Non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol is a predictor for cardiovascular mortality in patients with diabetes mellitus

Abstract

BACKGROUND

The non-high-density lipoprotein cholesterol (non-HDL-C)/HDL-C ratio (NHHR) is linked to cardiovascular event risk, but its prognostic value in diabetes mellitus (DM) patients remains unclear.

AIM

To explore the association of NHHR and cardiovascular mortality in patients with DM and generate predictive model.

METHODS

This cohort study analyzed data from 8425 DM patients in National Health and Nutrition Examination Survey. NHHR was calculated as (total cholesterol - HDL-C)/HDL-C. Cardiovascular mortality was determined via the National Death Index. Feature selection was performed using the Boruta algorithm and least absolute shrinkage and selection operator regression, followed by Cox proportional hazards models to evaluate NHHR’s relationship with cardiovascular mortality. Stratified and sensitivity analyses assessed the findings’ robustness. A nomogram was developed to predict cardiovascular mortality, with model performance evaluated using receiver operating characteristic curves and calibration plots.

RESULTS

Over an average follow-up of 94.2 months, 671 cardiovascular deaths (8.0%) occurred. Six key features including age, education, ethnicity, poverty-income ratio, history of heart failure, and NHHR, were selected. A non-linear association was found, with the highest NHHR quartile showing a 39% higher risk of cardiovascular death compared to the lowest quartile (Q4, hazard ratio: 1.39, 95% confidence interval: 1.11-1.73). Stratified analyses confirmed the increased risk across all subgroups. Sensitivity analyses supported the stability of the results. The nomogram for predicting cardiovascular mortality demonstrated high accuracy.

CONCLUSION

Elevated NHHR is associated with increased cardiovascular mortality risk in DM patients. NHHR could be a valuable prognostic marker, aiding in identifying high-risk patients and guiding targeted lipid management strategies.

Key Words: Lipid profile; Cardiovascular mortality; Diabetes; Predictive model; National Health and Nutrition Examination Survey

Core Tip: This study investigates the association between non-high-density lipoprotein cholesterol to high-density lipoprotein cholesterol ratio (NHHR) and cardiovascular mortality in patients with diabetes mellitus (DM). Analysis of 8425 DM patients from the National Health and Nutrition Examination Survey revealed a non-linear relationship between NHHR and cardiovascular mortality. Patients in the highest NHHR quartile (≥ 4.01) had a 39% higher risk of cardiovascular death compared to the lowest quartile. A predictive model for 5, 8, and 10-year cardiovascular mortality was developed, showing high accuracy. These findings suggest NHHR as a valuable prognostic marker for cardiovascular risk in DM patients.

INTRODUCTION

According to the International Diabetes Federation, the global number of patients with diabetes mellitus (DM) reached 537 million in 2021 and is projected to increase to 783 million by 2045[1]. In patients with DM, the risk of cardiovascular events and cardiovascular mortality is 2 to 4 times greater than that in patients without DM[2]. Therefore, it is urgent and crucial to identify and treat high-risk patients with DM to prevent cardiovascular events. Several cardiovascular risk stratification systems, such as the Framingham risk score and QRISK, have been developed. However, these stratification systems are limited by the use of a single race population (only white individuals), and stratification may not be possible for patients of other races[3,4]. Moreover, these two risk stratification systems focus on the general population instead of patients with DM, whose risk of cardiovascular mortality is high and whose pathological changes differ from those of the general population. In addition, the Framingham risk score only focuses on coronary artery disease rather than cardiovascular mortality, which may overlook residual risks. Another risk score, the Systematic Coronary Risk Evaluation 2-Diabetes score, is applied to perform 10-year cardiovascular risk stratification in patients with DM. Nevertheless, studies have also shown that this score is less predictive in Asian patients[5]. Thus, new predictive systems, developed by including DM patients of various races, are needed.

