A Composable Channel-Adaptive Architecture for Seizure Classification

TL;DR

The paper tackles cross-subject, multi-channel iEEG seizure detection by introducing a channel-adaptive architecture that processes channels independently, fuses them with a learnable HRR-based scheme, and stores long-term context in memory. It demonstrates that pre-training on large heterogeneous datasets combined with minimal subject-specific fine-tuning yields state-of-the-art performance with substantially fewer parameters than Transformer-based approaches. The approach is validated on short- and long-term iEEG datasets, showing improved F1-scores and data efficiency, while ablations confirm the importance of pre-training, HRR fusion, and memory. Practically, this CA framework enables scalable, plug-and-play deployment across diverse clinical setups with extended context windows suitable for real-world seizure monitoring.

Abstract

Objective: We develop a channel-adaptive (CA) architecture that seamlessly processes multi-variate time-series with an arbitrary number of channels, and in particular intracranial electroencephalography (iEEG) recordings. Methods: Our CA architecture first processes the iEEG signal using state-of-the-art models applied to each single channel independently. The resulting features are then fused using a vector-symbolic algorithm which reconstructs the spatial relationship using a trainable scalar per channel. Finally, the fused features are accumulated in a long-term memory of up to 2 minutes to perform the classification. Each CA-model can then be pre-trained on a large corpus of iEEG recordings from multiple heterogeneous subjects. The pre-trained model is personalized to each subject via a quick fine-tuning routine, which uses equal or lower amounts of data compared to existing state-of-the-art models, but requiring only 1/5 of the time. Results: We evaluate our CA-models on a seizure detection task both on a short-term (~20 hours) and a long-term (~2500 hours) dataset. In particular, our CA-EEGWaveNet is trained on a single seizure of the tested subject, while the baseline EEGWaveNet is trained on all but one. Even in this challenging scenario, our CA-EEGWaveNet surpasses the baseline in median F1-score (0.78 vs 0.76). Similarly, CA-EEGNet based on EEGNet, also surpasses its baseline in median F1-score (0.79 vs 0.74). Conclusion and significance: Our CA-model addresses two issues: first, it is channel-adaptive and can therefore be trained across heterogeneous subjects without loss of performance; second, it increases the effective temporal context size to a clinically-relevant length. Therefore, our model is a drop-in replacement for existing models, bringing better characteristics and performance across the board.

Figures (11)

• Figure 1: Architecture of the CA classifier. The classifier consists of three components. The Encoder component acts channel-wise and extracts short-term temporal features from each channel. The holographic (HRR) Fusion component combines the channel-wise features while preserving the spatial structure. Finally, the Memory component (TCN) extracts the relevant long-term temporal features and outputs the classification decision.
• Figure 2: Seizure classification performance of our CA models on the Long-term SWEC iEEG Dataset. The CA models are fine-tuned on a single seizure of the subject (LABOC), compared with all but one seizures for the baseline models (LOOC). The mean (dashed line), median (solid line), and quartile boxes are visible.
• Figure 3: Seizure classification performance of our CA models on the Short-term SWEC iEEG Dataset. The mean (dashed line), median (solid line), and quartile boxes are visible.
• Figure 4: The Fusion component of CA-EEGNet learns the spatial structure of the signal as training progresses, visualized through the cosine similarity between the generated key vectors. On the left, the component is initialized uniformly at the beginning of training and the relationship between the channels is simply linear. On the right, at the end of training, the relationship between the channels is more complex, representing a biologically plausible view of the connection between different electrodes and brain areas.
• Figure 5: Performance of CA-EEGNet when pre-trained with increasing number of subjects on the Long-term SWEC iEEG dataset. Performance increases up to 15 subjects but slows down, indicating that the scaling advantage is saturating. The mean (dashed line), median (solid line), and quartile boxes are visible.