A review of natural language processing in the identification of suicidal behavior

Background: Natural language processing (NLP) approaches offer powerful, flexible methods of identifying various psychological conditions from free-form verbal expression in both written and spoken form. These tools may be particularly useful for suicide risk detection and potentially easing human suffering on a global scale. The degree of their accuracy in discerning suicidal thoughts or behaviors, however, is currently unknown. Methods: English-language studies of NLP classification of suicidal ideation and/or behavior were reviewed. Searches were conducted using PsycInfo, PubMed, the Cochrane Library, and Google Scholar, which yielded 9,711 candidate articles. Of these, 31 were ultimately selected for inclusion and examined for identification accuracy. Results: Studies were diverse in terms of performance statistics reported, with inconsistencies across fields and publications. Regardless, all metrics suggested strong performance of NLP methods tested, which were heavily slanted toward accurate identification of positive cases (i.e., those at elevated risk for suicidal thoughts or behaviors). Limitations: Reports were non-standardized across studies, necessitating non-optimal combination of performance statistics through non-weight averages. Small sample sizes and biased test construction samples were also notable in some studies, and generalizability of study results was variable. Conclusions: Although formative, existing methods of NLP for suicide detection far surpass the accuracy of human raters. Use of these strategies in applied environments may enhance early identification and prevention, and future work to understand their dissemination and implementation is recommended.