Triaging Casual From Critical-Leveraging Machine Learning to Detect Self-Harm and Suicide Risks for Youth on Social...

Area of Science:

• Computational linguistics
• Digital mental health
• Youth psychology

Background:

• Youth (13-21 years) private Instagram conversations are analyzed for self-harm or suicide (SH-S) ideation.
• Existing automated mental health tools need improvement in identifying nuanced youth language related to SH-S.

Purpose of the Study:

• Develop interpretable machine learning models to detect the spectrum of SH-S expressions in youth conversations.
• Move beyond simple binary classification to understand varied SH-S language.

Main Methods:

• Analyzed youth-donated Instagram private conversations using traditional and transformer-based machine learning models.
• Incorporated features: psycholinguistic, sentiment, lexical, and conversational context (message to subconversation level).
• Evaluated models including Bidirectional Encoder Representations from Transformers and Distilled Bidirectional Encoder Representations from Transformers.

Main Results:

• Distilled Bidirectional Encoder Representations from Transformers achieved 99% accuracy for SH-S presence in individual messages.
• Fine-grained classification (self, other, hyperbole) accuracy was 89%, improving to 91% with expanded conversational context.
• Contextual understanding is crucial for distinguishing subtle SH-S discourse variations.

Conclusions:

• SH-S automatic detection systems must be sensitive to dynamic youth language on social media.
• Contextual and sentiment-aware models enhance detection and nuanced understanding of SH-S risk.
• Research provides a foundation for ethical interventions, requiring cross-platform and cross-population validation.