AAI_2025_Capstone_Chronicles_Combined

MENTAL HEALTH RISK DETECTION USING ML​

​

​

​

​

17

Procedure: Train Logistic Regression and Tabular Neural Network separately. Use VotingClassifier from scikit-learn with voting='soft'. Collect probabilities from each model. Average probabilities for final prediction Metrics: Accuracy, Macro F1-score, ROC-AUC

Initial optimization will focus on tuning the Tabular Neural Network, adjusting architectural elements such as step count, feature dimension, batch size, and learning rate to balance learning capacity and generalization. It is planned to use a grid search to find the optimal results. Results and Conclusion The machine learning models demonstrated strong capability in classifying individuals into low, moderate, or high mental health risk groups using structured survey data. TNN achieved the highest performance with 79% accuracy, a macro F1-score of 0.78, and ROC AUC scores of 0.94 (micro) and 0.93 (macro), indicating strong generalization. LR performed reliably with 73% accuracy and AUCs above 0.83, while SVE showed stable results but did not exceed TNN (Table 5). Overall, TNN was the most effective at capturing complex patterns in the data.

Table 5 Model Performance Comparison Across Metrics

Metric

Logistic Regression

TabNet-Inspired Model Soft Voting Ensemble

Accuracy

0.73

0.79

0.78

Macro F1-Score

0.73

0.78

0.78

Weighted F1-Score

0.73

0.79

0.78

217