Regulation and control monitoring information visualization method and system based on implantable medical chip

Abstract

The invention provides a regulation and control monitoring information visualization method and system based on an implantable medical chip, and relates to the technical field of data processing, and the method comprises the steps: obtaining an original living body neural signal, carrying out the signal feature abstraction of the original living body neural signal, obtaining an abstraction signal, and neglecting the limitation of the original living body neural signal, the method comprises the following steps: carrying out feature embedding on an original living neural signal to obtain a first signal feature representation with a small number of sampling points, carrying out feature embedding on an abstracted signal to obtain a second signal feature representation with a small number of sampling points, and carrying out noise addition on the first signal feature representation to obtain a noise neural signal; since the abstracted signal ignores the limitation on the original living neural signal, and the visual constraint parameter can provide different visual signal parameters, the visual neural signal constrained by the visual constraint parameter can be accurately obtained based on the visual constraint parameter, and the reliability of the visual neural signal is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and particularly relates to a regulation and monitoring information visualization method and system based on an implantable medical chip. BACKGROUND

[0002] With the rapid development of neuroscience and medical technology, implantable medical chips are increasingly widely used in the field of neural regulation and monitoring. Such chips can collect real-time in vivo neural signals in the human body, providing doctors with a new way to directly observe and analyze neural activity. The original in vivo neural signals can have high complexity and redundancy, and their high sampling point number and random fluctuation characteristics bring great challenges to subsequent signal processing and analysis. Traditional neural signal processing methods rely on specific signal features and pre-set analysis frameworks, which limits their versatility and flexibility for different types of neural signals. In particular, for neural signals from different individuals, different physiological states or different collection conditions, traditional methods are difficult to ensure the consistency and accuracy of the processing results. In addition, traditional visualization methods often focus on the direct display of signals, lack the ability to flexibly adapt to and individually customize the visualization needs of users, and are difficult to meet the needs of medical professionals for in-depth analysis of signals. SUMMARY

[0003] Therefore, the embodiments of the present application at least provide a regulation and monitoring information visualization method and system based on an implantable medical chip. The technical solution of the present application is realized as follows: In one aspect, the present application provides a regulation and monitoring information visualization method based on an implantable medical chip, comprising: acquiring an original in vivo neural signal, abstracting signal features of the original in vivo neural signal to obtain an abstracted signal, embedding features of the original in vivo neural signal to obtain a first signal feature representation, embedding features of the abstracted signal to obtain a second signal feature representation; the sampling point number of the original in vivo neural signal is greater than the sampling point number of the first signal feature representation and the second signal feature representation; adding noise to the first signal feature representation to obtain a noise-added neural signal; acquiring a visualization constraint parameter, mapping the visualization constraint parameter to a constraint implicit representation array, mapping the second signal feature representation to a signal implicit representation array, and the visualization constraint parameter is used to describe the visualization features of the visualization display result; removing noise from the noise-added neural signal according to the constraint implicit representation array and the signal implicit representation array to obtain a noise-cleaned transition signal, and restoring the noise-cleaned transition signal into a visualization neural signal conforming to a target visualization feature; the sampling point numbers of the original in vivo neural signal and the visualization neural signal are consistent.

[0004] In another aspect, the present application provides a computer system comprising a memory and a processor, wherein the memory stores a computer program capable of running on the processor, and the processor implements the steps of the above method when executing the program.

[0005] The beneficial effects of the present application include: the embodiment of the present application can ignore the limitation of the original living nerve signal based on obtaining the original living nerve signal, abstract the signal characteristics of the original living nerve signal, obtain the abstracted signal, and different original living nerve signals can be converted into visualized nerve signals, the original living nerve signal can be embedded into the first signal feature representation with less sampling points, the abstracted signal can be embedded into the second signal feature representation with less sampling points, noise can be added to the first signal feature representation, the noise can be added to obtain the noise signal, the visualized constraint parameter and the second signal feature representation can be mapped into the constraint implicit representation array and the signal implicit representation array based on the focusing strategy, which can be used as the guide array of the target machine learning algorithm, the noise signal can be removed based on the guide array to guide the target machine learning algorithm, and the noise cleaning transition signal more related to the abstracted signal and the visualized constraint parameter can be obtained, and the noise cleaning transition signal can be restored into the visualized nerve signal conforming to the visualized feature. Because the abstracted signal ignores the limitation of the original living nerve signal, and the visualized constraint parameter can provide different visualized signal parameters, the visualized nerve signal constrained by the visualized constraint parameter can be accurately obtained based on the visualized constraint parameter, and the reliability of the visualized nerve signal is improved. BRIEF DESCRIPTION OF DRAWINGS

[0006] The accompanying drawings, which are incorporated into and form a part of the specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the application.

[0007] Figure 1 An implementation process schematic diagram of a regulation monitoring information visualization method based on an implantable medical chip provided by the embodiment of the present application.

[0008] Figure 2 A hardware entity schematic diagram of a computer system provided by the embodiment of the present application. DETAILED DESCRIPTION

[0009] The embodiment of the present application provides a regulation monitoring information visualization method based on an implantable medical chip, which can be executed by a processor of a computer system. The computer system can be a server, a notebook computer, a tablet computer, a desktop computer or the like device with data processing capability. Figure 1 An implementation process schematic diagram of a regulation monitoring information visualization method based on an implantable medical chip provided by the embodiment of the present application is as follows: Figure 1As shown, the method comprises: Step S100: Obtain the original living body neural signal, abstract the signal features of the original living body neural signal to obtain an abstracted signal, embed the features of the original living body neural signal to obtain a first signal feature representation, and embed the features of the abstracted signal to obtain a second signal feature representation; the number of sampling points of the original living body neural signal is greater than the number of sampling points of the first signal feature representation and the second signal feature representation.

[0010] In step S100, the computer system first connects to the organism through the implantable medical chip and collects the original living body neural signal in real time. These signals can be composed of a series of electrical signals, reflecting the electrical activity of neurons in the organism. For example, in brain-machine interface (BMI) applications, an implantable electrode array can record local field potentials (LFP) or spike activity in the cerebral cortex region.

[0011] Before feature abstraction, the computer system can preprocess the original living body neural signal to remove noise and interference and improve signal quality. This can include steps such as filtering, detrending, baseline correction, etc. For example, a band-pass filter can be used to remove low-frequency drift and high-frequency noise, and only the frequency band related to neural activity (such as 0.1 Hz to 300 Hz) is retained.

[0012] Signal feature abstraction is the core step in step S100, which aims to extract useful features from the complex raw signal for subsequent processing and analysis, while ignoring unimportant details. This process can include the following procedures: the computer system first identifies action potentials (or called spikes) in the raw signal, which are direct manifestations of neuronal firing. For example, a threshold detection method can be set to identify signal segments that exceed the set threshold as potential action potentials. For each identified action potential, the system calculates its waveform feature values, such as peak amplitude, half-width, rise time, etc. In order to reduce the impact of noise on feature extraction, the computer system smoothes the raw signal or the identified action potentials. This can be achieved by moving average filtering, median filtering or Gaussian filtering, etc. The smoothed signal (called smoothed signal) will be smoother, which is beneficial for subsequent feature calculation. Then, the system calculates the time gradient and amplitude gradient of the smoothed signal, i.e. the rate of change of the signal with time and the rate of change of the amplitude. Based on these gradient information, the system can further perform local extreme value detection to more accurately locate the starting and ending points of the action potential. For example, if a signal segment shows a rapid change from positive to negative in the time gradient, and the amplitude gradient also changes significantly, then this segment may be an action potential. After identifying all important action potentials, the system performs feature abstraction based on the waveform features, time gradients and amplitude gradients of the action potentials, etc. This step may involve simplifying complex waveforms into a series of feature vectors, each containing multiple feature values such as peak amplitude, width, shape parameters of action potentials, etc. These feature vectors constitute the abstracted signal, which is more concise than the original signal, but retains enough information for subsequent analysis.

[0013] On the basis of feature abstraction, the computer system further performs feature embedding on the original live neural signals and the abstracted signals to generate first signal feature representations and second signal feature representations. Feature embedding maps high-dimensional raw data to a low-dimensional space while trying to preserve important information in the raw data. For the original live neural signals, the system uses target machine learning algorithms (such as autoencoders, variational autoencoders, etc.) for feature embedding. These algorithms learn how to effectively compress high-dimensional raw signals into low-dimensional first signal feature representations through unsupervised learning. For example, an autoencoder model may contain an encoder network and a decoder network, the encoder compresses the raw signal into a low-dimensional vector (first signal feature representation), and the decoder tries to reconstruct the original signal from the vector. The model is trained by minimizing the reconstruction error. Similarly, the system performs feature embedding on the abstracted signals to generate second signal feature representations. Since the abstracted signals have been preliminarily feature extracted and simplified, their feature embedding process may be more direct and efficient. For example, linear transformation, principal component analysis (PCA) or other dimensionality reduction techniques can be directly used to generate second signal feature representations. Through the feature embedding process, the number of sampling points of the original live neural signals and the abstracted signals is significantly reduced. This is because feature embedding aims to preserve the main features of the data while removing redundant and noisy information. For example, the original signal may contain millions of sampling points, while the first signal feature representation and the second signal feature representation after feature embedding may only have a few hundred or a few thousand elements. This reduction in the number of sampling points not only reduces the computational complexity of subsequent processing, but also improves the running efficiency of the algorithm.

[0014] Step S100 successfully transforms complex original live neural signals into more easily processed and visualized first signal feature representations and second signal feature representations through three sub-steps of signal preprocessing, feature abstraction and feature embedding. This process not only reduces the redundancy and noise of the data, but also preserves the key information.

[0015] Step S200: Add noise to the first signal feature representation to obtain a noisy neural signal.

[0016] In step S200, the computer system can choose to add different types of noise. Common noise types include Gaussian noise, salt and pepper noise, multiplicative noise, etc. Each type of noise has its specific distribution characteristics and application scenarios. Taking Gaussian noise as an example, the computer system can add noise to the first signal feature representation according to the following steps: First, determine the intensity of noise to be added, which can be controlled by a noise scaling factor (e.g., σ). The size of the noise scaling factor determines the amount of noise added to the signal, and it can be adjusted according to specific application scenarios and experimental requirements. For example, a small σ value can be set to simulate slight interference, or a large σ value can be set to simulate severe interference. According to the selected noise type and intensity, the computer system generates a noise vector with the same dimension as the first signal feature representation. For Gaussian noise, random numbers conforming to Gaussian distribution can be generated by a random number generator and filled into the noise vector. The generated noise vector is added element by element to the first signal feature representation (for multiplicative noise, it is multiplied element by element and then added to the original signal). In this way, each element will be disturbed by a certain degree of noise.

[0017] The first signal feature representation after noise addition processing (now called the noisy neural signal) will be used in subsequent steps, such as mapping of visual constraint parameters, noise removal, and signal restoration, etc. The presence of the noisy neural signal makes the entire processing flow more close to the actual application scenario, improving the practicality and reliability of the model.

[0018] For example, assume the first signal feature representation is a 100-length vector x = [x1, x2, …, x100], where each element xi represents the projection of the original living neural signal on a certain feature dimension. The computer system selects Gaussian noise as the noise type and sets the noise scaling factor σ = 0.1. Then it generates a Gaussian noise vector n = [n1, n2, …, n100] of the same length 100, where each element ni is randomly drawn from the standard normal distribution N(0, 1). Finally, the noise vector n is multiplied by the noise scaling factor σ and added to the first signal feature representation x, resulting in the noisy neural signal y = x + 0.1·n. 100 i ], where each element x 100 i is randomly drawn from the standard normal distribution N(0, 1). Finally, the noise vector n is multiplied by the noise scaling factor σ and added to the first signal feature representation x, resulting in the noisy neural signal y = x + 0.1·n.

[0019] Step S200 simulates the interference factors that the signal may be subjected to in the real world by adding noise to the first signal feature representation. This process enhances the robustness and generalization ability of subsequent processing steps, making the entire visualization process more close to the actual application scenario. By reasonably selecting the noise type and intensity, the computer system can flexibly control the characteristics of the noisy neural signal to meet the needs of different experiments and applications.

