This study used nationally representative data from the DEGS1 study, which is a part of the health monitoring program of the Robert Koch Institute, Berlin, Germany. It was conducted from 2008 to 2011 by means of interviews, examinations, and tests among the German population aged 18-79 years (n=8151). The DEGS1 data set, which is available for public use on request, included measurements for chronic stress among 5801 respondents aged 18 to 64 years [6,19].

Chronic stress was assessed using the 12-item German short version of TICS-SSCS (n=5850) [6]. It was developed by Schultz et al [8] based on the systemic-requirement-resource model of health [8,20]. The 12-item scale addresses 5 stress areas: chronic worrying, work overload, social overload, excessive demands of work, and lack of social recognition. Its internal consistency showed a Cronbach α of .91 and a good to very good reliability with values ranging from .84 to .91 (mean α=.87) [8]. All 12 questionnaire items use a 5-point Likert scale answer format (0=“never” to 4=“very often”) to measure chronic stress in the past 3 months [21,22]. A sum score (scale 0-48) was calculated for each participant, which is categorized in 3 classes based on a reference population with the TICS-SSCS: 1-11 (≤median)=low stress, 12-22=middle stress, and >22=high stress (≥90th percentile). This multiclass outcome is the recommended DEGS1 approach [6].

In addition, the DEGS1 data set included variables on sociodemographic characteristics, chronic diseases (eg, coronary heart disease, stroke, diabetes mellitus, depression, and anxiety disorder), living conditions, health-related behavior, preventive measures, and general health. Based on a literature review and using the Powershap feature selection method, 34 features were included in this analysis. Table 1 depicts descriptive information about the variables used.

To evaluate the method proposed in this study, we used the following most promising multiclass evaluation metrics: the area under the receiver operating characteristic curve (AUC), precision, recall, and F₁-score. Multiclass classification works on data sets in which all classes are mutually exclusive. In a multiclass classifier, the evaluation measures of individual classes are averaged out to determine the performance on overall system across the data. We applied the macroaverage approach [29].

The receiver operating characteristic (ROC) curve was used to evaluate the performance of the classifier. For different classification thresholds, the macro true-positive rate (equation 3) is plotted against the macro false-positive rate (equation 4). The AUC indicates the classifier’s ability to distinguish between classes. The value of the AUC is in the range (0,1), in which 1 is for a perfect classifier. In this study, the ROC curve is plotted for each class broken down into a series of binary problems using the One-vs-Rest approach. The macroaverage is computed by summing the individual values for true positive, true negative, false positive, and false negative. Then, macroaverage scores of true positive instances (precision; equation 5), true positive rate (recall; equation 6), true negative rate (specificity; equation 7), and the harmonic mean of the precision and recall computed on each class (F₁-score; equation 8) were computed. Mathematically, they are defined as follows:

We used Python 3.7 (Python Software Foundation) to implement our ML framework. In addition, several libraries from the python data science ecosystem were used to execute the experiments and the integrated development environment PyCharm. To implement the Powershap feature selection method, we used the Powershap Python library. The scikit-learn package (version 1.0.2) was used to train and evaluate the ML classifier. SHAP tool (version 0.40.0) was used to assess the explainability the model; that is, to identify factors protecting against chronic stress.

In addition to the performance evaluation, this study maximizes the interpretability of the underlying models. It focuses particularly on the explainability of the model, which can serve as an indispensable tool in the era of precision medicine.

Per our understanding, the interpretation of the prediction models is as crucial as the prediction accuracy because it extracts information that significantly affects outcomes and identifies the factors protecting against chronic stress from subjects with lower chronic stress. However, the ensemble learning method XGBoost represents a black-box model. To overcome this problem, Lundberg [30,31] proposes the SHAP approach for interpreting predictions of complex models created by different techniques; for example, NGBoost, CatBoost, XGBoost, LightGBM, and scikit-learn tree models. SHAP was initially developed by Shapley in 1953 and is based on the game theory [32]. It explains the prediction of a specific input (X) by calculating the impact of each feature on the prediction. The estimated Shapley values are calculated as follows (equation 9):

where is the prediction for x, but with a random number of feature values. TreeSHAP is used for gradient boosting models including XGBoost. It offers a rich visualization of each feature attribution and allows for partial dependence plots.

The TreeSHAP interaction values estimates as follows (equation 10):

where i ≠ j, δ_ij(S) = f_x(S ∪ {I,j} – f_x(S ∪ {i} – f_x(S ∪ {j} + f_x(S), M is the number of features, and S denotes all feature subsets. SHAP values advance the understanding of tree models by including feature importance, feature dependence plots, local explanations, and summary plots [30].

Ethics approval for the DEGS1 survey was obtained from the Charité – Universitätsmedizin Berlin Ethics Committee (EA2/047/08). All participants received written information and provided informed consent before the interview and examination. The analysis described here builds on a data set from the DEGS1 study, which was kindly provided by the Robert Koch Institute. This secondary analysis of anonymized data does not require a separate ethics vote.

