Towards Cross-Task Suicide Risk Detection via Speech LLM
Abstract
Suicide risk among adolescents remains a critical public health concern, and speech provides a non-invasive and scalable approach for its detection. Existing approaches, however, typically focus on one single speech assessment task at a time. This paper, for the first time, investigates cross-task approaches that unify diverse speech suicide risk assessment tasks within a single model. Specifically, we leverage a speech large language model as the backbone and incorporate a mixture of DoRA experts (MoDE) approach to capture complementary cues across diverse assessments dynamically. The proposed approach was tested on 1,223 participants across ten spontaneous speech tasks. Results demonstrate that MoDE not only achieves higher detection accuracy than both single-task specialised models and conventional joint-tuning approaches, but also provides better confidence calibration, which is especially important for medical detection tasks.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The main contribution of this paper is the development of a cross-task suicide risk detection model using speech LLMs, which integrates multiple assessment tasks into a single framework, enhancing detection accuracy and confidence calibration. This work represents a significant advancement in the application of machine learning for mental health, with the potential to improve early detection and intervention strategies.
Methodology
The paper introduces a novel approach for suicide risk detection using a speech large language model (LLM) combined with a mixture of DoRA experts (MoDE). This methodology is innovative as it unifies multiple speech assessment tasks into a single model, addressing the limitations of traditional single-task models. The use of weight-decomposed low-rank adaptation (DoRA) for parameter-efficient fine-tuning is particularly noteworthy, as it allows the model to dynamically capture both task-specific and cross-task information. The incorporation of load balancing and temperature scaling further enhances the model's performance and stability.
Results & evaluation
The experiments are well-structured, involving a dataset of 1,223 participants across ten spontaneous speech tasks. The results demonstrate that MoDE outperforms both single-task specialized models and conventional joint-tuning approaches, indicating the effectiveness of the proposed method. The evaluation metrics, including accuracy and confidence calibration measures, are appropriate for the medical context of the research. However, the paper could benefit from additional comparative analyses with more baseline models to strengthen the claims.
Reproducibility
The implementation details are sufficiently described, with clear information on the model architecture, training procedures, and hyperparameters. However, the absence of a publicly accessible code repository limits the reproducibility of the results. Providing a GitHub link or similar would enhance the ability for other researchers to replicate the findings.
Limitations
The study relies on the MINI-KID scale, a self-report tool that may not capture the full complexity of suicide risk. This limitation could affect the generalizability of the findings. Additionally, the study does not address potential biases in the dataset or the implications of using a single cultural context (Chinese adolescents) for the research.
Broader impact
The implications of this research are significant, as it addresses a critical public health issue by providing a scalable and non-invasive method for suicide risk detection. The potential for real-world applications in mental health assessment and intervention could lead to improved outcomes for adolescents at risk. However, ethical considerations regarding the use of AI in sensitive health contexts must be carefully managed. The main contribution of this paper is the development of a cross-task suicide risk detection model using speech LLMs, which integrates multiple assessment tasks into a single framework, enhancing detection accuracy and confidence calibration. This work represents a significant advancement in the application of machine learning for mental health, with the potential to improve early detection and intervention strategies.
Author of this paper? Send a correction.