Blood pressure assessment method based on relational attention mechanism

Abstract

The present invention discloses a blood pressure assessment method based on a relational attention mechanism, which belongs to the technical field of blood pressure assessment. It solves the problem that when the existing method combines photoplethysmography and electrocardiogram signals to assess blood pressure, the PPG signal is highly sensitive to noise, which hinders the signal recognition process and the final assessment performance. The method includes real-time acquisition of blood pressure-related signals, preprocessing of the blood pressure-related signals, extraction of signal key points in the blood pressure-related signals, construction of a blood pressure graph topology based on the signal key points, and a graph neural network model that identifies and processes the blood pressure graph topology based on the spatiotemporal dependency of physiological signals, and outputs a blood pressure assessment value. In an embodiment of the present invention, a graph neural network model based on a relational attention mechanism is provided. The graph neural network model combines signal preprocessing, graph construction and a relational attention mechanism, and extracts hidden information from graph topology data by enhancing feature extraction capabilities, thereby improving the accuracy of blood pressure estimation.

Description

Technical Field

[0001] The present invention belongs to the technical field of blood pressure assessment, and specifically relates to a blood pressure assessment method based on a relational attention mechanism. Background Art

[0002] Cardiovascular diseases (CVDs) remain one of the leading causes of high morbidity and mortality from chronic non-communicable diseases worldwide. As a key physiological indicator in health monitoring, accurate measurement of blood pressure (BP) is crucial for the early detection and management of CVDs. However, existing BP measurement methods have numerous limitations and are in urgent need of improvement and innovation.

[0003] Currently, cuff-based non-invasive blood pressure monitoring is the most widely used method for measuring blood pressure. Although its measurement results are relatively accurate, the repeated inflation and deflation of the cuff during 24-hour ambulatory blood pressure monitoring not only interferes with the patient's daily activities but also causes discomfort. Furthermore, cuff-based blood pressure monitors cannot provide continuous blood pressure data, making it difficult to meet the needs of continuous blood pressure monitoring in clinical and daily health monitoring. Therefore, the development of a cuff-free non-invasive blood pressure measurement method has become an urgent need in the current field of blood pressure monitoring.

[0004] In recent years, impedance-based measurement methods have been proposed as a potential solution. However, establishing reliable, durable, and stable contact quality between the electrode and the skin is a major challenge, requiring constant attention to sensor placement and maintenance. Furthermore, millimeter wave-based blood pressure estimation methods primarily rely on single-channel pulse wave analysis, but due to limitations in accuracy and portability, they have struggled to develop into a widely adopted technology. Pulse transit time (PTT) methods typically combine photoplethysmography (PPG) and electrocardiogram (ECG) signals to assess blood pressure. Although PPG signals are increasingly common in smart wearable devices, their high sensitivity to noise (such as motion artifacts and ambient light interference) can severely hinder signal recognition and ultimately assessment performance. Significant individual variability and the influence of physiological and environmental factors persist, limiting the accuracy of existing methods. To address these issues, we propose a blood pressure assessment method based on a relational attention mechanism. Summary of the Invention

[0005] The purpose of the present invention is to address the shortcomings of the existing technology and provide a blood pressure assessment method based on the relational attention mechanism, which solves the problem that when the existing method combines photoplethysmography and electrocardiogram signals to assess blood pressure, the PPG signal is highly sensitive to noise, which hinders the signal recognition process and the final evaluation performance.

[0006] The present invention is implemented as follows: a blood pressure assessment method based on a relational attention mechanism comprises:

[0007] Real-time acquisition of blood pressure-related signals, pre-processing of the blood pressure-related signals, and extraction of key signal points in the blood pressure-related signals, wherein the blood pressure-related signals include original PPG signals and original ECG signals;

[0008] Load the key points of the blood pressure-related signal and construct the blood pressure map topology based on the key points of the signal;

[0009] Pre-build a graph neural network model based on the relational attention mechanism, iteratively train the graph neural network model, and output a converged graph neural network model;

[0010] The blood pressure graph topology is loaded, and the graph neural network model identifies and processes the blood pressure graph topology based on the spatiotemporal dependencies of physiological signals, and outputs the blood pressure assessment value.

