NHANES is a series of cross-sectional, complex, multi-stage surveys conducted by the National Center for Health Statistics (NCHS) in the United States. It utilizes a complex, multi-stage probability sampling procedure to ensure the representation of participants from different geographical locations and sufficient racial/ethnic diversity. The purpose of NHANES is to assess the health and nutritional status of the civilian, non-institutionalized population in the United States (27). The survey collects data from approximately 5,000 individuals each year through interviews, physical examinations, and laboratory tests. The NHANES research protocol has been approved by the Ethics Review Board of the National Center for Health Statistics, and all procedures are conducted in accordance with relevant approved guidelines and regulations, with informed consent obtained from each participant. The specific analytical methods can be found in the NHANES Laboratory and Medical Technologists Procedures Manual. Our current study combined two survey cycles (2003–2006). A total of 20,470 study participants from the 2003–2006 NHANES cycles were included. These two cycles included participants with information on kidney function and the Food Frequency Questionnaire (FFQ_C).

In the NHANES surveys conducted in 2003–2004 and 2005–2006, 6,306 and 5,863 participants, respectively, responded to questions regarding the use of artificial sweeteners. After excluding 5,763 participants from 2003 to 2004 and 5,149 participants from 2005 to 2006 who lacked relevant measurement indicators, biochemical markers, and associated weights, the final analytical sample comprised 1,257 respondents from the National Health and Nutrition Examination Survey conducted from 2003 to 2006. The participants had complete anthropometric measurements and demographic characteristics data. The flowchart for this study is presented in Figure 1.

Dietary data in NHANES is collected through individual 24-h dietary recall interviews conducted by the Mobile Examination Center (MEC). The assessment of dietary intake is based on 24-h recalls obtained by trained interviewers during the participant’s visit to the MEC. The interviews are conducted using a computer-assisted dietary interview system with standardized probes, namely the Automated Multiple-Pass Method (AMPM) developed by the United States Department of Agriculture. The AMPM collects information on the types and quantities of all food and beverages consumed in the 24 h preceding the dietary interview. The AMPM is designed to ensure more comprehensive and accurate data collection while reducing the burden on participants. The data on the intake of artificial sweeteners in the diet is reported in the FFQ_C, which includes two 24-h dietary recall interviews and information on dietary supplement use, food security, and dietary behaviors obtained through the dietary questionnaire. The FFQ_C collects information on the frequency of food consumption over the past 12 months. All participants were asked whether artificial sweeteners were added to their diet. Our study includes complete FFQ_C and covariate data.

Detailed descriptions of biochemical analysis methods can be found in the NHANES Laboratory/Medical Technologists Procedures Manual. Renal function indicators are analyzed using the Beckman Synchron LX20, with the LX20 modular chemistry side employing the Jaffe rate method (kinetic alkaline picrate) to determine the creatinine (Cr) concentration in serum, plasma, or urine. The calibration of creatinine is traceable to the isotope dilution mass spectrometry reference method. Urinary creatinine and urinary albumin (evaluated using random urine samples with solid-phase fluorescence immunoassay) are used to calculate the urine albumin-to-creatinine ratio (UACR). Serum beta-2 microglobulin (β2M) and cystatin C (CYSC) are measured using immunoturbidimetric and immunonephelometric assays on the automated multichannel analyzer Siemens Dimension Vista 1,500.

