A genetically informed prediction model for suicidal and aggressive behaviour in teens.

Area of Science:

• Psychiatry
• Behavioral Genetics
• Machine Learning

Background:

• Suicidal and aggressive behaviors represent a significant public health concern.
• Overlapping risk factors suggest combined prediction approaches may enhance identification of at-risk individuals.
• Adolescent psychopathology requires robust predictive modeling for early intervention.

Purpose of the Study:

• To develop and validate a predictive model for suicidal behavior, aggressive behavior, or both in late adolescence.
• To assess the generalizability of the predictive model across different cohorts and geographical regions.
• To identify key genetic, environmental, and psychosocial predictors of these behaviors.

Main Methods:

• Utilized a large twin sample (N=5,974) from the Child and Adolescent Twin Study in Sweden (CATSS) for training and testing.
• Employed external validation with the Netherlands Twin Register (NTR; N=2,702).
• Developed a stacked ensemble machine learning model incorporating gradient boosted machines, random forests, elastic nets, and neural networks.

Main Results:

• The predictive model demonstrated good generalizability across CATSS and NTR cohorts (AUCs ~0.70-0.68).
• Key predictors included self-reported psychiatric symptoms, sex, and polygenic scores for psychiatric traits.
• Model performance for suicidal behaviors in the NTR was not significantly better than chance, indicating limitations.

Conclusions:

• A machine learning model integrating genetic and psychosocial data can predict combined suicidal and aggressive behaviors in adolescents.
• The findings suggest a potential role for genetic variables in adolescent psychopathology prediction.
• While comparable to existing methods, the model is not yet ready for clinical implementation.