Hypertension screening system, training method, hypertension screening method and electronic equipment

Abstract

The invention discloses a hypertension screening system, a training method, a hypertension screening method and electronic equipment, and belongs to the technical field of medical detection. The method comprises the following steps: firstly, decomposing an input BCG signal segment into signal components under a plurality of frequency bands so as to comprehensively capture multi-frequency band information related to hypertension; on the basis, the relative relation between different peaks in a single heart beat of a hypertensive patient and the significant difference between rhythms and modes between different heart beats compared with those of a non-hypertensive patient are considered, local complex features in the single heart beat are captured by using common convolution, and meanwhile, remote context information between the heart beats is extracted by using expansion convolution; a cross attention mechanism is introduced to realize interaction between local features and remote context information, so that key features related to hypertension diagnosis are focused more accurately; and 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] The present invention belongs to the technical field of medical detection, and more specifically, relates to a hypertension screening system, a training method, a hypertension screening method, and an electronic device. Background Art

[0002] In recent years, the early screening and detection of hypertension have been of great significance in preventing cardiovascular diseases. However, the existing hypertension monitoring technologies face many challenges, including the comfort of the user experience, detection accuracy, and the universality of the technology. Currently, the commonly used blood pressure detection methods include traditional mercury sphygmomanometers and electronic sphygmomanometers. However, these methods require the user to remain stationary and complete the measurement by cuff pressurization, which may cause obvious discomfort to users who need frequent detection, especially patients who need nocturnal monitoring. In addition, although existing continuous blood pressure monitoring technologies (such as photoplethysmography (PPG) and seismocardiogram (SCG)) can achieve non-invasive monitoring to a certain extent, they need to be in direct contact with the human body and are difficult to meet the requirements of completely non-intrusive monitoring.

[0003] BCG (Ballistocardiogram) is a technology that indirectly reflects cardiovascular activities by measuring the minute movements caused by cardiac pulsations. This technology realizes signal acquisition by embedding sensors in mattresses, chairs, or other daily necessities, and can non-invasively obtain physiological information related to cardiovascular activities, providing a new approach for the non-intrusive screening of hypertension. However, due to the complex composition and diverse waveforms of BCG signals, and their susceptibility to body movement artifacts and external noise interference, the hypertension detection technology based on BCG signals faces the problems of low signal quality and insufficient detection accuracy in practical applications.

[0004] Currently, the existing hypertension screening technologies based on BCG signals mainly adopt the method of combining feature engineering with machine learning. Starting from the morphological perspective, they explore the correlations between logarithmic energy, peak kurtosis, and other morphological statistical features in BCG waveforms and blood pressure. However, such methods have certain limitations: First, during the signal acquisition process, morphological features are easily interfered by other frequency components, resulting in insufficient robustness of feature extraction; Second, these methods often ignore the potential connection between high-frequency components in the signal and blood pressure, and are difficult to comprehensively reflect the complex characteristics of blood pressure changes. Therefore, the accuracy of the existing hypertension screening technologies based on BCG signals remains to be improved. Summary of the Invention

[0005] Aiming at the above defects or improvement requirements of the existing technology, the present invention provides a hypertension screening system, a training method, a hypertension screening method, and an electronic device to solve the technical problem of low accuracy of hypertension screening in the existing technology.

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

[0007] The signal decomposition module is configured to decompose the input BCG signal segment into N signal components x 1 , x 2 , …, x N , and input them into the N feature extraction modules one by one;

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

[0009] A conventional convolution module for extracting the local heartbeat feature x′ i of x i，s ;

[0010] A dilated convolution module for extracting the context feature x′ i of x i，b ;

[0011] The feature enhancement module is configured to calculate the correlation degree r i，s of x′ i，b relative to x′ i,sb and the correlation degree r i，b of x′ i，s relative to x′ i，bs based on the cross-attention mechanism; superimpose r i,sb onto x′ i，b to obtain the local heartbeat enhanced feature x i，s ; superimpose r i，bs onto x′ i，s to obtain the context enhanced feature x i,b ;

[0012] The second fusion module is configured to dynamically adjust the channel features in x i，s and x i，b respectively based on the channel attention mechanism, and fuse the dynamically adjusted x i，s and x i，b to obtain the joint feature z i in the i-th frequency band;

[0013] The first fusion module is configured to determine the weight coefficient for each frequency band based on the energy intensity of its signal component and weight the corresponding joint feature to obtain its weighted feature; fuse the weighted features of each frequency band to obtain the fused feature;

[0014] The classifier is configured to obtain the corresponding hypertension screening result based on the fused feature;

[0015] where \(i = 1, 2, \ldots, N\).

