Using deep learning to predict internalizing problems from brain structure in youth.

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Internalizing problems, such as anxiety and depression, are linked to adverse outcomes.
• Biological markers for internalizing problems remain poorly understood.
• Neurodevelopmental (ND) conditions frequently co-occur with internalizing problems.

Purpose of the Study:

• To utilize deep learning to predict internalizing problems using brain structure data.
• To assess the models' ability to predict cross-sectional internalizing problems.
• To evaluate the models' capacity for predicting longitudinal worsening trajectories of internalizing problems.

Main Methods:

• Deep learning models were developed using brain structure measures (thickness, surface area, volume).
• Models were trained and tested on four large-scale datasets (ABCD, HBN, HCP-D, ONPN).
• Performance was evaluated using area under the receiving operating characteristic curve (AUC) via stratified cross-validation.

Main Results:

• The cross-sectional model achieved an AUC of 0.80 for predicting internalizing problems.
• The longitudinal model showed sub-optimal performance in the general population (AUC=0.66).
• The longitudinal model performed well in external test sets of neurodevelopmental conditions (AUC=0.80) and across all ND conditions (AUC=0.73).

Conclusions:

• Deep learning utilizing brain structure features shows potential as a biomarker for internalizing problems.
• This approach is particularly promising for identifying individuals at higher risk within neurodevelopmental populations.
• Further research is warranted to refine these predictive models for clinical application.