A multi-classification detection method and system for arrhythmia based on pulse wave

Abstract

The present invention provides a pulse wave-based arrhythmia multi-classification detection method and system, which relate to the technical field of intelligent arrhythmia classification. A photoelectric volumetric pulse wave signal is processed by an arrhythmia type prediction model to obtain the type of arrhythmia. A waveform space feature extraction layer of the arrhythmia type prediction model converts a multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional space feature matrix through a plurality of groups of sequential effective size convolution kernels for a signal transmitted by a pulse wave signal acquisition device. A waveform time feature extraction layer receives the two-dimensional space feature matrix and outputs a multi-dimensional time feature vector in combination with a learning time dependency. An event occurrence prediction layer receives the multi-dimensional time feature vector and predicts the type of arrhythmia occurrence through a multi-layer perceptron. The photoelectric volumetric pulse wave signal is input into an end-to-end deep learning model. Online data processing and analysis predict the type of arrhythmia occurrence, thereby effectively improving the accuracy of arrhythmia classification results.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent arrhythmia classification, and in particular to a pulse wave-based arrhythmia multi-classification detection method and system. Background Art

[0002] Arrhythmia is a cardiovascular disease that seriously endangers human health. The "Summary of China's Cardiovascular Health and Disease Report 2019" shows that cardiovascular disease mortality still ranks first among the causes of death among Chinese residents. Arrhythmia can cause certain pathological changes in the patient's heart's basic function and tissue structure. Therefore, timely and accurate identification and intervention of arrhythmia can effectively protect the patient's life safety and has great practical significance.

[0003] Currently, conventional arrhythmia detection is still based on ECG. However, the ECG-based method has cumbersome signal acquisition and complex and difficult result analysis. Summary of the invention

[0004] The present invention discloses a pulse wave-based arrhythmia multi-classification detection method and system, which solves the problems of cumbersome signal collection and complex and difficult result analysis of the commonly used EGG method for existing arrhythmia detection. It collects short-time photoplethysmogram and deeply captures the spatiotemporal dual characteristics of the photoplethysmogram through an end-to-end learning model, accurately establishes the association between arrhythmia and the entire short-time signal, obtains multi-classification results with high accuracy, and has a simple data preprocessing process.

[0005] To achieve the above object, the technical solution of the present invention is specifically implemented as follows:

[0006] The present invention discloses a method for detecting arrhythmia by multi-classification based on pulse wave, comprising the following steps:

[0007] S1: Collecting photoelectric volumetric pulse wave signals from a target object through a pulse wave signal acquisition device to form an original signal data set, and recording the arrhythmia type of the corresponding target object as an original sample label, recording individual information to form an original data set, and transmitting the original signal data set to a data preprocessing unit;

[0008] S2: The data preprocessing unit receives the original data set, preprocesses the original data set, and divides the data set into a training data set, a verification data set, and a test data set;

[0009] S3: The pulse wave signal of the training data set obtained in step S2 is input into the arrhythmia type prediction model. The waveform space feature extraction layer converts the multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional space feature matrix through multiple groups of convolution kernels of sequential effective sizes on the pulse wave signal of the training data set; the waveform time feature extraction layer receives the two-dimensional space feature matrix, and outputs a multi-dimensional time feature vector in combination with the learning time dependency; the event occurrence prediction layer receives the multi-dimensional time feature vector, and predicts and classifies the arrhythmia occurrence type of the training data set through a multi-layer perceptron, and repeats the above steps in S3 for iterative training until the optimal arrhythmia type prediction model is obtained;

[0010] S4: Input the test data set in step S2 to the optimal arrhythmia type prediction model obtained in step S3 to obtain the arrhythmia type prediction value of the target object.

[0011] Furthermore, the specific implementation of the data preprocessing unit preprocessing the original data set includes:

[0012] S21: removing abnormal fragments from the final data set output in step S1;

[0013] S22: Detrending the data set after removing the abnormal segments in step S21;

[0014] S23: normalizing the data set after detrending in step S22;

[0015] S24: Divide the data set after the normalization process in step S23 into a training data set, a verification data set and a test data set.