The lipid profile is an important prognostic indicator in patients with DM. Traditional lipid indicators, such as low-density lipoprotein cholesterol (LDL-C) and high-density lipoprotein cholesterol (HDL-C) levels, cannot capture the complexity of lipid metabolism and are thus not able to assess the risk of individual cardiovascular events comprehensively[6,7]. Non-HDL-C measures the total cholesterol (TC) content of all lipoproteins except HDL-C. It quantifies all serum lipoproteins, including apolipoprotein B, LDL-C, very-LDL-C, chylomicrons, and lipoprotein(a)[8]. The American College of Cardiology expert consensus committee has suggested that non-HDL-C could be used as a metric equivalent to LDL-C in patients with DM and elevated triglyceride levels[9]. At present, the non-HDL-C/HDL-C ratio (NHHR), which is a recently developed composite indicator of atherogenic lipids, encompasses information pertaining to both atherogenic and anti-atherogenic lipid particles and has attracted much interest[10,11].

As a lipid parameter that reflects the lipid profile, the NHHR has superior predictive value in assessing the risks of atherosclerosis[12], insulin resistance[13], metabolic syndrome[14], non-alcoholic fatty liver disease[15], and chronic kidney disease[16]. However, relatively few studies have investigated the relationship between the NHHR and cardiovascular mortality in patients with DM, especially the prognostic value of the NHHR for predicting cardiovascular mortality. Moreover, risk stratification in patients with DM, whose risk of cardiovascular mortality is high, is crucial. Thus, in this study, we aimed to explore the association between the NHHR and cardiovascular mortality in patients with DM and generate a predictive model to predict the risk of cardiovascular mortality.

MATERIALS AND METHODS

Study population

This cohort study included patients from ten National Health and Nutrition Examination Survey (NHANES) cycles (1999-2018). We included patients with DM aged 18 years or older. Additionally, we excluded patients with missing data on TC, HDL-C, and other features included in this study. Patients without a reported follow-up time or with a follow-up time of 0 months were also excluded. Moreover, after calculating the NHHR values, patients with an NHHR value of 0 or lower, which may indicate an error in the measurement, were excluded. All procedures were performed following the guidelines of the Declaration of Helsinki. Written informed consent was obtained from all participants.

Definition of DM

Patients who met one of the following criteria were defined as DM patients: (1) A diagnosis of DM confirmed by a physician; (2) The use of insulin or oral hypoglycemic agents; (3) A hemoglobin A1c (HbA1c) level of 6.5% or higher; and (4) A fasting plasma glucose level of 126 mg/dL (7.0 mmol/L) or higher (after at least 8 hours of fasting) or a 2-hour blood glucose level of 200 mg/dL (11.1 mmol/L) or higher following an oral glucose tolerance test[17].

Measurement of NHHR

Serum samples were collected from participants, and the TC and HDL-C levels were measured at recruitment. The NHHR was further calculated via the following equation: NHHR = (TC - HDL-C)/HDL-C[18].

Ascertainment of mortality

The cardiovascular mortality data were obtained via linkage with the National Death Index to gather information on deaths up to December 31, 2019. Cardiovascular mortality was specified via International Statistical Classification of Diseases and Related Health Problems, Tenth Revision codes I00-I09, I11, I13, I20-I51, and I60-I69.

Feature selection

The covariates were retrieved from the NHANES database, and we further selected the most relevant features from the large dataset. We first applied the Boruta algorithm, which uses the random forest classifier. This algorithm performs all-relevant feature selection by comparing the importance of the original features with that of randomly permuted copies (shadow features). During each iteration, the algorithm first creates shadow features by randomly shuffling the values of the original variables, effectively breaking their associations with the outcome. A random forest classifier is then trained with 1000 trees, and the normalized importance measure (Z score) is calculated for both the original and shadow features. A feature is considered important if its Z score significantly exceeds the maximum Z score achieved by the shadow features (P < 0.05). To ensure robustness, we repeated this process across 100 iterations. Features were classified as ‘confirmed’ if they consistently demonstrated greater importance than their shadow counterparts, ‘rejected’ if they failed to exceed the importance of their shadow counterparts in 95% of iterations, and ‘tentative’ if their status remained uncertain[19]. Overall, 11 features were selected via the Boruta algorithm.

