Spinal cord field potential, electromyographic decoding system and device based on frequency band masking analysis

Abstract

This invention relates to the fields of neural engineering and electrical stimulation neuromodulation technology, and discloses a spinal cord field potential and electromyography decoding system and device based on frequency band masking analysis. The system includes: acquiring raw spinal cord field potential signals and denoising them to obtain LFP denoised signals; simultaneously acquiring raw electromyography signals of the target muscle group and extracting the EMG envelope to obtain baseline version data; performing power reduction on multiple frequency bands of the LFP denoised signal to obtain LFP masked signals, and combining them with the EMG envelope as masked version data; extracting LFP features and EMG envelope target values; inputting the LFP features of the baseline version data into a recurrent neural network model to fit and train the EMG envelope target values; obtaining the EMG envelope predicted by the masked version data and the baseline version data through the trained model, and comparing it with the EMG envelope target values ​​to obtain the correlation between the masked frequency bands and the EMG. This invention can effectively analyze the correlation between spinal cord field potentials and the EMG envelope of the target muscle group.

Description

Technical Field

[0001] This invention relates to the fields of neural engineering and electrical stimulation neuromodulation technology, and in particular to a spinal cord field potential and electromyography decoding system and device based on frequency band masking analysis. Background Technology

[0002] Brain-Machine Interface (BMI) aims to establish a communication link between the brain and external devices, providing basic motor functions for patients with motor dysfunction. In BMI systems, neural signal decoding is a core technological component, with the goal of extracting motor intention information from recorded neural signals. Local Field Potentials (LFPs), as a comprehensive signal of the electrical activity of neuronal populations near the recording electrodes, have frequency characteristics (δ, θ, α, β, γ, etc.) closely related to motor behavior. Existing technologies have explored using LFP decoding to identify motor intentions or control neuroprosthetics.

[0003] Epidural electrical stimulation (ES) is a neuromodulation technique that uses implanted electrodes to stimulate spinal cord neurons, restoring or improving motor function in patients with spinal cord injuries. The stimulation frequency of EES has a significant impact on the therapeutic effect, and different frequencies of electrical stimulation may activate different neural pathways. However, current selection of EES stimulation frequency relies heavily on experience or trial and error, lacking objective guidance based on the subject's own neural signals.

[0004] In summary, existing technologies have the following limitations in studying the relationship between different frequency bands of brain LFP and electromyography (EMG) activity of target muscle groups: 1. Decoding methods lack in-depth exploration of causal explanations and other aspects of neural signal encoding mechanisms. On the one hand, traditional methods typically employ correlation analysis or spectral power analysis to study the relationship between different frequency bands of the brain's LFP and the EMG activity of the target muscle group. However, the strength of the correlation only indicates a mathematical connection between the two. Spinal cord signals differ from brain signals; while spontaneously firing signals, they are controlled by higher centers such as the brain, just like muscles. Therefore, the high correlation detected between spinal cord LFP and EMG may still reflect the result of both being activated by higher centers, which is insufficient to explain the control effect of the signals encoded by the spinal cord itself on muscles, and cannot provide direct evidence for the coupling between spinal cord neurons and muscles. On the other hand, traditional research on neural signal decoding mainly focuses on improving the system's prediction accuracy of motor intentions, but lacks a technical framework to assess the necessity of specific neural signal features for the decoding process, thus neglecting the analysis of the underlying logic of "which neural signal components are necessary to drive muscle activity."

[0005] 2. Lack of physiological basis for EES frequency selection: Due to the lack of an efficient decoding system to correlate endogenous electrophysiological signals with exogenous electrical stimulation modulation signals, existing EES systems mostly use fixed frequencies obtained through multiple rounds of trial and error or empirical frequencies provided by previous studies for electrical stimulation modulation. They fail to make personalized adjustments to the stimulation frequency based on the differences in the modulation loops required by individual subjects, thus limiting the stimulation effect.

