A BCG signal blood pressure calculation method based on transfer learning and attention mechanism
By combining attention-enhanced deep separable convolutional neural networks and transfer learning, the problems of difficult data collection and noise interference in BCG signal blood pressure calculation are solved, and high-precision blood pressure monitoring of single-channel BCG signals is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LIAONING UNIVERSITY
- Filing Date
- 2026-04-08
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for calculating blood pressure using BCG signals suffer from difficulties in data collection, high signal quality requirements, poor robustness, and the high requirements of the subject's posture for multi-channel methods, making it difficult to effectively resist noise interference from breathing, body movement, and conversation.
By combining attention-enhanced deep separable convolutional neural networks with transfer learning, single-channel BCG signals are processed through filtering and exponential shift normalization. The model is trained on a heart rate dataset, and the weights are fine-tuned on a blood pressure dataset using transfer learning to achieve high-precision blood pressure monitoring of single-channel BCG signals.
Without strictly requiring the subject's posture, it can effectively resist noise interference and achieve high-precision monitoring of diastolic and systolic blood pressure in single-channel BCG signals, thus improving prediction accuracy.
Smart Images
Figure CN122096779A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of vital sign monitoring and relates to a method for calculating blood pressure based on BCG signals using transfer learning and attention mechanisms. Background Technology
[0002] A ballistocardiogram (BCG) is a signal generated by measuring minute changes in pressure on the human body surface using sensors. It contains information about heartbeat, respiration, and movement, and can be used as a non-contact method for monitoring cardiac activity.
[0003] Current methods for calculating blood pressure based on BCG signals still face several challenges: collecting blood pressure data for model training is difficult, high signal quality is required, and existing BCG signal-based blood pressure prediction methods are not robust and are susceptible to noise interference. Furthermore, most methods require multiple channels for blood pressure prediction, which places high demands on the subject's posture. Therefore, it is essential to develop a single-channel BCG signal-based blood pressure calculation method that can effectively resist noise interference from breathing, body movement, and conversation. Summary of the Invention
[0004] The purpose of this invention is to provide a method for calculating blood pressure based on BCG signals using transfer learning and attention mechanisms. This method uses attention-enhanced deep separable convolutional neural networks and transfer learning to calculate blood pressure more accurately.
[0005] This invention is achieved through the following technical solution: a method for calculating blood pressure based on BCG signals using transfer learning and attention mechanisms, the steps of which are as follows: Step 1) Filter the original BCG signal to obtain the filtered BCG signal BCG_F; The specific steps are as follows: BCG = {x(1),x(2),x(3),…,x(n)} BCG_F = Filter_BCG(BCG) The Filter_BCG function uses a Butterworth filter with a bandwidth of 0.8Hz-35Hz to filter the BCG signal, removing low-frequency signals below 0.8Hz and high-frequency signals above 35Hz. The result is denoted as BCG_F.
[0006] Step 2) Perform exponential shift normalization on the BCG_F signal to obtain BCG_stand; The specific steps are as follows: BCG_stand = scale_BCG(BCG_F) The scale_BCG function is composed of the following formula:
[0007] Wherein, EMA_current is the mean of the exponential moving average at the current time point, EMA_previous is the mean of the exponential moving average at the previous time point, EMA_sdc is the standard deviation of the exponential moving average at the current time point, EMA_sdp is the standard deviation of the exponential moving average at the previous time point, α is the decay factor, and value is the data value at the current time point. The resulting z_scores are arranged in order to form the BCG_stand sequence.
[0008] Step 3) Input BCG_stand into the attention-enhanced deep separable convolutional neural network, train it on the heart rate dataset, and save the optimal heart rate model weights; The specific structure of the attention-enhanced deep separable convolutional neural network is as follows: First, the BCG_stand signal is input into a one-dimensional feature extraction network, which includes an initial convolutional layer, consisting of a one-dimensional convolutional layer, a batch normalization layer, and an activation function connected in sequence. This layer is used to perform preliminary feature extraction on the above signal, and the result is denoted as BCG_CB. BCG_CB is input into the multi-level feature extraction module for further feature extraction. The multi-level feature extraction module contains six sequentially connected stages. The first, second, and third stages use Fused-MBConv blocks, while the fourth, fifth, and sixth stages use one-dimensional MBConv blocks. The above six stages ultimately increase the number of channels to 256. The 256-dimensional features output from the sixth stage are expanded to 512 dimensions through one-dimensional convolutional layers, batch normalization layers, and activation functions. Then, the resulting 512-dimensional features are pooled and flattened. The above features are then processed through two fully connected layers to obtain the final predicted value.
[0009] Attention-enhanced deep separable convolutional neural networks use Fused-MBConv blocks, specifically as follows: After receiving the input feature signal, the system first determines whether to enable the residual connection path based on the matching relationship between the number of input channels and the number of target output channels and the preset convolution stride value. Then, it enters the feature transformation stage: when the expansion coefficient is not 1, spatial feature extraction and channel dimension expansion are completed simultaneously through a fused convolution to generate an intermediate feature signal; when the expansion coefficient is 1, the input signal is directly transmitted. Then, the channel dimension of the intermediate feature signal is projected and compressed to the target output dimension through convolution. Finally, according to the initial determination result: if the residual connection condition is met, the original input signal and the projected output are added element by element; if the condition is not met, the projection result is directly output.
[0010] Attention-enhanced deep separable convolutional neural networks use one-dimensional MBConv blocks, specifically as follows: First, it checks if the residual connection condition is met: residual paths are enabled when the number of input channels equals the number of output channels and the convolution stride is 1. The processing flow begins with the channel expansion stage. When the expansion coefficient is greater than 1, the number of input channels is expanded to the hidden dimension through convolution; when the expansion coefficient is 1, the input features are directly passed. Next, depthwise separable convolution is performed to extract spatial features from the expanded features in channel grouping mode. Then, the channel attention weighting stage begins. After attention weighting is completed, the hidden dimensions are projected and compressed to the target number of output channels through convolution. Finally, based on the initial conditions, if residual connection is enabled, the original input signal is added to the processing result and output; otherwise, the processing result is output directly.
[0011] Attention-enhanced deep separable convolutional neural networks use SE attention in one-dimensional MBConv, specifically as follows: First, global average pooling is performed on each channel of the deep convolution output to obtain a channel-level statistical description. Then, the importance weights of each channel are learned through a bottleneck structure containing two linear layers. This structure first compresses the channel dimension to 1 / reduction and then restores it to the original dimension. During this process, an activation function is used to generate normalized weights between 0 and 1. Finally, the learned weights are multiplied with the original features channel by channel to achieve feature recalibration.
[0012] Step 4) In the blood pressure dataset, using the aforementioned heart rate model weights as initial weights, the model is fine-tuned through a transfer learning strategy to obtain the prediction model weights for diastolic and systolic blood pressure. Step 4) uses a transfer learning method, specifically: Pre-training was performed on the heart rate dataset, using the weights of the heart rate model as the initial weights for the blood pressure model. A diastolic blood pressure model was trained on the diastolic blood pressure dataset, and a systolic blood pressure model was trained on the systolic blood pressure dataset. During training, the feature extraction layer maintained a learning rate of 10. -4The fully connected layer maintains a learning rate of 10. -3 The learning rate is adjusted through the learning rate scheduler.
[0013] Step 5) Using the obtained prediction model weights, the single-channel BCG signal is input into the model with the obtained blood pressure prediction model weights for prediction, thereby obtaining the diastolic and systolic blood pressure corresponding to the single-channel BCG signal.
[0014] The beneficial effects of this invention are as follows: The present invention, through the above method, enables high-precision, non-intrusive blood pressure monitoring of single-channel BCG signals with noise interference from breathing, body movement, and conversation, even when there are no strict requirements on the subject's posture. This monitoring can be achieved using an attention-enhanced deep separable convolutional neural network that combines one-dimensional MBConv blocks with Fused-MBConv blocks and transfer learning methods. Both diastolic and systolic blood pressure can be monitored. Attached Figure Description
[0015] Figure 1 This is a flowchart of the present invention; Figure 2 This is a standard MBConv block structure diagram; Figure 3 For the structure diagram of the merged MBConv block (Fused-MBConv block); Figure 4 This is a diagram of the channel attention (SE attention) structure. Detailed Implementation
[0016] In the specific implementation process, this invention needs to be imported into a terminal that can perform analysis (such as Raspberry Pi 4). The terminal needs to input the collected BCG signal into this invention, with a time length of 15s as the division length. For prediction, a BCG signal with a time length of 15s is generally selected and stored in the memory. If the BCG signal time length is longer than 15s, continuous and uninterrupted heart rate monitoring can also be performed. The processor performs data processing on the stored BCG signal.
[0017] The specific process of the BCG signal heart rate calculation method based on Fast Fourier Transform and Convolutional Neural Network described in this invention is as follows: Figure 1 As shown, the specific methods include: Step 1: Filter the original BCG signal to obtain the filtered BCG signal BCG_F. BCG = {x(1),x(2),x(3),…,x(n)} BCG_F = Filter_BCG(BCG) = {x_f(1),x_f(2),x_f(3),…,x_f(n)} The Filter_BCG function uses a Butterworth filter to filter the BCG signal, removing low-frequency and high-frequency signals. The result is denoted as BCG_F. Step 2: Perform exponential shift normalization on the BCG_F signal to obtain BCG_stand. BCG_stand = scale_BCG(BCG_F) The scale_BCG function consists of the following three formulas:
[0018] Where EMA_current is the mean of the exponential moving average at the current time point, EMA_previous is the mean of the exponential moving average at the previous time point, EMA_sdc is the standard deviation of the exponential moving average at the current time point, EMA_sdp is the standard deviation of the exponential moving average at the previous time point, EMA is the mean of the exponential moving average, α is the decay factor, and value is the data value at the current time point. The resulting z_score sequence, arranged in order, is BCG_stand.
[0019] Step 3: Input BCG_stand into the attention-enhanced deep separable convolutional neural network, train it on the heart rate dataset, and save the optimal heart rate model weights.
[0020] First, the BCG_stand signal is input into the stem layer, which mainly includes: the initial convolutional layer, which is composed of a one-dimensional convolutional layer, a batch normalization layer and a SiLU activation function connected in sequence. It is used to perform preliminary feature extraction on the above signal, and the result is denoted as BCG_CB. The BCG_CB is input into a multi-level feature extraction module for further feature extraction. This module comprises six sequentially connected stages. The first and second stages utilize the Fused-MBConv block, which is structured as follows: First, a 3×1 convolution kernel is used to perform an expanded convolution operation, expanding the number of input channels to a predetermined multiple. Then, feature mapping is performed directly, omitting the depthwise separable convolution separation step found in traditional MBConv blocks. Finally, a 1×1 convolution kernel is used to perform a compression projection operation, adjusting the number of channels to the target output dimension. This design reduces computational complexity in high-resolution processing while maintaining sufficient feature extraction capabilities.
[0021] The third and fourth stages serve as transitional phases, gradually moving from Fused-MBConv blocks to standard MBConv blocks. The fourth stage introduces a complete standard MBConv block structure, comprising the following components: an extended convolutional layer, increasing the number of feature channels with a 1×1 convolutional kernel; a depthwise separable convolutional layer, using a 3×1 convolutional kernel to extract spatial features in the expanded feature space; a compressed projection layer, reducing the number of channels to the target dimension with a 1×1 convolutional kernel; and a conditional residual connection mechanism, which adds the original input features to the processed features only when the number of input and output channels are the same and the convolution stride is 1, forming a residual connection.
[0022] The fifth and sixth stages fully utilize the standard MBConv blocks, integrating a channel attention mechanism within each block. This attention mechanism first compresses the spatial dimension through a global average pooling layer, then learns the importance weights of each channel through two fully connected layers, and finally generates attention coefficients between 0 and 1 using a sigmoid activation function to recalibrate the feature channels.
[0023] After the sixth stage output, the network includes a head convolutional layer that expands the number of feature channels from 256 dimensions to 512 dimensions using a 1×1 convolutional kernel, significantly enhancing the expressive power of the feature space. Subsequently, a global average pooling layer averages the variable-length feature sequences after multiple downsamplings along the time dimension, aggregating them into a fixed-length 512-dimensional feature vector.
[0024] Table 1: Parameter Settings for Convolution Framework
[0025] The expansion factor is a temporary expansion factor for the number of input channels, used to create a high-dimensional feature representation space in the bottleneck structure; the repetition number refers to the number of layers in each stage.
[0026] During system operation, the main signal is processed through a feature extraction backbone network to extract 512-dimensional high-level features. The final heart rate prediction value is then generated through a fully connected network.
[0027] The first fully connected layer receives a 512-dimensional feature vector from the backbone network. This network employs a three-layer fully connected structure: the first layer maps the concatenated 512-dimensional features to a 256-dimensional hidden space; the second layer further compresses the 256-dimensional features to 128 dimensions; and the third layer compresses the 128-dimensional features to 32 dimensions. Each fully connected layer is followed by a SiLU activation function. A light regularization with a Dropout rate of 0.3 is applied after the first layer, and a light regularization with a Dropout rate of 0.2 is applied after the second layer, effectively preventing overfitting during training.
[0028] The second-stage fully connected layer receives 32-dimensional fused features from the first-stage fully connected layer. The network employs a three-layer fully connected structure: the first layer maps the 32-dimensional features to 64 dimensions; the second layer further compresses the 64-dimensional features back to 32 dimensions; and the third layer outputs the final predicted value. Each fully connected layer is followed by a SiLU activation function, with a light regularization of 0.1 after the first layer.
[0029] Table 2: Configuration of Fully Connected Layers after Adding Auxiliary Features
[0030] After 100 iterations, the weights of the optimal model for the heart rate task are saved as HR_0.
[0031] Step 4: In the blood pressure dataset, using the aforementioned heart rate model weights as initial weights, fine-tune the model through a transfer learning strategy to obtain prediction models for diastolic and systolic blood pressure.
[0032] The model loads weights from the training model of the BCG signal heart rate task, employs a selective parameter matching strategy, updates the model's parameters using diastolic and systolic blood pressure datasets, unfreezes all parameters, and uses a base learning rate of 10 for the feature extraction layer. -4 Applying a regularization of 0.01, the fully connected layer uses an enhanced learning rate of 10. -3 By applying a regularization of 0.01 and using the CosineAnnealingLR learning rate scheduler to decay the learning rate, the corresponding diastolic and systolic blood pressure results are obtained.
[0033] The blood pressure prediction method of this invention uses data collected by the inventors from a dataset of heart rate, blood pressure, and BCG signals of 392 individuals. In the heart rate dataset, each individual's data includes 15 minutes of BCG signals (including some body movement) and heart rate data in a resting state, and 15 minutes of BCG signals and heart rate data in a conversational state (including some body movement). In the blood pressure dataset, each individual's data includes 12 sets of blood pressure data (including diastolic and systolic blood pressure, with the subjects in a resting state during blood pressure collection). The data from the first 320 individuals collected were selected as the training set, and the data from the last 72 individuals were selected as the test set. The data collection process for both sets of data is consistent, with only the time order differing.
[0034] Table 3: Results of omitted steps
[0035] The MAE for the diastolic blood pressure BCG signal (including some body movement) in the test set was 8.001 mmHg, and the person correlation coefficient (the person correlation coefficient is the person correlation coefficient between the blood pressure value of this invention and the standard reference blood pressure, hereinafter referred to as the person correlation coefficient value) reached 0.356, which is significantly improved compared with no transfer learning. The MAE for the systolic blood pressure BCG signal (including some body movement) was 10.422 mmHg, and the person correlation coefficient reached 0.435, which is significantly improved compared with the prediction effect without using transfer learning. This shows the necessity of transfer learning in the above method.
Claims
1. A method for calculating blood pressure based on BCG signals using transfer learning and attention mechanisms, characterized in that, The steps are as follows: Step 1) Filter the original BCG signal to obtain the filtered BCG signal BCG_F: Step 2) Perform exponential shift normalization on the BCG_F signal to obtain BCG_stand: Step 3) Input BCG_stand into the attention-enhanced deep separable convolutional neural network, train it on the heart rate dataset, and save the optimal heart rate model weights; Step 4) In the blood pressure dataset, using the aforementioned heart rate model weights as initial weights, the model is fine-tuned through a transfer learning strategy to obtain the prediction model weights for diastolic and systolic blood pressure. Step 5) Using the obtained prediction model weights, the single-channel BCG signal is input into the model with the obtained blood pressure prediction model weights for prediction, thereby obtaining the diastolic and systolic blood pressure corresponding to the single-channel BCG signal.
2. The method for calculating blood pressure based on BCG signals according to claim 1, characterized in that, The specific steps in step 1) are as follows: BCG = {x(1),x(2),x(3),…,x(n)} BCG_F = Filter_BCG(BCG) The Filter_BCG function uses a Butterworth filter with a bandwidth of 0.8Hz-35Hz to filter the BCG signal, removing low-frequency signals below 0.8Hz and high-frequency signals above 35Hz. The result is denoted as BCG_F.
3. The method for calculating blood pressure based on BCG signals according to claim 1, characterized in that, The specific steps in step 2) are as follows: BCG_stand = scale_BCG(BCG_F) The scale_BCG function is composed of the following formula: Wherein, EMA_current is the mean of the exponential moving average at the current time point, EMA_previous is the mean of the exponential moving average at the previous time point, EMA_sdc is the standard deviation of the exponential moving average at the current time point, EMA_sdp is the standard deviation of the exponential moving average at the previous time point, α is the decay factor, and value is the data value at the current time point. The resulting z_scores are arranged in order to form the BCG_stand sequence.
4. The method for calculating blood pressure based on BCG signals according to claim 1, characterized in that, In step 3), the specific structure of the attention-enhanced deep separable convolutional neural network is as follows: First, the BCG_stand signal is input into a one-dimensional feature extraction network, which includes an initial convolutional layer, consisting of a one-dimensional convolutional layer, a batch normalization layer, and an activation function connected in sequence. This layer is used to perform preliminary feature extraction on the above signal, and the result is denoted as BCG_CB. BCG_CB is input into the multi-level feature extraction module for further feature extraction. The multi-level feature extraction module contains six sequentially connected stages. The first, second, and third stages use Fused-MBConv blocks, while the fourth, fifth, and sixth stages use one-dimensional MBConv blocks. The above six stages ultimately increase the number of channels to 256. The 256-dimensional features output from the sixth stage are expanded to 512 dimensions through one-dimensional convolutional layers, batch normalization layers, and activation functions. Then, the resulting 512-dimensional features are pooled and flattened. The above features are then processed through two fully connected layers to obtain the final predicted value.
5. The method for calculating blood pressure based on BCG signals according to claim 4, characterized in that, The attention-enhanced deep separable convolutional neural network uses the Fused-MBConv block, specifically as follows: After receiving the input feature signal, the system first determines whether to enable the residual connection path based on the matching relationship between the number of input channels and the number of target output channels and the preset convolution stride value. Then, it enters the feature transformation stage: when the expansion coefficient is not 1, spatial feature extraction and channel dimension expansion are completed simultaneously through a fused convolution to generate an intermediate feature signal. When the expansion coefficient is 1, the input signal is directly transmitted; then, the channel dimension of the intermediate feature signal is projected and compressed to the target output dimension through convolution; finally, according to the initial judgment result: if the residual connection condition is met, the original input signal and the projected output are added element by element. If the conditions are not met, the projection result will be output directly.
6. The method for calculating blood pressure based on BCG signals according to claim 4, characterized in that, The attention-enhanced deep separable convolutional neural network described above uses one-dimensional MBConv blocks, specifically as follows: First, determine if the residual connection condition is met: when the number of input channels is equal to the number of output channels and the convolution stride is 1, the residual path is enabled; the processing flow starts from the channel expansion stage, and when the expansion coefficient is greater than 1, the number of input channels is expanded to the hidden dimension through convolution; When the expansion coefficient is 1, the input features are directly passed; then depthwise separable convolution is performed to extract spatial features from the expanded features in channel grouping mode; Then, the channel attention weighting stage begins. After attention weighting is completed, the hidden dimensions are projected and compressed to the target number of output channels through convolution. Finally, based on the initial conditions, if residual connection is enabled, the original input signal is added to the processing result and output; otherwise, the processing result is output directly.
7. The method for calculating blood pressure based on BCG signals according to claim 6, characterized in that, The attention-enhanced deep separable convolutional neural network described above uses SE attention in one-dimensional MBConv. The specific method is as follows: First, global average pooling is performed on each channel of the deep convolution output to obtain a channel-level statistical description. Then, the importance weights of each channel are learned through a bottleneck structure containing two linear layers. This structure first compresses the channel dimension to 1 / reduction and then restores it to the original dimension. During this process, an activation function is used to generate normalized weights between 0 and 1. Finally, the learned weights are multiplied with the original features channel by channel to achieve feature recalibration.
8. The method for calculating blood pressure based on BCG signals according to claim 1, characterized in that, Step 4) uses a transfer learning method, specifically: Pre-training was performed on the heart rate dataset, using the weights of the heart rate model as the initial weights for the blood pressure model. A diastolic blood pressure model was trained on the diastolic blood pressure dataset, and a systolic blood pressure model was trained on the systolic blood pressure dataset. During training, the feature extraction layer maintained a learning rate of 10. -4 The fully connected layer maintains a learning rate of 10. -3 The learning rate is adjusted through the learning rate scheduler.