Machine learning based assessment of meditation for teen anger management

Abstract Background Anger dysregulation during adolescence can negatively affect emotional well-being, interpersonal relationships, and academic functioning. Although meditation and yoga-based practices have been studied for emotional regulation, predictive machine-learning approaches remain comparatively limited. This study evaluates the feasibility of predicting self-reported improvement in anger management using behavioral, lifestyle, meditation, and yoga-related characteristics.

Methods: A publicly available survey dataset of 400 observations is analyzed. The prediction target is the response to “Do you feel an improvement in anger management?” , with 331 “Yes” and 69 “No” responses. Random Forest (RF), Support Vector Machine (SVM), and XGBoost are evaluated using a stratified 60:40 train-test split and five-fold stratified cross-validation. Preprocessing included imputation, categorical encoding, feature selection, scaling where required, and SMOTE within the modelling pipeline. Hyperparameters are optimized using RandomizedSearchCV with weighted F1-score.

Results: Random Forest achieved the highest independent-test accuracy with Hold-out approach 86.25% and the highest mean five-fold accuracy 83.00 ± 1.90%. SVM and XGBoost achieved 83.75% test accuracy, while XGBoost obtained the highest mean Macro ROC-AUC 0.8447 ± 0.0489 and mean F1-score 83.01 ± 2.70% in cross-validation. Feature-importance and SHapley Additive exPlanations (SHAP) analyses highlighted social activity/confidence, concentration/focus, and mindfulness/presence as influential predictors. No statistically significant differences are observed between models.

Conclusions: Machine learning can feasibly predict self-reported improvement in anger management from meditation-, behavioral-, and lifestyle-related characteristics. RF showed the strongest overall accuracy and stability, while XGBoost demonstrated stronger cross-validated discrimination. These findings represent predictive associations rather than causal effects and require validation using larger, longitudinal, independently collected datasets.