[0016] Furthermore, in step S3, the step of iteratively training the arrhythmia type prediction model includes:

[0017] S31: inputting the training data set into the arrhythmia type prediction model for training;

[0018] S32: the training data set enters the spatial feature extraction layer of the arrhythmia type prediction model;

[0019] S33: The spatial feature extraction layer converts the multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional spatial feature matrix through multiple groups of sequential effective size convolution kernels on the pulse wave signal of the training data set, and then transmits the multi-layer two-dimensional spatial feature matrix to the temporal feature extraction layer;

[0020] S34: the time feature extraction layer receives the multi-layer two-dimensional spatial feature matrix, and combines the multi-layer two-dimensional spatial feature matrix with the learned time dependency to output a multi-dimensional time feature vector, and transmits it to the event occurrence prediction layer;

[0021] S35: The event occurrence prediction layer receives the multi-dimensional time feature vector and predicts the arrhythmia type of the sample in the training data set input in S31 through a multi-layer perceptron;

[0022] S36: Calculate weighted loss function;

[0023] S37: back propagating and updating the arrhythmia type prediction model parameters;

[0024] S38: inputting a validation data set into the updated arrhythmia type prediction model for validation, and calculating a weighted loss function;

[0025] S39: Determine whether the loss function of the validation data set no longer decreases after a certain number of rounds. If not, repeat the above steps S31-S39. Otherwise, the best arrhythmia type prediction model is obtained.

[0026] Furthermore, in step S3, the waveform space feature extraction layer is stacked by a number of one-dimensional convolutional feature extraction modules, and a random dropout layer is added at the end of the last one-dimensional convolutional feature extraction module. Each of the one-dimensional convolutional feature extraction modules is a convolution-layer normalization-activation layer structure, and the specific processing method is as follows:

[0027] Several [1, L] one-dimensional data are input into the first one-dimensional convolution feature extraction module;

[0028] According to the one-dimensional convolution formula:

[0029]

[0030] Use m convolution kernels of size [1, K], combined with the step size s and the number of zero values ​​p filled at the edge of the input, convolve the input data in order to obtain the output features of the convolution layer;

[0031] The size of the output feature of the convolutional layer is [m, (L-k+2*p) / s+1], and the obtained deep two-dimensional feature matrix is ​​length normalized;

[0032] Then add a random bias v to each element of the matrix, and obtain the module output matrix X through nonlinear activation;

[0033] Among them, L is the data length and K is the convolution kernel size.

[0034] Furthermore, in step S3, the waveform time feature extraction layer is composed of multiple unidirectional long short-term memory network feature extraction modules stacked together, and the input elements of each unidirectional long short-term memory network feature extraction module must go through three stages of memory cell selection, forgetting, and output. The specific expression is as follows:

[0035] C t =Zf ⊙C t-1 +Z i ⊙Z;

[0036] h t =Z o ⊙tanh(C t );

[0037] y t =(W′⊙h t );

[0038] Among them, Z f , Z i , Z o are the gating states corresponding to the three stages respectively, and t and t-1 represent the current and previous states respectively.

[0039] Furthermore, the weighting matrix of the weighted loss function in step S36 is determined by the frequency matrix of the input set. Assuming that the arrhythmia type and quantity of the input are (x1, n1), (x2, n2), ..., (x i ,n i ), among which the largest category is (x max ,n max ), then the weight vector is n i / n max They are concatenated in sequence, where i ≥ 3.

[0040] Another aspect of the present invention discloses a detection system applied to the above-mentioned arrhythmia multi-classification detection method based on pulse wave, comprising:

[0041] A pulse wave signal acquisition device, which acquires a photoplethysmographic pulse wave signal of a target object and transmits the acquired signal to a data preprocessing unit;

[0042] A data preprocessing unit, which receives the signal transmitted by the pulse wave signal acquisition device, preprocesses the received signal, and transmits the preprocessed signal to the arrhythmia type prediction model;

[0043] The arrhythmia type prediction model receives the signal transmitted by the data preprocessing unit, and predicts the arrhythmia type of the target object according to the received signal.

[0044] Furthermore, the arrhythmia type prediction model includes:

[0045] A waveform spatial feature extraction layer, which converts a multi-period one-dimensional pulse wave signal into a multi-layer two-dimensional spatial feature matrix by passing a plurality of groups of sequentially effective-size convolution kernels through the signal transmitted by the pulse wave signal acquisition device;

[0046] A waveform time feature extraction layer, which receives a two-dimensional spatial feature matrix and outputs a multi-dimensional time feature vector in combination with a learned time dependency;

[0047] The event occurrence prediction layer receives the multi-dimensional time feature vector and predicts and classifies the arrhythmia occurrence type of the training data set through a multi-layer perceptron.

