AAI_2025_Capstone_Chronicles_Combined

MENTAL HEALTH RISK DETECTION USING ML​

​

​

​

​

11

Neural Network

Clinical Data (Somepalli et al., 2021) Personalized Behavioral Risk Monitoring Hybrid Mental Health Classifier (Jain et al., 2025)

models for disease onset prediction using structured clinical data. Used deep learning to detect evolving mental health risk from behavioral and emotional changes. Combined Logistic Regression and neural networks to improve classification of psychiatric risk levels.

Soft Voting Ensemble

Note. Some study labels above are paraphrased for clarity.

Experimental Methods Three distinct machine learning models were implemented for mental health risk detection: Logistic Regression, a Tabular Neural Network (TNN), and a Soft Voting Ensemble. Each model was selected for its unique strengths in handling behavioral and psychological data. Logistic Regression (LR) serves as the baseline model due to its simplicity, computational efficiency, and widespread clinical acceptance for interpretable outcomes (Alkhamees & A.M., 2021). The TNN, a deep learning model tailored for tabular data, captures non-linear and high-dimensional feature interactions using components such as gated linear units (GLUs), batch normalization, and an attentive transformer (Arik & Pfister, 2021). To experiment further, we test a Soft Voting Ensemble. This integrated the prediction of the LR and TNN (Jain et al., 2025). The proposed modeling pipeline involves structured data preprocessing and stratification to ensure balanced representation. K-Modes clustering is used to derive unsupervised labels for mental health risk levels (low, medium, high). For classification, Logistic Regression with polynomial features and elastic net regularization is selected to enhance interpretability while

211