Method and system for classifying electrical stimulation monitoring data of wireless implantable nerve regulation

Abstract

The invention provides an electrical stimulation monitoring data classification method and system for wireless implantable nerve regulation, and relates to the field of data processing. When noise cleaning is performed on a first electrical stimulation monitoring signal, spectral analysis is performed on a first signal implicit representation of the first electrical stimulation monitoring signal to obtain a detail implicit representation and a global implicit representation; and integrating the detail implicit representation and the global implicit representation according to the influence variables to obtain a second signal implicit representation, because the detail influence variables and the global influence variables are obtained by training electrical stimulation monitoring signal samples, weighting the implicit representations according to the influence variables to obtain a second signal implicit representation. And detail features with higher analysis value in the electrical stimulation monitoring signal can be focused. The filtering matrix is also obtained by training according to the electrical stimulation monitoring signal sample, so that the filtering matrix can learn how to restore the error between the noise electrical stimulation monitoring signal and the de-noising electrical stimulation monitoring signal, and the de-noising quality of the electrical stimulation monitoring signal is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, specifically, a wireless implantable neuromodulation electrical stimulation monitoring data classification method and system. BACKGROUND

[0002] In the field of wireless implantable neuromodulation, accurate processing and classification of electrical stimulation monitoring signals are crucial for understanding neural activity and ensuring the effectiveness and safety of neuromodulation.

[0003] With the development of technology, in the process of electrical stimulation of neural tissue and monitoring its response, the electrical stimulation monitoring signals obtained often contain a variety of complex information. However, these signals are inevitably disturbed by noise. The sources of noise are diverse, such as thermal noise from the electronic components of the implanted device, interference from the electromagnetic environment, and interference from the complex physiological environment in the living body, etc. The existence of such noise seriously affects the accurate analysis and classification of electrical stimulation monitoring signals. Traditional electrical stimulation monitoring signal processing methods have many limitations in denoising and accurate classification. On the one hand, many traditional methods often use general filtering techniques in the denoising process, which do not fully consider the characteristics of electrical stimulation monitoring signals, especially the uniqueness of signals in the special application scenario of neuromodulation. For example, simple filtering may remove noise while also losing some key information in the signal, such as detailed information related to the small activity changes of neural cells, which is very important for accurately understanding the response of the nervous system to electrical stimulation. On the other hand, in terms of signal classification, due to the interference of noise and the failure of traditional methods to effectively extract and utilize key features in the signal, the accuracy of classification is greatly compromised. For example, in distinguishing different types of neural responses or different states of neural activity, if the noise in the electrical stimulation monitoring signal is not well handled, classification based on these signals is likely to produce false positives, which in turn affects the evaluation of neuromodulation effectiveness and subsequent treatment strategy adjustment. SUMMARY

[0004] The purpose of the present application is to provide a wireless implantable neuromodulation electrical stimulation monitoring data classification method and system. The present application is implemented as follows:

[0005] In a first aspect, the present application provides a wireless implantable neuromodulation electrical stimulation monitoring data classification method, the method comprising: performing feature implicit representation on a first electrical stimulation monitoring signal to obtain a first signal implicit representation; performing spectral analysis on the first signal implicit representation to obtain a first detail implicit representation and a first global implicit representation; integrating the first detail implicit representation and the first global implicit representation according to a detail influence variable and a global influence variable to obtain a second signal implicit representation, the detail influence variable and the global influence variable being obtained by training according to noise electrical stimulation monitoring signal samples and corresponding denoised electrical stimulation monitoring signal samples; filtering the second signal implicit representation through a first filter matrix to obtain a residual mapping signal, the residual mapping signal representing an error between the first electrical stimulation monitoring signal and the corresponding denoised electrical stimulation monitoring signal, the first filter matrix being obtained by training according to the noise electrical stimulation monitoring signal samples and the denoised electrical stimulation monitoring signal samples; integrating the first electrical stimulation monitoring signal and the residual mapping signal to obtain a second electrical stimulation monitoring signal, the signal-to-noise ratio of the second electrical stimulation monitoring signal being greater than that of the first electrical stimulation monitoring signal; and classifying the second electrical stimulation monitoring signal.

[0006] Optionally, the spectral analysis on the first signal implicit representation to obtain a first detail implicit representation and a first global implicit representation comprises:

[0007] classifying the first signal implicit representation into a second detail implicit representation and a second global implicit representation according to frequency spectrum, integrating the second detail implicit representation and the second global implicit representation to obtain a third signal implicit representation;

[0008] performing detail extraction on the third signal implicit representation to obtain the first detail implicit representation, and performing smoothing on the third signal implicit representation to obtain the first global implicit representation.

[0009] Optionally, the method further comprises:

[0010] filtering the third signal implicit representation through a second filter matrix to obtain a smoothing matrix, the second filter matrix being obtained by training according to the noise electrical stimulation monitoring signal and the denoised electrical stimulation monitoring signal;

[0011] performing matrix subtraction calculation on a unit filter matrix and the smoothing matrix to obtain a detail extraction matrix, the number of dimensions of the unit filter matrix being equal to that of the smoothing matrix, and the element values of one dimension of the unit filter matrix being all 1 and the element values of other dimensions being all 0;

[0012] The detail extraction on the third signal implicit representation is performed by using the detail extraction matrix to obtain the first detail implicit representation, and the smoothing on the third signal implicit representation is performed by using the smoothing matrix to obtain the first global implicit representation.

[0013] The detail extraction on the third signal implicit representation is performed by using the detail extraction matrix to obtain the first detail implicit representation, and the smoothing on the third signal implicit representation is performed by using the smoothing matrix to obtain the first global implicit representation.

[0014] Optionally, the number of the smoothing matrices is s, the number of dimensions of each smoothing matrix is m, the number of the detail extraction matrices is s, and the number of dimensions of each detail extraction matrix is m. The detail extraction on the third signal implicit representation is performed by using the detail extraction matrix to obtain the first detail implicit representation, and the smoothing on the third signal implicit representation is performed by using the smoothing matrix to obtain the first global implicit representation.

[0015] The third signal implicit representation is classified into s third signal sub-implicit representations in the dimension, wherein the number of sampling points covered by the smoothing matrix, the number of sampling points covered by the detail extraction matrix, and the number of sampling points covered by the third signal sub-implicit representation are consistent.

[0016] The detail extraction on the s third signal sub-implicit representations is performed by using the s detail extraction matrices to obtain s first detail sub-implicit representations, and the s first detail sub-implicit representations constitute the first detail implicit representation.

[0017] The smoothing on the s third signal sub-implicit representations is performed by using the s smoothing matrices to obtain s first global sub-implicit representations, and the s first global sub-implicit representations constitute the first global implicit representation.

[0018] Optionally, the frequency spectrum analysis on the first signal implicit representation comprises:

[0019] The frequency spectrum analysis on the first signal implicit representation is performed to obtain a pending detail implicit representation and a pending global implicit representation, and the pending detail implicit representation and the pending global implicit representation are integrated according to the detail influence variable and the global influence variable to obtain a pending signal implicit representation.

[0020] performing a new time of spectral analysis on the pending signal implicit representation, obtaining a new one of the pending detail implicit representation and a new one of the pending global implicit representation, integrating the new one of the pending detail implicit representation and the new one of the pending global implicit representation according to the detail influence variable and the global influence variable, obtaining a new one of the pending signal implicit representation, until when a time of the spectral analysis of a current round is not less than a first preset value, determining the pending detail implicit representation obtained at the current time as the first detail implicit representation, and determining the pending global implicit representation obtained at the current time as the first global implicit representation.

[0021] Optionally, the integrating the pending detail implicit representation and the pending global implicit representation according to the detail influence variable and the global influence variable, and obtaining the pending signal implicit representation, comprises:

[0022] If the time of the spectral analysis of the current round is not greater than a second preset value, fusing the pending detail implicit representation obtained at the current time and the pending global implicit representation according to the detail influence variable and the global influence variable, and obtaining a first control signal implicit representation, the second preset value is less than the first preset value;

[0023] performing up-sampling or down-sampling on the first electrical stimulation monitoring signal, and performing feature implicit representation on the first electrical stimulation monitoring signal after the up-sampling or down-sampling, and obtaining a second control signal implicit representation;

[0024] performing feature implicit representation on the first control signal implicit representation and the second control signal implicit representation by a weight focusing strategy, and obtaining the pending signal implicit representation at the current time;

[0025] The integrating the pending detail implicit representation and the pending global implicit representation according to the detail influence variable and the global influence variable, and obtaining the pending signal implicit representation, comprises:

[0026] If the time of the spectral analysis of the current round is greater than a third preset value, fusing the pending detail implicit representation obtained at the current time and the pending global implicit representation according to the detail influence variable and the global influence variable, and obtaining a third control signal implicit representation, the third preset value is not greater than the first preset value;

[0027] integrating one or more pending signal implicit representations obtained in iterations before the current time and the third control signal implicit representation, and obtaining the pending signal implicit representation at the current time.

[0028] Optionally, the wireless implantable neuromodulation electrical stimulation monitoring data classification method is implemented based on an electrical stimulation monitoring signal processing network, the electrical stimulation monitoring signal processing network comprising a feature implicit representation component, a spectrum analysis component, an adaptive fusion component, and a signal generation component; the feature implicit representation component is configured to perform feature implicit representation on the first electrical stimulation monitoring signal to obtain the first signal implicit representation; the spectrum analysis component is configured to perform spectrum analysis on the first signal implicit representation to obtain the first detail implicit representation and the first global implicit representation; the adaptive fusion component is configured to integrate the first detail implicit representation and the first global implicit representation according to the detail influence variable and the global influence variable to obtain the second signal implicit representation, the detail influence variable and the global influence variable being component parameters of the adaptive fusion component; and the signal generation component is configured to filter the second signal implicit representation through the first filter matrix to obtain the residual mapping signal, and integrate the first electrical stimulation monitoring signal and the residual mapping signal to obtain the second electrical stimulation monitoring signal, the first filter matrix being a component parameter of the signal generation component.

[0029] Optionally, the electrical stimulation monitoring signal processing network is trained through the following steps:

[0030] The feature implicit representation component performs feature implicit representation on the noise electrical stimulation monitoring signal sample to obtain a first signal training implicit representation;

[0031] The spectrum analysis component performs spectrum analysis on the first signal training implicit representation to obtain a first detail training implicit representation and a first global training implicit representation;

[0032] The adaptive fusion component integrates the first detail training implicit representation and the first global training implicit representation according to the detail influence variable and the global influence variable to obtain a second signal training implicit representation;

[0033] The signal generation component filters the second signal training implicit representation through the first filter matrix to obtain a residual training mapping signal, and integrates the noise electrical stimulation monitoring signal sample and the residual training mapping signal to obtain an estimated electrical stimulation monitoring signal;

[0034] The electrical stimulation monitoring signal processing network is trained according to the loss between the estimated electrical stimulation monitoring signal and the denoised electrical stimulation monitoring signal sample.

[0035] Optionally, the electric stimulation monitoring signal processing network comprises a plurality of the spectrum analysis components and a plurality of the adaptive fusion components, a≥1; the spectrum analysis of the first signal training implicit representation is performed by the spectrum analysis components to obtain the first detail training implicit representation and the first global training implicit representation, comprising:

[0036] The spectrum analysis of the first signal implicit representation is performed by the first spectrum analysis component to obtain the pending detail training implicit representation and the pending global training implicit representation; the integration of the pending detail training implicit representation and the pending global training implicit representation is performed by the first adaptive fusion component according to the detail influence variable and the global influence variable in the first adaptive fusion component to obtain the pending signal training implicit representation;

[0037] The new spectrum analysis of the pending signal training implicit representation is performed by a new spectrum analysis component to obtain a new pending detail training implicit representation and a new pending global training implicit representation; the integration of the new pending detail training implicit representation and the new pending global training implicit representation is performed by a new adaptive fusion component according to the detail influence variable and the global influence variable in the new adaptive fusion component to obtain a new pending signal training implicit representation, until the end spectrum analysis component outputs the pending detail training implicit representation and the pending global training implicit representation, the pending detail training implicit representation output by the end spectrum analysis component is taken as the first detail training implicit representation, and the pending global training implicit representation output by the end spectrum analysis component is taken as the first global training implicit representation;

[0038] The integration of the first detail training implicit representation and the first global training implicit representation is performed by the adaptive fusion components according to the detail influence variable and the global influence variable to obtain the second signal training implicit representation, comprising:

[0039] The integration of the first detail training implicit representation and the first global training implicit representation is performed by the last adaptive fusion component according to the detail influence variable and the global influence variable in the last adaptive fusion component to obtain the second signal training implicit representation;

[0040] The electric stimulation monitoring signal processing network comprises b signal generation components, 1≤b≤a; the filtering of the second signal training implicit representation by the first filter matrix is performed by the signal generation components to obtain a residual training mapping signal, the integration of the noise electric stimulation monitoring signal sample and the residual training mapping signal is performed to obtain an estimated electric stimulation monitoring signal, comprising:

[0041] The cth residual training mapping signal is obtained by filtering the pending signal training implicit representation output by the a-b+c th adaptive fusion component through the first filter matrix, and the cth estimated electric stimulation monitoring signal is obtained by integrating the noise electric stimulation monitoring signal sample and the cth residual training mapping signal; wherein, 1≤c≤b.

[0042] A computer system, comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method as described above is implemented.

[0043] The beneficial effects of the present application: In the process of noise cleaning of the first electric stimulation monitoring signal, the first signal implicit representation of the first electric stimulation monitoring signal is subjected to frequency spectrum analysis to obtain detailed implicit representation and global implicit representation, and the detailed implicit representation and the global implicit representation are integrated according to the influence variables to obtain the second signal implicit representation. Because the detailed influence variables and the global influence variables are obtained by training the electric stimulation monitoring signal sample, the implicit representation can be weighted according to these influence variables to focus on the more valuable detailed features in the electric stimulation monitoring signal. Because the filter matrix is also obtained by training the electric stimulation monitoring signal sample, the filter matrix can learn how to restore the error between the noise electric stimulation monitoring signal and the de-noised electric stimulation monitoring signal. By filtering the second signal implicit representation through the filter matrix, more detailed errors between the first electric stimulation monitoring signal and the de-noised electric stimulation monitoring signal can be restored to obtain the second electric stimulation monitoring signal with higher signal-to-noise ratio, increase the de-noising quality of the electric stimulation monitoring signal, and help more accurate signal classification. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a flowchart of a wireless implantable neural regulation electric stimulation monitoring data classification method provided by an embodiment of the present application.