The estimated glomerular filtration rate (eGFR) is the gold standard for assessing renal function. Traditionally, equations based on creatinine have been used to calculate the estimated glomerular filtration rate. However, the secretion of Cr in the renal tubules can systematically overestimate GFR and be influenced by factors other than renal function, such as low muscle mass, low activity level, vegetarian diet, frailty, lower limb amputation, end-stage heart failure or liver failure, and the use of drugs that inhibit proximal tubular secretion of Cr. In 2021, the American Kidney Society workgroup recommended using CYSC to assess GFR in adults with or at risk for CKD. Although CYSC is not influenced by muscle mass or diet, factors such as obesity, hypothyroidism, smoking, and systemic corticosteroid use were positively associated with CYSC values. Therefore, the equation for eGFRcr-cys based on Cr and CYSC can provide the most accurate eGFR values. The calculation equation is as follows: eGFR (mL/min/1.73 m²) = 135 × min (Bcr/κ, 1)^α × max (Bcr/κ,1)^−0.601 × min (Bcysc/0.8, 1)^−0.375 × max (Bcysc/0.8, 1)^−0.711 × 0.995^age[×0.969 if female] × 1[1.08 if black] where Bcr is blood Cr, Bcyst is blood CYSC, κ is 0.7 for females and 0.9 for males, α is −0.248 for females and − 0.207 for males, min indicates the minimum of Bcr/κ or 1, and max indicates the maximum of Bcr/κ or 1 (28).

In this study, participants were categorized into groups based on the frequency of artificial sweetener consumption reported in FFQ_C: never added (never), occasional (less than once a month, 1–3 times a month, once a week), regular (2–4 times a week, 5–6 times a week, once daily), and frequent (2–3 times daily, 4–5 times daily, 6 times or more daily). The renal function indicators included eGFR, UACR, CYSC, β2M, and serum creatinine (SCR). Blood pressure measurements were taken three times for all eligible individuals using a mercury sphygmomanometer at the MEC, including systolic blood pressure (SBP) and diastolic blood pressure (DBP). The average of the three blood pressure measurements, including SBP and DBP, was analyzed. Hypertension (HTN) was defined as a self-reported history of hypertension or SBP ≥140 mmHg or DBP ≥90 mmHg. CKD was defined as a self-reported history of CKD. Diabetes mellitus (DM) was a self-reported history of diabetes or fasting blood glucose ≥7.0 mmol/L. Gender was divided into male and female. Race/ethnicity was categorized as Mexican American, Other Hispanic, non-Hispanic white, non-Hispanic black, and other races (including multiple races). Education level was divided into less than 9th grade, 9th-11th grade (including 12th grade without a diploma), high school graduate, and college and above. The poverty income ratio (PIR) was expressed as the ratio of family income to the poverty line, with PIR values less than 1 indicating income below the poverty line and values greater than 1 indicating income above the poverty line. Self-reported smoking of less than 100 cigarettes was defined as a nonsmoker, while 100 or more cigarettes were defined as a smoker. Self-reported alcohol consumption of less than 12 drinks per year was defined as nondrinker. Body mass index (BMI, kg/m²) was calculated as weight (kg) divided by height squared (m²). All analysis models were adjusted for the following covariates: age, gender, race/ethnicity, BMI, education level, PIR, smoking status, alcohol consumption, fasting blood glucose, HTN, DM, CKD, triglycerides (TG), and high-density lipoprotein cholesterol (HDL).

In this study using NHANES data, mean and standard error of the mean (SEM) were used to compare groups for continuous variables using one-way analysis of variance (ANOVA). In contrast, percentages were used for comparing groups of categorical variables using chi-square tests. To assess the normality of the data, the Kolmogorov–Smirnov test, and variables deviating from a normal distribution were log-transformed. Consequently, all renal function markers (UACR, eGFR, CYSC, SCR, β2M) were log-transformed to achieve a normal distribution. Adjusted means for renal function markers (UACR, eGFR, SCR, CYSC, β2M) in artificial sweetener groups were evaluated using ANCOVA. These models were adjusted for age, gender, race, BMI, education level, PIR, smoking status, alcohol consumption, fasting blood glucose, HTN, DM, TG, and HDL. Logistic regression models were conducted with similar adjustment strategies, using three different levels of adjustment: model 1 (age, gender, race, BMI, education level, PIR), model 2 (age, gender, race, BMI, education level, PIR, smoking status, alcohol consumption, fasting blood glucose, HTN), and model 3 (age, gender, race, BMI, education level, PIR, smoking status, alcohol consumption, fasting blood glucose, HTN, DM, TG, HDL). ORs and 95% confidence intervals (CIs) for renal function markers associated with artificial sweetener groups were derived. The reference value was always the group that never added artificial sweeteners. Trend tests were conducted by treating artificial sweetener intake groups as categorical variables. Taking into account the masking variance, clustered design, and differential sampling probabilities, the sample weights provided by NHANES were applied in this observational study. Four-year dietary independent weights were used to assess all samples, and data analysis followed CDC guidelines for analyzing complex NHANES data. All statistical analyses were conducted using SPSS software version 26.0, and two-sided p-values less than 0.05 were considered statistically significant.

