TL;DR
This paper introduces a novel clustering approach to analyze the complex spatio-temporal interactions within the epileptic brain using intracranial EEG data. By leveraging a Self-Organizing Map-based similarity index (SOM-SI) and spectral clustering, the authors identify consistent patterns of regional brain activity during different seizure stages, offering potential insights into seizure focus and propagation.
What Problem Does This Paper Address?
Epileptic seizures are believed to arise from abnormal dynamic coupling between brain structures. While previous research has focused on analyzing temporal dynamics of EEG signals, understanding the spatio-temporal interactions and how they evolve during seizure events is crucial for identifying seizure foci and predicting seizure onset. Existing methods for analyzing these complex interactions often face challenges with computational complexity, sensitivity to noise, and parameter selection. This paper aims to address this by developing a robust and computationally efficient method to quantify and cluster these spatio-temporal interactions.
Methods
The core of the proposed method involves two main components:
- SOM-based Similarity Index (SOM-SI): This technique, an extension of the original Similarity Index (SI), uses a Self-Organizing Map (SOM) to create a quantized representation of the embedded time series data. This allows for a more computationally efficient estimation of the asymmetric dependency (influence) between different brain regions. The SOM-SI quantifies how the state-space trajectories of one signal correspond to the neighborhoods of another, providing a measure of functional connectivity.
- Spatio-Temporal Clustering Model: This model combines spectral clustering with a temporal quantification step using Hamming distance.
- Spatial Discretization: For short (10-second) time windows, the SOM-SI is computed for all channel pairs. These similarity indices are then used to form an affinity matrix, which is fed into a spectral clustering algorithm. This assigns each channel to one of a predefined number of clusters, representing spatial groupings at that specific time.
- Temporal Quantification: Over longer (30-minute) time segments, a cluster-similarity matrix is computed using Hamming distance. This matrix represents the probability that two channels share the same cluster label over that longer period.
- Hierarchical Clustering: Finally, hierarchical clustering is applied to the cluster-similarity matrix to identify stable spatio-temporal groupings of brain regions over extended periods.
Datasets
The study utilized intracranial EEG (iEEG) data from two patients with temporal lobe epilepsy. The data was recorded using depth and subdural electrodes, capturing activity from various brain regions including orbitofrontal, temporal, and subtemporal cortices, as well as hippocampal activity. The dataset comprised 8 complex partial seizures in total, with recordings spanning multiple days. The authors note that these patients had medically refractory epilepsy and underwent surgery.
Key Insights and Ideas
- Consistent Spatio-Temporal Patterns: The study found that despite variations in seizure states (interictal, preictal, ictal, postictal), the overall spatio-temporal groupings of brain regions remained remarkably consistent. This suggests a stable underlying functional architecture that is perturbed by seizures.
- Orbitofrontal Homologous Connectivity: The orbitofrontal regions consistently showed strong homologous connectivity, indicating a tightly coupled network within this area, irrespective of seizure activity.
- Left Hemisphere Ipsilateral Coupling: The left subtemporal and left temporal depth regions (considered nonfocal) exhibited strong ipsilateral connections, suggesting a robust intra-hemispheric functional relationship.
- Epileptic Focus Connectivity: The epileptic focus (right hippocampus depth region) maintained a relatively strong connection with the right subtemporal region.
- Potential for Seizure Blueprint: The consistent grouping patterns across seizures raise the intriguing question of whether these patterns could serve as a "blueprint" for epileptic seizures, potentially aiding in identifying seizure origins.
Key Results
- The proposed spatio-temporal clustering model was successfully applied to 8 complex partial seizures from 2 patients.
- Quantitative comparisons of SOM-SI profiles from different SOMs showed high correlation coefficients and low error percentages, validating the robustness of the SOM-SI measure.
- Qualitative analysis of the cluster groupings revealed consistent patterns across different seizure stages and between successive seizures. For instance, the LST and LTD channels in patient P093 consistently formed a group, as did the LOF and ROF channels.
- Statistical validation using surrogate analyses confirmed the veracity of the clustering results, with few exceptions noted for specific channel groups.
Limitations and Future Work
- Limited Data Size: The study was constrained by the availability of data, analyzing only 8 seizures from 2 patients. Future work should involve applying the approach to a larger dataset with more patients and a wider variety of seizure types (e.g., partial secondary generalized, subclinical seizures).
- Optimal Cluster Number: Determining the optimal number of clusters remains a challenge in clustering approaches. While the authors empirically fixed the number of clusters to 3 for consistency, they acknowledge the need for theoretical criteria to determine a fixed cluster size applicable to all datasets.
- Comparison with Normal Subjects: To confirm if the observed patterns are unique to epileptic brains, future research should involve applying the clustering approach to scalp EEG data from normal subjects and comparing the results.
- Exploring Alternative Tools: The authors suggest that exploring tools beyond clustering might be beneficial for unraveling hidden patterns in multidimensional time sequences.
- Statistical Significance of Connectivity: While strong ipsilateral connections were observed, statistical analysis to confirm significant differences in intrahemisphere coupling strengths is suggested.
Summary updated