A multi-scale transformer non-invasive blood pressure prediction method and system based on an enhanced graph neural network

Abstract

The present application relates to the technical field of intelligent medical monitoring and physiological signal processing, in particular to a multi-scale Transformer non-invasive blood pressure prediction method and system based on enhanced graph neural network. After preprocessing and quality screening of the photoplethysmogram signal, local details and cross-cycle features are extracted using multi-scale parallel convolution; a long-term dependence is modeled by a Transformer encoder that fuses rotational position encoding, relative position bias and gated feedforward network; blood pressure-related features are strengthened using a channel enhancement module based on GELU activation; the systolic pressure, diastolic pressure and mean arterial pressure three prediction tasks are constructed as a task graph, and after multi-head graph attention interaction aggregation, the continuous arterial blood pressure waveform and the three blood pressure parameters are output, and an end-to-end training is performed with a composite loss function containing consistency constraints. The present application improves the accuracy, stability and physiological consistency of PPG non-invasive blood pressure prediction, and is suitable for wearable and bedside continuous monitoring.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical monitoring and physiological signal processing technology, specifically to a multi-scale Transformer non-invasive blood pressure prediction method and system based on enhanced graph neural networks. Background Technology

[0002] Hypertension is a major risk factor for cardiovascular disease. Long-term, persistently high blood pressure can gradually damage bodily organs. Real-time, accurate blood pressure monitoring provides doctors and patients with detailed information on blood pressure fluctuations, profoundly impacting patient prognosis and quality of life. Traditional cuff-based blood pressure measurement has significant limitations: intermittent monitoring leads to a lack of information on blood pressure fluctuations, repeated inflation causes discomfort, and the measurement scenarios are limited, failing to reflect true blood pressure levels during daily activities or sleep. In recent years, continuous blood pressure measurement methods based on photoplethysmography (PPG) signals have received widespread attention. This method captures changes in blood flow volume using optical sensors and, combined with wearable devices, enables all-weather, high-frequency blood pressure monitoring.

[0003] Currently, continuous blood pressure measurement methods based on PPG signals can be mainly divided into two categories: traditional machine learning and deep learning. Traditional machine learning methods extract pulse wave parameters from PPG waveform feature points and establish blood pressure prediction models through methods such as multiple linear regression, random forest, or support vector regression. However, traditional machine learning methods are essentially shallow learning models, and their performance is highly dependent on the quality of manually designed features. When faced with massive high-dimensional data, they have inherent limitations such as limited feature representation ability and insufficient generalization performance, making it difficult to achieve optimal function representation.

[0004] In recent years, deep learning has demonstrated significant advantages in blood pressure prediction based on PPG signals. Researchers have proposed deep recurrent neural network models based on Long Short-Term Memory (LSTM) networks, improving prediction accuracy through bidirectional structure and residual connections. Others have combined convolutional neural networks (CNNs) and LSTMs, using CNNs to extract the optimal feature set and LSTMs to capture temporal feature dependencies, achieving an A rating for both systolic and diastolic blood pressure according to the British Hypertension Society criteria. Still others have constructed multi-stage deep learning architectures, using parallel CNNs to extract morphological features and combining them with LSTMs to model temporal dynamic changes. However, RNN models often struggle to effectively model the complex relationship between blood pressure fluctuations and physiological characteristics when processing long sequences due to gradient vanishing and insufficient ability to capture temporal dependencies. The self-attention mechanism of Transformer models has attracted attention due to its global feature capture capability. Researchers have proposed temporal convolutional networks based on the SE attention mechanism, dynamically calibrating multi-channel information weights through channel attention. Others have proposed a bidirectional gated recurrent unit architecture with an attention mechanism, dynamically weighting key feature vectors through the attention mechanism.

[0005] Although deep learning has made significant progress in the field of blood pressure prediction based on PPG signals, existing methods still have the following limitations: (1) Existing models usually use single-scale convolutional kernels for feature extraction, which makes it difficult to capture the high-frequency components reflecting small vascular fluctuations and the low-frequency components related to heart rate changes in PPG signals at the same time; (2) The standard Transformer has limitations in terms of positional information expression and sensitivity to local features. The ReLU activation function in the feedforward network is prone to neuron "death" and lacks explicit modeling of local features; (3) Most existing methods treat the three prediction tasks of systolic blood pressure (SBP), diastolic blood pressure (DBP) and mean arterial pressure (MAP) independently, ignoring the potential physiological correlation and information complementarity among the three; (4) Existing channel attention mechanisms have limited ability to preserve small feature changes and suppress noise features, making it difficult to achieve precise enhancement of fine-grained features. Summary of the Invention

[0006] The purpose of this invention is to propose a multi-scale Transformer non-invasive blood pressure prediction method and system based on enhanced graph neural networks, which realizes joint modeling of local and global features of PPG signals, improves the information coordination ability among multiple prediction tasks, and thus improves the accuracy and stability of non-invasive continuous blood pressure prediction.

[0007] According to a first aspect of the present disclosure, a multi-scale Transformer noninvasive blood pressure prediction method based on an enhanced graph neural network is provided, comprising the following steps: The raw photoplethysmography (PPG) signal is acquired, and preprocessing and quality screening are performed on the raw PPG signal to obtain standardized PPG segments. The standardized photoplethysmography (PPG) fragments are input into the multi-scale feature extraction module, and multi-scale features are extracted using a group of convolution kernels containing multiple parallel one-dimensional convolution branches. Channel fusion is then performed to obtain a multi-scale feature map. The multi-scale feature map is input into the dynamic feature encoding module, and temporal feature encoding is performed using a Transformer encoder that includes rotational position encoding, relative position offset and gated feedforward network to obtain encoded features; The encoded features are input into the feature perception enhancement module, and channel weights are generated through the channel attention mechanism and the encoded features are weighted to obtain the enhanced blood pressure-related features. The enhanced blood pressure-related features are input into the multi-task feature aggregation module to construct a task graph consisting of systolic blood pressure, diastolic blood pressure, and mean arterial pressure prediction tasks. The multi-head graph attention mechanism is used to perform feature interaction between tasks, and the final blood pressure prediction value is output. During the training phase, end-to-end joint training is performed based on a composite loss function.

[0008] In one embodiment, the multi-scale feature extraction module uses one-dimensional convolution kernels of different sizes to extract local rising edges, peak-to-trough transition regions, pulse cycle morphology, and low-frequency blood flow change features, and performs normalization, nonlinear activation, and channel fusion on the output of each convolution branch to obtain a multi-scale feature map.