[0045] Figure 2 is a composition schematic diagram of a computer system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0046] The execution subject of the wireless implantable neural regulation electric stimulation monitoring data classification method in the embodiment of the present application is a computer system, including but not limited to a server, a personal computer, a notebook computer, a tablet computer, etc. The embodiment of the present application provides a wireless implantable neural regulation electric stimulation monitoring data classification method, which is applied to a server, such as Figure 1As shown, the method comprises:

[0047] Step S100: performing feature implicit representation on the first electrical stimulation monitoring signal to obtain a first signal implicit representation.

[0048] In step S100, the computer system performs feature implicit representation on the first electrical stimulation monitoring signal to obtain a first signal implicit representation. The electrical stimulation monitoring signal refers to the relevant signal monitored during the process of electrical stimulation on the neural tissue in the wireless implantable neural regulation scenario. These signals contain rich information, such as the intensity, frequency, duration of electrical stimulation, and the response of the neural tissue to electrical stimulation.

[0049] Feature implicit representation is a method of representing the features in the original signal in an implicit way. From a technical point of view, the computer system can implement it in multiple ways. For example, a specific mathematical transformation can be used to achieve this purpose. One possible way is through the convolution layer operation in the convolutional neural network (CNN). Assuming that the first electrical stimulation monitoring signal is a discrete time series signal x(n), n = 1, 2, …, N, the computer system can define a convolution kernel k(m), m = 1, 2, …, M (where M is much smaller than N), and extract local features in the signal through convolution operation where y(n) is a preliminary feature implicit representation result after convolution operation. Of course, in actual situations, multiple convolution kernels may be operated simultaneously to extract more types of features.

[0050] In addition to the convolution operation in the convolutional neural network, the computer system can also use other linear or nonlinear transformations for feature implicit representation. For example, principal component analysis (PCA), which is a linear transformation method. For the first electrical stimulation monitoring signal X (which can be regarded as a matrix, each row representing a sample and each column representing a feature dimension), first calculate its covariance matrix Then solve the eigenvalue λ i and the corresponding eigenvector v i of the covariance matrix C. Sort the eigenvectors according to the size of the eigenvalues, and select the top k largest eigenvalues to form a projection matrix P. Then the feature implicit representation after PCA transformation is Y = XP.

[0051] By such a feature implicit representation process, the computer system converts the original first electrical stimulation monitoring signal into a first signal implicit representation. This first signal implicit representation is no longer a direct expression of the original signal, but a signal representation form that has been extracted and converted, which is more conducive to subsequent analysis and processing. It can highlight some important features in the original signal, while suppressing some unimportant or noise-related information, thereby laying the foundation for subsequent spectral analysis and other operations. For example, in neural regulation, if the original electrical stimulation monitoring signal contains weak neural response signals and strong background noise, after suitable feature implicit representation, the features corresponding to the neural response signals can be more prominent, and the features related to the background noise can be weakened, which helps to improve the accuracy and effectiveness of the entire data classification method.

[0052] Step S200: performing spectral analysis on the first signal implicit representation to obtain a first detail implicit representation and a first global implicit representation.

[0053] In step S200, the computer system performs spectral analysis on the first signal implicit representation obtained in step S100, and further obtains a first detail implicit representation and a first global implicit representation. Spectral analysis is a technical means of converting a signal from time domain to frequency domain for analysis. For the computer system, it helps to understand the frequency composition structure and other information of the signal. In actual operation, the computer system can use Discrete Fourier Transform (DFT) to realize spectral analysis. Assuming that the first signal implicit representation is a discrete signal x(n), n = 0, 1, …, N-1, its DFT is defined as where k = 0, 1, …, N-1, X(k) is the representation of x(n) in the frequency domain, which contains the amplitude and phase information of different frequency components of the signal.

[0054] The first detail implicit representation is an implicit representation of the signal details part extracted from the spectral analysis result. For example, in an electrical stimulation monitoring signal containing multiple frequency components, those frequency components that change rapidly or fluctuate locally may correspond to the subtle changes of neural activity or the transient characteristics of electrical stimulation. These information are summarized in the first detail implicit representation after spectral analysis. Technically, the computer system can identify these detail-related frequency components by setting a specific frequency range or frequency threshold, for example, considering the high frequency part (higher than a certain frequency threshold f th1 ) as the detail-related frequency component, and then performing special processing (such as weighting, normalization, etc.) on the spectral data corresponding to these frequency components to construct the first detail implicit representation.

[0055] The first global implicit representation is an implicit representation about the overall characteristics of the signal obtained from the spectral analysis result. It reflects the overall trend and main characteristics of the signal in a wide frequency range. For example, in the electrical stimulation monitoring signal, the overall energy distribution, the main frequency peak, etc. reflect the basic characteristics of the electrical stimulation and the overall response trend of the neural tissue to the electrical stimulation. The computer system can construct the first global implicit representation by smoothing the spectral data or performing special processing on the low frequency part (below a certain frequency threshold f th2 ) of the spectral data. For example, the spectral data is smoothed by using the moving average method, assuming that the spectral data is Y(k), and the moving average window size is M, then the smoothed spectral data is (when i is out of the range of 0 to N-1, appropriate boundary processing is performed), which can be used as the basis for constructing the first global implicit representation.

[0056] By obtaining the first detailed implicit representation and the first global implicit representation through spectral analysis, the first signal implicit representation can be analyzed from different angles, which helps to integrate the detailed influence variables and global influence variables in the subsequent process, and lays a foundation for further improving the accuracy of electrical stimulation monitoring data classification.

[0057] Step S300: integrating the first detailed implicit representation and the first global implicit representation according to the detailed influence variables and the global influence variables to obtain the second signal implicit representation, wherein the detailed influence variables and the global influence variables are obtained by training the noise electrical stimulation monitoring signal samples and the corresponding denoised electrical stimulation monitoring signal samples.

[0058] In step S300, the computer system integrates the first detailed implicit representation and the first global implicit representation obtained in step S200 according to the detailed influence variables and the global influence variables, thereby obtaining the second signal implicit representation.

[0059] The detailed influence variables and the global influence variables are obtained by training the noise electrical stimulation monitoring signal samples and the corresponding denoised electrical stimulation monitoring signal samples by the computer system. The noise electrical stimulation monitoring signal samples refer to the sample signals mixed with noise during the collection of the electrical stimulation monitoring signal, and the denoised electrical stimulation monitoring signal samples are ideal signal samples after removing the noise interference.

[0060] In the training process, the computer system can use machine learning or deep learning algorithms. For example, by constructing a regression model, assuming that the set of noise electrical stimulation monitoring signal samples is X = {x1, x2, …, x n}, and the set of corresponding denoised electrical stimulation monitoring signal samples is Y = {y1, y2, …, y n}. Assume a linear regression model y = β0 + β1, where β0 and β1 are parameters of the model. The computer system determines the values of β0 and β1 by minimizing a loss function (e.g., mean squared error ) to obtain model parameters that can reflect the relationship between the noise and the denoised signal, which can be further transformed into detail influence variables and global influence variables.

[0061] For the integration process of the first detail implicit representation and the first global implicit representation, assume that the first detail implicit representation is D, the first global implicit representation is G, the detail influence variable is α, and the global influence variable is β. The computer system can integrate in a weighted summation manner, i.e., the second signal implicit representation S = αD + βG.

[0062] To illustrate with a simple example of an electrical stimulation monitoring signal, suppose that the first detail implicit representation D contains the characteristics of small fluctuations in the electrical stimulation signal, which can be related to local activity of nerve cells; and the first global implicit representation G contains macroscopic characteristics such as overall intensity and frequency of the electrical stimulation signal. The detail influence variable α can represent the importance of the small fluctuation characteristics in the final integration, and the global influence variable β represents the importance of the macroscopic characteristics. Through the above weighted summation manner, the computer system integrates the two different levels of characteristic representations according to their importance, obtaining the second signal implicit representation S. This second signal implicit representation fuses the detail and global information of the signal, and can more comprehensively reflect the characteristics of the electrical stimulation monitoring signal than the first detail implicit representation or the first global implicit representation alone, providing a richer and more accurate signal representation basis for subsequent filtering and signal processing operations.

[0063] Step S400: filtering the second signal implicit representation through a first filter matrix to obtain a residual mapping signal, the residual mapping signal representing the error between the first electrical stimulation monitoring signal and the corresponding denoised electrical stimulation monitoring signal, the first filter matrix being obtained by training according to the noise electrical stimulation monitoring signal sample and the denoised electrical stimulation monitoring signal sample.

[0064] In step S400, the computer system filters the second signal implicit representation obtained in step S300 through the first filter matrix to obtain a residual mapping signal. The residual mapping signal represents the error between the first electrical stimulation monitoring signal and the corresponding denoised electrical stimulation monitoring signal.

[0065] The first filtering matrix is ​​obtained by training the computer system using both noisy and denoised electrical stimulation monitoring signal samples. During training, the computer system may employ various techniques. For example, the least squares method can be used to construct this filtering matrix. Assume that the noisy electrical stimulation monitoring signal samples form a matrix X (each column being a sample), the corresponding denoised electrical stimulation monitoring signal samples form a matrix Y, and there exists a filtering matrix A such that Y≈AX. Solving for A using the least squares method, i.e., minimizing the objective function... (where n is the number of samples, y i x is the i-th sample in Y. i (This is the i-th sample in X). By solving this optimization problem, the first filtering matrix A can be obtained.

[0066] The implicit representation of the second signal is a signal representation that integrates detailed and global information through the preceding steps. Assume the implicit representation of the second signal is a vector s = [s1, s2, ..., s...]. m (where m is the dimension of the vector), the first filtering matrix is ​​A = [a ij ] n×m (n is the number of rows in the filter matrix). The computer system performs filtering operations through matrix multiplication, obtaining the filtered result vector r = As = [r1, r2, ..., r n The resulting vector r is the residual mapping signal.

[0067] Taking a simple electrical stimulation monitoring signal scenario as an example, the first electrical stimulation monitoring signal may contain noise components due to various reasons (such as noise from the acquisition equipment, environmental interference, etc.), while the corresponding denoised electrical stimulation monitoring signal is an ideal signal free of noise interference. The second signal implicitly represents the integrated signal feature information. When the computer system uses the first filtering matrix to filter the implicit representation of the second signal, the resulting residual mapping signal reflects the difference between the two. For example, if a frequency component in the first electrical stimulation monitoring signal is shifted or its amplitude changes due to noise, this difference will be reflected in the residual mapping signal compared to the denoised electrical stimulation monitoring signal. This difference information is crucial for subsequent signal processing and data classification, helping the computer system to more accurately adjust and optimize processing strategies to improve the accuracy of electrical stimulation monitoring data classification.

[0068] Step S500: Integrate the first electrical stimulation monitoring signal with the residual mapping signal to obtain the second electrical stimulation monitoring signal. The signal-to-noise ratio of the second electrical stimulation monitoring signal is greater than that of the first electrical stimulation monitoring signal.

[0069] In step S500, the computer system integrates the first electrical stimulation monitoring signal and the residual mapping signal to obtain a second electrical stimulation monitoring signal, and a signal-to-noise ratio of the second electrical stimulation monitoring signal is greater than a signal-to-noise ratio of the first electrical stimulation monitoring signal.

[0070] The first electrical stimulation monitoring signal is an original electrical stimulation monitoring signal obtained in an initial monitoring process. This signal contains various information, including but not limited to actual parameter information of the electrical stimulation and possible noise components. For example, in an actual application scenario of neuromodulation, the first electrical stimulation monitoring signal can be a response signal of a neural tissue to an electrical stimulation collected by an implantable electrode. Due to the influence of factors such as the collection environment and the device itself, a certain degree of noise is inevitably mixed in, which can be manifested as a slight fluctuation in the signal, an abnormal frequency component, or an irregular amplitude change, etc.

[0071] The residual mapping signal is a result of filtering the second signal implicit representation by the first filter matrix in step S400, and it represents the error between the first electrical stimulation monitoring signal and the corresponding denoised electrical stimulation monitoring signal. In a mathematical sense, the residual mapping signal contains the part of information in the first electrical stimulation monitoring signal that deviates from the ideal denoised signal. For example, if the first electrical stimulation monitoring signal is regarded as being composed of an ideal signal (a signal after denoising) and a noise signal, then the residual mapping signal reflects the degree and manner of the influence of the noise signal on the ideal signal.

[0072] When performing the integration operation, the computer system can adopt various technical means. One feasible way is the linear addition method. Assuming that the first electrical stimulation monitoring signal is x(n) and the residual mapping signal is r(n) (where n represents the serial number of discrete sampling points), then the second electrical stimulation monitoring signal y(n) after integration can be represented as y(n) = x(n) + k × r(n), where k is a weighting coefficient. The determination of this weighting coefficient k depends on the specific signal characteristics and processing goals. If k = 1, it is a simple addition operation; if k is adjusted according to certain statistical characteristics or prior knowledge of the signal, for example, the value of k is determined according to the energy ratio of the first electrical stimulation monitoring signal and the residual mapping signal, then the signal integration can be more flexible.

[0073] From the perspective of improving the signal-to-noise ratio, since the residual mapping signal reflects the error part in the first electrical stimulation monitoring signal, through appropriate integration, the computer system can correct the noise in the original signal to some extent. For example, if there is obvious noise interference in a certain frequency band of the first electrical stimulation monitoring signal, resulting in a very low signal-to-noise ratio in that frequency band, and the residual mapping signal accurately reflects the noise error situation of this frequency band, then in the integration process, the residual mapping signal can be used to compensate for the noise in that frequency band in the opposite direction, thereby improving the signal-to-noise ratio of that frequency band and the entire signal.

[0074] For another example, assume that the first electrical stimulation monitoring signal is a complex signal containing multiple frequency components, in which the low-frequency part has some low-frequency noise due to power interference, and the high-frequency part has some high-frequency noise due to electromagnetic environmental interference. In the previous step, the residual mapping signal is obtained through a series of processing of the first electrical stimulation monitoring signal, which can accurately indicate the specific form of the low-frequency and high-frequency noise in the signal (such as amplitude, phase, etc. deviation). When the computer system performs integration, it can correct the noise in the first electrical stimulation monitoring signal according to this information. For the low-frequency part, the amplitude and phase of the low-frequency part can be adjusted according to the information about the low-frequency noise in the residual mapping signal to reduce the impact of low-frequency noise; the same applies to the high-frequency part. In this way, the signal-to-noise ratio of the second electrical stimulation monitoring signal in the low-frequency and high-frequency parts is improved, and the signal-to-noise ratio of the entire signal is also improved accordingly.

