Superior prognostic value of estimated small dense LDL cholesterol for cardiovascular risk assessment: evidence from international cohort studies

Abstract

Context

Small dense low-density lipoprotein cholesterol (sdLDL-C) refers to an atherogenic subclass of lipoproteins. Owing to limited access to direct measurement in routine clinical settings, several estimation formulas have been developed, with the equation proposed by Sampson et al. being the most widely adopted. Nevertheless, the clinical relevance of estimated sdLDL-C (E-sdLDL-C), particularly in predicting cardiovascular risk, remains to be validated in large-scale, diverse populations.

Methods

The international cohort analysis used two population-based datasets: the China Health and Retirement Longitudinal Study (CHARLS, n = 8,112; median follow-up: 7 years) and the UK Biobank (UKB, n = 321,310; median follow-up: 14.26 years). E-sdLDL-C was estimated using Sampson’s equation. Kaplan–Meier curves were employed to evaluate the incidence of cumulative event. Multivariable Cox models were applied to explore the predictive value of E-sdLDL-C for incident cardiovascular disease while adjusting for conventional lipid parameters and cardiovascular risk factors. Model performance was assessed using Harrell’s C-index and continuous net reclassification improvement to evaluate the incremental predictive value of E-sdLDL-C.

Results

An increase in E-sdLDL-C levels was strongly associated with a greater incidence of cardiovascular disease across both cohorts, independent of traditional lipid biomarkers, including low-density lipoprotein cholesterol, non-high-density lipoprotein cholesterol, triglyceride-rich lipoprotein cholesterol, and apolipoprotein B. In the UK Biobank cohort, both ischaemic heart disease and stroke showed a significant association with E-sdLDL-C (adjusted P < 0.05), whereas in the Chinese cohort, the association was evident for stroke but not for cardiac outcomes. Compared with conventional lipid measures, E-sdLDL-C showed stronger and more consistent associations. Notably, in sex-stratified analyses, the association between E-sdLDL-C and cardiovascular risk was stronger in males than in females, suggesting potential sex-specific differences in pathophysiology. Furthermore, E-sdLDL-C demonstrated significantly higher C-index and NRI in predicting cardiovascular outcomes compared with conventional lipid parameters. The findings were further validated through a series of sensitivity checks.

Conclusion

This international study demonstrated that E-sdLDL-C independently predicts cardiovascular risk and outperforms conventional lipid markers in terms of its prognostic value. In clinical settings where direct sdLDL-C measurement is not available, its estimation from routine lipid panels provides a practical and accessible alternative for cardiovascular risk stratification, potentially improving patient management and outcomes.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12944-025-02717-0.

Background

Cardiovascular disease (CVD) remains the primary cause of mortality globally, with ischaemic heart disease and stroke ranking first and third, respectively, according to 2024 WHO health statistics [1]. Given the sharp increase in CVD incidence with age [2], risk evaluation is essential for adults aged 40–75 years as part of primary prevention strategies [3].

In addition to traditional lipid markers such as low-density lipoprotein cholesterol (LDL-C), non-high-density lipoprotein cholesterol (non-HDL-C), apolipoprotein B (ApoB), and lipoprotein(a) [4, 5], emerging biomarkers such as small dense low-density lipoprotein cholesterol (sdLDL-C) and triglyceride-rich lipoprotein cholesterol (TRL-C) have shown strong associations with atherosclerotic cardiovascular outcomes [6–8]. SdLDL-C, the denser and more atherogenic LDL subclass [9], has been linked to CVD events, including stroke and coronary heart disease, in multiple studies [10–13].

Despite the clinical relevance of sdLDL-C, direct measurement techniques, including ultracentrifugation, electrophoresis, and nuclear magnetic resonance, are labour intensive and costly, limiting their utility in routine practice [14, 15]. Although homogeneous assays are available, they remain inaccessible in many settings, especially in primary hospitals, and the lack of standardized measurement protocols limits comparability across different reagents [16]. Several formulas have been established for estimating sdLDL-C [17–19]. Most of these methods require measurements of direct LDL-C and apolipoprotein B (ApoB), which are not available in some laboratories. Sampson et al. developed the first formula to derive sdLDL-C based on routinely tested lipid markers. These include total cholesterol (TC), HDL-C, and triglycerides (TG). The equation has undergone validation in diverse ethnic populations [20]. Elevated estimated sdLDL-C (E-sdLDL-C) has been linked to greater risk for developing CVD, diabetes, fatty liver disease, and kidney disease [21–23], but its predictive value in primary prevention requires further validation, especially as one study suggested that it may not be as effective for secondary coronary artery disease risk stratification as direct sdLDL-C [24]. Additionally, the agreement between estimated and direct sdLDL-C is only moderate, with reported R² values of approximately 0.7 [20, 25]. This highlights the need for population-based validation to assess predictive utility of E-sdLDL-C in diverse clinical contexts.

Cardiovascular risk increases substantially with age [2], highlighting the importance of accessible and reliable risk stratification tools in middle-aged and elderly populations. Considering the widespread burden of atherosclerotic disease in ageing populations and the limited availability of advanced lipid testing in many routine clinical settings, especially in resource-constrained environments, evaluating the predictive utility of E-sdLDL-C derived from standard lipid panels is of significant clinical and public health relevance. However, limited large-scale, population-based research has systematically evaluated the prognostic value of E-sdLDL-C, especially in older adults.

To address the current lack of evidence, this research explored the association between E-sdLDL-C and incident CVD in middle-aged and elderly individuals from both China and the United Kingdom. Data were obtained from two large-scale cohorts: the China Health and Retirement Longitudinal Study (CHARLS) and the UK Biobank. It was hypothesized that E-sdLDL-C would independently predict cardiovascular risk better than conventional lipid parameters and demonstrate consistent associations across diverse populations and health care settings.

Methods

Study population

CHARLS is a continuously conducted cohort investigation with national representativeness in China, enrolling Chinese participants aged 45 years or older, along with their spouses [26]. After excluding participants with cancer, those with CVD at baseline, individuals who provided nonfasting blood lipid samples, those without baseline or follow-up information on CVD, and those younger than 45 years (detailed in Fig. 1), a total of 8112 participants in wave 1 (2011) of the CHARLS were selected and formed the final analytic sample. Participants underwent follow-up assessments every two years. Follow-up assessments were conducted during 2013–2014, 2015–2016, and 2017–2018 (wave 2 to 4). Loss to follow-up was addressed by censoring.

Fig. 1.

Inclusion and exclusion process

UK Biobank is an extensive population-based longitudinal study, enrolling nearly 500,000 participants aged 40 to 69 years in the United Kingdom from 2006 to 2010 [27]. After excluding participants younger than 45 years, those with CVD at baseline, and those with cancer (as detailed in Fig. 1), a final sample of 321,310 subjects comprised the analytic cohort. They were followed for incident outcomes through July 1, 2023.

Data collection

For the CHARLS data, blood specimens were promptly stored at −20℃ after collection and transported to Beijing within two weeks. Upon arrival, they were preserved at −80℃ before analysis. Using an enzymatic colorimetric assay, fasting blood glucose, TC, HDL-C, direct LDL-C, and TG were determined. All assays were performed by certified laboratory technicians in strict accordance with standard protocols.

For the UK Biobank data, samples were taken randomly, and most of them were nonfasting samples. Biochemical indices, including fasting blood glucose, TC, TG, HDL-C, direct LDL-C, ApoB, and creatinine, were measured using a Beckman Coulter AU5800 clinical chemistry analyser [28]. In addition, Cholesterol in small LDL particles was measured by nuclear magnetic resonance (NMR) using a high-throughput metabolomics platform developed by Nightingale Health Ltd. (NMR-measured sdLDL-C). Among the 321,310 participants included in the present analysis, 179,315 individuals had available NMR-based sdLDL-C data (Data-Field 23547 in UK Biobank). Plasma samples were processed following standardized procedures and scanned using Bruker AVANCE IIIHD 500 MHz spectrometers [29].

Calculation of lipid parameters

Using the method of Sampson et al. [20], E-sdLDL-C was computed as follows: estimated LDL-C (E-LDL-C) minus estimated large buoyant LDL cholesterol (E-lbLDL-C).

