevant in myopic NTG and emphasized the importance of meticulous central visual field assessment in this population. For patients below the threshold, routine standard care with standard follow-up intervals is recommended. The DCA results demonstrate that this stratified approach yields superior net benefit compared with either universal referral of all highly myopic patients or passive monitoring without risk stratification. We wish to explicitly acknowledge a key limitation regarding the specificity of our findings. The three parameters identified as optimal predictors, namely inferior GCC thickness, temporal-inferior RNFL thickness, and inferior VD, are well-established markers of classical glaucomatous optic neuropathy and are not pathognomonic for high-myopia-associated NTG. Indeed, inferior and temporal-inferior RNFL thinning represent canonical patterns of glaucomatous axonal loss across diverse populations, and GCC thinning is a recognized early biomarker of ganglion cell injury in glaucoma regardless of refractive status. Similarly, reduced peripapillary VD has been consistently observed in non-myopic glaucoma cohorts. Therefore, we do not claim that these features represent high-myopia-specific glaucomatous changes. Rather, the contribution of this study lies in demonstrating that, within the challenging context of highly myopic eyes in which axial elongation, optic disc tilt, and peripapillary atrophy confound the interpretation of isolated structural parameters, the multimodal integration of these classical glaucomatous markers via machine learning can improve diagnostic discrimination between clinically diagnosed HM-NTG and non-glaucomatous HM. Our findings highlighted the inferior GCC and inferior-hemifield RNFL as robust diagnostic markers within the HM population. As noted above, these parameters mirror classical optic nerve damage patterns observed in glaucoma across diverse populations; their value in this study lies not in representing HM-specific pathology, but in their ability to withstand the structural complexities of high myopia when integrated into a multimodal machine learning framework. In individuals with HM, large and tilted optic discs, the myopic temporal crescent, and occasionally severe peripapillary staphyloma can alter the thickness and distribution of the fibers surrounding the optic nerve head, thereby affecting the ability of isolated RNFL measurements to detect glaucoma in this population. The marked reduction in macular thickness we observed further validates this, as the loss of the GCC remains a highly sensitive signal for glaucoma, regardless of the underlying myopic thinning [23]. With advances in modern OCT technology, automatic segmentation of the inner retinal layers became feasible, the macular GCC comprises the three innermost retinal layers that are potentially involved in glaucomatous damage, namely the RNFL, ganglion cell layer (GCL), and inner plexiform layer (IPL). Consequently, GCC measurements held substantial value in the prediction and diagnosis of glaucoma [24]. Kim et al. [25] suggested that GCC thickness might be a more suitable biomarker for the early diagnosis of glaucoma compared to RNFL thickness. Scuderi et al. [24] found that the most accurate GCC parameters were represented by the mean and inferior GCC thickness, a finding that aligns closely with our study. In a study by Soel et al. [26], it was found that GCC analysis may be the most effective parameter for detecting preperimetric glaucoma in myopic patients. In addition, inferior-hemifield RNFL and nasal–inferior RNFL also demonstrated important diagnostic value. This may be attributed to the fact that disc torsion and tilt in highly myopic eyes can induce localized scleral tension and mechanical stress on axons within the lamina cribrosa region [27]. Inferior disc tilting predominantly affects the structurally vulnerable inferior region, leading to preferential damage in this area. High myopia is characterized by axial elongation, which induces diffuse thinning of the entire RNFL and distorts optic disc morphology, thereby confounding the clinical differentiation of glaucomatous from non-glaucomatous optic neuropathy. However, we wish to clarify that this study was cross-sectional and did not measure longitudinal changes in axial length. Baseline AL did not differ significantly between the HM and HM-NTG groups (26.92 ± 1.17 mm vs. 27.25 ± 1.60 mm, P = 0.26), and sensitivity analyses confirmed that AL was not an independent predictor of HM-NTG status when added to the diagnostic model (OR = 1.08, 95% CI: 0.82–1.42), with negligible change in AUC (0.910 vs. 0.909). Furthermore, adjustment for age and sex did not materially alter the model’s performance or the stability of the three core predictors (Supplementary Table S2). Therefore, the observed OCT/OCTA differences between groups cannot be attributed to differential axial elongation, age, or sex. Rather, axial elongation should be understood as the shared anatomical background of both groups that complicates the interpretation of isolated structural parameters, not as the mechanistic cause of the discriminating features identified in this study. The reliability of conventional RNFL measurements is often compromised in highly myopic eyes [28]. Axial elongation triggers mechanical thinning of the retina and significant optic disc tilting, both of which introduce substantial “noise” into OCT-derived parameters [29]. In some HM populations, because current imaging systems cannot precisely distinguish physiological myopic stretching from true pathological thinning, myopia-related changes may lead to false-positive findings [30]. Therefore, relying solely on structural alterations may be insufficient for establishing a definitive diagnosis [11]. In this study, incorporating the microvessel density parameters calculated by OCTA into the diagnostic model could significantly improve the diagnostic performance. These findings suggest that OCTA provides hemodynamic evidence and may serve as a useful tool to improve the diagnosis of NTG in highly myopic eyes. The etiology of NTG remains a subject of intense controversy, yet there is an increasing shift toward the “vascular hypothesis.” [31]. Several previous studies have reported changes or dysfunction in the vascular structure of NTG patients [31]. Xu et al. [32] compared retinal vascular changes between patients with High tension glaucoma (HTG) and NTG using OCTA, revealing a significant reduction in retinal vascular perfusion density in NTG eyes compared to HTG eyes. OCTA provides detailed insights into the microvascular structure. Liu et al. [33] were the first to report that the peripapillary vascular density in glaucoma is lower than in normal eyes. Scripsema et al. [34] reported that vascular dropout and nerve fiber loss often occur together spatially, but they do not always progress at the same rate. This delay between the two makes VD an important marker. It helps us verify whether the structural thinning we see is a true pathological shift or just a benign consequence of myopic anatomy. In short, VD helps us see the disease in motion, rather than just looking at a static snapshot of atrophy. While structural atrophy remains the undisputed “gold standard” for glaucoma, our findings make a strong case for the inclusion of vascular data. The positive NRI tells us that vascular metrics act as a corrective filter, successfully reclassifying many individuals who would have been misdiagnosed if we had relied on structural thinning alone. Meanwhile, the significant IDI reflects a much sharper separation in predicted probabilities between HM-NTG and healthy HM eyes, effectively widening the “diagnostic gap” between the two groups. Recent longitudinal studies have highlighted that ongoing axial elongation in adulthood is not uniform among highly myopic eyes. Kong et al. [35] reported that baseline AL ≥ 28 mm, age < 40 years, and female sex are risk factors for continued axial elongation in young adults with nonpathologic high myopia. Inooka et al. [36] identified an AL threshold around 28.6 mm as a meaningful cutoff for prioritizing follow-up in adults, noting that elongation risk is concentrated in eyes with longer AL, greater interocular asymmetry, and myopic maculopathy. Similarly, Nishimura et al. [37], in a large population-based cohort, reported that adult axial elongation was associated with younger age, female sex, and pre-existing high myopia. These findings suggest that an AL threshold around 28 mm may carry clinically meaningful risk stratification value. In the present study, however, only a limited proportion of eyes (12.6%, n = 11) had AL ≥ 28 mm, precluding meaningful stratified analyses. Moreover, because our study was cross-sectional, we cannot evaluate whether ongoing axial elongation contributed to the observed microvascular or structural differences. Future longitudinal studies incorporating AL monitoring and systematic grading of myopic maculopathy, posterior staphyloma, and disc tilt would be necessary to elucidate the dynamic interplay between axial elongation and glaucomatous damage in this population. It is important to contextualize the contribution of OCTA microvascular parameters within the broader diagnostic framework. In this study, individual OCTA parameters demonstrated only modest discriminatory performance, with the best-performing parameter (inferior VD) achieving an AUC of 0.72, which was substantially lower than the top-performing OCT structural parameters (inferior GCC and inferior hemifield RNFL, both AUC ≈ 0.85). This hierarchy aligns with the established primacy of structural damage biomarkers in glaucoma diagnosis. When we formally compared the OCT-only model (AUC = 0.888) against the OCT + OCTA combined model (AUC = 0.909), the AUC difference of 0.021 was not statistically significant (DeLong test, P = 0.164). Therefore, we do not claim that OCTA “significantly improves” overall diagnostic discrimination in the traditional AUC-centric sense. Rather, the value of OCTA appears to lie in its complementary hemodynamic information, which enhanced risk reclassification as evidenced by significant NRI and IDI. Specifically, the addition of inferior VD helped reclassify patients who had intermediate structural profiles—those in whom isolated RNFL or GCC measurements fell into a diagnostic “gray zone”—by providing an independent microvascular dimension that is less susceptible to myopia-related anatomical distortion. This interpretation is consistent with the vascular hypothesis of NTG pathogenesis: while structural atrophy remains the gold standard for glaucoma diagnosis, microvascular dysfunction may represent an earlier or parallel pathophysiological process that is not fully captured by RNFL/GCC thinning alone [11]. In highly myopic