Association of Retinal Nerve Fiber Layer Thinning With Elevated High Density Lipoprotein Cholesterol in UK Biobank

Abstract

Purpose: The purpose of this study was to investigate the association between retinal nerve fiber layer (RNFL) thickness and high-density lipoprotein cholesterol (HDL-C) in a healthy population.

Methods: This cross-sectional study included 31,738 UK Biobank participants with high quality optical coherence tomography (OCT) images, excluding those with neurological or ocular diseases. The locally estimated scatterplot smoothing (LOESS) curve and multivariable piecewise linear regression models were applied to assess the association between HDL-C and RNFL thickness, and HDL-C subclasses were further analyzed using nuclear magnetic resonance (NMR) spectroscopy.

Results: Multivariate piecewise linear regression revealed that high HDL-C levels (>1.7 mmol/L in women or > 1.5 mmol/L in men) were associated with thinner RNFL thickness (women: β = −0.13, 95% confidence interval [CI] = −0.23 to −0.02, P = 0.017; male: β = −0.23, 95% CI = −0.37 to −0.10, P = 0.001). Conversely, a significant positive association between HDL-C and RNFL thickness was observed when HDL-C was between 1.4 and 1.7 mmol/L for female participants (β = 0.13, 95% CI = 0.02 to 0.24, P = 0.025). NMR analysis showed that these associations are potentially driven by distinct HDL-C subclasses.

Conclusions: This study revealed an association between HDL-C levels and retinal markers of neurodegenerative diseases, suggesting that elevated HDL-C may serve as a new risk factor for neurodegenerative conditions. These findings may contribute to the implementation of preventive interventions and improved patient outcomes.

The retina serves as an extension of the central nervous system, which is considered a potential biomarker for systemic neurodegenerative diseases, such as dementia.¹ Optical coherence tomography (OCT) has been widely used to measure retinal thickness, providing a visual window into the progression of systemic neurodegenerative diseases. Studies found that thinner retinal nerve fiber layer (RNFL) was associated with an increased risk of dementia.² Moreover, magnetic resonance imaging (MRI) of the brain, an important tool for the early diagnosis of neurodegenerative diseases, has shown associations with RNFL, suggesting concurrent neurodegenerative changes in the brain and eyes.³ As we know, the RNFL is widely recognized as a critical structural feature for diagnosing and managing glaucoma,⁴ and studies have demonstrated that individuals with glaucoma were at a higher risk of developing dementia.^5,6 

Systemic neurodegenerative disease and glaucoma share several risk factors or mechanisms,^7–9 and recent studies have explored their associations with serum lipid traits. A cohort study of healthy older adults has reported associations between high levels of high-density lipoprotein cholesterol (HDL-C) and an increased risk of dementia.¹⁰ Similarly, associations between high HDL-C levels and both IOP and glaucoma have been observed.^11,12 Traditionally, HDL-C has been considered the “good cholesterol” due to its association with a reduced risk of cardiovascular disease.¹³ However, emerging evidence suggests a potential link between elevated HDL-C levels and various adverse health conditions,^14,15 prompting a reconsideration of the HDL-C hypothesis and raising doubts about its widely recognized “good cholesterol” label. 

Given the potential association between HDL-C and adverse health outcomes and the established utility of OCT in the assessment of neurodegeneration, it is crucial to establish a link between HDL-C levels and retinal markers of neurodegenerative diseases. Therefore, this study aimed to explore the effect of HDL-C on RNFL thickness derived from OCT in a healthy population to determine its potential predictive value for future neurodegenerative change. Understanding this association could contribute to the development of early detection strategies for neurodegenerative conditions, which may lead to the implementation of early preventive interventions that can potentially improve patient prognosis. 

Methods

Study Population

More than 500,000 participants aged 40 to 69 years old were recruited into the UK Biobank from 2006 to 2010. The UK Biobank was approved by the North West Multicenter Research Ethics Committee by the principles of the Declaration of Helsinki. Written informed consent was obtained from all participants in 22 study assessment centers. Consenting participants provided baseline information via touchscreen questionnaires, verbal interviews, blood and urine tests, and physical measurements. Detailed information about the overall study and individual test protocols is available online (https://biobank.ndph.ox.ac.uk/ukb/index.cgi). 

