Introduction

Diabetes mellitus is one kind of metabolic disease characterized by hyperglycemia. It is a leading cause of death and shortened life expectancy and seriously affects the health of people worldwide. In 2017, the global prevalence of diabetes stood at 476.0 million, resulting in 1.37 million deaths and an estimated disability-adjusted life-years related to the disease was 67.9 million [1]. It is estimated that the number of people with diabetes will rise to approximately 642 million by 2040 [2].

Blood lipids is a collective term for cholesterol, triglycerides (TG), and lipids (such as phospholipids) in the serum, and the common test items for blood lipids include total cholesterol (TC), TG, low-density lipoprotein cholesterol (LDL-C), and high-density lipoprotein cholesterol (HDL-C). The role of dyslipidemia, an established risk factor for diabetes, in the prevalence and control of diabetes has been previously described [3, 4]. Additionally, lowering non-HDL-C levels, such as TG and LDL-C, could be used as a primary or secondary prevention strategy for cardiovascular disease [5,6,7]. The “lower is better” strategy has gained popularity in blood lipid management. Cholesterol is an essential nutrient for maintaining normal human physiological functions, such as cell membrane structure [8], cell signaling [9], and hormone production [10]. Complex mechanisms maintain blood lipid levels within a physiological range, and the dysregulation of these mechanisms can lead to elevated or decreased tissue cholesterol levels, resulting in embryonic or adult diseases [11]. Additionally, a number of real-world studies [12] and randomized controlled trials [13] have demonstrated a paradoxical link between diabetes and blood lipid levels. A meta-analysis of 13 trials related to statin demonstrated the promoting effect of statin treatment on the risk of diabetes and found that the incidence risk of diabetes expanded by 9% in the treatment group [14]. In another cross-sectional study, a correlation between significantly lower TC and LDL levels and a higher prevalence of diabetes was observed in older people, regardless of the use of lipid-lowering medications [15].

Although the cholesterol paradox has attracted increasing attention, the association between blood lipid levels and diabetes in older Chinese adults remains unclear, even though older adults are prone to dyslipidemia and diabetes [16, 17]. Moreover, previous studies only focused on one or two blood lipid indicators [18] or divided blood lipid status into two groups (normal or dyslipidemia) [3]. No systematic and detailed analyses have been conducted on traditional blood lipid parameters, particularly the dose–response relationships between blood lipids and diabetes and the contribution of different lipid components to the cholesterol paradox.

Therefore, this study aimed to explore the potential associations of blood lipids with diabetes among older Chinese adults and to provide a more comprehensive understanding of TC, TG, LDL-C, and HDL-C in association with diabetes.

Methods

Study population

China’s Basic Public Health Services Project is a public welfare project for primary healthcare aimed at all Chinese people, among which physical examination of the elderly is an important service. At the same time as providing free health services to the Chinese people through China’s Basic Public Health Services Project, a huge database called Residents’ Electronic Health Records has been established [19].

In total, 5,697,488 residents aged 65 years or older attending an annual physical examination in Zhejiang Province, China, in 2022 were initially included in this study. Participants who missed one or more research variables and/or had outliers (exceeding three times the standard deviation, clearly not meeting clinical standards) were excluded (n = 1,331,487). In addition, participants who had already been diagnosed with diabetes were excluded from the study (n = 1,097,073) to avoid the influence of drug use and behavioral interaction on levels of fasting blood glucose, blood lipids, and other variables. Ultimately, 3,268,928 participants with complete and logical parameters and without known diabetes were included.

Data collection

Demographic, clinical, and behavioral data were extracted from the Residents’ Electronic Health Records, 2022. The demographic data included age, sex, educational attainment, and marital status. The clinical data included body height, weight, waist circumference (WC), fasting blood glucose (FBG), systolic blood pressure (SBP), diastolic blood pressure (DBP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin (TBil), serum creatinine (Scr), blood urea nitrogen (BUN), TC, TG, LDL-C, and HDL-C. Behavioral data included smoking status, alcohol consumption status, and physical exercise.

Variables definition

Referring to the Chinese guidelines for type 2 diabetes, FBG ≥ 7.0 mmol/L was defined as diabetes in this study [20].

Body mass index (BMI) was calculated as body weight (kg) divided by body height (m) squared. BMI < 18.5 kg/m2, ≥ 24 kg/m2, and ≥ 28 kg/m2 was defined as low BMI, overweight and obesity, respectively [21]. WC ≥ 85 cm for men and ≥ 80 cm for women were defined as central obesity [21].