[0048] Furthermore, the pulse wave signal acquisition device comprises:

[0049] A photoelectric sensor for collecting photoplethysmographic signals of a target object in real time;

[0050] A signal transmitter, the signal transmitter is used to transmit the photoelectric volume pulse wave signal collected by the photoelectric sensor to a preprocessing unit;

[0051] The pressure-adaptive strap is used to apply a certain pressure to the target object to ensure that the photoelectric sensor can collect a suitable photoplethysmogram signal.

[0052] Beneficial technical effects:

[0053] 1. The present invention discloses a multi-classification detection method for arrhythmia based on pulse wave, which uses an arrhythmia type prediction model to process the collected photoelectric volume pulse wave signal to obtain the type of arrhythmia. The waveform space feature extraction layer of the arrhythmia type prediction model converts the multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional space feature matrix through multiple groups of sequential effective size convolution kernels for the signal transmitted by the pulse wave signal acquisition device. The waveform time feature extraction layer receives the two-dimensional space feature matrix and outputs a multi-dimensional time feature vector in combination with the learning time dependency. The event occurrence prediction layer receives the multi-dimensional time feature vector and predicts and classifies the arrhythmia occurrence type of the training data set through a multi-layer perceptron, which effectively improves the accuracy of the arrhythmia classification result, and the acquisition method of the photoelectric volume pulse wave signal is simple.

[0054] 2. The arrhythmia type prediction model used in the present invention is an end-to-end deep learning model, which directly learns the mapping relationship between input and output, bypasses many processing steps, can be automatically adjusted according to data, and saves prediction costs. In addition, the system can perform online analysis and processing on the input new PPG waveform segment based on the updated normalization standard and model parameters that have been trained and saved, and obtain instant arrhythmia prediction type results, which can greatly improve the efficiency of auxiliary diagnosis processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solution of the present invention, the drawings required for describing the embodiments are briefly introduced below.

[0056] Figure 1 A flowchart of the steps of a pulse wave-based arrhythmia multi-classification detection method disclosed in the present invention;

[0057] Figure 2 PPG waveforms for atrial fibrillation, tachycardia, premature beats, and no arrhythmia;

[0058] Figure 3 It is a structural schematic diagram of an arrhythmia type prediction model in a pulse wave-based arrhythmia multi-classification detection system according to the present invention;

[0059] Figure 4a It is a curve chart of the loss values ​​of the training set and the validation set;

[0060] Figure 4b is the f1-score value of the training set and the validation set;

[0061] Figure 4c This is the confusion matrix result diagram of the test set;

[0062] Figure 5 It is a structural schematic diagram of a pulse wave-based arrhythmia multi-classification detection system according to the present invention;

[0063] Figure 6 The present invention provides a schematic structural diagram of a pulse wave signal acquisition device in a pulse wave-based arrhythmia multi-classification detection system. DETAILED DESCRIPTION

[0064] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0065] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0066] Unless otherwise specifically stated, the relative arrangement of the parts and steps described in these embodiments, numerical expressions and numerical values ​​do not limit the scope of the present invention. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the accompanying drawings are not drawn according to the actual proportional relationship. The technology, method and equipment known to ordinary technicians in the relevant field may not be discussed in detail, but in appropriate cases, the technology, method and equipment should be regarded as a part of the authorization specification. In all examples shown and discussed here, any specific value should be interpreted as being merely exemplary, rather than as a limitation. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters represent similar items in the following drawings, so once a certain item is defined in an accompanying drawing, it does not need to be further discussed in subsequent drawings.

[0067] In the description of the present invention, it is necessary to understand that the directions or positional relationships indicated by directional words such as "front, back, up, down, left, right", "lateral, vertical, perpendicular, horizontal" and "top, bottom" are usually based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description. Unless otherwise specified, these directional words do not indicate or imply that the devices or elements referred to must have a specific direction or be constructed and operated in a specific direction. Therefore, they cannot be understood as limiting the scope of protection of the present invention. The directional words "inside and outside" refer to the inside and outside relative to the contours of each component itself.

[0068] For ease of description, spatially relative terms such as "above", "above", "on the upper surface of", "above", etc. may be used here to describe the spatial positional relationship between a device or feature and other devices or features as shown in the figure. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figure. For example, if the device in the accompanying drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" can include both "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatially relative descriptions used here are interpreted accordingly.

[0069] In addition, it should be noted that the use of terms such as "first" and "second" to limit components is only for the convenience of distinguishing the corresponding components. If not otherwise stated, the above terms have no special meaning and therefore cannot be understood as limiting the scope of protection of the present invention.

