1. Introduction
Lower limb amputation, whether due to accidents or illnesses such as diabetes, peripheral vascular disease, and gangrene [
1], can be traumatic both for the patient and caregiver. Lower limb amputation affects physical, psychological, emotional, and social aspects of individuals’ lives, especially activities related to the human leg, which is an important limb. The use of lower limb prosthesis for amputation can help increase life quality for patients. Previous studies showed higher scores in quality of life for amputation patients who used prostheses than patients who did not use prosthetics in the physical, psychological, and environmental domains [
2]. Amputation patients in Indonesia, where this study was conducted, mainly used conventional, low-cost, below-limb prosthetics [
3]. These conventional lower limb prosthetics commonly consist of fixed and passive structures where the ability to walk is still possible, although patients must overcome certain difficulties [
4]. This study mentions that the ability to successfully walk using prosthetics was an important factor contributing to amputation patients’ quality of life [
5]. Thus, it is very important to develop active lower limb prosthetics with ergonomic considerations to increase amputation patients’ quality of life. Motorized lower limb prosthetics can be achieved by using a bionic lower limb, incorporating a motor as the driving force. Motor control using signal measurement from electromyography (EMG) has been widely used for upper limb prosthesis. On the contrary, EMG control for bionic lower limb prosthesis is still in the early stage of research due to the practical use of motorized robotic lower limbs [
6].
This research was a preliminary study on the development of a bionic below-limb prosthesis. An EMG sensor called Myomes was used to record research participants’ muscle signals while walking on a treadmill at different speeds. EMG signals were obtained, then separated into three gaits and classified using the ANN machine learning method. Six training algorithms were used in this ANN method for gait classification. The preliminary results of this study will help researchers understand how the human gait works, and the best training method and classification results. In a future project, the best training method will be used for amputation patients training for bionic below-limb prosthetic customization. Patient signals will be recorded using Myomes and then fed back into the motor.
Research related to EMG signal recording and classification has been carried out extensively using different variations, methods, and purposes, i.e., rehabilitation and EMG data collections for various gaits [
7,
8,
9,
10,
11,
12]. Some of these studies recorded below-limb EMG signals statically where the participants were not moving or walking [
7,
8]. The current study contributes to the literature on EMG measurements that were conducted while participants were walking at various speeds on a treadmill. Another study [
9,
10,
12] collected EMG signals where all the participants were of the male gender. This research collected EMG samples from both male and female genders. A recent study noted that EMG recording activities were conducted during a ground-walking experiment in five environments, i.e., flat ground, upstairs, downstairs, uphill, and downhill [
12]; another study recorded EMG signals while participants walked on level ground [
11]. This study contributes to the knowledge of EMG signal recordings where participants walk in a controlled environment in which a treadmill runs at different speeds. Most of the previous research classified gait into two classes: stance and swing phase [
7,
8,
9,
10,
11,
12]. This study classified gait into three classes. Several methods were applied to human gait classification, including deep learning [
11,
12], utilizing an ANN combined with Levenberg–Marquardt as a training algorithm. This study classified other training algorithms (quasi-Newton method, Bayesian regulation backpropagation, gradient descent backpropagation, gradient descent with adaptive learning rate backpropagation, and one-step secant backpropagation), indicating that different algorithms result in different levels of accuracy.
3. Feature Extraction and Classification Method
The feature extraction step is carried out utilizing 12 features: 8 time-domain features and 4 frequency-domain features. The time domain features include the integrated EMG (IEMG), mean absolute value (MAV), the variance of EMG (VAR), root mean square (RMS), log detector (LOG), waveform length (WL), kurtosis, and skewness. The frequency-domain features include mean frequency (MNF), median frequency (MF), total power (TTP), and mean power (MNP).
Data collection videos were played and the raw EMG signal was segmented into three signals based on the selected phases. An illustration of the three phase categories of gait cycle for ANN classification is presented in
Figure 3 [
26]. Many previous studies adopted two phases of human gait partition, i.e., stance and swing phase [
7,
8,
9,
10,
11,
12]. However, other studies adopted a larger number of phases, such as three [
27,
28,
29,
30,
31], four [
32], five, and more [
33,
34,
35]. Future work will investigate the possible use of a motor control feedback signal for bionic below-limb prosthetics in which the EMG signals are segmented into three classes [
27].
Figure 4 shows the segmented EMG Myomes signals into three categories, i.e., (i) Initial contact, which is labeled as initial gait, (ii) Loading response to the terminal stance which is labeled as mid−gait and (iii) Preswing to terminal swing which is labeled as final gait.
