A disease classification method based on a multi-scale attention network and a ballistocardiogram

Abstract

The application relates to a disease classification method based on a multi-scale attention network and a ballistocardiogram, belongs to the technical field of human physiological signal processing, and extracts ballistocardiogram features through a convolutional neural network; multi-scale convolution and permutation attention mechanisms are used to construct a multi-scale attention unit; ballistocardiogram features are grouped and respectively pass through different small-size convolution paths; intrinsic feature maps of the ballistocardiogram signal are extracted in a multi-scale manner; a weight matrix is formed by mining the correlation between spatial features and channels through permutation attention; the multi-scale intrinsic maps are weighted and fused to enhance key features of the ballistocardiogram; and the probability that a subject belongs to four types of diseases, such as health or coronary artery, is predicted. The application can effectively improve the accuracy of ballistocardiogram-based disease classification, can reduce network parameters under the same performance condition, can avoid the problems of large depth network calculation and gradient disappearance, and is helpful to the practical application in the fields of home health monitoring and remote auxiliary medical treatment.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of machine learning and medical signal processing, and particularly relates to a disease classification method based on a multi-scale attention network and a ballistocardiogram. BACKGROUND

[0002] Respiratory disorders, hypertension, cardiovascular and cerebrovascular diseases, and other common diseases in the elderly often occur at night, accompanied by falls that can cause greater health damage. Therefore, real-time home monitoring of health status, prevention of sudden diseases, and timely treatment are very important for the healthy life of the elderly.

[0003] Non-contact physiological parameter measurement is the development trend of home health monitoring. The use of piezoelectric film sensors to measure ballistocardiograms to estimate heart rate, respiratory rate, and other physiological signals for health monitoring has made some progress. In terms of health monitoring, user body movement and physiological parameters are usually obtained, combined with measurement signal statistical characteristics, and threshold segmentation, random forest, K-means clustering, support vector machine, and other algorithms are used to evaluate heart-related disease identification. The Geng Duya team mainly studied the denoising method of ballistocardiogram signals and the use of a soft cushion composed of multiple piezoelectric film sensors to obtain a ballistocardiogram. The support vector machine was used to accurately identify four sleeping postures: supine, left lateral recumbency, right lateral recumbency, and prone. Wen et al. proposed a physiological signal database collected by a non-contact sensor. A method based on ballistocardiogram extraction of atrial fibrillation features was used to identify diseases. Nagura et al. proposed a template matching method to monitor heart rate variability. The ballistocardiogram was preprocessed by a low-pass filter to separate the respiratory signal, and then each heartbeat period was segmented as a template. Based on template matching, heart rate variability was identified. Yang Jiafeng et al. designed a non-invasive flexible pressure sleep breathing monitoring system. The respiratory baseline was extracted based on the sampling median, and the respiratory frame and apnea frame were extracted using frame data integral threshold judgment and maximum value threshold judgment. Then, the cross-correlation method was used to determine whether apnea occurred in real time. These traditional methods have less computational load, but due to the low signal-to-noise ratio of the ballistocardiogram, the physiological disease characteristics are not obvious, resulting in insufficient accuracy of physiological parameter estimation. Deep learning methods have strong feature extraction and classification capabilities, providing a powerful tool for physiological parameter estimation and disease identification. Support vector machines, convolutional neural networks, and recurrent neural networks have achieved some results in single disease identification based on ballistocardiogram heart rate estimation and atrial fibrillation identification. SUMMARY

[0004] The application provides a disease classification method based on a multi-scale attention network and a ballistocardiogram.

[0005] The application provides a disease classification method based on a multi-scale attention network and a ballistocardiogram.

