Method for establishing a cardiac defibrillation rhythm recognition model, device, medium

Abstract

The application discloses a heart defibrillation rhythm recognition model establishment method and application, equipment and a medium. The recognition model establishment method comprises the following steps: a deep convolutional neural network model structure is built, an initial convolution layer, a maximum pooling layer and a wavelet stack group layer are shared in a main path network and a secondary path network, the main path network further comprises a global average pooling layer and a full connection layer, and the secondary path network further comprises a decoding layer; the neural network model structure is trained and iterated, and corresponding data enhancement transformation is simulated and applied to an electrocardio impedance signal TTI and an electrocardio signal ECG in a first aid scene according to real interference physical characteristics. The heart defibrillation rhythm recognition model has excellent accuracy, super strong anti-interference ability and wide generalization in application, and the rescue efficiency is greatly improved, fast and accurate analysis ability is achieved, and thus the survival rate of patients is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cardiopulmonary resuscitation, in particular to a cardiac defibrillation rhythm recognition model establishment method and application thereof, a device and a medium. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute the prior art.

[0003] 85% to 90% of patients with cardiac arrest are ventricular fibrillation in the early stage, and the most effective method for treating ventricular fibrillation is to use an AED (Automated External Defibrillator) defibrillator as soon as possible. After the occurrence of ventricular fibrillation, an automated external defibrillator (AED) is a key device for rescuing patients with sudden cardiac death, and its core function is to automatically analyze the cardiac rhythm of the patient, determine whether it is a shockable rhythm (such as ventricular fibrillation-VF or pulseless ventricular tachycardia-VT), and guide the rescuer to perform defibrillation. Therefore, the accuracy and speed of the analysis of whether it is a shockable rhythm are directly related to the survival rate of the patient.

[0004] At present, the AED analysis shockable rhythm analysis algorithm of the AED device has at least the following technical problems in actual use:

[0005] 1. Known interference introduced by first aid device sensors and first aid measures: Unlike the cardiac rhythm analysis of ECG collected from general electrocardiogram monitoring devices, the ECG signal collected by the AED is often accompanied by noise introduced by other sensors of the device, such as electrode contact noise and implanted pacemaker pulses; in addition, first aid interventions will introduce noise during collection, such as muscle drift, artificial artifacts, etc. produced by chest compression during cardiopulmonary resuscitation (CPR).

[0006] 2. Random interference introduced by patients and the environment: Generally, the ECG signal collected by ordinary electrocardiogram monitoring devices needs to be measured under rest conditions, and repeated measurements are allowed in the case of signal quality interference. The AED rhythm analysis algorithm needs to adapt to the complex first aid environment, which may have a large number of unpredictable environmental disturbances. On the one hand, when the heart stops, the patient may have whole body convulsions or oxygen-related clinical symptoms; on the other hand, in order to perform necessary first aid, patient position adjustment, transportation and other operations may be performed. These disturbances are easily misidentified as electrocardiogram signals. The current AED analysis shockable rhythm analysis algorithm cannot effectively cope with these disturbances, and the anti-random interference ability is weak, which is prone to misjudgment.

[0007] 3. Balance of shockable rhythm discrimination speed and accuracy: AED rhythm analysis mainly discriminates the shockable rhythm in arrhythmia analysis. If the time taken by the algorithm is too long, it will cause interruption of CPR and affect the effect of first aid. The algorithm analysis usually needs to discriminate whether it is a shockable rhythm within 5-10 seconds. However, the current AED rhythm analysis algorithm cannot provide accurate discrimination of clinical level recognition accuracy within such a short time, or the accuracy is not enough, or the calculation time is too long.

[0008] Therefore, the traditional rhythm analysis algorithm does not enhance the feature expression of shockable arrhythmia mode, and in addition, it does not process the known interference introduced by first aid sensors and first aid measures, and the random interference introduced by patients and the environment in the first aid scene, causing the traditional algorithm to have a bottleneck in feature expression ability and complex interference mode learning in the first aid scene, and it is difficult to achieve clinical level recognition accuracy in a short time. Therefore, how to solve the above problems has become the subject to be studied and solved by the present application. SUMMARY

[0009] The purpose of the present application is to provide a method for establishing a cardiac defibrillation rhythm recognition model, its application, device and medium.

[0010] To achieve the above purpose, the first aspect of the present application provides a method for establishing a cardiac defibrillation rhythm recognition model, which comprises the following steps:

[0011] The deep convolutional neural network model structure is built, and the neural network model structure comprises a dual two-way network structure composed of a main path network and a secondary path network. The main path network and the secondary path network share an initial convolutional layer, a maximum pooling layer and a wavelet stack group layer. The main path network further comprises a global average pooling layer and a fully connected layer, and the secondary path network further comprises a decoding layer. A joint signal is composed of an ECG impedance signal TTI and a disturbed ECG signal ECG. The instantaneous frequency feature of the joint signal is captured and processed based on continuous wavelet transform, and the feature processing of adaptive different time and position scale is used to form a wavelet transform block. The wavelet stack group layer is stacked by a plurality of wavelet transform blocks;

[0012] The training and iteration of the neural network model structure, in view of the real interference physical characteristics of the first aid scene, synchronously apply corresponding data enhancement transformation to the electrocardio-impedance signal TTI and the electrocardio signal ECG, train the neural network model structure through the data-enhanced electrocardio-impedance signal TTI and the interfered electrocardio signal ECG, form the fusion of the main network and the secondary network for the adversarial training, perform classification and discrimination through the main network to determine whether to perform electric shock and calculate the discrimination loss function of the electric shockable heart rate, perform discrimination through the secondary network to generate the anti-interference electrocardio signal ECG and calculate the generation loss function of the anti-interference electrocardio signal, combine the discrimination loss function and the generation loss function to jointly build the joint adversarial loss function, generate the global optimization target iteration direction, retain the best model weight parameter after multiple iterations, and build the cardiac defibrillation rhythm recognition model through the best model weight parameter.

[0013] The second aspect of the present application proposes an application method of a cardiac defibrillation rhythm recognition model, which is an application of the cardiac defibrillation rhythm recognition model built by the recognition model building method of the first aspect of the present application. The application method of the recognition model comprises the following steps:

[0014] The deep convolutional neural network model structure is built, the neural network model structure comprises a dual two-way network structure composed of a main network and a secondary network, the initial convolutional layer, the maximum pooling layer and the wavelet stack group layer are shared in the main network and the secondary network, the global average pooling layer and the full connection layer are further included in the main network, and the decoding layer is further included in the secondary network; the joint signal is composed of the electrocardio-impedance signal TTI and the interfered electrocardio signal ECG, the wavelet transform block is composed of the instantaneous frequency feature of the joint signal captured and processed based on the continuous wavelet transform and the feature processing of different time, position and scale, and the wavelet stack group layer is stacked by a plurality of wavelet transform blocks;

[0015] The training and iteration of the neural network model structure, in view of the real interference physical characteristics of the first aid scene, synchronously apply corresponding data enhancement transformation to the electrocardio-impedance signal TTI and the electrocardio signal ECG, train the neural network model structure through the data-enhanced electrocardio-impedance signal TTI and the interfered electrocardio signal ECG, form the fusion of the main network and the secondary network for the adversarial training, perform classification and discrimination through the main network to determine whether to perform electric shock and calculate the discrimination loss function of the electric shockable heart rate, perform discrimination through the secondary network to generate the anti-interference electrocardio signal ECG and calculate the generation loss function of the anti-interference electrocardio signal, combine the discrimination loss function and the generation loss function to jointly build the joint adversarial loss function, generate the global optimization target iteration direction, retain the best model weight parameter after multiple iterations, and build the cardiac defibrillation rhythm recognition model through the best model weight parameter;

[0016] The deployment of the cardiac defibrillation rhythm recognition model, the model structure and parameters of the main path network in the cardiac defibrillation rhythm recognition model are extracted, the model is optimized in a light weight manner, is embedded into a processing chip of an AED device, and is converted into a light weight inference engine.

