Concha region vagus nerve stimulation device and pattern recognition method for output waveform thereof

Abstract

The present application relates to the technical field of pattern recognition, and particularly relates to a concha vagus nerve stimulation device and a pattern recognition method for output waveforms of the device, and specifically as follows: real-time collection of current waveform data output by the concha vagus nerve stimulation device; artificial labeling of original current waveform data; generation of reference current by the concha vagus nerve stimulation device according to hardware parameter configuration; generation of inertia component base matrix by a simulator guided by physical characteristics on the original current waveform data; construction of a pattern recognition model, input of original current waveform data in the data set into the pattern recognition model, pattern recognition prediction; alternating training of neural networks in the model; correction of classification results of the model based on the reference current, and finally generation of corrected classification probability. The present application can effectively distinguish inherent patterns of waveforms and distortions caused by inertia, can avoid loss of important effective features in the recognition process, and can further improve the accuracy of the recognition result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pattern recognition, in particular to a concha vagus nerve stimulation device and a pattern recognition method for output waveforms thereof. BACKGROUND

[0002] In the field of neuromodulation therapy and other medical fields, the use of stimulation devices can alleviate emotional tension and physical discomfort and other adverse reactions after certain medical operations, so accurate identification of the waveform pattern output by the stimulation device is of vital importance for optimizing stimulation parameters, improving treatment effects, and ensuring the reliability of clinical decision-making. However, existing waveform pattern recognition techniques have many limitations that are difficult to overcome when faced with these needs, and cannot meet the urgent need for high-precision identification in clinical applications; the waveform data output by the concha vagus nerve stimulation device has obvious non-stationary characteristics, which makes the waveform data extremely susceptible to the influence of biological tissue characteristics and device physical inertia, and further causes non-pathological distortion of the waveform form, such as the common overshoot or tailing phenomenon; the existence of these distortions seriously interferes with the effective features of the waveform, making the subsequent classification work face great challenges and making it difficult to guarantee classification accuracy; although existing preprocessing methods, such as band-pass filtering or simple normalization processing, can process waveform data to some extent, they cannot effectively distinguish between the inherent pattern of the waveform and the distortion caused by inertia, often losing important effective features in the processing process, thereby further reducing the accuracy of classification.

[0003] Chinese invention patent with publication number CN119316760A proposes a headset state detection method and device based on multiple sensors. The headset has a loudspeaker located in the ear canal, a first sound pickup sensor located in the ear canal and arranged near the loudspeaker, and a second sound pickup sensor located outside the ear canal. The method comprises: obtaining first headset state information according to the source audio signal input to the loudspeaker and the first audio signal picked up by the first sound pickup sensor; obtaining second headset state information according to the second audio signal picked up by the second sound pickup sensor and the first audio signal picked up by the first sound pickup sensor; and outputting a final detection result of the headset state based on the first headset state information and the second headset state information. This invention combines the feature relationship between the two types of signals, which can effectively classify the headset state in complex environments and improve the accuracy of headset state detection.

[0004] Existing technologies for processing signal data typically employ fixed thresholding or fixed gain compensation to remove inertial distortion. These methods lack adaptability, cannot accurately match distortion patterns under different waveforms, and easily disrupt effective waveform patterns, leading to reduced classification accuracy. Conventional feature fusion methods use simple concatenation or weighted averaging, which struggle to effectively capture the complex interaction between the original current waveform data and the de-inertialized features, thus limiting classification performance. They fail to fully utilize the global structural information of the original current waveform data and the advantages of the de-inertialized features, resulting in inaccurate classification results. Directly using fused features for classification is susceptible to residual inertial artifacts, and conventional classification layers such as fully connected layers... Layered softmax function calculations cannot explicitly decouple the essential waveform features from the inertial components, and ignore the multi-scale temporal dependence of features, resulting in limited classification accuracy and difficulty in meeting the requirements of high-precision classification. The probability vector output by the classifier in the existing technology does not take into account the individual differences in electrical properties of biological tissues, such as impedance distribution and dielectric constant. When the stimulus waveform passes through the concha tissue with different electrical conductance properties, the actual efficiency of neuronal activation is not monotonically related to the waveform morphology. Conventional post-processing methods, such as probability threshold truncation, cannot correct the classification bias caused by tissue characteristics, which can easily lead to misjudgment of the stimulus pattern of weakly responding individuals, affecting the effectiveness and safety of clinical applications.

[0005] Therefore, this invention proposes a vagus nerve stimulation device for the concha region and a pattern recognition method for its output waveform to solve the above problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention develops a vagus nerve stimulation device for the concha region and a method for pattern recognition of its output waveform. This invention effectively distinguishes between the inherent pattern of the waveform and the distortion caused by inertia, avoiding the loss of important effective features during the recognition process, thereby improving the accuracy of the recognition results.

[0007] On one hand, the technical solution of this invention to solve the technical problem is a method for pattern recognition of the output waveform of a vagus nerve stimulation device in the concha region, comprising the following steps:

[0008] S1. Apply electrical stimulation to the pre-set target locations in the human concha cavity and concha region using a vagus nerve stimulation device in the concha region, and capture the current waveform data output by the vagus nerve stimulation device in the concha region in real time through a high-precision bioelectric signal acquisition system.

[0009] S2. Based on the correlation between neurophysiological response characteristics and stimulation parameters, experts manually label the pattern categories of the collected raw current waveform data to form a dataset with category labels.

[0010] S3. The vagus nerve stimulation device in the concha region generates a reference current according to the hardware parameters and performs digital-to-analog conversion on the reference current to output a reference waveform.

[0011] S4, generating an inertia component base matrix by a simulator guided by physical characteristics for the original current waveform data in the data set;

[0012] S5, constructing a pattern recognition model including a dynamic inertia offset module, a residual calculation and feature enhancement module, a dual-channel feature fusion module, and a feature purification and classification module, inputting the original current waveform data in the data set into the pattern recognition model, and performing pattern recognition prediction;

[0013] S6, alternately training the coefficient prediction neural network in the dynamic inertia offset module and the backbone neural network in the dual-channel feature fusion module;

[0014] S7, correcting the classification results of the pattern recognition model based on the reference waveform of the reference current, and finally generating the corrected classification probability.

[0015] S4 is as follows:

[0016] Generating an inertia component base matrix by a simulator guided by physical characteristics for the original current waveform data in the data set, to represent the potential distortion pattern generated by the physical characteristics of the concha vagus nerve stimulation device and the biological tissue response;

[0017] Specifically, the numerical differentiation of the original current waveform data is used to calculate the acceleration related effect, and the convolution operation and the adaptively calculated trailing inertia coupling coefficient and the adaptively calculated acceleration inertia coupling coefficient are used to generate the simulated inertia component base matrix representing the potential distortion pattern, in combination with the trailing decay effect.

[0018] S5 is as follows:

[0019] Inputting the original current waveform data in the data set into the pattern recognition model, first constructing a local feature matrix through the dynamic inertia offset module, then dynamically generating a spatial variable coefficient matrix through a lightweight coefficient prediction neural network, and generating an inertia compensation signal with the inertia component base matrix.

[0020] Then, inputting the inertia compensation signal and the original current waveform data into the residual calculation and feature enhancement module, subtracting the inertia compensation signal from the original current waveform data to obtain an initial residual, and generating an enhanced residual feature vector through linear transformation and peak enhancement mechanism.

