Combining OCT and OCT-Angiography Longitudinal Data for the Detection of Visual Field Progression in Glaucoma

Abstract

PURPOSE:

To use longitudinal optical coherence tomography (OCT) and OCT-Angiography (OCTA) data to detect glaucomatous visual field (VF) progression with a supervised machine learning approach.

DESIGN:

Prospective cohort study.

METHODS:

110 eyes of glaucoma suspect (33.6%) and glaucoma (66.4%) patients with a minimum of five 24–2 VF tests and three optic nerve head and macula images over an average follow-up duration of 4.1 years were included. VF progression was defined using a composite measure including either a “likely progression event” on Guided Progression Analysis, a statistically significant negative slope of VF mean deviation or VF index, or a positive pointwise linear regression event. Feature-based gradient boosting classifiers were developed using different subsets of baseline and longitudinal OCT and OCTA summary parameters. Area under the receiver operating characteristic curve (AUROC) was used to compare the classification performance of different models.

RESULTS:

VF progression was detected in 28 eyes (25.5%). The model with combined baseline and longitudinal OCT and OCTA parameters at the global and hemifield levels had the best classification accuracy to detect VF progression (AUROC=0.89). Models including combined OCT and OCTA parameters had higher classification accuracy compared to those with individual subsets of OCT, or OCTA features alone. Including hemifield measurements significantly improved the models’ classification accuracy compared to using global measurements alone. Including longitudinal rates of change of OCT and OCTA parameters (AUROCs=0.80–0.89) considerably increased the classification accuracy of the models with baseline measurements alone (AUROCs=0.60–0.63).

CONCLUSIONS:

Longitudinal OCTA measurements complement OCT-derived structural metrics for the evaluation of functional VF loss in glaucoma patients.

Table of Contents Statement:

This prospective study demonstrated the utility of longitudinal OCT-Angiography measurements to enhance the performance of OCT-based structural parameters for the evaluation of glaucomatous visual field progression.

INTRODUCTION

Glaucoma is characterized by progressive damage to the retinal ganglion cells along with a commensurate visual field (VF) loss.^1,2 Because of the irreversible nature of glaucomatous damage, evaluation of disease progression is essential to prevent glaucoma-related vision loss. Various imaging modalities have been utilized for this task, including optic disc photography, scanning laser ophthalmoscopy, Optical Coherence Tomography (OCT), and OCT Angiography (OCTA) technology.^3–18 The advent of OCTA has provided clinicians with an unprecedented opportunity to assess retinal microcirculation in a non-invasive 3-D high-resolution manner.^19–21 Imaging the retinal microcirculation at the level of the optic nerve head (ONH), peripapillary retina, and the macula has enabled clinicians to detect and monitor microvascular damage from glaucoma with improved accuracy.^{8–18,22–30}

Given the newness of OCTA, the majority of currently available reports on the utility of this instrument in glaucoma are based on cross-sectional studies.¹³ Current evidence suggests good diagnostic accuracy of OCTA measurements in terms of discriminating between healthy individuals and glaucoma suspect and glaucoma patients.^9,13,17 Moreover, other investigations have shown that lower OCTA-measured vessel density is associated with faster rates of decline in retinal nerve fiber layer (RNFL) thickness,¹⁰ and higher probability of prior VF progression.¹⁸ Recently, two prospective studies have found associations between rates of change in ONH³¹ and macular³² OCTA measurements and VF progression in glaucoma patients. An essential, but less investigated, aspect of OCTA utility for glaucoma management is to evaluate whether including OCTA vessel density measurements enhances our ability to detect glaucoma or its progression over the standard of care OCT measurements alone. Shoji and colleagues demonstrated that serial macular OCTA measurements were able to detect microvascular loss in glaucoma eyes without apparent evidence of GCC thickness alteration over an average follow-up duration of less than 14 months.³⁰ Moreover, few studies have compared the performance of models using combinations of OCT and OCTA metrics to that of individual OCT or OCTA-based models^33–35 with promising findings for the evaluation of structure-function relationship.³⁵

The application of artificial intelligence techniques and, more specifically, machine learning methods enables the combination of different subsets of input parameters for developing robust predictive models and also gaining more insight into the role of different biomarkers of disease. These methods have gained increasing interest for the analysis of structural and functional measurements in ophthalmology over the last two decades. Notably, an automated machine learning-based photographic system for the diagnosis of diabetic retinopathy has recently gained Food and Drug Administration approval (IDx Technologies IDx-DR).³⁶ Previous studies have developed various machine learning models using structural and functional measurements to enhance the performance of classification tasks in glaucoma including diagnosis and monitoring of disease progression.^{33,34,37–41} Two recent studies investigated the use of machine learning models combining OCT and OCTA measurements to enhance the diagnostic accuracy of glaucoma.^33,34 Specifically, gradient boosting classifiers (GBCs) have previously demonstrated high diagnostic accuracy for discriminating between healthy subjects and glaucoma patients based on different combinations of OCT and OCTA metrics.³³ Accordingly, the application of machine learning strategies has the potential to determine whether OCTA measurements can improve the currently established OCT assessments for monitoring structural damage and change in glaucoma.

