Instrumenting Carotid Sonography Biomarkers and Polygenic Risk Score As a Novel Screening Approach for Retinal Detachment

Abstract

Purpose: Retinal detachment (RD) is a vision-threatening condition that manifests silently before abrupt disease onset; thus, most of the RD at-risk individuals are left unchecked until the first RD attack.

Methods: To establish an RD risk–informing system for a broader population, we utilized carotid ultrasonography (CUS) biometrics, RD polygenic risk score (PRS_RD), and clinical covariates (COVs) to assess RD risk predisposition factors. First, a backpropagation logistic regression model identified RD-associated CUS biomarkers and further incorporated them as a multivariable RD-risk nomogram. Next, a PRS_RD model was established with the selected single-nucleotide polymorphisms (SNPs) curated as high functional expression candidates in the retina single-cell RNA datasets. Finally, a three-component RD prediction model (CUS, PRS_RD, and COVs) was assembled by logistic cumulative analysis.

Results: Demographic analysis reported hypertension (HTN) status was associated with RD (odds ratio [OR] = 1.601). The CUS regression model revealed that the minimum flow of the right internal carotid artery (ICA-Q_min; OR = 1.04) and the time-averaged maximum velocity of the right common carotid artery (CCA-TAMAX; OR = 1.03) were associated with increased RD risk. Notably, genome-wide association studies (GWAS) identified three significant SNPs (IGFBPL1 rs117248428, OR = 1.63; CELF2 rs56168975, OR = 1.72; and PAX6 rs11825821, OR = 1.61; P < 5.00 × 10⁻⁶) that are highly expressed at the RD border of the retinal pigment epithelium and choroid. Finally, the three-component model demonstrated state-of-the-art RD prediction (AUC_HTN+ = 0.95, AUC_HTN– = 0.93).

Conclusions: Based on instrumenting CUS images and genetic PRS_RD, we are proposing a screening method for RD at-risk patients.

Translational Relevance: Results from this study demonstrated the combination of CUS and GWAS as a cost-effective, population-wide screening framework for identifying RD at-risk individuals.

Introduction

Retinal detachment (RD), with a 1/10,000 annual population incidence,¹ is an emerging eye disease that jeopardizes visual acuity when the detachment involves the fovea centralis. Although demographic factors have been associated with increased RD risk,² tools for pre-disease risk stratification and screening protocols for RD at-risk individuals remain underdeveloped. 

RD risk assessments have been widely reported for different clinical contexts, including age, sex, hypertension,³ laterality,⁴ seasonal changes,⁵ fellow eye RD entailing a primary RD,⁶ RD recurrence rate after surgical management,^7,8 RD risk after phacoemulsification surgery,^9–11 and RD risk with an open globe ocular injury.¹² However, RD assessment for potential cases before an RD attack has been an overlooked context that may have beneficial diagnostic benefits. 

Image diagnostics such as B-scans,¹³ color fundus photography,¹⁴ and optical coherence tomography¹⁵ have been the mainstay of RD evaluation, but these ocular-specific image modules are difficult to assess and require professional interpretation; thus, these ophthalmology diagnostics can be less applicable to population-wide RD screening programs. Therefore, this study attempted to instrument generalized examination modules to screen RD at-risk individuals. 

RD is a multifactorial disease that encompasses structural, molecular, and hemodynamic pathogenic components.^16,17 Retinal ischemia and its compromised perfusion have been extensively demonstrated as a hemodynamic stress that drives exudative retinal detachment in rat¹⁸ and rabbit¹⁹ models. Similarly, choroid ischemia induced by microsphere injection in the rhesus monkeys resulted in five out of nine choroidal ischemia eyes (45%) developing RD.²⁰ Therefore, investigating the hemodynamic and perfusion status of the retina may identify novel diagnostic approaches to RD assessments.²¹ 