Non-HDL-C was derived by subtracting HDL-C from TC, whereas TRL-C (mg/dL) was estimated by dividing TG by 5, following established methods [30].

Clinical outcomes

In the CHARLS cohort, CVD encompassing stroke and cardiac disease was the primary outcome of interest. CVD cases were identified among participants who had no history of such conditions at baseline, but reported a diagnosis of stroke or heart disease in any subsequent waves (wave 2, 3, or 4). These diagnoses were derived from participants’ survey responses indicating that a doctor had previously informed them any form of the condition, including myocardial infarction, coronary heart disease, angina, congestive heart failure, other heart problems, and stroke. This approach aligns with definitions used in other CHARLS-based studies [31].

In the UK Biobank database, CVD outcomes included newly diagnosed ischemic heart disease and/or stroke. Diagnoses of ischemic heart disease were identified based on the International Classification of Diseases 10th edition (ICD-10) codes, including I20, I21, I22, I24 and I25. Stroke was defined using ICD-10 diagnostic codes ranging from I60 to I64.

Covariates

For the CHARLS data, hypertension was identified if systolic pressure was ≥ 140 mmHg, diastolic pressure was ≥ 90 mmHg, antihypertensive treatment was ongoing, or the participant reported a physician’s diagnosis. Diabetes mellitus was defined if any of the following criteria were met: (a) fasting glucose ≥ 7.0 mmol/L, (b) HbA1c ≥ 6.5%, (c) random plasma glucose ≥ 11.1 mmol/L, (d) using glucose-lowering therapy, or (e) self-report of a physician-diagnosed type 2 diabetes. Other comorbidities, including kidney disease, liver disease, arthritis, digestive disorders, and asthma, were determined according to participants’ self-reported medical history. A person was considered a current smoker if they had smoked within the past 12 months and maintained a habit of smoking every day for at least a year. A person was considered a current alcohol drinker if they had consumed alcohol over the previous year and had consumed more than 100 millilitres per day (alcohol content greater than 50%) for more than a year.

In the UK Biobank, chronic kidney disease was identified when estimated glomerular filtration rate fell below 60 mL/min/1.73 m² [32]. Comorbidities, such as hypertension (I10, I15), chronic liver disease (K70 ~ K77, K83), arthritis (M00, M05, M06), asthma (J45, J46), and diabetes (E10, E11), were captured from health records. Participants who reported smoking or alcohol consumption at survey time were classified as current smokers or drinkers, respectively. Weekly moderate physical activity, measured in minutes, was categorized based on whether participants engaged in less than or at least 150 min per week. The Townsend Deprivation Index, a continuous indicator of socioeconomic status, was assigned to each participant based on the national census output area corresponding to their residential postcode.

Statistical analysis

To identify the most appropriate threshold of E-sdLDL-C for predicting incident CVD, receiver operating characteristic (ROC) curve was applied for each cohort. Cut-offs were chosen by locating the point closest to (0,1). Participants were categorized into low and high E-sdLDL-C groups accordingly. This binary classification was used for the primary analysis, whereas quartiles and continuous values were applied in sensitivity analyses to evaluate consistency. Continuous variables were summarized using medians with interquartile ranges, whereas categorical characteristics were described in terms of counts and proportions. Categorical variables were compared using the Pearson χ² test, and continuous variables were compared using the Wilcoxon rank-sum test.

To estimate the cumulative incidence of cardiovascular events, Kaplan–Meier survival analysis was conducted, with group comparisons evaluated via the log-rank test. Time-to-event outcomes were analyzed using Cox proportional hazards models, and results were reported as hazard ratios (HRs) with corresponding 95% confidence intervals (CIs). The proportional hazards assumption was formally evaluated using Schoenfeld residual tests (cox.zph function of the R survival package) in both CHARLS and the UK Biobank (see details in supplementary material). Separate models were developed for the CHARLS and UK Biobank datasets, incorporating available covariates and design-specific factors. A total of three adjusted models were specified for each cohort to evaluate robustness across different levels of covariable adjustment. In the CHARLS, model 1 included adjustments for age and sex; model 2 added body mass index (BMI), smoking status, and alcohol consumption; and model 3 further included comorbidities such as hypertension, diabetes, digestive diseases, asthma, kidney diseases, liver diseases, and arthritis. In the UK Biobank, model 1 included adjustments for age, sex, and ethnicity (White, Black, Asian, mixed/other); model 2 additionally included BMI, smoking status, and alcohol consumption; and model 3 incorporated further covariable adjustments, including physical activity, the Townsend Deprivation Index, the use of insulin, antihypertensive and cholesterol-lowering medications, and comorbidities, including hypertension, diabetes, chronic kidney diseases, arthritis, asthma, and chronic liver diseases. P-values for Cox regression coefficients were derived from Wald tests, as implemented in the summary.coxph() function of the R survival package.

To evaluate the incremental predictive value of E-sdLDL-C beyond conventional lipid markers, we assessed model performance using Harrell’s C-index and continuous net reclassification improvement (NRI) in both cohorts. All analyses were based on the multivariable Cox regression model (model 3) used in the primary outcome analysis. For each lipid parameter, we compared the base model with an extended model including the lipid variable. Differences in C-index between nested models were assessed using likelihood ratio tests, and C-index values were derived from the linear predictors of the Cox model. NRI and its 95% confidence interval were estimated using bootstrap resampling with 1,000 replications, while P values for NRI were calculated using normal approximation, based on the bootstrap-derived standard error and Z-statistics.

Multiple sensitivity analyses were conducted to verify the stability. Stratified Cox models were used to to assess interactions across key variables (e.g., sex, hypertension, and diabetes), and restricted cubic spline regression was applied to assess potential nonlinear associations between E-sdLDL-C levels and cardiovascular outcomes.

Statistical procedures were carried out in R (version 4.4.1), adopting a criterion of P < 0.05 (two-sided) to determine significance.

Results

Clinical outcomes and cut-off levels

In the CHARLS cohort, during the 7-year follow-up, 1624 participants had incident CVD, including 919 females and 705 males. A total of 1232 participants experienced cardiac disease, and 536 experienced stroke. In the UK Biobank cohort, 37,744 incident CVD cases were identified during the 14.26-year median follow-up. Among these participants, 28,173 were diagnosed with ischaemic heart disease, and 11,804 were diagnosed with stroke.

Given differences in measurement conditions, cohort-specific cut-off values were derived using ROC analysis within each dataset (Fig. 2). Cut-offs were defined as the point with the minimum distance to (0,1) on the ROC curve. In CHARLS, the values were E-sdLDL-C 34.44, direct LDL-C 115.79, non-HDL-C 139.56, and TRL-C 20.27. In the UK Biobank, they were E-sdLDL-C 44.45, direct LDL-C 139.52, ApoB 105.65, non-HDL-C 166.69, and TRL-C 27.78. All values are reported in mg/dL.

Fig. 2.

Optimal cut-off values of lipid-related markers for predicting cardiovascular outcomes. The optimal thresholds for lipid-related markers were determined by the minimum distance from point (0, 1) on the ROC curve. (A–D): CHARLS; (E–I): UK Biobank. Each vertical dashed line indicates the most appropriate cut-off value for corresponding biomarker

Baseline profiles

The participants’ baseline profiles were examined according to E-sdLDL-C stratification (low vs. high) in each cohort (Table 1). In the CHARLS cohort, the high E-sdLDL-C group had more females, higher BMIs, and higher lipid indices, except for HDL-C values. In the UK Biobank cohort, participants with elevated E-sdLDL-C were generally older, had increased BMI, and had adverse lipid profiles, including higher LDL-C, ApoB, non-HDL-C, and TRL-C. Notably, in the CHARLS cohort, higher prevalence of comorbidities—including hypertension and diabetes—was observed among individuals with elevated E-sdLDL-C levels. However, prevalence of hypertension and diabetes, along with the use of corresponding treatments, was lower among those with elevated E-sdLDL-C in the UK Biobank cohort. Across both populations, the proportion of cardiovascular disease events, including stroke and cardiac events, was substantially greater among participants with elevated E-sdLDL-C, compared to those with lower levels, supporting the potential role of the E-sdLDL-C level as a clinical risk indicator.