[0016] Further preferably, the feature enhancement module includes a convolutional unit and a cross-attention unit;

[0017] When the feature enhancement module calculates the correlation degree \(r\) i,sb it uses the features of \(x'\) i，s after passing through the convolutional unit as the Q value and the K value respectively, and uses \(x'\) i，b as the V value and inputs it into the cross-attention unit to obtain the correlation degree \(r\) i,sb ;

[0018] When the feature enhancement module calculates the correlation degree \(r\) i，bs it uses the features of \(x'\) i，b after passing through the convolutional unit as the Q value and the K value respectively, and uses \(x'\) is as the V value and inputs it into the cross-attention unit to obtain the correlation degree \(r\) i，bs .

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

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

[0021] When the first fusion module determines the weight coefficients of each frequency band, it uses the feature extraction unit to extract the features \(f\) e，1 of the energy intensity of the signal components in each frequency band respectively, \(f\) e，2 , \(\ldots\), \(f\) e，N , and normalizes \(f\) e，1 , \(f\) e，2 , \(\ldots\), \(f\) e，N to obtain the weight coefficients \(\sigma\) 1 , \(\sigma\) 2 , \(\ldots\), \(\sigma\) N .

[0022] Further preferably, the signal decomposition module is used to decompose the input BCG signal segment \(x\) into \(N\) signal components \(x\) 1 , \(x\) 2 , \(\ldots\), \(x\) N in different frequency bands by using convolutional kernels of different sizes, specifically including:

[0023] Taking the minimization of the difference between and \(x\) as the goal, adjusting the parameters of each convolutional kernel; after the adjustment is completed, the signal component \(x\) i in the \(i\)-th frequency band is obtained as \(x\) i =\(Conv\)

[0024] where \(Conv\) i (\(\cdot\)) represents the convolutional operation corresponding to the \(i\)-th convolutional kernel.

[0025] Further preferably, the above BCG signal segment is obtained by sequentially performing filtering processing and uniform partitioning processing on the originally collected BCG signal.

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

[0027] In a second aspect, the present invention provides a training method for the above hypertension screening system, including:

[0028] Input the BCG signal segments in the pre-collected training set into the hypertension screening system to obtain the corresponding hypertension screening results; calculate the classification loss based on the hypertension screening results of each BCG signal segment in the training set and the corresponding label indicating whether there is hypertension.

[0029] Construct a training loss including the classification loss; train the hypertension screening system based on the training loss.

[0030] Further preferably, the above training loss further includes a clustering loss;

[0031] The expression of the clustering loss is:

[0032]

[0033] wherein, the training set includes two types of BCG signal segments labeled with having hypertension and not having hypertension; N k is the number of BCG signal segments in the k-th category in the training set; f k,c is the fusion feature obtained by inputting the c-th BCG signal segment in the k-th category in the training set into the hypertension screening system via the first fusion module; Centr k is the clustering center of the k-th category, which is obtained by taking the average value of the fusion features of each BCG signal segment in the k-th category in the training set.

[0034] In a third aspect, the present invention provides a hypertension screening method, including: inputting the BCG signal segment to be detected into the hypertension screening system provided in the first aspect of the present invention to obtain the corresponding hypertension screening result.

[0035] In a fourth aspect, the present invention provides an electronic device, including: a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it executes the method provided in the second aspect or the third aspect of the present invention.

[0036] Fifth aspect, the present invention also provides a computer-readable storage medium, which includes a stored computer program. When the computer program is run by a processor, it controls the device where the storage medium is located to execute the method provided in the second aspect or the third aspect of the present invention.

[0037] Sixth aspect, the invention also provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the method provided in the second aspect or the third aspect of the present invention.

