Metabolic risk stratification of night shift workers in a large retail workplace through clustering and SHAP interpretation

The study was conducted based on the health examination index of night workers at a large distribution site located in Cheonan, South Chungcheong Province, and 320 employees worked at this site in 2024. Among them, 28 employees handling other harmful substances under the Korea Occupational Safety and Health Act were excluded from the study, and 73 people already under management due to hypertension, diabetes, and dyslipidemia were also excluded, so the final 219 people were selected as subjects for integrated management. According to the current Republic of Korea Occupational Safety and Health Act, night workers are defined as people (24) who work more than four times a month for 8 h for more than 6 months from 12 a.m. to 5 p.m., or work an average of 60 h a month from 10 p.m. to 6 p.m. Due to the nature of the workplace, the night shift schedule was regular rotational, the employees did not work in isolation, and the night shift was in compliance with the break time prescribed by law in Korea.

This study was conducted on large warehouse-type distribution businesses, and the facilities consist of various departments such as food courts, fresh food (wholesale, bakery, deli, seafood), product display and logistics, check-out and customer service, parking and cart management, and marketing. Looking at the characteristics of each department, first of all, employees of food-related departments generally work in a refrigerated or frozen environment and are responsible for shelf-life management and waste disposal. In addition, the actual working hours are often extended as the furloughed workers often have to perform inventory collection, processing, and cleaning tasks. The product display and logistics departments work mainly at dawn and have a high-intensity manual labor to repeatedly transport and organize heavy goods. The check-out and customer service departments have a high proportion of emotional labor as they handle direct cash transactions and customer complaints. In addition, the parking and cart collection departments may increase outdoor work or increase labor intensity depending on the season. The workplace complies with the Korean Labor Standards Act of 52 h a week, and the operating hours are from 10 a.m. to 10 p.m. Logistics, cooking, and cleaning tasks performed before and after business hours are the main causes of night and long working hours.

For the health screening data analysis, records were used that included demographic information (age), night shift work period (shift_years), anthropometric measurements (BMI), blood pressure (systolic [SBP] and diastolic [DBP]), fasting blood tests (glucose), lipid panels (total cholesterol, triglycerides, high-density lipoprotein [HDL] cholesterol, and low-density lipoprotein [LDL] cholesterol), and insomnia scores.

Height and body weight were measured using an automated stadiometer (InBody 520; Biospace Inc., Seoul, Korea) with participants wearing light clothing and no shoes. BMI was calculated as weight (kg) divided by height squared (m²) (25). Systolic and diastolic blood pressures were measured twice in a seated position after 5 min of rest using an automated sphygmomanometer (Omron HEM-7080IT, Kyoto, Japan); the mean of the two readings was recorded. Blood samples were collected after an overnight fast of at least 8 h and analyzed in a certified clinical laboratory following standard enzymatic procedures. Insomnia symptoms were assessed using the Insomnia Severity Index (ISI), a 7-item self-report questionnaire rated on a 5-point Likert scale (0–4). The total score ranges from 0 to 28, with higher scores indicating more severe insomnia. The ISI has demonstrated excellent internal consistency (Cronbach's α = 0.90) and strong convergent validity with objective sleep parameters (26).

The number of years at night work was also considered a key factor in this study. All variables were scaled using standardization (z-scores).

For data analysis in this study, Python 3.13 was used. For unsupervised learning, scikit-learn's K-means was applied as the primary algorithm. Uniform Manifold Approximation and Projection (UMAP) from umap-learn was utilized for understanding the structure and visualizing high-dimensional variables. A predictive model was built using RandomForestClassifier, and variable contributions were evaluated using a combination of Gini-based internal importance and permutation importance. To enhance interpretability, global and local Shapley Additive Explanations (SHAP) values were calculated using the shap package.