[0011] The method for preprocessing blood pressure-related signals comprises:

[0012] Loading blood pressure-related signals, performing denoising and baseline correction on the blood pressure-related signals, and obtaining a correction signal set;

[0013] Acquire a correction signal set, and segment the correction signal set based on a sliding window method to segment the correction signal set into at least one set of overlapping time windows;

[0014] Applying EMD decomposition to the signal within each set of time windows to separate the intrinsic mode functions;

[0015] Among them, when using EMD decomposition, the formula is expressed as:

[0016]

[0017] in, Indicates time The original signal, The signal Intrinsic Mode Function (IMF), in order to remove noise and motion artifacts, the middle IMFs are retained, while the IMFs related to noise and motion artifacts are discarded;

[0018] Load the correction signal set after EMD decomposition, reconstruct the correction signal set after decomposition and noise reduction, and obtain the signal reconstruction set;

[0019] The signal reconstruction set is expressed as:

[0020]

[0021] in is the set of modal functions IMFs retained for reconstruction.

[0022] The method for preprocessing the blood pressure-related signal further includes:

[0023] Key points of the signal are extracted within the time window. The AMPD algorithm is used to extract key points for the PPG signal in the signal reconstruction set, and the Neurokit2 library is used to extract key points for the ECG signal in the signal reconstruction set, thereby obtaining the key points of the PPG signal and the ECG signal.

[0024] The key points of the PPG signal are represented as a, b, c, d, e, and the key points of the ECG signal are represented as P, Q, R, S, T.

[0025] The time characteristics and amplitude characteristics of the PPG signal key points and ECG signal key points in the time window are expressed by the following formulas:

[0026]

[0027] in, represents the average time feature in the i-th window, is a function that calculates the time difference between consecutive key points, is the timestamp of the key point in the i-th window;

[0028] represents the average normalized amplitude feature within the i-th window, is a function that normalizes and averages the keypoint amplitudes, is the amplitude set of key points in the i-th window.

[0029] When collecting blood pressure-related signals in real time, the original PPG signal is obtained through the fingertip sensor, and the original ECG signal is recorded through the standard lead configuration. The blood pressure-related signals are sampled at a frequency of 125 Hz and 12-bit accuracy.

[0030] The method for constructing a blood pressure map topology based on signal key points includes:

[0031] Load the PPG signal key points and ECG signal key points, use the PPG signal key points and ECG signal key points as nodes of the blood pressure graph topology, and encode the time relationship between the PPG signal key points and ECG signal key points as edges of the blood pressure graph topology;

[0032] Among them, represents the blood pressure map topology of the i-th window, where is a collection of nodes (PPG signal key points and ECG signal key points), is a set of edges (PPG signal key points and ECG signal key points time relationship), each node With a feature vector Associated, each edge With a feature vector associated;

[0033] The characteristics of the blood pressure graph topology nodes and edges are defined as follows:

[0034]

[0035] in, is a node timestamp, is the amplitude of the key point is the time difference between consecutive key points, is a node timestamp;

[0036] Update the features of the topological nodes and edges of the blood pressure graph based on the graph-based relational attention mechanism;

[0037] node The features of the edge nodes are updated by aggregating the information of their neighbor nodes. The aggregation process calculates the mean, maximum, minimum and standard deviation of the neighbor node features to form a feature vector, which is then passed through the multi-layer perceptron MLP to obtain the updated node features. The features of are updated by aggregating the information of its connected nodes. The aggregation process calculates the mean, maximum, minimum and standard deviation of the connected node features to form a feature vector, which is also passed through the MLP to obtain the updated edge features.

[0038] The graph neural network model includes a first relational attention module, a second relational attention module, an average pooling layer, a softmax layer, and a batch normalization layer. The activation function of the graph neural network model is a Relu function. There are two groups of batch normalization layers, which are respectively arranged after the first relational attention module and the second relational attention module.