[0006] Besides the brain, the spinal cord is also part of the central nervous system. The relationship between different frequency bands of spinal cord LFP and the electromyographic activity of target muscle groups is of great significance. However, the existing research framework lacks effective analytical methods to decode the signal characteristics of spinal cord intrinsic neurons in motor control. Furthermore, even if methods used to study the relationship between different frequency bands of brain LFP and the electromyographic activity of target muscle groups are applied to study the relationship between different frequency bands of spinal cord LFP and the electromyographic activity of target muscle groups, the same limitations will remain. Summary of the Invention

[0007] Therefore, the technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a spinal cord field potential and electromyography decoding system and device based on frequency band masking analysis, which can realize the analysis of the correlation between spinal cord field potential and target muscle group electromyographic envelope, reveal the signal encoding mechanism of spinal cord intrinsic neurons, and improve the effect of EES stimulation parameter selection.

[0008] To address the aforementioned technical problems, this invention provides a spinal cord field potential and electromyography decoding system based on frequency band masking analysis, comprising: The data acquisition module simultaneously collects raw signals of spinal cord field potentials and raw electromyographic signals of the target muscle groups; The data preprocessing module removes power frequency interference from the original spinal cord field potential signal to obtain an LFP denoised signal, performs bandpass filtering and RMS envelope extraction on the original electromyography signal of the target muscle group to obtain an EMG envelope, and uses the corresponding LFP denoised signal and EMG envelope as baseline version data. The feature masking module uses a band-stop filter to reduce the power of multiple frequency bands of the LFP denoised signal to obtain multiple sets of LFP masking signals, which are then combined with the EMG envelope to serve as masked version data. The feature extraction module extracts the power values ​​of each frequency band of the LFP denoised signal in the baseline version data and the multiple sets of LFP masking signals in the masking version data as LFP features, and extracts the median of the EMG envelope as the target value of the EMG envelope. The decoder module constructs a recurrent neural network model. When training the recurrent neural network model, the LFP features of the baseline version data are used as the input of the recurrent neural network model. The EMG envelope output during the training process is fitted to the target value of the EMG envelope. The causal analysis module inputs the LFP features of the masked version data and the LFP features of the baseline version data into the trained recurrent neural network model to obtain the predicted EMG envelope of the masked version data and the predicted EMG envelope of the baseline version data. Combining the predicted EMG envelope of the masked version data, the predicted EMG envelope of the baseline version data, and the target value of the EMG envelope, the correlation between the masked frequency band and the EMG corresponding to the masked version data is obtained.

[0009] Furthermore, when removing power frequency interference from the original spinal cord field potential signal, a Butterworth notch filter is used to filter out 50Hz power frequency interference. The EMG envelope was obtained by bandpass filtering and RMS envelope extraction of the raw electromyographic signals of the target muscle group. The conventional electromyographic frequency band of 20 Hz-200 Hz was extracted using a Butterworth bandpass filter, and the EMG envelope was obtained by calculating the RMS value of a 300ms sliding window.

[0010] Furthermore, after using a band-stop filter to reduce the power of multiple frequency bands of the LFP denoised signal, multiple sets of LFP masking signals are obtained, which are then combined with the EMG envelope to form masked version data, including: The LFP denoised signal is divided into multiple frequency bands according to the EEG frequency band division rules. Butterworth band-stop filters are designed according to the intervals of each frequency band. The LFP denoised signal is subjected to multiple independent band-stop filters to obtain multiple sets of LFP masking signals. The multiple sets of LFP masking signals are combined with the EMG envelope to form multiple masking version data.

[0011] Furthermore, the LFP denoising signal in the baseline version data and the LFP masking signal in the masked version data correspond to the temporal relationship of the EMG envelope.

[0012] Furthermore, the step of designing Butterworth band-stop filters according to the respective intervals of the divided frequency bands includes: The cutoff frequency of the Butterworth band-stop filter is designed as follows: Wn = [x, y] / (Fs / 2), In the formula, Wn represents the cutoff frequency of the Butterworth band-stop filter, x is the lower limit frequency of the interval to be removed, y is the upper limit frequency of the interval to be removed, and Fs is the signal sampling frequency. The Butterworth band-stop filter is designed as follows: [b, a] = butter(2, Wn, 'stop'), In the formula, b and a are the output filter coefficients, butter represents the Butterworth filter, 2 is the order of the Butterworth band-stop filter, and 'stop' represents the band-stop filter.

[0013] Furthermore, the LFP masking signal is: LFP_ Filtered = filtfilt(b, a, LFP_Raw), In the formula, LFP_Raw represents the LFP denoised signal, LFP_Filtered represents the LFP masked signal, and filtfilt() represents zero-phase filtering.

