A hypertension screening system, training method, hypertension screening method and electronic device

Abstract

The application discloses a hypertension screening system, a training method, a hypertension screening method and an electronic device, and belongs to the technical field of medical detection. First, the input BCG signal segment is decomposed into signal components under multiple frequency bands to comprehensively capture multi-band information related to hypertension. On this basis, considering that the relative relationship between different peaks in a single heartbeat of a hypertension patient and the rhythm and mode between different heartbeats are significantly different from those of a non-hypertension patient, ordinary convolution is used to capture local complex features in a single heartbeat, at the same time, dilated convolution is used to extract long-distance context information between heartbeats, and a cross-attention mechanism is introduced to realize the interaction between local features and long-distance context information, so that the key features related to hypertension diagnosis can be more accurately focused. Finally, by fusing the features of different frequency bands, accurate and complete feature expression is realized, and the accuracy of hypertension screening is improved.

Description

Technical Field

[0001] This invention belongs to the field of medical testing technology, and more specifically, relates to a hypertension screening system, training method, hypertension screening method, and electronic equipment. Background Technology

[0002] In recent years, early screening and detection of hypertension have played a crucial role in preventing cardiovascular disease. However, existing hypertension monitoring technologies face numerous challenges, including user comfort, detection accuracy, and the versatility of the technology. Currently used blood pressure monitoring methods include traditional mercury sphygmomanometers and electronic sphygmomanometers, but these methods require the user to remain still and measure blood pressure via cuff application. This can be quite uncomfortable for users who require frequent monitoring, especially those who need nighttime monitoring. Furthermore, while existing continuous blood pressure monitoring technologies (such as photoplethysmography (PPG) and shock echocardiography (SCG)) can achieve non-invasive monitoring to some extent, they still require direct contact with the body, making it difficult to meet the need for completely painless monitoring.

[0003] BCG (Ballistocardiogram) is a technique that indirectly reflects cardiovascular activity by measuring the minute movements caused by the heartbeat. This technology acquires signals by embedding sensors in mattresses, chairs, or other everyday items, providing a non-contact way to obtain physiological information related to cardiovascular activity and offering a new approach for non-invasive hypertension screening. However, due to the complex composition and diverse waveforms of BCG signals, and their susceptibility to motion artifacts and external noise, BCG-based hypertension detection technologies face challenges in practical applications, including low signal quality and insufficient detection accuracy.

[0004] Currently, existing hypertension screening technologies based on BCG signals mainly employ feature engineering combined with machine learning to explore the correlation between blood pressure and logarithmic energy, kurtosis, and other morphological statistical features in the BCG waveform from a morphological perspective. However, these methods have certain limitations: firstly, during signal acquisition, morphological features are easily interfered with by other frequency components, leading to insufficient robustness in feature extraction; secondly, these methods often ignore the potential connection between high-frequency components in the signal and blood pressure, making it difficult to comprehensively reflect the complex characteristics of blood pressure changes. Therefore, the accuracy of existing hypertension screening technologies based on BCG signals needs further improvement. Summary of the Invention

[0005] In view of the above-mentioned defects or improvement needs of the prior art, the present invention provides a hypertension screening system, training method, hypertension screening method and electronic device to solve the technical problem of low accuracy of hypertension screening in the prior art.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a hypertension screening system, comprising: a signal decomposition module, N feature extraction modules, a first fusion module, and a classifier; N≥2;

[0007] The signal decomposition module is used to decompose the input BCG signal segment into N signal components x1, x2, ..., xn at different frequency bands. N And input them one by one into N feature extraction modules;

[0008] The i-th feature extraction module includes:

[0009] A standard convolutional module is used to extract x. i Local heartbeat features x′ i，s ;

[0010] Dilated convolution module, used to extract x i Contextual features x′ i，b ;

[0011] The feature enhancement module is used to compute x′ based on the cross-attention mechanism. i，s Relative to x′ i，b degree of correlation r i,sb and x′ i，b Relative to x′ i，s degree of correlation r i，bs ; will r i,sb Superimposed on x′ i，b The local heartbeat enhancement feature x is obtained above i，s ; will r i，bs Superimposed on x′ i，s The context-enhanced feature x is obtained from the above. i,b ;

[0012] The second fusion module is used to process x separately based on the channel attention mechanism. i，s and x i，b The channel features in the data are dynamically adjusted, and the dynamically adjusted x is then processed. i，s and x i，b By performing fusion, the joint feature z of the i-th frequency band is obtained. i ;

[0013] The first fusion module is used to determine the weighting coefficients of each frequency band based on the energy intensity of its signal components and to weight the corresponding joint features to obtain the weighted features; the weighted features of each frequency band are then fused to obtain the fused features.