The mean age of the 5801 DEGS1 study participants was 44 years, with more than half of the population being female (n=3080, 53.1%). The mean stress level of the total population was 12.00 (95% CI 11.79-12.20): 11% (n=625) of the participants had “high chronic stress” (category 2), while 38% (n=2188) had “middle” (category 1), and 52% (n=2988) of them had “low chronic stress” (category 0). Most participants reported their general state of health as very good or good (79.3%, n=4599). Table 1 shows the weighted demographic, clinical, and laboratory characteristics of the participants.

The evaluation metrics of the XGBoost model’s performance are presented in Table 3 differentiated by chronic stress classes. We see that the XGBoost model achieved the highest AUC score for class 2 with 0.89% and a good macroaverage AUC score of 81% for the overall model. The metrics for the 3 stress classes and the average results are reported in Table 3. The ROC curves for the multiclass chronic stress prediction of the XGBoost model are shown in Figure 1.

The result of the SHAP analysis is displayed in Figure 2. In this plot, the impact of a feature on the respective classes (stress classes 0-2) is stacked to illustrate the feature importance. This means that the features with large absolute Shapley values are more important than those with lower values. The plot shows that class 0 (low level of chronic stress) hardly uses the features gender, general state of health, satisfaction with living space, and social support. Class 2 as the high level of chronic stress uses the features number of sick days in the past 12 months, social support, sleeping hours per night in the past 4 weeks, gender, and general state of health. Interestingly, classes 0 and 2 use many identical features.

While the SHAP feature plot provides an overview of the role of each variable irrespective of the direction of these effects, the SHAP summary plot provides such additional information for classes. The impact distribution of each feature on the model output for classes with low and high levels of chronic stress is shown in Figures 3 and 4. Each row in this plot represents a single feature in order of their mean absolute SHAP values. It can be a negative or positive value and represents the importance of each feature. Each dot is a Shapley value for a particular feature and reflects its impact on a specific class for a given instance, and dots stack up to show density. It is color-coded in accordance with the magnitude to which the value contributes to the model impact (red=high and blue=low). The color is the actual feature value in the data set. For example, the red values for age as a continuous feature represent older people, while blue values represent younger people, and blue values for gender as a categorical feature (low value=1) represent males and red values (high value= 2) represent females. Overlapping points are jittered toward the y-axis, giving a sense of the distribution of the Shapley values per feature.

According to the SHAP summary plot result, gender is the most significant feature for class 0, and the number of sick days in the past 12 months has the highest impact on class 2. We note that the general state of health (shown in red) with high values has negative SHAP values and a relatively negative effect on the model for the low level of chronic stress and a positive impact (positive SHAP values) for class 2. Higher values on the social support scale have a positive impact on class 0 and negative effects on class 2, which means that chronic stress is less likely with strong social support.

To our knowledge, this is the first study to select the XGBoost algorithm as an ML multiclass classifier in the prediction of chronic stress as well as the SHAP method to interpret the model’s prediction. Based on nationally representative German data, chronic stress was predicted using 34 characteristics of adult participants. We identified male gender, a very good general state of health, high satisfaction with living space, strong social support, enough sleep, and more than 4 hours of sports activities per week as protective factors against chronic stress. These results are in line with those of other studies, which showed that resilience against chronic stress is promoted by social support, family connectedness, and friendship networks in the community [33-36]. For example, with a sample of 24,347 participants from the Canadian General Social Survey, Van der Horst et al [36] determined that good friendship networks are positively associated with less stress, better health, and more social support. A cross-sectional study of 538 nursing students from an Australian university showed that social support positively affect the psychological well-being [37].

Our ML approach allowed for the inclusion of a broad spectrum of individual characteristics, which comprised medical, lifestyle, living space, and social information, while other studies on chronic stress used multivariate models with fewer parameters only. For example, a large cross-sectional study with 34,129 participants from China, Ghana, India, Mexico, Russia, and South Africa showed positive associations of multimorbidity, stroke, depression, and hearing problems with perceived stress without assessing potential protective factors such as living space and social support [38]. A US cross-sectional telephone survey with 340,847 participants aged between 18 and 85 years documented that psychological well-being, especially stress, improved, but integrated only 5 parameters such as gender, employment status, partnership, and underage children in the household in their model analyzed [39]. In a study with 12,110 working adults from Minnesota, United States, a high level of perceived stress was associated with a higher-fat diet, less exercising, and being a smoker using a multivariate model with 6 variable topics but did not include medical and living circumstances [40].

This study used the population-based, representative DEGS1 data set, which implies a low risk of selection bias; yet, the results may not be transferrable to other settings. The DEGS1 data, which were collected from 2008 to 2011, may not fully describe current living conditions in Germany, especially the potential effects of the pandemic, which were shown in other studies, were not measured [41]. In our study, the SHAP methodology allowed for a detailed visualization of single feature attributions, which improved the understanding of the ML model.

In this study, we developed an XGBoost ML model to predict chronic stress in adults. The SHAP methodology identified various relevant factors protecting against chronic stress, which need to be considered when developing interventions for stress reduction and improving resilience.