Deep Capillary Nonperfusion on OCT Angiography Predicts Complications in Eyes with Referable Nonproliferative Diabetic Retinopathy

Abstract

Objective:

To evaluate the ability of capillary non-perfusion parameters on optical coherence tomography angiography (OCTA) to predict the development of clinically significant outcomes in eyes with referable non-proliferative DR (NPDR).

Design:

Prospective longitudinal observational study

Subjects:

59 patients (74 eyes) with treatment-naïve moderate and severe (referable) NPDR

Methods:

Patients were imaged with OCTA at baseline and then followed for 1 year. We evaluated two OCTA capillary nonperfusion metrics, vessel density (VD) and geometric perfusion deficits (GPD), in the superficial (SCP), middle (MCP), and deep capillary plexuses (DCP). We compared the predictive accuracy of baseline OCTA metrics for clinically significant DR outcomes at one year.

Main Outcome Measures:

Significant clinical outcomes at 1 year, defined as one or more of the following—vitreous hemorrhage, center-involving diabetic macular edema, and initiation of treatment with pan-retinal photocoagulation or anti-VEGF injections.

Results:

49 patients (61 eyes) returned for 1-year follow-up. GPD and VD in the MCP and DCP correlated with clinically significant outcomes at 1 year (p < 0.001). Eyes with these outcomes had lower VD and higher GPD, indicating worse nonperfusion of the deeper retinal layers, compared to those that remained complication-free. These differences remained significant (p = 0.046 to < 0.001) when OCTA parameters were incorporated into models that also considered sex, baseline corrected visual acuity, and baseline DR severity. Adjusted ROC curve for DCP GPD achieved an AUC of 0.929, with sensitivity of 89% and specificity of 98%. In a separate analysis focusing on high-risk PDR outcomes, MCP and DCP GPD and VD remained significantly predictive with comparable AUC and sensitivities to the pooled analysis.

Conclusions:

Evidence of deep capillary nonperfusion at baseline in eyes with clinically referable NPDR can predict short-term DR complications with high accuracy, suggesting that deep retinal ischemia has an important pathophysiologic role in DR progression. Our results suggest OCTA may provide additional prognostic benefit to clinical DR staging in high-risk eyes.

Precis

Deep capillary nonperfusion on optical coherence tomography angiography (OCTA) predicts short-term complications of diabetic retinopathy (DR) in eyes with clinically referable non-proliferative DR. OCTA may provide additional prognostic benefit to clinical staging in high-risk eyes.

Introduction

Diabetic retinopathy (DR), a leading cause of blindness in adults, affects over 150 million worldwide.¹ Vision-threatening complications include diabetic macular edema (DME) and proliferative DR (PDR).² The ability to predict DR progression and/or complications in an individual patient is an important unmet clinical need, as early identification of individuals at high risk of progression is critical to follow these subjects closer and administer early interventions before vision loss. Traditional risk factors associated with DR progression and DME³ include hemoglobin A1c and duration of diabetes, but these factors only account for about 11% of variance in DR risk.⁴ Clinical staging systems based on fundoscopy, like the Early Treatment of Diabetic Retinopathy Study (ETDRS) criteria⁵ and the related International Clinical Diabetic Retinopathy (ICDR) Disease Severity Scale,⁶ allow clinicians to stratify patients into stages that also correlate with DR progression and DME risk.^7,8 Patients with clinically referable non-proliferative DR (NPDR), defined as moderate or severe NPDR, are a particularly vulnerable group with a higher 1-year risk of progressing to PDR (moderate: 12%–27%; severe: 52%–75%) compared to less severe stages.⁷ However, conventional imaging for DR risk stratification, including fundus photography and fluorescein angiography (FA), has significant limitations. Although FA has historically been the gold standard for assessing retinal perfusion, it yields two-dimensional images that fail to visualize the deeper retinal capillaries, and dye leakage from compromised vessels may actually mask nonperfusion.^9,10 Compared to FA, fundus photos do not capture smaller retinal vessels. Two FDA-approved models (IDx-DR and EyeArt) sought to automate photographic DR staging using artificial intelligence, achieving > 85% sensitivity and > 90% specificity for detecting referable NPDR.^11,12 These models have several caveats, including their low sensitivity for DME and the fact that they do not actually identify patients who are at higher risk for DR progression.¹³ Smaller-scale photographic models for predicting DR progression have been proposed but achieved poor accuracy.¹⁴