[0014] The classifier is used to obtain the corresponding hypertension screening results based on the fused features;

[0015] Where i = 1, 2, ..., N.

[0016] More preferably, the feature enhancement module includes convolutional units and cross-attention units;

[0017] The feature enhancement module calculates the relevance r. i,sb When x′ i，s The features obtained after the convolution unit are used as the Q-value and K-value, respectively, x′ i，b The V value is input into the cross-attention unit to obtain the relevance r. i,sb ;

[0018] The feature enhancement module calculates the relevance r. i，bs When x′ i，b The features obtained after the convolution unit are used as the Q-value and K-value, respectively, x′ is The V value is input into the cross-attention unit to obtain the relevance r. i，bs .

[0019] More preferably, the convolutional unit is a depthwise separable convolutional unit.

[0020] More preferably, the first fusion module includes: a feature extraction unit;

[0021] When determining the weighting coefficients for each frequency band, the first fusion module uses a feature extraction unit to extract the energy intensity features f of the signal components in each frequency band. e，1 f e，2 , ..., f e，N and for f e，1 f e，2 , ..., f e，N After normalization, the weighting coefficients σ1, σ2, ..., σ of each frequency band are obtained. N .

[0022] More preferably, the signal decomposition module is used to decompose the input BCG signal segment x into N signal components x1, x2, ..., xn at different frequency bands using convolution kernels of different sizes. N Specifically, it includes:

[0023] To minimize Taking the difference between x and y as the target, the parameters of each convolution kernel are adjusted; after adjustment, the signal component x in the i-th frequency band is obtained. i =Conv i (x);

[0024] Among them, Conv i (·) represents the convolution operation corresponding to the i-th convolution kernel.

[0025] More preferably, the above-mentioned BCG signal segment is obtained by sequentially filtering and uniformly dividing the original acquired BCG signal.

[0026] More preferably, the length of the BCG signal segment is 5 to 30 seconds.

[0027] Secondly, the present invention provides a training method for the above-mentioned hypertension screening system, comprising:

[0028] The BCG signal segments from the pre-collected training set are input into the hypertension screening system to obtain the corresponding hypertension screening results; based on the hypertension screening results of each BCG signal segment in the training set and the corresponding label indicating whether hypertension exists, the classification loss is calculated.

[0029] A training loss including classification loss is constructed; the hypertension screening system is trained based on the training loss.

[0030] More preferably, the training loss also includes clustering loss;

[0031] The expression for clustering loss is:

[0032]

[0033] The training set includes two types of BCG signal segments labeled as those with and without hypertension; N k f is the number of BCG signal segments of the k-th class in the training set; k,c To input the c-th BCG signal fragment of the k-th category in the training set into the hypertension screening system via the fusion feature obtained by the first fusion module; Centr k The cluster center for the k-th category is obtained by averaging the fusion features of each BCG signal segment under the k-th category in the training set.

[0034] Thirdly, the present invention provides a method for hypertension screening, comprising: inputting a BCG signal fragment to be detected into the hypertension screening system provided in the first aspect of the present invention, and obtaining a corresponding hypertension screening result.

[0035] Fourthly, the present invention provides an electronic device comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the method provided in the second or third aspect of the present invention.

[0036] Fifthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls the device in which the storage medium is located to perform the method provided in the second or third aspect of the present invention.

[0037] In a sixth aspect, the invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method provided in the second or third aspect of the invention.

[0038] In summary, the above-described technical solutions conceived in this invention can achieve the following beneficial effects:

[0039] 1. This invention provides a hypertension screening system. First, the input BCG signal segment is decomposed into signal components across multiple frequency bands to comprehensively capture multi-frequency information related to hypertension, avoiding the problem of high-frequency signals being ignored due to their low amplitude in traditional methods, thus improving the completeness of feature representation. Based on this, considering the significant differences in the relative relationships between different peaks in a single heartbeat and the rhythm and patterns between different heartbeats between hypertensive patients and non-hypertensive patients, a dual-branch parallel feature extraction network is designed. Ordinary convolution is used to capture complex local features within a single heartbeat, while dilated convolution is used to extract long-distance contextual information between heartbeats. A cross-attention mechanism is introduced to achieve interaction between local features and long-distance contextual information, thereby more accurately focusing on key features related to hypertension diagnosis. Finally, by fusing features from different frequency bands, accurate and complete feature representation is achieved, improving the accuracy of hypertension screening.