[0070] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0071] On the one hand, the present invention discloses a multi-classification detection method of arrhythmia based on pulse wave, see Figure 1 , specifically including the following steps:

[0072] S1: Collecting photoelectric volumetric pulse wave signals from a target object through a pulse wave signal acquisition device to form an original signal data set, and recording the arrhythmia type of the corresponding target object as an original sample label, recording individual information to form an original data set, and transmitting the original signal data set to a data preprocessing unit;

[0073] Specifically, the pulse wave signal acquisition device collects N copies of photoelectric volumetric pulse wave signals of not less than 10 seconds to form an original signal data set as a sample, and simultaneously records the arrhythmia type of the corresponding target object as a sample label during the acquisition. Preferably, there are not less than three types of sample labels. In the embodiment of the present invention, there are four types of sample labels, namely atrial fibrillation, tachycardia, premature beats and no arrhythmia. There are slight differences in PPG waveforms. For example, see Figure 2 ; The pulse wave signal acquisition device sends the data set to the signal preprocessing unit;

[0074] S2: The data preprocessing unit receives the original data set, preprocesses the original data set, and divides the data set into a training data set, a verification data set, and a test data set;

[0075] Specifically, the data preprocessing unit preprocesses the original data set in the following ways:

[0076] S21: removing abnormal fragments from the final data set output in step S1;

[0077] The abnormal segment referred to in the present invention is that the amplitude of the next second signal suddenly changes to less than 1 / 2 times or more than 2 times the amplitude of the previous second signal, or a non-periodic interference signal segment appears;

[0078] S22: Detrending the data set after removing the abnormal segments in step S21;

[0079] In the embodiment of the present invention, detrending is achieved based on a sliding average process with a window length of 1 second;

[0080] S23: normalizing the data set after detrending in step S22;

[0081] In the embodiment of the present invention, normalization is achieved by min-max standardization so that the amplitude falls within the interval [0,1];

[0082] S24: Divide the data set after the normalization process in step S23 into a training data set, a verification data set and a test data set.

[0083] It is understandable that data enhancement can be performed when the number of samples is insufficient to enhance the generalization ability of the network. In one embodiment of the present invention, after waveform preprocessing, the numbers of waveform samples of atrial fibrillation, tachycardia, premature beats and no arrhythmia are 1953, 1451, 361 and 1151 respectively. Combined with the label frequency, the number of premature beat samples is increased to 1083 by adding random noise with a variance of 0.005, and the new sample set is divided into training, validation and test sets.

[0084] S3: The pulse wave signal of the training data set obtained in step S2 is input into the arrhythmia type prediction model. The waveform space feature extraction layer converts the multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional space feature matrix through multiple groups of convolution kernels of sequential effective sizes on the pulse wave signal of the training data set; the waveform time feature extraction layer receives the two-dimensional space feature matrix, and outputs a multi-dimensional time feature vector in combination with the learning time dependency; the event occurrence prediction layer receives the multi-dimensional time feature vector, and predicts and classifies the arrhythmia occurrence type of the training data set through a multi-layer perceptron, and repeats the above steps in S3 for iterative training until the optimal arrhythmia type prediction model is obtained;

[0085] Specifically, in step S3, the step of iteratively training the arrhythmia type prediction model includes:

[0086] S31: inputting the training data set into the arrhythmia type prediction model for training;

[0087] S32: the training data set enters the spatial feature extraction layer of the arrhythmia type prediction model;

[0088] S33: The spatial feature extraction layer converts the multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional spatial feature matrix through multiple groups of sequential effective size convolution kernels on the pulse wave signal of the training data set, and then transmits the multi-layer two-dimensional spatial feature matrix to the temporal feature extraction layer;

[0089] Specifically, the waveform space feature extraction layer is composed of several one-dimensional convolution feature extraction modules stacked together, and a random dropout layer is added at the end of the last one-dimensional convolution feature extraction module. Each of the one-dimensional convolution feature extraction modules is a convolution-layer normalization-activation layer structure. The specific processing method is:

[0090] Several [1, L] one-dimensional data are input into the first one-dimensional convolution feature extraction module;

[0091] According to the one-dimensional convolution formula:

[0092]

[0093] Use m convolution kernels of size [1, K], combined with the step size s and the number of zero values ​​p filled at the edge of the input, convolve the input data in order to obtain the output features of the convolution layer;

