Evaluating natural language processing derived linguistic features associated with current suicidal ideation, past...

Area of Science:

• Psychiatry and Mental Health
• Computational Linguistics
• Data Science in Healthcare

Background:

• Individuals with psychosis exhibit elevated suicide risk compared to the general population.
• Natural Language Processing (NLP) has been applied to psychosis research but not for predicting future suicidal behavior.
• This study investigates NLP-derived linguistic features for suicide risk assessment in psychosis.

Purpose of the Study:

• To determine if NLP-derived linguistic features from a dyadic task can predict suicidal ideation, past attempts, and future suicidal behavior in adults with psychotic disorders.
• To assess the predictive performance of NLP models for various suicide-related outcomes.
• To identify key linguistic features associated with suicide risk in this population.

Main Methods:

• 112 adults with psychotic disorders completed suicide severity scales and a dyadic role-play task.
• Linguistic features (lexical, diversity, sentiment) were extracted from task transcripts using NLP.
• Machine learning models (MLPRegressor) were trained to predict suicide outcomes, with SHAP for feature analysis.

Main Results:

• High prevalence of suicidal ideation (42.9%), past attempts (67.9%), and future suicidal behavior (13.3%) observed.
• NLP models demonstrated strong predictive performance for past attempts (F1=0.75) and current ideation (F1=0.74-0.79).
• Models for future suicidal behavior achieved the highest predictive accuracy (F1=0.86-0.93).

Conclusions:

• NLP-derived linguistic features from dyadic interactions show high predictive accuracy for future suicidal behavior in individuals with psychosis.
• These findings suggest NLP analysis of dyadic tasks can enhance the understanding of suicide risk.
• Further replication is needed to validate these promising results for clinical application.