Abstract
BACKGROUND: Crisis hotlines serve as a crucial avenue for the early identification of suicide risk, which is of paramount importance for suicide prevention and intervention. However, assessing the risk of callers in the crisis hotline context is constrained by factors such as lack of nonverbal communication cues, anonymity, time limits, and single-occasion intervention. Therefore, it is necessary to develop approaches, including acoustic features, for identifying the suicide risk among hotline callers early and quickly. Given the complicated features of sound, adopting artificial intelligence models to analyze callers' acoustic features is promising. OBJECTIVE: In this study, we investigated the feasibility of using acoustic features to predict suicide risk in crisis hotline callers. We also adopted a machine learning approach to analyze the complex acoustic features of hotline callers, with the aim of developing suicide risk prediction models. METHODS: We collected 525 suicide-related calls from the records of a psychological assistance hotline in a province in northwest China. Callers were categorized as low or high risk based on suicidal ideation, suicidal plans, and history of suicide attempts, with risk assessments verified by a team of 18 clinical psychology raters. A total of 164 clearly categorized risk recordings were analyzed, including 102 low-risk and 62 high-risk calls. We extracted 273 audio segments, each exceeding 2 seconds in duration, which were labeled by raters as containing suicide-related expressions for subsequent model training and evaluation. Basic acoustic features (eg, Mel Frequency Cepstral Coefficients, formant frequencies, jitter, shimmer) and high-level statistical function (HSF) features (using OpenSMILE [Open-Source Speech and Music Interpretation by Large-Space Extraction] with the ComParE 2016 configuration) were extracted. Four supervised machine learning algorithms (logistic regression, support vector machine, random forest, and extreme gradient boosting) were trained and evaluated using grouped 5-fold cross-validation and a test set, with performance metrics, including accuracy, F(1)-score, recall, and false negative rate. RESULTS: The development of machine learning models utilizing HSF acoustic features has been demonstrated to enhance recognition performance compared to models based solely on basic acoustic features. The random forest classifier, developed with HSFs, achieved the best performance in detecting the suicide risk among the models evaluated (accuracy=0.75, F(1)-score=0.70, recall=0.76, false negative rate=0.24). CONCLUSIONS: The results of our study demonstrate the potential of developing artificial intelligence-based early warning systems using acoustic features for identifying the suicide risk among crisis hotline callers. Our work also has implications for employing acoustic features to identify suicide risk in salient voice contexts.