[go: up one dir, main page]
More Web Proxy on the site http://driver.im/

WO2019119518A1 - 房颤信号的识别方法、装置和设备 - Google Patents

房颤信号的识别方法、装置和设备 Download PDF

Info

Publication number
WO2019119518A1
WO2019119518A1 PCT/CN2017/120106 CN2017120106W WO2019119518A1 WO 2019119518 A1 WO2019119518 A1 WO 2019119518A1 CN 2017120106 W CN2017120106 W CN 2017120106W WO 2019119518 A1 WO2019119518 A1 WO 2019119518A1
Authority
WO
WIPO (PCT)
Prior art keywords
signal
atrial fibrillation
network
layer
accuracy
Prior art date
Application number
PCT/CN2017/120106
Other languages
English (en)
French (fr)
Inventor
李烨
樊小毛
姚启航
尹丽妍
Original Assignee
中国科学院深圳先进技术研究院
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Priority to JP2020532967A priority Critical patent/JP7053836B2/ja
Publication of WO2019119518A1 publication Critical patent/WO2019119518A1/zh
Priority to US16/901,033 priority patent/US11538588B2/en

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A61B5/346—Analysis of electrocardiograms
    • A61B5/349—Detecting specific parameters of the electrocardiograph cycle
    • A61B5/352—Detecting R peaks, e.g. for synchronising diagnostic apparatus; Estimating R-R interval
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A61B5/346—Analysis of electrocardiograms
    • A61B5/349—Detecting specific parameters of the electrocardiograph cycle
    • A61B5/361—Detecting fibrillation
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present disclosure relates to the field of electrocardiographic monitoring technologies, and in particular, to a method, device and device for identifying atrial fibrillation signals.
  • Atrial fibrillation is the most common persistent chronic arrhythmia, often resulting from disordered atrial activity and irregular atrial compression. Atrial fibrillation signals in ECG signals can be identified by detecting variability in P-wave or RR interval, but the reliability and accuracy of these methods are difficult to meet clinical diagnostic requirements, resulting in no medical diagnostic or wearable monitoring devices. widely used.
  • an object of the present disclosure is to provide a method, device and device for identifying atrial fibrillation signal to improve the reliability and accuracy of atrial fibrillation signal recognition.
  • the present disclosure provides a method for identifying atrial fibrillation signal, the method comprising: acquiring an ECG signal to be identified; inputting an ECG signal to be identified into a pre-established AF signal recognition model, and outputting a room The recognition result of the tremor signal; wherein the AF signal recognition model is established by: acquiring a set number of ECG sample signals and corresponding identification information; wherein the identification information includes identification information of the normal signal and the AF signal; The number of normal signals is equalized by the SMOTE method; the network structure of the multi-channel convolutional neural network is established; each convolutional neural network is provided with a specific receptive field for identifying the corresponding granularity of the atrial fibrillation signal; The normal signal and the equalized processed AF signal are input into the network structure for training to generate atrial fibrillation signal recognition model.
  • the present disclosure provides an apparatus for identifying atrial fibrillation signal, the apparatus comprising: a signal acquisition module for acquiring an ECG signal to be identified; and a result output module, configured to input an ECG signal to be identified to
  • the pre-established AF signal recognition model outputs a recognition result of the atrial fibrillation signal; wherein the AF signal recognition model is established by acquiring a set number of ECG sample signals and corresponding identification information; wherein The identification information includes normal signal and identification information of the atrial fibrillation signal; according to the number of the normal signals, the atrial fibrillation signal is equalized by using a SMOTE method; and a network structure of the multi-channel convolutional neural network is established; each volume The neural network is provided with a specific receptive field for identifying a corresponding granularity of the AF signal; the normal signal and the equalized processed AF signal are input into the network structure for training to generate the AF signal recognition. model.
  • the present disclosure provides an identification device for atrial fibrillation signal, comprising a processor and a machine readable storage medium storing machine executable instructions executable by a processor, the processor executing the machine The instructions are executed to implement the above method for identifying atrial fibrillation signals.
  • the present disclosure provides a machine readable storage medium storing machine executable instructions that, when invoked and executed by a processor, cause the processor to implement the above A method of identifying atrial fibrillation signals.
  • the atrial fibrillation signal recognition model is established by multi-channel convolutional neural network training, and each convolutional neural network is provided with a specific receptive field, Identifying the corresponding granularity of the AF signal; by inputting the ECG signal to be identified into the AF signal recognition model, the recognition result of the AF signal can be obtained; the method uses multiple convolutional neural networks of different size receptive fields to identify The atrial fibrillation signal can identify the atrial fibrillation signals of different granularities, can comprehensively identify the atrial fibrillation signal, and improve the reliability and accuracy of the AF signal recognition.
  • the ECG signal required for this method has a short duration, and no manual intervention is required in the identification process, and the AF signal can be automatically and efficiently identified.
  • FIG. 1 is a flowchart of a method for identifying atrial fibrillation signal according to an embodiment of the present disclosure
  • FIG. 2 is a specific flowchart of establishing a model of atrial fibrillation signal recognition according to an embodiment of the present disclosure
  • FIG. 3 is a schematic structural diagram of a model of atrial fibrillation signal recognition according to an embodiment of the present disclosure
  • FIG. 4 is a flowchart of another method for identifying atrial fibrillation signal according to an embodiment of the present disclosure
  • FIG. 5 is a thermogram of atrial fibrillation signal recognition result according to an embodiment of the present disclosure
  • FIG. 6 is a schematic structural diagram of an apparatus for identifying atrial fibrillation signal according to an embodiment of the present disclosure
  • FIG. 7 is a schematic structural diagram of an apparatus for identifying atrial fibrillation signal according to an embodiment of the present disclosure.
  • Atrial fibrillation is the most common persistent chronic arrhythmia, often resulting from disordered atrial activity and irregular atrial compression. Atrial fibrillation is closely related to various diseases such as stroke, heart failure, coronary heart disease and thrombosis. Early identification of atrial fibrillation can help patients find heart abnormalities in time and reduce the disability and mortality caused by heart disease.
  • the AF signal in the ECG signal can be identified by detecting the variability of the P wave or RR interval.
  • the A-wave signal is identified based on the P-wave trigger signal equalization method, and the hypertension is identified according to the principle of P-wave disappearance.
  • the present disclosure provides a method, device and device for identifying atrial fibrillation signal, which can be applied to a server, and can also be applied to a wearable cardiac monitoring device.
  • a server In mobile phones, tablets, and other terminal devices, it is used to identify atrial fibrillation signals and to assist in the diagnosis of persistent chronic arrhythmias and other heart diseases.
  • the technique can be implemented in related software or hardware, which is described below by way of example.
  • Step S102 acquiring an ECG signal to be identified
  • Step S104 inputting the ECG signal to be identified into the pre-established AF signal recognition model, and outputting the recognition result of the AF signal;
  • the ECG signal to be identified generally includes a normal signal and atrial fibrillation signal; the ECG signal to be identified may be the original ECG signal collected by the device; in order to improve the accuracy of the recognition result, the original may be
  • the ECG signals are preprocessed, for example, filtering, noise reduction, RELU (Rectified Linear Units), activation function processing, regularization processing, and the like.
  • the pre-treated ECG signal can better adapt to the AF signal recognition model and improve the recognition effect.
  • Step S202 acquiring a set number of ECG sample signals and corresponding identification information; wherein the identification information includes identification information of a normal signal and atrial fibrillation signal;
  • Step S204 Perform equalization processing on the atrial fibrillation signal by using a SMOTE (Synthetic Minority Oversampling Technique) method according to the number of the normal signals;
  • SMOTE Synthetic Minority Oversampling Technique
  • Step S206 establishing a network structure of the multi-channel convolutional neural network; each convolutional neural network is provided with a specific receptive field for identifying a corresponding granularity of the atrial fibrillation signal;
  • the above step S204 uses SMOTE to equalize the atrial fibrillation signal, and the SMOTE can analyze a small number of samples and add new samples to the data set according to a few types of samples, so that SMOTE can be used to increase
  • the number of atrial fibrillation signals is such that the atrial fibrillation signal is roughly equal to the normal signal, facilitating the training of the network structure.
  • Step S208 the normal signal and the equalized processed AF signal are input into the network structure for training, and the atrial fibrillation signal recognition model is generated.
  • the atrial fibrillation signal originates from disordered atrial activity and irregular atrial compression. Therefore, the atrial fibrillation signal itself has various forms; for example, some of the atrial fibrillation signals have larger granularity, and the waveform is obvious or the wave height is higher; Some AF signals have smaller particle sizes, weaker waveforms or shorter wave heights.
