Privacy-preserving multi-source semi-supervised domain adaptation for seizure prediction.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neurology

Background:

• Inter-patient variability in EEG data poses challenges for seizure prediction models.
• Existing domain adaptation (DA) methods raise privacy concerns due to data access requirements.
• Current DA approaches often overlook variability among source patients, hindering adaptation.

Purpose of the Study:

• To develop a novel, privacy-preserving, multi-source-free semi-supervised domain adaptive seizure prediction model (MSF-SSDA-SPM).
• To address limitations of existing DA methods by avoiding direct access to source patient data and accounting for inter-patient variability.

Main Methods:

• The MSF-SSDA-SPM model treats each source patient as a distinct source, generating individual pretrained models.
• Adaptation is performed using pretrained source models and limited labeled target data, without accessing raw source data.
• A knowledge distillation strategy integrates knowledge from adapted source models into a single target model, optimizing feature extractors jointly while freezing classifiers.

Main Results:

• On the CHB-MIT dataset, MSF-SSDA-SPM achieved 88.6% sensitivity, 0.182/h false positive rate (FPR), and 0.856 AUC.
• On the Kaggle dataset, the model obtained 78.6% sensitivity, 0.178/h FPR, and 0.784 AUC.
• The results demonstrate significant improvements in seizure prediction performance and robust privacy protection.

Conclusions:

• The proposed MSF-SSDA-SPM model effectively overcomes privacy concerns associated with traditional DA methods in EEG-based seizure prediction.
• The model's ability to handle inter-patient variability among source patients leads to improved prediction accuracy.
• MSF-SSDA-SPM offers a promising solution for developing reliable and private seizure prediction systems.