Prediction of suicide using web based voice recordings analyzed by artificial intelligence.

Abstract

The integration of machine learning (ML) and deep learning models in suicide risk assessment has advanced significantly in recent years. In this study, we utilized ML in a case-control design, we predicted completed suicides using publicly available, web-based, real-world voice data, and treating speech as a biomarker. Our model demonstrated high accuracy in distinguishing between individuals who died by suicide and carefully matched controls achieving an area under the curve (AUC) of 0.74. This improved to an AUC of 0.85 and an accuracy of 76% when analyzing the subset of individuals who died by suicide within 12 months of the audio recording. The best predictive performance was observed with the Multilayer perceptron model, particularly when using the all Bene, Q + U Bene, and Q + U Raw feature sets-highlighting the importance of combining structured and unstructured paralinguistic features. The findings highlight the critical temporal proximity of voice biomarkers to suicide risk. The model's robustness is further evidenced by its resilience to perturbations in the analytical pipeline. This is the first study to successfully predict actual suicidal behavior rather than surrogate markers, marking a major step forward in suicide prevention. By demonstrating that speech can serve as a non-invasive and objective biomarker for suicide risk, this research opens new avenues for diagnostic and prognostic applications.