Resource-Efficient Continual Learning for Personalized Online Seizure Detection.

Area of Science:

• Neurology
• Artificial Intelligence
• Signal Processing

Background:

• Epilepsy diagnosis relies on time-consuming expert analysis of EEG signals.
• Automated seizure detection is challenging due to evolving patient-specific EEG patterns.
• Static deep learning models struggle with dynamic data, leading to information loss.

Purpose of the Study:

• To develop a personalized deep learning model for automated epileptic seizure detection.
• To address the challenge of integrating new data into models without catastrophic forgetting.
• To create a resource-efficient algorithm suitable for embedded systems.

Main Methods:

• Proposed a novel continual learning algorithm with a replay buffer mechanism for seizure detection.
• Integrated a replay buffer to retain past data information while incorporating new data.
• Evaluated the approach using the CHB-MIT EEG dataset.

Main Results:

• Achieved a 35.34% improvement in F1 score compared to standard fine-tuning.
• Demonstrated comparable F1 scores to resource-unlimited scenarios with a 1-hour replay buffer.
• Reduced the 24-hour False Alarm Rate by 33% compared to resource-unconstrained methods.

Conclusions:

• The proposed continual learning algorithm effectively enhances personalized epileptic seizure detection.
• The replay buffer mechanism is crucial for mitigating catastrophic forgetting in dynamic EEG data.
• The resource-efficient methodology is viable for real-world implementation in embedded systems.