[0039] The method for identifying and processing the topology of a blood pressure graph based on the spatiotemporal dependency of physiological signals using a graph neural network model includes:

[0040] Load the blood pressure graph topology and input it into the graph neural network model. The first relational attention module aggregates information from neighboring nodes while considering edge attributes to update node features. The updated features are normalized to stabilize training, and a nonlinear activation function is used to enhance the expressive power of the graph neural network model.

[0041] The second relational attention module further refines the representation of node features. The refined node features are obtained through the mean pooling operation of the average pooling layer. The pooling operation aggregates the node-level features into a single vector representation.

[0042] Load the vector representation, identify and process the vector representation through the softmax layer, and predict and generate the blood pressure assessment value;

[0043] The predicted blood pressure assessment value is calculated using the following formula:

[0044]

[0045] Among them, X represents the input node features, A is the adjacency matrix, and is a trainable weight matrix, represents the activation function, and Used to map the pooled representation to the final predicted blood pressure assessment value.

[0046] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0047] In an embodiment of the present invention, a graph neural network model based on a relational attention mechanism is provided. The graph neural network model combines signal preprocessing, graph construction and a relational attention mechanism. By enhancing feature extraction capabilities, it extracts hidden information from graph topology data, thereby improving the accuracy of blood pressure estimation. When predicting and generating blood pressure assessment values, the mean absolute errors (MAE) in diastolic blood pressure (DBP) and systolic blood pressure (SBP) estimation are 1.3520 mmHg and 2.9862 mmHg, respectively, which meet the standards of the American Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS) A-level standards.

[0048] In this embodiment of the present invention, the AMPD algorithm uses a multi-scale sliding window to detect local maxima, enabling it to handle complex signal structures. This makes it particularly useful for analyzing multiple frequency components and nonlinear features in PPG signals. Combining key point extraction with PPG and ECG signals enables more comprehensive physiological signal analysis.

[0049] In this embodiment of the present invention, by accurately extracting the key points of PPG and ECG signals, important features in the signals can be captured more accurately. Simultaneously, by constructing a blood pressure graph topology, the complex relationships between signals can be fully captured, improving the accuracy of signal analysis. The noise resistance and stability of the AMPD algorithm and the Neurokit2 library make key point extraction more robust. The graph's relational attention mechanism can effectively handle noise and outliers, further improving the robustness of signal processing. By constructing a blood pressure graph topology, more comprehensive physiological signal analysis can be achieved, providing strong support for the early detection and management of cardiovascular disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1This is a schematic diagram of the implementation flow of the blood pressure assessment method based on the relational attention mechanism provided by the present invention.

[0051] Figure 2 A Bland-Altman plot of DBP estimation for the graph neural network model in an embodiment of the present invention is shown.

[0052] Figure 3 A Bland-Altman plot of SBP estimation of the graph neural network model in an embodiment of the present invention is shown.

[0053] Figure 4 A histogram analysis of the DBP estimation error of the graph neural network model in an embodiment of the present invention is shown.

[0054] Figure 5 A histogram analysis of the SBP estimation error of the graph neural network model in an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meanings as commonly understood by those skilled in the art to which this application belongs. The terms used in the specification of the application are only for the purpose of describing specific embodiments and are not intended to limit this application. The terms "including" and "having" and any variations thereof in the specification and claims of this application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of this application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0056] When existing methods combine photoplethysmography and electrocardiogram signals to evaluate blood pressure, the high sensitivity of the PPG signal to noise hinders the signal recognition process and the final evaluation performance. To address the above problems, we propose a blood pressure evaluation method based on a relational attention mechanism. In short, when the method is implemented, the blood pressure-related signals are first collected in real time, the blood pressure-related signals are preprocessed, the signal key points in the blood pressure-related signals are extracted, the blood pressure graph topology is constructed based on the signal key points, and a graph neural network model based on a relational attention mechanism is preconstructed. Finally, the graph neural network model identifies and processes the blood pressure graph topology based on the spatiotemporal dependency of the physiological signals and outputs a blood pressure evaluation value. In an embodiment of the present invention, a graph neural network model based on a relational attention mechanism is provided. The graph neural network model combines signal preprocessing, graph construction, and a relational attention mechanism to enhance feature extraction capabilities, extract hidden information from graph topology data, and improve the accuracy of blood pressure estimation.