[0075] In addition, the computer system can also use an adaptive filtering-based method for integration. In this case, the weighting coefficient k is not a fixed value, but is dynamically adjusted according to the real-time characteristics of the signal. For example, define an adaptive algorithm to dynamically adjust the value of k according to the cross-correlation function of the first electrical stimulation monitoring signal and the residual mapping signal. If the cross-correlation function of the first electrical stimulation monitoring signal and the residual mapping signal shows that the noise correlation between the two is strong within a certain period of time, then the value of k is increased to enhance the correction effect of the noise; on the contrary, if the cross-correlation function shows that the noise correlation is weak, then the value of k is appropriately reduced to avoid overcorrection.

[0076] Step S500 effectively improves the signal-to-noise ratio of the second electrical stimulation monitoring signal by integrating the first electrical stimulation monitoring signal with the residual mapping signal, using the error information contained in the residual mapping signal, and provides a higher quality signal basis for subsequent electrical stimulation monitoring data classification operations, which helps to improve the accuracy and reliability of classification.

[0077] Step S600: Classify the second electrical stimulation monitoring signal to obtain a signal classification result.

[0078] In step S600, the computer system classifies the second electrical stimulation monitoring signal obtained in step S500. The classification of the electrical stimulation monitoring signal is a complex process, which involves dividing the signal into different categories according to various features of the signal. For the second electrical stimulation monitoring signal, it has been processed in the previous series of steps, and the signal-to-noise ratio has been improved, which makes the features related to neuromodulation more obvious, which is conducive to the accurate classification of the computer system.

[0079] The computer system can use various technical means to classify. One of the feasible methods is the classification technology based on machine learning algorithm. For example, the support vector machine (SVM) algorithm. Assuming that the second electrical stimulation monitoring signal can be represented as a feature vector x = (x1, x2, …, x n ), where x i may be the value of the signal in different dimensions (such as the amplitude, phase, and other features of different frequency components). In the training phase, the computer system has a set of training samples (x i , y i ) of known categories, where y i is the corresponding category label (for example, in neuromodulation, it may represent different excitation states of nerve cells or different types of neural responses, etc.). The goal of SVM is to find a hyperplane w·x+b=0 (where w = (w1, w2, …, w n ) is the normal vector of the hyperplane, and b is the bias term) that can separate samples of different categories as much as possible. Specifically, it is to maximize the minimum distance (called the margin) of two classes of samples to the hyperplane. This optimization problem can be expressed as:

[0080]

[0081] When classifying the second electrical stimulation monitoring signal, the computer system substitutes the signal feature vector x into the trained SVM model to calculate f(x) = sign(w·x+b). According to the value of f(x) (+1 or -1, etc., depending on the specific category definition), the category to which the signal belongs is determined.

[0082] Another classification technique is the neural network-based method. For example, a multilayer perceptron (MLP) is used. The feature vector of the second electrical stimulation monitoring signal is taken as the input layer of the MLP. Assuming that the input layer has n neurons, corresponding to the n features of the signal. After passing through several hidden layers, the number of neurons in each hidden layer and the activation function (such as the ReLU function f(x) = max(0, x)) can be set according to the specific situation. The number of neurons in the final output layer is the same as the number of categories to be classified. In the training phase, a set of training samples (x_i, y_i) is used to adjust the weights of the neural network through the backpropagation algorithm. The backpropagation algorithm is based on minimizing the loss function (such as the cross-entropy loss function L = -∑ i y i log(p(y i |x i )) where p(y i |x i ) is the probability of predicting y_i when the input is x_i) to optimize the network weights. When classifying the second electrical stimulation monitoring signal, the computer system inputs the signal into the trained MLP, and the category corresponding to the maximum activation value of the output layer neuron is the classification result of the signal.

[0083] For another example, in the actual scenario of neuromodulation, if the purpose of electrical stimulation is to distinguish the normal excitation state and the abnormal excitation state of nerve cells (the two states as different categories). For the second electrical stimulation monitoring signal, some of its features may be related to the membrane potential changes of nerve cells, ion channel activities, etc. If SVM classification is used, the computer system finds a best hyperplane to distinguish the signal features corresponding to the normal and abnormal states by analyzing a large number of training samples with known normal and abnormal states. When a new second electrical stimulation monitoring signal arrives, its feature vector is substituted into the SVM model, and the model output is used to determine whether the signal corresponds to the normal excitation state or the abnormal excitation state of the nerve cells.

[0084] Similarly, if MLP is used for classification, in the training process, the computer system learns the complex mapping relationship between signal features and classification results under normal and abnormal excitation states. For a new second electrical stimulation monitoring signal, its features are processed layer by layer through the MLP, and finally the probability distribution of normal and abnormal excitation states is obtained in the output layer. The category with the highest probability is the classification result of the signal.

[0085] In addition to the above two methods, a decision tree algorithm can also be used for classification. A decision tree is a tree-structured classification model that makes a series of conditional judgments based on different features of the signal to determine the class of the signal. For example, for the second electrical stimulation monitoring signal, the root node of the decision tree can make a judgment based on the amplitude of a certain main frequency component of the signal, if the amplitude is greater than a certain threshold, then continue to judge the next feature (such as the phase of another frequency component) along one branch, if it is less than the threshold, then continue to judge other features along another branch. Through such layer-by-layer branch judgment, finally reach the leaf node, the class corresponding to the leaf node is the classification result of the signal.

[0086] In step S600, the computer system accurately divides the second electrical stimulation monitoring signal into the corresponding category according to the features highlighted after processing, thereby providing an important basis for subsequent analysis, diagnosis or intervention of wireless implantable neural regulation, but the embodiments of the present application do not involve specific disease diagnosis.

[0087] In one embodiment, step S200, the spectral analysis of the first signal implicit representation, obtaining the first detail implicit representation and the first global implicit representation, can include:

[0088] Step S210: classifying the first signal implicit representation into a second detail implicit representation and a second global implicit representation according to the spectrum, integrating the second detail implicit representation and the second global implicit representation to obtain a third signal implicit representation;

[0089] Step S220: performing detail extraction on the third signal implicit representation to obtain the first detail implicit representation, and performing smoothing on the third signal implicit representation to obtain the first global implicit representation.

[0090] In step S210, the computer system classifies the first signal implicit representation into a second detail implicit representation and a second global implicit representation according to the spectrum, and then integrates the two to obtain a third signal implicit representation.

[0091] For the operation of "classifying according to the spectrum", the spectrum is a description of the characteristics of the signal in the frequency domain. The computer system can use spectral analysis techniques, such as Discrete Fourier Transform (DFT). Assuming that the first signal implicit representation is a discrete signal x(n), n = 0, 1, …, N-1, its DFT is defined as After obtaining the frequency spectrum of the signal by DFT, the computer system classifies the signal according to the characteristics of the frequencies. For example, the high-frequency part is classified as the second detailed implicit representation, because the high-frequency part often contains the fast-changing information in the signal, such as the subtle changes in the fast response of nerve cells to stimulation, etc. in the electrical stimulation monitoring signal; and the low-frequency part is classified as the second global implicit representation, which may more reflect the overall trend of the electrical stimulation, such as the basic frequency and intensity of the electrical stimulation, etc. macroscopic characteristics.

[0092] Assuming that the second detailed implicit representation is D_2 and the second global implicit representation is G_2 after classification, the computer system integrates them to obtain the third signal implicit representation S_3. The technical means of integration can be a simple weighted sum, for example, S3 = αD2 + βG2, where α and β are weighting coefficients, which can be determined by the computer system according to the prior knowledge of the signal or by statistical analysis of a large number of samples.

[0093] Then, in step S220, the computer system extracts the details of the third signal implicit representation to obtain the first detailed implicit representation, and smooths the third signal implicit representation to obtain the first global implicit representation.

[0094] In the aspect of detail extraction, the computer system can use technical means such as difference operation. For example, for the third signal implicit representation S3, if it is a discrete sequence s3(n), n = 0, 1, …, M-1, its first-order difference Δs3(n) = s3(n+1) - s3(n), n = 0, 1, …, M-2 can be calculated. This first-order difference sequence can be used as part of the first detailed implicit representation, which highlights the local change information in the signal. In the electrical stimulation monitoring signal, such local changes may correspond to subtle fluctuations in nerve cell activity, etc. detailed information.

[0095] For smoothing operation, the computer system can use moving average method. Assuming that the third signal implicit representation S3 is a sequence s3(n), and the moving average window size is L, then the smoothed sequence (appropriate boundary processing is performed when i is out of the range of 0 to M-1), this smoothed sequence can be used as the first global implicit representation. It can remove some small fluctuations in the signal, so as to more clearly show the overall trend of the signal, just like removing some fluctuations caused by noise or other small disturbances in the electrical stimulation monitoring signal, to highlight the overall characteristics of the electrical stimulation, such as the overall frequency range and intensity distribution, etc.

[0096] Through the operations of steps S210-S220, the computer system starts from the first signal implicit representation, classifies and integrates according to the frequency spectrum, and then respectively performs detail extraction and smoothing operation, so as to obtain the first detail implicit representation and the first global implicit representation, which provides an important basis for further processing in the subsequent wireless implantable neuromodulation electric stimulation monitoring data classification method.

[0097] In an implementation manner, the method can further include:

[0098] Step S201: filtering the third signal implicit representation by a second filter matrix to obtain a smoothing matrix, the second filter matrix being obtained by training according to the noise electric stimulation monitoring signal and the denoised electric stimulation monitoring signal;

[0099] Step S202: performing matrix subtraction calculation on the unit filter matrix and the smoothing matrix to obtain a detail extraction matrix, the dimension number of the unit filter matrix being equal to the dimension number of the smoothing matrix, and the element values of one dimension of the unit filter matrix being all 1 and the element values of other dimensions being all 0;

[0100] Based on this, step S220, the detail extraction is performed on the third signal implicit representation to obtain the first detail implicit representation, and the smoothing is performed on the third signal implicit representation to obtain the first global implicit representation, including:

[0101] Step S221: performing detail extraction on the third signal implicit representation by the detail extraction matrix to obtain the first detail implicit representation, and performing smoothing on the third signal implicit representation by the smoothing matrix to obtain the first global implicit representation.

[0102] In step S201, the computer system filters the third signal implicit representation by the second filter matrix, so as to obtain the smoothing matrix.

[0103] The third signal implicit representation is a signal representation form obtained through the previous steps (such as the spectral classification and integration of the first signal implicit representation). This representation form has extracted and converted the characteristics of the original signal to a certain extent, and contains information related to the original electric stimulation monitoring signal, but in an implicit and processed form.

[0104] The second filter matrix is obtained by training according to the noise electric stimulation monitoring signal and the denoised electric stimulation monitoring signal. In the training process, the computer system determines the element values of the second filter matrix by analyzing the relationship between a large number of noise electric stimulation monitoring signals (which contain unwanted interference components) and the corresponding denoised electric stimulation monitoring signals (which are ideal signals without noise interference). For example, assuming that there is a set of noise electric stimulation monitoring signal samples X n ={x n1x n2 ,…,x nm}(where m is the number of samples) and the corresponding denoised electrical stimulation monitoring signal samples X d ={x d1 ,x d2 ,…,x dm}. The computer system can employ an optimization algorithm such as least squares to find a matrix F (i.e., the second filter matrix) such that X d ≈FX n . Specifically, for least squares, the goal is to minimize The second filter matrix F is obtained by solving this optimization problem.

[0105] The smoothed matrix is the result matrix obtained after the filtering operation. Its role is to smooth the third signal implicit representation, and its essence is to weight, adjust, etc. some characteristics in the signal through filtering operation, so that the signal changes more smoothly in some aspects (such as frequency, amplitude, etc.), and removes some local irregular fluctuations that may be caused by noise.

[0106] Suppose the third signal implicit representation is a two-dimensional matrix where m and n represent the number of rows and columns of the matrix respectively, and each element s_{ij} represents a characteristic value of the signal in a certain dimension. If the second filter matrix (where q and p are the number of rows and columns respectively, and satisfy the dimension requirement of matrix multiplication), then the computer system obtains the smoothed matrix M by matrix multiplication M = FS. For example, if some columns in the third signal implicit representation correspond to high-frequency components in the signal (which may contain noise or small fluctuations that are not needed), after multiplication with the second filter matrix, the amplitude corresponding to these high-frequency components in the smoothed matrix may be reduced, thereby achieving the smoothing effect of the signal.

[0107] In step S202, the computer system performs matrix subtraction between the identity filter matrix and the smoothed matrix, thereby obtaining the detail extraction matrix.

[0108] The identity filter matrix has the same number of dimensions as the smoothed matrix, and the element value of one of its dimensions is 1, and the element value of the other dimensions is 0. The role of the identity filter matrix in this operation is to provide a reference for extracting information related to details from the smoothed matrix. For example, if the smoothed matrix is two-dimensional then the corresponding identity filter matrix can be (here is a simple form, the specific form is determined according to the actual dimension and operation requirement).

[0109] The detail extraction matrix is obtained by subtracting the smoothing matrix from the identity filter matrix, and its purpose is to be used for subsequent extraction of detail information from the third signal implicit representation. These detail information can be those rapidly changing, local features in the original signal, corresponding to the overall features after smoothing. For example, in the electrostimulation monitoring signal, the smoothed signal can represent the overall electrostimulation trend, while the detail information can be the tiny fluctuations generated by the instantaneous response of nerve cells to electrostimulation, etc.

[0110] Assuming the smoothing matrix The identity filter matrix Then the detail extraction matrix This detail extraction matrix can be used for subsequent operations to extract those detail features suppressed by the smoothing operation from the third signal implicit representation. For example, if the elements in the third signal implicit representation represent the energy values of the electrostimulation monitoring signal at different time points and frequencies, then the detail extraction matrix can help the computer system to recover the local energy fluctuation information that may be lost in the smoothing process, which is very important for a comprehensive understanding of the electrostimulation monitoring signal.

[0111] In step S221, the computer system performs detail extraction on the third signal implicit representation by the detail extraction matrix to obtain a first detail implicit representation, and performs smoothing on the third signal implicit representation by the smoothing matrix to obtain a first global implicit representation.

[0112] The first detail implicit representation is a form of implicit representation about the detail part of the signal extracted from the third signal implicit representation. These detail parts usually contain local changes, rapid fluctuations, etc. in the signal, which are very important for understanding the tiny changes in the electrostimulation monitoring signal (such as the instantaneous response of nerve cells to electrostimulation, etc.).