Moreover, we applied least absolute shrinkage and selection operator (LASSO) regression to select relative features in the cohort. To increase the robustness of feature selection, we implemented 10-fold cross-validation in the LASSO procedure. The LASSO procedure minimizes the residual sum of squares with an added penalty for the absolute values of the regression coefficients, leading to some coefficients being reduced to zero, thereby identifying crucial features. The optimal penalty parameter was determined by the value that minimized the mean cross-validated error. This procedure offers several advantages over traditional methods, particularly in its ability to select variables even in the presence of collinearity among the data. The covariates selected in both the Boruta algorithm and LASSO regression were defined as key features and adjusted for in further analyses.

Statistical analysis

Given the complex sampling design of the NHANES, all analyses incorporated sample weights. For normally distributed continuous variables, data are expressed as the mean ± SD, whereas nonnormally distributed continuous variables are reported as medians and interquartile ranges. The normality of continuous variables was assessed via the Shapiro-Wilk test. Categorical variables are presented as n (%). Normally distributed continuous variable differences were evaluated via Student’s t-test, nonnormally distributed continuous variable differences were evaluated via the Mann-Whitney U test, and categorical variable differences were evaluated via Pearson’s χ²-test. Hazard ratios (HRs) and 95% confidence intervals (CIs) for the relationship between the NHHR and cardiovascular mortality were calculated via Cox proportional hazards models. The proportional hazards assumption was verified via the Schoenfeld residuals test, and the results revealed P > 0.05, indicating the appropriateness of the Cox proportional hazards model. The features selected by both LASSO regression and the Boruta algorithm were selected as covariates for inclusion in the fully adjusted model. The reference group was set as the group with NHHR values in the lowest quartile, for which the HR was set to 1. Additionally, the HR and 95%CI were also calculated per SD increase in the NHHR.

Stratified analyses were conducted by sex (male, female), age (≥ 60 years, < 60 years), body mass index (BMI) (≥ 30 kg/m², < 30 kg/m²), and duration of DM (≥ 10 years, < 10 years). Moreover, a restricted cubic spline (RCS) model was constructed to explore the relationships between the NHHR and cardiovascular mortality, with the median value of the NHHR used as a reference. An adjustment of the RCS was made for the aforementioned key features. The number of knots selected for the RCS model was determined by Akaike’s information criterion (AIC), with the model with the lowest AIC value selected. Finally, the RCS model was fitted with 5 knots placed at the 5^th, 27.5^th, 50^th, 72.5^th, and 95^th percentiles of the NHHR distribution.

Moreover, a series of sensitivity analyses were also performed to evaluate the stability of the model: (1) We excluded patients who died due to cardiovascular reasons within 1 year of recruitment; (2) We included all the covariates included in feature selection in the multivariable Cox regression; (3) We excluded patients with extreme NHHR values (mean ± SD); (4) We excluded patients with a history of coronary heart disease; and (5) We excluded patients with a history of heart failure. To explore the prognostic value of the NHHR for cardiovascular mortality at 5, 8, and 10 years in patients with DM, we randomly divided patients with DM into training and validation cohorts at a ratio of 7:3 and constructed a prediction nomogram for the NHHR. The performance of the nomogram was assessed via receiver operating characteristic (ROC) curves and calibration curves, along with decision curve analysis (DCA), with the area under the ROC curve (AUC) ranging from 0.5 (not discriminant) to 1 (completely discriminant). Moreover, an interactive web-based dynamic nomogram application was built via the Shiny package. Statistical significance was set at a two-sided P < 0.05. Analyses were performed via R software (version 4.4.0), along with MSTATA software (www.mstata.com).

RESULTS

Baseline characteristics

A total of 8425 patients with DM were included in the analysis (Figure 1). Among these patients, 4365 (51.8%) were male. The median age was 60 years. During the mean follow-up time of 94.2 months, a total of 2371 patients died. The median NHHR value for all participants was 2.94, with median values of 3.14 for males and 2.72 for females. Compared with female patients, male patients had higher fasting plasma glucose and triglyceride levels. In addition, male patients had lower TC, HDL-C, and LDL-C levels and a lower prevalence of hypertension and hypercholesterolemia. The baseline characteristics are shown in Table 1.

