Retinal OCTA-derived microvascular remodeling is associated with coronary microvascular dysfunction in women with ischemia and no obstructive coronary artery disease: A pilot study

1. Introduction

Ischemia with no obstructive coronary artery disease (INOCA) disproportionately affects women and is associated with persistent angina, impaired quality of life, and adverse cardiovascular outcomes [1]. Coronary microvascular dysfunction (CMD), commonly diagnosed by impaired coronary flow reserve (CFR), is a major mechanism of INOCA [2]. Although cardiac positron emission tomography (PET) can non-invasively assess flow reserve, comprehensive CMD phenotyping often requires invasive coronary function testing, which is not readily scalable for early triage or longitudinal follow-up.

The retinal and coronary microcirculations share small-vessel caliber, autoregulatory physiology, and susceptibility to systemic vascular injury [3]. Optical coherence tomography angiography (OCTA) enables rapid, dye-free assessment of retinal microvascular architecture. Prior work from our group reported reduced retinal vessel density in women with CMD [4]. However, density alone may miss broader remodeling reflected by foveal avascular zone (FAZ) geometry, inter-vessel spacing, and tortuosity. We derived an OCTA Remodeling Index (ORI) and evaluated its association with CMD status and CFR in women with INOCA.

2. Methods

We retrospectively identified women with INOCA evaluated at Emory University. Of 24 women meeting the standard INOCA criterion, defined as <50% epicardial stenosis on coronary angiography, 21 had CMD ascertainment available and 18 had both CMD ascertainment and gradable bilateral 3 × 3-mm macular OCTA. These 18 women formed the analytic cohort. CMD was defined by impaired coronary microvascular function using CFR <2.5 when numeric CFR was available. Sixteen participants had numeric CFR assessment, and two were classified using cardiac PET-based microvascular assessment, one CMD and one non-CMD. This selection may enrich for women with more complete evaluation and higher symptom burden than the broader INOCA population. The study was approved by the Emory University IRB (STUDY2025P010279).

For contextual visualization only, we evaluated 16 healthy female controls from the public RASTA OCTA dataset, selected to have CHA₂DS₂-VASc score 1 attributable to female sex and no reported cardiometabolic comorbidity [5]. These controls were younger, leaner, and imaged on a different OCTA platform; therefore, they were not used for inferential between-cohort testing.

Macular OCTA was acquired after pharmacologic dilation on the Cirrus HD-OCT 500/5000 platform. The 3 × 3-mm superficial vascular complex slab was processed using deep learning-based segmentation of the retinal vasculature and FAZ [6]. Eye-level features quantified FAZ morphology, vessel density and length density, inter-branch spacing, tortuosity, and fractal complexity; bilateral features were averaged for each patient.

The ORI model, denoted M_ORI, was developed using nested leave-one-out cross-validation. Within each training fold, 19 candidate OCTA features were ranked by Cliff’s delta, highly correlated features were removed at |r|>0.90, retained features were quantile-normalized, and the score was calculated as an unweighted signed sum. Feature directionality was learned within the training fold so that higher M_ORI values reflected greater CMD association. The most stable features were maximum 3D box-counting fractal dimension of the FAZ, mean global branch distance, and mean parafoveal tortuosity. M_ORI was therefore treated as an internally derived retinal remodeling composite rather than an externally validated clinical classifier.

For benchmarking, M_ORI was compared with a clinical-variable model (M_clinical), a vessel-density model (M_density), and combined models incorporating M_ORI with clinical variables (M_ORI+clinical) or density metrics (M_ORI+density). M_clinical included age, body mass index, diabetes, hypertension, hyperlipidemia, reactive hyperemia index, augmentation index, augmentation index standardized to 75 beats/minute, and pulse-wave velocity and excluded CFR. M_density included extrafoveal avascular zone central density, inner and full-field vessel density, central perfusion, and inner and full-field perfusion density.

CMD versus non-CMD separation was assessed using the Mann–Whitney U test and Cliff’s delta with bootstrap 95% confidence intervals from 3000 resamples. ORI-CFR association was evaluated using Spearman correlation. Model discrimination was summarized using leave-one-out cross-validated AUC. Robustness analyses included exclusion of PET-classified participants, age- and BMI-adjusted partial Spearman correlation, leave-pair-out refitting, and full-pipeline label permutation testing with 1000 permutations. Threshold metrics were not reported because of the pilot sample size.