[0113] The first global implicit representation is a form of implicit representation reflecting the overall features of the signal after smoothing the third signal implicit representation. The global implicit representation focuses more on the overall trend, main frequency components, etc. of the signal macro features, such as the overall intensity, frequency range, etc. of the electrostimulation.

[0114] For detail extraction by the detail extraction matrix, assuming that the third signal implicit representation is S, and the detail extraction matrix is D, the computer system can perform the detail extraction operation by matrix multiplication. Specifically, the first detail implicit representation D I= DS. Here the matrix multiplication is performed according to the rules of matrix operations, and the elements in the third signal implicit representation are weighted and combined by the elements in the detail extraction matrix, so as to highlight the detail features in the signal. For example, if some elements in the detail extraction matrix are negative numbers, then when multiplied by the third signal implicit representation, the corresponding elements will be inversely weighted, so as to extract the detail features opposite to the smoothed features.

[0115] For the smoothing operation by the smoothing matrix, assuming that the third signal implicit representation is S, and the smoothing matrix is M, the computer system obtains the first global implicit representation by matrix multiplication G I = MS. In this process, the elements in the smoothing matrix weight the elements in the third signal implicit representation, so that the local fluctuations in the signal are smoothed out, thereby obtaining the first global implicit representation reflecting the overall features. For example, if the elements in the smoothing matrix have the characteristics of low-pass filtering (i.e., larger weight for low-frequency components and smaller weight for high-frequency components), then after multiplication with the third signal implicit representation, the high-frequency components (which may be noise or local minor fluctuations) in the signal will be suppressed, thereby obtaining a smoother first global implicit representation.

[0116] Assuming that the third signal implicit representation Detail extraction matrix

[0117] Smoothing matrix

[0118] Detail extraction by the detail extraction matrix, the first detail implicit representation Here it can be seen that the detail extraction matrix inversely weights the elements in the third signal implicit representation, highlighting the detail features in the signal (here it is a simple inverse relationship, and in actual cases it can be more complex).

[0119] Smoothing by the smoothing matrix, the first global implicit representation It can be seen that the smoothing matrix inversely weights the elements in the third signal implicit representation, so that the amplitude of the signal is reduced, which plays a smoothing role and better reflects the overall features of the signal.

[0120] Through the above steps S201-S202 and the operation of step S221, the computer system can extract the first detail implicit representation and the first global implicit representation from the third signal implicit representation, respectively, which is of great significance for further analyzing and processing the signal in the subsequent wireless implantable neural modulation electric stimulation monitoring data classification method. These representation forms can more comprehensively reflect the features of the electric stimulation monitoring signal, including both local detail features and overall macro features, thereby providing a richer information basis for accurate classification and other operations.

[0121] In an implementation, the number of smoothing matrices is s, the number of dimensions of each smoothing matrix is m, the number of detail extraction matrices is s, and the number of dimensions of each detail extraction matrix is m. Based on this, the step S221 of performing detail extraction on the third signal implicit representation by the detail extraction matrices to obtain a first detail implicit representation, and performing smoothing on the third signal implicit representation by the smoothing matrices to obtain a first global implicit representation, can include:

[0122] Step S2211: classifying the third signal implicit representation in dimensions into s third signal sub-implicit representations, wherein the number of sampling points covered by the smoothing matrices, the number of sampling points covered by the detail extraction matrices, and the number of sampling points covered by the third signal sub-implicit representations are consistent;

[0123] Step S2212: performing detail extraction on the s third signal sub-implicit representations by the s detail extraction matrices respectively to obtain s first detail sub-implicit representations, and the s first detail sub-implicit representations constitute the first detail implicit representation;

[0124] Step S2213: performing smoothing on the s third signal sub-implicit representations by the s smoothing matrices respectively to obtain s first global sub-implicit representations, and the s first global sub-implicit representations constitute the first global implicit representation.

[0125] In step S2211, the computer system classifies the third signal implicit representation in dimensions into s third signal sub-implicit representations. The third signal implicit representation is a signal representation form obtained after a series of previous operations (such as spectral classification, integration, etc.). This representation form contains the characteristic information of the processed electric stimulation monitoring signal in an implicit way. Compared with the original signal, it has undergone certain feature extraction and conversion, which is convenient for further analysis according to different needs. For example, when performing spectral analysis on the electric stimulation monitoring signal, the characteristics related to different frequency ranges or frequency components may be integrated into this representation form. It can be in the form of a matrix or a high-dimensional vector, and the elements in it represent the numerical values of the signal in different characteristic dimensions.

[0126] Classifying in dimensions means dividing the third signal implicit representation into several sub-portions according to the dimension characteristics of the signal representation. Here, the dimension can have various meanings. For example, if the third signal implicit representation is a matrix, the dimension can refer to the row or column; if it is a high-dimensional vector, the dimension can refer to different components of the vector. By classifying in dimensions, the computer system can decompose the entire signal representation into smaller sub-portions with certain independence according to certain rules for subsequent processing.

[0127] These are the sub-parts obtained by dimension-wise categorization. Each third signal sub-implicit representation inherits the partial features of the third signal implicit representation and can be regarded as a representation of the features of the original signal in a certain local dimension range. For example, if the third signal implicit representation is a matrix representing the frequency and amplitude features of the electrical stimulation monitoring signal, the third signal sub-implicit representations obtained by dimension-wise categorization can correspond to the signal features in different frequency bands or amplitude ranges, respectively.

[0128] Suppose the third signal implicit representation is a matrix A with dimensions m x n. If it is to be dimension-wise categorized into s third signal sub-implicit representations, the computer system can divide it by rows or columns. For example, if it is divided by rows, when m is divisible by s, each third signal sub-implicit representation A i (i = 1, 2, …, s) has dimensions The computer system can achieve this division through simple indexing operations, that is, Here, the indexing operation is performed according to the row indexing rule of the matrix, and the sub-matrix, that is, the third signal sub-implicit representation, is obtained by specifying the starting row and ending row.

[0129] Suppose the third signal implicit representation and s = 2. If it is dimension-wise categorized by rows, the first third signal sub-implicit representation and the corresponding index calculation is A_1 = A[1:2, :]; the second third signal sub-implicit representation and the corresponding index calculation is A2 = A[3:4, :]. Each sub-implicit representation here contains partial row information of the original matrix, reflecting the features of the original signal in different local dimension ranges. If the rows of the matrix in this example represent the features of the electrical stimulation monitoring signal at different time points, this division is equivalent to dividing the signal features in the entire time range into two sub-parts, facilitating subsequent detailed extraction and smoothing operations.

[0130] In step S2212, the computer system performs detailed extraction on the s third signal sub-implicit representations through s detailed extraction matrices, respectively, to obtain s first detailed sub-implicit representations, which constitute the first detailed implicit representation.

[0131] The detail extraction matrix is calculated by matrix subtraction between the unit filter matrix and the smoothing matrix in the previous step (e.g., step S202). The purpose of the detail extraction matrix is to extract the detail features from the signal representation, and its element values are determined according to the relationship between the noisy electrical stimulation monitoring signal and the denoised electrical stimulation monitoring signal. The elements in the detail extraction matrix weight different features in the signal to highlight those local, fast-changing features that can be suppressed by the smoothing operation, such as the tiny fluctuations in the electrical stimulation monitoring signal or the local changes caused by the instantaneous response of neural cells to electrical stimulation, etc.

[0132] The first detail sub-implicit representation is the result of detail extraction on the third signal sub-implicit representation by the detail extraction matrix. Each first detail sub-implicit representation reflects the detail features in the corresponding third signal sub-implicit representation and is a representation of the detail features of the original signal in the local dimension range. All the first detail sub-implicit representations are combined to obtain the first detail implicit representation, which can comprehensively reflect the detail information in the original signal.

[0133] The first detail implicit representation is an overall representation form composed of s first detail sub-implicit representations. It contains all the detail features extracted from the third signal implicit representation and is a comprehensive representation of the local changes, fast fluctuations, and other detail information in the original electrical stimulation monitoring signal, which helps to better understand the tiny changes and special situations (such as the instantaneous activity of neural cells) in the electrical stimulation monitoring signal.

[0134] Suppose the ith third signal sub-implicit representation is A i (i = 1, 2, …, s), and the corresponding detail extraction matrix is D i The computer system performs the detail extraction operation by matrix multiplication, i.e., the ith first detail sub-implicit representation B i = D i A i The matrix multiplication here is performed according to the conventional matrix operation rules, and the elements in the detail extraction matrix D i weight and combine the elements in the third signal sub-implicit representation A i to highlight the detail features therein.

[0135] Suppose s = 2, the first third signal sub-implicit representation corresponds to the detail extraction matrix Then the first first detail sub-implicit representation The second third signal sub-implicit representation corresponds to the detail extraction matrix Then the second first detail sub-implicit representation The two first detail sub-implicit representations B1 and B2 constitute a first detail implicit representation, which comprehensively reflects the detail characteristics in the original third signal implicit representation, for example, in this example, through the weighting operation of the detail extraction matrix, the inverse relationship of certain elements in the original signal or the local amplitude change and other detail information are highlighted.

[0136] In step S2213, the computer system smoothes the s third signal sub-implicit representations respectively through s smoothing matrices, to obtain s first global sub-implicit representations, which constitute a first global implicit representation.

[0137] The smoothing matrix is obtained by filtering the third signal implicit representation through the second filter matrix in the previous step (such as step S201). The function of the smoothing matrix is to smooth the signal, and the element value is determined according to the relationship between the noise electric stimulation monitoring signal and the denoised electric stimulation monitoring signal. The elements in the smoothing matrix weight different characteristics in the signal to reduce local fluctuations in the signal and highlight overall trends and main characteristics of the signal, such as overall intensity, frequency range and other macro characteristics of the electric stimulation.

[0138] The first global sub-implicit representation is obtained by smoothing the third signal sub-implicit representation through the smoothing matrix. Each first global sub-implicit representation reflects the overall characteristics in the corresponding third signal sub-implicit representation, and is a representation of the overall characteristics of the original signal in the local dimension range. Combining all the first global sub-implicit representations obtains the first global implicit representation, which can comprehensively reflect the overall information in the original signal.

[0139] The first global implicit representation is a global representation form composed of s first global sub-implicit representations. It contains all the overall characteristics extracted from the third signal implicit representation, and is a comprehensive representation of the overall trends, main frequency components and other macro information in the original electric stimulation monitoring signal. It is very important for understanding the overall characteristics of the electric stimulation monitoring signal (such as the basic mode of the electric stimulation, the overall response of the neural tissue to the electric stimulation, etc.).

[0140] Suppose the i-th third signal sub-implicit representation is A i (i = 1, 2, …, s), and the corresponding smoothing matrix is M i . The computer system performs smoothing operation through matrix multiplication, that is, the i-th first global sub-implicit representation C i = M i A i . The matrix multiplication here is performed according to the conventional matrix operation rules. The elements in the smoothing matrix M i weight the third signal sub-implicit representation A iThe elements in the first global implicit representation are weighted combined, so as to smooth out the local fluctuations and highlight the overall characteristics.

[0141] Assuming s = 2, the first third signal sub-implicit representation is The corresponding smoothing matrix is Then the first first global sub-implicit representation is The second third signal sub-implicit representation is The corresponding smoothing matrix is Then the second first global sub-implicit representation is The two first global sub-implicit representations C1 and C2 constitute the first global implicit representation, which comprehensively reflects the overall characteristics of the original third signal implicit representation, for example, in this example, through the weighted operation of the smoothing matrix, the amplitude of the original signal is reduced, the local small fluctuations are smoothed out, and the overall trend characteristics are highlighted.

[0142] Through the operations of steps S2211-S2213, the computer system can effectively extract the first detail implicit representation and the first global implicit representation from the third signal implicit representation, which is beneficial to comprehensively analyze the characteristics (including detail characteristics and overall characteristics) of the electrical stimulation monitoring signal, and provides a rich information basis for further processing and decision-making in the subsequent electrical stimulation monitoring data classification method of wireless implantable neural regulation.

[0143] In another embodiment, the step S200 of performing frequency spectrum analysis on the first signal implicit representation to obtain the first detail implicit representation and the first global implicit representation can include:

[0144] Step S200A: performing frequency spectrum analysis on the first signal implicit representation to obtain a pending detail implicit representation and a pending global implicit representation, and integrating the pending detail implicit representation and the pending global implicit representation according to the detail influence variable and the global influence variable to obtain a pending signal implicit representation;

[0145] Step S200B: performing a new frequency spectrum analysis on the pending signal implicit representation to obtain a new one pending detail implicit representation and a new one pending global implicit representation, and integrating the new one pending detail implicit representation and the new one pending global implicit representation according to the detail influence variable and the global influence variable to obtain a new one pending signal implicit representation, until the next round of frequency spectrum analysis is not less than the first preset value, and the pending detail implicit representation obtained at the next time is determined as the first detail implicit representation, and the pending global implicit representation obtained at the next time is determined as the first global implicit representation.

[0146] In step S200A, the computer system performs spectral analysis on the first signal implicit representation to obtain a pending detail implicit representation and a pending global implicit representation, and integrates the pending detail implicit representation and the pending global implicit representation according to the detail influence variable and the global influence variable to obtain a pending signal implicit representation.

[0147] In the spectral analysis of the first signal implicit representation, the computer system can use fast Fourier transform (FFT) technique. Assuming that the first signal implicit representation is a discrete time series x[n], n = 0, 1, …, N-1, the FFT calculation is With the spectral information obtained by FFT, the computer system can decompose it into the pending detail implicit representation and the pending global implicit representation. For example, in the spectrum, the higher frequency part can correspond to the pending detail implicit representation, because the higher frequency component often contains the detail information of the fast change, local fluctuation, etc. in the signal. In the electrical stimulation monitoring signal, this can represent the instantaneous, slight response change of the neural cells to the electrical stimulation. The lower frequency part can be classified as the pending global implicit representation, which reflects the overall trend of the signal, such as the basic frequency and overall intensity of the electrical stimulation, etc. macroscopic characteristics. The detail influence variable and the global influence variable are obtained by training according to the noise electrical stimulation monitoring signal sample and the corresponding denoised electrical stimulation monitoring signal sample before. For example, in the training process, the computer system can establish a regression model. Let the noise electrical stimulation monitoring signal sample set be {x i}, the corresponding denoised electrical stimulation monitoring signal sample set be {y i}, and assuming that a linear regression model y = β0+ β1x is used, the values of β0and β1are determined by minimizing the loss function (such as mean square error ), and these parameters can be converted into the detail influence variable and the global influence variable after processing.

