AAI_2025_Capstone_Chronicles_Combined

MENTAL HEALTH RISK DETECTION USING ML​

​

​

​

​

12

minimizing overfitting. A Tabular Neural Network is constructed with dense layers, batch normalization, and gated linear units (GLUs) to model non-linear interactions. These models are planned to be combined using a Soft voting Ensemble to enhance prediction stability across risk categories. The rationale for each model is summarized in Table 2 .

Table 3 Suitability of Selected Machine Learning Models

Method

Strengths

Why is it suitable for this project

Logistic Regression Interpretable, efficient, clinically trusted

Ideal for establishing a transparent and explainable baseline

Tabular Neural Network

Captures complex non-linear interactions in tabular data Balances interpretability and predictive power

Learns patterns in behavioral and environmental indicators

Soft Voting Ensemble

Improves stability and accuracy across risk categories

For Logistic Regression, we enhanced the model by adding interaction terms, new features based on pairs of existing variables, to help capture more complex relationships between factors. For example, the combined effect of stress and work hours may reveal patterns not apparent when considered independently. To avoid overfitting while still using all features, we applied elastic net regularization, which combines L1 (lasso) and L2 (ridge) penalties in equal proportion. This allows the model to remain accurate and generalizable without becoming too complex. We used the SAGA optimizer, which is specifically designed to work efficiently with elastic net and large datasets. Since our dataset was imbalanced across mental health risk categories, we used class weighting to ensure the model didn’t favor the majority class during

212