Explainable fuzzy clustering framework reveals divergent default mode network connectivity dynamics in schizophrenia

Charles A. Ellis, Robyn L. Miller, Vince D. CalhounFrontiers in Psychiatry 2024 · DOI

TL;DR

Researchers have developed a new "explainable fuzzy clustering" framework to analyze how brain networks shift over time in individuals with schizophrenia. By moving away from rigid "hard" clustering to a probabilistic model, the study reveals that schizophrenia is characterized by more rapid transitions between brain states and specific patterns of network anticorrelation that traditional methods often miss.

What Problem Does This Paper Address?

Neuroscientists use Dynamic Functional Network Connectivity (dFNC) to study how different regions of the brain "talk" to each other over time. Traditionally, this is done using "hard" clustering (like k-means), which assigns every moment of brain activity to a single discrete state.

However, the human brain doesn't usually exist in just one state; it’s often in a transition or a mixture of states. Hard clustering ignores this nuance, assuming all samples in a cluster are identical and forcing "in-between" data points into boxes where they might not fit. This makes it difficult to compare individuals—especially those with neuropsychiatric disorders like schizophrenia (SZ)—who may not manifest a specific brain state strongly enough to be "caught" by a hard cluster.

Methods

The authors propose an explainable fuzzy clustering framework that replaces binary assignments with probabilities.

  1. Fuzzy C-Means (FCM): Instead of saying a brain scan is 100% "State A," FCM might say it is 60% State A and 40% State B.
  2. Explainability Metrics: To understand why the model chooses a state, they used Global Permutation Percent Change (G2PC) and a novel Permutation-based Distribution Divergence (P2D). These methods perturb specific brain connections to see how much they influence the final clustering, effectively "opening the black box" of the model.
  3. Novel Summary Features: They introduced new ways to quantify brain dynamics, such as:
  4. Occupancy Rate (OCR): Percentage of time spent in a state.
  5. State Transitions (NST): How often the brain switches states.
  6. Fuzzy-specific features: Including Shannon entropy (to measure state uncertainty) and Kullback-Leibler divergence (to measure state stability).

Datasets

The study utilized the Functional Imaging Biomedical Informatics Research Network (FBIRN) dataset.

  • Participants: 151 individuals with schizophrenia (SZ) and 160 healthy controls (HC).
  • Modality: Resting-state functional magnetic resonance imaging (rs-fMRI) collected across 7 different sites.
  • Preprocessing: Used the Neuromark pipeline, an automated ICA-based approach to extract reproducible brain network components.

Key Insights and Ideas

  • The "In-Between" is Informative: By using fuzzy clustering, the researchers could capture subtle transitions that hard clustering obscures. This revealed that the stability of these memberships is a key differentiator in schizophrenia.
  • Divergent DMN Dynamics: The study focused on the Default Mode Network (DMN)—the brain's "autopilot" system. They found that the connections between the anterior cingulate cortex (ACC), posterior cingulate cortex (PCC), and the precuneus are the most critical features for defining these brain states.
  • Rapid Volatility: One of the most striking "aha" moments was the discovery that the SZ brain isn't just in "different" states; it moves between them differently, showing higher volatility between low-magnitude and high-magnitude connectivity patterns.

Key Results

  • State Preferences: Individuals with SZ spend significantly more time in states characterized by moderate anticorrelation between the anterior and posterior cingulate cortices and strong anticorrelation between the precuneus and the anterior cingulate.
  • Transition Speed: The SZ group showed a tendency to transition more rapidly between states than healthy controls.
  • Classification Performance: Using a Logistic Regression model with Elastic Net Regularization (LR-ENR), the researchers achieved an AUC of 0.68 using correlation-based features and average probabilities, proving these fuzzy features are statistically robust for distinguishing SZ from HC.
  • Symptom Correlation: They found a relationship between the "cumulative change" in fuzzy state probabilities and the severity of negative symptoms in SZ (though this was less significant after strict FDR correction).

Limitations and Future Work

  • Window Size: The dFNC was calculated using a 40-second sliding window. While standard, the authors acknowledge that different window sizes could shift the results.
  • Outlier Sensitivity: Like k-means, Fuzzy C-Means can be sensitive to outliers. Future iterations might explore "robust" fuzzy clustering variants.
  • Generalizability: The authors suggest testing the framework on other datasets and different disorders (like Bipolar Disorder or Autism) to see if these "fuzzy" dynamics are unique to schizophrenia or a broader hallmark of neuropsychiatric conditions.
  • Optimization: Future work will look into more robust ways to automatically determine the optimal number of fuzzy states (clusters) for a given dataset.

Summary updated

Abstract

We present a novel dFNC analysis framework and use it to identify effects of SZ upon network dynamics. Given the ease of implementing our framework and its enhanced insight into network dynamics, it has great potential for use in future dFNC studies.
BibTeX
@inproceedings{ellis2024explainable,
  author = {Charles A. Ellis and Robyn L. Miller and Vince D. Calhoun},
  title = {{Explainable fuzzy clustering framework reveals divergent default mode network connectivity dynamics in schizophrenia}},
  booktitle = {Frontiers in Psychiatry},
  year = {2024},
  doi = {10.3389/fpsyt.2024.1165424},
}
Figure