[0040] 2. Furthermore, in the hypertension screening system provided by this invention, the feature enhancement module includes a convolutional unit and a cross-attention unit; the cross-attention unit captures the correlation between local heartbeat features and global context features, and on this basis, the convolutional unit extracts complementary information between channels to strengthen the interaction between local heartbeat features and context features, thereby achieving more accurate feature expression and further improving the accuracy of hypertension screening.

[0041] 3. Furthermore, in the hypertension screening system provided by this invention, the convolution unit in the feature enhancement module is a depth-separated convolution unit, which performs convolution on each input channel separately without cross-channel calculation, significantly reducing the number of parameters and computational load, and improving the efficiency of hypertension screening.

[0042] 4. Furthermore, in the hypertension screening system provided by this invention, the signal decomposition module is used to decompose the input BCG signal segment into N signal components in different frequency bands using convolution kernels of different sizes. It can adaptively adjust the range and accuracy of frequency band division, flexibly adapt to the spectral characteristics of different BCG signals, and improve the accuracy of subsequent analysis.

[0043] 5. Furthermore, in the hypertension screening system provided by this invention, the original BCG signal is sequentially filtered and uniformly divided. The filtering process removes motion artifacts and signals from when the patient deviates from the sensor, thereby obtaining a high-quality signal input and further improving the accuracy of hypertension screening.

[0044] 6. Furthermore, considering that existing technologies rely on long signal acquisition times (e.g., 30 seconds to 5 minutes) when performing hypertension screening, which places high demands on users to maintain a stable posture in practical applications, the hypertension screening system provided by this invention, through the design of the above-mentioned signal decomposition module, feature extraction module and first fusion module, sets the length of the input BCG signal segment to 5 to 30 seconds, which can also acquire rich information, avoids the above-mentioned problems of existing technologies, and significantly improves detection efficiency and user experience.

[0045] 7. This invention provides a training method for a hypertension screening system. Based on the design of the hypertension screening system described above, the hypertension screening system can be trained using a pre-collected training set to obtain a hypertension screening system with high accuracy.

[0046] 8. Furthermore, the training method of the hypertension screening system provided by the present invention introduces classification loss and clustering loss. The classification loss constrains the accuracy of the classification results, while the clustering loss further improves the model's discriminative ability by enhancing feature clustering and category separation. This solves the learning bias problem caused by individual specificity, enabling the present invention to meet the needs of hypertension screening for unknown users in open set scenarios without relying on closed set scenarios (i.e., it does not require the test subjects to participate in the model training phase), thus enhancing the generalization ability in open scenarios. Attached Figure Description

[0047] Figure 1 The graphs show the signals of hypertension and non-hypertension, along with their corresponding frequency component analysis. (a) is the original signal graph of hypertension; (b) is the frequency component analysis graph of hypertension; (c) is the original signal graph of non-hypertension; and (d) is the frequency component analysis graph of non-hypertension.

[0048] Figure 2 This is a schematic diagram of the results of a hypertension screening system provided in an embodiment of the present invention;

[0049] Figure 3 The BCG signal segment and its signal decomposition diagram provided in the embodiments of the present invention are shown below; wherein, (a) is the original BCG signal segment diagram; (b) is the low-frequency component diagram; (c) is the mid-frequency component diagram; and (d) is the high-frequency component diagram.

[0050] Figure 4The above are t-SNE visualizations of the clustering results of the latent feature space generated by the hypertension screening system without and with clustering loss, provided in the embodiments of the present invention; wherein, (a) is the clustering result of the latent feature space generated by the hypertension screening system without clustering loss; and (b) is the clustering result of the latent feature space generated by the hypertension screening system with clustering loss. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0052] To achieve the above objectives, in a first aspect, the present invention provides a hypertension screening system. For example... Figure 1 The diagrams show the signals of hypertension and non-hypertension, along with their corresponding frequency component analysis. (a) is the original signal diagram of hypertension; (b) is the frequency component analysis diagram of hypertension; (c) is the original signal diagram of non-hypertension; and (d) is the frequency component analysis diagram of non-hypertension.