Assessment of HDL-C Levels

Serum HDL-C levels were measured in non-fasting blood samples obtained at baseline recruitment. Measurements were performed by immunoturbidimetric analysis using a Beckman automated hematology analyzer (AU5800; Beckman Coulter, USA). For a more detailed lipid profile analysis, we assessed 4 lipoprotein subclasses: small, medium, large, and very large HDL-C, using targeted high-throughput nuclear magnetic resonance (NMR) metabolomics on a random subset of 119,764 participants. Full details regarding the processing of blood samples and the methodology of NMR metabolomics have been described previously.^16,17 Additional information on the specific NMR metabolomics’ classifications is available online (https://biobank.ndph.ox.ac.uk/ukb/label.cgi?id=220). 

Assessment of RNFL Thickness

The spectral-domain OCT (SD-OCT; OCT-1000; Topcon, Japan) imaging was performed in a dark room without pupil dilation using the 3-dimensional 6 × 6 mm² macular volume scan mode (512 A scans per B scan; and 128 horizontal B scans in a raster pattern). Topcon Advanced Boundary Segmentation (TABS) software was used to automatically segment and calculate the mean layer thicknesses of individual retinal layers. Specifically, the RNFL refers to the area between the inner limiting membrane and the inner surface of the ganglion cell layer (Fig. 1). 

Figure 1.

 

Spectral-domain optical coherence tomography images. (A) Without segmentation lines: Displays the natural structure of the retinal layers. (B) With segmentation lines: Displays representative retinal nerve fiber layer. The top two lines represent the inner limiting membrane and the inner surface of the ganglion cell layer. The areas between these two lines represent the macular nerve fiber layer.

Patients were excluded from the analysis based on established quality control criteria.¹⁸ Eyes with signal strength below 45 and the poorest 20% of centration certainty, which assesses the accuracy of macula centering in the images, were excluded from the analyses to identify the subset of participants with high-quality, well-centered images and central, stable fixation during the OCT scan. To reduce confounding by other ocular parameters, participants were excluded if they had visual acuity of worse than 0.1 logarithm of the minimum angle of resolution (logMAR), high refractive error (>±6 diopters [D]) or abnormal IOP (≤5 mm Hg or ≥21 mm Hg). To account for the potential influence of neurodegenerative and ocular diseases on RNFL disruption, patients with a history of glaucoma, retinal disorders, dementia, Parkinson's disease, or multiple sclerosis were identified and excluded based on their baseline information (Supplementary Table S1). Finally, if both eyes of a patient met the inclusion criteria for this analysis, one eye was randomly selected for inclusion. 

Assessment of Covariates

To control for potential confounders, a wide range of covariates were included in this study, including demographic, socioeconomic, lifestyle, and medication history collected at the baseline visit. Self-reported ethnic background was categorized as White and non-White. The Townsend Deprivation Index was derived from the postal code of residence and calculated based on employment status, home and car ownership, and household condition, with a higher score representing a greater degree of deprivation.¹⁹ Educational qualifications were determined by self-reported highest level of qualification and divided into five levels according to the International Standard Classification of Education (Supplementary Table S2).²⁰ Smoking status (categorized as never, former, or current smoker), and alcohol status (categorized as never, former, or current drinkers) were determined through self-reporting. Body mass index (BMI) was calculated based on baseline height and weight measurements and categorized into underweight, normal weight, overweight, and obese using the World Health Organization criteria.²¹ Time spent on moderate to vigorous physical activity (MVPA) was transformed into metabolic equivalent task (MET) minutes/week and categorized into five quintiles based on adapted questions from the short International Physical Activity Questionnaire.²² Baseline systemic conditions, including high blood pressure, diabetes, heart attack, angina, and stroke, were identified using both hospital inpatient records (ICD-10) and self-reported data. Specifically, heart attack, angina, and stroke were categorized under cardiovascular diseases. Details on data sources and classification methods are provided in Supplementary Table S1. Spherical equivalent values were measured by automated refraction (RC-5000; Tomey, Japan) and calculated based on refraction results (sphere degree + 0.5 × cylinder degree). IOP was measured with the Ocular Response Analyzer (ORA; Reichert, USA). 