[0009] In one embodiment, the dynamic feature encoding module is implemented as follows: The multi-scale feature map is transformed from channel-first representation to time-series-first representation, and query vector, key vector and value vector are generated respectively; Rotational position encoding is applied to the query vector and key vector to obtain query vector and key vector that fuse absolute position information. A relative position bias matrix is ​​added to the attention score so that the attention weight simultaneously represents the numerical correlation and relative temporal distance of sampling points within the photoplethysmography segment. A gated feedforward network is set after multi-head self-attention. The gated feedforward network splits the linear transformation output into a gated branch and an information branch. Key features are filtered through Sigmoid gating, Swish activation and a second linear mapping. Grouped one-dimensional convolutions are set after the gated feedforward network to supplement local waveform pattern modeling, and gradient decay and overfitting are mitigated by residual connections and path dropping mechanisms to obtain encoded features. The above-mentioned method for acquiring multi-head self-attention is as follows: in, , These are the query vector and key vector, respectively, which incorporate location information. Let be the dimension of the key vector. It is a learnable bias matrix that relates to the relative positions of elements in the sequence. It is a value vector; The above gated feedforward network is represented as: in, These are the input features of the gated feedforward network. It is an intermediate feature tensor. , To transform linear The output is divided into two tensors along the feature dimension. This represents the Sigmoid function.

[0010] In one embodiment, the feature perception enhancement module is implemented as follows: Global average pooling is performed on the encoded features along the time dimension to obtain the global description vector corresponding to each channel; The global description vector is input into the first fully connected layer and the Gaussian error linear unit activation function to obtain compressed features that include nonlinear channel correlations; The compressed features are input into the second fully connected layer and the Sigmoid function to obtain channel weights with values ​​ranging from 0 to 1; The channel weights are multiplied channel by channel with the encoded features to enhance the feature response related to blood pressure changes and suppress irrelevant features caused by motion artifacts, sensor noise, or individual differences.

[0011] In one embodiment, the multi-task feature aggregation module is implemented as follows: Adaptive average pooling is performed on the enhanced blood pressure-related features to obtain sample-level global features, and continuous arterial blood pressure prediction waveforms are generated through the arterial blood pressure waveform prediction branch. Three initial blood pressure parameters are generated through systolic blood pressure prediction branch, diastolic blood pressure prediction branch and mean arterial pressure prediction branch respectively, and the three initial blood pressure parameters are used as the initial node features of three nodes in the task graph. The initial node features of each node are projected into a high-dimensional feature space through a linear mapping, and the attention score between the source node and the target node is calculated in multiple attention heads. The normalized attention weights are obtained through the Softmax function. : in, , Representing nodes respectively and nodes High-dimensional representation; , They represent the first Learnable weight parameters for the source and target nodes in an attention head; Indicates the first In each attention head, the node For nodes Attention score; Represents a node In the Attention to nodes Attention weights; The features of adjacent task nodes are weighted and aggregated according to the normalized attention weights. The aggregation results of multiple attention heads are then concatenated and input into a fully connected layer to obtain the final predicted values ​​of systolic blood pressure, diastolic blood pressure, and mean arterial pressure.

[0012] In one embodiment, the composite loss function is expressed as: L = λ1·MSE(Y, Y_hat) + λ2·MAE(SBP, SBP_hat) + λ3·MAE(DBP, DBP_hat)+ λ4·MAE(MAP, MAP_hat) + λ5·MAE(MAP_hat, DBP_hat + 2 / 3·(SBP_hat - DBP_hat)) Where Y is the actual ABP waveform, Y_hat is the predicted ABP waveform, SBP, DBP, and MAP are the actual systolic blood pressure, diastolic blood pressure, and mean arterial pressure, respectively, SBP_hat, DBP_hat, and MAP_hat are the corresponding predicted values, and λ1 to λ5 are preset weighting coefficients.

[0013] In one embodiment, a step of predicting confidence output is also included: A prediction confidence indicator is generated based on the quality screening results of photoplethysmography pulse wave segments, the consistency of output blood pressure parameters, and the degree of abnormality of continuous arterial blood pressure prediction waveforms. When the prediction confidence indicator is lower than the preset confidence condition, a re-acquisition prompt, a low confidence mark, or only blood pressure change trend information is output; when the prediction confidence indicator meets the preset confidence condition, the predicted waveforms of systolic blood pressure, diastolic blood pressure, mean arterial pressure, and continuous arterial blood pressure are output.

[0014] According to a second aspect of the present disclosure, a multi-scale Transformer noninvasive blood pressure prediction system based on an enhanced graph neural network is provided, comprising: The signal preprocessing module acquires the raw photoplethysmography (PPG) signal, performs preprocessing and quality screening on the raw PPG signal, and obtains standardized PPG segments. The multi-scale feature extraction module inputs the standardized photoplethysmography pulse wave segment into the multi-scale feature extraction module, extracts multi-scale features using a group of convolution kernels containing multiple parallel one-dimensional convolution branches, and performs channel fusion to obtain a multi-scale feature map. The dynamic feature encoding module inputs the multi-scale feature map into the dynamic feature encoding module, and uses a Transformer encoder containing rotational position encoding, relative position offset and gated feedforward network to perform temporal feature encoding to obtain encoded features; The feature perception enhancement module inputs the encoded features, generates channel weights through a channel attention mechanism, and weights the encoded features to obtain enhanced blood pressure-related features; The multi-task feature aggregation module inputs the enhanced blood pressure-related features into the multi-task feature aggregation module, constructs a task graph consisting of systolic blood pressure, diastolic blood pressure, and mean arterial pressure prediction tasks, and uses a multi-head graph attention mechanism to perform feature interaction between tasks, outputting the final blood pressure prediction value. The training module performs end-to-end joint training based on a composite loss function during the training phase.

[0015] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned multi-scale Transformer non-invasive blood pressure prediction method based on an enhanced graph neural network.

[0016] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned multi-scale Transformer noninvasive blood pressure prediction method based on an enhanced graph neural network.

[0017] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: (1) This invention sets up convolutional kernels of different sizes to simultaneously capture local details (rising edge, notch) and macroscopic change patterns (low-frequency blood flow changes across cycles) of the signal at the same level. After group normalization, GELU activation, and 1×1 convolution cross-scale fusion, a multi-scale feature map with rich information is formed. This mechanism effectively solves the technical defect that a single-scale convolutional kernel cannot simultaneously perceive high-frequency micro-fluctuations and low-frequency macroscopic trends in PPG signals.

