The China Health and Nutrition Survey (CHNS), established in 1989, is an ongoing prospective cohort study. A previous study has published a detailed study design (15). In brief, the multi-stage random cluster sampling method was used to select 4,400 households and 19,000 participants, which covers nine provinces (Guizhou, Guangxi, Heilongjiang, Henan, Hubei, Hunan, Liaoning, Jiangsu, and Shandong). So far, the CHNS has completed data collection from 11 waves (1989, 1991, 1993, 1997, 2000, 2004, 2006, 2009, 2011, 2015, and 2018). The CHNS has obtained clinical, dietary, anthropometric, and all other individual data from each household member. Because the collection and analysis of blood samples were only conducted and released in 2009, we therefore used data from the 2009 wave of the CHNS. A detailed study design has been described in a previous study (16).

We analyzed 12,178 individuals who agreed to participate and gave informed consent for analysis of their data. This study included participants aged ≥18 years who participated in the 2009 survey wave. Participants were excluded if they met any of the following criteria: (1) a history of diabetes (n = 131), hypertension (n = 1094), apoplexy (n = 38), or myocardial infarction (n = 28); and (2) incomplete data (n = 1963). After excluding these incomplete data, this study finally included 7,014 individuals (3,253 men and 3,761 women). In order to show that the sample size of this study has sufficient statistical power, sample size was calculated based on a previous study that determined the prevalence rate of MS [control group = 0.14 (514/3,556); experimental group = 0.39 (194/495)] in a Chinese population-based study (7). Statistical power for a two-sample proportion chi-squared test indicated that a total of 156 participants were included in this study, which could achieve 95% statistical power. Ethics approval was obtained from the Institutional Review Board of the University of North Carolina at Chapel Hill and the National Institute for Nutrition and Health, Chinese Center for Disease Control and Prevention.

Fasting blood samples were collected in blood collection tubes and tested in a national central lab in Beijing. Fasting blood glucose (FBG) and lipids [triglyceride (TG) and high-density lipoprotein cholesterol (HDL-C)] were analyzed using an automatic biochemical analyzer (Hitachi 7600 automated analyzer, Tokyo, Japan). Insulin was measured by an ELISA Kit (Millipore Corporation, Billerica, MA, USA). The homeostasis model assessment for insulin resistance (HOMA-IR) score was calculated using the following formula: fasting insulin (mU/L) * fasting glucose (mmol/L)/22.5 (17).

In order to determine all the above biochemical markers, FBG was measured by the glucose oxidase method. TG and HDL-C were measured by the enzymatic method.

Fasting blood samples were drawn from the cubital vein for routine blood detection. WBC count was analyzed using an automatic hematology analyzer (Beckman Coulter LH751, Beckman Coulter, USA). Generally, an individual who maintains a very low leukocyte count level is healthier. We grouped the subjects into three categories—low level (subnormal level: ≦3.9), middle level (normal level: 3.91–9.94), and high level (above normal level: ≧10.0). We aimed to examine the association between a subnormal WBC count and risk of MS.

Anthropometric measurements (height, body weight, and waist circumference) were made using standard protocols. Body mass index (BMI) was calculated as the ratio of weight (kg) and height (m) squared. BP was measured three times after participants quietly sat for at least 5 min, and the maximum values of the first two of three readings were considered as the final BP values. All participants were also asked to fill in a standard questionnaire, which included demographic variables [sex, age, residence (city, suburban, town or county, capital city, and rural village)] and lifestyle variables [smoking status (smoker or not) and drinking status (drinker or not)].

According to the criteria of the American Heart Association scientific statements of 2009 (1), if a participant suffered from three or more of the following syndromes, he/she will be defined as having MS:

All continuous variables were expressed as median [interquartile range (IQR)] because of non-normal distribution, and categorical variables were expressed as percentage. To examine differences in participants’ characteristics between WBC count category, Kruskal–Wallis test and chi-square test were applied for continuous variables and categorical variables, respectively. We examine the hypothesis of logistic regression analysis prior to its use. Firstly, to examine the independence of observation variables, interactions were tested by adding the respective multiplicative terms in the models simultaneously. There were significant interactions of WBC with age, residence, apoplexy, and insulin resistance on MS (p for interaction: <0.001, 0.038, <0.001, and 0.015, respectively) ( Appendix Table 1 ). Based on the above findings, we also carried out subgroup analysis based on sex, age (age <65 years, age ≥65 years), residence (city, suburb, town, and village), apoplexy (yes or no), and insulin resistance (yes or no). Secondly, variance inflation factors (VIFs) were used to examine multicollinearity among explanatory variables, and the result shows that the VIFs of explanatory variables range from 1.02 to 1.91, which means that there is no multicollinearity between explanatory variables ( Appendix Table 2 ). Thirdly, Cook’s distance was used to identify extreme outliers, and the results showed that Cook’s distance of each variable is less than or equal to 0.06, showing that no extreme outliers were found. Fourth, to initially assess whether there are linear relationships between continuous independent variables and dependent variable logit conversion values, the Box–Tidwell method was used (transformation of the independent variables). The results of the line test found nonlinear relationships between all continuous independent variables (age, BMI, and insulin resistance) and the dependent variable logit conversion values. To linearize relationships, all continuous variables were converted into ordinal categorical variables. To examine the association between WBC count and MS, multivariate logistic regression analyses were applied to calculate risk of MS in different WBC count categories when several potential confounding factors were adjusted. We adjusted for sex, age (continuous variable), BMI (continuous variable), residence (categorical variable), smoking and drinking status (categorical variable), and HOMA-IR.