Optical coherence tomography angiography (OCTA) provides non-invasive, depth-resolved visualization of discrete retinal capillary layers. Many cross-sectional studies have shown correlations between OCTA vascular parameters and DR severity,^15,16 suggesting potential utility for OCTA in staging DR. More recently, geometric perfusion deficit (GPD) has been proposed as a potentially more reliable OCTA proxy for nonperfusion than the more commonly used vessel density (VD) parameter. GPD defines pathologically ischemic zones based on the theoretical threshold for lateral diffusion of oxygen, namely 30 microns.¹⁷ Previous studies have not directly compared performance of GPD to other OCTA nonperfusion metrics such as nonperfusion area or intercapillary distance. However, because GPD defines ischemia based on a set distance from the center of vessels, it has the distinct advantage of being resilient to artifactual capillary breaks and variations in vessel diameter.

Few studies have directly investigated the relationship between baseline OCTA metrics and clinical outcomes of DR, which remains an important question. Several earlier studies have suggested that deep capillary foveal avascular zone (FAZ) area and nonperfusion correlate with DR progression,¹⁸ treatment requirement,¹⁹ and visual acuity decline.²⁰ However, these studies combined treated and untreated eyes, did not consider a uniform stage of DR, and only one study¹⁸ adjusted for the effects of baseline DR severity; thus, it is unclear whether their predictive models were entirely independent of clinical staging and treatment status, an important consideration. Furthermore, they used segmentation schemes that combined multiple capillary layers,^18,20 potentially missing subtle changes in individual layers, or used non-standard nonperfusion metrics that limit their generalizability.¹⁹ In this study, we explored the relationship between baseline OCTA parameters and clinically significant DR outcomes at 1 year in a cohort of treatment-naïve patients with referable NPDR at baseline. We analyzed the individual macular capillary plexuses—superficial (SCP), middle (MCP), and deep (DCP)—to quantify nonperfusion. We hypothesized that patients who experience vision-threatening complications during follow-up would have a higher burden of baseline ischemia characterized by lower VD and higher GPD.

Materials and methods

This prospective longitudinal study identified patients with DR who underwent OCTA imaging between June 2015 and July 2021 at the Department of Ophthalmology at Northwestern University in Chicago, Illinois. The study was approved by the Institutional Review Board of Northwestern University and conducted in accordance with the tenets of the Declaration of Helsinki and the regulations of the Health Insurance Portability and Accountability Act. Written informed consent was obtained from all subjects.

Inclusion criteria were age between 25 and 75 years, diagnosis of NPDR by a board-certified retina specialist, and eyes that were treatment-naïve at baseline (no history of pan-retinal photocoagulation (PRP), focal laser, intravitreal injection, or pars plana vitrectomy). We used ultra-widefield fundus photographs (Optomap Panoramic 200; Optos PLC, Scotland, U.K.) to confirm the baseline diagnosis of referable NPDR based on the ICDR Disease Severity Scale. Briefly, the ICDR scale categorizes NPDR as mild, moderate, and severe. Severe NPDR is characterized by four quadrants of hemorrhage or microaneurysm, two quadrants of venous beading, or one quadrant of intra-retinal microvascular abnormality exceeding standard reference photographs. Moderate NPDR is defined as hemorrhage or microaneurysm equal to or exceeding standard reference photographs, or the presence of exudates or vascular abnormalities not otherwise meeting severe NPDR criteria.⁶ Fundus photographs were staged by two independent graders (J.X.O. and H.F.), with disagreements decided by a third senior grader (A.A.F).

