Low-power-consumption pulse wave signal arrhythmia classification method and system

Abstract

The invention discloses a low-power-consumption pulse wave signal arrhythmia classification method and system, and the method comprises the steps: extracting the local features of a bottom layer of an original signal through a large convolution kernel, and halving the length of a sequence through pooling; adopting leakage integral-discharge neuron activation to pulse data; enlarging a receptive field by using a large convolution kernel and cavity convolution, and capturing local time sequence features in a longer range; dimension reduction is carried out on output features, batch normalization, leakage integral-discharge neurons and maximum pooling are matched, and local features are aggregated; calculating self-attention in each sub-sequence, combining a linear mapping layer, focusing a model to feature association in a local sub-sequence through an aggregated sub-attention graph, and extracting waveform dependence in a short time window in the pulse wave signal; qKV interaction is calculated for the whole sequence, a linear layer and spiking neurons are combined, long-distance global feature correlation is captured, and waveform dependence of different time periods in pulse wave signals is mined. According to the method, classification performance equivalent to that of a deep artificial neural network is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a low-power pulse wave signal arrhythmia classification method and system. BACKGROUND

[0002] Based on the photoplethysmogram (PPG, referred to as pulse wave) arrhythmia classification detection for its convenience and real-time, it shows a significant advantage in the field of large-scale screening, and provides a potential effective solution for long-term real-time monitoring of arrhythmia. In recent years, studies have shown that deep artificial neural networks (ANN) are effective in pulse wave arrhythmia classification. However, the learning and inference of deep neural network models involve dense floating point multiplication operations, which have high computational complexity and high power consumption, restricting their application in wearable portable devices. The pulse activation function in the spiking neural network (SNN) is driven by discrete pulses, which is mainly implemented through accumulation operation, with low computational complexity and low power consumption.

[0003] The current routine tools for diagnosing arrhythmia include 12-lead electrocardiogram and 24-hour Holter monitoring. However, the electrocardiogram-based recorder requires the patient to paste electrodes, causing burden and inconvenience. The pulse wave signal provides a simple, economical and non-intrusive means for continuous monitoring of the vascular system. Previous studies have shown that arrhythmia detection based on pulse wave signals has good application prospects.

[0004] Convenient arrhythmia detection and monitoring methods will greatly improve the prevention and management of arrhythmia-related complications. In the prior art, it includes a two-class task of identifying atrial fibrillation from pulse wave signals using a multi-task convolutional neural network; a deep convolutional neural network is first used to detect and classify multiple arrhythmias from pulse wave signals, which identifies six types of arrhythmias from pulse wave signals: sinus rhythm, premature ventricular contractions, premature atrial contractions, ventricular tachycardia, supraventricular tachycardia, and atrial fibrillation; a deep learning model based on convolutional neural network and Transformer attention module is proposed for a 3-class task of pulse wave signal, which identifies normal sinus rhythm, atrial fibrillation and ectopic beats (including premature ventricular contractions and premature atrial contractions).

[0005] However, the current pulse wave signal arrhythmia recognition method is mainly based on deep neural networks, which involves complex and dense floating point multiplication and addition operations, has high computational complexity, and deploying on hardware devices will result in high energy consumption. The present application uses a spiking neural network to construct a multi-class model of pulse wave signals. The spiking neuron in the spiking neural network uses a pulse activation function, which ensures that the model only involves sparse floating point addition, thereby avoiding the complex and dense floating point multiplication and addition operations in traditional artificial neural networks. The above characteristics ensure the low complexity and low power consumption of the model. SUMMARY

[0006] The embodiment of the present application provides a low-power pulse wave signal arrhythmia classification method and system, so as to guarantee low power of the model and achieve classification performance equivalent to a deep artificial neural network.

[0007] According to an embodiment of the present application, a low-power pulse wave signal arrhythmia classification method is provided, comprising the following steps: The original signal is extracted using a large convolution kernel to extract the bottom layer local feature of the original signal, and the sequence length is halved through pooling; and data is pulsed by adopting a leaky integral-discharge neuron activation; The receptive field is expanded by using a large convolution kernel and a hollow convolution to capture longer-range local time sequence features; The dimensionality of the output feature is reduced, and batch normalization, leaky integral-discharge neurons and maximum pooling are used to aggregate local features; Self-attention is calculated in each subsequence, combined with a linear mapping layer, and the model is focused on feature correlation in the local subsequence by aggregating sub-attention maps to extract waveform dependence in a short time window in the pulse wave signal; QKV interaction is calculated for the entire sequence, combined with a linear layer and a pulse neuron to capture long-distance global feature correlation and mine waveform dependence in different time periods in the pulse wave signal.

[0008] Further, in the step of extracting the bottom layer local feature of the original signal using a large convolution kernel, 1D convolution of the large convolution kernel is used to extract the bottom layer local feature of the original signal. In the step of pulsing the data, the data is pulsed to 0 and 1.

[0009] Further, in the step of expanding the receptive field by using a large convolution kernel and a hollow convolution to capture longer-range local time sequence features, a large convolution kernel is combined with an increasing expansion rate in the early stage to quickly cover a wide range of time sequence correlation and capture the overall pattern of the signal; and a small convolution kernel is used in the later stage to focus on local detailed features and provide discriminative clues for classification.

[0010] Further, in the step of calculating self-attention in each subsequence, QKV linear projection and pulse neuron activation are included; and the attention module adopts a calculation order optimized by Spikformer.

[0011] Further, in the step of calculating QKV interaction for the entire sequence, combined with a linear layer and a pulse neuron, long-distance global feature correlation is captured and waveform dependence in different time periods in the pulse wave signal is mined; and the pulse neuron introduces time sparsity to adapt to the dynamic characteristics of the time sequence signal.

[0012] Further, before the step of extracting the bottom layer local feature of the original signal using a large convolution kernel, the pulse wave signal data is standardized.

[0013] Further, dropout operation is adopted in each step to prevent model overfitting.

[0014] According to another embodiment of the application, a low-power pulse wave signal arrhythmia classification system is provided, comprising: A backbone network model is used to extract the bottom local features of the original signal using a large convolution kernel, and the sequence length is halved through pooling; a leaky integral-discharge neuron activation is adopted to pulse the data; A pulse residual hollow convolution module is used to expand the receptive field by using a large convolution kernel and a hollow convolution, and to capture longer-range local timing features; A down-sampling module is used to reduce the dimension of the output features, and is matched with batch normalization, leaky integral-discharge neurons and maximum pooling to aggregate local features; A local self-attention module is used to calculate self-attention within each sub-sequence, and is combined with a linear mapping layer to focus the model on the feature correlation within the local sub-sequence by aggregating sub-attention maps, and to extract waveform dependence in a short time window in the pulse wave signal; A global self-attention module is used to calculate QKV interaction on the entire sequence, and is combined with a linear layer and a pulse neuron to capture long-distance global feature correlation and mine waveform dependence in different time periods in the pulse wave signal.

[0015] A storage medium stores a program file capable of implementing any one of the low-power pulse wave signal arrhythmia classification methods described above.

[0016] A processor is used to run a program, wherein the program performs the low-power pulse wave signal arrhythmia classification method of any one of the above when running.

[0017] The low-power pulse wave signal arrhythmia classification method and system in the embodiments of the application replace the nonlinear activation function in the deep neural network with a pulse activation function, thereby replacing the dense floating-point multiplication and accumulation operation in the traditional artificial neural network with a sparse floating-point accumulation operation, and greatly reducing the computational complexity and power consumption of the algorithm. In addition, the model adopts a hollow convolution combined with a local-global attention mechanism, which can effectively model the long-time dependence relationship of the pulse wave signal, thereby ensuring low power consumption of the model while achieving classification performance comparable to that of a deep artificial neural network. BRIEF DESCRIPTION OF DRAWINGS

[0018] The drawings described herein are used to provide further understanding of the application, and form a part of the application. The illustrative embodiments of the application and their descriptions serve to explain the application, and do not constitute an improper limitation on the application. In the drawings: Figure 1 A pulse wave signal classification model based on a deep pulse neural network in the application; Figure 2 Figure for backbone network / down-sampling module and pulse residual hole convolution module in the application; Figure 3 Figure for attention module in the application. DETAILED DESCRIPTION

[0019] In order to enable persons skilled in the art to better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should fall within the scope of protection of the present application.

[0020] It should be noted that the terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0021] Embodiment 1 According to an embodiment of the present application, a low-power pulse wave signal arrhythmia classification method is provided, comprising the following steps: The bottom layer local features of the original signal are extracted using a large convolution kernel, and the sequence length is halved by pooling; the data is pulsed by using a leaky integral-discharge neuron activation; The receptive field is expanded by using a large convolution kernel and a hole convolution, and longer range local timing features are captured; The dimensionality of the output features is reduced, and batch normalization, leaky integral-discharge neurons and max pooling are used to aggregate local features; Self-attention is calculated within each sub-sequence, combined with a linear mapping layer, and the model is focused on the feature correlation within the local sub-sequence by aggregating sub-attention maps to extract waveform dependencies in short time windows in the pulse wave signal; QKV interaction is calculated for the entire sequence, combined with a linear layer and a pulse neuron, to capture long-distance global feature correlation and mine waveform dependencies in different time periods in the pulse wave signal.

[0022] The low-power pulse wave signal arrhythmia classification method in the embodiment of the application adopts a pulse activation function to replace a nonlinear activation function in a deep neural network, thereby replacing a dense floating-point multiplication and accumulation operation in a traditional artificial neural network with a sparse floating-point accumulation operation, and greatly reducing the calculation complexity and power consumption of the algorithm. In addition, the model adopts a hollow convolution combined with a local-global attention mechanism, which can effectively model the long-time dependence relationship of the pulse wave signal, thereby ensuring low power consumption of the model while achieving a classification performance comparable to that of a deep artificial neural network.

[0023] The low-power pulse wave signal arrhythmia classification method of the application will be described in detail below with specific embodiments: In view of the technical problems existing in the prior art, the application proposes a low-power pulse wave signal arrhythmia multi-classification method based on a pulse neural network, which adopts a pulse activation function to replace a nonlinear activation function in a deep neural network, thereby replacing a dense floating-point multiplication and accumulation operation in a traditional artificial neural network with a sparse floating-point accumulation operation, and greatly reducing the calculation complexity and power consumption of the algorithm. In addition, the model adopts a hollow convolution combined with a local-global attention mechanism, which can effectively model the long-time dependence relationship of the pulse wave signal, thereby ensuring low power consumption of the model while achieving a classification performance comparable to that of a deep artificial neural network.

[0024] The application proposes a pulse neural network pulse wave arrhythmia multi-classification method based on local-global attention, which adopts a hollow convolution and a self-attention model to effectively model the long-time dependence relationship of PPG time series data, improves the classification performance of the model, places the pulse activation operation before the convolution and linear transformation operation, ensures that the main data flow of the model does not involve dense floating-point multiplication operations, and reduces the power consumption of the model. Among them, the pulse residual hollow convolution can capture the long-time sequence dependence relationship of the pulse wave signal while enabling the model to be trained more effectively in a deep layer; the block local self-attention divides the sequence features, extracts attention features for the block subsequences respectively, and performs feature fusion, thereby capturing the internal correlation information of the subsequence; the global self-attention explores the long-range correlation relationship of the sequence data from a global perspective. Except for the input layer, the data flow is pulsed through the pulse activation neuron before the input convolution layer or the linear layer, thereby ensuring that the model avoids floating-point multiplication operations during the inference process, reducing the calculation complexity and power consumption of the algorithm.

[0025] The technical scheme of the application is described in detail as follows: Referring to Figure 1 The steps implemented by the embodiment of the application are as follows: Raw data preprocessing: standardizing the pulse wave signal data, and normalizing the values to the [0, 1] interval; The backbone network model ( Figure 2) using a 1D convolution with a large kernel size (kernel_size=31) to extract the underlying local features of the original signal, and through pooling to reduce the sequence length by half, to achieve preliminary dimension reduction and reduce the subsequent computational load. After the convolution operation, a leaky integrate-and-fire neuron (LIF) activation is used to pulse the data to 0 and 1, simplifying the subsequent calculation complexity.

[0026] The model further uses a pulse residual dilated convolution module ( Figure 2 ), which uses a large convolution kernel and dilated convolution to expand the receptive field and capture longer-range local temporal features (such as waveform changes in pulse wave signals). The residual connection alleviates the gradient vanishing, and the pulse neuron introduces time dimension sparsity. The dilation rate of the dilated convolution gradually increases, capturing local correlations at a longer distance through a gradually increasing dilation rate. The dilation rate gradually increases, with a change order of 2-4-6-8-10; the convolution kernel gradually decreases, with a change order of 21-21-13-13-8. In the early stage, a large convolution kernel is combined with a gradually increasing dilation rate to quickly cover a wide range of temporal correlations and capture the overall pattern of the signal. In the later stage, a small convolution kernel is used to focus on local detailed features, providing discriminative clues for classification.

[0027] The down-sampling module ( Figure 2 ) further reduces the dimensionality of the output features of the front-end pulse residual dilated convolution module (by half of the sequence length), and cooperates with batch normalization, leaky integrate-and-fire neuron, and max pooling to aggregate local features. It provides more compact feature input for the subsequent attention module.

[0028] The local self-attention module divides the sequence into blocks ( Figure 3 ), and calculates self-attention within each sub-sequence, including query-key-value (QKV) linear projection and pulse neuron activation, combined with a linear mapping layer, to focus on feature correlations within the local sub-sequence through aggregated sub-attention maps, and extract waveform dependencies in short time windows in pulse wave signals. The attention module uses the calculation order optimized by Spikformer[5], avoiding the exponential and division operations involved in the softmax operation, reducing the model calculation complexity.

[0029] The global self-attention module ( Figure 3 ) directly calculates QKV interaction on the entire sequence, combined with a linear layer and a pulse neuron. It captures long-distance global feature correlations and excavates waveform dependencies in different time periods in pulse wave signals. The pulse neuron introduces time sparsity, adapting to the dynamic characteristics of time series signals.

[0030] Dropout operation is adopted in the backbone network model, the down-sampling module, the pulse residual hollow convolution module, the local self-attention module and the global self-attention module to prevent model overfitting and ensure model training stability.

[0031] The key points and points to be protected of the present application are: The present application proposes a pulse wave signal classification model based on a deep pulse neural network, which can be used for population screening of heart diseases and deployed in a PPG device for long-term and continuous monitoring, so as to early discover high-risk patients and make them receive anticoagulant therapy as soon as possible, thereby preventing a large number of stroke events from occurring: (1) a deep pulse neural network model based on pulse residual hollow convolution and local-global self-attention mechanism is constructed to classify arrhythmia of pulse wave signals; (2) a series of pulse residual hollow convolution modules are adopted to gradually explore long-range dependence of pulse wave signal sequences, and meanwhile, stable training of the deep pulse neural network is ensured; (3) a local-global self-attention mechanism based on pulse neurons is adopted to further explore local and global information of pulse wave signals, and a spikformer attention map calculation method is adopted to avoid the complex attention map calculation process in the traditional artificial neural network; (4) data flow is pulsed through pulse activation neurons before inputting into a convolution layer or a linear layer, so that the model avoids floating point multiplication operation in the inference process, and the algorithm calculation complexity and power consumption are reduced.

[0032] Compared with the prior art, the present application has the following advantages: The proposed model is a deep neural network based on pulse neurons, the data flow is pulsed through pulse neuron activation, the model uses floating point addition instead of floating point multiplication, avoids the complex and dense floating point multiplication and addition operation in the traditional artificial neural network, effectively reduces the calculation complexity and power consumption of the algorithm. Through a series of pulse residual hollow convolution modules, long-range dependence of pulse wave signals is gradually explored, and the gradient attenuation problem in the deep network is avoided through residual connection, so that stable learning of the deep pulse neural network is ensured; the spikformer-based pulse attention map calculation method is adopted to avoid the complex calculation method of the attention map in the traditional self-attention module; the proposed model is the first deep pulse neural network model for arrhythmia multi-classification of pulse wave signals, which can reduce the calculation complexity and power consumption of the model while ensuring the heart disease classification performance, and has great social and economic value for the popularization of low-cost medical treatment.

[0033] The scheme flow of the present application is feasible in theoretical principle and actual operation. Experiments are carried out on real pulse wave signal data. The training data includes 228 patient PPG data records. These patient data are randomly divided into a training set, a validation set and a test set according to a ratio of 60%, 20% and 20%. Each patient has several PPG data segments with a length of 2500. The PPG data labels are 7 categories, including sinus rhythm, ventricular premature beat, atrial premature beat, ventricular tachycardia, supraventricular tachycardia, atrial fibrillation and noise. The 6 categories of PPG data records after eliminating the noise category are used for model learning.

[0034] In summary, the training set, the validation set and the test set respectively contain 66631, 22511 and 23989 PPG data segments. The model is trained for 50 epochs. The model with the highest accuracy on the validation set is saved, and the sensitivity and specificity of the model on the test set are reported. The sensitivity and specificity of the deep convolutional neural network VGG13 model on the test set are 74.78% and 97.14% respectively. The sensitivity and specificity of the present application scheme on the test set are 75.48% and 97.20% respectively. It can be seen that on the basis of maintaining low power consumption characteristics, the pulse wave arrhythmia classification performance of the proposed model is equivalent to the current best practice deep convolutional neural network pulse wave classification method, and even slightly better than the latter. The above experiments verify the effectiveness of the proposed model.

[0035] Embodiment 2 According to another embodiment of the present application, a low-power pulse wave signal arrhythmia classification system is provided, comprising: A backbone network model is used to extract the bottom layer local features of the original signal using a large convolution kernel, and the sequence length is halved through pooling. Leaky integral-discharge neuron activation is adopted to make the data pulsed. A pulse residual hollow convolution module is used to expand the receptive field by using a large convolution kernel and a hollow convolution, and to capture longer-range local time sequence features. A down-sampling module is used to reduce the dimension of the output features, and is matched with batch normalization, leaky integral-discharge neuron and maximum pooling to aggregate local features. A local self-attention module is used to calculate self-attention within each sub-sequence, and is combined with a linear mapping layer to focus the model on the feature correlation within the local sub-sequence by aggregating sub-attention maps, and to extract waveform dependence in a short time window in the pulse wave signal. A global self-attention module is used to calculate QKV interaction on the entire sequence, and is combined with a linear layer and a pulse neuron to capture long-distance global feature correlation and mine waveform dependence in different time periods in the pulse wave signal.

[0036] The low-power pulse wave signal arrhythmia classification system in the embodiment of the application adopts a pulse activation function to replace a nonlinear activation function in a deep neural network, thereby replacing a dense floating-point multiplication and accumulation operation in a traditional artificial neural network with a sparse floating-point accumulation operation, and greatly reducing the calculation complexity and power consumption of the algorithm.In addition, the model adopts a hollow convolution combined with a local-global attention mechanism, which can effectively model the long-time dependence relationship of the pulse wave signal, thereby ensuring low power consumption of the model while achieving a classification performance comparable to that of a deep artificial neural network The low-power pulse wave signal arrhythmia classification system of the application will be described in detail below with specific embodiments: In view of the technical problems existing in the prior art, the application proposes a low-power pulse wave signal arrhythmia multi-classification system based on a pulse neural network, which adopts a pulse activation function to replace a nonlinear activation function in a deep neural network, thereby replacing a dense floating-point multiplication and accumulation operation in a traditional artificial neural network with a sparse floating-point accumulation operation, and greatly reducing the calculation complexity and power consumption of the algorithm.In addition, the model adopts a hollow convolution combined with a local-global attention mechanism, which can effectively model the long-time dependence relationship of the pulse wave signal, thereby ensuring low power consumption of the model while achieving a classification performance comparable to that of a deep artificial neural network.

[0037] The application proposes a pulse neural network pulse wave arrhythmia multi-classification system based on local-global attention, which adopts a hollow convolution and a self-attention model to effectively model the long-time dependence relationship of PPG time series data, improves the classification performance of the model, places the pulse activation operation before the convolution and linear transformation operation, ensures that the main data flow of the model does not involve dense floating-point multiplication operations, and reduces the power consumption of the model.The pulse residual hollow convolution can capture the long-time sequence dependence relationship of the pulse wave signal while enabling the model to be trained more effectively at a deeper level;The block local self-attention divides the sequence features, extracts attention features for the block sub-sequences respectively, and performs feature fusion, thereby capturing the internal correlation information of the sub-sequences;The global self-attention explores the long-range correlation relationship of the sequence data from a global perspective.In addition to the input layer, the data flow is pulsed before the input convolution layer or the linear layer through the pulse activation neuron, thereby ensuring that the model avoids floating-point multiplication operations during the inference process, reducing the calculation complexity and power consumption of the algorithm.

[0038] The technical scheme of the application is described in detail as follows: Referring to Figure 1 The steps implemented by the embodiment of the application are as follows: Original data preprocessing: standardizing the pulse wave signal data, and normalizing the values to the [0, 1] interval; Main network model ( Figure 2) using a 1D convolution with a large kernel size (kernel_size=31) to extract the underlying local features of the original signal, and through pooling to reduce the sequence length by half, to achieve preliminary dimension reduction and reduce the subsequent computational load. After the convolution operation, a leaky integrate-and-fire neuron (LIF) activation is used to pulse the data to 0 and 1, simplifying the subsequent calculation complexity.

[0039] The model further uses a pulse residual dilated convolution module ( Figure 2 ), which uses a large convolution kernel and dilated convolution to expand the receptive field and capture longer-range local temporal features (such as waveform changes in pulse wave signals). The residual connection alleviates the gradient vanishing, and the pulse neuron introduces time dimension sparsity. The dilation rate of the dilated convolution gradually increases, capturing local correlations at a longer distance through a gradually expanding dilation rate. The dilation rate gradually increases, with a change order of 2-4-6-8-10; the convolution kernel gradually decreases, with a change order of 21-21-13-13-8. In the early stage, a large convolution kernel is combined with a gradually increasing dilation rate to quickly cover a wide range of temporal correlations and capture the overall pattern of the signal. In the later stage, a small convolution kernel is used to focus on local detailed features, providing discriminative clues for classification.

[0040] The down-sampling module ( Figure 2 ) further reduces the dimensionality of the output features of the front-end pulse residual dilated convolution module (by half of the sequence length), and cooperates with batch normalization, leaky integrate-and-fire neuron, and max pooling to aggregate local features. It provides more compact feature input for the subsequent attention module.

[0041] The local self-attention module divides the sequence into blocks ( Figure 3 ), and calculates self-attention within each sub-sequence, including query-key-value (QKV) linear projection and pulse neuron activation, combined with a linear mapping layer. Through aggregating sub-attention maps, the model focuses on feature correlations within local sub-sequences, extracting waveform dependencies in short time windows in pulse wave signals. The attention module uses the calculation order optimized by Spikformer[5], avoiding the exponential and division operations involved in the softmax operation, reducing the model calculation complexity.

[0042] The global self-attention module ( Figure 3 ) directly calculates QKV interaction on the entire sequence, combined with a linear layer and a pulse neuron. It captures long-distance global feature correlations and excavates waveform dependencies in different time periods in pulse wave signals. The pulse neuron introduces time sparsity, adapting to the dynamic characteristics of time series signals.

[0043] Dropout operation is adopted in the backbone network model, the down-sampling module, the pulse residual hollow convolution module, the local self-attention module and the global self-attention module to prevent model overfitting and ensure model training stability.

[0044] The key points and points to be protected of the application are: The application proposes a pulse wave signal classification model based on a deep pulse neural network, which can be used for population screening of heart diseases and deployed in a PPG device for long-term and continuous monitoring, so that high-risk patients can be found early and receive anticoagulant therapy as soon as possible, thereby preventing a large number of stroke events: (1) a deep pulse neural network model based on pulse residual hollow convolution and local-global self-attention mechanism is constructed to classify arrhythmia of pulse wave signals; (2) a series of pulse residual hollow convolution modules are adopted to gradually explore the long-range dependence relationship of the pulse wave signal sequence, and ensure that the deep pulse neural network can be stably trained; (3) a local-global self-attention mechanism based on pulse neurons is adopted to further mine local and global information of the pulse wave signal, and a spikformer attention map calculation method is adopted to avoid the complex attention map calculation process in the traditional artificial neural network; (4) the data stream is pulsed through the pulse activation neuron before inputting the convolution layer or the linear layer, so that the model avoids floating point multiplication operation in the inference process, and reduces the algorithm calculation complexity and power consumption.

[0045] Compared with the prior art, the application has the following advantages: The proposed model is a deep neural network based on pulse neurons, which makes the data stream pulsed through pulse neuron activation, replaces floating point multiplication with floating point addition, avoids complex and dense floating point multiplication and addition operations in the traditional artificial neural network, and effectively reduces the algorithm calculation complexity and power consumption. Through a series of pulse residual hollow convolution modules, the long-range dependence relationship of the pulse wave signal is gradually explored, and the gradient attenuation problem in the deep network is avoided through the residual connection, so that the stable learning of the deep pulse neural network is ensured. The spikformer-based pulse attention map calculation method is adopted to avoid the complex calculation method of the attention map in the traditional self-attention module. The proposed model is the first deep pulse neural network model for arrhythmia multi-classification of pulse wave signals, which can reduce the calculation complexity and power consumption of the model while ensuring the heart disease classification performance, and has great social and economic value for the popularization of low-cost medical treatment.

[0046] The scheme flow of the present application is feasible in theory and actual operation. Experiments are carried out on real pulse wave signal data. The training data includes 228 patient PPG data records. These patient data are randomly divided into a training set, a validation set and a test set according to the proportions of 60%, 20% and 20%. Each patient has several PPG data segments with a length of 2500. The PPG data labels are 7 categories, including sinus rhythm, ventricular premature beat, atrial premature beat, ventricular tachycardia, supraventricular tachycardia, atrial fibrillation and noise. The 6 categories of PPG data records after eliminating the noise category are used for model learning.

[0047] In summary, the training set, the validation set and the test set respectively contain 66631, 22511 and 23989 PPG data segments. The model is trained for 50 epochs. The model with the highest accuracy on the validation set is saved, and the sensitivity and specificity of the model on the test set are reported. The sensitivity and specificity of the deep convolutional neural network VGG13 model on the test set are 74.78% and 97.14% respectively. The sensitivity and specificity of the present application on the test set are 75.48% and 97.20% respectively. It can be seen that on the basis of maintaining low power consumption characteristics, the pulse wave arrhythmia classification performance of the proposed model is equivalent to the current best practice deep convolutional neural network pulse wave classification method, and even slightly better than the latter. The above experiments verify the effectiveness of the proposed model.

[0048] Embodiment 3 A storage medium, the storage medium stores a program file capable of realizing any one of the above low-power pulse wave signal arrhythmia classification methods.

[0049] Embodiment 4 A processor, the processor is used to run a program, wherein the program executes the low-power pulse wave signal arrhythmia classification method of any one of the above when running.

[0050] The above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0051] In the above embodiments of the present application, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0052] In several embodiments provided in the present application, it should be understood that the disclosed technology can be implemented by other ways. Among them, the system embodiments described above are only illustrative, for example, the division of units can be a logical function division, and actual implementation can have another division way, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or modules shown or discussed can be indirect coupling or communication connection between the units or modules through some interfaces, and can be electrical or other forms.

[0053] The units described as separate components can or can not be physically separated, and the components shown as units can or can not be physical units, that is, they can be located in one place or distributed to multiple units. Part or all of the units can be selected to achieve the purpose of the embodiment scheme according to actual needs.

[0054] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0055] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or all or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the embodiments of the present application. The foregoing storage medium includes: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0056] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A low-power pulse wave signal arrhythmia classification method, characterized in that, Includes the following steps: Large convolutional kernels are used to extract low-level local features of the original signal, and pooling is used to halve the sequence length; leak integral-discharge neuron activation is used to pulse the data. Large convolutional kernels and dilated convolutions are used to expand the receptive field and capture local temporal features over a longer range. Dimensionality reduction output features, combined with batch normalization, leaky integral-discharge neurons, and max pooling, aggregate local features; Self-attention is computed within each subsequence. Combined with a linear mapping layer, the model is focused on the feature associations within local subsequences by aggregating the sub-attention map, and waveform dependence within a short time window in the pulse wave signal is extracted. The QKV interaction is calculated for the entire sequence. By combining linear layers and spiking neurons, long-distance global feature associations are captured, and waveform dependencies at different time points in the pulse wave signal are mined.

2. The low-power pulse wave signal arrhythmia classification method according to claim 1, characterized in that, In extracting the low-level local features of the original signal using large convolution kernels, 1D convolution with large convolution kernels is used to extract the low-level local features of the original signal. In making the data pulsed, the data is pulsed to 0 and 1.

3. The low-power pulse wave signal arrhythmia classification method according to claim 1, characterized in that, In expanding the receptive field and capturing a longer range of local temporal features by using large convolutional kernels and dilated convolutions, the initial stage uses large convolutional kernels combined with gradually increasing dilation rates to quickly cover a wide range of temporal correlations and capture the overall pattern of the signal; the later stage uses small convolutional kernels to focus on local detailed features and provide discriminative clues for classification.

4. The low-power pulse wave signal arrhythmia classification method according to claim 1, characterized in that, Self-attention is computed within each subsequence, including query key-value QKV linear projection and spiking neuron activation; The attention module uses a Spikformer-optimized computation order.

5. The low-power pulse wave signal arrhythmia classification method according to claim 1, characterized in that, In computing QKV interactions across the entire sequence, combining linear layers and spiking neurons to capture long-distance global feature associations and mine waveform dependencies at different time points in pulse wave signals, spiking neurons introduce temporal sparsity to adapt to the dynamic characteristics of time-series signals.

6. The low-power pulse wave signal arrhythmia classification method according to claim 1, characterized in that, Before using large convolutional kernels to extract the low-level local features of the original signal, the pulse wave signal data is standardized.

7. The low-power pulse wave signal arrhythmia classification method according to claim 1, characterized in that, Dropout is used in each step to prevent model overfitting.

8. A low-power pulse wave signal arrhythmia classification system, characterized in that, include: The backbone network model is used to extract low-level local features of the original signal using large convolutional kernels and to halve the sequence length through pooling. Leakage integral-discharge neuron activation is used to pulse the data; The pulsed residual dilated convolution module is used to expand the receptive field by using large convolutional kernels and dilated convolutions to capture local temporal features over a longer range. The downsampling module is used to reduce the dimensionality of the output features, and in conjunction with batch normalization, leaky integral-discharge neurons and max pooling, it aggregates local features; The local self-attention module is used to compute self-attention within each subsequence. Combined with the linear mapping layer, it focuses the model on the feature associations within the local subsequence by aggregating the sub-attention map, and extracts the waveform dependence within a short time window in the pulse wave signal. The global self-attention module is used to compute QKV interactions over the entire sequence. It combines linear layers and spiking neurons to capture long-distance global feature associations and mine waveform dependencies at different time points in the pulse wave signal.

9. A storage medium, characterized in that, The storage medium stores a program file capable of implementing the low-power pulse wave signal arrhythmia classification method according to any one of claims 1 to 7.

10. A processor, characterized in that, The processor is used to run a program, wherein the program executes the low-power pulse wave signal arrhythmia classification method according to any one of claims 1 to 7.

Images