[0021] Then the residual feature and the original current waveform data are input into the dual-channel feature fusion module, and the feature extraction is performed through the backbone neural network. One channel uses the long short-term memory network to process the residual feature vector to capture the time sequence dependence relationship, and the other channel uses the convolutional neural network to process the original current waveform data vector to extract the local spatial mode. Then the extracted features are fused through the attention mechanism to generate a context weighted feature vector, which is normalized after being added to the residual feature to form a fusion feature vector.

[0022] Finally, the fusion feature is input into the feature purification and classification module. After generating a purification feature matrix from the fusion feature, multi-scale time sequence features are extracted through parallel hole convolution. The extracted features are output through a fully connected layer to generate classification probabilities. In addition, the model is optimized through classifier weight perception regularization.

[0023] In the dynamic inertia compensation module, a lightweight coefficient prediction neural network is used to dynamically generate a spatial variable coefficient matrix, and a Hadamard product is used to generate an inertia compensation signal that adapts to the current waveform distortion mode. The specific steps are as follows:

[0024] 1) The local mean, gradient intensity and energy features of the input original current waveform data are extracted through a moving average operator to construct a local feature matrix representing the dynamic characteristics of the waveform, and then the dynamic changes of the waveform are captured.

[0025] 2) A lightweight fully connected neural network is used as a coefficient prediction neural network to learn and predict a dynamic inertia compensation coefficient matrix from the local feature matrix.

[0026] 3) The dynamic coefficient matrix is combined with the inertia basis matrix through Hadamard product to generate an inertia compensation signal vector.

[0027] In the residual calculation and feature enhancement module, a feature enhancement function is used to generate a deformed and feature-enhanced residual feature vector through linear transformation and peak enhancement mechanism. The specific steps are as follows:

[0028] 1) Subtract the inertia compensation signal from the input original current waveform data to obtain an initial residual signal vector that has removed the inertia distortion, and then obtain a basic residual signal;

[0029] 2) The negative residual noise is suppressed through a learnable linear transformation, and the significant waveform events are enhanced through a peak screening mechanism to generate a residual feature vector.

[0030] In the dual-channel feature fusion module, a dual-channel fusion module is constructed to generate a classification feature vector through complementary feature extraction and adaptive fusion. The specific steps are as follows:

[0031] 1) The residual feature vector is processed using a long short-term memory network to capture its time dependence, and the input raw current waveform data vector is processed using a convolutional neural network to extract its local spatial pattern, generating two channel output vectors, which are the output vector of the residual feature channel and the output vector of the input raw current waveform data channel, respectively;

[0032] 2) The attention weight of the residual feature channel output to the raw current waveform data channel output is calculated using a normalized exponential function to generate a context weighted vector;

[0033] 3) The residual feature channel output is added to the context weighted vector through a residual connection, and a layer normalization operation is adopted to generate a fusion feature vector.

[0034] The final classification probability is generated in the feature purification and classification module through feature purification and multi-scale time series analysis, and the specific steps are as follows:

[0035] 1) Based on the gradient amplitude matrix of the inertial compensation signal of the fusion feature vector, a diagonal inertial sensitivity matrix is generated through time series maximum pooling, learnable weight mapping and Sigmoid activation function, then the diagonal inertial sensitivity matrix is operated with the unit matrix, and the fusion feature vector is acted on through Hadamard product to generate a purified feature matrix, and then the feature dimension related to residual inertial distortion is dynamically suppressed;

[0036] 2) Parallel hollow convolution is used to extract multi-scale time series patterns from the purified feature matrix, and after rectified linear unit activation, learnable weights are used to adapt the output of each scale and are spliced along the channel dimension, thereby capturing the hierarchical time dependence of waveform events;

[0037] 3) The multi-scale features are input into the bidirectional gated recurrent unit to extract time dependence, and the time attention pooling operator is used to aggregate key time step information, and finally the prediction probability vector is output through the fully connected layer and the normalized exponential function, and then the classification result is generated by combining the time dynamics and significant events;

[0038] 4) The gradient matrix of the classifier weight matrix to the inertial compensation signal vector is calculated through the chain rule approximation method, and the Frobenius norm square of the transpose and weight matrix product is taken as the penalty term, thereby quantifying the sensitivity of the classifier to inertial distortion;

[0039] The inertial sensitivity penalty term and the cross-entropy loss function are weighted and added to form the total loss function, thereby constraining the model's dependence on essential waveform features and reducing overfitting to individual device differences.

[0040] S6 adopts a two-stage cyclic update strategy to optimize the coefficient prediction neural network and the backbone neural network, and the specific steps are as follows:

[0041] 1) Phase one, update the coefficient prediction neural network parameters:

[0042] Fix the main neural network parameters, perform morphological opening and closing operation on the original current waveform data to generate pseudo-clean waveform, then update the coefficient prediction neural network parameters by minimizing the L2 distance between the compensated waveform and the pseudo-clean waveform;

[0043] 2) Phase two, update the main neural network parameters:

[0044] Fix the coefficient prediction neural network parameters, and update the main neural network parameters based on the total loss function.

[0045] In S7, the classification confidence is corrected combined with the electrical characteristics of biological tissues, and the specific steps are as follows:

[0046] 1) High-pass filter the residual feature vector, convert the tissue equivalent impedance spectrum vector through Fourier transform, reference current waveform normalization and tissue conductance characteristic parameter conversion, and then inverse the frequency-dependent impedance characteristics of biological tissues from the residual signal;

[0047] 2) Based on the deviation degree of the real part of each frequency point of the impedance spectrum vector from the ideal impedance, calculate the exponential decay weighted mean to obtain the scalar modulation factor, and then quantify the stimulation efficiency attenuation caused by tissue impedance mismatch;

[0048] 3) Convert the modulation factor into a logarithmic modulation vector, which acts on the output of the full connection layer of the classifier, generates a corrected classification probability vector through a normalized exponential function, and then adjusts the classification confidence according to the electrical characteristics of the tissue.

[0049] On the other hand, the application also provides an ear concha region vagus nerve stimulation device, which executes the mode recognition method of the ear concha region vagus nerve stimulation device output waveform, and the device comprises a base, one end of the base is provided with an antitragus fixing seat, and the other end is provided with an ear concha boat fixing seat; An antitragus electrode patch is arranged on the antitragus fixing seat, and an ear concha boat electrode patch is arranged on the ear concha boat fixing seat.

[0050] The effects provided in the summary of the invention are only the effects of the embodiments, not all the effects of the invention, and the above technical solutions have the following advantages or beneficial effects:

[0051] The application adopts a light-weight coefficient prediction neural network to dynamically generate a spatial variable coefficient matrix, and generates a compensation signal adapted to the distortion mode of the current waveform through Hadamard product, which can accurately match the distortion mode under different waveforms, can avoid destroying the effective mode of the waveform by directly subtracting a fixed proportion of inertia component base matrix, and effectively solves the inertia distortion problem caused by nonlinear changes of waveform amplitude, frequency and other factors;

[0052] The application can effectively capture the complex interaction between the original current waveform data and the de-inertia features by constructing a double-channel fusion module, extracting features in the residual feature vector and the original current waveform data vector respectively, and weighted fusion, integrating double-channel information and stabilizing feature distribution, and significantly improving the classification performance.

[0053] The application generates the final classification probability through feature purification and multi-scale time series analysis, uses parallel hollow convolution to extract multi-scale time series patterns of the purified feature matrix, captures the hierarchical time dependence of waveform events, uses a bidirectional gated recurrent unit to process long-term dependence, and uses time series attention pooling to aggregate key information, effectively solving the problems that direct classification using fusion features is easily disturbed by residual inertia artifacts, and that conventional classification layers cannot explicitly decouple waveform essential features and inertia components and ignore the multi-scale time series dependence of features.

