Machine learning predicts self-reported anger improvement in teens

A machine-learning analysis found that survey data describing meditation habits, daily behavior and lifestyle can predict whether adolescents report improved anger management, although the study offers no evidence that meditation caused the reported improvement. Random Forest produced the highest accuracy on an independent test set, while XGBoost led on measures of discrimination and minority-class detection. The results suggest that structured questionnaires may contain useful signals about emotional regulation, but their clinical and practical value remains uncertain because the analysis used a small, cross-sectional and self-reported dataset without external validation.

The observational study analyzed the publicly available “Yoga & Meditation Benefits Survey” dataset distributed through Kaggle. It included 400 records and 14 predictors covering demographic characteristics, meditation and yoga practices, behavioral habits, health-related information and perceived well-being. The outcome was a binary survey response to whether respondents felt an improvement in anger management: 331, or 82.75%, answered “Yes,” while 69, or 17.25%, answered “No.” Because the target was imbalanced toward perceived improvement, the researchers used class-sensitive metrics in addition to conventional accuracy.

Three supervised classification algorithms were compared: Random Forest, Support Vector Machine and Extreme Gradient Boosting, or XGBoost. Each was evaluated in two complementary ways: a stratified 60:40 split in which 40% of the observations formed an independent hold-out test set, and stratified five-fold cross-validation to assess performance across different partitions. The pipeline included missing-value imputation, categorical encoding, mutual-information feature selection and, where required, feature scaling. Synthetic Minority Oversampling Technique, or SMOTE, generated additional examples of the minority “No” class, but only within training folds; the independent test data and validation observations remained unsynthesized.

Hyperparameters for each algorithm were selected through randomized searches using weighted F1-score as the optimization criterion. Weighted F1 was intended to account for the unequal class sizes, while macro-averaged precision, recall and F1 assessed performance across both “Yes” and “No” rather than concentrating on the majority response. The researchers also used ROC-AUC, calculated as the mean area under the curves for the two classes, and confusion matrices to examine false-positive and false-negative patterns. Preprocessing and feature selection were fitted to training data only, a design intended to reduce information leakage and make the generalization estimates more credible.

On the 160-observation independent test set, Random Forest achieved the highest accuracy at 86.25%, correctly classifying 138 responses. Support Vector Machine and XGBoost each reached 83.75%. The ranking changed under ROC-AUC: XGBoost achieved 0.8650, Random Forest 0.8558 and Support Vector Machine 0.8259. Random Forest also had the highest weighted F1-score, 0.8558, and macro F1-score, 0.7393, indicating somewhat better recognition of the minority “No” responses. Support Vector Machine, however, recorded the highest macro recall at 0.7468. Random Forest correctly identified 14 of 28 “No” cases and 124 of 132 “Yes” cases in the test set, illustrating that its high overall accuracy still reflected difficulty with the less frequent class.

Five-fold cross-validation produced a broadly similar pattern, with less dramatic differences between algorithms. Random Forest recorded the highest mean accuracy at 83.00% plus or minus 1.90% and showed the lowest fold-to-fold variability among the three models. Support Vector Machine and XGBoost each averaged 82.50% accuracy, although Support Vector Machine showed greater variability at 4.42%, compared with 3.31% for XGBoost. XGBoost achieved the highest mean macro ROC-AUC at 0.8447 plus or minus 0.0489, mean F1-score at 83.01% plus or minus 2.70% and mean precision at 84.02% plus or minus 2.19%. Across all out-of-fold predictions, XGBoost was especially notable for detecting more “No” responses, with minority-class recall of 0.58, compared with 0.41 for Random Forest and 0.52 for Support Vector Machine.

Model-interpretation methods highlighted three recurring characteristics: becoming more socially active and confident, improvement in concentration and focus during study or work, and a mindfulness or presence score. Random Forest feature-importance analysis placed social activity and confidence first, followed by concentration and focus and then mindfulness or presence. Variables including meditation frequency, social-media use, sleep quality, meditation duration and age group had comparatively lower importance. SHapley Additive exPlanations, or SHAP, broadly supported this ranking by estimating how individual characteristics contributed to predictions, including the direction of each contribution for a given respondent.

Those influential characteristics should not be mistaken for mechanisms or treatment effects. The analysis did not randomize meditation, yoga or lifestyle practices, measure anger before and after an intervention, or establish whether respondents improved because of meditation. Some predictors may instead reflect shared response tendencies, overlapping survey wording, correlations among habits or characteristics that arise after people perceive better anger control. The study itself emphasizes that its findings are predictive associations rather than causal effects. Similarly, the use of two machine-learning explainability tools can show which variables a model relies upon, but it cannot demonstrate that changing one of those variables would change the outcome.

Several limitations constrain how far the results can be generalized. The sample was relatively small, the target classes were imbalanced and all key information was self-reported in a cross-sectional survey. Minority-class performance remained weaker than performance for “Yes” responses despite the use of SMOTE and balanced evaluation metrics. The models were tested internally through a hold-out split and cross-validation but were not evaluated on independently collected data from different adolescents, settings or time periods. The age, geographic and demographic distribution of the dataset was also not described in enough detail in the supplied article to establish how broadly the findings represent teenagers.

Pairwise Wilcoxon signed-rank tests found no statistically significant differences among the models at the predefined threshold of p less than 0.05. Random Forest versus Support Vector Machine and Random Forest versus XGBoost both produced p = 0.2482, while Support Vector Machine versus XGBoost produced p = 1.0000. The results therefore support the feasibility of prediction from structured survey characteristics, not the claim that one algorithm is definitively superior. Larger, longitudinal datasets incorporating objective behavioral or physiological measures, nested or participant-independent validation and external testing would be needed to determine whether the apparent predictive signal persists. Until then, the most defensible conclusion is that meditation-related and behavioral variables may help machines estimate self-reported anger improvement in adolescents, without showing that meditation or any other selected characteristic causes it.

Subject of Research: Machine-learning prediction of self-reported anger-management improvement in adolescents using meditation, behavioral and lifestyle survey data

Article Title: Machine learning based assessment of meditation for teen anger management

Article References: Rao, D. M., Gopi, K., & Panuganti, S. (2026). Machine learning based assessment of meditation for teen anger management. Discover Artificial Intelligence, 6(1), Article 1562. https://doi.org/10.1007/s44163-026-02565-6

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02565-6

Keywords: machine learning, adolescent mental health, anger management, meditation, predictive modeling, Random Forest, XGBoost, explainable AI, survey data, Machine, learning, based