Figure 1 Study flowchart. NHANES: National Health and Nutrition Examination Survey; TC: Total cholesterol; HDL-C: High-density lipoprotein cholesterol; NHHR: Non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio; DM: Diabetes mellitus.

Table 1 Baseline characteristics, median (interquartile range).

	Male	Female	All participants	P value
Sample size	4365	4060	8425
Age, years	60 (50-69)	60 (49-71)	60 (50-70)	0.39
Body mass index	31.1 (27.2-35.3)	32.6 (28.1-38.6)	31.7 (27.6-36.7)	< 0.01
Ethnicity, n (%)
Mexican American	858 (19.7)	868 (21.4)	1726 (20.5)	< 0.01
Other Hispanic	370 (8.5)	415 (10.2)	785 (9.3)
Non-Hispanic white	1712 (39.2)	1356 (33.4)	3068 (36.4)
Non-Hispanic black	1006 (23.0)	1059 (26.1)	2065 (24.5)
Other	419 (9.6)	362 (8.9)	781 (9.3)
Education, n (%)
Less than 9th grade	1033 (23.7)	1023 (25.2)	2056 (24.4)	< 0.01
9-11th grade	716 (16.4)	726 (17.9)	1442 (17.1)
High school graduate	1097 (25.1)	996 (24.5)	2093 (24.8)
College	864 (19.8)	890 (21.9)	1754 (20.8)
College graduate or above	632 (14.5)	405 (10.0)	1037 (12.3)
Unknown	23 (0.5)	20 (0.5)	43 (0.5)
Poverty-income ratio
< 1	797 (18.3)	978 (24.1)	1775 (21.1)	< 0.01
1-3	2224 (51.0)	2182 (53.7)	4406 (52.3)
> 3	1344 (30.8)	900 (22.2)	2244 (26.6)
Smoke, n (%)
Every day	686 (15.7)	466 (11.5)	1152 (13.7)	< 0.01
Some days	138 (3.2)	85 (2.1)	223 (2.6)
Not at all	1861 (42.6)	986 (24.3)	2847 (33.8)
Unknown	1680 (38.5)	2523 (62.1)	4203 (49.9)
Alcohol use
Non-drinker	1971 (45.2)	2465 (60.7)	4436 (52.7)	< 0.01
Drinker	2393 (54.8)	1592 (39.2)	3985 (47.3)
Unknown	1 (0.02)	3 (0.07)	4 (0.05)
Hypertension, n (%)	2674 (61.3)	2695 (66.4)	5369 (63.7)	< 0.01
Hypercholesterolemia, n (%)	2406 (55.1)	2269 (55.9)	4675 (55.5)	0.49
Plasma fasting glucose, mmol/L	7.55 (6.83-9.66)	7.11 (6.16-8.62)	7.33 (6.44-9.10)	< 0.01
HbA1c, %	6.7 (6.1-7.7)	6.6 (6.0-7.5)	6.7 (6.0-7.6)	0.06
HbA1c, mmol/mol	50 (43-61)	49 (42-58)	50 (42-60)	0.06
NHHR	3.14 (2.28-4.24)	2.72 (1.99-3.77)	2.94 (2.13-4.01)	< 0.01
TC	4.50 (3.83-5.35)	4.91 (4.19-5.69)	4.71 (3.98-5.53)	< 0.01
HDL-C	1.09 (0.91-1.27)	1.27 (1.06-1.53)	1.16 (0.98-1.42)	< 0.01
LDL-C	2.43 (1.81-3.13)	2.69 (2.02-3.41)	2.53 (1.91-3.26)	< 0.01
Triglycerides	0.67 (0.02-1.69)	0.56 (0.02-1.69)	0.67 (0.02-1.69)	0.91

TC: Total cholesterol; HDL-C: High-density lipoprotein cholesterol; LDL-C: Low-density lipoprotein cholesterol; NHHR: Non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio.

