The study is divided in two phases. In the first phase which called as Before Neurofeedback(BNF), experiments were performed in three groups of participants named as able-bodied (AB), paraplegic patients with (PWP) and with no (PNP) neuropathic pain. In all, thirty subjects participated in the BNF phase of this study (10 in each group). The age of participants was between 18 and 55 years (AB group: 3F, 7M, age 39.6 ± 10.2; PNP group: 2F, 8M, age 43.8 ± 9.1; PWP group: 3F, 7M, age 46.2 ± 9.4). In the second phase which called as After Neurofeedback(ANF), experiments were performed on five PWP participants (recruited from the first phase) who have completed neurofeedback training provided for the management of pain. Informed consent was obtained from all participants, and the ethical approval to the trials was granted from the University of Strathclyde Ethical Committee and from National Health Service Ethical Committee for Greater Glasgow and Clyde.
BNF data acquisition
Neuroscan EEG machine with 61 electrodes (according to standard 10–10 locations) was used to record task-related EEG at sampling rate of 1000 Hz. Participants were trained to perform three types of motor imaginary tasks which are imagination of the right hand waving (RH), left hand waving (LH), and tapping with both feet (F) for 3 sec excluding 1 sec resting and 1 sec preparation periods. Therefore, the length of each epoch is 5 sec long. In total, EEG data was recorded with 60 trials of each motor imaginary task (180 trials in total) carried out by each subject.
Neurofeedback training was provided to five PWP participants with the aim of reinforcing α band power, and inhibiting Ɵ and β. The alpha band power was reinforced because it was reported in some previous neurofeedback studies that increasing α band has promising results on chronic pain including CNP [29,30,31]. Additionally, the group of PWP participants in our study had the dominant α frequency on average 1 Hz lower than that of PNP participants. Therefore PWP participants were trained to increase the energy of higher α band (9–12 Hz) which does not include lowest α frequencies at 8Hz. The β and Ɵ power was inhibited because these frequencies bands are believed to be positively correlated with pain [21, 30].
Neurofeedback training was conducted firstly with an audio neurofeedback during which patients sat in front of a computer screen listening to music. The level of music was louder when α power was under the threshold, set at 110% of the baseline value, and quieter when α power was above the threshold. Following the audio neurofeedback, visual neurofeedback was provided to the patients in order to decrease β and Ɵ power to a value under the threshold set at 90% of the baseline value. In order to visualize α, β and Ɵ power changes during neurofeedback, three bars were presented on the computer screen. The changes of bars size represent a change of relative EEG power in three selected frequency bands as shown in Fig. 1. The bars color turned green when the EEG power of the chosen frequency band was in the desired range and turned red otherwise. The bar in the middle represented α band and turned green when the power was above the threshold set at 110% of the baseline value. Two sidebars represented β and Ɵ frequency bands and turned green when the power was below the threshold set at 90% of the baseline value.
As the five PWP participants, who completed 20 or more neurofeedback sessions, reported to have a sensation of pleasant warmth, the ANF EEG power in three targeted frequency bands (α, β and Ɵ) was compared to BNF EEG power. Throughout the comparison, it was noticed that there was a significant change in ANF EEG power as the β and Ɵ were suppressed and α power increased after neurofeedback. The change of EEG power after neurofeedback in two patients, (PWP1, PWP2), as represented in Fig. 2 was more noticeable.
Following neurofeedback training of five PWP participants, EEG data was obtained from them in order to test the accuracy of BCI-classifier developed, in first phase of this study, based on the pre-neurofeedback EEG. ANF data was obtained from the five participants in the same way and with the same EEG machine used in the first phase. The participants were instructed to perform MI of limbs during EEG recording in this phase as well.
EEG data was down-sampled to 250 Hz. An infinite impulse response high pass filter with 12db cutoff frequency was set to 1 Hz and a notch filter was applied between 48 and 52 Hz to eliminate line noise at 50 Hz. Before performing the analysis, EEG data were re-referenced to the average reference because the evoked activity, or signal of interest, has a more or less fixed time-delay to the stimulus, while the ongoing EEG activity behaves as additive noise. The averaging procedure will enhance the signal-to-noise ratio.
