XGBoost · 0.9610
The strongest AUC-ROC suggests the model was best at separating positive and negative cases across thresholds.
03 / Research / Study 02
Human-centered AI · Mental-health technology
Predictive analytics of adolescent behavioral patterns for mental-health forecasting.
A multimodal machine-learning framework that studies how academic, lifestyle, and psychological signals can support earlier recognition of adolescent mental-health risk.
01 / At a glance
Turning scattered signals into earlier support.
4models compared
0.961XGBoost AUC-ROC
5evaluation metrics
98%data retained after cleaning
The study addresses a practical problem: self-reports can be delayed, incomplete, or affected by bias, while changes in sleep, study load, academic pressure, and emotional well-being may appear together across daily life.
It frames prediction as a way to support timely human attention, not as a replacement for clinical judgment. The proposed system is a research framework and should not be used as a diagnostic tool.
02 / Data and method
A multimodal view of student well-being.
Inputs examined
- Demographics and academic context, including age, gender, academic pressure, CGPA, and study satisfaction.
- Lifestyle patterns, including sleep duration, screen time, physical activity, and work or study hours.
- Psychological indicators, including stress triggers, financial stress, and depression status.
- Visual exploration of relationships through heatmaps and scatter plots.
Model pipeline
- Screen the dataset and remove irrelevant or redundant variables.
- Impute missing values and normalize numerical features.
- Compare logistic regression, SVM, random forest, and XGBoost.
- Use stratified 5-fold cross-validation and evaluate accuracy, precision, recall, F1-score, and AUC-ROC.
03 / Reported results
Different metrics reveal different strengths.
| Model | Accuracy | Precision | Recall | F1-score | AUC-ROC |
|---|---|---|---|---|---|
| Logistic regression | 0.8550 | 0.9149 | 0.8037 | 0.8557 | 0.9218 |
| Random forest | 0.8375 | 0.8461 | 0.8831 | 0.8642 | 0.9116 |
| SVM | 0.8429 | 0.8507 | 0.8874 | 0.8687 | 0.9103 |
| XGBoost | 0.8319 | 0.8314 | 0.8319 | 0.8315 | 0.9610 |
Logistic regression
It reached the highest accuracy and precision in the reported comparison, making it a useful interpretable baseline.
SVM
SVM reported the highest recall and F1-score, highlighting a different trade-off for identifying possible positive cases.
04 / Feature insights
Academic pressure, sleep, and study time stood out.
The paper identifies academic pressure as the most important feature in the XGBoost analysis. Shorter sleep and longer study hours also appeared as meaningful signals, connecting the model's output to practical intervention questions in educational settings.
The study also reports weaker performance for extreme sleep-duration groups and notes that privacy, sampling bias, and demographic diversity remain important limitations before real-world deployment.
Read the source
Explore the complete dataset discussion, figures, and references.
05 / Figures from the paper
Selected views of the prediction framework.
