TL;DR
This paper introduces a novel statistical framework, Multilevel Integrative Clustering (MIC), for analyzing synchronized brain activity from EEG data across multiple individuals. It addresses the challenge of pooling information from heterogeneous EEG samples to infer population-level synchronicity patterns, offering a robust method for understanding brain network dynamics.
What Problem Does This Paper Address?
Analyzing brain activity using Electroencephalography (EEG) is crucial for understanding cognitive processes and neurological conditions. However, inferring synchronized brain activity patterns from a diverse group of individuals presents significant challenges. Simply pooling EEG data from different subjects is problematic due to inherent variability and the complex nature of brain signals. Existing methods often struggle to account for both within-subject and between-subject variations, making it difficult to draw reliable conclusions about population-level brain synchronicity. This paper aims to bridge this gap by developing a statistically rigorous framework that can effectively integrate information from multiple subjects to reveal these population-level patterns.
Methods
The core of the proposed method is Multilevel Integrative Clustering (MIC). It combines several key ideas:
- Spectral Synchronicity: The paper defines synchronicity based on the similarity of spectral densities of EEG signals, capturing coordinated neuronal activity within specific frequency bands.
- Eigen-Laplacian Representation: EEG signals are transformed into d-dimensional eigen-Laplacian vectors. This step acts as a form of non-parametric dimension reduction, summarizing the spectral information in a way that's amenable to clustering.
- Hierarchical Bayesian Modeling: A multilevel hierarchical model is employed to capture the complex relationships between epochs within a subject, subjects within a population, and electrodes within the brain. This model uses a hierarchical Dirichlet prior to define cluster membership probabilities, allowing for interpretable parameters that quantify "adherence" to population-level patterns.
- Graph-Partitioning Clustering: The clustering of electrodes is framed as a graph-partitioning problem, where electrodes are nodes and their synchronicity defines the edges. This approach leverages techniques from spectral clustering.
- MCMC for Inference: Markov Chain Monte Carlo (MCMC) methods are used to sample from the posterior distribution, enabling robust estimation of model parameters and cluster assignments.
What's novel is the integration of these techniques within a unified hierarchical Bayesian framework that explicitly models population-level synchronicity while accounting for individual differences and temporal dynamics.
Datasets
The paper uses simulated EEG data for its simulation studies. For the case study, it utilizes resting-state EEG data from two cohorts of children:
- Typically Developing (TD) children (9 participants, 29-60 months of age).
- Children diagnosed with Autism Spectrum Disorder (ASD) (10 participants, 27-99 months of age).
The EEG data was recorded at 250Hz using 129-channel geodesic nets.
Key Insights and Ideas
- Population-Level Inference is Achievable: The paper demonstrates that it's possible to infer meaningful population-level brain synchronicity patterns from heterogeneous EEG data, a significant step beyond individual-level analyses.
- Hierarchical Structure is Key: The multilevel model effectively captures the nested structure of the data (epochs within subjects, subjects within a population), allowing for borrowing of strength across levels and more robust inference.
- Adherence Parameters Offer Interpretability: The "adherence" parameters (α and β) provide a quantitative measure of how closely individual or epoch-level patterns conform to the overall population patterns, offering insights into individual variability.
- Spectral Synchronicity as a Robust Measure: Using spectral synchronicity as the basis for similarity allows the framework to be applied even when precise temporal alignment of signals across subjects is difficult, which is common in resting-state EEG.
Key Results
- Simulation Studies:
- The MIC method showed good performance in selecting the correct number of clusters (K) across various smoothing window sizes.
- Posterior estimates of adherence parameters (α) were generally close to the true values, with over 99% of credible intervals covering the true α's.
- Clustering accuracy was high (over 98%) at the subject level and approached 100% as subject-level adherence to population patterns increased.
- Group-level inference was robust to varying signal-to-noise ratio (SNR) and sample size configurations.
- Case Study (TD vs. ASD Children):
- The analysis identified 5 spectrally synchronized brain regions (cortical areas).
- Population-level cluster estimates were remarkably similar between TD and ASD cohorts, with some notable differences in the partition of occipital and parieto-temporal regions.
- The TD cohort exhibited more stable brain regions (lower entropy) compared to the ASD cohort, particularly in left-temporal and left-central areas.
- Higher clustering heterogeneity (higher entropy) was observed in ASD children, aligning with existing literature on abnormal left-hemispheric asymmetry.
- Subject's age was found to be a significant predictor of classification accuracy in a post-hoc analysis, more so than specific power bands.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the model offers a "simple representation of cluster variability." For applications with strong dynamic evolution of synchronicity patterns (e.g., stimulus-based EEG), more complex modeling might be needed.
- The current framework assumes a fixed number of clusters (K) and eigen-Laplacian components (d), which are selected using information criteria like BIC.
- While the method is robust to meta-processing strategies, the choice of these strategies can still influence results.
- Future Work:
- Formal treatment of group comparison and covariate adjustments.
- Extending the model to incorporate predictors through cohesion functions or regression on subject-level coherence parameters for simpler, more interpretable models.
- Investigating applications where strong dynamic evolution of synchronicity patterns is expected.
- A user-friendly R package implementation of the MIC method is available online.
Summary updated