Classifying Stress Using Behavioral and Psychological Data: A Comparison of Ensemble and Deep Learning Techniques

With the rise of digital health technology, monitoring mental health through computational methods has become highly significant. In this study, we looked at how well different machine learning algorithms can identify stress in two ways: by intensity levels (low, medium, and high) based on psychological and social factors, and by types of stress (academic stress, eustress, and distress) using behavioral indicators. We analyzed two datasets, one with $\mathbf{1, 1 0 0}$ samples and the other with $\mathbf{8 4 3}$ samples, which included 20 and 25 features, respectively. We put Random Forest, XGBoost, and Deep Neural Networks through some serious hyperparameter optimization and cross-validation tests to make our comparisons. We found that ensemble approaches were the most effective, with Random Forest achieving an accuracy of 88.18% for predicting stress intensity and XGBoost hitting $\mathbf{9 5. 8 6} \boldsymbol{\%}$ for classifying stress types. After running some statistical analysis, we figured out that the key predictors of stress outcomes were anxiety, self-esteem, and sleep quality. Our review includes detailed ablation experiments, feature importance rankings, and tests for statistical significance. The results from this study set realistic baselines for automated mental health assessment systems and show that ensemble methods outperform deep learning techniques on psychological datasets of moderate size. These findings align with established psychological theories and offer valuable insights for developing scalable stress monitoring technologies.