Assessing the association of healthy lifestyle indicators (life’s simple 7) with bacterial vaginosis in reproductive-age women: A cross-sectional analysis of a nationwide survey

Abstract

Bacterial vaginosis (BV) is a prevalent vaginal infection in reproductive-age women, associated with various behavioral and metabolic factors. Life’s simple 7 (LS7), a cardiovascular health metric from the American Heart Association, includes smoking, physical activity, diet, body mass index, blood pressure, blood lipids, and blood glucose. This study aimed to explore the link between LS7 scores and BV, offering a perspective on comprehensive risk assessment. This cross-sectional study used data from the National Health and Nutrition Examination Survey 2001–2004, including 1203 women aged 18 to 49 years (382 BV, 821 controls). Accounting for National Health and Nutrition Examination Survey’ complex design, we applied fasting weight (WTSAF2YR), strata (SDMVSTRA), and primary sampling unit (SDMVPSU). Weighted logistic regression models (Crude, Model 1, and Model 2) examined the LS7–BV association. Model 1 adjusted for age, race, poverty income ratio, marital status, and education; Model 2 further adjusted for white blood cell, lymphocyte, monocyte counts, and globulin. LS7 was analyzed categorically (<8.5 vs ≥ 8.5), with the cutoff determined by receiver operating characteristic curve. We also performed restricted cubic spline analysis, Extreme Gradient Boosting, and subgroup analyses. Weighted baseline analysis showed that the LS7 score in the BV group (8.31 ± 2.01) was significantly lower than that in the normal group (8.93 ± 2.02, P < .001). Weighted logistic regression revealed that each 1-unit increase in LS7 was associated with lower BV odds in the Crude Model (odds ratio [OR] = 0.86, 95% confidence interval [CI]: 0.80–0.92, P < .001) and remained significant in Model 2 (OR = 0.90, 95% CI: 0.82–0.98, P = .013). In categorical analysis, LS7 ≥ 8.5 was associated with lower BV odds than < 8.5 (Crude OR = 0.60, 95% CI: 0.43–0.85, P = .005). Restricted cubic spline analysis showed a significant inverse dose–response relationship (P for overall = .001). Subgroup and sensitivity analyses confirmed consistent inverse associations and robustness. Higher LS7 scores were associated with lower odds of BV. Due to the cross-sectional design, these findings are associational and cannot establish causality.

1. Introduction

Bacterial vaginosis (BV) is a prevalent vaginal infection among reproductive-age women globally, with an estimated prevalence of 26% (varying by region).^[1] It stems from vaginal microbiota imbalance caused by mixed infections involving Gardnerella vaginalis and anaerobes. Common symptoms include abnormal vaginal discharge, foul odor, and vulvar itching. Beyond physical discomfort, BV increases susceptibility to sexually transmitted infections (HIV, gonorrhea, chlamydia, and HSV-2),^[2] and contributes to pelvic inflammatory disease and adverse pregnancy outcomes.^[3,4] Diagnosis primarily relies on the Nugent score, and treatment involves antibiotics like metronidazole and clindamycin. However, managing BV is challenging due to a recurrence rate exceeding 50% and rising antibiotic resistance,^[5] underscoring the urgent need for effective preventive strategies.

Growing attention has focused on the role of lifestyle in BV development.^[6] The American Heart Association (AHA) introduced “Life’s Simple 7” (LS7), a comprehensive health assessment framework consisting of 7 modifiable metrics. This framework has been associated with chronic conditions, including cardiovascular disease, diabetes, and chronic obstructive pulmonary disease.^[7,8] Unlike traditional single-factor studies, LS7 offers a holistic approach to health evaluation. While some research has explored connections between specific LS7 components and BV,^[9,10] the overall association between LS7 and BV remains understudied.

This cross-sectional study examined the association between LS7 indicators and BV risk. The findings suggest a hypothesis that lifestyle interventions targeting LS7 components may influence BV risk; however, this hypothesis requires further validation in prospective or interventional cohort studies.

2. Materials and methods

2.1. Data source

The data for this study were derived from the National Health and Nutrition Examination Survey (NHANES). NHANES is a continuous, cross-sectional, nationally representative health survey conducted by the National Center for Health Statistics (NCHS) under the Centers for Disease Control and Prevention. This survey received approval from the NCHS Research Ethics Review Board, and informed consent was obtained from all participants. The survey employs a complex multi-stage probability sampling design, collecting data every 2 years through household interviews, physical examinations at the Mobile Examination Center, and laboratory tests. In this study, data from the 2001–2004 cycle of NHANES were utilized, focusing on women aged 18 to 49 years who participated in the fasting subgroup examination.

The complex sampling design of NHANES includes 3 essential components: the stratification variable (SDMVSTRA), which ensures the sample’s representativeness across various demographic subgroups; the clustering variable (SDMVPSU), which reflects the geographical clustering characteristics of primary sampling units; and the analysis weights, which correct for unequal probability sampling and nonresponse bias. For the data from 2 consecutive cycles (2001–2004) included in this study, we followed the recommendations of NCHS by dividing the fasting weight of each cycle by 2 (New_Weight = WTSAF2YR/2). This adjustment provides estimates that represent the average level of the noninstitutionalized civilian population across the United States during these 2 cycles over the 4-year period. In all analyses of this study, we fully considered the aforementioned sampling design characteristics. The survey package in R software constructed the survey design object to ensure the validity of statistical inferences.

2.2. Study population

We recruited 21,161 participants from the NHANES 2001–2004 cycle. Participants were sequentially excluded according to the criteria in Figure 1: 10,301 males; 7395 individuals not aged 18 to 49 years; 659 participants without BV test data; 1581 participants missing fasting weight (WTSAF2YR); and 22 participants missing LS7-related indicators. After this screening, 1203 female participants aged 18 to 49 years were included: 821 in the normal group (Nugent score: 0–3) and 382 in the BV group (Nugent score: 7–10).

Figure 1.

