Scientific Papers

Review of electromyography onset detection methods for real-time control of robotic exoskeletons | Journal of NeuroEngineering and Rehabilitation


Figure 1 presents the number of publications on EMG onset detection methods along the years. After analyzing all papers selected, the pre-processing and EMG onset detection categories were defined according to the methods used and their relevance in terms of papers in the literature that applied each of them.

Fig. 1
figure 1

Number of publications on EMG onset detection methods per year reviewed in this study (period: 1978–2022; total articles: 156)

Pre-processing methods

Pre-processing methods, used to improve the quality of EMG signals towards the extraction of meaningful information, usually add more computational time, which means a delay in real-time implementations. The pre-processing methods evaluated in this review were classified in the following categories: EMG Envelope, Teager-Kaiser Energy Operator (TKEO), Wavelet Transform and Others, which included those that did not fit in none of these categories. Different pre-processing methods were applied 90 times in the papers reviewed. Calculating the EMG envelope was the method most frequently used, followed by the TKEO method.

EMG envelope

According to the CEDE project, EMG envelope is a smooth curve that tracks changes in the amplitude of an EMG signal over time [8]. Calculating the EMG envelope is a pre-processing method that can be obtained in several ways, as shown in Fig. 2.

Fig. 2
figure 2

Comparison of the most common approaches to obtain the EMG envelope from the raw EMG signal

To obtain the EMG envelope from raw signals, two main options are available: (1) low-pass filtering of the rectified signal; (2) root-mean-square (RMS) on raw EMG signal.

Low-pass filtering of the rectified signal One of the most common approaches to calculate the EMG envelope is to use a discrete version of traditional low-pass filters such as Butterworth or Chebyshev on the rectified EMG signal (obtained by computing the absolute value of the raw signal).. These filters can be considered as Infinite Impulse Response filters [15]. This approach was applied in: [16,17,18,19,20,21,22,23,24,25,26], with the Butterworth being the most predominant filter used.

Moving average (MA) According to the CEDE project, MA is defined as a method to smooth EMG data, that acts as a low-pass filter, reducing random fluctuations in the rectified or squared EMG signal [8].

This method was first used in the context of EMG onset detection by Maple-Horvart and Gilbey in 1992 [27]. After that, MA was applied to calculate EMG envelopes for EMG onset detection in several other papers: [13, 28,29,30,31,32,33,34,35,36,37,38,39,40,41].

The MA is calculated with a series of averages from successive segments, with or without overlapping windows. The consequence of its use is the attenuation of rapid variations through local averaging, but retention of slow variations [28], smoothing the signal and acquiring its envelope.

Root-mean-square (RMS) on raw EMG signal This approach ([28, 33, 42,43,44,45,46,47,48,49,50,51]) computes the RMS value of the signal within a window that “moves” across the raw EMG signal.

The RMS value measures the square root of the signal’s power. Therefore, it has a physical meaning. RMS is useful in many other applications [42]. EMG envelopes can be calculated from the RMS according to Eq. 1.

$$\begin{aligned} \begin{aligned} X_{RMS}=\sqrt{\frac{1}{N}\sum _{i=1}^{N}x_i^{2}} \end{aligned} \end{aligned}$$

(1)

where \(x_{i}\) is the EMG value in the \(i^{th}\) sample and N is the number of samples.

Teager–Kaiser energy operator (TKEO)

The TKE operator method ([17, 25, 38,39,40, 48, 52, 53, 53,54,55,56,57,58,59,60,61,62,63,64,65,66,67]) was first proposed by Teager in 1982 [68,69,70]. The results obtained in these studies suggested that the production of speech involved nonlinear processes. As a result, Teager derived the TKE operator in the discrete-time domain to compute the energy of a sound. This method has been extended to cover other continuous signals such as EMG [53].

