Neuro-semantic graph fusion for explainable depression risk trajectory mapping using a graph-enhanced RoBERTa framework.

Introduction: Early identification of depression through social media analytics is frequently compromised by linguistic ambiguities, contextual complexities, and the opaque nature of conventional deep learning architectures. This study proposes Graph-RoBERTa-CL, a hybrid neuro-semantic framework that explicitly integrates three synergistic components: (i) RoBERTa-based Transformer semantic encoding for deep contextual representation, (ii) Graph Attention Networks (GAT) for relational contextual aggregation across semantically related posts, and (iii) Supervised Contrastive Learning (SCL) for latent space regularization under class imbalance, collectively enabling explainable depression risk trajectory mapping from social media corpora. Methods: The framework was trained and validated on the Suicide Detection Dataset ( Results: On the internal user-level test set, Graph-RoBERTa-CL achieved Accuracy 91.2%, Precision 89.7%, Recall 93.1%, F1-Score 0.914, and AUC 0.962, substantially outperforming BERT-base (F1: 0.798, AUC: 0.934) and LSTM-Attention (F1: 0.681, AUC: 0.856). On the external independent validation set, the model achieved F1 0.882 and AUC 0.941, with the modest cross-dataset reduction confirming genuine generalizability rather than dataset-specific overfitting. Ablation studies confirm that each architectural component - semantic encoding, graph attention aggregation, and contrastive optimization - contributes independently and synergistically to overall performance. Discussion: The proposed framework establishes a rigorous, interpretable, and ethically grounded mechanism for proactive mental health screening in digital ecosystems, designed to function as a human-in-the-loop clinical screening aid rather than a standalone diagnostic tool.