Machine learning predicts self-reported anger improvement in teens

A machine-learning study found that survey answers about meditation, daily behavior, and lifestyle can help predict whether teenagers say they have become better at managing anger. However, the study does not show that meditation caused this improvement. Random Forest was the most accurate model on a separate test set, while XGBoost performed best on some measures of how well models distinguish between responses and identify the less common group. The results suggest that questionnaires may contain useful clues about how people manage their emotions. But their value in healthcare and everyday life is still uncertain because the study used a small, one-time survey based on people's own answers and did not test the models on an independent dataset.

The researchers studied a publicly available dataset called the “Yoga & Meditation Benefits Survey,” shared through Kaggle. It contained 400 records and 14 input features covering age and other demographic details, meditation and yoga practices, behavior, health-related information, and people's views of their well-being. The outcome was a yes-or-no answer to whether respondents felt their anger management had improved. A total of 331 people, or 82.75%, answered “Yes,” while 69 people, or 17.25%, answered “No.” Because many more people reported improvement than did not, the researchers used measures that accounted for this imbalance as well as standard accuracy scores.

The researchers compared three machine-learning methods: Random Forest, Support Vector Machine, and Extreme Gradient Boosting, also called XGBoost. They tested each model in two ways. First, they divided the data using a stratified 60:40 split, with 40% of the records held back as a separate test set. Second, they used stratified five-fold cross-validation to check performance across different data splits. Their process included filling in missing values, converting category answers into numbers, selecting features using mutual information, and scaling features when needed. They also used the Synthetic Minority Oversampling Technique, or SMOTE, to create extra examples of the less common “No” group. This was done only within the training folds, so the separate test data and validation observations were not artificially increased.

The researchers chose each model's settings through randomized searches, using weighted F1-score to find the best combination. Weighted F1 takes the different sizes of the two groups into account. They also used macro-averaged precision, recall, and F1 to check how well the models handled both “Yes” and “No” answers, rather than mainly measuring success on the larger group. They calculated ROC-AUC, which measures how well a model distinguishes between classes, by taking the average area under the curves for both classes. Confusion matrices helped them examine false-positive and false-negative results. They fitted the data-preparation steps and feature selection using training data only. This was intended to prevent information from the test data leaking into the training process and to make the performance estimates more trustworthy.

On the independent test set of 160 observations, Random Forest achieved the highest accuracy, at 86.25%, correctly classifying 138 responses. Support Vector Machine and XGBoost each reached 83.75%. The ranking was different when measured by ROC-AUC: XGBoost scored 0.8650, Random Forest scored 0.8558, and Support Vector Machine scored 0.8259. Random Forest also achieved the highest weighted F1-score, at 0.8558, and the highest macro F1-score, at 0.7393. This suggests it was somewhat better at recognizing the less common “No” responses. However, Support Vector Machine had the highest macro recall, at 0.7468. In the test set, Random Forest correctly identified 14 of the 28 people who answered “No” and 124 of the 132 who answered “Yes.” This shows that even though its overall accuracy was high, it still struggled to identify the less common group.

Five-fold cross-validation showed a similar overall pattern, although the differences between the models were smaller. Random Forest had the highest average accuracy, at 83.00% plus or minus 1.90%, and its results varied the least across the five folds. Support Vector Machine and XGBoost both had an average accuracy of 82.50%. However, Support Vector Machine showed greater variation between folds, at 4.42%, compared with 3.31% for XGBoost. XGBoost achieved the highest average macro ROC-AUC, at 0.8447 plus or minus 0.0489, the highest average F1-score, at 83.01% plus or minus 2.70%, and the highest average precision, at 84.02% plus or minus 2.19%. When the researchers combined predictions made on held-out folds, XGBoost stood out for identifying more “No” responses. Its recall for this less common group was 0.58, compared with 0.41 for Random Forest and 0.52 for Support Vector Machine.

Methods used to understand the models pointed to three features that appeared repeatedly: becoming more socially active and confident, improving concentration and focus during study or work, and a score measuring mindfulness or being present in the moment. Random Forest's feature-importance analysis ranked social activity and confidence first, followed by concentration and focus, then mindfulness or presence. Other features, including how often people meditated, social-media use, sleep quality, meditation duration, and age group, had lower importance scores. SHapley Additive exPlanations, or SHAP, generally supported this ranking. SHAP estimates how individual features contribute to a model's predictions, including whether a feature pushes a particular person's prediction in one direction or the other.

These important features should not be treated as proof of how anger improves or as evidence that a treatment works. The researchers did not randomly assign people to meditation, yoga, or other lifestyle practices. They did not measure anger before and after an intervention, and they did not establish that meditation caused respondents to report better anger control. Instead, some features may reflect people's general tendency to give similar answers across a survey, overlapping wording in questions, links between different habits, or characteristics that develop after people feel they have better control of their anger. The study itself stresses that its findings show predictive relationships, not cause and effect. Likewise, the two methods used to explain the models can show which features influence predictions, but they cannot prove that changing a feature would change the outcome.

Several limitations make it difficult to apply the findings to other groups. The sample was relatively small, the two response groups were uneven in size, and all the main information came from people's own answers to a survey taken at one point in time. Even after using SMOTE and evaluation measures designed to account for class imbalance, the models performed worse on the less common “No” responses than on “Yes” responses. The researchers tested the models using a held-out sample and cross-validation within the same dataset, but they did not test them on new data collected independently from teenagers in other settings or at other times. The article provided also did not describe the dataset's age, geographic, and demographic makeup in enough detail to show how well the findings represent teenagers in general.

Pairwise Wilcoxon signed-rank tests found no statistically significant differences between the models at the chosen threshold of p less than 0.05. The comparison between Random Forest and Support Vector Machine gave p = 0.2482. The comparison between Random Forest and XGBoost also gave p = 0.2482, while the comparison between Support Vector Machine and XGBoost gave p = 1.0000. These results suggest that machine learning can potentially predict survey responses using structured information, but they do not establish that one algorithm is clearly better than the others. Larger studies that follow people over time, include objective measures of behavior or physical responses, use more rigorous validation that keeps participants separate between training and testing, and test models on independent datasets are needed to find out whether the predictive patterns hold up. For now, the most reasonable conclusion is that information about meditation and behavior may help computers estimate whether teenagers report better anger management. The findings do not show that meditation or any other selected feature causes that improvement.

Subject of Research: Using machine learning to predict self-reported improvement in anger management among teenagers from survey data about meditation, behavior, and lifestyle

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