To date, longitudinal studies on the utility of OCTA for monitoring glaucoma progression are limited,²⁹ and have not addressed whether the combination of OCT and OCTA measurements improves the performance of predictive models compared to sole reliance on the currently established OCT indices for monitoring glaucomatous VF progression. Given the previously reported high classification accuracy of GBCs combining OCT and OCTA data for the discrimination of healthy and glaucoma patients,³³ the principal aim of this study was to assess the performance of GBCs using different subsets of individual and combined OCT and OCTA features for the evaluation of VF progression in a longitudinal cohort of glaucoma suspect and glaucoma patients.

PATIENTS AND METHODS

This prospective, observational study included a cohort of glaucoma suspect and primary open-angle glaucoma patients from the Diagnostic Innovations in Glaucoma Study (DIGS) (ClinicalTrials.gov identifier: NCT00221897). The DIGS is an ongoing prospective, longitudinal study at the Hamilton Glaucoma Center, University of California, San Diego, designed to evaluate the anatomical structure in glaucoma. All the methods adhered to the tenets of the Declaration of Helsinki and the Health Insurance Portability and Accountability Act. The institutional review boards at the University of California, San Diego, approved the methods. The study protocols were explained to all participants and written informed consent was obtained.

DIGS protocol and eligibility criteria have been described in detail previously.⁴² In brief, all participants underwent comprehensive ophthalmological examination, including a review of medical history, assessment of best-corrected visual acuity (BCVA), slit-lamp biomicroscopy, Goldmann applanation tonometry, gonioscopy, ultrasound pachymetry, dilated fundus examination, stereophotography of the optic disc, and VF testing (Humphrey Field Analyzer; Carl Zeiss Meditec, Jena, Germany). All participants also completed Spectral-domain OCT (SD-OCT) [Avanti; Optovue, Inc.], and OCTA (Angiovue; Optovue, Inc., Fremont, CA, USA) imaging of the ONH and the macula.

Systemic measurements included systolic and diastolic blood pressure and pulse rate measured at the height of the heart with an automatic blood pressure instrument (model BP791IT; Omron Healthcare, Inc., Lake Forest, IL). Mean arterial pressure was calculated as 1/3 systolic blood pressure + 2/3 diastolic blood pressure.

Overall inclusion criteria at study entry were an age of more than 18 years, open angles on gonioscopy, and BCVA ≥ 20/40. Participants included in this study were required to have at least five 24–2 VF tests over a minimum of 2.5 years of follow-up. In addition, each participant was required to have at least three good quality (as described below) OCTA images of both the ONH and the macula within the follow-up duration. Participants with a history of intraocular surgery (except for uncomplicated cataract surgery or glaucoma surgery), an axial length of 27 mm or more, coexisting retinal pathologies, non-glaucomatous optic neuropathy, uveitis, or ocular trauma were excluded from the study. Participants were also excluded if they had a diagnosis of Parkinson’s disease, Alzheimer’s disease, dementia, or a history of stroke. Participants with systemic hypertension and diabetes mellitus were included unless they were diagnosed with diabetic or hypertensive retinopathy.

Included eyes were classified as glaucomatous if they had repeatable (≥2 consecutive) abnormal VF test results or evidence of glaucomatous optic neuropathy defined as excavation, the presence of focal thinning, notching of the neuroretinal rim, or localized or diffuse atrophy of the RNFL based on masked grading of optic disc photographs by 2 graders or clinical examination by a glaucoma specialist. An abnormal VF test was defined as a pattern standard deviation (PSD) outside of the 95% normal confidence limits or a Glaucoma Hemifield Test result outside normal limits. Glaucoma suspects were defined as those having elevated IOP (≥22 mm Hg) or suspicious-appearing optic discs without the presence of repeatable glaucomatous VF damage.

VISUAL FIELD TESTING

Visual field tests were performed using Swedish Interactive Threshold Algorithm standard 24–2 threshold test. All VFs were evaluated by the University of California, San Diego Visual Field Assessment Center personnel based on a standardized protocol.⁴³ Only reliable tests (≤ 33% fixation losses and false-negative errors, and ≤ 33% false-positive errors) were included in the analysis. Visual fields with the following artifacts were also excluded: evidence of rim and eyelid artifacts, inattention or fatigue effects, or VF damage caused by a disease other than glaucoma. The definition of VF progression required either of the following trend-based or event-based criteria: 1) a negative rate of 24–2 VF mean deviation (MD) change over the follow-up period with a P-value ≤ 0.05, 2) a negative rate of 24–2 VF index change over the follow-up with a P-value ≤ 0.05, 3) a rate of change in threshold sensitivity of < −1 dB/year detected at three or more test locations with a P-value < 0.01,^44–48 and 4) a significant change on >3 locations (ie, change greater than the test-retest variability) compared with 2 baseline examinations for ≥ 3 consecutive visits (ie, “likely progression” as reported in the Guided Progression Analysis) during the study follow-up with the changes being observed at the latest follow-up visit.⁴⁹

