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

Area of Science:

• Computational psychiatry
• Artificial intelligence in mental health
• Biomarkers for suicide risk

Background:

• Machine learning (ML) and deep learning models are increasingly used for suicide risk assessment.
• Speech analysis offers a potential non-invasive method for identifying psychological states.

Purpose of the Study:

• To predict completed suicides using publicly available, web-based voice data.
• To evaluate speech as a biomarker for suicide risk assessment.
• To identify the most effective ML models and features for suicide prediction.

Main Methods:

• A case-control study design was employed.
• Publicly available, web-based voice recordings were utilized.
• Machine learning models, including Multilayer Perceptron, were trained and validated.
• Paralinguistic features (structured and unstructured) were extracted and analyzed.

Main Results:

• The ML model achieved an Area Under the Curve (AUC) of 0.74 in distinguishing individuals who died by suicide from controls.
• Predictive accuracy improved to an AUC of 0.85 and 76% accuracy for suicides occurring within 12 months of audio recording.
• The Multilayer Perceptron model performed best, especially with combined structured and unstructured paralinguistic features.
• The model demonstrated robustness against analytical pipeline perturbations.

Conclusions:

• Speech analysis, using ML, can serve as an objective and non-invasive biomarker for suicide risk.
• Temporal proximity of voice biomarkers is critical for accurate prediction.
• This study represents a significant advancement in predicting actual suicidal behavior, paving the way for novel suicide prevention strategies.