Abstract
Schizophrenia is a chronic brain disorder associated with widespread alterations in functional brain connectivity. Although data-driven approaches such as independent component analysis are often used to study how schizophrenia impacts linearly connected networks, alterations within the underlying nonlinear functional connectivity structure remain largely unknown. Here we report the analysis of networks from explicitly nonlinear functional magnetic resonance imaging connectivity in a case-control dataset. We found systematic spatial variation, with higher nonlinear weight within core regions, suggesting that linear analyses underestimate functional connectivity within network centers. We also found that a unique nonlinear network incorporating default-mode, cingulo-opercular and central executive regions exhibits hypoconnectivity in schizophrenia, indicating that typically hidden connectivity patterns may reflect inefficient network integration in psychosis. Moreover, nonlinear networks including those previously implicated in auditory, linguistic and self-referential cognition exhibit heightened statistical sensitivity to schizophrenia diagnosis, collectively underscoring the potential of our methodology to resolve complex brain phenomena and transform clinical connectivity analysis.
Introduction
Functional magnetic resonance imaging (fMRI) studies typically rely on Pearson correlation to map brain connectivity. While effective, this method only captures linear relationships, potentially overlooking the complex, nonlinear dynamics of the brain. This "linear bias" may obscure critical biomarkers for psychiatric disorders like schizophrenia (SZ), often described as a "dysconnection syndrome." This study introduces a framework to isolate explicitly nonlinear (ENL) connectivity patterns and evaluates their utility in identifying neural signatures of schizophrenia.
Methods
The researchers developed a multi-step pipeline to isolate nonlinear signals from resting-state fMRI (rsfMRI) data. First, linear connectivity (LIN) was calculated using standard covariance. Next, total connectivity was measured using Distance Correlation (dCorr), which is sensitive to both linear and nonlinear dependencies. To isolate the purely nonlinear component, linear values were regressed out of the distance correlation values, yielding Explicitly Nonlinear (ENL) whole-brain functional connectivity. Group Spatial Independent Component Analysis (gr-sICA) was then applied to both LIN and ENL data to identify Intrinsic Connectivity Networks (ICNs).
The study utilized a large, multi-site dataset ($n=508$) aggregated from three major psychosis research projects: COBRE, FBIRN, and MPRC. The final pool included 315 healthy controls (HC) and 193 individuals diagnosed with schizophrenia (SZ).
Key Results
The Core-Periphery Gradient
A primary finding was the "core-periphery" gradient: ENL connectivity weight is systematically higher in the "core" regions of functional networks, whereas linear weight is higher in the periphery. This suggests that standard linear analyses underestimate connectivity at network centers.
Enhanced Sensitivity to Schizophrenia
ENL networks were significantly more sensitive to schizophrenia diagnosis than linear networks ($\chi^2 = 53.75, P < 0.00001$). In temporal regions associated with auditory and linguistic processing, ENL comparisons yielded an odds ratio of 22.63 for identifying SZ-related differences. For the posterior Default Mode Network (pDM), ENL analysis identified extensive clusters of alterations where linear analysis found only two significant voxels.
Unique Networks and Hidden Hypoconnectivity
The analysis revealed a unique ENL network that does not exist in the linear domain, bridging the default-mode, cingulo-opercular, and central executive regions. In schizophrenia patients, this network exhibited significant "hypoconnectivity" in the left anterior insula—a region critical for processing stimulus salience—which was invisible to traditional linear methods.
Discussion
These findings suggest that nonlinear patterns are robust, reproducible features of brain organization rather than noise, showing higher estimation stability than linear counterparts. While linear models explained approximately 53% of the variance in distance correlation, nearly half of the connectivity information remains untapped by standard methods. Future work should address demographic limitations, explore more complex non-parametric models for ENL isolation, and investigate the temporal dynamics of these networks. Establishing causality through interventions like TMS remains a critical next step for clinical application.
Summary updated