XGboost with multi-feature fusion for hemodynamic state prediction.

Area of Science:

• Neuroimaging
• Biophysics
• Machine Learning

Background:

• Existing blood oxygenation level-dependent (BOLD) functional magnetic resonance imaging (fMRI) methods struggle with unidimensional feature extraction and limited nonlinear modeling for hemodynamic state inversion.
• Accurate hemodynamic state inversion is crucial for understanding the relationship between neural activity and BOLD signals.

Purpose of the Study:

• To develop a novel multi-scale feature fusion-based XGBoost prediction framework for improved hemodynamic state inversion in BOLD fMRI time series.
• To overcome the limitations of existing methods by incorporating enhanced nonlinear modeling and feature extraction capabilities.

Main Methods:

• Constructed a hemodynamic state simulation dataset based on the balloon model, including parameters like vascular dilation and blood flow.
• Developed a time-frequency domain feature extraction system and utilized recursive feature elimination for optimal feature subset selection.
• Integrated selected features into a Bayesian-optimized XGBoost model for hemodynamic state prediction.

Main Results:

• The multi-scale feature fusion strategy significantly enhanced prediction accuracy compared to single-feature approaches.
• Successfully reconstructed hemodynamic state time-series curves using real-world BOLD fMRI data.
• Demonstrated the framework's capability in analyzing the nonlinear mapping between neural activity and hemodynamic responses.

Conclusions:

• The proposed framework offers innovative research perspectives and technical approaches for BOLD fMRI analysis.
• Multi-scale feature fusion is a key strategy for improving the accuracy of hemodynamic state inversion.
• The methodology provides a robust tool for understanding the complex relationship between brain activity and physiological responses.