[0020] Step S300: Obtain the visualization constraint parameters, map the visualization constraint parameters to the constraint implicit representation array, map the second signal feature representation to the signal implicit representation array, and the visualization constraint parameters are used to describe the visualization characteristics of the visualization display result. ​​

[0021] The visualization constraint parameters are a set of parameters that reflect the user's visualization needs, which can be represented by tags. In a computer system, these parameters can exist in the form of key-value pairs, JSON objects, or data structures in a specific format. For example, the user may specify a color mapping scheme (such as "display red for high action potential amplitude and blue for low") or specify a dynamic display method for the signal (such as "signal intensity changes over time in the form of a waveform").

[0022] The computer system first parses these visualization constraint parameters, understands the user's intentions, and converts them into internal representations. This conversion process may involve validity verification, type conversion, unit unification, and other operations to ensure that subsequent steps can be correctly processed.

[0023] In order to effectively apply the visualization constraint parameters to the subsequent signal processing and visualization generation process, the computer system maps these parameters to a constraint implicit representation array. This process can be achieved through a constraint embedding branch component in the target machine learning algorithm (such as a deep learning model).

[0024] The constraint embedding branch component is a network structure specially designed to process visualization constraint parameters. It receives visualization constraint parameters as input and, through a series of nonlinear transformations (such as fully connected layers, convolutional layers, activation functions, etc.), embeds these parameters into a high-dimensional space to generate a constraint implicit representation array. This array is a numerical vector, and each element represents the contribution or influence of the visualization constraint parameter in a specific dimension. For example, assuming the user specifies a color mapping scheme as the visualization constraint parameter, the constraint embedding branch component can convert this scheme into a vector containing multiple color coding values, each value corresponding to a specific signal intensity range or feature category. In this way, in the subsequent signal processing and visualization generation process, the computer system can determine how to map different parts of the signal to different colors according to this vector.

[0025] At the same time, the computer system also maps the second signal feature representation to a signal implicit representation array. This step is similar to the generation process of the constraint implicit representation array, but the input data is different. The signal implicit representation array is a further abstraction and compression of the second signal feature representation, which retains the main feature information of the signal while removing redundant and noisy parts.

[0026] The generation of the signal implicit representation array can also rely on a component in the target machine learning algorithm (which may be different from the constraint embedding branch component). This component receives the second signal feature representation as input and, through a series of transformation operations (such as the encoding part of the autoencoder, dimension reduction algorithm, etc.), maps it to a low-dimensional implicit space to generate the signal implicit representation array.

[0027] For example, assume there is a convolutional neural network (CNN) based target machine learning algorithm that contains a constraint embedding branch component and a signal embedding component. In step S300: The user specifies a color mapping scheme as "action potential amplitude between 0-50 μV is displayed as blue, 50-100 μV as green, and 100 μV and above as red". These parameters are encoded into a vector containing color coding values, for example [blue_code, green_code, red_code], where each color coding value is a three-dimensional vector (corresponding to the red, green, and blue components in the RGB color space). The constraint embedding branch component receives this color coding vector as input and maps it to a higher-dimensional implicit representation vector (e.g., 128-dimensional) through a series of convolutional layers, pooling layers, and fully connected layers, etc. Each element in this vector contains some abstract information about the color mapping scheme, but no longer directly corresponds to a specific color coding value. Meanwhile, the signal embedding component receives the second signal feature representation (assume it is a 64-dimensional vector) as input and maps it to a lower-dimensional implicit representation vector (e.g., 32-dimensional) through the encoding part of an autoencoder or similar dimension reduction algorithm. This vector compactly represents the main features of the original live neural signal.

[0028] Step S300 provides the necessary input data for the subsequent noise removal and signal restoration steps by mapping the visualization constraint parameters and the second signal feature representation into constraint implicit representation arrays and signal implicit representation arrays, respectively. These two implicit representation arrays not only contain the user's desired visualization feature information but also retain the main feature information of the signal, allowing the computer system to remove noise while preserving the original characteristics of the signal, ultimately generating a visualized neural signal that meets the desired visualization characteristics.

[0029] Step S400: According to the constraint implicit representation array and the signal implicit representation array, the noisy neural signal is de-noised to obtain a noise-cleared transitional signal, and the noise-cleared transitional signal is restored to a visualized neural signal that meets the target visualization characteristics; the sampling point number of the original live neural signal and the visualized neural signal is consistent.

[0030] In step S400, the goal of noise removal is to retain the main features of the signal while removing as much noise as possible. This process can rely on the de-noising component in the target machine learning algorithm or related algorithms. Taking a deep learning model as an example, the computer system can use a neural network containing a de-noising layer to achieve noise removal. The de-noising layer can be a convolutional layer, an autoencoder layer, or a specially designed de-noising network layer. In the training process, these layers learn how to recover a clean signal from a noisy signal.

[0031] In implementation, the computer system provides the constraint implicit representation array and the signal implicit representation array as conditional inputs or additional information to the denoising network. These implicit representation arrays are integrated into the denoising process in some way (e.g., concatenation, weighted sum, or conditional input) to guide the network on how to remove noise while preserving information related to the visualization constraints and signal features.

[0032] For example, assume the denoising network is a conditional generative adversarial network (cGAN) based architecture, where the generator (G) is responsible for generating a noise-cleaned transitional signal from the noisy neural signal, and the discriminator (D) is responsible for distinguishing whether the generated signal is close enough to the real one and complies with the visualization constraints. In this case, the constraint implicit representation array and the signal implicit representation array can be used as conditional inputs to the generator, influencing the style and content of the generated signal.

[0033] Through the denoising process, the computer system generates a noise-cleaned transitional signal. This signal is the result of the noisy neural signal after noise removal, which removes most of the noise components while preserving the main features of the signal and the user's desired visualization features.

[0034] The generation of the noise-cleaned transitional signal is an iterative or optimization process, which may involve multiple forward and backward propagations. In each iteration, the denoising network generates a candidate noise-cleaned signal based on the current input (noisy neural signal, constraint implicit representation array, signal implicit representation array) and the current network parameters. Then, the computer system evaluates the quality of this candidate signal through some loss function (such as mean square error loss, adversarial loss, feature matching loss, etc.), and updates the network parameters according to the gradient information of the loss function.

[0035] After obtaining the noise-cleaned transitional signal, the computer system further restores it to a visualized neural signal that complies with the target visualization features. This process may involve signal reconstruction, feature mapping, color mapping, etc. First, the computer system converts the noise-cleaned transitional signal from the implicit space back to the original signal space. This can be achieved through a decoder or inverse transformation process, which is the opposite of the feature embedding process in step S100. The decoder or inverse transformation function gradually recovers the sampling points and feature dimensions of the original signal based on the information in the noise-cleaned transitional signal. Then, the computer system further processes the restored signal according to the visualization constraint parameters, such as applying color mapping, adjusting dynamic effects, etc., to generate the final visualized neural signal. For example, if the visualization constraint parameters specify a color mapping scheme, the computer system will select appropriate colors for rendering based on the signal intensity or feature values in the noise-cleaned transitional signal.

[0036] For example, assume the noisy neural signal is a one-dimensional array of length N, representing noisy neural activity collected at T time points. The constraint implicit representation array is a vector of length M, containing visualization constraint information such as color mapping, dynamic effects, etc.; the signal implicit representation array is a vector of length K, summarizing the main features of the signal.

[0037] During the denoising process, the computer system uses a cGAN-based denoising network. The generator G receives the noisy neural signal, the constraint implicit representation array, and the signal implicit representation array as input, and generates a noise-cleared signal of length N. The discriminator D is responsible for evaluating the quality of the generated signal and optimizing the generator G through adversarial training.

[0038] During the training process, the computer system defines a composite loss function L that combines the mean square error loss (used to evaluate the accuracy of signal reconstruction) and the adversarial loss (used to encourage the generation of signals that meet the visualization constraints): L = λ1·MSE(x clean ,G(x noisy ,c,z))+λ2·log(1-D(G(x noisy ,c,z))); Where x clean is the true clean signal (available during the training phase), x noisy is the noisy neural signal, c is the constraint implicit representation array, z is the signal implicit representation array, and λ1 and λ2 are hyperparameters that balance the weights of different loss terms.

[0039] By minimizing the loss function L, the computer system continuously optimizes the denoising network until it generates a visualized neural signal that meets the target visualization characteristics. Ultimately, this signal not only removes the noise in the original noisy neural signal, but also accurately reflects the user's visualization needs, providing medical professionals with intuitive and clear neural activity monitoring results.

[0040] As an implementation, in step S100, an original living neural signal is obtained, and signal feature abstraction is performed on the original living neural signal to obtain an abstracted signal. Specifically, it can include: Step S110: Obtain waveform feature values corresponding to x action potentials in the original living neural signal, respectively, and perform smoothing operation on the original living neural signal according to the waveform feature values to obtain a smoothed signal; x≥1; Step S120: Obtain the time axis change rate and amplitude axis change rate corresponding to the x action potentials in the smoothed signal respectively, determine the intensity and angle corresponding to the x action potentials in the smoothed signal respectively according to the time axis change rate and amplitude axis change rate, perform local extremum detection on the smoothed signal based on the intensity and angle, and obtain an extremum nerve signal; Step S130: Generate an abstract signal according to the time axis change rate and amplitude axis change rate corresponding to the x action potentials in the extremum nerve signal.

[0041] Step S100 is the basis of the entire implantable medical chip-based regulation monitoring information visualization method, and its core task is to extract key features from the original living nerve signal, providing basic data for subsequent signal processing and visualization. In this embodiment, step S100 is refined into three specific steps: S110, S120 and S130, each step carrying a specific data processing task. The following is a detailed explanation of these steps, combined with examples for illustration.

[0042] In step S110, the computer system captures the waveform features of action potentials from the original living nerve signal, and performs smoothing processing on the signal based on these features to reduce noise and interference and improve signal quality.

[0043] Specifically, the computer system first analyzes the original living nerve signal, which can be a series of voltage values changing over time, reflecting the activity of neurons. By setting appropriate threshold values, the system can identify action potentials in the signal, i.e. those parts of the voltage values that deviate significantly from the baseline level. These action potentials are a direct manifestation of neuron firing and are crucial for subsequent feature extraction. For each identified action potential, the computer system calculates its waveform feature values. These feature values may include peak amplitude, half-width (i.e. the time required for the action potential to rise from the baseline to the peak and then fall back to the baseline), rise time, fall time, etc. These feature values quantify the shape and size of the action potential and are an important basis for subsequent smoothing processing.

[0044] Based on the calculated waveform feature values, the computer system performs smoothing processing on the original live neural signal. The purpose of smoothing processing is to reduce high-frequency noise and random fluctuations in the signal, making the signal smoother and facilitating subsequent feature extraction and analysis. Common smoothing methods include moving average filtering, median filtering, Gaussian filtering, etc. For example, a suitable window size (such as 5 or 7 sampling points) can be selected, and the signal values within the window are averaged or median calculated to obtain the smoothed signal values. This process will iterate through the entire original signal until a complete smoothed signal is generated. For example, assume that the original live neural signal is an array containing 1000 sampling points, with a sampling rate of 10 kHz. By setting the threshold to -50 μV (relative to the baseline level), the system identifies 50 action potentials. For each action potential, the system calculates its peak amplitude, half-width, rise time, and fall time, and other waveform feature values. Subsequently, a moving average filter with a window size of 7 sampling points is selected to smooth the original signal. The processed smoothed signal will be used as input for the next step of analysis.

[0045] After obtaining the smoothed signal, step S120 further analyzes the time gradient and amplitude gradient (i.e., the rate of change of the time axis and the rate of change of the amplitude axis) in the signal, based on which the intensity and angle of the action potential are determined, and local extremum detection is performed, so as to more accurately locate the features of the action potential.

[0046] The computer system first calculates the time gradient and amplitude gradient of each sampling point in the smoothed signal. The time gradient reflects the rate of change of the signal value with respect to time, while the amplitude gradient reflects the rate of change of the signal value with respect to amplitude. These gradient information is crucial for understanding the dynamic characteristics and local features of the signal. Gradient calculation can be achieved through difference method, i.e., calculating the difference between adjacent sampling points to approximate the gradient value.

[0047] Based on the calculated gradient information, the computer system can further determine the intensity and angle of each action potential. Here, "intensity" can be understood as the amplitude or energy level of the action potential, while "angle" may refer to the angle between the gradient vector and a certain reference direction, used to describe the shape or directional characteristics of the action potential. However, in practical applications, the term "angle" may not be directly applicable to the field of neural signal processing, and here it is more to illustrate the multidimensionality of gradient information in describing signal features. In actual operation, more attention may be paid to the absolute value or directionality (such as positive or negative) of the gradient, rather than the specific angle value.

