A method for decomposing static surface electromyography signals

Abstract

The present invention discloses a method for decomposing static surface electromyography (sEMG) signals. Aiming at the problem that the traditional sEMG decomposition method based on blind source separation cannot extract small motor units (MUs), the present invention decomposes MUs by combining multiple strategies. The present invention first uses an electromyography acquisition device to collect the electrical signals of the biceps brachii of the subject, performs preprocessing and channel extension; then accurately decomposes large MUs through cyclic CKC, and combines the improved cyclic CKC method with a strict iterative strategy (Post-Processor) and a "stripping" strategy to decompose small MUs. Compared with previous methods, the present invention greatly improves the decomposition yield and accuracy of MUs, and can also obtain good decomposition results in a high-noise environment, with strong robustness.

Description

Technical Field

[0001] The present invention relates to a method for decomposing static surface electromyogram signals, specifically a method for obtaining muscle neuron discharge information by decomposing static surface electromyogram signals. Background Art

[0002] In the past few decades, surface electromyography (sEMG) has received extensive attention due to its non-invasiveness, high yield of motor units (MUs), and application to high contraction levels of muscle force. Through significant advances in sEMG signal acquisition and processing, this technology has played a key role in understanding the neurophysiology of the neuromuscular system and the diagnosis of movement nerve diseases and neuromuscular diseases. The sEMG signal is the sum of active motor unit action potentials (MUAPs). sEMG decomposition is a technique that can decompose the sEMG signal into individual MUAPs and is essential for the study of MU discharge information and MUAP waveforms.

[0003] Researchers have made great efforts and proposed various decomposition algorithms. Previously, pattern recognition was one of the most commonly used sEMG signal decomposition techniques. By combining wavelet transform and ART network classification, Gazzoni et al. proposed an automatic decomposition algorithm for MUAP waveform detection and recognition in 2004. De Luca et al. extended their knowledge-based artificial intelligence framework to sEMG signals, which was originally developed for intramuscular electromyogram signal decomposition but was limited to low contraction force cases. Later, Nawab et al. improved this method in 2010 to make it applicable to high contraction force cases. Despite these efforts, the low-pass filtering effect caused by the separation of muscles from sEMG electrodes by skin and subcutaneous fat will lead to serious signal aliasing problems, inevitably greatly reducing the shape differences between MUAPs. All these factors together pose a major challenge to sEMG signal decomposition techniques based on pattern recognition.

[0004] To address the above technical challenges, sEMG signal decomposition techniques based on blind source separation (BSS) have been developed. Holobar and his colleagues developed a decomposition method based on convolutional kernel compensation (CKC) and stood out from the crowd. This method views the sEMG signal as a convolutional mixing model and uses the LMMSE framework to complete the decomposition work. Y Ning et al. proposed the KmCKC algorithm, which employs K-means clustering (KMC) to estimate the firing times of MUs by clustering the observation vectors in different time periods and reconstructs the optimal innervation pulse train (IPT) through continuous iteration. MQ Chen et al. developed a new framework for decomposing sEMG signals. This framework uses multi-constrained parameter FastICA to accurately extract the firing times of MUs and adopts a peeling strategy to avoid the local convergence problem of FastICA. The main limitation of the BBS method lies in that the ability to identify MUs is mainly determined by the energy of their action potentials: almost all MUs with high MUAP energy can be reliably extracted, while MUs with low-energy MUAPs are regarded as physiological noise and cannot be reliably separated. However, small MUs are of great significance for the diagnosis and research of diseases. For example, small MUs can supplement the library of representative single motor unit potentials (SMUPs) required by new motor unit number estimation (MUNE) techniques, which are key electrophysiological parameters for motor neuron diseases or neuromuscular diseases. In stroke patients with spastic hemiplegia, small MUs can also enhance the effect of local botulinum toxin injection on electrophysiological parameters. Summary of the Invention

[0005] In view of the fact that in the prior art, MUs with low-energy MUAPs are regarded as physiological noise and cannot be reliably separated, the present invention proposes a method for decomposing static surface electromyography signals, including the following steps:

[0006] Step 1. Use a device to collect static surface electromyography signals;

[0007] Step 2. Preprocess the signals;

[0008] Step 3. Perform cyclic CKC decomposition on the preprocessed signals to obtain the initial discharge sequences of MUs, and divide the discharge sequences of MUs into good sequences and poor sequences according to the index;

[0009] Step 4. Use spike detection for good sequences and use Post-Processor for bad sequences to obtain spike sequences;

[0010] Step 5. Use peak-triggered averaging technology to estimate the MUAP waveform and delete it from the static surface electromyography signal;

[0011] Step 6. Repeat steps 1-5 until no new MUs are generated.

[0012] Preferably, in step 3, the preprocessed signal is subjected to cyclic CKC decomposition, and the discharge sequences of MUs are divided into good sequences and poor sequences according to the indexes. The specific steps are as follows:

[0013] 3-1. Calculate the covariance matrix of the static surface electromyogram signal and its inverse matrix

[0014]

[0015] where is the preprocessed static surface electromyogram signal, E() represents the expectation, T represents the transpose operation, -1 represents the inverse operation.

[0016] 3-2. Calculate γ(n), and obtain the moment n0 corresponding to the maximum value in γ(n):

[0017]

[0018] Then calculate the j-th MU discharge sequence through n0

[0019]

[0020] 3-3. Add the moments corresponding to the first two maximum values in the discharge sequence to Ψ j , and update the MU discharge sequence according to the following formula

[0021]

[0022]

[0023] where card(Ψ j ) represents the number of elements in Ψ j . Repeat until card(Ψ j ) = r1;

[0024] 3-4. Empty Ψ j , and add the moments corresponding to the first two maximum values in it to Ψ through j , and remove the adjacent moments. Repeat this step until card(Ψ j ) = r2 to obtain the initial MU discharge sequence;

[0025] 3-5. Calculate the indexes of each MU discharge sequence, including: PNR, COV:

[0026]

[0027]

[0028] Among them, diff() represents the differential operation on the elements in the set, std() represents the standard deviation operation, and mean() represents the mean calculation.

[0029] 3 - 6. Classify the MU discharge sequences into good sequences and bad sequences through the PNR and COV indicators.

[0030] Preferably, in step 4, spike detection is performed on the good sequences, and a spike sequence is obtained from the bad sequences using a Post-Processor. The specific steps are as follows:

[0031] 4 - 1. Delete the part of the moment set Ψ of each bad sequence j where the overlapping moments with the good sequences are greater than 30%;

[0032] 4 - 2. Retain those bad sequences whose number of elements in Ψ after the deletion moments exceeds 40% of the original, and recalculate each bad sequence according to the retention result j and

[0033] 4 - 3. For each MU, add the moments corresponding to the top 2 maximum values obtained through to Ψ, removing adjacent moments and the part where the overlapping moments with the good sequences are greater than 30%. Repeat this step until card(Ψ j ) = r3 to obtain the small MU discharge sequence. j

[0034] Preferably, in step 5, the MUAP waveform is estimated using the peak-triggered averaging technique and deleted from the static surface electromyogram signal. The specific steps are as follows:

[0035] 5 - 1. Estimate the MUAP waveform using the peak-triggered averaging technique:

[0036]

[0037] where MUAP j is the waveform of the jth MU within ±20 moments of the discharge moment.

[0038] 5 - 2. From the static surface electromyogram signal for each moment in Ψ j , delete MUAP j at ±20 moments.