Statistical Analyses

Descriptive statistics are presented as mean (standard deviation) for continuous variables and number (percentage) for categorical variables. Given the potential bias of complete case analysis,²³ missing socioeconomic variables were imputed using the “missForest” algorithm. The proportion of missing data ranged from 0% to 13.2%. Following the imputation process, the minimal error observed for continuous variables (normalized root mean square error [NRMSE]) was 0.0001%, and the minimal error for categorical variables (proportion of falsely classified [PFC]) was 0.15%. Fitting smooth lines were performed to observe the changing trends of the average RNFL thickness with HDL-C using the locally estimated scatterplot smoothing (LOESS) method. Piecewise linear regression models were used to investigate the associations between RNFL and HDL-C. Multivariate models were adjusted for age, sex, assessment center, average total household income before tax, Townsend deprivation index, smoking status, drinking status, ethnic background, education achievement, BMI, MVPA time, sleep duration, statin use, diabetes, cardiovascular diseases, hypertension, spherical equivalent, and IOP. To ensure consistency between complete case analyses and results derived from imputed data, we conducted a sensitivity analysis using only complete cases (n = 27,095). Additionally, given the reported association between statin use and a reduced risk of dementia,²⁴ we performed a further sensitivity analysis excluding 4638 statin users. Another sensitivity analysis excluded 723 participants suspected of having glaucoma, identified by a cup-to-disc ratio ≥0.7.²⁵ Finally, we conducted an additional analysis focusing on lipid subfractions. All statistical analyses were performed using R software (version 4.3.1). 

Results

Of the 67,129 participants who completed OCT examinations at baseline, 51,710 were considered to have high-quality images after quality control. Of these, 36,736 had a refractive error within ±6 D, vision of 0.1 logMAR or better, normal IOP measurements, and no self-reported ocular or neurologic disease. After excluding participants with missing serum lipid data, a total of 31,738 participants were included in the final analysis (Fig. 2). 

Figure 2.

 

Flowchart of the study population. D, diopter; IOP, intraocular pressure; logMAR, logarithm of the minimum angle of resolution; NMR, nuclear magnetic resonance; OCT, optical coherence tomography.

The mean age of the participants included was 55.7 (8.25) years, with an age range of 40 to 70 years, and 52.7% being women. Most participants were White (91.7%). The mean refraction was −0.67 D, and the mean IOP was 15.35 mm Hg. Table 1 summarized the mean RNFL thickness and mean HDL-C concentrations, demographic, systemic, and ocular factors for the participants included. 

Table 1.

 

Demographic, Ocular, and Systemic Characteristics of Participants Included in the Study

 

The locally estimated scatterplot smoothing (LOESS) curves were used to investigate the association between RNFL and HDL-C and illustrate the changing trend of average RNFL thickness by different HDL-C levels (Fig. 3). The RNFL thickness was relatively flat until approximately 1.3 mmol/L of HDL-C, after which it started to increase, with the peak occurring at around 1.7 mmol/L and then showing a downward trend. Therefore, the piecewise linear regression was performed with the knots at 1.3 and 1.7 mmol/L. Univariate analyses found a consistent trend with the LOESS curve, but the model became inconsistent after the addition of sex into the model. Therefore, further sex-stratified analyses were conducted (Table 2). For female participants, HDL-C was associated with an increased RNFL thickness when HDL-C was between 1.4 and 1.7 mmol/L (β = 0.13, 95% CI = 0.02 to 0.24, P = 0.025). However, HDL-C was significantly negatively associated with RNFL thickness when HDL-C was >1.7 mmol/L (β = −0.13, 95% CI = −0.23 to −0.02, P = 0.017). For male participants, HDL-C was associated with a decreased RNFL thickness when HDL-C was >1.5 mmol/L (β = −0.23, 95% CI = −0.37 to −0.10, P = 0.001). 

Figure 3.

