Machine Learning Prediction of Self-Injurious Outcomes in Adolescents by Sexual and Gender Identity

Abstract

Objective Sexual and gender minority adolescents face elevated rates of self-injurious thoughts and behaviors (SITBs) relative to peers, yet fewer studies have examined risk in these youth, and reasons for higher risk remain unclear. Modeling SITBs using traditional statistical models has proven challenging. More complex machine learning approaches may offer better performance and insights. We explored and compared multiple machine learning models of suicide ideation, suicide attempts, and non-suicidal self-injury—both past-year frequency and dichotomous lifetime occurrence—among adolescents of diverse gender identities and sexual orientations. Method Data came from a large adolescent survey (N = 2,452) including psychological and demographic features. We compared prediction performance between generalized linear models, random forest models, and gradient boosting decision tree models using the full sample. Results Contrary to hypotheses, we found that these models generally performed comparably. We then selected the best-performing model families to run follow-up comparisons between cisgender and gender minority adolescents and between heterosexual and sexual minority adolescents. Depression was consistently the top-ranked feature across all models save one, in which discrimination was the top-ranked feature for lifetime occurrence of suicide attempt in the gender minority group. In addition, loneliness was more important in the gender minority group relative to the cisgender group for models of suicidal ideation. Conclusion Discrimination and loneliness emerged as important features in predicting SITBs amongst gender minorities. Future work should examine these factors both as possible statistical predictors of SITB risk and as treatment targets for gender minority youth.