[0014] Furthermore, by combining the EMG envelope predicted from the masked version data, the EMG envelope predicted from the baseline version data, and the target value of the EMG envelope, the correlation between the masked frequency band corresponding to the masked version data and the EMG is obtained, including: Calculate the MSE value of the EMG envelope predicted by the masked version data and the target value of the EMG envelope, and calculate the MSE value of the EMG envelope predicted by the baseline version data and the target value of the EMG envelope. The frequency band feature quantization index is calculated based on the MSE value of the EMG envelope predicted by the masked version data and the target value of the EMG envelope, and the MSE value of the EMG envelope predicted by the baseline version data and the target value of the EMG envelope. The larger the value of the frequency band feature quantization index, the higher the correlation between the masked frequency band corresponding to the masked version data and the EMG.

[0015] Furthermore, the frequency band characteristic quantification index is as follows: ΔMSE = MSE after shading - Baseline MSE In the formula, ΔMSE represents the frequency band feature quantization index, MSE after masking represents the MSE value of the EMG envelope predicted by the masked version data and the target value of the EMG envelope, and baseline MSE represents the MSE value of the EMG envelope predicted by the baseline version data and the target value of the EMG envelope.

[0016] Further, the power values ​​of each frequency band of the LFP denoised signal in the baseline version data and the multiple sets of LFP masking signals in the masking version data are extracted as LFP features, including: The LFP denoised signal in the baseline version data, the LFP masked signal in the masked version data, and the EMG envelope are synchronously windowed; the RMS power value of each frequency band of the LFP denoised signal in each window of the baseline version data is calculated, and the power value of each frequency band is used as the LFP feature of the baseline version data; the RMS power value of each frequency band of the LFP masked signal in each window of the masked version data is calculated, and the power value of each frequency band is used as the LFP feature of the masked version data.

[0017] The present invention also provides a spinal cord field potential and electromyography decoding device based on frequency band masking analysis, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it realizes the functions of each module in the spinal cord field potential and electromyography decoding system based on frequency band masking analysis.

[0018] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: This invention identifies the frequency characteristics of central nervous system signals that causally contribute to motor output through frequency band masking sensitivity analysis, thereby enabling the analysis of the correlation between spinal cord field potentials and the electromyographic envelope of target muscle groups. This can reveal the signal encoding mechanism of intrinsic spinal cord neurons. Furthermore, based on the neuronal signal encoding mechanism, personalized stimulation parameter optimization can be provided for spinal cord injury patients with individual differences, improving the efficiency and reliability of EES stimulation parameter selection. Attached Figure Description

[0019] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the operation of each module of the system in a preferred embodiment of the present invention.

[0020] Figure 2 This is a structural diagram of a recurrent neural network model in a preferred embodiment of the present invention.

[0021] Figure 3 The R index is used in the positive frequency band importance screening method. 2 Example graphs illustrating the importance of predicting joint angle changes individually across different frequency bands.

[0022] Figure 4 This is a visualization of the prediction performance of the recurrent neural network constructed and trained by the decoder module in the experiment of the preferred embodiment of the present invention on the training set and the validation set.

[0023] Figure 5 This is a visualization of the prediction performance of the recurrent neural network constructed and trained by the decoder module in the experiment of the preferred embodiment of the present invention on the test set.

[0024] Figure 6 This is a time-series diagram of the EMG envelope and the target value of the EMG envelope predicted by the recurrent neural network constructed and trained by the decoder module in the experiment of the preferred embodiment of the present invention for the baseline version data.

[0025] Figure 7 The results are the actual experimental results of the causal analysis module in the preferred embodiment of the present invention.

[0026] Figure 8 The MSE values ​​of the test set for the baseline version data and the 7 sets of masked version data in the preferred embodiment of the present invention are shown.