OCTA and SPECTRAL-DOMAIN OCT IMAGING

OCTA and spectral-domain OCT imaging were performed by the AngioVue imaging system (software version 2018.1.1.63). Using this platform, OCTA and spectral-domain OCT images are obtained from the same volumetric scans allowing precise automated registration of OCTA and OCT images and providing quantified metrics for the analysis of different layers of interest. The Avanti system for measuring vessel density and tissue thickness has been described previously.²⁷ Briefly, vessel density is defined as the proportion of measured area occupied by flowing blood vessels, defined as pixels having decorrelation values acquired by the split-spectrum amplitude-decorrelation angiography algorithm above the threshold level.⁵⁰ Capillary density (CD) which is another major metric of the OCTA instrument for the ONH images is measured after the automated removal of large vessels from the original en-face angiogram using the Angiovue software. In the present study, we analyzed OCTA images of both the ONH and the macula.

For the ONH OCTA images, the CD was calculated at the radial peripapillary capillary plexus in images comprised of 304 × 304 A-scans with a 4.5 × 4.5 mm² field of view centered on the optic disc. The retinal layers of each scan were automatically segmented by the AngioVue software in order to visualize the radial peripapillary capillary plexus layer in a slab from the internal limiting membrane (ILM) to RNFL posterior boundary. Circumpapillary CD was calculated over the region defined as a 750-μm-wide elliptical annulus extending from the optic disc boundary encircling 360-degree global area and was included in the analysis for the assessment of the global ONH microvascular structure. Superior and inferior hemifield CD measurements were calculated similarly but over the superior and inferior hemifields, respectively. The optic nerve cube scanning protocol was used to measure the circumpapillary RNFL thickness and hemifield measurements of the same scan slab as the OCTA scan. For the macular OCTA images, vessel density at the superficial macular slab was obtained from 3 × 3 mm² OCTA images comprised of 304×304 A-scans centered on the fovea. The retinal layers of each scan were segmented automatically by the AngioVue software to visualize the superficial retinal capillary plexus, as follows. Superficial vessel density was calculated in a macular slab extending from the ILM to 10 μm offset below the inner plexiform layer. Similarly, hemifield measurements were obtained using the entire corresponding hemifield in each image. The macula cube ganglion cell complex (GCC) thickness was calculated using the same volumetric scans as those of OCTA images. GCC thickness of the whole image as well as the corresponding hemifield measurements were calculated.

Only good-quality images were included in the analysis. Image quality review was completed on all OCTA and OCT images processed with standard AngioVue software (version 2018.1.1.63) according to a standard protocol established by the Imaging Data Evaluation and Analysis Reading Center. Expert reviewers evaluated all of the images and excluded those with poor quality, defined as images with any of the following: 1) low scan quality with quality index of less than 4, (2) poor clarity, (3) residual motion artifacts visible as irregular vessel pattern or disc boundary on the en-face angiogram, (4) image cropping or local weak signal resulting from the opacity of optic media, (5) signal devoid area as a result of blink, (6) poor centration on the ONH (for ONH images) or fovea (for macular images) and (7) the presence of segmentation errors that could not be corrected.⁵¹

GRADIENT BOOSTING CLASSIFIER MODELS

The GBC is an ensemble classifier that attempts to decrease error by resampling and varying the weights for individual weak learners to increase classification accuracy.⁵² In an empirical comparative study of supervised learning algorithms comparing random forests and boosted decision trees, the GBC had the best overall performance.⁵³ An advantage of GBC is that it provides a metric to assess the relative influence of each parameter included in the classifier.⁵⁴

We developed 8 separate GBC models using different combinations of baseline and longitudinal (rates of change for OCT/OCTA measurements were calculated within the entire duration of VF follow-up [± 6 months] using univariable linear regression) OCT/OCTA parameters as input features for the classification of glaucomatous VF progression. The individual inputs to the GBC models were as follows: 1) Baseline ONH and macular OCT and OCTA parameters along with their rates of change over time at the global and hemifield levels (Longitudinal Global + Hemifield Combined); 2) Baseline ONH and macular OCTA parameters along with their rates of change over time at the global and hemifield levels (Longitudinal Global + Hemifield OCTA); 3) Baseline ONH and macular OCT parameters along with their rates of change over time at the global and hemifield levels (Longitudinal Global + Hemifield OCT); 4) Baseline ONH and macular OCTA parameters at the global and hemifield levels (Baseline Global + Hemifield Combined); 5) Baseline global ONH and macular OCT and OCTA parameters along with their rates of change over time (Longitudinal Global Combined); 6) Baseline global ONH and macular OCTA parameters along with their rates of change over time (Longitudinal Global OCTA); 7) Baseline global ONH and macular OCT parameters along with their rates of change over time (Longitudinal Global OCT); and 8) Baseline global ONH and macular OCTA parameters (Baseline Global Combined). For comparative purposes, 8 separate logistic regression models were also developed using the same subsets of input features. Table 1 summarizes the included OCT/OCTA parameters and provides information about the input features used for developing each of the investigated models.