[0148] According to these variables, the pending detail implicit representation D t and the pending global implicit representation G t are integrated to obtain a pending signal implicit representation S t , for example, a weighted sum can be used S t = αD t + βG t , where α is the weight converted from the detail influence variable, and β is the weight converted from the global influence variable.

[0149] Then, step S200B, the computer system performs a new round of spectral analysis on the pending signal implicit representation, obtains a new pending detail implicit representation and a new pending global implicit representation, integrates the new pending detail implicit representation and the new pending global implicit representation according to the detail influence variable and the global influence variable, and obtains a new pending signal implicit representation, until the next round of spectral analysis is not less than the first preset value, the pending detail implicit representation obtained at the next time is determined as the first detail implicit representation, and the pending global implicit representation obtained at the next time is determined as the first global implicit representation.

[0150] When the spectral analysis is performed again, the FFT technique can also be used. Each time of spectral analysis can re-decompose the pending detail implicit representation and the pending global implicit representation. For example, with the increase of the round of spectral analysis, the decomposition of the pending signal implicit representation will be more refined. Assuming that in the first round, the pending detail implicit representation contains relatively broad high-frequency fluctuation information, with the increase of the round, the pending detail implicit representation in the subsequent round can further subdivide the more detailed fluctuation information of different frequency bands.

[0151] When integrating according to the detail influence variable and the global influence variable in each round, a similar weighted sum method is still used. For example, the weight of each round can be fine-tuned according to the analysis results of the previous round. If it is found in a certain round that some features in the pending detail implicit representation have a greater impact on the final result than previously estimated, the weight corresponding to the conversion of the detail influence variable can be increased.

[0152] When the round of spectral analysis reaches the first preset value, the pending detail implicit representation and the pending global implicit representation obtained at this time are determined as the first detail implicit representation and the first global implicit representation. The determination of this preset value can be based on prior knowledge of the electrical stimulation monitoring signal or through statistical analysis of a large amount of experimental data. For example, if it is found through multiple experiments that for a certain type of electrical stimulation monitoring signal, the result obtained after 5 rounds of spectral analysis performs best in terms of subsequent classification accuracy, etc., the first preset value can be set to 5.

[0153] Through this step-by-step iterative spectral analysis and integration process, the computer system can more accurately obtain the first detail implicit representation and the first global implicit representation from the first signal implicit representation, and provide more accurate and more representative signal feature information for subsequent processing in the electrical stimulation monitoring data classification method of wireless implantable neural modulation.

[0154] In an embodiment, in step S200A, integrating the pending detail implicit representation and the pending global implicit representation according to the detail influence variable and the global influence variable to obtain the pending signal implicit representation can include:

[0155] Step S200A1: If the next round of spectral analysis is not greater than the second preset value, the computer system fuses the pending detail implicit representation and the pending global implicit representation obtained in the next round according to the detail influence variable and the global influence variable, to obtain a first control signal implicit representation, the second preset value is less than the first preset value;

[0156] Step S200A2: The computer system up-samples or down-samples the first electric stimulation monitoring signal, and performs feature implicit representation on the up-sampled or down-sampled first electric stimulation monitoring signal to obtain a second control signal implicit representation.

[0157] Step S200A3: The computer system performs feature implicit representation on the first control signal implicit representation and the second control signal implicit representation by a weight focusing strategy to obtain a pending signal implicit representation in the next round.

[0158] In step S200A1, the pending detail implicit representation is a representation form related to signal details obtained by the computer system after performing spectral analysis on the first signal implicit representation. These details can include micro fluctuations, local changes or high-frequency components in the signal, etc. For example, in the electric stimulation monitoring signal, the pending detail implicit representation can include the micro changes of the electric signal produced by the reaction of the nerve cells to the electric stimulation, which can be represented as high-frequency parts or rapid fluctuations in the time domain in the frequency spectrum.

[0159] The pending global implicit representation is a representation form related to the overall characteristics of the signal obtained after spectral analysis. It reflects the overall trend of the signal, such as the basic frequency, overall intensity or low-frequency main component of the electric stimulation, etc. In the electric stimulation monitoring signal, it can represent the basic mode in the whole electric stimulation process, such as the overall amplitude under the constant electric stimulation frequency.

[0160] The detail influence variable and the global influence variable are obtained by training the computer system according to the noise electric stimulation monitoring signal samples and the corresponding denoised electric stimulation monitoring signal samples. They reflect the influence degree of noise on different parts (details and global) of the signal and how to adjust the weight of these parts in the denoising process. For example, by analyzing a large number of noise electric stimulation monitoring signals and denoised signals, it is found that the noise in certain frequency range has a greater influence on the signal details, so the relationship will be reflected when determining the detail influence variable.

[0161] The first control signal implicit representation is an intermediate representation form obtained by fusing the pending detail implicit representation and the pending global implicit representation in this step, which is used for further processing to obtain the pending signal implicit representation.

[0162] If the next round of spectrum analysis is not greater than the second preset value, the computer system fuses (such as weighted summation) the pending detail implicit representation and the pending global implicit representation obtained in the next round according to the detail influence variable and the global influence variable to obtain a first control signal implicit representation. Assuming that the pending detail implicit representation is D, the pending global implicit representation is G, the detail influence variable is a, and the global influence variable is β, then the first control signal implicit representation S_1 = aD + βG.

[0163] For example, in the processing of an electrical stimulation monitoring signal, the pending detail implicit representation D contains some high-frequency, rapidly changing signal characteristics, which may be caused by local activity of nerve cells or slight fluctuations in electrical stimulation. The pending global implicit representation G contains information such as the fundamental frequency and overall intensity of the electrical stimulation. If the detail influence variable a = 0.3 and the global influence variable β = 0.7, then the first control signal implicit representation S1 = 0.3D + 0.7G. This means that at this stage, the computer system gives more weight to global characteristics because the weight of the global influence variable is greater, and the contribution of global characteristics to the first control signal implicit representation is greater in the fusion process.

[0164] In step S200A2, the first electrical stimulation monitoring signal is the original monitoring signal in the entire wireless implantable neural regulation electrical stimulation monitoring data classification method. It contains various information in the electrical stimulation process, including parameter information of the electrical stimulation itself and response information of the neural tissue to the electrical stimulation, and may also contain noise and other interference information.

[0165] Upsampling or downsampling is a resampling operation on a signal. Upsampling is to increase the sampling rate of the signal, for example, by inserting new sampling points between the original sampling points of the signal through interpolation or other methods, thereby increasing the resolution of the signal. Downsampling is to reduce the sampling rate of the signal, usually by discarding some sampling points according to certain rules. In this step, the first electrical stimulation monitoring signal after upsampling or downsampling is processed, and the purpose is to obtain more different representations of the signal in order to further integrate with the first control signal implicit representation obtained above.

[0166] The second control signal implicit representation is the result obtained after feature implicit representation of the first electrical stimulation monitoring signal after upsampling or downsampling, which will be used together with the first control signal implicit representation for subsequent operations to obtain the pending signal implicit representation.

[0167] The computer system up-samples or down-samples the first electrical stimulation monitoring signal. For example, if up-sampling is used, assuming the first electrical stimulation monitoring signal is x(n), n=0,1,...,N-1, using a linear interpolation up-sampling method, the sampling rate is doubled. The new sampling points x(m) (m=0,1,...,2N-1) can be calculated by Then the up-sampled signal is characterized by an implicit representation. Assuming a wavelet transform based implicit representation method is used. For the up-sampled signal x(m), a wavelet transform is performed (where a is a scale parameter, b is a translation parameter, and ψ(t) is a wavelet function), the wavelet coefficients obtained are processed (e.g., selected combinations of coefficients at certain scales and translations) to obtain the second control signal implicit representation. For example, the wavelet coefficients at larger scales are selected to represent the overall trend of the signal, and the coefficients at smaller scales are selected to represent the details of the signal, and then the coefficients are combined to form the second control signal implicit representation.

[0168] In step S200A3, the weight focusing strategy is a method for determining the weights of different signal representations in the integration process. It adjusts the weights of the various signal representations (e.g., the first control signal implicit representation and the second control signal implicit representation) according to the characteristics of the signal, the previously trained variables (e.g., the detail influence variable and the global influence variable), and the current processing goal, etc., in order to more effectively utilize the information in the integration process.

[0169] The pending signal implicit representation is the result obtained after this step, which is a signal representation form obtained by comprehensively considering the pending detail implicit representation, the pending global implicit representation, and the related representation of the up-sampled or down-sampled first electrical stimulation monitoring signal, and is used for subsequent further spectral analysis or other processing.

[0170] The computer system performs feature implicit representation on the first control signal implicit representation S1 and the second control signal implicit representation S2 through the weight focusing strategy, and obtains the pending signal implicit representation S t For example, the weight focusing strategy can be to determine the weights according to the energy of the signal. Let the energy of the first control signal implicit representation S1 be E1=∑ i |S1(i)| 2 , and the energy of the second control signal implicit representation S2 be E2=∑ i |S2(i)| 2 .

[0171] Then the weights and are calculated, and the pending signal implicit representation S t= w1S1 + w2S2. For example, if E1 = 10, E2 = 5, then Then In this way, the computer system reasonably allocates the weight according to the energy of the signal, so that the information in the implicit representation of the two control signals can be better utilized in the integration process to obtain the pending signal implicit representation, which provides a more comprehensive and accurate signal representation for subsequent processing.

[0172] Through the operations of steps S200A1-S200A3, the computer system gradually constructs the pending signal implicit representation according to the detail influence variable and the global influence variable in the spectrum analysis process of different rounds, in combination with different representations of the original first electrical stimulation monitoring signal, which helps to more accurately analyze the signal characteristics and lays the foundation for finally obtaining accurate first detail implicit representation and first global implicit representation.

[0173] In another embodiment, in step S200A, the pending detail implicit representation and the pending global implicit representation are integrated according to the detail influence variable and the global influence variable to obtain the pending signal implicit representation, including:

[0174] Step S200Aa: If the round of the next spectrum analysis is greater than the third preset value, then the pending detail implicit representation and the pending global implicit representation obtained in the next round are fused according to the detail influence variable and the global influence variable to obtain the third control signal implicit representation, and the third preset value is not greater than the first preset value.

[0175] Step S200Ab: One or more pending signal implicit representations obtained in the iteration before the next round are integrated with the third control signal implicit representation to obtain the pending signal implicit representation of the next round.

[0176] In another embodiment of step S200A, steps S200Aa-S200Ab are included, which is another operation process of the computer system for integrating the pending detail implicit representation and the pending global implicit representation according to the detail influence variable and the global influence variable to obtain the pending signal implicit representation.

[0177] The pending detail implicit representation is a form of representation that the computer system obtains after performing a spectral analysis on the first signal implicit representation. It mainly embodies the detail characteristics in the signal. In the context of the electrical stimulation monitoring signal, these details can be the small and fast changes in the response of the nerve cells to the electrical stimulation. For example, when the nerve is stimulated by a short electrical pulse, the nerve cell membrane potential can produce some instantaneous and small fluctuations, which can be part of the pending detail implicit representation reflected in the signal. These detail characteristics can be manifested as high frequency components in the frequency spectrum or as rapid amplitude changes in the time domain.

[0178] The pending global implicit representation is a form of representation that reflects the overall characteristics of the signal, corresponding to the pending detail implicit representation, obtained by spectral analysis. It contains the macro characteristics of the electrical stimulation monitoring signal, such as the basic frequency of the electrical stimulation, the overall intensity, and the main low frequency components, etc. Just like in a continuous electrical stimulation process, the overall electrical stimulation frequency and intensity level are embodied in the pending global implicit representation, which reflects the basic framework and main trend of the electrical stimulation.

[0179] The detail influence variable and the global influence variable are obtained by training the computer system based on the noise electrical stimulation monitoring signal samples and the corresponding denoised electrical stimulation monitoring signal samples. They quantify the influence relationship of the noise on the detail and global parts of the signal. For example, when processing a large number of noise electrical stimulation monitoring signals and their corresponding denoised signals, it is found that the noise in certain frequency bands has a more significant impact on the detail part of the signal, so when determining the detail influence variable, this relationship will be reflected. Similarly, for the global part, the variable will also be adjusted according to the influence of the noise on the overall signal characteristics.

[0180] The third control signal implicit representation is an intermediate result obtained by fusing the pending detail implicit representation and the pending global implicit representation in this step, which is an important part of constructing the pending signal implicit representation.

[0181] If the next round of spectral analysis is greater than the third preset value, the computer system fuses (such as weighted sum) the pending detail implicit representation and the pending global implicit representation obtained in the next round according to the detail influence variable and the global influence variable, to obtain the third control signal implicit representation. Assuming that the pending detail implicit representation is D, the pending global implicit representation is G, the detail influence variable is α, and the global influence variable is β, then the third control signal implicit representation S3 = αD + βG.

[0182] For example, in a specific electrical stimulation monitoring signal processing, suppose the undetermined detail implicit representation D contains high-frequency, rapidly changing signal features resulting from local neuronal activity; these features may reflect the instantaneous response of neurons to electrical stimulation. The undetermined global implicit representation G contains macroscopic information such as the basic frequency and overall intensity of the electrical stimulation. If the detail influence variable α = 0.4 and the global influence variable β = 0.6, then the third control signal implicit representation S3 = 0.4D + 0.6G. The weighting coefficients here indicate that, in this case, global features have a relatively large weight in the fusion process, possibly because at this stage, the overall electrical stimulation features are more important for signal representation.

[0183] In step S200Ab, the implicit representation of the undetermined signal is a signal representation obtained after a series of operations. It integrates information from multiple aspects, including the undetermined detail implicit representation, the undetermined global implicit representation, and relevant information from previous iterations. This representation is intended to more comprehensively and accurately describe the electrical stimulation monitoring signal, facilitating subsequent in-depth spectral analysis or other processing operations. The integration here refers to the operation of combining the previously obtained third control signal implicit representation with one or more undetermined signal implicit representations obtained in the next previous iteration. This combination is not a simple addition, but rather, according to certain rules or algorithms, it fuses the information from various parts together to obtain a more complete and representative signal representation.

[0184] The computer system integrates one or more implicit representations of the undetermined signal obtained in previous iterations with a third implicit representation of the contrast signal to obtain the next implicit representation of the undetermined signal. Assume that two implicit representations S of the undetermined signal were obtained in previous iterations. t1 and S t2 And the currently obtained third control signal implicitly represented as S3. The computer system can integrate them using a weighted average method.