[0017] The use of the AED device deployed with the main path network model, the electrode pads of the AED device are applied to a body part of a patient, after an electrocardio signal ECG and an electrocardio impedance signal TTI are acquired, the identification of the shockable cardiac rhythm is performed by the main path discriminator, the electric shock is stopped when the non-shockable cardiac rhythm is identified, and the electric shock is performed when the shockable cardiac rhythm is identified, so that the misshock and the missed shock are avoided.

[0018] The third aspect of the present application provides an AED defibrillation device, and the model structure and parameters of the main path network which are extracted from the cardiac defibrillation rhythm recognition model according to the first aspect of the present application and are optimized in a light weight manner are embedded into a processing chip of the AED defibrillation device.

[0019] The fourth aspect of the present application provides a readable storage medium, and a control program is stored on the readable storage medium, and when the control program is executed by a detection module, the detection module executes the steps of the method according to the first aspect.

[0020] Alternatively, the model structure and parameters of the main path network which are extracted from the cardiac defibrillation rhythm recognition model according to the second aspect and are optimized in a light weight manner are stored on the readable storage medium, and the steps of the cardiac defibrillation rhythm recognition according to the second aspect are executed.

[0021] The relevant content of the present application is explained as follows:

[0022] 1. Through the implementation of the above technical solutions of the application, in view of the problem that the existing AED device cannot accurately identify various interference scenes during the process of applying electric shock defibrillation to a patient with ventricular fibrillation, resulting in missed electric shock, wrong electric shock, and other measures to rescue opportunities, endangering the life of the patient, and the like, a cardiac defibrillation rhythm recognition model establishment method, an application method of a cardiac defibrillation rhythm recognition model, an AED defibrillation device, and a readable storage medium are innovatively developed. In the cardiac defibrillation rhythm recognition model establishment method, in the step of building a deep convolutional neural network model structure, the ECG impedance signal TTI and the disturbed ECG signal ECG are jointly analyzed for the shockable mode representation of the cardiac rhythm, and two opposing networks are designed: the main network task is mainly to classify and distinguish whether electric shock is needed, and the secondary network is mainly to distinguish the interference and reconstruct the ECG signal without interference. The wavelet stack group layer in the dual two-way network structure composed of the main network and the secondary network processes the joint signal composed of the ECG impedance signal TTI and the disturbed ECG signal ECG; in the step of training and iteration of the neural network model structure, according to the real interference physical characteristics of the first-aid scene, the corresponding data enhancement transformation is applied to the ECG impedance signal TTI and the ECG signal ECG, and through training and iteration, on the one hand, the evaluation of the signal reconstruction ability of the secondary network is enhanced, and on the other hand, the discrimination ability of the main network for the shockable cardiac rhythm is iterated synchronously. The establishment of the cardiac defibrillation rhythm recognition model of the application makes the cardiac defibrillation rhythm recognition model have excellent accuracy, super strong anti-interference ability and wide generalization when applied, and the AED defibrillation device deploying the model structure and parameters of the main network extracted from the cardiac defibrillation rhythm recognition model and optimized, in which the rescue efficiency is greatly improved, the rapid and accurate analysis ability is improved, and the cardiac rhythm judgment is completed in a shorter interruption time, thereby increasing the effective CPR time and improving the survival rate of patients.

[0023] 2. In the first aspect of the above technical solution, in the neural network model structure, the initial convolutional layer, the maximum pooling layer, the wavelet stack unit, the global average pooling layer, and the fully connected layer are connected in stages to constitute a main discriminator of the main network, and the main discriminator is used to maximize the discrimination ability of the shockable heart rate. The discrimination loss function calculated by the main discriminator is based on the prediction probability of the shockable cardiac rhythm of the real data discrimination and the prediction probability of the anti-interference ECG signal data generated by the secondary network.

[0024] In the neural network model structure, the initial convolutional layer, the maximum pooling layer, the wavelet stacking unit, and the decoding layer are sequentially connected to form a secondary path generator of the secondary path network, the secondary path generator is used to maximize the misjudgment probability of the primary path discriminator for the generated anti-interference electrocardio signal data, and the generated loss function calculated by the secondary path generator is based on the reconstructed anti-interference electrocardio signal generated by the electrocardio impedance signal and the electrocardio signal data, and the prediction probability of the primary path discriminator for the anti-interference electrocardio signal data generated by the secondary path network.

[0025] The joint adversarial loss function is constructed by combining the discriminant loss function and the generation loss function in a weighted manner.

[0026] Through the further expansion of the composition and application of the primary path discriminator and the secondary path generator, the dual two-path network structure composed of the primary path network and the secondary path network shares the initial convolutional layer, the maximum pooling layer, and the wavelet stacking unit in actual application, and the discriminant loss function calculated by the primary path discriminator is based on the prediction probability of the shockable cardiac rhythm discriminated by the real data and the prediction probability of the anti-interference electrocardio signal data generated by the secondary path network, and the generation loss function calculated by the secondary path generator is based on the reconstructed anti-interference electrocardio signal generated by the electrocardio impedance signal and the electrocardio signal data, and the prediction probability of the primary path discriminator for the anti-interference electrocardio signal data generated by the secondary path network, so that the primary path network subsequently arranged in the AED defibrillation device can more accurately and efficiently identify the shockable cardiac rhythm.

[0027] 3. In the first aspect of the above technical solution, in the step of synchronously applying corresponding data enhancement transformation to the electrocardio impedance signal TTI and the electrocardio signal ECG, the electrocardio impedance signal TTI is used as direct physical information measurement of external interference signal, and the following data enhancement transformation is performed based on the real interference physical characteristics of the emergency scene:

[0028] CPR compression interference, periodically adding interference signals based on the real compression waveform model to the electrocardio signal ECG and the electrocardio impedance signal TTI;

[0029] Motion and vibration interference, simulating ambulance, helicopter, stretcher moving scenes, adding corresponding low-frequency vibration and high-frequency jitter noise;

[0030] Pacemaker pulse, adding pacemaker pulse signals with different amplitudes and frequencies to the electrocardio signal ECG;

[0031] Electrode contact noise, simulating baseline drift and high-frequency noise caused by poor electrode contact;

[0032] Signal quality variation, simulating changes in signal amplitude and morphology caused by difficult breathing, convulsions, or different skin conditions and electrode paste positions.

[0033] Through the above data enhancement transformation of multiple real interference physical characteristics, sufficient and comprehensive interference scenarios can be input during the training and iteration of the neural network model structure, and various rare emergency scenarios for different populations, different body types of patients, and various rare emergency scenarios can be simulated, so that the application value of the best model weight parameters retained after multiple iterations is higher, and the main road network laid in the AED defibrillation device in the subsequent layout can have more excellent accuracy, stronger anti-interference ability, and more extensive generalization.

[0034] 4. In the first aspect of the above technical solution, in the step of training the neural network model structure by the data-enhanced electrocardio impedance signal TTI and the disturbed electrocardio signal ECG, the model parameter update is performed in a small batch iteration manner during the training stage, and each training round includes the following steps:

[0035] reading a batch of samples;

[0036] forward propagation, classification and discrimination whether to perform electric shock and calculation of shockable heart rate discrimination loss function are performed through the main road network, and interference discrimination, reconstruction of anti-interference electrocardio signal ECG generation and calculation of anti-interference electrocardio signal generation loss function are performed through the secondary road network;

[0037] The joint adversarial loss function is constructed by combining the discrimination loss function and the generation loss function, and the global optimization objective function is generated to drive the model iteration direction;

[0038] The weight parameter update is performed through back propagation, the parameter gradient is calculated through the error back propagation algorithm, the network weight is adjusted through the adaptive optimization strategy, and the training round count and temporary cache cleaning are completed.