The discrete TKE operator \(\psi\) is defined in the time domain as:

$$\begin{aligned} \begin{aligned} \psi _{d}[x(n)]=x^{2}(n)-x(n+1)x(n-1) \end{aligned} \end{aligned}$$

(2)

where n is the sequence index and x the raw EMG signal. Considering a signal defined by Eq. 3:

$$\begin{aligned} \begin{aligned} x(n)=A cos[\omega _{0}(n)+\theta ] \end{aligned} \end{aligned}$$

(3)

where A is the amplitude, \(\omega _{0}(n)\) is the angular frequency, and \(\theta\) is the initial phase, the energy operator can be rewritten as defined in Eq. 4:

$$\begin{aligned} \begin{aligned} \psi _{d}[x(n)]\approx A^{2}sin^{2}(\omega _{0}) \end{aligned} \end{aligned}$$

(4)

Equation 4 shows that the TKEO is proportional to the instantaneous amplitude (A) and frequency (\(\omega _{0}\)) of the input signal. Therefore, TKEO is usually applied on EMG signals to extract motor unit activity by making the action potential spikes sharper and narrower, enhancing the muscle activation points [53].

Several studies have demonstrated that pre-processing using TKEO can improve the EMG onset detection with respect to different pre-processing methods [17, 52, 53, 57, 71].

Wavelet transform (WT)

Pre-processing of raw EMG using the wavelet transform was applied in the following papers: [32, 48, 52, 59, 62, 64, 67, 72,73,74,75,76,77,78].

The WT is one of many time-frequency representations used in signal processing. These transforms deconstruct a time domain signal into a sum of signals of different scales and time shifts, to produce a time-frequency representation of a time domain signal. WT is an effective tool to extract useful information from the EMG signal.[79].

Other pre-processing methods

The other pre-processing methods found in the literature were the Hilbert filter [9, 80,81,82], the Kalman filter [83], the Morphological Close Operator [38, 55], the Morphological Open Operator [38], the Multi Objective Optimization Genetic Algorithm [84], the Adaptative Linear Energy Detector [85], the use of an statistical criterion based on the amplitude distribution of EMG signal [86], the Constant False Alarm Rate method [87] and the Empirical Mode Decomposition [82].

EMG onset detection methods

EMG onset detection methods are those that, when applied to the EMG signal (raw or pre-processed signal), allow the identification of the beginning of muscle activation. In the pasts, the onset of muscle activation could be detected using mainly the following methods: Visual inspection, Threshold-based and Statistical. Recently, other studies have tested different methods (especially Machine Learning based) to determine muscle activation onset, reporting promising results. In our study, all EMG onset detection methods that do not fit into none of the previously mentioned categories were classified as ”Other EMG onset Detection Methods”. Figure 3 shows the number of papers that applied each of these categories within each different application domain (Robotics, Clinical, Research and Others). EMG onset detection has been applied in the application domain ’Research’ more than in all the other domains together.

Fig. 3
figure 3

Number of publications included in each of the different EMG onset detection categories (Visual inspection—black; Threshold—light gray; Statistical—dashed grey; Machine Learning—bold gray squares; and Others—dashed bold gray) within each different application domain. Application domains considered were Robotics, Clinical, Research and Others

Visual inspection

Visual inspection entails subjectivity and needs to be performed by an expert. There are no criteria established on how to carry out the visual inspection technique, although it is usually employed to detect the earliest rise in EMG activity above the steady-state (i.e., basal activity) [50, 88,89,90,91,92,93,94].

Despite being a subjective technique, visual inspection can be used to validate automatic EMG onset detection methods, serving as a gold-standard to develop computerized EMG onset detection methods. The visual detection of EMG onset has been widely referred in the literature: [11, 17, 21, 24,25,26, 32, 33, 36, 43, 49, 52, 57, 60, 60, 73, 95,96,97,98,99,100,101,102].

Threshold-based methods

The label ’THRESHOLD’ in Fig. 3 encompasses the use of one or more threshold-based methods, which are thoroughly described in this section, in each of the papers analyzed in this review. Threshold-based are the most common EMG onset detection methods found in the literature, being tested 253 times across the 156 papers analyzed (i.e., several papers tested and/or compared more than one different method based on EMG threshold).

In this approach, one attributes a threshold to discriminate between baseline activity and muscle activation. Thresholding is widely used due to its simplicity, speed and reliability. The simplicity of thresholding lies in its straightforward implementation. Additionally, thresholding is computationally efficient, making it suitable for real-time analysis of EMG signals and handling large datasets. Regarding its reliability, thresholding is a robust method that has been used in numerous studies in the literature. Although thresholding may not always provide the most accurate detection of EMG onset, it remains a popular choice in EMG signal processing. Nonetheless, there is lack of agreement among researchers on a standardized threshold method for EMG onset detection [46].