 

LOESS curves of RNFL thickness as a function of HDL-C levels for total (A), female patients (B) and male patients (C). The dashed lines delineate segments for regression analysis, indicating potential shifts in trends. HDL-C, high density lipoprotein cholesterol; RNFL, retinal fiber nerve layer.

Table 2.

 

Association Between HDL-C Levels and RNFL Thickness

 

To further explore whether the association between RNFL thickness and HDL-C was driven by specific lipid subfractions, we examined lipid subfractions with NMR spectroscopy (Supplementary Tables S3–S6). In female patients, large and very large HDL-C subclasses were associated with an increase in RNFL thickness when HDL-C was between 1.1 and 1.5 mmol/L (large HDL-C: β = 0.14, 95% CI = 0.01 to 0.28, P = 0.034; and very large HDL-C: β = 0.15, 95% CI = 0.02 to 0.29, P = 0.027), whereas medium HDL-C was significantly negatively associated with RNFL thickness when HDL-C was >1.5 mmol/L (β = −0.25, 95% CI = −0.38 to −0.13, P < 0.001). In male patients, small and medium HDL-C subclasses were associated with decreased RNFL thickness when HDL-C was >1.2 mmol/L (small HDL-C: β = −0.18, 95% CI = −0.32 to −0.04, P = 0.012; and medium HDL-C: β = −0.21, 95% CI = −0.34 to −0.07, P = 0.002). 

In sensitivity analysis using complete cases, excluding statin users, and excluding potential glaucoma cases, the association between RNFL thickness and HDL-C remained significant and no change in the direction of the association occurred (Supplementary Tables S7–S9). 

Discussion

This OCT-based study revealed a nonlinear relationship between RNFL thickness and HDL-C, as illustrated by the LOESS curve. Subsequently, a segmented linear regression analysis identified a potential association between high HDL-C levels (>1.7 mmol/L in female patients or >1.5 mmol/L in male patients) and thinner RNFL thickness. Interestingly, a positive and significant association between RNFL thickness and HDL-C was observed when HDL-C was between 1.4 and 1.7 mmol/L for female participants. By focusing on lipid subfractions from NMR spectroscopy, we revealed that these associations, which vary by gender and HDL-C levels, may be driven by distinct HDL-C subclasses. These findings establish a link between HDL-C levels and retinal markers of neurodegenerative diseases in a real-world healthy population. 

Compelling evidence supported an association between thinner RNFL and impaired cognitive function. Evidence from a meta-analysis of 25 OCT studies indicated that people with dementia had thinner RNFL compared with healthy controls.²⁶ A study using data from the UK Biobank revealed a significant association between RNFL thickness and cognitive function at baseline. Furthermore, individuals with thinner RNFL have a twofold increased risk of cognitive decline at a 3-year follow-up, suggesting that RNFL thinning precedes cognitive decline and can serve as a predictor of cognitive deterioration.²⁷ The prospective population-based Rotterdam Study has further strengthened this association, showing that thinner RNFL is associated with an increased risk of dementia.² A common pathogenic mechanism underlies the association between RNFL and dementia. Brain regions covering the visual tract are affected in patients with dementia, potentially disrupting neuronal connections of the visual tract and leading to retrograde degeneration of the optic nerve.² Thinner RNFL may serve as a preclinical biomarker for neurodegenerative processes. However, this parameter and its clinical significance remain controversial among studies and warrant further study with standardized diagnostic protocols.²⁸ 