[0185] First, calculate the weights of each signal representation. Let S be... t1 The energy is E t1 =∑ i |S t1 (i)| 2 S t2 The energy is E t2 =∑ i |S t2 (i)| 2 The energy of S3 is E3 = ∑ i |S3(i)| 2 Then calculate the weights.

[0186] The next undetermined signal is implicitly represented by S.t S t = w t1 S t1 + w t2 S t2 + w3S3, for example. If E t1 = 8, E t2 = 6, E3= 4, then Then In this way, the computer system can combine the information of previous iterations with the current results, gradually build a more accurate signal representation of the signal characteristics of the to-be-determined signal, which helps to more accurately analyze the signal in subsequent processing, and ultimately obtain an accurate first detail implicit representation and a first global implicit representation.

[0187] Through the operations of steps S200Aa-S200Ab, the computer system gradually builds a more perfect to-be-determined signal implicit representation by fusing the details and global characteristics of the signal according to different rounds of the spectrum analysis process, using the detail influence variable and the global influence variable, and combining the information of previous iterations, thereby providing a more accurate signal processing basis for the entire wireless implantable neuromodulation electrical stimulation monitoring data classification method.

[0188] In an embodiment, the wireless implantable neuromodulation electrical stimulation monitoring data classification method is implemented based on an electrical stimulation monitoring signal processing network, which includes a feature implicit representation component, a spectrum analysis component, an adaptive fusion component, and a signal generation component. The feature implicit representation component is used to perform feature implicit representation on a first electrical stimulation monitoring signal to obtain a first signal implicit representation. The spectrum analysis component is used to perform spectrum analysis on the first signal implicit representation to obtain a first detail implicit representation and a first global implicit representation. The adaptive fusion component is used to integrate the first detail implicit representation and the first global implicit representation according to a detail influence variable and a global influence variable to obtain a second signal implicit representation, wherein the detail influence variable and the global influence variable are component parameters of the adaptive fusion component. The signal generation component is used to filter the second signal implicit representation through a first filter matrix to obtain a residual mapping signal, integrate the first electrical stimulation monitoring signal and the residual mapping signal to obtain a second electrical stimulation monitoring signal, and the first filter matrix is a component parameter of the signal generation component.

[0189] In this embodiment of the wireless implantable neuromodulation electrical stimulation monitoring data classification method implemented based on the electrical stimulation monitoring signal processing network, each component cooperates to complete the entire data classification process.

[0190] First is the feature implicit representation component. The computer system performs feature implicit representation on the first electrical stimulation monitoring signal through the feature implicit representation component, thereby obtaining the first signal implicit representation. The first electrical stimulation monitoring signal is the original electrical stimulation monitoring data collected by a specific monitoring device in a wireless implantable neuromodulation scenario, which contains rich information such as the intensity, frequency, duration of electrical stimulation, and the response of neural tissue to electrical stimulation. The specific way in which the feature implicit representation component realizes this function can adopt various mathematical transformation methods. For example, it can use the convolution layer operation in the convolutional neural network (CNN). Assuming that the first electrical stimulation monitoring signal is a discrete time series signal x(n), n = 1, 2, …, N, the computer system can define a convolution kernel k(m), m = 1, 2, …, M (where M is usually much smaller than N), and extract the local features in the signal through convolution operation Here, y(n) is a preliminary feature implicit representation result after convolution operation. In addition, principal component analysis (PCA) is also a feasible way. For the first electrical stimulation monitoring signal X (which can be regarded as a matrix, each row representing a sample and each column representing a feature dimension), first calculate its covariance matrix Then solve the eigenvalues and corresponding eigenvectors of the covariance matrix C. According to the size of the eigenvalues, sort the eigenvectors, and select the eigenvectors corresponding to the first k largest eigenvalues to form the projection matrix P. Then the feature implicit representation after PCA transformation is Y = XP.

[0191] Then is the spectrum analysis component. The computer system performs spectrum analysis on the first signal implicit representation using the spectrum analysis component to obtain the first detail implicit representation and the first global implicit representation. Spectrum analysis is a technical means of analyzing signals from the time domain to the frequency domain. For example, the computer system can use discrete Fourier transform (DFT) to realize spectrum analysis. Assuming that the first signal implicit representation is a discrete signal x(n), n = 0, 1, …, N-1, its DFT is defined as where k = 0, 1, …, N-1, X(k) is the frequency domain representation of x(n) which contains the amplitude and phase information of different frequency components of the signal. In this frequency domain representation, the first detail implicit representation is an implicit representation of the signal detail part extracted from the frequency spectrum analysis result. For example, in an electrostimulation monitoring signal containing multiple frequency components, those frequency components with fast changes or local fluctuations can correspond to the subtle changes of neural activities or the transient characteristics of electrostimulation, which are summarized in the first detail implicit representation after frequency spectrum analysis. The computer system can identify these detail-related frequency components by setting a specific frequency range or frequency threshold, for example, regarding the high frequency part (higher than a certain frequency threshold) as the detail-related frequency component, and then performing special processing (such as weighting, normalization, etc.) on the frequency spectrum data corresponding to these frequency components to construct the first detail implicit representation. The first global implicit representation is an implicit representation of the overall characteristics of the signal obtained from the frequency spectrum analysis result. It reflects the overall trend and main characteristics of the signal in a wide frequency range. For example, in an electrostimulation monitoring signal, the overall energy distribution, the main frequency peak, etc. reflect the basic characteristics of electrostimulation and the overall response trend of neural tissue to electrostimulation. The computer system can construct the first global implicit representation by smoothing the frequency spectrum data or performing special processing on the low frequency part (lower than a certain frequency threshold). For example, using the moving average method to smooth the frequency spectrum data, assuming the frequency spectrum data is Y(k), and the moving average window size is M, then the smoothed frequency spectrum data (when i is out of the range of 0 to N-1, appropriate boundary processing is performed), which can be used as the basis for constructing the first global implicit representation.

[0192] Then there is the adaptive fusion component. The computer system integrates the first detail implicit representation and the first global implicit representation to obtain the second signal implicit representation by means of the adaptive fusion component according to the detail influence variable and the global influence variable. The detail influence variable and the global influence variable are component variables of the adaptive fusion component, which are obtained by the computer system through training based on the noise electrostimulation monitoring signal samples and the corresponding denoised electrostimulation monitoring signal samples. The noise electrostimulation monitoring signal samples refer to the sample signals mixed with noise during the collection of the electrostimulation monitoring signal, and the denoised electrostimulation monitoring signal samples are ideal signal samples after removing the noise interference. In the training process, the computer system can use machine learning or deep learning algorithms. For example, by constructing a regression model, assuming the set of noise electrostimulation monitoring signal samples is X = {x1, x2, …, xN}, and the corresponding set of denoised electrostimulation monitoring signal samples is Y = {y1, y2, …, yN}, the computer system can train a regression model to obtain the relationship between the noise electrostimulation monitoring signal samples and the corresponding denoised electrostimulation monitoring signal samples, and then use the trained regression model to obtain the second signal implicit representation based on the first detail implicit representation and the first global implicit representation. n n ​}. Assume a linear regression model y = β0 + β1x, where β0 and β1 are the parameters of the model. The computer system determines the values of β0 and β1 by minimizing a loss function (e.g., mean squared error from the noisy and denoised signal relationships, which can be further translated into detail influence variables and global influence variables. For the integration process of the first detail implicit representation and the first global implicit representation, assume the first detail implicit representation is D, the first global implicit representation is G, the detail influence variable is a, and the global influence variable is β. The computer system can integrate them using a weighted sum, i.e., the second signal implicit representation S = aD + βG. To illustrate with a simple example of an electrostimulation monitoring signal, suppose the first detail implicit representation D contains the characteristics of small fluctuations in the electrostimulation signal, which can be related to local activity of nerve cells; while the first global implicit representation G contains the macroscopic characteristics of the overall intensity and frequency of the electrostimulation signal. The detail influence variable a can represent the importance of small fluctuation characteristics in the final integration, and the global influence variable β represents the importance of macroscopic characteristics. Through the above weighted sum, the computer system integrates the two different levels of feature representations according to their importance, obtaining the second signal implicit representation S.

[0193] Finally, there is a signal generation component. The computer system uses the signal generation component to filter the second signal implicit representation through a first filter matrix to obtain a residual mapping signal, and then integrates the first electrostimulation monitoring signal with the residual mapping signal to obtain a second electrostimulation monitoring signal. The first filter matrix is a component variable of the signal generation component, which is obtained by the computer system training based on the noisy electrostimulation monitoring signal samples and the denoised electrostimulation monitoring signal samples. In the training process, the computer system can use the least squares method to construct this filter matrix. Assume that the noisy electrostimulation monitoring signal samples form a matrix X (each column is a sample), and the corresponding denoised electrostimulation monitoring signal samples form a matrix Y, and there is a filter matrix A such that Y ≈ AX. Solve A by the least squares method, i.e., minimize the objective function (where n is the number of samples, y_i is the i-th sample in Y, and x_i is the i-th sample in X). By solving this optimization problem, the first filter matrix A can be obtained. For the second signal implicit representation, it is a signal representation form that integrates the detail and global information through the previous steps. Assume that the second signal implicit representation is a vector s = [s1, s2, …, sm] (where m is the dimension of the vector), and the first filter matrix is A = [a1, a2, …, an] (where n is the number of rows of the filter matrix). The computer system performs a filtering operation through matrix multiplication to obtain a filtered result vector r = As = [r1, r2, …, rm]. m ](where m is the dimension of the vector), and the first filter matrix is A = [a ij ] n×m (n is the number of rows of the filter matrix). The computer system performs a filtering operation through matrix multiplication to obtain a filtered result vector r = As = [r1, r2, …, rm].n With a simple example, assume that the first electrical stimulation monitoring signal contains noise components due to various reasons (e.g. noise of the acquisition device, environmental interference, etc.), while the corresponding denoised electrical stimulation monitoring signal is the ideal one without noise interference. The second signal implicit representation contains the integrated signal feature information. When the computer system filters the second signal implicit representation using the first filter matrix, the resulting residual mapping signal reflects the difference between the two. For example, if a certain frequency component in the first electrical stimulation monitoring signal is shifted or changed in amplitude due to noise, then this difference from the denoised electrical stimulation monitoring signal will be reflected in the residual mapping signal. Then, the computer system integrates the first electrical stimulation monitoring signal x(n) with the residual mapping signal r(n), for example, in a linear addition manner y(n) = x(n) + kxr(n) (where k is a weighting coefficient), to obtain the second electrical stimulation monitoring signal y(n). This second electrical stimulation monitoring signal is of great significance in subsequent electrical stimulation monitoring data classification, as it integrates the information of the original signal and the residual mapping signal, providing a basis for more accurate classification.

[0194] In this embodiment of the electrical stimulation monitoring signal processing network, the first electrical stimulation monitoring signal is gradually converted into the second electrical stimulation monitoring signal through a series of processing steps, which fully utilizes various mathematical transformations, machine learning algorithms, and in-depth analysis of signal features, laying a solid foundation for the classification of electrical stimulation monitoring data for wireless implantable neuromodulation.

[0195] In one embodiment, the electrical stimulation monitoring signal processing network is trained by the following steps:

[0196] Step S10: Feature implicit representation of the noise electrical stimulation monitoring signal sample is performed by the feature implicit representation component to obtain the first signal training implicit representation;

[0197] Step S20: The first signal training implicit representation is subjected to spectral analysis by the spectral analysis component to obtain the first detail training implicit representation and the first global training implicit representation;

[0198] Step S30: The first detail training implicit representation and the first global training implicit representation are integrated by the adaptive fusion component according to the detail influence variable and the global influence variable to obtain the second signal training implicit representation;

[0199] Step S40: The second signal training implicit representation is filtered by the signal generation component through the first filter matrix to obtain the residual training mapping signal, and the noise electrical stimulation monitoring signal sample is integrated with the residual training mapping signal to obtain the estimated electrical stimulation monitoring signal;

[0200] Step S50: training the electrical stimulation monitoring signal processing network according to the loss between the estimated electrical stimulation monitoring signal and the denoised electrical stimulation monitoring signal sample.

[0201] In step S10, the computer system performs feature implicit representation on the noisy electrical stimulation monitoring signal sample through a feature implicit representation component to obtain a first signal training implicit representation. The noisy electrical stimulation monitoring signal sample is a signal sample mixed with noise when actually collecting the electrical stimulation monitoring signal due to various interference factors (such as equipment noise, environmental electromagnetic interference, etc.). The feature implicit representation component can use various technical means for feature implicit representation. For example, a wavelet transform method is used. For a discrete noisy electrical stimulation monitoring signal sample x(n), n = 0, 1, …, N-1, its wavelet transform is (where a is a scale parameter, b is a translation parameter, and ψ(t) is a wavelet function), and the first signal training implicit representation is obtained by selecting a combination of wavelet coefficients at appropriate scales and translation parameters. This representation converts the original signal into a form that better reflects its inherent characteristics, which can include frequency distribution, local variation, and other characteristics, providing a basis for subsequent processing.

[0202] Then in step S20, the computer system performs spectral analysis on the first signal training implicit representation through a spectral analysis component to obtain a first detail training implicit representation and a first global training implicit representation. Spectral analysis is an important means of converting a signal from the time domain to the frequency domain to reveal its frequency characteristics. The computer system can use discrete Fourier transform (DFT) to realize spectral analysis. Assuming that the first signal training implicit representation is y(n), n = 0, 1, …, M-1, its DFT is defined as

[0203] k = 0, 1, …, M-1. In the frequency domain, the first detail training implicit representation mainly focuses on the high-frequency part or rapidly changing frequency components in the signal, which may correspond to local fluctuations or minor changes in the noisy electrical stimulation monitoring signal, such as details. For example, in the electrical stimulation monitoring signal, it may be the signal change caused by the instantaneous minor response of nerve cells to electrical stimulation. The computer system can set a frequency threshold f thThe first detailed training implicit representation is constructed by specially processing (e.g., weighting) the frequency components above a certain threshold. The first global training implicit representation focuses on the low-frequency portion of the signal or the overall frequency trend, reflecting the overall characteristics of the noise-induced electrical stimulation monitoring signal, such as the overall energy distribution and major frequency peaks. The computer system can construct the first global training implicit representation by smoothing the low-frequency components (e.g., using a moving average method). For example, for a frequency domain signal Y(k), using a moving average method with a moving average window size of L, the smoothed signal... (Appropriate boundary handling is performed when i exceeds the range of 0 to M-1), and this smoothed signal can be used as part of constructing the first global training implicit representation.