Feature selection

Among all the features included in the analysis, the Boruta algorithm revealed that eleven important features had high Z values compared with the shadow features, indicating that they were more strongly associated with cardiovascular mortality among all the features (Figure 2A). Furthermore, to identify the features most closely related to cardiovascular mortality, we performed LASSO regression. The results revealed eight crucial features (Figure 2B and C). Combining the results of the Boruta algorithm and LASSO regression, six features were ultimately included in our subsequent analysis, including age, ethnicity, education level, poverty-income ratio, history of heart failure and NHHR value.

Figure 2 Feature selection by two machine learning method. A: Feature selection by Boruta algorithm, the vertical axis represents the Z-value of each variable, boxes with green are important features; B: The plot of the association between the cross-validation error and the log-transformed penalty parameter (λ) in the least absolute shrinkage and selection operator regression analysis; C: The least absolute shrinkage and selection operator regression screening path for the variables based on the optimal log of lambda. PIR: Poverty-income ratio; HYPER: Hypertension; CHD: Coronary heart disease; HF: Heart failure; NHHR: Non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio; BMI: Body mass index.

Association between the NHHR and cardiovascular mortality risk

A total of 671 of 8425 patients experienced cardiovascular death (8.0%) during the mean follow-up time of 94.2 months (Figure 3A). We first explored whether there was a linear association between the NHHR and cardiovascular mortality via the RCS model. The AIC indicated that 5 knots fit the model best (Supplementary Table 1), and according to the multivariable RCS model, we observed a nonlinear association between the NHHR and cardiovascular mortality (Figure 3B). We first categorized patients into four groups according to their NHHR, with Q1 representing the lowest NHHR values and Q4 representing the highest. When the selected features were included in the fully adjusted multivariable Cox regression model, we found that patients in Q4 (≥ 4.01) had a greater risk of cardiovascular mortality (HR: 1.39, 95%CI: 1.11-1.73). Furthermore, when the NHHR was treated as a continuous variable, we also discovered a greater risk of cardiovascular mortality in patients with greater NHHR values (per SD increase, HR: 1.14, 95%CI: 1.06-1.24) (Table 2).

Figure 3 Kaplan-Meier curve and restricted cubic splines of the analysis on non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio and cardiovascular mortality. A: Kaplan-Meier curve stratified by non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio (quartile 1 to quartile 4); B: Multiple-adjusted restricted cubic splines showing hazard ratios for the risk of incident cardiovascular mortality associated with remnant cholesterol. Blue solid lines represent hazard ratios, and shaded areas represent 95% confidence intervals. Analysis was adjusted for age, ethnicities, education levels, poverty-income ratio, history of heart failure and non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio. HR: Hazard ratios; CI: Confidence intervals; NHHR: Non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio.

Table 2 Association of non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio and cardiovascular mortality in patients with diabetes, hazard ratio (95% confidence intervals).

Cardiovascular mortality	Q1	Q2	Q3	Q4	Per SD
Model 1	Reference	0.86 (0.69-1.07)	0.92 (0.74-1.14)	0.86 (0.70-1.07)	0.94 (0.87-1.02)
Model 2	Reference	1.03 (0.82-1.28)	1.25 (1.01-1.55)	1.40 (1.12-1.75)	1.15 (1.06-1.24)
Model 3	Reference	1.02 (0.81-1.27)	1.24 (0.99-1.54)	1.39 (1.11-1.73)	1.14 (1.06-1.24)

Model 1: No adjustment. Model 2: Adjusted for age, ethnicities, education levels, poverty-income ratio. Model 3: Adjusted for age, ethnicities, education levels, poverty-income ratio, history of heart failure.

Subgroup analysis

To confirm the associations between the NHHR and cardiovascular mortality, we performed subgroup analyses stratified by sex, age, BMI, and duration of DM. We found that sex, age, and BMI did not affect the relationship between the NHHR and cardiovascular mortality. In contrast, the duration of DM was found to affect the relationship between the NHHR and cardiovascular mortality (P < 0.05) (Table 3).