[0053] This invention takes into account that persistent hypertension can lead to increased arterial wall stiffness, which in turn triggers more high-frequency simple harmonic vibrations. First, the BCG signal segment is decomposed into multiple frequency bands, and the features of each frequency band are extracted. This design effectively avoids the problem of high-frequency band signals being ignored due to their low amplitude in traditional methods, thereby enabling a more comprehensive capture of multi-frequency band information related to hypertension and improving the completeness of feature expression.

[0054] Furthermore, significant differences were observed between hypertensive patients and non-hypertensive patients in the relative relationships between different peaks within a single heartbeat, as well as the rhythm and pattern between different heartbeats. To address this, this invention designs a dual-branch parallel feature extraction network: on one hand, ordinary convolution is used to capture complex local features within a single heartbeat; on the other hand, dilated convolution is used to extract long-distance contextual information between heartbeats. Based on this, a cross-attention mechanism is introduced to achieve interaction between local features and long-distance contextual information, thereby more accurately focusing on key features related to hypertension diagnosis and achieving accurate hypertension screening.

[0055] Specifically, the hypertension screening system provided by the present invention includes: a signal decomposition module, N feature extraction modules, a first fusion module, and a classifier; N≥2;

[0056] The signal decomposition module is used to decompose the input BCG signal segment into N signal components x1, x2, ..., xn at different frequency bands.N And input them one by one into N feature extraction modules;

[0057] The i-th feature extraction module (i = 1, 2, ..., N) includes:

[0058] A standard convolutional module is used to extract x. i Local heartbeat features x′ i，s ;

[0059] Dilated convolution module, used to extract x i Contextual features x′ i，b ;

[0060] The feature enhancement module is used to compute x′ based on the cross-attention mechanism. i，s Relative to x′ i，b degree of correlation r i,sb and x′ i，b Relative to x′ i，s degree of correlation r i，bs ; will r i,sb Superimposed on x′ i，b The local heartbeat enhancement feature x is obtained above i，s ; will r i，bs Superimposed on x′ i，s The context-enhanced feature x is obtained from the above. i,b ;

[0061] The second fusion module is used to process x separately based on the channel attention mechanism. i，s and x i，b The channel features in the data are dynamically adjusted, and the dynamically adjusted x is then processed. i，s and x i，b By performing fusion, the joint feature z of the i-th frequency band is obtained. i ;

[0062] The first fusion module is used to determine the weighting coefficients of each frequency band based on the energy intensity of its signal components and to weight the corresponding joint features to obtain the weighted features; the weighted features of each frequency band are then fused to obtain the fused features.

[0063] The classifier is used to obtain the corresponding hypertension screening results based on the fused features. It should be noted that there are many types of classifiers that can be used, such as MLP, SVM, CNN, fully connected layers, softmax layers, etc., which are not limited here.

[0064] It should be noted that there are multiple methods for fusing features in the first and second fusion modules, such as splicing fusion, weighted summation, and gating mechanism fusion. No specific method is used here, but splicing fusion is preferred.

[0065] In one alternative implementation, the feature enhancement module includes a convolutional unit and a cross-attention unit;

[0066] The feature enhancement module calculates the relevance r. i,sb When x′ i，s The features obtained after the convolution unit are used as the Q-value and K-value, respectively, x′ i，b The V value is input into the cross-attention unit to obtain the relevance r. i,sb ;

[0067] The feature enhancement module calculates the relevance r. i，bs When x′ i，b The features obtained after the convolution unit are used as the Q-value and K-value, respectively, x′ i，s The V value is input into the cross-attention unit to obtain the relevance r. i，bs .

[0068] It should be noted that the type of convolutional unit is not limited, and can be a conventional convolutional unit, a depthwise separating convolutional unit, etc., with a depthwise separating convolutional unit being preferred.

[0069] It should be noted that the second fusion module includes a channel attention module, which is used to perform channel attention on x respectively. i，s and x i，b The channel features are dynamically adjusted. Various channel attention modules can be used, such as the SE module, CBAM module, and CA module, with the SE module being preferred.

