Gradient boosting decision-tree-based algorithm with neuroimaging for personalized treatment in depression.

Area of Science:

• Neuroimaging
• Machine Learning
• Precision Medicine

Background:

• Predicting depression remission is crucial for effective treatment.
• Current biomarkers lack clinical validity.
• Machine learning (ML) offers potential for identifying novel biomarkers from neuroimaging data.

Purpose of the Study:

• To identify predictive biomarkers of depression remission using ML and pretreatment FDG-PET/MRS neuroimaging.
• To reduce patient suffering and economic burden associated with ineffective antidepressant trials.

Main Methods:

• Simultaneous FDG-PET/MRS neuroimaging was performed on 60 major depressive disorder (MDD) patients.
• Metabolic data from 22 brain regions (PET) and neurotransmitter concentrations (MRS) were analyzed.
• An eXtreme Gradient Boosting classifier was trained using imaging features, demographics, and treatment assignment.

Main Results:

• The ML model achieved 62% sensitivity, 92% specificity, and 77% accuracy in predicting remission.
• Pretreatment glucose metabolism in the left hippocampus, identified via PET, was the strongest predictor of remission.

Conclusions:

• Pretreatment neuroimaging can potentially prevent weeks of failed antidepressant treatments.
• This study demonstrates an effective ML approach for neuroimaging analysis in MDD, addressing challenges like small sample size and high dimensionality.