[0057] The embodiment of the present invention provides a blood pressure assessment method based on the relational attention mechanism. Figure 1A schematic diagram of the implementation process of a blood pressure assessment method based on a relational attention mechanism is shown. The blood pressure assessment method based on a relational attention mechanism includes:

[0058] S10, collecting blood pressure-related signals in real time, preprocessing the blood pressure-related signals, and extracting key points in the blood pressure-related signals, wherein the blood pressure-related signals include original PPG signals and original ECG signals;

[0059] It should be noted that when collecting blood pressure-related signals in real time, the original PPG signal is obtained through the fingertip sensor, and the original ECG signal is recorded through the standard lead configuration. The blood pressure-related signal is sampled at a frequency of 125 Hz and 12-bit accuracy.

[0060] S20, loading the signal key points of the blood pressure related signal, and constructing a blood pressure map topology based on the signal key points;

[0061] S30: Pre-build a graph neural network model based on the relational attention mechanism, iteratively train the graph neural network model, and output a converged graph neural network model;

[0062] S40, loading the blood pressure graph topology, the graph neural network model identifies and processes the blood pressure graph topology based on the spatiotemporal dependency of physiological signals, and outputs a blood pressure assessment value.

[0063] In an embodiment of the present invention, the method for preprocessing a blood pressure-related signal includes:

[0064] S101, loading blood pressure related signals, performing denoising and baseline correction on the blood pressure related signals to obtain a correction signal set;

[0065] In this embodiment of the present invention, the raw PPG and ECG signals are preprocessed to remove noise and artifacts, ensuring high-quality input to the graph neural network model. This preprocessing step includes, but is not limited to, denoising, baseline correction, and exclusion of irregular or discontinuous signal segments. Furthermore, an EMD decomposition algorithm is applied to further reduce motion artifacts and enhance signal quality.

[0066] S102, obtaining a correction signal set, and segmenting the correction signal set based on a sliding window method to segment the correction signal set into at least one set of overlapping time windows;

[0067] It should be noted that the width of each time window is 3 seconds and the sliding step is 1.5 seconds.

[0068] S103, applying EMD decomposition to the signal within each group of the time windows to separate the intrinsic mode function;

[0069] Among them, when using EMD decomposition, the formula is expressed as:

[0070]

[0071] in, Indicates time The original signal, The signal Intrinsic Mode Function (IMF), in order to remove noise and motion artifacts, the middle IMFs are retained, while the IMFs related to noise and motion artifacts are discarded;

[0072] S104, loading the correction signal set after EMD decomposition, reconstructing the correction signal set after decomposition and noise reduction to obtain a signal reconstruction set;

[0073] The signal reconstruction set is expressed as:

[0074]

[0075] in is the set of modal functions IMFs retained for reconstruction.

[0076] S105, extracting key points from the signal within the time window, wherein the AMPD algorithm is used to extract key points from the PPG signal in the signal reconstruction set, and the Neurokit2 library is used to extract key points from the ECG signal in the signal reconstruction set, to obtain key points of the PPG signal and key points of the ECG signal;

[0077] It should be noted that in the embodiments of the present invention, the automatic multi-scale peak detection (AMPD) algorithm is used to extract key points in PPG signals. This algorithm is highly efficient and particularly suitable for periodic or quasi-periodic signals. This algorithm uses a multi-scale sliding window to detect local maxima, demonstrating excellent adaptability and noise immunity. For PPG signals, the AMPD algorithm can accurately extract key points (such as peaks and troughs), thereby improving the accuracy of signal analysis. Automatic multi-scale peak detection.

[0078] The key points of the PPG signal are represented as a, b, c, d, e, and the key points of the ECG signal are represented as P, Q, R, S, T.

