MAE-MI: masked autoencoder for self-supervised learning on motor imagery EEG data
Yifan Zhang, Xinyu Hu, Huijie Feng, Anqi Wu, Hao Li, Yang Yu · 2024
In this study, we propose a self-supervised method for learning universal neurocognitive representations from motor imagery electroencephalography (EEG) segments. Our model, named MAE-MI, is a masked autoencoder that learns robust and generic embeddings. The aim is to capture the essential features and recover the masked EEG signals during pretraining, rather than simply interpolating. Therefore, we design a connectivity guidance masking strategy. Furthermore, we optimize the encoder-decoder structure to suit the information density of EEG signals. We also introduce two finetuning modes for downstream tasks that are task-specific and subject-specific; they correspond to cross-subject and singlesubject evaluations, respectively. We assess the generalization performance of MAE-MI on a public motor imagery EEG dataset. The experimental results indicate that MAE-MI consistently outperforms the state-of-the-art methods with an average accuracy increase of 9.4% in single-subject prediction experiments; it also obtains competitive results in crosssubject experiments. We illustrate the feature extraction capability of MAE-MI by visualizing its reconstruction effects on corrupted EEG signals.