We excluded eyes with refractive error greater than 6.0 diopters, astigmatism greater than 3.0 diopters, significant media or lens opacities, previous retinal surgery, and other retinal or anterior-segment disease. Central macular thickness was automatically calculated for each eye as the mean thickness of the central 1 mm subfield centered on the fovea on Spectralis OCT (Heidelberg Engineering Inc., Heidelberg, Germany). To reduce segmentation errors and other imaging artifacts, we also excluded eyes with center-involving DME at baseline, defined as a central macular thickness greater than 320 μm for men and 305 μm for women.²¹

Patients underwent a comprehensive history and ophthalmologic exam at baseline and at 1-year follow-up. Baseline OCTA images were de-identified and the treating providers did not have access to the images during the study period. The study design was observational, and these patients were managed per standards of clinical care.²² Events assessed as vision-threatening complications at 1 year were initiation of treatment with PRP, focal laser, or anti-vascular endothelial growth factor (anti-VEGF) intravitreal injections, new vitreous hemorrhage, and development of center-involving DME. We defined the main outcome metric of “clinically significant outcomes” as occurrence of one or more of these events. We also recorded whether patients developed PDR on clinical exam at 1-year follow-up.

OCTA imaging

OCTA images were obtained using the RTVue-XR Avanti system (Optovue Inc., Fremont, California, USA) with split-spectrum amplitude-decorrelation angiography (SSADA) algorithm (version 2017.1.0.151).²³ Briefly, this system captures two consecutive B-scans, each containing 304 A-scans, over a 3 mm × 3 mm region centered on the fovea and extracts angiographic flow data with the SSADA algorithm. The A-scan rate was 70,000 scans/s, using a light source centered on 840 nm with a bandwidth of 45 nm. Projection artifacts were removed using the built-in Optovue software. We excluded images with quality index (Q-score) < 6 or signal strength index (SSI) ≤ 50. We also excluded images with significant motion, fixation, or shadow artifacts.

We segmented the full retinal slab from the inner limiting membrane (ILM) to 10 μm below the outer plexiform layer (OPL). The SCP was segmented from the ILM to 10 μm above the inner plexiform layer (IPL), the MCP from 10 μm above to 30 μm below the IPL, and the DCP from 30 μm below the IPL to 10 μm below the OPL.¹⁶ Segmentations were reviewed for errors and corrected manually as needed.

OCTA nonperfusion parameters

Image analysis was performed using FIJI, an open-source distribution of the program ImageJ.²⁴ The FAZ was traced from the full retinal slab and the FAZ area and circularity index, defined as the ratio of FAZ perimeter length to that of a circle of equivalent area, were calculated.²⁵ The SCP, MCP, and DCP OCTA slabs were binarized using two methods that we have previously described^15,16 and we calculated the intersection of both binarizations to improve noise removal. VD was calculated as percentage of the parafovea, defined as the annulus between concentric 1 mm- and 3 mm-diameter circles centered on the fovea, occupied by blood vessels. Vessel length density (VLD), which removes the effect of large SCP vessels, was calculated as ratio of vessel length in mm to total area in mm².¹⁶

We calculated GPD for each layer by modifying the approach from Chen et al.¹⁷ and implementing an automated FIJI macro to accomplish the following steps. We despeckled the binarized slabs by eliminating single noise pixels that could artificially appear like vessels. Vessels that became fragmented in despeckling were reconnected through morphologic operations, and the reconnected image was re-binarized using Mean thresholding. The binarized image was skeletonized and GPD was calculated as percentage of area >30 μm from the nearest vessel excluding the FAZ, as previously described.¹⁷

To assess the repeatability of OCTA quantification, a masked grader (N.J.K.) evaluated VD and GPD of the SCP, MCP, and DCP in a random subset of 20 eyes.

Statistical analysis

Statistical analysis was performed using IBM SPSS Statistics version 26 (IBM SPSS Statistics; IBM Corporation, Chicago, IL, USA) and R (R 4.1.1.; R Project for Statistical Computing, Vienna, Austria). Shapiro-Wilk tests were performed to determine whether continuous data were normally distributed. Demographic and clinical data were compared with independent-sample t-tests for normally distributed data, Mann-Whitney U-tests for non-parametric data, and Fisher’s exact tests for categorical data. We calculated absolute agreement intraclass correlation coefficients (ICC) to determine inter-rater reliability.

Receiver operating characteristic (ROC) curves were calculated for each OCTA parameter and used to determine sensitivity and specificity. We used DeLong’s method for non-parametric data for pairwise comparisons of ROC curves.²⁶ For deep capillary OCTA nonperfusion parameters, we performed logistic regression models adjusted for sex, baseline corrected visual acuity (VA), baseline NPDR severity, and image quality measured by Q-score as covariates and used generalized estimating equations (GEE) to account for the contribution of two eyes of the same patient. Because the proportion of eyes with and without clinically significant outcomes (30% and 70%, respectively) were unequal, we also calculated precision-recall (PR) curves for each ROC curve to account for potential effects of imbalanced group size when comparing model performance. Two-tailed p-values of 0.05 were considered statistically significant for all tests.

