Temporal Modeling of Social Media for Depression Forecasting: Deep Learning Approaches with Pretrained Embeddings

In the field of natural language processing, depression forecasting from social media has gained extensive attention, as platforms like X (formerly Twitter) offer real-time user-generated content that can reflect psychological states. Common approaches typically rely on static text analysis, which overlooks how users’ emotions change over time. To address this limitation, we propose a temporal modeling approach that applies deep learning models to capture both textual and temporal patterns in users’ tweet histories. Our experiments evaluated LSTM networks and Transformer architectures with pretrained embeddings on a dataset of over 3 million tweets. We demonstrate that incorporating temporal features significantly improved performance in depression forecasting. The best setting, which combines Llama 2 embeddings with personalized time-difference features, achieved 99.4% accuracy and 0.996 AUC. These results highlight the importance of modeling temporal dynamics for improving depression forecasting and suggest that personalized temporal signals provide capabilities beyond static content analysis.