A TIME-AWARE TEMPORAL BERT FRAMEWORK FOR LONGITUDINAL DETECTION OF DEPRESSIVE AND SUICIDE-RELATED RISK PATTERNS IN SOCIAL MEDIA

Abstract

In this paper, we introduce a time-aware deep learning model designed to identify and predict signs of depression and suicidal ideation across social networks. Standard static text classifiers typically analyze updates in isolation; however, our method tracks the long-term progression of a person's emotional state by merging contextual language embeddings with temporal encodings and specific psycholinguistic markers. We gathered our primary dataset from Twitter, Reddit, and Facebook, ensuring all user histories were strictly anonymized and organized chronologically. The study evaluates multiple neural network architectures, specifically Temporal BERT (Bidirectional Encoder Representations from Transformers), time-encoded BiLSTM (Bidirectional Long Short-Term Memory), and a temporal transformer utilizing positional features. Our experiments demonstrate that factoring in the chronological dimension substantially boosts classification accuracy, allowing for the earlier detection of declining mental health. The Temporal BERT model achieved the highest F1 score (harmonic mean of precision and recall) and AUC (Area Under the Receiver Operating Characteristic Curve) values on several datasets, outperforming both standard (static) BERT and basic recurrent models. Analysis of temporal trajectories also allowed us to identify clear clusters of user behavior: stable, improving, and deteriorating - this makes conclusions more interpretable and helps us understand personal emotional dynamics. The early-warning module was evaluated at 7-, 14-, and 21-day prediction horizons and showed that risk-related deterioration patterns could be identified in advance of the reference event. Across all evaluated horizons, Temporal BERT demonstrated the strongest Recall@k performance, meaning that it more consistently captured at-risk users among the top-ranked predictions.

This article emphasizes that depressive and suicide-related risk signals are often not evident in isolated posts but emerge through longitudinal behavioral patterns. The proposed approach may support earlier and more sensitive identification of elevated risk patterns in digital mental health monitoring settings. At the same time, such use requires strict ethical safeguards, rigorous anonymization, and human-in-the-loop oversight. Future research should extend the framework toward multimodal, multilingual, and socially contextualized modeling.