Machine learning for anxiety diagnosis using structural MRI does not generalize to unseen data: results from a large...

Area of Science:

• Neuroscience
• Psychiatry
• Machine Learning

Background:

• Anxiety disorders are prevalent in youth, necessitating early detection.
• Neuroimaging biomarkers may offer objective diagnostic tools.
• Machine learning can analyze complex neuroimaging data for predictive modeling.

Purpose of the Study:

• To investigate the predictive utility of structural MRI for anxiety disorders in youth.
• To apply machine learning algorithms to brain imaging data for classification.
• To assess the generalizability of predictive models in a high-risk cohort.

Main Methods:

• Utilized structural MRI data from the Brazilian High-Risk Cohort for Mental Conditions (BHRCS).
• Employed a random forest classifier trained on brain features extracted using FreeSurfer.
• Incorporated feature selection, ComBat harmonization, and cross-validation for model development and evaluation.

Main Results:

• The model achieved 64% accuracy and an AUC of 0.70 on the test set.
• Performance significantly dropped in the validation sample (AUC = 0.51), indicating poor generalizability.
• Structural MRI features showed some discriminative ability but were insufficient for reliable clinical prediction.

Conclusions:

• Structural MRI data alone has limited predictive value for classifying anxiety disorders in youth.
• Neurobiological differences captured by MRI may not translate directly into robust clinical prediction models.
• Further research is needed to identify more effective neuroimaging biomarkers for youth anxiety.