Flowchart of the sample selection from the NHANES 2001–2004. A total of 21,161 participants were initially screened. Following the sequential exclusion of males (n = 10,301), individuals aged < 18 or > 49 years (n = 7395), and participants with missing BV laboratory data (n = 659), 2806 eligible participants were identified. After further excluding those missing fasting weights (n = 1581) and missing LS7 data (n = 22), the final analytical sample comprised 1203 female participants (382 in the BV group and 821 in the normal control group). BV = bacterial vaginosis, LS7 = life’s simple 7, NHANES = National Health and Nutrition Examination Survey.

2.3. Diagnostic criteria for BV

BV was diagnosed using the Nugent scoring system (range 0–10), which quantifies 3 bacterial morphotypes on Gram‑stained vaginal smears: large Gram‑positive rods (Lactobacillus spp.), small Gram‑variable coccobacilli (Gardnerella spp. and Bacteroides spp.), and curved Gram‑variable rods (Mobiluncus spp.). With participants’ consent, vaginal swabs were self‑collected at the Mobile Examination Center. NHANES staff processed the samples for Gram staining and quantified the bacterial morphotypes under low‑power microscopy. Nugent scores were categorized as normal (0–3), intermediate (4–6), or BV‑positive (7–10).^[11]

2.4. Life’s simple 7

The AHA developed the LS7 index to improve cardiovascular health in the United States.^[12] LS7 evaluates 7 cardiovascular risk factors and behaviors: smoking, physical activity, diet, body mass index (BMI), blood pressure, total cholesterol, and fasting glucose. Each factor receives a rating of poor (0), intermediate (1), or ideal (2) according to AHA criteria. The total LS7 score ranges from 0 to 14, with higher scores reflecting better cardiovascular health. This study employed the LS7 total score as the primary variable, analyzed in 2 formats: continuous and categorical (<8.5 vs ≥8.5).

2.5. Covariates

Covariate selection followed a 2-stage process. First, candidate variables were evaluated using weighted univariate logistic regression and variance inflation factor (VIF) testing to remove redundancies and reduce multicollinearity. Next, the retained variables were reviewed for clinical plausibility and consistency with the literature. This refinement ensured that the final multivariable model rested on biological and physiological rationale rather than solely on data-driven choices. The final analysis used 2 models. Model 1 adjusted for demographic and socioeconomic covariates: age, race, poverty-income ratio, marital status, and educational level. Model 2 further adjusted for white blood cell count, lymphocyte count, monocyte count, and serum globulin to account for systemic inflammatory burden. From an epidemiological perspective, these inflammatory markers are not traditional confounders but potential mediators linking LS7-related lifestyle factors to BV status. We constructed Model 2 solely to determine whether the association persisted after controlling for systemic inflammation; we did not use it to claim a direct or independent association between LS7 and BV.

2.6. Data analysis and modeling strategies

2.6.1. Weighted logistic regression analysis

We used weighted logistic regression to assess the association between LS7 and BV. Because NHANES employed a complex sampling design, we combined the fasting weight (WTSAF2YR) with the stratification (SDMVSTRA) and clustering (SDMVPSU) variables to specify the survey design. A weighted logistic regression model estimated odds ratios (ORs) and 95% confidence intervals (95% CIs). LS7 entered the models both as a continuous variable and as a categorical variable (<8.5 vs ≥ 8.5). Results are reported for three models: the Crude Model (unadjusted), Model 1 (adjusted for age, race, poverty income ratio [PIR], marital status, and educational level), and Model 2 (Model 1 further adjusted for white blood cell count, lymphocyte count, monocyte count, and globulin).

2.6.2. Restricted cubic spline (RCS) analysis

RCS analysis investigates the dose–response relationship between LS7 metrics and the risk of BV. RCS serves as a nonparametric method for analyzing continuous variables within regression models. It employs piecewise cubic polynomials to smoothly capture complex nonlinear associations. In contrast to linear assumptions, RCS offers a more flexible fit to the data and mitigates overfitting by imposing linear restrictions at the spline endpoints. Three models were employed in the RCS analysis: the Crude Model, Model 1 (adjusted for demographic variables), and Model 2 (further adjusted for blood cell counts and globulin). These models aimed to evaluate the LS7–BV relationship across various adjustment levels. Predicted risk curves were generated to assess nonlinear trends or cutoff effects of LS7 metrics on BV.

2.6.3. Extreme Gradient Boosting

This study employed the machine learning algorithm Extreme Gradient Boosting (XGBoost) to evaluate the relative importance of candidate indicators in predicting BV risk. As an ensemble learning method based on gradient boosting techniques, XGBoost captures nonlinear relationships among features by combining multiple base learners, demonstrating excellent predictive performance and robustness. During model construction, this study focused on including continuous variables and age to enhance the stability and interpretability of the model features. The dataset was randomly divided into a training set and a test set at an 8:2 ratio. The learning rate was set at 0.08, and the maximum tree depth was limited to 6 to ensure predictive efficacy while effectively preventing overfitting. To improve model interpretability, the Shapley Additive exPlanations (SHAP) framework was introduced. The mean |SHAP value| (feature importance score) quantified the relative contribution of each clinical indicator to the model output, and SHAP summary plots visualized the direction of association between feature values and BV risk.

2.6.4. Subgroup analysis

Subgroup analyses evaluate the consistency of the association between LS7 metrics and the risk of BV across various population characteristics. The optimal cutoff value is determined using receiver operating characteristic analysis to classify the exposure factor into high and low groups. Logistic regression models calculate ORs and their 95% CIs within each subgroup. To quantify the association of different stratification variables, including age, race, PIR, smoking, marital status, education level, and hypertension, interaction tests (likelihood ratio tests) are conducted, and the interaction P-value (P for interaction) is calculated. All results are visualized using forest plots to directly compare the magnitude of effect estimates across different subgroups, thereby validating the robustness of LS7 metrics in predicting BV risk.

