Feasibility of using an artificial neural network model to estimate ... - kaist

Journal of Neuroscience Methods 194 (2011) 386–393

Contents lists available at ScienceDirect

Journal of Neuroscience Methods journal homepage: www.elsevier.com/locate/jneumeth

Feasibility of using an artificial neural network model to estimate the elbow flexion force from mechanomyography Wonkeun Youn 1 , Jung Kim ∗ School of Mechanical, Aerospace and Systems Engineering, Department of Mechanical Engineering, Korea Advanced Institute of Science and Technology, 373-1 Guseong-dong, Daejeon 305-701, Republic of Korea

a r t i c l e

i n f o

Article history: Received 26 June 2010 Received in revised form 13 October 2010 Accepted 8 November 2010 Keywords: Mechanomyogram Artificial neural network Force estimation Elbow flexion Multiple linear regression Cross-subject validation test

a b s t r a c t The goal of this study was to demonstrate the feasibility of using artificial neural network (ANN) models to estimate the elbow flexion forces from mechanomyography (MMG) under isometric muscle contraction and compare the performance of the ANN models with the performance from multiple linear regression (MLR) models. Five participants (mean ± SD age = 25.4 ± 2.96 yrs) performed ten predefined and ten randomly ordered elbow flexions from 0% to 80% maximal voluntary contractions (MVCs). The MMG signals were recorded from the biceps brachii (BR) and brachioradialis (BRD), both of which contribute to elbow flexion. Inputs into the model included the root-mean-square (RMS), a temporal characterization feature, which resulted in a slightly higher signal-to-noise ratio (SNR) than when using the mean absolute value (MAV), and the zero-crossing (ZC) as spectral characterization features. Additionally, how the RMS and the ZC as model inputs affected the estimation accuracy was investigated. A cross-subject validation test was performed to determine if the established model of one subject could be applied to another subject. It was observed that the ANN model provided a more accurate estimation based on the values of the normalized root mean square error (NRMSE = 0.141 ± 0.023) and the cross-correlation coefficient (CORR = 0.883 ± 0.030) than the estimations from the MLR model (NRMSE = 0.164 ± 0.030, CORR = 0.846 ± 0.033). The estimation results from the same-subject validation test were significantly better than those of the cross-subject validation test. Thus, using an ANN model on a subject-by-subject basis to quantify and track changes in the temporal and spectral responses of MMG signals to estimate the elbow flexion force is a reliable method. © 2010 Elsevier B.V. All rights reserved.

1. Introduction The surface of the skin above a contracting muscle will experience small oscillations that are generated by the lateral dimensional changes of the active muscle fibers. Mechanomyography (MMG) is the recording of these oscillations, which reflect the mechanical activities of the contracting muscle (Barry and Cole, 1990; Orizio, 1993). It has been reported that MMG may reflect three main physiological origins: (a) the gross lateral movement of the contracting muscle at the initiation of the muscle contraction, (b) smaller, subsequent oscillations occurring at the resonant frequency of the muscle, and (c) dimensional changes in the active fibers (Barry and Cole, 1990; Orizio, 1993; Orizio et al., 2003). Over the past two decades, the relationship between MMG and the corresponding force have been widely examined for various muscles (Beck et al., 2004a,b; Herda et al., 2009a,b; Orizio, 1993;

∗ Corresponding author. Tel.: +82 423503231; fax: +82 423505230. E-mail addresses: [email protected] (W. Youn), [email protected] (J. Kim). 1 Tel: +82 423504543; fax: +82 423504540. 0165-0270/$ – see front matter © 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.jneumeth.2010.11.003

