High-Performance Cross-Subject Decoding of Multiclass Rhythmic Motor Imagery Using EEG Data From 100 Subjects
Summary
This study enhances brain-computer interface (BCI) usability by improving cross-subject decoding for motor imagery (MI). Rhythmic MI with larger datasets and feature consistency boosts decoding accuracy, simplifying BCI systems.
Area of Science:
- Neuroscience
- Biomedical Engineering
- Signal Processing
Background:
- Brain-computer interfaces (BCIs) require effective cross-subject decoding to reduce calibration time and improve usability.
- Inter-subject variability in electroencephalography (EEG) features challenges motor imagery (MI) paradigms.
- Rhythmic MI induces steady-state movement-related rhythms (SSMRR), offering structured features for decoding.
Purpose of the Study:
- To explore high-performance cross-subject decoding using the rhythmic MI paradigm.
- To investigate the impact of model design and data characteristics on decoding performance.
- To identify optimal strategies for training data selection in cross-subject EEG decoding.
Main Methods:
- Utilized a multilayer perceptron (MLP)-based network for cross-subject decoding.
- Collected a dataset from 100 BCI-naïve participants.
- Analyzed the effects of training set size and EEG feature consistency on decoding accuracy.
Main Results:
- Achieved 72.94%±13.80% cross-subject four-class decoding accuracy.
- Demonstrated that MLP-based models perform comparably to state-of-the-art methods.
- Showed significant performance improvement with increased training data size and strong correlation between EEG feature consistency and accuracy.
Conclusions:
- Novel insights into cross-subject EEG decoding through model design, data scale, and quality.
- Feature-consistency-based data selection is more reliable than within-subject accuracy.
- Findings advance the development of practical and efficient BCIs.