Table 3 Subgroup analyses on non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio and cardiovascular mortality, hazard ratio (95% confidence intervals).

Cardiovascular mortality1	n	Q1	Q2	Q3	Q4	Per SD	P value for interaction
Gender
Male	4365	Reference	0.84 (0.61-1.14)	1.18 (0.89-1.57)	1.19 (0.88-1.59)	1.12 (1.01-1.25)	0.58
Female	4060	Reference	1.18 (0.85-1.63)	1.13 (0.80-1.59)	1.40 (0.99-1.97)	1.09 (0.95-1.24)
Age, years
≥ 60	5161	Reference	1.04 (0.83-1.32)	1.23 (0.97-1.54)	1.31 (1.04-1.67)	1.15 (1.04-1.26)	0.99
< 60	3264	Reference	1.03 (0.46-2.31)	1.46 (0.71-3.03)	1.89 (0.94-3.78)	1.10 (0.96-1.26)
Body mass index
≥ 30 kg/m2	4738	Reference	1.08 (0.79-1.48)	1.19 (0.87-1.63)	1.35 (0.98-1.85)	1.13 (1.02-1.26)	0.64
< 30 kg/m2	3687	Reference	0.93 (0.67-1.28)	1.30 (0.96-1.67)	1.44 (1.06-1.96)	1.16 (1.03-1.31)
Duration of diabetes
≥ 10 years	2865	Reference	1.06 (0.77-1.45)	1.17 (0.85-1.61)	1.60 (1.15-2.22)	1.26 (1.11-1.42)	0.04
< 10 years	5560	Reference	1.00 (0.72-1.37)	1.35 (1.00-1.82)	1.37 (1.01-1.86)	1.11 (1.00-1.24)

¹All adjusted for age, ethnicities, education levels, poverty-income ratio, history of heart failure.

Sensitivity analyses

To confirm the stability of our results, we conducted a series of sensitivity analyses. After excluding patients who experienced cardiovascular death within 3 years, a total of 8262 patients with DM were included, and the results remained unchanged (Q4, HR: 1.33, 95%CI: 1.03-1.71; HR: 1.13, 95%CI: 1.03-1.24, per SD increase) (Supplementary Table 2). When all the covariates were included in feature selection, the HRs of patients in Q4 remained significantly greater (Q4, HR: 1.24, 95%CI: 1.05-1.48; HR: 1.09, 95%CI: 1.01-1.18, per SD increase) (Supplementary Table 3). When excluding patients with extreme NHHR values (> 8.24) (n = 8331), patients with NHHR values in the highest quartile also had a higher risk of cardiovascular mortality (Q4, HR: 1.38, 95%CI: 1.10-1.72; HR: 1.16, 95%CI: 1.06-1.28, per SD increase) (Supplementary Table 4). Furthermore, when excluding patients with a history of heart failure or coronary heart disease, 7669 and 7576 patients were included, respectively, and the results remained unchanged (when excluding patients with heart failure, Q4, HR: 1.43, 95%CI: 1.11-1.85; HR: 1.16, 95%CI: 1.06-1.27, per SD; when excluding patients with coronary heart disease, Q4, HR: 1.46, 95%CI: 1.13-1.89; HR: 1.15, 95%CI: 1.05-1.25, per SD increase) (Supplementary Tables 5 and 6).

ROC curves for the NHHR and cardiovascular mortality

To assess the ability of the model to predict cardiovascular mortality, we included 6 independent predictors and developed a simple-to-use nomogram (Figure 4A, available online: https://pandorix.shinyapps.io/nhhr/). We generated ROC curves and calculated the AUCs for 5-, 8-, and 10-year cardiovascular mortality. According to the ROC curves, the AUCs of the predictive models were 0.832, 0.831, and 0.841 for 5-, 8-, and 10-year cardiovascular mortality, respectively, in the training cohort and 0.801, 0.794, and 0.787, respectively, in the validation cohort, indicating the high predictive accuracy of the models in estimating long-term cardiovascular mortality (Figure 4B and C). The calibration plots of the nomogram in the training and validation cohorts were also generated, which demonstrated good correlation between the observed and predicted risk (Figure 5). The results showed that the calibration curve of this model was relatively close to the ideal curve, which indicated that the predicted results were consistent with the actual findings.

