Ensemble Deep Learning for High-Accuracy Prediction of Mental Health States from Social-Media Behavioural Indicators

Abstract

Concerns about the effects of social-media usage on psychological well-being have spurred exponential growth in research?on how social media may relate to mental health and the development of automated mental health monitoring as an immediate research imperative. In?this study, we develop a one-class ensemble deep learning framework called MIND (Mental state Identification through Neural Detection) for healthy, At_Risk, and Stressed mental health prediction based on behaviour and interaction-based attributes obtained from 5,000 social-media users. We conducted extensive feature engineering to capture latent psycho-behavioral characteristics to?result in 23-dimensional input feature space. Stratified 5-fold cross-validation?was applied to train three heterogeneous neural architectures—Deep Feed-Forward, Wide-and-Deep, and Residual Fully Connected networks—which were subsequently fused through hard voting and probability averaging. The above mentioned averaging ensemble?outperformed all, achieving an overall accuracy of 99.46%, precision 0.9947, recall 0.9946, F1-score 0.9947 and strong positivity on reliability measures, Cohen’s kappa 0.9636 and MCC 0.9636. When evaluated class-wise, Stressed users were detected perfectly (F1 = 1.0), while Healthy users?were also well discriminated and At-Risk individuals achieved competitive performance. We also performed confidence and ensemble agreement analysis revealing decision?stability, with models agreeing in 99.28% of test samples. The results show?that ensemble deep learning is successful in detecting low signal-to-noise ratio behavioural risk and establishes the proposed system as a promising, highly accurate mental health prediction informatics solution for preventive digital health monitoring.