[0048] Next, using the gradient information and intensity information, the computer system performs local extremum detection. Local extremum refers to a point in the signal whose value is greater than its neighboring points (for maxima) or less than its neighboring points (for minima). In neural signal processing, action potentials can manifest as local maxima in the signal. By comparing the value of each sample point with its neighboring points, the system can identify all local maxima points, which correspond to the peak positions of action potentials.

[0049] For example, continuing with the smoothed signal in step S110. Suppose the smoothed signal is an array of length 1000. The computer system first calculates the temporal gradient and amplitude gradient for each sample point in the array. Then, based on the gradient information, it determines the "intensity" (simplified here as peak amplitude) and "directionality" (judged by comparing the signs of neighboring gradient values) of each action potential. Next, the system performs local extremum detection by comparing each sample point with its two neighboring points to identify local maxima points. Finally, the system obtains an array of indices containing the peak positions of action potentials, which correspond to the local maxima points in the smoothed signal.

[0050] In step S130, the computer system generates an abstracted signal based on the temporal gradient and amplitude gradient information in the extremum neural signal (i.e., the signal containing local extremum points). The abstracted signal is a simplified representation of the original signal that retains the main features while ignoring the minor details.

[0051] Based on the extremum neural signal, the computer system further analyzes the temporal gradient and amplitude gradient of each action potential. These gradient information reflects the dynamic characteristics and shape features of action potentials. However, due to the complexity and high dimensionality of the original signal, directly processing all gradient information can be both time-consuming and inefficient. Therefore, the system needs to select those features that are most valuable for subsequent analysis and perform dimensionality reduction to reduce computational load. There are many methods for dimensionality reduction, such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), t-SNE, etc. However, in the field of neural signal processing, since the features of action potentials are relatively simple and clear (such as peak amplitude, half-width, etc.), it may be more inclined to use feature selection-based methods rather than complex dimensionality reduction algorithms. The system can select a few key features (such as peak amplitude and half-width) to represent each action potential.

[0052] After selecting the key features, the computer system combines these features into a new array or vector to represent each action potential. This new representation is the abstracted signal. Compared with the original live neural signal, the abstracted signal has lower dimensionality and simpler structure, but it still contains enough information to reflect the main features of the original signal.

[0053] The specific form of the abstracted signal depends on the selected features and the requirements of the subsequent analysis. For example, if the system is mainly interested in the amplitude information of the action potentials without caring about the shape details, the abstracted signal can simply be an array containing the peak amplitude of all the action potentials. If the system also needs to consider the duration or shape features of the action potentials, the abstracted signal can contain more dimensions of features. Optionally, to improve the efficiency and stability of the subsequent processing, the computer system can perform standardization or normalization on the abstracted signal. Standardization is to scale the signal values into a certain fixed range (such as 0 to 1 or -1 to 1), while normalization is to adjust the mean and variance of the signal to conform to a certain statistical distribution (such as normal distribution). These processing steps help to reduce the impact of the differences between different signals on the subsequent analysis.

[0054] For example, suppose in step S120, the computer system has identified all the local maximum points in the smoothed signal and calculated the time gradient and amplitude gradient corresponding to these points. Now, the system selects the peak amplitude and half-width as the key features to represent each action potential. For each action potential, the system extracts its peak amplitude and half-width values and combines these values into a two-dimensional array (or vector). Assuming that a total of 50 action potentials are identified, the final abstracted signal will be a 50x2 array or list set.

[0055] Further, for the convenience of subsequent processing, the system can perform normalization on this abstracted signal. For example, the peak amplitude can be normalized to the range of 0 to 1 (achieved by subtracting the minimum value and dividing by the difference between the maximum and minimum values), while the half-width may not need to be normalized (as the value of the half-width can vary within a relatively stable range). The normalized abstracted signal will be more suitable for the input of machine learning models or further signal processing tasks.

[0056] As an implementation, step S130, generating an abstracted signal according to the time axis change rate and the amplitude axis change rate corresponding to the x action potentials in the extreme value neural signal, can include: Step S131: obtaining a first reference value and a second reference value, determining the change rate corresponding to the x action potentials in the extreme value neural signal according to the time axis change rate and the amplitude axis change rate corresponding to the x action potentials in the extreme value neural signal; Step S132: determining a significant action potential in the x action potentials of the extreme value neural signal, the change rate of the significant action potential being greater than or equal to the first reference value; Step S133: determining a fuzzy action potential in the x action potentials of the extreme value neural signal, the change rate of the fuzzy action potential being greater than or equal to the second reference value and less than the first reference value; Step S134: Determine the ambiguous action potentials adjacent to the salient action potential as candidate action potentials, and generate the abstracted signal according to the salient action potential and the candidate action potentials.

[0057] In step S131, the computer system sets two reference values, a first reference value and a second reference value, which will be used in the subsequent classification of action potentials. The selection of the reference values can be based on prior knowledge, experimental data or expert opinion, aiming to distinguish action potentials of different saliency.

[0058] Specifically, the first reference value and the second reference value can be preset according to experimental design or historical data analysis results. For example, in the setting of the first reference value, a certain percentile (such as 90%) of the absolute value of the time gradient or amplitude gradient of the action potential can be considered as a threshold, and the action potential exceeding this threshold is considered to have higher saliency. Similarly, the second reference value can be set to a lower percentile (such as 70%) to distinguish those ambiguous action potentials whose change rate is not as high as the salient action potential, but is still higher than the random noise level. For each action potential in the extreme neural signal, the computer system calculates its comprehensive change rate according to its time axis change rate and amplitude axis change rate (i.e. time gradient and amplitude gradient). Here, the comprehensive change rate can be a certain weighted sum or Euclidean distance of the time gradient and amplitude gradient, depending on the relative importance of the action potential features. For example, if the time gradient is more important for distinguishing action potentials, it can be given a higher weight.

[0059] For example, assume that the extreme neural signal contains 100 action potentials, and the time gradient and amplitude gradient of each action potential have been calculated in the previous step. The computer system sets the first reference value to be greater than 0.5 V / ms (assuming the unit) and the second reference value to be 0.2 V / ms. For each action potential, the computer system calculates the weighted sum of its time gradient and amplitude gradient as the comprehensive change rate, with the weights determined according to experimental design or expert opinion, for example, the time gradient weight is 0.7 and the amplitude gradient weight is 0.3.

[0060] After determining the reference values and calculating the change rates, in step S132, the computer system then identifies those action potentials whose change rates exceed the first reference value, i.e. the salient action potentials.

[0061] In particular, for each action potential in the extreme neural signal, the computer system compares its integrated rate of change to a first reference value. If the integrated rate of change of the action potential is greater than or equal to the first reference value, the action potential is labeled as a significant action potential. All segments of the signal that are labeled as significant action potentials are stored for later processing. These significant action potentials can represent the most significant, most meaningful portion of the neural activity. For example, continuing the previous example, the computer system iterates through the 100 action potentials and labels as significant action potentials those action potentials with an integrated rate of change greater than or equal to 0.5 V / ms. Assuming that 20 action potentials satisfy this condition, these action potentials are stored in a dedicated array for later analysis.

[0062] After identifying the significant action potentials, the computer system determines those action potentials whose rates of change are between the first reference value and a second reference value, i.e., ambiguous action potentials. These action potentials, while not as significant as the significant action potentials, can still contain useful information.

[0063] In particular, for the remaining portion of the extreme neural signal that is not labeled as a significant action potential, the computer system performs a comparison filter again. This time, the integrated rate of change of each action potential is compared to a second reference value. If the integrated rate of change of the action potential is greater than or equal to the second reference value and less than the first reference value, the action potential is labeled as an ambiguous action potential.

[0064] All segments of the signal that are labeled as ambiguous action potentials are likewise stored. These ambiguous action potentials can be important in some cases to understanding the complexity of the neural activity. For example, in the previous example, the computer system continues to iterate through the remaining 80 segments of the signal that are not labeled as significant action potentials. Action potentials with an integrated rate of change between 0.2 V / ms and 0.5 V / ms are labeled as ambiguous action potentials. Assuming that 30 action potentials satisfy this condition, these action potentials are stored in another array.

[0065] After determining the significant action potentials and the ambiguous action potentials, the computer system needs to generate an abstracted signal from these action potentials. This process can involve merging and simplifying the significant action potentials and the ambiguous action potentials adjacent to them.

[0066] In particular, for each significant action potential, the computer system checks whether there are ambiguous action potentials adjacent to it. If there are, these ambiguous action potentials are considered as candidate action potentials. The purpose of this step is to preserve the information of ambiguous action potentials that can be functionally associated with the significant action potentials.

[0067] For each significant action potential and its candidate set of action potentials, the computer system selects a series of key features to represent them. These features can include peak amplitude, half-width, rise time, fall time, and relative position with respect to the significant action potential, etc. Then, based on these features, the system generates an abstracted representation to replace the original set of action potentials. This abstracted representation can be a vector or an array containing the key feature values.

[0068] Finally, the computer system combines all the significant action potentials and their corresponding abstracted representations to construct the final abstracted signal. This signal is a lower-dimensional representation that contains the main features of the original signal, facilitating subsequent processing and analysis.

[0069] For example, in the previous example, the computer system iterates through each significant action potential and checks its neighboring action potentials for the presence of ambiguous action potentials. For each significant action potential and its candidate set of ambiguous action potentials (if any), the system selects peak amplitude, half-width, and relative position as key features. Then, for each significant action potential, it generates a vector containing these feature values as the abstracted representation. Finally, these vectors are combined in order to form a new array or list as the abstracted signal output.

[0070] Suppose there are 20 significant action potentials identified, each with an average of 1 to 2 candidate ambiguous action potentials adjacent to it. For each significant action potential and its candidate set, the system generates a vector containing 3 feature values (peak amplitude, half-width, relative position index). The final output abstracted signal will be an array or list of 60 to 120 feature values, depending on the number of candidate ambiguous action potentials.

[0071] As an implementation, in step S100, the feature embedding of the original living body neural signal to obtain the first signal feature representation can include: Step S140: loading the original living body neural signal into a target machine learning algorithm; the target machine learning algorithm includes a latent space mapping branch component, and the latent space mapping branch component includes a feature representation operator; Step S150: performing feature extraction on the original living body neural signal based on the feature representation operator to obtain a center array and a dispersion array of the original living body neural signal; Step S160: randomly sampling in the center array and the dispersion array of the original living body neural signal to obtain a latent space center array and a latent space dispersion array, and generating the first signal feature representation according to the latent space center array and the latent space dispersion array.

[0072] In step S140, the computer system loads the acquired raw bio- neural signals as input data into the target machine learning algorithm. The target machine learning algorithm is a pre-designed and trained model specifically for processing neural signal data. The algorithm contains multiple components, among which the latent space mapping branch component is the key part to complete the feature embedding task.

[0073] Before loading the raw signals, some preprocessing operations such as denoising, filtering, standardization, etc. can be performed to ensure the quality and consistency of the input data. However, the specific content of these preprocessing steps depends on the characteristics of the signal and the requirements of the algorithm, and may vary depending on the application scenario.

[0074] The appropriate machine learning algorithm is selected according to the task requirements. In the field of neural signal processing, commonly used algorithms include auto encoder, variational auto encoder (VAE), and variants of generative adversarial network (GAN) such as conditional GAN. These algorithms can learn effective feature representations from data through unsupervised learning or supervised learning.

[0075] The preprocessed raw bio- neural signals are loaded as input data into the target machine learning algorithm. This may involve formatting the data into a format acceptable to the algorithm and setting appropriate parameters and data flow.

[0076] For example, suppose a model based on variational auto encoder is chosen as the target machine learning algorithm. The model contains an encoder and a decoder, where the encoder is used to map the raw signal to the latent space (i.e. perform feature embedding), and the decoder attempts to reconstruct the raw signal from the latent space. When loading the raw bio- neural signals, format them as a two-dimensional array, where rows represent different time points and columns represent signal values recorded by different channels or electrodes. Then, pass this array as input to the encoder part of the variational auto encoder.

[0077] After the raw bio- neural signals are loaded into the target machine learning algorithm, feature representation operators are used to extract the features of the signals. Feature representation operators can be one or more neural network layers in the algorithm, which generate useful feature representations by learning the internal structure of the data.

