Computational methods for the identification of suicidal ideation: a systematic review

Area of Science:

• Computational linguistics
• Public health
• Artificial intelligence

Background:

• Suicide is a major public health concern, particularly among young people, with increased incidence during the COVID-19 pandemic.
• Natural Language Processing (NLP) offers tools to analyze text for predicting suicidal ideation.
• Early detection and intervention are crucial for suicide prevention strategies.

Purpose of the Study:

• To systematically review computational techniques for identifying suicidal ideation in natural language texts.
• To extract and synthesize findings on NLP methods applied to suicide risk detection.
• To identify current trends and limitations in computational approaches to suicidal ideation detection.

Main Methods:

• A systematic literature review following the PRISMA 2020 methodology.
• Searches conducted across major academic databases (Scopus, IEEE Xplore, ACM, Springer, Web of Science).
• Qualitative analysis using narrative synthesis; risk of bias assessed with AMSTAR 2.

Main Results:

• Identified 25 studies on computational methods for detecting suicidal ideation.
• Transformer-based models (e.g., BERT) and hybrid approaches (BERT with CNN/LSTM) are predominant.
• Techniques like TF-IDF and pre-trained embeddings are used to enhance performance.

Conclusions:

• Computational techniques show significant potential for early detection and prevention of suicidal ideation.
• Limitations include a lack of linguistic/cultural diversity and over-reliance on social media data.
• Further research should focus on diverse datasets and improving model interpretability for non-experts.