Detecting Suicidal Ideation in Adolescence Using Self-Reported Emotional and Behavioral Patterns: Comparing Machine Learning and Large Language Model Predictions

Abstract

Suicidal ideation in adolescents is a critical public health issue requiring early detection. This study examined whether machine learning (ML) and large language models (LLMs) can detect ideation in 1,197 students (ages 10–15) using self-reported Strengths and Difficulties Questionnaire (SDQ) data. Clinically relevant ideation was defined using Suicidal Ideation Questionnaire—Junior (SIQ-JR) cut-offs. Gemini 1.5 Pro and GPT-4o were prompted to estimate SIQ-JR scores from SDQ responses and demographics; Logistic Regression, Naive Bayes, and Random Forest models were trained on either SDQ data or LLM predictions. LLM predictions correlated with SIQ-JR (ρ = .61) and showed good discrimination across thresholds (area under the curve (AUC) ≥ .83), with item-level associations paralleling self-reports, revealing strong associations with emotional symptoms and peer problems. In cross-validated analyses, the best SDQ-based ML model reached sensitivity = .85 and specificity = .72; the best LLM-based model achieved .80 and .74. Notably, ML models trained directly on SDQ responses consistently outperformed those incorporating LLM predictions across all SIQ-JR thresholds. Nonetheless, LLMs demonstrated promising accuracy in identifying suicidal ideation based on SDQ and demographic data. Further refinement and validation are required before these approaches can be considered viable for clinical implementation.