Results

Of the 181 patients with NPDR who underwent OCTA imaging within the specified time frame, 59 patients (74 eyes) met baseline inclusion and exclusion criteria. 49 of these 59 patients (61 eyes) returned for 1-year follow-up. There were no significant differences in demographic, clinical, and baseline OCTA parameters apart from FAZ circularity (p = 0.004) between returning patients and those lost to follow-up (Supplemental Table 1; available at https://www.ophthalmologyretina.org/).

At 1 year, 18 eyes experienced one or more of the predefined clinically significant outcomes. 6 eyes developed new vitreous hemorrhage and were treated with either PRP alone (4 eyes) or PRP and anti-VEGF injection (2 eyes). 8 eyes received PRP for other high-risk PDR clinical characteristics, defined as neovascularization equal or greater than 1/3 of disc area on or within 1 diameter of the disc as per the Diabetic Retinopathy Study.²⁷ 5 eyes developed center-involving DME, of which 3 received anti-VEGF alone and 1 received both anti-VEGF and PRP. No eyes received focal laser, while 1 eye with center-involving DME was not treated because the eye had good visual acuity. Eyes that progressed to clinically significant outcomes were more likely to belong to male patients (p = 0.023) and had lower corrected VA at baseline (p = 0.004) but did not otherwise differ in demographic or clinical characteristics (Table 1). Eyes that developed clinically significant outcomes also had lower image Q-score (p = 0.035) and SSI (p = 0.015).

Table 1.

Demographic, clinical, and imaging characteristics of study patients

Subject characteristics	No complications (N = 35; 43 eyes)1	Complications (N = 16; 18 eyes)1	p
Age (y), mean ± SD	54.3 ± 11.5	51.9 ± 10.0	0.467
Male/female (% female)	10/25 (71)	11/5 (31)	0.023*
DM type, type I/type II (% Type 1)	8/27 (23)	2/14 (13)	0.474
DM duration (y), mean ± SD	19.5 ± 10.9	18.1 ± 8.9	0.760
HbA1c (%), mean ± SD	8.17 ± 1.88	9.04 ± 1.46	0.087
Insulin-dependent, no/yes (% insulin-dependent)	8/27 (77)	1/15 (94)	0.242
NPDR severity, moderate/severe (% severe)	23/20 (47)	5/13 (72)	0.093
CMT (μm), mean ± SD	265 ± 23	277 ± 30	0.132
HTN, no/yes (% yes)	13/22 (63)	5/11 (69)	0.803
Serum Cr, mean ± SD	1.01 ± 0.33	0.94 ± 0.25	0.646
Baseline VA (LogMAR), mean ± SD	0.084 ± 0.216	0.181 ± 0.212	0.004†
Refractive error (D), mean ± SD	−1.27 ± 1.86	−0.69 ± 1.75	0.334
Duration of follow-up (mo), mean ± SD	12.8 ± 2.4	12.9 ± 2.1	0.908
Complication/treat (%)
Vitreous hemorrhage	-	6 (33)
Central DME	-	5 (28)
PRP	-	14 (78)
Anti-VEGF	-	5 (28)
New PDR at follow-up (%)	1 (2)	16 (89)
High-risk characteristics	0 (0)	15 (83)
Q-score, mean ± SD	7.70 ± 0.86	7.17 ± 0.92	0.035†
Signal strength index, mean ± SD	66.1 ± 6.4	61.2 ± 6.8	0.015‡

^*Statistically significant at the 0.05 level (two-tailed) by Fischer’s Exact test

^†Statistically significant at the 0.05 level (two-tailed) by Mann-Whitney U test

^‡Statistically significant at the 0.05 level (two-tailed) by independent 2-sample t-test

Abbreviations: CMT = central macular thickness; Cr = creatinine; DM = diabetes; DME = diabetic macular edema; HTN = hypertension; NPDR = non-proliferative diabetic retinopathy; PDR = proliferative diabetic retinopathy; PRP = pan-retinal photocoagulation; SD = standard deviation; VA = corrected visual acuity

¹Two patients had one eye that experienced diabetic complications and one that did not and were analyzed as part of both groups.

