CN111832586A - Deep learning data preprocessing method and device and training system - Google Patents
Deep learning data preprocessing method and device and training system Download PDFInfo
- Publication number
- CN111832586A CN111832586A CN201910307795.3A CN201910307795A CN111832586A CN 111832586 A CN111832586 A CN 111832586A CN 201910307795 A CN201910307795 A CN 201910307795A CN 111832586 A CN111832586 A CN 111832586A
- Authority
- CN
- China
- Prior art keywords
- waveform
- neural network
- deep neural
- input
- interval
- Prior art date
- Legal status (The legal status 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 status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
Abstract
The invention discloses a deep learning data preprocessing method, a deep learning data preprocessing device and a deep learning data training system. The deep learning data preprocessing method comprises the following steps: collecting the Kth complete heart beat waveform of the ECG waveform, and resampling to N points; directly taking the N points in the step as 1 heartbeat waveform, or carrying out amplitude normalization processing on the N points in the step to obtain 1 heartbeat waveform; continuously collecting M heart beat waveforms as an input sample; and inputting the input sample into a CNN deep neural network model, or inputting the input sample into an RNN deep neural network model, or adopting a CRNN deep neural network model. According to the invention, through the principle, the collected ECG waveform is processed and then sent into the deep neural network for training, so that the purpose that the deep neural network can be used by adopting few training samples is realized, and the training difficulty of the deep neural network in the medical field is greatly reduced.
Description
Technical Field
The invention relates to the field of detection, in particular to a deep learning data preprocessing method, a deep learning data preprocessing device and a deep learning data preprocessing training system.
Background
In recent years, with the continuous development of information technology, Deep Neural Network (Deep Neural Network) based recognition methods have been successful in the classification field, and the application field is also expanded to various industries. The deep neural network training process generally includes inputting training samples into a deep neural network model, comparing classification results output through deep neural network processing with training sample values to obtain network loss calculated by a loss function, then returning the network loss to the deep neural network, correcting parameters of each layer on the deep neural network, repeating the steps until the network loss meets a certain convergence condition, considering that an optimization target of the deep neural network is achieved, and finishing training. Network losses are typically minimized as an optimization goal. When the deep neural network is trained in the existing medical field, the training data is mostly input one by one from the beginning of the whole collected data until the end input is completed, and the available effect can be realized only by a large number of training data samples by adopting the training data input mode. However, medical data needs to be acquired based on human body, especially some information for judging diseases is not easy to acquire, the workload of acquisition and marking is large due to the particularity of the medical data, and the difficulty is increased for the deep neural network in the medical field to train to be usable.
Disclosure of Invention
The invention aims to solve the technical problem that few training samples are adopted to enable a deep neural network to achieve the purpose of being usable, and aims to provide a data preprocessing method, a device and a training system for deep learning.
The invention is realized by the following technical scheme:
in a first aspect, the invention discloses a data preprocessing method for deep learning, and the data preprocessing method is used for processing
The later data is used for being sent into a deep neural network for training, and the classification of heart beats is realized, and the method comprises the following steps:
collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t; directly taking the N points in the step as 1 input waveform, or carrying out amplitude normalization processing on the N points in the step to obtain 1 input waveform; continuously acquiring M input waveforms as an input sample, wherein M is more than 0; and inputting the input sample into a CNN deep neural network model in an Nx 1 xM three-dimensional form, or decomposing the input sample into M Nx 1 point correspondences and sending the M Nx 1 point correspondences into M RNN basic neural units on the RNN deep neural network model, or decomposing the input sample into M Nx 1 x 1 point correspondences and sending the M N point correspondences into M identical CNN deep neural network models by adopting the CRNN deep neural network model and then sending the M RNN basic neural units on the RNN deep neural network model.
Preferably, the kth full heartbeat waveform of the ECG waveform is acquired and resampled to N points by the following method: acquiring a section of waveform in a first T1 time period and a second T2 time period of a Kth R position of the ECG waveform, resampling N points, wherein 0< N is less than or equal to Fs T, T is T1+ T2, the time period T1 before the R position at least comprises a section of waveform of P wave, and the time period T2 after the R position at least comprises a section of waveform of T wave.
In a second aspect, the invention discloses another deep learning data preprocessing method, wherein data processed by the method is sent to a deep neural network for training to realize cardiac rhythm classification and/or cardiac beat classification, and the method comprises the following steps:
s101: collecting the Kth complete heart beat waveform of ECG waveform, and resampling to N points, 0<N is less than or equal to Fs t; s102: acquiring the interval from the K-1 th to the K-1 th R on the ECG waveform or acquiring the interval from the K-1 th to the K +1 th R on the ECG waveform; s103: the 1 resampled input waveform in step S101 and the 1 RR interval in step S102 are comparedRespectively carrying out normalization treatment, and then combining into H to form a one-dimensional vector of N + 1; or directly combining the 1 resampled input waveform in the step S101 and the 1 RR interval in the step S102 into H to form a one-dimensional vector of N + 1; or performing normalization processing on the 1 heart beat waveform after resampling in the step S101 and the 1 RR interval alternative in the step S102, and combining the two waveforms into H to form a one-dimensional vector of N + 1; s104: taking continuous M H in the step S103 as an input sample, wherein M is more than 0; s105: inputting one input sample in the step S104 into the CNN deep neural network model in a three-dimensional form of (N +1) × 1 × M, or decomposing the input sample into M (N +1) × 1 points, and sending the points into M RNN basic neural units on the RNN deep neural network model; or adopting a CRNN deep neural network model to decompose the input sample into M input waveforms of Nx 1 x 1 and M RR intervals of 1 x 1, correspondingly sending the M input waveforms of Nx 1 x 1 into M identical CNN deep neural network models for processing, and obtaining M results PMM results PMCombining with M1 × 1 RR intervals to form M combinations, and sending into M RNN basic neural units on RNN deep neural network model, wherein each combination contains 1 result PMAnd 1 RR interval of 1 × 1.
