Diagnosis-Optimized Dynamic Feature Learning Reveals Altered Default Mode Network Connectivity in Schizophrenia

Masoud Seraji, Charles A. Ellis, Liang Ma, Jessica A. Turner, Adrian Preda, Daniel H. Mathalon, Theo G.m. van Erp, Mohammad S.e. Sendi, Vince D. CalhounbioRxiv 2025 · DOI

TL;DR

This study introduces a novel "explainability-guided" machine learning approach to analyze dynamic brain connectivity in schizophrenia (SZ). By iteratively identifying and removing the most influential connectivity features, the researchers uncovered subtle, yet significant, alterations in how the Default Mode Network (DMN) functions over time in SZ patients, suggesting these dynamic patterns could serve as valuable biomarkers for diagnosis and symptom tracking.

What Problem Does This Paper Address?

Schizophrenia (SZ) is a complex neuropsychiatric disorder characterized by widespread disruptions in brain connectivity, particularly within the Default Mode Network (DMN). The DMN is crucial for self-referential thought and mind-wandering, and its altered function is strongly linked to cognitive and social deficits seen in SZ. While previous studies have identified static connectivity differences, the dynamic, moment-to-moment fluctuations in brain network activity are thought to hold more nuanced information. However, standard dynamic functional connectivity (dFNC) analysis methods can be biased by dominant connectivity patterns, potentially masking subtler, yet clinically relevant, alterations. This paper aims to address this by developing a method that can uncover these less obvious, but potentially diagnostic, DMN connectivity states in SZ.

Methods

The core innovation here is an "explainability-guided iterative feature-removal clustering" approach. Here's a breakdown:

  1. Dynamic Functional Connectivity (dFNC) Extraction: Resting-state fMRI data was used to calculate pairwise correlations between different regions of the DMN over time using a sliding window approach. This generates a matrix representing how connectivity between DMN subnodes changes dynamically.
  2. Clustering into States: K-means clustering was applied to group these dynamic connectivity patterns into distinct "states" or "modes" of brain activity.
  3. Explainability-Guided Feature Removal: This is the novel part. Instead of just clustering, the researchers used a "Global Permutation Percent Change (G2PC)" method to identify which specific connectivity "edges" (connections between DMN subnodes) were most influential in defining these states. The most influential edge was then "ablated" (removed/masked), and the clustering process was repeated.
  4. Iterative Refinement: This feature removal and reclustering process was repeated until only a minimal set of features remained. This iterative process aims to reveal subtler connectivity patterns that might otherwise be overshadowed by dominant connections.
  5. State Occupancy Rate (OCR) Analysis: For each subject and each identified state, the percentage of time spent in that state (OCR) was calculated. These OCRs were then compared between SZ patients and healthy controls (HCs) and correlated with symptom severity.

Datasets

The study utilized two independent, well-established datasets:

  • FBIRN (Functional Imaging Biomedical Informatics Research Network): This dataset included 151 SZ patients and 160 healthy controls (HCs). Data was collected across seven sites.
  • COBRE (Center for Biomedical Research Excellence): This dataset included 68 SZ patients and 89 HCs. Data was collected at a single site.

The use of two independent datasets strengthens the generalizability and reproducibility of the findings.

Key Insights and Ideas

  • Subtle DMN Dynamics Matter: The iterative feature removal process revealed that even connections that aren't the "loudest" can carry crucial diagnostic information about SZ.
  • SZ Patients Occupy Aberrant States More: SZ patients spent significantly more time in certain DMN connectivity states characterized by atypical coupling and less time in states reflecting more "normative" or integrated DMN configurations.
  • Reproducible Diagnostic Signals: The feature removal order was remarkably similar across both independent datasets, indicating that the same DMN connections are driving the diagnostic signal for SZ, regardless of subtle differences in data acquisition.
  • Symptom-Connectivity Links: Specific DMN connectivity patterns were associated with different symptom dimensions (positive, negative, general), suggesting dynamic connectivity could inform clinical stratification.