[0038] Generally speaking, through the above technical solutions conceived by the present invention, the following beneficial effects can be achieved:

[0039] 1. The present invention provides a hypertension screening system. First, the input BCG signal segment is decomposed into signal components in multiple frequency bands to comprehensively capture multi-band information related to hypertension, avoiding the problem that high-frequency band signals are ignored due to their low amplitudes in traditional methods, thereby improving the integrity of feature expression. On this basis, considering that the relative relationship between different peaks in a single heartbeat of hypertensive patients and the rhythm and pattern between different heartbeats are significantly different from those of non-hypertensive patients, a dual-branch parallel feature extraction network is designed. Ordinary convolutions are used to capture local complex features within a single heartbeat, while at the same time, dilated convolutions are 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 as to more accurately focus on the key features related to hypertension diagnosis. Finally, by fusing features in different frequency bands, accurate and complete feature expression is achieved, improving the accuracy of hypertension screening.

[0040] 2. Further, in the hypertension screening system provided by the present invention, the feature enhancement module includes a convolution unit and a cross-attention unit. The cross-attention unit is used to capture the correlation between local heartbeat features and global context features. On this basis, the convolution unit extracts complementary information between channels to strengthen the interaction relationship between local heartbeat features and context features, realizing more accurate feature expression and further improving the accuracy of hypertension screening.

[0041] 3. Further, in the hypertension screening system provided by the present invention, the convolution unit in the feature enhancement module is a depthwise separable convolution unit, which performs convolution on each input channel separately without cross-channel calculation, significantly reducing the number of parameters and the amount of calculation, and improving the efficiency of hypertension screening.

[0042] 4. Further, in the hypertension screening system provided by the present invention, the signal decomposition module is used to decompose the input BCG signal segment into N signal components in different frequency bands by using convolution kernels of different sizes, which 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. Further, in the hypertension screening system provided by the present invention, it is obtained by sequentially performing filtering processing and uniform division processing on the originally collected BCG signal. The motion artifacts and signals when the patient deviates from the sensor are removed through the filtering processing, thereby obtaining a high-quality signal input and further improving the accuracy of hypertension screening.

[0044] 6. Further, considering that the prior art relies on long-term signal acquisition (such as 30 seconds to 5 minutes) for hypertension screening, which has relatively high requirements for the user to maintain a stable posture in practical applications. In the hypertension screening system provided by the present invention, through the design of the above-mentioned signal decomposition module, feature extraction module and first fusion module, the length of the input BCG signal segment is set to 5 seconds to 30 seconds, and rich information can also be collected, avoiding the above problems of the prior art and significantly improving the detection efficiency and user experience.

[0045] 7. The present invention provides a training method for a hypertension screening system. Based on the design of the above-mentioned hypertension screening system, the hypertension screening system is trained through a pre-collected training set, and a hypertension screening system with relatively high accuracy can be obtained.

[0046] 8. Further, the training method of the hypertension screening system provided by the present invention introduces classification loss and clustering loss. Among them, the classification loss constrains the accuracy of the classification result, and the clustering loss further improves the discrimination ability of the model by enhancing feature aggregation and class separation, solves the learning deviation problem caused by individual specificity, enables the present invention to not rely on a closed set scenario (that is, it does not require the test object to participate in the model training stage), can meet the need for hypertension screening of unknown users in an open set scenario, and enhances the generalization ability in an open scenario. Description of the Drawings

[0047] Figure 1 It is an analysis diagram of hypertension and non-hypertension signals and corresponding frequency components; among them, (a) is the original hypertension signal diagram; (b) is the frequency component analysis diagram of the hypertension signal; (c) is the original non-hypertension signal diagram; (d) is the frequency component analysis diagram of the non-hypertension signal;

[0048] Figure 2 It is a result schematic diagram of a hypertension screening system provided by an embodiment of the present invention;

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

[0050] Figure 4 The t-SNE visualization diagram of the clustering results of the latent feature space generated by the hypertension screening system when no clustering loss is introduced and when clustering loss is introduced according to the embodiments of the present invention; wherein, (a) is the clustering result of the latent feature space generated by the hypertension screening system without introducing clustering loss; (b) is the clustering result of the latent feature space generated by the hypertension screening system with the introduction of clustering loss. Detailed implementation manners

[0051] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present 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 only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0052] To achieve the above objective, in a first aspect, the present invention provides a hypertension screening system. As Figure 1 shown is the analysis diagram of hypertension and non-hypertension signals and their corresponding frequency components; wherein, (a) is the original hypertension signal diagram; (b) is the frequency component analysis diagram of the hypertension signal; (c) is the original non-hypertension signal diagram; (d) is the frequency component analysis diagram of the non-hypertension signal.

