Dataset distillation for efficient seizure prediction: tiny-scale data, comparable performance

Deep learning has shown significant potential in electroencephalogram (EEG)-based seizure prediction. However, translating these advances into practical applications faces critical challenges: the prohibitive time costs of reviewing accumulated historical data during frequent model updates, the need for rapid adaptation to diverse model architectures across hardware platforms, and the substantial burden of massive data storage. To this end, we introduce the first dataset distillation study tailored for seizure prediction, which condenses the original dataset into a compact, information-dense synthetic dataset. This tiny-scale proxy allows models to match the performance of those trained on the full-scale dataset.

Methodologically, we formulate the distillation as a min-max optimization task, introducing an EEG-aware alignment objective that jointly preserves latent feature distributions, temporal dynamics in the time-frequency domain, and inter-channel correlation structures. Through signal-compatible amplitude and phase decomposition, we optimize the frequency parameter to maximize the CF discrepancy via a sampling network, while simultaneously tuning the synthetic data to minimize this gap. Additionally, we integrate hard labels to refine soft labels to enhance the generalization of synthetic data.

Extensive validation across 37 public and 35 clinical subjects demonstrate the method's robustness across diverse scenarios and model architectures. With merely 1\% of the original data volume, our approach retains roughly 95\% of the original performance while reducing training duration by approximately 98\%.

This method addresses clinical translation bottlenecks by distilling massive EEG data into compact synthetic sets, offering a promising solution for efficient model maintenance, flexible model switching, and optimized storage.