Table 1.

OCT/OCTA Summary Parameters Used as Input to the Models for the Detection of Glaucomatous VF Progression

Imaging Modality	Location	Measurement Area		Model	Longitudinal Global + Hemifield Combined	Longitudinal Global + Hemifield OCTA	Longitudinal Global + Hemifield OCT	Baseline Global + Hemifield Combined	Longitudinal Global Combined	Longitudinal Global OCTA	Longitudinal Global OCT	Baseline Global Combined
			Parameters
OCT
	ONH	Global	Baseline Circumpapillary RNFL Thickness (μm)		✓		✓	✓	✓		✓	✓
		Hemifield	Baseline Superior Hemifield Circumpapillary RNFL Thickness (μm)		✓		✓	✓
		Hemifield	Baseline Inferior Hemifield Circumpapillary RNFL Thickness (μm)		✓		✓	✓
		Global	Circumpapillary RNFL Thickness Slope* (μm/yr)		✓		✓		✓		✓
		Hemifield	Superior Hemifield Circumpapillary RNFL Thickness Slope* (μm/yr)		✓		✓
		Hemifield	Inferior Hemifield Circumpapillary RNFL Thickness Slope* (μm/yr)		✓		✓
	Macula	Global	Baseline Macular GCC Thickness (μm)		✓		✓	✓	✓		✓	✓
		Hemifield	Baseline Superior Hemifield Macular GCC Thickness (μm)		✓		✓	✓
		Hemifield	Baseline Inferior Hemifield Macular GCC Thickness (μm)		✓		✓	✓
		Global	Macular GCC Thickness Slope* (μm/yr)		✓		✓		✓		✓
		Hemifield	Superior Hemifield Macular GCC Thickness Slope* (μm/yr)		✓		✓
		Hemifield	Inferior Hemifield Macular GCC Thickness Slope* (μm/yr)		✓		✓
OCTA
	ONH	Global	Baseline Circumpapillary Capillary Density (%)		✓	✓		✓	✓	✓		✓
		Hemifield	Baseline Superior Hemifield Circumpapillary Capillary Density (%)		✓	✓		✓
		Hemifield	Baseline Inferior Hemifield Circumpapillary Capillary Density (%)		✓	✓		✓
		Global	Circumpapillary Capillary Density Slope* (%/yr)		✓	✓			✓	✓
		Hemifield	Superior Hemifield Circumpapillary Capillary Density Slope* (%/yr)		✓	✓
		Hemifield	Inferior Hemifield Circumpapillary Capillary Density Slope* (%/yr)		✓	✓
	Macula	Global	Baseline Macular Vessel Density (%)		✓	✓		✓	✓	✓		✓
		Hemifield	Baseline Superior Hemifield Macular Vessel Density (%)		✓	✓		✓
		Hemifield	Baseline Inferior Hemifield Macular Vessel Density (%)		✓	✓		✓
		Global	Macular Vessel Density Slope* (%/yr)		✓	✓			✓	✓
		Hemifield	Superior Hemifield Macular Vessel Density Slope* (%/yr)		✓	✓
		Hemifield	Inferior Hemifield Macular Vessel Density Slope* (%/yr)		✓	✓

OCT: optical coherence tomography, OCTA: optical coherence tomography angiography, VF: visual field, ONH: optic nerve head, RNFL: retinal nerve fiber layer, GCC: ganglion cell complex.

^✓Indicates that the parameter was used as input to the model.

^*Rates of change were calculated using univariable linear regression.

TRAINING AND EVALUATION

Fivefold cross-validation was used to provide out-of-sample predictions for GBC and logistic regression models to avoid overly optimistic estimates of classification accuracy. Both stable and progressive eyes were randomly divided at the patient level into 5 subsets. Then, for each model, we used 4 subsets to train the model, and then we used the fifth subset to assess model performance. This sequence was repeated 5 times, with each subset serving as the test set 1 time, so that each tested eye was never part of its own training set and was tested only once.

STATISTICAL ANALYSES

Continuous and categorical data were presented as mean (95% confidence interval [CI]) and count (%). Statistically significant differences in characteristics between stable participants and those with VF progression were determined by 2-sample t-tests for continuous variables and the Fisher exact test for categorical variables. Eye characteristics were compared using linear mixed-effects models with random intercepts to account for within-subject variability.

Areas under receiver operating characteristic curves (AUROCs) were used to describe and compare the ability of different GBC and logistic regression models to distinguish between stable eyes and those with VF progression. Sensitivities at fixed specificities of 80%, 85%, 90%, and 95% were also reported for each GBC model. As measurements from both eyes of the same subject are likely to be correlated, the cluster of data for the study subject was considered as the unit of resampling and bias-corrected CIs and hypothesis tests for AUROCs were conducted using a clustered bootstrap with 2,000 resamples to take into account the potential bias of inter-eye correlation on the AUROCs and P-values. Due to the exploratory nature of this analysis, no type I error correction for multiple comparisons was applied (as recommended by Bender and Lange⁵⁵). GBC modeling was performed using the XGBoost package (version 1.5.2) on Python (version 3.9.6) for Windows. Statistical analyses were performed using Stata 17.0 (StataCorp LLC). P-values of less than 0.05 were considered statistically significant.