[0094] The size of the output feature of the convolutional layer is [m, (L-k+2*p) / s+1], and the obtained deep two-dimensional feature matrix is ​​length normalized;

[0095] Then add a random bias v to each element of the matrix, and obtain the module output matrix X through nonlinear activation;

[0096] Among them, L is the data length and K is the convolution kernel size;

[0097] In the embodiment of the present invention, the structure of the waveform spatial feature extraction layer is composed of four one-dimensional convolution modules connected in series, see Figure 3 The first three modules are composed of one-dimensional convolutional layers, layer normalization and activation functions respectively. The last module adds a random dropout layer with a dropout probability of 0.5. The stride and edge padding zero value of each module are 1. The one-dimensional convolutional layers of the four modules are composed of 32, 64, 128, and 256 convolution kernels of size [1,,3] respectively. The number of convolutions increases in multiples of 2, and the activation function is set to the rectified linear unit ReLU.

[0098] S34: the time feature extraction layer receives the multi-layer two-dimensional spatial feature matrix, and combines the multi-layer two-dimensional spatial feature matrix with the learned time dependency to output a multi-dimensional time feature vector, and transmits it to the event occurrence prediction layer;

[0099] Specifically, the waveform time feature extraction layer is composed of multiple unidirectional long short-term memory network feature extraction modules stacked together, and the input elements of each unidirectional long short-term memory network feature extraction module must go through three stages: memory cell selection, forgetting, and output. The specific expression is as follows:

[0100] C t =Z f ⊙C t-1 +Z i ⊙Z;

[0101] h t =Z o ⊙tanh(C t );

[0102] y t =(W′⊙h t );

[0103] Among them, Z f, Z i , Z o are the gating states corresponding to the three stages, t and t-1 represent the current and previous states respectively;

[0104] In the embodiment of the present invention, the structure of the waveform time feature extraction layer is a stack of two unidirectional long short-term memory network module extraction modules, see Figure 3 , the module structure remains consistent, including input gate, forget gate, memory cell and output gate. The input of the first module is connected to the output of the waveform space feature extraction layer. The LSTM input is a batch of one-dimensional vectors with a feature dimension of 256. The input gate, forget gate and output gate of LSTM can control the flow of information. The forget gate determines the information of the cell state at the previous moment retained to this moment. The input gate determines the content of the network input saved to the cell state at the next moment. The output gate controls the number of output values ​​output from the cell state to this moment. The memory cell stores all the information of the calculation process. The dimension of the hidden layer state of the two long short-term memory networks in this example is 128. The output information of the last step of the waveform time feature extraction layer is used as the input of the fully connected layer, and is mapped to the output space to obtain the prediction probability matrix after nonlinear changes with Relu as the activation function.

[0105] S35: The event occurrence prediction layer receives the multi-dimensional time feature vector and predicts the arrhythmia type of the sample in the training data set input in S31 through a multi-layer perceptron;

[0106] S36: Calculate weighted loss function;

[0107] Specifically, the weight matrix of the weighted loss function is determined by the frequency matrix of the input set. Assume that the arrhythmia type and number of the input are (x1,n1),(x2,n2),...,(x i ,n i ), among which the largest category is (x max ,n max ), then the weight vector is n i / n max They are spliced ​​in order, where i ≥ 3;

[0108] In the embodiment of the present invention, the one-dimensional weighted matrix of atrial fibrillation, tachycardia, premature beats and no arrhythmia is W=[1, 1.71, 1.80, 1.69] according to the number of each type of labels; the weighted multi-classification cross entropy loss function is as follows:

[0109]

[0110] where p=[p0,p1,...p C-1 ] is a probability distribution, each element p iIndicates the probability that the sample belongs to the i-th category, y=[y0,y1,...y C-1 ] is the onehot representation of the sample label. When the sample belongs to the i-th category, y i =1 otherwise 0, C is the sample label, in this example C = 3. The weighted cross entropy loss of the training batch will be back-propagated by the Adam algorithm, and the model parameters will be updated layer by layer to narrow the gap between the model prediction and the label.

[0111] It should be noted that, considering the category imbalance problem in the actual collected data, the weighted and loss function matrices are multiplied to achieve the effect of balanced compensation.

[0112] S37: back propagating and updating the arrhythmia type prediction model parameters;

[0113] S38: inputting a validation data set into the updated arrhythmia type prediction model for validation, and calculating a weighted loss function;

[0114] S39: Determine whether the loss function of the validation data set no longer decreases after a certain number of rounds. If not, repeat the above steps S31-S39. Otherwise, the best arrhythmia type prediction model is obtained.