[0006] 1.1 Constructing a multi-scale attention network: the multi-scale attention network is composed of a convolution module, a multi-scale attention module and a full connection module;

[0007] 1.1.1 The convolution module is sequentially stacked by a one-dimensional convolution unit, a linear rectifier activation ReLU unit, a batch normalization BN unit and a maximum pooling unit;

[0008] 1.1.2 The multi-scale attention module comprises eight multi-scale convolution attention layers, each multi-scale convolution attention layer is composed of a multi-scale convolution unit embedded into a permutation attention unit; the multi-scale convolution unit is a residual structure, and input features U are mapped to a feature map x∈R c×h×w , wherein c represents the number of feature map channels, h represents the height of the feature map, and w represents the width of the feature map; the feature map is averagely divided into three groups along the channel direction, and each group contains c / 3 channels, namely: x=[x1, x2, x3], R c / 3×h×w The first group of feature maps x1 is input into a dilated convolution with an expansion coefficient of 2, the second group of feature maps x2 is input into a 1×5 convolution, and the third group of feature maps x3 is added with the output y2 of the second group of convolution and input into a 1×5 convolution, to obtain three groups of intrinsic feature maps y1, y2 and y3 of different scales, which are input into a 1×1 convolution layer after being spliced by Concat, so as to fuse the different scale information of the three groups of intrinsic feature maps, and combine the global information and the detail features of the ballistocardiogram for disease classification;

[0009] The output of the convolution layer is the input of the permutation attention unit, and the permutation attention unit is divided into three parts of feature grouping, feature permutation and feature fusion; the input is averagely divided into g groups V=[v1, v2, v3, …, v g ] along the channel number by feature grouping, R c / g×h×w Then, for each group, feature permutation is performed: first, each group is averagely divided into two groups along the channel number, to obtain v K1 , v K2 ∈R C / 2g×h×w ; channel attention processing is performed on v K1 , global average pooling Fgp For v K1 Averaging each channel yields r = F gp (v K1 ), utilizing fully connected layer F c By learning channel correlations and utilizing the gating mechanism of the Sigmoid activation function σ to achieve feature selection, v′ is obtained. K1 =σ(F c (r))·v K1 ; for v K2 Spatial attention processing is performed using group normalization F. gn Calculate v K2 Normalize it using the mean and variance, and use the fully connected layer F c Feature enhancement is performed and the result is mapped to v′ using the Sigmoid activation function. K2 =σ(F c (F gn (v K2 )))·v K2 ; The concat operator is used to concatenate v′ K1 , v′ K2 Combining, we get v′ K =[v′ K1 , v′ K2 ]∈R c / g×h×w The average of all groups yields a feature V′ with the same dimension as the input. Feature fusion involves channel shuffling of V′, dividing it into d groups of e channels each. Each group is stacked vertically to obtain a d×e matrix, with each channel as an element. The matrix is ​​transposed and rearranged using the reshape operator, taking e elements at a time and stacking all rows horizontally to obtain V″, which has the same dimension as V′. The channels between groups have been rearranged, enabling information exchange between channels. V″ is added to the input feature U to obtain the output of the multi-scale convolutional attention layer, enhancing the key signal features of the cardiac impaction image.

[0010] 1.1.3 The fully connected module contains a fully connected layer and a normalized exponential function Softmax, which gives the probability that the input cardiac impact map belongs to various diseases;

[0011] 1.2 Construction of the training set: Cardiac impaction images were collected from 40 volunteers using a piezoelectric thin film. Each cardiac impaction image sample contained C sampling points, denoted as {S}. i The participants included individuals with tachycardia, hypertension, atrial fibrillation, coronary artery disease, and healthy individuals. A label vector L was set. im This indicates that the i-th sample corresponds to the m-th class; finally, the training set {S} is generated. i ,L im}, i = 1, 2, …, N, and divide them into a ballistocardiogram training set and a ballistocardiogram test set according to a ratio of 4:1;

[0012] 1.3 Learning the multi-scale attention network parameter Θ using the ballistocardiogram training set, including the following steps:

[0013] 1.3.1 Constructing a cross-entropy loss function, the calculation formula of which is:

[0014]

[0015] wherein: L m is the mth element of a label vector m , and L is 1 if the ballistocardiogram sample S belongs to the mth class, and is 0 for the rest; log is a logarithmic function;

[0016] 1.3.2 Obtaining a cost function using the ballistocardiogram training set, the calculation formula of which is:

[0017]

[0018] wherein: N is the number of samples in the ballistocardiogram training set;

[0019] 1.3.3 Optimizing the cost function using a batch method and a stochastic gradient descent method, and obtaining the optimal multi-scale attention network parameter Θ and the trained multi-scale attention network f when Γ(Θ) reaches a minimum value;

