Speech based suicide risk recognition for crisis intervention hotlines using explainable multi-task learning

Abstract

Background

Crisis Intervention Hotline can effectively reduce suicide risk, but suffer from low connectivity rates and untimely crisis response. By integrating speech signals and deep learning to assist in crisis assessment, it is expected to enhanced the effectiveness of crisis intervention hotlines.

Methods

In this study, a crisis intervention hotline suicide risk speech dataset was constructed, and the speech was labeled based on the Modified Suicide Risk Scale. On the dataset, the variability of speech duration between different callers and different speech high-level features were explored across callers. Finally, this study proposed a data-theoretically dual-driven, gender-assisted speech crisis recognition method based on multi-tasking and deep learning, and the results of the model were obtained through five-fold cross-validation.

Results

Analysis of the dataset demonstrated gender differences in callers, with male callers speaking more in crisis calls compared to females. Feature analysis revealed significant differences between crisis callers in terms of emotional intensity of speech, speech rate and texture. The proposed method outperformed other methods with an F1 score of 96 % on the validation data, and feature visualization of the model also demonstrated the validity of the method.

Limitations

The sample size of this study was limited and ignored information from other modalities.

Conclusion

These findings demonstrated the effectiveness of the proposed model in speech crisis recognition, and the statistical data analysis enhanced the Interpretability of the model, while showing that the integration of data and theoretical knowledge facilitates the effectiveness of the method.