Kernel Principal Component Analysis for dimensionality reduction in fMRI-based diagnosis of ADHD

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Automated diagnosis of Attention-Deficit Hyperactivity Disorder (ADHD) is challenging.
• Functional Magnetic Resonance Imaging (fMRI) and phenotypic data offer potential diagnostic markers.
• Existing methods for ADHD classification using neuroimaging data have limitations.

Purpose of the Study:

• To explore and compare various feature extraction methods for ADHD diagnosis using fMRI and phenotypic data.
• To evaluate the effectiveness of machine learning classifiers, specifically Support Vector Machines (SVMs), for ADHD classification.
• To determine the optimal combination of data types (phenotypic, imaging) and feature extraction techniques for improved diagnostic accuracy.

Main Methods:

• Utilized the ADHD-200 dataset, including resting-state fMRI scans and phenotypic information (age, gender, IQ, etc.).
• Applied machine learning (SVM) to classify participants into diagnostic groups: ADHD vs. controls, and ADHD subtypes (combined/inattentive) vs. controls.
• Tested feature extraction methods including Fast Fourier Transform (FFT), Principal Component Analysis (PCA) variants (PCA-t, PCA-st, kPCA-st), and their combinations on fMRI data.

Main Results:

• Phenotypic data alone achieved 72.9% accuracy for two-class diagnosis and 66.8% for three-class diagnosis.
• Imaging data alone performed less effectively than phenotypic data.
• Combining phenotypic and imaging data with FFT and kernelized PCA (kPCA-st) yielded the highest accuracies: 76.0% (two-class) and 68.6% (three-class), outperforming phenotypic-only approaches.

Conclusions:

• The combination of phenotypic data and resting-state fMRI, processed with FFT and kPCA-st feature extraction, significantly enhances automated ADHD diagnosis.
• These findings highlight the potential of integrated data and advanced feature extraction for improving the accuracy of ADHD classification systems.
• The developed approach offers a promising avenue for more reliable ADHD diagnosis, addressing challenges associated with existing datasets like ADHD-200.