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

TL;DR

The paper tackles the challenge of epilepsy seizure detection under data scarcity by leveraging cross-species and cross-modality EEG data (canine and human). It introduces a multi-space alignment (MSA) framework that unifies input normalization (Euclidean Alignment), input-space alignment (ResizeNet), feature-space alignment (domain adaptation), and output-space alignment (knowledge distillation). Across four public datasets spanning iEEG and sEEG, ResizeNet+MSA achieves over 90% AUC with as little as 5% labeled target data, demonstrating strong cross-species and cross-modality transfer. The approach suggests a scalable path to augment training data for large EEG models and potentially generalize to other brain-computer interface paradigms, with implications for translational neuroscience and clinical deployment in data-scarce populations.

Abstract

Epilepsy significantly impacts global health, affecting about 65 million people worldwide, along with various animal species. The diagnostic processes of epilepsy are often hindered by the transient and unpredictable nature of seizures. Here we propose a multi-space alignment approach based on cross-species and cross-modality electroencephalogram (EEG) data to enhance the detection capabilities and understanding of epileptic seizures. By employing deep learning techniques, including domain adaptation and knowledge distillation, our framework aligns cross-species and cross-modality EEG signals to enhance the detection capability beyond traditional within-species and with-modality models. Experiments on multiple surface and intracranial EEG datasets of humans and canines demonstrated substantial improvements in the detection accuracy, achieving over 90% AUC scores for cross-species and cross-modality seizure detection with extremely limited labeled data from the target species/modality. To our knowledge, this is the first study that demonstrates the effectiveness of integrating heterogeneous data from different species and modalities to improve EEG-based seizure detection performance. The approach may also be generalizable to different brain-computer interface paradigms, and suggests the possibility to combine data from different species/modalities to increase the amount of training data for large EEG models.

Figures (6)

• Figure 1: Evidences for cross-species and cross-modality feature transferability, along with gaps for successful knowledge transfer in algorithm design.a, iEEG from both canines and humans exhibits large fluctuations during epileptic seizures, indicating the transferability of time-domain features across species. b, The approximate entropy of iEEG from both species increases significantly during seizures, indicating the transferability of entropy features across species. c, Power spectral density spectrograms, derived from consecutive Fourier transforms for both species, show an increase in the power across all channels during seizures, suggesting the transferability of frequency-domain features. d, Input space disparity across species are highlighted by the discrepancy in electrode configurations between species. Sixteen intracranial electrodes were used for canines iEEG data collection in the Kaggle dataset, whereas only six were used for humans iEEG data collection in the Freiburg dataset. Canine iEEG signals were captured using intracranial electrodes linked to implanted devices, whereas human sEEG signals in the NICU and CHSZ datasets were collected via scalp electrodes, demonstrating the configuration differences. e, Feature distribution gaps between canines and humans.
• Figure 2: Ablation study, $t$-SNE feature visualization, and parameter sensitivity analysis. a, Ablation study on the Kaggle iEEG dataset, including two tasks of canine-to-human and human-to-canine, and two scenarios of unsupervised cross-species transfer and semi-supervised cross-species transfer. b, Feature $t$-SNE visualizations on the Canine-to-Human cross-species transfer task of Kaggle, CHSZ and NICU datasets. c, Parameter sensitivity analysis of $\lambda$ and $\beta$ on the Kaggle iEEG dataset and CHSZ sEEG dataset using the proposed ResizeNet+MSA approach. A point denotes the average, and the shadow denotes the standard deviation.
• Figure 3: Effect of Input Space Alignment via the Proposed ResizeNet on the Four Datasets (ResizeNet was not applied to the Kaggle H2 and H8 datasets, as their number of channels matches that of the canine dataset) a, The iEEG/sEEG signals before and after applying ResizeNet exhibit similar temporal structures, demonstrating its effectiveness. ResizeNet standardizes the number of channels across species, addressing input space discrepancies while preserving essential temporal structures and features. b, The mean approximate entropy of all channels before and after ResizeNet is shown for the four datasets, illustrating that ResizeNet retains critical entropy features of the iEEG/sEEG signals. c, Channel importance before and after applying ResizeNet, highlighting its ability to preserve significant spatial characteristics. Each column represents a different subject.
• Figure A1: Overview of the proposed cross-species and cross-modality epilepsy seizure detection framework.a, Training of the cross-species and cross-modality transfer network utilizes iEEG/sEEG data from both canines and humans. b, The proposed ResizeNet, which projects EEG signal of the species with higher dimensionality (collected with more EEG electrodes) to a lower dimensionality to match their feature spaces. c, The integration of knowledge distillation with ResizeNet for output-space alignment. d, Domain adaptation for distribution matching to achieve feature-space alignment. e, Illustration of the test phase on human iEEG/sEEG data.
• Figure A2: Illustration of three experiment scenarios, taking the Human-to-Canine transfer task as an example.