OCTA parameters showed excellent repeatability, with ICCs > 0.9 for all parameters, ranging from 0.902–0.978 for VD and VLD, and 0.916–0.983 for GPD.

Univariate ROC analysis of baseline OCTA parameters identified VD and GPD in the SCP, MCP, and DCP, as well as SCP VLD (all p < 0.030) as predicting clinically significant outcomes (Table 2, Fig. 1). Overall, eyes with clinically significant outcomes had lower VDs and higher GPDs at baseline than those without. Notably, DCP GPD and VD achieved AUCs of 0.885 and 0.873, respectively (all p < 0.001). A DCP GPD cutoff of > 12.8% maximized sensitivity at 89% and specificity at 84%, while a DCP VD cutoff of < 28.5% yielded a maximum sensitivity of 89% and specificity of 79% (Fig. 2).

Table 2.

Univariate ROC analysis of OCTA parameters for the entire study population

OCTA parameters	No complications (N = 35; 43 eyes)	Complications (N = 16; 18 eyes)	AUC (95% CI)	SN (%)	SP (%)	p
FAZ area (mm2), mean ± SD	0.371 ± 0.155	0.438 ± 0.148	0.621 (0.475 – 0.766)	100	26	0.103
FAZ circularity, mean ± SD	0.493 ± 0.142	0.479 ± 0.133	0.530 (0.372 – 0.687)	78	37	0.712
GPD (%), mean ± SD
SCP	16.83 ± 6.53	21.37 ± 7.15	0.669 (0.518 – 0.820)	44	91	0.028*
MCP	8.06 ± 3.68	13.02 ± 4.45	0.820 (0.700 – 0.940)	78	81	< 0.001*
DCP	10.22 ± 4.62	18.87 ± 6.17	0.885 (0.781 – 0.989)	89	84	< 0.001*
VD (%)
SCP	34.21 ± 5.44	30.35 ± 6.41	0.689 (0.534 – 0.843)	72	63	0.017*
MCP	35.65 ± 4.79	29.89 ± 4.25	0.822 (0.699 – 0.945)	67	91	< 0.001*
DCP	31.71 ± 4.66	24.61 ± 4.73	0.873 (0.764 – 0.982)	89	79	< 0.001*
SCP VLD (mm−1), mean ± SD	14.16 ± 2.35	12.58 ± 2.44	0.672 (0.516 – 0.827)	56	79	0.030*

^*Statistically significant at the 0.05 level

Abbreviations: CI = confidence interval; DCP = deep capillary plexus; FAZ = foveal avascular zone; GPD = geometric perfusion deficit; MCP = middle capillary plexus; OCTA = optical coherence tomography angiography; SCP = superficial capillary plexus; SD = standard deviation; SN = sensitivity; SP = specificity; VD = vessel density; VLD = vessel length density

Figure 1. Determination of geometric perfusion deficits (GPD) from OCTA images.

Columns from left to right represent superficial (SCP), middle (MCP), and deep capillary plexuses (DCP). GPDs (red areas) were determined as areas greater than 30 μm away from the center of the nearest blood vessel, excluding the foveal avascular zone as it represents a physiologic area of nonperfusion.

Figure 2. Cumulative frequency plots of deep capillary plexus (DCP) geometric perfusion deficits (GPDs) and vessel densities (VD) in patients with and without clinically significant outcomes at 1 year.

Gray dots indicate patients with clinically significant outcomes (“Complications”); black dots are those without (“No complications”). Each point corresponds to the percentile rank for the OCTA parameter value in the specified population. Red lines indicate the cutoff values for the OCTA parameters that maximized sensitivity and specificity, based on Youden’s index. (A) Plot for DCP GPD. (B) Plot for DCP VD.

Paired-sample analyses showed ROC curves for GPD and VD in the MCP and DCP were significantly more predictive (p = 0.027 to 0.006) than their corresponding SCP parameters (Fig. 3A–B). There were no significant differences between ROC curves for MCP vs. DCP GPD, MCP vs. DCP VD, and no significant differences between GPD vs. VD in each layer (all p > 0.05). PR curves showed similar trends, with DCP and to a lesser extent MCP parameters achieving greater areas under the PR curve (AUPRC) than their corresponding SCP parameters (Fig. 3C–D), supporting the conclusions from the ROC analysis.

