Predicting acute suicidal ideation on Instagram using ensemble machine learning models.

Area of Science:

• Computational psychiatry
• Digital phenotyping
• Machine learning in mental health

Background:

• Social networking (SN) data offers rich temporal and contextual insights for predicting suicidal thought and behavior.
• Current predictive modeling for acute suicidal ideation (SI) using SN data is underdeveloped.
• Machine learning algorithms combined with SN data present a promising avenue for SI prediction.

Purpose of the Study:

• To develop and evaluate an ensemble machine learning model for predicting acute suicidal ideation (SI) in adolescents using social networking data.
• To assess the predictive performance of language use and online activity predictors.
• To identify key predictors of SI through model introspection.

Main Methods:

• Applied an ensemble machine learning model to a dataset of 52 adolescents with a history of SI on Instagram.
• Utilized predictors capturing language use and activity within the social network.
• Employed out-of-sample, cross-validation and model explainer for performance evaluation and predictor importance analysis.

Main Results:

• The model accurately predicted acute SI with 0.702 accuracy (sensitivity=0.769, specificity=0.654, AUC=0.775).
• Social networking-derived predictors had a greater impact on prediction than linguistic predictors from structured interviews.
• Subject-specific analysis revealed informative trends for future acute SI risk prediction.

Conclusions:

• Ensemble learning on SN data can address complexities in modeling acute SI.
• Future research requires larger, diverse populations to refine digital biomarkers and validate external predictive capabilities.
• This approach may enhance the timely identification of individuals at risk for suicidal ideation.