[0115] Preferably, the arrhythmia type prediction model selected in the embodiment of the present invention is obtained by connecting four layers of one-dimensional convolution feature extraction modules and two layers of unidirectional long short-term memory network feature extraction modules in series, and the convolution kernel size of the former increases successively from 32, 64, 128, and 256.

[0116] In the embodiment of the present invention, the training batch size of the arrhythmia type prediction model is set to 32, the number of iterations is set to 200, the early stopping coefficient is set to 25, and the weight decay term is set to 1×10 -4 , the learning rate is initialized to 10 -3 After the hyperparameters are set, training can begin. When the training loss and validation loss values ​​decrease at the same time during model training, it means that the network prediction ability is gradually increasing. When the training loss and validation loss values ​​no longer decrease, it means that the network has converged and the training can be ended.

[0117] For the change of loss values ​​of the training set and the validation set in the embodiment of the present invention, see Figure 4a , the changes in f1-score values ​​of the training set and the validation set can be found in Figure 4b , both decrease at the same time, indicating that the network prediction ability is gradually increasing. When both stop decreasing, it means that the network has converged and the training can be ended. The ordinate is the loss value and the abscissa is the number of iterations. It can be seen that when the number of iterations exceeds 140, the loss value almost stops decreasing, and the f1-score of the prediction result exceeds 0.8. The confusion matrix results of the test set can be found in Figure 4c, the model performs well.

[0118] S4: Input the test data set in step S2 to the optimal arrhythmia type prediction model obtained in step S3 to obtain the arrhythmia type prediction value of the target object.

[0119] In an embodiment, the arrhythmia type prediction model input is 5332 preprocessed 10-second photoplethysmograms, including 4265 training set data and 1067 test set data. The arrhythmia type prediction model outputs the arrhythmia category label corresponding to the data to be classified. The possible category results are atrial fibrillation, tachycardia, premature beats and no arrhythmia.

[0120] It should be noted that the arrhythmia type prediction model in the present invention is an end-to-end deep learning model, which can perform online processing and analysis on the input photoelectric volume pulse waveform according to the saved and updateable normalization standard and model parameters to obtain the real-time arrhythmia prediction category without repeating the training process, thereby greatly improving the efficiency of auxiliary diagnosis processing.

[0121] Another aspect of the present invention discloses a multi-classification detection system for arrhythmia based on pulse wave, see Figure 5 , including a pulse wave signal acquisition device, a data preprocessing unit and an arrhythmia type prediction model. Specifically, the pulse wave signal acquisition device acquires the photoelectric volume pulse wave signal of the target object and transmits the acquired signal to the data preprocessing unit. As a preferred embodiment of the present invention, the pulse wave signal acquisition device includes a photoelectric sensor, a signal transmitter and a pressure adaptive strap, see Figure 6, wherein the photoelectric sensor is used to collect the photoelectric volumetric pulse wave signal of the target object in real time; the signal transmitter is used to transmit the photoelectric volumetric pulse wave signal collected by the photoelectric sensor to the preprocessing unit; the pressure adaptive bandage is used to apply a certain pressure to the target object to ensure that the photoelectric sensor can collect a suitable photoelectric volumetric pulse wave signal; the data preprocessing unit receives the signal transmitted by the pulse wave signal acquisition device and preprocesses the received signal, and transmits the preprocessed signal to the arrhythmia type prediction model; the arrhythmia type prediction model receives the signal transmitted by the data preprocessing unit, and predicts the arrhythmia of the target object according to the received signal. Arrhythmia type, as a preferred embodiment of the present invention, the arrhythmia type prediction model includes a waveform space feature extraction layer, a waveform time feature extraction layer and an event occurrence prediction layer, wherein the waveform space feature extraction layer converts the multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional space feature matrix through multiple groups of sequential effective size convolution kernels for the signal transmitted by the pulse wave signal acquisition device; the waveform time feature extraction layer receives the two-dimensional space feature matrix, and outputs a multi-dimensional time feature vector in combination with the learning time dependency; the event occurrence prediction layer receives the multi-dimensional time feature vector, and predicts and classifies the arrhythmia occurrence type for the training data set through a multi-layer perceptron.