[0054] The application combines biological tissue electrical characteristics to correct classification confidence, and through tissue impedance spectrum inversion and Bayesian probability correction, can significantly reduce the misjudgment risk caused by individual electrical characteristic differences, improve the accuracy and reliability of the classification result, and solve the problems that the conventional post-processing method cannot correct the classification deviation caused by tissue characteristics and is easy to cause misjudgment of the stimulation mode for weak response individuals.

[0055] In summary, the application can effectively distinguish the inherent mode of the waveform from the distortion caused by inertia, avoid losing important effective features in the recognition process, and improve the accuracy of recognition. BRIEF DESCRIPTION OF DRAWINGS

[0056] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application.

[0057] Figure 1 The figure is a schematic diagram of the method of the application.

[0058] Figure 2 The figure is a comparison chart of the damping oscillation wave processing effect of the application technology and the traditional method and ideal waveform.

[0059] Figure 3 The figure is a residual analysis comparison chart of the application technology and the traditional method.

[0060] Figure 4 The figure is a residual frequency spectrum analysis comparison chart of the application technology and the traditional method.

[0061] Figure 5 The figure is a waveform similarity index comparison chart of the application technology and the traditional method.

[0062] Figure 6 The figure is an inertia base matrix view of the application.

[0063] Figure 7 is a feature space distribution visualization diagram of the present technology and traditional methods.

[0064] Figure 8 is a column chart of the influence of different model components on performance indicators.

[0065] Figure 9 is a line chart of the influence of different model components on clinical indicators.

[0066] Figure 10 is a structural schematic diagram of the concha vagus nerve stimulation device.

[0067] Figure 11 is a control block diagram of the concha vagus nerve stimulation device.

[0068] wherein:

[0069] 1, base, 2, tragus fixing seat, 3, concha fixing seat, 4, tragus electrode patch, 5, concha electrode patch; 100, control unit; 110, main control module; 120, charging and discharging module; 130, output control module; 200, external control device. DETAILED DESCRIPTION

[0070] In order to clearly illustrate the technical features of the present scheme, the present application will be described in detail below with specific embodiments, and in conjunction with the accompanying drawings.

[0071] Example 1

[0072] A mode recognition method of the output waveform of the concha vagus nerve stimulation device, comprising the following steps:

[0073] S1, applying electrical stimulation to the preset target position in the human concha cavity and concha area through the concha vagus nerve stimulation device, and capturing the current waveform data output by the concha vagus nerve stimulation device in real time through a high-precision bioelectric signal acquisition system;

[0074] S2, based on the correlation between neural electrophysiological response characteristics and stimulation parameters, the expert manually labels the mode category of the collected original current waveform data to form a data set with category labels;

[0075] S3, the concha vagus nerve stimulation device generates a reference current according to the hardware parameter configuration, and outputs a reference waveform through digital-to-analog conversion of the reference current;

[0076] S4, generating an inertial component base matrix through a simulator guided by physical characteristics for the original current waveform data in the data set;

[0077] S5, a pattern recognition model is constructed, which includes a dynamic inertia offset module, a residual calculation and feature enhancement module, a dual-channel feature fusion module, and a feature purification and classification module. The original current waveform data in the data set is input into the pattern recognition model for pattern recognition prediction;

[0078] S6, the coefficient prediction neural network in the dynamic inertia offset module and the backbone neural network in the dual-channel feature fusion module are alternately trained;

[0079] S7, the classification results of the pattern recognition model are corrected based on the reference current waveform, and finally the corrected classification probability is generated.

[0080] In the specific implementation, S1 is specifically as follows:

[0081] Through a high-precision bioelectric signal acquisition system, real-time capture of original current waveform data output by the vagus nerve stimulation device in the ear and the ear is realized.

[0082] Specifically, during data acquisition, a multi-channel isolated amplifier is used to connect a clinical level stimulation electrode, and an electrical stimulation is applied to the preset target position in the human ear cavity and ear. The sampling rate is not less than 10kHz, and the stimulation waveform is recorded synchronously to obtain the original stimulation waveform data vector.

[0083] The collection process strictly follows the electromagnetic compatibility standard of medical equipment, and the environmental noise interference is eliminated through optical isolation technology to ensure that the original waveform data completely retains the current amplitude, phase and transient response characteristics of the stimulation device output.

[0084] In the specific implementation, S2 is specifically as follows:

[0085] The labeled categories cover four typical waveform patterns, namely standard single-phase square wave, damped oscillation wave, step decay wave and high-frequency pulse mutation wave.

[0086] In the specific implementation, S3 is specifically as follows:

[0087] The reference current waveform is directly generated from the hardware parameter configuration file of the stimulation device;

[0088] Specifically, according to the preset stimulation protocol, including pulse width 50-200μs, amplitude 0.1-5mA, frequency 1-100Hz, the ideal square wave, triangular wave or sine wave is output as the reference waveform through the digital-to-analog conversion module, and the reference waveform is used as the reference current waveform;

[0089] The reference current waveform is measured and verified by a high-precision oscilloscope under zero impedance load conditions to ensure that the rising / falling edge time is less than 10μs and there is no overshoot distortion, and the reference waveform data is stored in the form of discrete time sequence, and the length is strictly aligned with the original stimulation waveform.

[0090] In the specific implementation, the vagus nerve stimulation waveform data in the concha area has non-stationarity, is easily affected by biological tissue characteristics and device physical inertia, and causes non-pathological distortion of the waveform form, such as overshoot or tailing; the conventional pretreatment method adopts band-pass filtering or simple normalization processing on these waveforms, but cannot effectively distinguish the inherent mode of the waveform from the distortion caused by inertia, easily confuses effective features, and thus reduces the accuracy of subsequent classification, therefore, in the present application, the original current waveform data in the data set is subjected to a simulator guided by physical characteristics to generate an inertia component base matrix, and S4 is specifically as follows:

[0091] The original current waveform data in the data set is subjected to a simulator guided by physical characteristics to generate an inertia component base matrix, to represent potential distortion modes jointly caused by the physical characteristics of the concha area vagus nerve stimulation device and the biological tissue response; the acceleration-related effect is calculated based on the numerical differentiation of the original current waveform data, and the tailing inertia coupling coefficient and the acceleration inertia coupling coefficient are adaptively calculated through convolution operation and self-adaptive calculation, to generate a simulated inertia component base matrix representing the potential distortion mode, which can provide a target base for subsequent precise compensation of inertia distortion, distinguish the inherent mode of the waveform from the non-pathological distortion caused by physical inertia, and avoid confusing effective features;

[0092] The calculation formula is as follows:

[0093] ,

[0094] In the formula, represents the simulated inertia component base matrix corresponding to the i-th sample, representing the potential distortion mode jointly caused by the physical characteristics of the stimulation device and the biological tissue response; represents the original current waveform data vector of the i-th sample in the data set; represents the second-order numerical differentiation of , representing the inertia effect related to acceleration; represents convolution operation;

[0095] represents the acceleration inertia coupling coefficient, which is a scalar coefficient, and its value is adaptively calculated according to the input waveform, and the calculation method is represented as ;is the maximum value of the input original current waveform data vector ;is the minimum value of the input original current waveform data vector ; is the variance of the input original current waveform data vector ; is the variance of the input original current waveform data vector ; is the variance of the input original current waveform data vector is the length of the input raw current waveform data vector; is a positive integer;

[0096] is a scalar coefficient, which is adaptively calculated according to the input waveform, and the calculation method is represented as ; is the skewness of the input raw current waveform data vector ; is a natural exponential function; is a scale hyperparameter in the calculation of the trailing inertia coupling coefficient;

[0097] is an inertia decay kernel vector, which is used to represent the trailing decay effect caused by inertia, and the calculation method is represented as , that is, is represented as , ; is the th element of the inertia decay kernel vector ; is a natural constant; is a decay rate, which is estimated by the average decay time of the waveform data, and is set as ; is a trailing time constant, and is set as ; is a kernel length, which is matched with the waveform sampling rate, and is set as ; is a time index;

[0098] It should be noted that, controls the sensitivity of the skewness to the contribution of , and increasing will enhance the influence of skewness on the trailing effect, and, is obtained by grid search optimization on the training set, and the influence of different values on the classification accuracy is evaluated on the validation set, for example, under the condition of , the that makes the validation loss minimum is selected, and if there is no validation set, the default value is set as .

