High-Performance Cross-Subject Decoding of Multiclass Rhythmic Motor Imagery Using EEG Data From 100 Subjects

Effective cross-subject decoding is essential for reducing calibration time and enhancing the practical usability of brain-computer interfaces (BCIs). However, large inter-subject variability in EEG features poses a major challenge, particularly for motor imagery (MI) paradigms. Recent studies have shown that rhythmic MI can induce steady-state movement-related rhythms (SSMRR), which provide more structured electrophysiological features than conventional sensorimotor rhythms (SMR) and may offer a promising basis for efficient cross-subject decoding.

In this study, we comprehensively explored ways to achieve high-performance cross-subject decoding based on the rhythmic MI paradigm from both model and data perspectives.

We achieved an encouraging cross-subject four-class decoding accuracy of 72.94%±13.80% using a streamlined multilayer perceptron (MLP)-based network on a self-collected dataset comprising 100 BCI-naïve participants. From a model perspective, networks composed of simple MLP-based functional modules can achieve results comparable to, or even superior to, those of several state-of-the-art (SOTA) models. From a data perspective, increasing the training set size substantially improves cross-subject decoding performance (from 61.78% to 72.94%). Moreover, we revealed a strong positive correlation between EEG feature consistency and cross-subject decoding accuracy, providing a physiological explanation for why enlarging the training data scale enhances cross-subject generalization. Finally, we explored strategies for selecting high-quality training data. We found that feature-consistency-based selection serves as a more reliable criterion than within-subject decoding accuracy.

Overall, our study provides novel insights into cross-subject EEG decoding from the perspectives of model design, data scale and quality. The code is available in https://github.com/SJTUwyxuan/RhythmicMI-CrossSubject.