Behavioral variables were defined based on actual information provided by residents in the Residents’ Electronic Health Records, 2022. The smoking status was divided into three categories: never (participants who reported not smoking), regular smokers (participants who smoke at least one cigarette per day), and former smokers (participants who stopped smoking). Alcohol consumption was divided into three categories: never (participants who reported not drinking), drinkers (participants who reported drinking occasionally or regularly), and former drinkers (participants who stopped drinking). Physical exercise was defined as conscious engagement in at least 30 min of physical activity outside of work for > 3 days a week and was divided into two categories in this study: no and yes.

Statistical analysis

To analyze the association between blood lipids and diabetes, logistic regression models were used. TC, TG, LDL-C, and HDL-C levels were categorized into five groups based on their respective distributions. In the multivariate models, potential confounders were adjusted for age, sex, educational attainment, marital status, BMI, central obesity, SBP, DBP, ALT, AST, TBil, Scr, BUN, smoking status, physical exercise, and alcohol consumption. In addition, the TC, TG, LDL-C, and HDL-C were highly correlated. Therefore, to avoid multicollinearity in the statistical analysis, TC, TG, LDL-C and HDL-C were entered into the statistical models solely.

Restricted cubic splines (RCS) were used to flexibly model the associations of TC, TG, LDL-C, and HDL-C with diabetes. In spline models, four knots located at the 5th, 35th, 65th, and 95th percentiles were set up, and the comparison between the log-likelihood of a model with spline variables and the log-likelihood of a model with only a linear effect of the covariate was performed to test potential nonlinearity [22]. As the associations of TC, LDL-C, and HDL-C were approximately log-linear below and above their inflection points, the odds ratios (ORs) per 1 unit increase in TC, LDL-C, and HDL-C were calculated through logistic regression models. Given the similar shape between the TC-diabetes and LDL-C-diabetes relationships in the pre-phase analysis and the biological inclusion of LDL-C in TC, we examined how the shape of the TC-diabetes association changes after excluding data with low LDL-C (defined as those below the 2.5th, 5th, and 10th percentiles of total data).

Subgroup analysis was performed to study the associations between blood lipids and diabetes in various subgroups stratified by sex (male and female) and age (65-, 70-, 75- and 80-). In the subgroup analysis, multivariate logistic regression models were used, and confounding factors were fully adjusted, except for grouping factors. For sensitivity analysis, participants with glycosylated hemoglobin (HbA1c) data were extracted, and HbA1c ≥ 6.5% was defined as diabetes [23]. In total, 1,054,679 participants, including 67,949 patients with diabetes (HbA1c ≥ 6.5%), were included in the RCS models, with fully adjusted confounding factors to assess the relationship between blood lipids and diabetes.

All analyses were performed using R 4.3.0 for Windows 64-bit (R Foundation for Statistical Computing, Vienna, Austria. https://www.R-project.org/), and the significance level (alpha) was set at 0.05. RCS analysis for graphical displays was performed using the ‘rcssci’ package (Zhiqiang Nie (2023), R package version 0.4.0, https://cran.r-project.org/web/packages/rcssci/index.html).

Results

Study participants

In total, 3,268,928 older adults (≥ 65) were included in the analysis. The characteristics of participants were presented in Table 1, according to FBG levels (Normal [< 7 mmol/L] and Diabetes [≥ 7 mmol/L]). The average age of the participants was 72.60 years, with a standard deviation (SD) of 6.10 years. Additionally, 46.44% of the participants were male. The mean concentrations of TC, TG, LDL-C, and HDL-C in total participants population were 4.97 ± 1.06 mmol/L, 1.58 ± 1.00 mmol/L, 2.85 ± 0.87 mmol/L, and 1.44 ± 0.41 mmol/L, respectively, with corresponding SDs. FBG concentrations of 202,832 (6.20%) participants were ≥ 7 mmol/L, and compared with the normal FBG group, the levels of TC, TG, LDL-C were higher, and the levels of HDL-C were lower in the diabetes group (P < 0.001).

Table 1 Participant characteristics according to fasting blood glucose (electronic health records of residents aged 65 or older, 2022)
Full size table

1 Data are presented as n (%) or mean ± SD. Binary variables, central obesity and physical exercise, are listed the data of “Yes” only.

2 Differences of categorical variables between groups were examined with Pearson’s Chi-squared test, while numerical variables were examined with t-test.