Orizio et al., 1989). In their studies, it was hypothesized that MMG is reflective of two force production mechanisms: the recruitment and the firing rates of motor units (MUs). Electromyography (EMG) is also influenced by these two mechanisms. If the relationship between the MMG signals and the force was linear, an appropriate analytical approach could be easily obtained (Frank, 1966). Unfortunately, several studies have reported that the summation of the mechanical contributions of the recruited MUs in the MMG signal is nonlinear throughout the entire range of discharge frequencies (Orizio et al., 1996). Additionally, the relationship between the MMG signals and the corresponding forces can be highly nonlinear and complicated (Beck et al., 2005). For example, the MMG signal amplitude displays a nonlinear relationship because it first increases with force generation up to a certain maximum voluntary contraction (MVC) rate (from 60% to 80% MVC) and then plateaus or decreases to 100% MVC, depending on the type of muscle used during the concentric/eccentric muscle action (Akataki et al., 2004; Orizio et al., 2003). Furthermore, although several approaches have been investigated to describe the MMG amplitude versus the exerted force/torque relationship, a large variability exists in the results for different muscles (Maton

W. Youn, J. Kim / Journal of Neuroscience Methods 194 (2011) 386–393

387

et al., 1990; Orizio et al., 1989), and no quantitative models have been established to accurately describe the complex, nonlinear relationships between the MMG signals and force to estimate the exerted force. Therefore, we propose a novel method to estimate the elbow flexion force using MMG through an artificial neural network (ANN) model. The ANN model is capable of modeling complex, nonlinear relationships between input and output (Zhang and Eddy Patuwo, 1998) and has been outstanding in recognizing characteristics of non-stationary biological signals (Nigg and Herzog, 1999). Furthermore, unlike a biomechanical model (e.g., Hill’s muscle model), an ANN model, which acts as a black box, does not require a complex mathematical model of the underlying musculoskeletal system. Additionally, the ANN model is computationally efficient and has been applied to a wide range of diverse, real-time applications (Barniv et al., 2005; Chu et al., 2007). In particular, the ANN model has been extensively used to estimate the muscle force from EMG signals, resulting in satisfactory estimation performance (Arslan et al., 2009; Choi et al., 2010; Hahn, 2007; Song and Tong, 2005). Therefore the ANN model might be a reliable tool to establish a relationship between the MMG signals and the corresponding force.

izontal plane. The shoulder movement of the subject was restrained by shoulder braces. The elbow was stabilized at a 90◦ flexed position, and the forearm remained in a neutral position. The wrist was fixed with a metal plate to the force sensor, and the hand remained open. A one-directional force transducer (651AL, Maximum force 20 kgf, KTOYO Inc., Republic of Korea), which was coupled with a metal frame attached to the subject’s wrist, was used to measure the elbow flexion force. The value of the force was displayed on the screen to provide visual feedback to the subject in real time. The MMG was detected with two accelerometers (ADXL202JE, Analog Devices, USA), fixed to the skin surface above the muscle belly of the biceps brachii (BR) and brachioradialis (BRD) with double-sided foam tape, as shown in Fig. 1. The accelerometer is a dual-axis device that can measure the oscillation from two orthogonal directions (x and y) with a 3 dB bandwidth from 0 to 6 kHz. Since the belly of the brachialis (BRA) muscle, which also contributes to elbow flexion, lies just below the biceps brachii, it was not accessible to the accelerometers; thus, measurement of the BRA muscle was excluded.

2. Materials and methods

Each subject visited the laboratory twice, once for a familiarization session and once for the experimental trial. The familiarization session allowed the subjects to become comfortable with the experimental equipment and to practice the isometric experimental protocols. After the familiarization session, the subjects were asked to performed maximal voluntary contraction (MVC) for 5 s, and strong verbal encouragement was provided. The highest value of the measured force was recorded at a contraction level of 100% MVC. In each experiment, there were ten successive trials; each trial consisted of ANN training and ANN testing, which included a guided session and a non-guided session for validation. The subjects could relax between each trial to avoid muscular fatigue, which could affect the MMG signal in the subsequent trial. The subjects were not allowed to talk or move their body during the experiments to avoid motion artifacts. During ANN training, each subject was required to track their force generation on the computer screen in front of them. The screen displayed the digitized force signal 10 times over a period of 140 s in real time. The levels of the target force consisted of static levels (20%, 40%, 60%, and 80% MVC) with ramped muscle contraction (10% MVC/s) and dynamic levels that were generated by a ramp function with three different contraction velocities (10%, 20%, and 30% MVC/s), as shown in Fig. 2. The force trajectory was pro-