3. Results

The analytic cohort included 18 women with INOCA: 11 with CMD and 7 without CMD. Mean age was 54.7 ± 12.5 years. Median CFR was lower in CMD than non-CMD participants: 1.95 [IQR 1.70–2.10] versus 3.40 [IQR 3.24–3.58], p = 0.001. Age, BMI, diabetes, hypertension, hyperlipidemia, peripheral endothelial function, arterial stiffness measures, and OCTA signal strength did not differ statistically between groups.

ORI was substantially higher in women with CMD compared with non-CMD participants (Cliff’s δ=0.95; 95% CI 0.77–1.00; p < 0.001). ORI was also strongly inversely associated with CFR among participants with continuous CFR values available (Spearman ρ=−0.80; 95% CI −0.94 to −0.49; p < 0.001; n = 16). The leave-one-out cross-validated AUC for M_ORI was 0.97 (95% CI 0.88–1.00), interpreted as an internal cross-validation estimate.

The association between ORI and CMD was not explained by PET-classified participants. After excluding both PET-classified cases, CMD versus non-CMD separation remained essentially unchanged (Cliff’s δ=0.93; 95% CI 0.69–1.00; p = 0.001). The age- and BMI-adjusted partial Spearman association between M_ORI and CFR also remained strong (partial ρ=−0.82; p < 0.001; n = 15). Across all 77 leave-pair-out refits, the median Cliff’s δ was 0.93, with a range of 0.93–1.00. Full-pipeline label permutation testing yielded an add-one p-value of 0.021.

In comparator analyses, M_clinical showed little separation between CMD and non-CMD participants (Cliff’s δ=−0.12; p = 0.724; AUC=0.44), whereas M_density showed more modest separation than M_ORI (Cliff’s δ=0.64; p = 0.027; AUC=0.82). Adding clinical variables or vessel-density metrics to M_ORI did not improve performance in this pilot cohort. External controls had lower descriptive ORI values than both INOCA groups: −4.12±1.14 in controls, −2.51±1.55 in non-CMD INOCA, and 1.57±1.60 in CMD INOCA. The main findings are summarized in Fig. 1 and Tables 1 and 2.

Fig. 1.

Central illustration. Top schematic: Retinal and coronary microvascular beds share pathobiological mechanisms and are similarly susceptible to traditional cardiovascular risk factors (left); cohort flow diagram of 18 INOCA women included in the analysis who had CFR information and gradable, bilateral OCTA (middle); ORI construction pipeline (right). (A) ORI distribution in external healthy controls (n = 16; visual context only: see Methods), non-CMD INOCA (n = 7), and CMD INOCA (n = 11); PET-classified INOCA participants are highlighted with bold edges. The pairwise statistic shown is the within-INOCA primary analysis (Cliff’s δ = 0.95, p < 0.001); between-cohort comparisons with external controls are not inferentially tested. The PET-excluded comparison excludes both PET-classified INOCA participants, including one CMD and one non-CMD participant. (B) ORI is inversely associated with CFR (Spearman ρ = −0.80, p = 2 × 10⁻⁴, n = 16); shaded band is the bootstrap 95% CI for the regression line; dashed vertical line marks CFR = 2.5 used for the CMD definition.

Table 1.

Baseline characteristics and model performance for CMD identification.

Panel A. Demographics and clinical characteristics
Characteristics	All INOCA (N = 18)	CMD (n = 11)	non-CMD (n = 7)	p	External reference (n = 16)
Age, years	54.7 ± 12.5	56.5 ± 11.8	52.0 ± 13.9	0.62	39.6 ± 13.8
Female sex, n (%)	18 (100)	11 (100)	7 (100)	-	16 (100)
BMI, kg/m2	30.6 ± 7.3	28.1 ± 5.6	34.2 ± 8.4	0.13	21.9 ± 2.0
Diabetes mellitus, n (%)	5 (28)	3 (27)	2 (29)	1.00	0 (0)
Hypertension, n (%)	10 (56)	6 (55)	4 (57)	1.00	0 (0)
Hyperlipidemia, n (%)	14 (78)	8 (73)	6 (86)	1.00	0 (0)
CFR, median [IQR]	2.15 [1.77–3.24] *	1.95 [1.70–2.10]	3.40 [3.24–3.58]	0.001	-
Reactive hyperemia index	1.9 ± 0.7	2.0 ± 0.7	1.8 ± 0.7	0.60	-
Augmentation index, %	13.4 ± 18.3	17.8 ± 18.8	7.0 ± 16.7	0.36	-
AI@75, %	9.7 ± 18.6	14.0 ± 18.7	3.6 ± 18.0	0.43	-
Pulse-wave velocity, m/s	6.9 ± 1.4	7.0 ± 1.5	6.6 ± 1.5	1.00	-
OCTA signal strength	9.8 ± 0.5	9.7 ± 0.6	10.0 ± 0.0	0.28	-