[0078] Specifically, the raw signal is forward propagated through the feature representation operator. In this process, the signal data undergoes a series of linear transformations (such as fully connected layers or convolutional layers) and nonlinear activation functions (such as ReLU, sigmoid, etc.), gradually abstracting high-level features of the signal.

[0079] Under the action of the feature representation operator, the original signal is transformed into a series of feature vectors or feature maps. These features can include statistical information of the signal (such as mean, variance), time-frequency characteristics, spatial distribution, etc. In this embodiment, the central array (μ) and the dispersion array (σ) are mentioned, which refer to the mean and variance representation of the signal in the latent space, for subsequent probability distribution modeling.

[0080] For example, continuing with the variational autoencoder as an example. In the encoder, the original signal is first passed through one or more convolutional layers to extract local features (such as edges, textures, etc.), and then through a pooling layer to reduce the resolution of the feature map and increase the receptive field. Next, the feature map is converted to a one-dimensional vector through a fully connected layer, which is the preliminary representation of the central array (μ). At the same time, the encoder will also generate another one-dimensional vector as the preliminary representation of the dispersion array (σ), which describes the uncertainty or diversity of the features in the latent space. These two vectors will be used in the subsequent steps to generate the latent space central array and the latent space dispersion array.

[0081] After obtaining the central array (μ) and the dispersion array (σ) of the original living neural signal, the next step is to randomly extract (or called sample, decode) on the basis of these arrays to generate the first signal feature representation. This process can involve sampling the latent space distribution or generating feature vectors according to deterministic rules.

[0082] Specifically, in a generative model such as a variational autoencoder, the latent space can be modeled as a Gaussian distribution (or other types of probability distribution), whose mean and variance are given by the encoder (i.e. the central array μ and the dispersion array σ). In this step, one or more latent space points can be randomly sampled from this distribution as feature representations. This method can capture the diversity in the latent space and help generate diverse signal representations.

[0083] In some cases, in order to simplify the calculation or meet certain requirements, it is also possible to choose not to randomly sample but to directly use the central array (μ) as the feature representation (i.e. ignoring the information of the dispersion array σ). This method, although sacrificing diversity, can obtain a deterministic and repeatable feature representation.

[0084] Whether through random sampling or deterministic mapping, the latent space points will be further processed to generate the first signal feature representation. This can involve mapping the latent space points back to the dimension space of the original signal or converting them into another form of feature vector.

[0085] For example, the process of latent space sampling is illustrated by taking the variational autoencoder as an example. Suppose that the center array μ and the dispersion array σ output by the encoder represent the mean and the standard deviation vector of a Gaussian distribution in the latent space, respectively. To generate the first signal feature representation, a point z can be randomly sampled from this Gaussian distribution. Specifically, this can be achieved using the "reparameterization trick": first, a noise ε is sampled from the standard normal distribution N(0, I), and then z = μ + ε * exp(0.5 * σ 2 ) is calculated. Here, exp(0.5 * σ 2 ) is used to convert σ into the standard deviation form (since the encoder can output the logarithmic form of the variance, i.e., log(σ 2 )). The resulting z is the first signal feature representation, which is located in the latent space and captures the key features of the original living neural signal.

[0086] It should be noted that, in practical applications, σ can be constrained to ensure its non-negativity (e.g., using a soft plus function) for stability and numerical convenience. In addition, techniques such as batch normalization and layer normalization can be used to optimize the generation process of the feature representation to speed up the training and inference process.

[0087] As an implementation, the first signal feature representation is generated in a latent space mapping branch component in the target machine learning algorithm, and the latent space mapping branch component includes a normalization operator and a pooling operator. Based on this, in step S200, the first signal feature representation is added with noise to obtain a noisy neural signal, which can include: Step S210: Obtain a noise distribution and an iteration number s, and perform a normalization operation on the waveform feature values of the action potential in the first signal feature representation based on the normalization operator and the noise distribution; s ≥ 1; Step S220: Based on the pooling operator and the iteration number s, the normalized first signal feature representation is added with noise s times to obtain a noisy neural signal, wherein the noise distribution represents the confidence distribution information of the added noise, and the iteration number s represents the degree of adding noise to the first signal feature representation according to the noise distribution.

[0088] Step S210: Obtain the noise distribution and the iteration number, and perform a normalization operation In this step, the computer system obtains the noise distribution and the iteration number s, and uses these parameters to perform a normalization operation on the first signal feature representation, preparing for the subsequent noise addition process.

[0089] Noise distribution is a parameter that describes the type, intensity, and probability distribution of noise in time or space. In practical applications, the choice of noise distribution depends on various factors, such as application scenarios, signal characteristics, and desired simulation effects. Common noise distributions include Gaussian distribution, uniform distribution, Laplace distribution, etc. The parameters of noise distribution (such as mean and variance) can be obtained in various ways. One way is to determine it by analyzing historical data or prior knowledge. For example, if it is known that the noise level of neural signals under certain experimental conditions follows a Gaussian distribution with a mean of 0 and a variance of σ², these parameters can be directly used as the parameters of the noise distribution. Another method is to determine it through experimental design, that is, artificially introducing different types of noise during the experiment and observing its effect on signal quality, so as to select appropriate noise distribution and parameters.

[0090] The number of iterations s is a hyperparameter that controls the degree of noise addition. It determines the number of times or the intensity of noise addition to the first signal feature representation. The selection of the number of iterations s also depends on various factors, such as the signal-to-noise ratio of the signal, the complexity of the noise, and the tolerance of the subsequent processing steps to the noise, etc. In practical applications, the number of iterations s can be determined by cross-validation, grid search, and other optimization algorithms. Specifically, noise can be added to the first signal feature representation under different numbers of iterations, and the performance changes of subsequent processing steps (such as noise removal, signal restoration, etc.) can be observed. By comparing the performance indicators (such as signal reconstruction error, visualization effect, etc.) under different numbers of iterations, the number of iterations that makes the performance of subsequent processing steps optimal is selected as the value of s, which is not limited.

[0091] After determining the noise distribution and the number of iterations, the computer system uses the standardization operator to standardize the action potential waveform feature values in the first signal feature representation. The purpose of standardization is to scale the feature values to a unified scale, so as to facilitate subsequent noise addition and signal processing.

[0092] In the standardization operation, the mean and standard deviation of all action potential waveform feature values in the first signal feature representation can be calculated first; then each feature value is subtracted from the mean and divided by the standard deviation to obtain the standardized feature value. The feature values processed in this way will have zero mean and unit variance, thus eliminating the scale difference between different features.

[0093] For example, suppose the first signal feature representation is a vector f containing multiple action potential waveform feature values, each representing a certain attribute of an action potential (such as peak amplitude, half-width, etc.). First, the computer system calculates the mean μ and standard deviation σ of the vector f. Then, for each feature value f i in the vector f, its standardized value f i '=(f i- μ) / σ. The normalized vector f' obtained after this processing will serve as input for the subsequent noise addition process.

[0094] At the same time, suppose that the noise distribution has been determined to be Gaussian N(0, σ²_noise) through some means such as analyzing historical data or experimental design, where σ²_noise is the variance of the noise. The number of iterations s is then determined to be 3 through an optimization algorithm such as cross-validation.

[0095] Based on step S210, the computer system performs a noise addition operation on the normalized first signal feature representation using a pooling operator and a number of iterations s to generate a noisy neural signal.

[0096] The pooling operator is a downsampling module that aggregates feature values in local regions to reduce the number of parameters and computational load while preserving important feature information. In the noise addition process, the pooling operator can be used to control the distribution characteristics of the noise in space or time. One possible implementation is to perform some form of weighted summation or convolution operation between the noise distribution and the first signal feature representation in space or time to achieve local or global noise addition.

[0097] Based on the normalized first signal feature representation obtained in step S210 and the noise distribution parameters (such as mean, variance, etc.), the computer system begins to perform iterative noise addition operations. In each iteration, a certain number of noise samples are randomly drawn from the noise distribution and added to the normalized first signal feature representation. Through multiple iterations, the strength and complexity of the noise can be gradually enhanced to more realistically simulate the noise conditions in the actual environment.

[0098] It should be noted that after each iteration, the signal after adding noise can be processed (such as re-normalization, limiting noise intensity, etc.) to ensure that it is still within a reasonable range and meets the requirements of subsequent processing steps. For example, continuing the previous example, suppose that the normalized first signal feature representation vector f' and the noise distribution parameters (such as Gaussian distribution N(0, σ²_noise)) have been obtained in step S210. Now enter step S220 to perform iterative noise addition operations.

[0099] First, perform the first iteration: randomly draw a noise vector n1 of the same length as vector f' from the Gaussian distribution N(0, σ²_noise); then add the noise vector n1 to the vector f' to obtain the noise-added vector f'_1 = f' + n1. If you need to limit the noise intensity, you can scale the noise vector before adding (such as multiplying by a scaling factor less than 1).

[0100] A second and third iteration is then performed: the above process is repeated to generate noise vectors n2 and n3 respectively and add them to the vector f'_1 to obtain vectors f'_2 and f'_3 respectively. The final noisy neural signal obtained after three iterations can be represented as f'_3 or a vector that is further processed as needed.

[0101] In practical applications, the cumulative effect of noise can be considered, i.e. the noise intensity can gradually increase during multiple iterations, leading to signal distortion. In order to avoid this situation, certain processing can be performed on the signal after each iteration (such as limiting the noise intensity, adjusting the noise distribution parameters, etc.) to control the cumulative effect of noise.

[0102] As an implementation, in step S300, mapping the visualization constraint parameters to the constraint implicit representation array can include: Step S310: loading the visualization constraint parameters into the target machine learning algorithm, wherein the target machine learning algorithm includes a constraint embedding branch component; the constraint embedding branch component includes a feature embedding operator, a weight allocation operator, and a fusion operator; Step S320: performing feature embedding on the visualization constraint parameters based on the feature embedding operator to obtain y constraint array representations; y≥1; Step S330: generating y constraint array representations respectively corresponding to the focusing weights based on the weight allocation operator, and adjusting the constraint array representations based on the focusing weights to obtain y to-be-used focusing arrays; Step S340: performing weight fusion on the y to-be-used focusing arrays based on the fusion operator to obtain the constraint implicit representation array.

[0103] In the embodiments of the present application, step S300 converts the user-defined visualization constraint parameters into an internal representation form that can be processed by the computer system, i.e. the constraint implicit representation array.

[0104] In step S310, the computer system loads the visualization constraint parameters provided by the user or external system into the target machine learning algorithm. These visualization constraint parameters can include but are not limited to color mapping schemes, dynamic effect settings, view angle selections, label styles, etc., which collectively define the user's desired visualization display effect.

[0105] The target machine learning algorithm is pre-designed and trained, and is a model specially used for processing visualization constraint parameters and generating constraint implicit representation arrays. The algorithm can include multiple branch components, and the constraint embedding branch component is one of the core parts. The constraint embedding branch component is responsible for mapping the visualization constraint parameters to the implicit space of the algorithm for subsequent processing.

[0106] The constraint embedding branch component is further subdivided into a feature embedding operator, a weight assignment operator, and a fusion operator. These operators work together to convert complex visualization constraint parameters into concise and efficient constraint implicit representation arrays. For example, suppose the user defines a set of visualization constraint parameters, including a color mapping scheme (mapping signal intensity to red, yellow, and green colors), a dynamic effect setting (color gradient display when signal intensity changes), a view angle selection (top-down perspective), and a label style (displaying specific numerical values at signal peaks). The computer system loads these visualization constraint parameters into the target machine learning algorithm in the form of structured data (such as a JSON object).

[0107] After the visualization constraint parameters are loaded into the target machine learning algorithm, they are subjected to feature embedding using the feature embedding operator. Feature embedding is the process of converting raw data (in this case, visualization constraint parameters) into a numerical representation that is easier for algorithms to process.

[0108] The feature embedding operator maps the visualization constraint parameters into a high-dimensional feature space by analyzing their internal structure and semantic information. In this feature space, similar visualization constraint parameters will be mapped to close positions, facilitating subsequent weight assignment and fusion operations.

[0109] For example, taking the color mapping scheme as an example, the feature embedding operator may first convert color names (red, yellow, green) into RGB color codes (such as red corresponding to [255, 0, 0], yellow corresponding to [255, 255, 0], and green corresponding to [0, 255, 0]). Then, for the dynamic effect setting (such as color gradient), the feature embedding operator may further generate a series of intermediate color codes to represent each state in the color gradient process. These color codes will be embedded as part of the feature vector in the feature space.

