A Machine Learning-Based Case–Control Study on Suicide Risk Identification: Integrating Acoustic and Linguistic Features Under Stress Conditions

Abstract

Suicide is a significant global public health issue, with current risk assessment methods primarily relying on psychiatrists’ clinical judgment and scale-based evaluations, which can be challenging to implement. Recently, interest has increased in using vocal and linguistic features to identify suicide risk. This study investigates speech-based methods for assessing suicide risk in two phases involving 90 patients with major depressive disorder (MDD) or bipolar disorder (BD). In Phase 1, three types of question-answer materials with different emotional valences (positive, neutral, and negative) were employed. The model combining acoustic and word frequency features from negative emotional valence materials achieved the highest accuracy at 77.82%. Phase 2 introduced stress factors, highlighting that speech data collected under stress better reflects participants’ psychological states, providing more insights into suicide risk. These findings emphasize the potential of speech analysis in suicide prevention, while also calling for further research to validate and expand these results.