Logistic regression modeling to assess associations between blood lipid indicators and diabetes

Table 2 presents the associations between blood lipid parameters and diabetes. Multivariate-adjusted logistic regression models showed a positive relationship between TG and diabetes, whereas diabetes was negatively correlated with HDL-C. Compared with the lowest quintile of TG, the highest quintile had an OR of 2.56 for diabetes, with a 95% confidence interval (95% CI) of 2.52–2.60. For HDL-C, compared with the lowest quintile, an OR of 0.72 (0.71–0.73) was obtained among participants in the highest quintile. Notably, LDL-C showed a J-shaped association with diabetes in the multivariate-adjusted logistic regression models. There was a 3–5% lower prevalence risk of diabetes among older adults in the second and third quintiles, whereas those in the fourth and fifth quintiles had a 4–23% higher prevalence risk of diabetes than those in the lowest quintile of LDL-C. As for TC, when compared to those in the first quintile, the risk of diabetes prevalence did not exhibit a significant increase among participants in the second quintile. However, for those in the third to fifth quintiles, the risk of diabetes prevalence increased by 4 to 38%.

Table 2 OR (95% CI) of diabetes in relation to TC, TG, LDL-C and HDL-C (electronic health records of residents aged 65 or older, 2022)
Full size table

TC, TG, LDL-C and HDL-C were categorized into five groups based on their respective distributions.

Model1: adjusted for age (65-, 70-, 75, 80-), sex (male, female), educational attainment (illiterate and primary, junior and senior, college degree or above, unknown) and marital status (single, married, unknown).

Model2: model1 adjusted for BMI (normal, low BMI, overweight, obesity), central obesity (no, yes), SBP, DBP, ALT, AST, TBil, Scr and BUN.

Model3: model2 adjusted for smoking status (never, regular smoker, former smoker), physical exercise (no, yes) and alcohol consumption status (never, drinker, former drinker).

RCS to assess associations between blood lipid parameters and diabetes

RCS was used to model and visualize the associations of TC, TG, LDL-C, and HDL-C with diabetes in older adults, and these four blood lipid indicators all exhibited non-linear relationships with diabetes (P-non-linear < 0.001, Fig. 1). A positive relationship was observed between TG and the OR of diabetes (Fig. 1B), whereas an L-shaped relationship between HDL-C and diabetes was presented in the RCS (Fig. 1D). The OR of diabetes decreased rapidly within the lower range of HDL-C, and when the HDL-C was above 1.44 mmol/L, the reduction trend became relatively flat. Below 1.44 mmol/L, the OR per 1 unit higher HDL-C was 0.61 (0.59–0.63), and it was 0.90 (0.89–0.92) above 1.44 mmol/L (Fig. 1D; Table 3).

Fig. 1
figure 1

Association of TC, TG, LDL-C and HDL-C with diabetes. A: TC and diabetes; B: TG and diabetes; C: LDL-C and diabetes; D: HDL-C and diabetes. Odds ratios are indicated by red solid lines and border of 95% CIs by red dashed lines (left coordinate axis). Histograms represent the percentage of each group to the total population (right coordinate axis). Reference point is 20th centile of TG, inflection point of HDL-C and lowest value for each of TC and LDL-C, with knots placed at 5th, 35th, 65th, and 95th centiles of each TC, TG, LDL-C and HDL-C distribution. All models were adjusted for confounders in Table 2

Full size image

Strong J-shaped associations for TC and LDL-C between diabetes were observed using the RCS (Fig. 1A and C). A decrease in OR was observed until 4.04 mmol/L of TC, and the OR was 0.97 (0.95–1.00) per 1 mmol/L increase of TC; when TC was above 4.04 mmol/L, the OR was 1.17 (1.17–1.18) per 1 mmol/L increase of TC (Fig. 1A; Table 3). The plot demonstrated a considerable decrease in the OR within the lower range of LDL-C levels, reaching its nadir at approximately 2.33 mmol/L, before gradually increasing thereafter. Below 2.33 mmol/L, the OR for every 1 mmol/L increase of LDL-C was 0.88 (0.86–0.90), and it was 1.18 (1.17–1.18) after above 2.33 mmol/L of LDL-C (Fig. 1C; Table 3). The relationship between TC and diabetes was examined after excluding the participants with low LDL-C levels. When participants below the 2.5th percentile of LDL-C were excluded, the J-shaped relationship between TC and diabetes weakened. After excluding participants with low LDL-C concentrations (< 5th and 10th), the J-shaped association between TC and diabetes disappeared and transformed into a positive correlation (Fig. 2). However, a slight change was observed in the J-shaped association as participants with low HDL-C levels were gradually excluded; they did not disappear (Supplementary Figure A).