[0039] Through this training iteration method, the adaptive learning rate adjustment mechanism can effectively accelerate gradient convergence and reduce oscillation problems. The training loss curve, validation set indicators and self-supervised task performance are recorded in real time during the training process, which is used for later diagnosis of model training stability and task contribution, and provides a more efficient, accurate and over-fitting prevention training step for parameter update of the entire cardiac defibrillation rhythm recognition model.

[0040] 5. In the first aspect of the above technical solution, the main road network is composed of a global average pooling layer and a fully connected layer, and the main road network is composed of a main road discriminator. The main road discriminator maximizes the discrimination ability of the shockable heart rhythm, and the discrimination loss function of the main road discriminator is based on binary cross entropy, as follows:

[0041] ;

[0042] where D(x i ) represents the main road discriminator for real data x ia predicted probability of the shockable cardiac rhythm, D(G(z i ) represents a predicted probability of the main path discriminator for the clean ECG data G(z i ) generated by the secondary path network, y i represents a true data label;

[0043] The secondary path generator is constituted by the decoding layer of the secondary path network, and the secondary path generator maximizes the misjudgment probability of the main path discriminator for the generated anti-interference electrocardio signal ECG as a generation loss function, as follows:

[0044] ;

[0045] wherein, wherein G(zi) represents clean ECG data generated by the generator combined with the thoracic impedance signal TTI and ECG data z i , D(G(z i )) represents a predicted probability of the main path discriminator for the generated clean ECG data;

[0046] The joint adversarial loss function is constructed by combining the discriminant loss function and the generation loss function in a weighted manner:

[0047] ;

[0048] wherein and respectively represent predetermined coefficients of the discriminant loss function and the generation loss function.

[0049] Through the operation of the discriminant loss function of the above main path discriminator and the generation loss function of the secondary path generator, and the operation of constructing the joint adversarial loss function, the training and learning of the neural network model structure are strengthened, and the best model weight parameter of the cardiac defibrillation rhythm recognition model can be more accurate and effective.

[0050] 6. In the first aspect of the above technical solution, in the process of constructing the wavelet processing block based on the continuous wavelet transform to capture the instantaneous frequency characteristics of the joint signal and the adaptive characteristics of different time and position scales, the following steps are included:

[0051] The wavelet transform extracts the energy distribution of the signal at different scales and positions by performing an inner product operation on the signal and the wavelet basis function, and the time domain mathematical expression thereof is as follows: ;

[0052] wherein x(t) represents a time domain signal, t represents a current time, represents a wavelet function, and a and respectively represent a scaling factor and a translation factor of the wavelet function; the equivalent frequency domain expression of the wavelet transform block is as follows: ;

[0053] where x(ω) represents the frequency domain expression of the time domain signal;

[0054] Stacking the above wavelet transform blocks, for the stacked l layers of wavelet transform blocks, the extended form of wavelet transform convolution is:

[0055]

[0056] where h(j) represents the output value of the jth layer of stacked residual blocks, k represents the dimension of the hidden layer, and ω ij represents the connection weight of the hidden layer; x i represents the i-th element of the input signal x, a j and respectively represent the scaling factor and the translation factor of the jth layer of wavelet transform blocks, the ReLU activation function is used in the single layer of wavelet transform blocks, and the BatchNorm operation is used;

[0057] The calculation method of the ReLU activation function is:

[0058] The calculation method of the BatchNorm operation is:

[0059] where and respectively represent the mean and variance of a data batch, represents a very small constant to avoid division by zero.

[0060] Through the above improvement and application of wavelet transform, the instantaneous frequency characteristics of the ECG signal can be captured, such as the rapid change of QRS complex or the subtle fluctuation of P wave, and the time-frequency resolution is adaptive, the high-frequency part (such as QRS wave) has higher time resolution, and the low-frequency part (such as baseline drift) has higher frequency resolution, which is suitable for feature extraction of different time scales in ECG. Wavelet analysis can analyze the local time and frequency domain characteristics of the signal through the transformation of the wavelet basis function, and has the selective ability of signal direction in two-dimensional case, so it is particularly suitable for processing such electrocardio impedance signal TTI and disturbed electrocardio signal ECG, thereby providing a more specific core idea for the establishment of a cardiac defibrillation rhythm recognition model.

[0061] 7. In the first aspect of the above technical solution, the main road network comprises an initial convolution layer, a maximum pooling layer, a wavelet stacking group layer global average pooling layer, and a fully connected layer.

[0062] In the initial convolution layer, the input time series is first embedded through the initial convolution operation to perform preliminary feature extraction on the input signal, reduce the spatial resolution, and extract low-level features, as follows:​​​ ;

[0063] wherein X1 represents the initial feature, the input data wherein represents the number of input channels, D represents the dimension of the sample, and the overall process sequentially undergoes a convolution Conv operation with a convolution kernel k1 and a sliding step s1, a batch normalization BatchNorm operation, and a ReLU activation function operation;

[0064] In the max-pooling layer, the max-pooling layer is used to further reduce the spatial resolution of the feature map and reduce the calculation amount of the subsequent layer, as follows: ;

[0065] wherein X2 represents the output after max-pooling, and a max-pooling MaxPool operation with a convolution kernel k2 and a sliding step s2 is used;

[0066] In the global average pooling layer of the main path network, the global average pooling layer is used to compress the spatial dimension of the feature map and retain the channel information, as follows: ;

[0067] wherein the two-dimensional X L is flattened by the GAP operation, and the one-dimensional X G is outputted; G corresponds to the output of the GAP layer, and X L corresponds to the output of the wavelet stack group layer;

[0068] In the fully connected layer of the main path network, the fully connected layer is used to map the feature vector after the global average pooling layer to the target class, i.e., to judge the shockable cardiac rhythm classification class, as follows: ;

[0069] wherein Y represents the class output, and Linear represents the linear connection of the fully connected layer;

[0070] The secondary path network includes an initial convolution layer, a max-pooling layer, a wavelet stack group layer, and a decoding layer; wherein the initial convolution layer, the max-pooling layer, and the wavelet stack group layer are shared with the main path network;

[0071] In the decoding layer of the secondary path network, the decoding layer is used to use interpolation convolution to restore the feature signal to be consistent with the original electrocardiogram ECG input dimension, as follows: ;

[0072] wherein the interpolation algorithm Interpolation selects bilinear interpolation or nearest neighbor interpolation to enlarge the feature map, and then a convolution operation with a convolution kernel k3 and a sliding step s3 is used for result smoothing and to avoid the chessboard effect.

[0073] Through the more specific application of the dual two-way network structure, the main network is more prominent in maximizing the discrimination ability of the shockable rhythm, and the secondary network is more prominent in maximizing the ability of the discriminator to misjudge the generated ECG data, thereby making the application of the established joint adversarial loss function more accurate to generate a global optimization objective function to drive the model iteration direction from data enhancement to the beginning of the overall iteration process.