A Mendelian randomization analysis was employed as a causal inference approach to assess the association between artificial sweeteners and CKD as well as renal function. Single-nucleotide polymorphisms (SNPs) were used as instrumental variables, and summary statistics data from various genome-wide association studies (GWAS) were analyzed to evaluate the effect of exposure on the outcomes. Essentially, we analyzed summary statistics data for exposure (artificial sweeteners) and outcomes (renal function markers and CKD) from different GWAS to assess the impact of the former on the latter. Figure 2 provides an overview of the design and hypotheses of the MR study. This study was conducted following MR-STROBE guidelines to enhance the reporting of observational epidemiological studies.

The GWAS datasets related to artificial sweeteners were derived from the UK Biobank. Three European population GWAS datasets were compiled, including GWAS datasets for artificial sweetener intake in cereals, tea, and coffee. The combined datasets included 64,949 individuals and comprised 9,851,867 SNP loci. The genetic data mentioned can be downloaded from the MRC-IEU Open GWAS Project database, available at https://gwas.mrcieu.ac.uk/.

The GWAS dataset related to CKD was obtained from the FinnGen Biobank, which studied 500,000 individuals of Finnish descent and enriched the data on rare and late-onset diseases. The project included 1,932 clinical endpoints GWAS, identifying genome-wide significant associations at 2,491 loci, including 148 putative causal coding variants, with 62 of them being over twice enriched in the Finnish population and having a low minor allele frequency (<10%). This study retrieved summary statistics data for CKD in European populations, including 3,902 cases and 212,841 controls. The definition of CKD in this study strictly followed the GB61 classification in ICD11 (CKD Stage I: GB61.0; CKD Stage II: GB61.1; CKD Stage IIIa: GB61.2; CKD Stage IIIb: GB61.3; CKD Stage IV: GB61.4; CKD Stage V: GB61.5). Summary statistics data for eGFR were sourced from the GWAS summary statistics for eGFRcrea in the CKDGen Consortium (n = 765,348, predominantly of European ancestry), and the GWAS results for eGFRcrea generated in the UK Biobank (UKB) (Application Number 20272, n = 436,561, of European ancestry), with adjustments for age, sex, and other study-specific covariates. The summary statistics data for SCR and blood urea nitrogen (BUN) were obtained from a publicly available GWAS dataset, which involved a deep phenotyping GWAS (diseases, biomarkers, and drug use) in the Japan Biobank (n = 179,000) and a GWAS meta-analysis conducted in the UK Biobank and FinnGen (combined n = 628,000), with adjustments for age, age squared, sex, age × sex, age squared × sex, and the first 20 principal components as covariates. The summary statistics data for CYSC were collected and provided by the UKB, with the primary analysis restricted to individuals reporting European ancestry (459,327 participants, representing 94% of the cohort) and further limited to a genetically homogeneous and unrelated subset of 337,539 White British participants, excluding a small number of participants who withdrew from the UKB (149 individuals). The summary statistics data for UACR were derived from data on unrelated proteinuria patients of European ancestry (382,500 individuals) in the UKB, with exclusions based on gender mismatch, absence of reported relationships, gender chromosome aneuploidies, and excessive heterozygosity or missingness. The diagnosis of CKD was based on recognized histological and radiological criteria. Recommended diagnostic criteria include albuminuria [UACR ≥30 mg/g] and persistent reduction in kidney function [eGFR <60 mL·min⁻¹·1.73 m⁻²] for a duration of more than 3 months. The GWAS summary data used for renal function markers in this study were obtained from large-scale GWAS meta-analyses, including the DIAbetes Genetics Replication and Meta-analysis (DIAGRAM), UKB, BioBank Japan (BBJ), and FinnGen BioBank. The genetic information in the GWAS datasets pertaining to artificial sweeteners, CKD, and renal function markers (including UACR, eGFRcr, eGFRcysc, CYSC, SCR, and BUN) is summarized in Table 1.

