AAI_2025_Capstone_Chronicles_Combined

MENTAL HEALTH RISK DETECTION USING ML​

​

​

​

​

21

Note. ROC curves for the TabNet-Inspired Model showing high AUC scores across all risk classes, indicating strong class separability and well-calibrated predictions.

The primary goal was to develop a scalable, accessible tool to support early mental health risk screening. The consistency and fairness of our model outputs suggest they could inform decision-making in educational or workplace settings. Because the model performs reliably across all risk levels, it can be confidently applied in real-world screening systems where early and accurate detection is essential. Its sensitivity to high-risk cases ensures that those most in need are prioritized, making it especially useful in settings where mental health resources are limited. This improved generalization across risk categories and strengthened the model’s practical value. Unexpectedly, we found that even after simplifying our feature set by removing high-impact clustering variables like occupation, care options, and days indoors, the model’s performance remained high. This revealed strong predictive power in more subtle variables such as stress, social weakness, and help-seeking behavior, which we had initially assumed would be less influential. These findings suggest that nuanced behavioral and psychological features carry significant predictive weight in assessing mental health risk. Future improvements include SHAP-based explainability, secure API deployment, and the potential integration of richer contextual data. We also plan to explore alternative model architectures, such as transformer-based networks and large language models (LLMs), particularly if real-time behavioral or unstructured textual data becomes available. LLMs could enhance interpretability and expand the model’s ability to process qualitative inputs, such as journal entries or open-ended survey responses. Expanding the dataset to include clinically

221