The carotid axis and vascular health have drawn increasing attention from a diverse group of experts,²² including vision care and cardiovascular specialists.^23,24 Deficiencies in the carotid axis encompass carotid artery stenosis and retina ischemia,²⁵ carotid stiffness and retinal microvascular damage,²⁶ and carotid plaque length and retinal microvascular density.²⁷ To further investigate the carotid axis, clinical intervention studies, such as the INFLATE study,²⁸ have reported that the mean choroid thickness was significantly thinner in the eyes with ipsilateral carotid stenosis, but the choroid thickness was reversed within 1 week after the patient received carotid endarterectomy. Aligned with these endarterectomy findings, carotid artery embolization^29,30 and carotid artery ligation³¹ have been common surgical techniques used to study acute and chronic retina ischemia in animal models. With these carotid axis perfusion coupling effects, carotid ultrasonography (CUS) can serve as a proxy to interpolate ischemic events within the eye. 

The inherited genetic component poses significant risk for RD susceptibility; an individual suffers a twofold greater lifetime RD risk if a first-degree relatives has ever been diagnosed with RD.^32,33 The genetic linkage was further supported by a UK Biobank genome-wide association studies (GWAS) study, in which the single-nucleotide polymorphism (SNP)-deducted heritability (h²) of RD was around 23%,^33,34 indicating that RD is moderately inheritable. Moreover, a series of GWASs explored RD-associated SNPs for pathognomonic investigation purposes.^33,35–37 Notably, variants such as COL2A1 rs1793958, rs9651980, and rs1635532 are all extracellular matrix mutations that have been linked to increased familial-dominant rhegmatogenous retinal detachment (RRD).^33,38 Additional RD genetic variants have also been speculated to have a collective effect in disrupting retinal layers morphology (FAT3), proteoglycan organization (VCAN), and Norrin/Frizzled4 signal-dependent blood–retinal barrier integrity (LRP5, FZD4, TSPAN12, and NDP).³⁸ These genetic-driven ontology pathways reflect our up-to-date understanding of the mechanisms of rhegmatogenous and exudative RD. Our group previously leveraged a polygenic risk score (PRS) model using 879 SNPs to predict RD from a Taiwan Biobank (TWB) cohort (area under the curve [AUC] = 0.889).³⁷ Although ample evidence supports a genetic predisposition to RD, we intended to combine PRS prediction models with multimodal information to augment RD screens for at-risk individuals. 

The PRS constructs risk information from gene variants revealed by GWASs³⁹ Given a complex genetic background of specific diseases, the PRS has been a relative sensitive method to detect an individual's genetic susceptibility by swiftly computing the additive effects of common SNPs.^40–42 However, the PRS “Common Disease, Common Variant” (CDCV) hypothesis has some intrinsic constraints regarding various aspects of the disease etiology; hence, combining individual PRSs with medical imaging evidence is a current trend for improving the precision of evaluations in challenging diagnoses and underscreened clinical conditions.^43,44 For diseases with moderate SNP-based heritability, such as RD, incorporating appropriate medical imaging modalities could be an important goal for early detection. Two prior studies,^34,37 one of which was conducted by our team, have explored RD prediction using GWASs; however, no research has integrated genetic diagnostics with imaging modalities to enhance RD prediction. 

In this study, we proposed a novel, cost-effective, pre-symptom RD risk stratification screening approach by combining CUS and PRS with hypertension (HTN) classifications. By leveraging the established role of CUS in HTN management, our dual screening strategy aims to create a less invasive but more accessible model for clinical settings compared to using PRS alone (Fig. 1). 

Figure 1.

 

Study design overview. This two-segment study to evaluate the risk of RD included a multivariable logistic regression model for CUS features (such as Q_min and TAMAX of CCAs and ICAs, respectively), COVs (diabetes mellitus, myopia, and BMI), and a PRS model for GWASs. ECA, external carotid artery.

Methods

This two-segment study to evaluate the risk of RD included a multivariable logistic regression model for CUS and a PRS model for GWASs. It was conducted in accordance with the tenets of the Declaration of Helsinki and was approved by the Institutional Review Board of Taipei Veterans General Hospital (ID no. 2023-01-006AC). 