Table 1.

Baseline profiles of participants categorized by elevated versus reduced E-sdLDL-C levels

	CHARLS (n = 8112)			UK Biobank (n = 321310)
	E-sdLDL-C	E-sdLDL-C	P value	E-sdLDL-C	E-sdLDL-C	P value
	< 34.44 mg/dL	≥ 34.44 mg/dL		< 44.45 mg/dL	≥ 44.45 mg/dL
n	4057	4055		168,650	152,660
Female (%)	1874 (46.2)	2344 (57.8)	< 0.001	96,574 (57.3)	78,575 (51.5)	< 0.001
Age (years)	58.00 [51.00, 65.00]	58.00 [52.00, 64.00]	0.729	58.00 [51.00, 63.00]	59.00 [53.00, 63.00]	< 0.001
BMI (kg/m2)	22.15 [20.14, 24.46]	24.12 [21.72, 26.70]	< 0.001	25.86 [23.37, 29.03]	27.53 [25.11, 30.46]	< 0.001
Cholesterol (mg/dL)	170.88 [153.09, 190.59]	210.70 [190.98, 234.67]	< 0.001	198.17 [176.98, 220.10]	248.43 [227.32, 272.79]	< 0.001
Triglyceride (mg/dL)	74.34 [60.18, 93.81]	144.26 [110.62, 196.47]	< 0.001	97.53 [76.38, 129.03]	178.86 [138.75, 239.13]	< 0.001
HDL-C (mg/dL)	54.90 [45.62, 65.34]	45.62 [37.50, 54.12]	< 0.001	57.90 [47.83, 69.21]	51.92 [44.39, 60.99]	< 0.001
Direct LDL-C (mg/dL)	101.29 [84.67, 117.53]	132.22 [111.73, 154.25]	< 0.001	119.51 [104.22, 134.91]	160.77 [145.29, 178.83]	< 0.001
Estimated LDL-C (mg/dL)	100.08 [83.95, 116.69]	134.84 [116.74, 155.24]	< 0.001	119.66 [100.22, 138.88]	159.91 [140.57, 181.14]	< 0.001
ApoB (g/L)	-	-		0.90 [0.80, 1.00]	1.20 [1.09, 1.32]	< 0.001
Serum creatinine (µmol/L)	66.93 [56.94, 76.92]	66.93 [57.94, 78.91]	0.001	69.40 [60.80, 79.70]	71.20 [61.90, 81.40]	< 0.001
Non-HDL-C (mg/dL)	115.59 [100.90, 130.67]	162.76 [146.52, 184.02]	< 0.001	140.23 [122.63, 156.14]	193.93 [176.40, 215.77]	< 0.001
E-lbLDL-C (mg/dL)	73.25 [59.34, 88.17]	90.10 [74.18, 107.52]	< 0.001	85.17 [67.67, 102.44]	103.82 [86.65, 121.80]	< 0.001
E-sdLDL-C (mg/dL)	26.98 [22.67, 30.62]	43.36 [38.77, 50.75]	< 0.001	35.04 [29.56, 39.76]	54.66 [49.09, 62.52]	< 0.001
LDL-C/HDL-C	1.86 [1.47, 2.28]	2.84 [2.36, 3.44]	< 0.001	2.07 [1.71, 2.48]	3.12 [2.64, 3.66]	< 0.001
E-sdLDL-C/LDL-C	0.26 [0.22, 0.30]	0.34 [0.29, 0.40]	< 0.001	0.29 [0.25, 0.33]	0.35 [0.31, 0.40]	< 0.001
E-lbLDL-C/LDL-C	0.72 [0.68, 0.77]	0.68 [0.64, 0.73]	< 0.001	0.71 [0.67, 0.75]	0.65 [0.60, 0.69]	< 0.001
TRL-C (mg/dL)	12.76 [7.73, 19.33]	28.61 [18.94, 41.75]	< 0.001	19.51 [15.28, 25.81]	35.77 [27.75, 47.83]	< 0.001
CVD events (%)	706 (17.4)	918 (22.6)	< 0.001	17,830 (10.6)	19,914 (13.0)	< 0.001
Cardiac disease (%)	554 (13.7)	678 (16.7)	< 0.001	12,986 (7.7)	15,187 (9.9)	< 0.001
Stroke (%)	212 (5.2)	324 (8.0)	< 0.001	5993 (3.6)	5811 (3.8)	< 0.001
Current drinker (%)	1535 (37.9)	1306 (32.3)	< 0.001	155,308 (92.3)	140,984 (92.6)	0.048
Current smoker (%)	1376 (34.8)	1041 (26.2)	< 0.001	15,142 (9.0)	16,598 (10.9)	< 0.001
Townsend deprivation index	-	-		−2.19 [−3.68, 0.44]	−2.30 [−3.72, 0.19]	< 0.001
Physical activity (≥ 150 min/week)	-	-		66,755 (52.6)	57,528 (51.3)	< 0.001
Medical history	-	-
Hypertension (%)	1485 (36.8)	1973 (49.0)	< 0.001	45,148 (26.8)	37,375 (24.5)	< 0.001
Diabetes (%)	425 (10.6)	752 (18.8)	< 0.001	5994 (3.6)	1491 (1.0)	< 0.001
Chronic lung disease (%)	372 (9.2)	325 (8.1)	0.076
Arthritis (%)	1258 (31.1)	1344 (33.3)	0.036	2352 (1.4)	2063 (1.4)	0.300
Liver disease (%)	149 (3.7)	96 (2.4)	0.001	1805 (1.1)	1643 (1.1)	0.883
Kidney disease (%)	201 (5.0)	180 (4.5)	0.315	2244 (1.3)	1834 (1.2)	0.001
Digestive disease (%)	890 (22.0)	759 (18.9)	0.001
Asthma (%)	156 (3.9)	143 (3.6)	0.489	19,128 (11.3)	17,333 (11.4)	0.918
Medication
Cholesterol medication	-	-		5994 (3.6)	1491 (1.0)	< 0.001
Blood pressure medication	-	-		19,686 (11.7)	11,316 (7.4)	< 0.001
Insulin	-	-		1469 (0.9)	199 (0.1)	< 0.001

Data are represented as the median (Q1, Q3) or number (frequency). Categorical variables were compared using the Pearson χ² test, and continuous variables were compared using the Wilcoxon rank-sum test

Association of elevated E-sdldl-C with incident CVD

Kaplan–Meier curves demonstrated distinct differences in cumulative cardiovascular disease (CVD) incidence across lipid-related markers in both cohorts (Fig. 3). In the CHARLS cohort (panels A–D), participants with elevated E-sdLDL-C levels presented significantly greater cumulative CVD risk (log-rank P < 0.05). Similar trends were observed for non-HDL-C and TRL-C (P < 0.05), whereas direct LDL-C exhibited comparable values across the two categories (P = 0.13). In the UK Biobank cohort (panels E–I), all five markers—including E-sdLDL-C, direct LDL-C, ApoB, non-HDL-C, and TRL-C—showed strong associations with elevated cumulative CVD risk among participants with higher E-sdLDL-C concentrations, supported by log-rank significance P values below 0.0001 across all comparisons. These results were subsequently validated in terms of the risk of cardiac events and stroke separately (Figs. 4 and 5).

Fig. 3.

Kaplan–Meier curves for cumulative cardiovascular disease incidence by lipid-related markers across CHARLS and UK Biobank cohorts (A–D): CHARLS cohort; (E–I): UK Biobank. High vs. low groups were defined according to cut-off values specific to each lipid marker

Fig. 4.

Kaplan–Meier curves for the cumulative incidence of cardiac disease by lipid indices in the CHARLS and UK Biobank cohorts (A–D): CHARLS cohort; (E–I): UK Biobank cohort. High vs. low groups were defined according to cut-off values specific to each lipid marker

Fig. 5.