[0018] (2) This invention introduces three improvements to the standard Transformer encoder: First, it employs rotational position encoding combined with a relative position bias matrix, enabling attention weights to simultaneously represent the numerical correlation and relative temporal distance of sampling points. Second, it replaces ReLU with gated linear units and the Swish activation function, alleviating the neuron "death" problem and improving feature selection capabilities. Third, it adds grouped one-dimensional convolutions after the gated feedforward network to supplement local waveform pattern modeling. These improvements enable the model to possess both global dependency modeling and local feature perception capabilities.

[0019] (3) After performing global average pooling on the encoded features, the present invention generates channel weights through a two-level fully connected mapping activated by GELU and applies them to the original features channel by channel. GELU has a non-zero gradient in the negative interval, which can effectively preserve the small-amplitude blood pressure-related fluctuation features in the PPG signal and overcome the defect that weak key features are ignored due to the hard thresholding characteristics of ReLU.

[0020] (4) The present invention constructs three blood pressure prediction tasks as three nodes in the task graph. Through the multi-head graph attention mechanism, feature interaction and weighted aggregation are performed between nodes, which breaks the limitation of existing methods that independently predict three parameters, resulting in the failure to fully utilize information complementarity. This enables the model to automatically learn the physiological correlation between tasks and improves the accuracy of multi-task collaborative prediction.

[0021] (5) The composite loss function introduces a MAP consistency constraint term in addition to waveform error and parameter error to ensure that the three blood pressure parameters meet the physiological correlation. The credibility assessment module generates credibility identifiers based on signal quality and output consistency. When the credibility is low, it outputs prompts or trend information instead of specific values, thus avoiding false alarms caused by low-quality input. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0023] Figure 1 This is a schematic diagram of the overall structure of the MTEG network of the present invention; Figure 2 This is a schematic diagram of the multi-scale feature extraction module structure of the present invention; Figure 3 This is a schematic diagram of the dynamic feature encoding module structure of the present invention; Figure 4 This is a schematic diagram of the SE-Block feature perception enhancement module structure of the present invention; Figure 5 This is a schematic diagram of the Bland-Altman analysis of the present invention; Figure 6 This is a schematic diagram of the Pearson correlation analysis of the present invention; Figure 7 This is a schematic diagram of the arterial blood pressure waveform predicted by MTEG in this invention. Detailed Implementation

[0024] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0028] Example 1: like Figure 1As shown, this embodiment provides a multi-scale Transformer non-invasive blood pressure prediction method based on enhanced graph neural networks, including the following steps: S1. Acquire the raw photoplethysmography (PPG) signal, perform preprocessing and quality screening on the raw PPG signal to obtain standardized PPG segments. In this embodiment, the original photoplethysmography (PPG) signal can originate from a smartwatch, a clip-on optical sensor, a wristband-type health monitoring device, a patch sensor, or a bedside monitoring device. To meet the model processing requirements, each segment of the original PPG signal is converted into a time-series fragment of uniform length. When the sampling rates of different devices are inconsistent or there are missing sampling points, the signal can be resampled to a uniform sampling rate before generating a fixed-length input through interpolation, truncation, or zero-padding. The above processing does not change the trend of blood flow volume change reflected by the PPG, aiming to ensure that the subsequent neural network structure can operate stably under a uniform input size.

[0029] Quality screening is not for disease diagnosis, but rather to determine whether a current signal segment is suitable as model input. The quality screening criteria include, but are not limited to: prolonged saturation of continuous sample values ​​within the segment, significant lack of peak-valley structure, pre-normalized amplitude exceeding the device's measurement range, or abrupt changes in adjacent sampling points exceeding a preset threshold. If a segment does not meet these quality conditions, it is discarded or marked as a low-confidence segment to avoid noise interference causing drastic fluctuations in blood pressure prediction results.

[0030] S2. Input the standardized photoplethysmography (PPG) segment into the multi-scale feature extraction module, extract multi-scale features using a group of convolutional kernels containing multiple parallel one-dimensional convolutional branches, and perform channel fusion to obtain a multi-scale feature map, such as... Figure 2 As shown; S21. Construct multi-scale convolutional branch groups for the input standardized photoplethysmography pulse wave segments and perform parallel convolution operations.

[0031] To fully characterize the multi-scale physiological information contained in PPG signals, a basic data structure is first defined. The input PPG signal is represented as follows: ,in The number of signals. Represents the time series length of each signal record. The number of channels is represented (usually 1). Since each PPG signal contains rich multi-scale physiological information, encompassing features at different time scales and frequencies, it is crucial for accurate blood pressure prediction. For example, high-frequency components in a PPG signal may reflect subtle vascular fluctuations, while low-frequency components may be related to heart rate changes. Therefore, multi-scale feature extraction is a key step in blood pressure prediction.

[0032] This invention constructs a multi-scale convolutional branch group, employing convolutional kernels of different sizes (3, 5, 7, and 9) to perform parallel convolution operations on PPG signals to capture features at different time scales. Smaller kernels (3×3) capture local details in the PPG signal, such as rapid rises in systolic blood pressure or brief drops in diastolic blood pressure; medium-sized kernels (5×5 and 7×7) capture medium-scale features, such as the overall morphology of the pulse wave and the transition regions between peaks and troughs; larger kernels (9×9) can learn the overall state of the cardiovascular system and macroscopic changes in blood pressure. This multi-scale parallel design ensures that the model can simultaneously perceive subtle changes and overall trends in the signal.

[0033] The output of each convolutional branch is represented as: in, For the first The output of each branch For the first The sizes of the convolutional kernels (3, 5, 7, and 9, respectively) The number of channels output for each branch. The design of four parallel branches enables the model to process information at different scales simultaneously at the same level, avoiding the problem that traditional single-scale convolutions require stacking multiple layers to obtain multi-scale information, thereby improving computational efficiency and reducing the number of parameters.

[0034] S22. Perform fusion and activation processing on multi-branch features.

[0035] After obtaining the outputs of the four parallel convolutional branches, this invention processes the output of each branch using GroupNorm and Gaussian Error Linear Unit (GELU) activation functions to ensure the stability and nonlinear expressive power of the features. Then, a 1×1 convolution is used to concatenate and fuse the multi-scale features of the four branches along the channel dimension, thereby generating a multi-scale feature map: in, It is the fused multi-scale feature map. This represents the number of channels after merging.

[0036] Through the multi-scale feature extraction process described above, this invention can extract features at different time scales and frequencies from PPG signals. Different receptive fields correspond to differentiated perceptions of local features and physiological information at different frequencies, thereby effectively capturing key physiological information in PPG signals and providing rich foundational features for subsequent dynamic feature encoding and blood pressure prediction.

