Graph learning based suicidal ideation detection via tree-drawing test.

Area of Science:

• Psychology
• Computer Science
• Public Health

Background:

• Adolescent suicide is a critical global public health issue.
• Traditional detection methods for suicidal ideation lack accuracy and scalability.
• Artificial intelligence offers potential for accurate, scalable detection but often has strict data requirements.

Purpose of the Study:

• To introduce the Tree-Drawing Test (TDT) as an effective tool for detecting suicidal ideation.
• To propose a novel graph learning approach for automatic TDT application.
• To balance accuracy and scalability in suicidal ideation detection.

Main Methods:

• Constructing a semantic graph from psychological features automatically annotated from tree-drawing images.
• Utilizing a Graph Convolutional Network (GCN) model for individual suicidal ideation detection.
• Evaluating the method on a dataset of 806 students using metrics like macro-F1 and G-mean.

Main Results:

• The proposed graph learning method significantly outperforms traditional machine learning and convolutional neural network approaches.
• An ablation study confirmed the effectiveness of specific features (e.g., "leaves and fruits") in detection.
• The method demonstrated stability and robustness even with incomplete image data.

Conclusions:

• The novel approach effectively screens for suicidal ideation on a large scale.
• The method achieves high detection performance, model stability, and adaptability.
• This AI-driven TDT offers a flexible and scalable solution for early detection of suicidal ideation in adolescents.