[0204] In step S30, the computer system, through an adaptive fusion component, integrates the first detailed training implicit representation and the first global training implicit representation based on the detailed influence variables and the global influence variables to obtain the second signal training implicit representation. The detailed influence variables and the global influence variables are important parameters of the adaptive fusion component; they are obtained by the computer system through training based on the relationship between noisy electrical stimulation monitoring signal samples and their corresponding denoised electrical stimulation monitoring signal samples. For example, these variables can be determined by constructing a linear regression model based on the least squares method. Let the noisy electrical stimulation monitoring signal sample set be X = {x1, x2, ..., x...}. n The corresponding denoised electrical stimulation monitoring signal sample set is Y = {y1, y2, ..., y}. n Assume a linear regression model y = β0 + β1x, where β0 and β1 are the model parameters. The computer system minimizes the loss function (such as mean squared error). To determine the values ​​of β0 and β1, these parameters, after appropriate transformation, can be used to obtain the detailed influence variables and the global influence variables. Assume the first detailed training implicitly represents D. t The first global training implicit representation is G. t Let α be the variable affecting the details and β be the variable affecting the global aspects. The computer system can integrate these variables using a weighted summation method, i.e., training the implicit representation S with the second signal. t =αD t +βG t For example, if α = 0.3 and β = 0.7, this means that more emphasis is placed on the information of the global training implicit representation during the integration process. This may be because, in noisy electrical stimulation monitoring signals, global features are more important for the overall representation of the signal.

[0205] In step S40, the computer system filters the second signal training implicit representation through the first filter matrix by the signal generation component to obtain a residual training mapping signal, integrates the noise electrical stimulation monitoring signal sample with the residual training mapping signal to obtain an estimated electrical stimulation monitoring signal. The first filter matrix is an important component variable of the signal generation component, which is obtained by the computer system training according to the noise electrical stimulation monitoring signal sample and the de-noised electrical stimulation monitoring signal sample. For example, the principal component analysis (PCA) method is used to construct the first filter matrix. For the noise electrical stimulation monitoring signal sample matrix A (each row represents a sample, and each column represents a feature dimension), first, the covariance matrix C of A is calculated (where m is the number of samples), and then the eigenvalues λ i and the corresponding eigenvectors v i of the covariance matrix C are solved. The eigenvectors are sorted according to the size of the eigenvalues, and the eigenvectors corresponding to the first k largest eigenvalues are selected to form the first filter matrix F. Assuming that the second signal training implicit representation is a vector s = [s1, s2, …, s p ], the residual training mapping signal r = [r1, r2, …, r q ] is obtained by matrix multiplication r = Fs. Then the noise electrical stimulation monitoring signal sample x(n) and the residual training mapping signal r(n) are integrated, for example, a simple addition method y(n) = x(n) + r(n) is used to obtain the estimated electrical stimulation monitoring signal y(n). This integration method is based on the fact that the residual training mapping signal can reflect the error relationship between the noise electrical stimulation monitoring signal and the ideal signal (de-noised signal), and by adding the original noise electrical stimulation monitoring signal sample, an estimated signal close to the de-noised electrical stimulation monitoring signal can be obtained.

[0206] Finally, in step S50, the computer system trains the electrical stimulation monitoring signal processing network according to the loss between the estimated electrical stimulation monitoring signal and the de-noised electrical stimulation monitoring signal sample. The loss here is an index that measures the difference between the estimated electrical stimulation monitoring signal and the de-noised electrical stimulation monitoring signal sample. For example, the mean square error (MSE) can be used as the loss function, assuming that the estimated electrical stimulation monitoring signal is and the de-noised electrical stimulation monitoring signal sample is y d (n), n = 0, 1, …, N-1, then the mean square error is The computer system adjusts parameters of each component in the electrical stimulation monitoring signal processing network (such as transformation parameters in the feature implicit representation component, detail influence variables and global influence variables in the adaptive fusion component, the first filter matrix in the signal generation component, etc.) to minimize the loss function. For example, in a gradient descent-based optimization algorithm, the computer system calculates the gradient of the loss function with respect to each parameter, and then updates the parameters according to a certain learning rate. Assuming that the parameter is θ and the learning rate is η, the parameter update formula is By continuously iterating this process, the estimated electrical stimulation monitoring signal is closer and closer to the denoised electrical stimulation monitoring signal sample, thereby improving the processing capability of the electrical stimulation monitoring signal processing network for electrical stimulation monitoring data, so that it can play a better role in the actual wireless implantable neuromodulation electrical stimulation monitoring data classification task.

[0207] Through the training process of steps S10-S50, the computer system can effectively train the electrical stimulation monitoring signal processing network, so that the parameters of each component of the network are optimized, thereby accurately processing the electrical stimulation monitoring data and providing a reliable basis for subsequent classification and other operations.

[0208] In an embodiment, the electrical stimulation monitoring signal processing network includes a frequency spectrum analysis component and an adaptive fusion component, a≥1; step S20, by the frequency spectrum analysis component, the first signal training implicit representation is subjected to frequency spectrum analysis to obtain the first detail training implicit representation and the first global training implicit representation, which can include:

[0209] Step S21: by the first frequency spectrum analysis component, the first signal implicit representation is subjected to frequency spectrum analysis to obtain the undetermined detail training implicit representation and the undetermined global training implicit representation; by the first adaptive fusion component, the undetermined detail training implicit representation and the undetermined global training implicit representation are integrated according to the detail influence variables and the global influence variables in the first adaptive fusion component to obtain the undetermined signal training implicit representation;

[0210] Step S22: a new one of the spectrum analysis components performs a new one of the spectrum analyses on the pending signal training implicit representation, obtaining a new one of the pending detail training implicit representations and a new one of the pending global training implicit representations; a new one of the adaptive fusion components integrates the new one of the pending detail training implicit representations and the new one of the pending global training implicit representations according to the detail influence variable and the global influence variable in the new one of the adaptive fusion components, obtaining a new one of the pending signal training implicit representations, until the last one of the spectrum analysis components outputs the pending detail training implicit representation and the pending global training implicit representation, taking the pending detail training implicit representation output by the last one of the spectrum analysis components as the first detail training implicit representation and taking the pending global training implicit representation output by the last one of the spectrum analysis components as the first global training implicit representation;

[0211] Step S30: the adaptive fusion component integrates the first detail training implicit representation and the first global training implicit representation according to the detail influence variable and the global influence variable, obtaining the second signal training implicit representation, which can include:

[0212] Step S31: the last one of the adaptive fusion components integrates the first detail training implicit representation and the first global training implicit representation according to the detail influence variable and the global influence variable in the last one of the adaptive fusion components, obtaining the second signal training implicit representation.

[0213] In step S21, the first signal training implicit representation is obtained by the computer system in step S10 through the feature implicit representation component to perform feature implicit representation on the noise electrical stimulation monitoring signal sample. It is a feature converted representation form of the original noise electrical stimulation monitoring signal sample, which contains the relevant features of the original signal in an implicit manner, such as the local characteristics of the signal, the frequency distribution characteristics, etc., for subsequent spectrum analysis.

[0214] The spectrum analysis component is an important component in the electrical stimulation monitoring signal processing network, and its function is to perform spectrum analysis on the input signal (here, the first signal training implicit representation), convert the signal from the time domain to the frequency domain, and thus reveal the characteristics of the signal in the frequency domain, such as the frequency components of the signal, the amplitude and phase of each frequency component, etc.

[0215] The pending detail training implicit representation and the pending global training implicit representation are two representations obtained by the computer system after performing a spectral analysis on the first signal training implicit representation by the first spectral analysis component. The pending detail training implicit representation focuses on reflecting the characteristics of the details in the signal in the frequency domain. These details can be high frequency components, local fluctuations, or rapidly changing parts in the signal. For example, in an electrostimulation monitoring signal, these details can correspond to the signal changes in the frequency domain caused by the instantaneous, tiny reactions of nerve cells to electrostimulation. The pending global training implicit representation focuses on reflecting the overall characteristics of the signal in the frequency domain, such as the overall frequency distribution trend, the main frequency peak, etc. These overall characteristics reflect the macroscopic features of the electrostimulation monitoring signal.

[0216] The adaptive fusion component is a component of the electrostimulation monitoring signal processing network, which is used to perform fusion operations on different signal representations according to specific variables (such as the detail influence variable and the global influence variable). Here, the first adaptive fusion component is used to integrate the pending detail training implicit representation and the pending global training implicit representation.

[0217] The pending signal training implicit representation is an intermediate result obtained by the first adaptive fusion component after integrating the pending detail training implicit representation and the pending global training implicit representation according to the detail influence variable and the global influence variable in it. It is the basis for further processing.

[0218] The computer system first performs a spectral analysis on the first signal training implicit representation by the first spectral analysis component. For example, assuming that the first signal training implicit representation is a discrete-time sequence x(n), n = 0, 1, …, N-1, the computer system can use the Fast Fourier Transform (FFT) to perform the spectral analysis. The formula for FFT is where X(k) is the frequency domain representation of x(n). After obtaining the frequency domain signal by FFT, the computer system decomposes it into the pending detail training implicit representation and the pending global training implicit representation according to the frequency characteristics. For example, the part with a frequency higher than a certain threshold f th1 is considered as the pending detail training implicit representation, which can contain the rapidly changing information in the signal; the part with a frequency lower than f th2 (f th2 <f th1 is considered as the pending global training implicit representation, which reflects the overall trend of the signal.

[0219] Then, the first adaptive fusion component integrates the two representations according to the detail influence variable a and the global influence variable b. Assuming the pending detail training implicit representation is D, and the pending global training implicit representation is G, the pending signal training implicit representation S = aD + bG is obtained by weighted summation. For example, if a = 0.3, b = 0.7, it means that in this stage, the computer system considers the global features relatively more important, and gives the global training implicit representation a greater weight in the integration process.

[0220] In step S22, as the processing flow advances, the new spectral analysis component and the adaptive fusion component here are components used in turn in subsequent processing steps. Their functions are similar to the first spectral analysis component and the adaptive fusion component, but the objects of processing are the signal representation forms processed after the previous steps, and the signal features are gradually analyzed and fused through multiple such operations.

[0221] In each new round of processing, the computer system obtains the representation form by performing spectral analysis on the pending signal training implicit representation obtained in the previous round through a new spectral analysis component. Like the similar concept in step S21, they respectively focus on the detail features and overall features of the signal in the frequency domain, but as the number of rounds increases, these representation forms are more detailed and accurate in characterizing the signal features.

[0222] The pending signal training implicit representation (new round) is obtained by integrating the pending detail training implicit representation and the pending global training implicit representation of the new round through a new adaptive fusion component according to the corresponding detail influence variable and global influence variable. As the number of rounds increases, this result is constantly optimized and closer to the ideal signal representation form.

[0223] The computer system performs a new spectral analysis on the pending signal training implicit representation obtained in the previous round through a new spectral analysis component. For example, assuming that the pending signal training implicit representation obtained in the previous round is S prev , and still using FFT to perform spectral analysis to obtain a new frequency domain representation. Then, according to the new frequency characteristics, the new pending detail training implicit representation D new and the pending global training implicit representation G new are divided. For example, in this round, according to a more detailed frequency range division, the higher frequency part is further subdivided into a new pending detail training implicit representation, to more accurately capture the detailed features such as small fluctuations in the signal.

[0224] Then, through a new adaptive fusion component, the detail influence variable a new and the global influence variable b newa new one pending detail training implicit representation D new and a new one pending global training implicit representation G new and a new one pending signal training implicit representation S new = a new D new + β new G new For example, if in this round, according to the previous analysis results, it is found that the importance of the detail feature in the subsequent processing has increased, the value of a new may be adjusted to 0.4, and the corresponding β new = 0.6 to reflect this change. This process is repeated until the last spectrum analysis component outputs the pending detail training implicit representation and the pending global training implicit representation. For example, if a total of M rounds of such operations are set, when the Mth round is reached, the pending detail training implicit representation output by the last spectrum analysis component will be taken as the first detail training implicit representation, and the pending global training implicit representation output by the last spectrum analysis component will be taken as the first global training implicit representation. This means that after multiple rounds of spectrum analysis and fusion operations, the final representation that accurately reflects the detail and global features of the signal in the frequency domain is obtained, providing accurate signal feature information for subsequent processing.

[0225] In step S31, the first detail training implicit representation and the first global training implicit representation are the training implicit representations about the detail and global features of the signal in the frequency domain obtained after a series of operations in step S20. The first detail training implicit representation contains the frequency domain features related to the details of the signal, such as the high frequency components, local fluctuations, and other features as described above; the first global training implicit representation contains the overall frequency domain features of the signal, such as the overall frequency trend, main frequency peak, and other information.

[0226] The last adaptive fusion component is the last one in a series of adaptive fusion components, and it plays a key integration role in this stage of the entire signal processing flow. It integrates the first detail training implicit representation and the first global training implicit representation according to its own detail influence variable and global influence variable.

[0227] The second signal training implicit representation is the result obtained by the last adaptive fusion component after integrating the first detail training implicit representation and the first global training implicit representation. This representation form integrates the detail and global frequency domain features of the signal, and is a more comprehensive and accurate signal representation form after multiple rounds of processing and fusion, providing a suitable input for subsequent operations (such as filtering in the signal generation component).

[0228] Suppose the first detail training implicit representation is D final, the first global training implicit representation is G final , the detail influence variable in the last adaptive fusion component is a last , the global influence variable is b last . The computer system integrates the first detail training implicit representation and the first global training implicit representation by the last adaptive fusion component in a weighted summation manner to obtain a second signal training implicit representation S second = a last D final + b last G final . For example, if a last = 0.4 and b last = 0.6, it indicates that in this last fusion step, the weight of the global training implicit representation is still larger, but a certain weight is also given to the detail training implicit representation. This may be because in the whole signal processing process, although the global feature is always more important, the detail feature also has an important role that cannot be ignored for the final accurate processing of the signal. For example, in the electrical stimulation monitoring signal, the global feature may reflect the basic mode of the electrical stimulation, and the detail feature may reflect some subtle reactions of the nerve cells to the electrical stimulation, and the combination of the two can better represent the signal and provide a more accurate basis for subsequent processing.

[0229] In an embodiment, the electrical stimulation monitoring signal processing network includes b signal generation components, 1≤b≤a; based on this, step S40, filtering the second signal training implicit representation by the first filter matrix through the signal generation component to obtain a residual training mapping signal, and integrating the noise electrical stimulation monitoring signal sample and the residual training mapping signal to obtain an estimated electrical stimulation monitoring signal, can include:

[0230] Step S41: filtering the pending signal training implicit representation output by the a-b+c adaptive fusion component by the first filter matrix through the cth signal generation component to obtain the cth residual training mapping signal, and integrating the noise electrical stimulation monitoring signal sample and the cth residual training mapping signal to obtain the cth estimated electrical stimulation monitoring signal; wherein 1≤c≤b.