[0079] In this embodiment of the present invention, the AMPD algorithm uses a multi-scale sliding window to detect local maxima, enabling it to process complex signal structures. This makes it particularly useful for analyzing multiple frequency components and nonlinear features in PPG signals. Combining keypoint extraction with PPG and ECG signals enables more comprehensive physiological signal analysis. For example, by analyzing keypoints in both PPG and ECG signals, the functional status of the cardiovascular system can be more accurately assessed.

[0080] In the embodiment of the present invention, the time characteristics and amplitude characteristics of the key points of the PPG signal and the ECG signal in the time window are expressed by the following formulas:

[0081]

[0082] in, represents the average time feature in the i-th window, is a function that calculates the time difference between consecutive key points, is the timestamp of the key point in the i-th window;

[0083] represents the average normalized amplitude feature within the i-th window, is a function that normalizes and averages the keypoint amplitudes

[0084] An embodiment of the present invention provides a method for constructing a blood pressure map topology based on signal key points. The method for constructing a blood pressure map topology based on signal key points specifically includes:

[0085] S201, loading PPG signal key points and ECG signal key points, using the PPG signal key points and ECG signal key points as nodes of the blood pressure graph topology, and encoding the time relationship between the PPG signal key points and the ECG signal key points as edges of the blood pressure graph topology;

[0086] Among them, represents the blood pressure map topology of the i-th window, where is a collection of nodes (PPG signal key points and ECG signal key points), is a set of edges (PPG signal key points and ECG signal key points time relationship), each node With a feature vector Associated, each edge With a feature vector associated;

[0087] The characteristics of the blood pressure graph topology nodes and edges are defined as follows:

[0088]

[0089] in, is a node timestamp, is the amplitude of the key point is the time difference between consecutive key points, is a node timestamp;

[0090] S202, updating the features of the topological nodes and edges of the blood pressure graph based on the graph relational attention mechanism;

[0091] Among them, the node The features of the edge nodes are updated by aggregating the information of their neighbor nodes. The aggregation process calculates the mean, maximum, minimum and standard deviation of the neighbor node features to form a feature vector, which is then passed through the multi-layer perceptron MLP to obtain the updated node features. The features of are updated by aggregating the information of its connected nodes. The aggregation process calculates the mean, maximum, minimum and standard deviation of the connected node features to form a feature vector, which is also passed through the MLP to obtain the updated edge features.

[0092] It should be noted that the blood pressure map topology construction process involves mapping key points identified from PPG and ECG signals to nodes and encoding the temporal relationships between these points as edges in the temporal graph. At the same time, the graph-based relational attention mechanism updates the features of the blood pressure map topology nodes and edges. Unlike the standard transformer attention, which only calculates the query (Q), key (K), and value (V) vectors from node features, the relational attention incorporates edge features into the calculation of Q, K, and V. This is achieved by concatenating the edge vector with the node vector before applying the linear transformation.

[0093] In this embodiment of the present invention, by accurately extracting the key points of PPG and ECG signals, important features in the signals can be captured more accurately. Simultaneously, by constructing a blood pressure graph topology, the complex relationships between signals can be fully captured, improving the accuracy of signal analysis. The noise resistance and stability of the AMPD algorithm and the Neurokit2 library make key point extraction more robust. The graph's relational attention mechanism can effectively handle noise and outliers, further improving the robustness of signal processing. By constructing a blood pressure graph topology, more comprehensive physiological signal analysis can be achieved, providing strong support for the early detection and management of cardiovascular disease.

[0094] In an embodiment of the present invention, the graph neural network model includes a first relational attention module, a second relational attention module, an average pooling layer, a softmax layer, and a batch normalization layer. The activation function of the graph neural network model is a Relu function. There are two groups of batch normalization layers, which are respectively arranged after the first relational attention module and the second relational attention module.

[0095] Specifically, if Figure 1 As shown, the first relational attention module is connected to the batch normalization layer, the batch normalization layer is connected to the second relational attention module through the Relu function, the second relational attention module is connected to another group of batch normalization layers, the batch normalization layer is connected to the average pooling layer, and the average pooling layer is connected to the softmax layer.