[0037] S3. Input the multi-scale feature map into the dynamic feature encoding module, and use a Transformer encoder containing rotational position encoding, relative position offset, and gated feedforward network to perform temporal feature encoding to obtain the encoded features, such as... Figure 3 As shown; S31. Perform sequence transformation and positional encoding on multi-scale features.

[0038] The multi-scale feature map obtained after the multi-scale feature extraction module As input to the dynamic feature encoding process. To adapt to the input requirements of the Transformer, [the following is used]: Convert to sequence format: To enhance the expressive power of location information, this invention introduces Rotary Position Embedding (ROPE) and a relative position bias mechanism. Position information is encoded into the query vector through a rotation operation. and key vector This enhances the model's sensitivity to positional information, thereby helping the model understand the relative positional relationships of different elements in the sequence. S32. Perform self-attention computation based on rotational position encoding and relative position bias.

[0039] Attention is calculated on the position-encoded vector: in, , These are the query vector and key vector, respectively, which incorporate location information. Let be the dimension of the key vector. It is a learnable bias matrix that relates to the relative positions of elements in the sequence. It is a value vector.

[0040] By join in and In the dot product, the model can integrate relative position information into the calculation process of attention weights, so that the attention distribution not only depends on the numerical content of the sequence elements, but also reflects the relative positional relationship between the elements in the PPG signal, thereby improving the model's ability to understand the temporal structure.

[0041] S33, Perform gated feedforward network and convolution fusion processing.

[0042] Compared to the ReLU activation function, the Swish and GLU functions, while retaining nonlinear modeling capabilities, further optimize the continuity and stability of gradient propagation, especially demonstrating stronger training stability and feature representation capabilities in deep networks. Therefore, this invention improves the activation function of the Feed-Forward Network (FFN) module in the Transformer model by replacing ReLU with the GLU and Swish functions: here, For the input features of the FFN module, It is an intermediate feature tensor. , To transform linear The output is divided into two tensors along the feature dimension. This represents the Sigmoid function. This indicates element-wise multiplication.

[0043] The purpose of introducing GLU and Swish activation functions is to introduce gating mechanisms while maintaining nonlinear expressive power, allowing for more flexible control of information flow. Compared to ReLU, the Sigmoid function still has a non-zero derivative in the negative interval, which helps alleviate the gradient vanishing problem. GLU dynamically adjusts the importance of information between channels through gating mechanisms, while Swish provides smoother activation, which is beneficial for modeling complex periodic and nonlinear features in PPG signals, such as heart rate changes and pulse waveform structures.

[0044] In addition, to further extract local features from the PPG signal, this invention adds a convolutional fusion module (Conv Fusion) after the FFN: in, It is the output feature tensor of the FFN module and its layers after normalization. It involves grouping and performing one-dimensional convolution. This is the local feature tensor after convolution. Using grouped convolution can extract local features, enhancing the model's ability to capture local patterns (such as waveform shape and periodic changes) in PPG signals. Residual connections ensure effective information transfer, avoiding the gradient vanishing problem and thus improving the model's training stability.

[0045] Finally, to improve the model's generalization ability, this invention introduces a DropPath mechanism after the residual connection. This mechanism enhances the model's robustness by randomly discarding parts of the path, preventing overfitting during training and resulting in more stable performance on test data. in, It is the final output of the entire dynamic feature encoding module.

[0046] Through the above improvements, the dynamic feature encoding module of this invention can more effectively capture long-term dependencies in PPG signals and enhance the modeling ability of dynamic blood pressure changes. RoPE and relative position bias matrix enhance the modeling ability of positional information, GLU and Swish activation functions filter important features and improve nonlinear expression ability, convolutional fusion module further extracts local features, and DropPath regularization improves the generalization ability of the model.

[0047] S4. Input the encoded features into the feature perception enhancement module, generate channel weights through a channel attention mechanism, and weight the encoded features to obtain the enhanced blood pressure-related features, such as... Figure 4 As shown; S41. Perform global average pooling and channel compression on the encoded features.

[0048] After multi-scale feature extraction and dynamic feature encoding, considering that subtle fluctuations in the PPG signal may contain crucial blood pressure information, this invention replaces the original ReLU activation function in the SE-Block with GELU, and names the improved SE-Block the Feature Perception Enhancement Module. Compared to ReLU, GELU has a smoother activation response, which helps to better preserve subtle feature changes, thereby improving the channel attention mechanism's ability to perceive and express fine-grained information with greater accuracy.

[0049] Specifically, in the feature perception enhancement module, firstly... Perform global average pooling to compress the time dimension: in, It is the output of the dynamic feature encoding module. for The number of channels, pooling operations along the time series dimension This process yields the global feature representation for each channel. .

[0050] S42. Perform channel weight learning based on Gaussian error linear units.

[0051] Next, the correlation between channels is learned through the two fully connected layers: in, This is the channel compression ratio, used to control the dimensions of the intermediate layer. This is the compressed low-dimensional channel representation. The final importance weights for each channel are determined by GELU and Sigmoid, which are used to introduce nonlinear activation and generate channel weights, respectively.

[0052] Finally, the learned channel weights are applied to the original features to obtain the channel-weighted enhanced features: in, This indicates multiplication by channel.

[0053] This feature perception enhancement module enhances the feature channels related to blood pressure changes through a channel attention mechanism, while suppressing irrelevant noise features, thereby improving the model's ability to perceive key information about blood pressure changes.

[0054] S5. Input the enhanced blood pressure-related features into the multi-task feature aggregation module to construct a task graph consisting of systolic blood pressure, diastolic blood pressure, and mean arterial pressure prediction tasks. Then, utilize a multi-head graph attention mechanism to perform feature interaction between tasks and output the final predicted blood pressure value, such as... Figure 5 As shown; S51, Perform multi-task initial prediction and graph structure construction.

[0055] After completing the feature-aware enhancement, the features of each sample in the sequence dimension are first compressed to a global average value through adaptive average pooling, resulting in the feature representation of each sample: Subsequently, arterial blood pressure (ABP) signals and their corresponding systolic blood pressure (SBP), diastolic blood pressure (DBP), and mean arterial pressure (MAP) were predicted using a 1D convolutional layer and three independent fully connected layers, respectively: in, This is the feature representation after pooling. For the predicted ABP waveform, , and These are the initial predicted values ​​for SBP, DBP, and MAP, respectively.

[0056] Considering the potential physiological correlation among SBP, DBP, and MAP, this invention further designs a multi-task feature aggregation module to share information and collaboratively optimize predicted blood pressure values. Dynamic correlations are established among the three prediction tasks (SBP, DBP, and MAP) to promote information interaction between multiple tasks, thereby improving the model's predictive ability for different blood pressure indicators.

