Incremental Learning for Patient-Specific EEG-Based Seizure Detection

Abstract

In recent years, advancements in wearable device technology have facilitated the shift from clinical epilepsy monitoring units to home-based EEG systems. This transition has driven the increasing demand for personalized precision treatments in ultra-long-term dynamic EEG data. However, existing EEG decoding models employ static architectures (with fixed parameters after training), which struggle to adapt to data distribution shifts caused by physiological, pharmacological, and environmental factors, limiting their applicability in long-term monitoring. To address this issue, we propose SAR-LTS, a novel incremental learning framework for seizure detection. Based on the significant similarity exhibited in short-term neural activity, SAR-LTS utilizes the similarity-aware mechanism with local temporal sampling to construct an experience pool for dynamically storing representative EEG samples. These samples are then replayed periodically via stratified random (SR) retrieval to reinforce historical knowledge. Experimental results on CHB-MIT and Siena datasets demonstrate that SAR-LTS significantly outperforms baseline models in patient-specific incremental learning scenarios. Compared to frozen models, it achieves an average accuracy improvement of 9.0-20.8%. This study provides a promising technical solution for long-term personalized epilepsy management.