Subgroup and sensitivity analysis

The J-shaped association between TC and diabetes was stronger in females than in males and stronger in the 65- than in higher age groups (70-, 75-, and 80-); similar associations for TG, LDL-C, and HDL-C were found across all subgroups in the subgroup analysis (Supplementary Tables A and B). The findings of this study remained robust in the sensitivity analysis. After adjusting for the confounding factors presented in Table 2, the RCS curves of these four traditional blood lipid parameters were similar to those shown in Fig. 1 (Supplementary Figure B).

Table 3 OR (95% CI) of diabetes for every 1 mmol/L increase of TC, LDL-C and HDL-C, stratified by the inflection point of HDL-C and lowest value for each of TC and LDL-C according to RCS (electronic health records of residents aged 65 or older, 2022)
Full size table

According to RCS, when TC and LDL-C are 4.04 and 2.33 mmol/L respectively, their respective ORs are the lowest; the inflection point of HDL-C is at 1.44 mmol/L. Adjusted for confounding factors shown in Table 2.

Fig. 2
figure 2

ORs (95% CI) of diabetes according to TC (exclusion of low LDL-C). A: excluded participants with LDL-C below 2.5th centile; B: excluded participants with LDL-C below 5th centile; C: excluded participants with LDL-C below 10th centile. Odds ratios are indicated by red solid lines and border of 95% CIs by red dashed lines (left coordinate axis). Histograms represent the percentage of each group to the total population (right coordinate axis). Reference point is 20th centile of each TC after exclusion of low LDL-C, with knots placed at 5th, 35th, 65th, and 95th centiles. All models were adjusted for confounders in Table 2

Full size image

Discussion

In this cross-sectional study of 3,268,928 older Chinese adults, the shapes of the associations between common blood lipid parameters (TC, TG, LDL-C, and HDL-C) and diabetes were determined. The relationship between the four blood lipid indicators and diabetes was non-linear with different shapes. A strong positive association was observed between TG levels and diabetes. In contrast, a negative relationship was observed between HDL-C levels and diabetes, and the OR curve was L-shaped. Both TC and LDL-C levels showed J-shaped associations with diabetes, with an inverse correlation before the inflection point and a positive correlation after the inflection point.

Many epidemiological studies have explored the association between blood lipid parameters and diabetes, and Low HDL-C and high TG have been proven to be established risk factors for diabetes in several previous studies [4, 18, 24, 25]. An observational study of 5,012 participants, which was based on two cross-sectional surveys, reported that the ORs of hypertriglyceridemia for diabetes were 1.54 (1.01–2.35) in men and 2.02 (1.49–3.10) in women [4]. A nationwide population-based study of over 5 million adults without known diabetes showed that the hazard ratio for incident diabetes in the low HDL-C and high variability group was 1.40 (1.38–1.42) compared with the high HDL-C and low variability group [24]. The reduction in HDL levels may be caused by a mechanism initiated by elevated plasma TGs against a backdrop of insulin resistance. This mechanism involves the facilitation of the transfer of cholesteryl esters from HDL to TG-rich particles through increased catabolism and the activity of cholesteryl ester transfer protein [26]. This study found that the growth trend of the OR was relatively stable with an increase in TG. However, a strong L-shaped relationship was observed between HDL-C and diabetes, and the reduction in the OR tended to be gradual after the plasma HDL-C level reached 1.44 mmol/L. Significant changes in the slope of the association between HDL-C levels and diabetes may be related to a negative feedback model of pancreatic islets. Evidence suggests that HDL plays a direct role in glycemic control by acting on pancreatic beta cells [27]. Hormonal secretion by the pancreatic islets is highly regulated. Insulin secretion lowers blood glucose levels, and decreasing glucose levels would shut down insulin secretion of pancreatic beta cells, which relieves the local inhibition of glucagon secretion of pancreatic alpha cells [28]. The complementary action of hormones secreted by pancreatic beta and alpha cells allows the maintenance of blood glucose at physiological levels.

