Spread of social media platforms digital footprints of users capturing emotional and behavioural footprints resulting in unprecedented opportunity for novel data-driven techniques for mental health assessment. Abstract: In this work we present a deep learning framework to predict mental state - either Healthy, Stressed, or At Risk - with behavioral metrics, including screen time, interaction patterns, sleep duration, and physical activity. A total of 18 features were prepared after non-informative attributes were removed, categorical encoding and feature scaling was performed, and then a Multi-Layer Perceptron model with dropout regularization was trained. Results cause a trained model to obtain very high predictive performance, namely 99.70% accuracy with test loss of 0.0055, and learning-curve patterns support stable generalization without overfitting. Class-wise AUC analysis (AUC=1.000) evidenced boundless precision, recall and F1-scores of the evaluation metrics for all the mental-state categories. Further calibration and rejection-curve experiments confirmed the well-calibrated probability estimates of the model and that accuracy could be improved even more by filtering low-confidence predictions. We used sub-group error analysis to show there were only slight performance differences across social-media platforms and genders, and Kolmogorov–Smirnov testing to show that there was no statistically significant data drift between training and test distributions. ConclusionThe proposed model shows high robustness and generalizability to a wide range of digital mentalhealth monitoring and early-risk detection scenarios to be offered at scale.

Keywords:  Deep Learning, Mental State Prediction, Behavioural Metrics Analysis, Digital Mental-Health Monitoring