WO2019100565A1 - 基于人工智能自学习的动态心电图分析方法和装置 - Google Patents

基于人工智能自学习的动态心电图分析方法和装置 Download PDF

Info

Publication number
WO2019100565A1
WO2019100565A1 PCT/CN2018/072359 CN2018072359W WO2019100565A1 WO 2019100565 A1 WO2019100565 A1 WO 2019100565A1 CN 2018072359 W CN2018072359 W CN 2018072359W WO 2019100565 A1 WO2019100565 A1 WO 2019100565A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
heartbeat
analysis
wave
lead
Prior art date
Application number
PCT/CN2018/072359
Other languages
English (en)
French (fr)
Inventor
刘畅
胡传言
周位位
卢海涛
汪嘉雨
曹君
Original Assignee
乐普（北京）医疗器械股份有限公司
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 乐普（北京）医疗器械股份有限公司 filed Critical 乐普（北京）医疗器械股份有限公司
Priority to EP18881678.9A priority Critical patent/EP3692904A4/en
Priority to US16/651,912 priority patent/US11234629B2/en
Publication of WO2019100565A1 publication Critical patent/WO2019100565A1/zh

Links

Images

Classifications

    • 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
    • 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
    • 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
    • A61B5/364—Detecting abnormal ECG interval, e.g. extrasystoles, ectopic heartbeats
    • 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
    • A61B5/366—Detecting abnormal QRS complex, e.g. widening
    • 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
    • A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
    • 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
    • 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
    • A61B5/7271—Specific aspects of physiological measurement analysis