  • the single-channel convolutional neural network usually only can sense the atrial fibrillation signal in a narrow range of granularity due to the single field of receptive field. It is difficult to learn and recognize various forms of atrial fibrillation signals, resulting in incomplete recognition and low recognition accuracy.
  • the present embodiment uses a multi-channel convolutional neural network to establish the above-mentioned atrial fibrillation signal recognition model, and each convolutional neural network is provided with a specific receptive field for identifying the corresponding granularity of the atrial fibrillation signal.
  • Receptively small convolutional neural networks can identify small-scale atrial fibrillation signals, receptively large convolutional neural networks, and can identify large-scale atrial fibrillation signals.
  • each neural network is constructed by a multi-layer convolutional layer, and the convolution kernel receptive field of each convolution layer can be set separately; therefore, the convolution kernel receptive field of each convolution layer in each neural network can be the same , can also be different.
  • the convolution kernel receptive fields of each convolution layer may be increased layer by layer, or may be reduced layer by layer, and other setting manners may be adopted.
  • the convolutional neural network may be an MS-CNN (Multi-scale Convolutional Neural Network) or other types of neural networks.
  • a method for identifying atrial fibrillation signal wherein the atrial fibrillation signal recognition model is established by multi-channel convolutional neural network training, and each convolutional neural network is provided with a specific receptive field to identify the corresponding granularity.
  • Atrial fibrillation signal; the recognition result of the atrial fibrillation signal can be obtained by inputting the ECG signal to be identified into the AF signal recognition model; the method uses multiple convolutional neural networks of different size receptive fields to identify the AF signal. It can identify the atrial fibrillation signals of different granularities, can comprehensively identify the AF signal, and improve the reliability and accuracy of AF signal recognition.
  • the ECG signal required for this method has a short duration, and no manual intervention is required in the identification process, and the AF signal can be automatically and efficiently identified.
  • the network structure of the multi-channel convolutional neural network includes two neural networks as an example, which are respectively a first network and a second network;
  • the first network and the second network are both VGG-16 convolutional neural networks, and of course, other types of convolutional neural networks such as CaffeNet and ResNet.
  • Each of the first network and the second network includes a multi-layer convolution layer and a multi-layer maximum pooling layer; as shown in FIG. 3, the first network includes 13 convolution layers (convolution layer, ie, ConvNet layer in FIG. 3) and 5 layers of the largest pooling layer (the largest pooling layer is the Maxpool layer in Figure 3); 13 layers of convolutional layers are arranged in sequence, 5 layers of the largest pooling layer are interspersed between 13 layers of convolutional layers; the second network also includes 13 The layer convolution layer and the 5 layer maximum pooling layer; 13 layers of convolution layers are sequentially arranged, and 5 layers of the largest pooling layer are interspersed between the 13 layers of convolution layers.
  • the first network includes 13 convolution layers (convolution layer, ie, ConvNet layer in FIG. 3) and 5 layers of the largest pooling layer (the largest pooling layer is the Maxpool layer in Figure 3); 13 layers of convolutional layers are arranged in sequence, 5 layers of the largest pooling layer are interspersed between 13 layers of
  • the receptive field of the convolution kernel in the first network is less than or equal to the receptive field of the convolution kernel in the second network; as shown in FIG. 3, the 13-layer volume of the first network
  • the layered convolution kernels have the same receptive field, all of which are 3; of course, the receptive field can be adjusted to other values.
  • the convolution kernel receptive field of the 13-layer convolutional layer in the second network can be set to different values. For example, in Figure 3, from left to right, the convolution kernel receptive fields of the first four layers of convolutional layers are set to 7, respectively.
  • the convolution kernel receptive field of the layer is set to 3; of course, the convolution kernel receptive field of the first 4 layers of convolutional layer can also be set to 3, 5, 9, etc.; convolution of other layers except the first 4 layers of convolutional layer
  • the nuclear receptive field can also be flexibly set to other values.
  • Table 1 below shows various configuration parameters of the AF signal recognition model; wherein, in the MS-CNN (3, 3) configuration, the convolution kernel receptive fields of each convolution layer of the first network are all 3; The convolution kernel receptive field of each layer of convolutional layer is also 3; in the MS-CNN (3, 5) configuration, the convolution kernel receptive field of each convolutional layer of the first network is 3; The convolution kernel receptive field of the first four layers of convolutional layer is 5, and the convolution kernel receptive field of other layers is 3; in the MS-CNN (3,7) configuration, the convolution kernels of the convolutional layers of the first network The wild is 3; the convolution kernel receptive field of the first 4 layers of the second network is 7, and the convolution kernel receptive field of the other layers is 3; in the MS-CNN (3, 9) configuration, the first network The convolution kernel receptive field of each convolution layer is 3; the convolution kernel receptive field of
  • the first pool and the largest pooled layer of the last layer of the second network are connected to each other; the interconnected maximum pooled layer is also connected with multiple layers of fully connected layers (the full connection layer is a map) The Full connect layer in 3) and the Softmax layer. As shown in FIG. 3, the first pool and the largest pooled layer of the last layer of the second network are connected in parallel, and then the four layers of the full connection layer and the first layer of the Softmax layer are sequentially connected.
  • an ECG (Electrocardiogram) signal (corresponding to the ECG signal) is input to the first layer convolution layer of the first network and the second network, respectively, through each convolution layer, After the maximum pooling layer and the fully connected layer are processed, the classification result is output by the Softmax layer, and the result includes the AF signal and the normal signal Normal, thereby realizing the recognition of the atrial fibrillation signal.
  • ECG Electrocardiogram
  • the inventor In the establishment process of the above-mentioned atrial fibrillation signal recognition model, the inventor usually needs to repeatedly adjust various configuration parameters to optimize the recognition effect of the model; based on this, during the establishment of the above-mentioned atrial fibrillation signal recognition model, normal signals and equalization will be performed.
  • the processed AF signal is input into the network structure for training, and the steps of generating the AF signal recognition model include:
  • Step 1 input the normal signal and the equalized processed AF signal into the network structure for training to generate an initial model
  • Step 2 calculate the sensitivity of the AF signal recognition model
  • Step 3 calculate the specificity of the AF signal recognition model
  • Step 4 Calculate the accuracy of the AF signal recognition model
  • Step 5 Calculate the accuracy of the AF signal recognition model
  • # represents the number
  • TP is the correct atrial fibrillation signal
  • FP is the wrong AF signal
  • TN is the correct normal signal
  • FN is the normal signal of the recognition error
  • Step 6 determining whether the sensitivity, specificity, accuracy, and accuracy respectively meet the corresponding threshold, and if not, adjusting the configuration parameters in the network structure until the sensitivity, specificity, accuracy, and accuracy satisfy the corresponding threshold;
  • step 7 the initial model is determined as a model of atrial fibrillation signal recognition.
  • Step S402 acquiring an original ECG signal
  • Step S404 pre-processing the original ECG signal to generate an ECG signal to be identified; the pre-processing includes filtering processing and regularization processing.
  • the frequency of the ECG signal collected by the ECG collector or other cardiac testing equipment is wider than 150 Hz; and the inventors have repeatedly tested and found that the frequency contributing to the identification of the AF signal is concentrated at 0.5 Hz to 40 Hz. Therefore, in order to reduce the interference signal recognized by the model while reducing the amount of data and improve the recognition efficiency, the ECG signal needs to be filtered before being input to the model; for example, FIR (Finite Impulse Response, finite-length unit impulse response) can be used.
  • FIR Finite Impulse Response, finite-length unit impulse response
  • the filter downsamples the original ECG signal to filter out high frequency noise; the parameters of the FIR filter can be set to a low pass filter with a cutoff frequency of 60hz and 512 steps, and the original ECG signal can be downsampled. Up to 120Hz.
  • the atrial fibrillation signal recognition model needs to use an activation function for activation processing before performing recognition processing; since the model in this embodiment is established by a two-way convolutional neural network, the model established by the two-way convolutional neural network is not suitable.
  • the above RELU algorithm therefore, the regularization process is used in this embodiment to prevent the fitting of the model, which is beneficial to the diversity of the gradient of the model node.
  • Step S406 the ECG signal to be identified is input into the AF signal recognition model, and the recognition result of the AF signal is output.
  • a method for identifying atrial fibrillation signal provided by an embodiment of the present disclosure, wherein the atrial fibrillation signal recognition model is established by two-way convolutional neural network training, and each convolutional neural network is provided with a specific receptive field to identify the corresponding granularity.
