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

📰 News: Our paper has been accepted for publication in National Science Review (IF=16.3).

This repository contains the original Python code for our paper Canine EEG Helps Human: Cross-Species and Cross-Modality Epileptic Seizure Detection via Multi-Space Alignment (National Science Review, 2025).

Overview

This work addresses the challenge of cross-species (canine/human) and cross-modality (scalp/intracranial) EEG-based epileptic seizure detection. Traditional models rely on within-species and within-modality data, limiting their generalizability. We introduce ResizeNet+Multi-Space Alignment (MSA), a multi-space alignment framework that facilitates knowledge transfer across species and modalities, overcoming dataset heterogeneity and enhancing seizure detection performance.

This study represents a pioneering effort in multi-species and multi-modality EEG integration, offering a scalable solution to train large brain models with diverse EEG data.

---

Challenges

Our study focuses on a scenario where the target species has little or no labeled data, leveraging auxiliary labeled data from another species or modality. The key challenges include:

• Electrode configuration and signal heterogeneity: Differences in electrode placements, sampling rates, and signal properties hinder direct transfer across species and modalities.
• Distributional shifts: EEG feature distributions vary significantly across species, datasets, and subjects, posing challenges for feature alignment.
• Limited labeled data: Scarcity of labeled seizure events in the target species constrains model training.

Cross-Species Similarity

Despite biological differences, EEG seizure patterns exhibit cross-species similarities across multiple feature domains:

• Temporal domain: Both canine and human EEG signals show pronounced fluctuations during seizures.
• Entropy domain: The approximate entropy of intracranial EEG increases significantly during seizures for both species, highlighting potential transferability.
• Spectral domain: Power spectral density analysis reveals similar increases in seizure-related frequency components across species.

Figure 1: Evidence for cross-species and cross-modality feature transferability.

Cross-Species Discrepancy

While similarities exist, significant discrepancies remain:

• Input space differences: Variations in electrode placement and device types introduce modality-specific biases.
  • Canine EEG data is acquired via implanted intracranial electrodes, whereas human scalp EEG is recorded non-invasively.
  • Even within the same modality, electrode configurations differ significantly (e.g., 16 intracranial electrodes for canines vs. 6 for humans).
• Feature distribution gaps: Distinct seizure characteristics across species lead to feature misalignment.

Figure 2: Gaps for successful cross-species knowledge transfer in algorithm design.

Proposed Framework

We propose the ResizeNet+MSA approach to enable epilepsy pattern transfer across species and modalities (see Figure 3).
ResizeNet is highly adaptable for cross-headset/cross-dataset transfer in BCI tasks, such as cross-headset motor imagery (MI) classification. We are actively exploring its applications in broader domains.

Figure 3: The framework of cross-species and cross-modality transfer network utilizes intracranial/scalp EEG data from canines and humans (left). ResizeNet, which projects EEG signals of the species with higher dimensionality to a lower dimensionality to match their feature spaces (right).

Key Results

We validate ResizeNet+MSA on four clinical EEG datasets (Kaggle, Freiburg, CHSZ, NICU), demonstrating significant improvements in cross-species seizure detection:

• Unsupervised transfer scenario: With no labeled data in the target domain, ResizeNet+MSA achieves 85.4% accuracy, outperforming non-alignment methods by 17%.
• Limited labeled data scenario: When the target domain has <5% labeled data, ResizeNet+MSA achieves an AUC of 92.8%, surpassing the within-species baseline by 18.7%.
• Feature preservation: ResizeNet retains essential EEG signal characteristics, see Figure 4.
• Effective feature alignment: Post-alignment, category-related features cluster more distinctly across species, improving classification robustness, see Figure 5.

Figure 4: Significant characteristic preservation after ResizeNet transformation.

Figure 5: Improved feature alignment across species using ResizeNet+MSA.

---

Citation

If you find this work useful, please consider citing our paper:

@article{wang2025canine,
  title={Canine EEG Helps Human: Cross-Species and Cross-Modality Epileptic Seizure Detection via Multi-Space Alignment},
  author={Wang, Ziwei and Li, Siyang and Wu, Dongrui},
  journal={National Science Review},
  pages={nwaf086},
  year={2025},
  publisher={Oxford University Press},
}

Contact

For any questions or collaborations, please feel free to reach out via [email protected] or open an issue in this repository.