XGboost with multi-feature fusion for hemodynamic state prediction

This study presents a multi-scale feature fusion-based XGBoost prediction framework to address the hemodynamic state inversion problem in blood oxygenation level-dependent (BOLD) functional magnetic resonance imaging time series, addressing the limitations of existing methods, particularly unidimensional feature extraction and insufficient nonlinear modeling capability. Based on the balloon model, a hemodynamic state simulation dataset was constructed, incorporating vascular dilation, blood flow changes, and other physiological parameters. This study developed a time-frequency domain feature extraction system and employed recursive feature elimination to select an optimal feature subset. The selected features were then integrated into a Bayesian-optimized XGBoost model for state prediction. The experimental results indicate that the multi-scale feature fusion strategy enhances prediction accuracy compared with single-feature methods and reconstructs hemodynamic state time-series curves in real-world data. The proposed systematic feature extraction and fusion methodology presents innovative research perspectives and technical approaches for analyzing the nonlinear mapping relationship between neural activity and hemodynamic responses.