[0053] The present invention takes into account that continuous hypertension will lead to an increase in arterial wall stiffness, which in turn causes the generation of more high-frequency harmonic vibrations. First, the BCG signal segment is decomposed into multiple frequency bands, and the features of each frequency band are extracted respectively. This design effectively avoids the problem that the high-frequency band signals are ignored due to their low amplitudes in traditional methods, so as to be able to capture more comprehensive multi-band information related to hypertension and improve the integrity of feature expression.

[0054] In addition, it is observed that there are significant differences in the relative relationships between different peaks in a single heartbeat of hypertensive patients, as well as the rhythm and pattern between different heartbeats, compared with non-hypertensive patients. For this reason, the present invention designs a dual-branch parallel feature extraction network: on the one hand, ordinary convolution is used to capture the local complex features within a single heartbeat; on the other hand, dilated convolution is used to extract the long-distance context information between heartbeats. On this basis, a cross-attention mechanism is introduced to realize the interaction between local features and long-distance context information, so as to more accurately focus on the key features related to hypertension diagnosis and achieve 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 x 1 , x 2 , …, x N , and input them into the N feature extraction modules one by one;

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

[0058] A conventional convolution module for extracting the local heartbeat feature x′ i of x i，s ;

[0059] A dilated convolution module for extracting the context feature x′ i of x i，b ;

[0060] The feature enhancement module is used to calculate the correlation degree r i，s of x′ i，b relative to x′ i,sb and the correlation degree r i，b of x′ i，s relative to x′ i，bs based on the cross-attention mechanism; superimpose r i,sb onto x′ i，b to obtain the local heartbeat enhanced feature x i，s ; superimpose r i，bs onto x′ i，s to obtain the context enhanced feature x i,b ;

[0061] The second fusion module is used to dynamically adjust the channel features in x i，s and x i，b respectively based on the channel attention mechanism, and fuse the dynamically adjusted x i，s and x i，b to obtain the joint feature z i in the i-th frequency band;

[0062] The first fusion module is used to determine the weight coefficient for each frequency band based on the energy intensity of its signal component and weight the corresponding joint feature to obtain its weighted feature; fuse the weighted features of each frequency band to obtain the fusion feature;

[0063] The classifier is used to obtain the corresponding hypertension screening result based on the fused features; it should be noted that there are various classifiers that can be used, such as MLP, SVM, CNN, fully connected layer, softmax layer, etc., which are not limited here.

[0064] It should be noted that there are various methods for fusing features in the first fusion module and the second fusion module. Methods such as concatenation fusion, weighted summation, and gated mechanism fusion can be used for fusion, which are not limited here, and concatenation fusion is preferably used.

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

[0066] When the feature enhancement module calculates the correlation degree r i,sb it takes the features of x' i，s after passing through the convolutional unit as the Q value and the K value respectively, and x' i，b as the V value and inputs them into the cross-attention unit to obtain the correlation degree r i,sb ;

[0067] When the feature enhancement module calculates the correlation degree r i，bs it takes the features of x' i，b after passing through the convolutional unit as the Q value and the K value respectively, and x' i，s as the V value and inputs them into the cross-attention unit to obtain the correlation degree r i，bs .

[0068] It should be noted that the type of the convolutional unit is not limited, and it can be a conventional convolutional unit, a depthwise separable convolutional unit, etc., and a depthwise separable convolutional unit is preferably used.

[0069] It should be noted that the second fusion module includes a channel attention module to dynamically adjust the channel features in x i，s and x i，b respectively based on the channel attention mechanism. There are various channel attention modules here, such as SE module, CBAM module, CA module, etc., and the SE module is preferably used.