To explore health risk profiles, we used a two-step procedure. First, UMAP was applied to standardized data (Z-score transformation) to reduce dimensions while preserving local structures. Second, we performed K-means clustering on 2D UMAP embeddings and applied the commonly used settings n_neighbors = 15, min_dist = 0.1, and random_state = 42 (27 –29). This combination of parameters is known to be effective in reliably preserving inter-cluster structures and in deriving reproducible results, which have been conventionally adopted in several prior studies.

The same parameter settings were maintained in all analyses to ensure consistency and reproducibility of the analyses. Whether the cluster-specific analysis reflected the differentiation of the actual clinically significant health risk groups (e.g., blood pressure, BMI, triglycerides, etc.) and evaluated mathematically good analysis.

To investigate the variables that strongly distinguished the clusters, a Random Forest classifier was trained to predict cluster membership using the original standardized variables as input. In this study, considering that the total number of samples is limited to 200, model performance and feature-importance stability were evaluated using a repeated 5-fold cross-validation scheme (10 repetitions). A Random Forest classifier (n_estimators = 500, class_weight = "balanced") was employed. Cross-validated accuracies, precisions, recalls, and F1-scores were computed for each repetition and summarised in terms of sample mean, standard deviation and range (30, 31).

To interpret the relative importance of the variables that contributed to the prediction, the SHAP value was calculated, and a SHAP summary plot was created for each cluster to visualize the contribution distribution of the main variables. This allowed us to identify the key variables that contributed to distinguishing each cluster.

A total of 219 participants were included in the final analysis. The mean age was 38.7 years, and 44% were female. Average BMI was 24.1, and the mean shift work tenure was 8.3 years. Full descriptive statistics are presented in Table 1.

Using the number of clusters (k) from 2 to 11 silhouette scores as an evaluation indicator by applying the K-means algorithm (Figure 1), the silhouette coefficient at k = 2 was the highest at 0.5614, but we looked at the case of k = 4 with the next highest silhouette coefficient of 0.4974. For K = 4, it was the best fit for the validity of the cluster structure as it was close to 0.5 with k = 2. This study ultimately set both k = 2 and k = 4 clusters as analysis units, and subsequent cluster characteristic analysis and predictive modeling were conducted.

Figure 2 shows the clustering results for k = 2 and k = 4. For k = 2, the cluster was visually well separated with minimal overlap along both embedding axes (Figure 2A). For k = 4, we classified the entire dataset into four distinct subgroups (Figure 2B). At k = 4, the cluster showed distinguishable patterns in key features such as shift year, age, and metabolic profile (e.g., blood pressure, cholesterol), demonstrating the interpretability of subgroups by more clinical factors. These results suggest that the multi-group classification can more accurately reflect the physiological and working time factors of the subjects than simple binary segmentation.

For the two clusters, a comparison of the mean Z-scores of the main variables first showed that cluster 0 showed a markedly lower negative value in the shift work year (Z ≈ −0.48, average = 5.83 years) and was below average age (Figure 3A). This could be seen as a young, short-term shift work group, and other metabolic and sleep-related indicators did not show any significant deviation from the mean. Cluster 1 showed a markedly high positive value in the shift work year (Z ≈ +1.75, average = 12.5 years), and the age was also relatively high (Figure 3B). This cluster could be interpreted as an older, longer-term shift group, with total cholesterol, HDL, and LDL above average, while triglycerides remained around the population mean. Conversely, BMI, blood pressure (SBP and DBP), fasting blood glucose, triglyceride, and insomnia scores were generally below average or showed no significant difference (Figure 3A). In summary, the k = 2 cluster was largely separated by the difference between shift work period and age, which exhibited distinct characteristics between the long-term and early-shift groups.