The MR analysis relies on genetic instrumental variables that must satisfy three fundamental assumptions: (1) the assumption of association, indicating a close intertwining between the instrumental variables and the exposure factor; (2) the assumption of independence, stating that the instrumental variables are detached from any potential confounding factors that might impact the exposure-outcome relationship; and (3) the assumption of exclusivity, asserting that the instrumental variables influence the outcome solely through their effect on the exposure. Only when these three core assumptions are fulfilled can the MR study effectively counterbalance latent sources of confounding. Based on the initial assumption in MR analysis, this study proposes selecting instrumental variables as SNPs from the GWAS dataset of the exposure variable. These SNPs should display significant genomic associations (p < 5 × 10⁻⁸), exhibit no linkage disequilibrium (LD) (r² < 0.001 within a 1 Mb distance), and possess a minimum minor allele frequency (MAF) greater than 1%. Unfortunately, it is unfortunate that a limited number of SNPs significantly associated with artificial sweetener intake in relation to CKD and renal function markers meet the p < 5 × 10⁻⁸ criterion. To further delve into the connection between artificial sweetener intake and CKD, as well as renal function markers, and obtain more comprehensive findings, the genetic instrument association threshold is thus adjusted to p < 1 × 10⁻⁵ (29). After identifying each corresponding significant SNP, LD analysis is carried out using the European genotype data from the 1,000 Genomes Project as the reference template. Only the SNP with the lower p-value is retained from the SNPs displaying LD (R²) above the specified threshold (R² = 0.001). When extracting information on the instrumental variables from the outcome variable dataset, palindrome SNPs exhibiting inconsistent allele frequencies between the exposure and outcome datasets are removed. For missing SNPs in the exposure and outcome variable datasets, proxy SNPs with strong LD (R² > 0.8) are utilized as substitutes. The effects on the outcome, along with the corresponding allele values, are reported for the proxy SNP and the target SNP. The association assumption is evaluated using the F-statistic. Typically, an instrumental variable is considered strong when the F-statistic exceeds 10, calculated as F = (R² × (N – K − 1))/((1 − R²) × k), where N represents the sample size of the exposure database and K represents the number of SNPs, while R² denotes the proportion of variance explained by the SNPs in the exposure database. For fewer than 10 instrumental variables, the formula for calculating R² is R² = 2 × EAF × (1 − EAF) × β², where EAF signifies the effect allele frequency, and β is the allele effect. For 10 or more instrumental variables, the formula for calculating R² is R² = (2 × EAF × (1 − EAF) × β²)/[(2 × EAF × (1 − EAF) × β²) + (2 × EAF × (1 − EAF) × N × SE(β)²)] (30), where SE represents the standard error of the β estimate. Blindly excluding SNPs with F < 10 could introduce further biases. A search for these SNPs related to artificial sweetener intake in the PhenoScanner database is conducted, while excluding known SNPs associated with CKD risk factors (including BMI, diabetes, hypertension, heart failure, smoking, alcohol consumption, and lipid-related SNPs). The selected instrumental variables should include information regarding their reference alleles, effect alleles, effect allele frequencies, effect sizes (beta), standard errors (SE), and p-values from both the exposure and outcome dataset. To accommodate the second and third assumptions, we employed the MR-Egger method, with a horizontal pleiotropy test conducted based on the intercept and its corresponding p-value from MR-Egger regression.