[0074] Compared with the prior art, the application has the following advantages and effects due to the use of the above scheme:

[0075] In view of the problem that the existing AED device cannot accurately identify various interference scenes during the process of applying electric shock defibrillation to a patient with ventricular fibrillation, resulting in missed electric shock, incorrect electric shock, and other measures to rescue opportunities, endangering the life of the patient, and the like, a cardiac defibrillation rhythm recognition model establishment method, an application method of a cardiac defibrillation rhythm recognition model, an AED defibrillation device, and a readable storage medium are innovatively developed. In the cardiac defibrillation rhythm recognition model establishment method, in the step of building a deep convolutional neural network model structure, the ECG impedance signal TTI and the disturbed ECG signal ECG are jointly analyzed for the representation of the shockable mode of the cardiac rhythm, and two adversarial networks are designed: the main network task is mainly to classify and determine whether electric shock is needed, and the secondary network is mainly to determine the interference and to reconstruct the ECG signal without interference. The wavelet stack group layer in the dual two-way network structure formed by the main network and the secondary network processes the combined signal composed of the ECG impedance signal TTI and the disturbed ECG signal ECG; in the step of training and iteration of the neural network model structure, the corresponding data enhancement transformation is applied to the ECG impedance signal TTI and the ECG signal ECG according to the real interference physical characteristics of the first-aid scene, and through training and iteration, on the one hand, the evaluation of the signal reconstruction ability of the secondary network is enhanced, and on the other hand, the discrimination ability of the main network for the shockable cardiac rhythm is iterated synchronously. The establishment of the cardiac defibrillation rhythm recognition model by the application makes the cardiac defibrillation rhythm recognition model have excellent accuracy, super strong anti-interference ability, and wide generalization when applied, and the AED defibrillation device deploying the model structure and parameters of the main network extracted from the cardiac defibrillation rhythm recognition model and optimized, which greatly improves the rescue efficiency, has fast and accurate analysis ability, and allows the cardiac rhythm to be determined in a shorter interruption time, thereby increasing the effective CPR time and improving the survival rate of patients. Specifically:

[0076] 1. Excellent accuracy: through large-scale clinical backtracking test verification, the algorithm has a recognition sensitivity of 96.9% and a specificity of 99.7% for shockable cardiac rhythm within a 5-second analysis window, and the performance index reaches the leading level in the industry.

[0077] 2. Strong anti-interference ability: can effectively suppress various strong interferences including manual / mechanical CPR compression, vehicle vibration, patient convulsion, pacemaker pulse, etc., greatly reducing the risk of false shock and missed shock;

[0078] 3. Wide generalization: based on global, multi-ethnic large-scale real data and unique data enhancement technology training, it shows stable high precision for different populations, patients of different body types and various rare emergency scenes.

[0079] 4. Improve rescue efficiency: fast and accurate analysis capability allows rhythm judgment to be completed in a shorter interruption time, thereby increasing effective CPR time and improving patient survival rate. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 The overall framework schematic diagram of the deep convolutional neural network model structure of the embodiment of the application is shown in the figure.

[0081] Figure 2 The schematic diagram of the dual two-way network structure in the embodiment of the application is shown in the figure.

[0082] Figure 3 The schematic diagram of the shockable and non-shockable rhythm of the ECG signal and the TTI signal when subjected to CPR interference is shown in the figure.

[0083] Figure 4 The schematic diagram of the simulation and addition of various interference signals in the data enhancement technology of the application is shown in the figure (one).

[0084] Figure 5 The schematic diagram of the simulation and addition of various interference signals in the data enhancement technology of the application is shown in the figure (two).

[0085] Figure 6 The specific implementation flow of the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0086] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below. In the following description, a lot of specific details are set forth in order to fully understand the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without departing from the spirit of the present application, therefore the present application is not limited by the specific embodiments disclosed below.

[0087] The terms used herein are only for describing specific embodiments and are not intended to limit the present application. The singular forms such as "a", "this", "this", "this" and "the" as used herein also include the plural forms.

[0088] As used herein, the terms "first", "second", etc. do not particularly refer to order or sequence, nor are they used to limit the present application, but are merely used to distinguish components or operations described by the same technical terms.

[0089] As used herein, the terms "connected" or "positioned" can refer to two or more components or devices being directly in physical contact with each other, or indirectly in physical contact with each other, or can refer to two or more components or devices being in operative or communicative contact with each other.

[0090] As used herein, the terms "connected", "including", "having", etc. are all open terms, i.e. meaning including but not limited to.

[0091] As used herein, the terms "connected", "including", "having", etc. are all open terms, i.e. meaning including but not limited to.

[0092] The present application aims to provide a heart defibrillation rhythm recognition model establishment method, an application method of a heart defibrillation rhythm recognition model, an AED defibrillation device and a readable storage medium, which can accurately identify various interference scenes during the process of applying electric shock defibrillation to a patient with ventricular fibrillation using existing AED equipment, thereby avoiding missed electric shock, incorrect electric shock and other measures to rescue opportunities, endangering the life of the patient, etc. The heart defibrillation rhythm recognition model has excellent accuracy, super strong anti-interference ability and wide generalization when applied, and the AED defibrillation device deployed from the model structure and parameters of the main road network extracted from the heart defibrillation rhythm recognition model and optimized, which greatly improves the rescue efficiency, has fast and accurate analysis ability, allows the heart rhythm to be judged in a shorter interruption time, thereby increasing the effective CPR time and improving the survival rate of patients.

[0093] In embodiment one, the first aspect of the present application proposes a heart defibrillation rhythm recognition model establishment method, which comprises the following steps:

[0094] The deep convolutional neural network model structure is built, and the neural network model structure comprises a dual two-way network structure composed of a main road network and a secondary road network (see Figure 2), the main path network and the secondary path network share the initial convolutional layer, the maximum pooling layer and the wavelet stack group layer, the main path network further comprises a global average pooling layer and a fully connected layer, and the secondary path network further comprises a decoding layer; the joint signal is composed of the ECG impedance signal TTI and the disturbed ECG signal ECG, the wavelet transform block is composed of the instantaneous frequency feature of the joint signal captured based on the continuous wavelet transform and the feature processing of different time and position scales, and the wavelet stack group layer is stacked by a plurality of wavelet transform blocks;

[0095] The training and iteration of the neural network model structure, the corresponding data enhancement transformation is simulated and applied to the ECG impedance signal TTI and the ECG signal ECG in synchronization according to the real interference physical characteristics of the emergency scene, the neural network model structure is trained by the ECG impedance signal TTI and the disturbed ECG signal ECG after data enhancement, and the fusion of the main path network and the secondary path network is formed through the adversarial training; whether the shock needs to be performed is determined by the main path network, and the discrimination loss function of the shockable rhythm is calculated, the interference is discriminated by the secondary path network, the anti-interference ECG signal ECG is generated, and the generation loss function of the anti-interference ECG signal is calculated, the joint adversarial loss function is constructed by combining the discrimination loss function and the generation loss function, and the global optimization target iteration direction is generated, and the best model weight parameter is reserved after multiple iterations, and the cardiac defibrillation rhythm recognition model is constructed by the best model weight parameter.

[0096] In the cardiac defibrillation rhythm recognition model establishment method, reference is made to Figure 1 The overall framework of the cardiac defibrillation rhythm recognition model is shown, in the step of building the deep convolutional neural network model structure, the ECG impedance signal TTI and the disturbed ECG signal ECG are jointly analyzed for the shockable rhythm mode representation, and two adversarial networks are designed: the main path network mainly performs classification and discrimination to determine whether the shock needs to be performed, and the secondary path network mainly discriminates the interference and focuses on reconstructing the ECG signal without interference, and the wavelet stack group layer in the dual two-path network structure formed by the main path network and the secondary path network processes the joint signal composed of the ECG impedance signal TTI and the disturbed ECG signal ECG; in the step of training and iteration of the neural network model structure, the corresponding data enhancement transformation is simulated and applied to the ECG impedance signal TTI and the ECG signal ECG in synchronization according to the real interference physical characteristics of the emergency scene, through training and iteration, on the one hand, the evaluation of the signal reconstruction ability of the secondary path network is enhanced, and on the other hand, the discrimination ability of the main path network to the shockable rhythm is iterated in synchronization. Thus, the cardiac defibrillation rhythm recognition model has more practical value.