[0039] The beneficial effects of the present invention are as follows:

[0040] A new static surface electromyogram (sEMG) decomposition framework was developed by combining the cyclic CKC, Post-Processor, and 'peeling' strategies. The cyclic CKC method was used to extract the discharge sequences of large MUs, while the Post-Processor and 'peeling' strategies were specifically used for the decomposition of small MUs. The high yield and accuracy of the decomposed MUs from simulated and experimental signals demonstrated the reliability and ability of the proposed new framework in decomposition. Brief Description of the Drawings

[0041] Figure 1 Flowchart for the cyclic CKC and result classification;

[0042] Figure 2 Flowchart for the Post-Processor;

[0043] Figure 3 Overall flowchart for the new framework;

[0044] Figure 4 Example diagrams for the MU discharge sequence and MUAP waveform. Detailed Implementation Manner

[0045] The present invention will be further described below in conjunction with specific embodiments. The following description is only for demonstration and explanation, and does not impose any formal limitation on the present invention.

[0046] Step 1. Use an electromyogram acquisition device to collect the surface muscle electrical signals of the biceps brachii, with a sampling frequency of 2000 Hz, 64 channels, and a sampling time of 10 seconds.

[0047] Step 2. Perform band-pass filtering on the collected signals from 20 to 500 Hz and perform delay extension with a delay coefficient of 10. The extended signal matrix should be 640x20000.

[0048] Step 3. Perform cyclic CKC decomposition on the preprocessed signals to obtain the preliminary discharge sequences of MUs, and classify the discharge sequences of MUs into good sequences and poor sequences according to the indicators. The structure is as Figure 1 shown. The specific steps are as follows:

[0049] 3-1. Calculate the covariance matrix of the static surface electromyogram signals and its inverse matrix The matrix size is 640x640:

[0050] 3-2. Calculate γ(n), obtain the moment n0 corresponding to the maximum value in γ(n), and then calculate the jth MU discharge sequence through n0 The sequence length is 1x20000;

[0051] 3 - 3. Obtain the moments corresponding to the first two maximum values in the discharge sequence and add them to Ψ j , and update the MU discharge sequence according to the following formula Repeat until card(Ψ j ) = r1, where the number of elements r1 in the moment set Ψ j is set to 20;

[0052] 3 - 4. Empty Ψ j , obtain the moments corresponding to the first two maximum values among them and add them to Ψ , and remove adjacent moments. Repeat this step until card(Ψ j ) = r2, where the number of elements r2 in the moment set Ψ j is set to 50, to obtain the initial MU discharge sequence; j

[0053] 3 - 5. Calculate the indicators of each MU discharge sequence, including: PNR, COV. Finally, classify the MU discharge sequences into good sequences and bad sequences according to the PNR and COV indicators. The threshold ranges of PNR and COV are 35 and 0.3 respectively.

[0054] Step 4. Use spike detection for good sequences and use Post - Processor for bad sequences to obtain spike sequences, the structure is as Figure 2 shown. The specific steps are as follows:

[0055] 4 - 1. Delete the part in the moment set Ψ j of each bad sequence where the overlapping moments with good sequences are more than 30%;

[0056] 4 - 2. Retain the bad sequences whose number of elements in Ψ j exceeds 40% of the original after deleting the moments, and recalculate the length of each bad sequence to be 1x20000;

[0057] 4 - 3. For each MU, obtain the moments corresponding to the first two maximum values among them and add them to Ψ j , remove adjacent moments and the part where the overlapping moments with good sequences are more than 30%. Repeat this step until card(Ψ j ) = r3, where the number of elements r2 in the moment set Ψ j is set to 70, to obtain the small MU discharge sequence.

[0058]

[0059] Step 5. Estimate the MUAP waveform using peak - triggered averaging technique and delete it from the static surface electromyogram signal. The specific steps are as follows:

[0059] 5-1. Estimate the MUAP waveform using peak-triggered averaging technique, and each waveform should be a 640*41 matrix;

[0060] 5-2. From the static surface electromyogram signals Among them, for Ψ j Delete the MUAP for each of the ±20 moments at each moment in j .

[0061] Step 6. Repeat Steps 1-5 until no new MUs are generated.

