Lymphocytes and inflammation markers linked to metabolic syndrome risk in shift workers using machine learning on large population data

The study population was derived from the NHANES, a comprehensive survey that collects nationally representative health and nutritional data from participants in the United States. Extensive household interviews were conducted to obtain demographic and health history information. Physical examinations and blood sample collections were performed in a Mobile Examination Center (MEC). Serum samples were analyzed by the Laboratory Sciences Division of the National Center for Environmental Health at the Centers for Disease Control and Prevention.

This cross-sectional study used publicly available data from NHANES for the years 2005–2010 and 2017–2018 and was approved by the National Center for Health Statistics Research Ethics Review Board. Further details are available on the official NHANES website (https://www.cdc.gov/nchs/nhanes/about_nhanes.htm↗). The original study protocol and formal approval from the Research Ethics Review Board (protocol #2005−06; #2011−17) can be accessed on the NHANES Research Ethics Review Board website (https://www.cdc.gov/nchs/nhanes/irba98.htm↗). All participants provided written informed consent at the time of registration. Our analysis included 3,079 employed individuals aged 20 years or older who completed both interviews and MEC assessments. Participants were required to have a shift work schedule to be included in the study. We excluded pregnant women and multiracial individuals to minimize potential bias. Pregnancy induces significant physiological changes that can affect MetS assessment. Similarly, participants from different racial backgrounds were excluded because MetS risk factors can vary by ethnicity. As the MetS definition used in this study is primarily suited to European and American populations, we focused on individuals from these groups to maintain consistency.

Daytime work was defined as standard weekday work between 9:00 a.m. and 5:00 p.m. Shift work included any schedule that differed from these regular daytime hours [22]. Shift work status was determined using questionnaire items such as: “Which best describes your hours worked?” and “What was your overall work schedule for the past 3 months?” The response options included “evening or nights, “early mornings,” or “variable schedule (early mornings, days, and nights).” For analysis, participants who reported standard daytime work were classified as the “Regular” group, and those who reported any nonstandard schedule were classified as “Irregular.”

MetS was defined according to the harmonized criteria proposed in the joint interim statement of the International Diabetes Federation Task Force on Epidemiology and Prevention, National Heart, Lung, and Blood Institute, American Heart Association, World Heart Federation, International Atherosclerosis Society, and the International Association for the Study of Obesity [23]. Participants were diagnosed with MetS if they met three or more of the following criteria: (1) waist circumference: ≥ 102 cm in men or ≥ 88 cm in women; (2) triglycerides: ≥ 150 mg/dL or current use of lipid-lowering medication; (3) High-density lipoprotein cholesterol (HDL-C): HDL-C < 40 mg/dL for men or < 50 mg/dL for women; (4) Blood pressure: systolic ≥ 130 mmHg and/or diastolic ≥ 85 mmHg, or current use of antihypertensive medication; and (5) Fasting glucose: elevated levels or current use of antihyperglycemic agents.

Fasting venous blood samples were collected from all participants to measure leukocyte, neutrophil, lymphocyte, monocyte, and platelet counts, expressed as thousands of cells per microliter (1000 cells/μL). The following complete blood count (CBC)-derived inflammatory indicators were calculated: Systemic Immune-Inflammation Index (SII), Systemic Inflammatory Response Index (SIRI), Systemic Inflammatory Composite Index (AISI), Monocyte-to-Lymphocyte Ratio (MLR), Neutrophil-to-Monocyte Ratio (NMLR), Platelet-to-Lymphocyte ratio (PLR), and Neutrophil-to-Lymphocyte Ratio (NLR) [24]. The formulas were defined as follows:

In this study, we examined several sociodemographic factors, including age, sex, and race (non-Hispanic Black, non-Hispanic white, Mexican American and other Hispanic). Education level was characterized as less than high school, high school diploma/GED, or college and above. The poverty-income ratio (PIR) and alcohol consumption (non-drinker, 1–5 drinks/month, 5–10 drinks/month, or ≥10 drinks/month) were also assessed. Smoking status was classified as never smoked, former smoker, or current smoker. Body mass index (BMI) and self-reported diet quality (excellent, very good, good, fair, or poor) were included.

