Machine Learning-Based Identification of Functional Dysregulation Characteristics in Core Brain Networks of Adolescents with Bipolar Disorder Using Task-fMRI.

Background and

Objective

Adolescent bipolar disorder (BD) has substantial symptom overlaps with other psychiatric disorders. Identifying its distinctive candidate neuroimaging markers may be helpful for exploratory early differentiation and to inform future translational studies after independent validation.

Methods

This cross-sectional study enrolled adolescents with BD and age- and sex-matched healthy controls. Assessments included clinical/behavioral scales and an emotional Go/NoGo task-based fMRI (Go trials require a response; NoGo trials require response inhibition) acquired across three mood states (depression, mania, and remission) and matched controls. We applied several conventional machine learning classifiers to task-fMRI data to classify BD versus healthy controls and to identify the most relevant neuroimaging predictors.

Results

A total of 43 adolescents with BD (15 in remission, 11 with depression, and 17 with mania) and 43 matched healthy controls were included. Under the Go - NoGo condition, activation-derived features in the remission state showed the strongest discrimination, with RF achieving the best performance (accuracy = 94.29%, AUC = 98.57%). These findings suggest that task-evoked functional alterations may remain detectable during remission. In addition, activation patterns in regions within the limbic system, prefrontal cortex, and default mode network were significantly correlated with clinical scales and behavioral measures implicating these regions in emotion regulation and cognitive functioning in adolescents with BD.

Conclusions

This study showed that adolescents with BD during remission without manic and depressive symptoms may still have aberrant neural activity in the limbic system, prefrontal cortex, and default mode network, which may serve as a potential candidate neuroimaging signature of adolescent BD.