[0027] Figure 9 In the preferred embodiment of the present invention, the experiment was conducted on... Figure 7 Statistical analysis chart of the results. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0029] This invention discloses a spinal cord field potential and electromyography decoding system based on frequency band masking analysis, including a data acquisition module, a data preprocessing module, a feature masking module, a dataset partitioning module, a feature extraction module, a decoder module, and a causal analysis module. When each module studies the relationship between different frequency bands of spinal cord LFP and the electromyographic activity of the target muscle group (i.e., decoding the influence of spinal cord intrinsic neurons on the signal characteristics of motor innervation), the workflow of each module is as follows: Figure 1 As shown, the specific steps include: S1: The data acquisition module synchronously acquires the raw spinal cord field potential signal (i.e., the LFP signal of the spinal cord) and the raw electromyographic signal of the target muscle group (i.e., the EMG signal of the target muscle group).

[0030] S2: The data preprocessing module removes power frequency interference from the original spinal cord field potential signal to obtain the LFP denoised signal. The original electromyography signal of the target muscle group is bandpass filtered and RMS (Root Mean Square) envelope is extracted to obtain the EMG envelope. The corresponding LFP denoised signal and EMG envelope are used as the baseline version data.

[0031] S2-1: When removing power frequency interference from the original spinal cord field potential signal, specifically, a Butterworth notch filter is used to filter out 50 Hz power frequency interference.

[0032] S2-2: Bandpass filtering and RMS envelope extraction are performed on the raw electromyography (EMG) signals of the target muscle group to obtain the EMG envelope. Specifically, the Butterworth bandpass filter is used to extract the conventional EMG frequency band of 20 Hz - 200 Hz, and the RMS value of a 300ms sliding window is calculated to obtain the EMG envelope.

[0033] S2-3: Use the corresponding LFP denoised signal and EMG envelope as the baseline version data.

[0034] S3: The feature masking module uses a band-stop filter to reduce the power of multiple frequency bands of the LFP denoised signal, resulting in multiple sets of LFP masked signals. These are then combined with the EMG envelope to form masked version data. When reducing the power of multiple frequency bands, the frequency bands can be adjusted according to actual conditions, or other neural signal feature indicators of interest (such as firing phase) can be selected to design appropriate filtering methods for power reduction, thereby achieving masking.

[0035] S3-1: According to the rules for dividing EEG frequency bands, the LFP denoising signal is divided into multiple frequency bands. In this embodiment, it is divided into 7 frequency bands: δ band (1-4 Hz), θ band (4-8 Hz), α band (8-12 Hz), low β band (12-22 Hz), high β band (22-30 Hz), low γ band (30-60 Hz), and high γ band (60-90 Hz).

[0036] S3-2: Based on the frequency bands they belong to, design Butterworth band-stop filters for each band. Perform multiple (7) independent band-stop filters on the LFP denoised signal to obtain multiple (7) sets of LFP masking signals. Combine these multiple (7) sets of LFP masking signals (with the corresponding frequency band removed) with the EMG envelope to form multiple (7) masked version data sets. The timing relationship between the LFP denoised signal in the baseline version data and the LFP masking signal in the masked version data remains consistent with the EMG envelope.

[0037] The cutoff frequency of the Butterworth band-stop filter is designed as follows: Wn = [x, y] / (Fs / 2), In the formula, Wn represents the cutoff frequency of the Butterworth band-stop filter, x is the lower limit frequency of the interval to be removed, y is the upper limit frequency of the interval to be removed, and Fs is the signal sampling frequency.

[0038] The Butterworth band-stop filter is designed as follows: [b, a] = butter(2, Wn, 'stop'), In the formula, b and a are the output filter coefficients, butter represents the Butterworth filter, 2 is the order of the Butterworth band-stop filter, and 'stop' represents the band-stop filter.

[0039] The LFP masking signal is: LFP_ Filtered = filtfilt(b, a, LFP_Raw), In the formula, LFP_Raw represents the LFP denoised signal, LFP_Filtered represents the LFP masked signal, and filtfilt() represents zero-phase filtering. That is, band-stop filtering is performed on the LFP denoised signal in the x-y frequency band to generate the LFP masked signal with the x-y frequency band removed.

[0040] S4: The dataset partitioning module divides the baseline version data and the masked version data into training, validation, and test sets according to a ratio of 70%, 15%, and 15%, respectively. All three sets of the baseline version data (training, validation, and test sets) are retained for subsequent training, while only the test set of the masked version data is retained for subsequent comparative analysis.

[0041] S5: The feature extraction module extracts the power values ​​of each frequency band of the LFP denoised signal in the baseline version data and the multiple sets of LFP masked signals in the masked version data as LFP features, and extracts the median of the EMG envelope as the target value of the EMG envelope.