Dietary quality was further evaluated using the Healthy Eating Index-2020 (HEI-2020), which measures adherence to the Dietary Guidelines for Americans based on 13 food and nutrient components. Furthermore, we considered health complications (including weak/failing kidneys, stroke, asthma, anemia, arthritis, congestive heart failure, coronary heart disease, chronic bronchitis, cancer, or malignancy) and Patient Health Questionnaire (PHQ) scores. These covariates were selected based on previous research [25,26].

Four standard ML models were developed using the shift worker dataset, which served as the model development cohort. A total of 32 variables collected in this study were included in the modeling process. The models applied in this analysis were random forest (RF), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and linear regression (LR).

The dataset was randomly divided into training and test sets following the widely accepted Pareto principle, with a 70:30 split. Specifically, 70% of the data was used for model selection and tuning, while the remaining 30% was reserved for evaluating performance. During the training, 10-fold cross-validation was employed to ensure robust model validation. In this procedure, the training set was divided into 10 subsets: nine were used to train the model, and one served as the validation set. This process was repeated 10 times so that each subset was used once for validation. The algorithm’s final accuracy was then calculated as the average across the 10 iterations. Following this validation process, the test set (30% of the data) was used to evaluate how well the model generalized to unseen data.

The analytic sample was randomly split into a training dataset accounting for 70% and an independent test dataset accounting for 30%. The training dataset was used for model development and internal validation, and the test dataset was reserved solely for evaluating generalization to unseen data. For all ML algorithms, performance within the training dataset was assessed using cross-validated out-of-fold predictions, or out-of-bag (OOB) predictions for RF. The primary performance metric was AUC.

For LR, XGBoost, and LightGBM, we implemented 10-fold cross-validation using the training dataset. In each iteration, nine folds were used to fit the model and the remaining fold served as the validation fold, so that every observation was predicted only from models that had not been trained on it. Hyperparameters for the XGBoost and LightGBM, such as learning rate, maximum depth, and the number of boosting iterations, were tuned to maximize cross-validated AUC [27]. The LR model was fitted without penalization as a baseline parametric classifier and did not require extensive hyperparameter tuning. After cross-validation, each algorithm was refitted on the full training dataset using the optimal hyperparameters, and its generalization performance was evaluated on the independent 30% test dataset.

The RF model was implemented using bootstrap aggregation with OOB estimation. Each tree was grown on a bootstrap sample of approximately 80% of the training observations, with the remaining observations serving as OOB samples for that tree. For each participant, the OOB prediction was obtained by aggregating votes from trees that did not include that participant in their bootstrap sample, providing an internal cross-validation. To reduce overfitting, we restricted tree complexity by using 1,000 trees, sampling 80% of observations per tree, setting the number of variables considered at each split to the square root of the total number of predictors, and enforcing a minimum terminal node size of 5 observations [28,29].

All models were implemented using the “RandomForest,” “lightgbm,” “treeshap,” “xgboost,” and “fastshap” packages.

To thoroughly evaluate model performance, we utilized standard binary classification metrics, including area under the curve (AUC), accuracy (ACC), sensitivity (recall), and specificity. Furthermore, we analyzed the ROC curve to gain further insight into model behavior. The bootstrap method was used to refine estimates of model performance.

A major limitation of complex ML models is their inherent lack of interpretability, often referred to as the “black box” nature [30,31]. To enhance interpretability, we used SHapley Additive ExPlanations (SHAP) values from the ‘shap’ package, which are based on game theory, to visualize feature importance. This approach enhances model transparency and interpretability, thereby improving the reliability of its predictions [32].

In our study, we utilized the permutation importance method to identify the most significant features for predicting MetS in shift workers. This method evaluates feature importance by measuring how randomly shuffling a feature’s values affects the model’s predictive performance. To ensure reliability and minimize errors, we conducted 1,000 permutations for each feature across all models, yielding importance values that were averaged and ranked to determine their relative contributions. This methodology enables researchers and practitioners to better understand the factors influencing model outcomes by analyzing feature importance, recognizing interdependencies, and considering key elements for informed decision-making [30].

Continuous variables were presented as mean, median, minimum, and maximum values, whereas categorical variables were expressed as frequencies and percentages. Participants were divided into MetS and non-MetS groups according to diagnostic criteria. Differences between the two groups were analyzed using the Wilcoxon rank-sum or Kruskal–Wallis test for continuous variables and the Fisher exact or chi-squared test for categorical variables.