To examine the robustness of the above-mentioned results, a sensitivity analyses was further conducted as follows: we excluded participants who was excessively thin and obese [BMI (<18 or >40 kg/m²)]. We also used receiver operating characteristic (ROC) curve analysis, precision, specificity, and sensitivity through ML including multilayer perceptron (MLP) and logistic regression (LR) to compare the predictive performance of the MS risk model. In all two-sided tests, p-values of <0.05 were defined as statistically significant. Stata 16.0 software (Stata Corp LP, College Station, TX, USA) was used for all tests.

Participants’ characteristics according to WBC category are shown in Table 1 ; 3.4% of participants have a higher WBC level. Participants with a higher WBC level were more likely to be male, to be younger, and to live in a rural village. In addition, BMI and HOMA-IR increased statistically across different WBC categories. The proportion of participants with above-normal WBC count, who were cigarette smokers and alcohol drinkers, was statistically higher. However, there was no significant difference for other participants’ characteristics between WBC categories.

In all 7,014 participants, 1,479 were diagnosed with MS. The prevalence rates of MS in the low level, middle level, and high level were 12.1% (48 of 397), 21.4% (1,368 of 6,380), and 26.6% (63 of 237), respectively. Figure 1 shows a positive association between WBC count and risk of MS and its components. After adjustment for a potential confounder, the adjusted ORs for MS across WBC category were 1.00 (reference) for low level, 1.65 (95%CI: 1.18, 2.31) for middle level, and 2.18 (95%CI: 1.36, 3.50) for high level (p for trend, 0.001). Similar significant relationships were also found when components of MS were analyzed respectively ( Figure 1 ). In addition, a higher WBC level was also associated with increased risk of MS in both male and female ( Figure 2 ).

Another purpose of this study was to examine the associations of WBC category and MS by the following subgroups: age (<65 or ≧65 years), residence (city, suburban, town, or rural village), smoking status (smoker or non-smoker), drinking status (drinker or non-drinker), BMI (<25 kg/m² or ≧25 kg/m²), and insulin resistance (yes or no). The strength of the association between WBC count and MS increased with age (OR, 1.42 [95%CI, 1.10-1.83] for adults aged <65 years; 1.85 [95%CI, 1.06, 3.21] for those aged ≧65 years). The strength of the association between WBC count and MS was weaker in participants who resided in a rural village (OR, 1.22 [95% CI, 0.90–1.66]) compared with other residences. The association between WBC count and MS was stronger in non-smokers (OR, 1.55 [95% CI, 1.19–2.03]) and non-drinkers (OR, 1.56 [95% CI, 1.19–2.04]) compared with smokers (OR, 1.34 [95%CI, 0.83-2.14))/drinkers (OR, 1.35 [95% CI, 0.86–2.12]). The association between WBC count and MS was stronger in normal/underweight participants compared with overweight/obese participants (OR, 1.71 [95% CI, 1.27–2.30] vs. 1.20 [95% CI, 0.82–1.74]) and was equivalent whether there was insulin resistance or not (OR, 1.54 [95% CI, 1.12–2.10] vs. 1.42 [95% CI, 1.01–1.99]). The detailed results has been showed in Figures 3–5. Distribution of components of MS according to WBC category is also shown in violin plots ( Figure S1 ).

The aforementioned findings remained stable in a sensitivity analysis ( Table S1 ). The results remained significant with the exclusion of excessively thin and obese participants [BMI (<18 or >40 kg/m²)].

To assess the predictive performance of predictive models, we selected MLP and logistic regression to compare their accuracy. The AUROC was calculated both in the 70% of training set and in the 30% of test set. It is generally accepted that an AUROC curve greater than 0.5 indicated having an adequate predictive performance. The results showed that the MPL algorithm represented a more prominent predictive performance (AUROC in the training set: 0.814; AUROC in the testing set: 0.842) compared to other algorithms ( Table S2 ).