In this study time–frequency decomposition was performed in a frequency range 3–55 Hz using a sinusoidal wavelet with minimum 3 wavelet cycles per data window. Following this, ERD/ ERS was averaged for the theta (4–8 Hz), alpha (8–12 Hz), and beta (16–24 Hz) frequency bands in four time windows (0.4–0.8s, 0.8–1.2s, 1.2–1.6s, 1.6–2.4s). These frequency bands and time windows were selected on the basis of results reported in our previous study .
Data normalization considered to be one of the main processing steps for learning tasks, and it can be defined as scaling technique that helps finding new range from an existing one range. In this study, Min-Max normalization technique was applied to rescale data to [0, 1] range.
The classification was performed on EEG data recorded in the first phase of this study (BNF). Decision tree(DT), Naïve Bayes(NB), Support vector machine(SV), K-nearest neighbor(KNN), and Artificial Neural Network(ANN) algorithms are employed in this study as classifiers in order to find the most appropriate algorithm to build the predictive model for studying the efficacy of NF training. The classification process was carried out with four different combinations of electrodes which were called as C1, C2 and C3 that included 61, 27 and 9 electrodes respectively as shown in Fig. 3. The combination C1 was chosen because studies reported global effect of pain while C2 was chosen on the basis of definition of NP reported lesion to the somatosensory cortex. The combination C3 were chosen on the basis of somatotopoical representation of upper and lower limbs and cortical shift of paralyzed limbs reported in the literature. In this study, the efficiency of each classifier was evaluated by taking 70% and 30% data for training and testing respectively.
DT algorithm builds a classification tree with nodes, branches and leaves (leaf defined as a single class which doesn’t split any more) based on selected predictors set. The DT classifier in this study was grown using fitctree function that splits categorical predictors applying the exact search algorithm. The number of discriminative predictors and their levels decide the number of the nodes that determines the size of tree. The numbers of nodes were 21, 101, and 43 in case of classification of MI data of RH for three group combinations AB vs PWP, AB vs PNP, and PNP vs PWP respectively. However, in case of data of MI of F, the numbers of nodes were 77, 69, 67 for all three groups’ combinations AB vs PWP, AB vs PNP, and PNP vs PWP respectively. Similarly, the numbers of nodes were 33, 25 and 7 in case of classification of MI data of LH. The reason for choosing DT classifier for this study is its effectiveness in implementation and it has been used for analyzing EEG data in many other studies as well .
The Normal distribution was employed to estimate the distribution of the data since it is the easiest to work with because it is only required to estimate the mean and the standard deviation from training data. Based on this, NB classifier was selected for this study as it is powerful and fast learning algorithm. Moreover, NB is reliable approach to analyzing EEG .
K-NN classifier does not use any model to fit but is only depending on memory, and a new instance is classified based on the closest training samples present in the feature space. When a test data is entered, it is assigned to the class that is most common amongst its k nearest neighbors .The optimal value of k (number of nearest neighbors) in this study was 3; and this value found through experiments. KNN classifier was chosen for this study as it was reported that it gained high accuracy in EEG classification based studies .
SVM classifier in this study was used for binary classification with linear Kernal Function and supports sequential minimal optimization. SVM classifier is used on a large scale in EEG based studies for being powerful classifier based on pattern recognition [36, 37]. Additionally, SVM for binary classification is considered to be more effective for EEG classification based on pattern recognition approach .
The ANN classifier applied in this study as a classifier was selected with the following parameters. The number of neurons in the hidden layer was chosen according to experiments that revealed that the optimal number of neurons in our case of EEG classification is 20 neurons. The number of neurons in the output layer was one, because the ANN classifier was selected to perform binary classification. The network training function was trainscg that updates weight and bias values according to the scaled conjugate gradient method. The ANN trained on back propagation algorithm. ANN classifier was selected for this study based on literature review showed that ANN classifier gained high accuracy in classification of EEG patterns . Moreover, ANN was applied successfully for classification of EEG patterns related to MI .