Integrating speech biomarkers and large language models for adolescent suicide risk detection with mobile application for real-world evaluation

Adolescent suicide is a significant public health issue, highlighting the need for efficient methods to detect suicide risk. Here, we develop and validate a speech-based suicide risk detection framework grounded in large language models (LLMs). Two independent cohorts of adolescents aged 10-18 years are analyzed: a development cohort (n = 1,223), with voice recordings collected in structured interview settings for model training and internal evaluation, and an external validation cohort (n = 460), collected through a mobile application to assess feasibility in naturalistic settings. An integrated model combining a speech encoder and an LLMs-based text-processing branch achieves its best performance on the self-introduction task. The model yields an accuracy of 0.808 and a macro-F1 score of 0.807 for suicide risk detection and remains effective under naturalistic mobile assessment. These findings support integrating LLMs with speech-derived markers for scalable adolescent suicide risk detection.