Figure 4 Nomogram predicting cardiovascular mortality in patients with diabetes mellitus. A: Heart failure 1 indicates heart failure history, and 2 indicates no heart failure history. Race 1 indicates Mexican American, 2 indicates other Hispanic, 3 represents non-Hispanic white, 4 indicates non-Hispanic black, and 5 represents other ethnicities; Education (1-4 not shown in the nomogram but available online) 1 indicates less than 9^th grade, 2 indicates 9-11^th grade, 3 represents high school graduate, 4 indicates college, 5 represents college graduate or above, and 6 indicates unknown; B: Receiver operating characteristic curves of non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio (NHHR) in predicting cardiovascular mortality for 60, 96, and 120 months in training cohort; C: Receiver operating characteristic curves of NHHR in predicting cardiovascular mortality for 60, 96, and 120 months in validation cohort. PIR: Poverty-income ratio; HF: Heart failure; NHHR: Non-high-density lipoprotein cholesterol/high-density lipoprotein cholesterol ratio; AUC: Area under the curve.

Figure 5 Calibration curves of the nomogram prediction model for cardiovascular mortality in patients with diabetes mellitus. A-C: Calibration curve for 60, 96, and 120 months in the training set; D-F: Calibration curve for 60, 96, and 120 months in the validation set.

DISCUSSION

In this study, we analyzed data from 8425 patients with DM to investigate the association between the NHHR and cardiovascular mortality and generate a predictive model for the risk of cardiovascular mortality. Over a mean follow-up duration of 94.2 months, we found that there was a nonlinear relationship between the NHHR and cardiovascular mortality, with patients in the highest quartile (NHHR ≥ 4.01) having a 39% greater risk of cardiovascular death than patients in the lowest quartile. The results remain robust in subgroup analyses and a series of sensitivity analyses. Moreover, we generated a predictive model for cardiovascular mortality at 5, 8, and 10 years, which showed high predictive accuracy. These findings underscore the potential of the NHHR as a valuable prognostic marker for cardiovascular risk in patients with DM.

HDL-C plays a crucial role in reverse cholesterol transport, a process in which excess cholesterol is removed from peripheral tissues and transported back to the liver for excretion. This mechanism helps prevent the accumulation of cholesterol in arterial walls, thereby reducing the risk of cardiovascular mortality[20]. Compared with that in the general population, the NHHR in patients with DM is greater. Studies have shown that patients with diabetes have lipid metabolism dysfunction, characterized by elevated non-HDL-C and reduced HDL-C levels even in patients with well-controlled glucose levels[21,22]. A previous study indicated that in patients with DM, the NHHR was approximately 30% higher than that in patients without DM, and patients with higher NHHR values had a greater risk of developing DM (odds ratio: 1.08, 95%CI: 1.01-1.15), which reflects the underlying metabolic disturbances associated with DM[23]. This increased NHHR in patients with DM can be attributed to several factors, including insulin resistance, which is a feature of DM and contributes to dyslipidemia. Low HDL-C levels are one of the most prevalent lipid metabolism abnormalities in patients with DM[24]. Insulin resistance leads to increased hepatic production of triglyceride-rich lipoproteins, resulting in elevated levels of non-HDL-C while simultaneously inhibiting HDL-C metabolism, thereby increasing the NHHR[25]. Additionally, chronic inflammation, which further influences lipid metabolism, promotes atherogenic processes and increases the risk of cardiovascular mortality, is often observed in patients with DM[26]. Chronic inflammation is a driving factor in the development of atherosclerosis and DM, and a higher HDL-C levels leads to decreased cardiovascular risk[27]. Low HDL-C levels are also correlated with the enhancement of low-grade inflammation and the apoptosis of pancreatic beta cells[28-30]. With respect to the mechanism, studies have indicated that HDL-C levels are positively associated with paraoxonase 1 (PON1) levels, which are crucial in the protection of HDL-C from oxidative modification. PON1 has been found to have antioxidant, anti-inflammatory, and antiadhesion properties in patients with atherogenic dyslipidemia[31]. Moreover, PON1 facilitates cholesterol efflux from macrophages to HDL, enhancing the reverse cholesterol transport process and resulting in reduced cardiovascular risk[32]. Thus, with a decreased HDL-C level, an elevated NHHR further indicates chronic inflammation and atherosclerosis in patients with DM, which increases the risk of cardiovascular mortality, as we reported in this study.