Figure 3. OCTA vessel parameters in middle (MCP) and deep capillary plexuses (DCP) are more predictive of 1-year clinically significant outcomes than superficial (SCP) parameters.

ROC curves for each parameter are presented in the top row, with corresponding precision-recall (PR) curves in the bottom row. (A and C) ROC and PR curves for geometric perfusion deficits in superficial (SCP; red), middle (MCP; blue), and deep capillary plexuses (DCP; black). (B and D) ROC and PR curves for vessel densities in SCP (red), MCP (blue), and DCP (black). Reference lines assuming random assignment are indicated as gray dashed lines. P-values indicate pairwise comparisons of ROC curves with Bonferroni correction for multiple comparisons. AUC = area under ROC curve; AUPRC = area under PR curve; SN = sensitivity; SP = specificity.

Eyes with clinically significant outcomes also tended to have larger FAZ area and lower circularity at baseline, but these parameters did not achieve significance.

We also assessed whether MCP and DCP nonperfusion parameters identified from univariate analysis remained significant in multivariate models after adjusting for the contribution of both eyes of the same patient and incorporating demographic and clinical variables, particularly sex, baseline corrected VA, and image quality, which were significantly different between groups. We also incorporated baseline NPDR severity into each model. Although baseline NPDR severity in our sample did not differ significantly between eyes with and without clinically significant outcomes, we chose to include this parameter because of its plausible association with developing DR complications. In the multivariate models, all OCTA parameters remained significant (p = 0.046 to < 0.001) for predicting complications and treatment requirement (Supplemental Table 2; available at https://www.ophthalmologyretina.org/). AUCs ranged from 0.866 to 0.929, with the DCP parameters having higher specificities (95+% vs. 86+%) and sensitivities (89% vs. 78%) compared to MCP parameters (Fig. 4). Although AUCs and AUPRCs were higher for DCP than MCP parameters, these differences remained non-significant, as did GPD vs. VD within each layer (all p > 0.05).

Figure 4. After adjusting significant OCTA parameters for sex, baseline corrected VA, and baseline NPDR severity, models show improved accuracy in predicting clinically significant outcomes.

ROC curves for each parameter are presented in the top row, with corresponding precision-recall (PR) curves in the bottom row. Columns correspond to each OCTA parameter, with middle capillary plexus (MCP) geometric perfusion deficits (GPD), deep capillary plexus (DCP) GPD, MCP vessel density (VD), and DCP VD from left to right. AUC = area under ROC curve; AUPRC = area under PR curve; SN = sensitivity; SP = specificity.

We further explored the ability of OCTA parameters to predict complications specifically associated with progression to PDR. Of the 18 eyes that experienced clinically significant outcomes, 15 were noted to have PDR with high-risk features on follow-up. Of the 5 eyes that developed center-involving DME, 2 eyes had concurrent high-risk PDR outcomes (vitreous hemorrhage), and 1 eye was noted to have PDR but without high-risk features and was therefore not treated. Overall, analysis of high-risk PDR outcomes revealed similar results to our pooled outcomes analysis. MCP and DCP GPD and VD remained significantly predictive of PDR-only outcomes on univariate analysis (all p < 0.001; Supplemental Table 3; available at https://www.ophthalmologyretina.org/) and in adjusted models (all p < 0.037; Supplemental Table 4 and Supplemental Figure 1; available at https://www.ophthalmologyretina.org/). DCP GPD and VD achieved comparable AUCs (univariate: 0.848 and 0.836; multivariate: 0.909 and 0.906) and sensitivities (all 87%) for predicting high-risk PDR outcomes as for pooled clinically significant outcomes. There were no significant differences between the ROC curves generated for pooled clinically significant outcomes vs. high-risk PDR outcomes (all p < 0.05).

