Predicting short-term suicidal thoughts in adolescents using machine learning: developing decision tools to identify...

Area of Science:

• Psychiatry
• Digital Health
• Adolescent Health

Background:

• Mobile technology offers novel methods for monitoring short-term suicide risk.
• Assessing theoretically informed risk factors daily can predict near-term suicidal ideation in adolescents.
• This approach can inform algorithms for real-time suicide-focused interventions.

Purpose of the Study:

• To predict next-day suicidal ideation in adolescent inpatients post-discharge.
• To develop and evaluate decision algorithms for identifying elevated daily suicide risk.
• To inform the development of timely, suicide-focused interventions.

Main Methods:

• Adolescents (N=78) completed daily text-based surveys for 4 weeks post-discharge.
• Multi-level classification and regression trees (CARTS) with cross-validation were used.
• Models predicted next-day suicidal ideation using daily risk factors, including person-specific means and deviations.

Main Results:

• The best model (AUC=0.86) incorporated ideation duration, hopelessness, burdensomeness, and self-efficacy.
• Excluding ideation duration resulted in acceptable performance (AUC=0.78).
• Models using only previous-day scores showed weaker performance (AUCs 0.75-0.82).

Conclusions:

• Dynamic risk factors assessed daily in adolescents show promise for predicting next-day suicidal thoughts.
• These findings support the development of decision tools for identifying short-term suicide risk.
• The approach can guide interventions sensitive to proximal increases in suicide risk.