Based on the established association of RNFL with neurodegenerative diseases and cognition, this study provides evidence that elevated HDL-C may serve as a new risk factor for neurodegenerative diseases. Previous cross-sectional studies generally reported that HDL-C levels were significantly lower in participants with dementia.²⁹ However, the findings from prospective cohort studies are contradictory, possibly for two reasons. On the one hand, this could be explained by reverse causation. Reverse causation occurs when the disease process induces changes in modifiable traits of interest.³⁰ Consequently, reverse causation is an important limitation in cross-sectional studies, which cannot indicate whether risk factors predict future events. We speculate that high HDL levels increase the risk of neurodegenerative and retinal diseases, but after onset, the original risk factors may change, such as a decrease in HDL-C, leading to confounding effects on the real impact of risk factors. Therefore, this study was conducted in a healthy population to examine the association between RNFL and HDL-C, avoiding the confounding effect of reverse causation. On the other hand, our study found that high HDL-C levels (>1.7 mmol/L in female patients or >1.5 mmol/L in male patients) were associated with thinner RNFL thickness, which may indicate a potential threshold effect of HDL-C at higher levels. Most previous studies categorized HDL-C into tertiles or quartiles without specifically identifying individuals with very high HDL-C levels. However, recent studies focused on patients with very high levels of HDL-C and identified an association between HDL and dementia risk. A large cohort study of 18,668 participants revealed that high HDL-C levels (>2 mmol/L) were associated with the risk of incident dementia in both older male patients and female patients (hazard ratio [HR] = 1.27, 95% CI = 1.03–1.58).¹⁰ In addition, observational and genetic studies suggested that, for any dementia, men with HDL-C 2.2 to 2.7 mmol/L and >2.7 mmol/L had HRs of 1.66 (95% CI = 1.30–2.11) and 2.00 (95% CI = 1.35–2.98) compared with men with HDL-C 1.3 to 1.5 mmol/L. Similarly, compared with women with HDL-C 1.6 to 1.9 mmol/L, women with HDL-C 2.7 to 3.2 mmol/L and >3.2 mmol/L had HRs for dementia of 0.94 (95% CI = 0.74–1.18) and 1.45 (95% CI = 1.03–2.05).³¹ These findings suggest that the adverse effect of HDL may manifest only beyond a certain threshold. Conversely, moderate HDL-C levels, as observed in previous studies, might exhibit a certain degree of neuroprotection. This aligns with our finding of a significant positive correlation between RNFL thickness and HDL-C when the HDL-C was between 1.4 and 1.7 mmol/L in female participants. 

The relationship between HDL-C and RNFL may extend beyond the established role of HDL in retinal lipid efflux,³² suggesting a deeper link with neurodegenerative mechanisms. Plasma HDL-C levels do not always reflect the true functional dynamics of lipid transport, and dysfunction is possible at exceptionally high levels.³³ Dysregulated lipid particles could lead to accumulation of cholesterol deposits in cells, causing adverse effects.³⁴ Notably, a large cohort study from the EYE-RISK and European Eye Epidemiology Consortia³⁵ found that higher HDL levels were associated with an increased risk of age-related macular degeneration (AMD; OR = 1.21 per 1 mmol/L increase; 95% CI = 1.14–1.29). The disruption of physiological homeostasis induced by elevated HDL-C levels may impact cholesterol transport and esterification mechanisms, thereby facilitating the formation of lipid-rich deposits that are characteristic of early AMD.³⁶ Additionally, a recent cross-sectional study has reported an association between high HDL cholesterol and diabetic retinopathy,³⁷ which may be related to impaired antioxidant functions of HDL.³⁸ Similarly, the protective role of HDL against amyloid-beta aggregation, crucial in dementia prevention,³⁹ may be compromised at high HDL-C levels. These pathological changes are not uncommon, HDL can shift from anti-inflammatory to pro-inflammatory in individuals with chronic conditions that promote systemic oxidative stress and inflammation.⁴⁰ This is due to reduced activities of apolipoprotein A-I (apoA-I) and paraoxonase-1 (PON1).^41,42 However, the exact mechanisms and cellular pathways of lipid metabolism in neurodegenerative and retinal diseases remain unclear and further research is warranted to elucidate potential therapeutic targets and preventive strategies. 

Lipoproteins can be further classified into subclasses based on size, and the difference in lipid content leads to particle-specific functions.⁴³ For example, changes in lipid composition of HDL can affect its atheroprotective capacity.⁴⁴ We speculated that the effects of blood lipid metabolites on the RNFL thickness would vary depending on the specific lipid subfraction. Consequently, we investigated the relationship between HDL-C subclasses and RNFL thickness using NMR spectroscopy to examine specific lipid subfractions. Our findings revealed that the relationship between HDL-C levels and RNFL thickness is influenced by both gender and HDL-C concentration, potentially driven by distinct HDL-C subclasses. Specifically, when HDL-C levels are at high levels (>1.7 mmol/L in female patients or >1.5 mmol/L in male patients), the observed negative association between HDL-C levels and RNFL thickness may be attributed to medium HDL-C, with small HDL-C also playing a role in male patients. Although the positive association of RNFL thickness with HDL-C levels in female patients, when HDL-C was between 1.4 and 1.7 mmol/L, may be influenced by large and very large HDL-C subclasses. This underscores the complex interactions among specific lipid subfractions. Understanding these mechanisms could lead to targeted interventions that modify lipid profiles to prevent or slow the progression of neurodegenerative and retinal diseases. 