Kaplan–Meier curves for the cumulative incidence of stroke by lipid indices in the CHARLS and UK Biobank cohorts (A–D) CHARLS cohort; (E–I) UK Biobank cohort. High vs. low groups were defined according to cut-off values specific to each lipid marker

The Cox models revealed a significant association between elevated E-sdLDL-C and CVD outcomes in both CHARLS and UK Biobank cohorts (Table 2). In the CHARLS cohort, the multivariable-adjusted hazard ratios (HRs) for high vs. low E-sdLDL-C were 1.30 (95% CI 1.18–1.44) in model 1, 1.22 (95% CI 1.10–1.35) in model 2, and 1.21 (95% CI 1.09–1.34) in model 3. The association with stroke remained significant across all the models (model 3 h 1.42; 95% CI 1.17–1.71; P < 0.001), whereas no significant association was observed for cardiac events after full adjustment (HR 1.10; 95% CI 0.98–1.24; P = 0.110) (Table 2). In the UK Biobank cohort, individuals with elevated E-sdLDL-C consistently demonstrated elevated risk for CVD, cardiac outcomes, as well as stroke. The HRs were 1.24 (1.21–1.27) for CVD, 1.29 (1.26–1.33) for cardiac events, and 1.09 (1.04–1.14) for stroke in model 3.

Table 2.

Association between E-sdLDL-C and cardiovascular outcomes

	CHARLS				UK Biobank
	Case/Total		HR (95% CI)	P value	Case/Total		HR (95% CI)	P value
CVD	918/4055	Model 1	1.3 (1.18, 1.44)	< 0.001	19,914/152,660	Model 1	1.19 (1.17, 1.22)	< 0.001
		Model 2	1.22 (1.1, 1.35)	< 0.001		Model 2	1.14 (1.12, 1.16)	< 0.001
		Model 3	1.21 (1.09, 1.34)	< 0.001		Model 3	1.24 (1.21, 1.27)	< 0.001
Cardiac disease	678/4055	Model 1	1.18 (1.06, 1.33)	0.003	15,187/152,660	Model 1	1.25 (1.22, 1.28)	< 0.001
		Model 2	1.09 (0.97, 1.23)	0.131		Model 2	1.19 (1.16, 1.21)	< 0.001
		Model 3	1.1 (0.98, 1.24)	0.110		Model 3	1.29 (1.26, 1.33)	< 0.001
Stroke	324/4055	Model 1	1.62 (1.36, 1.93)	< 0.001	5811/152,660	Model 1	1.03 (0.99, 1.06)	0.169
		Model 2	1.53 (1.28, 1.84)	< 0.001		Model 2	1 (0.96, 1.04)	0.956
		Model 3	1.42 (1.17, 1.71)	< 0.001		Model 3	1.09 (1.04, 1.14)	< 0.001

HRs and 95% CIs were derived using Cox regression to compare individuals with elevated versus reduced E-sdLDL-C levels, based on cohort-specific thresholds (34.44 mg/dL for CHARLS; 44.45 mg/dL for UK Biobank)

Model 1 was adjusted for age, sex, and ethnicity (UK Biobank only)

Model 2 was additionally included BMI, smoking, and alcohol

Model 3 further included cohort-specific comorbidities and socioeconomic factors (see Methods)

To further evaluate the independent prognostic utility of E-sdLDL-C, additional models were constructed by incorporating each traditional lipid parameter (LDL-C, non-HDL-C, TRL-C, or ApoB) into model 3 (Table 3). In both cohorts, the associations between E-sdLDL-C and CVD remained statistically significant after these adjustments, although the effect sizes were slightly attenuated (HRs ranging from 1.08 to 1.30).

Table 3.

Relationship of E-sdLDL-C with cardiovascular risk after additional adjustment for traditional lipid markers

	CHARLS				UK Biobank
	CVD		Stroke		CVD		Cardiac disease		Stroke
	HR (95% CI)	P value	HR (95% CI)	P value	HR (95% CI)	P value	HR (95% CI)	P value	HR (95% CI)	P value
Model 3	1.21 (1.09, 1.34)	< 0.001	1.42 (1.17, 1.71)	< 0.001	1.24 (1.21, 1.27)	< 0.001	1.29 (1.26, 1.33)	< 0.001	1.09 (1.04, 1.14)	< 0.001
Model 3 + direct LDL-C	1.22 (1.08, 1.37)	0.001	1.43 (1.16, 1.75)	0.001	1.12 (1.09, 1.16)	<0.001	1.15 (1.11, 1.2)	<0.001	1.04 (0.98, 1.1)	0.221
Model 3 + non-HDL-C	1.3 (1.13, 1.5)	< 0.001	1.35 (1.06, 1.73)	0.017	1.08 (1.04, 1.12)	<0.001	1.1 (1.05, 1.15)	<0.001	1.02 (0.96, 1.09)	0.527
Model 3 + TRL-C	1.24 (1.1, 1.39)	< 0.001	1.37 (1.12, 1.69)	0.002	1.19 (1.15, 1.22)	<0.001	1.22 (1.18, 1.26)	<0.001	1.07 (1.02, 1.13)	0.009
Model 3 + ApoB	-	-	-	-	1.08 (1.04, 1.12)	<0.001	1.09 (1.05, 1.14)	<0.001	1.02 (0.96, 1.08)	0.587

HRs and 95% CIs were derived from Cox regression models. Each model included E-sdLDL-C and one additional lipid marker (LDL-C, non-HDL-C, TRL-C, or ApoB) from each cohort and was based on model 3. Model 3 was adjusted for age, sex, ethnicity, BMI, smoking status, alcohol consumption, cohort-specific comorbidities and socioeconomic factors (see Methods)

Each lipid parameter was also evaluated in a separate model containing the same covariates to allow direct comparison of their predictive strength (Table 4). E-sdLDL-C had the highest hazard ratio for CVD risk among all the lipid markers in both cohorts, indicating that E-sdLDL-C outperformed traditional lipid markers in terms of predictive value.

Table 4.

Association between individual lipid markers and cardiovascular risk

	CHARLS			UK Biobank
	Cases/Total	HR (95% CI)	P value	Cases/Total	HR (95% CI)	P value
E-sdLDL-C	918/4055	1.21 (1.09, 1.34)	< 0.001	19,914/152,660	1.24 (1.21, 1.27)	< 0.001
direct LDL-C	827/3976	1.05 (0.94, 1.16)	0.390	19,071/158,425	1.21 (1.18, 1.24)	< 0.001
Non-HDL-C	866/3963	1.12 (1.01, 1.25)	0.025	19,288/155,651	1.22 (1.19, 1.25)	< 0.001
TRL-C	913/4148	1.13 (1.02,1.26)	0.019	21,084/148,665	1.17 (1.14, 1.20)	< 0.001
ApoB	-	-	-	18,800/149,179	1.23 (1.20, 1.26)	< 0.001

Separate Cox proportional hazards models were constructed for each lipid marker (E-sdLDL-C, LDL-C, non-HDL-C, TRL-C, or ApoB) using covariates from each cohort. The model included adjustments for age, sex, and ethnicity (UK Biobank only), BMI, smoking, alcohol, and cohort-specific comorbidities and socioeconomic indicators (see Methods)

Subgroup analyses

Subgroup evaluations were conducted to evaluate the stability of the association between E-sdLDL-C concentrations with CVD within different strata (Table 5). In the CHARLS cohort, a stronger association was observed among male participants (HR 1.35; 95% CI 1.15–1.59; P < 0.001) than in female participants (HR 1.08; 95% CI 0.94–1.24; P = 0.29), with a statistically significant interaction (P for interaction = 0.042). In the UK Biobank cohort, similar trends were observed, where risk estimates were HR: 1.26 (95% CI 1.22–1.30) in men and HR: 1.19 (95% CI 1.14–1.24) in women, with both P < 0.001 and an interaction P value of 0.02. After age stratification, the association remained significant across age groups, with a stronger effect size observed in younger participants in the UK Biobank cohort (HR: 1.40 for age < 60 vs. 1.15 for age ≥ 60; P for interaction < 0.001), while the CHARLS cohort showed no significant interaction (P for interaction = 0.549). Subgroup analyses by hypertension, diabetes, and cholesterol-lowering medication use showed consistent associations, with no significant interactions observed in the CHARLS cohort. However, significant interactions were noted for hypertension status in the UK Biobank cohort (P for interaction < 0.001), where the relationship was slightly stronger in those without hypertension (HR 1.29) than in those with hypertension (HR 1.14).

Table 5.