Preferably, the kth full heartbeat waveform of the ECG waveform is acquired and resampled to N points by the following method: acquiring a section of waveform in a first T1 time period and a second T2 time period of a Kth R position of the ECG waveform, resampling N points, wherein 0< N is less than or equal to Fs T, T is T1+ T2, the time period T1 before the R position at least comprises a section of waveform of P wave, and the time period T2 after the R position at least comprises a section of waveform of T wave.
In a third aspect, the invention discloses a data preprocessing device, wherein data processed by the device is used for being sent to a deep neural network for training to realize heart beat classification, and the data preprocessing device comprises a waveform acquisition and processing unit: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t; a normalization processing unit: directly taking N points obtained by the waveform acquisition unit as 1 input waveform, or carrying out amplitude normalization processing on the N points obtained by the waveform acquisition unit to obtain 1 input waveform; a sample forming unit: continuously acquiring M input waveforms as an input sample, wherein M is more than 0; an input sample format processing unit: and inputting the input sample into a CNN deep neural network model in an Nx 1 xM three-dimensional form, or decomposing the input sample into M Nx 1 point correspondences and sending the M Nx 1 point correspondences into M RNN basic neural units on the RNN deep neural network model, or decomposing the input sample into M Nx 1 x 1 point correspondences and sending the M N point correspondences into M identical CNN deep neural network models by adopting the CRNN deep neural network model and then sending the M RNN basic neural units on the RNN deep neural network model.
Preferably, the kth full heartbeat waveform of the ECG waveform is acquired and resampled to N points by the following method: acquiring a section of waveform in a first T1 time period and a second T2 time period of a Kth R position of the ECG waveform, resampling N points, wherein 0< N is less than or equal to Fs T, T is T1+ T2, the time period T1 before the R position at least comprises a section of waveform of P wave, and the time period T2 after the R position at least comprises a section of waveform of T wave.
In a fourth aspect, the present invention discloses another data preprocessing device, where data processed by the device is used to be sent to a deep neural network for training to realize cardiac rhythm classification and/or cardiac beat classification, and the data preprocessing device includes a waveform acquisition processing unit: for acquiring the Kth complete cardiac waveform of ECG waveform, resampling to N points, 0<N is less than or equal to Fs t; an R interval acquisition unit: for acquiring the interval from the K-1 th to the Kth R on the ECG waveform, or the interval from the K-1 th to the K +1 th R on the ECG waveform, or for acquiring the interval from the K-1 th to the K +1 th R on the ECG waveform; a normalization processing unit: the device is used for respectively carrying out normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval in the RR interval acquisition unit, and then combining the normalized input waveform and the RR interval into H to form a one-dimensional vector of N + 1; or directly combining 1 resampled input waveform in the waveform acquisition processing unit and 1 RR interval in the RR interval acquisition unit into H to form a one-dimensional vector of N + 1; or performing normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval alternative in the R interval acquisition unit, and combining the waveforms into H to form a one-dimensional vector of N + 1; a sample forming unit:for taking M consecutive H's in the normalization processing unit as an input sample, M > 0; an input sample format processing unit: the system comprises M RNN basic neural units, a CNN deep neural network model, a sample forming unit, a data processing unit and a data processing unit, wherein the CNN basic neural network model is used for inputting one input sample in the sample forming unit into the CNN deep neural network model in a three-dimensional form of (N +1) multiplied by 1 multiplied by M, or correspondingly sending the input sample into the M RNN basic neural units on the RNN deep neural network model by decomposing the input sample into M points of (N +1) multiplied; or adopting a CRNN deep neural network model to decompose the input sample into M input waveforms of Nx 1 x 1 and M RR intervals of 1 x 1, correspondingly sending the M input waveforms of Nx 1 x 1 into M identical CNN deep neural network models for processing, and obtaining M results PMM results PMCombining with M1 × 1 RR intervals to form M combinations, and sending into M RNN basic neural units on RNN deep neural network model, wherein each combination contains 1 result PMAnd 1 RR interval of 1 × 1.
In a fifth aspect, the present invention discloses a training system comprising
A data preprocessing device: the method is used for extracting characteristic points on an ECG waveform to form an input sample, and then processing the input sample into a specific form; and the training module is used for processing the data preprocessing device into input samples in a specific form, sending the input samples into the corresponding deep neural network model for training, and adjusting the filter weight in the deep neural network model until the loss function judges that the adjusted filter weight meets the requirement.
Compared with the prior art, the invention has the following advantages and beneficial effects:
the invention selects the waveform which can reserve the necessary characteristics to carry out the resampling normalization processing, or selects the waveform which can reserve the necessary characteristics to carry out the resampling normalization processing and the normalization processing combination on the RR interphase, or directly selects the waveform which can reserve the necessary characteristics, the RR interphase and the combination of the two, then further processes the data into a specific form, sends the data into a corresponding deep neural network to carry out the training, and can adopt the same deep neural network model even with different sampling rates without replacing the deep neural network model, thus realizing the automatic analysis of the electrocardiogram, such as the heart rhythm classification or the heart beat classification, eliminating the interference of unnecessary waveforms and unnecessary characteristics to the deep neural network training, accelerating the training speed of the deep neural network, and realizing the purpose of adopting few training samples to enable the deep neural network to be available, the training difficulty of the deep neural network in the medical field is greatly reduced.
Drawings
The accompanying drawings, which are included to provide a further understanding of the embodiments of the invention and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the invention and together with the description serve to explain the principles of the invention. In the drawings:
FIG. 1 is a flow chart of the present invention;
FIG. 2 is another flow chart of the present invention;
FIG. 3 is a block diagram of a data preprocessing apparatus;
FIG. 4 is a block diagram of another structure of the data preprocessing apparatus;
FIG. 5 is a block diagram of another configuration of a training system;
FIG. 6 is a schematic diagram of an ECG waveform extraction waveform connection as an input;
FIG. 7 is a schematic block diagram of a CNN type deep neural network structure;
FIG. 8 is a schematic block diagram of a CNN structure with a heart beat waveform as a model input;
FIG. 9 is a schematic block diagram of an RNN architecture with a heart beat waveform as a model input;
FIG. 10 is a schematic block diagram of a CRNN architecture with a heart beat waveform as the model input;
FIG. 11 is a schematic diagram of a CNN structure with a combination of cardiac waveforms and RR intervals as inputs to the model;
FIG. 12 is a schematic block diagram of the RNN structure with cardioversion waveform and RR interval combinations as model inputs;
FIG. 13 is a schematic block diagram of a CRNN structure with a combination of heartbeat waveforms and RR intervals as model inputs;
FIG. 14 is a full heartbeat waveform.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to examples and accompanying drawings, and the exemplary embodiments and descriptions thereof are only used for explaining the present invention and are not meant to limit the present invention.