Missing data were addressed using multiple imputation, a flexible, simulation-based statistical method commonly applied to handle missing values or nonresponse in the NHANES dataset [33]. To ensure nationally representative estimates, analyses were weighted per NHANES guidelines.

Logistic regression analysis was used to examine the correlation between inflammatory markers and MetS, with results expressed as odds ratios (OR) and 95% confidence intervals (95% CIs). Given that the relationships between inflammatory indicators and MetS may vary among participants due to multiple influencing factors, we conducted stratified analyses to assess these differences in participants with MetS. We also applied a generalized additive model (GAM) to explore potential non-linear relationships. When a non-linear association was detected, a two-piecewise linear regression model was constructed to calculate the threshold effect of inflammation levels on MetS using the resulting smoothing plots. When the smoothing curve revealed a relationship between MetS status and inflammation levels, a recursive method was used to automatically identify the inflection point for subsequent analyses.

All statistical analyses were performed using the R software (http://www.r-project.org↗) and EmpowerStats (http://www.empowerstats.com↗), with statistical significance defined as p < 0.05.

We screened the shift worker population and included 3,079 individuals in the study cohort. Participants were divided into two groups: MetS (unweighted n = 529, weighted n = 9,291,315) and non-MetS (unweighted n = 2,250, weighted n = 26,216,750). A detailed flowchart of the selection process is presented in Fig 1. Table 1 shows the baseline characteristics of the two groups. Participants with MetS were generally older (median age: 45 vs. 38, p < 0.001), more likely to be overweight (34.2% vs. 30.5%) or obese (61.6% vs. 32.2%), and exhibited higher levels of inflammatory markers, including neutrophils (median: 4.00 vs. 3.60, p < 0.001), lymphocytes (median: 2.10 vs. 2.00, p = 0.024), monocytes (median: 0.60 vs. 0.50, p = 0.009), SIRI (median: 0.91 vs. 1.02, p = 0.048) and AISI (median: 223 vs. 246, p = 0.027). No significant differences were observed in shift work patterns or other inflammatory indicators between the groups.

Of all participants, 2,166 were randomly assigned to the training set and 913 to the test set. All variables were included in training the four ML models, and the AUC was used to evaluate their discriminative performance.

In the LightGBM model (Fig 2A, red line), the AUC for the training dataset was 0.944, which was considerably higher than that for the test dataset (Fig 2B, red line; 0.722). In comparison, the RF model demonstrated an AUC of 0.741 in the training dataset (Fig 2A, purple line) and 0.729 in the test dataset (Fig 2B, purple line). Similarly, the XGBoost model achieved an AUC of 0.818 for the training dataset (Fig 2A, blue line) and 0.747 for the test dataset (Fig 2B, blue line). LR yielded AUC values, 0.763 for the training dataset (Fig 2A, green line) and 0.699 for the test dataset (Fig 2B, green line). All models showed some degree of performance decline when applied to the test dataset; however, the decrease was less significant in RF and XGBoost than in LightGBM. Evaluation of calibration (Fig 2C) and DCA (Fig 2E) for all four models on the training set confirmed that LR (Fig 2C, blue line; Fig 2E, blue-green line) performed acceptable. followed closely by XGBoost (Figs 2C and 2E, purple line), LightGBM (Figs 2C and 2E, brown and green lines), and RF (Fig 2C and 2E, orange and blue line). When evaluated on the test dataset (Figs 2D and 2F), XGBoost (Figs 2D and 2F, purple line) was the best-performing algorithms, followed closely by LR (Figs 2D and 2F, blue and blue-green line), LightGBM (Figs 2D and 2F, brown and green lines), and RF (Figs 2D and 2F, orange and blue line). Per-model performance metrics, including AUC values with 95% CIs estimated via bootstrap resampling, as well as sensitivity, specificity, and calibration data at clinically relevant thresholds, are provided in the S1 and S2 Table.

The bar plot highlights the importance of various features across four models: LightGBM, LR, XGBoost, and RF. In the LightGBM model, the top 15 features identified include BMI, age, sex, and several inflammatory ratios such as MLR, PLR, AISI, SII, SIRI, NMLR and NLR, among others (Fig 3A). Fig 3B compares individual incidence risk (f(x) = −3,21) with the overall average risk (E[f(x)] = −2.64), showing that the individual has a lower risk than the population mean. Factors such as PIR, sex, and AISI contribute to this elevated risk.

