Revealing Hidden Patterns in Schizophrenia rs-fMRI via Novel Unsupervised Learning

609 

Abstract Submission 

Masoud Seraji¹, Charles Ellis¹, Mohammad Sendi¹, Vince Calhoun¹

¹Center for Translational Research in Neuroimaging and Data Science (TReNDS), Atlanta, GA

Masoud Seraji  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Charles Ellis  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Mohammad Sendi  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Vince Calhoun  
Center for Translational Research in Neuroimaging and Data Science (TReNDS)
Atlanta, GA

Resting-state functional magnetic resonance imaging (rs-fMRI), especially with dynamic functional network connectivity (dFNC) analysis, has provided insights into neuropsychiatric disorders. In this field, clustering techniques are commonly used to identify dFNC states, uncovering connections between features and disorder pathology. However, in multidimensional feature spaces, influential features may overshadow others relevant to the condition of interest. This is evident in schizophrenia (SZ), where whole brain dFNC analyses failed to detect SZ effects on the default mode network (DMN) [Sendi et al., 2020] that were later revealed in DMN-specific analyses [Sendi et al., 2021]. This raises questions about the potential for obscured effects in whole-DMN analyses. Our study aims to mitigate the potential obscuring influence of dominant features in clustering analyses by introducing a novel feature learning-based approach that identifies a maximally significant, minimally complex (MSMC) subset of DMN-related dFNC features associated with SZ.

Figure 1 shows our methodology, using the Functional Imaging Biomedical Informatics Research Network (FBIRN) dataset with 151 individuals with schizophrenia (SZs) and 160 healthy controls (HCs). Five initial mock scans were excluded before preprocessing. We employed statistical parametric mapping (SPM12) for motion correction and normalization to the echo-planar imaging template in the standard Montreal Neurological Imaging space (3x3x3 mm³). The GIFT toolbox was then used to extract 7 independent components from the default mode detwork (DMN): three precuneus (PCu), two anterior cingulate cortex (ACC), and two posterior cingulate cortex (PCC) nodes. Subsequently, we applied a sliding tapered window to estimate dynamic functional network connectivity (dFNC) and extracted 21 connectivity features within the DMN. To identify an MSMC subset of features, we iteratively assigned samples to five clusters using k-means clustering with correlation distance, used Global Permutation Percent Change (G2PC) feature importance [Ellis et al., 2021] to identify and remove the most important feature, and repeated the process until correlation distance became inapplicable. We evaluated the impact of feature removal on the occupancy rate (OCR, i.e., the percentage of time each subject spent in each state), for SZs and HCs. Two-sample t-tests compared OCR differences between SZs and HCs at each iteration. Further analysis focused on the iteration with the most significant OCRs and the fewest features (i.e., the MSMC iteration).

Figure 2 shows the OCR analysis, confirming our hypothesis that top features may overshadow more relevant dynamic features in SZ during clustering. Intra-PCC and intra-ACC features were removed, leaving some intra-PCN and inter-region features. SZ individuals primarily occupied states 0 and 4, characterized by high PCN/PCC correlation and moderate PCC1/ACC anticorrelation. In contrast, healthy controls (HCs) favored state 3, the least common state, marked by moderate to high PCN/PCC correlation and high PCC1/ACC correlation. Our PCN/PCC and PCC1/ACC findings fit with those of [Ellis et al., 2022] and [Ellis et al., 2023], respectively.

Prior research has indicated that conventional clustering methods used for identifying dFNC states might overlook crucial features associated with neuropsychiatric disorders. In this investigation, we introduced a novel methodology to explore this potential issue in the context of the DMN in SZ. Notably, we observed that the largest number of significant distinctions between individuals with SZ and HC only became apparent after eliminating several of the top features, thereby supporting the validity of our initial hypothesis. We anticipate that our innovative approach will enhance the comprehensive understanding of neuropsychiatric and neurological disorders in future studies.

Psychiatric (eg. Depression, Anxiety, Schizophrenia) ¹

Connectivity (eg. functional, effective, structural)

fMRI Connectivity and Network Modeling ²

Methods Development

Task-Independent and Resting-State Analysis

ADULTS

Affective Disorders

Computational Neuroscience

Data analysis

DISORDERS

FUNCTIONAL MRI

Machine Learning

Modeling

Psychiatric

Schizophrenia