A Novel Knowledge Distillation-Based Feature Selection for the Classification of ADHD.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Attention Deficit Hyperactivity Disorder (ADHD) affects millions globally, characterized by inattention and hyperactivity.
• Accurate ADHD diagnosis is challenging due to the lack of a gold standard test.
• Current diagnostic methods struggle to reliably differentiate ADHD patients from healthy individuals.

Purpose of the Study:

• To develop and validate a Knowledge Distillation-based approach for identifying discriminating neuroimaging features in ADHD.
• To improve the accuracy and reliability of ADHD diagnosis using machine learning on functional connectivity data.
• To outperform existing state-of-the-art methods in ADHD classification.

Main Methods:

• Utilized a Knowledge Distillation framework with a neural network trained on functional connectivity features.
• Employed an Autoencoder to reproduce learned embeddings from the neural network.
• Applied a forward feature selection algorithm to identify the most discriminating features between ADHD and healthy control groups.
• Validated the approach across five different independent datasets (KKI, Peking, NYU, NI, OHSU).

Main Results:

• Achieved promising classification results across all five individual sites.
• Reported a combined accuracy of up to 81% and individual site accuracies as high as 73%.
• The extracted features demonstrated superior performance compared to state-of-the-art methods in the literature.

Conclusions:

• The proposed Knowledge Distillation approach effectively identifies discriminating features for ADHD diagnosis.
• This method shows significant potential for improving the accuracy and efficiency of ADHD diagnosis.
• The findings validate the efficacy of using machine learning and functional connectivity for ADHD research.