The beeswarm plot presents the overall importance of features in the LightGBM, LR, XGBoost, and RF models. A color scale on the right indicates the relative value of each feature, where yellow dots represent high values and purple dots indicate low values. The violin plot, aligned at the midline, aggregates the dots representing each case in the internal validation set; the vertical spread reflects the number of cases sharing identical SHAP values. Fig 3C displays the SHAP values of the 15 features and their association with MetS risk among shift workers in the LightGBM model. Higher values of BMI, age, MLR, and platelet counts were associated with increased risk, while higher values of PLR, PIR, SII, SIRI, and other inflammatory markers were associated with decreased risk.

Figs 3D and 3F present the top 15 features contributing to predictions in the LR model, Figs 3G and 3I show the top 15 predictive features in the XGBoost model, and Figs 3J and 3L depict the top 15 influential features identified by the RF model. The findings indicate that, in addition to established risk factors for MetS such as BMI and age, inflammatory predictors, including lymphocytes, neutrophils, monocytes, and platelets, are also significant [34,35]. Moreover, previous studies have reported associations between these inflammatory factors and both shift work and MetS [15,36]. The risk contribution of each feature closely mirrors that observed in the LightGBM model. When comparing individual and overall incidence risks, the LR model (Fig 3E, f(x) = −589415, E[f(x)] = 0), the XGBoost model (Fig 3H, f(x) = 0.133, E[f(x)] = 0.172), and the RF model (Fig 3K, f(x) = −552, E[f(x)] = 0) all demonstrated lower individual risks than the overall average risk. Across all models, the individual risk consistently falls below the overall risk.

The relationship between hub feature levels and SHAP values (Fig 4.) showed that the varying trends of different inflammatory factors influenced the model’s predictive performance. The SHAP value of neutrophils increased sharply with neutrophil count (Fig 4A), peaking within the range of 3.5–3.7, and then exhibited a U-shaped trend, first decreasing and then increasing, with the lowest value observed within the range of 10–12. The SHAP value of lymphocytes (Fig 4B) showed an overall stable upward trend, with a slight U-shaped pattern within the range of 1.6–2.5. In contrast, the SHAP value of monocytes (Fig 4C) showed an opposite trend to that of neutrophils, decreasing sharply to its lowest point (0.6–0.7), followed by an inverted U-shaped trend. The SHAP value of platelets (Fig 4D) exhibited two inverted U-shaped patterns: the first occurred within the range of 100–250, reaching a trough around 250–260, followed by the second inverted U-shaped trend. Platelet counts within the range of approximately 220–300 and > 500 indicated a negative contribution to the model predictions.

To facilitate interpretation of feature effects and potential interactions, SHAP dependence plots were generated exclusively for the LR model. Unlike tree-based models such as LightGBM, RF, and XGBoost, which can capture complex non-linear relationships [29,31], LR offers a more transparent and interpretable framework for visualizing the relationship between individual predictors and model outputs. The inherently linear structure of LR enables clearer identification of both the direction and magnitude of feature contributions, as well as more interpretable interactions via color-coded modifiers in SHAP plots. In the LR model, age and lymphocyte count exhibited a positive and approximately linear association with SHAP values, while sex emerged as a prominent interaction factor. In contrast to non-linear models that often produce unstable SHAP dependence plots, the LR model provides more stable and interpretable visualizations. This clarity makes LR particularly valuable for examining feature relationships in analyses based on SHAP values [28].

The SHAP dependence plot displays the distribution of output values for a single feature and uses color coding to introduce a second feature, enhancing the analysis of potential feature interactions. In the LR model (Fig 5.), age (Fig 5A), and lymphocyte count (Fig 5B) were positively and linearly correlated with SHAP values. Chronic bronchitis and shift schedule were used for color classification; however, their effects on the SHAP values of age and lymphocytes were relatively low. Neutrophils (Fig 5C) also demonstrated a positive linear correlation with SHAP values, with alcohol consumption as an interaction factor significantly influencing the SHAP values. Conversely, platelet count (Fig 5D) exhibited a negative linear correlation with SHAP values, and PIR, as an interaction factor, significantly influenced SHAP values. Monocyte counts (Fig 5E) were negatively correlated with the model predictions. Among the categorical variables, participants who were overweight or obese demonstrated significantly higher SHAP values than those in other groups (Fig 5F), suggesting that the interaction between BMI positively influenced model predictions.

