Subject-Adaptive Transfer Learning Using Resting State EEG Signals for Cross-Subject EEG Motor Imagery Classification

TL;DR

The paper tackles inter-subject variability in EEG motor imagery by leveraging resting-state EEG signals to adapt models to unseen subjects without collecting target-subject task EEG. It introduces ResTL, a three-stage framework that first disentangles task- and subject-dependent features, then calibrates RS-EEG signals to embed task information via inverse-synthesis-style optimization, and finally fine-tunes the model on calibrated signals to simulate target-subject TS processing. Key contributions include a novel RS-EEG calibration mechanism guided by prototypes and triplet losses, and a three-stage training/adaptation pipeline that achieves state-of-the-art accuracy on three public benchmarks, with comprehensive analyses including t-SNE visualizations and ablations. The approach reduces dependence on time-consuming TS data from new subjects, enhancing practicality for real-world brain-computer interfaces while delivering robust cross-subject MI performance.

Abstract

Electroencephalography (EEG) motor imagery (MI) classification is a fundamental, yet challenging task due to the variation of signals between individuals i.e., inter-subject variability. Previous approaches try to mitigate this using task-specific (TS) EEG signals from the target subject in training. However, recording TS EEG signals requires time and limits its applicability in various fields. In contrast, resting state (RS) EEG signals are a viable alternative due to ease of acquisition with rich subject information. In this paper, we propose a novel subject-adaptive transfer learning strategy that utilizes RS EEG signals to adapt models on unseen subject data. Specifically, we disentangle extracted features into task- and subject-dependent features and use them to calibrate RS EEG signals for obtaining task information while preserving subject characteristics. The calibrated signals are then used to adapt the model to the target subject, enabling the model to simulate processing TS EEG signals of the target subject. The proposed method achieves state-of-the-art accuracy on three public benchmarks, demonstrating the effectiveness of our method in cross-subject EEG MI classification. Our findings highlight the potential of leveraging RS EEG signals to advance practical brain-computer interface systems. The code is available at https://github.com/SionAn/MICCAI2024-ResTL.

Figures (2)

• Figure 1: Overview of the proposed ResTL that comprises three stages: (1) Initial classifier training using cross-entropy loss $\mathcal{L}_{CE}$, task loss $\mathcal{L}_{task}$ and subject loss $\mathcal{L}_{sub}$ for disentanglement. (2) After training, RS EEG signals (violet) are calibrated to contain task-dependent features while retaining subject-dependent features by updating the initial signals with a fixed model. Here, different colors correspond to the different class labels. (3) Finally, the calibrated signals are used to fine-tune the pre-trained model for the target subject, minimizing $\mathcal{L}_{CE}$.
• Figure 2: t-SNE of calibrated signals on BCI IV-2b. (Blue: Right TS EEG signals, Red: Left TS EEG signals, Green: Calibrated signals to right, Violet: Calibrated signals to left). Calibrated signals generally have similar task-dependent features as TS EEG signals corresponding to their respective target classes i.e., (Blue, Green) and (Red, Violet) are closely clustered. While all subject-dependent features are clustered in the same group, indicating similar subject-dependent features. This clearly highlights that subject characteristics in RS EEG signals are preserved during calibration.