[0042] S5-1: Perform windowing to synchronize the LFP denoised signal in the baseline version data, the LFP masking signal in the masked version data, and the EMG envelope. In this embodiment, each window can be set to 50-150 ms, and a sequence is constructed with a 50% overlap ratio. The sequence length is adjusted appropriately according to the decoding purpose and model performance.

[0043] S5-2: Calculate the RMS power values ​​of each frequency band of the LFP denoised signal for each window in the baseline version data, and use the power values ​​of each frequency band as the LFP features of the baseline version data. Calculate the RMS power values ​​of each frequency band of the LFP masked signal for each window in the masked version data, and use the power values ​​of each frequency band as the LFP features of the masked version data. Extract the median of the EMG envelope for each window as the target value of the EMG envelope, and use the target value of the EMG envelope as the target value of the subsequent recurrent neural network model.

[0044] S5-3: Normalize the LFP features and the target value of the EMG envelope respectively.

[0045] S6: Decoder module construction as follows Figure 2 The recurrent neural network model shown uses the LFP features of the baseline version data as input to the recurrent neural network model during training, and the EMG envelope output during the training process is fitted to the target value of the EMG envelope.

[0046] S7: The causal analysis module inputs the LFP features of the masked version data and the LFP features of the baseline version data into the trained recurrent neural network model to obtain the predicted EMG envelope of the masked version data and the predicted EMG envelope of the baseline version data. Combining the predicted EMG envelope of the masked version data, the predicted EMG envelope of the baseline version data, and the target value of the EMG envelope, the correlation between the masked frequency band and the EMG corresponding to the masked version data is obtained.

[0047] S7-1: Calculate the MSE (Mean Square Error) values ​​of the EMG envelope and EMG envelope target value predicted by the masked version data, and calculate the MSE values ​​of the EMG envelope and EMG envelope target value predicted by the baseline version data.

[0048] S7-2: The quantization index for calculating frequency band characteristics is: ΔMSE = MSE after shading - Baseline MSE In the formula, ΔMSE represents the frequency band feature quantization index, MSE after masking represents the MSE value of the EMG envelope and the target value of the EMG envelope predicted by the masked version data, and baseline MSE represents the MSE value of the EMG envelope and the target value of the EMG envelope predicted by the baseline version data.

[0049] S7-3: ΔMSE can represent the degree of influence of the masked frequency band on the predicted EMG. The larger the value of ΔMSE, the higher the correlation between the masked frequency band corresponding to the masked version data and the EMG, that is, the higher the importance of the masked frequency band. This can be used to generate a ranking of frequency band importance.

[0050] The present invention also discloses a spinal cord field potential and electromyography decoding device based on frequency band masking analysis, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it realizes the functions of each module in the spinal cord field potential and electromyography decoding system based on frequency band masking analysis.

[0051] This invention, by constructing an analytical framework of "training baseline, occlusion features, and quantization error," offers the following advantages compared to existing technologies: 1. Possesses the ability to make causal inferences and reveals neural coding mechanisms.

[0052] This invention breaks through the limitations of traditional correlation analysis by using frequency masking sensitivity analysis to quantify the necessity of various neural signal components in the decoding process. This allows for the identification of LFP frequency features that contribute causally to electromyographic activity rather than merely having a correlation, providing a technical tool for understanding how spinal cord neurons encode motor commands.

[0053] 2. It can guide the selection of neural regulation parameters.

[0054] Decoding the inherent neural signal encoding frequency characteristics of different spinal cord and muscle circuits in normal populations provides an objective physiological basis for selecting stimulation frequencies for neuromodulation techniques such as epidural electrical stimulation. Therefore, it can provide personalized stimulation parameter optimization for the neural circuits that need to be regulated for spinal cord injury patients with individual differences.

[0055] 3. The analytical framework is universal.

