Psychological distress, digital behavioral risks, and physical activity as predictors of subjective wellbeing: a machine learning study in a large sample of community-dwelling adults

Background Regular physical activity is a key determinant of psychological wellbeing, yet its interaction with emerging digital behavioral risks remains insufficiently understood. Compulsive internet use and nomophobia have been linked to psychological distress, which negatively affects life satisfaction and happiness. However, few studies have examined these factors simultaneously. Machine learning offers a promising approach for improving predictive accuracy and clarifying the relative contributions of these variables to subjective wellbeing. Aim This study examined the combined predictive roles of physical activity, psychological distress, compulsive internet use, and nomophobia in subjective wellbeing (life satisfaction and happiness) and identified the optimal machine learning models for each outcome. Methods A cross-sectional study was conducted among 1,479 community-dwelling adults (51.05% males) recruited between September and December 2024. Participants completed validated Arabic versions of the IPAQ-SF, DASS-21, CIUS, NMP-Q, SWLS, and SHS. Six machine learning algorithms (linear regression, random forests, support vector machines, XGBoost, k-nearest neighbors, and LASSO) were trained using an 80/20 train–test split with 5-fold cross-validation. Performance was evaluated using R 2 , RMSE, and MAE. Results Psychological distress was the strongest negative predictor of wellbeing, showing the largest associations with happiness ( r = −0.40) and life satisfaction ( r = −0.37). Compulsive internet use was moderately associated with distress ( r = 0.48), whereas nomophobia showed negligible relationships with both outcomes. For subjective happiness, random forest achieved the best performance ( R 2 = 0.154, RMSE = 0.869), slightly outperforming linear regression and LASSO ( R 2 = 0.153). For life satisfaction, linear regression and LASSO performed best ( R 2 = 0.129, RMSE = 1.317), while support vector machines showed the lowest accuracy. Higher physical activity levels, particularly vigorous activity, were consistently associated with more favorable psychological profiles. Conclusion Prediction of subjective wellbeing is outcome-specific, with different machine learning models performing optimally for happiness and life satisfaction. Psychological distress emerged as the strongest negative predictor, whereas physical activity demonstrated a consistent protective association. Longitudinal studies are needed to clarify causal pathways linking digital behavioral risks to wellbeing.