2.6.5. Sensitivity analysis

To ensure the robustness of the study findings, researchers conducted sensitivity analyses. Specifically, they repeated the comparisons of baseline characteristics and logistic regression analyses using unweighted data, comparing the consistency between weighted and unweighted results. Additionally, they addressed missing data using the Classification and Regression Trees algorithm within the multivariate imputation by chained equations (MICE) framework. Five iterations of imputation were performed using the MICE package in R software. The analysis preserved the distributional characteristics of the original data, as verified by the Kolmogorov–Smirnov test (P > .05). Missing values in variables with a missing rate of ≤ 40% were imputed using the observed values at the terminal nodes of the decision trees.^[13] This algorithm effectively imputes both categorical and continuous variables, ensuring the stability of the imputed results. Researchers handled missing values through multiple imputation and compared the consistency of data distributions before and after imputation (see Table S1, Supplemental Digital Content 1).

To further account for behavioral and reproductive confounders known to be relevant to BV, we constructed 2 additional weighted logistic regression models (Models 3 and 4). Model 3 added sexual activity status, number of sexual partners, unprotected intercourse, history of vaginal douching, systemic antibiotic use, pregnancy history, and hormonal contraceptive use to the covariates already included in Model 2. Model 4 included only these reproductive and behavioral risk factors, without adjustment for demographic or inflammatory variables. The full regression results for Models 3 and 4 are presented in Table S6, Supplemental Digital Content 2.

2.7. Statistical analysis

All data processing and statistical analyses utilized R software (version 4.5.0). To address NHANES’ complex sampling design, survey design objects were created with the survey package. Weighted means and 95% CIs described continuous variables, with between-group comparisons conducted using the weighted Wilcoxon rank-sum test. Categorical variables were depicted by weighted frequencies and percentages, and between-group comparisons were made using the Rao-Scott χ² test. The optimal cutoff value (8.5 points) for the LS7 categorical variable was established through receiver operating characteristic curve analysis, selecting the LS7 score corresponding to the maximum Youden index as the optimal cutoff point (refer to Fig. S1, Supplemental Digital Content 3). Prior to multivariate analysis, VIF testing was executed to eliminate multicollinearity.

To ensure data robustness, missing values were handled using MICE within the Classification and Regression Trees algorithm framework. Five iterations of imputation were performed using the mice package in R software. The imputation process preserved the distributional characteristics of the original data, as verified by the Kolmogorov–Smirnov test (P > .05). For variables with a missing rate ≤ 40%, missing values were imputed using observed values at the terminal nodes of the decision trees.^[13] The RCS analysis was performed using the rms package, while the xgboost package was employed to construct the XGBoost model. A P-value < .05 was deemed statistically significant.

2.8. Ethics statement

This study utilized de-identified data from the publicly available NHANES. The survey underwent review and approval by the Research Ethics Review Committee of the NCHS, and all participants provided informed consent.^[14] Ethical approval was granted an exemption with opinion number 2026009101 from the Ethics Committee of The Affiliated Yongchuan Hospital of Chongqing Medical University. This exemption aligns closely with Item 1 of Article 32 of the Measures for Ethical Review of Life Science and Medical Research Involving Human Subjects (2023), issued by the National Health Commission of the People’s Republic of China on February 18, 2023. This regulation permits exemptions for research employing legally obtained public data or data derived from nonintrusive observation of public behaviors.

3. Results

3.1. Weighted baseline characteristics of the study subjects

This study included 1203 female participants aged 18 to 49 years from NHANES 2001–2004, with 821 cases (68%) in the normal group and 382 cases (32%) in the BV group. Table 1 presents the weighted baseline characteristics. The distributions of each variable before and after multiple imputation remained consistent, validating the imputation’s reliability (refer to Table S1, Supplemental Digital Content 1). Significant disparities between the BV and normal groups were evident in various aspects: race (P < .001), showing a notably higher percentage of non-Hispanic Black individuals in the BV group (24.48%) compared to the normal group (9.50%); PIR (P = .007), indicating a greater prevalence of low-income individuals in the BV group (34.00%) than in the normal group (26.65%); smoking status (P = .020), revealing a higher percentage of current smokers in the BV group (33.54%) than in the normal group (24.90%); and marital status (P = .006), with a higher proportion of unmarried individuals in the BV group (51.65%) than in the normal group (38.86%).

Table 1.

Baseline characteristics of BV weighted.