[0070] In one optional implementation, the first fusion module includes: a feature extraction unit;

[0071] When determining the weighting coefficients for each frequency band, the first fusion module uses a feature extraction unit to extract the energy intensity features f of the signal components in each frequency band. e，1 f e，2 , ..., f e，N and for f e，1 f e，2 , ..., f e，N After normalization, the weighting coefficients σ1, σ2, ..., σ of each frequency band are obtained. N .

[0072] It should be noted that the feature extraction unit can be a fully connected layer, CNN, RNN, autoencoder, encoder in Transformer, etc., without limitation, but a fully connected layer is preferred.

[0073] It should be noted that there are various methods for the signal decomposition module to decompose the input BCG signal segment x into N signal components in different frequency bands, such as bandpass filtering and wavelet decomposition algorithms, which are not limited here. Preferably, convolution kernels of different sizes are used to decompose the input BCG signal segment x into N signal components x1, x2, ..., x in different frequency bands. N Specifically, it includes:

[0074] To minimize Taking the difference between x and y as the target, the parameters of each convolution kernel are adjusted; after adjustment, the signal component x in the i-th frequency band is obtained. i =Conv i (x);

[0075] Among them, Conv i (·) represents the convolution operation corresponding to the i-th convolution kernel.

[0076] In one optional implementation, the aforementioned BCG signal segment is obtained by sequentially filtering and uniformly dividing the original acquired BCG signal. Preferably, the length of the aforementioned BCG signal segment is 5 to 30 seconds.

[0077] To further illustrate the hypertension screening system provided by the present invention, a specific embodiment is described in detail below:

[0078] In this embodiment, N=3 is used as an example. The specific hypertension screening system is as follows: Figure 2 As shown, BCG signal segments are used as detection signals. Through signal decomposition and feature extraction, combined with a deep learning model, hypertension screening is performed using continuous fixed-length signal segments. The specific process involves the following steps:

[0079] Signal acquisition:

[0080] The user's raw BCG signal is acquired. In this embodiment, a pressure sensor is used to acquire the BCG signal, and each user's BCG signal is stored independently as a pair of time and amplitude values. The amplitude of the raw BCG signal is normalized to the [0,1] interval. Then, a root mean square (RMS) filter is used to detect invalid signals, remove motion artifacts and signals when the patient deviates from the sensor, thereby obtaining a high-quality signal input. The RMS filter formula is as follows:

[0081]

[0082] Where S[n] represents the filtered signal and X[n] represents the unfiltered signal, with a window length of W = 50. The root mean square of the local energy of the signal is calculated by summing the squares, thus smoothing the signal and highlighting high-energy regions for easier abnormal detection. Based on empirical values, the lower threshold for the effective signal is set to 0.1, and the upper threshold to 0.25. When the amplitude of S[n] is less than 0.1, it is considered that the patient has deviated from the sensor and is marked as an invalid signal segment; when the signal amplitude is greater than 0.25, it is considered a motion artifact and is also marked as an invalid signal segment.

[0083] After removing invalid signal segments from the original signal, the remaining segments longer than 5s are cut into multiple 5s BCG signal segments to facilitate subsequent work.

[0084] Signal decomposition:

[0085] This embodiment decomposes BCG signal segments using multi-scale convolution kernels, dividing them into low-frequency, mid-frequency, and high-frequency signal components. The low- and mid-frequency components mainly reflect the morphological characteristics of the BCG signal segments, such as the overall waveform trend and amplitude changes; the high-frequency components contain features closely related to hypertension, such as the enhancement of high-frequency harmonics caused by arterial wall sclerosis.

[0086] Let the BCG signal segment be x∈R D (where D represents the length of the BCG signal segment), the decomposition process is as follows:

[0087]

[0088] Among them, Conv l Conv m Conv h These represent the convolution operations corresponding to low-frequency, mid-frequency, and high-frequency convolution kernels, respectively, used to capture the dynamic changes of signals at different receiving field scales; (symbols omitted) This indicates point-by-point addition. In this embodiment, Conv l Conv m Conv h The corresponding convolutional kernel sizes were set to 201, 51, and 11 to cover different scales of the receiving field, and the frequency separation effect was further enhanced by stacking two layers of convolutional blocks. This was achieved through reconstruction and smoothing loss, i.e., minimizing... Targeting the difference between x and y, the parameters of each convolution kernel are adjusted to achieve Conv l Conv m Conv h The learner acquired the ability to extract different frequency components from the original BCG signal segment. After adjustment, the low-frequency component x was extracted from the BCG signal segment x.l =Conv l (x)∈R D , intermediate frequency component x m =Conv m (x)∈R D and high-frequency components x h =Conv h (x)∈R D .