  • the atrial fibrillation signal after the ECG signal is filtered and regularized, it is input into the AF signal recognition model to obtain the recognition result of the atrial fibrillation signal; the method uses multiple convolutional neural networks of different size receptive fields. Identifying the atrial fibrillation signal can identify different sizes of atrial fibrillation signals, can comprehensively identify the AF signal, and improve the reliability and accuracy of AF signal recognition.
  • the ECG signal required for this method has a short duration, and no manual intervention is required in the identification process, and the AF signal can be automatically and efficiently identified.
  • the inventors In order to verify the recognition effect of the above-mentioned method for identifying the atrial fibrillation signal, the inventors also analyzed the data outputted by the above-mentioned atrial fibrillation signal recognition model, see the thermogram of the atrial fibrillation signal recognition result shown in FIG. 5; the above-mentioned atrial fibrillation signal recognition In the model, the 256-dimensional feature data output by the Softmax layer is subjected to principal component analysis, and 4 principal component data are intercepted for display; as shown in FIG. 5, the thermal powers of the five main component data of five atrial fibrillation signals and five normal signals are shown. Figure; It can be seen from the figure that the gray value of the AF signal and the normal signal Normal are significantly different, and the difference is obvious, indicating that the recognition method of the AF signal is better.
  • Table 2 shows the accuracy of AF signal recognition in each receptive field configuration; four configuration parameters are counted in Table 2 (MS-CNN(3,3), MS-CNN(3,5), MS -CNN (3,7) and MS-CNN (3,9)), Sensi, Specific Spe, Accuracy Pre and Accuracy Acc data corresponding to the input length of the four ECG signals; Table 2 It shows that the accuracy of the data input length of the model is not less than 96% for 5 seconds, 10 seconds, 20 seconds and 30 seconds; among them, in the MS-CNN (3, 7) configuration, the 20-second input length corresponds to The accuracy rate reached 98.13%.
  • FIG. 6 is a schematic structural diagram of an apparatus for identifying atrial fibrillation signal; the apparatus includes:
  • a signal acquisition module 60 configured to acquire an ECG signal to be identified
  • the result output module 61 is configured to input the ECG signal to be identified into the pre-established AF signal recognition model, and output the recognition result of the AF signal; wherein the AF signal recognition model is established by: acquiring the setting a quantity of ECG sample signals and corresponding identification information; wherein the identification information includes normal signal and identification information of the atrial fibrillation signal; according to the number of normal signals, the AMO signal is equalized by the SMOTE method; and the multi-channel convolutional nerve is established The network structure of the network; each convolutional neural network is provided with a specific receptive field for identifying the corresponding granularity of the AF signal; the normal signal and the equalized AF signal are input into the network structure for training to generate atrial fibrillation Signal recognition model.
  • the network structure of the multi-channel convolutional neural network includes a first network and a second network; the first network and the second network are both VGG-16 convolutional neural networks; the first network and the second network each comprise a multi-layer convolution layer And a multi-layer maximum pooling layer; the receptive field of the convolution kernel in the first network is less than or equal to the receptive field of the convolution kernel in the second network; and the largest pooled layer of the last layer of the first network and the second network are interconnected The largest pooled layer after interconnection is also connected with multiple layers of fully connected layers and Softmax layers.
  • the signal acquisition module is further configured to: acquire an original ECG signal; preprocess the original ECG signal to generate an ECG signal to be identified; and perform preprocessing including filtering processing and regularization processing.
  • the result output module is further configured to: input the normal signal and the equalized processed AF signal into the network structure for training, and generate an initial model;
  • TP is to identify the correct atrial fibrillation signal
  • FP is to identify the wrong AF signal
  • TN is to identify the correct normal signal
  • FN is the normal signal to identify the error
  • Corresponding threshold if not, adjust the configuration parameters in the network structure until the sensitivity, specificity, accuracy, and accuracy meet the corresponding threshold; determine the initial model as the AF signal recognition model.
  • the device for identifying atrial fibrillation signal uses a multi-channel convolutional neural network of different size receptive fields to identify atrial fibrillation signals, can identify atrial fibrillation signals of different granularities, and can comprehensively identify atrial fibrillation signals. Improve the reliability and accuracy of AF signal recognition.
  • the ECG signal required for this method has a short duration, and no manual intervention is required in the identification process, and the AF signal can be automatically and efficiently identified.
  • the present embodiment provides an identification device for atrial fibrillation signal corresponding to the above method embodiment.
  • 7 is a schematic structural diagram of the identification device. As shown in FIG. 7, the device includes a processor 1201 and a memory 1202. The memory 1202 is configured to store one or more computer instructions, and one or more computer instructions are executed by the processor. To achieve the above identification method of atrial fibrillation signal.
  • the identification device shown in FIG. 7 further includes a bus 1203 and a forwarding chip 1204.
  • the processor 1201, the forwarding chip 1204, and the memory 1202 are connected by a bus 1203.
  • the identification device can be a network edge device.
  • the memory 1202 may include a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk storage.
  • the bus 1203 may be an ISA bus, a PCI bus, or an EISA bus.
  • the bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one double-headed arrow is shown in Figure 7, but it does not mean that there is only one bus or one type of bus.
  • the forwarding chip 1204 is configured to connect to the at least one user terminal and other network elements through the network interface, and send the encapsulated IPv4 packet or the IPv6 packet to the user terminal through the network interface.
  • the processor 1201 may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the foregoing method may be completed by an integrated logic circuit of hardware in the processor 1201 or an instruction in a form of software.
  • the processor 1201 may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP processor, etc.), or a digital signal processor (DSP). ), Application Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component.
  • the methods, steps, and logical block diagrams disclosed in the embodiments of the invention may be implemented or carried out.
  • the general purpose processor may be a microprocessor or the processor or any conventional processor or the like.
  • the steps of the method disclosed in connection with the embodiments of the present invention may be directly implemented by the hardware decoding processor, or may be performed by a combination of hardware and software modules in the decoding processor.
  • the software module can be located in a conventional storage medium such as random access memory, flash memory, read only memory, programmable read only memory or electrically erasable programmable memory, registers, and the like.
  • the storage medium is located in the memory 1202, and the processor 1201 reads the information in the memory 1202 and, in conjunction with its hardware, performs the steps of the method of the foregoing embodiment.
  • Embodiments of the present disclosure also provide a machine readable storage medium storing machine executable instructions that, when invoked and executed by a processor, cause the processor to implement
  • a processor executes machine executable instructions that, when invoked and executed by a processor, cause the processor to implement
  • the method, device and device for identifying atrial fibrillation signal do not require manual feature engineering and have the function of automatically extracting ECG features; two-way VGG-16 network is adopted, but the two networks have different granularities.
  • the convolution kernel can capture more feature information of different granularity.
  • each block of the flowchart or block diagram can represent a module, a program segment, or a portion of code that includes one or more of the Executable instructions. It should also be noted that, in some alternative implementations, the functions noted in the blocks may also occur in a different order than those illustrated in the drawings.
  • each block of the block diagrams and/or flowcharts, and combinations of blocks in the block diagrams and/or flowcharts can be implemented in a dedicated hardware-based system that performs the specified function or function. Or it can be implemented by a combination of dedicated hardware and computer instructions.
  • each functional module or unit in each embodiment of the present invention may be integrated to form a separate part, or each module may exist separately, or two or more modules may be integrated to form a separate part.
  • the functions may be stored in a computer readable storage medium if implemented in the form of a software functional unit and sold or used as a standalone product. Based on such understanding, the portion of the technical solution of the present disclosure that contributes in essence or to the prior art or the portion of the technical solution may be embodied in the form of a software product stored in a storage medium, including The instructions are used to cause a computer device (which may be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present disclosure.
  • the foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and the like. .