[0099] In the specific implementation, S5 is specifically as follows:

[0100] The original current waveform data in the data set is input into the pattern recognition model, first passes through the dynamic inertia offset module to construct a local feature matrix, and then adopts a lightweight coefficient prediction neural network to dynamically generate a spatial variable coefficient matrix, and generates an inertia compensation signal with the inertia component base matrix;

[0101] Then, the inertial compensation signal and the original current waveform data are input into the residual calculation and feature enhancement module. The initial residual is obtained by subtracting the inertial compensation signal from the input original current waveform data. Then, the enhanced residual feature vector is generated through linear transformation and peak enhancement mechanism.

[0102] Next, the residual features and the original current waveform data are input into the dual-channel feature fusion module. Feature extraction is performed through the backbone neural network. One channel uses a long short-term memory network to process the residual feature vector to capture the temporal dependency, while the other channel uses a convolutional neural network to process the original current waveform data vector to extract local spatial patterns. The extracted features are then fused through an attention mechanism to generate a context-weighted feature vector, which is then added to the residual features and normalized to form a fused feature vector.

[0103] Finally, the fused features are input into the feature purification and classification module. After generating a purified feature matrix based on the fused features, multi-scale temporal features are extracted through parallel dilated convolution. The extracted features are then output as classification probabilities through a fully connected layer. In addition, the model is optimized through classifier weight-aware regularization.

[0104] In a specific implementation, during the processing of vagus nerve stimulation waveform data in the concha region, the inertial distortion intensity changes nonlinearly with factors such as waveform amplitude and frequency. Directly subtracting a fixed proportion of the inertial component basis matrix would destroy the effective waveform pattern. Conventional processing methods, such as fixed thresholding or fixed gain compensation, lack adaptability and cannot accurately match the distortion patterns under different waveforms. Therefore, this invention employs a lightweight coefficient prediction neural network to dynamically generate a spatially variable coefficient matrix in the dynamic inertial cancellation module, and generates an inertial compensation signal adapted to the current waveform distortion pattern through Hadamard product. The specific steps are as follows:

[0105] 1) Constructing the local feature matrix:

[0106] The local mean, gradient strength, and energy characteristics of the input raw current waveform data are extracted by the moving average operator, and a local feature matrix characterizing the dynamic properties of the waveform is constructed to capture the dynamic changes of the waveform.

[0107] The calculation formula is as follows:

[0108] ,

[0109] In the formula, Indicates the first The local feature matrix of each sample; Represents the moving average operator; The energy vector representing the waveform is obtained by... It is obtained by squaring each element; The window width for the moving average of the mean; The window width is the moving average of the gradient intensity. The window width for the energy moving average; This represents the matrix transpose operation;

[0110] Indicates the first The absolute value vector of the first-order difference of the waveforms of each sample is calculated first, then the first-order difference vector is calculated. ,in ,and Boundary values Then take the absolute value to get the first... The absolute value vector of the first-order difference of the waveform of the n samples element And thus obtain the first The absolute value vector of the first-order difference of the waveform of each sample; For the first The first-order difference vector of each sample; For the first The first-order difference vector of the n samples is... One element; For the first The first element of the first-order difference vector of each sample; For the first The first sample of the original stimulus waveform data vector One element; For the first The first sample of the original stimulus waveform data vector One element;

[0111] It should be noted that, in , and During the processing, the moving average operator applies the input vector to... , or Perform smoothing processing, in order to For example, regarding window width Each element of the output vector ,in, It is a positive integer. For positive integers, boundary values ​​are padded with zeros. for The One element, This indicates rounding down. For the first The first sample of the original stimulus waveform data vector One element;

[0112] 2) Predicting the dynamic coefficient matrix using a coefficient prediction neural network:

[0113] By using a lightweight fully connected neural network as the coefficient prediction neural network, the dynamic inertia cancellation coefficient matrix is ​​learned and predicted from the local feature matrix, thereby enabling fine-grained control of the compensation amount at each time point to adapt to nonlinear changes in waveform amplitude and frequency.

[0114] The calculation formula is as follows:

[0115] ,

[0116] In the formula, Indicates the first The dynamic inertia cancellation coefficient matrix of each sample; The coefficient prediction neural network adopts a three-layer fully connected architecture, namely, input layer - hidden layer - output layer; This represents the set of learnable parameters for a coefficient prediction neural network, including the weight matrix and bias vector.

[0117] It should be noted that the role of the coefficient prediction neural network is to... Mapped to , The dimension is The output layer is mapped through a fully connected layer, with a dimension of [dimensional value missing]. That is, the output for matrix;

[0118] 3) Generate inertial compensation signal:

[0119] By combining the dynamic coefficient matrix with the inertial basis matrix through the Hadamard product, an inertial compensation signal vector is generated, thereby producing a compensation signal that precisely matches the current waveform distortion mode, thus avoiding the destruction of the effective waveform mode.

[0120] The calculation formula is as follows:

[0121] ,

[0122] In the formula, Indicates the first The inertial compensation signal vector of each sample; This represents the Hadamard product, which is the element-wise multiplication of matrices.

[0123] It should be noted that, During the calculation process, for matrix, for Vectors, before operations Broadcast for Matrix, i.e., copying each column Then multiply element by element to get The matrix is ​​then averaged over each row and compressed into a single matrix. dimensional vector .

[0124] In a specific implementation, although inertial cancellation of the original current waveform data can remove some inertial distortion, residual noise and weak features still affect classification accuracy. Conventional methods directly input the residual signal into the classifier, which is difficult to effectively highlight key waveform features such as the peak of the stimulus pulse and cannot suppress residual noise. Therefore, the residual calculation and feature enhancement module of this invention uses a feature enhancement function to generate distortion-reduced and feature-enhanced residual feature vectors through linear transformation and peak enhancement mechanisms. The specific steps are as follows:

[0125] 1) Calculate the initial residual:

[0126] The inertial compensation signal is subtracted from the input raw current waveform data to obtain the initial residual signal vector after the inertial distortion has been initially removed, and then the basic residual signal is obtained.

[0127] The waveform after removing inertial distortion is shown below:

[0128] ,

[0129] In the formula, Indicates the first The initial residual signal vector of each sample represents the waveform after removing inertial distortion;

[0130] 2) Feature enhancement processing:

[0131] By suppressing negative residual noise through learnable linear transformation and enhancing significant waveform events by combining peak filtering mechanism, residual feature vectors are generated, which can highlight effective waveform features and suppress residual noise, ensuring that the signal input to subsequent networks retains key information and reduces distortion interference.

