PAIR Comparison between Two Within-Group Conditions of Resting-State fMRI Improves Classification Accuracy

Area of Science:

• Neuroimaging
• Machine Learning
• Brain Function Analysis

Background:

• Classification using resting-state functional magnetic resonance imaging (RS-fMRI) aims to differentiate individuals.
• Previous studies show high accuracy within single datasets but struggle with multi-dataset generalization due to high dimensionality and variability.
• Challenges include feature selection and accounting for subject variability across different scanning conditions (eyes open/closed).

Purpose of the Study:

• To develop and validate a novel feature extraction method (PAIR) for RS-fMRI data.
• To improve classification accuracy and cross-dataset generalizability compared to traditional methods.
• To assess the effectiveness of treating eyes-closed (EC) and eyes-open (EO) conditions as paired rather than independent.

Main Methods:

• Utilized two independent RS-fMRI datasets with both EC and EO conditions.
• Applied Amplitude of Low-Frequency Fluctuation (ALFF) metric for feature extraction.
• Introduced the PAIR method, creating EC-EO and EO-EC difference maps for classification using Support Vector Machine (SVM).

Main Results:

• The PAIR method achieved high within-dataset classification accuracy (91.40% and 92.75%).
• Cross-dataset validation demonstrated robust performance (94.93% and 90.32%).
• The UNPAIR method showed substantially lower accuracy in both within-dataset and cross-dataset validation.

Conclusions:

• The PAIR method is recommended for feature extraction in within-group RS-fMRI studies with paired conditions.
• Dimensionality reduction incorporating prior knowledge of brain regions is crucial for feature selection.
• The findings highlight the importance of paired analysis for enhancing classification accuracy and generalizability in RS-fMRI.