2.1. Subjects The experiment was conducted on five healthy, young, right-handed male subjects (mean ± SD age = 25.4 ± 2.96 yrs; stature = 175.8 ± 2.49 cm; mass = 67.6 ± 3.58 kg). The relatively small range in age of the male subjects was intended to minimize the potential, confounding effects of age and gender because there have been reports of age- and gender-related changes from MMG during force generation (Akataki et al., 2002; Nonaka et al., 2006). All of the participants were free of neuromuscular and musculoskeletal pathology. All of the subjects were given sufficient information about the purpose and procedures, and before participation, informed consent was obtained from each subject. The KAIST Institutional Review Board approved the proposed experimental protocol of this study. 2.2. Research design The setup used for conducting the experiment on the isometric elbow flexion is illustrated in Fig. 1. The subject was asked to sit comfortably in a chair with his right upper limb positioned on a hor-

Fig. 1. Experimental setup.

2.3. Experimental procedures

Fig. 2. Guided target force trajectories for the ANN model training.

388

W. Youn, J. Kim / Journal of Neuroscience Methods 194 (2011) 386–393

grammed using the Matlab software (version 7.0, The MathWorks, USA). A combination of the static level and dynamic level with different velocities was employed to consolidate the feasibility of the model in various experimental conditions. The ANN testing part was divided into two consecutive sessions; the guided session used the same force trajectory as training to observe if the established ANN model could accurately estimate the target force trajectory. The non-guided session was the random force trajectory, where subjects were allowed to exert any force trajectory of their own choosing for 100 s. The non-guided session was included to observe if the ANN model could successfully estimate an untrained force trajectory.

Although the root-mean-sqaure (RMS) and the mean-absolutevalue (MAV) processing have been frequently used (Cescon et al., 2004, 2007, 2008) for temporal characterization features of MMG, no quantitative comparison study of the signal-to-noise ratio (SNR, where noise is defined as the variability in the signals according to (Clancy and Hogan, 1999)) between the RMS and the MAV, has been conducted. In contrast, SNR has been used to evaluate the quality of the EMG processing, and it has been reported that MAV processing results in a higher SNR than RMS signal processing by 2.0–6.5% for amplitude estimation (Clancy and Hogan, 1999). Thus, we performed the elbow flexion experiment under isometric muscle contraction as described previously during 10, 20, 30, and 40% MVC over a period of 7 s, and the RMS and the MAV processing of the MMG were computed for every 100-ms segment step length. The equations for the RMS and the MAV processing can be defined as follows:

N

k

(1)

k=1

1 |xk |, for k = 1, ..., N, N N

MAV =

ZC =

N 

(2)

k=1

where xk are the raw signals of each segment, and N is the number of samples. The results are illustrated in Fig. 3. For the entire MVC contraction, the RMS had a slightly higher SNR than the MAV (P < 0.05). A higher SNR value indicates less variation about the same mean value of the amplitude estimate, which will lead to a greater accuracy the force estimation model. Therefore, the RMS was selected as the input for the model. For the spectral characterization feature, in past studies, the mean frequency (MNF) and the median frequency (MDF) were usu-

sgn(−xk xk+1 ), for k = 1, ..., N

(3)

k=1

 sgn(x) =

2.4. MMG feature selection

  N−1 1 RMS =  x2 , for k = 1, ..., N

ally used to describe the spectral features of MMG (Cescon et al., 2004; Madeleine et al., 2002). However, due to the computationally heavy transformation of MNF or MDF, which requires a Fourier transformation, the ZC was selected to describe the spectral feature because it needs only a relatively simple calculation and has similar characteristics of the MNF and the MDF (Hägg, 1991). The equations for the ZC can be defined as follows:

1 if x > 0 , 0 otherwise

where xk are the raw signals of each segment, and N is the number of samples. The ZC is the number of times that the signal passes the zero amplitude axis (Zecca et al., 2002). 2.5. Signal processing The MMG signals were band-pass filtered (5–100 Hz) using the designed analog filter, and the force signal was low-pass filtered with a cutoff frequency of 10 Hz in a commercially available signal conditional amplifier (ST-AM 100, Senstech Inc., Republic of Korea). The MMG and the force signals were recorded simultaneously with a data acquisition system (NI 6063E, National Instrument, USA). The MMG and force signals were then sampled at 1000 Hz. The sampling frequency was adequate for obtaining sufficient information of the MMG signals because the most relevant information was within the range of 5–100 Hz. The recorded MMG signals and force were used to construct the model. The MMG signals were digitally fully-rectified and the features (i.e., RMS and ZC) were extracted and then normalized with respect to the MVC value. The normalized features were used as inputs into the model. In the feature extraction procedure, the RMS and the ZC values were used because they represent the power and spectral changes in the signals, respectively, and are indications of the muscle activation level (Fig. 4). 2.6. Artificial neural network An artificial neural network (ANN) with three feed-forward layers (an input layer, a hidden layer, and an output layer) was used to construct the model describing the MMG and the corresponding force relationship. The inputs into the model were the RMS and the ZC values of the MMG signals, which were obtained from the two selected muscles (BR and BRD). The output was the exerted force during the isometric elbow flexion. To improve the

Fig. 3. Signal-to-noise ratio (SNR) from the RMS and the MAV processing during 10, 20, 30, and 40% MVC. *P < 0.05; **P < 0.001; NS: not significant.

W. Youn, J. Kim / Journal of Neuroscience Methods 194 (2011) 386–393

389

Fig. 4. Activities of the two contributed muscles (BR and BRD) during isometric elbow flexion from subject C. (A) MMG, (B) normalized RMS, (C) normalized ZC, and (D) exerted elbow flexion force.

estimation performance of the ANN model, the output force was normalized by the MVC and ranged between 0 and 1 (Arslan et al., 2009). A linear summation function and a hyperbolic tangent sigmoid function were employed as the neural transfer functions. The number of hidden layers was determined based on the universal approximation theorem (Haykin, 2008), which states that a single hidden layer is sufficient to guarantee network convergence. The number of hidden neurons was determined adaptively by testing various numbers of neurons until satisfactory results were obtained (Arslan et al., 2009). The optimum number of hidden neurons was determined for each subject, which varied from 8 to 13 for the different subjects. An epoch number of 200 was chosen for the learning stage. The back propagation algorithm was used as the training algorithm using the Levenberg–Marquart method (Levenberg, 1944; Marquardt, 1963) because it provides the fastest convergences for medium-sized neural networks of up to hundreds of neurons (Song and Tong, 2005). The algorithm can be represented as follows: w = w + w T

w = [J J + I] e = R − z,

(4) −1 T

J e

(5) (6)

where w is the weight vectors, w is the difference between the weight vectors, J is the Jacobian matrix that contains the first deriva-

tives of the network errors with respect to the weight,  is a scale parameter, I is the identity matrix, R is a vector of the measured force, z is a vector of the estimated force, and e is a vector of the network errors. The ANN model was designed and implemented using the Matlab Neural Network Toolbox (version 7.0, The MathWorks, USA). 2.7. Performance evaluation with the multiple linear regression model Because a nonlinear relationship has been commonly observed between the MMG and the force for various muscles under isometric muscle contraction (Akataki et al., 2003, 2004; Ebersole et al., 1998; Orizio et al., 2003), linear regression may not be a good approach. Although several approaches, including polynomial regression using linear, quadratic, and cubic models (Herda et al., 2009a) and log-transformed regression (Herda et al., 2009b), have been investigated, the quality of the regression model varied greatly, and only temporal features (i.e., RMS) were considered in the model. Alternatively, the ANN model is known to be a flexible, mathematical structure that is capable of establishing complex, nonlinear relationships between input and output data (Hsu et al., 1995). Thus, in this study, we sought to determine whether using an ANN model that used temporal and spectral characterization features (i.e., RMS and ZC) is a reliable method that can model the nonlinear aspects of a complicated biological mechanism.

390

W. Youn, J. Kim / Journal of Neuroscience Methods 194 (2011) 386–393

Fig. 5. Mean and standard deviations of the force estimation results in the guided session and non-guided session by the ANN model from MMG with RMS + ZC, RMS, and ZC inputs.

