Source-Free Domain Adaptation (SFDA) for Privacy-Preserving Seizure Subtype Classification

Area of Science:

• Medical Informatics
• Machine Learning
• Neurology

Background:

• Electroencephalogram (EEG) based seizure subtype classification is crucial for clinical diagnosis.
• Source-free domain adaptation (SFDA) enables privacy-preserving transfer learning by using pre-trained models instead of source data.
• SFDA reduces the need for extensive labeled data from new patients, crucial for privacy-sensitive medical applications.

Purpose of the Study:

• To introduce and evaluate a novel boosting-based SFDA approach for seizure subtype classification.
• To extend the proposed method to an unsupervised setting, eliminating the need for any labeled data from the target patient.
• To assess the performance of the proposed methods against existing machine learning techniques in challenging cross-dataset and cross-patient scenarios.

Main Methods:

• Development of semi-supervised transfer boosting (SS-TrBoosting) for SFDA in seizure classification.
• Extension of SS-TrBoosting to unsupervised transfer boosting (U-TrBoosting) for unsupervised SFDA.
• Experimental validation using three public seizure electroencephalogram (EEG) datasets.

Main Results:

• Both SS-TrBoosting and U-TrBoosting demonstrated superior performance compared to classical and state-of-the-art machine learning methods.
• The proposed methods achieved high accuracy in cross-dataset and cross-patient seizure subtype classification tasks.
• Effective transfer learning was achieved without direct access to source patient data, ensuring privacy.

Conclusions:

• SS-TrBoosting and U-TrBoosting represent effective SFDA strategies for EEG-based seizure subtype classification.
• The unsupervised approach (U-TrBoosting) is particularly promising for scenarios with limited or no labeled data for new patients.
• These methods offer a privacy-preserving and data-efficient solution for improving seizure diagnostics.