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Cardiology (AREA)
  • Pathology (AREA)
  • Theoretical Computer Science (AREA)
  • Veterinary Medicine (AREA)
  • Mathematical Physics (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • Evolutionary Computation (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Data Mining & Analysis (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • General Physics & Mathematics (AREA)
  • Physiology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Fuzzy Systems (AREA)
  • Signal Processing (AREA)
  • Psychiatry (AREA)
  • Primary Health Care (AREA)
  • Epidemiology (AREA)
  • Databases & Information Systems (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

本公开提供了一种房颤信号的识别方法、装置和设备；其中，该方法包括：获取待识别的心电信号；将待识别的心电信号输入至预先建立的房颤信号识别模型中，输出房颤信号的识别结果；其中，该房颤信号识别模型通过下述方式建立：获取设定数量的心电样本信号和对应的标识信息；其中，标识信息包括正常信号和房颤信号的标识信息；根据正常信号的数量，采用SMOTE方式对房颤信号进行均衡处理；建立多路卷积神经网络的网络结构；每路卷积神经网络设置有特定的感受野，用于识别相应粒度的房颤信号；将正常信号和均衡处理后的房颤信号输入至网络结构中进行训练，生成房颤信号识别模型。本公开可以提高房颤信号识别的可靠性和准确性。

