Acoustic and machine learning methods for speech-based suicide risk assessment: A systematic review.
Level 1 - systematic review of randomized trials
Systematic review of observational and diagnostic machine learning studies.
PubMed 41203082 · doi:10.1016/j.jad.2025.120569
What was done
This PRISMA-compliant systematic review evaluated artificial intelligence and machine learning methods using acoustic speech analysis to assess suicide risk. Investigators searched PubMed, Cochrane, Scopus, and Web of Science through February 2025 for studies comparing acoustic features between individuals at risk of suicide and those not at risk. Risk of bias was assessed using the PROBAST tool, and findings across 33 included articles were synthesized narratively.
What was found
Across 33 included studies (29 reporting classifier models), significant acoustic differences between risk and non-risk cohorts were identified in fundamental frequency (F0), jitter, Mel-frequency cepstral coefficients (MFCC), and power spectral density (PSD). Reported classifier performance spanned AUCs from 0.62 to 0.985 and accuracies from 60% to 99.85%. Multimodal models integrating acoustics, linguistic features, and metadata performed better than unimodal acoustic models.
Why it matters
This review synthesizes the current landscape of acoustic biomarkers for suicide risk assessment, demonstrating technical feasibility while identifying pervasive methodological gaps that currently prevent clinical translation.
Limits
The review was not prospectively registered. Primary studies were constrained by small sample sizes, severe dataset class imbalances favoring non-risk controls, heterogeneous speech elicitation protocols, and rare reporting of disaggregated group-level performance metrics.
Cited by
- supports Machine learning algorithms analyzing speech patterns can predict suicidality.