ENHANCING SUICIDE IDEATION DETECTION WITH MULTIMODAL DEEP LEARNING: ETHICAL, EXPLAINABLE, AND PRIVACY-PRESERVING FRAMEWORK

Source abstract

Suicide prevention is a critical global challenge, and early detection of suicidal ideation can save lives. This study proposes a multimodal deep learning framework integrating text, image, and speech data to enhance detection accuracy and reliability. Using advanced models such as BERT, ResNet, and BiLSTM, combined with Explainable AI (SHAP, LIME) and Federated Learning, the system ensures interpretability, ethical compliance, and user privacy. Experiments on a real-world social media dataset of 42,000 posts (including 8,400 labeled as suicidal) show that the proposed model achieves 90.5% F1-score, outperforming strong text-only baselines by 15% and unimodal models by 8–12%. A pilot deployment in a mental health support platform validated the system’s practical utility, with 85.8% of AI-flagged posts confirmed by clinicians. This research offers a scalable, interpretable, and privacy-preserving AI solution for early suicide risk identification and intervention.