Characteristic	Overall (Unweighted N = 1203)	Normal (Unweighted N = 821)	BV (Unweighted N = 382)	P-value
Age	33.83 (9.35)	33.57 (9.32)	34.45 (9.41)	.2
Race				<.001
Mexican American	296.00 (8.99%)	205.00 (9.04%)	91.00 (8.88%)
Non-Hispanic Black	277.00 (13.93%)	137.00 (9.50%)	140.00 (24.48%)
Non-Hispanic White	536.00 (67.13%)	412.00 (72.17%)	124.00 (55.12%)
Other Hispanic	44.00 (4.73%)	32.00 (4.64%)	12.00 (4.95%)
Other Race	50.00 (5.22%)	35.00 (4.65%)	15.00 (6.57%)
PIR				.007
≤ 1.3	430.00 (28.82%)	274.00 (26.65%)	156.00 (34.00%)
1.3–3.5	433.00 (35.32%)	286.00 (34.10%)	147.00 (38.22%)
> 3.5	340.00 (35.86%)	261.00 (39.25%)	79.00 (27.78%)
Smoke				.020
Current	294.00 (27.45%)	177.00 (24.90%)	117.00 (33.54%)
Former	181.00 (14.88%)	127.00 (14.45%)	54.00 (15.89%)
Never	728.00 (57.67%)	517.00 (60.65%)	211.00 (50.57%)
Marital status				.006
Unmarried	523.00 (42.64%)	315.00 (38.86%)	208.00 (51.65%)
Married	680.00 (57.36%)	506.00 (61.14%)	174.00 (48.35%)
Education				.084
Non-higher education	577.00 (39.86%)	374.00 (37.85%)	203.00 (44.66%)
Higher education	626.00 (60.14%)	447.00 (62.15%)	179.00 (55.34%)
Height	163.55 (6.68)	163.43 (6.68)	163.84 (6.67)	.4
Weight	74.67 (20.24)	73.65 (19.89)	77.12 (20.87)	.002
BMI	27.88 (7.24)	27.55 (7.15)	28.68 (7.39)	.009
Waist circumstance	91.48 (16.28)	90.68 (16.11)	93.41 (16.54)	.014
SBP	112.18 (14.11)	111.78 (13.84)	113.14 (14.68)	.2
DBP	69.36 (10.42)	69.03 (10.48)	70.15 (10.27)	.2
PP	42.80 (11.23)	42.72 (10.94)	43.00 (11.91)	.9
Alanine aminotransferase (ALT)	19.86 (11.29)	19.69 (10.44)	20.27 (13.11)	>.9
Albumin	41.32 (3.24)	41.34 (3.21)	41.27 (3.29)	.5
Albumin urine	14.66 (53.20)	12.24 (24.29)	20.44 (90.23)	.011
Aspartate aminotransferase (AST)	21.06 (10.09)	20.77 (8.10)	21.74 (13.71)	.13
Blood urea nitrogen	10.37 (3.52)	10.46 (3.52)	10.15 (3.52)	.2
CRP	0.45 (0.63)	0.44 (0.65)	0.46 (0.60)	.061
Chloride	104.09 (2.37)	104.06 (2.45)	104.14 (2.19)	.9
Fasting glucose	92.71 (16.76)	92.18 (15.02)	93.98 (20.27)	.13
Globulin	30.63 (3.96)	30.40 (3.73)	31.18 (4.42)	.017
Glycated hemoglobin	5.25 (0.54)	5.22 (0.47)	5.31 (0.68)	.007
Hematocrit	39.90 (3.38)	39.92 (3.25)	39.86 (3.66)	.9
Hemoglobin	13.56 (1.25)	13.59 (1.19)	13.48 (1.39)	.4
Insulin	10.44 (8.24)	10.03 (7.99)	11.43 (8.75)	.048
LDL C	111.31 (33.35)	111.53 (33.52)	110.80 (32.99)	.7
lymphocyte count	1.98 (0.58)	1.94 (0.56)	2.09 (0.62)	.004
Mean cell hemoglobin	30.22 (2.38)	30.33 (2.20)	29.96 (2.75)	.4
Mean Cell Hgb Conc	33.94 (0.84)	34.01 (0.81)	33.78 (0.90)	.004
Mean platelet volume	8.23 (0.88)	8.21 (0.87)	8.28 (0.92)	.5
Mononuclear cell count	0.49 (0.16)	0.48 (0.15)	0.52 (0.16)	.018
Neutrophil count	4.20 (1.66)	4.15 (1.69)	4.32 (1.58)	.053
Percentage of neutrophils	59.68 (8.66)	59.80 (8.54)	59.40 (8.93)	.8
Platelet count	285.94 (70.10)	285.53 (69.83)	286.92 (70.81)	.6
Potassium	4.01 (0.30)	4.00 (0.31)	4.02 (0.29)	.3
Red blood cell count	4.49 (0.37)	4.49 (0.36)	4.51 (0.39)	.2
Sodium	138.43 (1.96)	138.44 (1.98)	138.40 (1.91)	.8
Total calcium	9.34 (0.33)	9.33 (0.32)	9.36 (0.36)	.2
Total cholesterol	190.21 (40.26)	190.35 (38.17)	189.86 (44.89)	.5
Total protein	7.19 (0.47)	7.17 (0.45)	7.24 (0.51)	.046
Triglycerides	106.34 (76.90)	107.47 (79.54)	103.63 (70.22)	>.9
Uric acid	4.49 (1.02)	4.47 (0.99)	4.54 (1.10)	.5
White blood cell count	6.90 (2.01)	6.80 (2.01)	7.14 (1.98)	.014
Hypertension				.039
No	1027.00 (84.45%)	709.00 (86.19%)	318.00 (80.30%)
Yes	176.00 (15.55%)	112.00 (13.81%)	64.00 (19.70%)
Diabetes				.029
No	1160.00 (96.67%)	795.00 (96.99%)	365.00 (95.92%)
Borderline	6.00 (0.37%)	1.00 (0.06%)	5.00 (1.11%)
Yes	37.00 (2.96%)	25.00 (2.96%)	12.00 (2.97%)
Antihypertensive medication				.2
No	1101.00 (90.07%)	759.00 (90.80%)	342.00 (88.34%)
Yes	102.00 (9.93%)	62.00 (9.20%)	40.00 (11.66%)
Lipid lowering medication				.2
No	1169.00 (96.42%)	803.00 (96.94%)	366.00 (95.18%)
Yes	34.00 (3.58%)	18.00 (3.06%)	16.00 (4.82%)
Glucose lowering medication				.6
No	1199.00 (99.65%)	818.00 (99.60%)	381.00 (99.78%)
Yes	4.00 (0.35%)	3.00 (0.40%)	1.00 (0.22%)
Sex intercourse				.8
No	312.00 (15.10%)	201.00 (15.35%)	111.00 (14.50%)
Yes	891.00 (84.90%)	620.00 (84.65%)	271.00 (85.50%)
Partners num				.7
No	377.00 (21.57%)	243.00 (21.09%)	134.00 (22.72%)
Yes	826.00 (78.43%)	578.00 (78.91%)	248.00 (77.28%)
Condom unprotected				.059
No	1118.00 (91.66%)	772.00 (93.07%)	346.00 (88.30%)
Yes	85.00 (8.34%)	49.00 (6.93%)	36.00 (11.70%)
Ever pregnant				.071
No	351.00 (27.80%)	252.00 (29.51%)	99.00 (23.72%)
Yes	852.00 (72.20%)	569.00 (70.49%)	283.00 (76.28%)
Hormone contraception				.3
No	1089.00 (90.04%)	740.00 (89.42%)	349.00 (91.50%)
Yes	114.00 (9.96%)	81.00 (10.58%)	33.00 (8.50%)
Douche history				<.001
No	917.00 (76.64%)	675.00 (81.40%)	242.00 (65.27%)
Yes	286.00 (23.36%)	146.00 (18.60%)	140.00 (34.73%)
Antibiotic use				>.9
No	1160.00 (96.09%)	793.00 (96.05%)	367.00 (96.19%)
Yes	43.00 (3.91%)	28.00 (3.95%)	15.00 (3.81%)
LS7	8.74 (2.04)	8.93 (2.02)	8.31 (2.01)	<.001