To validate this claim, a simulation study was performed to observe which model, the ANN model or the multiple linear regression (MLR) model, resulted in a better estimation. The MLR models were constructed using MMG data from both the BR and BRD muscles and the force data obtained from the experiments. A one-way ANOVA test with a significance level of P < 0.05 was conducted to compare the normalized root mean square (NRMSE) and the crosscorrelation coefficient (CORR) between the estimated force and the measured force of the ANN and the MLR models. The NRMSE and the CORR can be defined as follows:

 (

NRMSE = CORR =

(x1,i − x2,i )2 /N)

(7)

xmax − xmin

  N

N



  x1,i x2,i − x1,i x2,i    

x21,i − (

x1,i )

2

N

x22,i − (

, x2,i )

(8)

2

where x1,i , x2,i , and N indicate the measured force, the estimated force, and the total number of data points, respectively; the subscript i indicates the ith data point. 2.8. ANN and MLR model with three different inputs (RMS + ZC, RMS, and ZC) In general, if more relevant characterization features are included in the model, the accuracy of the force estimation is expected to improve. However, little is known about how the temporal and spectral representative characterization features (i.e., RMS and ZC values) affect the accuracy of the models. Thus, the roles of the RMS and ZC values in the accuracy of the force estimation were investigated. Three models were compared: one model used the RMS value, another model used the ZC value, and the other model used both the RMS and the ZC values as the inputs into the ANN models. The same procedure was repeated for the MLR models. 2.9. Cross-subject validation A cross-subject validation study was conducted to explore whether the ANN model constructed for one subject could be used to estimate the exerted elbow flexion force of other subjects. Each subject performed the ten trials, and the measured MMG signals and the force were used in the cross-subject validation study. First, for each subject, ten ANN models were constructed using ten training datasets. Then, each ANN model was tested using 40 datasets

composed of 10 datasets from 4 subjects who had not trained with the ANN model. The same ANN model was also tested using ten testing datasets from one subject who had trained with the ANN model for the evaluation comparison. The same procedure was repeated for the MLR models.

3. Results The estimation results of the ANN and the MLR model with three different inputs (i.e., RMS + ZC, RMS, and ZC) for the guided and the non-guided session are shown in Figs. 5 and 6. In Fig. 5, although the NRMSEs and the CORRs of the ANN model with RMS + ZC inputs were less than and greater than, respectively, than those of the ANN model with RMS inputs, the differences were not statically significant (P > 0.05). The estimated results of the ANN model with the RMS input were greater than the ANN model with the ZC input (P < 0.001). In Fig. 6, the estimation results of the MLR model with RMS + ZC inputs were greatest among the three inputs (P < 0.001), and the estimation performance of the MLR model with the RMS the RMS input was greater than that of the MLR model with the ZC input (P < 0.001). To verify the feasibility of the ANN model versus the MLR model, the average performance of five subjects based on the ANN model with RMS + ZC input was evaluated by comparing them with those based on the MLR model. Fig. 7 summarizes the results, which includes the average NRMSEs and CORRs of the five subjects of the non-guided and guided sessions, and the total. In Fig. 7, the average NRMSEs and CORRs between the estimated force and the measured force were 0.137 ± 0.023 and 0.885 ± 0.041 for the ANN model, respectively, and 0.159 ± 0.025 and 0.849 ± 0.036 for the MLR model, respectively. The NRMSEs and the CORRS of the ANN model were statistically less than and greater than, respectively, than those of the MLR model, indicating that the ANN model resulted in greater accuracy than the MLR model for all five subjects (P < 0.001). Tables 1 and 2 summarize the estimation results by the ANN and the MLR model of the cross-subject validation test (the training and testing data were from different subjects) along with the samesubject simulation performance (the training and testing data were from the same subjects). For all the subjects, the estimation of the same-subject validation was significantly better than that of crosssubject validation (P < 0.05).

W. Youn, J. Kim / Journal of Neuroscience Methods 194 (2011) 386–393

391

Fig. 6. Mean and standard deviations of the force estimation results in the guided session and non-guided session by the MLR model from MMG with RMS + ZC, RMS, and ZC inputs.