Panel B. Model performance for CMD identification (N = 18; LOOCV unless noted)
Model	Cliff's delta	MW p	Spearman rho vs CFR	rho p	AUC (95% CI)
MORI	0.95	<0.001	−0.80	<0.001	0.97 [0.88–1.00]
Mclinical	−0.12	0.724	0.44	0.085	0.44 [0.15–0.71]
Mdensity	0.64	0.027	−0.26	0.331	0.82 [0.57–1.00]
MORI+clinical	0.87	0.001	−0.65	0.006	0.94 [0.79–1.00]
MORI+density	0.77	0.006	−0.52	0.039	0.88 [0.69–1.00]

Values are mean ± SD, median [IQR], or n (%). p values reflect CMD versus non-CMD comparisons only.

^⁎Continuous CFR values were available in 16/18 analytic participants; two additional participants were classified by cardiac PET (one CMD, one non-CMD) and were included in group analyses but excluded from continuous CFR correlations. External healthy controls are contextual only.

Table 2.

Three-group descriptive context and sensitivity/robustness analyses.

Panel A. ORI distribution by group (descriptive only; no between-cohort inference)
Group	n	Mean ± SD	Median [IQR]
Normal	16	−4.12 ± 1.14	−4.22 [−4.60, −3.39]
nonCMD	7	−2.51 ± 1.55	−2.74 [−3.11, −1.43]
CMD	11	1.57 ± 1.60	1.43 [0.75, 2.39]

Panel B. Sensitivity and robustness analyses
Analysis	n	Effect	Estimate [95% CI]	p value
Primary MORI CMD vs non-CMD	18	Cliff's delta	0.95 [0.77, 1.00]	p < 0.001
Primary MORI vs CFR	16	Spearman rho	−0.80 [−0.94, −0.49]	p < 0.001
PET-excluded CMD vs non-CMD	16	Cliff's delta	0.93 [0.69, 1.00]	p = 0.001
PET-excluded MORI vs CFR	16	Spearman rho	−0.80	p < 0.001
Age/BMI partial MORI vs CFR	15	Partial Spearman rho	−0.82	p < 0.001
Leave-pair-out refits (77 CMD/non-CMD pairs)	16 per refit	Median delta	0.93 [0.93, 1.00]	-
Full-pipeline label permutation	18	Add-one p	-	0.021

Panel C. ORI feature stability across LOOCV folds
Feature	Remodeling domain	Fold inclusion
Max 3D box-counting fractal	FAZ structural complexity	100.0%
Mean global branch distance	Inter-vessel spacing heterogeneity	94.4%
Mean parafoveal tortuosity	Geometric distortion	66.7%
Mean global tortuosity	Tortuosity	27.8%
Maximum global tortuosity	Tortuosity	5.6%
Global density	Density	5.6%

Abbreviations: AUC, area under the receiver-operating-characteristic curve; CFR, coronary flow reserve; CMD, coronary microvascular dysfunction; FAZ, foveal avascular zone; INOCA, ischemia with no obstructive coronary artery disease; LOOCV, leave-one-out cross-validation; OCTA, optical coherence tomography angiography; ORI, OCTA Remodeling Index.

4. Discussion

In this pilot study of women with INOCA, an OCTA-derived retinal remodeling composite was strongly associated with CMD status and inversely correlated with CFR. The key observation is that a retinal structural phenotype tracked with coronary vasodilatory reserve, the physiologic abnormality that defines CMD in many women with INOCA. This finding is biologically plausible because CMD reflects impaired vasodilatory reserve and small-vessel remodeling, not simply reduced vascular density. A retinal signal integrating FAZ complexity, inter-vessel spacing, and parafoveal tortuosity may therefore capture microvascular alterations not fully represented by density-based OCTA metrics alone.

These findings should be interpreted as preliminary evidence requiring prospective validation. M_ORI demonstrated stronger CMD separation than clinical variables or vessel-density metrics, and the association persisted across PET-exclusion, age/BMI-adjusted, leave-pair-out, and permutation analyses. OCTA-derived retinal metrics may eventually help identify women with INOCA who warrant earlier coronary physiologic testing or intensified preventive risk-factor management, but clinical utility requires larger representative cohorts and assessment of downstream decision change.