Subgroup analysis of the associations between E-sdLDL-C and cardiovascular events in the CHARLS and UK biobank cohorts

Characteristic	CHARLS				UK Biobank
	No. of participants	HR (95% CI)	P value	P for interaction	No. of participants	HR (95% CI)	P value	P for interaction
Sex				0.042				0.02
Female	4218	1.08 (0.94, 1.24)	0.29		175,149	1.19 (1.14, 1.24)	< 0.001
Male	3894	1.35 (1.15, 1.59)	< 0.001		146,161	1.26 (1.22, 1.30)	< 0.001
Age (years)				0.549				< 0.001
≥ 60	3431	1.15 (0.99, 1.33)	0.07		144,150	1.15 (1.11, 1.18)	< 0.001
< 60	4681	1.23 (1.06, 1.43)	0.006		177,160	1.40 (1.34, 1.46)	< 0.001
Hypertension				0.205				< 0.001
No	4606	1.28 (1.09, 1.49)	0.002		238,787	1.29 (1.25, 1.33)	< 0.001
Yes	3458	1.10 (0.95, 1.27)	0.185		82,523	1.14 (1.10, 1.19)	< 0.001
Diabetes				0.086				0.617
No	6846	1.23 (1.10, 1.38)	< 0.001		313,825	1.24 (1.21, 1.27)	< 0.001
Yes	1177	0.95 (0.72, 1.25)	0.722		7485	1.28 (1.12, 1.48)	< 0.001
Cholesterol medication								0.617
No	-				313,825	1.24 (1.21, 1.27)	< 0.001
Yes	-				7485	1.28 (1.12, 1.48)	< 0.001

HRs and 95% CIs were derived from Cox regression models. The models included adjustment for age, sex, BMI, smoking status, drinking status, and other relevant covariates unless the variable was used for stratification. P values for interactions were derived from multiplicative interaction terms between E-sdLDL-C status and the stratified variable

Incremental predictive value of E-sdldl-C over conventional lipid parameters

Model performance was evaluated using Harrell’s C-index and continuous net reclassification improvement (NRI) in both cohorts (Table 6). In the UK Biobank, adding E-sdLDL-C to the model significantly improved C-index for CVD (0.689 to 0.692), cardiac events (0.698 to 0.703), and stroke (0.680 to 0.680), all P < 0.001. Continuous NRI further supported the incremental value of E-sdLDL-C, with significant improvement for CVD (15.29%, 95% CI 13.93–16.56), cardiac events (17.65%, 95% CI 16.13–19.10), and stroke (6.61%, 95% CI 4.48–8.90); all P < 0.001. In comparison, direct LDL-C and TRL-C showed minimal or inconsistent improvement in C-index and NRI, while ApoB demonstrated modest predictive value, generally lower than that of E-sdLDL-C.

Table 6.

Comparison of C-index and net reclassification improvement (NRI) for E-sdLDL-C and conventional lipid parameters

	CHARLS				UK Biobank
	Harrell’s C-index (95% CI)	P values	Continuous NRI, % (95% CI)	P values	Harrell’s C-index (95% CI)	P values	Continuous NRI, % (95% CI)	P values
CVD
Basic model	0.632 (0.617, 0.646)				0.689 (0.686, 0.692)
Basic model + E-sdLDL-C	0.633 (0.618, 0.647)	0.046	11.61 (5.73, 17.08)	< 0.001	0.692 (0.689, 0.695)	<0.001	15.29 (13.93, 16.56)	<0.001
Basic model + direct LDL-C	0.632 (0.617, 0.646)	0.267	2.49 (−2.95, 7.96)	0.378	0.691 (0.688, 0.694)	<0.001	13.07 (11.80, 14.42)	<0.001
Basic model + TRL-C	0.632 (0.617, 0.646)	0.5343	7.11 (1.67, 12.26)	0.007	0.691 (0.688, 0.694)	< 0.001	11.64 (10.42, 12.95)	<0.001
Basic model + ApoB	-	-	-	-	0.692 (0.689, 0.695)	< 0.001	14.17 (12.96, 15.48)	<0.001
Cardiac disease
Basic model	0.631 (0.615, 0.648)				0.698 (0.695, 0.702)
Basic model + E-sdLDL-C	0.631 (0.615, 0.648)	0.8237	8.18 (2.05, 14.38)	0.009	0.703 (0.699, 0.706)	< 0.001	17.65 (16.13, 19.10)	< 0.001
Basic model + direct LDL-C	0.631 (0.615, 0.648)	0.4094	2.37 (−3.51, 8.23)	0.444	0.701 (0.698, 0.705)	< 0.001	14.77 (13.33, 16.33)	< 0.001
Basic model + TRL-C	0.631 (0.615, 0.648)	0.4879	−4.08 (−10.09, 1.66)	0.193	0.701 (0.697, 0.704)	< 0.001	13.55 (12.14, 14.95)	< 0.001
Basic model + ApoB	-	-	-	-	0.702 (0.699, 0.706)	< 0.001	16.24 (14.87, 17.70)	< 0.001
Stroke
Basic model	0.672 (0.65, 0.694)				0.68 (0.674, 0.685)		7.07 (4.88, 9.20)
Basic model + E-sdLDL-C	0.677 (0.654, 0.699)	< 0.001	16.49 (7.27, 25.6)	< 0.001	0.68 (0.675, 0.686)	< 0.001	6.61 (4.48, 8.90)	< 0.001
Basic model + direct LDL-C	0.672 (0.65, 0.695)	0.1559	4.66 (−4.07, 13.48)	0.309	0.68 (0.675, 0.686)	< 0.001	6.18 (4.03, 8.42)	< 0.001
Basic model + TRL-C	0.674 (0.652, 0.697)	0.031	13.61 (4.97, 22.94)	0.003	0.68 (0.674, 0.685)	0.001	6.18 (4.03, 8.42)	< 0.001
Basic model + ApoB	-	-	-	-	0.68 (0.675, 0.686)	< 0.001	6.85 (4.65, 9.18)	< 0.001

The basic model included age, sex, ethnicity, BMI, smoking status, alcohol consumption, cohort-specific comorbidities and socioeconomic factors (see Methods), consistent with model 3 used in the previous analysis. P values for C-index reflect model improvement based on likelihood ratio tests comparing nested Cox models with and without each lipid parameter. NRI and its 95% confidence interval were estimated using bootstrap resampling with 1,000 replications

Similar patterns were observed in CHARLS, where E-sdLDL-C modestly improved C-index for CVD (0.632 to 0.633, P = 0.046), with no significant changes for cardiac disease or stroke. Continuous NRI remained positive for all outcomes, particularly for stroke (16.49%, 95% CI 7.27–25.6, P < 0.001). In contrast, conventional lipid markers provided limited or no improvement in model performance. These findings indicate that E-sdLDL-C offers incremental predictive value beyond traditional lipids in cardiovascular risk stratification.

Sensitivity analysis

Cross-classification models

As a sensitivity analysis, cross-classification of E-sdLDL-C and traditional lipid indices (direct LDL-C, non-HDL-C, and TRL-C) was conducted by categorizing them into four groups based on high/low levels, with the low/low group used as the reference. The joint associations of E-sdLDL-C and conventional lipid indexes with incident CVD were explored by four-category classification (Table 7). In both two cohorts, participants with concurrently increased E-sdLDL-C and conventional lipid levels presented the highest CVD risk compared with participants presenting simultaneously low levels.

Table 7.