Discussion

In this prospective longitudinal study, we investigated the relationship between baseline OCTA parameters and the development of DR complications in treatment-naive eyes. We demonstrate that deep capillary (MCP and DCP) nonperfusion metrics were significantly different in eyes that went on to develop clinically significant outcomes at 1 year (Table 2). At baseline, these eyes had worse perfusion in the deeper capillaries than those that remained complication-free (Fig. 1). These differences in OCTA parameters remained significant in multivariate models incorporating the contributions of sex, baseline corrected VA, baseline DR severity, and image quality, which improved the overall model specificity especially for the DCP (Fig. 4). We also showed that deep capillary nonperfusion metrics retained similar predictive power when limited to high-risk PDR outcomes.

By excluding previously treated eyes, evaluating all three macular capillary plexuses using two different metrics of nonperfusion, and adjusting for baseline DR severity, our study rigorously addresses an important unmet need. Earlier longitudinal studies did not distinguish eyes with respect to their treatment status^18–20 and enrolled eyes with a wide range of baseline DR severity without correcting for DR stage,^19,20 limiting their ability to assess the prognostic capability of OCTA parameters independent of baseline DR severity. Another strength of our study is the focus on clinically referable (moderate and severe) NPDR eyes, a high-risk population that has been poorly represented in previous longitudinal studies.^18–20 While referable NPDR eyes are more likely to develop DR complications, these patients, especially those who are treatment-naïve, do not typically receive as frequent follow-up as patients who have already experienced complications or developed PDR.²² We believe these eyes represent a key population where DR prognostic tools may be especially useful in developing personalized clinical algorithms facilitating timely follow-up and treatment of complications.

Interestingly, deep capillary nonperfusion appears to predict a variety of DR complications and outcomes. Our results are consistent with longitudinal studies where DCP nonperfusion predicted worsening by two or more stages on the ETDRS scale,¹⁸ treatment requirement,¹⁹ and loss of visual acuity.²⁰ A recent study by our group found that DCP nonperfusion showed high specificity and sensitivity for identifying eyes with clinically referable DR in a sample of 150 eyes ranging from diabetes without DR to PDR, irrespective of treatment status.²⁸ These results suggest DCP nonperfusion may also be useful for baseline screening of DR to identify high risk eyes. The association of deep capillary nonperfusion with multiple aspects of DR progression suggests this metric is a robust indicator for eyes at high risk of short-term complications.

It is intriguing to speculate on the unique role of the deep capillaries in prognosticating DR complications. Histologic studies suggest the DCP may show earlier signs of DR vascular damage such as microaneurysms and vessel hyalinization.²⁹ Retinal oxidative stress and ischemia lead to upregulation of VEGF, an angiogenic factor associated with development of DME and PDR.^30,31 Experimental animal studies have suggested that photoreceptor cells may be the main source of reactive oxygen species contributing to oxidative stress in diabetic eyes.³² Oxidative stress and inflammation also promote capillary degeneration and formation of DR vascular lesions.³³ The DCP, as the closest layer to the photoreceptors, may be the most susceptible to photoreceptor-related oxidative stress, linking DCP compromise to VEGF-induced DME- and PDR-associated complications. Alternately, DCP nonperfusion may itself drive outer retinal ischemia and progression to DME and PDR. The outer retina, as a watershed zone between the retinal capillary and choroidal circulations, is more vulnerable to ischemia resulting from diabetic vascular insult to the deep capillary layers.³⁴ Since the DCP partially contributes to the metabolic needs of the photoreceptor layer, especially during dark adaptation, DCP ischemia from capillary closure in DR may compound photoreceptor metabolic stress, further upregulating VEGF and increasing DR complication risk in a vicious cycle.^34,35

Because fundoscopy and FA do not capture the deep vasculature, our findings suggest OCTA may have a distinct and important role in DR. We are encouraged that the predictive power of deep capillary nonperfusion parameters, particularly in the DCP, was preserved in adjusted and unadjusted prediction models, suggesting DCP GPD and VD may be practically useful and robust indicators independent of DR severity. From our data, we propose a DCP GPD cutoff of > 12.8% and DCP VD cutoff of < 28.5%, which achieved good sensitivity and specificity (89% and up to 98%, respectively) in identifying eyes that developed clinically significant outcomes. Although previous longitudinal studies did not report sensitivity and specificity,^18–20 precluding direct comparison with our results, our values are comparable to those achieved by conventional screening tools for referable DR, such as color fundus photographs evaluated by manual or automatic graders.^36,37 However, these screening tools only detect referable DR at baseline, and unlike our current study, do not directly predict the course of DR. We believe OCTA may assist short-term risk stratification of eyes already identified as high-risk through clinical staging.