Example 1:
as shown in FIG. 1, the present invention proposes a deep learning data preprocessing method by which
The processed data is sent into a deep neural network for training to realize heart beat classification, which comprises the following steps
The method comprises the following steps:
as shown in fig. 6, step S1: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t; fs is a sampling rate capable of storing necessary waveform characteristics, Fs can be obtained by a person skilled in the art according to empirical values, K is a positive integer, T has a value range including time for a complete heartbeat waveform, and a complete heartbeat waveform is a segment of waveform including a P wave, a QRS complex, and a T wave, as shown in fig. 14; wherein the ECG waveform can be a raw ECG waveform or a filtered ECG waveform. The resampling is to change the sampling rate of the selected waveform, so that the number of sampling points of each input waveform is the same, and the deep neural network can identify according to the selected waveform sample.
Step S2: directly taking the N points in the step as 1 input waveform, or carrying out amplitude normalization processing on the N points in the step to obtain 1 input waveform; the amplitude normalization processing can be selected as dividing the value of the sampling point by the value corresponding to the 1mv voltage, and other methods can also be adopted for normalization processing.
Step S3: m of the input waveforms are continuously collected as one input sample, M > 0.
As shown in fig. 7-10, step S4: inputting the one input sample into the CNN deep neural network model in a three-dimensional form of N × 1 × M, where N represents waveforms of N points, and M represents M waveforms having N points, and outputting a two-dimensional result 1 × M, as shown in fig. 8; or decomposing the input sample into M N × 1 points, sending the M N basic neural units to the RNN deep neural network model, where each RNN basic neural unit corresponds to one N × 1 point, and outputting M1 × 1 values, as shown in fig. 9; or, using a CRNN deep neural network model, decomposing the input sample into M N × 1 × 1 points, sending the M N × 1 points into M identical CNN deep neural network models, processing the M identical CNN deep neural network models, and sending the M RNN basic neural units on the RNN deep neural network model, where each identical CNN deep neural network model correspondingly inputs an N × 1 point, each RNN basic neural unit on the RNN deep neural network model correspondingly inputs a CNN deep neural network model output value, and outputs M two-dimensional 1 × 1 values, as shown in fig. 10. N in step S4 each represents a waveform of N points, and M each represents M waveforms having N points.
The method for acquiring a complete heartbeat waveform is various, wherein preferably, the Kth complete heartbeat waveform of the ECG waveform is acquired and resampled to N points, and the method comprises the following steps: acquiring a section of waveform in a first T1 time period and a second T2 time period of a Kth R position of the ECG waveform, resampling N points, wherein 0< N is less than or equal to Fs T, T is T1+ T2, the time period T1 before the R position at least comprises a section of waveform of P wave, and the time period T2 after the R position at least comprises a section of waveform of T wave. The R position of the ECG waveform refers to the peak position of the R wave, Fs is a sampling rate that can preserve the necessary waveform characteristics, Fs can be obtained by one skilled in the art from empirical values, and K and N are positive integers. The method selects the waveform, and is simple and convenient to operate.
The invention selects the waveform capable of retaining the necessary characteristics and carries out resampling normalization processing or directly selects the waveform capable of retaining the necessary characteristics, processes the data into a specific form and then sends the specific form into the corresponding deep neural network for training, and even if the sampling rates are different, the same deep neural network model can be adopted, the deep neural network model does not need to be replaced, the automatic analysis of the electrocardiogram can be realized, and the heart beat classification can be realized.
Example 2:
as shown in fig. 2, the present invention further provides another deep learning data preprocessing method, in which the data processed by the method is sent to a deep neural network for training to implement cardiac rhythm classification and/or cardiac beat classification, and the method includes the following steps:
s101: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t; fs is a sampling rate capable of storing necessary waveform characteristics, Fs can be obtained by a person skilled in the art according to empirical values, K is a positive integer, T has a value range including the time for a complete heartbeat waveform, and the complete heartbeat waveform is a segment of waveform including a P wave, a QRS complex and a T wave; wherein the ECG waveform can be a raw ECG waveform or a filtered ECG waveform. The resampling is to change the sampling rate of the selected waveform, so that the number of sampling points of each input waveform is the same, and the deep neural network can identify according to the selected waveform sample.
S102: acquiring the interval from the K-1 th to the K-1 th R on the ECG waveform or acquiring the interval from the K-1 th to the K +1 th R on the ECG waveform; wherein the RR interval refers to the peak-to-peak time difference between two adjacent R waves. Wherein M is a positive integer greater than 0.
S103: respectively carrying out normalization processing on the 1 resampled input waveform in the step S101 and the 1 RR interval in the step S102, and then combining the input waveforms into H to form a one-dimensional vector of N + 1; or directly combining the 1 resampled input waveform in the step S101 and the 1 RR interval in the step S102 into H to form a one-dimensional vector of N + 1; or normalizing the 1 resampled input waveform in the step S101 and the 1 RR interval in the step S102, and combining the normalized input waveforms into H to form a one-dimensional vector of N + 1;
wherein, the N points in the step S101 are processed by amplitude normalization, and the processing mode can be selected as dividing the value of the sampling point by the value corresponding to the 1mv voltage. The M RR intervals in step S102 are normalized by dividing the RR intervals by the sampling rate Fs. Fs is a sampling rate that preserves the necessary waveform characteristics, and Fs can be derived empirically by one skilled in the art.