Our study aimed to assess the genetic predictive impact of artificial sweetener intake on CKD and various renal function-related indicators (eGFRcr, eGFRcysc, BUN, UACR, CYSC, and SCR). CKD serves as the primary outcome, while the secondary outcomes encompass different renal function indices. The genetic data selected for analysis exclusively come from European populations to mitigate the biases resulting from population stratification. Four methods are employed in this study to evaluate the causal relationship between artificial sweetener intake and renal function markers, as well as the risk of CKD: inverse-variance weighted method (IVW), weighted median estimator (WME), MR-Egger regression, and mode-based estimation (simple and weighted modes). The IVW method assumes that all SNPs serve as valid genetic instruments, assuming no horizontal pleiotropy. The WME method allows for the presence of horizontal pleiotropy in the instrumental variables, but requires that over 50% of the SNPs function as valid instruments. The MR-Egger method permits causal inference even when all SNPs are considered invalid instruments. Given the consideration for potential weak instrumental variables in the analysis, we employed the recently proposed MR-RAPS method to enhance the reliability of our results. MR-RAPS enables robust adjustment for pleiotropy using a robust adjusted profile score and provides reliable inferences when conducting MR analysis with a large number of weak instrumental variables, particularly when both the exposure and outcome variables are complex (31).

To satisfy the second and third assumptions of MR analysis, tests for instrument heterogeneity and outliers are conducted using Cochran’s Q test (32) and the MR-PRESSO (33) (Mendelian Randomization Pleiotropy RESidual Sum and Outlier) global test. Cochran’s Q test is employed to assess heterogeneity in the IVW model, with a significance threshold of p < 0.05 indicating the presence of heterogeneity. A random-effects model is utilized for causal inference within the IVW framework in such cases. In the MR-PRESSO analysis, not only can the magnitude of instrument pleiotropy be estimated, but the effect size between the exposure and outcome variables can also be calculated after removing potential outliers, thereby allowing for an assessment of the results before and after correction. Sensitivity analysis is performed using a leave-one-out approach, where each instrumental variable SNP is excluded in turn, and the remaining SNPs are used to assess causal associations through the IVW method. The impact of each SNP on the MR analysis is determined by observing whether the effect estimate crosses the null value with a significance threshold of α = 0.05. Causal estimates are presented as ORs along with their corresponding 95% CIs. To enhance the visualization of IVW results, we utilized a radial plot and the RadialMR approach. RadialMR eliminates the need for genetic data re-encoding and allows for direct detection of outliers or potential pleiotropic SNPs, thereby providing more reliable causal estimates (34). Considering multiple testing, the two-sided p-values are adjusted using the Benjamini-Hochberg method to control the false discovery rate (FDR), a widely used approach for multiple hypothesis testing correction. The P.adjust function from the stats package (v4.2.3) is employed to control the FDR at a level of 5%. Adjusted p-values <0.05 are considered statistically significant. All statistical analyses are performed using R version 4.2.3 and the packages “MendelianRandomization,” “MR-PRESSO,” “TwoSampleMR,” “mr-raps,” and “RadialMR.”

The data utilized in this study were derived entirely from published research or publicly available GWAS summary data. As the study did not involve the collection of original data, ethical approval from an ethics committee was not required. Ethical approval for the publicly available summary statistics data was previously obtained and documented in the original GWAS publications. Each study included in our analysis received approval from its respective institutional ethics review board, and all participants provided written informed consent.

