Acoustic and machine learning methods for speech-based suicide risk assessment: A systematic review

Area of Science:

• * Computational psychiatry and clinical informatics.
• * Speech signal processing and machine learning applications in mental health.

Background:

• * Suicide is a critical public health issue requiring enhanced detection and intervention strategies.
• * Acoustic analysis of speech presents a potential avenue for objective suicide risk assessment.

Purpose of the Study:

• * To systematically review the existing literature on Artificial Intelligence (AI) and Machine Learning (ML) for suicide risk assessment using speech acoustics.
• * To evaluate the effectiveness of various acoustic features and machine learning models in differentiating individuals at risk of suicide (RS) from those not at risk (NRS).

Main Methods:

• * Systematic review adhering to PRISMA guidelines, analyzing 33 articles from major scientific databases (PubMed, Cochrane, Scopus, Web of Science).
• * Inclusion criteria focused on studies analyzing acoustic features in relation to suicide risk, excluding those lacking acoustic data or methodological rigor.
• * Risk of bias was assessed using the PROBAST tool.

Main Results:

• * Consistent identification of significant acoustic variations between RS and NRS groups, notably in jitter, fundamental frequency (F0), Mel-frequency cepstral coefficients (MFCC), and power spectral density (PSD).
• * Machine learning classifiers demonstrated variable performance (AUC 0.62-0.985, accuracy 60%-99.85%), with multimodal approaches showing superior results.
• * Prevalence of imbalanced datasets favoring NRS and limited reporting of group-specific performance metrics were noted.

Conclusions:

• * AI and ML hold potential for suicide risk detection via acoustic speech analysis, highlighting specific acoustic markers.
• * Methodological inconsistencies, sample size limitations, and class imbalance necessitate further research for improved clinical validity and generalizability.
• * Future studies should focus on addressing these limitations to enhance the reliability and applicability of AI-driven suicide risk assessment tools.