Noninvasive blood pressure monitoring method based on electrocardiosignal and volume pulse wave characteristics

Abstract

The invention provides a non-invasive blood pressure monitoring method based on electrocardiosignal and volume pulse wave characteristics. The method comprises the steps that electrocardiosignals and volume pulse wave signals are collected; preprocessing, and inputting into a feature extraction module network to obtain a monitoring result; the feature extraction module network comprises two branches, namely an electrocardio branch and a volume pulse wave branch; each of the two branches comprises a feature extraction layer and a plurality of levels of structure re-parameterization modules; the later-stage structure re-parameterization module is connected with the previous-stage structure re-parameterization module through the down-sampling module; the structural re-parameterization modules corresponding to all levels are subjected to feature fusion through feature fusion modules, and the output of the previous feature fusion module is input into the next feature fusion module and fused with the output of the structural re-parameterization modules of the two branches of the next level. According to the method, richer physiological signal features can be captured, the robustness and generalization ability of the model are enhanced through cross-modal information complementation, and the detection accuracy is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical electronic detection technology, and more specifically, relates to a non-invasive blood pressure monitoring method based on electrocardiogram signals and volume pulse wave characteristics. Background Art

[0002] As a core physiological indicator of the human cardiovascular system, noninvasive and continuous blood pressure monitoring has become a key tool for preventing and diagnosing cardiovascular diseases such as hypertension and heart failure. With the iterative advancements in biomedical engineering and microelectronics, health assessment systems based on ambulatory blood pressure monitoring are becoming a standard solution for cardiovascular disease prevention.

[0003] Currently, mainstream blood pressure collection technologies mainly present two technical routes: invasive arterial catheter measurement and non-invasive cuff measurement. The invasive arterial catheter measurement method involves inserting an invasive arterial catheter into the radial artery or aorta, allowing the pressure sensor to directly contact the blood and monitor blood pressure changes in real time. It can achieve medical-grade accuracy of ±2mmHg (compliant with ISO 81060-2 standard), but there is a risk of endothelial damage and is only applicable in intensive care unit (ICU) scenarios. The cuff measurement method flattens the brachial artery by inflating the cuff, obstructing blood flow, and then gradually releases the pressure. Professionals then use a stethoscope to identify the process of inflation and deflation. Although the cuff measurement method has the advantage of being non-invasive, it is limited by the principle of Korotkoff sound detection. A single measurement requires an interval period of more than 5 minutes, making it difficult to capture the transient fluctuation characteristics of blood pressure.

[0004] In recent years, wrist-worn devices based on photoplethysmography (PPG) have enabled continuous monitoring. Their core technology relies on a linear regression model between pulse transit time (PTT) and systolic / diastolic blood pressure. However, due to individual physiological differences such as changes in vascular elastic modulus and postural compensatory responses, dynamic measurement errors can reach ±15 mmHg. While the current mainstream CNN-LSTM hybrid neural network can extract joint time-frequency domain features of volumetric pulse wave signals (sampling rate ≥ 100 Hz), it faces bottlenecks in embedded deployment, such as a large number of model parameters (up to 21M in the ResNet-34 architecture) and high inference latency (over 200 ms for RNN units).

[0005] Existing AI models for blood pressure monitoring are relatively simple and cannot effectively mine the classification features of volumetric pulse wave signals. The relevant features of volumetric pulse wave signals, such as waveform amplitude, period, and morphology, are closely related to the body's systolic and diastolic blood pressure. Furthermore, traditional feature engineering struggles to effectively capture the nonlinear relationship between blood pressure and multimodal physiological signals (such as volumetric pulse waves and electrocardiogram signals).

[0006] Deep learning technology offers a new path to addressing these issues. Through an end-to-end neural network architecture, blood pressure mapping relationships can be learned directly from raw physiological signals. For example, a three-dimensional convolutional network can simultaneously process volumetric pulse waveforms, ECG rhythms, and accelerometer data. Multimodal fusion methods can reduce prediction errors to ±5 mmHg. However, the model's computational overhead reaches 3.2 GFLOPs, making it difficult to run in real time. Summary of the Invention

[0007] In order to overcome the shortcomings and deficiencies in the prior art, the purpose of the present invention is to provide a non-invasive blood pressure monitoring method based on electrocardiogram (ECG) signals and volume pulse wave features. This method combines the multimodal fusion technology of ECG signals and volume pulse wave signals to capture richer physiological signal features. It also enhances the robustness and generalization ability of the model through cross-modal information complementation, significantly improving the detection accuracy.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a non-invasive blood pressure monitoring method based on electrocardiogram (ECG) and volume pulse wave features, wherein the ECG and volume pulse wave signals of a monitored person are collected; the ECG and volume pulse wave signals are preprocessed separately and then simultaneously input into a feature extraction module network to obtain monitoring results;

[0009] The feature extraction module network includes two branches, namely the electrocardiogram branch and the volume pulse wave branch; both branches include a feature extraction layer and several levels of structural reparameterization modules; the latter level structural reparameterization module is connected to the previous level structural reparameterization module through a downsampling module to gradually reduce the feature dimension; the number of structural reparameterization modules in the two branches is the same and corresponds one to one; the structural reparameterization modules corresponding to each level are respectively subjected to feature fusion through the feature fusion module, and the output of the previous feature fusion module is input into the latter feature fusion module and fused with the output of the structural reparameterization modules of the two branches of the latter level; the fusion feature output by the last feature fusion module is processed by the pooling layer and the regularization layer, and classified by the classifier to obtain the monitoring result.

[0010] Preferably, the feature extraction layer includes two layers of feature convolution one, a ReLU activation function layer one connected between the two layers of feature convolution one, and a feedforward network one connected to the back side of the second layer of feature convolution one.

[0011] Preferably, each level of the structural reparameterization module includes a RepDSABlock block and a RepNetBlock block connected in sequence;

[0012] The RepDSABlock block includes a depthwise separable convolution module 1, an attention enhancement module with a depthwise separable attention mechanism, and a feedforward network 2; the RepNetBlock block includes a depthwise separable convolution module 2 and a feedforward network 3;

[0013] Each level of downsampling module includes: depth convolution layer two, feature convolution layer two and feedforward network four.

[0014] Preferably, the feedforward network 1, feedforward network 2, feedforward network 3 and feedforward network 4 all include: two layers of 1×1 feature convolution three, and a ReLU activation function layer two connected between the two layers of feature convolution three.

[0015] Preferably, the depthwise separable convolution layer 1 of the RepDSABlock block and the depthwise separable convolution layer 2 of the RepNetBlock block both refer to a multi-branch structure consisting of a parallel 3×3 depthwise convolution layer 1, a 1×1 point-by-point convolution layer and an identity mapping branch.

[0016] Preferably, the feature extraction module network refers to a trained feature extraction module network;

[0017] During training, the weights of the multi-branch structure are initialized to empty;

[0018] 3×3 depth convolution layer - extract convolution weights W 3×3 and bias b 3×3 ; 1×1 point-by-point convolution layer extracts convolution weight W 1×1 and bias b 1×1 , pad the 1×1 convolution kernel to a 3×3 format:

[0019] W 1×1 ← Zero padding (W 1×1 ,3×3);

[0020] The identity mapping branch constructs the identity mapping convolution kernel:

[0021] W identity ←Zero-fill(I,3×3),b identity =0

[0022] Where I is the identity matrix;

[0023] After the fusion of 3×3 depth convolution layer 1, 1×1 point-by-point convolution layer and identity mapping branch, the convolution kernel weight W final and bias b final Expressed as:

[0024] W final ←W 3×3 +W 1×1 +W identity ;

[0025] b final ←b 3×3 +b 1×1 +b identity ;

[0026] After training, the multi-branch structure is converted into a single-branch inference structure containing only a 3×3 depth convolution layer, and the weight W final and bias b final Set as the weights and bias of the 3×3 depthwise convolutional layer one.

[0027] Preferably, the feature fusion module includes a 1×1 feature convolution layer 4, a ReLU activation function layer 3 and a BN layer connected in sequence.

[0028] Preferably, the preprocessing includes: signal downsampling, signal normalization, and filter noise reduction;

[0029] The signal downsampling refers to: uniformly downsampling the electrocardiogram signal and the volume pulse wave signal to the same sampling frequency;

[0030] The signal normalization refers to: unifying the electrocardiogram signal and the volume pulse wave signal to the same scale;

[0031] The filter noise reduction refers to using a second-order Butterworth filter algorithm to remove baseline drift and power frequency interference noise from the original electrocardiogram signal and volume pulse wave signal respectively.

[0032] A readable storage medium, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the non-invasive blood pressure monitoring method based on electrocardiogram signals and volume pulse wave characteristics.

[0033] A computer device includes a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the non-invasive blood pressure monitoring method based on electrocardiographic signals and volume pulse wave characteristics is implemented.

[0034] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0035] 1. This paper uses a structurally reparameterized network based on a hybrid CNN and Transformer architecture, combined with multimodal fusion technology of electrocardiogram (ECG) and volume pulse wave (PPG) signals. Compared with traditional single-channel ECG signal detection methods, this technology significantly improves detection accuracy. This technology not only captures richer physiological signal features, but also enhances the robustness and generalization ability of the model through cross-modal information complementation, providing more reliable support for clinical diagnosis.

[0036] 2. The present invention introduces a hierarchical feature fusion mechanism to fuse the multi-scale feature information of the ECG signal and the volume pulse wave signal, thereby improving the expression ability and classification performance of the model. First, the model extracts the initial features from the ECG signal and the volume pulse wave signal respectively, and through multi-level downsampling operations, gradually reduces the feature dimension to reduce the amount of calculation while retaining key semantic information. After each level of downsampling, the feature fusion module fuses the local features from different scales, the global representation and the output of the previous semantic layer to achieve information complementarity between features. Among them, the global branch introduces a self-attention mechanism to model long-range dependencies in the spatial and temporal dimensions, thereby enhancing the model's perception of the overall signal structure. Subsequently, the fused features are weighted and optimized through the channel attention mechanism, so that the model can adaptively focus on the discriminative feature channels, further improving the expression effect of specific semantic information. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a principle block diagram of the feature extraction module network of the present invention;

[0038] Figure 2 It is a principle block diagram of the preprocessing of the present invention;

[0039] Figure 3 is a comparison diagram of the ECG signal before and after preprocessing of the present invention;

[0040] Figure 4 This is a principle block diagram of the high-precision non-invasive blood pressure monitoring system of the present invention;

[0041] Figure 5 is a circuit schematic diagram of the power management module of the present invention;

[0042] Figure 6 It is a circuit schematic diagram of the electrocardiogram data acquisition module of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0044] Example 1

[0045] This embodiment is a non-invasive blood pressure monitoring method based on ECG signal and volume pulse wave features, which collects the ECG signal and volume pulse wave signal of the monitored person; the ECG signal and volume pulse wave signal are pre-processed separately and then simultaneously input into the feature extraction module network to obtain the monitoring results. Figure 1 shown.

[0046] Preprocessing includes: signal downsampling, signal normalization, filter noise reduction, such as Figure 2 shown.

[0047] Signal downsampling refers to uniformly downsampling the ECG signal and volume pulse wave signal to the same sampling frequency (for example, 1 kHz). This sampling rate effectively reduces the amount of data while preserving the morphological characteristics of the original ECG and volume pulse wave signals, shortening the computational complexity and training convergence time of the model, making it more conducive to model deployment and application.

[0048] Signal normalization refers to unifying the ECG signal and the volume pulse wave signal to the same scale. Since the amplitude distribution of data collected by different hospitals varies greatly, this can lead to difficulties in fitting the model during training, which in turn affects the convergence speed. Therefore, the present invention uses normalization processing, and its formula is as follows:

[0049]

[0050] Where x_norm is the raw signal data, x_min and x_max are the minimum and maximum values of the input signal, respectively. Normalization can bring data from different sources to the same scale, reducing the impact of data distribution differences on model training and improving the model's generalization and convergence speed.

[0051] The filter noise reduction refers to: using the second-order Butterworth filtering algorithm to remove the baseline drift and power frequency interference noise from the original ECG signal and volume pulse wave signal respectively. Baseline drift is usually caused by respiratory movement or poor electrode contact, while power frequency interference comes from 50Hz or 60Hz AC interference from the power system. Through the application of the Butterworth filter, these noises can be effectively filtered out, allowing doctors to make clearer diagnoses of the ECG and pulse wave on the signal display UI interface. Comparison of ECG signals before and after the preprocessing module is shown in Figure 3 As shown in the figure, the preprocessed signal significantly reduces noise interference and improves signal readability and diagnostic accuracy.

[0052] This invention is based on and improves the deep learning network framework RepVGG. RepVGG is an efficient convolutional neural network architecture. Its core concept is to transform the multi-branch structure during training into a single-branch inference structure by fusing the convolutional layers in the model with BN (batch normalization) layers, thereby significantly improving the model's inference speed and computational efficiency. In this patent, we have made adaptive improvements to RepVGG by changing the pooling layer in the network to a one-dimensional structure, enabling it to better handle feature extraction tasks for one-dimensional signals, thereby constructing a backbone processing model suitable for ECG signals and volume pulse wave signals.

[0053] In addition, most existing models only use a single-channel ECG signal and fail to fully utilize the physiological features in the volume pulse wave signal. To make up for this shortcoming, this patent proposes a dual-channel signal fusion strategy, which is to introduce a Transformer module in the classification layer of the dual-channel model of ECG signal and volume pulse wave signal. Transformer can capture long-range dependencies in the signal through the self-attention mechanism, thereby significantly improving the feature extraction capability. This not only enhances the model's ability to mine signal features, but also adapts to the real-time reasoning requirements during deployment, providing a more efficient and accurate solution for blood pressure detection.

[0054] The Transformer module introduced in this paper is the HiFuse module, or the feature fusion module of the present invention. This module serves as the core component for fusing ECG and volumetric pulse wave signals, and constructs a dual-path backbone extraction network, RepVGG-FuseNet. It extracts multi-level information from ECG and volumetric pulse wave signals from global and local feature blocks, and effectively fuses these features through a hierarchical feature fusion block. It adaptively fuses local features from different levels, global representations, and semantic information from previous levels, thereby improving the richness and accuracy of feature representation. During the downsampling step, the model reduces computational complexity by reducing feature dimensionality while retaining key information. The global feature block in the feature fusion module utilizes a self-attention mechanism to capture global information in space and time, further enhancing the model's overall understanding of the signal. Furthermore, the feature fusion module feeds global features into a channel attention mechanism, optimizing feature representations for specific semantics by leveraging the interdependencies between channel mappings. This design significantly improves the fusion of ECG and volumetric pulse wave features, solving the subsequent problem of high-quality feature output for classification.

[0055] The input of the feature extraction module network is the ECG signal and the volume pulse wave signal. The input tensor size is B×C×L, where B represents the batch size, C is the number of channels, and L is the signal length.

[0056] The feature extraction module network structure consists of two branches: an electrocardiogram (ECG) branch and a volume pulse wave (VPW) branch. Both branches include a feature extraction layer and several levels of structural reparameterization modules. The latter level of structural reparameterization modules is connected to the previous level of structural reparameterization modules via a downsampling module to gradually reduce the feature dimension. The present invention has four levels of structural reparameterization modules, and through downsampling operations, the sequence length is gradually compressed to L / 4, L / 8, L / 16, and L / 32, while the number of channels is increased to extract multi-scale features.

[0057] The number of structural reparameterization modules of the two branches is the same and corresponds one to one; the structural reparameterization modules corresponding to each level are respectively fused through the feature fusion module, and the output of the previous feature fusion module is input into the next feature fusion module and fused with the output of the structural reparameterization modules of the two branches of the next level; the fusion features output by the last feature fusion module are processed by the pooling layer and the regularization layer, and classified by the classifier to obtain the monitoring results.

[0058] The feature extraction layer includes two layers of feature convolution, a ReLU activation function layer connected between the two layers of feature convolution, and a feedforward network (FFN) connected to the back side of the second layer of feature convolution.

[0059] Each level of structural reparameterization module includes a RepDSABlock block and a RepNetBlock block connected in sequence;

[0060] The RepDSABlock block includes a depthwise separable convolutional module (DW Conv), an attention enhancement module (DSA) with a depthwise separable attention mechanism, and a feedforward network (FFN). The RepNetBlock block includes a depthwise separable convolutional module (DW Conv) and a feedforward network (FFN). The attention enhancement module introduces the deep separable attention mechanism (DSA) and uses 3×3 and 1×1 DW convolutions combined with the FFN to perform feature dimensionality reduction and reorganization, improving the model's ability to model both local details and global features.

[0061] Each level of downsampling module includes: depth convolution layer 2 (DW Conv), feature convolution layer 2 and feed-forward network 4 (FFN).

[0062] Feedforward network 1, feedforward network 2, feedforward network 3, and feedforward network 4 all include: two layers of 1×1 feature convolution three, and a ReLU activation function layer two connected between the two layers of feature convolution three.

[0063] The depthwise separable convolution layer 1 of the RepDSABlock block and the depthwise separable convolution layer 2 of the RepNetBlock block both refer to a multi-branch structure consisting of a parallel 3×3 depthwise convolution layer 1, a 1×1 point-by-point convolution layer, and an identity mapping branch.

[0064] The feature extraction module network refers to a trained feature extraction module network;

[0065] During training, the weights of the multi-branch structure are initialized to empty;

[0066] 3×3 depth convolution layer - extract convolution weights W 3×3 and bias b 3×3 ; 1×1 point-by-point convolution layer extracts convolution weight W 1×1 and bias b1×1 , pad the 1×1 convolution kernel to a 3×3 format:

[0067] W 1×1 ← Zero padding (W 1×1 ,3×3);

[0068] The identity mapping branch constructs the identity mapping convolution kernel:

[0069] W identity ←Zero-fill(I,3×3),b identity =0

[0070] Where I is the identity matrix;

[0071] After the fusion of 3×3 depth convolution layer 1, 1×1 point-by-point convolution layer and identity mapping branch, the convolution kernel weight W final and bias b final Expressed as:

[0072] W final ←W 3×3 +W 1×1 +W identity ;

[0073] b final ←b 3×3 +b 1×1 +b identity ;

[0074] After training, the multi-branch structure is converted into a single-branch inference structure containing only a 3×3 depth convolution layer, and the weight W final and bias b final Set as the weights and bias of the first 3×3 depthwise convolution layer. In the multi-branch structure design, the present invention draws on the depthwise separable convolution of MobileNetV3 and moves it up to the multi-branch part, specifically including 3×3 depthwise convolution, 1×1 pointwise convolution and identity mapping branch. Through structural reparameterization, an efficient conversion from multi-branch during training to single branch during inference is achieved, significantly reducing the computational complexity. In addition, an attention mechanism is introduced after each branch to enhance the feature expression capability, enabling the model to more accurately capture the key information in the signal. At the same time, the module retains the advantages of the original residual connection, effectively prevents model degradation through identity mapping, and ensures the stability of training. The improved RepDSABlock module not only inherits the efficient inference characteristics of the traditional convolutional network, so that only a single-branch 3×3 depthwise convolution needs to be retained during inference, but also improves the feature extraction capability through the multi-branch attention mechanism during the training stage. This feature makes the module particularly suitable for modeling the complex nonlinear relationship between electrocardiogram signals and volume pulse wave signals.

[0075] The feature fusion module (HiFuse) consists of four 1×1 feature convolutions, three ReLU activation function layers, and a batch normalization layer connected in sequence. The feature fusion module integrates multi-scale feature information from the ECG and volumetric pulse wave signals by introducing a hierarchical feature fusion mechanism, improving the model's expressiveness and classification performance. First, the model extracts initial features from the ECG and volumetric pulse wave signals respectively. Through multiple levels of downsampling, the feature dimensionality is gradually reduced to reduce computational complexity while preserving key semantic information. After each level of downsampling, the feature fusion module fuses local features from different scales, the global representation, and the output of the previous semantic layer to achieve information complementarity between features. The global branch introduces a self-attention mechanism to model long-range dependencies in spatial and temporal dimensions, enhancing the model's perception of the overall signal structure. The fused features are then weighted and optimized using a channel-wise attention mechanism, enabling the model to adaptively focus on discriminative feature channels, further improving the representation of specific semantic information.

[0076] It is based on a high-precision non-invasive blood pressure monitoring system; the system, such as Figure 4 As shown, it mainly includes:

[0077] Power management module: provides stable power supply;

[0078] ECG data acquisition module: collects analog biometric signals;

[0079] MCU main control module: converts analog signals into digital signals;

[0080] AI algorithm module: integrates the features of digital signals through neural networks;

[0081] Human-computer interaction module: Matching the features classified by the classifier with the database;

[0082] Data display module: Prints the processed information through the external interface.

[0083] The functions of each module of the system are as follows:

[0084] Power management module: aims to provide stable power supply support for the acquisition equipment. Specifically, the system uses an external battery as the power input, and converts AC power into DC power through the power management circuit. First, the input AC voltage is converted into a 12V battery voltage; then, the TPS564201DCDC chip is used as a step-down module to gradually reduce the battery voltage from 12V to 5V, and further to 2.4V to meet the power supply requirements of the acquisition chip. In addition, the system uses the LM1117 LDO linear voltage regulator chip to stabilize the voltage at 3.3V, providing reliable power for peripherals such as the MCU main control chip, Bluetooth module, buttons and LED lights. In order to improve the power quality, the present invention also introduces large-capacitance capacitors to effectively filter out noise in the input voltage and ensure the stability and reliability of the system operation. The specific circuit modules are as follows: Figure 5 shown.

[0085] ECG data acquisition module: ADS1293 is used as the ECG signal processing chip. This chip is a 12-bit analog-to-digital converter (ADC) with 5-channel analog signal input capability and supports three-lead ECG detection. It has the advantages of low power consumption (0.3mW per channel), low input noise (7μVpp) and high transmission frequency (25.6ksps), and can efficiently collect and process ECG signals. In addition, the volume pulse wave signal processing adopts the high-precision AFE4490 blood oxygen acquisition chip, which has significant advantages such as low noise (50pA RMS), high sampling time (50μs to 4ms) and wide dynamic range (110dB), ensuring the accurate acquisition of blood oxygen signals. Through the stable voltage provided by the power management module, the above chips can work together to achieve high-precision, low-noise multi-modal physiological signal acquisition, providing a reliable data foundation for subsequent signal processing and analysis. Specific circuit modules such as Figure 6 shown.

[0086] MCU Main Control Module: In practical applications, sampling frequency, interference levels, and signal amplitudes vary significantly across different hospital environments, adversely affecting the collected physiological signals. To address this issue, the present invention utilizes an STM32F103RCT6 microcontroller as the main signal processing control unit, achieving efficient signal processing while ensuring low cost. The MCU main control module performs preprocessing on the input ECG and pulse wave signals, including signal downsampling, signal normalization, and filter noise reduction.

[0087] The preprocessed signal is input to the edge inference computing card via the SPI (Serial Peripheral Interface) or I2C (Integrated Circuit Bus) transmission protocol for real-time on-board processing. SPI and I2C are two commonly used high-speed serial communication protocols that enable efficient data transmission and are suitable for embedded systems and edge computing scenarios. The high-computing power of the inference computing card not only solves the problem of slow data processing but also improves the real-time performance and response speed of the model, thereby meeting the real-time diagnosis requirements of medical equipment.

[0088] The AI algorithm module executes the feature extraction module network of the present invention to achieve feature extraction.

[0089] Human-computer interaction module: As the back-end display core module of the system, it enables doctors to send instructions to the system background through the UI interface. It also displays the electrocardiogram, pulse wave and patient-related information, as well as the model's prediction results, to assist doctors in pre-excitation bypass localization diagnosis. The human-computer interaction module automatically parses the electrocardiogram and pulse wave data, and presents it in the form of a waveform on the data display module UI interface. Doctors can zoom in, out or pan the waveform using the mouse wheel or touch operation to observe the signal characteristics in more detail. In addition, doctors can trigger the classification function, and the system will automatically call the pre-trained deep learning model to analyze the electrocardiogram, and display the confidence of the diagnosis result and the probability distribution of the model for each category in real time.

[0090] Example 2

[0091] This embodiment provides a readable storage medium, wherein the readable storage medium stores a computer program. When the computer program is executed by a processor, the processor executes the non-invasive blood pressure monitoring method based on electrocardiogram signals and volume pulse wave characteristics described in Example 1.

[0092] Example 3

[0093] This embodiment provides a computer device, including a processor and a memory for storing a program executable by the processor. When the processor executes the program stored in the memory, the non-invasive blood pressure monitoring method based on electrocardiogram signals and volume pulse wave characteristics described in the first embodiment is implemented.

[0094] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A non-invasive blood pressure monitoring method based on electrocardiographic signals and volume pulse wave characteristics, characterized by: Collecting the ECG signal and volume pulse wave signal of the monitored person; pre-processing the ECG signal and volume pulse wave signal separately, and then simultaneously inputting them into the feature extraction module network to obtain the monitoring results; The feature extraction module network includes two branches, namely the electrocardiogram branch and the volume pulse wave branch; both branches include a feature extraction layer and several levels of structural reparameterization modules; the latter level structural reparameterization module is connected to the previous level structural reparameterization module through a downsampling module to gradually reduce the feature dimension; the number of structural reparameterization modules in the two branches is the same and corresponds one to one; the structural reparameterization modules corresponding to each level are respectively subjected to feature fusion through the feature fusion module, and the output of the previous feature fusion module is input into the latter feature fusion module and fused with the output of the structural reparameterization modules of the two branches of the latter level; the fusion feature output by the last feature fusion module is processed by the pooling layer and the regularization layer, and classified by the classifier to obtain the monitoring result.

2. The non-invasive blood pressure monitoring method based on electrocardiographic signal and volume pulse wave characteristics according to claim 1, characterized in that: The feature extraction layer includes two layers of feature convolution one, a ReLU activation function layer one connected between the two layers of feature convolution one, and a feedforward network one connected to the back side of the second layer of feature convolution one.

3. The non-invasive blood pressure monitoring method based on electrocardiographic signal and volume pulse wave characteristics according to claim 2, characterized in that: Each level of structural reparameterization module includes a RepDSABlock block and a RepNetBlock block connected in sequence; The RepDSABlock block includes a depthwise separable convolution module 1, an attention enhancement module with a depthwise separable attention mechanism, and a feedforward network 2; the RepNetBlock block includes a depthwise separable convolution module 2 and a feedforward network 3; Each level of downsampling module includes: depth convolution layer two, feature convolution layer two and feedforward network four.

4. The non-invasive blood pressure monitoring method based on electrocardiographic signal and volume pulse wave characteristics according to claim 3, characterized in that: The feedforward network 1, feedforward network 2, feedforward network 3 and feedforward network 4 all include: two layers of 1×1 feature convolution three, and a ReLU activation function layer two connected between the two layers of feature convolution three.

5. The non-invasive blood pressure monitoring method based on electrocardiographic signal and volume pulse wave characteristics according to claim 3, characterized in that: The depthwise separable convolution layer 1 of the RepDSABlock block and the depthwise separable convolution layer 2 of the RepNetBlock block both refer to a multi-branch structure consisting of a parallel 3×3 depthwise convolution layer 1, a 1×1 point-by-point convolution layer, and an identity mapping branch.

6. The non-invasive blood pressure monitoring method based on electrocardiographic signal and volume pulse wave characteristics according to claim 5, characterized in that: The feature extraction module network refers to a trained feature extraction module network; During training, the weights of the multi-branch structure are initialized to empty; 3×3 depth convolution layer - extract convolution weights W 3×3 and bias b 3×3 ; 1×1 point-by-point convolution layer extracts convolution weight W 1×1 and bias b 1×1 , pad the 1×1 convolution kernel to a 3×3 format: W 1×1 ← Zero padding (W 1×1 ,3×3); The identity mapping branch constructs the identity mapping convolution kernel: W identity ←Zero-fill(I,3×3),b identity =0 Where I is the identity matrix; After the fusion of 3×3 depth convolution layer 1, 1×1 point-by-point convolution layer and identity mapping branch, the convolution kernel weight W final and bias b final Expressed as: IN final ←In 3×3 +W 1×1 +W identity ; b final ←b 3×3 +b 1×1 +b identity ; After training, the multi-branch structure is converted into a single-branch inference structure containing only a 3×3 depth convolution layer, and the weight W final and bias b final Set as the weights and bias of the 3×3 depthwise convolutional layer one.

7. The non-invasive blood pressure monitoring method based on electrocardiographic signal and volume pulse wave characteristics according to claim 1, characterized in that: The feature fusion module includes a 1×1 feature convolution layer 4, a ReLU activation function layer 3 and a BN layer connected in sequence.

8. The non-invasive blood pressure monitoring method based on electrocardiographic signal and volume pulse wave characteristics according to claim 1, characterized in that: The preprocessing includes: signal downsampling, signal normalization, and filter noise reduction; The signal downsampling refers to: uniformly downsampling the electrocardiogram signal and the volume pulse wave signal to the same sampling frequency; The signal normalization refers to: unifying the electrocardiogram signal and the volume pulse wave signal to the same scale; The filter noise reduction refers to using a second-order Butterworth filter algorithm to remove baseline drift and power frequency interference noise from the original electrocardiogram signal and volume pulse wave signal respectively.

9. A readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, enables the processor to perform the non-invasive blood pressure monitoring method based on electrocardiographic signals and volume pulse wave features according to any one of claims 1 to 8.

10. A computer device comprising a processor and a memory for storing a program executable by the processor, characterized in that: When the processor executes the program stored in the memory, the non-invasive blood pressure monitoring method based on electrocardiographic signal and volume pulse wave characteristics according to any one of claims 1 to 8 is implemented.

Images