[0096] The UCI-BP dataset was used as the training dataset for the pre-built graph neural network model based on the relational attention mechanism and the iterative training of the graph neural network model. PPG signals were acquired via fingertip sensors, while ECG signals were recorded using a standard lead configuration. All signals were sampled at 125 Hz and 12-bit precision. The dataset underwent preliminary preprocessing, including denoising, baseline correction, and exclusion of irregular or discontinuous signal segments. The performance of the proposed graph neural network model was validated using this dataset, including a comparison of physiological data preprocessing results and a discussion of the effectiveness of the blood pressure estimation algorithm to support the validity of the proposed method.

[0097] The performance of the proposed graph neural network model is evaluated through regression analysis, Bland-Altman analysis and error distribution histogram.

[0098] Specifically, the graph neural network model is used to estimate DBP estim and SBP estim .

[0099] Regression analysis:

[0100] DBP estim and SBP estim Perform regression analysis to compare them with the reference values. The results of a one-dimensional linear fit of the data are shown below:

[0101]

[0102] Compared with the ideal linear equation, it can be observed that the DBP estim It showed the strongest linear correlation, with a proportionality coefficient (kkk) of 0.9709. It was followed by SBP estim , its kkk value is 0.9232. Table 1 shows the performance evaluation comparison results of the graph neural network model of the embodiment of the present invention and the benchmark model.

[0103] Table 1

[0104]

[0105] Bland-Altman analysis:

[0106] To assess the consistency and variability of blood pressure measurements, Figure 2-Figure 3 The results of the Bland-Altman analysis are shown, where Figure 2 FIG4 shows a Bland-Altman diagram of DBP estimation of the graph neural network model in an embodiment of the present invention. Figure 3 The Bland-Altman plot of the SBP estimation of the graph neural network model in the embodiment of the present invention is shown. The labels in the figure are:

[0107]

[0108] and, Mean blood pressure ( ) and blood pressure differences ( ) is calculated as follows:

[0109]

[0110] The statistical values ​​of the errors in the blood pressure estimates show that the standard deviation (STD) follows the pattern:

[0111]

[0112] This is consistent with the correlation trend observed in blood pressure measurements. The Bland-Altman plot further verifies the consistency of the estimated blood pressure values, with most points concentrated in Nearly 95.9% of the samples for DBP and SBP were within the mean error plus 1.96 standard deviations, indicating strong agreement between the estimated and reference blood pressure values. Furthermore, the percentages of samples within the mean error plus 1.96 standard deviations for DBP and SBP, respectively, were 95.94% and 95.09%, exceeding the 95% standard deviation. Furthermore, the results obtained using the UCI-BP dataset met the AAMI criteria, which stipulate that the mean absolute error (MAE) should not exceed 5 mmHg and the standard deviation should be less than 8 mmHg.

[0113] Estimation error histogram analysis:

[0114] In this paper, the BP estimation error distribution of 1200 data points is analyzed by histogram, and the results are as follows: Figure 4-Figure 5 As shown, Figure 4 The histogram analysis of the DBP estimation error of the graph neural network model in the embodiment of the present invention is shown. Figure 5 A histogram analysis of the SBP estimation error of the neural network model in the figure embodiment of the present invention is shown. In addition, Table 1 provides the cumulative percentage (CP) of the absolute value of the estimation error, evaluated according to the British Hypertension Society (BHS) criteria. According to the BHS criteria, the proposed BP estimation method is consistent with the Class A accuracy standard. Specifically, for SBP and DBP estimation, the CP of the method exceeds 60% within the range of ±5 mmHg, exceeds 85% within the range of ±10 mmHg, and exceeds 95% within the range of ±15 mmHg. The error percentages of SBP and DBP are significantly higher than the values ​​specified by the BHS criteria.