RESULTS

This prospective cohort study included 110 eyes (37 glaucoma suspects, 73 glaucomas) of 71 participants over an average follow-up duration of 4.1 years (95% CI: 4.0, 4.2) with a mean of 6.8 (95% CI: 6.5, 7.1) 24–2 VF tests. After image quality evaluation, 253 (18.6%) out of 1363 captured pairs of OCT/OCTA images were excluded because of having poor quality due to the presence of artifacts resulting in a total of 1110 pairs of OCT/OCTA images to be included in the final analysis. The average number of included pairs of OCT/OCTA measurements was 5.5 (95% CI: 5.2, 5.8) for ONH imaging, and 4.6 (95% CI: 4.3, 4.9) for macular imaging. Visual field progression was detected in 28 eyes (25.5%) at the end of the follow-up. Eyes with and those without VF progression had similar baseline demographic and clinical characteristics in terms of gender, race, history of diabetes, history of hypertension, systolic and diastolic blood pressure, age, axial length, CCT, IOP, 24–2 VF MD, and 24–2 VF PSD. Eyes with VF progression had worse 24–2 VF MD (−7.80 dB versus −3.51 dB, P-value = 0.001), and higher 24–2 VF PSD (6.91 dB versus 4.13 dB, P = 0.002) compared to those without VF progression at the end of the follow-up. Table 2 provides a summary of the characteristics of the study population.

Table 2.

Characteristics of the study population (71 patients, 110 eyes)

Variables	No VF Progression	VF Progression	P-value*
Patient level characteristics	50 (70.4%)	21 (29.6%)
Baseline age (years)	69.1 (65.8, 72.3)	68.2 (62.8, 73.6)	0.777
Gender (females, %)	23 (46.0%)	11 (52.4%)	0.795
Race (African American, %)	14 (28.0%)	3 (14.3%)	0.361
Diabetes	7 (14.0%)	2 (9.5%)	0.716
Hypertension	30 (60%)	10 (47.6%)	0.433
Systolic BP (mmHg)	126.7 (121.5, 131.9)	131.6 (123.7, 139.4)	0.303
Diastolic BP (mmHg)	77.9 (74.9, 80.9)	78.8 (75.4, 82.1)	0.736
Mean arterial BP (mmHg)	94.2 (90.9, 97.5)	96.4 (92.7, 100.1)	0.431
Eye level characteristics	82 (74.5%)	28 (25.5%)
Axial length (mm)	24.5 (24.2, 24.7)	24.3 (23.9, 24.7)	0.248
CCT (μm)	534.0 (525.0, 543.1)	527.4 (513.1, 541.6)	0.257
Baseline IOP (mmHg)	14.6 (13.6, 15.6)	15.1 (13.2, 17.0)	0.626
Mean IOP during follow-up (mmHg)	14.8 (14.0, 15.6)	15.3 (13.5, 17.0)	0.655
VF Measurements
Baseline MD (dB)	−3.18 (−4.25, −2.10)	−4.17 (−6.36, −1.98)	0.506
Baseline PSD (dB)	3.93 (3.16, 4.69)	5.69 (3.85, 7.52)	0.067
Last MD (dB)	−3.51 (−4.67, −2.34)	−7.80 (−10.64, −4.96)	0.001
Last PSD (dB)	4.13 (3.31, 4.94)	6.91 (5.05, 8.76)	0.002

VF: visual field, BP: blood pressure, CCT: central corneal thickness, IOP: intraocular pressure, MD: mean deviation, PSD: pattern standard deviation.

^*P-value < 0.05 was considered as statistically significant and is shown in bold.

Table 3 and Figure 1 demonstrate the classification accuracy of the GBC models (in terms of AUROCs) with different subsets of OCT/OCTA parameters as input features for the detection of glaucomatous VF progression. Supplemental Table 1 demonstrates the statistical significance for pairwise comparisons of AUROCs between different GBC models. The sensitivities at fixed specificities of 80%, 85%, 90%, and 95% are shown in Table 3. The GBC models combining OCT and OCTA parameters as input features outperformed their respective models that used either OCT or OCTA measurements alone. Among the GBC models that used global OCT/OCTA measurements, the model with the combination of OCT and OCTA parameters had a higher AUROC (0.84, 95% CI: 0.80, 0.88) compared to the AUROCs of the OCT-based (0.81, 95% CI: 0.77, 0.85) and the OCTA-based (0.80, 95% CI: 0.77, 0.84) models (all P-values < 0.05). Similarly, among the GBC models that included OCT/OCTA measurements at the global and hemifield levels, the combination of OCT and OCTA measurements resulted in a higher classification accuracy (AUROC = 0.89, 95% CI: 0.87, 0.92) compared to the OCT-based (AUROC = 0.85, 95% CI: 0.81, 0.88), and the OCTA-based (AUROC = 0.85, 95% CI: 0.82, 0.88) models (all P-values < 0.05).