[0070] In an alternative embodiment, the first fusion module includes: a feature extraction unit;

[0071] When the first fusion module determines the weight coefficients of each frequency band, it uses the feature extraction unit to extract the features f e，1 of the energy intensity of the signal components in each frequency band, f e，2 , …, f e，N , and normalizes f e，1 , f e，2 , …, f e，N to obtain the weight coefficients σ of each frequency band1 , σ 2 , …, σ N .

[0072] It should be noted that the feature extraction unit can be a fully connected layer, CNN, RNN, autoencoder, encoder in Transformer, etc., which is not limited here, and preferably a fully connected layer.

[0073] It should be noted that there are various methods for the signal decomposition module to decompose the input BCG signal segment x into signal components in N different frequency bands. The band-pass filtering method, wavelet decomposition algorithm, etc. can be used, which is not limited here. Preferably, the input BCG signal segment x is decomposed into signal components x 1 , x 2 , …, x N , specifically including:

[0074] Taking the minimization of the difference between and x as the goal, adjust the parameters of each convolutional kernel; after the adjustment is completed, the signal component x in the i-th frequency band is obtained i = Conv i (x);

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

[0076] In an alternative embodiment, the above BCG signal segment is obtained by sequentially performing filtering processing and uniform partitioning processing on the originally collected BCG signal. Preferably, the length of the above BCG signal segment is 5 seconds to 30 seconds.

[0077] To further illustrate the hypertension screening system provided by the present invention, the following will be described in detail with a specific embodiment:

[0078] In this embodiment, N = 3 is taken as an example. The specific hypertension screening system is as Figure 2 shown. Using the BCG signal segment as the detection signal, through signal decomposition and feature extraction, combined with a deep learning model, hypertension screening is performed through continuous fixed-length signal segments. Specifically, the following processes are involved:

[0079] Signal acquisition:

[0080] Collect the original BCG signals of users. In this embodiment, a pressure sensor is used to obtain BCG signals, and the BCG signals of each user are independently stored in the form of a binary tuple of time and amplitude. Normalize the amplitude of the original BCG signals, standardize them to the interval [0, 1], and then use a root mean square (RMS) filter to detect invalid signals, removing motion artifacts and signals when the patient deviates from the sensor, so as to obtain high-quality signal input. The RMS filtering formula is as follows:

[0081]

[0082] Among them, S[n] represents the filtered signal, X[n] is the signal before filtering, and the window length W = 50 is taken. Square-sum the local energy of the signal and then take the root mean square to smooth the signal and highlight the high-energy regions for easy detection of abnormalities. Set the lower threshold of the valid signal to 0.1 and the upper threshold to 0.25 according to empirical values. When the amplitude of S[n] is less than 0.1, it is regarded that the patient deviates from the sensor and is marked as an invalid signal segment; when the signal amplitude is greater than 0.25, it is regarded as a motion artifact and is also marked as an invalid signal segment.

[0083] After removing the invalid signal segments from the original signal, perform a cutting operation on the remaining segments with a length longer than 5s, cutting them into multiple BCG signal segments with a length of 5s for subsequent work.

[0084] Signal decomposition:

[0085] In this embodiment, a multi-scale convolutional kernel is used to decompose the BCG signal segments, decomposing the BCG signal segments into low-frequency signal components, intermediate-frequency signal components, and high-frequency signal components. The low- and intermediate-frequency components mainly reflect the morphological characteristics of the BCG signal segments, such as the overall waveform trend and amplitude change; the high-frequency component contains features closely related to the hypertension state, such as the enhancement of high-frequency harmonics caused by arterial wall sclerosis.

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

[0087]

[0088] Among them, Conv l 、Conv m 、Conv h respectively represent the convolution operations corresponding to the convolutional kernels of low frequency, intermediate frequency, and high frequency, used to capture the dynamic changes of the signal at different receptive field scales; the symbol represents element-wise addition. In this embodiment, Conv l 、Conv m 、Convh The corresponding convolutional kernel sizes are set to 201, 51, and 11 respectively to cover receptive fields of different scales, and the frequency separation effect is further enhanced by stacking two convolutional blocks. By reconstructing and smoothing the loss, that is, aiming to minimize the difference between and x, the parameters of each convolutional kernel are adjusted, so that Conv l 、Conv m 、Conv h learn the ability to extract different frequency components from the original BCG signal segment. After the adjustment, the low-frequency component x l =Conv l (x)∈R D 、the middle-frequency component x m =Conv m (x)∈R D and the high-frequency component x h =Conv h (x)∈R D are extracted from the BCG signal segment x.

