ECHO-NET: AN EXPLAINABLE HYBRID CONTEXTUAL OPTIMIZATION NETWORK FOR GENERALIZED SUICIDE RISK DETECTION FROM SOCIAL MEDIA USING BERTWEET-BIGRU ATTENTION LEARNING

Abstract

The rapid proliferation of social media platforms has created an emerging need for intelligent systems that can detect suicidal ideation from user generated content at scale. Existing machine learning and deep learning approaches often suffer from limited contextual understanding, limited cross-domain generalizability, high false negative rates and limited interpretableness which are all critical concern in clinical mental health settings. This paper introduces ECHO-Net (Explainable Hybrid Contextual Optimization Network) as a novel architecture that unifies tweet based transformer embeddings using BERT, Bidirectional Gated Recurrent Units BiGRU), Dilated Temporal Bi-directional Temporal Convolutional Networks (Bi-TCN), an Emotion-Aware Psychological Attention Module and Contrastive Semantic Alignment to provide generalized suicide risk detection. A curated dataset comprising of approximately 57,306 instances of social media data were drawn from Twitter and suicidal text repositories was used to evaluate the performance of the proposed model based on transparent feature attributions identifying dominant suicidal semantic patterns. The ECHO-Net framework reduces false negative predictions by a large amount and outperforms all evaluated transformer and recurrent baselines offering a scalable and clinically accountable solution for early suicide risk assessment in digital mental health ecosystems.