Figure 4a presents the segmented EMG signal on initial contact phase. It represents a gait cycle initiation and contact point that the gravity body center is maximum in this phase which shows an existence of the muscle activity. In addition,
Figure 4a shows the Myomes EMG signal amplitude of approximately 18%.
Figure 4b presents the segmented EMG signal from loading response to a terminal stance where the foot plantar surface touched the treadmill. There were minimal activities to the muscle due to the single limb stance. The highest EMG activities were recorded from preswing to terminal swing which amplitude approximately of 35% as presented in
Figure 4c. The segmentation EMG signal is also based on previous research [
27] that used an adaptive controller for active ankle foot prosthetics signals.
In ANN gait classification, the EMG data are divided as follows: 70% for training, 15% for validation, and 15% for testing. The present study used two layers: a feed-forward network with a tangent−sigmoid transfer function for the hidden layer and a softmax transfer function for the output layer, as presented in
Figure 5.
According to
Figure 5, the hidden layer can be presented in Equation (1), as follows:
where
a1 denotes the output vector in the hidden layer, x is an n-length input vector, and
IW denotes the weight matrix input layer. The transfer function of the input layer and hidden layer bias vector is stated in
tf1 and
b1. Equation (2) illustrates the output layer of the first output neuron.
where
a2 is the hidden layer output vector,
LW is weight matrix output layer.
There are six training methods used as the training algorithm, i.e., Levenberg−Marquardt back propagation with the second-order training speed approach, the quasi-Newton method, Bayesian regulation backpropagation, gradient descent backpropagation, gradient descent with adaptive learning rate backpropagation and one-step secant backpropagation. The prediction error was calculated using mean square error (MSE). A smaller MSE value indicates that the ANN classification accuracy is higher.
Figure 6 illustrated the ANN model with 12 features calculated as input, 25 neurons in the hidden layer, and 3 outputs which represent each class, i.e., the initial gait, mid-gait and final gait.
5. Discussion
Gait classification and pattern recognition was conducted with different machine learning methods and different types of sensor. Some research used gait analysis to detect people with gait disorders [
36], understand the biomechanics of muscle for pre-disease diagnosis, and design lower limb prosthetics [
37]. The present study analyzed human gait with the aim of obtaining a better understanding and a better design for lower limb prosthetics. Various sensors were used as data acquisition kits, such as accelerometers [
36,
37,
38]. In the present study, an EMG meter named Myomes was used to collect the participant data.
Table 6 summarizes previous research in EMG signal classification compared to the present study. Most studies using gait signals from EMG measurements classified the signals into two classes using various machine learning methods [
10,
11]; meanwhile, another study classified human gait into eight classes using only one machine learning method, linear discriminant analysis (LDA), with Bayesian information criteria as the optimization method [
39]. The present study classified human gait into three classes using ANN with various training algorithms to find which training algorithm had the best result. A comparison of previous research in
Table 6 could be a consideration in future improvements to the research.
The paper [
11] showed that ANN, with a number of architectural variations, affected the system accuracy. This was consistent with this study, where variations in the training algorithm affected the system accuracy, although both studies utilized different variations in the experiment. Paper [
11] showed the accuracy of unseen subjects (US) consisting of a single subject and learned subject (LS), with 22 unlearned subjects. According to the accuracy results, LS subjects have a better value, at 94.49% (further information regarding this accuracy result can be found in Ref. [
11]). This is also consistent with this study, where overall accuracy was higher than each subject’s accuracy and both studies suggest that walking patterns might be unique, as presented in
Table 6.
Table 6 also shows that a number of machine learning methods have been used in EMG pattern recognition studies. The machine learning model with fewer classes has higher accuracy compared to those with a higher number of classes.
Study [
10] showed higher accuracy than previous research [
11]; however, the research [
10] only classified below-limb EMG signals into two classes. A further consideration of this study could be to improve data in the future by embedding a number of muscles where EMG signals are recorded. Another machine learning method could also be considered. However, the ANN has been proven as a potential embedded machine learning method for hardware implementation. This is the main reason why the ANN is selected in present study. The future work of the present study is to develop an embedded system in which the ANN will be embedded into a certain microcontroller for bionic below-limb prosthesis. In previous research, the EMG signals were segmented into eight classes where the human gait was identified [
39]. This will be taken into consideration for future studies, as more classes for human gait might result in a better ergonomic design.
A previous study concluded that the ANN training algorithm’s determination depends on many factors, such as the complexity of problems, amount of data in the training set, weight and biases in the networks, the error goal, and machine learning purposes, i.e., for pattern recognition or approximation [
40].
