Distinct neuroimaging subtypes of ADHD among adolescents based on semi-supervised learning

Attention deficit hyperactivity disorder (ADHD) is a childhood-onset neurodevelopmental disorder diagnosed and subtyped solely based on clinical traits, which are prone to subjective judgment and lack of reliability. Also, the clinical subtyping does not offer a clear indication of the patient outcome. Here, we propose to use a neuroimaging data-driven approach for subtyping, using a semi-supervised learning method to decipher the heterogeneity among ADHD patients. We identified three distinct subtypes of ADHD with abnormal cortical thickness (CT) compared to the controls, namely, the under-developed (lower CT), over-developed (higher CT), and mixed subtypes, based on 6509 adolescents from the Adolescent Brian Cognitive Development (ABCD) study. The findings were reliably repeated in external datasets. Interestingly, we found significantly lower cognitive scores together with worse socioeconomic status in the under-developed subtype, and the over-developed subtype had the worst response to stimulant medication. We further revealed significant differences in gene expressions and neurotransmitter distributions among the subtypes, pointing out that the upregulation of the dopamine and other excitatory pathways may play a strong role in the under-developed and mixed subtypes but not the over-developed subtype, which may explain their difference response to stimulant medication. Our study suggested that neuroimaging-based ADHD subtyping may uncover the disease heterogeneity in clinical presentations, treatment response, genetics, and neurobiology, and thereby, may potentially guide personalized therapy.