[0057] Specifically, the three prediction tasks—SBP, DBP, and MAP—are treated as three nodes in a graph, and a graph structure is constructed. The initial feature of each node is its corresponding predicted value: S52, Perform inter-task feature interaction and aggregation based on multi-head graph attention.

[0058] To adapt the feature representation to the graph structure, the node features are first projected into a high-dimensional space using a linear mapping: in, The mapped feature dimensions, For the first High-dimensional feature representation of each task .

[0059] Next, a multi-head attention mechanism is introduced to enhance feature interaction capabilities. For the first... Attention head, source node With the target node The attention score between them is: Then, the attention weights are obtained by normalizing using the Softmax function: in, , Representing nodes respectively and nodes High-dimensional representation; , They represent the first Learnable weight parameters for the source and target nodes in an attention head; Indicates the first In each attention head, the node For nodes Attention score; Represents a node In the Attention to nodes Attention weights.

[0060] Subsequently, a weighted aggregation is performed based on the attention weights to obtain the updated representation of each node: The outputs of multiple attention heads are concatenated to form the final feature representation: Finally, the aggregated features are mapped back to the predicted blood pressure value through a fully connected layer, yielding the final outputs of SBP, DBP, and MAP: in, , , These represent the final aggregate representations of the three task nodes: SBP, DBP, and MAP, respectively.

[0061] The multi-task feature aggregation module achieves information fusion and collaborative optimization among the three blood pressure prediction tasks by constructing a task graph structure and introducing an attention mechanism, thereby improving the model's prediction accuracy for different blood pressure indicators and enhancing the model's overall robustness.

[0062] S6. During the training phase, end-to-end joint training is performed based on the composite loss function.

[0063] This invention employs an end-to-end joint training strategy, designing a composite loss function to simultaneously optimize sequence prediction accuracy and blood pressure parameter accuracy. The composite loss function comprises the mean squared error between the predicted and actual sequences, systolic blood pressure error, diastolic blood pressure error, and mean arterial pressure error; wherein systolic and diastolic blood pressure are taken from the maximum and minimum values ​​of the predicted sequences, respectively, and mean arterial pressure is calculated based on the physiological relationship between systolic and diastolic blood pressure. This joint training strategy ensures that the prediction results both conform to the temporal variation pattern and possess clear physiological significance.

[0064] The model was trained using the AdamW optimizer with an initial learning rate of 1×10⁻⁶. -6 The weight decay coefficient is 1×10 -4The batch size is 64. The learning rate is scheduled using a cosine annealing strategy with an initial warm-up period of 10 rounds and a period doubling mechanism. When the validation loss fails to decrease for 5 consecutive rounds, the learning rate is reduced to one-quarter of its current value. An early stopping mechanism is also employed during training; training terminates when the validation loss fails to improve for 40 consecutive rounds.

[0065] like Figure 6-7 As shown, to further illustrate the technical effects of the present invention, the MTEG model described in this invention is compared with existing blood pressure prediction models on a dataset constructed based on the MIMIC-II waveform database. Evaluation metrics include Mean Absolute Error (MAE) and Standard Deviation (STD), where MAE measures the average deviation between the predicted value and the reference value, and STD measures the dispersion of the prediction error. The comparative experimental results are as follows: Table 1. Statistics of prediction results for different models on the MIMIC dataset, in mmHg. As shown in Table 1, on the MIMIC dataset, the MAE of the MTEG model for SBP, DBP, and MAP are 4.92 mmHg, 2.68 mmHg, and 2.59 mmHg, respectively. Compared with RDAE, ABPM, BP, PPG2ABP, and MSA-ResNet, this invention has lower overall prediction errors for SBP and DBP, and maintains a lower STD in MAP prediction. In particular, the STD of MTEG's SBP is 8.42 mmHg, lower than RDAE's 10.37 mmHg, PPG2ABP's 10.69 mmHg, and MSA-ResNet's 9.79 mmHg, indicating that it has better stability for different samples.

[0066] To verify the contributions of each key module in this invention, an ablation experiment was conducted: TEG represents replacing the multi-scale feature extraction module with a single-scale convolution module; MEG represents replacing the dynamic feature encoding module with the original Transformer encoder; MTE represents removing the multi-task feature aggregation module and retaining only the independent blood pressure prediction branch. The ablation experiment results are as follows: Table 2. Statistics of ablation experimental results from the MIMIC dataset, in mmHg. As shown in Table 2, the complete MTEG model outperforms the TEG, MEG, and MTE models after removing their respective modules in terms of MAE metrics for SBP, DBP, and MAP, respectively. This indicates that the multi-scale feature extraction module, dynamic feature encoding module, and multi-task feature aggregation module all contribute to the prediction accuracy. Meanwhile, the complete MTEG model also maintains a low level in STD metrics for SBP and DBP, suggesting that the collaboration of each module helps to improve prediction stability.

[0067] Example 2: This embodiment provides a multi-scale Transformer non-invasive blood pressure prediction system based on enhanced graph neural networks, including: The signal preprocessing module acquires the raw photoplethysmography (PPG) signal, performs preprocessing and quality screening on the raw PPG signal, and obtains standardized PPG segments. The multi-scale feature extraction module inputs the standardized photoplethysmography pulse wave segment into the multi-scale feature extraction module, extracts multi-scale features using a group of convolution kernels containing multiple parallel one-dimensional convolution branches, and performs channel fusion to obtain a multi-scale feature map. The dynamic feature encoding module inputs the multi-scale feature map into the dynamic feature encoding module, and uses a Transformer encoder containing rotational position encoding, relative position offset and gated feedforward network to perform temporal feature encoding to obtain encoded features; The feature perception enhancement module inputs the encoded features, generates channel weights through a channel attention mechanism, and weights the encoded features to obtain enhanced blood pressure-related features; The multi-task feature aggregation module inputs the enhanced blood pressure-related features into the multi-task feature aggregation module, constructs a task graph consisting of systolic blood pressure, diastolic blood pressure, and mean arterial pressure prediction tasks, and uses a multi-head graph attention mechanism to perform feature interaction between tasks, outputting the final blood pressure prediction value. The training module performs end-to-end joint training based on a composite loss function during the training phase.