[0110] For other visualization constraint parameters such as view angle selection and label style, the feature embedding operator will also use similar methods to process them into numerical feature vectors. Finally, for a given set of visualization constraint parameters, the feature embedding operator will generate y constraint array representations (where y is the number of constraint parameters or the number of categories), each containing the feature information of the constraint parameter. The feature embedding operator can be an encoder.

[0111] After obtaining y constraint array representations, step S330 uses the weight assignment operator to generate corresponding focus weights and adjust the constraint array representations. The purpose of focus adjustment is to adjust the influence weight of constraint parameters in subsequent processing according to their importance and relevance.

[0112] The weight assignment operator is an attention module that determines the importance of each constraint parameter by analyzing the similarities and differences among the constraint array representations. For constraint parameters that are more critical or highly correlated with other constraint parameters, the weight assignment operator assigns them higher weights; otherwise, it assigns them lower weights.

[0113] For example, suppose three constraint array representations are generated in step S320: color mapping scheme A, dynamic effect setting B, and view angle selection C. The weight assignment operator first analyzes the similarities and differences among these three constraint array representations. For example, if color mapping scheme A and dynamic effect setting B are highly correlated in visual effects (e.g., color gradient needs to rely on a specific color mapping), the weight assignment operator can assign them higher mutual weights.

[0114] Then, the weight assignment operator adjusts the focus of the constraint array representations according to these weight information. The specific operations may include weighted summation of feature values in each array representation, application of nonlinear transformation, or adjustment of the direction of feature vectors, etc. After focus adjustment, the y focus arrays to be used will highlight the key features and reduce redundant information.

[0115] After obtaining the y focus arrays to be used, step S340 uses a fusion operator to weight fuse these arrays to generate the final constraint implicit representation array. The purpose of weight fusion is to integrate the effects of multiple constraint parameters into a unified internal representation form.

[0116] The fusion operator is a computational network layer that generates the constraint implicit representation array by considering the weights of each focus array to be used and their interactions. For example, continuing the previous example, suppose three focus arrays are obtained in step S330: focused color mapping scheme A', focused dynamic effect setting B', and focused view angle selection C'. The fusion operator first reads the weight information of each focus array. These weights may be determined by the weight assignment operator in step S330 or obtained through other mechanisms (such as user input).

[0117] Then, the fusion operator performs fusion operations on the focus arrays according to these weight information. The specific operations include, for example, weighted summation of feature values in each focus array, application of specific fusion functions (such as weighted average, max pooling, etc.), or inference using deep learning models, etc. The final generated constraint implicit representation array will be a comprehensive representation form that contains the effects of all visualization constraint parameters, which will serve as input data for subsequent steps (such as noise removal, signal restoration, and visualization rendering, etc.).

[0118] In practical applications, each sub-step in step S300 can be adjusted and optimized according to specific requirements and algorithm design. For example, the feature embedding operator can use different types of neural network structures (such as fully connected networks, convolutional neural networks, autoencoders, etc.) to achieve more efficient feature extraction; the weight assignment operator can use attention mechanisms or graph neural networks to capture complex relationships between constraint parameters; the fusion operator can combine domain knowledge and optimization algorithms to ensure the accuracy and robustness of the fusion results.

[0119] As an implementation, the noise-added neural signal is obtained based on s times of noise addition to the first signal feature representation, and the first signal feature representation is obtained in a hidden space mapping branch component in the target machine learning algorithm, and the hidden space mapping branch component includes a residual operator and an interpolation operator. Based on this, step S400, according to the constraint implicit representation array and the signal implicit representation array, the noise-added neural signal is de-noised to obtain a noise-cleaning transition signal, which can include: Step S410: combining the constraint implicit representation array and the signal implicit representation array into a guide array based on the residual operator; Step S420: performing first noise removal on the noise-added neural signal based on the interpolation operator and the guide array to obtain a noise-removed signal G s-1 , the noise-removed signal G s-1 is repeatedly de-noised until s times of noise removal, and the noise-removed signal G0 obtained by the s-th noise removal is determined as the noise-cleaning transition signal.

[0120] In step S410, the computer system effectively fuses the constraint implicit representation array extracted from the user's visualization requirements and the signal implicit representation array extracted from the signal features to form a guide array that guides the noise removal process. This fusion process relies on the residual operator in the hidden space mapping branch component, which is good at capturing and emphasizing the difference between the input data, so as to preserve the key information in the fusion process.

[0121] The residual operator can alleviate the gradient vanishing problem in deep network training, and optimize network performance by learning the difference (residual) between input and output. In step S410, the role of the residual operator is slightly different, which is used to effectively fuse the two implicit representation arrays, the constraint implicit representation array and the signal implicit representation array. Specifically, the residual operator can combine the information in the two arrays by element-wise addition or more complex weighted sum to form a guide array that guides noise removal. This fusion method can ensure that the guide array contains both user-defined visualization constraints and signal feature information.

[0122] For example, assume the constraint implicit representation array is a vector C representing a color mapping scheme, where each element corresponds to the color encoding of a different signal intensity interval; the signal implicit representation array is a vector S containing the main features of the signal, such as the encoding of statistical properties like signal intensity, frequency, etc. The computer system fuses these two vectors using a residual operator to form the guide array G_guide. The fusion process can be similar to C + λS, where λ is a weight factor balancing the importance of constraint information and signal features. By adjusting the value of λ, the system can optimize the fusion effect according to the actual application scenario.

[0123] In step S420, after obtaining the guide array, the computer system enters the core part of noise removal, that is, using the interpolation operator and the guide array to iteratively process the noisy neural signal, gradually removing the noise components in the signal. The role of the interpolation operator in this process is to intelligently interpolate or reconstruct the noisy signal according to the information provided by the guide array, to approximate the state of the original pure signal.

[0124] The interpolation operator can estimate the value of unknown data points from known data points (such as sampling points). In step S420, the interpolation operator intelligently interpolates or reconstructs each point in the noisy neural signal according to the constraint information and signal feature information in the guide array, to reduce the impact of noise. This interpolation process can be local (only considering neighboring points) or global (considering the feature distribution of the entire signal).

[0125] Since the noisy neural signal is obtained by adding s times of noise, in order to recover the pure state of the signal as much as possible, the computer system needs to perform the same number of noise removal iterations on the signal. In each iteration, the system will use the interpolation operator and the current guide array to process the signal to generate a noise-removed signal after one iteration. As the number of iterations increases, the noise component gradually decreases, and the signal gradually approaches its pure state.

[0126] For example, assume the noisy neural signal is a one-dimensional array N with a length of L, where each element represents the signal intensity value (possibly containing noise) at a certain time point. In the first iteration, the computer system uses the interpolation operator and the guide array G_guide to process N to generate a noise-removed signal G_1. The processing process may involve local or global interpolation of each element in N according to the color mapping scheme and signal feature information in G_guide to reduce the impact of noise. Subsequently, the system takes G_1 as the new input and repeats the above process to generate G_2, G_3,..., until G_s. After the s-th iteration, G_s (denoted as G0 to avoid confusion) is considered to be a noise-cleared signal, which contains less noise components and retains the main features of the signal and the user's visualization constraints.

[0127] It is worth noting that the guide array G guide can remain unchanged during the iteration process (i.e., the same guide information is used for each iteration), or it can be dynamically adjusted based on the iteration results (e.g., regenerated based on the characteristics of the current noise-removed signal). In addition, the specific implementation of the interpolation operator can vary depending on the application scenario, such as using linear interpolation, spline interpolation, machine learning-based interpolation methods, etc.

[0128] During the execution of step S400, the computer system can evaluate the effectiveness of noise removal by monitoring the changes in these performance indicators, and adjust the algorithm parameters or optimize the algorithm implementation as needed to improve performance.

[0129] As an implementation, in step S420, the first noise removal of the noisy neural signal is performed based on the interpolation operator and the guide array, to obtain a noise-removed signal G s-1 , including: Step S421: Obtain the center array and dispersion array of the noisy neural signal based on the interpolation operator, and randomly extract in the center array and dispersion array of the noisy neural signal to obtain an iteration center array and an iteration dispersion array; Step S422: Obtain a random array of noise distribution, adjust the iteration center array and the iteration dispersion array according to the guide array and the random array, and obtain a noise-removed signal G s-1 .

[0130] In step S421, the computer system analyzes the noisy neural signal using the interpolation operator to obtain the center array (which can represent the main trend or mean of the signal) and the dispersion array (which describes the fluctuations or variance of the signal relative to the center). Then, the system randomly extracts (or samples, perturbs) based on these arrays to generate an iteration center array and an iteration dispersion array for iterative processing.

[0131] For example, assume that the noisy neural signal has been converted to a domain suitable for processing by the interpolation operator (e.g., converted to the frequency domain by Fourier transform). In this domain, the interpolation operator can identify the main frequency components of the signal (corresponding to the center array) and the fluctuations around these components (corresponding to the dispersion array). Then, the system generates an iteration center array and an iteration dispersion array by randomly extracting (e.g., slightly adjusting or perturbing certain frequency components in the frequency domain) on these components. Assuming that the interpolation operator is actually a parameterized function or model (such as a neural network layer, a Gaussian process, etc.), it generates the center array and the dispersion array by learning the statistical characteristics of the noisy neural signal. These arrays define a probability distribution region of the signal in the latent space, and random extraction is a sampling within this region to generate a new signal representation (i.e., the iteration center array and the iteration dispersion array).

[0132] For example, assume that the noisy neural signal has been mapped into a high-dimensional latent space, and forms a distribution centered at a certain point with a certain scatter in this space. The interpolation operator generates a center array and a scatter array by analyzing the geometric properties of this distribution in the latent space (e.g., mean vector, covariance matrix, etc.). Then, the system randomly draws within the region of this distribution (e.g., by sampling from a Gaussian distribution) to generate an iteration center array and an iteration scatter array. These arrays will then be used to generate the noise-removed signal G s-1 .

[0133] After obtaining the iteration center array and the iteration scatter array, step S422 further adjusts these iteration arrays in combination with a random array of the noise distribution and a guidance array extracted from the user's visualization requirements to generate the noise-removed signal G s-1 .

[0134] The random array of the noise distribution is an array containing random noise samples that reflect the statistical properties of the noise component in the noisy neural signal. This array can be generated in various ways, such as sampling from a pre-defined noise model (e.g., Gaussian noise, salt and pepper noise, etc.). In practical applications, the random array of the noise distribution can be closely related to the generation process of the noisy neural signal, i.e., it can be generated based on the noise model and parameters used when adding noise in step S200.

[0135] The guidance array is an implicit representation array extracted from the user-defined visualization constraint parameters, which contains the user's desired visualization feature information (e.g., color mapping scheme, dynamic effect setting, etc.). In the noise removal process, the guidance array serves as a global reference framework, guiding the adjustment direction of the iteration center array and the iteration scatter array. By continuously aligning the iteration arrays with the guidance array and optimizing the matching degree between them (possibly through a certain form of loss function), the system can gradually generate a noise-removed signal that meets the user's expectations.

[0136] In step S422, the computer system obtains the random array of the noise distribution and the guidance array. Then, it combines these arrays with the iteration center array and the iteration scatter array for comprehensive adjustment. This adjustment process may involve various operations such as weighted summation, nonlinear transformation, optimization algorithm iteration, etc., depending on the algorithm design and application scenario requirements.

[0137] One way is to preliminarily adjust the iteration center array according to the guidance array to ensure that it generally meets the user's desired visualization features; then fine-tune the iteration scatter array according to the random array of the noise distribution to introduce certain randomness and uncertainty; finally, combine the adjusted iteration center array and iteration scatter array to generate the noise-removed signal Gs-1 .

[0138] In step S422, the computer system first adjusts the iteration center array according to the guidance array. For example, if a certain element value in the guidance array is larger than the corresponding element value in the iteration center array, the system can increase the value of the element in the iteration center array to make it closer to the value in the guidance array.

[0139] Next, the system fine-tunes the iteration dispersion array using the random array of the noise distribution. This fine-tuning process may involve combining elements in the random array with elements in the iteration dispersion array in some way (such as weighted sum) to introduce randomness. However, since it is desired to remove noise while preserving the naturalness and diversity of the signal, this fine-tuning should be moderate and should not destroy the main features of the signal.

