Integrating speech biomarkers and large language models for adolescent suicide risk detection with mobile application for real-world evaluation.
Level 3 - non-randomized controlled study
Multi-cohort diagnostic prediction model development with independent external validation cohort
PubMed 42155450 · doi:10.1016/j.xcrm.2026.102823
What was done
Researchers developed and validated a multimodal suicide risk detection framework combining acoustic speech biomarkers with large language model (LLM) text processing for adolescents aged 10–18 years. The framework was trained and internally evaluated on a development cohort of 1,223 adolescents using voice recordings from structured interviews, and then externally tested on 460 adolescents who provided audio via a mobile smartphone application in naturalistic settings.
What was found
The integrated model combining a speech encoder and an LLM-based text processing branch achieved its best performance on a self-introduction task. The model demonstrated an accuracy of 0.808 and a macro-F1 score of 0.807 for suicide risk detection, maintaining effectiveness in the naturalistic mobile cohort. Specific sensitivity, specificity, and positive predictive values were not provided in the abstract.
Why it matters
Combining acoustic speech features and language models in a mobile application offers a potentially scalable, non-invasive method to assist in early detection of adolescent suicide risk in real-world environments.
Limits
The abstract omits separate performance breakdowns between the internal and external validation cohorts, confidence intervals, and standard diagnostic performance metrics (sensitivity, specificity, positive/negative predictive values). Clinical characteristics, ground-truth suicide risk definitions, and the demographic diversity of the participants were not detailed.
Cited by
- supports Machine learning algorithms analyzing speech patterns can predict suicidality.