A machine learning approach predicts future risk to suicidal ideation from social media data

Area of Science:

• Computational psychiatry
• Digital phenotyping
• Artificial intelligence in mental health

Background:

• Social media data offers insights into longitudinal environmental influences on suicidal thoughts and behaviors.
• Existing methods for suicide risk assessment have limitations in capturing dynamic, real-time risk factors.

Purpose of the Study:

• To develop and validate an algorithm, Suicide Artificial Intelligence Prediction Heuristic (SAIPH), for predicting future suicidal ideation (SI) risk using Twitter data.
• To assess the algorithm's performance in identifying individuals at elevated risk and its correlation with real-world suicide rates.

Main Methods:

• Trained neural networks on Twitter data associated with psychological constructs (e.g., depression, anxiety, stress).
• Utilized a random forest model incorporating neural network outputs to predict SI status.
• Analyzed 512,526 tweets from 283 SI cases and 3,518,494 tweets from 2,655 controls for model training and validation.

Main Results:

• The SAIPH model achieved an AUC of 0.88 (95% CI 0.86-0.90) in predicting SI status.
• Identified a ~7-fold increased risk for SI within 10 days when algorithm scores exceeded an individual-specific threshold (OR=6.7).
• Validated the model with regional Twitter data, showing significant associations between algorithm SI scores and county-wide suicide death rates, particularly in younger individuals.

Conclusions:

• Machine learning analysis of social media, exemplified by SAIPH, shows promise for identifying individuals at future risk of suicidal ideation.
• The SAIPH algorithm can be adapted as a clinical decision tool for suicide screening and risk monitoring.
• Digital phenotyping via social media analysis represents a novel approach to proactive mental health surveillance and intervention.