Knowledge-Data Fusion Based Source-Free Semi-Supervised Domain Adaptation for Seizure Subtype Classification

TL;DR

This work tackles privacy-preserving cross-subject EEG seizure subtype classification using source-free semi-supervised domain adaptation. It introduces Knowledge-Data Fusion (KDF) with mutual learning between a knowledge-driven Soft Decision Tree and a data-driven Vision Transformer, guided by Jensen–Shannon divergence $JSD$, followed by a target-domain adaptation stage called MutualSHOT that uses a consistency-based pseudo-labeling strategy and information maximization to align distributions without accessing source data. By fusing 41 handcrafted expert features with raw EEG processed by ViT, the approach achieves state-of-the-art cross-subject performance on CHSZ and TUSZ, even with only a single labeled target sample per class. The proposed framework advances privacy-preserving, data-efficient seizure subtype diagnosis with practical implications for clinical decision support and surgical planning.

Abstract

Electroencephalogram (EEG)-based seizure subtype classification enhances clinical diagnosis efficiency. Source-free semi-supervised domain adaptation (SF-SSDA), which transfers a pre-trained model to a new dataset with no source data and limited labeled target data, can be used for privacy-preserving seizure subtype classification. This paper considers two challenges in SF-SSDA for EEG-based seizure subtype classification: 1) How to effectively fuse both raw EEG data and expert knowledge in classifier design? 2) How to align the source and target domain distributions for SF-SSDA? We propose a Knowledge-Data Fusion based SF-SSDA approach, KDF-MutualSHOT, for EEG-based seizure subtype classification. In source model training, KDF uses Jensen-Shannon Divergence to facilitate mutual learning between a feature-driven Decision Tree-based model and a data-driven Transformer-based model. To adapt KDF to a new target dataset, an SF-SSDA algorithm, MutualSHOT, is developed, which features a consistency-based pseudo-label selection strategy. Experiments on the public TUSZ and CHSZ datasets demonstrated that KDF-MutualSHOT outperformed other supervised and source-free domain adaptation approaches in cross-subject seizure subtype classification.

Figures (5)

• Figure 1: Traditional machine learning and deep learning approaches for seizure subtype classification.
• Figure 2: Source-free domain adaptation.
• Figure 3: KDF-MutualSHOT in pre-training and fine-tuning stages. In the fine-tuning stage, the source EEG data and features are unavailable, and the source models are used to initialize the target models.
• Figure 4: The training loss calculation processing of MutualSHOT.
• Figure 5: Performance w.r.t. the number of labeled target samples. (a) TUSZ$\rightarrow$CHSZ; and, (b) CHSZ$\rightarrow$TUSZ.