Task-driven ICA feature generation for accurate and interpretable prediction using fMRI.

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Data Analysis

Background:

• Functional Magnetic Resonance Imaging (fMRI) holds promise for predicting disease progression and drug effects.
• Multivariate techniques enhance fMRI predictive power but face overfitting issues due to high dimensionality.
• Robust methods are needed to transform fMRI data into signals reflecting neural dynamics.

Purpose of the Study:

• To demonstrate a task-specific Independent Component Analysis (ICA) procedure for identifying functional brain network signals.
• To show that these ICA-derived signals can be used for accurate and interpretable prediction from fMRI data.
• To evaluate the performance of task-specific ICA parcellations against other feature generation methods.

Main Methods:

• Implemented a task-specific Independent Component Analysis (ICA) approach to parcellate fMRI data.
• Utilized two independent datasets to test the predictive performance of ICA-derived signals.
• Compared task-specific ICA parcellations with resting-state and anatomical parcellations.

Main Results:

• Task-specific ICA parcellations outperformed other feature generation methods in predictive accuracy across both datasets.
• The identified ICA signals were robust to non-neural artifacts and informative for prediction.
• Feature selection strategies highlighted the functional relevance of ICA-derived networks for discriminative tasks.

Conclusions:

• Task-specific ICA parcellation is a powerful technique for generating predictive and informative signals from fMRI time series.
• This method effectively addresses overfitting in high-dimensional fMRI data.
• The findings support the functional validity of ICA-based parcellations for neuroimaging research.