[0068] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0069] Example 3: An electronic device is provided for running the aforementioned "a multi-scale Transformer non-invasive blood pressure prediction method based on an enhanced graph neural network". The electronic device includes a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S6 of the method described in Embodiment 1, specifically including but not limited to: S1. Acquire the raw photoplethysmography (PPG) signal, perform preprocessing and quality screening on the raw PPG signal to obtain standardized PPG segments. S2. Input the standardized photoplethysmography pulse wave segment into the multi-scale feature extraction module, extract multi-scale features using a group of convolution kernels containing multiple parallel one-dimensional convolution branches, and perform channel fusion to obtain a multi-scale feature map; S3. Input the multi-scale feature map into the dynamic feature encoding module, and use the Transformer encoder, which includes rotational position encoding, relative position offset and gated feedforward network, to perform temporal feature encoding to obtain the encoded features; S4. Input the encoded features into the feature perception enhancement module, generate channel weights through the channel attention mechanism, and weight the encoded features to obtain the enhanced blood pressure related features; S5. Input the enhanced blood pressure-related features into the multi-task feature aggregation module to construct a task graph consisting of systolic blood pressure, diastolic blood pressure, and mean arterial pressure prediction tasks, and use the multi-head graph attention mechanism to perform feature interaction between tasks, and output the final blood pressure prediction value. S6. During the training phase, end-to-end joint training is performed based on the composite loss function.

[0070] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage via a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0071] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S6 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0072] Application Example 1: Continuous Blood Pressure Monitoring Scenarios with Smart Wearable Devices This invention is applied to wearable devices such as smartwatches or health patches. In scenarios where these devices collect user PPG signals via optical sensors, a multi-scale feature extraction module is used to extract local features at different time scales from the PPG signal in parallel. Specifically, when a user wears a smartwatch, the watch's built-in optical heart rate sensor continuously emits green light to illuminate the skin and receives the reflected light signal, converting it into a PPG signal. This PPG signal is first preprocessed (including filtering, denoising, and normalization) and then input into the MTEG model of this invention. The multi-scale feature extraction module simultaneously captures local details and macroscopic variation patterns in the signal through four parallel convolutional branches (with kernel sizes of 3, 5, 7, and 9), generating a multi-scale feature map. The dynamic feature encoding module performs sequence transformation and positional encoding on the multi-scale feature map, capturing long-term dependencies in blood pressure changes through an improved Transformer encoder. The feature perception enhancement module strengthens the response to channels related to blood pressure changes and suppresses noise features. The multi-task feature aggregation module achieves information coordination among the three prediction tasks of systolic blood pressure, diastolic blood pressure, and mean arterial pressure, outputting the final predicted blood pressure value.

[0073] Compared to traditional single-scale feature extraction and independent task prediction methods, this application significantly improves the accuracy and stability of non-invasive blood pressure prediction. Experimental results show that this invention achieves BHS Grade A and AAMI standards for diastolic blood pressure and mean arterial pressure prediction, and BHS Grade B for systolic blood pressure prediction. This provides reliable technical support for long-term ambulatory blood pressure monitoring, enabling users to continuously monitor blood pressure changes in daily life and promptly detect abnormal blood pressure patterns.

[0074] Application Example 2: Assisted Diagnostic Scenarios in Clinical Intensive Care Units When this invention is applied to a continuous blood pressure monitoring system in a hospital intensive care unit (ICU), in scenarios where long-term, high-frequency blood pressure monitoring of patients is required, the system extracts PPG signals from multimodal physiological signals using the method of this invention, and then sequentially processes them through multi-scale feature extraction, dynamic feature encoding, feature perception enhancement, and multi-task feature aggregation to output continuous arterial blood pressure waveforms and predicted values ​​of SBP, DBP, and MAP.

[0075] Specific application scenarios include: for postoperative patients, close monitoring of blood pressure changes is necessary to prevent complications caused by hypotension or hypertension; for shock patients, real-time assessment of blood pressure recovery is needed to guide fluid resuscitation; for cardiovascular disease patients, monitoring of blood pressure fluctuations is required to assess changes in their condition and treatment effectiveness. In these scenarios, traditional cuff-based blood pressure measurement not only causes discomfort to patients but also fails to provide continuous dynamic blood pressure information. The non-invasive continuous blood pressure monitoring method of this invention can provide medical staff with continuous and accurate dynamic blood pressure trend information without increasing patient discomfort, assisting in the diagnosis of abnormal blood pressure patterns such as morning hypertension or nocturnal hypotension, helping to adjust treatment plans in a timely manner and improve patient prognosis.

[0076] In this application example, after integrating the MTEG model, the ICU monitoring system collects the patient's PPG signal through wearable sensors near the bed, outputs real-time blood pressure prediction results, and displays a blood pressure trend graph on the monitoring screen at the nurse's station. When abnormal blood pressure fluctuations are detected, the system automatically triggers an alarm to remind medical staff to pay attention. Experimental results show that this invention can provide dynamic blood pressure trends and parameter prediction results in continuous monitoring scenarios in the ICU, and can serve as an auxiliary signal processing solution for bedside monitoring systems.

[0077] Application Example 3: Remote Monitoring Scenarios for Family Health Management This invention is applied to a home health management platform. In daily health monitoring scenarios for the elderly or hypertensive patients, users collect PPG signals through home wearable devices (such as smart bracelets or health patches) and upload them to the cloud. The blood pressure prediction model of this invention, deployed in the cloud, processes and analyzes the signals, generates a blood pressure trend report, and pushes it to the user's mobile phone or a family member's terminal. This embodiment fully utilizes the physiological correlation between SBP, DBP, and MAP through a multi-task feature aggregation module, improving the model's generalization ability across different individuals, making remote blood pressure monitoring more accurate and reliable, and contributing to early warning and timely intervention for hypertension.

[0078] The specific application process is as follows: Users wear a smart bracelet every morning after waking up. The bracelet automatically collects PPG signals and calculates blood pressure values. Through continuous monitoring for a week, the system establishes a personal blood pressure baseline model for each user, identifying their normal blood pressure fluctuation range. When abnormally high blood pressure (such as systolic blood pressure consistently above 140 mmHg) or abnormally low blood pressure is detected, the system pushes alerts to the user and their family via a mobile application. For hypertensive patients, the system can also assist in assessing drug efficacy based on medication records and blood pressure trends, reminding users to take their medication on time.

[0079] In this application example, the family health management platform can also integrate other health data (such as heart rate, blood oxygen saturation, sleep quality, etc.) to form a comprehensive health record. Through the accumulation of long-term blood pressure monitoring data, the platform can identify the user's blood pressure change trends, predict cardiovascular disease risk, and provide personalized health management suggestions. Clinical trials have shown that the accuracy of blood pressure prediction in a home setting meets the needs of daily health management and provides an effective technical means for chronic disease management.