Study Population

Approximately 180,000 individuals were enrolled in the nationwide TWB1 and TWB2 studies between 2016 and 2023 (https://www.twbiobank.org.tw/).⁴⁵ After providing written informed consent, the participants were interviewed by well-trained researchers to fill out a structured questionnaire that documented demographics, lifestyle behavior, environmental exposures, dietary habits, family history, and health-related information. Study participants were divided into subgroups according to self-reported hypertension conditions (HTN⁺ and HTN^–). 

Multivariable Logistic Regression Model for Carotid Ultrasound

The GE LOGIQ P5 ultrasound machine (GE HealthCare, Chicago, IL) was used to examine CUS metrics in the TWB study, in which 426 RD cases and 21,015 controls received carotid CUS examinations. The multivariable logistic regression model was used to train the primary prediction model by using the features of common and internal carotid arteries (CCAs and ICAs, respectively), including the presence of plaque, intima–media thickness (IMT), flow rate, pulsatility index (PI), vascular resistance index, and diameter. Continuous variables were analyzed using an independent samples t-test and categorical variables using the χ² test. Model performance among full, stepwise, and backward selection was assessed by area under the curve (AUC), Akaike information criterion (AIC), Bayesian information criterion (BIC), and residual deviance. AUC evaluated the discriminative power, AIC and BIC assessed model quality by penalizing complexity, and residual deviance measured the discrepancy between observed and predicted results. 

The predictability of selected features was further visualized by nomograms⁴⁶ based on multivariate cumulative analysis. Although logistic regression analysis is commonly employed for such predictions, its interpretation can be challenging. In contrast, nomograms offer a straightforward and visual approach to estimating event probabilities through clearly defined components represented as lines. A typical nomogram includes four primary lines: 

• • Point line—This line assigns a scale ranging from 0 to 100 points, with points corresponding to specific risk categories based on the values of the associated factors.
• • Risk factor line—This line is derived from the linear predictor values of a logistic regression model (CUS, PRS, and COV).
• • Total point line—This line represents the cumulative total of all points (combinations of CUS, PRS, and COV) and connects the total points to the probability values.
• • Probability line—This line links the total points to the predicted probability of an event (in our case, the occurrence of retinal detachment), expressed as values between 0 and 1.

All statistical analyses were performed using R 4.1.1 (R Foundation for Statistical Computing, Vienna, Austria). The tests were two tailed, and P  < 0.05 was taken as statistically significant. 

Polygenic Risk Score Model for GWASs

This GWAS focused on the linkages between SNPs and disease phenotypes (RD) in all participants with adjustment for multiple potential confounders (age and sex). SNP data were generated using a custom Axiom Genome-Wide Array Plate system, with more than 653,000 SNPs in TWB1 and 750,000 SNPs in TWB2, and were further visualized by the Manhattan plot using the R package qqman (https://github.com/stephenturner/qqman).⁴⁷ Quality control (QC) was applied to identify the following: (1) individuals missing more than 5% SNP data; (2) individuals with heterozygosity more than 3 SD; (3) individuals with identity by descent of kinship coefficient greater than 0.1875; (4) SNPs missing in more than 2% of participants; (5) SNPs deviating from Hardy–Weinberg equilibrium (P < 0.05); and (6) SNPs with minor allele frequency under 0.05. Ultimately, 18,148 independent samples (344 RD cases and 17,804 controls) and 2,584,644 SNPs that passed QC were used for subsequent analysis. All SNP discoveries were performed on PLINK 1.9 (http://www.cog-genomics.org/plink/1.9/).⁴⁸ 

PRS_RD models were constructed based on GWAS results with thresholds of hyperparameter and linkage disequilibrium (LD) to estimate an individual's genetic risk of developing RD. SNPs within 25 Mb of the index SNP were considered for clumping with the P threshold values of hyperparameters (P < 0.05, P < 0.01, P < 0.005, P < 0.0025, P < 0.001, P < 0.0001, and P < 0.00001), SNPs were screened by pairwise squared correlation (r² < 0.1 and r² < 0.01), and LD was estimated with a window size of 4000 bp. 