Fig. 7. Mean and standard deviation of the force estimation results in the guided session and non-guided session by the ANN and the MLR model from MMG with RMS + ZC inputs.

Table 1 Performance comparison of the ANN model for the same-subject validation and cross-subject validation. Significant differences from the same-subject validation are indicated as follows: *P < 0.05; **P < 0.001; NS: not significant. Values are reported as the mean ± standard deviation. Subjects

Same-subject validation NRMSE

S1 S2 S3 S4 S5 Total

0.130 0.201 0.145 0.167 0.172 0.163

± ± ± ± ± ±

Cross-subject validation CORR

0.017 0.074 0.021 0.070 0.054 0.058

0.904 0.858 0.898 0.844 0.801 0.863

NRMSE ± ± ± ± ± ±

0.021 0.059 0.023 0.091 0.102 0.076

0.296 0.282 0.439 0.233 0.301 0.310

± ± ± ± ± ±

CORR 0.136** 0.104** 0.130** 0.064** 0.100** 0.130**

0.836 0.782 0.696 0.801 0.749 0.773

± ± ± ± ± ±

0.054** 0.130** 0.155** 0.068** 0.142** 0.126**

Table 2 Performance comparison of the MLR model for the same-subject validation and cross-subject validation. Significant differences from the same-subject validation are indicated as follows: *P < 0.05; **P < 0.001; NS: not significant. Values are reported as the mean ± standard deviation. Subjects

Same-subject validation NRMSE

S1 S2 S3 S4 S5 Total

0.148 0.219 0.182 0.170 0.186 0.181

± ± ± ± ± ±

Cross-subject validation CORR

0.015 0.059 0.037 0.053 0.046 0.051

0.870 0.819 0.847 0.855 0.810 0.841

NRMSE ± ± ± ± ± ±

0.022 0.028 0.029 0.030 0.033 0.036

0.233 0.531 0.474 0.221 0.298 0.354

± ± ± ± ± ±

CORR 0.064** 0.355** 0.264** 0.047** 0.151** 0.240**

0.829 0.801 0.832 0.820 0.795 0.816

± ± ± ± ± ±

0.037** 0.054** 0.037** 0.042** 0.082* 0.053**

392

W. Youn, J. Kim / Journal of Neuroscience Methods 194 (2011) 386–393

4. Discussion The goal of this study was to demonstrate the feasibility of using ANN models to estimate the elbow flexion forces from MMG under isometric muscle contraction and compare its performance with the performance from MLR models. The experimental results demonstrated that the ANN model with a back-propagation learning algorithm can be used to estimate the elbow flexion force accurately from MMG under isometric muscle contraction. When compared with the estimation performance of the MLR model, the ANN model results were significantly more accurate. These results suggest that using an ANN model may be a more reliable method to establish the complex, nonlinear relationship between the MMG and the corresponding force than the MLR model. In this study, we also demonstrated that the estimation accuracy of the cross-subject validation test was significantly less than that of the same-subject validation test. This finding indicates that the ANN and the MLR model should be constructed on a subject-bysubject basis. This may be due to many factors, such as different fiber type compositions, muscle morphologies, and the physical milieu, such as muscle stiffness, intra-muscular pressure, and intramuscular temperature according to the individual (Orizio, 1993), which affects the temporal and spectral responses of the MMG during the force generation. For all the subjects, the model with the RMS input resulted in a better estimation performance than that with the ZC input. This might be because the increase in the RMS value with increasing force was more pronounced than that in the ZC value and also, could possibly be due to the large oscillation with the same corresponding force found with the ZC value. Although the ANN model with both RMS and ZC inputs does not show a statistical improvement against the ANN model with the RMS input, the MLR model with RMS and ZC inputs resulted in a significantly better performance. This experimental result indicates that the spectral feature (i.e., ZC) may provide additional quantitative information about force generation. Therefore, in light of this study’s findings, one might consider combining both the temporal and spectral characterization features to establish a model between the MMG and force. The limitations of the present investigation included that only the established models from BR and BRD were verified, and the experimental setup in this study was limited to elbow flexion under isometric muscle contraction. Hence, future studies should examine a wide range of possible movement observed in daily activities with various muscles, other machine learning methods, such as support vector machines, and other temporal and spectral characterization features to further investigate and improve the estimation performance. To the best of our knowledge, no quantitative evaluation of an ANN model that establishes a nonlinear relationship between the MMG and the corresponding force versus the MLR model has been investigated. Moreover, no quantitative evaluation of the RMS and the MAV signal processing of the MMG based on the SNR has been conducted, though they have been frequently used. In addition, the cross-subject validation tests were conducted to investigate if the model should be established on a subject-by-subject basis. Also, the roles of temporal and spectral characterization features were investigated. Therefore, this study has made significant progress in terms of demonstrating the feasibility of an ANN model versus the MLR model from MMG to estimate the elbow flexion force and showing that the RMS processing resulted in a slightly higher SNR than the MAV processing. This study has also made a significant improvement by suggesting that the established model should be a subject-specific model and that the spectral characterization feature of MMG can be inputs into the model.