Definitions

  • the invention relates to the technical field of artificial intelligence data analysis, in particular to a dynamic electrocardiogram analysis method and device based on artificial intelligence self-learning.
  • Non-invasive ECG is an important tool for cardiovascular disease (CVD) analysis and screening.
  • DCG Dynamic Electrocardiography
  • the dynamic electrocardiogram can continuously record the whole process of ECG activities for more than 24 hours, including ECG information in different situations such as rest, activities, meals, work, study and sleep.
  • Dynamic electrocardiogram can detect non-sustained arrhythmia that is difficult to detect by conventional electrocardiography, especially the detection rate of transient arrhythmia and transient myocardial ischemic attack.
  • dynamic electrocardiogram can determine whether the patient's heart palpitations, dizziness, fainting and other symptoms are related to arrhythmia, such as sinus bradycardia, conduction block, rapid atrial fibrillation, paroxysmal supraventricular tachycardia, persistent ventricular Tachycardia, etc., is one of the most important and widely used cases of current 24-hour Holter.
  • the dynamic electrocardiogram is mainly composed of three parts: ECG signal acquisition recorder, lead system and computer analysis software.
  • the signal acquisition recorder is responsible for recording ECG data that stores the patient for 24 hours or more. Because the patient's daily activities in the daily activities are relatively large, and the interference situation is relatively high, the frequency response, sampling frequency, resolution, impedance, anti-interference performance and other requirements of the signal acquisition are relatively high.
  • the high-performance ECG signal recorder's ECG signal waveform has high signal-to-noise ratio and strong signal fidelity, which is very helpful for subsequent analysis and calculation.
  • the lead system includes electrode pads and lead wires, and the analysis and calculation software performs arrhythmia analysis, long intermittent pause, flutter and tremor, conduction block, premature beat and escape, bradycardia, and tachycardia on the collected digital signals. , ST segment change detection, analysis and classification of ECG events, help doctors to analyze reports.
  • the common ML simplified lead system measures the data of four leads of I, II, V2, and V5, or I, II, V1, V5, etc., and can calculate III, aVR, aVL, aVF according to the relationship of ECG vector.
  • the data of the lead shows that the three lead data of II, V5 and aVF are used; the ML simplified lead system Holter is the mainstream product on the market; however, because the lead data cannot accurately analyze the ST segment changes, part of it Doctor organizations also require 12-lead Holter for inspection; with the widespread use of mobile medical and wearable devices, single-lead heart patches with real-time monitoring capabilities are also becoming popular in certain populations, such as patients after PCI. , pacemaker users, long-distance runners, etc.
  • the analysis of dynamic electrocardiogram mainly includes the following steps: first, performing signal filtering processing, and then detecting the most significant P-QRS-T wave group features in the heartbeat; secondly, identifying the QRS wave, P wave, T wave, RR
  • the characteristic signals such as interval, PR interval and ST segment are classified according to cardiac electrophysiological activity.
  • Heart rhythm event The electrocardiographer analyzes and describes the patient's dynamic electrocardiogram based on the heartbeat classification and heart rate events.
  • the frequency range of the normal ECG signal is 0.01 Hz-100 Hz, wherein 90% of the energy is concentrated in the 0.25 Hz-35 Hz range. Therefore, in the 24-hour long-term continuous wear process of the dynamic electrocardiogram, the signal is easily interfered.
  • the main interference includes: Power Line Interference (50 Hz and its harmonics generated by human body capacitance, equipment circuit, lead, etc.) Interference), electromyography (normally 5-500Hz high frequency interference caused by muscle activity), baseline drift (Baseline Drift, caused by poor electrode contact, skin impedance, respiratory motion, etc., frequency
  • Power Line Interference 50 Hz and its harmonics generated by human body capacitance, equipment circuit, lead, etc.
  • electromyography normally 5-500Hz high frequency interference caused by muscle activity
  • baseline drift Baseline Drift, caused by poor electrode contact, skin impedance, respiratory motion, etc., frequency
  • the range is generally between 0.05 Hz and 2.00 Hz, which is very close to the ST segment and Q wave frequency components of the electrocardiogram signal).
  • the main problems of the above method are as follows: First, due to the influence of the interference signal, the anti-interference ability is relatively poor. Heart beat classification recognition cannot actively eliminate the influence of interference fragments. Second, in the heartbeat feature extraction, P and T cannot be accurately identified. Heartbeat testing often has multiple tests and missed tests. For some special ECG signals, such as the tall T wave of patients with slow heart rhythm, or the signal of T wave hypertrophy, it is often easy to check. Third, the classification of heartbeat basically stays in the three types of sinus, supraventricular and ventricular, which is far from meeting the complex and comprehensive analysis requirements of clinical electrocardiographers. Fourth, it is not possible to accurately identify atrial flutter, pacemaker heartbeat, and ST-T changes.
  • the object of the present invention is to provide a dynamic electrocardiogram analysis method and device based on artificial intelligence self-learning, which can automatically, quickly and completely analyze the measurement data output by the electrocardiograph device, and obtain the required report data.
  • a first aspect of the embodiments of the present invention provides a dynamic electrocardiogram analysis method based on artificial intelligence self-learning, including:
  • Receiving dynamic electrocardiogram data output by the dynamic electrocardiogram monitoring device comprising: converting the data format of the dynamic electrocardiogram data into a preset standard data format, and performing first filtering on the converted dynamic electrocardiogram data of the preset standard data format deal with;
  • each of the heartbeat data corresponding to one heart cycle, including a corresponding P wave, QRS Wave group, T wave amplitude and start and end time data;
  • the heartbeat analysis data of the specific heartbeat in the primary classification information result is input to the trained ST segment and the T wave change model to identify the ST segment and the T wave evaluation information;
  • the detailed feature information includes amplitude, direction, Form and start and end time data
  • the heartbeat classification information is analyzed and matched according to the basic rule reference data of the electrocardiogram, and the result of the analysis matching is processed according to the signal quality evaluation parameter to generate electrocardiogram event data;
  • the determining the detection confidence includes:
  • the detection confidence of each heartbeat data is determined based on the estimated value of the noise and the maximum amplitude in each heartbeat data.
  • the interference identification of the heartbeat data according to the interference recognition two-class model obtained by the training specifically includes:
  • the heartbeat data is cut and sampled by the first data amount, and the data obtained by the cut sampling is input to the interference recognition two-class model for interference recognition;
  • the interference identification is performed for each of the sampled data segments.
  • the generating the heartbeat analysis data according to the heartbeat time series data is specifically:
  • Each lead heartbeat data in the heartbeat time series data is cut according to a preset threshold, thereby generating heartbeat analysis data of each lead.
  • the heartbeat classification model obtained by the training performs feature extraction and analysis on the amplitude and time characterization data of the heartbeat analysis data, and the first classification information of the heartbeat analysis data is specifically included:
  • the heartbeat analysis data is input to the trained single-lead heart beat classification model, and the second data amount is used for feature extraction and analysis of amplitude and time characterization data. Obtaining the primary classification information of the single lead.
  • the heartbeat classification model obtained by the training performs feature extraction and analysis on the amplitude and time characterization data of the heartbeat analysis data, and the first classification information of the heartbeat analysis data is specifically included:
  • the heartbeat analysis data of each lead is subjected to feature extraction and analysis of the amplitude and time characterization data by the third data amount, and the classification information of each lead is obtained;
  • the classified voting decision calculation is performed according to the classification information of each lead and the lead weight value reference coefficient, and the primary classification information is obtained.
  • the heartbeat classification model obtained by the training performs feature extraction and analysis on the amplitude and time characterization data of the heartbeat analysis data, and the first classification information of the heartbeat analysis data is specifically included:
  • the heartbeat analysis data of each lead is subjected to feature extraction and analysis of the synchronous amplitude and time characterization data by the fourth data amount, and the primary classification information of the heartbeat analysis data is obtained.
  • the signal quality analysis is performed on the heartbeat analysis data, and the signal quality of the heartbeat analysis data is evaluated, and the signal quality evaluation parameter of the heartbeat analysis data is obtained as follows:
  • the signal-to-noise ratio determines the signal quality evaluation parameters.
  • the heartbeat analysis data of the specific heartbeat in the result of the primary classification information is input to the trained ST segment and the T wave change model to identify the ST segment and the T wave evaluation information, and specifically includes:
  • the specific heart beat data in the primary classification information is sequentially input into the ST segment and the T wave change model according to the training according to the lead, and the feature extraction of the amplitude and time characterization data is performed on the specific heart beat data of each lead. And analyzing, obtaining ST segment and T wave change information of each lead, and determining ST segment and T wave evaluation information; the ST segment and T wave evaluation information are specifically ST segment and T wave corresponding to the heart beat segment data Changed lead position information.
  • the performing the P-wave and the T-wave feature detection on the heartbeat analysis data, and obtaining the detailed feature information specifically includes:
  • Performing a QRS complex signal elimination process on the heartbeat analysis data performing third filtering on the heartbeat analysis data after the QRS complex signal elimination processing, and performing data separation on the heartbeat analysis data through an independent component analysis algorithm, Obtaining independent component sequence data, according to the distribution characteristics of the peaks of the independent component sequence data and the position of the QRS complex, selecting the independent component with the highest probability as the corresponding P wave and T wave component, and simultaneously determining the P wave and the T wave.
  • the performing the P-wave and the T-wave feature detection on the heartbeat analysis data, and obtaining the detailed feature information specifically includes:
  • Performing peak detection on the signal characteristic data determining a T wave detection interval between two adjacent QRS groups, and determining a data point of a maximum amplitude in the T wave detection interval as a T wave;
  • Determining a data point of a maximum amplitude in an interval other than the T wave detection interval between adjacent two QRS groups is a P wave