[0097] In one of the preferred modes of the embodiment one of the present application, in the neural network model structure, the initial convolutional layer, the maximum pooling layer, the wavelet stacking unit, the global average pooling layer, and the fully connected layer are connected in stages to form a main path discriminator of the main path network, the main path discriminator is used to maximize the discrimination ability of the shockable heart rate, and the discrimination loss function calculated by the main path discriminator is based on the prediction probability of the real data discrimination of the shockable heart rate and the prediction probability of the anti-interference electrocardio signal data generated by the secondary path network;

[0098] In the neural network model structure, the initial convolutional layer, the maximum pooling layer, the wavelet stacking unit, and the decoding layer are connected in stages to form a secondary path generator of the secondary path network, the secondary path generator is used to maximize the misjudgment probability of the generated anti-interference electrocardio signal data by the main path discriminator, and the generation loss function calculated by the secondary path generator is based on the reconstructed anti-interference electrocardio signal generated by the joint electrocardio impedance signal and electrocardio signal data, and the prediction probability of the anti-interference electrocardio signal data generated by the secondary path network by the main path discriminator;

[0099] The discrimination loss function and the generation loss function are combined using a weighted method to jointly construct a joint adversarial loss function.

[0100] Through the above further expansion of the composition and application of the main path discriminator and the secondary path generator, the dual two-path network structure composed of the main path network and the secondary path network shares the initial convolutional layer, the maximum pooling layer, and the wavelet stacking unit in actual application, and the discrimination loss function calculated by the main path discriminator is based on the prediction probability of the real data discrimination of the shockable heart rate and the prediction probability of the anti-interference electrocardio signal data generated by the secondary path network, and the generation loss function calculated by the secondary path generator is based on the reconstructed anti-interference electrocardio signal generated by the joint electrocardio impedance signal and electrocardio signal data, and the prediction probability of the anti-interference electrocardio signal data generated by the secondary path network by the main path discriminator, so that the main path network subsequently laid in the AED defibrillation device can more accurately and efficiently identify the shockable heart rate.

[0101] In another preferred mode of the embodiment one of the present application, in the step of synchronously applying corresponding data enhancement transformation to the electrocardio impedance signal TTI and the electrocardio signal ECG, the electrocardio impedance signal TTI is used as direct physical information measurement of external interference signal, and based on the real interference physical characteristics of the first-aid scene, the following data enhancement transformation is performed:

[0102] CPR compression interference, periodically adding a real compression waveform model based interference signal to the electrocardio signal ECG and the electrocardio impedance signal TTI; (ECG signal and TTI signal under CPR interference Shockable heart rhythm and non-shockable heart rhythm Figure 3 )

[0103] Motion and vibration interference, simulate ambulance, helicopter, stretcher moving scene, add the corresponding low-frequency vibration and high-frequency jitter noise;

[0104] Pacemaker pulse, add a pacing pulse signal with different amplitudes and frequencies in the ECG signal;

[0105] Electrode contact noise, simulate the baseline drift and high-frequency noise caused by poor electrode contact;

[0106] Signal quality variation, simulate the signal amplitude and morphology changes caused by difficult breathing, convulsions, or different skin conditions, electrode placement.

[0107] Through the above data enhancement transformation of multiple real interference physical characteristics, enough and comprehensive interference scenarios can be input in the training and iteration process of the neural network model structure, and various rare emergency scenes for different populations, different body types of patients can be simulated, so that the application value of the best model weight parameter retained after multiple iterations is higher, and the main road network of the subsequent layout to the AED defibrillation device can have more excellent accuracy, stronger anti-interference ability, and more extensive generalization. The simulation of adding various interference signals in the data enhancement technology is shown in the following Figure 4 and Figure 5 .

[0108] Specifically, after data enhancement, the model parameter update is performed in the training stage in a small batch iteration manner. The optimizer is Adam, and its adaptive learning rate adjustment mechanism can effectively accelerate gradient convergence and reduce oscillation problems. Each round of training includes the following steps: reading a batch of samples; forward propagation to generate two paths (predictable values of shockable arrhythmia, clean ECG signals after ECG reconstruction); respectively, the loss of the main road network and the loss of the secondary road network and the joint loss The early stopping mechanism is used to monitor the average loss value on the validation set. If the validation performance does not improve significantly for several consecutive rounds, such as exceeding the set threshold, the training is automatically stopped to prevent overfitting. In addition, the training loss curve, validation set indicators and self-supervised task performance are recorded in real time during the training process for later diagnosis of model training stability and contribution of each task.

[0109] In another preferred mode of the embodiment one of the application, in the step of training the neural network model structure from the data-enhanced ECG impedance signal TTI and the disturbed ECG signal ECG, the model parameter update is performed in the training stage in a small batch iteration manner, and each round of training includes the following steps:

[0110] Reading a batch of samples;

[0111] The forward propagation is respectively performed through the main path network to determine whether to perform electric shock and to calculate a determination loss function of a shockable heart rate, and through the secondary path network to determine interference, reconstruct an anti-interference electrocardio signal ECG and calculate a generation loss function of the anti-interference electrocardio signal;

[0112] The determination loss function and the generation loss function are combined to construct a joint adversarial loss function, and a global optimization target function is generated to drive the model iteration direction;

[0113] The backward propagation is performed and the weight parameter is updated, an error back propagation algorithm is used to calculate the parameter gradient, an adaptive optimization strategy is applied to adjust the network weight, and the training round count and temporary cache cleaning are completed.

[0114] By using the training iteration mode, the adaptive learning rate adjustment mechanism can effectively accelerate the gradient convergence and reduce the oscillation problem. The training loss curve, the validation set index and the self-supervised task performance are recorded in real time during the training process, which is used for the stability of the diagnosis model training and the contribution of each task in the later period, so as to provide a more efficient, accurate and over-fitting prevention training step for the parameter update of the whole cardiac defibrillation rhythm recognition model.

[0115] In another preferred mode of the embodiment one of the application, the global average pooling layer and the full connection layer of the main path network constitute the main path determinator, the main path determinator maximizes the discrimination ability of the shockable heart rhythm, and the determination loss function of the main path determinator is based on the binary cross entropy, as follows:

[0116] ;

[0117] Wherein D(x i ) represents the prediction probability of the shockable heart rhythm determined by the main path determinator for the real data x i , D(G(z i )) represents the prediction probability of the main path determinator for the clean ECG data G(z i ) generated by the secondary path network, and y i represents the real data label;

[0118] The decoding layer of the secondary path network constitutes the secondary path generator, the secondary path generator maximizes the misjudgment probability of the main path determinator for the generated anti-interference electrocardio signal ECG as the generation loss function, as follows:

[0119] ;

[0120] Wherein, wherein G(zi) represents the clean ECG data generated by the generator through the thoracic impedance signal TTI joint ECG data z i , and D(G(z i )) represents the prediction probability of the main path determinator for the generated clean ECG data.

[0121] The joint adversarial loss function is constructed by combining the discriminative loss function and the generative loss function using a weighted manner:

[0122] ;

[0123] Wherein and respectively represent predetermined coefficients of the discriminative loss function and the generative loss function.

[0124] Through the operation of the discriminative loss function of the above primary road judge and the operation of the generative loss function of the secondary road generator, and the operation of constructing the joint adversarial loss function, the training and learning of the neural network model structure are strengthened, and the best model weight parameter of the cardiac defibrillation rhythm recognition model can be more accurate and effective.

[0125] In one preferred mode of the embodiment one of the application, in the process of constructing the wavelet processing block based on the continuous wavelet transform to capture the instantaneous frequency characteristics of the joint signal and the adaptive characteristics of different time and position scales, the following steps are included:

[0126] The wavelet transform extracts the energy distribution of the signal at different scales and positions by performing an inner product operation between the signal and the wavelet basis function, and the time domain mathematical expression thereof is: ;

[0127] Wherein x(t) represents a time domain signal, t represents a current time, represents a wavelet function, and respectively represent a scaling factor and a translation factor of the wavelet function; and the equivalent frequency domain expression of the wavelet transform block is: ;

[0128] Wherein x(ω) represents a frequency domain expression of the time domain signal.