Description

房颤信号的识别方法、装置和设备
相关申请的交叉引用
本申请要求于2017年12月19日提交中国专利局的申请号为201711380466.9、名称为“房颤信号的识别方法、装置和设备”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本公开涉及心电监测技术领域，尤其是涉及一种房颤信号的识别方法、装置和设备。
背景技术
心房颤动是最常见的持续性慢性心律失常，常源于无序的心房活动和不规则的心房压缩。通过检测P波或者RR间期的变异性可以识别心电信号中的房颤信号，但是这些方法识别的可靠性和准确性难以达到临床诊断需求，导致在医学诊断或可穿戴监测设备中没有得到广泛应用。
随着人工智能的快速发展，还出现了应用机器学习识别房颤信号的方法，但这种识别方法需要较长的心电信号，且对数据的处理要求较高，通常需要人工提取信号特性，进行降噪处理；对于穿戴式单导联设备采集的动态心电图信号，该方法的识别可靠性和准确性依然难以达到临床诊断需求。
发明内容
有鉴于此，本公开的目的在于提供一种房颤信号的识别方法、装置和设备，以提高房颤信号识别的可靠性和准确性。
为了实现上述目的，本公开采用的技术方案如下：
第一方面，本公开提供了一种房颤信号的识别方法，该方法包括：获取待识别的心电信号；将待识别的心电信号输入至预先建立的房颤信号识别模型中，输出房颤信号的识别结果；其中，房颤信号识别模型通过下述方式建立：获取设定数量的心电样本信号和对应的标识信息；其中，标识信息包括正常信号和房颤信号的标识信息；根据正常信号的数量，采用SMOTE方式对房颤信号进行均衡处理；建立多路卷积神经网络的网络结构；每路卷积神经网络设置有特定的感受野，用于识别相应粒度的房颤信号；将正常信号和均衡处理后的房颤信号输入至网络结构中进行训练，生成房颤信号识别模型。
第二方面，本公开提供了一种房颤信号的识别装置，该装置包括：信号获取模块，用于获取待识别的心电信号；结果输出模块，用于将待识别的心电信号输入至预先建立的房颤信号识别模型中，输出房颤信号的识别结果；其中，所述房颤信号识别模型通过下述方式建立：获取设定数量的心电样本信号和对应的标识信息；其中，所述标识信息包括正常信号和房颤信号的标识信息；根据所述正常信号的数量，采用SMOTE方式对所述房颤信号进行均衡处理；建立多路卷积神经网络的网络结构；每路卷积神经网络设置有特定的感受野，用于识别相应粒度的房颤信号；将所述正常信号和均衡处理后的房颤信号输入至所述网络结构中进行训练，生成所述房颤信号识别模型。
第三方面，本公开提供了一种房颤信号的识别设备，包括处理器和机器可读存储介质，机器可读存储介质存储有能够被处理器执行的机器可执行指令，处理器执行机器可执行指令以实现上述房颤信号的识别方法。
第四方面，本公开提供了一种机器可读存储介质，机器可读存储介质存储有机器可执行指令，机器可执行指令在被处理器调用和执行时，机器可执行指令促使处理器实现上述房颤信号的识别方法。
上述房颤信号的识别方法、装置、设备和机器可读存储介质，其房颤信号识别模型通过多路卷积神经网络训练而建立，且每路卷积神经网络设置有特定的感受野，以识别相应粒度的房颤信号；通过将待识别的心电信号输入至该房颤信号识别模型中，可以获得房颤信号的识别结果；该方式采用多路不同大小感受野的卷积神经网络识别房颤信号，可以识别不同粒度的房颤信号，能较为全面地识别房颤信号，提高了房颤信号识别的可靠性和准确性。
同时，该方式所需的心电信号时长较短，且识别过程中无需人工干预，可以自动、高效地识别房颤信号。
本公开的其他特征和优点将在随后的说明书中阐述，或者，部分特征和优点可以从说明书推知或毫无疑义地确定，或者通过实施本公开的上述技术即可得知。
为使本发明的上述目的、特征和优点能更明显易懂，下文特举较佳实施方式，并配合所附附图，作详细说明如下。
附图说明
为了更清楚地说明本公开具体实施方式或现有技术中的技术方案，下面将对具体实施方式或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图是本公开的一些实施方式，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本公开实施方式提供的一种房颤信号的识别方法的流程图；
图2为本公开实施方式提供的建立房颤信号识别模型的具体流程图；
图3为本公开实施方式提供的房颤信号识别模型的结构示意图；
图4为本公开实施方式提供的另一种房颤信号的识别方法的流程图；
图5为本公开实施方式提供的房颤信号识别结果的热力图；
图6为本公开实施方式提供的一种房颤信号的识别装置的结构示意图；
图7为本公开实施方式提供的一种房颤信号的识别设备的结构示意图。
具体实施方式
为使本公开实施方式的目的、技术方案和优点更加清楚，下面将结合附图对本公开的技术方案进行清楚、完整地描述，显然，所描述的实施方式是本公开一部分实施方式，而不是全部的实施方式。基于本公开中的实施方式，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施方式，都属于本公开保护的范围。
心房颤动是最常见的持续性慢性心律失常，常源于无序的心房活动和不规则的心房压缩。心房颤动与卒中、心力衰竭、冠心病和血栓等多种疾病密切相关，早期的房颤识别可以帮助患者及时发现心脏异常，降低心脏疾病引发的致残率和致死率。
可以通过检测P波或者RR间期的变异性来识别心电信号中的房颤信号，例如，基于P波触发信号均衡的方法识别房颤信号，根据P波消失的原理识别高血压人群中的房颤信号，基于随机性、变异性和复杂性的RR间期识别房颤信号等；然而，由于这些方法识别的可靠性和准确性难以达到临床诊断需求，导致在医学诊断或可穿戴监测设备中没有得到广泛应用。
随着人工智能的快速发展，现有技术中还出现了应用机器学习识别房颤信号的方式，例如，采用扩展非线性贝叶斯网络识别单通道单导联的房颤信号；但这种识别方法需要较长的心电信号(通常为30秒以上)，且对数据的处理要求较高，通常需要人工提取信号特性，并进行降噪处理；对于穿戴式单导联设备采集的动态心电图信号，该方法的识别可靠性和准确性依然难以达到临床诊断需求。
针对上述房颤信号识别的准确率较低的问题，本公开实施方式提供了一种房颤信号的识别方法、装置和设备；该技术可以应用于服务器中，还 可以应用于穿戴式心脏监测设备、手机、平板电脑等终端设备中，用于识别房颤信号，辅助诊断持续性慢性心律失常等心脏疾病。该技术可以采用相关的软件或硬件实现，下面通过实施例进行描述。
参见图1所示的一种房颤信号的识别方法的流程图；该方法包括如下步骤：
步骤S102，获取待识别的心电信号；
步骤S104，将待识别的心电信号输入至预先建立的房颤信号识别模型中，输出房颤信号的识别结果；
上述待识别的心电信号通常包含正常信号和房颤信号；该待识别的心电信号可以为设备采集的原始心电信号；为了提高识别结果的准确性，在进行识别之前，可以对原始的心电信号进行预处理，例如，滤波、降噪、采用RELU(Rectified Linear Units，修正的线性单位)激活函数处理、正则化处理等。预处理后的心电信号可以更好地适应房颤信号识别模型，提高识别效果。
参见图2所示的建立房颤信号识别模型的具体流程图；该房颤信号识别模型通过下述方式建立：
步骤S202，获取设定数量的心电样本信号和对应的标识信息；其中，所述标识信息包括正常信号和房颤信号的标识信息；
步骤S204，根据所述正常信号的数量，采用SMOTE(Synthetic Minority Oversampling Technique，合成少数过采样技术)方式对所述房颤信号进行均衡处理；
步骤S206，建立多路卷积神经网络的网络结构；每路卷积神经网络设置有特定的感受野，用于识别相应粒度的房颤信号；
对于一段心电信号，通常仅存在数个房颤信号，正常信号的数量远大于房颤信号的数量，二者数量不均衡；如果使用这样的数据对神经网络的网络结构进行训练，可能会影响模型的识别效果；基于此，上述步骤S204 采用SMOTE对房颤信号进行均衡处理，该SMOTE可以对少数类样本进行分析并根据少数类样本人工合成新样本添加到数据集中，因此，可以采用SMOTE增加房颤信号的数量，以使房颤信号与正常信号大致均衡，便于网络结构的训练。
步骤S208，将所述正常信号和均衡处理后的房颤信号输入至所述网络结构中进行训练，生成所述房颤信号识别模型。
由上述可知，房颤信号源于无序的心房活动和不规则的心房压缩，因此，房颤信号本身具有多种形态；例如，部分房颤信号粒度较大，波形明显或波高较高；而部分房颤信号粒度较小，波形微弱或波高较矮。单路的卷积神经网络由于感受野单一，通常仅能感受较窄粒度范围内的房颤信号，难以学习并识别出各种形态的房颤信号，导致识别不全面，识别准确性较低。
基于此，本实施方式采用多路卷积神经网络建立上述房颤信号识别模型，每路卷积神经网络设置有特定的感受野，用于识别相应粒度的房颤信号。感受野较小的卷积神经网络，可以识别粒度较小的房颤信号，感受野较大的卷积神经网络，可以识别粒度较大的房颤信号。
由于每路神经网络通过由多层卷积层搭建，且每层卷积层的卷积核感受野可以单独设置；因此，每路神经网络中每层卷积层的卷积核感受野可以相同，也可以不同。当每层卷积层的卷积核感受野不同时，各层卷积层的卷积核感受野可以逐层变大，也可以逐层减小，还可以采用其他设置方式。另外，上述卷积神经网络可以为MS-CNN(Multi-scaleConvolutional Neural Network，多尺度卷积神经网络)，也可以为其他类型的神经网络。
本公开实施方式提供的一种房颤信号的识别方法，其房颤信号识别模型通过多路卷积神经网络训练而建立，且每路卷积神经网络设置有特定的感受野，以识别相应粒度的房颤信号；通过将待识别的心电信号输入至该房颤信号识别模型中，可以获得房颤信号的识别结果；该方式采用多路不 同大小感受野的卷积神经网络识别房颤信号，可以识别不同粒度的房颤信号，能较为全面地识别房颤信号，提高了房颤信号识别的可靠性和准确性。
同时，该方式所需的心电信号时长较短，且识别过程中无需人工干预，可以自动、高效地识别房颤信号。
参见图3所示的房颤信号识别模型的结构示意图；图3中以多路卷积神经网络的网络结构包括两路神经网络为例进行说明，分别为第一网络和第二网络；该实施方式中，第一网络和第二网络均为VGG-16卷积神经网络，当然，还可以为CaffeNet、ResNet等其他类型的卷积神经网络。
上述第一网络和第二网络均包括多层卷积层和多层最大池化层；如图3中，第一网络包括13层卷积层(卷积层即图3中的ConvNet层)和5层最大池化层(最大池化层即图3中的Maxpool层)；13层卷积层依次排列，5层最大池化层穿插在13层卷积层之间；第二网络也包括13层卷积层和5层最大池化层；13层卷积层依次排列，5层最大池化层穿插在13层卷积层之间。
为了识别不同粒度的房颤信号，可以设置为第一网络内卷积核的感受野小于或等于第二网络内卷积核的感受野；如图3中所示，第一网络的13层卷积层的卷积核感受野相同，均为3；当然该感受野还可以调整为其他数值。第二网络中的13层卷积层的卷积核感受野可以设置为不同数值，例如图3中，从左至右，前4层卷积层的卷积核感受野分别设置为7，其他层的卷积核感受野分别设置为3；当然，前4层卷积层的卷积核感受野还可以设置为3、5、9等；除前4层卷积层的其他层的卷积核感受野也可以灵活设置为其他数值。
如下述表1为房颤信号识别模型的多种配置参数；其中，MS-CNN(3，3)配置中，第一网络的各卷积层的卷积核感受野均为3；第二网络的各层卷积层的卷积核感受野也均为3；MS-CNN(3，5)配置中，第一网络的各卷积层的卷积核感受野均为3；第二网络的前4层卷积层的卷积核感受野为 5，其他层的卷积核感受野为3；MS-CNN(3，7)配置中，第一网络的各卷积层的卷积核感受野均为3；第二网络的前4层卷积层的卷积核感受野为7，其他层的卷积核感受野为3；MS-CNN(3，9)配置中，第一网络的各卷积层的卷积核感受野均为3；第二网络的前4层卷积层的卷积核感受野为9，其他层的卷积核感受野为3。
表1
Figure PCTCN2017120106-appb-000001
该房颤信号识别模型中，第一网络和第二网络的最末层的最大池化层相互连接；相互连接后的最大池化层还依次连接有多层全连接层(全连接层即图3中的Full connect层)和Softmax层。如图3中，第一网络和第二 网络的最末层的最大池化层并列连接，再依次连接4层全连接层和1层Softmax层。