[0140] Finally, the system combines the adjusted iteration center array and the iteration dispersion array to generate the noise-removed signal Gs-1. This signal should be closer to the user's desired position represented by the guidance array in the latent space than the original noisy neural signal, while also retaining some randomness and naturalness to avoid over-smoothing or distortion.

[0141] As an implementation, the noise-cleaning transition signal is generated in a latent space mapping branch component in the target machine learning algorithm, which includes a feature representation operator and an interpolation operator. Based on this, in step S400, the noise-cleaning transition signal is restored into a visual neural signal conforming to the target visual feature, including: Step S430: based on the feature representation operator, feature extraction is performed on the noise-cleaning transition signal to obtain a center array and a dispersion array of the noise-cleaning transition signal, random sampling is performed in the center array to obtain a latent space center array, and random sampling is performed in the dispersion array to obtain a latent space dispersion array; Step S440: based on the interpolation operator, the latent space center array, and the latent space dispersion array, a visual neural signal conforming to the target visual feature is generated.

[0142] In step S430, the computer system uses the feature representation operator to deeply analyze the noise-cleaning transition signal to extract its key features and map these features into the latent space. Through feature extraction and latent space parameter extraction, the system can more accurately control the generation process of the signal, thereby ensuring that the finally generated visual neural signal not only conforms to the target visual feature, but also retains important information of the original signal.

[0143] In step S430, the feature representation operator extracts a center array and a dispersion array from the noise-cleansed transition signal. These two arrays represent the dominant trend (or mean) and the fluctuations around the dominant trend (or variance) of the signal, respectively, which together characterize the distribution of the signal in the latent space.

[0144] After extracting the center array and the dispersion array, the system further performs random draws (or sampling, perturbation) on these arrays to generate a latent space center array and a latent space dispersion array. This process increases the diversity and randomness of the signal, which helps to avoid overfitting and improve the generalization ability of the model.

[0145] For example, suppose the noise-cleansed transition signal has been mapped into a high-dimensional latent space and formed a distribution centered at a certain point with a certain dispersion in this space. The feature representation operator generates a center array and a dispersion array by analyzing the geometric characteristics (such as mean, variance, shape, etc.) of this distribution in the latent space. These arrays represent the position and directionality of the signal in the latent space in the form of numerical vectors.

[0146] To generate the visual neural signal that conforms to the target visual feature, random draws are performed on these arrays. Specifically, the system can fine-tune the elements of the center array within a small range (such as adding a small random perturbation) to explore the subtle changes of the signal features in the latent space; at the same time, the system can also perform scaling or rotation operations on the elements of the dispersion array to adjust the dispersion or direction of the signal in the latent space. These operations together constitute the extraction process of the latent space parameters.

[0147] After obtaining the latent space center array and the latent space dispersion array, step S440 uses an interpolation operator to generate a visual neural signal that conforms to the target visual feature according to these parameters. The interpolation operator maps the parameters in the latent space back to the original signal space (or the user-defined visualization space), thereby generating a signal representation that can be visually displayed.

[0148] Under the action of the interpolation operator, the latent space center array and the latent space dispersion array are mapped back to the original signal space (or the visualization space) to generate a preliminary representation of the visual neural signal. However, this preliminary representation may still need to go through a series of post-processing steps (such as smoothing, normalization, color mapping, etc.) before it can be finally presented to the user. These post-processing steps aim to further optimize the visual effect of the signal and improve its interpretability.

[0149] To meet the target visualization features, the system will fully consider the user's definition and expectations in the process of generating the visualized neural signal. For example, if the user defines a specific color mapping scheme to distinguish different intensity signal regions, the system will apply this mapping scheme when generating the signal to assign the corresponding color value to each time point or spatial point. Similarly, if the user expects to see dynamic changes in the visualization of the signal (such as flickering, gradient, etc.), the system will also introduce the corresponding dynamic effect parameters when generating the signal to achieve this expectation.

[0150] For example, continuing the previous example, it is assumed that the latent space center array and the latent space dispersion array have been obtained. Now the system uses an interpolation operator to map these parameters back to the original signal space (or visualization space) to generate the visualized neural signal. Specifically, the interpolation operator can determine the main position of the signal on the time axis or spatial axis according to the value of each element in the latent space center array; at the same time, according to the element value in the latent space dispersion array, adjust the fluctuation range or shape change of the signal at these positions.

[0151] If the user expects to highlight the areas with higher signal intensity in the visualization, the system can apply an enhancement filter to amplify the signal values in these areas; if the user defines a specific color mapping scheme, the system will assign the corresponding color value to each time point or spatial point according to the size of the signal value to form an intuitive color-coded map; if the user expects to see dynamic changes in the visualization of the signal, the system can introduce a change parameter in the time dimension (such as periodic change, random jitter, etc.) when generating the signal to achieve the display of dynamic effects.

[0152] Finally, through the mapping of the interpolation operator and the optimization of the post-processing step, the system successfully generates a visualized neural signal that meets the target visualization features. This signal not only removes most of the noise components in the original noisy neural signal, but also retains the main features of the signal and the user-defined visualization constraints to provide medical professionals with intuitive and clear neural activity monitoring results.

[0153] The present application also provides a training process for the target machine learning algorithm, which can specifically include the following steps: Step S10: Obtain a learning example library; the learning example library includes original living neural signal samples, visualized neural signal samples, and visualized neural signal parameters; the visualized neural signal parameters are used to describe the visualization features of the visualization display results; the number of sampling points of the original living neural signal samples and the visualized neural signal samples is consistent.

[0154] The learning example library is a collection of samples used to train the target machine learning algorithm. The raw live neural signal examples are directly derived from real-time neural signal data collected by implantable medical chips. Each example is a series of voltage values over time, representing the activity of neurons during a specific time period. For example, a raw live neural signal example might be a one-dimensional array containing 1000 sample points, with a sampling rate of 1000 Hz, covering 1 second of neural activity. Visualized neural signal examples are the results of raw signals after some visualization processing. They have been converted according to user's visualization needs (such as color mapping, dynamic effects, etc.). Each visualized neural signal example not only contains the processed signal values, but also may contain metadata related to visualization (such as color coding scheme, view parameters, etc.). Importantly, these examples strictly correspond to the original signal examples in time, i.e., they have the same number of sample points, ensuring data consistency and comparability.

[0155] Visualized neural signal parameters are used to describe specific features of the visualization display results, such as color mapping rules, dynamic effect parameters, view scaling ratios, etc. They exist in the form of structured data (such as JSON objects, XML files, etc.), which is convenient for computer systems to parse and process. For example, a visualized neural signal parameter may specify a color mapping rule that "when the signal intensity is greater than the threshold, display red, and when it is less than the threshold, display blue".

[0156] The computer system obtains these learning examples through various means. On the one hand, it can extract historical data from existing databases as examples; on the other hand, it can also cooperate with medical professionals to generate new examples through actual collection and annotation. During the acquisition process, the system will strictly screen and preprocess the examples to ensure their quality meets the training requirements.

[0157] Before training the target machine learning algorithm, it is crucial to ensure the diversity and representativeness of the learning example library. This means that the example library should contain data from different patients, different time periods, and different physiological states, so that the algorithm can learn more extensive and general feature representations and mapping rules. At the same time, accurate labeling and classification of examples are also essential steps, which will directly affect the training effect and generalization ability of the algorithm.

[0158] Step S20: Abstract the signal features of the raw live neural signal examples to obtain abstracted signal examples, and load the raw live neural signal examples and the abstracted signal examples to the initial machine learning algorithm.

[0159] In step S20, the computer system conducts in-depth analysis on the original living body neural signal samples in the learning sample library to extract and abstract the key features in the signals. These features can include waveform features, frequency components, energy distribution, etc. of the signals, which collectively describe the inherent laws and characteristics of neural activity.

[0160] The computer system can first preprocess the original living body neural signal samples to remove noise, standardize signal amplitude, adjust sampling rate, etc. This step helps to reduce interference factors in the subsequent feature extraction process and improve the accuracy and stability of the features. Then, the system uses feature extraction algorithms (such as time-frequency analysis, wavelet transform, principal component analysis, etc.) to extract key features from the preprocessed signals. These features can be statistical quantities (such as mean, variance, skewness, kurtosis, etc.), waveform parameters (such as peak value, half-width, rise time, etc.), frequency domain features (such as spectral energy distribution, dominant frequency, etc.), or time-frequency features (such as instantaneous frequency, instantaneous energy, etc.) of the signals.

[0161] After extracting a large number of original features, the system needs to further abstract these features (refer to the process of step S100 described above).

[0162] Step S30: Feature embedding is performed on the original living body neural signal samples to obtain first signal feature representation samples, and feature embedding is performed on the abstracted signal samples to obtain second signal feature representation samples; the number of sampling points of the original living body neural signal samples is greater than the number of sampling points of the first signal feature representation samples and the second signal feature representation samples.

[0163] In step S30, the computer system uses feature embedding technology to convert the original living body neural signal samples and the abstracted signal samples into lower-dimensional feature representations, i.e., the first signal feature representation samples and the second signal feature representation samples. This process can refer to the description of step S100 described above.

[0164] For the original living body neural signal samples, the computer system first inputs them into a pre-trained feature embedding model, which is, for example, a deep neural network (such as an autoencoder, a convolutional neural network, etc.), to map the high-dimensional original signals into a low-dimensional feature space by learning the correlation between the internal structure of the signals and the features. Specifically, each layer in the model transforms the input signal to extract features at different levels, and finally outputs a low-dimensional feature vector as the first signal feature representation sample.

[0165] For the abstracted signal samples, the feature embedding process can be relatively simple since they are already an abstraction and simplification of the original signals. The computer system can directly input the abstracted signal samples into the same feature embedding model (or a model specifically designed for abstracted signals) for further feature extraction and dimensionality reduction. Since the abstracted signal samples already contain the main feature information of the original signals, this step is mainly to further refine and compress these features.

[0166] For the abstracted signal samples (assuming a 10-dimensional feature vector), the computer system can choose to use a lighter feature embedding model (such as a single-layer fully connected network) for further processing. This model will fine-tune the abstracted signal samples to generate a second signal feature representation sample with a lower dimension (e.g., 5-dimensional). This feature representation not only contains the main feature information of the abstracted signal, but also further reduces the redundancy of the data and the computational complexity.

[0167] Through the feature embedding process, the number of sampling points of the original living neural signal samples is reduced from the original 1000 to 10 (or less) in the first signal feature representation sample, and the dimension of the abstracted signal samples is also reduced from the original possibly higher dimension to 5 (or less) in the second signal feature representation sample. This reduction in the number of sampling points not only reduces the storage and transmission cost of the data, but also improves the efficiency of subsequent processing and analysis.

[0168] Step S40: Add noise to the second signal feature representation sample to obtain a noise-added neural signal sample.

[0169] Step S50: Map the visualized neural signal parameter to a constraint implicit representation array, map the second signal feature representation sample to a signal implicit representation array, remove noise from the noise-added neural signal sample according to the constraint implicit representation array and the signal implicit representation array, obtain a noise-cleaned transitional signal sample, and restore the noise-cleaned transitional signal sample to a predicted visualized neural signal that conforms to the target visualized feature.

[0170] The processes of steps S40 and S50 can refer to the descriptions of the aforementioned steps S200-S400, and the principles are consistent, which will not be described here.

[0171] Step S60: Determine the algorithm debugging error based on the predicted visualized neural signal and the learning sample library, and optimize the algorithm parameters of the initial machine learning algorithm based on the algorithm debugging error until the initial machine learning algorithm meets the preset debugging stop requirement, to obtain a target machine learning algorithm; the target machine learning algorithm is used to map the original living neural signal to a visualized neural signal that conforms to the target visualized feature.

[0172] In step S60, the computer system first evaluates the similarity between the predicted visualized neural signal and the true visualized neural signal in the learning example library, quantifying this difference by calculating an algorithm tuning error. Then, based on this error signal, the system adjusts the parameters of the initial machine learning algorithm in order to reduce the error in subsequent iterations and improve the prediction accuracy of the algorithm. This process will continue until the algorithm performance meets the preset tuning stop requirement, thus obtaining the final target machine learning algorithm.

[0173] The algorithm tuning error is an important indicator for measuring the difference between the algorithm prediction result and the true result in the learning example library. This error can involve multiple dimensions, including but not limited to the similarity of signal shape, the accuracy of color mapping, the consistency of dynamic effect, etc. In order to comprehensively evaluate these differences, the computer system can use a composite loss function that combines multiple error measurement standards.

