2.1 Motor Imagery Classification
5-layer CNN extract features and classify motor imagery EEG with left hand and right-hand movements (Zhichuan Tang, et. al, 2016). Conventional methods like SVM with power, SVM with common spatial pattern and SVM with autoregression are used in motor imagery classification. CNN performs better than the conventional classification methods. Combining deep learning with data augmentation for 2-way motor imagery classification (Zhiwen Zhang et. al, 2018). EMD decomposes EEG, mixes their IMFs to form new artificial frames of EEG, applied as inputs to complex morlet wavelets. CNN with wavelet NN helps in obtaining higher classification accuracy. STFT trains EG and give frequency domain representation. CNN and LSTM are deep NN used for EEG motor imagery classification with promising results (ZijianWang et.al, 2017).
Spatial distributions, β and µ rhythms help in imagery activities classification of EEG signals. Gradient descent, MLP are used for training the neural network which may lead to less accuracy, speed of convergence, PSO-GSA attains better accuracy, convergence speed in motor imagery classification (Sajjad Afrakhteh et. al, 2018), (Rahul Kala et. al, 2011). EEG muscular motor imagery is approximated based on RBF, further conventional MLP-NN and asynchronous NN are applied to increase accuracy, speed of control for the EEG classification (I. E. Shepelev et. al, 2018). Unclean EEG filtering, low SNR problems can be addressed using traditional BPNN, however an improved BPNN with weight splitting technique and PSO for appropriately training the low weights help in better motor imagery classification (Long Liu, 2019). Table 1 shows various techniques on motor imagery EEG
Table 1. Various motor imagery EEG aspects and techniques used
| Authors | Year | Type of Disease | Approaches | Achievement |
| Sunny T. D. et. al | 2016 | Motor imagery classification | Bayesian spatio-spectral filters | Classification performance |
| Rami Alazrai et. al | 2018 | Decoding finger movements-EEG | Choi William, quadratic T-F | 2-way classification |
| Akara Supratak et. al | 2017 | Sleep stage scoring | CNN, bidirectional LSTM | Accuracy, F1 score |
| Nicola Michielli et. al | 2019 | Sleep stage classification | Cascaded LSTM RNN | Neuro cognitive performance |
| Arnaud Sors et. al | 2017 | Sleep stage scoring | CNN | Cohort study, class wise patterns |
| Irene Sturm et. al | 2016 | Motor imagery classification | Layer-wise propagation | Neural activity complex perception |
2.1.1. Inferences
- •
Motor imagery classification with left hand, right hand movement can be done using conventional methods, CNN.
- •
Gradient descent, MLP are conventional training NN methods, PSO-GSA can give better convergence speed in motor imagery classifications.
- •
MLP-NN, asynchronous NN can increase accuracy, speed of control in muscular motor imagery EEG signal classification.
- •
Improved BPNN with weight splitting and PSO for appropriate weights training achieve better accuracy for motor imagery EEG classification.
- •
Deep EEG features and classification for motor imagery can be done using STFT, CNN and LSTM.