S104: taking continuous M H in the step S103 as an input sample, wherein M is more than 0;
as shown in fig. 11 to 13, S105: inputting one input sample in the step S104 into the CNN deep neural network model in a three-dimensional form of (N +1) × 1 × M, with the output size being one of three cases, i.e., 1 × M, 1 × 1 or 1 × M +1, as shown in fig. 11; or decomposing the input sample into M points (N +1) x 1, sending the points into M RNN basic neural units on the RNN deep neural network model, inputting one point (N +1) x 1 to each RNN basic neural unit, outputting M values (1 x 1) or outputting Y2 with size of 1 x 1, and outputting Y2MY2 is the waveform output value and RR interval output value, and Y can be output simultaneouslyMAnd Y2, as shown in fig. 12; or adopting a CRNN deep neural network model, decomposing the input sample into M Nx 1 x 1 input waveforms in a three-dimensional form and M1 x 1 RR intervals, correspondingly sending the M Nx 1 x 1 input waveforms in the three-dimensional form into M identical CNN deep neural network models for processing, and obtaining M results PMM results PMCombining with M1 × 1 RR intervals to form M combinations, and sending into M RNN basic neural units on RNN deep neural network model, wherein each combination contains 1 result PMAnd 1 RR interval of 1 × 1, inputting a combination for each RNN basic neural unit on the RNN deep neural network model, outputting M1 × 1 values representing the corresponding output of each heartbeat, and recording as Y1……YM(ii) a Or 1 value of 1 × 1 is output, which represents the output of a segment of waveform corresponding to the input heartbeat and is marked as Y2(ii) a Y can also be output simultaneously1……YMAnd Y2As shown in fig. 13. In step S105, (N +1) represents N to represent the number of input waveform points corresponding to one heartbeat, and M represents the number of corresponding heartbeats. The basic RNN unit structure includes conventional RNN, LSTM or gated cyclic unit GRU.
Preferably, the kth full heartbeat waveform of the ECG waveform is acquired and resampled to N points by the following method: acquiring a section of waveform in a first T1 time period and a second T2 time period of a Kth R position of the ECG waveform, resampling N points, wherein 0< N is less than or equal to Fs T, T is T1+ T2, the time period T1 before the R position at least comprises a section of waveform of P wave, and the time period T2 after the R position at least comprises a section of waveform of T wave. The R position of the ECG waveform refers to the peak position of the R wave, Fs is a sampling rate that can preserve the necessary waveform characteristics, Fs can be obtained by one skilled in the art from empirical values, and K and N are positive integers. The method selects the waveform, and is simple and convenient to operate.
The invention selects the waveform which can keep necessary characteristics to carry out resampling normalization processing and carry out normalization processing combination on RR interphase, or directly selecting the combination of waveform and RR interval capable of retaining necessary characteristics, further processing the data into specific form, sending into corresponding deep neural network for training, even if different sampling rates are adopted, the same deep neural network model can be adopted without replacing the deep neural network model, the automatic analysis of the electrocardiogram can be realized, such as the classification of heart rhythm or heart beat, the invention eliminates the interference of unnecessary waveforms and unnecessary features on deep neural network training, accelerates the training speed of the deep neural network, realizes the purpose of using the deep neural network by adopting few training samples, and greatly reduces the difficulty of deep neural network training in the medical field.
Example 3:
as shown in fig. 3, the invention also discloses a data preprocessing device, the data processed by the method is used for being sent to a deep neural network for training, so as to realize heart beat classification, and the data preprocessing device comprises a processor, a memory and a waveform acquisition and processing unit: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t; fs is a sampling rate capable of storing necessary waveform characteristics, Fs can be obtained by a person skilled in the art according to empirical values, K is a positive integer, T has a value range including the time for a complete heartbeat waveform, and the complete heartbeat waveform is a segment of waveform including a P wave, a QRS complex and a T wave; wherein the ECG waveform can be a raw ECG waveform or a filtered ECG waveform. The resampling is to change the sampling rate of the selected waveform, so that the number of sampling points of each input waveform is the same, and the deep neural network can identify according to the selected waveform sample.
A normalization processing unit: directly taking N points obtained by the waveform acquisition unit as 1 input waveform, or carrying out amplitude normalization processing on the N points obtained by the waveform acquisition unit to obtain 1 input waveform;
a sample forming unit: continuously acquiring M input waveforms as an input sample, wherein M is more than 0;
an input sample format processing unit: and inputting the input sample into a CNN deep neural network model in an Nx 1 xM three-dimensional form, or decomposing the input sample into M Nx 1 point correspondences and sending the M Nx 1 point correspondences into M RNN basic neural units on the RNN deep neural network model, or decomposing the input sample into M Nx 1 x 1 point correspondences and sending the M N point correspondences into M identical CNN deep neural network models by adopting the CRNN deep neural network model and then sending the M RNN basic neural units on the RNN deep neural network model. The basic RNN unit structure includes conventional RNN, LSTM or gated cyclic unit GRU.
The device selects the waveform capable of retaining the necessary characteristics and carries out resampling normalization processing or directly selects the waveform capable of retaining the necessary characteristics, processes data into a specific form and then sends the specific form into the corresponding deep neural network for training, even if the sampling rates are different, the same deep neural network model can be adopted, the deep neural network model does not need to be replaced, the automatic analysis of the electrocardiogram can be realized, the heart beat classification is realized, the interference of unnecessary waveforms on the deep neural network training is eliminated, the training progress of the deep neural network is accelerated, the automatic analysis of the electrocardiogram is realized, the invention realizes the purpose that the deep neural network can be used by adopting few training samples, the deep neural network training difficulty in the medical field is greatly reduced, and a doctor classifies the heart beats through automatic analysis without manually classifying the heart beats, the detection of heart beat related diseases can be realized, such as premature ventricular contraction (PVC for short), left bundle branch block, right bundle branch block and the like.