[0089] like Figure 3 The diagram shows a BCG signal segment and its decomposition; where (a) is the original BCG signal segment diagram; (b) is the low-frequency component diagram; (c) is the mid-frequency component diagram; and (d) is the high-frequency component diagram.

[0090] Feature extraction:

[0091] In the feature extraction stage, this invention designs a feature extraction method that integrates local and contextual information. For each frequency band signal, conventional convolution is used to extract local beat features, and dilated convolution is used to extract long-range contextual features. A local-global cross-attention mechanism is used to enhance feature interaction and optimize the correlation between frequency domain components, thereby generating a more comprehensive feature representation. Specifically:

[0092] For each frequency component x of the BCG signal segment x l x m x h The processing utilizes both regular convolutional modules and dilated convolutional modules to extract local heartbeat features and long-range contextual features. Regular convolution is primarily used to capture short-range heartbeat waveform features, while dilated convolution captures a wider range of contextual relationships by expanding the receptive field. Specifically, for each frequency component, a local feature x′ is extracted using a regular convolution Conv(·). s The dilated convolution DConv(·) is used to extract long-range contextual features x′. b In this embodiment, each regular convolutional module and dilated convolutional module contains three convolutional layers, each followed by a batch normalization layer and an activation layer (preferably using the ReLU activation function). The kernel size of the regular convolutional layer is 3, and the kernel size of the dilated convolutional layer is 5, with a dilation rate of 3. The number of channels in each layer is set to 64, 128, and 128 respectively. The output of the convolution can be expressed as:

[0093] The interaction between features is enhanced through a local-global cross-attention mechanism, further optimizing local features and contextual features. In this embodiment, x′ s and x′ bThe inputs are respectively fed into the cross-attention module CrossAttn(·,·) to capture the correlation between local and global features. The depthwise separable convolution DSConv(·) is used to extract complementary information between channels, generating attention weights to strengthen the interaction between local and contextual features. The attention mechanism focuses on the important information interaction between different frequency components by constructing a correlation matrix. The cross-attention operation is defined as follows:

[0094] x s =CrossAttn(x′) s DSConv(x′) b ))+x s

[0095] x b =CrossAttn(x′) b DSConv(x′) s ))+x b

[0096] Enhanced local features x based on channel attention mechanism s and contextual features x b In this embodiment, Squeeze-and-Excitation processing is performed on x. The specific process is as follows: s and x b Global average pooling is performed separately to extract global information for each channel. Then, a fully connected layer is used to calculate the weight of each channel, dynamically adjusting the importance of different channels. The processed local features and context features are concatenated to obtain the joint features for the corresponding frequency band. In this embodiment, the joint features for low frequency, mid frequency, and high frequency are denoted as z. l z m and z h .

[0097] Feature fusion:

[0098] By employing a frequency domain energy attention mechanism, the joint features of each frequency band are effectively fused to generate a latent representation, specifically:

[0099] Obtain the low-frequency signal component x of the BCG signal segment l Intermediate frequency signal component x m and high-frequency signal components x h Energy intensity e l e m e h .

[0100] Energy intensity e l e m e hThe input is a fully connected layer, and then the energy intensity is normalized along the channel dimension using the SoftMax function to obtain the weight coefficients σ of the frequency band features at low, mid, and high frequencies. l , σ m , σ h Weighted fusion of frequency band features is performed based on weight coefficients. The features of low-frequency, mid-frequency, and high-frequency components are multiplied by their corresponding weight coefficients through element-wise multiplication to obtain the weighted features. and The weighted features are then concatenated to form a fused feature; the formula for obtaining the weighting coefficients is as follows:

[0101] σ l , σ m , σ h =SoftMax(concat(FC(e l ), FC(e m ), FC(e h )))

[0102] Where FC(·) represents a fully connected layer.

