Table 2.

Techniques used for prediction of onset of movement and main findings of the studies reviewed.

Reference	Preprocessing techniques	Classifiers	Performance	Latency (ms)	Offline or online system	Single-trial analysis	Limitations
(Yom-Tov and Inbar, 2003) [43]	Low-pass filter (10 Hz) using 8th-order Chebyshev	Simple threshold element, support vector machine (SVM), and linear vector quantiser 3-feature reduction with 1-nearest neighbor (1-NN)	Using hybrid detector 25% improvement in performance was achieved as compared to Mason-Birch low frequency asynchronous detector (LFASD)	25 decisions s−1	Offline	—	Detector fails to work correctly partly due to MRPs related to other limbs and imagined movements
(Haw et al., 2006) [60]	Building a specific template during 3 or 4 training sessions for each subject	Thresholding based on correlation and error	Accuracy was 70% with a false positive rate (FPR) of (5/24)	—	—	Yes	Variability in performance between users
(Bai et al., 2007) [61]	Low pass filter (100 Hz) using 3rd-order Butterworth filter	Linear Mahalanobis Distance (MD), Quadratic MD, Bayesian Classifier (BC), Multilayer Perceptron (MLP) Neural Network, Probabilistic Neural Networks, and SVM	Accuracy was 75%	—	Offline	Yes	Large number of electrodes (122)
(Boye et al., 2008) [53]	Downsampling from 500 Hz to 20 Hz, with antialiasing prefiltering (0–5 Hz) and PCA and Locality Preserving Projection (LPP)	A variation of kNN and SVM	Sensitivity for SVM = 96.3 ± 2.0% for kNN = 84.5 ± 5.1%; specificity for SVM = 94.8 ± 2.7% and for kNN = 98.9 ± 1.2%	—	—	Yes	Method was tested on segmented data rather than ongoing EEG traces with only 1 subject
(Kato et al., 2011) [34]	Low pass filter (35 Hz) and high pass filter (0.05 Hz) for EEG and 0.1 Hz for EOG	SVM	Detection rate (intention to switch = 99.3% and (not to switch = 2.1%)	—	Both	Yes	Online system cannot differentiate between intend to switch and do not intend to switch
(Niazi et al., 2011) [42]	Band pass filter (0.05–10 Hz) with Optimized Spatial Filter (OSF)	Neyman Pearson Lemma	For healthy subject's movement execution TPR = 82.5 ± 7.81% and for movement imagination TPR = 64.5 ± 5.33%	−66.6 ± 121	Offline	Yes	Small sample size (patients) and no online detection due to instrumentational limitation
			For stroke patients TPR = 55.01 ± 12.01%	−56.8 ± 139
(Lew et al., 2012) [63]	Narrow band zero phase noncausal IIR filter with cutoff frequencies of 0.1 and 1 Hz	Linear Discriminant Analysis (LDA)	TPR = 76 ± 7% (healthy)	−167 ± 68 (healthy)	Offline	Yes	Large number of electrodes (34)
			For stroke and control subjects TPR = 81 ± 11% (left hand) versus (right hand) TPR = 79 ± 12%	Right hand = −140 ± 92 versus left hand = −162 ± 105
(Niazi et al., 2012) [19]	Band pass filter (0.1–100 Hz) and OSF	Matched Filter	TPR = 67.15 ± 7.87% and FPR = 22.05 ± 9.07%	−125 ± 309 (offline)	Online	—	Different aspects of triggered stimulations were not fully considered
(Niazi et al., 2013) [65]	Band pass filter (0.05–10 Hz) and OSF to maximize SNR	Matched Filter	For motor execution (healthy) TPR = 69 ± 21% and FPR = 2.8 ± 1.7	−196 ± 162	Offline	Yes	—
			For stroke patients TPR = 58 ± 11% and FPR = 4.1 ± 3.9	152 ± 239
			For motor imagery (healthy) TPR = 65 ± 22% and FPR = 4.0 ± 1.7	—
(Ahmadian et al., 2013) [64]	Filtering data between 0.1 Hz and 70 Hz	Independent component analysis (ICA)	Computation time for constraint blind source extraction (CBSE) algorithm was 0.26 s and blind source separation (BSS) algorithm took 51.90 s	260	—	Yes	Large number of electrodes (128) with small number of subjects
(Jochumsen et al., 2013) [39]	Band-pass filter (0.05–10 Hz) using 2nd-order Butterworth in forward and reverse direction with three spatial filters, large Laplacian spatial filter (LLSF), OSF, and common spatial patterns (CSP)	SVM	TPR = ~80% and FPR <1.5 accuracy = 80 ± 10% (speed) and 75 ± 9% (force)	317 ± 73	Offline	Yes	Inclusion of only healthy subjects
(Jiang et al., 2015) [66]	ICA followed by LSF to enhance SNR	ICA	TPR = 76.9 ± 8.97% and FPR = 2.93 ± 1.09 per minute	−180 ± 354	Offline	Yes	Prediction of gait initiation was not done
(Xu et al., 2014) [20]	Band-pass filter (0.05–3 Hz) and large LSF to enhance SNR	LPP followed by LDA	LPP-LDA TPR = 79 ± 12% FPR = 1.4 ± 0.8 per minute	315 ± 165	Online	—	Inclusion of only healthy subjects and classifier did not work for training trials less than 15