Example 4:
as shown in fig. 4, the invention also discloses another data preprocessing device, the data processed by the device is used for being sent into a deep neural network for training to realize cardiac rhythm classification and/or cardiac beat classification, and the device comprises a processor, a memory, a data acquisition module, a data processing module and a data processing module,
A waveform acquisition processing unit: the method is used for collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, and enabling 0< N to be less than or equal to Fs t; an R interval acquisition unit: for acquiring the interval from the K-1 th to the Kth R on the ECG waveform, or the interval from the K-1 th to the K +1 th R on the ECG waveform, or for acquiring the interval from the K-1 th to the K +1 th R on the ECG waveform;
a normalization processing unit: the device is used for respectively carrying out normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval in the RR interval acquisition unit, and then combining the normalized input waveform and the RR interval into H to form a one-dimensional vector of N + 1; or directly combining 1 resampled input waveform in the waveform acquisition processing unit and 1 RR interval in the RR interval acquisition unit into H to form a one-dimensional vector of N + 1; or performing normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval alternative in the R interval acquisition unit, and combining the waveforms into H to form a one-dimensional vector of N + 1;
a sample forming unit: for taking M consecutive H's in the normalization processing unit as an input sample, M > 0;
an input sample format processing unit: the system comprises M RNN basic neural units, a CNN deep neural network model, a sample forming unit, a data processing unit and a data processing unit, wherein the CNN basic neural network model is used for inputting one input sample in the sample forming unit into the CNN deep neural network model in a three-dimensional form of (N +1) multiplied by 1 multiplied by M, or correspondingly sending the input sample into the M RNN basic neural units on the RNN deep neural network model by decomposing the input sample into M points of (N +1) multiplied; or adopting a CRNN deep neural network model to decompose the input sample into M input waveforms of Nx 1 x 1 and M RR intervals of 1 x 1, correspondingly sending the M input waveforms of Nx 1 x 1 into M identical CNN deep neural network models for processing, and obtaining M results PMM results PMCombining with M1 × 1 RR intervals to form M combinations, and sending into M RNN basic neural units on RNN deep neural network model, wherein each combination contains 1 result PMAnd 1 RR interval of 1 × 1. Wherein the RNN basic nerve unit structureIncluding conventional RNNs, LSTM or gated loop units GRU, etc.
The device selects the waveform capable of retaining the necessary characteristics to carry out resampling normalization processing and carry out normalization processing combination on RR interphase, or directly selects the combination of the waveform capable of retaining the necessary characteristics and the RR interphase, then further processes the data into a specific form, and sends the data into the corresponding deep neural network for training, even if the sampling rates are different, the same deep neural network model can be adopted, the deep neural network model does not need to be replaced, the automatic analysis of the electrocardiogram can be realized, the heart rhythm classification or the heart beat classification can be realized, the invention eliminates the interference of unnecessary waveforms and unnecessary characteristics on the deep neural network training, accelerates the training speed of the deep neural network, realizes the purpose of using few training samples to enable the deep neural network to be usable, and greatly reduces the training difficulty of the deep neural network in the medical field, the doctor can realize the detection of the diseases related to the heart beat, the diseases related to the rhythm and the two diseases simultaneously without manually classifying the heart rate or classifying the heart beat through automatically analyzing the heart rate or the heart beat, wherein the heart beat classification corresponds to the diseases related to the heart beat, and the heart rate classification corresponds to the diseases related to the rhythm. Heart beat related diseases, such as Premature Ventricular Contraction (PVC), left bundle branch block, right bundle branch block, etc. Rhythm-related diseases such as atrial fibrillation, atrial flutter, arrhythmia, etc.
Example 5:
as shown in fig. 5: the invention discloses a training system, which comprises a processor, a memory, a data preprocessing device and a training module, wherein
A data preprocessing device: the method is used for extracting characteristic points used for judging diseases on an ECG waveform to form an input sample, and then processing the input sample into a specific form;
and the training module is used for processing the data preprocessing device into input samples in a specific form, sending the input samples into the corresponding deep neural network model for training, and adjusting the filter weight in the deep neural network model until the loss function judges that the adjusted filter weight meets the requirement.
Wherein the deep neural network model can be a CNN model, an RNN model or a CRNN model. As shown in fig. 7-13, for example, a CNN model inputs X × 1 × Z input samples, which are transmitted from a data preprocessing apparatus one by one, through an input layer, where X is a positive integer according to actual needs, Z is the number of layers, and 4 layers are taken here; then, the first convolution layer f is 5 × 1 × Z × 32, that is, the number of filters is 32, the size of the filters is 5 × 1 × Z, the step length s of movement is 1, and the output result is a three-dimensional form of X × 1 × 32; then, the first maximum pooling layer f is 2 × 1, namely, the window size is 2 × 1, the moving step length s is 2, and the output pooling result is in a three-dimensional form of X/2 × 1 × 32; then, the second convolution layer f is 5 × 1 × 32 × 64, that is, the number of filters is 64, the size of the filters is 5 × 1 × 32, the step length s of movement is 1, and the output result is a three-dimensional form of X/2 × 1 × 64; then, the second maximum pooling layer f is 2 × 1, namely the window size is 2 × 1, the moving step length s is 2, and the output pooling result is in a three-dimensional form of X/4 × 1 × 64; and finally, unfolding the three-dimensional form into a one-dimensional form, sequentially obtaining 256 one-dimensional vectors and data with the size of VxW through two full-connection layers, wherein V represents the classification number, and W represents the number of output results, and then outputting the data with the final result of 1 xW after maximum value or threshold value judgment, wherein the convolutional layer, the maximum pooling layer and the full-connection layers have the same effect as each layer in the existing deep neural network, belong to the prior art, and are not repeated here. A RNN recurrent neural network is provided, the basic neural unit is RNN unit, there are M RNN units, M is positive integer. A CRNN model structure comprises a convolutional neural network CNN and a cyclic neural network RNN, wherein the convolutional neural network comprises 2 convolutional layers, maximum pooling is carried out after each convolutional layer, and the convolutional layers are transmitted to the cyclic neural network RNN through a full connection layer after the last maximum pooling is carried out, as shown in a part, not in a dotted frame, of the CNN network structure in fig. 7.