Joint associations of E-sdLDL-C and conventional lipid parameters with cardiovascular risk in the CHARLS and UK biobank cohorts

		CHARLS			UK Biobank
		Cases/Total	HR (95% CI)	P value	Cases/Total	HR (95% CI)	P value
LDL-C	Low E-sdLDL-C/Low LDL-C	520/2937	REF		15,019/136,509	REF
	Low E-sdLDL-C/High LDL-C	186/1121	0.93 (0.79, 1.11)	0.441	2789/31,879	1.07 (1.02, 1.12)	0.010
	High E-sdLDL-C/Low LDL-C	277/1199	1.2 (1.02, 1.4)	0.023	3602/25,909	1.17 (1.12, 1.22)	< 0.001
	High E-sdLDL-C/High LDL-C	641/2856	1.18 (1.04, 1.33)	0.009	16,282/126,546	1.31 (1.27, 1.35)	< 0.001
Non-HDL-C	Low E-sdLDL-C/Low non-HDL-C	619/3493	REF		15,816/146,077	REF
	Low E-sdLDL-C/High non-HDL-C	87/564	0.88 (0.7, 1.11)	0.272	2014/22,573	1.06 (1, 1.12)	0.033
	High E-sdLDL-C/Low non-HDL-C	139/656	1.13 (0.93, 1.37)	0.223	2640/19,582	1.13 (1.08, 1.19)	< 0.001
	High E-sdLDL-C/High non-HDL-C	779/3399	1.2 (1.07, 1.34)	0.002	17,274/133,078	1.3 (1.27, 1.34)	< 0.001
TRL-C	Low E-sdLDL-C/Low TRL-C	559/3288	REF		12,521/134,295	REF
	Low E-sdLDL-C/High TRL-C	147/769	1.08 (0.89, 1.31)	0.439	5309/34,355	1.04 (1, 1.08)	0.068
	High E-sdLDL-C/Low TRL-C	152/676	1.26 (1.05, 1.52)	0.013	4139/38,350	1.2 (1.15, 1.25)	< 0.001
	High E-sdLDL-C/High TRL-C	766/3379	1.22 (1.08, 1.37)	0.001	15,775/114,310	1.3 (1.26, 1.34)	< 0.001
ApoB	Low E-sdLDL-C/Low ApoB	-	-	-	15,166/141,758	REF
	Low E-sdLDL-C/High ApoB	-	-	-	2586/26,160	1.12 (1.07, 1.18)	< 0.001
	High E-sdLDL-C/Low ApoB	-	-	-	3574/28,859	1.14 (1.1, 1.2)	< 0.001
	High E-sdLDL-C/High ApoB	-	-	-	16,214/123,019	1.32 (1.29, 1.36)	<0.001

Individuals were classified into four groups based on the combination of E-sdLDL-C (low or high) and each conventional lipid parameter (LDL-C, non-HDL-C, TRL-C, and ApoB) via study-specific cut-off values derived from ROC analysis (cut-off values detailed in Results). The group with both low E-sdLDL-C and low lipid parameters served as the reference (REF). The multivariable Cox regression models were adjusted for the covariates defined in model 3

In the CHARLS cohort, CVD risk was notably higher among individuals exhibiting high E-sdLDL-C alongside reduced LDL-C, as well as in those with elevated E-sdLDL-C and low TRL-C. In contrast, elevated traditional lipid levels in the context of low E-sdLDL-C were unrelated to significant risk increases, highlighting the potentially greater predictive contribution of E-sdLDL-C in this population.

In the UK Biobank cohort, a consistent increasing trend in hazard ratios was observed across the four groups. The greatest risk appeared in individuals with both high E-sdLDL-C and elevated conventional lipids, followed by those with high E-sdLDL-C but low levels of traditional lipids, and then by participants with low E-sdLDL-C but increased conventional lipid concentrations, compared with the reference group (low E-sdLDL-C/low conventional lipid). This pattern was observed across all lipid strata, indicating that elevated E-sdLDL-C may confer a greater cardiovascular risk than elevated levels of traditional lipid markers alone.

Quartile-based analysis

Participants were stratified into quartiles according to E-sdLDL-C concentrations or subsequent analysis, as detailed below: CHARLS: Q1 (< 26.97 mg/dL), Q2 (26.97–34.44 mg/dL), Q3 (34.44–43.36 mg/dL), and Q4 (≥ 43.36 mg/dL); UK Biobank: Q1 (< 34.58 mg/dL), Q2 (34.58–43.55 mg/dL), Q3 (43.55–54.00 mg/dL), and Q4 (≥ 54.00 mg/dL). The results derived from E-sdLDL-C concentration quartiles were consistent with the main findings. A distinct stepwise pattern was evident in the UK Biobank cohort (Fig. 6), showing substantially increased CVD incidence in those with elevated E-sdLDL-C concentrations compared to individuals in the lower quartiles (log-rank P < 0.0001) (panels D–F). In the CHARLS cohort, although the trend was generally consistent, separation was observed mainly between the highest quartiles (Q3 and Q4) and the lowest quartiles (Q1) (panels A–C).

Fig. 6.

Kaplan-Meier curves for the cumulative incidence of cardiovascular outcomes across E-sdLDL-C quartiles. Panels A–C: CHARLS; panels D–F: UK Biobank. A and D: Cardiovascular disease; B and E: cardiac disease; C and F: stroke. E-sdLDL-C was categorized into quartiles (Q1–Q4)

Using multivariable Cox regression models with adjustment for covariates defined in model 3 (Table 8), it was observed that, in the CHARLS cohort, hazard ratios for CVD and stroke were significantly greater in Q3 and Q4 than in Q1, whereas no significant associations were found for cardiac diseases across quartiles. In the UK Biobank cohort, hazard ratios generally increased with higher quartiles, and significant associations were observed for Q3 and Q4 for most outcomes. For ischaemic heart disease, no significant association was found in Q2 compared with Q1. These findings remained generally in accordance with the initial results, supporting the stability of the observed associations.

Table 8.

Association between E-sdLDL-C quartiles and the risk of cardiovascular outcomes in the CHARLS and UK biobank cohorts

Events	Quartile of E-sdLDL-C	CHARLS				UK Biobank
		Cases/Total	Incidence rate, per 1000 person-years	HR (95% CI)	P value	Cases/Total	Incidence rate, per 1000 person-years	HR (95% CI)	P value
CVD	Q1	345/2028	27.3	Ref		8216/80,328	7.5	Ref
	Q2	360/2028	28.6	1.00 (0.86, 1.17)	0.965	8680/80,327	7.9	1.04 (1.01, 1.08)	0.017
	Q3	454/2028	36.1	1.22 (1.05, 1.42)	0.009	9612/80,327	8.8	1.17 (1.13, 1.21)	< 0.001
	Q4	465/2028	37	1.21 (1.04, 1.4)	0.014	11,236/80,328	10.4	1.38 (1.34, 1.43)	< 0.001
Cardiac disease	Q1	279/2028	22.0	Ref		2780/80,328	2.4	Ref
	Q2	274/2028	21.6	0.92 (0.78, 1.1)	0.359	2896/80,327	2.5	1.05 (0.99, 1.12)	0.115
	Q3	349/2028	27.6	1.12 (0.95, 1.32)	0.194	3033/80,327	2.7	1.09 (1.03, 1.16)	0.006
	Q4	330/2028	26.1	1.01 (0.85, 1.19)	0.938	3095/80,328	2.7	1.15 (1.08, 1.23)	< 0.001
Stroke	Q1	92/2028	7.06	Ref		5968/80,328	5.4	Ref
	Q2	120/2028	9.19	1.33 (1, 1.77)	0.051	6328/80,327	5.7	1.05 (1.01, 1.1)	0.02
	Q3	138/2028	10.5	1.44 (1.09, 1.92)	0.011	7165/80,327	6.5	1.2 (1.15, 1.25)	< 0.001
	Q4	186/2028	14.3	1.88 (1.43, 2.47)	< 0.001	8712/80,328	7.9	1.47 (1.41, 1.53)	< 0.001

HRs and 95% CIs for Q2-Q4 quartiles were derived using Cox regression models to evaluate the associations between E-sdLDL-C quartiles and incident cardiovascular disease, cardiac disease, and stroke, with Q1 used as the reference. For each cohort, model 3 was applied with adjustments for covariates defined separately in the CHARLS and UK Biobank cohorts, as detailed in the Methods section

Restricted cubic spline (RCS) models

The P-spline curve was plotted to evaluate potential nonlinear associations between E-sdLDL-C and CVD outcomes via restricted cubic spline regression analyses, with covariable adjustments according to model 3 specifications (Fig. 7). In the CHARLS cohort, the association between E-sdLDL-C and CVD (panel A) showed a significant nonlinear pattern (P for nonlinearity = 0.003), with the highest HR observed at approximately 45 mg/dL, followed by a decline at higher concentrations; however, the confidence intervals were wide in the higher range. No significant associations were found for cardiac diseases (panel B) (P for overall = 0.066). For stroke (panel C), a nonlinear trend emerged (P for nonlinearity = 0.032), with HRs increasing with increasing E-sdLDL-C. In the UK Biobank cohort, E-sdLDL-C showed a positive correlation with CVD (panel D), exhibiting a significant nonlinear association (P for nonlinearity = 0.023), with the risk rising steadily across the exposure range. For cardiac diseases (panel E), a significant overall association was observed (P < 0.001), but the trend was linear (P for nonlinearity = 0.127). A comparable linear relationship was observed for stroke (panel F) (P for nonlinearity = 0.567). The results further support the consistent relationship between E-sdLDL-C and CVD across populations and analytic strategies.