  • the direction and morphological characteristics of the P wave and the T wave are determined based on the peak and position data of the P wave and the T wave.
  • the evaluating the ECG event data according to the signal quality evaluation parameter to generate report graphic data specifically includes:
  • the signal quality evaluation parameter the data segment of each electrocardiogram event is evaluated, and the data segment with the highest signal quality evaluation parameter is selected as a typical data segment in the ECG event, and report graphic data is generated.
  • the method further includes:
  • the modified data is used as training sample data for model training in the ECG automatic analysis method based on artificial intelligence self-learning.
  • the method for dynamic electrocardiogram analysis based on artificial intelligence self-learning includes data preprocessing, heart beat feature detection, interference signal detection and heart rate classification based on deep learning method, signal quality assessment and guidance, and heart
  • the dynamic electrocardiogram analysis method of the invention can also record the modification information of the automatic analysis result, and collect the modified data feedback to the deep learning model to continue training, and continuously improve the accuracy of the automatic analysis method.
  • a second aspect of the embodiments of the present invention provides an apparatus, including a memory and a processor, the memory is configured to store a program, and the processor is configured to execute the method in the first aspect and the implementation manners of the first aspect.
  • a third aspect of the embodiments of the present invention provides a computer program product comprising instructions for causing a computer to perform the methods of the first aspect and the implementations of the first aspect when the computer program product is run on a computer.
  • a fourth aspect of the embodiments of the present invention provides a computer readable storage medium.
  • the computer readable storage medium stores a computer program.
  • the computer program is executed by the processor, the first aspect and the method in each implementation manner of the first aspect are implemented. .
  • FIG. 1 is a system structural diagram of dynamic electrocardiogram analysis based on artificial intelligence self-learning according to an embodiment of the present invention
  • FIG. 2 is a flowchart of a method for analyzing a dynamic electrocardiogram based on artificial intelligence self-learning according to an embodiment of the present invention
  • FIG. 3 is a schematic diagram of a two-classification model for interference recognition according to an embodiment of the present invention.
  • FIG. 4 is a schematic diagram of a heart beat classification model according to an embodiment of the present invention.
  • FIG. 5 is a schematic diagram of a ST segment and T wave change model according to an embodiment of the present invention.
  • FIG. 6 is a schematic diagram of a ST segment and T wave change positioning analysis process according to an embodiment of the present invention.
  • FIG. 7 is a schematic structural diagram of an apparatus for dynamic electrocardiogram analysis based on artificial intelligence self-learning according to an embodiment of the present invention.
  • the basic principles of the artificial intelligence model especially the convolutional neural network model, are first introduced.
  • the artificial intelligence convolutional neural network (CNN) model is a supervised learning method in deep learning. It is a multi-layer network (hidden layer) connection structure that simulates a neural network. The input signal passes through each hidden layer in turn, and one of them is carried out.
  • a series of complex mathematical processing Convolution convolution, Pooling pooling, Regularization regularization, prevention of overfitting, Dropout temporary discarding, Activation activation, generally using Relu activation function
  • Activation activation generally using Relu activation function
  • CNN belongs to the supervised learning method in artificial intelligence.
  • the input signal passes through multiple hidden layer processing to reach the final fully connected layer.
  • the classification result obtained by softmax logistic regression is related to the known classification result (label label).
  • label label There will be an error.
  • a core idea of deep learning is to continually minimize this error through a large number of sample iterations, and then calculate the parameters that connect the hidden layer neurons. This process generally requires constructing a special cost function, using a nonlinearly optimized gradient descent algorithm and a backpropagation algorithm (BP) to quickly and efficiently minimize the entire depth (the number of layers in the hidden layer). ) and breadth (dimension dimension) are all complex parameters in the neural network structure.
  • BP backpropagation algorithm
  • Deep learning inputs the image to be recognized into the training model, passes through the first hidden layer, the second hidden layer, the third hidden layer, and finally outputs the recognition result.
  • Each layer extracts features of different levels of abstraction and ultimately identifies what category the image is, such as a car, person or animal.
  • the algorithm model of deep learning is mathematically complex. Developing a complete algorithmic program requires a strong background in professional knowledge and extensive work experience. In recent years, companies such as Google, Microsoft, Baidu, Facebook, and some famous universities (such as the University of California, Berkeley, and the University of Montreal, Canada) have also developed and launched various artificial intelligence development open source platforms to help some research and development in the field of deep learning. The company quickly mastered this cutting-edge technology. Among them, Berkeley's Caffe and Google's Tensorflow are the two most widely used framework tools.
  • the deep learning model is extremely complex, and the required training data from hundreds of thousands, millions to tens of millions, plus repeated iterations, the amount of nonlinear optimization is very large.
  • the calculation is performed by the central processing unit of a common computer, which often takes from ten hours to several days or even longer. In this case, switching to a Graphics Processing Unit (GPU) calculation can greatly speed up the calculation.
  • GPU Graphics Processing Unit
  • the GPU card provided by Nvidia, due to powerful graphics and computer vision computing capabilities, a large number of linear algebra and other computing libraries, supports parallel processing, can meet the various methods of deep learning needs, and become the current artificial intelligence high-performance training and The underlying hardware of reasoning.
  • the dynamic electrocardiogram analysis method based on artificial intelligence self-learning of the invention is implemented based on the CNN model.
  • the dynamic electrocardiogram analysis method based on artificial intelligence self-learning is implemented based on the system architecture shown in FIG. 1 .
  • the system architecture includes an ECG monitoring device 1, a data storage transmission device 2, an analysis system hardware module 3, a model training module 4, and an analysis system execution module 5.
  • the analysis system execution module 5 specifically includes: a data preprocessing module 501, a heartbeat The detection module 502, the interference recognition module 503, the lead heartbeat merge module 504, the confidence calculation module 505, the heart beat classification module 506, the heart beat audit module 507, the P wave and T wave feature detection module 508, and the signal quality evaluation module 509
  • the invention realizes the dynamic electrocardiogram analysis process of artificial intelligence self-learning through the analysis system hardware module 3, the model training module 4 and the analysis system execution module 5 in the system architecture.
  • the method mainly includes the following steps:
  • Step 100 Receive dynamic electrocardiogram data output by the dynamic electrocardiogram monitoring device
  • the dynamic electrocardiogram monitoring device converts the dynamic electrocardiogram electrical signal into a digital signal output, which may be a time series data of a single-lead or multi-lead dynamic electrocardiograph device for measuring the collected output, or may be obtained through a database or other file method.
  • the single-lead or multi-lead dynamic electrocardiogram data is stored by the data storage transmission device, and can be transmitted to the analysis system hardware through WIFI, Bluetooth, USB, 3G/4G/5G mobile communication network, Internet of Things, etc.
  • the module is input as an input signal to the analysis system execution module.
  • the generated dynamic electrocardiogram time series data is in the lead mark, the data encoding format. , gain, accuracy, data length per second, baseline position, etc. are very different, data preprocessing must be performed, all input dynamic ECG time series data are processed according to the requirements of the analysis process of the present invention, according to the pre-processing Set the standard data format for storage.
  • the preset standard data format refers to a data format capable of supporting data identification for the entire process of data processing.
  • Step 101 The data format of the dynamic electrocardiogram data is resampled into a preset standard data format, and the first filtering process is performed on the ECG data of the converted preset standard data format.
  • the format adaptive reading, resampling, filtering, and standard data format saving of the dynamic electrocardiogram data are performed by the data preprocessing module.
  • the format of the dynamic electrocardiogram data is read and read. Different readings are implemented for different devices. After reading, the baseline needs to be adjusted and converted into millivolt data according to the gain.
  • Dynamic ECG data is resampled to convert the data into a sampling frequency that can be processed by the entire process. Then, high frequency, low frequency noise interference and baseline drift are removed by filtering to improve the accuracy of artificial intelligence analysis.
  • the processed dynamic electrocardiogram data is saved in a preset standard data format.
  • the resampling frequency can be 200 Hz, ensuring that the data length of each fixed time range is consistent, performing calculation, training, and reasoning, and obtaining satisfactory analysis results without excessively increasing the complexity and time of the calculation training.
  • the sampling rate of the ECG device itself gradually increases, the GPU computing performance continues to rise and the cost decreases, and the artificial intelligence algorithm continues to innovate and optimize, the sampling rate can be improved.
  • Digital signal filtering can use high-pass filter, low-pass filter and median filter to eliminate power frequency interference, myoelectric interference and baseline drift interference, and avoid the impact on subsequent analysis.
  • a low-pass, high-pass Butterworth filter can be used for zero-phase shift filtering to remove baseline drift and high-frequency interference, retaining an effective ECG signal; median filtering can utilize a sliding window of preset duration The median of the point voltage magnitude replaces the amplitude of the center sequence of the window. The low frequency baseline drift can be removed.
  • Step 102 Perform heartbeat detection processing on the first filtered processed electrocardiogram data, and identify multiple heartbeat data included in the dynamic electrocardiogram data;
  • Each heart beat data corresponds to a heart cycle, including corresponding P wave, QRS wave group, T wave amplitude data and start and stop time data; heart beat detection performed by the heart beat detection module includes QRS wave group detection and P wave, T wave detection.
  • QRS complex detection consists of two processes, one is the signal processing process, extracting the characteristic frequency band of the QRS complex from the first filtered ECG data; the second is to determine the occurrence time of the QRS complex by setting a reasonable threshold .