[0089] As Figure 3 shown is the BCG signal segment and its signal decomposition diagram; among them, (a) is the original BCG signal segment diagram; (b) is the low-frequency component diagram; (c) is the middle-frequency component diagram; (d) is the high-frequency component diagram.

[0090] Feature extraction:

[0091] In the feature extraction stage, the present invention designs a feature extraction method that fuses local and context information. For each frequency band signal, a conventional convolution is used to extract local heartbeat features, and a dilated convolution is used to extract long-distance context features. The local-global cross-attention mechanism is used to enhance feature interaction and optimize the correlation between frequency domain components, so as to generate a more comprehensive feature representation. Specifically:

[0092] For each frequency component x l 、x m 、x h of the BCG signal segment x, they are respectively processed by a conventional convolution module and a dilated convolution module to extract local heartbeat features and long-distance context features. The conventional convolution is mainly used to capture the short-distance heartbeat waveform features, while the dilated convolution captures the context relationship in a wider range by expanding the receptive field. Specifically, for each frequency component, the local feature x′ s is extracted by using the conventional convolution Conv(·), and the long-distance context feature x′ b is extracted by using the dilated convolution DConv(·). In this embodiment, each conventional convolution module and dilated convolution module contains three layers of convolution operations, and a batch normalization layer and an activation layer (the activation function is preferably the ReLU activation function) are attached after each layer. The convolution kernel size of the conventional convolution layer is 3, the convolution kernel size of the dilated convolution layer is 5, and the dilation rate is set to 3. The number of channels in each layer is set to 64, 128, and 128 in sequence. The output of the convolution can be expressed as:

[0093] Enhance the interaction between features through the local-global cross-attention mechanism, and further optimize the local features and context features. In this embodiment, x' s and x' b are respectively input into the cross-attention module CrossAttn(·, ·) to capture the correlation between the local and the global, use the depthwise separable convolution DSConv(·) to extract the complementary information between channels, generate the attention weights, and strengthen the interaction relationship between the local features and the context 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] Based on the channel attention mechanism, process the enhanced local feature x s and context feature x b . In this embodiment, Squeeze-and-Excitation processing is performed. The specific process is as follows: Perform global average pooling on x s and x b respectively, extract the global information of each channel, and then calculate the weights of each channel through a fully connected layer to dynamically adjust the importance of different channels. Concatenate the processed local features and context features to obtain the joint feature in the corresponding frequency band. In this embodiment, the joint features in the low frequency, medium frequency, and high frequency are respectively denoted as z l , z m and z h .

[0097] Feature fusion:

[0098] Effectively fuse the joint features of each frequency band through the frequency domain energy attention mechanism to generate a latent representation. Specifically:

[0099] Obtain the low-frequency signal component x of the BCG signal segment l , the intermediate-frequency signal component x m and the high-frequency signal component x h of the energy intensity e l , e m , e h .

[0100] Input the energy intensity e l , e m , e h into the fully connected layer, and then normalize the energy intensity along the channel dimension through the SoftMax function to obtain the weight coefficients σ of the band features at low, intermediate, and high frequencies l , σ m , σ h . Based on the weight coefficients, perform weighted fusion on the band features, multiply the features of the low-frequency, intermediate-frequency, and high-frequency components by their corresponding weight coefficients through element-wise multiplication operations respectively to obtain the weighted features and and splice the weighted features to form the fused feature; among them, the formula for obtaining the weighted coefficient 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 the fully connected layer

[0103] Classification:

[0104] Take the fused feature as the input, and further process it through a multi-layer perceptron (MLP) to generate a classification label for output. In this embodiment, the multi-layer perceptron adopted consists of two fully connected layers, with sizes of 384×128 and 128×2 respectively. A Sigmoid activation function is used between the hidden layer and the output layer to improve the non-linear ability of classification and the expression effect of the model. The output layer generates the final binary classification label, which is used to indicate whether there is a hypertension risk

[0105] In summary, in this embodiment, the signal is decomposed into low-frequency, medium-frequency, and high-frequency components through multi-scale convolutional kernels, and the high-frequency harmonic features related to hypertension are captured emphatically. At the same time, through the frequency-domain energy attention mechanism, the features of each frequency band are effectively fused to generate a latent representation, thereby greatly improving the accuracy of the screening results. Using this method, the detection can be completed only by collecting 5 seconds of short-time signals, breaking through the limitation of the traditional method that requires 30 seconds to 5 minutes of signal collection time, and significantly improving the detection efficiency and user experience.

