Affective and Cognitive Distortions-Aided Suicide Risk Prediction for Long-Form Speech in Psychological Support Hotlines.

Speech-based suicide risk prediction is vital for psychological support hotlines but remains challenging because existing methods often insufficiently incorporate clinically relevant prior cues and have difficulty identifying sparse high-risk signals in long-form speech. We propose the Affective & Cognitive Distortions-assisted Speech Suicide Risk Prediction Network (ACD-SSRNet) to address these challenges. First, we construct a multi-view feature system that integrates general acoustic-textual features with affective and cognitive-distortion cues motivated by clinical knowledge. Second, a hierarchical cascaded decoupling module is developed to reduce heterogeneous feature redundancy while preserving task-critical information. Finally, we design a prior-guided multi-path graph attention structure to locate sparse high-risk segments and capture long-range temporal dependencies. Experiments on a real-world hotline dataset show that ACD-SSRNet outperforms state-of-the-art baselines, achieving a 2.79% improvement in F1-score and a 2.57% improvement in accuracy. We further conducted an expert evaluation on five representative de-identified hotline cases, showing that the model can capture key affective and cognitive-distortion segments associated with suicide risk.