For the four clusters, to find out about the mean Z-scores of the key variables, Cluster 0 showed below-average values for shift work tenure, age, and most metabolic indicators, with particularly low values for blood pressure and blood glucose (Figure 4A). This can be viewed as a low-risk group or a group with relatively good health indicators. Cluster 1 showed distinct positive values for shift work tenure (Z ≈ +2.0, Mean = 13.3 years) and age (Z ≈ +0.3, Mean = 40.9 years) with no significant deviation in other metabolic indicators. This cluster is interpreted as an older group with long-term shift work experience (Figure 4B). Cluster 2 had a slightly higher age and showed an increase in some lipid indicators but had low shift work tenure. Additionally, total cholesterol and LDL were below average. This can be interpreted as a group with a moderate level of metabolic risk and age distribution (Figure 4C). Cluster 3 had low shift work tenure and age, but showed above-average values for BMI, blood pressure (SBP, DBP), blood glucose, and lipids (total cholesterol, triglycerides, and LDL). This cluster can be characterized as a young shift work group with high metabolic risk (Figure 4D).

In summary, the k = 4 clusters were categorized as (1) a low-risk, healthy group, (2) a long-term shift work, relatively older group, (3) a moderate-risk group, and (4) a relatively young shift work group with high metabolic risk. This shows that employee groups can be distinctly characterized by a combination of shift work tenure, age, and metabolic indicators. Accordingly, the results of the supervised modeling and interpretation will be explained below, based on the k = 4 solution.

Cross-validated evaluation of the Random Forest classifier showed that both clustering solutions (k = 2 and k = 4) achieved stable and reproducible performance. For the k = 2 model, the mean accuracy across 10 × 5 repeated cross-validation was 0.976 ± 0.0027 (range 0.968–0.977), and macro-averaged precision, recall, and F1-scores were likewise consistent (0.971 ± 0.0017, 0.960 ± 0.0069, and 0.964 ± 0.0045, respectively). For the k = 4 solution, the classifier also demonstrated robust performance, with an accuracy of 0.897 ± 0.016 (range 0.872–0.918) and stable macro-precision, recall, and F1-scores (0.904 ± 0.014, 0.901 ± 0.016, and 0.899 ± 0.016, respectively).

Cluster-specific performance metrics for both k = 2 and k = 4 including mean ± SD and the observed minimum–maximum ranges are summarized in Table 2. Overall, the model consistently reproduced the underlying cluster structure across repeated resampling, indicating strong classification stability and supporting the interpretability of the derived subgroup patterns.

A SHAP analysis was used to identify the variables contributing to the classification of each cluster. Overall, shift work experience (Shift_Exp) emerged as a consistently important variable across all clusters, showing that the type and duration of work are key factors in cluster differentiation. In Cluster 0, SBP and DBP had the greatest impact on the model's output, with age and shift work experience also playing significant roles (Figure 5A). Cluster 1 showed an overwhelmingly large contribution from shift work experience, confirming that long-term tenure is a major factor in determining cluster characteristics (Figure 5B). In Cluster 2, metabolic indicators such as LDL, total cholesterol (TC), and BMI had a relatively large influence, combining with shift work experience to distinguish the cluster (Figure 5C). Finally, in Cluster 3, LDL, BMI, and blood pressure indicators were important factors, with a particularly high contribution from lipid metabolism-related indicators (Figure 5D).

As such, the SHAP analysis shows that while each cluster is commonly influenced by shift work tenure, they are also specifically differentiated by a variety of physiological factors such as blood pressure, lipids, BMI, and age.

Although this study looked at metabolic health risks in shift workers through cluster analysis and SHAP interpretation, it has some limitations. First, since it is a cross-sectional study design, it is difficult to directly determine the causal relationship between shift work and metabolic abnormalities. Second, there is a limit to generalizing results for other industries or occupations because it was only targeted at workers in one workplace. Third, because it included self-report questionnaire data such as work type and lifestyle, there is a possibility of bias in response. Fourth, although the number of clusters was determined by the silhouette factor, this method may not fully reflect the clinical implications of all clusters. Fifth, the number of samples is limited, so the accuracy of 87.9%, which is the result of holdout validation of the random forest model, needs to be interpreted carefully. Sixth, since only night workers were included in the study, comparative analysis with day workers or shift workers was impossible. In the future, research including various types of work is needed. Finally, due to the lack of long-term follow-up results such as metabolic syndrome or cardiovascular disease incidence, it was not possible to assess whether the health risk of each cluster led to actual disease.