Several limitations are important. The cohort was small, retrospective, single-center, and limited to women who completed both coronary physiologic assessment and gradable bilateral OCTA. M_ORI was derived and evaluated within the same dataset, so performance estimates are internal and should not be interpreted as external diagnostic accuracy. CMD ascertainment included invasive testing in most participants and cardiac PET in two participants, although PET-exclusion sensitivity analysis yielded similar results. Finally, the external healthy cohort was used only for contextual visualization because demographic, cardiometabolic, and platform differences preclude diagnostic or causal inference.

5. Conclusions

In women with INOCA, OCTA-derived retinal microvascular remodeling was associated with CMD status and lower CFR in this pilot cohort. These findings do not establish retinal OCTA as a diagnostic test for CMD. Rather, they support prospective evaluation of OCTA-derived remodeling metrics as candidate non-invasive triage markers for identifying women with INOCA who may warrant earlier physiologic testing or intensified preventive risk-factor management.

Sources of funding

This work was supported by the American Heart Association (grant 18POST3408033; recipient: Ahmed AlBadri). Research reported in this publication was supported by the National Cancer Institute under award numbers - R01CA249992 - 01A1, R01CA257612 - 01A1, R01CA264017 - 01, R01CA268287 - 01A1, U01CA113913 - 16A1, U01CA269181 - 01, U24CA274494–01; U54CA302465–01; R01CA268207–01; the National Heart, Lung and Blood Institute under award numbers - R01HL158071 - 01A1; the National Institute of Allergy and Infectious Diseases under award number - R01AI175555 - 01A1 and R01HL151277 - 01A1, the National Institute of Biomedical Imaging and Bioengineering under award number 75N92022D00015; the National Library of Medicine under award number - R01LM013864 - 01A1; the National Institute on Aging under award number - R01AG089759; the National Institute of Diabetes and Digestive and Kidney Diseases under award number - R01DK118431; the Office of the Director, National Institutes of Health under award number 1OT2OD038065–01; the Kidney Mapping and Atlas Project (KMAP) under award number - U01DK133090 - 01; the United States Department of Veterans Affairs VA Merit Review award under award number - IBX004121; the VA Biomedical Laboratory Research and Development Service under award numbers - I01CX002622, I01CX002776, IK6BX006185 the VA Research and Development Office through the Lung Precision Oncology Program - LPOP-L0021; the Advanced Research Projects Agency for Health (ARPA-H) under award number D25CA00140–00; and sponsored research agreements from Astrazeneca, Bristol Myers Squibb, the Prevent Cancer Foundation, Innovation in Cancer Informatics, the Institute for Technology in Healthcare, the Breast Cancer Research Foundation and the Scott Mackenzie Foundation. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health, the U.S. Department of Veterans Affairs, or the United States Government. The funder had no role in study design, data collection, analysis, interpretation, or manuscript preparation.

Ethical considerations

This study was approved by the Emory University Institutional Review Board (STUDY2025P010279). All procedures were performed in accordance with institutional guidelines and the Declaration of Helsinki. The external reference cohort was drawn from the publicly available RASTA OCTA dataset.

Data availability

The INOCA cohort data analyzed in this study contain protected health information and cannot be shared publicly. The RASTA OCTA dataset is available. Analysis code is available from the corresponding author upon a reasonable request.

CRediT authorship contribution statement

Dharini Raghavan: Writing – original draft, Visualization, Validation, Methodology, Formal analysis. Sudeshna Sil Kar: Writing – review & editing, Methodology. Sakshi Shiromani: Writing – review & editing, Data curation, Conceptualization. Ahmed AlBadri: Writing – review & editing, Funding acquisition, Data curation. Nieraj Jain: Writing – review & editing, Visualization, Validation, Data curation. Puja K. Mehta: Writing – review & editing, Validation, Supervision, Investigation, Data curation. Gourav Modanwal: Writing – review & editing, Supervision, Methodology, Investigation, Conceptualization. Anant Madabhushi: Writing – review & editing, Supervision, Project administration, Investigation, Funding acquisition.

Declaration of competing interest

The authors declare the following financial interests/personal relationships which may be considered as potential competing interests:

Ahmed Albadri reports financial support was provided by American Heart Association Inc. Dr. Madabhushi is a Research Career Scientist at the Atlanta Veterans Affairs Medical Center. Dr. Madabhushi is an equity holder in Picture Health, Elucid Bioimaging, and Inspirata Inc. Currently he serves on the advisory board of Picture Health. He currently consults at Johnson & Johnson. He also has sponsored research agreements with AstraZeneca and Bristol Myers-Squibb. His technology has been licensed to Picture Health and Elucid Bioimaging. All other authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.