Machine learning prediction of conduct problems in children using the longitudinal ABCD study.

BACKGROUND

Children with conduct problems are at elevated risk for negative psychosocial, educational, and behavioral outcomes. Identifying at-risk children can aid in providing timely intervention and prevention, ultimately improving their long-term outcomes. There is a need to develop screening tools to better identify at-risk children who may benefit from early intervention.

METHODS

Data were collected from the longitudinal Adolescent Brain Cognitive Development (ABCD) Study. Children completed a baseline visit at age 9-10, then returned annually for 3 years (n = 3,517). We used machine learning classifiers (logistic regression, Naïve Bayes, support vector machine, and random forest) to predict conduct problems (i.e., conduct disorder or oppositional defiant disorder) in children after 1, 2, and 3 years.

RESULTS

The best-performing model (the random forest classifier) predicted children at risk for conduct problems with an accuracy of 90% or greater (AUC = 0.98 at 1 year, AUC = 0.97 at 2 years, AUC = 0.97 at 3 years). A random forest classifier simplified to include only 10 features was able to predict conduct problems nearly as well (AUC = 0.97 at 1 year, AUC = 0.96 at 2 years, AUC = 0.97 at 3 years).

CONCLUSIONS

Using factors previously linked to conduct problems, we built machine learning models to identify predictors of conduct problems in children over a 3-year period. A small number of self-report features can be used to predict persistent conduct problems with 90% or greater specificity and sensitivity up to 3 years after initial assessment. This suggests that parent and child self-report data, along with machine learning, can identify children at risk for persistent conduct problems.