[0115] An embodiment of the present invention further provides a method for using a graph neural network model to identify and process the topology of a blood pressure graph based on the spatiotemporal dependencies of physiological signals. The method for using a graph neural network model to identify and process the topology of a blood pressure graph based on the spatiotemporal dependencies of physiological signals specifically includes:

[0116] S301: Load the blood pressure graph topology and input it into the graph neural network model. The first relational attention module aggregates information from neighboring nodes while considering edge attributes to update node features. The updated features are normalized to stabilize training, and a nonlinear activation function is used to enhance the expressive power of the graph neural network model.

[0117] S302: The second relational attention module further refines the node features. The refined node features are obtained through the average pooling operation of the average pooling layer. The pooling operation aggregates the node-level features into a single vector representation.

[0118] S303, loading the vector representation, identifying and processing the vector representation through a softmax layer, and predicting and generating a blood pressure assessment value;

[0119] The predicted blood pressure assessment value is calculated using the following formula:

[0120]

[0121] Among them, X represents the input node features, A is the adjacency matrix, and is a trainable weight matrix, represents the activation function, and Used to map the pooled representation to the final predicted blood pressure assessment value.

[0122]

[0123] The first and second relational attention modules iteratively refine node embeddings by considering structural and relational dependencies, ensuring that temporal and contextual information is effectively captured. In an embodiment of the present invention, a method for processing blood pressure graph topology based on spatiotemporal dependencies of physiological signals using a graph neural network model is proposed. This method is used for cuffless blood pressure estimation using synchronized PPG and ECG signals. The proposed graph neural network model utilizes graph-based representations and relational attention mechanisms to effectively capture spatiotemporal dependencies, achieving superior accuracy and robustness under the AAMI standard.

[0124] In summary, the present invention provides a blood pressure assessment method based on a relational attention mechanism. In an embodiment of the present invention, a graph neural network model based on a relational attention mechanism is provided. The graph neural network model combines signal preprocessing, graph construction and relational attention mechanism, and extracts hidden information from graph topology data by enhancing feature extraction capabilities, thereby improving the accuracy of blood pressure estimation. When predicting and generating blood pressure assessment values, the mean absolute errors (MAE) in diastolic pressure (DBP) and systolic blood pressure (SBP) estimation are 1.3520 mmHg and 2.9862 mmHg, respectively, which meet the standards of the American Association for the Advancement of Medical Instrumentation (AAMI) and the British Hypertension Society (BHS) A-level standards.

[0125] It should be noted that for the aforementioned embodiments, for simplicity of description, they are all expressed as a series of action combinations. However, those skilled in the art should be aware that the present invention is not limited by the order of the actions described, because according to the present invention, certain steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention.

[0126] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the scope of protection of the invention. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on these embodiments, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in this field can still combine, add, delete or make other adjustments to the features in the various embodiments of the present invention according to the circumstances without conflict, without making creative work, so as to obtain different other technical solutions that do not deviate from the concept of the present invention in essence, and these technical solutions also fall within the scope of protection of the present invention.

Claims

1. A blood pressure assessment method based on relational attention mechanism, characterized in that: include: Real-time acquisition of blood pressure-related signals, pre-processing of the blood pressure-related signals, and extraction of key signal points in the blood pressure-related signals, wherein the blood pressure-related signals include original PPG signals and original ECG signals; Load the key points of the blood pressure-related signal and construct the blood pressure map topology based on the key points of the signal; Pre-build a graph neural network model based on the relational attention mechanism, iteratively train the graph neural network model, and output a converged graph neural network model; The blood pressure graph topology is loaded, and the graph neural network model identifies and processes the blood pressure graph topology based on the spatiotemporal dependencies of physiological signals, outputting the blood pressure assessment value. The method for constructing a blood pressure map topology based on signal key points includes: Load the PPG signal key points and ECG signal key points, use the PPG signal key points and ECG signal key points as nodes of the blood pressure graph topology, and encode the time relationship between the PPG signal key points and ECG signal key points as edges of the blood pressure graph topology; Among them, represents the blood pressure map topology of the i-th window, where is a collection of nodes, is a set of edges, each node With a feature vector Associated, each edge With a feature vector associated; The characteristics of the blood pressure graph topology nodes and edges are defined as follows: in, is a node timestamp, is the amplitude of the key point, is the time difference between consecutive key points, is a node timestamp; Update the features of the topological nodes and edges of the blood pressure graph based on the graph-based relational attention mechanism; node The features of the edge nodes are updated by aggregating the information of their neighbor nodes. The aggregation process calculates the mean, maximum, minimum and standard deviation of the neighbor node features to form a feature vector, which is then passed through the multi-layer perceptron MLP to obtain the updated node features. The features of are updated by aggregating the information of its connected nodes. The aggregation process calculates the mean, maximum, minimum and standard deviation of the connected node features to form a feature vector, which is also passed through the MLP to obtain the updated edge features.

