Using deep learning to classify pediatric posttraumatic stress disorder at the individual level.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Medical Imaging

Background:

• Children exposed to natural disasters face a high risk of developing posttraumatic stress disorder (PTSD).
• Previous research identified altered brain network topology in pediatric PTSD using functional neuroimaging.
• Graph-based network metrics have shown differences between pediatric PTSD patients and healthy controls (HC).

Purpose of the Study:

• To investigate the utility of deep learning (DL) models in classifying pediatric PTSD using neuroimaging markers.
• To assess the potential of DL-driven classification for aiding in the diagnosis of pediatric PTSD.
• To identify specific brain regions and networks crucial for distinguishing pediatric PTSD from HC.

Main Methods:

• Functional connectivity was analyzed in 33 pediatric PTSD patients and 53 HC using resting-state fMRI data.
• A whole-brain functional connectome was constructed based on partial correlation coefficients between 90 brain regions.
• Graph theory analysis examined topological properties, and a DL algorithm was employed for classification.

Main Results:

• Deep learning models utilizing graph topological measures achieved 71.2% overall accuracy in differentiating pediatric PTSD from HC.
• Frontoparietal areas (central executive network), cingulate cortex, and amygdala were key contributors to the DL model's classification performance.
• These findings suggest the potential clinical utility of neuroimaging-based DL classifiers.

Conclusions:

• Graph topological measures derived from fMRI data can form the basis of clinically useful imaging models for distinguishing pediatric PTSD.
• Deep learning models show potential as valuable tools for identifying the brain mechanisms underlying PTSD in pediatric populations.
• This approach may enhance diagnostic capabilities and understanding of PTSD neurobiology in children.