Using smartphone surveys to predict next-week suicide attempts.

Area of Science:

• Psychiatry and Mental Health
• Digital Health
• Machine Learning in Healthcare

Background:

• Accurate prediction of suicidal behavior remains a significant clinical challenge.
• Existing methods lack the precision to forecast near-term suicide attempts (SA).
• Novel approaches are needed to identify individuals at immediate risk.

Purpose of the Study:

• To evaluate the efficacy of smartphone-based surveys and metadata in predicting near-term suicidal behavior.
• To assess the accuracy of machine learning models in forecasting suicide attempts (SA) and suicide-related events (SRE) within a 7-day window.
• To identify key predictors of imminent suicidal behavior from passively collected digital data.

Main Methods:

• Utilized brief, 20-item smartphone surveys sent 6 times daily to 619 high-risk patients over 3 months.
• Collected survey responses (N = 79,448) and passive metadata (e.g., time since last submission).
• Employed machine learning models, including bidirectional long short-term memory and lasso-penalized logistic regression, to predict next-week SA and SRE.

Main Results:

• The best model (bidirectional LSTM for SRE) achieved an AUC of 0.94 (sensitivity 0.87, specificity 0.90, PPV 0.30).
• Prediction for SA yielded an AUC of 0.90 (sensitivity 0.74, PPV 0.16).
• Prediction accuracy was higher for SREs than SAs, and improved with more data sources, adult participants, and personalized training data.

Conclusions:

• Brief smartphone surveys combined with metadata can predict next-week suicide attempts and related events with notable accuracy.
• Within-study history of SREs and self-reported agitation were strong predictors of SA.
• Future research should focus on enhancing accuracy and developing just-in-time interventions for high-risk periods.