[0106] In a second aspect, the present invention provides a training method for the above-mentioned hypertension screening system, including:

[0107] Input the BCG signal segments in the pre-collected training set into the hypertension screening system to obtain the corresponding hypertension screening results; calculate the classification loss based on the hypertension screening results of each BCG signal segment in the training set and the corresponding label indicating whether there is hypertension.

[0108] Construct a training loss including the classification loss; train the hypertension screening system based on the training loss.

[0109] The above-mentioned training set includes BCG signal segments of different subjects, specifically including BCG signal segments of normal people and BCG signal segments of hypertensive patients. The acquisition method of the above-mentioned BCG signal segments is the same as the acquisition method of the BCG signal segments adopted in the first aspect of the present invention.

[0110] It should be noted that the above-mentioned classification loss is used to measure the difference loss between the hypertension screening result of the BCG signal segment and the corresponding label; there are various measurement methods, and it can be measured by cross-entropy loss, L2 loss, hinge loss, exponential loss, etc., which are not limited here.

[0111] In an optional implementation manner, a deep hierarchical clustering strategy is further introduced to train the hypertension screening system, and the above-mentioned training loss further includes a clustering loss;

[0112] The expression of the clustering loss is:

[0113]

[0114] where the training set includes two types of BCG signal segments with labels of having hypertension and not having hypertension; N k is the number of BCG signal segments in the kth category in the training set; f k,c is the fusion feature obtained by inputting the cth BCG signal segment in the kth category in the training set into the hypertension screening system through the first fusion module; Centr k is the clustering center of the kth category, which is obtained by taking the average of the fusion features of each BCG signal segment in the kth category in the training set.

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

[0116] Specifically, as Figure 4 shown is the t-SNE visualization diagram of the clustering results of the latent feature space generated by the hypertension screening system without introducing the clustering loss and with introducing the clustering loss; among them, (a) is the clustering result of the latent feature space generated by the hypertension screening system without introducing the clustering loss; (b) is the clustering result of the latent feature space generated by the hypertension screening system with introducing the clustering loss.

[0117] The present invention optimizes the model by introducing the joint classification loss and the clustering loss, solves the learning bias problem caused by individual specificity, and enhances the generalization ability in the open scenario. The classification loss constrains the accuracy of the classification result, and the clustering loss further improves the discrimination ability of the model by enhancing the feature aggregation and class separability.

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

[0119] In the third aspect, the present invention provides a hypertension screening method, including: inputting the BCG signal segment to be detected into the hypertension screening system provided in the first aspect of the present invention to obtain the corresponding hypertension screening result.

[0120] The acquisition method of the above BCG signal segment is the same as the acquisition method of the BCG signal segment adopted in the first aspect of the present invention.

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

[0122] In the fourth aspect, the present invention provides an electronic device, including: a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, it executes the method provided in the second aspect or the third aspect of the present invention.

[0123] The related technical solutions are the same as those of 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 elaborated here.

[0124] In the fifth aspect, the present invention further provides a computer-readable storage medium, the computer-readable storage medium includes a stored computer program, wherein when the computer program is run by a processor, it controls the device where the storage medium is located to execute the method provided in the second aspect or the 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, which will not be elaborated here.

[0126] In a sixth aspect, the invention further provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the methods provided in the second aspect or the third aspect of the present 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, which will not be elaborated here.