[0129] Stack the above wavelet transform block, and for a wavelet transform block stacked by l layers, the extended form of the wavelet transform convolution is:

[0130] ;

[0131] Wherein h(j) represents an output value of the jth layer stacked residual block, k represents the dimension of the hidden layer, and ij represents the connection weight value of the hidden layer; x i represents the ith element of the input signal x, a j and respectively represent the scaling factor and the translation factor of the jth layer wavelet transform block, the ReLU activation function is used in the single layer wavelet transform block, and the BatchNorm operation is used.

[0132] The calculation method of ReLU activation function is as follows: ;

[0133] The calculation method of BatchNorm operation is as follows: ;

[0134] Wherein and respectively represent the mean and variance of a data batch, represents a very small constant to avoid division by zero (such as 0.00001).

[0135] Through the improvement and application of wavelet transform above, the instantaneous frequency characteristics of ECG signal can be captured, such as the rapid change of QRS complex or the subtle fluctuation of P wave, and the time-frequency resolution is adaptive, the high-frequency part (such as QRS wave) has higher time resolution, and the low-frequency part (such as baseline drift) has higher frequency resolution, which is suitable for feature extraction of different time scales in ECG. Wavelet analysis can analyze the local time and frequency domain characteristics of the signal through the transformation of wavelet basis function, and has the selective ability of signal direction in two-dimensional case, so it is particularly suitable for processing such ECG impedance signal TTI and disturbed ECG signal, thereby providing a more specific core idea for the establishment of a cardiac defibrillation rhythm recognition model.

[0136] More specifically, the gradient update method for solving the back propagation of the wavelet block is as follows:

[0137] ;

[0138] ;

[0139] ;

[0140] Wherein, represents the number of rounds of update iteration, is the update gradient of the wavelet block coefficient, as follows.

[0141] ;

[0142] ;

[0143] ;

[0144] Wherein represents the set learning rate.

[0145] In another preferred mode of the embodiment one of the application, the main network comprises an initial convolutional layer, a maximum pooling layer, a wavelet stack group layer global average pooling layer and a fully connected layer.

[0146] In the initial convolution layer, the input time series is first embedded through the initial convolution operation to preliminarily extract features of the input signal, reduce the spatial resolution and extract low-level features, as follows: ;

[0147] wherein X1 represents the preliminary features, and the input data wherein represents the number of input channels, and D represents the dimension of the sample, and the whole process sequentially passes through the convolution Conv operation with the convolution kernel k1 and the sliding step s1, the batch normalization BatchNorm operation, and the ReLU activation function operation;

[0148] In the max-pooling layer, the max-pooling layer is used to further reduce the spatial resolution of the feature map and reduce the calculation amount of the subsequent layer, as follows: ;

[0149] wherein X2 represents the output after the max-pooling, and the max-pooling MaxPool operation with the convolution kernel k2 and the sliding step s2 is used;

[0150] In the global average pooling layer of the main path network, the global average pooling layer is used to compress the spatial dimension of the feature map and retain the channel information, as follows: ;

[0151] wherein the two-dimensional X L is flattened by the GAP operation, and the one-dimensional X G is outputted; G corresponds to the output of the GAP layer, and X L corresponds to the output of the wavelet stack group layer;

[0152] In the fully connected layer of the main path network, the fully connected layer is used to map the feature vector after the global average pooling layer to the target class, i.e., to judge the shockable arrhythmia classification class, as follows: ;

[0153] wherein Y represents the class output, and Linear represents the linear connection of the fully connected layer;

[0154] The secondary path network comprises an initial convolution layer, a max-pooling layer, a wavelet stack group layer, and a decoding layer; wherein the initial convolution layer, the max-pooling layer, and the wavelet stack group layer are shared with the main path network;

[0155] In the decoding layer of the secondary path network, the decoding layer is used to restore the feature signal to the same dimension as the original ECG input signal through the interpolation convolution, as follows: ;

[0156] Wherein, the Interpolation interpolation algorithm selects bilinear interpolation or nearest neighbor interpolation to enlarge the feature map, and then uses a convolution kernel with k3 and a sliding step s3 to perform convolution operation to smooth the result and avoid the chessboard effect.

[0157] In the second embodiment, the application provides a method for applying the cardiac defibrillation rhythm recognition model.

[0158] The deep convolutional neural network model structure is built, and the neural network model structure includes a dual two-way network structure composed of a main network and a secondary network.

[0159] The neural network model structure is trained and iterated, and the real interference physical characteristics of the emergency scene are simulated to apply corresponding data enhancement transformation to the electrocardio impedance signal TTI and the disturbed electrocardio signal ECG.

[0160] The cardiac defibrillation rhythm recognition model is deployed, the model structure and parameters of the main network in the cardiac defibrillation rhythm recognition model are extracted, the model is optimized, and the optimized model is embedded into the processing chip of the AED device to convert the model into a lightweight inference engine.

[0161] The AED device with the main network model is used, the electrode pads of the AED device are applied to the patient's body part, the electrocardio signal ECG and the electrocardio impedance signal TTI are obtained, and the main discriminator is used to identify the shockable cardiac rhythm.

[0162] In the third embodiment, the AED defibrillation device has a processing chip, and a model structure and parameters of a main path network extracted from the cardiac defibrillation rhythm recognition model in the first embodiment are embedded in the processing chip.

[0163] In the fourth embodiment, a readable storage medium has a control program stored thereon, and the control program, when executed by a detection module, causes the detection module to perform the steps of the method in the first embodiment.

[0164] Alternatively, the readable storage medium has a model structure and parameters of a main path network extracted from the cardiac defibrillation rhythm recognition model in the second embodiment and stored thereon, and the steps of the cardiac defibrillation rhythm recognition in the second embodiment are executed.

[0165] Next, the training and deployment of the cardiac defibrillation rhythm recognition model in two different stages are described as shown in FIG. 1. Figure 5

[0166] Step 1, initialization: the model structures of the main path network and the secondary path network are initialized, including the time steps and input dimensions of the ECG and TTI signals, the dimensions of each layer of the network, the loss functions of the main path network and the secondary path network, the adversarial loss function, and the like; the model training hyperparameters are initialized, including the loss function weighting coefficient, the maximum number of training rounds, and the learning rate, and the like.

[0167] Step 2, signal preprocessing: the ECG and TTI signals are subjected to necessary preprocessing methods, including interpolation to fill in missing points of the device signal, band-pass filtering to remove baseline drift and high-frequency noise, dynamic window standardization method to eliminate baseline offset, heart rhythm cycle recognition, and construction of an ECG and TTI signal segment of a fixed cycle as an input signal of the model (three heart rhythm cycles as a sample).

[0168] Step 3, model training: using the prepared large-scale labeled data set, the data enhancement technology is applied to train the dual-path network in an end-to-end manner.

[0169] Step 4, calculation of the adversarial loss of the model: the main path network predicts the discriminant analysis of the shockable rhythm, and maximizes the accuracy of the discrimination; the secondary path model generates an anti-noise ECG signal, and the adversarial generation loss jointly maximizes the discrimination accuracy and maximizes the generation of clean anti-noise ECG signals.

[0170] ​Step 5, model's adversarial training loop iteration: joint discriminative loss and production loss, generate global optimization objective function to drive model iteration direction, from data augmentation to the whole iteration process. Weight update: perform error backpropagation algorithm to calculate parameter gradient, apply adaptive optimization strategy to adjust network weight, complete training round count and temporary cache cleaning. The iteration process of the model can be optimized using technical means including early stopping judgment, for example, after each training round, the system calculates the average loss of the current model on the validation set; if the performance does not improve for several consecutive rounds, the early stopping mechanism is triggered to terminate training early to avoid overfitting.

[0171] Step 6, model test evaluation: the model outputs the prediction results on the independent test set, and the discriminative results of the main network for the shockable heart rhythm are used as the evaluation indicators of the final model. Including calculating Accuracy, AUC and F1, etc. Classification indicators, quantifying model performance and visualizing analysis.

