Canine EEG Helps Human: Cross-Species and Cross-Modality Epileptic Seizure Detection via Multi-Space Alignment

1. Problem Statement

Epilepsy affects approximately 65 million people globally, yet accurate diagnosis and seizure detection remain challenging due to the transient nature of seizures and the heavy reliance on labor-intensive, expert-labeled EEG data. Current deep learning models for seizure detection face three critical bottlenecks:

• Data Scarcity: High-quality, labeled EEG data is difficult to obtain for specific patients or new datasets.
• Domain Shift: Models trained on one patient or dataset often fail to generalize to others due to non-stationarity and individual differences (cross-patient problem).
• Cross-Species/Modality Heterogeneity: While animal models (specifically canines) share significant neurological similarities with humans, existing algorithms cannot effectively transfer knowledge between species or modalities (e.g., intracranial EEG vs. scalp EEG) due to:
  • Input Heterogeneity: Differences in electrode configurations (e.g., 16 channels for canines vs. up to 72 for humans), sampling rates, and acquisition devices.
  • Distribution Discrepancies: Significant shifts in feature distributions and signal characteristics between species and modalities.
  • Label Scarcity: The target species (e.g., humans in a new clinical setting) often has little to no labeled data available for training.

2. Methodology: Multi-Space Alignment (MSA)

The authors propose a Multi-Space Alignment (MSA) framework, specifically ResizeNet+MSA, which performs simultaneous alignment in the input, feature, and output spaces to bridge the gap between source (auxiliary) and target domains.

A. Input Space Alignment

To handle channel dimensionality mismatches (e.g., 16 canine channels vs. 18 human scalp channels):

1. Euclidean Alignment (EA): A preprocessing step that standardizes the mean covariance matrices of EEG trials for each subject individually. This normalizes data distributions across different subjects and species.
2. ResizeNet: A novel neural network module designed to unify channel dimensions.
   • It utilizes a Transformer Encoder (with multi-head self-attention) to capture temporal dependencies and spatial relationships between channels.
   • A Linear Layer projects the high-dimensional input (e.g., human sEEG) to a lower-dimensional space (e.g., matching canine iEEG channels) or vice versa.
   • This ensures the input to the feature extractor has a consistent dimensionality without discarding valuable information via simple truncation.

B. Feature Space Alignment

To address distribution shifts in the extracted features:

• Domain Adaptation (DA): The framework employs unsupervised domain adaptation techniques (specifically using Maximum Mean Discrepancy, MMD) to minimize the distributional gap between the labeled source domain (e.g., Canine) and the unlabeled (or partially labeled) target domain (e.g., Human). This forces the model to learn domain-invariant features.

C. Output Space Alignment

To ensure predictive consistency:

• Knowledge Distillation (KD): The framework uses a "Teacher-Student" approach.
  • The "Teacher" model processes the resized signal (via ResizeNet).
  • The "Student" model processes a signal obtained via simple channel selection (baseline).
  • The Kullback-Leibler (KL) divergence is used to align the probability distributions (logits) of the two models. This refines the alignment by ensuring the resized signal preserves the essential classification information of the original signal.

D. Optimization Objective

The total loss function combines three components:
$$L = L_{CE} + \lambda \cdot L_{KD} + \beta \cdot L_{DA}$$
Where $L_{CE}$ is the supervised cross-entropy loss, $L_{KD}$ is the knowledge distillation loss, and $L_{DA}$ is the domain adaptation loss. $\lambda$ and $\beta$ are hyperparameters balancing the contributions.

3. Key Contributions

1. First Cross-Species/Modality Study: This is the first study to demonstrate that integrating heterogeneous EEG data from different species (Canine and Human) and modalities (iEEG and sEEG) can significantly improve seizure detection performance.
2. ResizeNet Architecture: Introduction of a Transformer-based projection network that effectively handles channel dimensionality mismatches while preserving spatial and temporal features, outperforming naive channel truncation.
3. Multi-Space Alignment Framework: A unified approach that simultaneously aligns input (dimensionality), feature (distribution), and output (prediction) spaces, addressing the complex heterogeneity of cross-species data.
4. Pilot for Large Brain Models: Provides a proof-of-concept for utilizing vast amounts of animal EEG data to train large-scale brain models, addressing the critical bottleneck of data scarcity in human neurology.

4. Experimental Results

The framework was evaluated on four public datasets: Kaggle (Canine iEEG), Freiburg (Human iEEG), CHSZ (Human sEEG), and NICU (Human sEEG).

• Performance Metrics: The primary metric was the Area Under the Curve (AUC).
• Unsupervised Transfer: In scenarios where the target species had zero labeled data, the MSA framework achieved substantial improvements over baselines. For example, in Canine-to-Human transfer, it achieved an average AUC of 75.53% (vs. ~66-71% for baselines).
• Semi-Supervised Transfer: With as little as 5% labeled data from the target species:
  • Canine-to-Human (iEEG): Achieved >90% AUC (e.g., 93.63% average on Kaggle).
  • Human-to-Canine: Achieved >90% AUC in most scenarios.
  • Cross-Modality (sEEG): Achieved >90% AUC when transferring from Human sEEG to Canine iEEG with 10% target labels.
• Ablation Studies:
  • ResizeNet was crucial; models without it (using simple channel truncation) performed significantly worse.
  • Multi-space alignment (Input + Feature + Output) consistently outperformed single-space alignment strategies.
  • EA (Euclidean Alignment) was identified as a mandatory preprocessing step for normalization.
• Feature Preservation: SHAP value analysis confirmed that ResizeNet preserved the importance of critical EEG channels and maintained temporal structures and entropy features after dimensionality reduction.

5. Significance and Implications

• Clinical Impact: This approach offers a viable solution for clinical settings where labeled data is scarce. It allows clinicians to leverage existing animal data or data from other modalities to build robust diagnostic tools for new patients.
• Biological Validation: The success of the transfer validates the hypothesis that epileptic mechanisms and signal characteristics are highly conserved across species (Canine and Human), supporting the use of canines as a biological model for human epilepsy.
• Future of AI in Neurology: The study paves the way for "Large Brain Models" by demonstrating that massive, heterogeneous datasets (combining human and animal data) can be effectively aligned and utilized to train deep learning models, overcoming the traditional barrier of data silos.
• Generalizability: The methodology is not limited to epilepsy; it can be adapted to other brain-computer interface (BCI) paradigms and neurological disorders where cross-species or cross-modality data integration is required.