Wherein the data preprocessing device comprises
A waveform acquisition processing unit: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t;
a normalization processing unit: directly taking N points obtained by the waveform acquisition unit as 1 input waveform, or carrying out amplitude normalization processing on the N points obtained by the waveform acquisition unit to obtain 1 input waveform;
a sample forming unit: continuously acquiring M input waveforms as an input sample, wherein M is more than 0;
an input sample format processing unit: and inputting the input sample into a CNN deep neural network model in an Nx 1 xM three-dimensional form, or decomposing the input sample into M Nx 1 point correspondences and sending the M Nx 1 point correspondences into M RNN basic neural units on the RNN deep neural network model, or decomposing the input sample into M Nx 1 x 1 point correspondences and sending the M N point correspondences into M identical CNN deep neural network models by adopting the CRNN deep neural network model and then sending the M RNN basic neural units on the RNN deep neural network model. N each represents a waveform of N points, and M each represents M waveforms having N points.
In another preferred embodiment, the data preprocessing device comprises
A waveform acquisition processing unit: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein N is more than 0 and less than or equal to Fs t, and Fs is the sampling rate;
an R interval acquisition unit: for acquiring the interval from the K-1 th to the Kth R on the ECG waveform, or the interval from the K-1 th to the K +1 th R on the ECG waveform, or for acquiring the interval from the K-1 th to the K +1 th R on the ECG waveform;
a normalization processing unit: the device is used for respectively carrying out normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval in the RR interval acquisition unit, and then combining the normalized input waveform and the RR interval into H to form a one-dimensional vector of N + 1; or directly combining 1 resampled input waveform in the waveform acquisition processing unit and 1 RR interval in the RR interval acquisition unit into H to form a one-dimensional vector of N + 1; or performing normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval alternative in the R interval acquisition unit, and combining the waveforms into H to form a one-dimensional vector of N + 1;
a sample forming unit: for taking M consecutive H's in the normalization processing unit as an input sample, M > 0;
an input sample format processing unit: the system comprises M RNN basic neural units, a CNN deep neural network model, a sample forming unit, a data processing unit and a data processing unit, wherein the CNN basic neural network model is used for inputting one input sample in the sample forming unit into the CNN deep neural network model in a three-dimensional form of (N +1) multiplied by 1 multiplied by M, or correspondingly sending the input sample into the M RNN basic neural units on the RNN deep neural network model by decomposing the input sample into M points of (N +1) multiplied; or adopting a CRNN deep neural network model to decompose the input sample into M N multiplied by 1 heart beat waveforms and M1 multiplied by 1 RR intervals, correspondingly sending the M N multiplied by 1 input waveforms into M same CNN deep neural network models for processing, and obtaining M results PMM results PMCombining with M1 × 1 RR intervals to form M combinations, and sending into M RNN basic neural units on RNN deep neural network model, wherein each combination contains 1 result PMAnd 1 RR interval of 1 × 1. N each represents a waveform of N points, and M each represents M waveforms having N points. The basic RNN unit structure includes conventional RNN, LSTM or gated cyclic unit GRU.
The system selects the waveform capable of retaining the necessary characteristics to perform resampling normalization processing, or selects the waveform capable of retaining the necessary characteristics to perform resampling normalization processing and perform normalization processing combination on RR interphase, or directly selects the waveform capable of retaining the necessary characteristics, RR interphase and combination of the waveform and the RR interphase, then further processes data into a specific form, sends the specific form into a corresponding deep neural network to perform training, and achieves automatic analysis on an electrocardiogram. And (3) corresponding the final output result of each model to the corresponding cardiac rhythm classification or cardiac beat classification, and detecting the heart beat related diseases, the rhythm related diseases and the two types of diseases simultaneously according to the obtained cardiac rhythm classification or cardiac beat classification by a doctor, wherein the heart beat related diseases correspond to the cardiac beat classification, and the rhythm related diseases correspond to the cardiac rhythm classification. Heart beat related diseases, such as Premature Ventricular Contraction (PVC), left bundle branch block, right bundle branch block, etc. Rhythm-related diseases such as atrial fibrillation, atrial flutter, arrhythmia, etc.
It will be understood by those skilled in the art that all or part of the steps in the method for implementing the above embodiments may be implemented by hardware, or may be implemented by a program to instruct relevant hardware, where the program may be stored in a computer-readable storage medium, and when executed, may include the procedures of the embodiments of the methods as described above. The storage medium may be a read-only memory, a random access memory, a magnetic disk or an optical disk, etc.
The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are merely exemplary embodiments of the present invention, and are not intended to limit the scope of the present invention, and any modifications, equivalent substitutions, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (10)
1. A deep learning data preprocessing method is used for sending data processed by the method into a deep neural network for training to realize heart beat classification, and is characterized by comprising the following steps of:
collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t;
directly taking the N points in the step as 1 input waveform, or carrying out amplitude normalization processing on the N points in the step to obtain 1 input waveform;
continuously acquiring M input waveforms as an input sample, wherein M is more than 0;
and inputting the input sample into a CNN deep neural network model in an Nx 1 xM three-dimensional form, or decomposing the input sample into M Nx 1 point correspondences and sending the M Nx 1 point correspondences into M RNN basic neural units on the RNN deep neural network model, or decomposing the input sample into M Nx 1 x 1 point correspondences and sending the M N point correspondences into M identical CNN deep neural network models by adopting the CRNN deep neural network model and then sending the M RNN basic neural units on the RNN deep neural network model.
2. The deep learning data preprocessing method as claimed in claim 1, wherein the kth complete heartbeat waveform of the ECG waveform is acquired and resampled to N points by the following method: acquiring a section of waveform in a first T1 time period and a second T2 time period of a Kth R position of the ECG waveform, resampling N points, wherein 0< N is less than or equal to Fs T, T is T1+ T2, the time period T1 before the R position at least comprises a section of waveform of P wave, and the time period T2 after the R position at least comprises a section of waveform of T wave.