[0103] Classification:

[0104] The fused features are used as input and further processed by a multilayer perceptron (MLP) to generate classification labels for output. In this embodiment, the MLP consists of two fully connected layers with dimensions of 384×128 and 128×2, respectively. A sigmoid activation function is used between the hidden and output layers to improve the non-linearity of the classification and the model's expressive power. The output layer generates the final binary classification label to indicate the presence or absence of hypertension risk.

[0105] In summary, this embodiment decomposes the signal into low-frequency, mid-frequency, and high-frequency components using multi-scale convolutional kernels, focusing on capturing high-frequency harmonic features associated with hypertension. Simultaneously, a frequency-domain energy attention mechanism effectively fuses features from each frequency band to generate a latent representation, thereby significantly improving the accuracy of screening results. Using this method, detection can be completed with only a 5-second signal acquisition, overcoming the limitation of traditional methods requiring 30 seconds to 5 minutes of signal acquisition time, significantly improving detection efficiency and user experience.

[0106] Secondly, the present invention provides a training method for the above-mentioned hypertension screening system, comprising:

[0107] The BCG signal segments from the pre-collected training set are input into the hypertension screening system to obtain the corresponding hypertension screening results; based on the hypertension screening results of each BCG signal segment in the training set and the corresponding label indicating whether hypertension exists, the classification loss is calculated.

[0108] A training loss including classification loss is constructed; the hypertension screening system is trained based on the training loss.

[0109] The training set includes BCG signal fragments from different subjects, specifically BCG signal fragments from normal individuals and BCG signal fragments from hypertensive patients. The method for obtaining these BCG signal fragments is the same as the method used in the first aspect of this invention.

[0110] It should be noted that the classification loss mentioned above is used to measure the difference loss between the hypertension screening results of BCG signal segments and the corresponding labels; there are various measurement methods, such as cross-entropy loss, L2 loss, hinge loss, exponential loss, etc., which are not limited here.

[0111] In one alternative implementation, a deep hierarchical clustering strategy is further introduced to train the hypertension screening system, and the training loss also includes clustering loss.

[0112] The expression for clustering loss is:

[0113]

[0114] The training set includes two types of BCG signal segments labeled as those with and without hypertension; N k f is the number of BCG signal segments of the k-th class in the training set; k,c To input the c-th BCG signal fragment of the k-th category in the training set into the hypertension screening system via the fusion feature obtained by the first fusion module; Centr k The cluster center for the k-th category is obtained by averaging the fusion features of each BCG signal segment under the k-th category in the training set.

[0115] The hypertension screening system is trained by simultaneously minimizing classification loss and clustering loss.

[0116] Specifically, such as Figure 4 The figure shows a t-SNE visualization of the clustering results of the latent feature space generated by the hypertension screening system without and with clustering loss; where (a) is the clustering result of the latent feature space generated by the hypertension screening system without clustering loss; and (b) is the clustering result of the latent feature space generated by the hypertension screening system with clustering loss.

[0117] This invention addresses the learning bias caused by individual specificity by introducing a joint classification loss and clustering loss optimization model, thereby enhancing generalization ability in open scenarios. The classification loss constrains the accuracy of the classification results, while the clustering loss further improves the model's discriminative ability by enhancing feature clustering and class separation.

[0118] The related technical solutions are the same as those provided in the first aspect of this invention for the hypertension screening system, and will not be described in detail here.

[0119] Thirdly, the present invention provides a method for hypertension screening, comprising: inputting a BCG signal fragment to be detected into the hypertension screening system provided in the first aspect of the present invention, and obtaining a corresponding hypertension screening result.

[0120] The method for obtaining the BCG signal segment described above is the same as the method for obtaining the BCG signal segment used in the first aspect of this invention.

[0121] The related technical solutions are the same as those provided in the first aspect of this invention for the hypertension screening system, and will not be described in detail here.

[0122] Fourthly, the present invention provides an electronic device comprising: a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the method provided in the second or third aspect of the present invention.

[0123] The related technical solutions are the same as the training method of the hypertension screening system provided in the second aspect of the present invention and the hypertension screening method provided in the third aspect of the present invention, and will not be described in detail here.

[0124] Fifthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when executed by a processor, controls the device in which the storage medium is located to perform the method provided in the second or third aspect of the present invention.

[0125] The related technical solutions are the same as the training method of the hypertension screening system provided in the second aspect of the present invention and the hypertension screening method provided in the third aspect of the present invention, and will not be described in detail here.

[0126] In a sixth aspect, the invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the method provided in the second or third aspect of the invention.

