Optimizing functional brain network analysis by incorporating nonlinear factors and frequency band selection with machine learning models.

The accurate assessment of the brain's functional network is seen as crucial for the understanding of complex relationships between different brain regions. Hidden information within different frequency bands, which is often overlooked by traditional linear correlation-based methods such as Pearson correlation (PC) and partial correlation, fails to be revealed, leading to the neglect of more intricate nonlinear factors. These limitations were aimed to be overcome in this study by the combination of fast continuous wavelet transform and normalized mutual information (NMI) to develop a novel approach. Original time-domain signals from resting-state functional magnetic resonance imaging were decomposed into different frequency domains using fast continuous wavelet transform, and adjacency matrices were constructed to enhance feature separation across brain regions. Both linear and nonlinear aspects between brain regions were comprehensively considered through the integration of complex correlation coefficient and NMI. The construction of functional brain networks was enabled by the adaptive selection of optimal frequency band combinations. The construction of the model was facilitated by feature extraction using tree models with extreme gradient boosting. It was demonstrated through comparative analysis that the method outperformed baseline methods such as PC and NMI, achieving an area under the curve of 0.9054. The introduction of nonlinear factors was found to increase precision by 14.25% and recall by 17.14%. Importantly, the approach optimized the original data without significantly altering the feature topology. Overall, this innovation advances the understanding of brain function, offering more accurate potential for future research and clinical applications.