3. A deep learning data preprocessing method is used for sending data processed by the method into a deep neural network for training to realize cardiac rhythm classification and/or cardiac beat classification, and is characterized by comprising the following steps:
s101: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t;
s102: acquiring the interval from the K-1 th to the K-1 th R on the ECG waveform or acquiring the interval from the K-1 th to the K +1 th R on the ECG waveform;
s103: respectively carrying out normalization processing on the 1 resampled input waveform in the step S101 and the 1 RR interval in the step S102, and then combining the input waveforms into H to form a one-dimensional vector of N + 1; or directly combining the 1 resampled input waveform in the step S101 and the 1 RR interval in the step S102 into H to form a one-dimensional vector of N + 1; or normalizing the 1 resampled input waveform in the step S101 and the 1 RR interval alternative in the step S102, and combining the normalized input waveforms into H to form a one-dimensional vector of N + 1;
s104: taking continuous M H in the step S103 as an input sample, wherein M is more than 0;
s105: inputting one input sample in step S104 into the CNN deep neural network model in a three-dimensional form of (N +1) × 1 × M, or decomposing the input sample into M (N +1) samplesThe point of 1 corresponds to M RNN basic neural units sent to an RNN deep neural network model; or adopting a CRNN deep neural network model to decompose the input sample into M input waveforms of Nx 1 x 1 and M RR intervals of 1 x 1, correspondingly sending the M input waveforms of Nx 1 x 1 into M identical CNN deep neural network models for processing, and obtaining M results PMM results PMCombining with M1 × 1 RR intervals to form M combinations, and sending into M RNN basic neural units on RNN deep neural network model, wherein each combination contains 1 result PMAnd 1 RR interval of 1 × 1.
4. The deep learning data preprocessing method as claimed in claim 3, wherein the Kth complete heart beat waveform of the ECG waveform is acquired and re-sampled to N points by the following method: acquiring a section of waveform in a first T1 time period and a second T2 time period of a Kth R position of the ECG waveform, resampling N points, wherein 0< N is less than or equal to Fs T, T is T1+ T2, the time period T1 before the R position at least comprises a section of waveform of P wave, and the time period T2 after the R position at least comprises a section of waveform of T wave.
5. The data preprocessing device is used for sending the data processed by the device into a deep neural network for training to realize heart beat classification, and is characterized by comprising
A waveform acquisition processing unit: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t;
a normalization processing unit: directly taking N points obtained by the waveform acquisition unit as 1 input waveform, or carrying out amplitude normalization processing on the N points obtained by the waveform acquisition unit to obtain 1 input waveform;
a sample forming unit: continuously acquiring M input waveforms as an input sample, wherein M is more than 0;
an input sample format processing unit: and inputting the input sample into a CNN deep neural network model in an Nx 1 xM three-dimensional form, or decomposing the input sample into M Nx 1 point correspondences and sending the M Nx 1 point correspondences into M RNN basic neural units on the RNN deep neural network model, or decomposing the input sample into M Nx 1 x 1 point correspondences and sending the M N point correspondences into M identical CNN deep neural network models by adopting the CRNN deep neural network model and then sending the M RNN basic neural units on the RNN deep neural network model.
6. The data pre-processing device of claim 5, wherein the Kth complete heartbeat waveform of the ECG waveform is acquired and re-sampled to N points by: acquiring a section of waveform in a first T1 time period and a second T2 time period of a Kth R position of the ECG waveform, resampling N points, wherein 0< N is less than or equal to Fs T, T is T1+ T2, the time period T1 before the R position at least comprises a section of waveform of P wave, and the time period T2 after the R position at least comprises a section of waveform of T wave.
7. The data preprocessing device is used for sending the data processed by the method into a deep neural network for training to realize cardiac rhythm classification and/or cardiac beat classification, and is characterized by comprising
A waveform acquisition processing unit: the method is used for collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, and enabling 0< N to be less than or equal to Fs t;
an R interval acquisition unit: for acquiring the interval from the K-1 th to the Kth R on the ECG waveform, or the interval from the K-1 th to the K +1 th R on the ECG waveform, or for acquiring the interval from the K-1 th to the K +1 th R on the ECG waveform;
a normalization processing unit: the device is used for respectively carrying out normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval in the RR interval acquisition unit, and then combining the normalized input waveform and the RR interval into H to form a one-dimensional vector of N + 1; or directly combining 1 resampled input waveform in the waveform acquisition processing unit and 1 RR interval in the RR interval acquisition unit into H to form a one-dimensional vector of N + 1; or performing normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval alternative in the R interval acquisition unit, and combining the waveforms into H to form a one-dimensional vector of N + 1;
a sample forming unit: for taking M consecutive H's in the normalization processing unit as an input sample, M > 0;
an input sample format processing unit: the system comprises M RNN basic neural units, a CNN deep neural network model, a sample forming unit, a data processing unit and a data processing unit, wherein the CNN basic neural network model is used for inputting one input sample in the sample forming unit into the CNN deep neural network model in a three-dimensional form of (N +1) multiplied by 1 multiplied by M, or correspondingly sending the input sample into the M RNN basic neural units on the RNN deep neural network model by decomposing the input sample into M points of (N +1) multiplied; or adopting a CRNN deep neural network model, decomposing the input sample into M input waveforms of Nx 1 x 1 and M input waveforms of 1 x 1, correspondingly sending the M input waveforms of Nx 1 x 1 into M identical CNN deep neural network models for processing, and obtaining M results PMM results PMCombining with M1 × 1 RR intervals to form M combinations, and sending into M RNN basic neural units on RNN deep neural network model, wherein each combination contains 1 result PMAnd 1 RR interval of 1 × 1.
8. The training system is characterized by comprising
A data preprocessing device: the method is used for extracting characteristic points on an ECG waveform to form an input sample, and then processing the input sample into a specific form;
and the training module is used for processing the data preprocessing device into input samples in a specific form, sending the input samples into the corresponding deep neural network model for training, and adjusting the filter weight in the deep neural network model until the loss function judges that the adjusted filter weight meets the requirement.