  • a P wave, a QRS complex, a T wave component, and a noise component are generally included.
  • the QRS complex has a frequency range of 5 to 20 Hz, and the QRS complex signal can be proposed by a bandpass filter within this range.
  • the specific detection process is a process based on peak detection.
  • the threshold is judged for each peak sequence in the signal.
  • the QRS complex judgment process is entered, and more features are detected, such as RR interval and shape.
  • the threshold adjustment needs to be dynamically performed according to the change of the data characteristics in the time domain.
  • QRS complex detection mostly adopts the dual amplitude threshold combined with the time threshold.
  • the high threshold has a higher positive rate
  • the lower threshold has a higher sensitivity rate
  • the RR interval exceeds a certain period.
  • the low threshold is used for detection to reduce the missed detection. Since the low threshold is low due to the low threshold, it is susceptible to T wave and myoelectric noise, and it is easy to cause multiple tests. Therefore, high threshold is preferred for detection.
  • the lead parameters of the heartbeat data are also determined.
  • Step 103 Determine a detection confidence level of each heart beat according to the heart beat data
  • the confidence calculation module can provide an estimate of the confidence of the QRS complex detection based on the amplitude of the QRS complex and the amplitude ratio of the noise signal during the RR interval.
  • Step 104 Perform interference recognition on the heartbeat data according to the interference recognition two-class model obtained by training, obtain whether the heartbeat data has interference noise, and a probability value for determining interference noise;
  • the dynamic electrocardiogram detection is prone to interference caused by various effects during the long-term recording process, the acquired data is invalid or inaccurate, which cannot correctly reflect the condition of the subject, and also increases the difficulty and workload of the doctor; and the interference data also leads to The main factor that intelligent analysis tools cannot work effectively. Therefore, it is especially important to minimize external signal interference. All along, in the computer automatic analysis method of dynamic electrocardiogram, the accurate identification of the interference signal has not been satisfactorily solved, resulting in the overall accuracy of the dynamic ECG computer analysis algorithm is not high.
  • This step is based on the end-to-end two-category recognition model with deep learning algorithm as the core. It has the characteristics of high precision and strong generalization performance, which can effectively solve the disturbance problem caused by the main interference sources such as electrode stripping, motion interference and static interference. It overcomes the problem that the traditional algorithm has poor recognition effect due to the diversity and irregularity of the interference data.
  • the interference identification process is completed by the interference recognition module, and the main steps may include:
  • Step A using interference recognition two-class model for heart rate data for interference recognition
  • Step B identifying a data segment in the heartbeat data that the heartbeat interval is greater than or equal to the preset interval determination threshold
  • Step C performing signal abnormality determination on the data segment whose heartbeat interval is greater than or equal to the preset interval determination threshold, and determining whether it is an abnormal signal;
  • the identification of the abnormal signal mainly includes whether the electrode piece is off, low voltage, and the like.
  • Step D if it is not an abnormal signal, determine a starting data point and a ending data point of the sliding sample in the data segment according to the set time value by a preset time width, and start sliding sampling the data segment from the starting data point, Obtaining a plurality of sampled data segments until the data point is terminated;
  • step E interference identification is performed for each sampled data segment.
  • the above steps A-E will be described with a specific example.
  • the heartbeat data of each lead is cut and sampled by the first data amount, and then input to the interference recognition two-classification model for classification, and a probability value of the interference recognition result and the corresponding result is obtained; the heartbeat interval is greater than or equal to 2 Second heart beat data, first judge whether it is signal overflow, low voltage, electrode off; if it is not the above, according to the first data amount, starting from the left heart beat, to the right continuously with the first data amount without overlapping sliding sampling, Identify.
  • the input may be the first data volume heartbeat data of any lead, and then the interference identification two-class model is used for classification, and the direct output is the interference classification result, and the obtained result is fast, the accuracy is high, the stability is good, and the subsequent Analysis provides more effective and quality data.
  • the structure of the above-mentioned interference recognition two-classification model is an end-to-end two-class identification system based on the artificial intelligence deep learning convolutional neural network LeNet-5 and AlexNet.
  • the sampling rate is 200 Hz
  • the data length is a segment D[300] of 300 ECG voltage values (millivolts)
  • the input data is: InputData(i,j), where i is the ith Lead, j is the jth segment D of the lead i. All the input data is randomly dispersed to start training, which ensures the convergence of the training process. At the same time, it controls the collection of too many samples from the same patient's ECG data, improving the generalization ability of the model, and the accuracy of the real scene. After the training converges, using 1 million independent test data for testing, the accuracy rate can reach 99.3%. Another specific test data is shown in Table 1 below.
  • the interference data is often caused by the external disturbance factor, there are mainly the electrode piece falling off, low voltage, static interference and motion interference.
  • the interference data generated by different disturbance sources is different, and the interference data generated by the same disturbance source is different.
  • the diversity is wide, but it is very different from the normal data, it is also possible to ensure the diversity when collecting the interference training data, and take the sliding sampling of the moving window. It is possible to increase the diversity of the interference data so that the model is more robust to the interference data. Even if the future interference data is different from any previous interference, the similarity with the interference is greater than the normal data. The ability of the model to identify interfering data is enhanced.
  • the interference identification two-class model used in this step can be as shown in Figure 3.
  • the network first uses two layers of convolutional layers.
  • the convolution kernel size is 1x5, and each layer is followed by a maximum pool.
  • the number of convolution kernels starts at 128, and the number of convolution kernels doubles each time the largest pooling layer is passed.
  • the convolutional layer is followed by two fully connected layers and a softmax classifier. Since the classification number of the model is 2, softmax has two output units, which in turn correspond to the corresponding categories, and uses cross entropy as the loss function.
  • Step 105 determining the validity of the heartbeat data according to the detection confidence, and, according to determining the valid heartbeat data and the lead parameter thereof, combining the result of the interference recognition and the time rule to generate the heartbeat time series data, and according to Heart beat time series data generates heart beat analysis data;
  • each lead heart rate data can be cut using a preset threshold to generate heartbeat analysis data of each lead required for specific analysis.
  • the heartbeat data merging process performed by the lead heartbeat merging module is as follows: according to the refractory period of the reference data of the basic rule of the electrocardiogram, the time characterization data combination of the different lead heartbeat data is obtained, and the heartbeat data with large deviation is discarded.
  • the above-mentioned time characterization data combination voting generates a merged heart beat position, the merged heart beat position is added to the combined cardiac beat time sequence, and the next set of heart beat data to be processed is moved, and the loop execution is performed until all the heart beat data are merged.
  • the ECG activity refractory period may preferably be between 200 milliseconds and 280 milliseconds.
  • the acquired time characterization data combination of the different lead heartbeat data should satisfy the following condition: the time characterization data of the heartbeat data combination contains at least one time characterization data of the heartbeat data.
  • the number of leads using the detected heartbeat data is determined as a percentage of the effective number of leads; if the time of the heartbeat data is indicative of the position of the corresponding lead is a low voltage
  • the lead, the interfering segment, and the electrode are considered to be inactive leads for this heartbeat data when the electrode is detached.
  • the time average of the heart beat data can be used to obtain the average value of the data.
  • this method sets a refractory period to avoid false merges.
  • a unified heartbeat time series data is output through the merging operation.
  • This step can simultaneously reduce the multi-detection rate and missed detection rate of the heart beat, and effectively improve the sensitivity and positive predictive rate of heart beat detection.
  • Step 106 Perform feature extraction and analysis on amplitude and time characterization data of the heartbeat analysis data according to the heartbeat classification model obtained by the training, and obtain primary classification information of the heartbeat analysis data;
  • the classification method of the heartbeat performed in the heartbeat classification module of the present invention may be based on specific conditions.
  • the multi-lead classification method includes two methods: lead voting decision classification method and lead synchronization association classification method.
  • the lead voting decision classification method is based on the heartbeat analysis data of each lead to conduct independent lead classification, and then the result vote is merged to determine the voting result decision method of the classification result; the lead synchronous association classification method uses the heart beat for each lead Analyze data for simultaneous association analysis.
  • the single-lead classification method is to analyze the heartbeat analysis data of the single-lead device, and directly use the corresponding lead model for classification, and there is no voting decision process. The following describes several classification methods described above.
  • Single-lead classification methods include:
  • the single-lead heartbeat data is cut by the second data amount to generate single-lead heartbeat analysis data, and input to the trained heartbeat classification model corresponding to the lead for amplitude and The feature extraction and analysis of the time representation data yields the primary classification information of the single lead.
  • the lead voting decision classification method may specifically include:
  • the lead heartbeat data is cut by the third data amount, thereby generating heartbeat analysis data of each lead;
  • the second step according to the heartbeat classification model corresponding to each lead of the training, the feature extraction and analysis of the amplitude and time characterization data of each lead heartbeat data are obtained, and the classification information of each lead is obtained;
  • the classification voting decision calculation is performed according to the classification information of each lead and the reference weight value reference coefficient, and the primary classification information is obtained.
