NUST Institutions Library Catalogue

Social media has emerged as a tool for exploring the human psyche, offering exceptional
access to real-time behavioral signals that are transforming the landscape of computational
psychiatry. Among these, bipolar disorder is difficult to detect because of its volatility,
requiring robust modeling of emotion, behavior, and timing. Prior studies have largely
focused on text-based sentiment analysis and linguistic features to classify mental health
conditions; however, these approaches often neglect non-verbal markers such as circadian
rhythms and affective variability. Existing models primarily depend on static textual cues,
limiting their ability to capture the dynamic, multimodal nature of psychiatric expression.
This study addresses these limitations by integrating temporal rhythms, emotional
dynamics, and behavioral signals extracted from Reddit user histories to develop predictive
models of bipolar disorder and high-risk psychological states. We propose a series of
interpretable multimodal architectures employing classical machine learning (Logistic
regression, Random Forest, and XGBoost), deep sequence models (LSTM, GRU), and
transformer-based frameworks (Roberta, GPT). Our approach incorporates temporal
posting features, emotional entropy, and community-level interaction structures.
Compared to benchmark studies, our models demonstrate significant improvements in both
classification (F1 > 0.99) and regression (R² > 0.89), highlighting the predictive power of
fused behavioral signals. This work advances the field by providing a scalable, languageindependent framework for the early detection of psychiatric risk. It also holds broader
implications for public health by offering a foundation for real-time, ethically deployable
digital mental health tools.