[0127] The related technical solutions are the same as the training method of the hypertension screening system provided in the second aspect of the present invention and the hypertension screening method provided in the third aspect of the present invention, and will not be described in detail here.

[0128] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A hypertension screening system, characterized in that, include: The system consists of a signal decomposition module, N feature extraction modules, a first fusion module, and a classifier. N≥2； The signal decomposition module is used to decompose the input BCG signal segment into N signal components in different frequency bands. And input them one by one into N feature extraction modules; No. i The feature extraction module includes: A standard convolutional module is used for extraction. Local heartbeat characteristics ; Dilated convolution module, used for extraction Contextual features ; The feature enhancement module is used for computation based on the cross-attention mechanism. Compared to degree of correlation as well as Compared to degree of correlation ;Will Overlay Local heartbeat enhancement features were obtained. ;Will Overlay Context-enhanced features are obtained from the above. ; The second fusion module is used to perform separate processing based on the channel attention mechanism. and The channel features in the data are dynamically adjusted, and the dynamically adjusted data is then processed. and By merging, we obtain the first... i Joint features across frequency bands ; The first fusion module is used to determine the weighting coefficient of each frequency band based on the energy intensity of its signal components and to weight the corresponding joint features to obtain its weighted features; and to fuse the weighted features of each frequency band to obtain the fused features. The classifier is used to obtain the corresponding hypertension screening results based on the fusion features; in, ; The feature enhancement module includes convolutional units and cross-attention units; The feature enhancement module calculates the relevance. At that time, The features obtained after passing through the convolutional unit are used as Q-values ​​and K-values, respectively. The V value is input into the cross-attention unit to obtain the relevance. ; The feature enhancement module calculates the relevance. At that time, The features obtained after passing through the convolutional unit are used as Q-values ​​and K-values, respectively. The V value is input into the cross-attention unit to obtain the relevance. .

2. The hypertension screening system according to claim 1, characterized in that, The convolutional unit is a depthwise separable convolutional unit.

3. The hypertension screening system according to any one of claims 1-2, characterized in that, The first fusion module includes: a feature extraction unit; When determining the weighting coefficients for each frequency band, the first fusion module uses the feature extraction unit to extract the energy intensity features of the signal components in each frequency band. and to Normalization is performed to obtain the weighting coefficients for each frequency band. .

4. The hypertension screening system according to any one of claims 1-2, characterized in that, The signal decomposition module is used to decompose the input BCG signal segments using convolution kernels of different sizes. Decomposed into N signal components at different frequency bands Specifically, it includes: To minimize and Taking the differences between them as the target, the parameters of each convolution kernel are adjusted; after adjustment, the first... i Signal components in each frequency band ; in, Indicates the first i The convolution operation corresponding to each convolution kernel.

5. The hypertension screening system according to claim 1, characterized in that, The BCG signal segment is obtained by sequentially filtering and uniformly dividing the original acquired BCG signal. The length of the BCG signal segment is 5 to 30 seconds.

6. The training method for the hypertension screening system according to any one of claims 1-5, characterized in that, include: The BCG signal segments from the pre-collected training set are input into the hypertension screening system to obtain the corresponding hypertension screening results. Based on the hypertension screening results of each BCG signal segment in the training set and the corresponding labels indicating whether hypertension exists, the classification loss is calculated. Construct a training loss that includes classification loss; The hypertension screening system is trained based on the training loss.

7. The training method according to claim 6, characterized in that, The training loss also includes clustering loss; The expression for the clustering loss is: The training set includes two types of BCG signal segments labeled as having hypertension and not having hypertension. For the training set, the first k The number of BCG signal segments in each category; To make the training set the first k The first category c The fusion features obtained by the first fusion module of the hypertension screening system when a BCG signal segment is input into the hypertension screening system; For the first k The cluster centers for the nth category are determined by analyzing the training set. k The fusion features of each BCG signal segment under each category are averaged to obtain the result.

8. A method for hypertension screening, wherein the hypertension screening method is executed by a computer, characterized in that, include: The BCG signal segment to be detected is input into the hypertension screening system according to any one of claims 1-5 to obtain the corresponding hypertension screening result.

9. An electronic device, characterized in that, include: A memory and a processor, the memory storing a computer program, the processor executing the method according to any one of claims 6-8.

Images