Incremental Learning for Patient-Specific EEG-Based Seizure Detection.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Wearable technology enables home-based electroencephalography (EEG) monitoring, increasing demand for personalized epilepsy treatments.
• Static EEG decoding models fail to adapt to dynamic data shifts in long-term monitoring.
• This limits the effectiveness of current precision epilepsy management strategies.

Purpose of the Study:

• To develop a novel incremental learning framework for adaptive, patient-specific seizure detection.
• To address the limitations of static models in handling ultra-long-term dynamic EEG data.
• To enhance the accuracy and applicability of EEG-based epilepsy monitoring.

Main Methods:

• Proposed SAR-LTS, a similarity-aware incremental learning framework for seizure detection.
• Utilized local temporal sampling to create a dynamic experience pool of representative EEG samples.
• Employed stratified random (SR) retrieval for periodic replay to reinforce historical knowledge.

Main Results:

• SAR-LTS significantly outperformed baseline models in patient-specific incremental learning scenarios.
• Achieved an average accuracy improvement of 9.0-20.8% compared to static, frozen models.
• Demonstrated superior performance on the CHB-MIT and Siena EEG datasets.

Conclusions:

• SAR-LTS offers a promising solution for long-term, personalized epilepsy management through adaptive EEG analysis.
• The framework effectively addresses data distribution shifts in dynamic EEG data.
• Enables more robust and accurate seizure detection in real-world, long-term monitoring settings.