Table 5 showed that the Levenberg−Marquardt backpropagation training algorithm resulted in the best and higher accuracy among other training algorithms, at 96% overall accuracy. This result is consistent with a previous study [
41] that concluded that the Levenberg−Marquardt backpropagation training algorithm showed good potential in gait mechanics’ estimation. The Levenberg−Marquardt backpropagation training algorithm performed well with a low weight because convergent results can be obtained in a short time and a lower error rate can be generated [
40].
ANN training algorithm variations were used in other studies, as presented in
Table 7.
Table 7 shows a comparison between the other studies that used the ANN training algorithm variation and the present study. According to
Table 7, it can be concluded that different datasets and classes result in different accuracies for ANN classification. Further details of the classification accuracy can be found in the references presented in
Table 7.
Table 7 shows the ANN method machine learning method used in various dataset, including non-biomedical datasets such as car evaluation [
42] and biomedical datasets, i.e., Parkinson’s, cardiotography I, cardiotography II [
42], length of stay (LOS) for hospitalized patients with COVID-19 [
40], and heart disease patients [
43].
Some research [
40,
42] found that the Bayesian regulation backpropagation training algorithm has the highest accuracy. The results related to how the Bayesian regulation backpropagation training algorithm updates the weights and biases in ANN networks using the Levenberg−Marquardt backpropagation training algorithm as optimisation method. The Bayesian regulation backpropagation training algorithm leads to a high-generalizability model [
40] that is suitable for function approximation and did not perform well in classification or pattern recognition. This finding is consistent with the results of the present study, where the Bayesian regulation backpropagation training algorithm had a lower accuracy than other training algorithm, at 89% overall accuracy, which can be seen in
Table 5.
Previous research [
43] is consistent with the present study, and study [
41] reached the same conclusion, that the Levenberg−Marquardt backpropagation training algorithm performed best when used in the ANN model for classification and pattern recognition, although study [
43] used a different dataset to the present study and study [
41].
The Levenberg−Marquardt backpropagation training algorithm used a second-order training speed without Hessian matrix computation so that the algorithm is fast and efficient. However, the Levenberg−Marquardt backpropagation training algorithm is not suitable for a very large number of data. One of the setbacks of the Levenberg−Marquardt backpropagation training algorithm is that the algorithm needs a large amount of storage for the derived matrices, which can be quite large for some complex problems and data [
43]. To date, the Levenberg−Marquardt backpropagation training algorithm has performed well in pattern recognition [
41,
43].
In the future, this study will be used as the basis for the development of a robotic lower limb prosthesis conducted in CBIOM3S, Universitas Diponegoro, in collaboration with a neurosurgeon affiliated with Dr. Kariadi General Hospital, Central Java, Indonesia. The signal results for muscle strength measurements using Myomes will be recorded to train the robotic lower limb prosthesis user and customize them to create an ergonomic product. Comfortable prosthesis may increase amputees’ quality of life.
6. Conclusions
A study on EMG signal pattern recognition, acquired from three general lower limb human gait movements, was presented. A two-layer feed-forward ANN network was selected as the machine learning method to classify three classes. Three classes, i.e., initial contact, which is labeled as initial gait, loading response to the terminal stance, which is labeled as mid-gait, and pre-swing to terminal swing, which is labeled as final gait, were associated with the segmented EMG signals. The ANN algorithm used and compared six training accuracy methods, i.e., the Levenberg−Marquardt backpropagation training algorithm, quasi−Newton training method, Bayesian regulation backpropagation training method, gradient descent backpropagation, gradient descent with adaptive learning rate backpropagation and one-step secant backpropagation. The machine learning study performs well, classifying three classes of human walking gait with an overall accuracy (training, testing, and validation) of 96% for Levenberg−Marquardt backpropagation, respectively. According to the accuracy results, Levenberg-Marquardt backpropagation outperformed other methods. This result can be used as a preliminary study for lower-limb prosthesis to improve its ergonomic factor. In future works, other machine learning methods will be examined and compared with the presently proposed method.
Human gait classification using EMG signals, as presented in this study, has potential application as a form of motor control related to robotic lower limb prosthesis. Gait classification and EMG signal measurement can be used to develop a customized lower limb prosthesis for each subject, as people walk at different speeds.
Future works related to this research will include another means of gait measurement using myomes, such as ease of walking, climbing up and down the stairs and walking on an uphill terrain. This study will also evaluate myomes using the future myomes channel and filter, aiming to perfect these so they can be used as a measurement tool for robotic lower limb prosthetics. These have been developing in CBIOM3S.