[0122] The present invention discloses a pulse wave-based arrhythmia multi-classification detection method and system, which analyze data of continuously collected short-time photoelectric volume pulse waves, and the collection method is simple; and the learning features are not the traditional numerical features extracted from segmented periodic signals, which ensures the objectivity of the prediction; the proposed arrhythmia type prediction model can deeply capture the dual temporal and spatial characteristics of the photoelectric volume pulse wave, and accurately establish the association between the arrhythmia and the entire short-time signal, so that the obtained arrhythmia multi-classification results are accurate and effective.

[0123] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0124] The above embodiments are only descriptions of the preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary engineering and technical personnel in the field should fall within the protection scope determined by the claims of the present invention.

Claims

1. A multi-classification detection method for arrhythmia based on pulse wave, characterized in that: The following steps are involved: S1: Collecting photoelectric volumetric pulse wave signals from a target object through a pulse wave signal acquisition device to form an original signal data set, and recording the arrhythmia type of the corresponding target object as an original sample label, recording individual information to form an original data set, and transmitting the original signal data set to a data preprocessing unit; S2: The data preprocessing unit receives the original data set, preprocesses the original data set, and divides the data set into a training data set, a verification data set, and a test data set; S3: The pulse wave signal of the training data set obtained in step S2 is input into the arrhythmia type prediction model. The waveform space feature extraction layer converts the multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional space feature matrix through multiple groups of convolution kernels of sequential effective sizes on the pulse wave signal of the training data set; the waveform time feature extraction layer receives the two-dimensional space feature matrix, and outputs a multi-dimensional time feature vector in combination with the learning time dependency; the event occurrence prediction layer receives the multi-dimensional time feature vector, and predicts and classifies the arrhythmia occurrence type of the training data set through a multi-layer perceptron, and repeats the above steps in S3 for iterative training until the optimal arrhythmia type prediction model is obtained; In step S3, the step of iteratively training the arrhythmia type prediction model includes: S31: inputting the training data set into the arrhythmia type prediction model for training; S32: the training data set enters the spatial feature extraction layer of the arrhythmia type prediction model; S33: The spatial feature extraction layer converts the multi-cycle one-dimensional pulse wave signal into a multi-layer two-dimensional spatial feature matrix through multiple groups of sequential effective size convolution kernels on the pulse wave signal of the training data set, and then transmits the multi-layer two-dimensional spatial feature matrix to the temporal feature extraction layer; The waveform space feature extraction layer is composed of several one-dimensional convolution feature extraction modules stacked together, and a random dropout layer is added at the end of the last one-dimensional convolution feature extraction module. Each of the one-dimensional convolution feature extraction modules is a convolution-layer normalization-activation layer structure. The specific processing method is as follows: Several [1, L] one-dimensional data are input into the first one-dimensional convolution feature extraction module; According to the one-dimensional convolution formula: Use m convolution kernels of size [1, K], combined with the step size s and the number of zero values ​​p filled at the edge of the input, convolve the input data in order to obtain the output features of the convolution layer; The size of the output feature of the convolutional layer is [m, (L-k+2*p) / s+1], and the obtained deep two-dimensional feature matrix is ​​length normalized; Then add a random bias v to each element of the matrix, and obtain the module output matrix X through nonlinear activation; Among them, L is the data length and K is the convolution kernel size; Specifically, the structure of the waveform spatial feature extraction layer is composed of four one-dimensional convolutional modules connected in series. The first three modules are composed of one-dimensional convolutional layers, layer normalization, and activation functions in turn. The last module adds a neuron random dropout layer with a dropout probability of 0.

