AutoEncoder Filter Bank Common Spatial Patterns to decode Motor Imagery from EEG

• Published in: IEEE journal of biomedical and health informatics Vol. 27; no. 5; pp. 1 - 12
• Main Authors: Mammone, Nadia, Ieracitano, Cosimo, Adeli, Hojjat, Morabito, Francesco C.
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.05.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; AutoEncoders; Bioinformatics; Brain Computer Interface; Brain-Computer Interfaces; Classification; Datasets; Deep Learning; EEG; Electroencephalography; Electroencephalography - methods; Feature extraction; Filter banks; Humans; Imagery; Imagination; Information filters; Mental task performance; Motor Imagery; Neural networks; Neural Networks, Computer; Signal Processing, Computer-Assisted; Statistical analysis; Task analysis; Training; Transfer learning
• ISSN: 2168-2194; 2168-2208; 2168-2208
• DOI: 10.1109/JBHI.2023.3243698

The present paper introduces a novel method to decode imagined movement from electroencephalographic (EEG) signals. Decoding the imagined movement with good accuracy is a challenging topic in motor imagery (MI) BCIs, poor accuracy may indeed hinder the application of such systems in practice. The present paper introduces an extension of the well-established Filter Bank Common Spatial Patterns (FBCSP) algorithm, named AutoEncoder(AE)-FBCSP, to benefit from the ability of AE to learn how to map data from the feature space onto a latent space where information relevant for classification are is embedded. The proposed method is based on a global (cross-subject) and subsequent transfer learning subject-specific (intra-subject) approach. A multi-way extension of AE-FBCSP is also introduced in this paper. The proposed methodology consists of recording high-density EEG (64 electrodes). Features are extracted by means of FBCSP and used to train a custom AE, in an unsupervised way, to project the features into a compressed latent space. Latent features then are used to train a supervised classifier (feed forward neural network) to decode the imagined movement. The algorithm was tested using a dataset of EEG extracted from a publicly available database of data collected from 109 subjects. AE-FBCSP was extensively tested in the 3-way (right-hand vs left-hand motor imagery vs resting) classification and also in the 2-way, 4-way and 5-way ones, both in cross- and intra-subject analysis. AE-FBCSP outperformed standard FBCSP in a statistically significant way (p <inline-formula><tex-math notation="LaTeX">< </tex-math></inline-formula> 0.05) and outperformed also comparable methods in the literature applied to the same dataset. AE-FBCSP achieved an average accuracy of 89.09% in the 3-way subject-specific classification. With AE-FBCSP, 71.43% of subjects achieved a very high accuracy (<inline-formula><tex-math notation="LaTeX">></tex-math></inline-formula> 87.68%) whereas no subject achieved an accuracy <inline-formula><tex-math notation="LaTeX">></tex-math></inline-formula> 87.68% with FBCSP. One of the most interesting outcomes is that AE-FBCSP remarkably increased the number of subjects that responded with a very high accuracy, which is a fundamental requirement for BCI systems to be applied in practice.