Gene-Set Enrichment Analysis

Gene-set enrichment analysis (GSEA) included gene expression analysis from single-cell RNA (scRNA) datasets and protein function analysis via snpXplorer. Extracting genes from the PRS_RD model SNP list with hyperparameter thresholds of P < 1 × 10⁻⁴, the gene expression analysis was queried by the Human Eye Transcriptome Atlas⁴⁹ by obtaining normalized reads in the tissue of posterior ocular segments and further visualization with the Scanpy toolkit⁵⁰ and Python 3.10.13. The protein function analysis identified the likely affected pathway with snpXplorer⁵¹ based on the combined annotation-dependent depletion (CADD) score, expression quantitative trait loci (eQTL), and the position of the variant. GSEA was also achieved using the snpXplorer portal. P < 0.05 was considered statistically significant. 

Construction of a Three-Component RD Prediction Model

A three-component RD prediction model was assembled by logistic cumulative analysis with the combination of the total point of CUS multivariate cumulative analysis (CUS risk), the final PRS_RD score, and adjustment against selected COVs, including diabetes mellitus,⁵² myopia,⁵³ and body mass index (BMI).⁵⁴ 

Results

Participant Characteristics

This study enrolled 77,090 participants (1265 RD cases and 75,825 controls) with a subset of 18,148 individuals (344 RD cases and 17,804 controls) containing complete genome-wide SNP data (2,584,644 SNPs). In agreement with previous reported RD studies, the older population (P < 0.001) and male sex (odds ratio [OR] = 1.43, P < 0.001) demographic characters were associated with higher RD incidence in our cohort (Table 1, Supplementary Table S1). 

Table 1.

 

Demographic Information for the Participants

 

Hypertension Raises the Susceptibility of RD

In addition, we analyzed underlying systemic factors and noted that self-reported HTN⁺ patients were associated with increased RD risk (OR = 1.601; P < 0.0001; 95% confidence interval [CI], 1.497–1.712) when compared to the HTN^– patients. Given that high blood pressure may affect carotid biometric patterns, we investigated the feasibility of instrumenting carotid biomarker as potential RD predictors and adjusted the presence of HTN as an independent confounder in the following RD risk stratification experiments. 

Carotid Sonography Metrics Stratified RD Risk

Univariate logistic regression analysis showed that the laterality and segmental position of CUS biomarkers conferred distinct RD risks (Table 2), which was substantiated by the finding that ICA plaque in the left but not the right side of carotid artery was associated with increase RD risk, whereas the reduced CCA normality on both sides was associated with RD risks. After subgrouping self-reported hypertension conditions (HTN⁺ and HTN^–), CCA features remained an effective RD predictor in the HTN^– group, whereas no stand-alone CUS metrics were predictive of RD risk in the HTN⁺ group. 

Table 2.

 

Univariate Logistic Regression Between Carotid Sonography Metrics and Retinal Detachment Risk

 

Table 3.

 

Backward Multivariable Logistic Regression Between Carotid Sonography Metrics and Retinal Detachment Risk

 

To further examine factor-specific predictive values, we performed full (AUC = 0.633) (Supplementary Table S2), stepwise (AUC = 0.616) (Supplementary Table S3), and backward (AUC = 0.623) (Table 3) multivariate Cox regression. As a result, the backward model derived from the overall population indicated that right CCA metrics, including the presence of plaque (OR = 0.80), minimum velocity (Q_min, OR = 0.96), time-averaged peak velocity (TAMAX, OR = 1.03), and IMT (OR = 1.35), contributed to the major predicting effects. On the other hand, left ICA metrics also played an important role in the HTN⁺ model, including Q_min (OR = 0.83), TAMAX (OR = 1.13), and PI (OR = 11.24). 

To further construct a clinically applicable scoring system, we instrumented the backward model-selected CUS biometrics to construct RD risk estimation nomograms (Figs. 2A, 2B). By adding the points associated with each CUS metric, the CUS-based RD risk was distributed between RD cases, with mean HTN⁺ (µ_HTN+) = 278, mean HTN^– (µ_HTN–) = 200, median HTN⁺ (m_HTN+) = 279, and mean HTN^– (m_HTN– = 201, and control, with µ_HTN+ = 275, µ_HTN– = 199, m_HTN+ = 275, m_HTN– = 199 (Fig. 2C). For the pure CUS predictive mode, after HTN confounder adjustment, the overall performance was AUC_HTN+ = 0.627 and AUC_HTN– = 0.632 (Figs. 2D–H). 

