Large-scale nonlinear Granger causality (lsNGC) analysis of functional MRI data for schizophrenia classification

Alterations in the connectivity of the brain network play a crucial role in the pathogenesis of schizophrenia. We introduce large-scale Non-Linear Granger Causality (lsNGC), a novel method for analyzing effective connectivity in high-dimensional systems with short time-series, such as fMRI data. lsNGC incorporates dimension reduction into a nonlinear time-series prediction approach, enabling directed, multivariate causality analysis in complex networks. We apply lsNGC to resting-state fMRI data from the Centers of Biomedical Research Excellence (COBRE) repository to classify individuals with schizophrenia. Using brain connections estimated by lsNGC as features, we perform classification with support vector machines. Our method significantly outperforms competing approaches, achieving a mean AUC of 0.876 ± 0.094, compared to correlation (0.765 ± 0.163), local models (0.847 ± 0.139), and antisymmetric correlation (0.794 ± 0.115). Similarly, lsNGC achieves a mean F1 score of 0.831±0.088, outperforming correlation (0.694 ± 0.158), local models (0.764±0.128), and antisymmetric correlation (0.744 ± 0.120). Statistical analysis using Mann-Whitney U tests and Wilcoxon signed rank tests confirms the superior performance of lsNGC (p < 10−10) against all competing methods. These results suggest the potential of lsNGC as an imaging biomarker for schizophrenia and its applicability in clinical fMRI analysis for disease-related classification tasks.