An integrated feature ranking and selection framework for ADHD characterization.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Current attention deficit hyperactivity disorder (ADHD) diagnosis relies on subjective methods, leading to misdiagnosis and untreated cases.
• Neuroimaging, particularly magnetic resonance imaging (MRI), offers potential for objective ADHD biomarkers but generates high-dimensional data.
• Traditional machine learning struggles with high-dimensional neuroimaging data and small sample sizes, causing overfitting and computational issues.

Purpose of the Study:

• To develop an efficient framework for identifying discriminative brain features for ADHD diagnosis.
• To utilize normalized brain cortical thickness from MRI data for ADHD subject discrimination.
• To address challenges of high dimensionality and small sample sizes in neuroimaging-based ADHD research.

Main Methods:

• Proposed a novel integrated framework for feature ranking and selection.
• Combined information theoretic criteria and the Least Absolute Shrinkage and Selection Operator (Lasso) method.
• Applied the framework to normalized brain cortical thickness features from MRI data to distinguish ADHD subjects from healthy controls.

Main Results:

• The proposed framework achieved high/comparable ADHD prediction accuracy.
• It identified a minimal set of highly informative features, reducing model complexity.
• Selected brain regions of interest align with existing ADHD neuroimaging-behavioral studies.

Conclusions:

• The novel framework effectively identifies objective biomarkers for ADHD diagnosis using MRI data.
• This approach enhances diagnostic accuracy and aids in identifying relevant brain regions.
• The method offers a robust solution for feature selection in high-dimensional neuroimaging datasets for ADHD research.