Machine Learning Screening for Work Addiction Risk

This is part of the global project on work addiction, which was financed by the National Science Centre, Poland, under grant no. 2020/39/D/HS6/00198 ("The role of macro-, meso-, and micro-level factors in work addiction and related health problems"). The project is led by the consortium of the University of Silesia in Katowice, Poland (Prof. Edyta Charzyńska) and the University of Gdańsk, Poland (Prof. Paweł A. Atroszko). The dataset “data” (available in CSV UTF-8 and XLSX formats) contains data collected online from 34,899 full-time employees from 91 countries or territories spanning six continents. Data were collected online between autumn 2022 and winter 2023, with data collection in Cambodia taking place in autumn 2024. After applying exclusion criteria (missing values on predictor or outcome variables, fewer than 30 working hours per week, fewer than 150 observations per country or territory), the final analytical sample comprised employees from 85 countries or territories. Two analytical datasets were constructed: Dataset 1 (psychosocial model, n = 22,136), including demographic, occupational, and psychosocial variables (job satisfaction, job stress); and Dataset 2 (demographic model, n = 23,219), including only demographic and occupational variables. The data were used to compare two machine learning-based screening strategies for work addiction risk: a psychosocial model incorporating job satisfaction and job stress, and a demographic-only model relying solely on readily available organizational data. Work addiction risk was operationalized using the five-item International Work Addiction Scale (IWAS-5; Charzyńska et al., 2025), with participants scoring 18 or above classified as at-risk. Analyses included: Elastic Net regularization for feature selection (α = 0.005, l1_ratio = 0.7); stratified 5-fold cross-validation (k = 5, random state = 42); and evaluation of five machine learning algorithms applied to both full and Elastic Net-selected feature sets — Logistic Regression, Random Forest, XGBoost, Neural Networks, and an Ensemble Voting classifier — yielding 10 model configurations per dataset. Performance was assessed using balanced accuracy, AUC-ROC, sensitivity, specificity, Youden's J statistic, Matthews Correlation Coefficient (MCC), F1-score, PPV, and NPV. Analyses were conducted in Python 3.13 using scikit-learn 1.7, TensorFlow 2.20, XGBoost, Pandas 2.3, NumPy 2.3, and Matplotlib 3.10. The results indicate that the psychosocial model demonstrated superior screening performance. The best-balanced psychosocial model (Logistic Regression with Elastic Net selection) achieved balanced accuracy of 0.72, AUC-ROC of 0.80, and Youden's J of 0.45, correctly identifying 72.7% of at-risk individuals while maintaining 72.2% specificity. The demographic-only model retained meaningful screening utility, detecting 57.0% of at-risk cases (Ensemble Voting, balanced accuracy 0.61, AUC-ROC 0.65), supporting its use as a low-burden preliminary screening tool in organizational settings. Full List of Funding Sources (Including Those Related to Data Collection in Specific Cultures) This work was supported by the National Science Centre, Poland [grant number 2020/39/D/HS6/00198]. Data collection in Armenia was supported by the Science Committee of the Republic of Armenia in the frames of the research project 25RG-5A025. Data collection in the Czech Republic was supported by financing from NPO “Systemic Risk Institute” no. LX22NPO5101, funded by European Union - Next Generation EU (Ministry of Education, Youth and Sports, NPO: EXCELES). Data collection in Estonia was supported by the Estonian Research Council grant (PRG2190). Data collection in Hungary was supported by the Hungarian National Research, Development, and Innovation Office (Grant number: FK134807) and the János Bolyai Research Scholarship of the Hungarian Academy of Sciences granted to Bernadette Kun. Acknowledgments The authors are deeply grateful to the members of the team involved in the Global Research on Work Addiction for their contributions to data collection and their overall involvement in the project. A list of international collaborators and their bios is available on the project’s website: https://workaddiction.org/team. Files Attached (i) data — raw survey data (CSV/XLSX format)(ii) codebook — variable names, labels, measurement levels, response categories, and missing value codes (CSV/XLSX format)(iii) config.py — Python configuration file defining all file paths, data processing parameters, feature definitions, and model settings(iv) preprocessing.py — Python script for data loading, cleaning, filtering, feature engineering, and export of the analysis-ready dataset(v) main.py — Python script implementing the full machine learning pipeline: Elastic Net feature selection, 5-fold cross-validation, model training and evaluation, visualization, and report generation(vi) README — documentation describing all files, methodology, and instructions for reproducing the analyses