Figure 2.

 

Backward regression models validate the utility of CUS metrics in RD prediction. (A, B) Nomograms for RD risk prediction based on CUS metrics are presented for patients with (A) and without (B) self-reported HTN. The points of each variable were summed to obtain the total points (CUS risk), which can be transformed into the predicted value (RD risk) via the scale bar shown. (C) Density distribution for CUS risk. (D–H) Receiver operating characteristic (ROC) curves (D) for the CUS model with performance measuring AUC (E), AIC (F), BIC (G), and residual deviance (H).

Characterize the Functional Significance of RD Risk Loci

To develop a genetic model on top of the CUS prediction model, we conducted a GWAS PRS experiment from the multimodal TWB dataset. After adjusting for age and sex, 277 RD risk loci were extracted from the GWAS analysis with a set threshold of adjusted P < 1.0 × 10⁻³ (Supplementary Table S4). Three significant genomic loci were identified (Fig. 3): IGFBPL1 rs117248428 (insulin-like growth factor-binding protein-like 1, adjusted P = 3.30 × 10⁻⁷, OR = 1.630), CELF2 rs56168975 (CUGBP Elav-like family member 2, adjusted P = 7.98 × 10⁻⁶, OR = 1.716), and PAX6 rs11825821 (paired box 6, adjusted P = 9.18 × 10⁻⁶, OR = 1.608). The complete list of all 277 SNPs and their mapped functional genes is displayed in Supplementary Table S3. The statistical summary of our GWAS is accessible via GCP ID GCP000990. 

Figure 3.

 

Manhattan plot of the TWB RD GWAS. The horizontal red and blue lines represent adjusted P values of 5.0 × 10⁻⁸ and 1.0 × 10⁻⁵, respectively.

To aggregate the functional aspect of the RD SNPs, we prioritized the SNP analysis with a preprocessing expression analysis that included scRNA database expression levels of the posterior ocular segments and protein function analysis via snpXplorer. We then performed GSEA of the selected SNPs, and the results showed that SNP functional genes were highly expressed at the RD border across the retinal pigment epithelium and choroid, and previous reported RD risk loci (FAT3, COL22A1, VIPR2, and TMEM132D) were also expressed (Fig. 4). Subsequent functional SNP analysis revealed enriched pathways including cell–cell adhesion via plasma-membrane adhesion molecules, nervous system development, regulation of neuron projection development, and cell junction assembly (Fig. 5). 

Figure 4.

 

Gene expression analysis in posterior ocular segments. (A) Dotplot of mean expression. (B) Stacked violin plot of median expression and distribution. The labeled genes were also identified by previous studies. CNV, choroidal neovascularization; ILM, internal limiting membrane; RPE, retinal pigment epithelium.

Figure 5.

 

Top 20 enhanced biological pathways in protein function analysis. Protein function analysis was conducted on the online portal snpXplorer by annotating the significant SNP list of PRS_RD with the hyperparameter threshold of P < 1 × 10⁻³.

RD Polygenic Risk Score Model

To investigate the cumulative RD predicting power of the RD-derived SNPs, we established an array of PRS models using different GWAS SNP enrolling thresholds (P < 0.05, P < 0.01, P < 0.005, P < 0.0025, P < 0.001, P < 0.0001, and P < 0.00001) and LD thresholds (r² < 0.1 and r² < 0.01) based on the GWAS results (Supplementary Fig. S1). After balancing AUC performance and clinical significance, we opted for a 277-SNP (P < 1.0 × 10⁻³ and r² < 0.01) PRS condition for our final model. The predicted AUCs for RD in the HTN⁺ and HTN^– subgroups were 0.946 and 0.913, respectively (Fig. 6A). PRS scores were higher in RD cases (µ_HTN+ = 25.2, µ_HTN– = 25.2) than in controls (µ_HTN+ = 7.3, µ_HTN– = 7.7) (Fig. 6B). After refining the PRS model by adjusting RD clinical covariates, rarely did AUC performance increase in both the HTN⁺ (from 0.793 to 0.796) and HTN^– (from 0.769 to 0.768) subgroups (Supplementary Table S7 and Supplementary Fig. S2). 

