Suicide ideation detection based on documents dimensionality expansion.

Area of Science:

• Computational linguistics
• Artificial intelligence
• Mental health informatics

Background:

• Classifying informal mental disorder documents is difficult due to informal language, data noise, and cultural nuances.
• Traditional deep learning models struggle with informal text, missing long-range dependencies and literal interpretations.
• Non-standard inputs like emojis pose challenges for conventional natural language processing.

Purpose of the Study:

• To develop an accurate and secure method for classifying informal documents related to mental disorders.
• To overcome the limitations of conventional deep learning models in processing informal text.
• To enhance privacy and explainability in mental health document classification.

Main Methods:

• Expanded data dimensionality by transforming and fusing textual data and signs from 1D to 2D.
• Utilized pre-trained 2D Convolutional Neural Network (CNN) models (AlexNet, Restnet-50, VGG-16).
• Applied the approach to a social media dataset for suicide-related content classification.

Main Results:

• Achieved high classification accuracy exceeding 99% for suicide-related content.
• Demonstrated the effectiveness of 2D data representation for capturing complex patterns in informal text.
• Validated the approach on a real-world social media dataset.

Conclusions:

• The 2D data transformation method significantly improves the accuracy of classifying informal mental health-related documents.
• Pre-trained 2D CNN models offer an efficient solution without requiring new model design.
• The 2D visual representation enhances data privacy and aids in model explainability.