Subjects
Eleven healthy subjects (7 male and 4 female; Mean age = 27.4 (±2.55) years; Mean weight = 68.7 (±3.56) kg; and Mean height = 169.3 (±6.2) cm) volunteered to participate in this study. The healthy participants exclusion criterion was; (i) no history of myo or neuropathology, and (ii) no evident abnormal motion restriction. The technique was tested on one amputee participant (Female; Age = 42 Years) who volunteered to take part in this study. Amputee volunteer had trans-radial one-third proximal amputation of the right forearm (with 9 cm long stump).
SEMG recording procedures
Bipolar electrodes (DELSYS, Boston, MA, USA) were placed on the forearm (FDS) muscle for the healthy participants (Figure 1a) in accordance with standard procedures [29] to record surface electromyogram (sEMG). These are active electrodes, with the preamplifier and two electrodes built into a single unit. The electrodes are self-adhesive, and have two silver bars; each of 1 mm thickness, 10 mm length and the fixed inter-electrode distance of 10 mm. Electrolyte Gel (Sigma) were used on the electrodes prior to affixing them on the skin.
Experiments were repeated twice for the able-bodied volunteers, with electrodes placed on the proximal end in the first experiment- EXP1, and on the distal end for the second experiment – EXP2 (Figure 1a). This was done to determine the effect of variation in the electrode location on the outcome of the experiments. The ground electrode was placed on the volar aspect of the wrist. For the amputee participant, the electrodes were placed on the remaining stump of the participant as shown in Figure 1b.
DELSYS (Boston, MA, USA) sEMG acquisition system was used to record the signal. The system gain was 1000, CMRR was 92 dB, and bandwidth was 20–450 Hz, with 12 dB/ octave roll-off. The input impedance of the system was 115 Pico-farad in parallel with 1 K-ohm. The sampling rate of the system was 1024 samples/ second for each channel and the resolution were 16 bits/ sample. Prior to the placement of electrodes, the skin of the participant was prepared by shaving (if required) and exfoliation to remove dead skin. Skin was cleaned with 70% v/v alcohol swab to remove any oil or dust from the skin surface.
Experimental protocol
Experiments were conducted where the sEMG was recorded while the participants performed four sets of generic finger flexion actions (Figure 2) labelled and described below.
-
Background activity: All fingers resting.
-
Class 1: Flexion of little (pinkie) finger
-
Class 2: Flexion of ring finger
-
Class 3: Flexion of middle finger
-
Class 4: Flexion of index finger
These generic actions were selected for the following reasons:
-
these actions will allow the user to control individual fingers in the recently available advanced robotic/ prosthetic hand [7, 10, 19] and utilise these devices to the maximum advantage.
-
these actions can be used for communication and control commands that can be used for several different devices and applications.
The participants performed the flexion without any resistance and as was convenient and easily reproducible by them. SEMG was recorded through the experiment. The examiner gave on-screen and oral commands to the participant to perform the action without any fixed order of the fingers. Each flexion was maintained for about 7–8 seconds and was repeated twelve times. The experiments were repeated twice with changed electrode location.
The experimental protocol was approved by the RMIT University Human Ethics Committee and Alfred Health Ethics Committee and performed in accordance with the Declaration of Helsinki 1975, as revised in 2004. Prior to the recording, the participants were encouraged to familiarize themselves with the experimental protocol and with the equipment. For the experiment with the amputee, bilateral action training modality was performed [30]. The amputee subject performed the finger flexions with the healthy hand while attempting to repeat the same flexion with the phantom limb [30].
Data analysis
Removal of noise and background activity
As a first step, the signal corresponding to class 1 for each subject on each day was normalised to the root mean square (RMS) of the recording of the same subject. The next step was the removal of the background activity from the signal (Figure 3). This becomes more challenging because when signal strength is low, the noise magnitude becomes comparable with the signal itself.
The pilot analysis of the recordings showed large inter-experimental variations in the spectrum of the background activity, making stationary spectral filtering unsuitable because of the variations in the noise characteristics [31]. For this reason, adaptive spectral subtraction filters were used to remove the background activity from the recorded signal. The template of the spectrum of the noise was obtained for each subject using a bandpass filter (4th order Butterworth, frequency 10 to 450 Hz) and averaging over 20 windows (>20).
Decomposing sEMG to obtain relative strength of contraction
The first step of the proposed method required the identification of the temporal location of the MUAPs. For this purpose, sEMG signal was decomposed using wavelet transform (bi-orthogonal wavelet ‘bior3.3’) and the maxima were identified based on the change of sign [32].
Local maxima, Wf(s,x
n
) can be described mathematically as follows [32]:
(1)
Where Wf(s,x) is the wavelet transform of the function f(x) at a scale s and n = 2 to N-1, N is the total number of coefficients at any given scale s.
The wavelet maxima that were present in each of the scales and travelled from finest scale to coarsest scale were considered [33, 34]. Other wavelet maxima were rejected as being random transients. The data of each flexion was segmented into 300 ms windows, and the magnitudes at the lowest (finest) scale of the accepted maxima were obtained. Based on the magnitude, the magnitude and number of maxima points of every windowed recording (density of peaks) corresponding to each flexion were the two dimensional representation of the recording. These were separated into four (number of different actions) groups using cluster analysis. The centroid and corresponding density of each cluster was determined for each time segment, and averaged for the duration of each flexion. This resulted in one value of the magnitude and the corresponding density of peaks for each action. This two dimensional feature set was used to train and test the system.
Classification
Twin support vector machines (TSVM) linear kernel classifier [35] was used to classify the features. The advantages of using TSVM is that it solves two related SVMs, one for each class, and generates two separate hyper planes without the assumption that patterns in each class arise from similar distributions. It allows the use of a different kernel for each class which can be separately optimized based on the data. This data dependent kernel optimization for each class is particularly useful in this study. For more details, please refer to [15, 35].
The vectors corresponding to the four densities and four magnitudes were the input to the classifier. During the training phase the associated finger flexion was the target. After the system was trained, the system accuracy was validated using ten-fold cross – validation and tested using Type I error (Specificity) and Type II error (Sensitivity) [15]. The training dataset had 100 data points, and testing was done using 30 data points. The training data set and test data set was chosen randomly using random sub-sampling method. Ten-cross validation method was used to determine the accuracy, sensitivity and specificity of the system. Classification was performed on the data from the individual subject. The system was tested individually for the two experiments to determine the inter-experimental variations.