Table 3.

Estimated Areas Under the Receiver Operating Characteristic Curves (AUROC) and Sensitivities at Fixed Specificities for All Investigated Gradient Boosting Classifier Models for the Detection of Glaucomatous Visual Field Progression

		Sensitivity at
Model	AUROC (95% CI)	80% Spec.	85% Spec.	90% Spec.	95% Spec.	P-value*
Longitudinal Global + Hemifield Combined	0.89 (0.87, 0.92)	79.2	76.9	72.2	53.8	-
Longitudinal Global + Hemifield OCTA	0.85 (0.82, 0.88)	71.8	68.0	64.9	55.8	0.040
Longitudinal Global + Hemifield OCT	0.85 (0.81, 0.88)	71.9	63.6	51.7	39.2	0.003
Baseline Global + Hemifield Combined	0.63 (0.58, 0.69)	39.2	30.4	18.2	9.0	<0.001
Longitudinal Global Combined	0.84 (0.80, 0.88)	77.1	73.7	68.3	56.7	0.013
Longitudinal Global OCTA	0.80 (0.77, 0.84)	68.5	61.1	51.9	34.8	<0.001
Longitudinal Global OCT	0.81 (0.77, 0.85)	72.0	66.3	58.9	44.9	<0.001
Baseline Global Combined	0.60 (0.55, 0.65)	36.9	30.0	21.3	11.0	<0.001

AUROC: area under the receiver operating characteristic curve, Spec: specificity, OCTA: optical coherence tomography angiography, OCT: optical coherence tomography.

^*Statistical significance was determined using a paired bootstrap test and P-values compare the AUROC of the models to the AUROC of the “Longitudinal Global + Hemifield Combined” model. Bolded P-values indicate a statistically significant difference with P ≤ 0.05.

Figure 1.

Area under the receiver operating characteristic curve (AUROC) of different optical coherence tomography (OCT)-based, optical coherence tomography angiography (OCTA)-based, and combined OCT and OCTA-based GBC models for the detection of glaucomatous visual field progression. As demonstrated, models that include OCT and/or OCTA measurements at the global and hemifield levels (B) outperformed their respective models that include only the global measurements (A). Also, the combination of OCT and OCTA measurements improved the classification accuracy compared to including only the OCT or OCTA measurements. At last, the addition of longitudinal rates of change of OCT and/or OCTA measurements to the baseline values substantially improved the classification performance of the models. The Longitudinal Global + Hemifield Combined GBC model showed the highest classification accuracy for the detection of glaucomatous visual field progression compared to the other models (AUROC: 0.89 [0.87, 0.92]).

The addition of OCT/OCTA rates of change to the models resulted in considerable improvement in the classification accuracy of the GBC models. The AUROCs for the GBC models combining baseline OCT and OCTA measurements were 0.60 (95% CI: 0.55, 0.65) after including the global measurements, and 0.63 (95% CI: 0.58, 0.69) after including both the global and the hemifield parameters. After the addition of OCT/OCTA rates of change to their corresponding baseline values the AUROCs increased to 0.84 (95% CI: 0.80, 0.88) for the GBC model with global measurements, and to 0.89 (95% CI: 0.87, 0.92) for the GBC model with the global and hemifield measurements.

The addition of OCT/OCTA measurements at the hemifield level to the global measurements improved the classification accuracy of the GBC models. After the addition of hemifield measurements, the AUROCs increased from 0.84 (95% CI: 0.80, 0.88) to 0.89 (95% CI: 0.87, 0.92) for the GBC models combining OCT and OCTA measurements, from 0.81 (95% CI: 0.77, 0.85) to 0.85 (95% CI: 0.81, 0.88) for the OCT-based GBC models, and from 0.80 (95% CI: 0.77, 0.84) to 0.85 (95% CI: 0.82, 0.88) for the OCTA-based GBC models (all P-values < 0.05).

Supplemental Table 2 demonstrates the classification accuracy and the statistical significance for pairwise AUROC comparisons of different investigated logistic regression models. The longitudinal model with combined OCT and OCTA parameters at the global and hemifield levels had the highest classification accuracy (AUROC = 0.79 [95% CI: 0.76, 0.82]) for the detection of glaucomatous VF progression which was significantly higher than those of the OCT-based models (P-values < 0.05). Similar to the GBC models, the logistic regression models with longitudinal OCT/OCTA parameters (AUROCs ranging from 0.72 to 0.79) significantly outperformed the models with baseline OCT/OCTA parameters (AUROCs ranging from 0.60 to 0.62) for the detection of glaucomatous VF progression (all P-values < 0.05). However, the difference in models’ classification accuracy after the addition of hemifield to the global measurements did not reach statistical significance (all P-values > 0.05).