This study established a direct connection between HDL-C levels and retinal markers of neurodegenerative diseases in a real-world healthy population. A retrospective study found similar results to ours, indicating elevated HDL-C levels were negatively associated with RNFL thickness in patients with multiple sclerosis (P = 0.008). However, it is important to note that patients with multiple sclerosis are often accompanied by optic nerve inflammation, which may lead to reduced RNFL thickness through retrograde trans-synaptic degeneration.⁴⁵ Therefore, the identified association in patients with multiple sclerosis cannot be directly attributed to the association between RNFL thickness and HDL-C. In our study, we specifically excluded participants who reported neurological and ocular diseases, avoiding the potential confounding effects of these conditions on RNFL. Consequently, our results are more representative of a premorbid population, further strengthening the direct association between HDL-C levels and thinner RNFL. 

The strengths of this study include its large sample size that provided sufficient statistical power to detect the changing trends in RNFL thickness with HDL-C, and widely available diagnostic data that allowed us to conduct the study in a healthy population, avoiding confounding by neurodegenerative and ocular diseases. However, there are still several limitations in this study. First, serum samples were collected only once and from non-fasting individuals, which may have been influenced by recent dietary intake. Second, OCT examinations were conducted by technicians at different assessment centers, potentially introducing measurement errors that weaken possible associations with RNFL. However, all technicians underwent standardized training for OCT examination and the inclusion of assessment center as a covariate aimed to control for potential bias. Third, existing studies have been confirmed for the impact of age on RNFL thickness and HDL-C levels. Despite adjusting for age in multivariate regression models and the relatively narrow age range of UK Biobank participants, it is difficult to completely eliminate the possibility of age acting as a mediating factor. Last, the study population was predominantly of European descent, which may limit the generalizability of the findings to other ethnic groups, although this does not undermine the internal validity of the study. 

In conclusion, this real-world study revealed an association between high HDL-C levels and thinner RNFL thickness. Our findings suggest the potential existence of a threshold effect of HDL-C at higher levels. Moreover, the study revealed an association between HDL-C levels and retinal markers of neurodegenerative diseases, suggesting that elevated HDL-C may serve as a novel risk factor for conditions such as dementia. These findings may contribute to the implementation of preventive interventions and improved patient outcomes. 

Acknowledgments

The authors are grateful to the UK Biobank participants. This research was conducted using the UK Biobank Resource under Application Number 87083. 

Supported by the Guangzhou Basic Research Program, City & University (Institute) Joint Funding Project (2023A03J0174 and SL2023A03J00514), National Natural Science Foundation of China of Guangdong Province (2021A1515011673 and 2022A1515011181) and National Natural Science Foundation of China (81873673 and 81900841). 

Author Contributions: G.M.J., Z.Z.L., and D.Y.Z. designed the study. Y.Y.M., Y.W., L.J., and G.M.J. conducted the data analysis. Y.Y.M. and Y.W. drafted the manuscript. All authors critically revised the manuscript for important intellectual content. All authors read and approved the final manuscript. 

Availability of Data and Materials: The data used in this current study are available from the UK Biobank data resources. Permissions are required in order to gain access to the UK Biobank data resources, subject to successful registration and application process. Further information can be found on the UK Biobank website (https://www.ukbiobank.ac.uk/). 

Disclosure: Y. Ma, None; Y. Wu, None; L. Jin, None; L. Hu, None; W. Chen, None; C. A. Young, None; X. Zhang, None; D. Zheng, None; Z. Liu, None; G. Jin, None