采用该房颤信号识别模型进行识别时，ECG(Electrocardiogram，心电图)信号(相当于上述心电信号)分别输入至第一网络和第二网络的首层卷积层中，经各卷积层、最大池化层、全连接层处理后，由该Softmax层输出分类结果，该结果中包括房颤信号AF和正常信号Normal，进而实现房颤信号的识别。
上述房颤信号识别模型在建立过程中，发明人通常需要反复调节各种配置参数，以使模型的识别效果最佳；基于此，上述房颤信号识别模型的建立过程中，将正常信号和均衡处理后的房颤信号输入至网络结构中进行训练，生成房颤信号识别模型的步骤，具体包括：
步骤1，将正常信号和均衡处理后的房颤信号输入至网络结构中进行训练，生成初始模型；
步骤2，计算房颤信号识别模型的灵敏度
Figure PCTCN2017120106-appb-000002
步骤3，计算房颤信号识别模型的特异性
Figure PCTCN2017120106-appb-000003
步骤4，计算房颤信号识别模型的精确度
Figure PCTCN2017120106-appb-000004
步骤5，计算房颤信号识别模型的准确度
Figure PCTCN2017120106-appb-000005
其中，#代表数量；TP为识别正确的房颤信号；FP为识别错误的房颤信号；TN为识别正确的正常信号；FN为识别错误的正常信号；
步骤6，判断灵敏度、特异性、精确度和准确度是否分别满足对应的阈值，如果否，调整网络结构中的配置参数，直至灵敏度、特异性、精确度和准确度满足对应的阈值；
步骤7，将初始模型确定为房颤信号识别模型。
参见图4所示的另一种房颤信号的识别方法的流程图；该方法应用上述图3中所示的房颤信号识别模型实现；该方法包括如下步骤：
步骤S402，获取原始心电信号；
步骤S404，对原始心电信号进行预处理，生成待识别的心电信号；该预处理包括滤波处理和正则化处理。
通过心电采集器或其他心脏检测设备采集到的心电信号的频率宽度较宽，超过150Hz；而经发明人反复试验发现，对识别房颤信号贡献较大的频率集中在0.5Hz至40Hz之间；因此，为了减少模型识别的干扰信号同时减少数据量，提高识别效率，心电信号在输入至模型之前，需要进行滤波处理；例如，可以采用FIR(Finite Impulse Response，有限长单位冲激响应)滤波器对原始心电信号进行降采样，滤除高频噪声；该FIR滤波器的参数可以设置为截止频率为60hz、512阶的低通滤波器，进而可以将上述原始心电信号降采样至120Hz。
通常，上述房颤信号识别模型在进行识别处理之前，需要采用激活函数进行激活处理；由于该实施方式中的模型由两路卷积神经网络建立，而两路卷积神经网络建立的模型不适合上述RELU算法，因此该实施方式中采用正则化处理，以防止模型的拟合，有利于模型节点梯度下降的多样性。
步骤S406，将待识别的心电信号输入至房颤信号识别模型中，输出房颤信号的识别结果。
本公开实施方式提供的一种房颤信号的识别方法，其房颤信号识别模型通过两路卷积神经网络训练而建立，且每路卷积神经网络设置有特定的感受野，以识别相应粒度的房颤信号；心电信号经滤波处理和正则化处理后，输入至该房颤信号识别模型中，可以获得房颤信号的识别结果；该方式采用多路不同大小感受野的卷积神经网络识别房颤信号，可以识别不同粒度的房颤信号，能较为全面地识别房颤信号，提高了房颤信号识别的可靠性和准确性。
同时，该方式所需的心电信号时长较短，且识别过程中无需人工干预，可以自动、高效地识别房颤信号。
为了验证上述房颤信号的识别方法的识别效果，发明人还在上述房颤信号识别模型输出的数据进行了分析，参见图5所示的房颤信号识别结果的热力图；上述房颤信号识别模型中，Softmax层输出的256维特征数据经主成分分析后，截取4个主成分数据进行显示；如图5中显示了5个房颤信号和5个正常信号的4个主成分数据的热力图；从图中可以看出，房颤信号AF与正常信号Normal的灰度值呈现明显不同的特征，区别较为明显，说明上述房颤信号的识别方法的识别效果较好。
下述表2所示为各个感受野配置下的房颤信号识别准确率；该表2中统计了四种配置参数(MS-CNN(3，3)、MS-CNN(3，5)、MS-CNN(3，7)和MS-CNN(3，9))下，四种心电信号输入长度(Input length)对应的灵敏度Sen、特异性Spe、精确度Pre和准确度Acc数据；表2表明，该模型对于5秒、10秒、20秒以及30秒的数据输入长度，准确率均不低于96％；其中，在MS-CNN(3，7)配置下，20秒输入长度对应的准确率达到了98.13％。
表2
Figure PCTCN2017120106-appb-000006
对应于上述方法实施方式，参见图6所示的一种房颤信号的识别装置的结构示意图；该装置包括：
信号获取模块60，用于获取待识别的心电信号；
结果输出模块61，用于将待识别的心电信号输入至预先建立的房颤信号识别模型中，输出房颤信号的识别结果；其中，房颤信号识别模型通过下述方式建立：获取设定数量的心电样本信号和对应的标识信息；其中，标识信息包括正常信号和房颤信号的标识信息；根据正常信号的数量，采用SMOTE方式对房颤信号进行均衡处理；建立多路卷积神经网络的网络结构；每路卷积神经网络设置有特定的感受野，用于识别相应粒度的房颤信号；将正常信号和均衡处理后的房颤信号输入至网络结构中进行训练，生成房颤信号识别模型。
上述多路卷积神经网络的网络结构包括第一网络和第二网络；第一网络和第二网络均为VGG-16卷积神经网络；第一网络和第二网络均包括多层卷积层和多层最大池化层；第一网络内卷积核的感受野小于或等于第二网络内卷积核的感受野；第一网络和第二网络的最末层的最大池化层相互连接；相互连接后的最大池化层还依次连接有多层全连接层和Softmax层。
上述信号获取模块，还用于：获取原始心电信号；对原始心电信号进行预处理，生成待识别的心电信号；预处理包括滤波处理和正则化处理。
上述结果输出模块，还用于：将正常信号和均衡处理后的房颤信号输入至网络结构中进行训练，生成初始模型；
计算初始模型的灵敏度
Figure PCTCN2017120106-appb-000007
计算初始模型的特异性
Figure PCTCN2017120106-appb-000008
计算初始模型的精确度
Figure PCTCN2017120106-appb-000009
计算初始模型的准确度
Figure PCTCN2017120106-appb-000010
其中，TP为识别正确的房颤信号；FP为识别错误的房颤信号；TN为识别正确的正常信号；FN为识别错误的正常信号；判断灵敏度、特异性、精确度和准确度是否分别满足对应的阈值，如果否，调整网络结构中的配置参数，直至灵敏度、特异性、精确度和准确度满足对应的阈值；将初始模型确定为房颤信号识别模型。
本公开实施方式提供的一种房颤信号的识别装置，采用多路不同大小感受野的卷积神经网络识别房颤信号，可以识别不同粒度的房颤信号，能较为全面地识别房颤信号，提高了房颤信号识别的可靠性和准确性。同时，该方式所需的心电信号时长较短，且识别过程中无需人工干预，可以自动、高效地识别房颤信号。
本实施方式提供了一种与上述方法实施方式相对应的房颤信号的识别设备。图7为该识别设备的结构示意图，如图7所示，该设备包括处理器1201和存储器1202；其中，存储器1202用于存储一条或多条计算机指令，一条或多条计算机指令被处理器执行，以实现上述房颤信号的识别方法。
图7所示的识别设备还包括总线1203和转发芯片1204，处理器1201、转发芯片1204和存储器1202通过总线1203连接。该识别设备可以是网络边缘设备。
其中，存储器1202可能包含高速随机存取存储器(RAM，Random Access Memory)，也可能还包括非不稳定的存储器(non-volatile memory)，例如至少一个磁盘存储器。总线1203可以是ISA总线、PCI总线或EISA总线等。所述总线可以分为地址总线、数据总线、控制总线等。为便于表示，图7中仅用一个双向箭头表示，但并不表示仅有一根总线或一种类型的总线。
转发芯片1204用于通过网络接口与至少一个用户终端及其它网络单元连接，将封装好的IPv4报文或IPv6报文通过网络接口发送至用户终端。
处理器1201可能是一种集成电路芯片，具有信号的处理能力。在实现过程中，上述方法的各步骤可以通过处理器1201中的硬件的集成逻辑电路或者软件形式的指令完成。上述的处理器1201可以是通用处理器，包括中央处理器(Central Processing Unit，简称CPU)、网络处理器(Network Processor，简称NP)等；还可以是数字信号处理器(Digital Signal Processing，简称DSP)、专用集成电路(Application Specific Integrated Circuit，简称ASIC)、现成可编程门阵列(Field-Programmable Gate Array，简称FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以实现或者执行本发明实施方式中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。结合本发明实施方式所公开的方法的步骤可以直接体现为硬件译码处理器执行完成，或者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器1202，处理器1201读取存储器1202中的信息，结合其硬件完成前述实施方式的方法的步骤。
本公开实施方式还提供了一种机器可读存储介质，该机器可读存储介质存储有机器可执行指令，该机器可执行指令在被处理器调用和执行时，机器可执行指令促使处理器实现上述房颤信号的识别方法，具体实现可参见方法实施方式，在此不再赘述。
本公开实施方式提供的一种房颤信号的识别方法、装置和设备，不需要人工的特征工程，具有自动提取心电图特征的功能；采用两路VGG-16网络，但是两路网络具有不同粒度的卷积核，可以更多的捕获不同粒度的特征信息。
在本申请所提供的几个实施方式中，应该理解到，所揭露的装置和方法，也可以通过其它的方式实现。以上所描述的装置实施方式仅仅是示意 性的，例如，附图中的流程图和框图显示了根据本发明的多个实施方式的装置、方法和计算机程序产品的可能实现的体系架构、功能和操作。在这点上，流程图或框图中的每个方框可以代表一个模块、程序段或代码的一部分，所述模块、程序段或代码的一部分包含一个或多个用于实现规定的逻辑功能的可执行指令。也应当注意，在有些作为替换的实现方式中，方框中所标注的功能也可以以不同于附图中所标注的顺序发生。例如，两个连续的方框实际上可以基本并行地执行，它们有时也可以按相反的顺序执行，这依所涉及的功能而定。也要注意的是，框图和/或流程图中的每个方框、以及框图和/或流程图中的方框的组合，可以用执行规定的功能或动作的专用的基于硬件的系统来实现，或者可以用专用硬件与计算机指令的组合来实现。
另外，在本发明各个实施方式中的各功能模块或单元可以集成在一起形成一个独立的部分，也可以是各个模块单独存在，也可以两个或两个以上模块集成形成一个独立的部分。
所述功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本公开的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本公开各个实施方式所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
最后应说明的是：以上所述实施方式，仅为本公开的具体实施方式，用以说明本公开的技术方案，而非对其限制，本公开的保护范围并不局限于此，尽管参照前述实施方式对本公开进行了详细的说明，本领域的普通 技术人员应当理解：任何熟悉本技术领域的技术人员在本公开揭露的技术范围内，其依然可以对前述实施方式所记载的技术方案进行修改或可轻易想到变化，或者对其中部分技术特征进行等同替换；而这些修改、变化或者替换，并不使相应技术方案的本质脱离本公开实施方式技术方案的精神和范围，都应涵盖在本公开的保护范围之内。因此，本公开的保护范围应所述以权利要求的保护范围为准。

