An in-depth survey on Deep Learning-based Motor Imagery Electroencephalogram (EEG) classification

被引:8
作者
Wang, Xianheng [1 ]
Liesaputra, Veronica [1 ]
Liu, Zhaobin [2 ]
Wang, Yi [3 ]
Huang, Zhiyi [1 ]
机构
[1] Univ Otago, Dept Comp Sci, Dunedin, New Zealand
[2] Dalian Maritime Univ, Coll Informat Sci & Technol, Dalian, Liaoning, Peoples R China
[3] RIKEN Ctr Brain Sci, Lab Circuit & Behav Physiol, Wako, Saitama, Japan
关键词
Motor imagery electroencephalogram; classification; Deep learning; Survey; BRAIN-COMPUTER INTERFACES; SINGLE-TRIAL EEG; NEURAL-NETWORKS; SIGNALS; FRAMEWORK; TASKS;
D O I
10.1016/j.artmed.2023.102738
中图分类号
TP18 [人工智能理论];
学科分类号
081104 ; 0812 ; 0835 ; 1405 ;
摘要
Electroencephalogram (EEG)-based Brain-Computer Interfaces (BCIs) build a communication path between human brain and external devices. Among EEG-based BCI paradigms, the most commonly used one is motor imagery (MI). As a hot research topic, MI EEG-based BCI has largely contributed to medical fields and smart home industry. However, because of the low signal-to-noise ratio (SNR) and the non-stationary characteristic of EEG data, it is difficult to correctly classify different types of MI-EEG signals. Recently, the advances in Deep Learning (DL) significantly facilitate the development of MI EEG-based BCIs. In this paper, we provide a systematic survey of DL-based MI-EEG classification methods. Specifically, we first comprehensively discuss several important aspects of DL-based MI-EEG classification, covering input formulations, network architectures, public datasets, etc. Then, we summarize problems in model performance comparison and give guidelines to future studies for fair performance comparison. Next, we fairly evaluate the representative DL-based models using source code released by the authors and meticulously analyse the evaluation results. By performing ablation study on the network architecture, we found that (1) effective feature fusion is indispensable for multi-stream CNN-based models. (2) LSTM should be combined with spatial feature extraction techniques to obtain good classification performance. (3) the use of dropout contributes little to improving the model performance, and that (4) adding fully connected layers to the models significantly increases their parameters but it might not improve their performance. Finally, we raise several open issues in MI-EEG classification and provide possible future research directions.
引用
收藏
页数:20
相关论文
共 133 条
[1]   A Review of Brain-Computer Interface Games and an Opinion Survey from Researchers, Developers and Users [J].
Ahn, Minkyu ;
Lee, Mijin ;
Choi, Jinyoung ;
Jun, Sung Chan .
SENSORS, 2014, 14 (08) :14601-14633
[2]   Deep learning for motor imagery EEG-based classification: A review [J].
Al-Saegh, Ali ;
Dawwd, Shefa A. ;
Abdul-Jabbar, Jassim M. .
BIOMEDICAL SIGNAL PROCESSING AND CONTROL, 2021, 63
[3]   A Deep Learning Framework for Decoding Motor Imagery Tasks of the Same Hand Using EEG Signals [J].
Alazrai, Rami ;
Abuhijleh, Motaz ;
Alwanni, Hisham ;
Daoud, Mohammad, I .
IEEE ACCESS, 2019, 7 :109612-109627
[4]   A Multibranch of Convolutional Neural Network Models for Electroencephalogram-Based Motor Imagery Classification [J].
Altuwaijri, Ghadir Ali ;
Muhammad, Ghulam .
BIOSENSORS-BASEL, 2022, 12 (01)
[5]   Deep Learning for EEG motor imagery classification based on multi-layer CNNs feature fusion [J].
Amin, Syed Umar ;
Alsulaiman, Mansour ;
Muhammad, Ghulam ;
Mekhtiche, Mohamed Amine ;
Hossain, M. Shamim .
FUTURE GENERATION COMPUTER SYSTEMS-THE INTERNATIONAL JOURNAL OF ESCIENCE, 2019, 101 :542-554
[6]   Multilevel Weighted Feature Fusion Using Convolutional Neural Networks for EEG Motor Imagery Classification [J].
Amin, Syed Umar ;
Alsulaiman, Mansour ;
Muhammad, Ghulam ;
Bencherif, Mohamed A. ;
Hossain, M. Shamim .
IEEE ACCESS, 2019, 7 :18940-18950
[7]   A Deep Learning Method for Classification of EEG Data Based on Motor Imagery [J].
An, Xiu ;
Kuang, Deping ;
Guo, Xiaojiao ;
Zhao, Yilu ;
He, Lianghua .
INTELLIGENT COMPUTING IN BIOINFORMATICS, 2014, 8590 :203-210
[8]  
Ang KK, 2008, IEEE IJCNN, P2390, DOI 10.1109/IJCNN.2008.4634130
[9]  
[Anonymous], 2023, Comparing the runtime of a PyTorch model vs a TensorFlow model
[10]   On the effects of data normalization for domain adaptation on EEG data [J].
Apicella, Andrea ;
Isgro, Francesco ;
Pollastro, Andrea ;
Prevete, Roberto .
ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE, 2023, 123