[0062] The flowchart of the whole work is as Figure 3 shown. Here, 6 subjects were used as experimental objects, and each subject had 1 experiment. Compared with KMCKC, regardless of the signal-to-noise ratio (5, 10, 15, 20 dB) of the simulated sEMG signals, the new framework can extract more MUs. On average, 8.25±0.95 (a total of 10 units, 3 more units than KMCKC) were extracted, corresponding to an estimated decomposition accuracy of 100%. From the experimental sEMG signals, on average, 14±2.5 MUSTs (4 more MUs than KMCKC) were identified, and the average COV was 0.16±0.04. Figure 4 is an example diagram of the MU discharge sequence and the MUAP waveform decomposed;

[0063] Compared with the previous KMCKC algorithm, the present invention uses a cyclic convolution kernel compensation (CKC) method (a new iterative CKC method) to extract large MUs, and uses a Post-Processor and a "stripping" strategy to extract the remaining MUs with low energy. Using such a method, the MU yield and recognition rate have been significantly improved, and the robustness in a high-noise environment has also been enhanced. It can be seen that the algorithm proposed by the present invention has a certain improvement in the ability to decompose static surface electromyogram signals.

Claims

1. A method for decomposing static surface electromyography signals, characterized in that, It includes the following steps: Step 1. Use a device to collect static surface electromyography signals; Step 2. Preprocess the signals; Step 3. Perform cyclic CKC decomposition on the preprocessed signals to obtain the initial discharge sequences of MUs, and classify the discharge sequences of MUs into good sequences and poor sequences according to the indicators; Step 3 specifically includes the following sub-steps: Sub-step 3-1, calculate the covariance matrix of the static surface electromyography signal and its inverse matrix where is the preprocessed static surface electromyography signal, E() represents the expectation, T represents the transpose operation, and -1 represents the inverse operation; Sub-step 3-2. Calculate γ(n) and obtain the moment n0 corresponding to the maximum value in γ(n): Then, calculate the discharge sequence of the j-th MU through n0 Sub-step 3-3, add Ψ at the moments corresponding to the first two maximum values in the discharge sequence j , update the MU discharge sequence according to the following formula where card(Ψ j ) represents the number of elements in Ψ j ; repeat until card(ψ j ) = r1; Sub-step 3-4, set ψ j to be empty, and through obtain the moments corresponding to the first two maximum values among them and add them to ψ j , and remove adjacent moments. Repeat this step until card(ψ j ) = r2 to obtain the initial discharge sequence of MU; Sub-step 3-5. Calculate the indicators of each MU discharge sequence, including: PNR, COV: Among them, diff() represents the difference operation on the elements in the set, std() represents the standard deviation operation, and mean() represents the mean calculation; Sub-step 3-6. Classify the MU discharge sequences into good sequences and poor sequences through the PNR and COV indicators; Step 4. Use spike detection for good sequences and use Post-Processor for bad sequences to obtain spike sequences; Step 5. Estimate the MUAP waveform using the peak-triggered averaging technique and delete it from the static surface electromyography signals; Step 6. Repeat Steps 1-5 until no new MUs are generated.

2. The static surface electromyogram signal decomposition method according to claim 1, wherein The specific steps of Step 4 include the following sub-steps: Sub-step 4-1: Delete the set of moments ψ of each bad sequence j where the part that coincides with the good sequence has a coincidence time greater than 30%; Sub-step 4-2, retain those ψ after the deletion moment j bad sequences with the number of elements exceeding 40% of the original, and recalculate each bad sequence according to the retention result Sub-step 4-3: For each MU, by obtain the moments corresponding to the first two maximum values therein and add them to ψ j , remove adjacent moments and the part where the coincidence time with the good sequence is greater than 30%; repeat this step until card(Ψ j ) = r3 to obtain the small MU discharge sequence.

3. A method for decomposing static surface electromyography signals according to claim 1, characterized in that The specific steps of Step 5 include the following sub-steps: Sub-step 5-1. Estimate the MUAP waveform using the peak-triggered averaging technique: where MUAP j is the waveform of the j-th MU within 20 instants before and after the discharge instant; Sub-step 5-2. From the static surface electromyogram signals For Ψ in j Delete MUAP for each moment in ±20 moments j .

Images