Comparison of dimensionality reduction and feature selection for cognitive task decoding using functional...

Area of Science:

• Neuroimaging
• Machine Learning
• Data Science

Background:

• Functional magnetic resonance imaging (fMRI) enables brain study but generates excessive features, risking data overfitting.
• Feature selection (FS) and dimensionality reduction (DR) offer solutions by simplifying models for cognitive task decoding.

Purpose of the Study:

• To systematically compare the performance of various FS and DR methods in cognitive task decoding using machine learning classifiers.
• To identify optimal strategies for managing high-dimensional neuroimaging data.

Main Methods:

• Utilized two open-source fMRI datasets for comparative analysis.
• Applied a suite of machine learning classifiers with different FS and DR techniques.
• Conducted statistical tests to evaluate the contributions of DR/FS methods and classifiers to decoding accuracy.

Main Results:

• No single DR or FS method was universally superior; performance varied across datasets and tasks.
• Optimal predictive performance was achieved using a small fraction (0.05-0.10%) of the original features.
• Some methods outperformed baseline approaches (using all features or random selection) and even deep learning.

Conclusions:

• A 'sweet spot' exists for balancing feature retention and predictive accuracy in fMRI analysis.
• Careful application of FS and DR is crucial for effective cognitive task decoding.