Mean (sd) or Frequency (%). Wilcoxon rank sum test; Pearson Chi-squared test. Continuous variables are presented as the mean and 95% confidence interval. Category variables are described as the percentage and 95% confidence interval.

BMI = body mass index, BV = bacterial vaginosis, CRP = C‑reactive protein, DBP = diastolic blood pressure, LDL C = low‑density lipoprotein cholesterol, LS7 = life’s simple 7, PIR = poverty income ratio, PP = pulse pressure, SBP = systolic blood pressure.

Significant differences between the BV group and the normal group were observed in various laboratory indicators: weight (P = .002), BMI (P = .009), waist circumference (P = .014), urine albumin (P = .011), globulin (P = .017), glycated hemoglobin (P = .007), lymphocyte count (P = .004), mean corpuscular hemoglobin concentration (P = .004), monocyte count (P = .018), white blood cell count (P = .014), total protein (P = .046), and insulin (P = .048).

The LS7 score was significantly lower in the BV group (8.31 ± 2.01) compared to the normal group (8.93 ± 2.02; P < .001). The weighted percentage distribution of specific LS7 scores within each group further highlighted that participants in the BV group were clustered at lower scoring levels compared to those in the normal group (refer to Fig. S2, Supplemental Digital Content 4). Additionally, the BV group exhibited a significantly higher prevalence of hypertension (P = .039) and diabetes (P = .029) than the normal group.

Based on these baseline differences, we next performed univariate and multivariate analyses to identify factors associated with BV.

3.2. Univariate and multivariate analyses

We conducted a weighted univariate logistic regression analysis to identify candidate variables linked to BV. Variables with a P-value of <.05 were then included in the multivariate logistic regression model. Detailed results from both the univariate and multivariate analyses are provided in Table S2, Supplemental Digital Content 5.

Univariate analysis identified several factors significantly associated with BV. Race showed a notable difference, with non-Hispanic Black individuals having higher odds compared to Mexican Americans (OR = 2.30, 95% CI: 1.64–3.24, P < .001). Marital status also played a role, as married individuals had lower odds (OR = 0.52, 95% CI: 0.41–0.67, P < .001). Higher education level was associated with lower odds (OR = 0.74, 95% CI: 0.58–0.94, P = .015). Additionally, increased globulin levels (OR = 1.06, 95% CI: 1.03–1.10, P < .001) and lymphocyte count (OR = 1.58, 95% CI: 1.28–1.94, P < .001) were linked to higher odds. Elevated AST levels (OR = 1.02, 95% CI: 1.00–1.03, P = .017) and a higher LS7 score (OR = 0.86, 95% CI: 0.81–0.91, P < .001) were also significant factors.

Multiple regression analysis identified multiple factors related to BV after adjustment. Non-Hispanic Black individuals had an increased OR (OR = 2.13, 95% CI: 1.43–3.17, P < .001), while non-Hispanic White individuals had a decreased OR (OR = 0.58, 95% CI: 0.38–0.90, P = .015). Being married was also associated with lower odds (OR = 0.67, 95% CI: 0.50–0.88, P = .005). Additionally, AST levels (OR = 1.02, 95% CI: 1.00–1.03, P = .037), lymphocyte count (OR = 1.31, 95% CI: 1.02–1.69, P = .036), and borderline diabetes (OR = 9.90, 95% CI: 1.02–96.37, P = .048) were significant factors.

We conducted VIF testing on the variables in the multivariate model. The results indicated that all VIF values were below 2 (see Table S3, Supplemental Digital Content 6), suggesting no significant multicollinearity.

Given that LS7 showed a significant inverse association in the multivariate analysis, we next performed a focused weighted logistic regression to quantify its association with BV.

3.3. Weighted logistic regression analysis of LS7 and BV

Weighted logistic regression models assessed the relationship between LS7 and BV, with findings detailed in Table 2. Treating LS7 as a continuous variable, each 1-unit increase in LS7 was associated with lower odds of BV in the Crude Model (OR = 0.86, 95% CI: 0.80–0.92, P < .001). This negative association persisted after adjusting for age, race, PIR, marital status, and education level in Model 1 (OR = 0.89, 95% CI: 0.82–0.97, P = .009). Model 2 further incorporated systemic inflammatory markers (white blood cell count, lymphocyte count, monocyte count, and globulin) to examine whether the correlation still existed after considering the systemic inflammatory burden. The results showed that there was still a significant negative correlation: each 1-unit increase in LS7 was associated with lower odds of BV (OR = 0.90,95% CI 0.82–0.98, P = .013).

Table 2.

Weighted multivariable logistic regression analysis of the association between LS7 and BV.

Characteristic	Case (%)	Crude Model			Model 1			Model 2
		OR	95% CI	P-value	OR	95% CI	P-value	OR	95% CI	P-value
LS7 continuous		0.86	0.80–0.92	<.001	0.89	0.82–0.97	.009	0.90	0.82–0.98	.013
LS7 categorical
< 8.5	210/561 (37.43%)	1 (Ref.)			1 (Ref.)			1 (Ref.)
≥ 8.5	172/642 (26.79%)	0.60	0.43–0.85	.005	0.71	0.52–0.97	.033	0.72	0.52–1.00	.048

The Crude Model was non-adjusted. The Model 1 was adjusted by age, race, PIR, marital status, and education. The Model 2 was adjusted by age, race, PIR, marital status, education, white blood cell count, lymphocyte count, mononuclear cell count, and globulin.