Claims (10)

  1. 一种房颤信号的识别方法，其特征在于，所述方法包括：
    获取待识别的心电信号；
    将所述待识别的心电信号输入至预先建立的房颤信号识别模型中，输出房颤信号的识别结果；
    其中，所述房颤信号识别模型通过下述方式建立：
    获取设定数量的心电样本信号和对应的标识信息；其中，所述标识信息包括正常信号和房颤信号的标识信息；
    根据所述正常信号的数量，采用SMOTE方式对所述房颤信号进行均衡处理；
    建立多路卷积神经网络的网络结构；每路卷积神经网络设置有特定的感受野，用于识别相应粒度的房颤信号；
    将所述正常信号和均衡处理后的房颤信号输入至所述网络结构中进行训练，生成所述房颤信号识别模型。
  2. 根据权利要求1所述的方法，其特征在于，所述多路卷积神经网络的网络结构包括第一网络和第二网络；
    所述第一网络和第二网络均为VGG-16卷积神经网络；所述第一网络和第二网络均包括多层卷积层和多层最大池化层；所述第一网络内卷积核的感受野小于或等于所述第二网络内卷积核的感受野；
    所述第一网络和所述第二网络的最末层的最大池化层相互连接；相互连接后的所述最大池化层还依次连接有多层全连接层和Softmax层。
  3. 根据权利要求1所述的方法，其特征在于，所述获取待识别的心电信号的步骤，包括：
    获取原始心电信号；
    对所述原始心电信号进行预处理，生成待识别的心电信号；所述预处理包括滤波处理和正则化处理。
  4. 根据权利要求1所述的方法，其特征在于，所述将所述正常信号和均衡处理后的房颤信号输入至所述网络结构中进行训练，生成所述房颤信号识别模型的步骤，包括：
    将所述正常信号和均衡处理后的房颤信号输入至所述网络结构中进行训练，生成初始模型；
    计算所述初始模型的灵敏度
    Figure PCTCN2017120106-appb-100001
    计算所述初始模型的特异性
    Figure PCTCN2017120106-appb-100002
    计算所述初始模型的精确度
    Figure PCTCN2017120106-appb-100003
    计算所述初始模型的准确度
    Figure PCTCN2017120106-appb-100004
    其中，TP为识别正确的房颤信号；FP为识别错误的房颤信号；TN为识别正确的正常信号；FN为识别错误的正常信号；
    判断所述灵敏度、所述特异性、所述精确度和所述准确度是否分别满足对应的阈值，如果否，调整所述网络结构中的配置参数，直至所述灵敏度、所述特异性、所述精确度和所述准确度满足对应的阈值；
    将所述初始模型确定为所述房颤信号识别模型。
  5. 一种房颤信号的识别装置，其特征在于，所述装置包括：
    信号获取模块，配置成获取待识别的心电信号；
    结果输出模块，配置成将所述待识别的心电信号输入至预先建立的房颤信号识别模型中，输出房颤信号的识别结果；
    其中，所述房颤信号识别模型通过下述方式建立：
    获取设定数量的心电样本信号和对应的标识信息；其中，所述标识信息包括正常信号和房颤信号的标识信息；
    根据所述正常信号的数量，采用SMOTE方式对所述房颤信号进行均衡处理；
    建立多路卷积神经网络的网络结构；每路卷积神经网络设置有特定的感受野，配置成识别相应粒度的房颤信号；
    将所述正常信号和均衡处理后的房颤信号输入至所述网络结构中进行训练，生成所述房颤信号识别模型。
  6. 根据权利要求5所述的装置，其特征在于，所述多路卷积神经网络的网络结构包括第一网络和第二网络；
    所述第一网络和第二网络均为VGG-16卷积神经网络；所述第一网络和第二网络均包括多层卷积层和多层最大池化层；所述第一网络内卷积核的感受野小于或等于所述第二网络内卷积核的感受野；
    所述第一网络和所述第二网络的最末层的最大池化层相互连接；相互连接后的所述最大池化层还依次连接有多层全连接层和Softmax层。
  7. 根据权利要求5所述的装置，其特征在于，所述信号获取模块，还配置成：
    获取原始心电信号；
    对所述原始心电信号进行预处理，生成待识别的心电信号；所述预处理包括滤波处理和正则化处理。
  8. 根据权利要求5所述的方法，其特征在于，所述结果输出模块，还配置成：
    将所述正常信号和均衡处理后的房颤信号输入至所述网络结构中进行训练，生成初始模型；
    计算所述初始模型的灵敏度
    Figure PCTCN2017120106-appb-100005
    计算所述初始模型的特异性
    Figure PCTCN2017120106-appb-100006
    计算所述初始模型的精确度
    Figure PCTCN2017120106-appb-100007
    计算所述初始模型的准确度
    Figure PCTCN2017120106-appb-100008
    其中，TP为识别正确的房颤信号；FP为识别错误的房颤信号；TN为识别正确的正常信号；FN为识别错误的正常信号；
    判断所述灵敏度、所述特异性、所述精确度和所述准确度是否分别满足对应的阈值，如果否，调整所述网络结构中的配置参数，直至所述灵敏度、所述特异性、所述精确度和所述准确度满足对应的阈值；
    将所述初始模型确定为所述房颤信号识别模型。
  9. 一种房颤信号的识别设备，其特征在于，包括处理器和机器可读存储介质，所述机器可读存储介质存储有能够被所述处理器执行的机器可执行指令，所述处理器执行所述机器可执行指令以实现权利要求1至4任一项所述的方法。
  10. 一种机器可读存储介质，其特征在于，所述机器可读存储介质存储有机器可执行指令，所述机器可执行指令在被处理器调用和执行时，所述机器可执行指令促使所述处理器实现权利要求1至4任一项所述的方法。
PCT/CN2017/120106 2017-12-19 2017-12-29 房颤信号的识别方法、装置和设备 WO2019119518A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP2020532967A JP7053836B2 (ja) 2017-12-19 2017-12-29 心房細動信号の識別方法、装置及び機器
US16/901,033 US11538588B2 (en) 2017-12-19 2020-06-15 Atrial fibrillation signal recognition method, apparatus and device

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711380466.9 2017-12-19
CN201711380466.9A CN108113666B (zh) 2017-12-19 2017-12-19 房颤信号的识别方法、装置和设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/901,033 Continuation US11538588B2 (en) 2017-12-19 2020-06-15 Atrial fibrillation signal recognition method, apparatus and device

Publications (1)

Publication Number Publication Date
WO2019119518A1 true WO2019119518A1 (zh) 2019-06-27

Family

ID=62229643

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/120106 WO2019119518A1 (zh) 2017-12-19 2017-12-29 房颤信号的识别方法、装置和设备

Country Status (4)

Country Link
US (1) US11538588B2 (zh)
JP (1) JP7053836B2 (zh)
CN (1) CN108113666B (zh)
WO (1) WO2019119518A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113749668A (zh) * 2021-08-23 2021-12-07 华中科技大学 一种基于深度神经网络的可穿戴心电图实时诊断系统