[0056] This invention is not only applicable to decoding the frequency characteristics of spinal cord LFP signals, but can also be extended to neural signals from other sources (such as cortical EEG and deep EEG) and other neural signal indicators to be analyzed (such as firing phase and event window). It can systematically evaluate the contribution of the encoding characteristics of various neural signals to behavioral output. It can be used to identify key frequency characteristics under different motor states, thereby revealing the neural encoding mechanism of spinal cord motor control circuits and providing a powerful tool at the signal decoding level for studying the neural activity and plasticity changes of the central nervous system. It can guide the selection of neuromodulation stimulation parameters: the frequency characteristics of the decoded central nervous signals have physiological significance. Combined with experimental verification, the stimulation frequency of neuromodulation methods such as electrical stimulation can be efficiently selected, improving the therapeutic effect of neuromodulation. It can promote the development of brain-computer interfaces: based on the frequency characteristics of the decoded central nervous signals, the neural signal decoding algorithm can be optimized, improving the control accuracy of neural prostheses.

[0057] To further demonstrate the beneficial effects of this invention, an experiment was conducted on mouse spinal cord LFP-EMG decoding and frequency importance analysis. The experimental design and procedure are as follows: (1) Animal model: 9 C57BL / 6 mice.

[0058] (2) Signal acquisition: The LFP signal in the gray matter of the spinal cord lumbar enlargement and the EMG of the tibialis anterior muscle of the hind limb were simultaneously acquired by the Alpha Omega electrophysiological acquisition system and multi-channel electrodes. The sampling rate was 937.5 Hz. Continuous rhythmic walking data for more than 1 minute were recorded for each animal. The data of 9 mice were summarized into corresponding LFP and EMG arrays.

[0059] (3) Data preprocessing: The LFP signal of each mouse was subjected to a 4th order Butterworth notch filter at 50 Hz to remove power frequency interference and obtain the LFP denoised signal; the EMG signal was subjected to a 4th order Butterworth bandpass filter at 60-200 Hz, and the EMG envelope was calculated by selecting a 300 ms sliding window RMS according to the mouse stepping frequency (2-3 steps / s).

[0060] (4) Generation of different versions: Eight versions of LFP data were generated for each mouse, namely the baseline version data composed of the original unmasked LFP denoised signal and the EMG envelope. Seven sets of LFP masked signals were generated after masking the δ / θ / α / low β / high β / low γ / high γ frequency bands through a 4th-order Butterworth band-stop filter, and then combined with the EMG envelope to form seven masked versions of data.

[0061] (5) Feature extraction: The LFP denoised signal from the baseline version data of 9 mice, the LFP occlusion signal from the occlusion version data, and the EMG envelope were segmented into windows, each window being 50 ms long. The subsequent training sequence was constructed with a 50% overlap ratio, and each sequence contained 40 consecutive windows. The signal components of 7 frequency bands in each window were extracted using a second-order Butterworth bandpass filter, and the RMS power was calculated as the 7-dimensional features of LFP in that window. A 7-dimensional feature vector sequence was constructed as the input value for subsequent prediction. At the same time, the median of the EMG envelope in each window was extracted and used as the target value for subsequent prediction.

[0062] (6) Data set partitioning: The corresponding 7-dimensional feature vector sequence and EMG envelope median are divided according to the sample order of 9 mice. Each sample is randomly partitioned into 70%, 15%, and 15% segments to ensure that the subsequent training set, validation set, and test set contain different sample data.

[0063] (7) Model training: Construct a two-layer GRU (128→64 units) network architecture, use the Adam optimizer for training, set the initial learning rate to 1e-3, decrease the learning rate by 50% every 20 epochs, and train for a total of 40 epochs; fit the median of the EMG envelope to the LFP feature vector of the input baseline version training set, and evaluate the performance in the validation set by MSE, RMSE, and R²; record the MSE of the test set at this time as the baseline value for subsequent comparisons.

[0064] (8) Frequency masking sensitivity analysis: Input the LFP feature vector sequence of each masked version data test set into the recurrent neural network model trained in the previous step, record the MSE of each masked version data at this time and calculate the difference between it and the baseline value, sort them in descending order, and generate the frequency band importance ranking: the larger the ΔMSE, the more important the frequency band is for predicting EMG distribution.