Notably, we found no significant difference between GPD and VD. Originally, we hypothesized that GPD would outperform VD, as GPD has been shown to have lower inter-image variability because it is less susceptible to artifactual variations in vessel diameter.¹⁷ As VD is a more commonly understood and available OCTA metric, VD may be the preferable metric in a clinical setting. Of note, because we performed our calculations of GPD and VD using FIJI, the cutoff values we obtained may not be interchangeable with the machine-determined values from the proprietary software. Previously, our group found that automated VD calculated by the AngioVue (Optovue Inc., Fremont, California, USA) software overestimates DCP VD in eyes with substantial nonperfusion in more severe DR, which prompted us to develop an alternate binarization method.¹⁶ Therefore, efforts to optimize existing non-perfusion tools built into the OCTA software, as well as potentially incorporating a physiologically relevant metric of ischemia, such as the GPD, may facilitate the clinical application of OCTA metrics in prognosticating DR. Future studies in larger cohorts will also be necessary to further validate these parameters across different imaging platforms as well as refine the specific cutoff values for risk-stratifying eyes.

We found no significant correlation of FAZ area or circularity with clinically significant outcomes in our cohort. We expected that larger FAZ area and lower circularity would predict clinically significant outcomes.^18,20 It important to note that previous studies evaluated SCP and DCP FAZ areas separately, finding DCP FAZ area to predict DR progression and visual acuity change.^18,20 In contrast, we measured the FAZ not from individual layers, but from the full retinal slab, which is more robust, physiologically accurate, and less susceptible to segmentation errors.³⁸ Several factors may contribute to the relatively poor correlation of FAZ metrics with DR clinical outcomes. The FAZ area shows substantial variation in healthy individuals at baseline.³⁹ Additionally, FAZ areas measured in different capillary plexuses likely do not represent valid physiologic entities, as the trilaminar macular capillaries collapse into a single vascular layer directly surrounding the FAZ.³⁸ The apparent predictive value of the “DCP FAZ” in previous studies may indeed be a reflection of peri-foveal DCP capillary dropout, in line with our results.

Limitations of our study include the smaller sample size, which prompted us to consider a composite outcome of high-risk complications. Although VEGF upregulation is seen in both PDR and DME, the exact pathophysiologic mechanisms are thought to differ, with PDR resulting from progressive retinal ischemia and angiogenesis, while DME reflects the consequences of breakdown of the blood-retinal barrier.² In a separate analysis of high-risk PDR outcomes, we found deep capillary nonperfusion parameters remained significantly predictive in both unadjusted and adjusted models with comparable AUC and sensitivity, which further supports the potential role of OCTA in prognosticating DR progression. However, due to the smaller sample size of the center-involving DME sub-group, as well as the substantial overlap between DME and PDR in this group, analysis of DME-only outcomes was not possible in this study. Separate investigation of DME-related complications and validation with a larger dataset would be important future studies. While patients were seen at the same center and received treatment according to standard of care, they were seen by different providers, introducing the possibility of inter-clinician variation. Although we included only referable eyes and incorporated the effects of DR staging into our statistical models, we staged DR severity using the ICDR Disease Severity Scale, which does not distinguish the severe and very severe NPDR groups defined by the ETDRS criteria.⁵ Future studies with larger datasets may allow incorporation of the ETDRS scale to better stratify such eyes. We excluded eyes with center-involving DME at baseline to reduce the effects of edema on OCTA image segmentation; however, this criterion may still miss smaller cystoid spaces that could distort segmentation and artifactually reduce flow signals.⁴⁰ We also focused on short-term follow-up and clinical outcomes, leaving the long-term predictive value of baseline OCTA parameters an open question for future studies.

In conclusion, we found that deep capillary nonperfusion was significantly predictive of 1-year complications in eyes with treatment-naïve, clinically referable NPDR. Our findings highlight the importance of deep capillary nonperfusion as an independent risk factor for diabetic complications and suggest a distinct role for OCTA in DR prognostication alongside conventional imaging.