Cavanagh et al. were the first to propose the use of a threshold-based method [103]: authors investigated the dependence of electromechanical delay in the human elbow flexor group upon selected initial conditions at the time of muscle activation. Most common strategies followed to set threshold values are based on the baseline amplitude characteristics of the EMG signal, such as the mean or standard deviation. Some researchers name this strategy as the Shewhart protocol [16, 30, 104].

Some of the signal characteristics that can be considered to select the threshold are the following:

Threshold-based methods can be classified in three different categories: Single Threshold (ST), Double Threshold (DT) and Adaptive Threshold (AT).

Single threshold (ST) ST method is the most predominant EMG onset detection method found in the literature: [2, 11, 13, 16,17,18,19, 22,23,24, 27, 29,30,31,32, 35,36,37,38,39,40,41, 43, 45,46,47,48, 51,52,53, 55,56,57, 59, 60, 63, 66, 67, 74, 77, 81,82,83,84, 86, 95, 96, 99, 103, 105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133].

ST compares the amplitude of the EMG signal (raw or EMG envelope) with a previously selected threshold. The onset is detected when the EMG amplitude is bigger than the threshold.

This method can be considered the most intuitive and standard computer-based method of time-locating the onset of muscle contraction activity [2].

ST can be useful to overcome some of the problems related to visual inspection. However, results of applying ST strongly depend on the choice of the threshold [134], which can lead to false positives in noisy signals. In theses cases, it is advisable to work on the EMG envelope, which smooths the signal and improves the onset detection.

Double threshold (DT): The DT method was applied in the following studies: [2, 10, 19,20,21, 24, 33, 34, 42, 51, 55, 65, 73, 82, 83, 87, 94, 95, 100, 115, 119,120,121, 123, 127, 128, 135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151].

To overcome some of the problems associated with ST, Lidierth et al. introduced the DT method in 1986 [135]. This method adds a second threshold to determine the muscle activation onset time, with the final goal of avoiding false positives and enhance EMG onset detection precision. A common strategy when applying DT method is to define an amplitude threshold, similar to what is done in ST. If the signal amplitude is higher than this threshold for a certain amount of time or samples (second threshold), then muscle activation is detected with DT. Due to the stochastic characteristics of the EMG signal, it is normally necessary to use a pre-processing method to obtain the signal envelope and then apply the second threshold.

Adaptive threshold (AT) The AT method can be applied directly on the raw EMG signal. AT segments the signal using the signal-to-noise (SNR) [61] or the energy value [85] to adapt the threshold of muscle activation by windows. As the SNR is the relative power of wanted EMG to unwanted signal components that are contained in the overall signal [8], this threshold method can be considered as an improvement of ST method, as it adapts its threshold value according to the EMG window being analyzed, which might enable a more precise EMG onset detection over time. AT was applied in the following works: [11, 52, 54, 55, 85, 97, 118, 121, 152,153,154,155,156].

Statistical methods

The onset of muscle activation can be detected by evaluating the statistical properties of the EMG signal before and after a possible change in model parameters [115]. Two main statistical approaches can be identified in the literature: the Approximated Generalized Likelihood Ratio (AGLR) and the Cumulative Sum (CUSUM).

Approximated generalized likelihood ratio (AGLR): The AGLR method was applied in the following publications: [10, 17, 37, 52, 57, 58, 67, 72, 77, 83, 93, 94, 100, 115, 121, 128, 157,158,159,160].

Sometimes also referred to as “Maximum Likelihood Estimator”, this method was first proposed as a change detection algorithm, with its first use in the context of muscle activity detection being presented in Hogan et al., in 1980 [157]. In short, the AGLR algorithm calculates an estimate of muscle activity as a function of the mean and variance of the activity level [121].

By using a log-likehood ratio test g(k) [66], the AGLR method detects if there is muscle contraction or not.

The log-likelihood ratio test is defined by the following equation:

$$\begin{aligned} \begin{aligned} g(k)=ln\left( \prod _{k=1}^{r} \frac{p1(Y_{n}|H_{1})}{p0(Y_{n}|H_{0})}\right) \end{aligned} \end{aligned}$$

(5)

where ln represents the natural logarithm, Y(n) represents the series of EMG samples, k the index of the product, r is the total length of the series, p1 and p0 represent the probability density function corresponding to the alternative hypothesis \(H_{1}\) (i.e., there are changes in the statistical properties of the EMG sequence) and the null hypothesis \(H_{0}\) (i.e., there are no changes in the statistical properties of the EMG sequence), respectively.