[0020] 1.4 Given a ballistocardiogram Y of a subject in a ballistocardiogram test set, classifying the given ballistocardiogram based on the trained multi-scale attention network, and determining whether the subject is healthy or has hypertension, tachycardia, atrial fibrillation, or coronary artery disease, including the following steps:

[0021] 1.4.1 Inputting a ballistocardiogram Y with a length of C, and the multi-scale attention network f predicting the probability of the subject of the ballistocardiogram belonging to health, hypertension, tachycardia, atrial fibrillation, or coronary artery disease as:

[0022] p m = f(Y, Θ), m = 1, 2, …, 5, wherein: Θ is a multi-scale attention network parameter;

[0023] 1.4.2 Calculating the maximum probability and classifying the ballistocardiogram as the class with the maximum probability.

[0024] The technical problem solved by the present application: ballistocardiogram is an important basis for non-invasive and non-contact diagnosis of heart-related diseases. However, the non-contact collection method of ballistocardiogram makes the disease-related characteristics blurred, and the waveform division point is not clear, which increases the uncertainty of correct identification of the disease, and the disease classification effect is not good only by using single scale information. In view of this problem, the present application adopts deep convolutional neural network for ballistocardiogram disease identification, proposes to use multi-scale convolution to mine different scale features of ballistocardiogram, and combines attention mechanism to extract key features related to ballistocardiogram disease, thereby improving the accuracy of ballistocardiogram disease classification.

[0025] Advantages of the present application: the multi-scale attention network proposed in the present application uses multi-scale convolution module to realize multi-scale feature extraction of ballistocardiogram, which has fewer network parameters compared with the convolution network of layer-by-layer multi-scale feature extraction. This lightweight design reduces the computational cost in the training process and is more suitable for the lightweight requirements of actual application. In addition, the present application further combines the lightweight permutation attention to mine the internal correlation of multi-scale features, which can extract the slight changes of disease characteristics and make the disease classification more accurate. The present application provides a feasible and efficient disease classification method for home health monitoring, which can be applied to remote medical treatment, wearable health monitoring equipment and non-contact health monitoring system for home care. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 It is a multi-scale convolution attention layer structure diagram;

[0027] Figure 2 It is a permutation attention structure diagram;

[0028] Figure 3 It is a ballistocardiogram sample schematic diagram;

[0029] Figure 4 It is a ballistocardiogram classification result schematic diagram. DETAILED DESCRIPTION

[0030] The present application will be described below in conjunction with the drawings.

[0031] A disease classification method based on multi-scale attention network and ballistocardiogram of the present application comprises the following steps:

[0032] 1.1 Constructing a multi-scale attention network: the multi-scale attention network is composed of a convolution module, a multi-scale attention module and a full connection module;

[0033] 1.1.1 The convolution module is composed of one-dimensional convolution unit, linear rectifier activation ReLU unit, batch normalization BN unit and maximum pooling unit which are sequentially stacked;

[0034] 1.1.2 Multi-scale attention module contains 8 multi-scale convolution attention layers, each of which is composed of a multi-scale convolution unit embedded in a permutation attention unit; the multi-scale convolution unit is a residual structure, and the input feature U is mapped to the feature map x ∈ R c×h×w , where c represents the number of feature map channels, h represents the height of the feature map, and w represents the width of the feature map; the feature map is evenly divided into three groups along the channel direction, each group containing c / 3 channels, i.e.: x = [x1, x2, x3], R c / 3×h×w . The first group of feature maps x1 inputs a dilated convolution with an expansion coefficient of 2, the second group of feature maps x2 inputs a 1x5 convolution, and the third group of feature maps x3 adds the second group of convolution outputs y2 and inputs a 1x5 convolution, obtaining three groups of intrinsic feature maps y1, y2, y3 of different scales, which are concatenated to obtain [y1, y2, y3] input a 1x1 convolution layer, and the different scale information of the three groups of intrinsic feature maps is fused to combine the global information and detailed features of the cardiac impact signal for disease classification;