[0065] This application does not consider system performance issues such as improving model decoding accuracy and cross-sample generalization ability. Its goal is to rapidly perform personalized causal analysis on small-scale collected spinal cord and muscle coupled signals to screen neural signal features. However, this causal analysis method has not yet been reported at the level of central nervous system-muscle signal coupling, especially since both the spinal cord and muscles are inferior innervations of the brain. Ordinary correlation analysis can only analyze the mathematical relationship between the two signals in terms of consistency, and cannot determine the contribution of specific spinal cord LFP frequency bands to EMG transmission. Furthermore, ... Figure 3 The positive frequency band importance screening method shown (see the paper "Jie F, Xiaocheng L, Peiyu W, et al. A Hyperflexible ElectrodeArray for Long-Term Recording and Decoding of Intraspinal Neuronal Activity.[J]. Advanced science, 2023") uses the index R. 2 The importance of different frequency bands in predicting joint angle changes was measured. However, this method has limited resolution. The system may have learned mixed behavioral patterns during hindlimb movement, and there may be frequency bands that contain similar information but are redundant with each other, ultimately resulting in the simultaneous existence of several frequency bands with very high R². Figure 3 The horizontal axis represents the frequency band, and the vertical axis represents R. 2 .

[0066] Figures 4-6 The actual training and testing results of the decoder module in the experiment are as follows: Figures 4-6 It can be seen that the test set R 2 Approximately matching the training and validation sets, the predicted EMG envelope values ​​show a trend largely synchronized with the target values. This allows for the effective learning of individual-specific signal coupling patterns across different samples using a small-scale in vivo electrophysiological dataset with one minute of data collection per sample.

[0067] Figure 7 The actual experimental results of the causal analysis module in the experiment are presented in the form of a heatmap: each row is the frequency band importance data of one mouse, with a total of 9 samples; each column is the difference in MSE change after the corresponding frequency band is masked. Figure 8The MSE values ​​are the baseline version data and the test set MSE values ​​of the 7 sets of occluded version data in the experiment. Figure 9 To Figure 7 Statistical analysis of the results. From Figure 8 and Figure 9 It can be seen that the 12-22 Hz frequency band in the LFP of the mouse spinal cord makes an important contribution to the EMG firing of the tibialis anterior muscle during rhythmic walking, and the features of this frequency band were effectively separated in multiple frequency bands.

[0068] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0070] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A spinal cord field potential and electromyography decoding system based on frequency band masking analysis, characterized in that, include: The data acquisition module simultaneously collects raw signals of spinal cord field potentials and raw electromyographic signals of the target muscle groups; The data preprocessing module removes power frequency interference from the original spinal cord field potential signal to obtain an LFP denoised signal, performs bandpass filtering and RMS envelope extraction on the original electromyography signal of the target muscle group to obtain an EMG envelope, and uses the corresponding LFP denoised signal and EMG envelope as baseline version data. The feature masking module uses a band-stop filter to reduce the power of multiple frequency bands of the LFP denoised signal to obtain multiple sets of LFP masking signals, which are then combined with the EMG envelope to serve as masked version data. The feature extraction module extracts the power values ​​of each frequency band of the LFP denoised signal in the baseline version data and the multiple sets of LFP masking signals in the masking version data as LFP features, and extracts the median of the EMG envelope as the target value of the EMG envelope. The decoder module constructs a recurrent neural network model. When training the recurrent neural network model, the LFP features of the baseline version data are used as input, and the EMG envelope output during the training process is fitted to the target value of the EMG envelope. The causal analysis module inputs the LFP features of the masked version data and the LFP features of the baseline version data into the trained recurrent neural network model to obtain the predicted EMG envelope of the masked version data and the predicted EMG envelope of the baseline version data. Combining the predicted EMG envelope of the masked version data, the predicted EMG envelope of the baseline version data, and the target value of the EMG envelope, the correlation between the masked frequency band and the EMG corresponding to the masked version data is obtained.

2. The spinal cord field potential and electromyography decoding system based on frequency band masking analysis according to claim 1, characterized in that: When removing power frequency interference from the original spinal cord field potential signal, a Butterworth notch filter is used to filter out 50Hz power frequency interference. The EMG envelope was obtained by bandpass filtering and RMS envelope extraction of the raw electromyographic signals of the target muscle group. The conventional electromyographic frequency band of 20 Hz-200 Hz was extracted using a Butterworth bandpass filter, and the EMG envelope was obtained by calculating the RMS value of a 300ms sliding window.

