AI and Psychology: Examining the Influence of Social Media on Mental Health [Leveraging Machine Learning to Analyze Digital Behavior and Psychological Well-being]

Abstract: This paper presents a machine learning pipeline for predicting binary mental health risk from social media usage patterns. Using a survey dataset of 474 valid responses covering four clinical domains, ADHD-related attentional difficulties, anxiety, self-esteem disruption, and depression, we introduce a custom non-linear Likert scoring system that converts raw ordinal responses into weighted composite domain scores. A binary outcome variable is constructed by calibrating a total score threshold against WHO prevalence benchmarks, producing a 35.2% at-risk prevalence. Eight scikit-learn classifiers and a Keras feedforward network are evaluated under identical stratified conditions. An MLP Neural Network is selected for deployment, achieving 97.9% test accuracy and ROC-AUC = 0.9985 with 0.97 recall on the at-risk class. Pearson correlation analysis confirms ADHD Score (r = 0.71) and Depression Score (r = 0.68) as the strongest predictors, while all demographic features fall below |r| = 0.21, indicating demographic-invariant risk distribution. The trained model is deployed as an interactive Streamlit application. We discuss implications for AI-driven mental health monitoring and the limitations of cross-sectional survey-based classification.