If the log-likelihood g(k) value is smaller than a pre-defined threshold, it indicates that the muscle is relaxed, whereas EMG onset is detected if g(k) value exceeds the threshold.

Cumulative sum (CUSUM): This method was used in: [67, 72, 95, 109]. CUSUM was first proposed by Ellaway in 1978 [161] with applications on the analysis of histograms.

The first study to propose the use of CUSUM to detect EMG onset was Chanaud et al., in 1991 [109], which used this method to determine how the different regions of the biceps femoris activated in a cat during a broad range of limb movements.

The CUSUM method works as follows [161]: a reference level (k), dependent on the task to be performed and selected in a previous training phase, is subtracted from each of the series of points on the signal (x1, x2, …, xi, …, Xn). The result of these subtractions, shown in Eq. 6, is a new series of points (Si) which are formed by adding up these differences consecutively.

$$\begin{aligned}S1 &= (x1-k) \\S2 &= (x1-k)+(x2-k) \end{aligned}$$

(6)

The CUSUM chart is defined as the sequential plot of the values of Si, expressed by the Eq. 7:

$$\begin{aligned} \begin{aligned} Si = \sum (xi-k) \end{aligned} \end{aligned}$$

(7)

The CUSUM technique has a smoothing action on the data [161] and the EMG onset detection is determined by a previous threshold, which can be defined by a training phase (see [72] for more details).

Other statistical methods Other statistical methods were also used in the following papers: [25, 49, 50, 60, 61, 75, 145, 147, 160, 162, 163].

Machine learning methods

Machine learning (ML), which is a discipline within the field of Artificial Intelligence, has recently gained increasing popularity due to the ability to extract patterns and information from complex and high-dimensional datasets. In the context of EMG onset detection, ML algorithms can automatically learn and adapt to the characteristics of the EMG signals, enabling the development of highly accurate and efficient detection methods. Machine learning-based algorithms were found in [33, 65, 164, 165].

Di Nardo et al. [165] evaluated a novel machine-learning-based approach (DEMANN) for detecting muscle activation onset/offset timing from sEMG signals. The study trained a neural network and evaluated DEMANN’s performance on simulated and real sEMG signals. DEMANN was validated against different reference algorithms, including the DT method. The study found that DEMANN provided a reliable prediction of muscle onset/offset and was minimally affected by SNR variability.

Trigili et al.[164] presented a ML-based algorithm able to detect users’ motion intention based on EMG signals and assessed its applicability towards the control of an upper-limb exoskeleton for people with severe arm disabilities. The algorithm was able to detect the onset of muscle activation before the actual movement, and its computational load was compatible with real-time applications. The study concluded that the proposed algorithm was promising for controlling upper-limb exoskeletons in real-time applications, and for assisting people with severe arm disabilities in performing functional tasks.

Dow et al. [33] presented the development of an algorithm to detect inspiratory events from EMG signals. A state-machine was utilized for classification and inspirations were detected with  98% accuracy in anesthetized and awake rats. The proposed algorithm can be explored in humans, as it may be useful for individuals requiring assisted ventilation.

Ghislieri et al. [65] introduced and validated a new approach to detect muscle activation intervals from sEMG signals using long short-term memory (LSTM) recurrent neural networks. The performance of the proposed LSTM-based muscle activity detector was compared with two other widely used approaches: TKEO and DT method. The study included simulated and real sEMG signals from healthy individuals, orthopedic patients, and neurological patients. Results showed that the LSTM outperformed the other approaches. The proposed algorithm overcomes the main limitations of other tested approaches and works directly on sEMG signals without the need for background-noise and SNR estimation.

Despite the growing interest in using ML techniques for EMG onset detection, there is currently no consensus on a reference method. As noted by Di Nardo et al. [165], a standardized approach for evaluating and comparing the performance of ML algorithms for EMG onset detection is still lacking. While several studies have proposed different ML approaches and achieved promising results, the absence of a reference method makes it challenging to classify these methods.

Other EMG onset detection methods

Other methods can be classified as: Energy-based methods [34, 85, 92], Entropy-based methods [12, 25, 38, 125, 146], Mathematical/numerical techniques [9, 16, 55, 166, 167], Computer Vision [127], Slope / discontinuities detectors [51, 62, 76] and External stimulation [168, 169].



Source link