Deep neural networks detect suicide risk from textual facebook posts

Area of Science:

• Computational psychiatry
• Artificial intelligence in mental health
• Social media analytics

Background:

• Detecting suicide risk is crucial yet challenging, with traditional methods yielding limited accuracy (AUCs 0.56-0.58).
• Existing research has not significantly improved prediction beyond chance levels over decades.

Purpose of the Study:

• To develop and evaluate Artificial Neural Network (ANN) models for predicting suicide risk using social media language.
• To compare the predictive performance of a Single Task Model (STM) versus a Multi-Task Model (MTM) incorporating hierarchical risk factors.

Main Methods:

• Constructed ANN models using 83,292 Facebook posts from 1002 authenticated users with psychosocial data.
• Employed Deep Contextualized Word Embeddings for text representation.
• Developed an STM (texts → suicide) and an MTM (texts → traits → risks → disorders → suicide).

Main Results:

• The MTM achieved significantly higher prediction accuracy (AUCs 0.697–0.746) compared to the STM (AUCs 0.621–0.629).
• Effect sizes for the MTM were substantially larger (d = 0.729–0.936).
• Content analysis revealed predictions were based on diverse text features, not explicit suicide mentions.

Conclusions:

• Machine learning analysis of social media text offers a promising avenue for enhancing suicide risk detection.
• The MTM approach, integrating multiple layers of risk factors, demonstrates superior predictive power.
• Findings support the development of practical tools for identifying individuals at risk through social media activity.