[0132] The calculation formula is as follows:

[0133] ,

[0134] In the formula, Indicates the first The residual feature vector of each sample; This represents the activation function of the rectified linear unit, used to suppress negative residuals; Represents the learnable linear transformation weight matrix; Represents a learnable bias vector; This represents the element-wise addition operator;

[0135] represents a peak value screening function, the output is the same dimension as the input vector calculation method is represented as ; represents the first feature value of the initial residual error signal vector of the first sample; represents the global mean of the initial residual error signal vector of the first sample; represents the global standard deviation of the initial residual error signal vector of the first sample; represents a peak determination coefficient, which controls the sensitivity of peak screening, and is set to .

[0136] In the embodiment, using only the de-inertia residual feature may lose the global structure information of the original current waveform data, such as the overall envelope shape. The conventional feature fusion method uses simple splicing or weighted average, which is difficult to effectively capture the complex interaction between the original current waveform data and the de-inertia feature, thereby limiting the classification performance. Therefore, in the dual-channel feature fusion module of the present application, a dual-channel fusion module is constructed to generate a classification feature vector through complementary feature extraction and adaptive fusion. The specific steps are as follows:

[0137] 1) Extract dual-channel features:

[0138] The residual feature vector is processed using a long short-term memory network to capture its time sequence dependence, and the input original current waveform data vector is processed using a convolutional neural network to extract its local spatial pattern, generating two channel output vectors, which are the output vector of the residual feature channel and the output vector of the input original current waveform data channel, respectively. Then, the complementary expression of the waveform on the de-distortion feature and the original form can be obtained.

[0139] The calculation formula is as follows:

[0140] ,

[0141] ,

[0142] In the formula, represents the output vector of the residual feature channel of the first sample, which is generated by the LSTM network and captures the time sequence dependence of the de-inertia feature; represents the output vector of the original current waveform data channel of the first sample, which is generated by the CNN network and extracts the local spatial pattern of the original current waveform data; represents a long short-term memory network, and the parameter is ; This is the set of weights and bias parameters for the LSTM network. Represents a convolutional neural network with parameters as follows: ; This is the set of weights and bias parameters for a CNN network.

[0143] 2) Attention mechanism fusion:

[0144] The attention weight of the residual feature channel output to the original current waveform data channel output is calculated by using the normalized exponential function, and a context weighted vector is generated to adaptively highlight the important parts of the original current waveform data affected by the distortion removal feature.

[0145] The calculation formula is as follows:

[0146] ,

[0147] In the formula, Indicates the first The context-weighted vector of each sample reflects China Key characteristics of the impact; This is the weight matrix for the query vector; The weight matrix is ​​the weight matrix of the key vectors; This represents the scaling factor used for calculating the stable gradient; for Transpose of; This represents a normalized exponential function, used to calculate attention weights along the row direction;

[0148] 3) Generate feature vectors:

[0149] The residual feature channel output is added to the context weighted vector by residual connection, and a layer normalization operation is used to generate a fused feature vector, thereby integrating dual-channel information and stabilizing the feature distribution.

[0150] The calculation formula is as follows:

[0151] ;

[0152] In the formula, Indicates the first The fusion feature vector of each sample; Presentation layer normalization operation.

[0153] In the embodiment, if the fusion feature is directly used for classification, it is easy to be disturbed by residual inertial artifacts, the conventional classification layer such as the full connection layer plus the Softmax function calculation cannot explicitly decouple the waveform essential feature and the inertial component, and the multi-scale time sequence dependency of the feature is ignored, therefore, in the feature purification and classification module of the application, the final classification probability is generated through the feature purification and multi-scale time sequence analysis, and the specific steps are as follows:

[0154] 1) Inertial mask generation and feature purification:

[0155] Based on the gradient amplitude matrix of the inertial compensation signal on the fusion feature vector, a diagonal inertial sensitivity matrix is generated through the time sequence maximum pooling, the learnable weight mapping and the Sigmoid activation function, then, after the operation of the diagonal inertial sensitivity matrix and the unit matrix, the fusion feature vector is acted on through the Hadamard product to generate a purified feature matrix, and then the feature dimension related to the residual inertial distortion is dynamically inhibited;

[0156] The calculation formula is as follows:

[0157] ,

[0158] ,

[0159] In the formula, is the Sigmoid activation function; is the learnable mask weight matrix, the dimension is ; is the maximum pooling compression rate, set ; is the dimension of the fusion feature vector ; is the unit matrix, the dimension corresponds to , the dimension is ; is the time sequence maximum pooling operator; is the gradient amplitude matrix of the feature to the inertial component, the dimension is , and the calculation mode is represented as ; represents the gradient matrix of the fusion feature vector to the inertial compensation signal vector , which is obtained in an approximate calculation manner, and the calculation mode is represented as ; represents the gradient matrix of the fusion feature vector to the residual feature vector , which is specifically the Jacobian matrix; represents the conversion of the vector to the diagonal matrix; is the purified feature matrix; The inertial sensitivity matrix, representing the inertial mask, is a diagonal matrix with dimension 1. ;

[0160] 2) Extracting multi-scale temporal features:

[0161] Parallel dilated convolution is used to extract multi-scale temporal patterns from the cleaned feature matrix. After activation by rectified linear units, learnable weights are used to adapt the output at each scale and splice them along the channel dimension to capture the hierarchical temporal dependence of waveform events.

[0162] The calculation formula is as follows:

[0163] ,

[0164] ,

[0165] In the formula, For the first Convolutional feature output at each scale, dimension ; Represents the activation function of the rectified linear unit; Number of output channels One-dimensional dilated convolution; This indicates that the kernel width is 3; Indicates the void ratio ; This represents the number of output channels for the convolution; For example, the scale number, ; It is a positive integer; This indicates splicing along the channel dimension, with a scale number of... ; Weights are adapted to a learnable scale;

[0166] 3) Generate classification probabilities:

[0167] Multi-scale features are input into a bidirectional gated recurrent unit to extract temporal dependencies. A temporal attention pooling operator is used to aggregate key time step information. Finally, a predicted probability vector is output through a fully connected layer and a normalized exponential function. This process combines temporal dynamics and significant events to generate classification results.

[0168] The calculation formula is as follows:

[0169] ,

[0170] ,

[0171] In the formula, It is a bidirectional gated loop unit with 128 hidden units; Represents the predicted probability vector; is the set of learnable parameters of the bidirectional gated recurrent unit; is the normalized exponential function; is the classifier weight matrix, with dimension ; is the dimension of the bidirectional gated recurrent unit hidden state, i.e., the length of the feature vector output at each time step, when the number of hidden units of the bidirectional gated recurrent unit is 128, ; is the number of classes;

[0172] is the time sequence attention pooling operator, and the calculation method is represented as ; is the output vector of the bidirectional gated recurrent unit at the time step, with dimension ; is the kernel length, and the length of the output sequence of the bidirectional gated recurrent unit is defined to match the kernel length, and is set to ; is the time index;

[0173] is the attention weight at the time step, and the calculation method is represented as ; is the attention score calculation vector, with dimension ; is the hidden layer dimension of the attention mechanism, i.e., the projection space dimension in attention calculation, and is set to ; is the transpose of ; is the learnable attention enhancement weight matrix; is the learnable attention enhancement bias vector; is the output vector of the bidirectional gated recurrent unit at the time step, with dimension ;

[0174] 4) Custom regularization:

[0175] Since the model is prone to overfitting to specific inertial distortion patterns in the training set during training, leading to a decline in the generalization performance for new device individuals, and conventional regularization methods such as L2 regularization cannot specifically suppress the sensitivity of the classifier to the inertial component, it is difficult to improve the recognition ability of the model to the essential waveform features. Therefore, by using the chain rule approximation method, the gradient matrix of the classifier weight matrix to the inertial compensation signal vector is calculated, and the Frobenius norm square of the transpose of the weight matrix multiplied by the weight matrix is taken as the penalty term, which can quantify the sensitivity of the classifier to the inertial distortion;

[0176] The calculation formula is as follows:

[0177] ,

[0178] In the formula, represents an inertial sensitivity penalty term, used to measure the sensitivity of the classifier to the inertial component; represents a classifier weight matrix, with dimensions , is the feature dimension, is the number of categories; is the transpose of ; represents a fusion feature vector , the gradient matrix of the inertial compensation signal vector is obtained in an approximate calculation manner, and the calculation manner is represented as ; represents a fusion feature vector , the gradient matrix of the residual feature vector is obtained in an approximate calculation manner, and the calculation manner is represented as ;

[0179] The inertial sensitivity penalty term and the cross-entropy loss function are weighted and added to form a total loss function, which further constrains the model to depend on the essential waveform features, reduces the overfitting to the individual differences of the equipment, and improves the generalization ability of the model.

[0180] The calculation formula is as follows:

[0181] ,

[0182] In the formula, represents a total loss function; represents a cross-entropy loss function; represents a real label vector; represents a predicted probability vector; represents a classifier bias vector; represents a regularization strength coefficient, which controls the weight of the penalty term, and is set to .

[0183] In the specific implementation, if the coefficient prediction neural network and the backbone neural network are jointly trained, the coefficient prediction neural network is easy to be dominated and degenerate, for example, outputting a full zero matrix. The conventional alternating optimization method lacks clear supervision target guidance, resulting in that the inertial modeling and the classification task cannot be effectively cooperated. Therefore, the present application adopts a two-stage cyclic updating strategy to cooperatively optimize the coefficient prediction neural network and the backbone neural network, and the specific steps of S6 are as follows:

[0184] 1) In the first stage, the parameters of the coefficient prediction neural network are updated:

[0185] The original current waveform data is subjected to morphological opening and closing operations to generate pseudo-clean waveforms, and then the coefficient prediction neural network parameters are updated by minimizing the L2 distance between the compensated waveforms and the pseudo-clean waveforms.

[0186] The updated coefficient prediction neural network parameters are represented as:

[0187] wherein, represents a set of learnable parameters of the coefficient prediction neural network, including a weight matrix and a bias vector; is an L2 norm; represents a loss function that is minimized with respect to .

[0188] represents a pseudo-clean waveform vector, which is estimated by morphological filtering, and the calculation method is represented as . represents morphological opening operation; represents morphological closing operation; is the size of the opening operation structure element; is the size of the closing operation structure element;

[0189] It should be noted that the morphological opening operation and the morphological closing operation are implemented by one-dimensional morphological operation, and the structure element is a one-dimensional kernel with a length of / , such as a rectangular kernel, which is subjected to sliding window processing according to elements.

[0190] 2) Phase two, updating the main neural network parameters:

[0191] The coefficient prediction neural network parameters are fixed, the main neural network parameters are updated based on the total loss function, and then focused on the classification task optimization.

[0192] The updated main neural network parameters are represented as .

[0193] wherein, represents a set of trainable parameters of the main neural network, represents a loss function that is minimized with respect to .

[0194] ​In a specific implementation, since the probability vector output by the classifier does not consider the individual differences in electrical properties of biological tissues, such as impedance distribution and dielectric constant, when the stimulus waveform passes through concha tissues with different electrical conductance properties, the actual efficiency of neuron activation is not monotonically correlated with the waveform morphology. Conventional post-processing methods, such as probability threshold truncation, cannot correct the classification bias caused by tissue characteristics, which can easily lead to misjudgment of stimulus patterns for individuals with weak responses. Therefore, this invention combines the electrical properties of biological tissues to correct the classification confidence. The specific steps in S7 are as follows:

[0195] 1) Tissue impedance spectroscopy inversion:

[0196] The residual eigenvector is high-pass filtered, and the tissue equivalent impedance spectrum vector is generated by Fourier transform, reference current waveform normalization and tissue conductivity parameter conversion. Then, the frequency-varying impedance characteristics of biological tissue are inverted from the residual signal.

[0197] The calculation formula is as follows:

[0198] ,

[0199] In the formula, Fourier transform; This is the inverse Fourier transform; The equivalent impedance spectrum vector of the organization is a complex vector; Indicates the first The residual feature vector of each sample; The reference current waveform is preset by the parameters of the stimulation device;

[0200] For the high-pass filter mask matrix, extract Components, calculated as follows , ; The first high-pass filter mask Each element has a value of 0 or 1; These are the frequency components after the Fourier transform; It is a positive integer; Set the cutoff frequency. If so, the cutoff frequency is set to 1000Hz;

[0201] The conductivity-impedance conversion coefficient is approximated by the Helmholtz equation, and the calculation method is expressed as follows: ; The average conductivity is calculated as follows: ; The equivalent dielectric constant is calculated as follows: ; Let be the vacuum permittivity, set ; is the sampling interval, set ; is the angular frequency, set ; denotes the L1 norm of , which is calculated as ; is the th element of the residual feature vector of the th sample; is a positive integer; is the length of the original stimulus waveform data vector;

[0202] 2) Calculate the stimulation intensity modulation factor:

[0203] Based on the deviation of the real part of each frequency point of the impedance spectrum vector from the ideal impedance, calculate the scalar modulation factor by the exponentially decaying weighted mean, and then quantify the stimulation efficiency attenuation caused by the impedance mismatch of the tissue;

[0204] The calculation formula is as follows:

[0205] ,

[0206] wherein is the ideal tissue impedance, set ; is the complex impedance value of the tissue equivalent impedance spectrum vector at the th frequency point; is a positive integer; is the stimulation intensity modulation factor, which is a scalar, and , used to quantify the stimulation efficiency attenuation caused by the deviation of the tissue impedance from the ideal state, when the optimal impedance matching is indicated; is the impedance mismatch penalty coefficient, which is optimized by grid search, such as setting ; denotes taking the real part of a complex number;

[0207] 3) Bayesian probability correction:

[0208] Convert the modulation factor into a logarithmic modulation vector, which acts on the output of the full connection layer of the classifier, generate a corrected classification probability vector through the normalized exponential function, and then adjust the classification confidence according to the electrical characteristics of the tissue, which can suppress the risk of misjudgment under impedance mismatch;

[0209] The calculation formula is as follows:

[0210] ,

[0211] In the formula, Let be the logarithmic modulation vector, defined as , dimension ; For stimulus intensity modulation factor The exponentiation is a scalar exponentiation operation of the stimulus intensity modulation factor; Represents a logarithmic function with base to the natural constant, in In China, Take the natural logarithm element by element, and The dimension is ; For the modulation vector, , Will Expand to The first element of the dimension is always 1 to preserve the baseline class probability; Number of categories; To correct the classification probability vector The category sensitivity index is a scalar value, set as follows: ;

[0212] It should be noted that when When the impedance approaches 0, it indicates a severe mismatch in tissue impedance. This makes the logits of non-baseline categories tend towards negative infinity, thereby suppressing the risk of misclassification. Logits refer to the unnormalized scores before Softmax.