Fig. 7.

Restricted cubic spline regression curves for the associations between E-sdLDL-C levels and cardiovascular outcomes. Panels A–C: CHARLS; panels D–F: UK Biobank. A and D: Cardiovascular disease; B and E: cardiac disease; C and F: stroke. The models were adjusted for covariates in model 3 and incorporated four knots positioned at the 5th, 35th, 65th, and 95th values of E-sdLDL-C distribution, using the lowest value as reference. Solid lines represent HRs, and shaded areas represent 95% CI

Sensitivity analysis using NMR-measured sdldl-C

In the UK Biobank, a subset of participants (n = 179,315) had available data on sdLDL-C measured by nuclear magnetic resonance (NMR) spectroscopy, referred to as “Cholesterol in Small LDL”. As a sensitivity analysis, NMR-measured sdLDL-C was evaluated using a cohort-specific cutoff of 7.15 mg/dL derived from ROC analysis. Kaplan-Meier curves demonstrated significantly higher cumulative incidence of CVD and cardiac disease among participants with elevated NMR-measured sdLDL-C (log-rank P < 0.001). For stroke, the log-rank P value was 0.18, indicating a less distinct separation (Fig. 8). In Cox regression analyses adjusted according to model 3, elevated NMR-measured sdLDL-C was significantly associated with increased risk of cardiovascular disease (HR 1.25; 95% CI 1.21–1.29), cardiac disease (HR 1.29; 95% CI 1.24–1.34), and stroke (HR 1.12; 95% CI 1.05–1.18), with P values below 0.001 (Table 9). These results are consistent with the main analysis based on estimated sdLDL-C, indicating that the observed associations are not dependent on the method of sdLDL-C quantification.

Fig. 8.

Kaplan–Meier curves for cumulative incidence of cardiovascular disease, cardiac disease, and stroke according to NMR-measured sdLDL-C levels in the UK Biobank High vs. low groups were defined according to a cut-off of 7.15 mg/dL for NMR-measured sdLDL-C (Cholesterol in Small LDL). A: Cardiovascular disease; B: cardiac disease; C: stroke

Table 9.

Association of NMR-measured sdLDL-C with risk of cardiovascular disease, cardiac disease, and stroke in the UK biobank cohort

	Cases/Total		HR (95% CI)	P value
CVD	10,549/83,027	Model 1	1.15 (1.12, 1.19)	< 0.001
		Model 2	1.14 (1.11, 1.17)	< 0.001
		Model 3	1.25 (1.21, 1.29)	< 0.001
Cardiac disease	8031/83,027	Model 1	1.19 (1.15, 1.23)	< 0.001
		Model 2	1.17 (1.13, 1.21)	< 0.001
		Model 3	1.29 (1.24, 1.34)	< 0.001
Stroke	3075/83,027	Model 1	1.02 (0.98, 1.08)	0.333
		Model 2	1.02 (0.97, 1.07)	0.378
		Model 3	1.12 (1.05, 1.18)	< 0.001

Multivariable Cox regression analyses were performed using a cutoff of 7.15 mg/dL for NMR-measured sdLDL-C (Cholesterol in Small LDL) 

Model 1 was adjusted for age, sex, and ethnicity

Model 2 additionally included BMI, smoking, and alcohol

Model 3 further included comorbidities and socioeconomic factors (see Methods)

Discussion

In this large-scale, prospective analysis incorporating data from both CHARLS and the UK Biobank, a significant relationship was identified between higher E-sdLDL-C levels and elevated CVD risk. This link remained robust after adjusting for conventional lipid measures, including directly measured LDL-C, non-HDL-C, TRL-C, and ApoB, although hazard ratios were modestly attenuated. These findings indicate that E-sdLDL-C represents an independent indicator of cardiovascular risk, providing predictive value beyond that of standard lipid parameters. Moreover, compared with traditional lipid measures, E-sdLDL-C demonstrated consistently stronger associations with incident CVD, as evidenced by higher hazard ratios in fully adjusted models. This incremental predictive value was further supported by significant improvement in model performance, reflected by higher C-index and continuous NRI in both cohorts.

Notably, across CHARLS and the UK Biobank, higher E-sdLDL-C was more strongly associated with cardiovascular risk in men than in women, with significant sex interaction. Hormonal regulation and lipid metabolism likely contribute. Men in midlife tend to have a more atherogenic profile and higher sdLDL-C/LDL-C ratios, whereas women show a sharp rise after menopause [33]. The menopausal decline in estrogen impairs hepatic LDL receptor–mediated clearance, raises triglycerides, and increases the proportion of small, dense LDL [34]. Higher hepatic lipase activity in men promotes conversion of larger LDL to sdLDL and a smaller HDL particle profile, amplifying the atherogenic potential of sdLDL [35]. Transcriptomic evidence also points to sex-specific pathways: immune and inflammatory responses may be more prominent in females, whereas energy-metabolism pathways (e.g., porphyrin metabolism) appear more involved in males [36]. These sex-specific regulatory mechanisms may underlie the differential impact of E-sdLDL-C on cardiovascular outcomes observed in our study.

Besides, in analyses of subgroups stratified by cholesterol-lowering medication use, the link between elevated E-sdLDL-C levels and CVD remained consistent across statin treatment groups. This suggests that E-sdLDL-C may help identify residual cardiovascular risk not fully addressed by conventional lipid-lowering therapy, which aligns with previous reports showing that reductions in sdLDL-C in response to statin therapy are generally less pronounced than those observed for total LDL-C [37].

Multiple sensitivity analyses demonstrated broadly consistent associations, suggesting the robustness of the findings. Subgroup analyses using a four-category classification further revealed that individuals with high E-sdLDL-C had increased CVD risk even when conventional lipid levels were within the normal range, whereas elevated conventional lipid markers in the presence of low E-sdLDL-C did not confer similar risk. The restricted cubic spline method further confirmed a favorable link of E-sdLDL-C levels with CVD risk in both study populations.

Furthermore, we conducted a sensitivity analysis using NMR-based sdLDL-C, which was available only in the UK Biobank cohort. This biomarker was not included in the main analysis due to several practical limitations: the method is not widely accessible in clinical settings, involves higher cost, and lacks standardized reference ranges [14]. Moreover, the UK Biobank reported sample dilution in NMR assays, which may affect absolute concentrations. In our data, the correlation between E-sdLDL-C and NMR-sdLDL-C was moderate (Spearman r = 0.782) (Figure S4), likely reflecting both methodological differences and dilution effects. Nevertheless, the NMR-based analysis showed associations with cardiovascular outcomes that were consistent with the main findings, thereby supporting the robustness of our results.

These findings align with previous evidence indicating that sdLDL particles are more atherogenic due to their small size, prolonged circulation time in the bloodstream, heightened susceptibility to oxidation and glycation, and enhanced arterial wall penetration [38, 39]. Notably, the pattern of associations differed between stroke and cardiac outcomes across the two cohorts. In the CHARLS cohort, the association between E-sdLDL-C and stroke appeared stronger than that with cardiac disease, whereas in the UK Biobank cohort, the association was more apparent for cardiac outcomes. Differences in outcome definitions may partially account for this discrepancy. In the CHARLS cohort, “cardiac disease” was defined broadly, encompassing myocardial infarction, coronary heart disease, angina, congestive heart failure, and other heart problems, which include heterogeneous conditions with varying aetiologies. In contrast, in the UK Biobank study, the outcome focused specifically on ischaemic heart disease, a more homogeneous and mechanistically relevant endpoint for assessing atherosclerotic risk. Given the well-established role of sdLDL in promoting atherosclerosis, the more specific definition of cardiac outcomes in the UK Biobank study may explain the stronger observed associations.