Families Citing this family (24)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110897629A (zh) * 2018-09-14 2020-03-24 杭州脉流科技有限公司 基于深度学习算法的心电特征提取方法、装置、系统、设备和分类方法
CN109171712B (zh) * 2018-09-28 2022-03-08 东软集团股份有限公司 心房颤动识别方法、装置、设备及计算机可读存储介质
CN109871808B (zh) * 2019-02-21 2021-03-02 天津惊帆科技有限公司 房颤模型训练、检测方法和装置
CN110037685A (zh) * 2019-04-01 2019-07-23 上海数创医疗科技有限公司 具有改进卷积神经网络识别算法的便携式心电图仪
CN110037683B (zh) * 2019-04-01 2020-12-29 上海数创医疗科技有限公司 用于识别心律类型的改进卷积神经网络及其训练方法
CN110037682B (zh) * 2019-04-01 2020-07-21 上海数创医疗科技有限公司 基于改进卷积神经网络的识别心律类型的方法
CN110037684A (zh) * 2019-04-01 2019-07-23 上海数创医疗科技有限公司 基于改进卷积神经网络的识别心律类型的装置
CN110493072A (zh) * 2019-07-11 2019-11-22 网宿科技股份有限公司 基于深度学习的带宽滤波方法、装置、服务器及存储介质
US11005689B2 (en) 2019-07-11 2021-05-11 Wangsu Science & Technology Co., Ltd. Method and apparatus for bandwidth filtering based on deep learning, server and storage medium
CN110688942B (zh) * 2019-09-25 2023-05-26 西安邮电大学 一种基于InResNet网络的心电信号联合识别模块及方法
CN110619322A (zh) * 2019-10-18 2019-12-27 武汉大学 一种基于多流态卷积循环神经网络的多导联心电异常信号识别方法及系统
US11596344B2 (en) * 2019-10-30 2023-03-07 Tencent America LLC Deep neural network on ECG Poincare plot for atrial fibrillation classification
CN112826514B (zh) * 2019-11-22 2022-07-22 华为技术有限公司 一种房颤信号的分类方法、装置、终端以及存储介质
CN111329445B (zh) * 2020-02-20 2023-09-15 广东工业大学 基于组卷积残差网络和长短期记忆网络的房颤识别方法
CN111671416B (zh) * 2020-05-29 2022-11-29 中国科学院深圳先进技术研究院 心电信号滤波方法及装置
CN111700609A (zh) * 2020-07-27 2020-09-25 郑州大学 基于短时心电信号的房颤检测方法、装置及设备
CN111904411B (zh) * 2020-08-25 2022-09-23 浙江工业大学 基于多尺度特征提取的多导联心跳信号分类方法及装置
CN112932497A (zh) * 2021-03-10 2021-06-11 中山大学 一种不平衡单导联心电数据分类方法及系统
CN113080992A (zh) * 2021-03-30 2021-07-09 北京芯动卫士科技有限公司 一种基于dcgan和深度神经网络方法的房颤识别方法
CN113425272B (zh) * 2021-08-02 2024-02-20 北京雪扬科技有限公司 一种通过穿戴设备采集数据分析房颤、室颤的方法
CN114098754B (zh) * 2021-10-29 2024-07-16 西安交通大学 一种房颤信号预处理方法、检测系统、设备及存储介质
CN114469133B (zh) * 2021-12-14 2023-10-03 中国科学院深圳先进技术研究院 一种无扰式房颤监测方法
CN114176600B (zh) * 2021-12-28 2023-10-20 上海交通大学 基于因果分析的心电图st段异常判别系统
CN117257324B (zh) * 2023-11-22 2024-01-30 齐鲁工业大学(山东省科学院) 基于卷积神经网络和ecg信号的房颤检测方法

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105726018A (zh) * 2016-02-06 2016-07-06 河北大学 一种与rr间期无关的房颤自动检测方法
US20170032221A1 (en) * 2015-07-29 2017-02-02 Htc Corporation Method, electronic apparatus, and computer readable medium of constructing classifier for disease detection
CN107391900A (zh) * 2017-05-05 2017-11-24 陈昕 房颤检测方法、分类模型训练方法及终端设备

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9775533B2 (en) * 2013-03-08 2017-10-03 Singapore Health Services Pte Ltd System and method of determining a risk score for triage
CN104970789B (zh) * 2014-04-04 2017-12-19 中国科学院苏州纳米技术与纳米仿生研究所 心电图分类方法及系统
CN106570564B (zh) 2016-11-03 2019-05-28 天津大学 基于深度网络的多尺度行人检测方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20170032221A1 (en) * 2015-07-29 2017-02-02 Htc Corporation Method, electronic apparatus, and computer readable medium of constructing classifier for disease detection
CN105726018A (zh) * 2016-02-06 2016-07-06 河北大学 一种与rr间期无关的房颤自动检测方法
CN107391900A (zh) * 2017-05-05 2017-11-24 陈昕 房颤检测方法、分类模型训练方法及终端设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
BRUSER, C. ET AL.: "Automatic Detection of Atrial Fibrillation in Cardiac Vibration Signals", IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, vol. 17, no. 1, 16 October 2012 (2012-10-16), pages 162 - 171, XP011494169, ISSN: 1089-1089 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113749668A (zh) * 2021-08-23 2021-12-07 华中科技大学 一种基于深度神经网络的可穿戴心电图实时诊断系统
CN113749668B (zh) * 2021-08-23 2022-08-09 华中科技大学 一种基于深度神经网络的可穿戴心电图实时诊断系统

Also Published As

Publication number Publication date
CN108113666A (zh) 2018-06-05
US11538588B2 (en) 2022-12-27
CN108113666B (zh) 2019-06-21
JP7053836B2 (ja) 2022-04-12
US20200312459A1 (en) 2020-10-01
JP2021506421A (ja) 2021-02-22

Similar Documents

Publication Publication Date Title
WO2019119518A1 (zh) 房颤信号的识别方法、装置和设备
KR102451795B1 (ko) 심전도 신호의 검출 방법
US10869610B2 (en) System and method for identifying cardiac arrhythmias with deep neural networks
WO2021017313A1 (zh) 房颤的检测方法、装置、计算机设备及存储介质
CN111163690B (zh) 心律失常的检测方法、装置、电子设备及计算机存储介质
CN109171712B (zh) 心房颤动识别方法、装置、设备及计算机可读存储介质
CN104970789B (zh) 心电图分类方法及系统
WO2019100562A1 (zh) 基于人工智能的心电图心搏自动识别分类方法
CN104706349B (zh) 一种基于脉搏波信号的心电信号构建方法
CN106377247B (zh) 基于特征选择的心律失常分类方法
CN109009084B (zh) 多导联心电信号的qrs波群校验方法、装置及设备、介质
CN113080990B (zh) 一种基于CycleGAN和BiLSTM神经网络方法的心博异常检测方法
JP6583892B2 (ja) Ｅｃｇ信号の並行解析装置、方法及び移動端末
JP7119012B2 (ja) 不整脈に関係する心拍を分類するためのシステム
CN106580307A (zh) 心电图的质量判断方法及质量判断系统
CN115177262B (zh) 一种基于深度学习的心音心电联合诊断装置和系统
CN106485213A (zh) 一种利用心电信号进行自动身份识别的特征提取方法
US10743789B2 (en) ECG signal parallel analysis apparatus, method and mobile terminal
KR102437349B1 (ko) 어텐션 메커니즘을 이용한 부정맥 탐지 장치
WO2020155169A1 (zh) 房颤筛查的方法和装置
CN114519370A (zh) 一种基于深度学习早停机制的左心室肥大检测方法及系统
CN116528763A (zh) 生物信号解析系统
WO2021068107A1 (zh) 基于心冲击信号的身份识别方法、电子设备及存储介质
Shaik et al. A novel approach for QRS delineation in ECG signal based on chirplet transform
CN114652324B (zh) 一种单心拍分类方法、装置、设备及计算机可读存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17935463

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2020532967

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 19.11.2020)

122 Ep: pct application non-entry in european phase

Ref document number: 17935463

Country of ref document: EP

Kind code of ref document: A1