Data collection
A custom rig was used to measure hand force and torque exerted by the volunteers. The rig (see Figure 1) consisted of a force sensor (Futek LCM-300) which measured contraction force. This sensor was placed between two plastic halves, which formed together a semi-sphere to enable the volunteers to comfortably hold the rig with their hand. These two plastic halves were connected to a metallic platform through a torque sensor (Transducer Techniques TRT-100) that recorded torque produced by the volunteer while performing ulnar or radial deviation movements.
Guidelines presented in the sEMG for the non-invasive assessment of muscles (SENIAM) project [33] were followed to obtain a fine skin contact with the electrodes. According to these guidelines, the skin was cleaned with an alcohol swab and electrodes were placed at the locations shown in Figure 2. sEMG electrodes were attached to the volunteers' forearms using medical adhesive bands that made the electrodes' active faces adhere the skin.
sEMG signals were recorded from the following four muscles in order to detect movement of wrist and fingers [34]: Extensor Digitorum (ED), Palmaris Longus (PL), Flexor Carpi Ulnaris (FCU) and Extensor Carpi Radialis(ECR). Function of each muscle is summarized in Table 1. sEMG signals were acquired through a Noraxon system (Myosystem 1400L). A data acquisition board from National Instruments (USB-6289) was used in this study for acquiring both the sEMG signals and the data obtained from the custom rig used to measure hand force and torque. Since the EMG signal has usable energy in the 0-500 Hz range [35], the acquired sEMG signal was digitized at 1024 samples per second and stored on a computer through an application developed in LAbVIEW software. The developed LabVIEW application also had a graphical interface to enable volunteers visualizing force they were exerting during the tests. For each participant, the maximum force exerted to the rig was used to define the participant's maximum voluntary contraction (MVC). According to [36], the applied force should not exceed 40-50% of the MVC in order to prevent upper extremity musculoskeletal injuries. For this reason, all the protocols were defined to prevent exceeding this limit.
Protocol
12 seniors (70 years old on average) and 7 young volunteers (27 years old on average) participated in this study. The Office of Research Ethics, Simon Fraser University approved this study and each senior signed a consent form. Each volunteer followed the eight predefined protocols summarized in Table 2. These protocols were defined to simulate simple activities of daily living involving the wrist and fingers such as opening and closing the screw cap of a jar or grasping an object. The identified protocols considered a combination of several hand movements including grasping, finger pinching, wrist ulnar/radial deviation and forearm pronation/supination. Each volunteer started at rest position as shown in Figure 3-a.
In protocol A, as shown in Figure 3-b, the volunteer was asked to squeeze the custom rig with maximum force in pronation position of the arm for two times. The recorded maximum force was used to define MVC for squeezing.
In protocol B, as shown in Figures 3c-d, the volunteer was asked to apply maximum torque in ulnar and radial deviation for two times (pronation position of the arm). Maximum torques for ulnar and radial deviations were used to identify ulnar/radial MVCs.
In protocol C, the volunteer was asked to squeeze the custom rig at 50% of her/his MVC for 5 seconds (pronation position of the arm). The volunteer repeated this protocol three times. Using the graphical interface of the developed LabVIEW application, the volunteer had visual feedback for the force applied to the custom rig.
In protocol D, the volunteer was asked to alternate radial and ulnar deviation for 5 seconds at 50% of MVC (pronation position of the arm). The volunteer repeated this procedure three times.
In protocol E, as shown in Figures 3e-h, the volunteer pinched the force sensor firstly with thumb and index finger, secondly with thumb and middle finger, thirdly with thumb and ring finger, and finally with thumb and little finger (pronation position of the arm). The pinching was repeated two times for each combination of fingers.
In Protocols FC, FD and FE (see Figures 4a-h), each volunteer started at rest position and repeated protocols C, D and E but with their arm in supinated position. Figure 5 presents the output recorded by the force and torque sensors for one of the volunteers following protocols A, B, C and D. Figure 6 presents a sample output of the force and torque sensors related to protocols E, FC, FD and FE.
