Early Mental Health Detection in Adults using Temporal and Linguistic Analysis of Social Media Data

Conventional screening methods for mental-health conditions, like depression, are based on self-report and clinical evaluation and scale poorly, making early detection essential for prompt intervention. This paper suggests a hybrid approach that combines temporal behavioral patterns and linguistic representations to detect depression in adults at an early stage, based on Reddit data. The temporal cues are represented by a two-layer LSTM (long short-term memory network) whose output is fused with linguistic features extracted by fine-tuned Roberta transformer (transformer with Roberta pre-training) by a cross-modal attention fusion layer, which is then fed to a classification layer. We evaluate the model on the publicly available Depression: Reddit Cleaned dataset (7,732 posts; 53% non-depression, 47% depression), partitioned via stratified sampling into 70% training (5,412 posts), 15% validation (1,160 posts) and 15% testing (1,160 posts). The proposed fusion model attains 92.7% accuracy, 91.8% precision, 93.4% recall, 92.6% macro-F1 and 96.8% AUC, outperforming the strongest text-only Roberta baseline by +3.3% accuracy, +3.5% F1 and +2.2% AUC, and surpassing the behavior-only LSTM by +7.7% accuracy and +7.9% F1. An ablation confirms the contribution of each component: removing the temporal branch drops F1 by 3.5 points, removing the linguistic branch drops F1 by 7.9 points, replacing cross-modal attention with simple concatenation drops F1 by 2.0 points, and freezing the Roberta encoder drops F1 by 4.1 points. Improvements over baselines are statistically significant (paired t-test, p