  • the lead weight value reference coefficient is based on the Bayesian statistical analysis of the electrocardiogram data to obtain the voting weight coefficient of each lead for different heart beat classifications.
  • the lead synchronization association classification method may specifically include:
  • each lead heartbeat data is cut with a fourth data amount to generate heartbeat analysis data of each lead; and then according to the trained multi-lead synchronous correlation classification model for each lead
  • the heartbeat analysis data is used for feature extraction and analysis of the synchronized amplitude and time characterization data, and the primary classification information of the heartbeat analysis data is obtained.
  • the synchronous correlation classification method input of the heart beat data is all the lead data of the dynamic electrocardiogram device.
  • the data points of the same position and a certain length on each lead are intercepted and synchronously transmitted to the trained
  • the artificial intelligence deep learning model performs computational analysis, and the output is that each heartbeat position point comprehensively considers all the lead ECG signal characteristics, and the accurate heart beat classification of the heartbeat associated with the heart rhythm characteristics in time.
  • the method fully considers that the different lead data of the dynamic electrocardiogram actually measures the information flow transmitted by the cardiac electrical signal in different directions of the electrocardiogram axis, and integrates the multi-dimensional digital features of the dynamic electrocardiogram signal transmitted in time and space.
  • the analysis greatly improves the traditional method of relying solely on a single lead independent analysis, and then summarizes the results to conduct some statistical voting methods and is relatively easy to get the classification error, which greatly improves the accuracy of the heartbeat classification.
  • the heartbeat classification model in this step we used a training set containing 17 million data samples from 300,000 patients for training. These samples are generated by accurately labeling the data according to the requirements of dynamic electrocardiogram analysis and diagnosis.
  • the annotations are mainly for common arrhythmia, conduction block and ST segment and T wave changes, which can meet the model training of different application scenarios.
  • the marked information is saved in a preset standard data format.
  • a small sliding is performed on the classification with less sample size to amplify the data. Specifically, it is based on each heart beat and according to a certain step. (For example, 10-50 data points) move 2 times, which can increase the data by 2 times, and improve the recognition accuracy of the classified samples with less data. After the actual results are verified, the generalization ability has also been improved.
  • the length of the interception of the training data may be 1 second to 10 seconds.
  • the sampling rate is 200 Hz, with a sampling length of 2.5 s, and the obtained data length is a segment D [500] of 500 ECG voltage values (millivolts)
  • the input data is: InputData (i, j), where i is The i-th lead, j is the j-th segment D of the lead i. All the input data is randomly dispersed to start training, which ensures the convergence of the training process. At the same time, it controls the collection of too many samples from the same patient's ECG data, improving the generalization ability of the model, and the accuracy of the real scene.
  • the segment data D corresponding to all the leads is synchronously input, and the lead data of multiple spatial dimensions (different cardiac axis vectors) of each time position is synchronously learned according to the multi-channel analysis method of image analysis, thereby Get a more accurate classification result than the conventional algorithm.
  • the heartbeat classification model used in this step can be as shown in FIG. 4, and specifically can be an end-to-end multi-label classification model inspired by the model of convolutional neural network AlexNet, VGG16, Inception and the like based on artificial intelligence deep learning.
  • the model's network is a 7-layer convolutional network, with each convolution followed by an activation function.
  • the first layer is a convolutional layer of two different scales, followed by six convolutional layers.
  • the convolution kernels of the seven-layer convolution are 96, 256, 256, 384, 384, 384, 256, respectively. Except for the first-level convolution kernel, which has two scales of 5 and 11, respectively, the other layer has a convolution kernel scale of 5.
  • the third, fifth, sixth and seventh layers of the convolutional layer are followed by the pooling layer. Finally followed by two fully connected layers.
  • Step 107 input the heartbeat analysis data of the specific heartbeat in the primary classification information result into the trained ST segment and the T wave change model, and determine the ST segment and T wave evaluation information;
  • the ST segment and the T wave evaluation information are specifically the lead position information in which the ST segment and the T wave corresponding to the heartbeat analysis data are changed. Because clinical diagnosis requires localization of changes to ST segments and T waves to specific leads.
  • the specific heartbeat data of the primary classification information refers to the heartbeat analysis data including the sinus beat (N) and other heartbeat types that may contain ST changes.
  • the ST segment and T wave change lead positioning module inputs the specific heart beat data of the primary classification information into each lead according to each lead into an artificial intelligence deep learning training model for identifying the ST segment and the T wave change, and performs calculation analysis and output.
  • the results indicate whether the characteristics of the lead segment conform to the conclusions of the ST segment and the T wave change, so that the information of the ST segment and the T wave change occurring in the specific lead, that is, the ST segment and the T wave evaluation information can be determined.
  • the specific method may be: inputting the data of each lead heartbeat analysis of the sinus heartbeat in the primary classification information, inputting the ST segment and the T wave change model, and identifying and analyzing the sinus beat analysis data one by one to determine whether the sinus beat analysis data is There are ST segment and T wave changing characteristics and specific lead position information that occurs, and ST segment and T wave evaluation information is determined.
  • a schematic diagram of the specific execution process can be shown in FIG. 6.
  • the heartbeat with ST segment and T wave changes is relatively low in all heart beats.
  • select ST-segment and T-wave changes and ST-segment and The training data ratio of T wave change is about 2:1, which guarantees the model's good generalization ability in the classification process and does not appear to have a tendency to train more data.
  • the shape of the heart beat is diverse, the forms of different individuals are not the same. Therefore, in order to better estimate the distribution of each classification, the characteristics can be effectively extracted.
  • the training samples are collected from individuals of different ages, weights, genders, and areas of residence.
  • the ECG data of a single individual in the same time period is often highly similar, in order to avoid over-learning, when acquiring data of a single individual, a small number of samples of different time periods are randomly selected from all the data; The patient's heartbeat morphology has large differences between individuals and high intra-individual similarity. Therefore, when dividing training and test sets, different patients are divided into different data sets to avoid the same individual data appearing in the training set at the same time. With the test set, the resulting model test results are closest to the real application scenario, ensuring the reliability and universality of the model.
  • the ST segment and T wave change model used in this step can be as shown in Fig. 5. Specifically, it can be an end-to-end classification model inspired by models such as artificial neural network deep learning based convolutional neural networks AlexNet and VGG16. Specifically, the model is a 7-layer network with 7 convolutions, 5 pools, and 2 full connections. The convolution kernel used for convolution is 1x5, and the number of filters for each layer is different.
  • the number of layer 1 convolution filters is 96; the second layer convolution is combined with the third layer convolution, the number of filters is 256; the fourth layer convolution is combined with the fifth layer convolution, and the number of filters is 384; the number of the sixth layer convolution filter is 384; the number of the seventh layer convolution filter is 256; the first, third, fifth, sixth, and seventh layers of the convolution layer are pooled. Then there are two full connections, and finally the results are divided into two categories using the Softmax classifier. In order to increase the nonlinearity of the model and extract the features of higher dimensionality of the data, two convolutional modes are used.
  • Step 108 Perform P wave and T wave feature detection on the heartbeat analysis data according to the heartbeat time series data, and determine detailed feature information of the P wave and the T wave in each heart beat;
  • detailed feature information includes amplitude, direction, shape and start and stop time; in the analysis of heartbeat signals, the characteristics of P wave, T wave and QRS wave are also important basis in ECG analysis.
  • the P-wave, T-wave and QRS complexes are extracted by calculating the position of the segmentation point in the QRS complex and the position of the P- and T-wave segmentation points. . It can be realized by QRS group cut point detection, single lead PT detection algorithm and multi-lead PT detection algorithm.
  • QRS group segmentation point detection According to the QRS group segment power maximum point and the start and end points provided by the QRS complex detection algorithm, the R point, R' point, S point and S' point of the QRS group in a single lead are searched. When there is multi-lead data, the median of each segmentation point is calculated as the last segmentation point position.
  • P-wave and T-wave detection algorithms are relatively low in amplitude relative to QRS complex, and the signal is gentle, which is easy to be submerged in low-frequency noise, which is a difficult point in detection.
  • QRS complex detection the method uses a low-pass filter to perform third filtering on the signal after eliminating the influence of the QRS complex on the low frequency band, so that the relative amplitude of the PT wave is increased.
  • the T wave is then found between the two QRS complexes by peak detection. Because the T wave is a wave group generated by ventricular repolarization, there is a clear lock-time relationship between the T wave and the QRS wave group.
  • the midpoint between each QRS complex and the next QRS complex (such as the range between 400ms and 600ms after the first QRS complex) is used as the T-wave detection.
  • the largest peak is selected as the T wave in this interval.
  • the direction and morphological characteristics of the P wave and the T wave are determined based on the peak and position data of the P wave and the T wave.
  • the cutoff frequency of the low pass filtering is set between 10-30 Hz.
  • Multi-lead P-wave and T-wave detection algorithms In the case of multi-lead, since the generation time of each wave in the heart beat is the same, the spatial distribution is different, and the temporal and spatial distribution of noise is different, the traceability algorithm can be used to perform P, Detection of T waves.
  • the signal is first subjected to QRS complex elimination processing and the signal is third filtered using a low pass filter to remove interference.
  • Each individual component in the original waveform is then calculated by an independent component analysis algorithm. Among the separated independent components, according to the distribution characteristics of the peaks and the position of the QRS complex, the corresponding components are selected as the P-wave and T-wave signals, and the direction and morphological characteristics of the P-wave and the T-wave are determined.