[0035] The convolution layer output is the input of the permutation attention unit, which is divided into three parts: feature grouping, feature permutation, and feature fusion. The input is evenly divided into g groups along the channel number, i.e. V = [v1, v2, v3, …, v g ], R c / g×h×w Then, for each group, feature permutation is performed: first, each group is evenly divided into two groups along the channel number, obtaining v K1 , v K2 ∈ R C / 2g×h×w ; Channel attention processing is performed on v K1 , and global average pooling F gp is used to average each channel of v K1 to obtain

[0036]

[0037] The channel correlation is learned using a fully connected layer F c , and the feature selection is realized using the gate mechanism of the Sigmoid activation function σ to obtain v′ K1 = σ(F c (r))·v K1 = σ(W1r + b1)·v K1

[0038] Where: the weight W1 ∈ R c / 2g×1×1 , and the bias b1;

[0039] Spatial attention processing is performed on v K2 , and group normalization F gn is used to calculate vK2 Normalize it using the mean and variance, and use the fully connected layer F c Feature enhancement is performed and the result is mapped to v′ using the Sigmoid activation function. K2 =σ(F c (F gn (v K2 )))·v K2 ; The concat operator is used to concatenate v′ K1 , v′ K2 Combining, we get v′ K =[v′ K1 , v′ K2 ]∈R c / g×h×w The average of all groups yields a feature V′ with the same dimension as the input. Feature fusion involves channel shuffling of V′, dividing it into d groups of e channels each. Each group is stacked vertically to obtain a d×e matrix, with each channel as an element. The matrix is ​​transposed and rearranged using the reshape operator, taking e elements at a time and stacking all rows horizontally to obtain V″, which has the same dimension as V′. The channels between groups are now rearranged, enabling information exchange between channels. V″ is added to the input feature U to obtain the output of the multi-scale convolutional attention layer, enhancing the key signal features of the cardiac impaction image.

[0040] 1.1.3 The fully connected module contains a fully connected layer and a normalized exponential function Softmax, which gives the probability that the input cardiac impact map belongs to various diseases;

[0041] 1.2 Construction of the training set: Cardiac impaction images were collected from 40 volunteers using a piezoelectric thin film. Each cardiac impaction image sample contained C sampling points, denoted as {S}. i The participants included individuals with tachycardia, hypertension, atrial fibrillation, coronary artery disease, and healthy individuals. A label vector L was set. im This indicates that the i-th sample corresponds to the m-th class; finally, the training set {S} is generated. i ,L im}, i = 1, 2, ..., N, divided into a cardiac impaction training set and a cardiac impaction test set in a 4:1 ratio;

[0042] 1.3 Learning the parameters Θ of a multi-scale attention network using the cardiac impact map training set includes the following steps:

[0043] 1.3.1 Construct the cross-entropy loss function, the calculation formula of which is as follows:

[0044]

[0045] Where: L m For label vectors the mth element of the L m 1, and the rest of the elements are 0; log is a logarithmic function;

[0046] 1.3.2 Obtain the cost function using the set of ballistocardiogram training, and the calculation formula is:

[0047]

[0048] Wherein: N is the number of samples in the set of ballistocardiogram training;

[0049] 1.3.3 Use the batch processing method and the stochastic gradient descent method to optimize the cost function, and when the Γ(Θ) reaches the minimum value, the optimal multi-scale attention network parameter Θ and the trained multi-scale attention network f are obtained;

[0050] 1.4 Given the ballistocardiogram Y of the subject in the set of ballistocardiogram test, based on the trained multi-scale attention network, the given ballistocardiogram is classified, and whether the subject is healthy or has hypertension, tachycardia, atrial fibrillation, coronary artery disease is judged, including the following steps:

[0051] 1.4.1 Input the ballistocardiogram Y with a length of C, and the multi-scale attention network f predicts that the subject of the ballistocardiogram belongs to the probability of being healthy, having hypertension, tachycardia, atrial fibrillation, and coronary artery disease:

[0052] p m =f(Y, Θ), m = 1, 2, …, 5, wherein: Θ is the multi-scale attention network parameter;

[0053] 1.4.2 Calculate the maximum probability The ballistocardiogram is classified as the class with the maximum probability.

