<![CDATA[Advances in technology have led to an increase in the frequency of smartphone, social media, and various digital app usage in daily life. These digital activities give rise to various digital behaviors—such as screen time, number of notifications, social media usage, and sleep patterns—which can be evaluated to understand users mental health. High digital device usage is suspected to be associated with increased anxiety levels among users. Therefore, this study aims to analyze the impact of digital behaviors on anxiety levels by utilizing explainable machine learning approaches and AI. This study uses the dataset, comprising 500 data points obtained from Kaggle. The research stages consist EDA, feature engineering, target leakage evaluation, comparison of several machine learning algorithms, cross-validation, and model interpretation using SHAP. Based on the EDA results, the variables social_media_time_min, notification_count, and digital_addiction_score showed a positive relationship with anxiety_level. During the model-building process, the variable digital_wellbeing_score was removed because it had a very high correlation with anxiety_level -0.84, which could lead to potential target leakage. The algorithm comparison revealed that the Ridge Regression model performed best compared to Random Forest, SVR, and XGBoost Regression, with an R² score of 0.221 and an RMSE of 1.627. Additionally, the results of 5-Fold Cross Validation showed an average R² Score of 0.158 with a standard deviation of 0.062, indicating that the model demonstrated fairly consistent performance. The SHAP interpretation reveals that notification_count and social_media_time_min are the variables that most strongly influence the prediction of anxiety_level. The results of this study indicate that digital behavior affects users anxiety levels, although the relationships among the variables remain quite complex. This study also emphasizes the importance of evaluating the target leakage and understanding the model in the development of machine learning-based mental health analysis to ensure that the prediction results are more objective and clear.]]>