9. Training system according to claim 8, characterized in that the data preprocessing means comprise
A waveform acquisition processing unit: collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, wherein 0< N is less than or equal to Fs x t;
a normalization processing unit: directly taking N points obtained by the waveform acquisition unit as 1 input waveform, or carrying out amplitude normalization processing on the N points obtained by the waveform acquisition unit to obtain 1 input waveform;
a sample forming unit: continuously acquiring M input waveforms as an input sample, wherein M is more than 0;
an input sample format processing unit: and inputting the input sample into a CNN deep neural network model in an Nx 1 xM three-dimensional form, or decomposing the input sample into M Nx 1 point correspondences and sending the M Nx 1 point correspondences into M RNN basic neural units on the RNN deep neural network model, or decomposing the input sample into M Nx 1 x 1 point correspondences and sending the M N point correspondences into M identical CNN deep neural network models by adopting the CRNN deep neural network model and then sending the M RNN basic neural units on the RNN deep neural network model.
10. Training system according to claim 8, characterized in that the data preprocessing means comprise
A waveform acquisition processing unit: the method is used for collecting the Kth complete heart beat waveform of the ECG waveform, resampling to N points, and enabling 0< N to be less than or equal to Fs t;
an R interval acquisition unit: for acquiring the interval from the K-1 th to the Kth R on the ECG waveform, or the interval from the K-1 th to the K +1 th R on the ECG waveform, or for acquiring the interval from the K-1 th to the K +1 th R on the ECG waveform;
a normalization processing unit: the device is used for respectively carrying out normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval in the RR interval acquisition unit, and then combining the normalized input waveform and the RR interval into H to form a one-dimensional vector of N + 1; or directly combining 1 resampled input waveform in the waveform acquisition processing unit and 1 RR interval in the RR interval acquisition unit into H to form a one-dimensional vector of N + 1; or performing normalization processing on 1 resampled input waveform in the waveform acquisition and processing unit and 1 RR interval alternative in the R interval acquisition unit, and combining the waveforms into H to form a one-dimensional vector of N + 1;
a sample forming unit: for taking M consecutive H's in the normalization processing unit as an input sample, M > 0;
an input sample format processing unit: for inputting one input sample in the sample shaping unit into the CNN deep neural network model in three-dimensional form of (N +1) × 1 × M, or inputting the above one input sampleThe decomposition is that M (N +1) multiplied by 1 points correspond to M RNN basic neural units sent to an RNN deep neural network model; or adopting a CRNN deep neural network model to decompose the input sample into M input waveforms of Nx 1 x 1 and M RR intervals of 1 x 1, correspondingly sending the M input waveforms of Nx 1 x 1 into M identical CNN deep neural network models for processing, and obtaining M results PMM results PMCombining with M1 × 1 RR intervals to form M combinations, and sending into M RNN basic neural units on RNN deep neural network model, wherein each combination contains 1 result PMAnd 1 RR interval of 1 × 1.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910307795.3A CN111832586A (en) | 2019-04-16 | 2019-04-16 | Deep learning data preprocessing method and device and training system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910307795.3A CN111832586A (en) | 2019-04-16 | 2019-04-16 | Deep learning data preprocessing method and device and training system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111832586A true CN111832586A (en) | 2020-10-27 |
Family
ID=72914879
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910307795.3A Pending CN111832586A (en) | 2019-04-16 | 2019-04-16 | Deep learning data preprocessing method and device and training system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111832586A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113171102A (en) * | 2021-04-08 | 2021-07-27 | 南京信息工程大学 | ECG data classification method based on continuous deep learning |
-
2019
- 2019-04-16 CN CN201910307795.3A patent/CN111832586A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113171102A (en) * | 2021-04-08 | 2021-07-27 | 南京信息工程大学 | ECG data classification method based on continuous deep learning |
| CN113171102B (en) * | 2021-04-08 | 2022-09-02 | 南京信息工程大学 | ECG data classification method based on continuous deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN106725428B (en) | Electrocardiosignal classification method and device | |
| CN109770860B (en) | Electrocardiosignal processing device and electrocardio equipment | |
| CN109907752B (en) | Electrocardiogram diagnosis and monitoring system for removing motion artifact interference and electrocardio characteristic detection | |
| CN111772619B (en) | Heart beat identification method based on deep learning, terminal equipment and storage medium | |
| CN109480824B (en) | Method and device for processing electrocardio waveform data and server | |
| CN111990989A (en) | Electrocardiosignal identification method based on generation countermeasure and convolution cyclic network | |
| CN108647565A (en) | A kind of data preprocessing method classified to electrocardiosignal based on deep learning model | |
| CN110717416B (en) | Neural network training method for ST segment classification recognition based on feature selection | |
| CN110897633A (en) | Electrocardiosignal segmentation method based on deep neural network | |
| CN110367968B (en) | Right bundle branch retardation detection method, device, equipment and storage medium | |
| CN110717415B (en) | ST-segment classification convolutional neural network based on feature selection and using method thereof | |
| CN110367936B (en) | Electrocardiosignal detection method and device | |
| CN113057648A (en) | ECG signal classification method based on composite LSTM structure | |
| Gawande et al. | Heart diseases classification using convolutional neural network | |
| CN110327039A (en) | Atrial fibrillation classification method based on convolutional neural networks | |
| CN110236529A (en) | A kind of multi-lead arrhythmia cordis intelligent diagnosing method based on MODWT and LSTM | |
| CN110693488A (en) | Electrocardiosignal processing system, electrocardiosignal processing method, electronic equipment and readable storage medium | |
| CN108647584B (en) | Arrhythmia identification and classification method based on sparse representation and neural network | |
| CN111832537A (en) | Abnormal electrocardiosignal identification method and abnormal electrocardiosignal identification device | |
| CN111832586A (en) | Deep learning data preprocessing method and device and training system | |
| CN111528833B (en) | Rapid identification and processing method and system for electrocardiosignals | |
| CN111956202B (en) | Training method and model of convolutional neural network based on sample distribution improvement | |
| CN110537907B (en) | Electrocardiosignal compression and identification method based on singular value decomposition | |
| CN110811608A (en) | Atrial fibrillation monitoring method based on ECG (ECG) signals | |
| CN116712083A (en) | QRS complex wave detection method based on U-Net network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |