Blood pressure measurement system and method based on oscillatory wave and PPG signal collaborative learning

Abstract

The invention provides a blood pressure measurement system and method based on oscillatory wave and PPG signal collaborative learning, and relates to the technical field of wearable medical health monitoring. According to the method, the advantages of the CNN in local feature extraction, the Transform in time sequence modeling and the PPG signal in assisting blood pressure prediction are brought into full play, and the accuracy of blood pressure estimation can be remarkably improved. Complex features are automatically extracted through deep learning, dependence on manual feature engineering is reduced, traditional machine learning is adopted, model interpretation is enhanced, and the precision and robustness of blood pressure estimation are improved by combining the advantages of the two. The method is wide in applicability, can be applied to various scenes such as family health monitoring, clinical monitoring and wearable equipment, adopts multi-task learning, can predict systolic pressure and diastolic pressure at the same time, and shares a feature extraction network. Moreover, the algorithm is a lightweight algorithm, can be deployed on wearable equipment to realize low-power-consumption and high-efficiency real-time blood pressure estimation, and provides powerful support for health monitoring.

Description

Technical Field

[0001] The present invention relates to the technical field of wearable medical health monitoring, and particularly to a blood pressure measurement system and method based on collaborative learning of oscillatory waves and PPG signals. Background Art

[0002] Hypertension is one of the most common chronic diseases globally. According to the World Health Organization, approximately 1.8 billion adults worldwide are affected by it. Accurate and convenient blood pressure monitoring is crucial for early detection of hypertension, assessment of treatment effects, and prevention of complications. Although the traditional mercury sphygmomanometer is the gold standard, it relies on manual operation and cannot provide real-time monitoring. Electronic blood pressure monitors achieve non-invasive measurement based on the oscillometric method, but their accuracy is significantly affected by factors such as individual differences and measurement postures. With the popularization of wearable devices, continuous blood pressure monitoring technology based on photoplethysmography (PPG) has become a research hotspot, but existing methods generally suffer from problems such as insufficient measurement accuracy and weak anti-interference ability.

[0003] Traditional oscillometric blood pressure monitors rely on the Korotkoff sound method as the calibration standard, but the Korotkoff sound method itself has defects such as strong subjectivity and being greatly affected by environmental noise. Research shows that the difference in the judgment of the starting point of the Korotkoff sound by different doctors can reach 5 - 10 mmHg. Existing machine learning models are mostly trained based on the mapping relationship between oscillatory waves and Korotkoff sound method blood pressure, and their accuracy is limited by the reliability of the reference standard.

[0004] Existing methods mainly rely on manual features (such as oscillatory wave peaks, areas, etc.) or single deep learning models (such as CNN) to extract features. Manual features are difficult to capture complex hemodynamic changes, while pure deep learning models have problems such as poor interpretability and overfitting to small sample data. In addition, existing research rarely fuses multi-modal features of oscillatory waves and PPG signals, and fails to fully utilize the complementary information of different signal modalities.

[0005] Existing models are usually trained based on specific populations (such as healthy adults) and have poor adaptability to special populations such as obesity and arteriosclerosis. Clinical data shows that the systolic blood pressure measurement error of traditional methods in the elderly population can reach more than 15 mmHg.

[0006] Patent CN117678987A proposes a blood pressure measurement method. It acquires the PPG signal and accelerometer (ACC) signal of the person to be detected, processes them to obtain R-wave peak data for calculating PTT data, and simultaneously extracts the PPG signal features. Then, both are input into a pre-trained model to obtain the blood pressure detection result. This method avoids manual measurement through signal processing and reduces costs. However, this method has limitations. It relies only on PPG and ACC signals, is prone to domain shift, has large measurement errors in special populations, and it is difficult to ensure high precision and reliability. In addition, the source and features of the training data are not clear, resulting in insufficient generalization ability of the model and inability to adapt to different scenarios and populations.

[0007] Therefore, there is an urgent need for a combined measurement method of oscillometric wave and PPG signal applicable to home health management that can solve the small sample problem and enhance clinical credibility. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a blood pressure measurement system and method based on collaborative learning of oscillometric wave and PPG signal to achieve accurate blood pressure estimation in view of the above-mentioned deficiencies of the prior art.

[0009] To solve the above technical problem, the technical solution adopted by the present invention is as follows: On the one hand, the present invention provides a blood pressure measurement system based on collaborative learning of oscillometric wave and PPG signal, including an oscillometric wave and PPG signal database construction module, a data preprocessing module, a feature extraction module, and a blood pressure estimation module; The oscillometric wave and PPG signal database stores the oscillometric wave and PPG signal of 3000 subjects, as well as the corresponding diastolic blood pressure and systolic blood pressure; The data preprocessing module is used to filter the acquired oscillometric wave signal and PPG signal to remove common noise; The feature extraction module constructs a CNN-Transformer network, including a CNN feature extractor, a position encoder, multiple layers of Transformer encoders, a global average pooling layer, a fully connected layer, and a feature fusion module. The CNN feature extractor is used to extract the local features of the oscillometric wave signal and PPG signal, and the multiple layers of Transformer encoders are used to extract the temporal features of the oscillometric wave signal; on this basis, the position encoder adds position encoding to the extracted features to retain sequence information and enhance the model's perception ability of temporal features; the global average pooling layer performs global average pooling on the output of the multiple layers of Transformer encoders and then inputs it into the fully connected layer; the feature fusion module combines time-frequency analysis and statistical features to extract manual features, performs feature fusion with the features extracted by deep learning, and divides the training set and test set; The blood pressure estimation module uses the XGBoost model to estimate blood pressure based on the extracted features, achieving accurate prediction of systolic and diastolic blood pressure.

[0010] Further, the oscillometric wave and PPG signal database construction module includes an oscillometric wave signal device, a cuff, a PPG sensor, a blood pressure measurement device, and a database; first, the oscillometric wave signal device is correctly placed on the upper arm of the subject and fixed by the cuff; the cuff will gradually apply pressure and record the oscillometric wave data of the subject at different pressures and store it in the database; then, the PPG sensor is correctly placed at the fingertip. During the acquisition process, the subject is required to remain stationary to reduce the influence of motion artifacts on the PPG signal, and the systolic and diastolic blood pressures of the subject are measured by the blood pressure measurement device and stored in the database.

[0011] Further, the data preprocessing module includes an oscillometric wave data preprocessing module and a PPG data preprocessing module; The oscillometric wave data preprocessing module first uses a low-pass filter to denoise the original oscillometric wave signal, removing the high-frequency noise in the signal, thereby retaining the low-frequency components in the oscillometric wave signal; then performs high-pass filtering to remove the low-frequency components in the signal, thereby retaining the high-frequency components in the oscillometric wave signal; The PPG data preprocessing module first uses low-pass filtering to remove the high-frequency noise of the PPG signal and retain the useful components in the PPG signal; then removes the low-frequency drift or baseline drift through high-pass filtering to retain the important information in the PPG signal; adopts a waveform quality evaluation method based on wavelet transform to detect and remove motion artifacts or bad signal segments, and its calculation formula is as shown in Equation (1): (1); Where, a and b are the scale and translation parameters, ψ is the mother wavelet function; f(t) is the original signal, which is a function of t.

[0012] Smooth the signal through the Savitzky-Golay filter technology to reduce high-frequency fluctuations, and its calculation formula is as shown in Equation (2): (2); Where, c j are the Savitzky-Golay coefficients, m is half of the window length, x i+j is the original data value of the input signal at the position i + j, where j is the offset index within the window.

[0013] Further, the specific implementation method of the blood pressure estimation module is as follows: First, set the hyperparameters of XGBoost, including the number of trees, learning rate, and maximum depth; Then, use the training set to train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean squared error, as shown in the following formula: (9); Where, θ are the parameters of the XGBoost model, y i is the true blood pressure value, is the blood pressure value predicted by the model; Use the trained XGBoost model to predict the test set to obtain the predicted blood pressure value ; Then, use evaluation metrics to measure the performance of the model. The evaluation metrics include mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE): (10); (11); (12); Where, n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0014] On the other hand, the present invention also provides a blood pressure measurement method based on the collaborative learning of oscillometric waves and PPG signals, which is implemented by the above-mentioned blood pressure measurement system based on the collaborative learning of oscillometric waves and PPG signals, and includes the following steps: Step 1: Construct a database of oscillometric wave signals and PPG signals, including oscillometric wave signals, PPG signals, and their corresponding systolic blood pressure and diastolic blood pressure; Step 2: Preprocess the oscillometric wave signals in the database, including low-pass filtering and high-pass filtering, to optimize the noise, baseline drift, and amplitude differences between different waveforms in the oscillometric wave signals; Step 3: Preprocess the PPG signals in the database, including low-pass filtering, high-pass filtering, artifact removal, and smoothing, to remove the noise, baseline drift, and amplitude differences between different waveforms in the PPG signals; Step 4: Perform feature extraction and blood pressure estimation modeling; establish a CNN-Transformer network, including a CNN feature extractor, positional encoding, multiple layers of Transformer encoders, global average pooling, and a fully connected layer; use CNN to extract local features of the oscillometric signal and PPG signal and splice them to form joint features; then use multiple layers of Transformer encoders to extract the temporal features of the oscillometric signal; add positional encoding to the extracted features to retain sequence information and enhance the model's perception ability of temporal features; perform global average pooling on the output of the multiple layers of Transformer encoders and then input it into the fully connected layer; Step 5: Perform blood pressure prediction based on newly collected oscillometric signal data and PPG signal, including preprocessing of the oscillometric signal and PPG signal and blood pressure estimation based on machine learning and deep learning models; fuse the features extracted by deep learning with the manual features extracted by time-frequency analysis and statistical features, and then input them into the XGBoost model to learn the mapping relationship between the oscillometric signal, PPG signal and systolic blood pressure, diastolic blood pressure, and achieve accurate estimation of diastolic blood pressure and systolic blood pressure.

[0015] Further, the specific method of step 1 is as follows: Step 1.1: Correctly place the oscillometric signal device on the upper arm of the subject and fix it with a cuff; Step 1.2: Gradually pressurize the cuff, record the oscillometric data of the subject under different pressures, and store it; Step 1.3: Correctly place the PPG sensor at the fingertip; during the acquisition process, require the subject to remain stationary to reduce the influence of motion artifacts on the PPG signal; Step 1.4: Measure the systolic blood pressure and diastolic blood pressure of the subject through a blood pressure measuring device.

[0016] Further, the specific method of step 2 is as follows: Step 2.1: Use a low-pass filter to remove high-frequency noise in the original oscillometric signal, thereby retaining the low-frequency components in the oscillometric signal; Step 2.2: Use a high-pass filter to remove the influence of baseline drift on the oscillometric signal, thereby retaining the high-frequency components in the oscillometric signal.

[0017] Further, the specific method of step 3 is as follows: Step 3.1: Use a low-pass filter to remove high-frequency noise and retain the useful components in the PPG signal; Step 3.2: Use a high-pass filter to remove low-frequency drift or baseline drift and retain the important information in the PPG signal; Step 3.3: Adopt a wavelet transform waveform quality assessment method to detect and remove motion artifacts or bad signal segments; Step 3.4: Smooth the signal through Savitzky-Golay filter technology to reduce high-frequency fluctuations.

[0018] Further, the specific method of step 4 is as follows: Step 4.1: Perform data preparation, load the oscillatory wave signal, PPG signal, and their corresponding diastolic and systolic blood pressure values, and divide them into a training set and a test set; Step 4.2: Build a deep learning model, that is, build a CNN-Transformer network, including a CNN feature extractor, positional encoding, multiple Transformer encoders, global average pooling, and a fully connected layer; Step 4.2.1: Input the oscillatory wave signal and PPG signal; Step 4.2.2: Construct the first CNN network to extract features from the oscillatory wave signal. The specific method is as follows: Construct convolutional layer 1 with a kernel size of 5 and 16 kernels, and use the ReLU activation function; Pooling layer 1 with max pooling and a pool size of 2; Construct convolutional layer 2 with a kernel size of 5 and 32 kernels, and use the ReLU activation function; Pooling layer 2 with max pooling and a pool size of 2; Output layer 1 to output the features extracted from the oscillatory wave signal; Step 4.2.3: Construct the second CNN network to extract features from the PPG signal. The specific method is as follows: Construct convolutional layer 3 with a kernel size of 3 and 16 kernels, and use the ReLU activation function; Pooling layer 3 with max pooling and a pool size of 2; Construct convolutional layer 4 with a kernel size of 5 and 32 kernels, and use the ReLU activation function; Pooling layer 4 with max pooling and a pool size of 2; Construct convolutional layer 5 with a kernel size of 7 and 64 kernels, and use the ReLU activation function; Pooling layer 5 with average pooling and a pool size of 2; Output layer 2 to output the features extracted from the PPG signal; Step 4.2.4: Concatenate the feature vectors extracted from the oscillatory wave signal and PPG signal to form a joint feature, then transform the input shape to convert the output of the CNN part into the shape required by the Transformer network; Step 4.2.5: Construct a Transformer network to capture the temporal dependence of the oscillatory wave signal. The specific method is as follows: Step 4.2.5.1: Input embedding, which converts the input features into high-dimensional feature vectors; Step 4.2.5.2: Positional encoding, adding positional encoding to retain temporal information and introducing the temporal information into the feature vectors. The calculation formula is as shown in Equation (5): (5); where, pos is the position, i is the dimension, d model is the dimension of the model; Step 4.2.5.3: Multi-head self-attention mechanism, calculating attention scores and performing weighted summation based on these scores to capture the dependencies between different positions in the input sequence. The calculation process of self-attention is as shown in Equation (6): (6); where, Q , K and V are obtained by linear transformation of the input features, d k is K 's dimension; Step 4.2.5.4: Layer normalization to stabilize and accelerate the training process; Step 4.2.5.5: Feed-forward neural network, further processing the features through a non-linear activation function (ReLU) to enhance the expressive power of the model. The calculation process is as shown in Equation (7): (7); where max(0, xW1 + b1) is the non-linear activation function, W 1 and W 2 are weight matrices, b 1 and b 2 are biases; Step 4.2.5.6: Layer normalization to stabilize and accelerate the training process; Step 4.2.5.7: Repeat steps 4.2.5.3 to 4.2.5.6 six times; Step 4.2.6: Perform global average pooling on the output of the Transformer network; Step 4.2.7: Input the result of global average pooling into a fully connected layer.

[0019] Furthermore, the specific method of step 5 is as follows: Step 5.1: Preprocess the newly collected oscillatory wave signal and PPG signal data, and filter the oscillatory wave signal and PPG signal; Step 5.2: Extract features from the processed oscillatory wave signal and PPG signal; Step 5.2.1: Use CNN to extract time-frequency features from the oscillatory wave signal and PPG signal to capture local pattern information; Step 5.2.2: Add positional encoding to the extracted features to retain sequence information and enhance the model's ability to model temporal dependencies; Step 5.2.3: Utilize the multi-head self-attention mechanism and feed-forward neural network to capture the long-range dependencies of the signal and improve the feature expression ability; Step 5.2.4: Perform average pooling on the extracted feature vectors to obtain fixed-length feature vectors, reduce redundant information, and improve computational efficiency; Step 5.3: Fuse the pooled feature vectors with the manually extracted time-frequency features and statistical features, and input them into the XGBoost model to learn the mapping relationship between the oscillatory wave signal, PPG signal, and blood pressure value, and achieve accurate estimation of diastolic and systolic blood pressure; The specific method is as follows: Step 5.3.1: Input the fused features; Step 5.3.2: Data preprocessing, including data cleaning and data standardization; The data standardization formula is as follows: (8); where, x ij is the j-th feature value of the i-th sample, and x ij * is the j-th feature value of the i-th sample after standardization, μ j is the mean of the j-th feature, σ j is the standard deviation of the j-th feature; Step 5.3.3: Divide the preprocessed dataset into a training set Dtrain and a test set Dtest ; Step 5.3.4: Set the hyperparameters of XGBoost, including the number of trees, learning rate, and maximum depth; Step 5.3.5: Use the training set Dtrain to train the XGBoost model; The objective function of XGBoost is to minimize the loss function, and the loss function is the mean squared error: (9); where, θ is the parameter of the XGBoost model, and y iis the real blood pressure value, is the blood pressure value predicted by the model; Step 5.3.6: Use the trained XGBoost model to predict the test set Dtest to obtain the predicted blood pressure value ; Step 5.3.7: Use evaluation metrics to measure the performance of the model. The evaluation metrics include mean squared error MSE, root mean squared error RMSE, and mean absolute error MAE: (10); (11); (12); where n test is the number of samples in the test set, and y test,i is the real blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0020] The beneficial effects of adopting the above technical solutions are as follows: The blood pressure measurement system and method based on collaborative learning of oscillometric waves and PPG signals provided by the present invention can effectively extract local features in the oscillometric wave signal and PPG signal and capture the temporal dependence of the signals by combining the advantages of CNN and Transformer. During the feature extraction process, CNN is used to extract local features of the oscillometric wave signal and PPG signal, and Transformer is used to capture the temporal features of the oscillometric wave signal, thereby obtaining high-dimensional and highly expressive deep features. These deep features are combined with manually extracted features of time-frequency analysis and statistical feature extraction and trained through the XGBoost model to learn the mapping relationship between the oscillometric wave signal, PPG signal and systolic blood pressure, diastolic blood pressure, improving the accuracy of blood pressure prediction. The training data of this model comes from different individuals, effectively reducing the influence of individual differences on the blood pressure estimation results and enhancing the applicability and generality of this model. Description of the Drawings

[0021] Figure 1 is the structural block diagram of the blood pressure measurement system based on collaborative learning of oscillometric waves and PPG signals provided by the embodiment of the present invention; Figure 2 is the implementation block diagram of the oscillometric wave data preprocessing module provided by the embodiment of the present invention; Figure 3 is the implementation block diagram of the PPG data preprocessing module provided by the embodiment of the present invention; Figure 4 is the implementation block diagram of the CNN-Transformer network module provided by the embodiment of the present invention; Figure 5 It is a block diagram of a CNN structure for extracting the characteristics of oscillatory wave signals provided by an embodiment of the present invention; Figure 6 It is a block diagram of a CNN structure for extracting the characteristics of PPG signals provided by an embodiment of the present invention; Figure 7 It is a block diagram for implementing a Transformer network model provided by an embodiment of the present invention; Figure 8 It is a block diagram of the structure of an XGBoost model provided by an embodiment of the present invention. Specific implementation manners

[0022] The following will further describe in detail the specific implementation manners of the present invention in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0023] In this embodiment, a blood pressure measurement system based on collaborative learning of oscillatory waves and PPG signals, as Figure 1 shown, includes the construction of an oscillatory wave database, a data preprocessing module, a feature extraction and fusion module, and a blood pressure estimation module.

[0024] The oscillatory wave database stores the oscillatory wave signals of 3,000 subjects and the corresponding diastolic and systolic blood pressures; the data preprocessing module is used to remove common noises in the oscillatory wave signals through filtering. First, the oscillatory wave signal device is correctly placed on the upper arm of the subject and fixed with a cuff to ensure accuracy and comfort. The cuff will gradually be pressurized, and the oscillatory wave data of the subject at different pressures will be recorded and stored. Then, the PPG sensor is correctly placed at the fingertip. During the acquisition process, the subject is required to remain still, avoid large movements or postural changes, and reduce the influence of motion artifacts on the PPG signal. The systolic and diastolic blood pressures of the subject are measured by a blood pressure measuring device.

[0025] The data preprocessing module is used to perform filtering on the acquired oscillatory wave signals and PPG signals to remove common noises. As Figure 2 shown, first, a low-pass filter is used to denoise the original oscillatory wave signal. The main purpose of low-pass filtering is to remove high-frequency noises in the signal, such as power interference and high-frequency random noises, so as to retain the low-frequency components in the oscillatory wave signal; then, high-pass filtering is performed. The main purpose of high-pass filtering is to remove the low-frequency components in the signal, such as baseline drift and respiratory artifacts and other low-frequency noises, so as to retain the high-frequency components in the oscillatory wave signal. As Figure 3As shown, first, low-pass filtering is used to remove high-frequency noise in the PPG signal, such as light source changes or motion artifacts, and retain the useful components in the PPG signal; then, high-pass filtering is performed: to remove low-frequency drift or baseline drift (such as slow changes caused by breathing or body position changes) to retain important information in the PPG signal; a waveform quality assessment method based on wavelet transform is adopted to detect and remove motion artifacts or bad signal segments, and its calculation formula is as shown in (1); the signal is smoothed by Savitzky-Golay filter technology to reduce high-frequency fluctuations, and its calculation formula is as shown in Equation (2).

[0026] (1); Wherein, a and b are scale and translation parameters, ψ is the mother wavelet function, f(t) is the original signal, which is a function of t.

[0027] (2); Wherein, c j are Savitzky-Golay coefficients, m is half of the window length, x i+j is the original data value of the input signal at the position i + j, where j is the offset index within the window.

[0028] The feature extraction and fusion module uses deep learning (CNN and Transformer networks) to extract high-level features, and at the same time combines two traditional machine learning methods of time-frequency analysis and statistical features to obtain manual features and perform feature fusion to enhance the prediction ability of the model. First, data preparation is carried out, the oscillatory wave signal, PPG signal and their corresponding diastolic and systolic blood pressure values are loaded, and the training set and test set are divided. Then, a feature extraction module is established, including a CNN feature extractor, a position encoder, a multi-layer Transformer encoder, a global average pooling layer and a fully connected layer, and a CNN-Transformer network is constructed, as shown in Figure 4 As shown. In addition, two traditional machine learning methods of time-frequency analysis and statistical features are combined to extract manual features and fuse them with the features extracted by deep learning to enhance the prediction ability of the model. Finally, evaluation is carried out through the XGBoost model, the performance of the model is verified using the test set, overfitting phenomena are detected and the model is tuned to ensure the accuracy and generalization ability of the model. The construction process of the CNN-Transformer network is specifically as follows: Step 1: Input layer: Input the oscillatory wave signal.

[0029] Step 2: The first CNN network: Extract features from the oscillatory wave signal, and the network structure is as shown inFigure 5 As shown below. The specific steps are as follows: Construct the first convolutional layer with a convolutional kernel size of 5, 16 convolutional kernels, and use the ReLU activation function. The output of the first convolutional layer is as shown in Equation (3): (3); Where, x i+m,j+n is the input signal, w m,n is the convolutional kernel, b is the bias.

[0030] The first pooling layer, with the pooling method being max pooling and the pooling kernel size being 2. The max pooling process is as shown in Equation (4): (4); Construct the second convolutional layer with a convolutional kernel size of 5, 32 convolutional kernels, and use the ReLU activation function.

[0031] The second pooling layer, with the pooling method being max pooling and the pooling kernel size being 2.

[0032] The first output layer outputs the features extracted from the oscillatory wave signal.

[0033] Step 3: The second CNN network: Extract features from the PPG signal. The network structure is as Figure 6 shown below. The specific steps are as follows: Construct the third convolutional layer with a convolutional kernel size of 5, 16 convolutional kernels, and use the ReLU activation function.

[0034] The third pooling layer, with the pooling method being max pooling and the pooling kernel size being 2.

[0035] Construct the fourth convolutional layer with a convolutional kernel size of 5, 32 convolutional kernels, and use the ReLU activation function.

[0036] The fourth pooling layer, with the pooling method being max pooling and the pooling kernel size being 2.

[0037] Construct the fifth convolutional layer with a convolutional kernel size of 7, 64 convolutional kernels, and use the ReLU activation function; The fifth pooling layer, with the pooling method being average pooling and the pooling kernel size being 2; The second output layer outputs the features extracted from the PPG signal.

[0038] Step 4: Feature vector concatenation and transformation of the input shape: Concatenate the feature vectors extracted from the oscillatory wave signal and the PPG signal to form a joint feature, and then transform the input shape to convert the output of the CNN part into the shape required by the Transformer network.

[0039] Step 5: Transformer Network: Capture the temporal dependence of the oscillatory wave signal. The network structure is as Figure 7 shown. The specific steps are as follows: Step 5.1: Input Embedding, convert the input features into high-dimensional feature vectors.

[0040] Step 5.2: Positional Encoding, add positional encoding to retain the temporal information and introduce the temporal information into the feature vectors. The calculation formula is as (5).

[0041] (5); where, pos is the position, i is the dimension, d model is the dimension of the model.

[0042] Step 5.3: Multi-Head Self-Attention Mechanism, calculate the attention scores and perform weighted summation based on these scores to capture the dependencies between different positions in the input sequence. The calculation process of self-attention is as formula (6): (6); where, Q , K and V are obtained by linear transformation of the input features, d k is K 's dimension.

[0043] Step 5.4: Layer Normalization, stabilize and accelerate the training process.

[0044] Step 5.5: Feed-Forward Neural Network, further process the features through a non-linear activation function to enhance the model's expressive ability. The calculation process is as formula (7): (7); where max(0,xW1+b1) is the non-linear activation function, W 1 and W 2 are weight matrices, b 1 and b 2 are biases.

[0045] Step 5.6: Layer Normalization, stabilize and accelerate the training process.

[0046] Step 5.7: Repeat steps 5.3 to 5.6 six times.

[0047] Step 6: Global Average Pooling: Perform global average pooling on the output of the Transformer.

[0048] Step 7: Fully connected layer: Input the result of global average pooling into the fully connected layer.

[0049] The blood pressure estimation module uses the XGBoost machine learning model to learn the mapping relationship between oscillometric waves, PPG signals, and blood pressure values based on training data, so as to achieve accurate estimation of diastolic and systolic blood pressures. First, preprocess the newly collected oscillometric wave signal data. Then, input the processed oscillometric wave data into machine learning and deep learning models. First, use CNN and Transformer to extract features, and combine time-frequency analysis and statistical feature extraction to extract manual features, and add positional encoding to the features to retain sequence information. Then, use the multi-head self-attention mechanism and feed-forward neural network to capture long-range dependencies and further enhance the expressive power of features. Next, compress the feature vector into a vector of fixed length through global average pooling. Finally, combine the traditional machine learning model XGBoost to train and map the pooled feature vector to achieve the estimation of diastolic and systolic blood pressures.

[0050] The specific steps of constructing the XGBoost network are as follows: Step 1: Input the fused features.

[0051] Step 2: Data preprocessing, including data cleaning and data standardization. The data standardization formula is as follows: (8); where x ij is the j-th feature value of the i-th sample, is the j-th feature value of the i-th sample after standardization, μ j is the mean of the j-th feature, σ j is the standard deviation of the j-th feature.

[0052] Step 3: Divide the preprocessed data set into a training set Dtrain and a test set Dtest , usually divided according to 80% training set and 20% test set.

[0053] Step 4: Set the hyperparameters of XGBoost. The number of trees (n_estimators) is equal to 100, the learning rate (learning_rate) is equal to 0.01, and the maximum depth (max_depth) is equal to 5.

[0054] Step 5: Train the model, use the training set Dtrain to train the XGBoost model. The objective function of XGBoost is to minimize the loss function, and the loss function is the mean squared error (MSE): (9); Among them, θ is the parameter of the model, and y i is the real blood pressure value, is the blood pressure value predicted by the model.

[0055] Step 6: Use the trained XGBoost model to predict the test set Dtest to obtain the predicted blood pressure value .

[0056] Step 7: Use evaluation metrics to measure the performance of the model. The evaluation metrics include mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE): (10); (11); (12); Among them, n test is the number of samples in the test set, y test,i is the real blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0057] Through the above steps, the blood pressure estimation module can effectively extract features from the oscillometric wave and PPG signals and predict the blood pressure value.

[0058] A blood pressure measurement method based on the collaborative learning of oscillometric wave and PPG signals is also provided, which is implemented through the above-mentioned blood pressure measurement system based on the collaborative learning of oscillometric wave and PPG signals, and includes the following steps: Step 1: Construct an oscillometric wave signal and PPG signal database, including oscillometric wave signals, PPG signals, and their corresponding systolic blood pressure and diastolic blood pressure. The specific method is as follows: Step 1.1: Place the oscillometric wave signal device correctly on the upper arm of the subject and fix it with a cuff to ensure accuracy and comfort.

[0059] Step 1.2: Gradually pressurize the cuff, record the oscillometric wave data of the subject under different pressures, and store it.

[0060] Step 1.3: Place the PPG sensor correctly at the fingertip. During the acquisition process, the subject is required to remain stationary, avoid large movements or postural changes, and reduce the influence of motion artifacts on the PPG signal.

[0061] Step 2: Preprocess the oscillatory wave signals in the database, including low-pass filtering and high-pass filtering, to optimize the noise, baseline drift, and amplitude differences between different waveforms in the oscillatory wave signals. The specific method is as follows: Step 2.1: Use a low-pass filter to remove the high-frequency noise in the original oscillatory wave signal, thereby retaining the low-frequency components in the oscillatory wave signal.

[0062] Step 2.2: Use high-pass filtering to remove the influence of baseline drift on the oscillatory wave, thereby retaining the high-frequency components in the oscillatory wave signal.

[0063] Step 3: Preprocess the PPG signals in the database, including low-pass filtering, high-pass filtering, artifact removal, and smoothing, to remove the noise, baseline drift, and amplitude differences between different waveforms in the PPG signals. The specific method is as follows: Step 3.1: Use low-pass filtering to remove high-frequency noise, such as light source changes or motion artifacts, and retain the useful components in the PPG signal.

[0064] Step 3.2: Use high-pass filtering to remove low-frequency drift or baseline drift, such as the slow changes caused by breathing and body position changes, and retain the important information in the PPG signal.

[0065] Step 3.3: Adopt the wavelet transform waveform quality assessment method to detect and remove motion artifacts or bad signal segments.

[0066] Step 3.4: Smooth the signal through the Savitzky-Golay filter technology to reduce high-frequency fluctuations.

[0067] Step 4: Perform feature extraction and blood pressure estimation modeling; use CNN to extract the local features of the oscillatory wave signal and the PPG signal, and use Transformer to extract the temporal features of the oscillatory wave signal to obtain high-dimensional and highly expressive deep features. The specific method is as follows: Step 4.1: Prepare the data, load the oscillatory wave signal, the PPG signal, and their corresponding diastolic and systolic blood pressure values, and divide the training set and the test set.

[0068] Step 4.2: Establish a deep learning model, that is, establish a CNN-Transformer network, including a CNN feature extractor, positional encoding, multiple layers of Transformer encoders, global average pooling, and a fully connected layer.

[0069] Step 4.2.1: Input the oscillatory wave signal and the PPG signal.

[0070] Step 4.2.2: Construct the first CNN network to extract features from the oscillatory wave signal. The specific method is as follows: Construct the first convolutional layer with a kernel size of 5 and 16 kernels, and use the ReLU activation function; The first pooling layer with max pooling and a pool size of 2; Construct the second convolutional layer with a kernel size of 5 and 32 kernels, and use the ReLU activation function; The second pooling layer with max pooling and a pool size of 2; The first output layer, outputting the features extracted from the oscillatory wave signal.

[0071] Step 4.2.3: Construct the second CNN network to extract features from the PPG signal. The specific method is as follows: Construct the third convolutional layer with a kernel size of 3 and 16 kernels, and use the ReLU activation function; The third pooling layer with max pooling and a pool size of 2; Construct the fourth convolutional layer with a kernel size of 5 and 32 kernels, and use the ReLU activation function; The fourth pooling layer with max pooling and a pool size of 2; Construct the fifth convolutional layer with a kernel size of 7 and 64 kernels, and use the ReLU activation function; The fifth pooling layer with average pooling and a pool size of 2; The second output layer, outputting the features extracted from the PPG signal.

[0072] Step 4.2.4: Concatenate the feature vectors extracted from the oscillatory wave signal and the PPG signal to form a joint feature, and then transform the input shape to convert the output of the CNN part into the shape required by the Transformer network.

[0073] Step 4.2.5: Construct the Transformer network to capture the temporal dependencies of the oscillatory wave signal. The specific method is as follows: Step 4.2.5.1: Input embedding, converting the input features into high-dimensional feature vectors; Step 4.2.5.2: Positional encoding, adding positional encoding to retain the temporal information and introduce the temporal information into the feature vectors. The calculation formula is as shown in Equation (5): (5); where, pos is the position, i is the dimension, d model is the dimension of the model; Step 4.2.5.3: The multi-head self-attention mechanism calculates the attention scores and performs weighted summation based on these scores to capture the dependencies between different positions in the input sequence; the calculation process of self-attention is as shown in Equation (6): (6); where, Q 、 K and V are obtained by linear transformation of the input features, d k is K the dimension of; Step 4.2.5.4: Layer normalization to stabilize and accelerate the training process; Step 4.2.5.5: The feed-forward neural network further processes the features through a non-linear activation function to enhance the model's expressive power, and the calculation process is as shown in Equation (7): (7); where max(0, xW1 + b1) is the non-linear activation function, W 1 and W 2 are weight matrices, b 1 and b 2 are biases; Step 4.2.5.6: Layer normalization to stabilize and accelerate the training process; Step 4.2.5.7: Repeat Steps 4.2.5.3 to 4.2.5.6 six times.

[0074] Step 4.2.6: Perform global average pooling on the output of the Transformer network.

[0075] Step 4.2.7: Input the result of global average pooling into the fully connected layer.

[0076] Step 5: Perform blood pressure prediction based on the newly collected oscillatory wave signal data and PPG signal, including preprocessing of the oscillatory wave signal and PPG signal and blood pressure estimation based on machine learning and deep learning models; fuse the features extracted by deep learning with the manual features extracted by time-frequency analysis and statistical feature extraction, and then input them into the XGBoost model to learn the mapping relationship between the oscillatory wave signal, PPG signal and systolic blood pressure, diastolic blood pressure, so as to achieve accurate estimation of diastolic blood pressure and systolic blood pressure. The specific method is as follows: Step 5.1: Preprocess the newly collected oscillatory wave signal and PPG signal data, and filter the oscillatory wave signal and PPG signal.

[0077] Step 5.2: Extract features from the processed oscillatory wave signal and PPG signal. The specific method is as follows: Step 5.2.1: Use CNN to extract time-frequency features from the oscillatory wave signal and the PPG signal to capture local pattern information.

[0078] Step 5.2.2: Add positional encoding to the extracted features to preserve sequence information and enhance the model's ability to model temporal dependencies.

[0079] Step 5.2.3: Utilize the multi-head self-attention mechanism and the feed-forward neural network to capture the long-range dependencies of the signal and improve the feature expression ability; the long-range dependencies refer to the mutual correlations between elements at relatively distant time steps or spatial positions in the signal. For example, within a relatively long time period, there are connections between the waveform features of the oscillatory waves and the PPG signals generated by each heartbeat of the heart. The amplitude of the oscillatory wave at the current moment is related to the heart contraction intensity and blood vessel state several seconds ago, and this kind of correlation across a relatively long time interval is the long-range dependency.

[0080] Step 5.2.4: Perform average pooling on the extracted feature vectors to obtain feature vectors of a fixed length, reduce redundant information, and improve computational efficiency.

[0081] Step 5.3: Fuse the pooled feature vectors with the manually extracted time-frequency features and statistical features, and input them into the XGBoost model to learn the mapping relationship between the oscillatory wave signal, the PPG signal, and the blood pressure value, and achieve accurate estimation of diastolic and systolic blood pressures. As Figure 8 shown, the specific method is as follows: Step 5.3.1: Input the fused features.

[0082] Step 5.3.2: Perform data preprocessing, including data cleaning and data standardization. The data standardization formula is as follows: (8); where, x ij is the j-th feature value of the i-th sample, and x ij * is the j-th feature value of the i-th sample after standardization, μ j is the mean of the j-th feature, σ j is the standard deviation of the j-th feature.

[0083] Step 5.3.3: Divide the preprocessed dataset into a training set Dtrain and a test set Dtest .

[0084] Step 5.3.4: Set the hyperparameters of XGBoost, including the number of trees, the learning rate, and the maximum depth.

[0085] Step 5.3.5: Use the training setDtrain Train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean squared error: (9); where θ are the parameters of the XGBoost model, y i is the true blood pressure value, is the blood pressure value predicted by the model.

[0086] Step 5.3.6: Use the trained XGBoost model to predict the test set Dtest to obtain the predicted blood pressure value .

[0087] Step 5.3.7: Use evaluation metrics to measure the performance of the model. The evaluation metrics include the mean squared error MSE, the root mean squared error RMSE, and the mean absolute error MAE: (10); (11); (12); where n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

[0088] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present invention.

Claims

1. A blood pressure measurement system based on collaborative learning of oscillometric waves and PPG signals, characterized in that: It includes an oscillometric wave and PPG signal database construction module, a data preprocessing module, a feature extraction module, and a blood pressure estimation module; The oscillometric wave and PPG signal database stores the oscillometric waves and PPG signals of 3,000 subjects, as well as their corresponding diastolic and systolic blood pressures; The data preprocessing module is used to filter the acquired oscillometric wave signals and PPG signals to remove common noises; The feature extraction module constructs a CNN-Transformer network, including a CNN feature extractor, a positional encoder, multiple Transformer encoders, a global average pooling layer, a fully connected layer, and a feature fusion module. The CNN feature extractor is used to extract the local features of the oscillometric wave signals and PPG signals, and the multiple Transformer encoders are used to extract the temporal features of the oscillometric wave signals. On this basis, the positional encoder adds positional encoding to the extracted features to retain sequence information and enhance the model's perception ability of temporal features. The global average pooling layer performs global average pooling on the output of the multiple Transformer encoders and then inputs it into the fully connected layer. The feature fusion module combines time-frequency analysis and statistical features to extract manual features, performs feature fusion with the features extracted by deep learning, and divides the training set and the test set; The blood pressure estimation module uses the XGBoost model to estimate blood pressure based on the extracted features, achieving accurate prediction of systolic and diastolic blood pressures.

2. The blood pressure measurement system based on collaborative learning of oscillometric waves and PPG signals according to claim 1, characterized in that: The oscillometric wave and PPG signal database construction module includes an oscillometric wave signal device, a cuff, a PPG sensor, a blood pressure measurement device, and a database. First, the oscillometric wave signal device is correctly placed on the upper arm of the subject and fixed by the cuff. The cuff will gradually apply pressure and record the oscillometric wave data of the subject at different pressures and store it in the database. Then, the PPG sensor is correctly placed at the fingertip. During the acquisition process, the subject is required to remain still to reduce the influence of motion artifacts on the PPG signal. The systolic and diastolic blood pressures of the subject are measured by the blood pressure measurement device and stored in the database.

3. The blood pressure measurement system based on collaborative learning of oscillometric waves and PPG signals according to claim 1, characterized in that: The data preprocessing module includes an oscillometric wave data preprocessing module and a PPG data preprocessing module; The oscillometric wave data preprocessing module first uses a low-pass filter to denoise the original oscillometric wave signal, removing the high-frequency noise in the signal to retain the low-frequency components in the oscillometric wave signal. Then, it performs high-pass filtering to remove the low-frequency components in the signal to retain the high-frequency components in the oscillometric wave signal; The PPG data preprocessing module first uses low-pass filtering to remove the high-frequency noise in the PPG signal and retain the useful components in the PPG signal. Then, it removes the low-frequency drift or baseline drift through high-pass filtering to retain the important information in the PPG signal. It adopts a waveform quality assessment method based on wavelet transform to detect and remove motion artifacts or bad signal segments, and its calculation formula is as shown in Equation (1): （1）； wherein, a and b are scale and translation parameters, ψ is the mother wavelet function; f(t) is the original signal, which is a function of t; The signal is smoothed by the Savitzky-Golay filter technology to reduce high-frequency fluctuations, and its calculation formula is as shown in Equation (2): （2）； Among them, c j are Savitzky-Golay coefficients, m is half of the window length, x i+j is the original data value of the input signal at position i + j, where j is the offset index within the window.

4. The blood pressure measurement system based on collaborative learning of oscillometric waves and PPG signals according to claim 1, characterized in that: The specific implementation method of the blood pressure estimation module is: First, set the hyperparameters of XGBoost, including the number of trees, learning rate, and maximum depth; Then, use the training set to train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean squared error, as shown in the following formula: （9）； Among them, θ are the parameters of the XGBoost model, and y i is the actual blood pressure value, is the blood pressure value predicted by the model; Use the trained XGBoost model to predict the test set and obtain the predicted blood pressure values ; Then, use evaluation metrics to measure the performance of the model. The evaluation metrics include mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE): （10）； （11）； （12）； where n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, is the predicted blood pressure value of the i-th sample in the test set.

5. A blood pressure measurement method based on collaborative learning of oscillometric waves and PPG signals, characterized in that: It is implemented by the blood pressure measurement system based on the collaborative learning of oscillatory waves and PPG signals described in claim 1, including the following steps: Step 1: Construct a database of oscillatory wave signals and PPG signals, including oscillatory wave signals, PPG signals, and their corresponding systolic blood pressure and diastolic blood pressure; Step 2: Preprocess the oscillatory wave signals in the database, including low-pass filtering and high-pass filtering, to optimize the noise, baseline drift, and amplitude differences between different waveforms in the oscillatory wave signals; Step 3: Preprocess the PPG signals in the database, including low-pass filtering, high-pass filtering, artifact removal, and smoothing, to remove the noise, baseline drift, and amplitude differences between different waveforms in the PPG signals; Step 4: Perform feature extraction and blood pressure estimation modeling; establish a CNN-Transformer network, including a CNN feature extractor, positional encoding, multi-layer Transformer encoders, global average pooling, and a fully connected layer; use CNN to extract the local features of oscillatory wave signals and PPG signals and splice them to form joint features; then use multi-layer Transformer encoders to extract the temporal features of oscillatory wave signals; add positional encoding to the extracted features to retain sequence information and enhance the model's perception ability of temporal features; perform global average pooling on the output of the multi-layer Transformer encoders and then input it into the fully connected layer; Step 5: Perform blood pressure prediction based on newly collected oscillatory wave signal data and PPG signals, including preprocessing of oscillatory wave signals and PPG signals and blood pressure estimation based on machine learning and deep learning models; fuse the features extracted by deep learning with the manual features extracted by time-frequency analysis and statistical feature extraction, and then input them into the XGBoost model to learn the mapping relationship between oscillatory wave signals, PPG signals, and systolic blood pressure and diastolic blood pressure, so as to achieve accurate estimation of diastolic blood pressure and systolic blood pressure.

6. The blood pressure measurement method based on collaborative learning of oscillometric waves and PPG signals according to claim 5, wherein: The specific method of step 1 is as follows: Step 1.1: Correctly place the oscillatory wave signal device on the upper arm of the subject and fix it with a cuff; Step 1.2: Gradually pressurize the cuff, record the oscillatory wave data of the subject under different pressures, and store them; Step 1.3: Correctly place the PPG sensor at the fingertip; during the acquisition process, require the subject to stay still to reduce the influence of motion artifacts on the PPG signal; Step 1.4: Measure the systolic blood pressure and diastolic blood pressure of the subject through a blood pressure measurement device.

7. The blood pressure measurement method based on collaborative learning of oscillation wave and PPG signal according to claim 5, characterized in that: The specific method of step 2 is as follows: Step 2.1: Use a low-pass filter to remove the high-frequency noise in the original oscillatory wave signal, so as to retain the low-frequency components in the oscillatory wave signal; Step 2.2: Use high-pass filtering to remove the influence of baseline drift on the oscillatory wave, thereby retaining the high-frequency components in the oscillatory wave signal.

8. The blood pressure measurement method based on collaborative learning of oscillometric wave and PPG signal according to claim 5, characterized in that: The specific method of step 3 is as follows: Step 3.1: Use low-pass filtering to remove high-frequency noise and retain the useful components in the PPG signal; Step 3.2: Use high-pass filtering to remove low-frequency drift or baseline drift and retain the important information in the PPG signal; Step 3.3: Adopt the wavelet transform waveform quality assessment method to detect and remove motion artifacts or bad signal segments; Step 3.4: Smooth the signal through the Savitzky-Golay filter technology to reduce high-frequency fluctuations.

9. The blood pressure measurement method based on collaborative learning of oscillometric waves and PPG signals according to claim 5, wherein: The specific method of step 4 is as follows: Step 4.1: Perform data preparation, load the oscillatory wave signal, PPG signal, and their corresponding diastolic and systolic blood pressure values, and divide the training set and test set; Step 4.2: Establish a deep learning model, that is, establish a CNN-Transformer network, including a CNN feature extractor, positional encoding, multi-layer Transformer encoders, global average pooling, and a fully connected layer; Step 4.2.1: Input the oscillatory wave signal and PPG signal; Step 4.2.2: Construct the first CNN network to extract features from the oscillatory wave signal. The specific method is as follows: Construct convolutional layer 1 with a kernel size of 5, 16 convolutional kernels, and use the ReLU activation function; Pooling layer 1 with max pooling and a pooling kernel size of 2; Construct convolutional layer 2 with a kernel size of 5, 32 convolutional kernels, and use the ReLU activation function; Pooling layer 2 with max pooling and a pooling kernel size of 2; Output layer 1 to output the features extracted from the oscillatory wave signal; Step 4.2.3: Construct the second CNN network to extract features from the PPG signal. The specific method is as follows: Construct convolutional layer 3 with a kernel size of 3, 16 convolutional kernels, and use the ReLU activation function; Pooling layer 3 with max pooling and a pooling kernel size of 2; Construct convolutional layer 4 with a kernel size of 5, 32 convolutional kernels, and use the ReLU activation function; Pooling layer 4 with max pooling and a pooling kernel size of 2; Construct convolutional layer 5 with a kernel size of 7, 64 convolutional kernels, and use the ReLU activation function; Pooling layer 5 with average pooling and a pooling kernel size of 2; Output layer 2 to output the features extracted from the PPG signal; Step 4.2.4: Concatenate the feature vectors extracted from the oscillatory wave signal and PPG signal to form a joint feature, and then transform the input shape to convert the output of the CNN part into the shape required by the Transformer network; Step 4.2.5: Construct a Transformer network to capture the temporal dependence of the oscillatory wave signal. The specific method is as follows: Step 4.2.5.1: Input embedding to convert the input features into high-dimensional feature vectors; Step 4.2.5.2: Positional encoding, add positional encoding to retain the temporal information and introduce the temporal information into the feature vectors. The calculation formula is as shown in Equation (5): （5）； Among them, pos is the position, i is the dimension, d model is the dimension of the model; Step 4.2.5.3: The multi-head self-attention mechanism calculates attention scores and performs weighted summation based on these scores to capture the dependencies between different positions in the input sequence; the calculation process of self-attention is as shown in Equation (6): （6）； Among them, Q , K and V are obtained by linear transformation of the input features, d k is K 's dimension; Step 4.2.5.4: Layer normalization to stabilize and accelerate the training process; Step 4.2.5.5: The feed-forward neural network further processes the features through non-linear activation functions to enhance the model's expressive power, and the calculation process is as shown in Equation (7): （7）； Among them, max(0, xW1 + b1) is a non-linear activation function, W 1 and W 2 are weight matrices, b 1 and b 2 are biases; Step 4.2.5.6: Layer normalization to stabilize and accelerate the training process; Step 4.2.5.7: Repeat the construction from Step 4.2.5.3 to Step 4.2.5.6 six times; Step 4.2.6: Perform global average pooling on the output of the Transformer network; Step 4.2.7: Input the result of global average pooling into the fully connected layer.

10. The blood pressure measurement method based on collaborative learning of oscillometric waves and PPG signals according to claim 9, characterized in that: The specific method of Step 5 is as follows: Step 5.1: Preprocess the newly collected oscillatory wave signal and PPG signal data, and filter the oscillatory wave signal and PPG signal; Step 5.2: Extract features from the processed oscillatory wave signal and PPG signal; Step 5.2.1: Use CNN to extract time-frequency features from the oscillatory wave signal and PPG signal to capture local pattern information; Step 5.2.2: Add positional encoding to the extracted features to retain sequence information and enhance the model's ability to model temporal dependencies; Step 5.2.3: Utilize the multi-head self-attention mechanism and the feed-forward neural network to capture the long-range dependencies of the signal and improve the feature expression ability; Step 5.2.4: Perform average pooling on the extracted feature vectors to obtain fixed-length feature vectors, reduce redundant information, and improve computational efficiency; Step 5.3: Fuse the pooled feature vectors with the manually extracted time-frequency features and statistical features, and input them into the XGBoost model to learn the mapping relationship between the oscillatory wave signal, PPG signal, and blood pressure value, and achieve accurate estimation of diastolic and systolic blood pressure; the specific method is as follows: Step 5.3.1: Input the fused features; Step 5.3.2: Data preprocessing, including data cleaning and data standardization; the data standardization formula is as follows: （8）； where, x ij is the j-th eigenvalue of the i-th sample, and x ij * is the j-th eigenvalue of the standardized i-th sample, μ j is the mean of the j-th feature, σ j is the standard deviation of the j-th feature; Step 5.3.3: Divide the preprocessed dataset into a training set Dtrain and a test set Dtest ; Step 5.3.4: Set the hyperparameters of XGBoost, including the number of trees, learning rate, and maximum depth; Step 5.3.5: Use the training set Dtrain to train the XGBoost model; the objective function of XGBoost is to minimize the loss function, and the loss function is the mean squared error: （9）； Among them, θ is the parameter of the XGBoost model, and y i is the real blood pressure value, is the blood pressure value predicted by the model; Step 5.3.6: Use the trained XGBoost model to predict the test set Dtest to obtain the predicted blood pressure values ; Step 5.3.7: Use evaluation metrics to measure the performance of the model, and the evaluation metrics include mean squared error MSE, root mean squared error RMSE, and mean absolute error MAE: （10）； （11）； （12）； where n test is the number of samples in the test set, y test,i is the true blood pressure value of the i-th sample in the test set, and is the predicted blood pressure value of the i-th sample in the test set.

Images