eyes, where axial elongation and disc tilt confound structural interpretation, OCTA may serve as a corrective filter for false-positive or equivocal structural findings, rather than as a standalone diagnostic modality. Given the relatively small sample size in this study, the SVM method, which is suitable for small datasets, was chosen [38]. One advantage of using SVM to build diagnostic models is its ability to reduce the risk of overfitting in small sample cohorts [39]. By identifying the optimal hyper-plane that maximizes the margin between diagnostic classes, SVM maintains superior generalization performance and robustness even when the number of input features is large relative to the number of subjects [40]. The interpretation of inferior OCTA vessel density in highly myopic eyes warrants careful consideration. Recent work by Takahashi et al. [41] demonstrated that inferior microvasculature dropout in open-angle glaucoma is associated with peripapillary atrophy-β area, suggesting a mechanical stress component, whereas superior dropout correlates more closely with reduced ocular blood flow. This implies that inferior VD reflects structural alterations in addition to glaucomatous vascular compromise. Furthermore, Xiong et al. [42] showed that highly myopic eyes exhibit complex lamina cribrosa abnormalities detectable by retro-mode scanning laser ophthalmoscopy, highlighting the intricate optic nerve head architecture that may confound isolated structural or vascular measurements. While our machine learning model integrates multiple parameters to mitigate the impact of any single confounding anatomical factor, we did not systematically grade disc tilt, PPA extent, posterior staphyloma, or myopic maculopathy. Future studies incorporating these anatomical variables. It is important to reiterate that the present study was conducted exclusively within a highly myopic population and did not include non-highly myopic control groups. Consequently, our findings cannot address whether the identified parameter combination is specific to HM-NTG or merely reflects general glaucomatous damage that happens to occur in highly myopic eyes. The clinical utility demonstrated here is context-specific: the model assists ophthalmologists in deciding whether a highly myopic eye with suspicious optic disc appearance and normal IOP has developed NTG or remains non-glaucomatous. Future studies incorporating non-myopic NTG and healthy control groups would be necessary to determine whether HM confers any unique modifications to the glaucomatous damage pattern. This study has several limitations. (1) All samples came from a single clinical center with a similar ethnic and geographic population. This may limit the model’s generalizability. (2) This study only showed the correlation between indicator characteristics and disease status at one point in time. Longitudinal studies in the future are needed to confirm the model’s predictive value. (3) To show the clinical value of the model, external validation with multicenter datasets is needed. (4) Due to the small sample size, we used only two machine learning methods to build the model. Other methods with potentially better performance could not be evaluated. (5) The study did not include longitudinal axial length measurements or systematic grading of myopia-related anatomical features (e.g., posterior staphyloma, disc tilt angle, peripapillary atrophy extent, myopic maculopathy). Consequently, we cannot determine whether the observed OCT/OCTA differences were causally related to axial elongation or merely coexisted within the shared anatomical substrate of high myopia. (6) Only a small subset of eyes had AL ≥ 28 mm, limiting our ability to evaluate the potential threshold effect suggested by recent longitudinal studies. (7) Goldmann applanation tonometry was not used in this study. Although non-contact tonometry and Corvis ST biomechanically corrected IOP were rigorously applied, the absence of Goldmann applanation tonometry as a reference standard for IOP measurement may be perceived as a methodological limitation. Future studies should ideally incorporate Goldmann tonometry to further validate IOP measurements in highly myopic eyes. (8) The significant NRI and IDI values were derived from the same dataset used for model development, which may inflate reclassification metrics in small cohorts. Although bootstrap confidence intervals were calculated, the precision of these estimates is limited by n = 87, and independent external validation is required to confirm the incremental clinical utility of OCTA beyond structural OCT. Conclusion Within highly myopic eyes, the combination of inferior GCC thickness, temporal-inferior RNFL thickness, and inferior VD may help distinguish clinically diagnosed HM-NTG from non-glaucomatous HM. The addition of OCTA-derived inferior vessel density did not significantly increase the AUC but improved patient risk reclassification, suggesting that OCTA serves as a complementary adjunct that may assist in resolving diagnostic uncertainty in eyes with equivocal structural findings. This multimodal machine learning approach offers a promising tool for diagnostic risk stratification in highly myopic patients, though future longitudinal studies are needed to validate the prognostic value of microvascular changes beyond structural assessment. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1