[0080] Application Example 4: Real-time Blood Pressure Monitoring in Sports and Fitness Scenarios This invention is applied to sports and fitness equipment, allowing users to monitor blood pressure changes in real time via a smartwatch or fitness tracker during aerobic exercises such as running and cycling. Blood pressure monitoring during exercise is crucial for assessing exercise intensity and preventing exercise-induced hypertension or hypotension. Traditional cuff blood pressure monitors cannot be used during exercise, while the non-invasive continuous blood pressure monitoring method of this invention can output blood pressure values ​​in real time, helping users understand the impact of exercise on their blood pressure.

[0081] In practical applications, the fitness tracker continuously collects PPG signals during user exercise and calculates blood pressure values ​​in real time using a built-in MTEG model. When an abnormally high blood pressure is detected (such as systolic blood pressure exceeding 200 mmHg), the tracker vibrates to remind the user to reduce exercise intensity; when low blood pressure is detected, the tracker reminds the user to rest appropriately. After exercise, the system generates a blood pressure change curve during the exercise period, helping users understand the impact of different exercise intensities on blood pressure, thereby enabling them to develop a more scientific exercise plan.

[0082] This application example also includes blood pressure monitoring in special environments such as high-altitude sports and diving. In these environments, blood pressure changes are more drastic, and real-time monitoring is crucial to ensuring user safety. The MTEG model of this invention, through multi-scale feature extraction and dynamic feature encoding, can effectively handle PPG signal changes caused by exercise, ensuring the accuracy of blood pressure prediction.

[0083] Application Example 5: Correlation Analysis of Fall Risk Warning and Blood Pressure in the Elderly This invention is applied to a home-based elderly care monitoring system, using a smart mattress or wearable device to continuously monitor blood pressure changes in the elderly, and combined with fall detection functionality to provide early warning of fall risks. Studies have shown that abnormal blood pressure fluctuations (such as orthostatic hypotension) are a significant contributing factor to falls in the elderly. By continuously monitoring blood pressure changes, risk factors can be identified before a fall occurs, allowing for preventative measures to be taken in advance.

[0084] In practical applications, the optical sensors built into the smart mattress continuously collect PPG signals during the elderly person's sleep, and calculate the nighttime blood pressure value using the MTEG model of this invention. The system analyzes nighttime blood pressure change patterns and identifies abnormal patterns such as nighttime hypotension and morning hypertension. When it detects a blood pressure abnormality that may lead to a fall (such as a sudden drop in blood pressure upon waking in the morning), the system reminds the elderly person to get up slowly via a mobile app or smart speaker to avoid falls caused by orthostatic hypotension.

[0085] This application example can also be combined with data from other sensors (such as accelerometers and gyroscopes) to automatically analyze blood pressure changes before and after a fall, providing healthcare professionals with information to analyze the causes of falls. The accumulation of long-term blood pressure monitoring data also helps assess the cardiovascular health of older adults, predict fall risk, and develop personalized prevention strategies.

[0086] Application Example 6: Monitoring of Hypertensive Disorders During Pregnancy This invention is applied to health monitoring of pregnant women, using wearable devices to continuously monitor changes in the pregnant woman's blood pressure, enabling early detection of gestational hypertension disorders (such as gestational hypertension and preeclampsia). Gestational hypertension disorders are one of the leading causes of maternal and perinatal morbidity and mortality; early identification and timely intervention are crucial for ensuring maternal and infant safety.

[0087] In practical applications, pregnant women wear smart bracelets or health patches. The devices periodically (e.g., every 15 minutes) collect PPG signals and calculate blood pressure values. The system establishes a personal blood pressure baseline model for the pregnant woman, identifying blood pressure variation patterns during normal pregnancy. When abnormally high blood pressure is detected (e.g., systolic blood pressure ≥140 mmHg or diastolic blood pressure ≥90 mmHg) or signs of preeclampsia such as proteinuria appear, the system provides timely warnings, reminding the pregnant woman to seek medical attention.

[0088] The advantage of this application is that it enables continuous and unobtrusive blood pressure monitoring, avoiding the discomfort caused to pregnant women by traditional cuff-type blood pressure measurements. It also allows for continuous monitoring in daily life, capturing blood pressure abnormalities that might be missed by intermittent measurements. Clinical trials have shown that the accuracy of this invention in blood pressure monitoring during pregnancy meets clinical needs, providing an effective technical means for the early screening of hypertensive disorders of pregnancy.

[0089] The experimental data described above are only used to illustrate the implementation effect of the technical solution described in this invention and the synergistic effect between the modules, and do not constitute a limitation on the scope of protection of the claims. Under different sampling rates, different segment lengths, different device forms, or different training datasets, those skilled in the art can make equivalent adjustments to the model parameters and training strategies according to the actual application conditions.

[0090] It should be noted that the above experimental parameters and evaluation data are used to illustrate the feasibility and technical effects of the present invention, and do not constitute a limitation on the scope of protection of the claims. Those skilled in the art can make equivalent adjustments to the input segment length, the number of convolutional kernels, the number of attention heads, the hidden dimension, the learning rate, or the confidence threshold without departing from the core technical concept of the present invention.

Claims

1. A multi-scale Transformer noninvasive blood pressure prediction method based on enhanced graph neural networks, characterized in that, Includes the following steps: The raw photoplethysmography (PPG) signal is acquired, and preprocessing and quality screening are performed on the raw PPG signal to obtain standardized PPG segments. The standardized photoplethysmography (PPG) fragments are input into the multi-scale feature extraction module, and multi-scale features are extracted using a group of convolution kernels containing multiple parallel one-dimensional convolution branches. Channel fusion is then performed to obtain a multi-scale feature map. The multi-scale feature map is input into the dynamic feature encoding module, and temporal feature encoding is performed using a Transformer encoder that includes rotational position encoding, relative position offset and gated feedforward network to obtain encoded features; The encoded features are input into the feature perception enhancement module, and channel weights are generated through the channel attention mechanism and the encoded features are weighted to obtain the enhanced blood pressure-related features. The enhanced blood pressure-related features are input into the multi-task feature aggregation module to construct a task graph consisting of systolic blood pressure, diastolic blood pressure, and mean arterial pressure prediction tasks. The multi-head graph attention mechanism is used to perform feature interaction between tasks, and the final blood pressure prediction value is output. During the training phase, end-to-end joint training is performed based on a composite loss function.

2. The multi-scale Transformer non-invasive blood pressure prediction method based on enhanced graph neural networks according to claim 1, characterized in that, The multi-scale feature extraction module uses one-dimensional convolution kernels of different sizes to extract local rising edges, peak-to-trough transition regions, pulse cycle morphology, and low-frequency blood flow change features, and normalizes, nonlinearly activates, and fuses the outputs of each convolution branch to obtain a multi-scale feature map.

