From Childhood to Adulthood: An Attempt to Model Long-Term Outcomes of ADHD Using Machine Learning

Background. Childhood-onset attention-deficit/hyperactivity disorder (ADHD) persists in 28% to 58% of individuals, and is associated with adverse long-term outcomes, like additional psychiatric diagnoses and lower educational attainment. Knowing early life predictors for adverse long-term outcomes might help develop preventive measures, but too few have been identified using conventional statistical models. The current study investigated the potential of machine learning in identifying predictors for ADHD persistence and long-term functioning of adults with childhood-onset ADHD. Method. A total of 133 adults ( M _age = 29.53, SD _age = 2.83, 75% males) from an 18-year prospective cohort study were assessed on ADHD persistence and long-term functioning. Machine learning models were developed using cross-validation, with and without additional synthetic data, and compared to a conventional logistic regression model. With SHapley Additive exPlanations (SHAP) we explored predictor importance. Childhood predictors encompassed ADHD symptoms and treatment, other psychopathology, cognition, somatic characteristics, ADHD polygenic risk score, and parental demographics and psychopathology. Results. Accuracy of all models for ADHD persistence and functioning was low. SHAP suggested childhood anxiety symptoms as the most influential predictor for both outcomes. No machine learning model outperformed logistic regression for ADHD persistence (AUC’s _{machine learning} = .50–.56; AUC _{logistic regression} = .56) or long-term functioning (AUC’s _{machine learning} = .47–.59; AUC _{logistic regression} = .59). Adding synthetic data did not improve the models. Conclusions. Findings highlight the complexity of predicting long-term outcomes of ADHD. While conventional models have performed best so far, model improvement is crucial before models can inform clinical practice, e.g., by increasing sample size and investigating additional predictors.