Conflict of interest None. Acknowledgements This research was supported by the Happy tech. program through the National Research Foundation of Korea (NRF) funded by the Ministry of Education, Science and Technology (No. 20100020449). References Akataki K, Mita K, Watakabe M. Electromyographic and mechanomyographic estimation of motor unit activation strategy in voluntary force production. Electromyography and Clinical Neurophysiology 2004;44:489. Akataki K, Mita K, Watakabe M, Ito K. Age-related change in motor unit activation strategy in force production: a mechanomyographic investigation. Muscle & Nerve 2002;25:505–12. Akataki K, Mita K, Watakabe M, Itoh K. Mechanomyographic responses during voluntary ramp contractions of the human first dorsal interosseous muscle. European Journal of Applied Physiology 2003;89:520–5. Arslan Y, Adli M, Akan A, Baslo M. Prediction of externally applied forces to human hands using frequency content of surface EMG signals. Computer Methods and Programs in Biomedicine 2009. Barniv Y, Aguilar M, Hasanbelliu E, Branch H, Aeronaut N, Adm S, et al. Using EMG to anticipate head motion for virtual-environment applications. IEEE Transactions on Biomedical Engineering 2005;52:1078–93. Barry D, Cole N. Muscle sounds are emitted at the resonant frequencies of skeletalmuscle. IEEE Transactions on Biomedical Engineering 1990;37: 525–31. Beck T, Housh T, Cramer J, Weir J, Johnson G, Coburn J, et al. Mechanomyographic amplitude and frequency responses during dynamic muscle actions: a comprehensive review. Biomedical Engineering Online 2005;4:67. Beck T, Housh T, Johnson G, Weir J, Cramer J, Coburn J, et al. Mechanomyographic amplitude and mean power frequency versus torque relationships during isokinetic and isometric muscle actions of the biceps brachii. Journal of Electromyography and Kinesiology 2004a;14:555–64. Beck T, Housh T, Johnson G, Weir J, Cramer J, Coburn J, et al. Mechanomyographic and electromyographic time and frequency domain responses during submaximal to maximal isokinetic muscle actions of the biceps brachii. European Journal of Applied Physiology 2004b;92:352–9. Cescon C, Farina D, Gobbo M, Merletti R, Orizio C. Effect of accelerometer location on mechanomyogram variables during voluntary, constant-force contractions in three human muscles. Medical and Biological Engineering and Computing 2004;42:121–7. Cescon C, Madeleine P, Farina D. Longitudinal and transverse propagation of surface mechanomyographic waves generated by single motor unit activity. Medical and Biological Engineering and Computing 2008;46:871–7. Cescon C, Madeleine P, Graven-Nielsen T, Merletti R, Farina D. Two-dimensional spatial distribution of surface mechanomyographical response to single motor unit activity. Journal of Neuroscience Methods 2007;159:19–25. Choi C, Kwon S, Park W, Lee H, Kim J. Real-time pinch force estimation by surface electromyography using an artificial neural network. Medical Engineering & Physics 2010. Chu J, Moon I, Lee Y, Kim S, Mun M. A supervised feature-projection-based real-time EMG pattern recognition for multifunction myoelectric hand control. IEEE/ASME Transactions on Mechatronics 2007;12:282–90. Clancy E, Hogan N. Probability density of the surface electromyogram and its relation to amplitude detectors. IEEE Transactions on Biomedical Engineering 1999;46:730–9. Ebersole K, Housh T, Johnson G, Evetovich T, Smith D, Perry S. The effect of leg flexion angle on the mechanomyographic responses to isometric muscle actions. European Journal of Applied Physiology and Occupational Physiology 1998;78:264–9. Frank R. Use of transformations. Journal of Marketing Research 1966;3:247–53. Hägg G. Comparison of different estimators of electromyographic spectral shifts during work when applied on short test contractions. Medical and Biological Engineering and Computing 1991;29:511–6. Hahn M. Feasibility of estimating isokinetic knee torque using a neural network model. Journal of Biomechanics 2007;40:1107–14. Haykin S. Neural networks: a comprehensive foundation. Prentice Hall; 2008. Herda T, Housh T, Weir J, Ryan E, Costa P, DeFreitas J, et al. Mechanomyographic amplitude versus torque relationship. Physiological Measurement 2009a;30:115–28. Herda T, Weir J, Ryan E, Walter A, Costa P, Hoge K, et al. Reliability of absolute versus log-transformed regression models for examining the torque-related patterns of response for mechanomyographic amplitude. Journal of Neuroscience Methods 2009b;179:240–6. Hsu K, Gupta H, Sorooshian S. Artificial neural network modeling of the rainfallrunoff process. Water Resources Research 1995;31:2517–30.

