Multimodal neuromarkers in schizophrenia via cognition-guided MRI fusion

Jing Sui, Shile Qi, Theo G.m. van Erp, Juan Bustillo, Rongtao Jiang, Dongdong Lin, Jessica A. Turner, Eswar Damaraju, Andrew R. Mayer, Yue Cui, Zening Fu, Yuhui Du, Jiayu Chen, Steven G. Potkin, Adrian Preda, Daniel H. Mathalon, Judith M. Ford, James Voyvodic, Bryon A. Mueller, Ayşenil Belger, Sarah McEwen, Daniel S. OʼLeary, Mary McMahon, Tianzi Jiang, Vince D. CalhounNature Communications 2018 · DOI

Abstract

Cognitive impairment is a feature of many psychiatric diseases, including schizophrenia. Here we aim to identify multimodal biomarkers for quantifying and predicting cognitive performance in individuals with schizophrenia and healthy controls. A supervised learning strategy is used to guide three-way multimodal magnetic resonance imaging (MRI) fusion in two independent cohorts including both healthy individuals and individuals with schizophrenia using multiple cognitive domain scores. Results highlight the salience network (gray matter, GM), corpus callosum (fractional anisotropy, FA), central executive and default-mode networks (fractional amplitude of low-frequency fluctuation, fALFF) as modality-specific biomarkers of generalized cognition. fALFF features are found to be more sensitive to cognitive domain differences, while the salience network in GM and corpus callosum in FA are highly consistent and predictive of multiple cognitive domains. These modality-specific brain regions define—in three separate cohorts—promising co-varying multimodal signatures that can be used as predictors of multi-domain cognition.

Introduction

Cognitive dysfunction is a core feature of schizophrenia, yet identifying reliable biological markers (neuromarkers) remains a challenge. Most neuroimaging studies focus on a single modality (e.g., just gray matter volume) or a single cognitive task. However, the brain is a complex, integrated system. This research addresses the lack of multimodal biomarkers that are both replicable across different clinics and capable of predicting cognitive performance in new, unseen individuals. Specifically, the study investigates the "triple network" model—the interaction between the Salience Network (SN), Central Executive Network (CEN), and Default Mode Network (DMN)—to determine if its disruption serves as a generalized signature for cognitive decline.

Methods and Data

The study employs a supervised learning strategy called MCCAR + jICA (Multimodal Canonical Correlation Analysis with Reference + joint Independent Component Analysis). Unlike "blind" fusion, which looks for any pattern of variation, this "guided" approach uses cognitive scores as a reference to steer the fusion process. It integrates three types of MRI data:

  1. sMRI (Structural): Gray Matter (GM) density.
  2. dMRI (Diffusion): Fractional Anisotropy (FA), measuring white matter integrity.
  3. rs-fMRI (Functional): Fractional Amplitude of Low-Frequency Fluctuation (fALFF), measuring local spontaneous brain activity.

By using cognitive scores as the "anchor," the algorithm identifies co-varying brain regions across all three modalities that specifically relate to cognitive performance. The researchers validated their model across three independent datasets: the Discovery Cohort (FBIRN, n=294), Replication Cohort 1 (UNM, n=83), and Replication Cohort 2 (COBRE, n=88). Notably, the cohorts used different neuropsychological batteries (CMINDS vs. MCCB), testing the model's ability to generalize across different clinical assessment tools.

Key Results

The model identified four key regions of interest as "neuromarker signatures": the Salience Network (GM), the Corpus Callosum (FA), the Central Executive Network (fALFF), and the Default Mode Network (fALFF).

Multimodal brain signatures correlate with cognition and differentiate schizophrenia patients from healthy controls.

Individuals with schizophrenia showed significant decreases across all identified multimodal signatures compared to healthy controls ($p < 10^{-7}$ for all modalities in the discovery cohort). Furthermore, the model trained on the FBIRN cohort successfully predicted cognitive scores in the independent UNM ($r = 0.231$) and COBRE ($r = 0.406$) cohorts. Despite differences in imaging sites and cognitive tests, the spatial maps of these neuromarkers showed high similarity ($r > 0.5$) across the different cohorts.

Discussion

A key finding was the relative stability of structural versus functional markers. Structural markers (GM and FA) were highly consistent across different cognitive domains (attention, memory, etc.), whereas functional markers (fALFF) were more "domain-specific," showing unique patterns depending on the specific cognitive task being measured.

The results strongly support the triple network model of psychopathology. Specifically, decreased gray matter in the Salience Network (insula and anterior cingulate) and reduced white matter integrity in the Corpus Callosum were the most consistent predictors of generalized cognitive impairment. The fusion approach revealed that deficits in one modality (like white matter tracts) often "connect" affected regions in another modality (like functional hubs in the CEN), providing a more holistic view of the disease's pathology.

While the method uses pre-calculated "feature maps" rather than raw 4D fMRI data, it provides a tractable and robust framework for biomarker discovery. Future models could incorporate Functional Network Connectivity (FNC) to further boost the accuracy of individualized cognitive predictions.

DOI: 10.1038/s41467-018-05432-w

Summary updated

Abstract

Cognitive impairment is a feature of many psychiatric diseases, including schizophrenia. Here we aim to identify multimodal biomarkers for quantifying and predicting cognitive performance in individuals with schizophrenia and healthy controls. A supervised learning strategy is used to guide three-way multimodal magnetic resonance imaging (MRI) fusion in two independent cohorts including both healthy individuals and individuals with schizophrenia using multiple cognitive domain scores. Results highlight the salience network (gray matter, GM), corpus callosum (fractional anisotropy, FA), central executive and default-mode networks (fractional amplitude of low-frequency fluctuation, fALFF) as modality-specific biomarkers of generalized cognition. FALFF features are found to be more sensitive to cognitive domain differences, while the salience network in GM and corpus callosum in FA are highly consistent and predictive of multiple cognitive domains. These modality-specific brain regions define-in three separate cohorts-promising co-varying multimodal signatures that can be used as predictors of multi-domain cognition.
BibTeX
@inproceedings{sui2018multimodal,
  author = {Jing Sui and Shile Qi and Theo G.M. van Erp and Juan Bustillo and Rongtao Jiang and Dongdong Lin and Jessica A. Turner and Eswar Damaraju and Andrew R. Mayer and Yue Cui and Zening Fu and Yuhui Du and Jiayu Chen and Steven G. Potkin and Adrian Preda and Daniel H. Mathalon and Judith M. Ford and James Voyvodic and Bryon A. Mueller and Ayşenil Belger and Sarah McEwen and Daniel S. OʼLeary and Mary McMahon and Tianzi Jiang and Vince D. Calhoun},
  title = {{Multimodal neuromarkers in schizophrenia via cognition-guided MRI fusion}},
  booktitle = {Nature Communications},
  year = {2018},
  doi = {10.1038/s41467-018-05432-w},
}
Figure