Key Results

  • Stable DMN States: The identified DMN connectivity states (cluster centroids) were highly similar across both the FBIRN and COBRE datasets, suggesting a stable underlying topology of DMN configurations.
  • Differential State Occupancy: SZ patients showed significant differences in their occupancy rates (OCR) for several DMN states compared to HCs. For instance, HCs spent more time in a state characterized by robust, well-balanced connectivity (State 3), while SZ patients spent more time in other states.
  • Feature Importance Consistency: The iterative feature removal process yielded a highly consistent ranking of connectivity features across both datasets, with features eliminated early in one dataset also being eliminated early in the other.
  • Symptom Correlations: In the COBRE dataset, associations between symptom severity and state occupancy were more pronounced. For example, greater time in hyperconnected posterior DMN states was linked to positive symptoms, while reduced occupancy of integrated states related to negative/general psychopathology.

Limitations and Future Work

The authors acknowledge several limitations:

  • Dataset Differences: Variations in data acquisition (e.g., eyes open vs. eyes closed), scanner models, and sample sizes across sites could influence findings. While multi-site analyses were performed, these differences might explain some dataset-specific divergences.
  • Medication Effects: Most participants were on antipsychotic medication, making it difficult to disentangle illness-related connectivity changes from medication effects.
  • Limited Scope: The focus on the DMN means other brain networks implicated in SZ were not examined.
  • Cross-Sectional Design: The study is cross-sectional, limiting causal interpretations. Longitudinal studies are needed to understand how DMN dynamics change with treatment or illness progression.

Future work suggested includes:

  • Broader Application: Applying the methodology to other psychiatric and neurological disorders to identify transdiagnostic biomarkers.
  • Improved Explainability: Incorporating more sophisticated AI techniques for feature importance analysis.
  • Refined Feature Removal: Exploring alternative methods for feature removal, such as binning correlated features or grouping edges by network pairs.
  • Longitudinal and Real-Time Analysis: Investigating how DMN dynamics change over time and with treatment, and exploring real-time applications.
  • Multimodal Integration: Combining fMRI with EEG or MEG to explore spectral and electrophysiological correlates of DMN states.
  • Preclinical Models: Translating findings to animal models to investigate causal mechanisms.

Summary updated

Abstract

Schizophrenia (SZ) is characterized by widespread neural dysconnectivity, with particularly pronounced alterations in the default mode network (DMN)- a set of brain regions involved in self-referential thought and mind-wandering. We investigated dynamic functional connectivity within the DMN in SZ by analyzing resting-state fMRI data from two independent cohorts (FBIRN and COBRE). Using a novel iterative feature-removal clustering approach focusing on DMN independent components, we identified distinct recurring connectivity states while iteratively removing dominant connectivity features to reveal subtler network patterns. This approach iteratively refines the set of features considered, ensuring a more balanced representation and facilitating the identification of significant interactions that would otherwise be overlooked. Cluster centroids for the four connectivity states were highly similar across both datasets, reflecting stable shared DMN patterns. State occupancy was compared between participants with SZ and healthy controls, and associations with clinical symptom severity were examined. The SZ group exhibited significant alterations in DMN connectivity dynamics, spending more time than controls in certain connectivity states characterized by atypical DMN coupling and less time in a state reflecting a normative DMN configuration. A comparison between the FBIRN and COBRE datasets reveals both similarities and differences in occupancy rate (OCR) states. Overall, the pattern of group effects converged while state-specific divergences likely reflected protocol/sample differences rather than conflicting biology. Crucially, the explainability pipeline yielded a highly similar feature-removal order across datasets, indicating that the same DMN edges drive the diagnostic signal in both cohorts. Several OCR effects also tracked symptom ratings, linking abnormal DMN dynamics to clinical expression. Our findings suggest that SZ is characterized by reproducible disturbances in the temporal organization of DMN connectivity. These dynamic DMN features may serve as potential biomarkers for SZ, offering diagnostic and clinical utility by capturing network dysfunction that static connectivity measures could overlook.
BibTeX
@misc{seraji2025diagnosis,
  author = {Masoud Seraji and Charles A. Ellis and Liang Ma and Jessica A. Turner and Adrian Preda and Daniel H. Mathalon and Theo G.M. van Erp and Mohammad S.E. Sendi and Vince D. Calhoun},
  title = {{Diagnosis-Optimized Dynamic Feature Learning Reveals Altered Default Mode Network Connectivity in Schizophrenia}},
  year = {2025},
  doi = {10.1101/2025.09.30.679557},
  howpublished = {bioRxiv},
}
Figure