Figure 6.

 

Predictive performances of the PRS_RD with different tuning parameters. (A) ROC curves with different clumping P threshold of hyperparameters. (B) Density distribution for PRS_RD in patients with (1) and without (2) self-reported HTN.

Construction of a Three-Component RD Prediction Model

Finally, we assembled a three-segment (CUS, PRS_RD, COV) logistic cumulative-based model, and the resultant model outperformed any other single-module models (Fig. 7). In the PRS_RD model that was established with SNPs with a threshold of P < 0.0001, the add-on of CUS metrics resulted in an AUC increase in both HTN⁺ (from 0.796 to 0.824), and HTN^– (from 0.768 to 0.796) subgroups (Supplementary Table S7, Supplementary Fig. S2). As for the P < 0.001 PRS_RD model, incorporating CUS metrics improved the AUC from 0.944 to 0.947 in the HTN⁺ group. In the HTN^– population, the PRS model using SNPs with a threshold < 0.001 did not acquire prediction gain when CUS metrics were incorporated. This indicates that the prediction power of some selected SNPs is context dependent, with a particular intersecting relationship with the hypertensive and carotid occlusive populations. Furthermore, the involvement of available clinical covariates from our datasets did not undermine the state-of-the-art performance of our hybrid model. 

Figure 7.

 

Three-component RD prediction model combining total CUS metrics, COVs, and PRS_RD. (A) Nomogram. (B) ROC curves. (C, D) AUCs of the RD prediction model in patients with (1) and without (2) self-reported HTN.

Discussion

Our research aimed to address the scarcity of information on primary RD prediction by leveraging CUS parameters to identify high-risk individuals within the general population. By comparing CUS results with those from standard eye disease surveys, we sought to evaluate the effectiveness and accuracy of CUS in predicting RD, complementing our previous work on constructing a predictive PRS model for primary RD. Consequently, we demonstrated that the hemodynamic metrics of CUS are associated with RD risk. Moreover, the RD polygenic risk score (PRS_RD) model was further boosted by integrating CUS biomarkers through logistic cumulative analysis. Together, this study delineated the carotid hemodynamic status as a novel screening target for at-risk individuals before disease onset. 

Although CUS has been recognized as a useful diagnostic tool for predicting systemic diseases such as stroke⁵⁵ and coronary artery disease,⁵⁶ CUS has also been valuable for screening potential ischemic ocular diseases, such as ocular ischemia, amaurosis fugax, Hollenhorst plaques, retinal artery occlusion, and bilateral temporary visual loss.^57,58 Moreover, the hypoperfusion or embolic effects⁵⁹ of atherosclerotic carotid arteries in the terminal retina include transient blindness, venous stasis retinopathy, ocular ischemic syndrome, iris neovascularization, secondary glaucoma, and age-related macular degeneration.^60–62 

Despite the anatomy connection between carotid vascular supply and the downstream ocular tissues, CUS initially was not a standard eye disease survey. We demonstrated the potential of CUS examination in predicting RD risk based on the presence of plaques, Q_min, TAMAX, PI, resistive index (RI), and IMT. Plaques, RI,⁶³ and IMT⁶⁴ have been well-known parameters for atherosclerotic severity and ischemic pathologies,^65,66 including transient ischemic attack, embolic stroke, and retinal artery occlusion. However, the variables selected by the backward regression algorithm for the final model were not identical in the HTN⁺ and HTN^– subgroups, suggesting the involvement of different pathomechanisms. 