[0054] 1. Working conditions

[0055] The experimental platform of the application adopts an Intel(R) Core(TM) i5-7500 CPU@3.40GHz, an NVIDIA GTX1080Ti, a display memory of 16GB, a PC running Windows 10, and a language of python. The running environment is python == 3.7, torch == 1.0.1, scipy == 1.3.1, and matplotlib.

[0056] 2. Experimental content and result analysis

[0057] The experimental effect of the application will be illustrated by the experiment on the ballistocardiogram database as follows:

[0058] ​The experiment of the application uses an open source database (Bed-Based Ballistocardiography Dataset), and selects 40 heart impulse graphs of subjects in the database, each of which is labeled with health and tachycardia, hypertension, atrial fibrillation, and coronary artery disease. The trained multi-scale attention network is applied to 3325 samples in the test set, including 737 healthy samples, 662 hypertension samples, 641 tachycardia samples, 652 atrial fibrillation samples, and 643 coronary artery disease samples, each with a length of 10 seconds and a sampling frequency of 500Hz. Figure 3 A heart impulse graph sample diagram is given, and from the diagram, it can be seen that the heart impulse waveform characteristics are not obvious, and the respiratory signal and some interference components are superimposed, which brings great difficulty to disease recognition.

[0059] The method of the application is applied to the test set, and the classification results are shown in Table 1. As can be seen from Table 1, the application has fewer misclassified samples.

[0060] Table 1 Classification results of the method of the application

[0061] health hypertension tachycardia atrial fibrillation coronary artery disease health 736 0 0 0 1 hypertension 0 662 0 0 0 tachycardia 1 0 640 0 0 atrial fibrillation 1 0 0 651 0 coronary artery disease 0 0 0 1 642

[0062] In order to evaluate the classification effect, the application calculates the accuracy, recall rate, specificity and other indicators, and compares them with one-dimensional convolutional neural network and multilayer perception. For fair comparison, the training data set of the one-dimensional convolutional neural network and the multilayer perception is the same as that of the application, and the optimal hyperparameters are selected after multiple experiments. The comparison results are shown in Figure 4 . By comparing the classification results, the application has obvious advantages in the three evaluation indicators compared with the other two methods, which verifies the accuracy of the heart impulse graph disease classification in the application.

[0063] The heart impulse graph sample is added to white noise to verify the anti-interference ability of the application to noise. Table 2 shows the F1-score of each class when the signal-to-noise ratio is 15dB, which comprehensively evaluates the accuracy and recall rate. The data in Table 2 shows that compared with the other two methods, the F1 score of the classification result of the application is the largest, all above 90%, and significantly higher than the first two methods. This shows that the method of the application has better classification effect and good stability than the other two methods.

[0064] Table 2 Comparison of F1 scores of classification results of three methods

[0065] F1-score multi-layer perception network one-dimensional convolution network the invention health 69.5 89.8 93.3 hypertension 60 93.2 95.7 tachycardia 53.2 87.3 92.2 atrial fibrillation 50.8 66.9 90.1 coronary artery disease 63.7 87.9 92.3

