Machine learning of clinical and neural data predicts future homicide in high-risk youth

Homicide is one of the most severe forms of violence and is the single most costly crime in the United States. Despite its high financial cost and incalculable emotional toll, we know very little about predicting who will commit such crimes. Here we followed a large (n = 202) sample of incarcerated youth for 16 years after release from a maximum-security juvenile correctional facility. Official records and clinical interviews indicated n = 35 committed a homicide during this follow-up period. Clinical (psychopathic traits, age of first arrest) and neural variables (gray matter volume of the amygdala and temporal pole) collected at baseline while boys were housed at the correctional facility significantly differed between homicide and non-homicide (n = 167) groups. Classification through machine learning models using these clinical and neural data predicted which formerly incarcerated youth committed a future homicide as adults with high accuracy. The model which included a priori clinical (psychopathic traits) and neural variables (bilateral amygdala, insula, parahippocampal gyrus, middle temporal pole, superior temporal pole, and orbitofrontal cortex) achieved top performance with 76% accuracy (sensitivity = 86%; specificity = 75%). The implications of these results are discussed as they relate to intervention and prevention efforts.