HuMamba: Adolescent Suicide Risk Detection with HuBERT and Mamba-Based Temporal Modeling

Early identification of adolescent suicide risk is crucial for timely mental health interventions. This study introduces a speech-based detection framework that integrates the HuBERT pre-trained model with Mamba, a temporal state-space module, to effectively capture nuanced emotional patterns and temporal dynamics in adolescent speech. Comparative evaluations reveal the superior sensitivity of HuBERT over Wav2Vec2 in extracting fine-grained emotional features. Additionally, Mamba outperforms conventional long short-term memory (LSTM) architectures in sequential modeling. The unified HuBERT-Mamba framework achieves an average accuracy of $\mathbf{6 3. 7 5 \%}$ and an F1Score of $\mathbf{6 2. 8 4 \%}$ across multi-task evaluations, demonstrating robust performance and generalization capability. This research provides a practical and effective approach for leveraging speech signals in suicide risk assessment, offering methodological support for developing intelligent mental health screening systems.