Sparse knowledge sharing (SKS) for privacy-preserving domain incremental seizure detection.

Area of Science:

• Neurology
• Machine Learning
• Data Privacy

Background:

• Epilepsy affects millions globally, necessitating accurate electroencephalogram (EEG)-based seizure detection for diagnosis and monitoring.
• Current seizure detection models are often patient-specific due to data distribution shifts, requiring individualized training.
• Privacy-preserving domain incremental learning (PP-DIL) addresses sequential learning from different patient domains while protecting data privacy.

Purpose of the Study:

• To develop a privacy-preserving domain incremental learning approach for EEG-based seizure detection.
• To address challenges of catastrophic forgetting, privacy protection, and domain distribution shifts in PP-DIL.
• To propose a rehearsal-free method that balances model plasticity and stability.

Main Methods:

• Proposed Sparse Knowledge Sharing (SKS) approach for PP-DIL.
• Utilized Euclidean alignment to standardize data across domains.
• Implemented adaptive pruning for SKS to create domain-specific and shared parameters.
• Incorporated supervised contrastive learning to improve feature discrimination.

Main Results:

• SKS demonstrated superior performance in privacy-preserving domain incremental learning tasks.
• Experiments conducted on two public seizure datasets validated the effectiveness of SKS.
• The approach achieved a favorable balance between learning new information (plasticity) and retaining old information (stability).

Conclusions:

• SKS offers an effective rehearsal-free and privacy-preserving solution for sequential learning in seizure detection.
• The method successfully mitigates catastrophic forgetting and handles domain shifts.
• SKS provides a robust framework for personalized, privacy-conscious seizure monitoring systems.