[0172] Model deployment and application stage:

[0173] Step 7, deployment system of the main network: after training, save the best model weight file, extract the model structure and parameters of the main network, optimize the model, embed it into the processing chip of the AED device, and convert it into a lightweight inference engine. The device collects signals in real time, inputs the preprocessed signals into the model, and outputs the classification results, and the device decides whether to recommend defibrillation based on the results;

[0174] Step 8, weight saving of the whole network: integrate the iWaveNet full model containing adversarial training into the medical cloud platform, keep the complete model, and use it for incremental training and metabolic feature adaptation package distribution.

[0175] To evaluate the application performance of the heart defibrillation rhythm recognition model of the application, the applicant conducted a comprehensive test based on a large-scale, multi-source clinical electrocardiogram database. The database is constructed in accordance with the AHA guidelines, covering various shockable and non-shockable heart rhythms and intermediate heart rhythms, and containing various interference scenarios to ensure the rigor and comprehensiveness of the test.

[0176] The test database integrates real emergency data (PDB) collected by Pumian (Jiangsu) Medical Technology Co., Ltd. from globally recognized authoritative electrocardiogram databases (including VFDB, CUDB, MITDB, etc.), a total of 13 databases from different sources. The database contains a total of 1913 rigorously annotated clinical electrocardiogram rhythm segments. The determination and annotation of heart rhythm types follow the AHA guidelines, and the specific classification and data volume are as follows:

[0177] Shockable rhythm (433 cases): including coarse ventricular fibrillation (356 cases) and ventricular tachycardia (77 cases);

[0178] Non-shockable rhythms (1391 cases): including normal sinus rhythm (824 cases), asystole (108 cases), and other complex rhythms (such as supraventricular tachycardia, atrial fibrillation / flutter, pacemaker rhythm, etc., a total of 459 cases);

[0179] Intermediate rhythms (89 cases): including fine ventricular fibrillation (58 cases) and other ventricular tachycardia (31 cases).

[0180] The core indicators of performance evaluation are sensitivity (the ability to identify shockable rhythms) and specificity (the ability to avoid misshocking non-shockable rhythms), as shown in Table 1 below.

[0181]

[0182] Table 1

[0183] The test results show that the application performance of the heart defibrillation rhythm recognition model of the present application is excellent and stable in all rhythm categories, and is far superior to the AHA guideline target.

[0184] The above test results fully verify the advancement and reliability of the algorithm of the present application. The test performance on a large-scale, diversified adult clinical database not only fully meets the performance requirements in the AHA guideline, but also reaches an industry-leading level in key indicators. This proves that the network structure and data enhancement training technology described in the present application can effectively improve the electrocardiogram rhythm analysis capability of AED in real-world complex scenarios, providing a solid technical guarantee for rescuing lives, thus achieving the purpose of the present application.

[0185] The above examples are only for illustrating the technical concepts and characteristics of the present application, and the purpose is to enable those skilled in the art to understand the content of the present application and implement it, and cannot limit the protection scope of the present application. Any equivalent changes or modifications made in accordance with the spirit and essence of the present application should be covered within the protection scope of the present application.

Claims

1. A method for establishing a cardiac defibrillation rhythm recognition model, characterized in that, The recognition model building method includes the following steps: The deep convolutional neural network model structure is constructed, which includes a dual dual-path network structure consisting of a main path network and a secondary path network. The main path network and the secondary path network share an initial convolutional layer, a max pooling layer, and a wavelet stacking group layer. The main path network also includes a global average pooling layer and a fully connected layer, while the secondary path network includes a decoding layer. The joint signal is composed of the transelectrocardiogram (TTI) signal and the interfered electrocardiogram (ECG) signal. Based on continuous wavelet transform, the instantaneous frequency characteristics of the joint signal are captured and processed, and the wavelet transform blocks are constructed by adaptive feature processing at different time and location scales. The wavelet stacking group layer is composed of multiple wavelet transform blocks stacked together. The training and iteration of the neural network model structure, targeting the real interference physical characteristics of emergency scenarios, simultaneously applies corresponding data augmentation transformations to the cardiac impedance signal TTI and the electrocardiogram signal ECG. The neural network model structure is trained by the data-augmented cardiac impedance signal TTI and the interfered electrocardiogram signal ECG, forming a fusion of adversarial training between the main network and the secondary network. The main network classifies and determines whether defibrillation is needed and calculates the discriminant loss function for defibrillable rhythms. The secondary network identifies interference, reconstructs the generation of anti-interference ECG signals, and calculates the generation loss function for anti-interference ECG signals. The discriminant loss function and the generation loss function are combined to construct a joint adversarial loss function, which is used to generate the global optimization target iteration direction. After multiple iterations, the optimal model weight parameters are retained, and the cardiac defibrillation rhythm recognition model is constructed from these optimal model weight parameters. In the step of simultaneously applying corresponding data augmentation transformations to the cardiac impedance signal TTI and the electrocardiogram signal ECG, the cardiac impedance signal TTI is used as a direct physical information measure of external interference signals, and based on the real physical characteristics of interference in emergency scenarios, the following data augmentation transformations are performed: CPR compression interference involves synchronously adding periodic interference signals based on a real compression waveform model to the electrocardiogram (ECG) signal and the transcranial impedance signal (TTI). Motion and vibration interference are used to simulate the movement of ambulances, helicopters, and stretchers, with corresponding low-frequency vibration and high-frequency jitter noise added. Pacemaker pulses are pacing pulse signals with different amplitudes and frequencies added to the electrocardiogram (ECG) signal. Electrode contact noise, simulating baseline drift and high-frequency noise caused by poor electrode contact; Signal quality variations simulate changes in signal amplitude and morphology caused by patient breathing difficulties, convulsions, or different skin conditions and electrode placement. In the training process of the neural network model structure using data-enhanced trans-tachycardia (TTI) and interfered electrocardiogram (ECG) signals, the training phase employs a mini-batch iterative approach to update the model parameters. Each training round includes the following steps: Read a batch of samples; Forward propagation involves classifying and determining whether an electric shock is necessary via the main network and calculating the discriminant loss function for shockable heart rates. The secondary network then identifies interference, reconstructs the generation of the interference-resistant ECG signal, and calculates the generation loss function for the interference-resistant ECG signal. A joint adversarial loss function is constructed by combining the discriminative loss function and the generative loss function, and a global optimization objective function is generated to drive the model iteration direction. Backpropagation is performed and weight parameters are updated. The backpropagation algorithm is executed to calculate the parameter gradient. An adaptive optimization strategy is applied to adjust the network weights. Training round count and temporary cache cleanup are completed. The main path discriminator consists of a global average pooling layer and a fully connected layer in the main path network. This discriminator maximizes the ability to distinguish shockable heart rhythms. The discriminative loss function of the main path discriminator is based on binary cross-entropy, as follows: ； Where D(x) i This indicates that the main path discriminator correctly identifies the real data x. i The predictive probability of a shockable rhythm, D(G(z) i )) indicates that the primary path discriminator considers the clean ECG data G(z) generated by the secondary path network. i The predicted probability of y i Indicates real data labels; The decoding layer of the secondary network constitutes the secondary generator. The secondary generator maximizes the misclassification probability of the main discriminator for the generated anti-interference ECG signal as the generation loss function, as follows: ； Wherein, G(zi) represents the generator's trans-thoracic impedance signal TTI combined with ECG data z. i The generated clean ECG data, D(G(z) i )) represents the prediction probability of the main path discriminator for the generated clean ECG data; A joint adversarial loss function is constructed by combining the discriminative loss function and the generative loss function using a weighted approach. ； in and These represent the predetermined coefficients of the discriminant loss function and the generator loss function, respectively.