W. Youn, J. Kim / Journal of Neuroscience Methods 194 (2011) 386–393 Levenberg K. A method for the solution of certain nonlinear problems in least squares. Quarterly of Applied Mathematics 1944;2:164–8. Madeleine P, Farina D, Merletti R, Arendt-Nielsen L. Upper trapezius muscle mechanomyographic and electromyographic activity in humans during low force fatiguing and non-fatiguing contractions. European Journal of Applied Physiology 2002;87:327–36. Marquardt D. An algorithm for least-squares estimation of nonlinear parameters. Journal of the Society for Industrial and Applied Mathematics 1963;11:431–41. Maton B, Petitjean M, Cnockaert J. Phonomyogram and electromyogram relationships with isometric force reinvestigated in man. European Journal of Applied Physiology and Occupational Physiology 1990;60:194–201. Nigg B, Herzog W. Biomechanics of the musculo-skeletal system. Wiley; 1999. Nonaka H, Mita K, Akataki K, Watakabe M, Itoh Y. Sex differences in mechanomyographic responses to voluntary isometric contractions. Medicine & Science in Sports & Exercise 2006;38:1311. Orizio C. Muscle sound: bases for the introduction of a mechanomyographic signal in muscle studies. Critical Reviews in Biomedical Engineering 1993;21: 201.

393

Orizio C, Gobbo M, Diemont B, Esposito F, Veicsteinas A. The surface mechanomyogram as a tool to describe the influence of fatigue on biceps brachii motor unit activation strategy. Historical basis and novel evidence. European Journal of Applied Physiology 2003;90:326–36. Orizio C, Liberati D, Locatelli C, De Grandis D, Veicsteinas A. Surface mechanomyogram reflects muscle fibres twitches summation. Journal of Biomechanics 1996;29:475–81. Orizio C, Perini R, Veicsteinas A. Muscular sound and force relationship during isometric contraction in man. European Journal of Applied Physiology and Occupational Physiology 1989;58:528–33. Song R, Tong K. Using recurrent artificial neural network model to estimate voluntary elbow torque in dynamic situations. Medical and Biological Engineering and Computing 2005;43:473–80. Zecca M, Micera S, Carrozza M, Dario P. Control of multifunctional prosthetic hands by processing the electromyographic signal. Critical Reviews in Biomedical Engineering 2002;30:459. Zhang G, Eddy Patuwo B. Forecasting with artificial neural networks: the state of the art. International Journal of Forecasting 1998;14:35–62.