When LS7 was divided into 2 groups using a cutoff of 8.5 points, the proportion of BV was lower in the LS7 ≥ 8.5 group compared to the LS7 < 8.5 group (26.79% vs 37.43%). In the Crude Model, the LS7 ≥ 8.5 group had significantly lower odds of BV (OR = 0.60, 95% CI: 0.43–0.85, P = .005). This association remained statistically significant in Model 1 (OR = 0.71, 95% CI: 0.52–0.97, P = .033) and was of borderline statistical significance in Model 2 (OR = 0.72, 95% CI: 0.52–1.00, P = .048). These inflammatory markers should be considered intermediate mediators rather than traditional confounders. Therefore, the results from Model 2 cannot confirm a direct, independent association between LS7 and BV.

To further characterize the nature of this association, we then used RCS analysis to explore the dose–response relationship between LS7 and BV.

3.4. Dose–response relationship between LS7 and BV

We used RCS analysis to examine the dose–response relationship between LS7 and BV (Fig. 2). In the crude model, LS7 was significantly associated with BV (P for overall = .001), and the test for nonlinearity was not significant (P for nonlinear = .300). After adjustment in Model 1, the association remained significant (P for overall = .028; P for nonlinear = .249). Further adjustment in Model 2 produced consistent results (P for overall = .038; P for nonlinear = .220). All three models indicated a linear inverse trend in BV odds as LS7 increased, with the largest decrease in odds occurring at an LS7 score of about 8.97.

Figure 2.

Dose–response relationship between LS7 scores and the risk of BV. Three models were examined in this study: the Crude Model (unadjusted), Model 1 (adjusted for age, race, PIR, marital status, and education level), and Model 2 (further adjusted for white blood cell count, lymphocyte count, monocyte count, and globulin). The Crude Model showed a significant association between LS7 and BV odds (P for overall = .001), with no significant nonlinearity detected (P for nonlinear = .300). Upon adjustment in Model 1, the association remained significant (P for overall = .028; P for nonlinear = .249). Subsequent adjustment in Model 2 produced consistent results (P for overall = .038; P for nonlinear = .220). All three models demonstrated a linear inverse trend in BV odds as the LS7 score increased, with the most pronounced lower odds observed at an LS7 score of approximately 8.97. BV = bacterial vaginosis, CI = confidence interval, LS7 = life’s simple 7, PIR = poverty income ratio.

As a complementary analysis, we applied XGBoost with SHAP to rank the relative importance of all candidate predictors, including LS7, in a non‑parametric framework.

3.5. XGBoost-based analysis of BV-related factors

We used an XGBoost model to evaluate the relative contributions of multiple variables to BV (Fig. 3). SHAP analysis showed that globulin (mean SHAP = 0.078) was the most important predictor of BV, followed by lymphocyte count (0.076) and LS7 score (0.045). Age (0.033), white blood cell count (0.032), and monocyte count (0.007) also contributed to the model. The SHAP swarm plot indicated that higher globulin and lymphocyte counts were associated with increased BV risk (positive SHAP values), whereas a higher LS7 score was associated with lower BV odds (negative SHAP values), consistent with the weighted logistic regression results.

Figure 3.

Machine learning analysis and feature interpretation in the NHANES cohort. (A) The SHAP summary plot depicts the mean absolute SHAP values for the primary predictors of BV. Globulin emerges as the most critical predictor, with a mean SHAP value of 0.078, followed by lymphocyte count (0.076) and LS7 score (0.045). Age (0.033), white blood cell count (0.032), and mononuclear cell count (0.007) also contributed to the model predictions. (B) The SHAP swarm plot (or dependence plot) illustrates the association direction: higher globulin and lymphocyte counts correlate with positive SHAP values (indicating increased BV odds), while a higher LS7 score correlates with negative SHAP values (indicating decreased BV odds). BV = bacterial vaginosis, LS7 = life’s simple 7, NHANES = National Health and Nutrition Examination Survey, SHAP = Shapley additive explanations.

To assess whether this inverse association was consistent across different population strata, we next conducted subgroup analyses.

3.6. Subgroup analysis

We conducted stratified analyses across clinical subgroups to further explore the association between LS7 and BV (Fig. 4). The inverse association between LS7 and BV was observed in most subgroups. In the age subgroup, the association was more pronounced among women aged ≥ 30 years (OR = 0.55, 95% CI: 0.38–0.79, P = .003) but did not reach statistical significance in women aged < 30 years (OR = 0.75, 95% CI: 0.46–1.22, P = .254). In the marital-status subgroup, the association reached significance in married women (OR = 0.46, 95% CI: 0.28–0.77, P = .006) but not in unmarried women (OR = 0.81, 95% CI: 0.51–1.30, P = .392). A significant inverse association was also observed in participants with lower education level (OR = 0.56, 95% CI: 0.34–0.91, P = .027) and in those without hypertension (OR = 0.65, 95% CI: 0.45–0.95, P = .036). Interaction tests for all subgroups did not reach statistical significance (P for interaction > .05), suggesting that the inverse association of LS7 with BV was observed consistently across the examined population characteristics.

Figure 4.