Although restricted cubic spline analysis revealed statistically significant nonlinearity, the overall association between E-sdLDL-C and CVD risk remained visually positive and monotonic, especially in the UK Biobank cohort. In the CHARLS cohort, the apparent deviation from linearity may be attributed to wider confidence intervals at higher E-sdLDL-C levels owing to the limited sample size. The differences highlight the influence of population distribution on risk curve interpretation.

Considering differences in lipid distributions and sample collection protocols between the CHARLS and UK Biobank cohorts (fasting vs. nonfasting samples), cohort-specific cut-off values for defining elevated E-sdLDL-C were applied based on ROC analysis within each dataset. This approach improves internal validity by optimizing the discrimination of cardiovascular outcomes in each cohort. However, it may reduce the direct comparability of effect estimates between populations. To address this, sensitivity analyses were performed using quartile-based classification of E-sdLDL-C. The consistent positive trends observed across quartiles in both cohorts suggest that the observed associations are not solely dependent on the specific thresholds used and support the robustness of the findings under alternative exposure definitions.

These findings are in agreement with findings reported in earlier literature, which consistently highlights the predictive role of sdLDL-C in CVD risk assessment. As demonstrated in a longitudinal analysis by Ikezaki et al. [40], sdLDL-C emerged as the strongest lipid predictor for incident atherosclerotic cardiovascular disease, outperforming conventional lipid parameters in risk prediction, including LDL triglycerides, triglycerides, TRL-C, and direct LDL-C. A prospective case‒cohort study by Duran et al. [7] reported that sdLDL-C was significantly related to incident myocardial infarction but not to ischaemic stroke. This vascular bed–specific association is partly echoed in the present findings: although E-sdLDL-C showed a notable correlation with both ischaemic heart disease and stroke in the UK Biobank cohort, the effect size was lower for stroke (HR 1.09) than for ischaemic heart disease (HR 1.29). These findings suggest that E-sdLDL-C may exert more pronounced atherogenic influence on coronary arteries than on cerebral vessels.

As most previous prospective studies have used direct laboratory assays to quantify sdLDL-C, the present study is distinct in that it uses an estimation equation reported by Sampson et al. [20] to derive E-sdLDL-C. Although several equations have been developed for estimating sdLDL-C from routine lipid measures [17, 19], the Sampson formula offers distinct advantages. It was derived from and validated in a large and diverse population (n = 20,171) and does not require direct measurements of LDL-C, ApoB, or particle size, which enhances its applicability in large epidemiological cohorts. Recently, Endo et al. [41] conducted a 10-year cohort study in 17,963 Japanese individuals and demonstrated that higher E-sdLDL-C levels were significantly related to incident ischaemic heart disease, irrespective of LDL-C concentration. The study revealed that, compared with participants with concurrently low sdLDL-C and LDL-C, those with elevated sdLDL-C—regardless of their LDL-C status—had a nearly 1.5-fold greater risk of ischaemic heart disease. Similar patterns were observed in the UK Biobank cohort in the present study, where E-sdLDL-C demonstrated a more pronounced link to ischaemic cardiac events than did other lipid markers. This concordance supports the validity and potential clinical applicability of E-sdLDL-C as a reliable substitute for direct sdLDL-C in large-scale settings.

Strengths and limitations

This research features several notable merits. First, two extensive, population-based prospective cohorts—the CHARLS cohort in China and the UK Biobank cohort in the United Kingdom—were analysed, allowing evaluation of the link between E-sdLDL-C levels and incident CVD across diverse ethnicities and health care systems. Second, multiple cardiovascular endpoints, including composite CVD, ischaemic heart disease, and stroke, were examined, enabling a comprehensive assessment of E-sdLDL-C across various vascular outcomes. Third, the validated Sampson equation was applied to estimate sdLDL-C using routine lipid parameters, enhancing the feasibility and clinical applicability of the findings. This approach allows for practical and cost-effective assessment of sdLDL-C, as it requires only standard lipid panel measurements without the need for specialized equipment or additional reagents, making it especially suitable for use in resource-limited settings. The equation has been shown to outperform direct sdLDL-C in atherosclerotic cardiovascular disease risk prediction while requiring no specialized assays [20]. Finally, a broad range of sensitivity analyses—including stratified subgroup analyses, quartile-based categorization, cross-classification with traditional lipid markers, restricted cubic spline modelling, and additional analysis using NMR-measured sdLDL-C—were conducted and consistently confirmed the robustness of E-sdLDL-C as an independent predictor of cardiovascular events.

Several limitations should also be considered. The CHARLS dataset lacked data on ApoB concentrations and information on medications, especially lipid-lowering medications, which may lead to residual confounding. The outcomes differed between the cohorts: the CHARLS cohort relied on self-reported physician diagnoses, whereas the UK Biobank cohort used ICD-coded hospital records, potentially introducing classification inconsistencies. Furthermore, the fasting status varied across cohorts—with fasting samples in the CHARLS cohort and predominantly nonfasting samples in the UK Biobank cohort—which may influence lipid levels. Nonetheless, the consistent predictive associations observed in both settings support the generalizability and robustness of the findings.

Conclusion

This large-scale prospective analysis in two ethnically and geographically distinct cohorts demonstrated that higher E-sdLDL-C, derived from Sampson’s formula, was an independent predictor of CVD, irrespective of traditional lipid indices. This relationship was further validated by multiple complementary sensitivity analyses.

Given that E-sdLDL-C can be derived from standard lipid panels without additional testing, it offers a convenient, cost-effective, and scalable tool for enhancing cardiovascular risk assessment, particularly in settings where access to direct sdLDL-C measurement is limited.

These findings suggest that integrating E-sdLDL-C into routine lipid evaluations may help clinicians better identify high-risk individuals, including those with normal conventional lipid levels, among middle-aged and older adults, thereby supporting earlier and more targeted preventive interventions. Future studies should explore its clinical implementation and evaluate whether E-sdLDL-C–guided strategies can improve patient outcomes in both early intervention and post-diagnosis settings.

Abbreviations

ApoB

Apolipoprotein B

BMI

Body Mass Index

CVD

Cardiovascular disease

CHARLS

China Health and Retirement Longitudinal Study

CI

Confidence interval

E-sdLDL-C

Estimated small dense low-density lipoprotein cholesterol

HDL-C

High-density lipoprotein cholesterol

HR

Hazard ratio

ICD

International Classification of Diseases

LDL-C

Low-density lipoprotein cholesterol

Non-HDL-C

Non-high-density lipoprotein cholesterol

ROC

Receiver operating characteristic

sdLDL-C

Small dense low-density lipoprotein cholesterol

TG

Triglyceride

TC

Total cholesterol

TRL-C

Triglyceride-rich lipoprotein cholesterol

NRI

Net reclassification improvement

NMR

Nuclear magnetic resonance

Authors’ contributions

YM was responsible for designing the research, conducting data analysis, and drafting the manuscript. XZ contributed to the study design and provided methodological support in data processing. DM, YG, and YL contributed to data collection and assisted in manuscript preparation. XC conceived and supervised the entire study, critically revised the manuscript, and provided major intellectual input. KL and LQ provided critical insights, methodological guidance, and contributed significantly to revision. All authors examined and approved the final version.

Funding

The research received financial support from the Joint Research Funding Program between the Macau Science and Technology Development Fund and the Department of Science and Technology of Guangdong Province (0009/2024/AGJ) and the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2024ZD0533200).

Data availability

Data for analysis were obtained from publicly accessible sources, namely the CHARLS and UK Biobank databases. The CHARLS data is available for access at the following URL: https://charls.charlsdata.com/. Access to the UK Biobank dataset requires an approved application via https://www.ukbiobank.ac.uk/.

Declarations

Ethics approval and consent to participate

The original CHARLS study was reviewed and approved by the Ethics Committee at Peking University (IRB00001052-11015). UK Biobank received ethical clearance from the North West Multi-Centre Research Ethics Committee (REC: 11/NW/0382). And all individuals gave written consent before taking part.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.