Protocols A and B (see Table 2) were followed to record the maximum torque produced by the user. Protocols C, D, E, FC, FD, and FE were instead used to generate data for the formation of the different hand gesture classes summarized in Table 3. Specifically, protocols C, D and E enabled extracting data for classification purpose in the pronation position of the arm (classes 2-8 in Table 3) whereas protocols FC, FD and FE were used to extract data for classification in the supination position of the arm (classes 9-15 in Table 3).
Feature extraction and classification
The proposed sEMG signal classification scheme is presented in Figure 7. As shown in this figure, signals recorded from the Noraxon measurement system were processed in MATLAB R2009a for feature extraction in order to reduce the dimensionality of the raw sEMG input.
Pattern recognition accuracy is influenced by the selection of extracted features and features cannot be extracted from the individual samples as the structural detail of the signal is lost [37]. In fact, the features need to be calculated by segmenting the raw sEMG signal and calculating a set of features from each segment. For this reason, the recorded data was segmented into 250 ms intervals corresponding to 256 samples in each segment and features were extracted from each segment. Then, for the next feature extraction, the segment window was incremented by 125 ms including 128 samples.
Waveform length, time windowed RMS and AR models were used to extract six features for each of the four sEMG channels. Specifically, waveform length and RMS provided one feature each, whereas AR models provided four features in total as explained in the following paragraphs.
The waveform length, which measures the waveform complexity in each segment, was computed as:
(1)
where t
r
is the amplitude of the rth sample and N is the number of samples.
The time windowed RMS value of the raw sEMG signal was used in order to provide information regarding the amplitude of the signal. This feature is mathematically presented as:
(2)
where m
i
is the amplitude of the ith sample in the time domain, and n is the number of samples. In our case n was equal to 256.
The last feature used in this study was based on AR models. AR models can be defined as a linear combination of previous samples and noise. The mathematical representation of current value is given by (3):
(3)
where w is the additive noise and {q for i = 1, ..., p } are AR model coefficients. Four AR model coefficients were selected as adequate for modelling EMG signals as discussed in [38].
Six seconds of data per person per protocol was extracted. In order to train and test the pattern recognition model, the gathered data was divided into training and testing sets (see Figure 7) [39]. The testing set was limited to 3807 data segments, namely 90% of the gathered data, as the use of a higher number of segments did not significantly improve the classification accuracy. The remaining 10% of the gathered data, corresponding to 423 data segments, was used as testing set.
SVM [40] was chosen as classifier in this study. SVM was selected among all the other possible pattern recognition tools, as it is a well-known robust classifier, which has extensively and successfully been used to process bio-information signals [41–43]. In addition, SVM works well in high dimensional spaces and has shown good classification results in many practical applications [44–49].
In its general formulation, the SVM [40] requires solving the following optimization problem:
(4)
where w is the vector representing adaptive model parameters, c>0 is the penalty factor, N is the total number of data points, a
n
is the label associated with a data point, ξ
n
is the slack variable, z is the learned model, x
n
is the vector representing a data point, and n is the index associated to a data point.
In this study, the LibSVM tool [50] was used in MATLAB R2009a environment. LibSVM has an implementation for multi class SVM using one-versus-one strategy, whose details are presented in [51]. The LibSVM supports well-known kernels such as the radial basis function (RBF), polynomial, sigmoid and Gaussian kernels.
Following guidelines presented in [52], the RBF was selected as it nonlinearly maps the samples and has limited numbers of hyper parameters thus reducing the complexity of model selection. The mathematical representation of the RBF kernel is:
(5)
Eight fold cross validation along with grid search was used to select the pattern recognition optimal parameters c and γ. Figure 8 shows an illustrative example of results obtained for a single participant. It can be seen that the cross validation accuracy does occur in the interval (0,100) for c and (0,3) for γ. This interval was selected for the identification of the optimal parameters for all participants.