3. The multi-scale Transformer noninvasive blood pressure prediction method based on enhanced graph neural networks according to claim 1, characterized in that, The dynamic feature encoding module is implemented as follows: The multi-scale feature map is transformed from channel-first representation to time-series-first representation, and query vector, key vector and value vector are generated respectively; Rotational position encoding is applied to the query vector and key vector to obtain query vector and key vector that fuse absolute position information. A relative position bias matrix is ​​added to the attention score so that the attention weight simultaneously represents the numerical correlation and relative temporal distance of sampling points within the photoplethysmography segment. A gated feedforward network is set after multi-head self-attention. The gated feedforward network splits the linear transformation output into a gated branch and an information branch. Key features are filtered through Sigmoid gating, Swish activation and a second linear mapping. Grouped one-dimensional convolutions are set after the gated feedforward network to supplement local waveform pattern modeling, and gradient decay and overfitting are mitigated by residual connections and path dropping mechanisms to obtain encoded features. The above-mentioned method for acquiring multi-head self-attention is as follows: in, , These are the query vector and key vector, respectively, which incorporate location information. Let be the dimension of the key vector. It is a learnable bias matrix that relates to the relative positions of elements in the sequence. It is a value vector; The above gated feedforward network is represented as: in, These are the input features of the gated feedforward network. It is an intermediate feature tensor. , To transform linear The output is divided into two tensors along the feature dimension. This represents the Sigmoid function.

4. The multi-scale Transformer non-invasive blood pressure prediction method based on augmented graph neural networks according to claim 1, characterized in that, The feature perception enhancement module is implemented as follows: Global average pooling is performed on the encoded features along the time dimension to obtain the global description vector corresponding to each channel; The global description vector is input into the first fully connected layer and the Gaussian error linear unit activation function to obtain compressed features that include nonlinear channel correlations; The compressed features are input into the second fully connected layer and the Sigmoid function to obtain channel weights with values ​​ranging from 0 to 1; The channel weights are multiplied channel by channel with the encoded features to enhance the feature response related to blood pressure changes and suppress irrelevant features caused by motion artifacts, sensor noise, or individual differences.

5. The multi-scale Transformer noninvasive blood pressure prediction method based on enhanced graph neural networks according to claim 1, characterized in that, The multi-task feature aggregation module is implemented as follows: Adaptive average pooling is performed on the enhanced blood pressure-related features to obtain sample-level global features, and continuous arterial blood pressure prediction waveforms are generated through the arterial blood pressure waveform prediction branch. Three initial blood pressure parameters are generated through systolic blood pressure prediction branch, diastolic blood pressure prediction branch and mean arterial pressure prediction branch respectively, and the three initial blood pressure parameters are used as the initial node features of three nodes in the task graph. The initial node features of each node are projected into a high-dimensional feature space through a linear mapping, and the attention score between the source node and the target node is calculated in multiple attention heads. The normalized attention weights are obtained through the Softmax function. : in, , Representing nodes respectively and nodes High-dimensional representation; , They represent the first Learnable weight parameters for the source and target nodes in an attention head; Indicates the first In each attention head, the node For nodes Attention score; Represents a node In the Attention to nodes Attention weights; The features of adjacent task nodes are weighted and aggregated according to the normalized attention weights. The aggregation results of multiple attention heads are then concatenated and input into a fully connected layer to obtain the final predicted values ​​of systolic blood pressure, diastolic blood pressure, and mean arterial pressure.

6. The multi-scale Transformer noninvasive blood pressure prediction method based on enhanced graph neural networks according to claim 1, characterized in that, The composite loss function is expressed as follows: L = λ1·MSE(Y, Y_hat) + λ2·MAE(SBP, SBP_hat) + λ3·MAE(DBP, DBP_hat) + λ4·MAE(MAP, MAP_hat) + λ5·MAE(MAP_hat, DBP_hat + 2 / 3·(SBP_hat - DBP_hat)) Where Y is the actual ABP waveform, Y_hat is the predicted ABP waveform, SBP, DBP, and MAP are the actual systolic blood pressure, diastolic blood pressure, and mean arterial pressure, respectively, SBP_hat, DBP_hat, and MAP_hat are the corresponding predicted values, and λ1 to λ5 are preset weighting coefficients.

7. The multi-scale Transformer noninvasive blood pressure prediction method based on enhanced graph neural networks according to claim 1, characterized in that, It also includes the step of predicting the confidence output: A prediction confidence indicator is generated based on the quality screening results of photoplethysmography pulse wave segments, the consistency of output blood pressure parameters, and the degree of abnormality of continuous arterial blood pressure prediction waveforms. When the prediction confidence indicator is lower than the preset confidence condition, a re-acquisition prompt, a low confidence mark, or only blood pressure change trend information is output; when the prediction confidence indicator meets the preset confidence condition, the predicted waveforms of systolic blood pressure, diastolic blood pressure, mean arterial pressure, and continuous arterial blood pressure are output.

8. A multi-scale Transformer non-invasive blood pressure prediction system based on an enhanced graph neural network, characterized in that, include: The signal preprocessing module acquires the raw photoplethysmography (PPG) signal, performs preprocessing and quality screening on the raw PPG signal, and obtains standardized PPG segments. The multi-scale feature extraction module inputs the standardized photoplethysmography pulse wave segment into the multi-scale feature extraction module, extracts multi-scale features using a group of convolution kernels containing multiple parallel one-dimensional convolution branches, and performs channel fusion to obtain a multi-scale feature map. The dynamic feature encoding module inputs the multi-scale feature map into the dynamic feature encoding module, and uses a Transformer encoder containing rotational position encoding, relative position offset and gated feedforward network to perform temporal feature encoding to obtain encoded features; The feature perception enhancement module inputs the encoded features, generates channel weights through a channel attention mechanism, and weights the encoded features to obtain enhanced blood pressure-related features; The multi-task feature aggregation module inputs the enhanced blood pressure-related features into the multi-task feature aggregation module, constructs a task graph consisting of systolic blood pressure, diastolic blood pressure, and mean arterial pressure prediction tasks, and uses a multi-head graph attention mechanism to perform feature interaction between tasks, outputting the final blood pressure prediction value. The training module performs end-to-end joint training based on a composite loss function during the training phase.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the multi-scale Transformer non-invasive blood pressure prediction method based on any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the multi-scale Transformer non-invasive blood pressure prediction method based on an enhanced graph neural network as described in any one of claims 1-7.

Images