Subgroup analysis of the association between LS7 scores and the risk of BV. Subgroups comprise age (<30 vs ≥30 years), race (Mexican American, non-Hispanic Black, non-Hispanic White, other Hispanic, and other race), smoking status (current, former, and never), marital status (married and unmarried), education level (non-higher vs higher education), hypertension (yes and no), antihypertensive medication (yes and no), and lipid-lowering medication use (yes and no). Each subgroup analysis was adjusted for age, race, PIR, marital status, education level, white blood cell count, lymphocyte count, monocyte count, and globulin, except for the stratification variable itself. Squares depict odds ratios (ORs), horizontal lines show 95% confidence intervals (CIs), and the dashed vertical line denotes the null (OR = 1). P-values for interaction (P for interaction) are displayed for each subgroup. A P for interaction > .05 indicates that the inverse association between LS7 and BV remained consistent across strata, with no significant modification by these factors. BV = bacterial vaginosis, CI = confidence interval, LS7 = life’s simple 7, OR = odds ratio, PIR = poverty income ratio.

Finally, we performed sensitivity analyses using unweighted data to assess whether the weighting strategy influenced our findings.

3.7. Sensitivity analysis

To assess whether the findings were influenced by the weighting strategy, we performed sensitivity analyses using unweighted data. The unweighted baseline characteristics (see Table S4, Supplemental Digital Content 7) showed that the LS7 score in the BV group (8.10 ± 2.00) remained significantly lower than in the normal group (8.71 ± 1.98, P < .001), and the direction of differences for other variables was consistent with the weighted analysis. Unweighted logistic regression analyses (see Table S5, Supplemental Digital Content 8) also demonstrated a significant negative association between LS7 and BV (Crude Model: OR = 0.86, 95% CI: 0.81–0.91, P < .001; Model 2: OR = 0.89, 95% CI: 0.83–0.96, P = .001). These results were consistent with those from the weighted analyses.

We then conducted a sensitivity analysis adjusting for all available behavioral and reproductive confounders related to BV; the full regression results are in Table S6, Supplemental Digital Content 2. Treating LS7 as a continuous variable, the inverse association – higher scores linked to lower BV risk – remained statistically significant in Model 3 (OR = 0.89, 95% CI: 0.82–0.97, P = .011) and Model 4 (OR = 0.89, 95% CI: 0.83–0.96, P = .005). By contrast, when LS7 was dichotomized at 8.5, the association attenuated after full adjustment and lost statistical significance (Model 4, OR = 0.70, 95% CI: 0.48–1.04, P = .076), suggesting that reproductive and behavioral confounders partly explained the effect seen in the unadjusted model.

4. Discussion

This study provides the first epidemiological evidence of a significant inverse correlation between the AHA’s LS7 metrics and the risk of BV. Analyzing data from the NHANES (2001–2004), we systematically evaluated the association between LS7 scores and BV risk using weighted logistic regression, RCS analysis, XGBoost machine learning, and sensitivity analyses. The results show a inverse association between higher LS7 scores and BV prevalence. When LS7 was treated as a continuous variable, a negative correlation remained after adjusting for demographic characteristics and systemic inflammatory markers in Model 2 (OR = 0.90, 95% CI: 0.82–0.98, P = .013). It is important to note that the inflammatory indicators included in Model 2 are likely intermediate mediators rather than conventional confounders. Therefore, this model demonstrates the association’s stability after controlling for systemic inflammation, but cannot confirm whether there is a direct independent relationship between LS7 and BV. Using a cutoff of 8.5 points, participants with LS7 ≥ 8.5 had lower odds of BV, although the association was of borderline statistical significance (OR = 0.72, 95% CI: 0.52–1.00, P = .048). The XGBoost model further confirmed LS7 as a significant predictor of BV, with a mean SHAP value of 0.045.

Our study advances beyond prior research focused on isolated behavioral factors – such as diet, BMI, smoking, and glucose levels – by integrating multidimensional indicators encompassing nutrition, metabolism, and behavior. This holistic approach underscores the inverse association of LS7. After adjusting for race, PIR, marital status, and education level, the inverse association remained consistent across demographic and socioeconomic subgroups in the main analyses. Subgroup analyses further revealed that the inverse association of LS7 was broadly applicable across most population strata, with no significant interactions observed (P for interaction > .05 for all subgroups), suggesting that cardiovascular health is associated with lower BV prevalence across different population characteristics. To better understand the biological basis, it is worth noting that BV is characterized by a decrease in Lactobacillus spp. and an overgrowth of anaerobic bacteria, including Gardnerella spp., Bacteroides spp., and Mobiluncus spp., in accordance with the Nugent scoring criteria utilized in this study. The composition of the vaginal microbiota is influenced by various host factors, such as lifestyle and metabolic status.^[15]

To contextualize our findings within existing literature, we first compared them with studies focusing on individual lifestyle factors. Previous studies on lifestyle factors and BV have primarily examined individual behaviors, such as diet and smoking. For example, a case-control study conducted in Iran reported significant associations between high dietary glycemic index/load and BV, with ORs of 2.99 and 4.01, respectively.^[9] Another study from Tehran indicated that a vegetarian diet was linked to a reduced risk of BV, while an unhealthy diet increased the risk.^[16] However, these isolated indicators may not fully capture the cumulative effect of various lifestyle dimensions. In contrast, the composite metric LS7 integrates 7 modifiable factors and may more accurately reflect the overall impact of lifestyle on BV. After adjusting for socioeconomic factors, our analysis demonstrated that higher LS7 scores were significantly associated with a lower risk of BV: the mean LS7 score in the BV group was 8.31 ± 2.01, compared to 8.93 ± 2.02 in the control group (P < .001). Furthermore, each 1-unit increase in LS7 score corresponded to a 10% lower odds of BV (OR = 0.90, 95% CI: 0.82–0.98), suggesting that LS7 may be an important correlate of BV.

Beyond dietary factors, previous studies have also linked individual anthropometric measures to BV. Previous studies indicate a link between higher BMI and increased BV risk, with 1 study showing an OR of 1.03 among 5428 participants, particularly in non-Hispanic white women (OR 1.0327).^[17] Another study with 5918 participants found overweight and class I obese women had higher Nugent scores, supporting the risk association.^[18] Conversely, a cohort study of 1946 sex workers in Kenya reported obesity might lower BV risk (adjusted risk ratio 0.82). Factors like ethnicity and socioeconomic status may influence these findings.^[19] In contrast, a high LS7 score, reflecting multiple health indicators, showed an inverse association with BV, with the LS7 score in the BV group being significantly lower than that in the normal group (8.31 vs 8.93, P < .001), suggesting LS7 may better represent lifestyle associations with BV.

