A low-power pulse wave signal arrhythmia classification method and system

Abstract

The low-power pulse wave signal arrhythmia classification method and system of the application comprises the following steps: using a large convolution kernel to extract the bottom local features of the original signal, and reducing the sequence length by half through pooling; adopting a leaky integral-discharge neuron activation to make the data pulsed; using a large convolution kernel and a hollow convolution to expand the receptive field and capture longer-range local timing features; dimensionally reducing the output features, cooperating with batch normalization, leaky integral-discharge neurons and maximum pooling, and aggregating the local features; calculating self-attention in each subsequence, combining a linear mapping layer, focusing the model on the feature correlation in the local subsequence through the aggregation of the sub-attention map, and extracting the waveform dependence in the short time window of the pulse wave signal; calculating QKV interaction on the whole sequence, combining a linear layer and a pulse neuron, capturing long-distance global feature correlation, and mining the waveform dependence in different time periods of the pulse wave signal. The application can achieve a classification performance equivalent to that of a deep artificial neural network.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a low-power pulse wave signal arrhythmia classification method and system. Background Technology

[0002] Photoplethysmography (PPG)-based arrhythmia classification and detection has demonstrated significant advantages in large-scale screening due to its convenience and real-time performance, providing a potentially effective solution for long-term real-time monitoring of arrhythmias. In recent years, studies have demonstrated the effectiveness of deep artificial neural networks (ANNs) in pulse wave arrhythmia classification; however, the learning and inference of deep neural network models involve dense floating-point multiplication operations, resulting in high computational complexity and power consumption, which limits their application in wearable and portable devices. Spiking neural networks (SNNs), on the other hand, employ discrete pulse-driven activation functions, primarily implemented through accumulation operations, resulting in low computational complexity and low power consumption.

[0003] Currently, routine tools for diagnosing arrhythmias include 12-lead electrocardiography (ECG) and 24-hour Holter monitoring. However, ECG-based recorders require electrodes to be attached to the patient, causing burden and inconvenience. Pulse wave signals provide a simple, economical, and interference-free means for continuous monitoring of the vascular system. Previous studies have shown that pulse wave signal-based arrhythmia detection methods have promising application prospects.

[0004] Convenient methods for detecting and monitoring arrhythmias will greatly improve the prevention and management of arrhythmia-related complications. Existing technologies include a binary classification task using multi-task convolutional neural networks to identify atrial fibrillation in pulse wave signals; the first application of deep convolutional neural networks to detect and classify multiple types of arrhythmias in pulse wave signals, identifying six types of arrhythmias: sinus rhythm, premature ventricular contractions (PVCs), premature atrial contractions (PACs), ventricular tachycardia, supraventricular tachycardia, and atrial fibrillation; and a proposed three-classification task based on a deep learning model using convolutional neural networks and a Transformer attention module to identify normal sinus rhythm, atrial fibrillation, and ectopic beats (including PVCs and PACs).

[0005] However, current methods for identifying abnormal heart rates in pulse wave signals mainly rely on deep neural networks, involving complex and dense floating-point accumulation, multiplication, and addition operations. This results in high computational complexity and high energy consumption when deployed on hardware. This invention employs a spiking neural network to construct a multi-classification model for pulse wave signals. The spiking neurons in the spiking neural network use a spiking activation function, whose characteristics ensure that the model only involves sparse floating-point accumulation, thus avoiding the complex and dense floating-point accumulation, multiplication, and addition operations found in traditional artificial neural networks. These features guarantee the model's low complexity and low power consumption. Summary of the Invention

[0006] This invention provides a low-power pulse wave signal arrhythmia classification method and system to ensure low power consumption of the model while achieving classification performance comparable to deep artificial neural networks.

[0007] According to an embodiment of the present invention, a low-power pulse wave signal arrhythmia classification method is provided, comprising the following steps:

[0008] 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.

[0009] Large convolutional kernels and dilated convolutions are used to expand the receptive field and capture local temporal features over a longer range.

[0010] Dimensionality reduction output features, combined with batch normalization, leaky integral-discharge neurons, and max pooling, aggregate local features;

[0011] 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.

[0012] 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.

[0013] Furthermore, 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.

[0014] In making the data pulsed, the data is pulsed to 0 and 1.

[0015] Furthermore, in expanding the receptive field and capturing a longer range of local temporal features by using large convolutional kernels and dilated convolutions, the large convolutional kernels combined with gradually increasing dilation rates are used in the early stage to quickly cover a wide range of temporal correlations and capture the overall pattern of the signal; in the later stage, small convolutional kernels are used to focus on local detailed features and provide discriminative clues for classification.

[0016] Furthermore, self-attention is computed within each subsequence, including query key-value QKV linear projection and spiking neuron activation; the attention module employs a Spikformer-optimized computation order.

[0017] Furthermore, in the process of calculating QKV interactions across the entire sequence, combining linear layers and spiking neurons to capture long-distance global feature associations and mining waveform dependencies at different time points in the pulse wave signal, spiking neurons introduce temporal sparsity to adapt to the dynamic characteristics of time-series signals.

[0018] Furthermore, the pulse wave signal data is standardized before using large convolutional kernels to extract the low-level local features of the original signal.

[0019] Furthermore, dropout is used in each step to prevent model overfitting.

[0020] According to another embodiment of the present invention, a low-power pulse wave signal arrhythmia classification system is provided, comprising:

[0021] 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.

[0022] 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.

[0023] 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;

[0024] 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.

[0025] 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.

[0026] A storage medium storing a program file capable of implementing any of the above-mentioned low-power pulse wave signal arrhythmia classification methods.

[0027] A processor for running a program, wherein the program executes any of the above-mentioned low-power pulse wave signal arrhythmia classification methods during runtime.

[0028] The low-power pulse wave signal arrhythmia classification method and system in this invention employs a pulse activation function instead of the nonlinear activation function in deep neural networks. This replaces the dense floating-point multiplication and accumulation operations in traditional artificial neural networks with sparse floating-point accumulation operations, significantly reducing the computational complexity and power consumption of the algorithm. Furthermore, the model uses dilated convolution combined with a local-global attention mechanism to effectively model the long-term dependencies of pulse wave signals, thus ensuring low power consumption while achieving classification performance comparable to deep artificial neural networks. Attached Figure Description

[0029] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0030] Figure 1 This is a diagram of the pulse wave signal classification model based on a deep pulse neural network in this invention;

[0031] Figure 2 This is a diagram of the backbone network / downsampling module and the pulse residual dilated convolution module in this invention;

[0032] Figure 3 This is a diagram of the attention module in this invention. Detailed Implementation

[0033] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0034] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0035] Example 1

[0036] According to an embodiment of the present invention, a low-power pulse wave signal arrhythmia classification method is provided, comprising the following steps:

[0037] 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.

[0038] Large convolutional kernels and dilated convolutions are used to expand the receptive field and capture local temporal features over a longer range.

[0039] Dimensionality reduction output features, combined with batch normalization, leaky integral-discharge neurons, and max pooling, aggregate local features;

[0040] 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.

[0041] 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.

[0042] The low-power pulse wave signal arrhythmia classification method in this invention uses a pulse activation function instead of the nonlinear activation function in a deep neural network, thereby replacing the dense floating-point multiplication and accumulation operation in a traditional artificial neural network with a sparse floating-point accumulation operation, significantly reducing the computational complexity and power consumption of the algorithm. Furthermore, the model employs dilated convolution combined with a local-global attention mechanism, which can effectively model the long-term dependencies of the pulse wave signal, thus ensuring low power consumption while achieving classification performance comparable to deep artificial neural networks.

[0043] The low-power pulse wave signal arrhythmia classification method of the present invention will be described in detail below with specific embodiments:

[0044] To address the technical problems of existing technologies, this invention proposes a low-power multi-classification method for cardiac arrhythmias based on spiking neural networks. The model uses a spiking activation function instead of the nonlinear activation function in deep neural networks, thereby replacing the dense floating-point multiplication and accumulation operations in traditional artificial neural networks with sparse floating-point accumulation operations, significantly reducing the computational complexity and power consumption of the algorithm. Furthermore, the model employs dilated convolution combined with a local-global attention mechanism, which can effectively model the long-term dependencies of pulse wave signals, thus ensuring low power consumption while achieving classification performance comparable to deep artificial neural networks.

[0045] This invention proposes a pulse wave arrhythmia multi-classification method based on local-global attention spiking neural networks. It employs dilated convolution and self-attention models to effectively model long-term dependencies in PPG time-series data, improving model classification performance. Placing the pulse activation operation before convolution and linear transformation operations ensures that the main data flow of the model does not involve dense floating-point multiplication operations, reducing model power consumption. Specifically, pulse residual dilated convolution captures long-term dependencies in pulse wave signals while enabling more efficient deep training; block-based local self-attention segments sequence features, extracts attention features from each sub-sequence, and fuses these features to capture internal correlation information within the sub-sequences; global self-attention uncovers long-range correlations in the sequence data from a global perspective. Except for the input layer, the data flow is pulsed by pulse activation neurons before entering the convolutional or linear layers, ensuring that the model avoids floating-point multiplication operations during inference, reducing computational complexity and power consumption.

[0046] The technical solution of this invention is described in detail below:

[0047] Reference Figure 1 The implementation steps of this invention are as follows:

[0048] Raw data preprocessing: The pulse wave signal data is standardized to the [0,1] interval;

[0049] Backbone network model ( Figure 2 We use 1D convolution with a large kernel (kernel_size=31) to extract the low-level local features of the original signal. Pooling halves the sequence length to achieve initial dimensionality reduction and reduce subsequent computation. After the convolution operation, we use Leaky Integrate-and-Fire Neurons (LIF spiking neurons) to activate the data, making it pulsed into 0s and 1s, further simplifying the computational complexity.

[0050] The model further employs a pulsed residual dilated convolution module ( Figure 2 This approach utilizes large convolutional kernels and dilated convolutions to expand the receptive field, capturing longer-range local temporal features (such as waveform changes in pulse waves). Residual connections alleviate gradient vanishing, and spiking neurons introduce temporal sparsity. The dilation rate of dilated convolutions gradually increases, capturing more distant local correlations through this progressively larger dilation rate, with the order of increase being 2-4-6-8-10; while the convolutional kernel gradually decreases, with the order of increase being 21-21-13-13-8. In the early stages, large convolutional kernels combined with progressively increasing dilation rates quickly cover a wide range of temporal correlations, capturing the overall pattern of the signal. In the later stages, small convolutional kernels focus on local detailed features, providing discriminative cues for classification.

[0051] downsampling module ( Figure 2 The output features of the front-end pulse residual dilated convolution module are further reduced in dimensionality (sequence length is halved), and combined with batch normalization, leaky integral-discharge neurons, and max pooling, local features are aggregated. This provides a more compact feature input for the subsequent attention module.

[0052] The local self-attention module divides the sequence into blocks ( Figure 3 The model computes self-attention within each subsequence, including query-key-value (QKV) linear projection and spiking neuron activation. Combined with a linear mapping layer, the aggregated sub-attention map focuses the model on feature associations within local subsequences, extracting waveform dependencies within short time windows in the pulse wave signal. The attention module employs a Spikformer-optimized computation order, avoiding the exponential and division operations involved in the softmax operation, thus reducing the model's computational complexity.

[0053] Global self-attention module ( Figure 3 This approach directly computes QKV interactions across the entire sequence, combining linear layers and spiking neurons. It captures long-range global feature associations and mines waveform dependencies across different time periods in the pulse wave signal. Spiking neurons introduce temporal sparsity, adapting to the dynamic characteristics of time-series signals.

[0054] Dropout was used in the backbone network model, downsampling module, pulse residual dilated convolution module, local self-attention module, and global self-attention module to prevent overfitting and ensure model training stability.

[0055] The key points and areas to be protected in this invention are:

[0056] This invention proposes a pulse wave signal classification model based on a deep pulsating neural network, which can be used for population screening of heart disease and for long-term, continuous monitoring deployed in PPG devices. This model can identify high-risk patients early, enabling them to receive anticoagulation therapy as soon as possible, thereby preventing a large number of stroke events.

[0057] (1) Construct a deep spiking neural network model based on pulse residual dilated convolution and local-global self-attention mechanism to classify arrhythmias in pulse wave signals;

[0058] (2) A series of pulse residual dilated convolution modules are used to gradually explore the long-range dependency of the pulse wave signal sequence, while ensuring that the deep spiking neural network can be trained stably.

[0059] (3) A local-global self-attention mechanism based on spiking neurons is adopted to further mine the local and global information of the pulse wave signal. The Spikformer attention map calculation method is used to avoid the high complexity of attention map calculation process in traditional artificial neural networks.

[0060] (4) Before the data stream is input to the convolutional or linear layer, it is pulsed by the pulse activation neurons, so that the model avoids floating-point multiplication during inference, thereby reducing the computational complexity and power consumption of the algorithm.

[0061] Compared with the prior art, the advantages of the present invention are:

[0062] The proposed model is a deep neural network based on spiking neurons. By activating spiking neurons, the data stream is pulsed, allowing the model to replace floating-point multiplication with floating-point addition. This avoids the complex and dense floating-point accumulation and addition operations found in traditional artificial neural networks, effectively reducing the computational complexity and power consumption of the algorithm. Through a series of pulse residual dilated convolutional modules, the long-range dependencies of the pulse wave signal are gradually extracted, and the gradient decay problem in deep networks is avoided through residual connections, thus ensuring the stable learning of the deep spiking neural network. A spikeformer-based pulse attention map calculation method is adopted, avoiding the complex calculation method of attention maps in traditional self-attention modules. The proposed model is the first deep spiking neural network model for multi-class classification of arrhythmias based on pulse wave signals that can be directly trained. It can reduce the computational complexity and power consumption of the model while maintaining its cardiac disease classification performance, which has significant socio-economic value for the popularization of low-cost medical care.

[0063] The proposed solution is theoretically and practically feasible. Experiments were conducted using real pulse wave signal data. The training data included PPG data records from 228 patients, which were randomly divided into training, validation, and test sets at a ratio of 60%, 20%, and 20%. Each patient had several PPG data segments of 2500 characters each. The PPG data were tagged with seven categories: sinus rhythm, premature ventricular contractions (PVCs), premature atrial contractions (PACs), ventricular tachycardia, supraventricular tachycardia, atrial fibrillation, and noise. The remaining six categories of PPG data records, after removing the noise category, were used for model learning.

[0064] In summary, the training, validation, and test sets contain 66,631, 22,511, and 23,989 PPG data fragments, respectively. The model was trained for 50 epochs, and the model with the highest accuracy on the validation set was saved. The sensitivity and specificity of this model on the test set were also reported. The sensitivity and specificity of the deep convolutional neural network VGG13 model on the test set were 74.78% and 97.14%, respectively. The sensitivity and specificity of the proposed solution on the test set were 75.48% and 97.20%, respectively. Therefore, while maintaining low power consumption, the proposed model's pulse wave arrhythmia classification performance is comparable to, or even slightly better than, the current best practice deep convolutional neural network pulse wave classification method. The above experiments verify the effectiveness of the proposed model.

[0065] Example 2

[0066] According to another embodiment of the present invention, a low-power pulse wave signal arrhythmia classification system is provided, comprising:

[0067] 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.

[0068] 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.

[0069] 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;

[0070] 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.

[0071] 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.

[0072] The low-power pulse wave signal arrhythmia classification system in this invention uses a pulse activation function instead of the nonlinear activation function in a deep neural network, thereby replacing the dense floating-point multiplication and accumulation operations in traditional artificial neural networks with sparse floating-point accumulation operations, significantly reducing the computational complexity and power consumption of the algorithm. Furthermore, the model employs dilated convolution combined with a local-global attention mechanism, which can effectively model the long-term dependencies of pulse wave signals, thus ensuring low power consumption while achieving classification performance comparable to deep artificial neural networks.

[0073] The low-power pulse wave signal arrhythmia classification system of the present invention will be described in detail below with specific embodiments:

[0074] To address the technical problems of existing technologies, this invention proposes a low-power pulse wave signal arrhythmia multi-classification system based on spiking neural networks. The model uses a spiking activation function instead of the nonlinear activation function in deep neural networks, thereby replacing the dense floating-point multiplication and accumulation operations in traditional artificial neural networks with sparse floating-point accumulation operations, significantly reducing the computational complexity and power consumption of the algorithm. Furthermore, the model employs dilated convolution combined with a local-global attention mechanism, which can effectively model the long-term dependencies of pulse wave signals, thus ensuring low power consumption while achieving classification performance comparable to deep artificial neural networks.

[0075] This invention proposes a pulse wave arrhythmia multi-classification system based on local-global attention spiking neural networks. It employs dilated convolution and self-attention models to effectively model long-term dependencies in PPG time-series data, improving classification performance. Placing the spiking activation operation before convolution and linear transformation operations ensures that the main data flow of the model does not involve dense floating-point multiplication operations, reducing power consumption. Specifically, spiking residual dilated convolution captures long-term dependencies in pulse wave signals while enabling more efficient deep training; block-based local self-attention segments sequence features, extracts attention features from each sub-sequence, and fuses these features to capture internal correlations within the sub-sequences; global self-attention uncovers long-range correlations in the sequence data from a global perspective. Except for the input layer, the data flow is pulsed by spiking activation neurons before entering the convolutional or linear layers, ensuring that the model avoids floating-point multiplication operations during inference, reducing computational complexity and power consumption.

[0076] The technical solution of this invention is described in detail below:

[0077] Reference Figure 1 The implementation steps of this invention are as follows:

[0078] Raw data preprocessing: The pulse wave signal data is standardized to the [0,1] interval;

[0079] Backbone network model ( Figure 2 We use 1D convolution with a large kernel (kernel_size=31) to extract the low-level local features of the original signal. Pooling halves the sequence length to achieve initial dimensionality reduction and reduce subsequent computation. After the convolution operation, we use Leaky Integrate-and-Fire Neurons (LIF spiking neurons) to activate the data, making it pulsed into 0s and 1s, further simplifying the computational complexity.

[0080] The model further employs a pulsed residual dilated convolution module ( Figure 2This approach utilizes large convolutional kernels and dilated convolutions to expand the receptive field, capturing longer-range local temporal features (such as waveform changes in pulse waves). Residual connections alleviate gradient vanishing, and spiking neurons introduce temporal sparsity. The dilation rate of dilated convolutions gradually increases, capturing more distant local correlations through this progressively larger dilation rate, with the order of increase being 2-4-6-8-10; while the convolutional kernel gradually decreases, with the order of increase being 21-21-13-13-8. In the early stages, large convolutional kernels combined with progressively increasing dilation rates quickly cover a wide range of temporal correlations, capturing the overall pattern of the signal. In the later stages, small convolutional kernels focus on local detailed features, providing discriminative cues for classification.

[0081] downsampling module ( Figure 2 The output features of the front-end pulse residual dilated convolution module are further reduced in dimensionality (sequence length is halved), and combined with batch normalization, leaky integral-discharge neurons, and max pooling, local features are aggregated. This provides a more compact feature input for the subsequent attention module.

[0082] The local self-attention module divides the sequence into blocks ( Figure 3 Self-attention is computed within each subsequence, including query-key-value (QKV) linear projection and spiking neuron activation. Combined with a linear mapping layer, the model focuses 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 attention module adopts the Spikformer[5] optimized computation order, avoiding the exponential and division operations involved in the softmax operation, and reducing the computational complexity of the model.

[0083] Global self-attention module ( Figure 3 This approach directly computes QKV interactions across the entire sequence, combining linear layers and spiking neurons. It captures long-range global feature associations and mines waveform dependencies across different time periods in the pulse wave signal. Spiking neurons introduce temporal sparsity, adapting to the dynamic characteristics of time-series signals.

[0084] Dropout was used in the backbone network model, downsampling module, pulse residual dilated convolution module, local self-attention module, and global self-attention module to prevent overfitting and ensure model training stability.

[0085] The key points and areas to be protected in this invention are:

[0086] This invention proposes a pulse wave signal classification model based on a deep pulsating neural network, which can be used for population screening of heart disease and for long-term, continuous monitoring deployed in PPG devices. This model can identify high-risk patients early, enabling them to receive anticoagulation therapy as soon as possible, thereby preventing a large number of stroke events.

[0087] (1) Construct a deep spiking neural network model based on pulse residual dilated convolution and local-global self-attention mechanism to classify arrhythmias in pulse wave signals;

[0088] (2) A series of pulse residual dilated convolution modules are used to gradually explore the long-range dependency of the pulse wave signal sequence, while ensuring that the deep spiking neural network can be trained stably.

[0089] (3) A local-global self-attention mechanism based on spiking neurons is adopted to further mine the local and global information of the pulse wave signal. The Spikformer attention map calculation method is used to avoid the high complexity of attention map calculation process in traditional artificial neural networks.

[0090] (4) Before the data stream is input to the convolutional or linear layer, it is pulsed by the pulse activation neurons, so that the model avoids floating-point multiplication during inference, thereby reducing the computational complexity and power consumption of the algorithm.

[0091] Compared with the prior art, the advantages of the present invention are:

[0092] The proposed model is a deep neural network based on spiking neurons. By activating spiking neurons, the data stream is pulsed, allowing the model to replace floating-point multiplication with floating-point addition. This avoids the complex and dense floating-point accumulation and addition operations found in traditional artificial neural networks, effectively reducing the computational complexity and power consumption of the algorithm. Through a series of pulse residual dilated convolutional modules, the long-range dependencies of the pulse wave signal are gradually extracted, and the gradient decay problem in deep networks is avoided through residual connections, thus ensuring the stable learning of the deep spiking neural network. A spikeformer-based pulse attention map calculation method is adopted, avoiding the complex calculation method of attention maps in traditional self-attention modules. The proposed model is the first deep spiking neural network model for multi-class classification of arrhythmias based on pulse wave signals that can be directly trained. It can reduce the computational complexity and power consumption of the model while maintaining its cardiac disease classification performance, which has significant socio-economic value for the popularization of low-cost medical care.

[0093] The proposed solution is theoretically and practically feasible. Experiments were conducted using real pulse wave signal data. The training data included PPG data records from 228 patients, which were randomly divided into training, validation, and test sets at a ratio of 60%, 20%, and 20%. Each patient had several PPG data segments of 2500 characters each. The PPG data were tagged with seven categories: sinus rhythm, premature ventricular contractions (PVCs), premature atrial contractions (PACs), ventricular tachycardia, supraventricular tachycardia, atrial fibrillation, and noise. The remaining six categories of PPG data records, after removing the noise category, were used for model learning.

[0094] In summary, the training, validation, and test sets contain 66,631, 22,511, and 23,989 PPG data fragments, respectively. The model was trained for 50 epochs, and the model with the highest accuracy on the validation set was saved. The sensitivity and specificity of this model on the test set were also reported. The sensitivity and specificity of the deep convolutional neural network VGG13 model on the test set were 74.78% and 97.14%, respectively. The sensitivity and specificity of the proposed solution on the test set were 75.48% and 97.20%, respectively. Therefore, while maintaining low power consumption, the proposed model's pulse wave arrhythmia classification performance is comparable to, or even slightly better than, the current best practice deep convolutional neural network pulse wave classification method. The above experiments verify the effectiveness of the proposed model.

[0095] Example 3

[0096] A storage medium storing a program file capable of implementing any of the above-mentioned low-power pulse wave signal arrhythmia classification methods.

[0097] Example 4

[0098] A processor for running a program, wherein the program executes any of the above-mentioned low-power pulse wave signal arrhythmia classification methods during runtime.

[0099] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0100] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0101] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The system embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of units or modules may be electrical or other forms.

[0102] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0103] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0104] If the integrated unit is implemented as 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 solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0105] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

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; residual connections alleviate gradient vanishing, and spiking neurons introduce sparsity in the temporal dimension; the dilation rate of dilated convolutions gradually increases, capturing local connections at greater distances through the gradually increasing dilation rate, with the order of change being 2-4-6-8-10; the convolutional kernels gradually decrease in size, with the order of change being 21-21-13-13-8. 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. QKV interactions are calculated for the entire sequence, and by combining linear layers and spiking neurons, long-distance global feature associations are captured to mine waveform dependencies at different time points in the pulse wave signal. 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; 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. 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.

2. 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.

3. 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.

4. 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.

5. A low-power pulse wave signal arrhythmia classification system, said system being used to implement the low-power pulse wave signal arrhythmia classification method according to claim 1, 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.

6. 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 4.

7. 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 4.

Images