Table 6 presents the MR estimates for the causal relationship between adding artificial sweeteners in cereals, tea, and coffee and CKD. Using the IVW and MR-RAPS methods as primary analyses, no significant associations were found between the addition of artificial sweeteners in cereals, tea, and coffee and CKD in the European population. Specifically, no significant associations were observed for the addition of artificial sweeteners in cereals (IVW: OR = 1.37, β = 0.317, 95% CI = −0.571 to 1.204, p = 0.484; MR-RAPS: OR = 1.38, β = 0.325, 95% CI = −0.616 to 1.266, p = 0.499), tea (IVW: OR = 0.83, β = −0.185, 95% CI = −0.744 to 0.374, p = 0.517; MR-RAPS: OR = 0.83, β = −0.190, 95% CI = −0.779 to 0.398, p = 0.525), and coffee (IVW: OR = 1.17, β = 0.157, 95% CI = −0.390 to 0.703, p = 0.574; MR-RAPS: OR = 1.18, β = 0.163, 95% CI = −0.411 to 0.736, p = 0.579). The corrected p-values using the Benjamin-Hochberg method also did not indicate a causal relationship between adding artificial sweeteners in cereals, tea, and coffee and CKD (p > 0.05). The MR-Egger intercepts did not provide statistical evidence of directional pleiotropy (all p-values >0.05). The MR-PRESSO test did not identify potential pleiotropy among the included SNPs. However, the RadialMR plot suggested the presence of horizontal pleiotropy, indicating a potential influential SNP driving the causal relationship between exposure and outcome. Therefore, caution is warranted when removing all outliers since, even though classified as outliers, the SNPs should also be considered for their association with a separate phenotype representing a pleiotropic pathway of the outcome. The scatter plot demonstrates the estimated effects of SNPs on exposure (AS_cereal, AS_coffee, AS_tea) and CKD outcome (Supplementary Figure S4). The funnel plot indicates positions where directional pleiotropy exists for each outcome (Supplementary Figure S1). The leave-one-out analysis is depicted in Supplementary Figure S2, and the RadialMR plot is shown in Supplementary Figure S3.

Supplementary Tables S2–S7 present the results of the MR analysis methods for the relationship between the addition of artificial sweeteners in cereals, tea, and coffee and renal function indicators. Using the IVW and MR-RAPS methods as primary analyses, we observed a positive correlation between the addition of artificial sweeteners in cereals (IVW: OR = 1.02, β = 0.024, 95% CI = 0.004 to 0.044, p = 0.017; MR-RAPS: OR = 1.03, β = 0.025, 95% CI = 0.004 to 0.046, p = 0.019) and eGFRcysc, as well as a negative correlation between the addition of artificial sweeteners in cereals (IVW: OR = 0.91, β = −0.091, 95% CI = −0.165 to −0.017, p = 0.017; MR-RAPS: OR = 0.91, β = −0.094, 95% CI = −0.172 to −0.016, p = 0.019) and SCR. However, after correction using the Benjamin-Hochberg method, the adjusted p-values were greater than 0.05 (adjusted p-value: 0.057 and 0.056), indicating suggestive evidence of potential correlations. Moreover, we found a negative correlation between the addition of artificial sweeteners in coffee (IVW: OR = 0.94, β = −0.063, 95% CI = −0.104 to −0.023, p = 0.002; MR-RAPS: OR = 0.94, β = −0.065, 95% CI = −0.108 to −0.022, p = 0.003) and UACR, and the adjusted p-value after Benjamin-Hochberg correction remained below 0.05 (adjusted p-value: 0.009). However, there was no causal relationship between adding artificial sweeteners in cereals, tea, and coffee and other renal function indicators (p > 0.05). Scatter plots (Supplementary Figures S4, S8, S12) were used to visualize the effect sizes of each MR analysis. MR-Egger, weighted median, MR-RAPS, and IVW estimates showed consistent results. Furthermore, sensitivity analyses using the MR-PRESSO test and RadialMR plot to remove potential pleiotropic SNPs indicated no evidence of pleiotropy or heterogeneity (Supplementary Tables S2–S7 and Supplementary Figures S3, S7, S11). Funnel plots depicted the balanced distribution of each SNP (Supplementary Figures S1, S5, S9). The leave-one-out analysis results demonstrated that excluding any single SNP in the genetic variation hardly biased the results (Supplementary Figures S2, S6, S10; see also Figures 3–5).