5. The step size and edge padding zero value of each module are both 1. The one-dimensional convolutional layers of the four modules are composed of 32, 64, 128, and 256 convolution kernels of size [1,3] in turn. S34: the time feature extraction layer receives the multi-layer two-dimensional spatial feature matrix, and combines the multi-layer two-dimensional spatial feature matrix with the learned time dependency to output a multi-dimensional time feature vector, and transmits it to the event occurrence prediction layer; S35: The event occurrence prediction layer receives the multi-dimensional time feature vector and predicts the arrhythmia type of the sample in the training data set input in S31 through a multi-layer perceptron; S36: Calculate weighted loss function; The weighting matrix of the weighted loss function in step S36 is determined by the frequency matrix of the input set. Assume that the arrhythmia types and quantities of the input are (x1, n1), (x2, n2), ..., (x i ,n i ), among which the largest category is (x max ,n max ), then the weight vector is n i / n max They are spliced ​​in order, where i ≥ 3; The weighted multi-class cross entropy loss function is calculated as follows: where p=[p0,p1,...p c-1 ] is a probability distribution, each element p i Indicates the probability that the sample belongs to the i-th category, y=[y0,y1,...y c-1 ] is the onehot representation of the sample label. When the sample belongs to the i-th category, y i =1 otherwise 0, C is the sample label; W is the one-dimensional weighting matrix; S37: back propagating and updating the arrhythmia type prediction model parameters; S38: inputting a validation data set into the updated arrhythmia type prediction model for validation, and calculating a weighted loss function; S39: Determine whether the loss function of the validation data set no longer decreases after a certain number of rounds. If not, repeat the above steps S31-S39. Otherwise, obtain the best arrhythmia type prediction model. S4: inputting the test data set in step S2 into the best arrhythmia type prediction model obtained in step S3 to obtain the arrhythmia type prediction value of the target object; The input of the arrhythmia type prediction model is any 10-second photoelectric volume pulse wave, and the output of the arrhythmia type prediction model is one of the common arrhythmia subtypes: atrial fibrillation, tachycardia, premature beats, and no arrhythmia; The arrhythmia type prediction model is obtained by connecting four layers of one-dimensional convolution feature extraction modules and two layers of unidirectional long short-term memory network feature extraction modules in series; During the training process, the training batch size of the arrhythmia type prediction model was set to 32, the number of iterations was set to 200, the early stopping coefficient was set to 25, and the weight decay term was set to 1×10 -4 , the learning rate is initialized to 10 -3 .

2. The method for detecting arrhythmia based on pulse wave according to claim 1, characterized in that: The specific implementation of the data preprocessing unit preprocessing the original data set includes: S21: removing abnormal fragments from the final data set output in step S1; S22: Detrending the data set after removing the abnormal segments in step S21; S23: normalizing the data set after detrending in step S22; S24: Divide the data set after the normalization process in step S23 into a training data set, a verification data set and a test data set.

3. The method for detecting arrhythmia based on pulse wave according to claim 1, characterized in that: In step S3, the waveform time feature extraction layer is composed of multiple unidirectional long short-term memory network feature extraction modules stacked together, and the input elements of each unidirectional long short-term memory network feature extraction module must go through three stages: memory cell selection, forgetting, and output. The specific expression is as follows: C y =Z f ⊙C t-1 +Z i ⊙Z; h t =Z o ⊙tanh(C t ); y t =(W′⊙h t ); Among them, Z f , Z i , Z o are the gating states corresponding to the three stages respectively, and t and t-1 represent the current and previous states respectively.

4. A detection system applied to the pulse wave-based arrhythmia multi-classification detection method as described in any one of claims 1 to 3, characterized in that: include: A pulse wave signal acquisition device, which acquires a photoplethysmographic pulse wave signal of a target object and transmits the acquired signal to a data preprocessing unit; A data preprocessing unit, which receives the signal transmitted by the pulse wave signal acquisition device, preprocesses the received signal, and transmits the preprocessed signal to the arrhythmia type prediction model; The arrhythmia type prediction model receives the signal transmitted by the data preprocessing unit, and predicts the arrhythmia type of the target object according to the received signal.

5. The detection system of the pulse wave-based arrhythmia multi-classification detection method according to claim 4, characterized in that: The arrhythmia type prediction model includes: A waveform spatial feature extraction layer, which converts a multi-period one-dimensional pulse wave signal into a multi-layer two-dimensional spatial feature matrix by passing a plurality of groups of sequentially effective-size convolution kernels through the signal transmitted by the pulse wave signal acquisition device; A waveform time feature extraction layer, which receives a two-dimensional spatial feature matrix and outputs a multi-dimensional time feature vector in combination with a learned time dependency; The event occurrence prediction layer receives the multi-dimensional time feature vector and predicts and classifies the arrhythmia occurrence type of the training data set through a multi-layer perceptron.

6. The detection system of the pulse wave-based arrhythmia multi-classification detection method according to claim 5, characterized in that: The pulse wave signal acquisition device comprises: A photoelectric sensor for collecting photoplethysmographic signals of a target object in real time; A signal transmitter, the signal transmitter is used to transmit the photoelectric volume pulse wave signal collected by the photoelectric sensor to a preprocessing unit; The pressure-adaptive strap is used to apply a certain pressure to the target object to ensure that the photoelectric sensor can collect a suitable photoplethysmogram signal.

Images