[0174] For example, suppose the mean square error (MSE) is used to measure the prediction accuracy of signal intensity, while the cross-entropy loss is used to evaluate the correctness of color mapping. The composite loss function can be represented as: L=α·MSE(y pred ,y true )+β·Cross Entropy(p pred ,p true ); Where y pred and y true represent the predicted signal intensity and the true signal intensity, respectively, p pred and p true represent the predicted color distribution and the true color distribution, respectively, and α and β are hyperparameters that balance the weights of different loss terms.

[0175] For example, suppose the learning example library contains the original live neural signal of a specific patient and its corresponding visualized neural signal. During the training process, the computer system generates a predicted visualized neural signal and calculates the MSE and Cross Entropy loss between it and the true visualized neural signal. For example, the MSE may calculate the average deviation of signal intensity, while the Cross Entropy evaluates the accuracy of color mapping (assuming that the visualized signal uses color coding to represent signal intensity).

[0176] After calculating the algorithm tuning error, the computer system optimizes the parameters of the initial machine learning algorithm based on this error signal. This process can be achieved through optimization algorithms such as gradient descent (or its variants such as Adam, RMS prop, etc.), which can automatically adjust algorithm parameters according to error gradients to minimize the error function.

[0177] In the optimization process, the system first calculates the gradient of the error function with respect to the algorithm parameters, i.e. the partial derivative of the error function with respect to each parameter. Then, the system updates the parameter values based on these gradient information, which can be achieved by subtracting a step proportional to the gradient from the current parameter value. This step (learning rate) is an important hyperparameter that needs to be carefully adjusted to balance the convergence speed and stability.

[0178] For example, suppose the system employs the Adam optimization algorithm to adjust the weights and biases in the neural network. In each iteration, the system first forward propagates the predicted visualized neural signals, then calculates the MSE and Cross Entropy loss, and backward propagates these losses to calculate the gradient of each parameter in the network. Then, the system updates the network parameters, including weights and biases, according to the rules of the Adam algorithm.

[0179] The optimization process will continue until the preset debugging stop requirements are met. These requirements may include reaching the maximum number of iterations, reducing the error below a certain threshold, no longer significantly improving performance on the validation set, etc. In the context of regulation monitoring information visualization based on implantable medical chips, the debugging stop requirements can be particularly considered for the special nature of medical applications, such as ensuring the stability and reliability of the algorithm, avoiding overfitting to improve generalization ability, etc.

[0180] For example, during the training of the system, a maximum number of iterations (such as 100 epochs) and an early stopping criterion on the validation set (such as stopping training if the validation loss does not decrease for 10 consecutive epochs) can be set. When the training reaches these stop requirements, the system considers that the algorithm has converged to a sufficiently good solution, and saves the algorithm parameters at this time as the final target machine learning algorithm model.

[0181] As an implementation, in step S50, the visualized neural signal parameter is mapped to a constraint implicit representation array, the second signal feature representation sample is mapped to a signal implicit representation array, and the noise removal is performed on the noise signal sample according to the constraint implicit representation array and the signal implicit representation array, to obtain a noise-cleared transition signal sample. The noise-cleared transition signal sample is restored to a predicted visualized neural signal conforming to the target visualized feature, including: Step S51: dividing the visualized neural signal parameter in the learning sample library into u visualized labels, determining v visualized difference parameters according to the corresponding uniform visual elements of the u visualized labels, 1≤u≤v; Step S52: mapping the v visual difference parameters into v constraint implicit representation arrays, mapping the second signal feature representation example into a signal implicit representation array, de-noising the noisy neural signal example according to the v constraint implicit representation arrays and the signal implicit representation array, obtaining v diversified neural signals, calculating the parameter mean value of the v diversified neural signals, obtaining a de-noised transitional signal example, and restoring the de-noised transitional signal example into a predicted visual neural signal conforming to the target visual feature.

[0182] In step S51, the computer system finely divides the visual neural signal parameters in the learning example library, and determines the visual difference parameters based on the divisions, so as to be mapped into constraint implicit representation arrays subsequently.

[0183] The visual neural signal parameters can contain various expectations of the user on the visual result, such as color mapping schemes, dynamic effect settings, view angle selections, etc. These parameters can exist in various forms, such as color coding tables, dynamic effect parameter tables, view angle conversion matrices, etc. The computer system first needs to uniformly process these parameters, and divide them into multiple visual tags. The visual tag is a high-level abstraction of the visual neural signal parameter, and each tag represents a class of visual characteristics. For example, the color mapping scheme can be a visual tag, which contains the mapping relationship between the signal intensity and the color; the dynamic effect setting can be another visual tag, which describes the display effect (such as flickering, gradient, etc.) of the signal over time.

[0184] After dividing the visual tags, the computer system further analyzes the specific visual elements represented by these tags and their differences. The uniform visual element refers to the common or similar part in different visual tags, such as the basic color in the color wheel, the time period in the dynamic effect, etc. Through the analysis of these uniform visual elements, the system can identify the differences between different visual tags, i.e., the visual difference parameters.

[0185] The visual difference parameters are used to quantify the specific differences between different visual tags, and they are crucial for subsequent generation of diversified visual signals. For example, in the color mapping scheme, the visual difference parameters can include the difference of the RGB values corresponding to different color intervals; in the dynamic effect setting, the visual difference parameters can include the variation range of the parameters such as the flickering frequency and the gradient speed.

[0186] For example, suppose the learning example library contains two color mapping schemes as visualized neural signal parameters. The first scheme (labeled as Tagl) maps the part of signal intensity in the range of 0-50 to blue and the part in the range of 50-100 to red; the second scheme (labeled as Tag2) maps the part of 0-50 to green and the part of 50-100 to yellow. In this example, “color mapping” is a unified visual element, while “blue and red” and “green and yellow” are difference parameters between the two color mapping schemes.

[0187] In step S52, the computer system maps the visualized difference parameters determined in step S51 into constraint implicit representation arrays and maps the second signal feature representation example into a signal implicit representation array. Then, based on these implicit representation arrays, the computer system performs noise removal processing on the noisy neural signal example, generates a diversified noise-cleaned transitional signal example, and finally restores it into a predicted visualized neural signal that conforms to the target visualized feature.

[0188] Constraint implicit representation arrays are numerical expressions of visualized difference parameters, which exist in the form of vectors, matrices, or other data structures, and are used to pass users’ visualized requirements within algorithms. The computer system generates corresponding constraint implicit representation arrays according to the specific values of each visualized difference parameter. These arrays may contain information of color coding, dynamic effect parameters, view transformation matrices, and other elements.

[0189] Signal implicit representation arrays are another numerical expression of the second signal feature representation example, which retains the main feature information of the original signal but reduces the dimension and complexity of the data. This array can be extracted from the original live neural signal example through steps such as feature embedding and dimension reduction processing.

[0190] After obtaining the constraint implicit representation arrays and the signal implicit representation array, the computer system uses these arrays as guidance information to perform noise removal processing on the noisy neural signal example. This processing process may involve parallel running or alternating optimization strategies of multiple noise removal models to ensure that the generated signal both removes noise and retains diversified features.

[0191] Specifically, the system may use a generative model based on conditional generative adversarial network (cGAN) or variational autoencoder (VAE) to generate a diversified noise-cleaned transitional signal example according to the constraint implicit representation array. These models can introduce visualized differences specified by the constraint implicit representation array while maintaining the main features of the signal, thereby achieving the diversified generation of the signal.

[0192] For the generated multiple noise-cleaning transition signal samples, the system can employ methods such as parametric mean calculation to integrate their common features and reduce individual differences. This process helps to obtain a more stable and user-desired predicted visualized neural signal.

[0193] Finally, the system restores the integrated noise-cleaning transition signal samples to the final signal representation that meets the target visualization features according to the visualization difference parameters in the constraint implicit representation array. This process may involve multiple steps such as color mapping, dynamic effect rendering, view transformation, etc., depending on the specific content of the visualization requirements.

[0194] For example, continuing the previous color mapping scheme example. Suppose the system has generated two noise-cleaning transition signal samples based on different color mapping schemes (corresponding to Tag1 and Tag2 color mapping, respectively). In the parametric mean calculation process, the system may find that the two signal samples have larger color differences in the low signal intensity part (0-50 range) but smaller color differences in the high signal intensity part (50-100 range). Therefore, the system may choose to perform weighted average processing on the color of the low signal intensity part to reduce the color difference; and for the high signal intensity part, directly use one of the color mapping schemes (such as Tag1 red) as the final representation.

[0195] After such processing, the system obtains a predicted visualized neural signal that not only removes noise but also meets the user-defined color mapping scheme. This signal can be further used for neural activity visualization, disease diagnosis auxiliary analysis, and other fields.

[0196] Step S50 successfully combines user-defined visualization requirements with signal features processed by the algorithm by a series of complex operations such as detailed division of visualization neural signal parameters, determination of visualization difference parameters, mapping of constraint implicit representation array and signal implicit representation array, and noise removal and signal restoration, etc., to generate a predicted visualized neural signal that meets the target visualization features. This process not only relies on advanced machine learning algorithms and techniques, but also requires in-depth understanding of neural signal processing, visualization theory, and user requirements, etc.

[0197] As an implementation, the learning example library further includes a parameter representation label; the parameter representation label is semantic information associated with the visualized neural signal parameter; the initial machine learning algorithm includes an initial constraint embedding branch component and an initial noise adding branch component. Based on this, in step S60, the algorithm debugging error is determined according to the predicted visualized neural signal and the learning example library, and the algorithm parameter of the initial machine learning algorithm is optimized based on the algorithm debugging error until the initial machine learning algorithm meets the preset debugging stop requirement, and the target machine learning algorithm is obtained, which can include: mapping the constraint implicit representation array to the to-be-executed semantic feature, determining the algorithm debugging error of the initial noise adding branch component and the initial constraint embedding branch component according to the predicted visualized neural signal and the action potential difference value of the visualized neural signal example, and the semantic difference value of the parameter representation label and the to-be-executed semantic feature, optimizing the parameter variable in the initial noise adding branch component and the initial constraint embedding branch component based on the algorithm debugging error and the constraint condition, until the initial machine learning algorithm meets the constraint condition, and the target machine learning algorithm is obtained.

[0198] In step S60, the computer system maps the constraint implicit representation array (which is generated by step S50 to guide the noise removal and signal restoration process) to the to-be-executed semantic feature. The to-be-executed semantic feature is an abstraction and interpretation of the constraint implicit representation array, which not only contains numerical constraint information, but also contains the semantic meaning behind these constraints. For example, suppose a color mapping scheme is defined in the visualized neural signal parameter, which divides the signal intensity into three levels and uses red, yellow and green to represent low, medium and high intensity respectively. The corresponding constraint implicit representation array may contain a set of color coding values. In the mapping process, the computer system interprets these color coding values as the to-be-executed semantic feature of "low intensity signal should be displayed in red, medium intensity signal should be displayed in yellow, and high intensity signal should be displayed in green".

[0199] Next, the computer system calculates the algorithm debugging error to evaluate the quality of the current predicted visualized neural signal. This process not only considers the action potential difference between the predicted signal and the real visualized neural signal example (i.e. the direct comparison of signal intensity), but also includes the semantic difference between the parameter representation label and the to-be-executed semantic feature. The calculation of semantic difference is an important indicator to measure the degree of algorithm understanding of user visualization needs.

[0200] For example, when comparing the predicted signal with the real signal, the computer system calculates the action potential difference at each time point, obtaining a series of error values. Then, it further analyzes whether these error values are inconsistent with the semantic information in the parameter representation label. For example, if the real signal should be displayed in green in a certain high intensity area, but the predicted signal is incorrectly displayed in yellow, then a larger semantic difference will be generated.

[0201] The algorithm debugging error can be a composite loss function that combines the quantification error of the action potential difference and the qualitative assessment of the semantic difference. A simplified loss function example is as follows: ; where N is the number of time points of the signal, and are the action potential values of the predicted signal and the real signal at the i-th time point, and are the color mapping schemes (or other visualization features) of the predicted signal and the real signal, λ1 and λ2 are weight coefficients that balance different loss terms, and the SD (Semantic Difference) function is used to calculate the semantic difference.