Different models were used to adjust for covariates in the multivariate logistic regression analyses examining the association between inflammatory factors and MetS risk (Table 2). For neutrophils, Models I (unadjusted) and II (adjusted for age, sex, and race) showed a significant positive association with MetS risk; however, this association was no longer significant in Model III (further adjusted for shift schedule, PIR, alcohol consumption, smoking, BMI, self-reported diet quality, HEI2020, weak/failing kidneys, stroke, asthma, anemia, arthritis, coronary heart disease, chronic bronchitis, cancer or malignancy, and PHQ score). Similarly, lymphocytes remained consistently associated with elevated MetS risk across all models, with the strongest associations observed in Models II and III, suggesting their potential as reliable predictors. In Model I, monocytes were significantly associated with increased risk. However, the association became non-significant in Model III, indicating that additional covariates may mediate this relationship. When analyzed categorically, participants in the third quartile (Q3: 0.563–0.6) had significantly lower odds of MetS across all models (Model I: OR: 0.02; 95% CI: 0.00–0.08, p < 0.001; Model II: OR: 0.02; 95% CI: 0.00–0.09, p < 0.001; Model III: OR: 0.01, 95% CI: 0.00–0.06, p < 0.001) compared with the reference group. For platelet count, no significant associations were observed in any model (all ORs = 1.00, all p > 0.05). However, in the categorical analysis using quartiles, significant associations were observed for Q3 [255–287) across all models (Model I: OR: 0.53; 95% CI: 0.40–0.70, p < 0.001; Model II: OR: 0.65; 95% CI: 0.49–0.87, p = 0.004; Model III: OR: 0.56; 95% CI: 0.41–0.75, p < 0.001). Additional analyses of other inflammatory markers are presented in S3 Table.

These findings suggest that inflammatory factors are associated with prevalent MetS among shift workers and may be informative when assessing MetS-related risk profiles. However, these effects were attenuated after comprehensive covariate adjustments, warranting further investigation.

Fig 6 illustrates that neutrophils (Fig 6A), lymphocytes (Fig 6B), and monocytes (Fig 6C) were significantly associated with an increased risk of MetS (neutrophils: OR: 1.09; 95% CI: 1.03–1.16, p = 0.002, lymphocytes: OR: 1.21; 95% CI: 1.06–1.38, p = 0.006, monocytes: OR: 1.71; 95% CI: 1.02–2.88, p = 0.044). Among non-Hispanic White participants, both neutrophils (OR: 1.13; 95% CI: 1.03–1.23, p = 0.009) and monocytes (OR: 3.87; 95% CI: 1.72–-8.73, p = 0.001) demonstrated strong associations with MetS risk. Moreover, a negative correlation was observed between alcohol consumption and inflammatory factors.

To better understand how variations in inflammatory factor counts influence MetS risk, we conducted a GAM analysis to identify non-linear relationships and potential threshold effects (Fig 7 and Table 3). The results indicated that the risk of MetS among shift workers increased steadily with higher neutrophil counts (Fig 7A). For lymphocytes (Fig 7B), a non-linear relationship was observed, with an inflection point at 2.2. Below this threshold, the association was inverse but not significant (OR: 0.8; 95% CI: 0.6–1.0, p = 0.083), whereas above the threshold, higher counts were significantly associated with an increased risk of MetS (OR: 1.5; 95% CI: 1.2–1.8, p < 0.001; likelihood ratio test, p = 0.002). Monocytes (Fig 7C) demonstrated a linear association with MetS risk, showing no evidence of a threshold effect. Platelet counts (Fig 7D) exhibited a more complex, S-shaped pattern, rising at low levels, plateauing between 160 and 350, and then increasing again after minor fluctuations. Two-piecewise regression identified an inflection point at 257. The effect estimates were 1.0 (95% CI: 1.0–1.0, p < 0.001) on the left and 1.0 (95% CI: 1.0–1.0, p = 0.001) on the right. Although statistically significant, the OR of 1.0 indicated no meaningful association between platelet counts and MetS. Additional analyses of other imflammatory markers are presented in S4 Table.

All relevant data are within the manuscript and itsfiles. Supporting Information