There is no clear evidence indicating linkage between CUS metrics and RD pathomechanisms to date. However, the potential usefulness of CUS metrics in RD screening is considerable because CUS metrics indirectly reflect some plaque-prone systemic conditions, such as hypertension, diabetes mellitus, cigarette smoking, and elevated serum low-density lipoprotein.⁶⁷ At the genetic level, although carotid plaque is strongly linked to coronary artery diseases and stroke at the SNP level,⁶⁸ there are common SNP functional genes between our RD SNP list and previous GWASs for carotid IMT and carotid plaque (Supplementary Table S5), including ADAMTS9, APOB, APOE, ALPL ATG4B, EDNRA, and LDLR. Together, in light of these carotid–eye connections, CUS can be can be used to evaluate ischemic ocular diseases. 

In ophthalmologic diagnostics, GWAS PRS has proven feasible in risk stratification for dry eye disease,⁶⁹ glaucoma,⁷⁰ and age-related macular degeneration⁷⁰ in the general population; however, the value of PRS in RD risk grading has not been accessed before. Notably, the genetic features of RD have been extensively explored in the past decade.^{33,35,71–73} Previous studies⁷⁴ identified several RRD risk loci, shown in Supplementary Table S7. Boutin et al.³³ explored the self-reported RD population from the UK biobank and revealed six highly reproducible RD risk loci.^33,35 Except for the SNPs related to patients’ underlying conditions, such as myopia and previous cataract operation, FAT3 and COL22A1 were mapped from the unique SNPs associated with primary RD. Loss-of-function mutation of FAT3 resulted in an aberrant retinal neuronal architecture that may be more susceptible to future retinal break.⁷⁵ Also, COL22A1 encodes for the retina-enriched collagen type XXII alpha 1, whose role in RD remains unclear. In our work, GWAS analysis based on individuals from the TWB study revealed several significant RD risk loci, including CELF2 rs56168975, IGFBPL1 rs117248428, and PAX6 rs11825821. Using a P threshold of 1 × 10⁻⁵, the three RD risk loci identified were novel when compared to previous works. PAX6 encodes a transcription factor responsible for ocular development. Despite the unclear pathophysiological mechanisms, several dysregulated PAX6-related phenotypes, such as aniridia, WAGR syndrome (Wilms tumor, aniridia, genitourinary abnormalities, and mental retardation), isolated foveal hypoplasia, MAC (microphthalmia, anophthalmia, and coloboma), Gillespie syndrome, and anterior segment dysgenesis–Peters anomaly were reported.⁷⁶ Despite a few case reports describing the comorbid RD in aniridia⁷⁷ and Peters anomaly,⁷⁸ the pathophysiological linkage of surgery-unrelated RD in these diseases remains unclear. Second, the insulin-like growth factor binding protein-like 1 (IGFBPL1) encodes an extracellular protein stabilizing IGF-1. IGFBPL1 is a neuroprotective agent in the retina because it suppresses the senescence marker p21 via calcium ion flux regulation.⁷⁹ Moreover, the lack of IGFBP1 leads to retinal ganglion cell (RGC) apoptosis.⁸⁰ Instead of being a direct predisposing factor to RD, RGC death has been well recognized as a terminal manifestation of several ocular diseases, such as glaucoma and traumatic optic neuropathy. Interestingly, the IGFBP1 signaling pathway was found to be upregulated in individuals with pathological myopia, a known risk factor for RRD.⁸¹ 

CUGBP Elav-like family member 2 (CELF2) encodes an RNA binding protein that exerts a tumor-suppressive effect via several signaling transduction pathways (e.g., PI3K/Akt, MAPK, Wnt/β‑catenin, endoplasmic reticulum–associated degradation, autophagy) based on posttranscriptional editing.⁸² CELF2 is expressed in several retina compartments, especially the optic nerve and microglia.⁴⁹ CELF2 may control early ocular development⁸³ through RNA alternative splicing.⁸⁴ Unfortunately, no direct evidence linking it to any specific retinal pathology exists. 