  • Step 109 Perform secondary classification processing on the heartbeat analysis data according to the basic rule of the electrocardiogram, the detailed feature information of the P wave and the T wave, and the ST segment and the T wave evaluation information in one classification information, and finally obtain the heartbeat classification information;
  • the basic rule of ECG reference data is generated by following the basic rules describing the electrophysiological activity of cardiomyocytes and the clinical diagnosis of electrocardiogram in authoritative ECG textbooks, such as the minimum time interval between two heart beats, and the minimum of P and R waves.
  • Interval is used to subdivide the classification information after heartbeat classification; mainly based on the inter-heart rate RR interval and the medical saliency of different heartbeat signals on each lead; the heartbeat review module is based on the electrocardiogram
  • the basic rule reference data combined with the classification and recognition of a plurality of continuous heartbeat analysis data, and the detailed feature information of the P wave and the T wave, the ventricular heart beat classification is divided into finer heart beat classification, including: ventricular premature beat (V) , ventricular escape (VE), accelerated ventricular premature beats (VT), subventricular subtypic breakdown into supraventricular premature beats (S), atrial escape (SE), borderline escape (JE ) and atrial accelerated premature beats (AT) and so on.
  • V ventricular premature beat
  • VE ventricular escape
  • VT accelerated ventricular premature beats
  • S supraventricular premature beats
  • SE atrial escape
  • JE borderline escape
  • AT atrial accelerated premature beats
  • the secondary classification process it is also possible to correct the misclassification identification of the reference data that does not conform to the basic rules of the electrocardiogram occurring in the classification.
  • the subdivided heart beats are classified according to the reference data of the basic rules of the electrocardiogram, and the classification and identification which does not conform to the basic data of the electrocardiogram basic rules are found, and the classification is corrected according to the RR interval and the classification marks before and after.
  • a variety of heartbeat classifications can be output, such as: normal sinus beat (N), complete right bundle branch block (N_CRB), complete left bundle branch block (N_CLB), indoor resistance Hysteresis (N_VB), first degree atrioventricular block (N_B1), pre-excitation (N_PS), premature ventricular contraction (V), ventricular escape (VE), accelerated ventricular premature beat (VT), supraventricular premature beats ( S), atrial escape (SE), borderline escape (JE), accelerated atrial premature beats (AT), atrial flutter (AF), artifacts (A) and other classification results.
  • N normal sinus beat
  • N_CRB complete right bundle branch block
  • N_CLB complete left bundle branch block
  • N_VB indoor resistance Hysteresis
  • N_PS first degree atrioventricular block
  • V premature ventricular contraction
  • VE ventricular escape
  • VT accelerated ventricular premature beat
  • S supraventricular premature beats
  • SE atrial escape
  • JE borderline escape
  • AF accelerated
  • the heart rate parameters of the basic calculation include: RR interval, heart rate, QT time, QTc time and other parameters.
  • Step 110 performing signal quality analysis on the heartbeat analysis data, evaluating signal quality of the heartbeat analysis data, and obtaining signal quality evaluation parameters of the heartbeat analysis data;
  • the signal quality evaluation module performs a second filtering process and an envelope calculation on the RR interval signal by extracting the RR interval signal in the heartbeat analysis data, determines the noise signal strength, and calculates a relative heartbeat time sequence.
  • the signal-to-noise ratio of the maximum amplitude of the data is determined by the signal quality evaluation parameters.
  • the specific steps may include:
  • Step A extracting position information and width information of the QRS complex in the heartbeat analysis data
  • Step B extracting a signal of an RR interval between two adjacent QRS group signals
  • Step C filtering the signal of the RR interval, and performing envelope calculation on the filtered signal to obtain an average power of the noise signal of the RR interval; wherein, the average power of the noise signal of the RR interval is between the RRs The mean value of the envelope amplitude during the period.
  • Step D obtaining a signal quality evaluation parameter according to an average power of the noise signal and a power of the QRS group signal;
  • the signal quality evaluation parameter can be characterized by the noise level relative to the QRS complex during the RR interval. Specifically, it is calculated based on the power of the QRS complex and the average power of the noise signal.
  • S i is the amplitude of the i-th QRS complex signal
  • N i,t is the amplitude of the t-th sample point in the i-th RR interval interval
  • T is the RR interval length.
  • Step 111 Analyze and match the heartbeat classification information according to the reference data of the basic rule of the electrocardiogram, and process the result of the analysis and matching according to the signal quality evaluation parameter to generate ECG event data;
  • the event classification module performs pattern matching according to the heartbeat secondary classification result according to the basic rule of the electrocardiogram, and obtains the ECG event data in the heartbeat data in combination with the signal quality evaluation parameter.
  • the event classification module may classify the following typical ECG events corresponding to the ECG event data, including but not limited to:
  • Second degree type II (2:1) atrioventricular block
  • Step 112 Perform statistical analysis on the ECG event data to obtain corresponding report conclusion data and report item data; evaluate the ECG event data according to the signal quality evaluation parameter, generate report graphic data; and output report item data and report graphic data. And report conclusion data.
  • the report generation module removes the untrustworthy event in the ECG event data according to the reference signal quality evaluation parameter, calculates various heart rate parameters, counts the number of events and the time of occurrence, and obtains report conclusion data and report item data.
  • the specific heart rate parameter calculation includes calculating the average heart rate, and the maximum and minimum heart rate.
  • the maximum and minimum heart rate the fixed-length segment is used as the statistical unit, and the whole process scan and statistical comparison are performed one by one.
  • the length of the clip is generally 8-10 seconds, and can be freely set as needed.
  • the statistical calculation method for the heart rate of the sinus heart rate and the ectopic heart rate is different.
  • the maximum and minimum heart rate for the sinus heart rate in the main electrocardiogram, only the sinus classification heart rate is calculated; for the atrial flutter atrial fibrillation main body ECG, only the atrial flutter atrial fibrillation heartbeat and sinus beats are calculated; Other ectopic beats of atrial fibrillation accounted for the main electrocardiogram, and all types of heart beats were involved in calculations except for artifacts.
  • the report generation module can calculate the following heart rate parameters:
  • Atrial flutter atrial fibrillation as a percentage of total heart rate
  • the quality evaluation is performed according to the signal quality evaluation parameters, and the event fragment with the highest quality of the data signal is selected, and the number of event types included in the segment is analyzed, and the most representative segment is preferentially selected, and the starting position of the segment is preferred.
  • the event heartbeat is guaranteed to be in the middle of the selected segment to generate report graphic data.
  • the selection rule of the electrocardiogram event can be specifically as follows:
  • the segment starting point position is 0.3 second from the first heartbeat distance.
  • the starting position of the segment is 0.37 seconds away from the first heartbeat distance.
  • the interference-free signal ratio reaches the set threshold (preferably at 60%)
  • the threshold is determined in the -95% range, and the segment meets the screening conditions from which the optimal typical data segment is obtained.
  • report item data report graph data, and report conclusion data are output.
  • the present invention can also perform manual test based on the obtained classification result, and correct the heartbeat classification data which is inconsistent with the classification result of the automatic analysis output, and then input it as training sample feedback to the training.
  • Model for artificial intelligence self-learning The deep learning model can continuously improve the three classification models (interference two classification model, heart beat classification model, ST segment and T wave change model) used in the present invention by inputting new training data for continuous iterative loop training. Accuracy.
  • the specific implementation manners of the three classification models are specifically described in the above-mentioned execution steps, the specific implementation manners are not limited to the specific implementation manners. All three models can be implemented using the LeNet-5 model, one or more of the AlexNet model, the VGG16 model, and the Inception model. Other models can also be used, such as the ResNet model, the RNN-related model, the Auto-Encoding model, or Models such as SeqToSeq perform training and classification reasoning. At the same time, the number of categories of model classification is not limited, and the types of ECG events identified and the heart rate parameters of statistical analysis are not limited.
  • FIG. 7 is a schematic structural diagram of a device according to an embodiment of the present invention, where the device includes: a processor and a memory.
  • the memory can be connected to the processor via a bus.
  • the memory may be a non-volatile memory such as a hard disk drive and a flash memory in which software programs and device drivers are stored.
  • the software program can perform various functions of the above method provided by the embodiments of the present invention; the device driver can be a network and an interface driver.
  • the processor is configured to execute a software program, and when the software program is executed, the method provided by the embodiment of the present invention can be implemented.
  • the embodiment of the present invention further provides a computer readable storage medium.
  • the computer readable storage medium stores a computer program, and when the computer program is executed by the processor, the method provided by the embodiment of the present invention can be implemented.
  • Embodiments of the present invention also provide a computer program product comprising instructions.
  • the processor is caused to perform the above method.
  • the method and device for dynamic electrocardiogram analysis based on artificial intelligence self-learning provided by the embodiments of the present invention, through data preprocessing, heartbeat feature detection, interference signal detection using depth learning method, signal quality evaluation and guidance, and depth
  • the learning method performs heartbeat classification, heartbeat audit, ECG events and parameter analysis and calculation, and finally automatically outputs an automated analysis method of the report data in a complete and fast process.
  • the automatic analysis method of the invention can also record the modification information of the automatic analysis result, and collect the modified data feedback to the deep learning model to continue training, and continuously improve and improve the accuracy of the automatic analysis method.
  • the steps of a method or algorithm described in connection with the embodiments disclosed herein can be implemented in hardware, a software module executed by a processor, or a combination of both.
  • the software module can be placed in random access memory (RAM), memory, read only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or technical field. Any other form of storage medium known.