[0213] Example 2

[0214] like Figures 10 to 11 As shown, this embodiment proposes a vagus nerve stimulation device for the concha region, used to perform a pattern recognition method for the output waveform of the vagus nerve stimulation device for the concha region. The device includes a base 1, with a tragus fixation seat 2 at one end and a concha cymba fixation seat 3 at the other end. Tragus electrode patches 4 are disposed on the tragus fixation seat 2, and concha cymba electrode patches 5 are disposed on the concha fixation seat 3. A control unit 100 is provided inside the base, including a main control module, a charging and discharging module, and an output control module. The main control module 110 is used for signal interaction, managing charging and discharging, and controlling the output stimulation signal. The charging and discharging module 120 is connected to the main control module 110 and charges the battery according to the control signal. The output control module 130 is connected to the main control module 110 and outputs a stimulation signal with frequency, pulse width, and current intensity that meet the specified requirements according to the control signal.

[0215] The main control module 110 may further include: a communication module for signal interaction with the external control device 200, sending and receiving control signals; and a management module for sending the received control signals to the charging / discharging module 120 and the output control module 130, and managing and controlling the corresponding modules to work collaboratively according to the control signals. The main control module 110 can be paired and bound to the external control device 200, then receive control signals from the external control device 200 and feed back the current stimulation state to the external control device 200. Furthermore, the external control device 200 can display the received stimulation state through a display component, facilitating the patient's or therapist's understanding of the current stimulation state.

[0216] Example 3

[0217] like Figures 2 to 6 As shown, the effect of processing the vagus nerve stimulation waveform in the concha region using the method of the present invention is verified. By comparing with conventional processing methods and ideal waveforms, the ability of the dynamic inertia cancellation module to suppress waveform distortion is verified.

[0218] The experiment selected a typical damped oscillation wave as the object of analysis. Figure 2 The comparison view of the damped oscillation wave processing effect clearly shows the overshoot distortion and tail distortion of the original current waveform data. The waveform processed by the traditional method still has obvious distortion (blue curve), while the waveform processed by the technology of this invention (red curve) almost coincides with the ideal waveform (green dashed line). Figure 3 The residual analysis view shown quantifies that the residual amplitude of the present invention is significantly smaller than that of the traditional method, and the energy distribution is more uniform. Figure 4 The residual spectrum analysis shown attempts to further reveal that the residual energy of the present invention is concentrated in the low-frequency region, indicating that it effectively filters out high-frequency distortion components; Figure 5 In the waveform similarity index comparison view shown, the technology of this invention is superior to the traditional method in all three indicators: root mean square error, correlation coefficient and peak error, which intuitively reflects the synergistic advantages of dynamic inertia cancellation and feature enhancement module. Figure 6 The inertial basis matrix view shown is visualized through a frequency-time heatmap, demonstrating the inertial component modeling effect guided by the physical properties of this invention, providing a theoretical basis for waveform purification.

[0219] Example 4

[0220] like Figure 7 As shown, the t-SNE dimensionality reduction method is used to reveal the discriminative ability of different methods in extracting features through two-dimensional feature space projection. Figure 7In the left side of the figure, the feature space of the traditional method, the feature points of the four types of waveforms (standard single-phase square wave, damped oscillation wave, step decay wave, and high-frequency pulse mutation wave) are severely overlapped, especially the damped oscillation wave and the step decay wave form a diffuse cross zone, indicating that the traditional features cannot effectively classify similar waveform patterns. Figure 7 In the right side of the figure, the feature space of the present invention, the four types of feature points form compact clusters, and the boundaries between the clusters are clearly visible. The high-frequency pulse mutation wave maintains the maximum distance from the other categories. The experimental results show that the significant separation is the result of the combined action of feature purification and multi-scale time sequence classification module. The parallel hollow convolution captures the hierarchical time dependence of pulse events, the bidirectional gate recurrent unit integrates long-term context, and the time attention pooling focuses on the key stimulus points. In the feature space of the present invention, the point groups of each category are distributed radially, reflecting their retention of the continuous change characteristics of waveform amplitude and phase. In contrast, the feature points of the traditional method are randomly aggregated, which confirms the feature confusion caused by inertial distortion interference.

[0221] Example 5

[0222] As Figures 8 to 9 shown, through progressive ablation design, the contribution of each technical component to performance and clinical indicators is quantitatively evaluated, Figure 8 The bar chart shows the improvement trajectory of accuracy, precision, recall, and F1 score when adding core modules (fixed inertia compensation, dynamic inertia offset, dual-channel fusion, feature purification, and classification correction) to the baseline model. The experimental results show that the dynamic inertia offset module brings the largest performance jump, followed by the dual-channel fusion, proving that the dynamic inertia offset module and the dual-channel fusion operation are the core carriers of technical advantages.

[0223] Figure 9 The line chart analyzes the promoting effect of each technical component on clinical effect. The blue curve (circle marker) represents the stimulation efficiency, which measures the proportion of effective neuroregulation. Specifically, it is defined as the proportion of stimulation pulses that successfully trigger the target physiological response, which directly determines the effectiveness of treatment. The higher the value, the better the treatment effect. The red curve (square marker) represents the mis-stimulation rate, which reflects the proportion of incorrect activation. Specifically, it is defined as the proportion of mis-triggering of target nerve fibers that do not accurately hit the target, which represents the size of treatment safety. The lower the value, the lower the risk of side effects. The stimulation efficiency curve (blue) increases quasi-linearly with the increase of technical components, while the mis-stimulation rate curve (red) decreases rapidly. The complete model achieves a clinical-level performance with a stimulation efficiency of over 90% and a mis-stimulation rate of less than 5%. The experimental results verify the clinical value of the classification result correction module. Through impedance spectrum inversion and Bayesian probability correction, it significantly reduces the risk of misjudgment caused by individual electrical characteristic differences. In particular, after adding the classification correction module, the mis-stimulation rate drops sharply, proving its protective effect on weak response individuals.

[0224] The above describes the specific embodiments of the application in conjunction with the drawings, but is not a limitation on the protection scope of the application. Various modifications or variations made by those skilled in the art on the basis of the technical solutions of the application without creative labor are still within the protection scope of the application.