3. The spinal cord field potential and electromyography decoding system based on frequency band masking analysis according to claim 1, characterized in that: The process involves using a band-stop filter to reduce the power of multiple frequency bands of the LFP denoised signal to obtain multiple sets of LFP masking signals, which are then combined with the EMG envelope to form masked version data, including: The LFP denoised signal is divided into multiple frequency bands according to the EEG frequency band division rules. Butterworth band-stop filters are designed according to the intervals of each frequency band. The LFP denoised signal is subjected to multiple independent band-stop filters to obtain multiple sets of LFP masking signals. The multiple sets of LFP masking signals are combined with the EMG envelope to form multiple masking version data.

4. The spinal cord field potential and electromyography decoding system based on frequency band masking analysis according to claim 1, characterized in that: The LFP denoising signal in the baseline version data and the LFP masking signal in the masking version data correspond to the temporal relationship of the EMG envelope.

5. The spinal cord field potential and electromyography decoding system based on frequency band masking analysis according to claim 3, characterized in that: The step of designing Butterworth band-stop filters according to the respective intervals of the divided frequency bands includes: The cutoff frequency of the Butterworth band-stop filter is designed as follows: Wn = [x, y] / (Fs / 2), In the formula, Wn represents the cutoff frequency of the Butterworth band-stop filter, x is the lower limit frequency of the interval to be removed, y is the upper limit frequency of the interval to be removed, and Fs is the signal sampling frequency. The Butterworth band-stop filter is designed as follows: [b, a] = butter(2, Wn, 'stop'), In the formula, b and a are the output filter coefficients, butter represents the Butterworth filter, 2 is the order of the Butterworth band-stop filter, and 'stop' represents the band-stop filter.

6. The spinal cord field potential and electromyography decoding system based on frequency band masking analysis according to claim 5, characterized in that: The LFP masking signal is: LFP_ Filtered = filtfilt(b, a, LFP_Raw), In the formula, LFP_Raw represents the LFP denoised signal, LFP_Filtered represents the LFP masked signal, and filtfilt() represents zero-phase filtering.

7. The spinal cord field potential and electromyography decoding system based on frequency band masking analysis according to claim 1, characterized in that: By combining the EMG envelope predicted from the masked version data, the EMG envelope predicted from the baseline version data, and the target value of the EMG envelope, the correlation between the masked frequency band corresponding to the masked version data and the EMG is obtained, including: Calculate the MSE value of the EMG envelope predicted by the masked version data and the target value of the EMG envelope, and calculate the MSE value of the EMG envelope predicted by the baseline version data and the target value of the EMG envelope. The frequency band feature quantization index is calculated based on the MSE value of the EMG envelope predicted by the masked version data and the target value of the EMG envelope, and the MSE value of the EMG envelope predicted by the baseline version data and the target value of the EMG envelope. The larger the value of the frequency band feature quantization index, the higher the correlation between the masked frequency band corresponding to the masked version data and the EMG.

8. The spinal cord field potential and electromyography decoding system based on frequency band masking analysis according to claim 7, characterized in that: The frequency band characteristic quantification index is: ΔMSE = MSE after shading - Baseline MSE In the formula, ΔMSE represents the frequency band feature quantization index, MSE after masking represents the MSE value of the EMG envelope predicted by the masked version data and the target value of the EMG envelope, and baseline MSE represents the MSE value of the EMG envelope predicted by the baseline version data and the target value of the EMG envelope.

9. The spinal cord field potential and electromyography decoding system based on frequency band masking analysis according to any one of claims 1-8, characterized in that: Extracting the LFP denoised signal from the baseline version data and the power values ​​of each frequency band of multiple sets of LFP masking signals from the masking version data as LFP features, including: The LFP denoised signal in the baseline version data, the LFP masked signal in the masked version data, and the EMG envelope are synchronously windowed; the RMS power value of each frequency band of the LFP denoised signal in each window of the baseline version data is calculated, and the power value of each frequency band is used as the LFP feature of the baseline version data; the RMS power value of each frequency band of the LFP masked signal in each window of the masked version data is calculated, and the power value of each frequency band is used as the LFP feature of the masked version data.

10. A spinal cord field potential and electromyography decoding device based on frequency band masking analysis, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the functions of each module in the spinal cord field potential and electromyography decoding system based on frequency band masking analysis as described in any one of claims 1-9.

Images