Multi-environment prediction of suicidal beliefs

Area of Science:

• Computational psychiatry
• Machine learning in mental health
• Suicidology

Background:

• Suicide rates are disproportionately high in military and veteran populations.
• Predicting suicidal behavior is challenging due to complex risk factors.
• Existing machine learning models often lack generalizability to new environments.

Purpose of the Study:

• To investigate environment-wise domain adaptation for predicting suicidal beliefs.
• To adapt machine learning models trained on individuals without suicidal ideation (SI) to predict suicidal beliefs in those with prior SI.
• To explore potential causal relationships between Suicide Cognitions Scale (SCS) and suicidal ideation/behavior.

Main Methods:

• Utilized environment-wise domain adaptation techniques.
• Adapted invariance-based machine learning models.
• Trained models on a sample without prior SI and tested on a sample with prior SI (N=2744 primary care patients, 17 risk/protective factors).

Main Results:

• The adapted models demonstrated some generalizability to a new environment (sample with prior SI).
• Results suggest that suicidal ideation and suicidal behavior are likely causally linked to SCS.
• The study highlights the potential of domain adaptation in predicting suicidal beliefs across different populations.

Conclusions:

• Environment-wise domain adaptation shows potential for improving suicide risk prediction in diverse populations.
• The Suicide Cognitions Scale (SCS) appears to have a causal relationship with suicidal ideation and behavior.
• Further research is needed to refine these methods for clinical application in identifying at-risk individuals.