Analyzing Suicide Risk From Linguistic Features in Social Media: Evaluation Study

Area of Science:

• Computational linguistics
• Machine learning
• Mental health informatics

Background:

• Effective suicide risk assessment is crucial for prevention.
• Individuals at risk may not seek professional help.
• Machine learning (ML) and computational linguistics offer tools for risk analysis.

Purpose of the Study:

• To explore the use of computerized language analysis for assessing suicide risk on social media.
• To understand suicide risk through statistical and ML-based text analysis.

Main Methods:

• Analysis of the University of Maryland Suicidality Dataset (N=866 Reddit users).
• Linguistic Inquiry and Word Count lexicon for sentiment, thinking styles, and part of speech; TextStat for readability.
• Statistical tests (Mann-Whitney U, Kruskal-Wallis) and ML models (gradient boost, random forest, SVM) with 10-fold cross-validation.

Main Results:

• Statistically significant linguistic differences between at-risk and no-risk users (P<.05).
• At-risk users showed higher authenticity, first-person pronouns, and negation; lower clout.
• Random forest and gradient boost models achieved higher F1-scores (0.65, 0.62) than SVM (0.52).

Conclusions:

• Linguistic features like social posturing (authenticity, clout), pronoun use, and negation are associated with suicide risk.
• This research enhances understanding of at-risk users' thought patterns and ML model mechanisms.
• Demonstrated ML's potential to assist healthcare professionals in suicide risk assessment.