Detecting Suicidal Ideation in Adolescence Using Self-Reported Emotional and Behavioral Patterns: Comparing Machine...

Area of Science:

• Adolescent mental health
• Computational psychiatry
• Public health surveillance

Background:

• Suicidal ideation in adolescents is a significant public health concern.
• Early detection is crucial for intervention and prevention efforts.
• Existing screening tools require validation with novel computational approaches.

Purpose of the Study:

• To evaluate the efficacy of machine learning (ML) and large language models (LLMs) in detecting adolescent suicidal ideation.
• To compare the performance of ML models trained on questionnaire data versus LLM-generated predictions.
• To identify key indicators associated with suicidal ideation in adolescents.

Main Methods:

• Utilized self-reported Strengths and Difficulties Questionnaire (SDQ) data from 1,197 students (ages 10-15).
• Employed LLMs (Gemini 1.5 Pro, GPT-4o) to predict Suicidal Ideation Questionnaire-Junior (SIQ-JR) scores from SDQ and demographic data.
• Trained and cross-validated ML models (Logistic Regression, Naive Bayes, Random Forest) using SDQ data and/or LLM predictions.

Main Results:

• LLM predictions showed moderate correlation (ρ = .61) and good discrimination (AUC ≥ .83) with SIQ-JR scores.
• ML models trained directly on SDQ data consistently outperformed LLM-based models in detecting suicidal ideation.
• The best SDQ-based ML model achieved 85% sensitivity and 72% specificity; the best LLM-based model achieved 80% sensitivity and 74% specificity.

Conclusions:

• While LLMs show potential for identifying suicidal ideation from SDQ and demographic data, direct ML analysis of SDQ responses is currently more effective.
• Emotional symptoms and peer problems were strongly associated with suicidal ideation.
• Further research and validation are necessary for clinical implementation of these AI-driven detection methods.