Claims

1. A method of pattern recognition of an output waveform of a concha vagus nerve stimulation device, characterized by, The method comprises the following steps: S1, applying electric stimulation to the preset target position in the human body's concha cavity and concha area through a concha area vagus nerve stimulation device, and capturing the current waveform data output by the concha area vagus nerve stimulation device in real time through a high-precision bioelectric signal acquisition system; S2, based on the correlation between the neural electrophysiological response characteristics and the stimulation parameters, manually labeling the mode category of the collected original current waveform data by an expert to form a data set with category labels; S3, the concha area vagus nerve stimulation device generates a reference current according to the hardware parameter configuration, and outputs a reference waveform after digital-to-analog conversion of the reference current; S4, generating an inertial component base matrix for the original current waveform data in the data set through a simulator guided by physical characteristics; S4 is as follows: Generate an inertial component base matrix for the original current waveform data in the data set through a simulator guided by physical characteristics, to represent the potential distortion mode generated by the physical characteristics of the concha area vagus nerve stimulation device and the biological tissue response; Specifically, based on the numerical differentiation calculation of the original current waveform data, the acceleration-related effect is calculated, and the convolution operation and the adaptively calculated trailing inertia coupling coefficient and the adaptively calculated acceleration inertia coupling coefficient are combined to generate the simulated inertial component base matrix representing the potential distortion mode; S5, constructing a pattern recognition model, which includes a dynamic inertia offset module, a residual calculation and feature enhancement module, a dual-channel feature fusion module, and a feature purification and classification module, and inputting the original current waveform data in the data set into the pattern recognition model for pattern recognition prediction; S5 is as follows: Input the original current waveform data in the data set into the pattern recognition model, first pass through the dynamic inertia offset module to construct a local feature matrix, then use a lightweight coefficient prediction neural network to dynamically generate a spatial variable coefficient matrix, and generate an inertia compensation signal with the inertia component base matrix; Then input the inertia compensation signal and the original current waveform data into the residual calculation and feature enhancement module, subtract the inertia compensation signal from the original current waveform data to obtain the initial residual, and generate the enhanced residual feature vector through linear transformation and peak enhancement mechanism; Next, input the residual feature and the original current waveform data into the dual-channel feature fusion module, extract features through the backbone neural network, one channel uses a long short-term memory network to process the residual feature vector to capture the time sequence dependence, and the other channel uses a convolutional neural network to process the original current waveform data to extract local spatial patterns, then fuse the extracted features through an attention mechanism to generate a context weighted feature vector, add the residual feature after normalization to form a fusion feature vector; Finally, input the fusion feature into the feature purification and classification module, generate a purified feature matrix according to the fusion feature, extract multi-scale time sequence features through parallel hole convolution, output classification probability through a fully connected layer, and further optimize the model through classifier weight perception regularization; S6, alternately train the coefficient prediction neural network in the dynamic inertia offset module and the backbone neural network in the dual-channel feature fusion module. S7. Correct the classification result of the pattern recognition model based on the reference current-based reference waveform, and finally generate the corrected classification probability.

2. The method of claim 1, wherein the pattern recognition is based on a pattern of output waveforms of the concha vagus nerve stimulation device. In the dynamic inertia compensation module, a lightweight coefficient prediction neural network is used to dynamically generate a spatial variable coefficient matrix, and a Hadamard product is used to generate an inertia compensation signal that adapts to the current waveform distortion mode. The specific steps are as follows: 1) Extract the local mean, gradient intensity and energy features of the input original current waveform data through the moving average operator, construct the local feature matrix representing the dynamic characteristics of the waveform, and then capture the dynamic changes of the waveform; 2) Use a lightweight fully connected neural network as a coefficient prediction neural network to learn and predict a dynamic inertia compensation coefficient matrix from the local feature matrix; 3) Combine the dynamic coefficient matrix with the inertia basis matrix through Hadamard product to generate an inertia compensation signal vector.

3. The method of claim 1, wherein the pattern recognition is based on a mode of the output waveform of the concha vagus nerve stimulation device. In the residual calculation and feature enhancement module, a feature enhancement function is used to generate a de-distortion and feature-enhanced residual feature vector through linear transformation and peak enhancement mechanism. The specific steps are as follows: 1) Subtract the inertia compensation signal from the input original current waveform data to obtain an initial residual signal vector that has initially removed inertia distortion, and then obtain a basic residual signal; 2) Suppress negative residual noise through a learnable linear transformation, and enhance significant waveform events through a peak selection mechanism to generate a residual feature vector.

4. The method of claim 1, wherein the pattern recognition of the output waveform of the concha vagus nerve stimulation device is characterized by, In the dual-channel feature fusion module, a dual-channel fusion module is constructed to generate a classification feature vector through complementary feature extraction and adaptive fusion. The specific steps are as follows: 1) Use a long short-term memory network to process the residual feature vector to capture its temporal dependence, and use a convolutional neural network to process the input original current waveform data vector to extract its local spatial pattern, generating two channel output vectors, namely the residual feature channel output vector and the input original current waveform data channel output vector; 2) Use a normalized exponential function to calculate the attention weight of the residual feature channel output on the original current waveform data channel output to generate a context weighted vector; 3) Add the residual feature channel output and the context weighted vector through a residual connection, and use layer normalization to generate a fusion feature vector.

5. The method of claim 1, wherein the pattern recognition is based on a feature of the output waveform of the concha vagus nerve stimulation device. In the purification and classification module, the final classification probability is generated through feature purification and multi-scale temporal analysis. The specific steps are as follows: 1) Based on the gradient amplitude matrix of the inertia compensation signal, generate a diagonal inertia sensitivity matrix through temporal maximum pooling, learnable weight mapping and Sigmoid activation function, then operate the diagonal inertia sensitivity matrix and the identity matrix, and then apply the Hadamard product to the fusion feature vector to generate a purified feature matrix, which dynamically suppresses the feature dimensions related to residual inertia distortion; 2) Use parallel atrous convolution to extract multi-scale temporal patterns from the purified feature matrix, activate through a rectified linear unit, use learnable weights to adapt the output of each scale, and then concatenate along the channel dimension to capture the hierarchical temporal dependence of waveform events. 3) The multi-scale features are input into the bidirectional gated recurrent unit to extract the time-dependent relationship, the key time step information is aggregated using the time attention pooling operator, and finally the prediction probability vector is output through the fully connected layer and the normalized exponential function, and then the classification results are generated by combining the time dynamics and significant events; 4) Custom regularization: By using the chain rule approximation method, the gradient matrix of the classifier weight matrix with respect to the inertial compensation signal vector is calculated, and the Frobenius norm square of the transpose of the product of the weight matrix is calculated as a penalty term, thereby quantifying the sensitivity of the classifier to inertial distortion; The inertial sensitivity penalty term and the cross-entropy loss function are weighted and added to form the total loss function, thereby constraining the model's dependence on the essential waveform features and reducing overfitting to individual device differences.

6. The method of claim 1, wherein S6 The two-stage cyclic update strategy is used to optimize the coefficient prediction neural network and the backbone neural network, and the specific steps are as follows: 1) Phase one, update the coefficient prediction neural network parameters: Fix the backbone neural network parameters, perform morphological opening and closing operations on the original current waveform data to generate pseudo-clean waveforms, and then update the coefficient prediction neural network parameters by minimizing the L2 distance between the compensated waveforms and the pseudo-clean waveforms; 2) Phase two, update the backbone neural network parameters: Fix the coefficient prediction neural network parameters, and update the backbone neural network parameters based on the total loss function.

7. The method of claim 6, wherein the pattern recognition of the output waveform of the concha vagus nerve stimulation device is characterized by, In S7, the classification confidence is corrected based on the electrical properties of biological tissues, and the specific steps are as follows: 1) High-pass filter the residual feature vector, generate the tissue equivalent impedance spectrum vector through Fourier transform, reference current waveform normalization and tissue conductance characteristic parameter conversion, and then inverse the frequency-dependent impedance characteristics of biological tissues from the residual signal; 2) Calculate the exponential decay weighted mean based on the deviation of the real part of each frequency point of the impedance spectrum vector from the ideal impedance to obtain the modulation factor, and then quantify the stimulation efficiency attenuation caused by tissue impedance mismatch; 3) Convert the modulation factor into a logarithmic modulation vector, apply it to the classifier fully connected layer output, generate a corrected classification probability vector through a normalized exponential function, and then adjust the classification confidence based on the electrical properties of the tissue.

8. A vagus nerve stimulation device for the concha region, performing a pattern recognition method for the output waveform of the vagus nerve stimulation device for the concha region as described in any one of claims 1-7, characterized in that: The device comprises a base, one end of the base is provided with an antitragus fixing seat, and the other end is provided with a cymba concha fixing seat; an antitragus electrode patch is arranged on the antitragus fixing seat, and a cymba concha electrode patch is arranged on the cymba concha fixing seat.

Images