Figure 2 shows the relative influence of different OCT/OCTA parameters for the detection of glaucomatous VF progression in the best GBC model. As demonstrated, longitudinal OCT/OCTA measurements had a higher overall relative influence compared to baseline measurements, and a combination of ONH and macular parameters contributed to the overall model performance with the inferior hemifield circumpapillary RNFL thickness slope showing the highest relative influence compared to the other included parameters. Figure 3 shows the baseline and end of follow-up macular vessel density and GCC thickness maps along with the corresponding 24–2 VF tests of a representative eye with VF progression during the follow-up.

Figure 2.

Mean relative influence over each cross-validation fold in the Longitudinal Global + Hemifield Combined GBC model. As demonstrated, the rates of change of optical coherence tomography (OCT) and optical coherence tomography angiography (OCTA) measurements showed higher relative influence compared to their respective baseline measurements for the detection of glaucomatous visual field progression.

Figure 3.

Baseline and end of follow-up macular optical coherence tomography angiography (OCTA), and 24–2 visual field (VF) assessments of one of the study participants with VF progression. “A1” shows the en-face OCTA images segmented at the superficical macular layer. “A2” and “A3” demonstrate the corresponding color-coded OCTA-derived vessel density and optical coherence tomography-derived ganglion cell complex thickness maps, respectively. The image pairs reveal inferior hemifield macular damage during the follow-up. “B” shows the corresponding superior hemifield VF progression in the same eye during the study follow-up.

DISCUSSION

In the present study, different subsets of baseline and longitudinal OCT and OCTA structural measurements were used as input features to develop and evaluate the classification performance of GBC machine learning models for the detection of glaucomatous VF progression. The combination of OCT and OCTA measurements increased the classification performance of the GBC models compared to using either OCT or OCTA measurements alone. Moreover, the addition of longitudinal OCT/OCTA features to the baseline measurements substantially improved the classification accuracy of the GBC models. Finally, we demonstrated that including OCT/OCTA structural information at the hemifield level improves the GBC models’ classification accuracy compared to using global indices alone. Our findings support the utility of OCTA to provide complementary and clinically relevant information along with OCT for the evaluation of VF deterioration in glaucoma patients.

To the best of our knowledge, this is the first prospective longitudinal study to integrate OCT and OCTA structural measurements for the detection of glaucomatous VF progression over several years (mean follow-up duration of 4.1 years in this study). Recent investigations have evaluated whether OCTA measurements improve or complement OCT structural metrics for the diagnosis and evaluation of structure-function relationship in glaucoma.^17,33–35 Bowd and colleagues demonstrated that the combination of OCT and OCTA parameters of the ONH and the macula using artificial intelligence provides high diagnostic accuracy (AUROC = 0.93) for discriminating between healthy and mild to moderate glaucoma patients. Although the combination of OCT and OCTA metrics outperformed the diagnostic performance of OCTA metrics alone, this combination showed a similar diagnostic accuracy compared to that of the model using OCT measurements alone.³³ This finding was later confirmed by another study that employed several machine learning strategies to integrate OCT and OCTA parameters for the diagnosis of mild to moderate glaucoma patients.³⁴ In contrast, two recent studies Wong and associates revealed that OCTA-provided vascular measurements demonstrate a better structure-function relationship compared to OCT-derived metrics²³, and the combination of OCT and OCTA structural indices improves the modeling of focal VF defects in early glaucoma.³⁵ Notably, all of the mentioned studies have been cross-sectional given the novelty of the OCTA technology and lack of enough longitudinal follow-up data for the detection of VF progression in glaucoma patients.

The addition of longitudinal OCT/OCTA measurements to the baseline metrics substantially improved the classification accuracy of the detection of VF progression. This finding highlights the role of longitudinal structural monitoring with OCT and OCTA technologies in glaucoma patients. Previous studies have reported a range for the accuracy of global and more localized baseline structural parameters to predict future glaucomatous VF progression from no statistically significant association to a statistically significant, but weak to relatively moderate classification accuracy.^{31,37,56–62} Moreover, given the relatively recent introduction of OCTA, published longitudinal studies on the evaluation of glaucoma patients using this technology are limited.²⁹ In this regard, prior investigations on healthy, glaucoma suspect and glaucoma patients have demonstrated a longitudinal reduction in macular OCT and OCTA structural parameters over the follow-up time^11,63,64 and also an association between baseline OCTA metrics and prior rates of VF progression.¹⁸ Two recent longitudinal investigations have found an association between rates of change in OCTA measurements and concurrent glaucomatous VF deterioration. In a prospective study of glaucoma suspect and glaucoma patients, Nishida and colleagues demonstrated that faster OCTA measured macular vessel density loss is associated with accelerated concurrent and subsequent rates of VF progression in these patients.⁶⁵ In another study, Shin et al. demonstrated that faster OCTA derived circumpapillary capillary density loss is associated with a higher probability of concurrent VF progression regardless of disease stage. Notably, circumpapillary RNFL thickness was associated with VF progression only in eyes with early-stage glaucoma, and baseline circumpapillary RNFL thickness and capillary density were not predictive future of VF progression.³¹ These findings are in agreement with our results as the addition of longitudinal rates of change of OCT and OCTA measurements considerably improved the predictive performance of the investigated models for the detection of VF progression. Even though the accuracy of baseline OCT/OCTA parameters for the prediction of future VF progression was limited, the present study included only the global and hemifield level measurements. Future studies combining more localized structural parameters in addition to the currently investigated metrics might enhance the prediction accuracy of subsequent VF progression.