2. The method for establishing a cardiac defibrillation rhythm recognition model according to claim 1, characterized in that: In the neural network model structure, the initial convolutional layer, the max pooling layer, the wavelet stacking unit, the global average pooling layer, and the fully connected layer are connected in sequence to form the main path discriminator of the main path network. The main path discriminator is used to maximize the ability to distinguish shockable heart rates. The discriminant loss function calculated by the main path discriminator is based on the prediction probability of shockable heart rhythms discriminated from real data and the prediction probability of anti-interference ECG signal data generated by the secondary path network. In the neural network model structure, the initial convolutional layer, the max pooling layer, the wavelet stacking unit, and the decoding layer are connected in sequence to form the secondary path generator of the secondary path network. The secondary path generator is used to maximize the misjudgment probability of the main path discriminator for the generated anti-interference ECG signal data. The generation loss function calculated by the secondary path generator is based on the reconstructed anti-interference ECG signal generated by the ECG signal data combined with the cardiac impedance signal, and the prediction probability of the main path discriminator for the anti-interference ECG signal data generated by the secondary path network. A joint adversarial loss function is constructed by combining the discriminative loss function and the generative loss function using a weighted approach.

3. The method for establishing a cardiac defibrillation rhythm recognition model according to claim 1 or 2, characterized in that, The process of constructing wavelet processing blocks by capturing and processing the instantaneous frequency features of the joint signal based on continuous wavelet transform and adaptive feature processing at different time and location scales includes the following steps: Wavelet transform extracts the energy distribution of a signal at different scales and locations by performing an inner product operation between the signal and wavelet basis functions. Its time-domain mathematical expression is: ; Where x(t) represents the time-domain signal, and t represents the current time. Denotes the wavelet function, α and Let represent the scaling factor and translation factor of the wavelet function, respectively; the equivalent frequency domain expression of the wavelet transform block is: ; Where x(ω) represents the frequency domain representation of the time-domain signal; Stacking the above wavelet transform blocks, for a wavelet transform block with l layers stacked, the expanded form of its wavelet transform convolution is: ； Where h(j) represents the output value of the j-th stacked residual block, k represents the dimension of the hidden layer, and ω ij Indicates the connection weights of the hidden layer; x i Let a represent the i-th element of the input signal x. j and These represent the scaling factor and translation factor of the j-th wavelet transform block, the ReLU activation function used in a single wavelet transform block, and the BatchNorm operation. The ReLU activation function is calculated as follows: ; The BatchNorm operation is calculated as follows: ; in and Let these represent the mean and variance of a data batch, respectively. This represents a very small constant that is avoided by division by zero.

4. The method for establishing a cardiac defibrillation rhythm recognition model according to claim 3, characterized in that: The main network consists of an initial convolutional layer, a max pooling layer, a wavelet stacked group global average pooling layer, and a fully connected layer; In the initial convolutional layer, the input time series is first embedded through an initial convolution operation, which performs preliminary feature extraction on the input signal, reduces the spatial resolution, and extracts low-level features, as follows: ; Where X1 represents the preliminary feature, input data ,in The input channel number is represented by D, and the dimension of the sample is represented by D. The whole process sequentially goes through the Conv operation with a kernel of k1 and a stride of s1, the BatchNorm operation, and the ReLU activation function operation. In the max pooling layer, the max pooling layer is used to further reduce the spatial resolution of the feature map and reduce the computational cost of subsequent layers, as follows: ; Where X2 represents the output after max pooling, using MaxPool operation with convolution kernel k2 and sliding stride s2; In the global average pooling layer of the main network, the global average pooling layer is used to compress the spatial dimension of the feature map while preserving channel information, as follows: ; The GAP operation selects to use the flattening operation to transform the two-dimensional X... L The output is flattened into a one-dimensional X. G Towards, X G The output corresponding to the GAP layer, X L The output of the corresponding wavelet stacking group; In the fully connected layers of the main network, the fully connected layers are used to map the feature vectors after the global average pooling layer to the target category, that is, to determine the category of shockable heart rhythm, as follows: ; Where Y represents the category output, and Linear represents the linear connection of the fully connected layer; The secondary network consists of an initial convolutional layer, a max pooling layer, a wavelet stacking group layer, and a decoding layer; the initial convolutional layer, the max pooling layer, and the wavelet stacking group layer are shared with the main network. In the decoding layer of the secondary network, the decoding layer is used to use interpolation convolution to restore the feature signal to the same dimension as the original ECG signal input, as follows: ; The Interpolation algorithm selects bilinear interpolation or nearest neighbor interpolation to amplify the feature map, and then uses a convolution operation with a kernel of k3 and a stride of s3 to smooth the result and avoid the checkerboard effect.

5. An application method for a cardiac defibrillation rhythm recognition model, characterized in that, The application method of this recognition model is the application of the cardiac defibrillation rhythm recognition model constructed by the cardiac defibrillation rhythm recognition model establishment method according to any one of claims 1 to 4. The application method of the recognition model includes the following steps: The deep convolutional neural network model structure is constructed, which includes a dual dual-path network structure consisting of a main path network and a secondary path network. The main path network and the secondary path network share an initial convolutional layer, a max pooling layer, and a wavelet stacking group layer. The main path network also includes a global average pooling layer and a fully connected layer, while the secondary path network includes a decoding layer. The joint signal is composed of the transelectrocardiogram (TTI) signal and the interfered electrocardiogram (ECG) signal. Based on continuous wavelet transform, the instantaneous frequency characteristics of the joint signal are captured and processed, and the wavelet transform blocks are constructed by adaptive feature processing at different time and location scales. The wavelet stacking group layer is composed of multiple wavelet transform blocks stacked together. The training and iteration of the neural network model structure, targeting the real interference physical characteristics of emergency scenarios, simultaneously applies corresponding data augmentation transformations to the cardiac impedance signal TTI and the electrocardiogram signal ECG. The neural network model structure is trained by the data-augmented cardiac impedance signal TTI and the interfered electrocardiogram signal ECG, forming a fusion of adversarial training between the main network and the secondary network. The main network classifies and determines whether shock is needed and calculates the discriminant loss function for shockable heart rate. The secondary network identifies interference, reconstructs the generation of anti-interference ECG signals, and calculates the generation loss function for anti-interference ECG signals. The discriminant loss function and the generation loss function are combined to construct a joint adversarial loss function, which is used to generate the global optimization target iteration direction. After multiple iterations, the optimal model weight parameters are retained, and the cardiac defibrillation rhythm recognition model is constructed from these optimal model weight parameters. The deployment of the cardiac defibrillation rhythm recognition model involves extracting the model structure and parameters of the main network in the cardiac defibrillation rhythm recognition model, performing lightweight optimization of the model, embedding it into the processing chip of the AED device, and converting it into a lightweight inference engine. Cardiac defibrillation rhythm recognition: The electrode pads of the AED device are applied to the patient's body parts, and after acquiring the electrocardiogram (ECG) signal and the transcatheter impedance (TTI) signal, the main circuit discriminator identifies the shockable rhythm. If an unshockable rhythm is identified, the defibrillation is stopped; if a shockable rhythm is identified, the defibrillation is performed to avoid accidental and missed defibrillations.

6. An AED defibrillator, characterized in that: The processing chip of the AED defibrillator is embedded with the model structure and parameters of the lightweight optimized main network extracted from the cardiac defibrillation rhythm recognition model in the application method of the cardiac defibrillation rhythm recognition model according to any one of claims 1 to 4.

7. A readable storage medium, characterized in that: The readable storage medium stores a control program, which, when executed by the detection module, causes the detection module to perform the steps of the application method of the cardiac defibrillation rhythm recognition model as described in any one of claims 1 to 4. Alternatively, the readable storage medium stores the model structure and parameters of the lightweight optimized main network extracted from the cardiac defibrillation rhythm recognition model in the application method of the cardiac defibrillation rhythm recognition model according to any one of claims 1 to 4, and performs the steps of the application method of the cardiac defibrillation rhythm recognition model according to claim 5.

Images