[0128] Those skilled in the art can easily understand that the above are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A hypertension screening system, characterized in that: Includes: Signal decomposition module, N feature extraction modules, first fusion module and classifier; N≥2； The signal decomposition module is used to decompose the input BCG signal segment into N signal components x1, x2, ..., x N , and input them into N feature extraction modules one by one; The i-th feature extraction module includes: Conventional convolution module, used to extract x i The local heartbeat feature x′ i，s ; The dilated convolution module is used to extract x i The context feature x′ i，b ; The feature enhancement module is used to calculate x′ based on the cross-attention mechanism i，s Relative to x′ i，b The correlation degree r i,sb and x′ i，b Relative to x′ i，s The correlation degree r i，bs ; r i,sb Superimposed to x′ i，b The local heart beat enhancement feature x is obtained i，s ; r i，bs Superposition to x′ i，s The context-enhanced feature x is obtained i,b ; The second fusion module is used to separately analyze x based on the channel attention mechanism. i，s and x i，b The channel features in are dynamically adjusted, and the dynamically adjusted x i，s and x i，b Fusion is performed to obtain the joint feature z under the i-th frequency band i ; The first fusion module is used to determine the weight coefficient of each frequency band based on the energy intensity of its signal component and weight the corresponding joint features to obtain its weighted features; fuse the weighted features of each frequency band to obtain a fused feature; The classifier is used to obtain corresponding hypertension screening results based on the fusion features; Where i=1,2,…,N.

2. The hypertension screening system according to claim 1, characterized in that: The feature enhancement module includes a convolution unit and a cross attention unit; The feature enhancement module calculates the correlation degree r i,sb When x′ i，s The features after the convolution unit are used as Q value and K value respectively, x′ i，b As the V value input to the cross attention unit, the correlation degree r is obtained i,sb ; The feature enhancement module calculates the correlation degree r i，bs When x′ i，b The features after the convolution unit are used as Q value and K value respectively, x′ i，s As the V value input to the cross attention unit, the correlation degree r is obtained i，bs .

3. The hypertension screening system according to claim 2, characterized in that: The convolution unit is a depth-wise separation convolution unit.

4. The hypertension screening system according to any one of claims 1 to 3, characterized in that: The first fusion module includes: a feature extraction unit; When determining the weight coefficient of each frequency band, the first fusion module uses the feature extraction unit to respectively extract the feature f of the energy intensity of the signal component under each frequency band. e，1 , f e，2 , …, f e，N , and f e，1 , f e，2 , …, f e，N Normalization is performed to obtain the weight coefficients σ1, σ2, …, σ of each frequency band. N .

5. The hypertension screening system according to any one of claims 1 to 3, characterized in that: The signal decomposition module is used to decompose the input BCG signal segment x into N signal components x1, x2, ..., x3 in different frequency bands by using convolution kernels of different sizes. N , specifically including: To minimize The difference between and x is taken as the goal, and the parameters of each convolution kernel are adjusted; after the adjustment is completed, the signal component x in the i-th frequency band is obtained i =Conv i (x); Among them, Conv i (·) represents the convolution operation corresponding to the i-th convolution kernel.

6. The hypertension screening system according to claim 1, characterized in that: The BCG signal segments are obtained by sequentially performing filtering and uniform division processing on the originally collected BCG signals; The length of the BCG signal segment is 5 seconds to 30 seconds.

7. The training method for a hypertension screening system according to any one of claims 1 to 6, characterized in that: include: Inputting the pre-collected BCG signal segments in the training set into the hypertension screening system to obtain corresponding hypertension screening results; Calculate the classification loss based on the hypertension screening results of each BCG signal segment in the training set and the corresponding label indicating whether there is hypertension; Construct training loss including classification loss; Based on the training loss, the hypertension screening system is trained.

8. The training method according to claim 7, characterized in that: The training loss also includes clustering loss; The expression of the clustering loss is: The training set includes two types of BCG signal segments labeled as having hypertension and not having hypertension; N k is the number of BCG signal fragments in the kth category in the training set; f k,c is the fusion feature obtained by the first fusion module in the hypertension screening system when the cth BCG signal segment under the kth category in the training set is input into the hypertension screening system; Centr k is the cluster center of the k-th category, which is obtained by averaging the fusion features of each BCG signal segment under the k-th category in the training set.

9. A method for screening for hypertension, characterized in that: Includes: The BCG signal segment to be detected is input into the hypertension screening system according to any one of claims 1 to 6 to obtain the corresponding hypertension screening result.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the memory stores a computer program, and the processor executes the method according to any one of claims 7 to 9.

Images