Furthermore, we incorporated tissue-specific human scRNA sequencing data of the genes mapped from the core RD risk loci we identified (Fig. 3). The results supported the relatively abundant intraocular distribution of our mapped genes compared to previous RD GWASs,^33,35 suggesting a more suitable future clinical utility. On the other hand, we found several shared genes between carotid artery disease GWAS studies and our RD-related risk loci (Supplementary Table S8). The similarity suggests potential pathophysiological crosstalk between these diseases and further strengthens the utility of CUS in RD evaluation. At the functional genomics level, we used an online portal⁸⁵ to annotate and functionally analyze the RD-related polymorphisms to understand the potential impact of genome-wide variants on the biomolecular pathways. Among the significantly enriched pathways, cell junction organization, regulation of cell projection organization, cell communication, cell junction assembly, and cell–cell adhesion via plasma membrane adhesion molecules are similar to the RD-related pathways (e.g., cell adhesion and extracellular matrix organization) revealed by Wang et al.⁸⁶ 

Some limitations of this study should be noted. First, due to the observational nature of the Taiwan Biobank study, the causality of RD cannot be inferred from the CUS metrics and gene polymorphisms. Second, a selection bias may exist as we intentionally selected individuals with previous CUS records. Generally, Taiwanese National Health Insurance supports those with past cerebrovascular disease histories receiving CUS examinations. For this reason, we have summarized the prevalence of systemic diseases that may lead to the clinical advice of CUS in Supplementary Table S3. Third, the strict threshold of GWAS significance (5 × 10⁻⁸) was not reached in the process of risk loci selection. The selected SNPs for our population have not been reported in previous GWAS studies adopting a stricter threshold; therefore, interpretation of the associations between the RD risk loci we identified should be treated with caution, and future validation from external datasets should be performed. Fourth, given that CUS is not a standard tool for eye disease surveys, future publication of additional image-based predictive models for primary RD would warrant inter-study comparisons to validate and refine their utility and accuracy. Sixth, external validation of our two-component RD prediction algorithm is currently limited due to the lack of accessible datasets that simultaneously include genetic and carotid sonographic (CUS) data. Although we identified several GWAS datasets containing RD-related genetic data (e.g., IEU Open GWAS project⁸⁷; UK biobank-based pheweb, https://pheweb.org/UKB-Neale; GWAS catalog⁸⁸), none provided carotid sonographic results. This limitation makes a comprehensive validation of our combined model impractical at present. To address this, we plan to expand our discussion to highlight the unmet need for such datasets and advocate for cautious interpretation of the clinical utility of our model. Additionally, we are recruiting a larger RD cohort with both genetic and sonographic data to enable thorough future validation. 

Overall, the outcomes suggest that CUS can be considered as an initial tool to screen individuals with high RD risk due to its easy accessibility and non-invasiveness. GWAS PRS is an accurate approach for RD risk stratification but expensive. Our study demonstrated that leveraging CUS and GWASs could offer significant cost-effectiveness in RD screening and could guide further clinical decision-making. 

Acknowledgments

This study is based in part on data from the Big Data Center, Taipei Veterans General Hospital. The interpretations and conclusions contained herein do not represent the position of Taipei Veterans General Hospital. The authors thank the Department of Statistics, Tamkang University, for technical support. 

Supported by grants from the National Science and Technology Council (NSTC 111-2314-B-075-036-MY3, NSTC 111-2314-B-075-039-MY3, NSTC 112-2321-B-A49-007), Taipei Veterans General Hospital (V113E-002-3, V113C-108, V113C-164), Big Data Center of Taipei Veterans General Hospital, and Center for Intelligent Drug Systems and Smart Bio-devices (IDS2B) as part of the Featured Areas Research Center Program within the framework of the Higher Education Sprout Project of the Ministry of Education (MOE) in Taiwan. 

Disclosure: K.-J. Chang, None; C.-Y. Wang, None; H.-Y. Wu, None; P.-Y. Weng, None; C.-H. Lu, None; W. Chiu, None; W.-C. Fang, None; C.-E. Kao, None; C.-Y. Li, None; Y.-T. Chung, None; Y.-C. Chen, None; A.-R. Hsieh, None; S.-H. Chiou, None; C.-C. Hsu, None; T.-C. Lin, None; S.-J. Chen, None; D.-K. Hwang, None