Notably, there are significant differences between studies regarding the effects of TC and LDL-C on diabetes [4, 15, 29,30,31]. In a multicenter retrospective cohort study of 28,476 patients with coronary heart disease, high TC and LDL-C levels were risk factors for diabetes, with ORs of 1.08 (1.06–1.11) and 1.06 (1.03–1.10), respectively [29]. In contrast, increased TC and LDL-C levels were beneficial to reduce the prevalence of diabetes in a cross-sectional cohort study of 3,688 participants aged ≥ 50 years [15]. Another observational study of 9,892 patients with hypertension provided completely different results (a U-shaped relation between LDL-C levels and diabetes prevalence) [30], and RCS performed in this study found similar results that TC and LDL-C showed J-shaped associations with diabetes among older adults ≥ 65 years. Moreover, a new insight was added: the observed J-shaped association between TC and diabetes can be explained when the two different shapes of diabetes prevalence risk for LDL-C and HDL-C are combined. The increased prevalence risk of diabetes within the lower TC range (< 4.04 mmol/L) could be mainly attributed to a combination of the high prevalence risk in older adults with low LDL-C levels and the increase in prevalence risk under low HDL-C conditions may have a synergistic effect on the occurrence of this phenomenon. The increase in diabetes prevalence risk in the higher TC range (≥ 4.04 mmol/L) may be due to the reversal of the relationship between LDL-C and diabetes, and when the promoting effect of high LDL-C on diabetes prevalence exceeds the protective effect of high HDL-C levels, the prevalence risk of diabetes increases with the increase of TC.

The cholesterol paradox regarding serum cholesterol levels and the risk of diabetes has been preliminarily discussed in clinical trials of statins [14]. The patterns observed in this study suggest that HDL-C and LDL-C jointly lead to the cholesterol paradox, and the two opposite effects of low and high LDL-C on diabetes seem to play a crucial role. A post-hoc analysis of an intervention trial demonstrated a positive protective effect of rosuvastatin on coronary artery disease or cerebrovascular disease events, while the risk of diabetes increased with a reduction in LDL-C levels. In Everett BM et al.’s study, the risk of type 2 diabetes among participants with an LDL-C < 30 mg/dl increased by 56% compared with participants with LDL-C ≥ 30 mg/dl [32]. Genetic studies have shown that LDL-C-lowering variants of several genes, including NPC1L1, HMGCR, and PCSK9, are directly associated with an increased risk of diabetes [33]. In addition, overexpression of NPC1L1 in the liver can suppress gluconeogenesis and lower fasting blood glucose and blood glucose levels, whereas inhibiting NPC1L1 with ezetimibe, a lipid-lowering agent, may promote gluconeogenesis [34]. Furthermore, the free-radical theory of aging and disrupted cholesterol homeostasis were used to account for the inverse association between LDL-C levels and cardiovascular mortality in older people [35]. However, it is unclear whether this hypothesis can explain the negative relationship between LDL-C and diabetes in geriatric populations with low LDL-C levels.

Strengths and limitations

This study has several strengths. First, the associations between different lipid parameters and diabetes prevalence were fully explored, and the contributions of different cholesterol components to the cholesterol paradox were further elaborated. Second, China’s basic public health services projects and the mature information system can provide a large amount of health-related data. Third, the non-linear relationship between blood lipids and diabetes found in this study, particularly the J-shaped relationship between LDL-C, TC, and diabetes, provides a reference for the management of blood lipid levels in older adults.

However, this study had some limitations. First, participants with diabetes were only differentiated by fasting blood glucose rather than by doctors’ diagnoses, which may have caused some bias. Second, type 1, type 2, and other types of diabetes were not distinguished. Type 2 diabetes accounts for more than 90% of all diabetes cases in China [36]. Third, The data analyzed in this study were cross-sectional, thus preventing any conclusive evidence of a causal relationship between risk factors and diseases. Further prospective studies are required to validate the findings of this study. Fourth, this study only included older Chinese adults and the findings from this population should be cautiously generalized to other populations. Fifth, owing to the limited information contained in the database, no data were collected on the use of lipid-lowering drugs, which may be a confounding factor affecting the results.

Conclusions

In this large-scale cross-sectional study of older Chinese adults, the relationship between blood lipid parameters and diabetes prevalence was not a simple linear correlation, and low TC and LDL-C levels were risk factors for diabetes prevalence among older Chinese adults. The positive correlation between low LDL-C levels and the risk of diabetes may be related to the cholesterol paradox. Therefore, extremely low LDL-C levels should be avoided in clinical lipid management, particularly in patients without diabetes. These findings may provide a new reference for the development of lipid management guidelines. The makers of blood lipid management strategies need to fully consider the role of the cholesterol paradox to avoid adverse events that may be caused by low blood lipid parameters.