Landscapes

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

Abstract

一种基于人工智能自学习的动态心电图分析方法,包括：数据预处理，心搏特征检测，基于深度学习方法的干扰信号检测和心搏分类，信号质量评估与导联合并，心搏的审核，心电图事件和参数的分析计算，最终自动输出报告数据。该方法实现了一个完整快速流程的动态心电图的自动化分析方法，还可以记录对自动分析结果的修改信息，并收集这些修改的数据反馈给深度学习模型继续训练，不断地改进和提升自动分析方法的准确率。还公开了一种基于人工智能自学习的动态心电图分析装置。

Description

基于人工智能自学习的动态心电图分析方法和装置
本申请要求于2017年11月27日提交中国专利局、申请号为201711203048.2、发明名称为“基于人工智能自学习的动态心电图分析方法和装置”的中国专利申请的优先权。
技术领域
本发明涉及人工智能数据分析的技术领域，尤其涉及一种基于人工智能自学习的动态心电图分析方法和装置。
背景技术
无创心电图检查是心血管疾病(CVD)分析和筛查的重要工具。在无创心电图检查之中，动态心电图(Dynamic Electrocardiography，DCG)技术于1957年由Holter首先应用于监测心脏电生理活动的研究，所以又称Holter心电图仪，目前已成为临床心血管领域中无创检查的重要分析方法之一。动态心电图可连续记录24小时以上的心电活动全过程，包括休息、活动、进餐、工作、学习和睡眠等不同情况下的心电图信息。动态心电图能够发现常规心电图检查不易发现的非持续性心律失常，尤其能够提高对一过性心律失常及短暂的心肌缺血发作的检出率,是临床分析病情、确立分析、判断疗效的重要客观依据。另外，动态心电图可确定病人的心悸、头晕、昏厥等症状是否与心律失常有关，如窦性心动过缓、传导阻滞、快速心房颤动，阵发性室上性心动过速，持续性室性心动过速等，这是目前24小时动态心电图最重要、应用最广泛的情况之一。
与普通心电图相比，动态心电图扩大了心电图临床运用的范围，在医 院、家庭、体检中心和社区中有着广泛的应用空间。
动态心电图主要由心电信号采集记录仪、导联系统和计算机分析软件三部分组成。信号采集记录仪负责记录存储患者24小时甚至多天的心电图数据。由于患者日常活动中身体经常性变化比较大，受到的干扰情况比较多，所以对信号采集的频率响应、采样频率、分辨率、阻抗、抗干扰性能等要求比较高。高性能的心电信号记录仪采集的心电图信号波形信噪比高、信号保真强，这对于后续的分析计算帮助非常大。导联系统包括电极片和导联线，分析计算软件对采集的数字信号进行心律失常分析、长间歇停搏，扑动和颤动，传导阻滞，早搏和逸搏，心动过缓，心动过快，ST段改变检测、心电事件的分析与归类、帮助医生进行分析报告。
与临床的静态心电图使用的标准12导联体系不同，Holter一般使用Mason-Likar导联体系，为了进一步改进动态心电图的适应性，医学专家发明了M-L简化导联体系(Reduced Mason-Likar Lead System)，常见的M-L简化导联体系测量I，II，V2，和V5，或者I，II，V1，V5等四个导联的数据，可以根据心电向量关系计算出III，aVR，aVL，aVF四个导联的数据，分析显示使用II，V5，aVF三个导联数据；M-L简化导联体系Holter是市场上的主流产品；但是，由于其导联数据无法准确分析ST段的改变，所以部分医生机构也要求使用12导联Holter进行检查；随着移动医疗和可穿戴设备的普及应用，现在具有实时监护功能的单导联心贴片也在一些特定人群中开始流行，比如PCI术后病人，起搏器使用者，长跑运动员等等。
动态心电图的分析主要包括如下步骤：首先，进行信号滤波处理，然后检测识别心搏中的最显著P-QRS-T波群特征；其次，针对识别出来的QRS波、P波、T波、R-R间期、P-R间期、ST段等特征信号，根据心脏电生理活动对心搏进行分类；最后，结合心搏分类和心搏的节律性特征，将一些有规律的连续心搏活动进一步概括为心律事件。心电图专业医生根据心搏 分类和心律事件，对患者的动态心电图检查进行分析描述。
由于心脏细胞电生理活动产生的信号是一种弱信号(毫伏级)，所以正常心电图信号的频率范围是0.01Hz-100Hz，其中，90％的能量集中在0.25Hz-35Hz区间。因此，动态心电图在24小时长期连续的戴机过程中，信号很容易被干扰，主要干扰包括：工频干扰(Power Line Interference，由人体电容、设备电路、导联等产生的50Hz及其谐波构成的一种干扰)、肌电干扰(Electromyography,通常是肌肉活动引起的不规则的5-2000Hz高频干扰)、基线漂移(Baseline Drift，由电极接触不良、皮肤阻抗、呼吸运动等引起，频率范围一般在0.05Hz-2.00Hz之间，与心电图信号ST段和Q波频率分量非常接近)。
传统的动态心电图数字信号分析是基于Pan和Tompkins在1985年发表的论文(1)，该方法设计了一个信号分析计算流程：信号滤波；信号放大；信号差分；根据心电图信号和心脏电生理活动的一些基本规律，通过一些自适应阈值的计算，寻找每个P-QRS-T波群的特征；根据找到的每个P-QRS-T波以及前后的波群，对每个心搏进行分类识别；在心搏分类识别的基础上，通过分析心搏变化的规律，生成各种心律失常和传导阻滞等事件；医生通过使用一些客户端的分析工具，认真审核心搏和事件，最后给出分析描述。同时，从24小的动态心电图数据中，抓取每个事件的典型8秒片段，生成分析报告。针对MIT心律失常标准数据库的测试结果表明，该方法的识别正确率到达99.3％，成为动态心电图机器识别方法中的经典分析方法。市场上的主流动态心电图设备的分析方法，基本上都在Pan-Tompkins论文的基础上发展而来。
然而，上述方法的主要问题在于如下几点：第一，受到干扰信号的影响，抗干扰能力比较差。心搏分类识别无法主动排除干扰片段的影响。第二，在心搏特征提取中，无法对P、T准确识别。心搏检测经常存在多检和漏检。对一些特殊心电图信号，例如心律比较缓慢患者的高大T波，或者 T波肥大的信号，都经常容易多检。第三，对心搏的分类基本上停留在窦性、室上性和室性这三种，远远无法满足临床心电图医生复杂全面的分析要求。第四，无法准确识别房扑房颤、起搏器心搏和ST-T改变。因此，无法帮助房扑房颤类患者，对于帮助评价起搏器功能意义不大，也无法准确分析ST-T改变对心肌缺血分析的帮助。第五，对心搏和心电事件的识别还不够准确全面，受到前面许多因素的影响，很容易漏掉，同样会影响医生的判读。第六，现有的分析方法没有对事件片段的信号质量进行客观评估，没有对24小时数据进行全面的分析统计，报告摘要和事件截图仍然依靠医生的经验和能力，很容易造成数据没有反映整体检测的错误，报告事件缺失，患者报告截图质量不高或者不典型。第七，由于上述这些问题存在，无法做到从自动分析至最后的自动报告。医生还是需要花费大量的宝贵时间认真阅读动态心电图数据，无法从根本上帮助医生提高分析分析的能力，无论是质量还是效率。
当前，虽然许多厂家不同程度地修改了核心方法，提高了整体识别的准确率，但是由于患者动态心电图信号本身的复杂变异性，特别是前面提到的各种极端的干扰存在，导致在实际的应用过程中，Pan-Tompkins方法和后续的各厂家改进方法很难取得MIT标准数据库的测试结果。实际情况下，识别准确率不到60％。许多研究者也在尝试使用其他的方法来提高P-QRS-T波群特征的提取与识别，其中比较有影响的是Cuiwei Li等人1995年发表的小波变换方法。
关于心律事件分类，还没有一个类似于Pan-Tompkins的经典方法。常见的方法基本上是在QRS检测基础上，进行更加复杂的信号形态学分析，且根据动态心电图的一些基本原理，来完成心搏事件的归纳。如前所述，不可避免的大量干扰存在，各个导联信号采集的质量差异，以及患者本身的个体差异因素，导致这些方法很难取得满意的效果。近年来，不断有新的方法在试图提高分类准确率，诸如使用数学上比较复杂的小波变换方法、 Hilbert-Huang变换方法。然而，从实际使用情况看，仍没有太大的突破。这其中的主要问题仍然是识别准确率比较低，能够识别的心搏事件分类不足以覆盖动态心电图检查分析的需要。
虽然市场上大多数的动态心电图分析软件都可以对数据进行自动分析，但由于动态心电图信号本身的复杂与变异性，很容易受到戴机过程中各种干扰。另外，针对24小时平均10万左右的心搏数，目前自动分析软件的准确率远远不够，无法帮助求医生在比较短的时间内正确分析心电图数据并且给出正确的分析报告。
因此，如何帮助医院，特别是广大的各级基础医院，在没有足够多的专业医生解读动态心电图的困难条件下，通过使用最前沿的科技方法-人工智能技术，有效提高动态心电图的自动分析水平，是本发明所要解决的难题和挑战。
