A Machine Learning Approach to Identifying Changes in Suicidal Language.

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning in healthcare

Background:

• Early identification and intervention are crucial for preventing suicidal deaths.
• Machine learning (ML) models have shown promise in detecting suicidal language.
• The persistence of suicidal language after clinical discharge requires further investigation.

Purpose of the Study:

• To examine the stability of suicidal language in individuals up to 30 days after discharge from care.
• To assess if machine learning algorithms can detect persistent suicidal language over time.
• To compare the stability of language-based markers with traditional assessment measures.

Main Methods:

• A multi-center study enrolled 253 subjects into suicidal or control cohorts.
• Machine learning algorithms analyzed language from standardized instruments and interviews.
• Subjects were re-interviewed around 30 days post-discharge, with language compared to initial assessments.

Main Results:

• Language characteristics indicative of suicidality at initial encounter remained detectable 30 days later (AUC = 89%).
• ML algorithms trained on later interviews could identify subjects from earlier interviews (AUC = 85%).
• Patient language demonstrated stability over 30 days, contrasting with changes in standard measure responses.

Conclusions:

• Suicidal language exhibits significant persistence for at least 30 days post-discharge.
• Advanced computational methods reveal stable language-based phenotypes in patients.
• This stability in language provides a potential data-driven approach for identifying individuals at risk.