Data-Driven Demo Application for Adolescent Mental Health Screening with Decision Support

Abstract

We present a data-driven application for screening adolescent mental health status with decision-support functionalities. It focuses on substance use and uses National Survey on Drug Use and Health (NSDUH) 2021–2022 data. We follow four-steps (data extraction, statistical analysis, machine-learning analytics, AI application) emphasizing reusable implementation and clinical relevance. Which includes K-nearest neighbor imputation, variance, correlation analysis, training, and test sets split before computing chi-square statistics and Mutual Information Analysis (MIA) to avoid data leakage. This provided 25 features (e.g., depressive symptoms, sleep patterns, energy levels, restlessness, and appetite) and 16 labels. Top-performing models were chosen based on the F1-score, employing a 5-fold stratified cross-validation strategy and Synthetic Minority Over-sampling Technique for Nominal (SMOTEN). The resulting interactive Artificial Intelligence (AI) application integrates these models to screen substance use adolescents for potential mental health concerns. Clinicians can input responses, view predicted mental status, and interact with corresponding decision-support functionalities. This includes displaying model prediction probability, severity level, nearest-neighbor patient instances, and the distribution and count of similar features and labels. Moreover, users can also explore the co-occurrence of features linked to specific labels. Although clinical judgment supersedes algorithmic predictions, this application serves solely as a research and demonstration tool and is not clinically validated. It is hosted at: https://huggingface.co/spaces/pantdipendra/AdolescentsMentalHealthPrediction.