Smoking represents a significant risk factor for BV. Research in Sweden revealed that 52% of BV patients were smokers, compared to only 32% in the control group.^[20] A Danish study found that smoking over 10 cigarettes daily at conception raised BV risk during pregnancy (OR = 1.59, 95% CI: 1.29–1.93).^[21] Our findings confirm that BV prevalence is higher in smokers (33.54% in the BV group vs 24.90% in the normal group, P = .020), further supporting the association between smoking and BV.

Beyond behavioral factors such as smoking, metabolic status is also associated with BV risk, as suggested by our subgroup and baseline comparisons. Considering metabolic status is crucial in assessing BV risk factors, yet studies linking metabolic issues like diabetes and hypertension to BV are limited. Research indicates women with diabetes, particularly those with poor blood sugar control, face higher risks of reproductive infections.^[22] Additionally, women with gestational diabetes show severe vaginal microbiota imbalances postpartum, linked to BV.^[23] Pregnant women with hypertension also exhibited greater vaginal microbiota diversity, with reduced Lactobacillus and increased BV-associated bacteria.^[24] These findings imply that metabolic changes are associated with disrupted vaginal microbiota and higher BV risk. In our study, the BV group exhibited a significantly higher prevalence of hypertension (P = .039) and diabetes (P = .029) than the normal group. Our research employed the LS7 score to assess health through various lifestyle factors affecting metabolic status. A higher LS7 score indicates better metabolic health, and this may be associated with lower BV odds, potentially through improved metabolic conditions and vaginal microbiota balance, thus serving as a predictive marker and a potential target for future research.

Although a direct link between total cholesterol and BV has not been established, our XGBoost analysis shows the LS7 score significantly predicts BV. The SHAP analysis ranked LS7 as the third most important predictor (mean SHAP = 0.045), following globulin (0.078) and lymphocyte count (0.076). The LS7 score reflects the dose–response relationship between health behaviors and BV risk, as demonstrated by our RCS analysis, which revealed a linear inverse trend in BV odds with increasing LS7 score (P for overall = .038 in Model 2; P for nonlinear = .220), with the most pronounced lower odds observed at an LS7 score of approximately 8.97. This underscores the potential value of composite scoring models in future studies exploring lifestyle factors in relation to BV risk.

The study identifies significant clinical implications for BV risk assessment through 7 modifiable LS7 indicators. The RCS analysis showed a notable BV risk trend with a P for overall < .05 across all three models, particularly with lower odds becoming more evident when LS7 scores reached approximately 8.97, suggesting an association between cardiovascular health and vaginal microbiota composition, and supporting integrated prevention strategies.

An inverse association between LS7 and BV was observed across most subgroups and remained evident in sensitivity analyses using unweighted data. However, because this study was cross-sectional, these findings should be interpreted as associative rather than causal. They do not support using LS7 as a clinical risk‑stratification tool in current practice. Instead, they provide a hypothesis-generating basis for future research – specifically, prospective cohort or interventional studies are needed to determine whether lifestyle modifications targeting LS7 components could ultimately inform BV prevention or management strategies.

Previous studies have shown that recurrence rates of BV treated with antibiotics alone are relatively high.^[25,26] However, because this study was cross-sectional, it could not be used to formulate clinical intervention plans or confirm that modifying LS7-related lifestyle factors reduces BV recurrence risk. The association between the LS7 threshold observed here and disease risk therefore serves only as a preliminary hypothesis. Prospective cohort studies and intervention trials are still needed to determine whether the LS7 metric can help assess risk of BV onset and recurrence.

This study presents several limitations. Firstly, the cross-sectional design limits causal inferences. Secondly, potential inaccuracies may arise from relying on self-reported LS7 components. Thirdly, although a significant association between LS7 scores and BV was identified, the direction of this relationship remains ambiguous, and separate clinical investigations were not conducted for validation. Nevertheless, the consistency of the inverse association across weighted and unweighted analyses, as well as across different model specifications, supports the robustness of our findings. Fourthly, this study could not fully eliminate residual confounding. Although we included all extractable NHANES risk factors related to sexual behavior, reproduction, and antibacterial use in the sensitivity model, the survey did not collect more detailed data on reproductive tract hygiene and sexual practices during the study period. These unmeasured confounders may bias the observed associations and likely explain why the binary LS7 lost statistical significance after adjustment. Future studies should include more comprehensive questionnaires on reproductive tract hygiene and sexual behavior to further validate our findings. Lastly, while the NHANES sample mirrors the U.S. population, further confirmation is needed to generalize our results to non-American populations. Future studies should incorporate prospective cohort designs and interventional approaches to investigate the role of LS7 in the prevention and management of BV. Additionally, researchers should explore the potential mediating role of inflammatory markers, such as globulin and lymphocyte count, in the LS7–BV relationship, as our SHAP analysis identified these factors as top predictors.

5. Conclusion

In this cross-sectional analysis of the NHANES cohort, LS7 scores were inversely associated with BV prevalence, and this inverse trend persisted across most subgroups. The study only established an association and did not confirm causality. Further prospective studies are needed to verify these findings.

Acknowledgments

The authors acknowledge the assistance provided by the staff of the Department of Gynecology at the Affiliated Yongchuan Hospital of Chongqing Medical University during the research. Additionally, the authors express gratitude to their families for their unwavering support and encouragement throughout this research journey.

Author contributions

Conceptualization: Yang Feng.

Data curation: Xiaoqian Zhang.

Formal analysis: Yang Feng.

Validation: Yang Feng.

Writing – original draft: Yang Feng.

Writing – review & editing: Xiaoqian Zhang.