The addition of OCT/OCTA measurements at the ONH and the macular hemifield levels improved the diagnostic accuracy of the GBC models for the detection of glaucomatous VF progression. A combination of ONH and macular measurements contributed to the model performance of the GBC with the highest classification accuracy with the inferior circumpapillary RNFL thickness slope demonstrating the highest relative influence compared to the other included parameters. In a previous longitudinal study of glaucoma patients (baseline MD = −8.5 dB) with an average follow-up duration of 3.7 years, Leung and colleagues found the inferotemporal sector as the most frequent location where RNFL progression was detected.⁶⁶ Moreover, recent evidence suggests that the combination of ONH and macular structural information improves the accuracy to detect the presence of glaucomatous damage^33,67,68 and to monitor disease progression and understanding of its patterns in glaucoma patients.^69,70 Although global structural indices are commonly used metrics for the assessment of structural change in glaucoma, sole reliance on these measurements has the potential to miss subtle defects usually occurring at the earlier stages of the disease or failure to detect further structural damage at the advanced stage where these indices are approaching their measurement floor. Using a region of interest approach and integration of different structural parameters into statistical models are among the proposed solutions to overcome this potential limitation.^37,71–74 Moreover, the advent of artificial intelligence and machine learning strategies have enabled more accurate modeling and simultaneous incorporation of different structural information for a variety of classification tasks including the discrimination of stable versus progressive glaucoma eyes. Bowd and colleagues recently developed a deep learning autoencoder algorithm trained on SD-OCT-derived RNFL thickness map data of non-glaucomatous changes (e.g., aging, variability) in a longitudinal follow-up of glaucoma patients. Subsequently, they used a Markov-based segmentation algorithm to classify the RNFL thickness map into regions of “no change”, “not likely progression”, and “likely progression”. They found that their proposed model identified 40% more eyes with glaucomatous progression compared with using the average circumpapillary RNFL thickness of the same optic disc cube scans.⁷¹ In another study on moderate to advanced glaucoma eyes, Nouri-Mahdavi and associates developed machine learning models using different combinations of baseline and longitudinal ONH and macular OCT data for the detection of VF progression with a clinically relevant classification accuracy (AUROC = 0.81).³⁷ These reports show promise for the application of artificial intelligence algorithms to enhance the accuracy of monitoring functional glaucoma deterioration using multimodal and different combinations of structural data.

This study has some possible limitations. Previous evidence shows that the OCTA artifacts with a potential influence on the instrument metrics are frequent.⁵¹ To address this issue, all OCTA images were carefully evaluated by expert reviewers according to a previously published systematic quality control criterion⁵¹ and poor-quality images were excluded from the analysis. In addition, all eyes were required to have concurrent longitudinal follow-up data for a minimum of 5 reliable VF tests along with good quality OCT and OCTA images of both the ONH and the macula for at least 3 visits. The use of strict quality control and inclusion criteria resulted in relatively small sample size and the results of the GBC models were not tested on an external, independent test dataset. However, 5-fold cross-validation with independent test and training sets was used to control for the overfitting of models and to avoid overly optimistic estimates of accuracy.^33,37,75 Future longitudinal studies with increased sample size and longer follow-up duration may provide a more comprehensive representation of the complex structure-function relationship in glaucoma patients, increase the accuracy of prognostic models using structural data by the inclusion of multimodal structural and vascular information, and allow for the evaluation of the prognostic significance of longitudinal OCT/OCTA measurements for the prediction of VF progression over extended follow-up durations.

In conclusion, we demonstrated that the combination of OCT and OCTA measurements using feature-based GBC machine learning models improves the accuracy of the detection of glaucomatous VF progression compared to using individual OCT or OCTA metrics alone. The classification accuracy of the models considerably improved after the addition of longitudinal OCT/OCTA features to the baseline measurements. OCTA measurements have the potential to provide a complementary dimension to established OCT indices for monitoring VF progression in glaucoma patients.

Abbreviations and Acronyms:

AUROC

area under the receiver operating characteristic curve

BCVA

best-corrected visual acuity

CD

capillary density

CI

confidence interval

DIGS

diagnostic innovations in glaucoma study

GBC

gradient boosting classifier

GCC

ganglion cell complex

ILM

internal limiting membrane

MD

mean deviation

OCT

optical coherence tomography

OCTA

optical coherence tomography angiography

ONH

optic nerve head

PSD

pattern standard deviation

RNFL

retinal nerve fiber layer

SD-OCT

spectral-domain optical coherence tomography

VF

visual field