[0202] After obtaining the algorithm debugging error, the computer system will optimize the parameters of the initial machine learning algorithm based on this error signal. The goal of optimization is to reduce the error, so that the predicted visual neural signal is closer to the real signal, and better meets the user's visualization needs.

[0203] Since the initial machine learning algorithm contains an initial constraint embedding branch component and an initial noise adding branch component, the optimization process will be carried out for these two components respectively. For the constraint embedding branch component, optimization may involve adjusting the weights of feature embedding operators, changing the parameters of hidden space mapping, etc.; for the noise adding branch component, optimization may focus on adjusting the parameters of noise distribution, optimizing the noise adding strategy, etc.

[0204] For example, during the optimization process, the computer system can use gradient descent (or its variants such as Adam optimizer) to adjust the algorithm parameters. For the constraint embedding branch component, it can fine-tune the weight distribution in the feature embedding process according to the feedback of the semantic difference, to ensure that the semantic features to be executed can be accurately mapped to the hidden space; for the noise adding branch component, it can adjust the noise intensity and noise type according to the feedback of the action potential difference, to reduce the impact of noise on signal quality.

[0205] Finally, the computer system regularly checks whether the preset debugging stopping conditions are met. These conditions may include reaching the maximum number of iterations, the algorithm debugging error falling below the predetermined threshold, the performance on the validation set no longer significantly improving, etc. Only when all conditions are met, the training process will be terminated, and the algorithm parameters at this time are considered as the optimal solution, and are used to generate the target machine learning algorithm.

[0206] Through the implementation of the step S60, the computer system comprehensively considers the quantization error between the predicted signal and the real signal and the semantic understanding degree of the user's visualization demand, and finely optimizes the variable of the initial machine learning algorithm. This process not only improves the performance and stability of the algorithm, but also ensures that the finally generated target machine learning algorithm can accurately map the original living body neural signal to the visualized neural signal meeting the user's expectation Figure 2 A hardware entity schematic diagram of a computer system provided for an embodiment of the present application is shown in FIG. 10. The hardware entity of the computer system 1000 includes a processor 1001 and a memory 1002, wherein the memory 1002 stores a computer program executable on the processor 1001, and the processor 1001 implements the steps in the method of any of the above embodiments when executing the program. Figure 2

[0207] The above is only an embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed in the present application, which should be covered within the protection scope of the present application.​

Claims

1. A method for visualizing monitoring information based on an implanted medical chip, characterized in that, The method comprises the following steps: obtaining an original living nerve signal, abstracting signal features of the original living nerve signal to obtain an abstracted signal, embedding features of the original living nerve signal to obtain a first signal feature representation, and embedding features of the abstracted signal to obtain a second signal feature representation; the number of sampling points of the original living nerve signal is greater than the number of sampling points of the first signal feature representation and the second signal feature representation; adding noise to the first signal feature representation to obtain a noise-added nerve signal; obtaining a visual constraint parameter, mapping the visual constraint parameter to a constraint implicit representation array, mapping the second signal feature representation to a signal implicit representation array, and using the visual constraint parameter to describe the visual features of a visual display result; removing noise from the noise-added nerve signal according to the constraint implicit representation array and the signal implicit representation array to obtain a noise-cleaned transition signal, and restoring the noise-cleaned transition signal to a visual nerve signal that conforms to a target visual feature; the number of sampling points of the original living nerve signal is consistent with the number of sampling points of the visual nerve signal.

2. The method of claim 1, wherein, The method comprises the following steps: obtaining waveform feature values corresponding to x action potentials in the original living nerve signal, performing smoothing operation on the original living nerve signal according to the waveform feature values to obtain a smoothed signal; x≥1; obtaining time axis change rates and amplitude axis change rates corresponding to the x action potentials in the smoothed signal, determining intensities and angles corresponding to the x action potentials in the smoothed signal according to the time axis change rates and the amplitude axis change rates, and performing local extreme value detection on the smoothed signal based on the intensities and the angles to obtain an extreme value nerve signal; generating an abstracted signal according to the time axis change rates and the amplitude axis change rates corresponding to the x action potentials in the extreme value nerve signal.

3. The method of claim 2, wherein, The method comprises the following steps: obtaining a first reference value and a second reference value, determining change rates corresponding to the x action potentials in the extreme value nerve signal according to the time axis change rates and the amplitude axis change rates corresponding to the x action potentials in the extreme value nerve signal; determining a significant action potential in the x action potentials of the extreme value nerve signal, wherein the change rate of the significant action potential is greater than or equal to the first reference value; determining a fuzzy action potential in the x action potentials of the extreme value nerve signal, wherein the change rate of the fuzzy action potential is greater than or equal to the second reference value and less than the first reference value; determining a candidate action potential adjacent to the significant action potential, and generating an abstracted signal according to the significant action potential and the candidate action potential.

4. The method of claim 1, wherein, The method comprises the following steps: loading the original living body nerve signal to a target machine learning algorithm; the target machine learning algorithm comprises a hidden space mapping branch component, and the hidden space mapping branch component comprises a feature representation operator; performing feature extraction on the original living body nerve signal based on the feature representation operator to obtain a center array and a dispersion array of the original living body nerve signal; randomly extracting in the center array and the dispersion array of the original living body nerve signal to obtain a hidden space center array and a hidden space dispersion array, and generating a first signal feature representation according to the hidden space center array and the hidden space dispersion array; the first signal feature representation is generated in a hidden space mapping branch component in the target machine learning algorithm, and the hidden space mapping branch component comprises a normalization operator and a pooling operator; the noise addition on the first signal feature representation to obtain a noise-added nerve signal comprises: obtaining a noise distribution and an iteration number s, performing a normalization operation on a waveform feature value of an action potential in the first signal feature representation based on the normalization operator and the noise distribution; s≥1; performing s times of noise addition on the normalized first signal feature representation based on the pooling operator and the iteration number s to obtain the noise-added nerve signal, wherein the noise distribution represents confidence distribution information of noise addition, and the iteration number s represents a degree of noise addition to the first signal feature representation according to the noise distribution.

5. The method of claim 1, wherein, the mapping of the visual constraint parameter into a constraint implicit representation array comprises: loading the visual constraint parameter to a target machine learning algorithm, wherein the target machine learning algorithm comprises a constraint embedding branch component; the constraint embedding branch component comprises a feature embedding operator, a weight distribution operator and a fusion operator; performing feature embedding on the visual constraint parameter based on the feature embedding operator to obtain y constraint array representations; y≥1; generating focusing weights corresponding to the y constraint array representations respectively based on the weight distribution operator, and performing focusing adjustment on the constraint array representations according to the focusing weights to obtain y to-be-used focusing arrays; performing weight fusion on the y to-be-used focusing arrays based on the fusion operator to obtain a constraint implicit representation array; the noise-added nerve signal is obtained based on s times of noise addition on the first signal feature representation, the first signal feature representation is obtained in a hidden space mapping branch component in the target machine learning algorithm, and the hidden space mapping branch component comprises a residual operator and an interpolation operator; the noise removal on the noise-added nerve signal according to the constraint implicit representation array and the signal implicit representation array to obtain a noise-cleaned transition signal comprises: combining the constraint implicit representation array and the signal implicit representation array into a guide array based on the residual operator; performing first noise removal on the noise-added neural signal based on the interpolation operator and the guide array to obtain a noise-removed signal G s-1 repeating the noise removal on the noise-removed signal G s-1 until s times of noise removal are performed, and determining a noise-removed signal G0 obtained by s times of noise removal as a noise-cleaned transition signal.

6. The method of claim 5, wherein, performing a first noise removal on the noisy neural signal based on the interpolation operator and the guide array to obtain a noise removed signal G s-1 comprising: obtaining a center array and a dispersion array of the noise-added nerve signal based on the interpolation operator, and randomly extracting in the center array and the dispersion array of the noise-added nerve signal to obtain an iteration center array and an iteration dispersion array. obtaining a random array of noise distribution, adjusting the iteration center array and the iteration dispersion array according to the guide array and the random array, and obtaining a noise-removed signal G s-1 .

7. The method of claim 1, wherein, The noise-cleaning transition signal is generated in a hidden space mapping branch component in a target machine learning algorithm, and the hidden space mapping branch component includes a feature representation operator and an interpolation operator; and the restoring of the noise-cleaning transition signal into the visual neural signal conforming to the target visual feature includes: performing feature extraction on the noise-cleaning transition signal based on the feature representation operator to obtain a center array and a dispersion array of the noise-cleaning transition signal, randomly extracting in the center array to obtain a hidden space center array, and randomly extracting in the dispersion array to obtain a hidden space dispersion array; generating the visual neural signal conforming to the target visual feature based on the interpolation operator, the hidden space center array, and the hidden space dispersion array.

8. The method according to any one of claims 4 to 7, characterized in that, The method further includes: obtaining a learning sample library; the learning sample library includes original living body neural signal samples, visual neural signal samples, and visual neural signal parameters; the visual neural signal parameters are used to describe visual features of visual display results; the original living body neural signal samples and the visual neural signal samples have consistent sampling point numbers; performing signal feature abstraction on the original living body neural signal samples to obtain abstracted signal samples, and loading the original living body neural signal samples and the abstracted signal samples into an initial machine learning algorithm; performing feature embedding on the original living body neural signal samples to obtain first signal feature representation samples, and performing feature embedding on the abstracted signal samples to obtain second signal feature representation samples; the original living body neural signal samples have a sampling point number greater than sampling point numbers of the first signal feature representation samples and the second signal feature representation samples; adding noise to the second signal feature representation samples to obtain noise-added neural signal samples; mapping the visual neural signal parameters into a constraint hidden representation array, mapping the second signal feature representation samples into a signal hidden representation array, removing noise from the noise-added neural signal samples according to the constraint hidden representation array and the signal hidden representation array to obtain a noise-cleaning transition signal sample, and restoring the noise-cleaning transition signal sample into a predicted visual neural signal conforming to a target visual feature; determining an algorithm debugging error according to the predicted visual neural signal and the learning sample library, optimizing algorithm parameters of the initial machine learning algorithm based on the algorithm debugging error, and obtaining a target machine learning algorithm when the initial machine learning algorithm meets a preset debugging stop requirement; the target machine learning algorithm is used to map original living body neural signals into visual neural signals conforming to a target visual feature.

9. The method of claim 8, wherein, The visual neural signal parameter is mapped into a constraint implicit representation array, the second signal feature representation sample is mapped into a signal implicit representation array, noise removal is performed on the noisy neural signal sample according to the constraint implicit representation array and the signal implicit representation array, a noise-cleared transitional signal sample is obtained, and the noise-cleared transitional signal sample is restored into a predicted visual neural signal conforming to the target visual feature, including: The visual neural signal parameters in the learning sample library are divided into u visual markers, v visual difference parameters are determined according to the uniform visual elements corresponding to the u visual markers, 1≤u≤v; The v visual difference parameters are mapped into v constraint implicit representation arrays, the second signal feature representation sample is mapped into a signal implicit representation array, noise removal is performed on the noisy neural signal sample according to the v constraint implicit representation arrays and the signal implicit representation array, v diverse neural signals are obtained, parameter mean calculation is performed on the v diverse neural signals, a noise-cleared transitional signal sample is obtained, and the noise-cleared transitional signal sample is restored into a predicted visual neural signal conforming to the target visual feature; The learning sample library further includes a parameter representation marker; the parameter representation marker is semantic information associated with the visual neural signal parameter; the initial machine learning algorithm includes an initial constraint embedding branch component and an initial noise adding branch component, the algorithm debugging error is determined according to the predicted visual neural signal and the learning sample library, the algorithm parameter of the initial machine learning algorithm is optimized based on the algorithm debugging error, until the initial machine learning algorithm meets the preset debugging stop requirement, a target machine learning algorithm is obtained, including: The constraint implicit representation array is mapped into a to-be-executed semantic feature, the algorithm debugging error of the initial noise adding branch component and the initial constraint embedding branch component is determined according to the predicted visual neural signal, the difference value of the action potential of the visual neural signal sample, and the semantic difference value between the parameter representation marker and the to-be-executed semantic feature, the parameter variable in the initial noise adding branch component and the initial constraint embedding branch component is optimized based on the algorithm debugging error and the constraint condition, until the initial machine learning algorithm meets the constraint condition, a target machine learning algorithm is obtained.

10. A computer system comprising a memory and a processor, said memory storing a computer program operable on the processor, characterised in that, The processor implements the steps in the method of any one of claims 1 to 9 when executing the program.

Images