[0231] The cth signal generation component is a component responsible for a specific operation in the electrical stimulation monitoring signal processing network, and in this step, it undertakes the task of filtering a specific input signal by the first filter matrix and integrating the noise electrical stimulation monitoring signal sample and the residual training mapping signal obtained by filtering. Here, c represents a specific serial number in multiple signal generation components, 1≤c≤b, and each signal generation component has its specific role in the whole processing flow.

[0232] The pending signal training implicit representation outputted by the a-b+c adaptive fusion component is a kind of signal representation form obtained after a series of operations, especially after being processed by the a-b+c adaptive fusion component. The role of the adaptive fusion component is to integrate different signal representations (such as the detail training implicit representation and the global training implicit representation) according to the detail influence variable and the global influence variable, and obtain the pending signal training implicit representation. This pending signal training implicit representation contains signal feature information after integration, which is the result of the previous steps of signal spectrum analysis, feature extraction, fusion and other operations, and is used as the input signal of the cth signal generation component.

[0233] The first filter matrix is an important component variable of the signal generation component, which is obtained by the computer system through training based on the noise electrical stimulation monitoring signal sample and the denoised electrical stimulation monitoring signal sample. This matrix plays a filtering role in signal processing, and through specific mathematical operations (such as matrix multiplication) with the input signal (here, the pending signal training implicit representation outputted by the a-b+c adaptive fusion component), it can extract or adjust certain features in the signal, thereby obtaining the residual training mapping signal. For example, the computer system can use the least squares method to construct this filter matrix. Assuming that the noise electrical stimulation monitoring signal sample forms a matrix X (each column is a sample), the corresponding denoised electrical stimulation monitoring signal sample forms a matrix Y, and there is a filter matrix A such that Y ≈ AX. Solve A by the least squares method, that is, minimize the objective function (where n is the number of samples, y i is the i-th sample in Y, x i is the i-th sample in X). By solving this optimization problem, the first filter matrix A can be obtained.

[0234] The residual training mapping signal is the result obtained by the computer system through the cth signal generation component using the first filter matrix to filter the pending signal training implicit representation outputted by the a-b+c adaptive fusion component. It represents a mapping relationship, reflecting the difference information between the pending signal training implicit representation and the ideal signal (in an ideal case, the signal without noise and other interference factors) after filtering. For example, in the electrical stimulation monitoring signal, if the pending signal training implicit representation contains signal deviation caused by noise or other interference, then the residual training mapping signal can reflect the characteristics of this deviation, such as amplitude change or phase shift at a certain frequency component.

[0235] The estimated electrical stimulation monitoring signal is a signal obtained by the computer system after integrating the noisy electrical stimulation monitoring signal sample with the cth residual training mapping signal. It is an estimated result of the de-noised electrical stimulation monitoring signal, and by combining the original signal containing noise (the noisy electrical stimulation monitoring signal sample) with the residual training mapping signal, it attempts to restore a more true electrical stimulation monitoring signal without noise interference.

[0236] The computer system first performs a filtering operation on the pending signal training implicit representation output by the a-b+cth adaptive fusion component using a first filter matrix through the cth signal generation component. Assuming that the pending signal training implicit representation output by the a-b+cth adaptive fusion component is a vector s = [s1, s2, …, s m (where m is the dimension of the vector), and the first filter matrix is A = [a1, a2, …, am], the computer system performs the filtering operation through matrix multiplication to obtain a filtered result vector r = As = [r1, r2, …, rm], and this result vector r is the cth residual training mapping signal. ij ] n×m (n is the number of rows of the filter matrix). The computer system performs the filtering operation through matrix multiplication to obtain a filtered result vector r = As = [r1, r2, …, rm], and this result vector r is the cth residual training mapping signal. n

[0237] Then, the computer system integrates the noisy electrical stimulation monitoring signal sample with the cth residual training mapping signal to obtain the cth estimated electrical stimulation monitoring signal. For example, assuming that the noisy electrical stimulation monitoring signal sample is x(n), and the cth residual training mapping signal is r_c(n), the computer system can use a simple addition method y c (n) = x(n) + r c (n) to obtain the cth estimated electrical stimulation monitoring signal y c (n). This integration method is based on the fact that the residual training mapping signal can reflect the error relationship between the noisy electrical stimulation monitoring signal and the ideal signal (the de-noised signal), and by adding it to the original noisy electrical stimulation monitoring signal sample, an estimated signal close to the de-noised electrical stimulation monitoring signal can be obtained.

[0238] ​For example, assume that in an electrostimulation monitoring scenario, the noisy electrostimulation monitoring signal sample is a signal containing the response of a nerve cell to an electrostimulation and some background noise interference. After the foregoing steps, the pending signal training implicit representation output by the a-b+c adaptive fusion component contains information about the signal after feature fusion, such as the comprehensive features of different frequency components in the signal, the mixed features of possible nerve cell response and noise, and the like. When the computer system filters the pending signal training implicit representation using the first filter matrix, it is like using a filter to screen out the part deviating from the ideal signal, obtaining a residual training mapping signal. If in this process, the first filter matrix finds that a certain high-frequency component in the pending signal training implicit representation has a large deviation (possibly because the noise interference causes the amplitude of the high-frequency component to be abnormal), the residual training mapping signal will reflect the information that the high-frequency component needs to be adjusted. Then, when the computer system integrates the noisy electrostimulation monitoring signal sample with the residual training mapping signal, it is equivalent to correcting the original noisy electrostimulation monitoring signal sample according to the deviation information, and the estimated electrostimulation monitoring signal obtained will be closer to the true electrostimulation monitoring signal without noise interference.

[0239] Through the operation of step S41, the computer system can gradually construct the estimated electrostimulation monitoring signal by using the output of the signal generation component, the first filter matrix, and the adaptive fusion component, which is very important for the training process of the entire electrostimulation monitoring signal processing network. It helps the computer system to adjust the parameters of each component in the network according to the relationship between the estimated electrostimulation monitoring signal and the denoised electrostimulation monitoring signal sample (such as training the network according to the loss between the two in subsequent step S50), thereby improving the processing ability and accuracy of the entire network for electrostimulation monitoring data.

[0240] The embodiment of the present application provides a computer system, as shown in the accompanying drawings, the computer system 100 comprises a processor 101 and a memory 103. Wherein, the processor 101 and the memory 103 are connected, such as connected through a bus 102. Optionally, the computer system 100 can further comprise a transceiver 104. It should be noted that in actual application, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation to the embodiment of the present application. Figure 2 The embodiment of the present application provides a computer system, as shown in the accompanying drawings, the computer system 100 comprises a processor 101 and a memory 103. Wherein, the processor 101 and the memory 103 are connected, such as connected through a bus 102. Optionally, the computer system 100 can further comprise a transceiver 104. It should be noted that in actual application, the transceiver 104 is not limited to one, and the structure of the computer system 100 does not constitute a limitation to the embodiment of the present application.

Claims

1. A method of wireless implantable neuromodulation electrical stimulation monitoring data classification, comprising: The method comprises: characteristic implicit representation is carried out on the first electric stimulation monitoring signal to obtain a first signal implicit representation; spectrum analysis is carried out on the first signal implicit representation to obtain a first detail implicit representation and a first global implicit representation; the first detail implicit representation and the first global implicit representation are integrated according to a detail influence variable and a global influence variable to obtain a second signal implicit representation, the detail influence variable and the global influence variable being obtained by training according to a noise electric stimulation monitoring signal sample and a corresponding denoised electric stimulation monitoring signal sample; the second signal implicit representation is filtered through a first filter matrix to obtain a residual mapping signal, the residual mapping signal representing an error between the first electric stimulation monitoring signal and the corresponding denoised electric stimulation monitoring signal, the first filter matrix being obtained by training according to the noise electric stimulation monitoring signal sample and the denoised electric stimulation monitoring signal sample; the first electric stimulation monitoring signal and the residual mapping signal are integrated to obtain a second electric stimulation monitoring signal, a signal-to-noise ratio of the second electric stimulation monitoring signal being greater than that of the first electric stimulation monitoring signal; the second electric stimulation monitoring signal is classified.

2. The method of claim 1, wherein, The spectrum analysis on the first signal implicit representation to obtain a first detail implicit representation and a first global implicit representation comprises: the first signal implicit representation is classified into a second detail implicit representation and a second global implicit representation according to a frequency spectrum, the second detail implicit representation and the second global implicit representation are integrated to obtain a third signal implicit representation; the third signal implicit representation is subjected to detail extraction to obtain the first detail implicit representation, and the third signal implicit representation is smoothed to obtain the first global implicit representation.

3. The method of claim 2, wherein, The method further comprises: the third signal implicit representation is filtered through a second filter matrix to obtain a smoothing matrix, the second filter matrix being obtained by training according to the noise electric stimulation monitoring signal and the denoised electric stimulation monitoring signal; a unit filter matrix and the smoothing matrix are subjected to matrix subtraction calculation to obtain a detail extraction matrix, a dimension number of the unit filter matrix being equal to that of the smoothing matrix, and an element value of one dimension of the unit filter matrix being 1 and element values of other dimensions being 0; the third signal implicit representation is subjected to detail extraction through the detail extraction matrix to obtain the first detail implicit representation, and the third signal implicit representation is smoothed through the smoothing matrix to obtain the first global implicit representation. ​ 4. The method of claim 3, wherein, The number of the smoothing matrices is s, and the number of dimensions of each smoothing matrix is m. The number of the detail extraction matrices is s, and the number of dimensions of each detail extraction matrix is m. The detail extraction of the third signal implicit representation through the detail extraction matrix and the smoothing of the third signal implicit representation through the smoothing matrix to obtain the first detail implicit representation and the first global implicit representation include: The third signal implicit representation is classified into s third signal sub-implicit representations in the dimension. The number of sampling points covered by the smoothing matrix, the number of sampling points covered by the detail extraction matrix, and the number of sampling points covered by the third signal sub-implicit representation are consistent. The s third signal sub-implicit representations are subjected to detail extraction through the s detail extraction matrices to obtain s first detail sub-implicit representations. The s first detail sub-implicit representations constitute the first detail implicit representation. The s third signal sub-implicit representations are subjected to smoothing through the s smoothing matrices to obtain s first global sub-implicit representations. The s first global sub-implicit representations constitute the first global implicit representation.

5. The method of claim 1, wherein, The frequency spectrum analysis of the first signal implicit representation to obtain the first detail implicit representation and the first global implicit representation includes: The frequency spectrum analysis of the first signal implicit representation to obtain a pending detail implicit representation and a pending global implicit representation. The pending detail implicit representation and the pending global implicit representation are integrated according to the detail influence variable and the global influence variable to obtain a pending signal implicit representation. The frequency spectrum analysis of the first signal implicit representation to obtain a pending detail implicit representation and a pending global implicit representation. The pending detail implicit representation and the pending global implicit representation are integrated according to the detail influence variable and the global influence variable to obtain a pending signal implicit representation.

6. The method of claim 5, wherein, The integration of the pending detail implicit representation and the pending global implicit representation according to the detail influence variable and the global influence variable to obtain a pending signal implicit representation includes: If the round of the next frequency spectrum analysis is not greater than a second preset value, the second preset value is less than the first preset value, the pending detail implicit representation and the pending global implicit representation obtained in the next round are fused according to the detail influence variable and the global influence variable to obtain a first control signal implicit representation. The first electric stimulation monitoring signal is up-sampled or down-sampled. The feature implicit representation of the first electric stimulation monitoring signal after up-sampling or down-sampling is obtained. obtaining a current pending signal implicit representation by feature implicit representation of the first control signal implicit representation and the second control signal implicit representation through a weight focusing strategy; obtaining a current pending signal implicit representation by feature implicit representation of the first control signal implicit representation and the second control signal implicit representation through a weight focusing strategy; obtaining a third control signal implicit representation by fusing the current pending detail implicit representation and the current pending global implicit representation according to the detail influence variable and the global influence variable if the current round of spectrum analysis is greater than a third preset value, the third preset value being not greater than the first preset value; obtaining a current pending signal implicit representation by integrating one or more pending signal implicit representations obtained in iterations before the current round and the third control signal implicit representation.

7. The method according to any one of claims 1 to 6, characterized in that, The wireless implantable neuromodulation electric stimulation monitoring data classification method is implemented based on an electric stimulation monitoring signal processing network, the electric stimulation monitoring signal processing network comprising a feature implicit representation component, a spectrum analysis component, an adaptive fusion component and a signal generation component; the feature implicit representation component is configured to perform feature implicit representation on the first electric stimulation monitoring signal to obtain the first signal implicit representation; the spectrum analysis component is configured to perform spectrum analysis on the first signal implicit representation to obtain the first detail implicit representation and the first global implicit representation; the adaptive fusion component is configured to integrate the first detail implicit representation and the first global implicit representation according to the detail influence variable and the global influence variable to obtain the second signal implicit representation, the detail influence variable and the global influence variable being component parameters of the adaptive fusion component; the signal generation component is configured to filter the second signal implicit representation through the first filter matrix to obtain the residual mapping signal, and integrate the first electric stimulation monitoring signal and the residual mapping signal to obtain the second electric stimulation monitoring signal, the first filter matrix being a component parameter of the signal generation component.

8. The method of claim 7, wherein, The electric stimulation monitoring signal processing network is trained through the following steps: performing feature implicit representation on the noise electric stimulation monitoring signal sample through the feature implicit representation component to obtain a first signal training implicit representation; performing spectrum analysis on the first signal training implicit representation through the spectrum analysis component to obtain a first detail training implicit representation and a first global training implicit representation; integrating the first detail training implicit representation and the first global training implicit representation according to the detail influence variable and the global influence variable through the adaptive fusion component to obtain a second signal training implicit representation; filtering the second signal training implicit representation through the first filter matrix through the signal generation component to obtain a residual training mapping signal, and integrating the noise electric stimulation monitoring signal sample and the residual training mapping signal to obtain an estimated electric stimulation monitoring signal; The electrical stimulation monitoring signal processing network is trained based on a loss between the estimated electrical stimulation monitoring signal and the denoised electrical stimulation monitoring signal sample.

9. A computer system, characterized by Comprise: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, the one or more computer programs are executed by the processor, and the method as claimed in any one of claims 1-8 is implemented.

Images