Multimodal Ensemble Deep Learning to Predict Disruptive Behavior Disorders in Children.

Area of Science:

• Neuroscience
• Child Psychiatry
• Artificial Intelligence

Background:

• Disruptive behavior disorders (DBDs), including oppositional defiant disorder and conduct disorder, are common in children.
• Early diagnosis is critical to mitigate risks of comorbid mental health and substance use disorders.
• Diagnosing DBDs is challenging due to frequent co-occurrence with conditions like ADHD, anxiety, and depression.

Purpose of the Study:

• To develop and evaluate a multimodal deep learning model for classifying children with DBDs.
• To investigate the utility of combining diffusion, structural, and resting-state functional MRI data for DBD classification.
• To identify brain regions critical for DBD prediction using advanced visualization techniques.

Main Methods:

• A multimodal ensemble 3D CNN deep learning model was employed.
• Input data included diffusion, structural, and rs-fMRI from 1100 children (ages 108-131 months) in the ABCD Study.
• Classification performance was assessed, and brain regions were identified using gradient-weighted class activation mapping.

Main Results:

• The 3D CNN model achieved 72% accuracy in classifying children with DBDs.
• The model demonstrated 70% sensitivity, 72% specificity, and an F1-score of 70.
• Gradient-weighted class activation mapping highlighted key cortical and subcortical regions involved in DBD prediction.

Conclusions:

• Multimodal neuroimaging data integrated into a 3D CNN model shows promise for classifying DBDs in children.
• This approach can aid in early identification and potentially inform targeted interventions.
• Further research can refine these models for improved diagnostic accuracy and understanding of DBD neurobiology.