Current guidelines emphasize the role of controlling LDL-C levels in the prevention of clinical events. Moreover, evidence on other lipid measures is also accumulating[33]. Our findings highlight the significance of HDL-C in lipid metabolism and the importance of HDL-C in predicting clinical outcomes. Cheng et al[34] reported that in patients with hypertension, increased non-HDL-C levels (over 190 mg/dL) are associated with increased risks of all-cause mortality and cardiovascular mortality. Another study revealed that the NHHR is an independent risk factor for acute ST-segment elevation myocardial infarction in both male and female patients[35]. Wang et al[36] analyzed the relationship between the NHHR and carotid plaque stability and reported that a higher NHHR was correlated with unstable carotid plaques, suggesting that the NHHR could be a useful indicator for identifying patients at high risk of carotid plaques. In addition, the NHHR is also a stronger index for observing lipid metabolism, as evidenced by its value in predicting new-onset nonalcoholic fatty liver disease[37]. Here, we also revealed that the NHHR is associated with increased risk in patients with DM, and the results remained unchanged when patients were stratified by sex, age, and BMI. Moreover, we first generated a predictive model and confirmed the high accuracy of the model, which provides a new index for the prediction of risk in in patients with DM in clinical settings.

Our study has several strengths. First, we utilized data from the NHANES, which has a large sample size, and focused on patients with DM, providing new robust and accurate evidence for patients with DM. Second, we selected important features from the data via two rigorous machine learning models, the Boruta algorithm and LASSO regression model, which further provided crucial features for the multivariable Cox regression model and predictive model. Third, we developed a simple-to-use nomogram and generated three ROC curves and calibration plots to determine and validate the accuracy of the predictive model. Moreover, we have also published a model for clinical application. However, some limitations should also be noted. First, mortality outcomes were established via linking with the National Death Index records, which could result in misclassification. Nevertheless, a previous validation study confirmed the accuracy of the matching method, thereby reducing the likelihood of such misclassification[38]. Second, we performed the study based on NHANES data, and we were not able to present external validation results of the use of the model in different populations. However, we randomly divided patients at a 7:3 ratio and utilized the validation cohort to validate our conclusions. The results from the validation cohort indicated that the AUC of our predictive model was approximately 0.8, which demonstrated the capability of the model. Third, we analyzed data from the NHANES database, which is a large representative database, but the patients were mainly from the United States, which may cause selection bias; therefore, studies with patients from more regions or countries are warranted in the future. In addition, studies with a broader age range and younger cohorts are also needed. Fourth, this was a cohort study with a large sample size, and the demographic characteristics, lifestyle habits, and medication use of patients with DM may differ, which may cause potential bias in the results. In addition, specific information on medications (including antihypertensive and cholesterol-lowering drugs) was not available in the NHANES database, and more detailed information is needed for further studies.

CONCLUSION

In conclusion, our study provides robust evidence that an elevated NHHR is associated with an increased risk of cardiovascular mortality in patients with DM. In addition, we generated a predictive model that can be used for risk stratification in clinical practice. Future research should focus on longitudinal studies to further elucidate the mechanisms by which the NHHR influences cardiovascular outcomes, and studies with patients from more countries and regions are warranted to validate our findings.

ACKNOWLEDGEMENTS

We would like to appreciate the support by participants involved in the NHANES study.