AAI_2025_Capstone_Chronicles_Combined

MENTAL HEALTH RISK DETECTION USING ML​

​

​

​

​

9

Tabular Neural Networks (TNNs), such as TabNet, have recently demonstrated competitive performance on structured datasets (Kim, 2025). TNNs utilize embedding layers, gated linear units (GLUs), and attention-based feature selection to model complex, non-linear relationships in structured data (Arik & Pfister, 2021). Somepalli et al. (2021) found that these models outperformed traditional tree-based models such as Random Forest in clinical prediction tasks, including disease onset and classification. TNNs are thus well-suited for modeling multifaceted mental health data where relationships between coping strategies, stress, and behavioral patterns are not linearly separable. Jain et al. (2025) demonstrated that combining classifiers using a soft voting strategy can enhance the sensitivity and specificity of mental health disorder detection. Motivated by this, we implemented a soft voting ensemble, shown in Figure 3 , that integrates predictions from both Logistic Regression and Tabular Neural Network models to strike a balance between interpretability and predictive performance.

Figure 3 Selected Models for Mental Health Risk Detection

Note. The selected architecture integrates Logistic Regression for interpretability and a Tabular Neural Network for non-linear pattern recognition. A Soft Voting Ensemble combines its outputs to enhance robustness and predictive accuracy.

209