2. The blood pressure assessment method based on the relational attention mechanism according to claim 1, wherein: The method for preprocessing blood pressure-related signals comprises: Loading blood pressure-related signals, performing denoising and baseline correction on the blood pressure-related signals, and obtaining a correction signal set; Acquire a correction signal set, and segment the correction signal set based on a sliding window method to segment the correction signal set into at least one set of overlapping time windows; Applying EMD decomposition to the signal within each set of time windows to separate the intrinsic mode functions; Among them, when using EMD decomposition, the formula is expressed as: in, Indicates time The original signal, The signal In order to remove noise and motion artifacts, the middle IMFs are retained, while the IMFs related to noise and motion artifacts are discarded; Load the correction signal set after EMD decomposition, reconstruct the correction signal set after decomposition and noise reduction, and obtain the signal reconstruction set; The signal reconstruction set is expressed as: in is the set of modal functions IMFs retained for reconstruction.

3. The blood pressure assessment method based on the relational attention mechanism according to claim 2, wherein: The method for preprocessing the blood pressure-related signal further includes: Key points of the signal are extracted within the time window. The AMPD algorithm is used to extract key points for the PPG signal in the signal reconstruction set, and the Neurokit2 library is used to extract key points for the ECG signal in the signal reconstruction set, thereby obtaining the key points of the PPG signal and the ECG signal. The key points of the PPG signal are represented as a, b, c, d, e, and the key points of the ECG signal are represented as P, Q, R, S, T.

4. The blood pressure assessment method based on the relational attention mechanism according to claim 3, wherein: The time characteristics and amplitude characteristics of the PPG signal key points and ECG signal key points in the time window are expressed by the following formulas: in, represents the average time feature in the i-th window, is a function that calculates the time difference between consecutive key points, is the timestamp of the key point in the i-th window; represents the average normalized amplitude feature within the i-th window, is a function that normalizes and averages the keypoint amplitudes, is the amplitude set of key points in the i-th window.

5. The blood pressure assessment method based on the relational attention mechanism according to claim 2, wherein: When collecting blood pressure-related signals in real time, the original PPG signal is obtained through the fingertip sensor, and the original ECG signal is recorded through the standard lead configuration. The blood pressure-related signals are sampled at a frequency of 125 Hz and 12-bit accuracy.

6. The blood pressure assessment method based on the relational attention mechanism according to claim 4, wherein: The graph neural network model includes a first relational attention module, a second relational attention module, an average pooling layer, a softmax layer, and a batch normalization layer. The activation function of the graph neural network model is a Relu function. There are two groups of batch normalization layers, which are respectively arranged after the first relational attention module and the second relational attention module.

7. The blood pressure assessment method based on the relational attention mechanism according to claim 6, wherein: The method for identifying and processing the topology of a blood pressure graph based on the spatiotemporal dependency of physiological signals using a graph neural network model includes: Load the blood pressure graph topology and input it into the graph neural network model. The first relational attention module aggregates information from neighboring nodes while considering edge attributes to update node features. The updated features are normalized to stabilize training, and a nonlinear activation function is used to enhance the expressive power of the graph neural network model. The second relational attention module further refines the representation of node features. The refined node features are obtained through the mean pooling operation of the average pooling layer. The pooling operation aggregates the node-level features into a single vector representation. The vector representation is loaded, and the vector representation is recognized and processed through the softmax layer to predict and generate the blood pressure assessment value.

Images