AAI_2025_Capstone_Chronicles_Combined

MENTAL HEALTH RISK DETECTION USING ML​

​

​

​

​

18

ROC AUC (Micro)

0.84

0.94

0.93

ROC AUC (Macro)

0.84

0.94

0.93

Class 0 AUC

0.83

0.94

0.93

Class 1 AUC

0.83

0.93

0.92

Class 2 AUC

0.85

0.95

0.93

Note. Tabular Neural Network shows it outperforms the other two models in all categories.

To further refine the model, we conducted a grid search on the TabNet architecture. We tested various combinations of step count, feature dimension, batch size, and learning rate. The best-performing configuration used 4 steps, 32 feature dimensions, a batch size of 256, and a learning rate of 0.001, achieving a validation accuracy of 78.7%. However, test performance improved by only 0.5%, indicating the model had already reached near-optimal performance with its original configuration. Model training behavior met expectations across all architectures. The Tabular Neural Network (TNN) demonstrated consistent learning with minimal signs of overfitting and early stopping (Arik & Pfister, 2021). Logistic Regression performed reliably without underfitting, aided by elastic net regularization. Figure 5 illustrates the full infrastructure of our system, from user input and data preprocessing to model inference and clinical interpretation, showcasing how the pipeline could function in a real-world deployment. Our research aimed to determine whether behavioral, psychological, and demographic survey features could accurately predict mental health risk. The results strongly support this hypothesis. The models performed consistently well using features such as stress, coping

218