Claims

1. A disease classification method based on a multi-scale attention network and a ballistocardiogram, characterized by comprising the following steps: 1.1 Constructing a multi-scale attention network: the multi-scale attention network is composed of a convolution module, a multi-scale attention module and a full connection module; 1.1.1 The convolution module is composed of a one-dimensional convolution unit, a linear rectifier activation ReLU unit, a batch normalization BN unit, and a maximum pooling unit stacked in sequence; 1.1.2 The multi-scale attention module comprises 8 multi-scale convolution attention layers, each of which is composed of a multi-scale convolution unit embedded with a permutation attention unit; the multi-scale convolution unit is of a residual structure, and an input feature U is mapped to a feature map x e R c×h×w wherein: c denotes the number of feature map channels, h denotes the height of the feature map, and w denotes the width of the feature map; the feature map is evenly divided into three groups along the channel direction, each group containing c / 3 channels, that is: x = [x1, x2, x3], R c / 3×h×w The first group of feature maps x1 is input into a dilated convolution with a dilation factor of 2, the second group of feature maps x2 is input into a 1x5 convolution, and the third group of feature maps x3 is added with the output y2 of the second group and input into a 1x5 convolution, to obtain three groups of intrinsic feature maps y1, y2, and y3 of different scales, which are input into a 1x1 convolution layer after being spliced by Concat, so as to fuse the information of the three groups of intrinsic feature maps of different scales, in order to combine the global information and the detailed features of the heart shock signal for disease classification. The convolutional layer outputs are input to a permutation attention unit, which is divided into three parts: feature grouping, feature permutation, and feature fusion. The feature grouping divides the input into g groups V = [v1, v2, v3, …, v g ], R c / g×h×w K1 K2 ∈ R C / 2g×h×w ; channel attention processing is performed on v K1 . Global average pooling F gp is used to average each channel of v K1 to obtain r = F gp (v K1 ). A fully connected layer F c is used to learn channel correlation and a Sigmoid activation function σ is used to realize feature selection to obtain v′ K1 = σ(F c (r))·v K1 . Spatial attention processing is performed on v K2 . Group normalization F gn is used to calculate the mean and variance of v K2 for normalization. A fully connected layer F c is used for feature enhancement and a Sigmoid activation function is used to map to v′ K2 = σ(F c (F gn (v K2 )))·v K2 . Concat splicing operator is used to combine v′ K1 and v′ K2 to obtain v′ K = [v′ K1 , v′ K2 ] ∈ R c / g×h×w , and all groups are averaged to obtain features V′ of the same dimension as the input.​​ The feature fusion performs channel mixing on V', divides it into d groups according to the number of channels, and each group has e channels. Each group is stacked in the vertical direction to obtain a d x e matrix. Each channel is an element in the matrix. After transposing the matrix, the reshape operator is used to rearrange the elements. Each time, e elements are arranged in a row, and all rows are stacked horizontally to obtain V'', which has the same dimension as V'. The channels between groups have been rearranged, realizing the information exchange between channels. V'' is added to the input feature U to obtain the output of the multi-scale convolution attention layer, enhancing the key signal features of the heart shock map; 1.1.3 The fully connected module includes a fully connected layer and a Softmax function, which gives the probability of the input heart shock map belonging to each disease; 1.2 Constructing the training set: 40 volunteers are used to collect ballistocardiogram samples by piezoelectric film, each ballistocardiogram sample contains C sampling points, denoted as {S i}, the subjects are respectively people with tachycardia, hypertension, atrial fibrillation, coronary artery disease and healthy people, and a label vector L im is set, indicating that the ith sample corresponds to the mth class; finally, the training set {S i ,L im}, i = 1, 2, …, N, is generated, and is divided into a ballistocardiogram training set and a ballistocardiogram test set in a ratio of 4:

1. 1.3 Learning the multi-scale attention network parameters Θ using the heart shock map training set, including the following steps: 1.3.1 Construct a cross-entropy loss function, whose calculation formula is: wherein: L m is the label vector is the m-th element of the label vector L, the ballistocardiogram sample S belongs to the m-th class if L m is 1, and the rest of the elements are 0; log is the logarithm function; 1.3.2 Obtain the cost function using the heart shock map training set, whose calculation formula is: Wherein: N is the number of samples in the heart shock map training set; P im is the probability that the i-th sample predicted by the multi-scale attention network belongs to the m-th class, L im is the label of the i-th sample; 1.3.3 Optimize the cost function using the batch processing method and the stochastic gradient descent method. When Γ(Θ) reaches the minimum value, the optimal multi-scale attention network parameters Θ and the trained multi-scale attention network f are obtained; 1.4 Given the heart shock map Y of the subject in the heart shock map test set, classify the given heart shock map based on the trained multi-scale attention network, and determine whether the subject is healthy or has hypertension, tachycardia, atrial fibrillation, and coronary artery disease, including the following steps: 1.4.1 Input the heart shock map Y with a length of C. The multi-scale attention network f predicts the probability of the heart shock map belonging to healthy, having hypertension, tachycardia, atrial fibrillation, and coronary artery disease: p m = f(Y, Q), m = 1, 2,..., 5, where: Q is a multi-scale attention network parameter; 1.4.2 Computing the Maximum Probability The cardiogram is classified as the class m with the maximum probability.

Images