发明内容
本发明的目的是提供一种基于人工智能自学习的动态心电图分析方法和装置，能够对心电图设备输出的测量数据进行自动、快速、完整的分析，得到所需的报告数据。
为实现上述目的，本发明实施例第一方面提供了一种基于人工智能自学习的动态心电图分析方法，包括：
接收动态心电监测设备输出的动态心电图数据；将所述动态心电图数据的数据格式经过重采样转换为预设标准数据格式，并对转换后的预设标准数据格式的动态心电图数据进行第一滤波处理；
对所述第一滤波处理后的心电图数据进行心搏检测处理，识别所述心电图数据包括的多个心搏数据，每个所述心搏数据对应一个心搏周期，包括相应的P波、QRS波群、T波的幅值和起止时间数据；
根据所述心搏数据确定每个心搏的检测置信度；
根据训练得到的干扰识别二分类模型对所述心搏数据进行干扰识别，得到心搏数据是否存在干扰噪音，以及用于判断干扰噪音的一个概率值；
根据所述检测置信度确定心搏数据的有效性，并且，根据确定有效的心搏数据的导联参数和心搏数据，基于所述干扰识别的结果和时间规则合并生成心搏时间序列数据；根据所述心搏时间序列数据生成心搏分析数据；
根据训练得到的心搏分类模型对所述心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到所述心搏分析数据的一次分类信息；
对所述一次分类信息结果中的特定心搏的心搏分析数据输入到训练好的ST段和T波改变模型进行识别，确定ST段和T波评价信息；
根据所述心搏时间序列数据，对所述心搏分析数据进行P波和T波特征检测，确定每个心搏中P波和T波的详细特征信息，详细特征信息包括幅值、方向、形态和起止时间的数据；
对所述心搏分析数据在所述一次分类信息下根据所述心电图基本规律参考数据、所述P波和T波的详细特征信息以及所述ST段和T波评价信息进行二次分类处理，最终得到心搏分类信息；
对所述心搏分析数据进行信号质量分析，评估所述心搏分析数据的信号质量，得到所述心搏分析数据的信号质量评价参数；
根据心电图基本规律参考数据对所述心搏分类信息进行分析匹配，根据信号质量评价参数对所述分析匹配的结果进行处理，生成心电图事件数据；
对所述心电图事件数据进行统计分析，得到相应的报告结论数据和报告表项数据；根据信号质量评价参数对所述心电图事件数据进行评估，生成报告图形数据；并且输出所述报告表项数据、报告图形数据和报告结论数据。
优选的，所述确定检测置信度具体包括：
根据所述心搏数据确定RR间期并计算所述RR间期内噪声的估计值；
根据所述噪声的估计值和每个心搏数据中的最大幅值确定每个心搏数据的检测置信度。
优选的，所述根据训练得到的干扰识别二分类模型对所述心搏数据进行干扰识别具体包括：
以第一数据量对所述心搏数据进行切割采样，并将切割采样得到的数据输入到干扰识别二分类模型进行干扰识别；
识别心搏数据中，心搏间期大于等于预设间期判定阈值的数据片段；
对所述心搏间期大于等于预设间期判定阈值的数据片段进行信号异常判断，确定是否为异常信号；
如果不是异常信号，则以预设时间宽度，根据设定时值确定所述数据片段中滑动取样的起始数据点和终止数据点，并由所述起始数据点开始对所述数据片段进行滑动取样，至所述终止数据点为止，得到多个取样数据段；
对每个所述取样数据段进行所述干扰识别。
优选的，所述根据所述心搏时间序列数据生成心搏分析数据具体为：
根据预设阈值，对所述心搏时间序列数据中各导联心搏数据进行切割，从而生成各导联的心搏分析数据。
优选的，所述根据训练得到的心搏分类模型对所述心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到所述心搏分析数据的一次分类信息具体包括：
对于单导联的心搏分析数据，将所述心搏分析数据输入到训练得到的对应所述单导联心搏分类模型，以第二数据量进行幅值和时间表征数据的特征提取和分析，得到所述单导联的一次分类信息。
优选的，所述根据训练得到的心搏分类模型对所述心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到所述心搏分析数据的一次分类信息具体包括：
根据训练得到的各导联对应的心搏分类模型对各导联的心搏分析数据以第三数据量进行幅值和时间表征数据的特征提取和分析，得到各导联的分类信息；
根据各导联的分类信息和导联权重值参考系数进行分类投票决策计算，得到所述一次分类信息。
优选的，所述根据训练得到的心搏分类模型对所述心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到所述心搏分析数据的一次分类信息具体包括：
根据训练得到的多导联同步关联分类模型对各导联的心搏分析数据以第四数据量进行同步幅值和时间表征数据的特征提取和分析，得到心搏分析数据的一次分类信息。
优选的，所述对所述心搏分析数据进行信号质量分析，评估所述心搏分析数据的信号质量，得到所述心搏分析数据的信号质量评价参数具体为：
提取所述心搏分析数据中的RR间期信号，对所述RR间期信号进行第二滤波处理和包络计算，确定噪声信号强度，并计算相对于相应的心搏时间序列数据的最大振幅的信号噪声比，确定所述信号质量评价参数。
优选的，所述对所述一次分类信息结果中的特定心搏的心搏分析数据输入到训练好的ST段和T波改变模型进行识别，确定ST段和T波评价信息，具体包括：
将所述一次分类信息中的特定心搏数据，按导联依次输入到根据训练得到的ST段和T波改变模型，对各导联的特定心搏数据进行幅值和时间表征数据的特征提取和分析，得到各导联的ST段和T波改变信息，确定ST段和T波评价信息；所述ST段和T波评价信息具体为所述心搏片段数据对应的ST段和T波发生改变的导联位置信息。
优选的，所述对所述心搏分析数据进行P波和T波特征检测，得到详细特征信息具体包括：
对所述心搏分析数据进行QRS波群信号消除处理，并对所述QRS波群信号消除处理后的心搏分析数据进行第三滤波，通过独立成分分析算法对心搏分析数据进行数据分离，得到各独立成分序列数据，依据所述独立成分序列数据的峰值的分布特征以及QRS波群的位置，选取概率最大的独立成分作为相应的P波、T波成分，同时确定P波和T波的方向和形态特征。
优选的，所述对所述心搏分析数据进行P波和T波特征检测，得到详细特征信息具体包括：
对所述心搏分析数据进行P波、T波信号特征提取处理；
对所述信号特征数据进行峰值检测，确定相邻两个QRS波群之间的T波检测区间，并在所述T波检测区间内确定最大幅值的数据点为T波；
确定相邻两个QRS波群之间除T波检测区间之外的区间内最大幅值的数据点为P波；
根据P波和T波的峰值与位置数据，确定P波和T波的方向和形态特征。
优选的，所述根据信号质量评价参数对所述心电图事件数据进行评估，生成报告图形数据，具体包括：
根据所述信号质量评价参数，对每种心电图事件的数据片段进行评估，选择信号质量评价参数最高的数据片段为所述心电图事件中的典型数据片段，生成报告图形数据。
优选的，所述方法还包括：
接收对所述心搏分类信息的修改信息；
将修改后的数据作为训练样本数据，用于所述基于人工智能自学习的心电图自动分析方法中的模型训练。
本发明实施例提供的基于人工智能自学习的动态心电图分析方法，包括数据的预处理，心搏特征检测，基于深度学习方法的干扰信号检测和心搏分类，信号质量评估与导联合并，心搏的审核，心电图事件和参数的分 析计算，最终自动输出报告数据的一个完整快速流程的自动化分析方法。本发明的动态心电图分析方法，还可以记录对自动分析结果的修改信息，并收集这些修改的数据反馈给深度学习模型继续训练，不断地提升自动分析方法的准确率。
本发明实施例第二方面提供了一种设备，该设备包括存储器和处理器，存储器用于存储程序，处理器用于执行第一方面及第一方面的各实现方式中的方法。
本发明实施例第三方面提供了一种包含指令的计算机程序产品，当计算机程序产品在计算机上运行时，使得计算机执行第一方面及第一方面的各实现方式中的方法。
本发明实施例第四方面提供了一种计算机可读存储介质，计算机可读存储介质上存储有计算机程序，计算机程序被处理器执行时实现第一方面及第一方面的各实现方式中的方法。
附图说明
图1为本发明实施例提供的基于人工智能自学习的动态心电图分析的系统结构图；
图2为本发明实施例提供的基于人工智能自学习的动态心电图分析方法的流程图；
图3为本发明实施例提供的干扰识别二分类模型的示意图；
图4为本发明实施例提供的心搏分类模型的示意图；
图5为本发明实施例提供的ST段和T波改变模型的示意图；
图6为本发明实施例提供的ST段和T波改变定位分析流程的示意图；
图7为本发明实施例提供的基于人工智能自学习的动态心电图分析的设备结构示意图。
具体实施方式
下面通过附图和实施例，对本发明的技术方案做进一步的详细描述。
为便于理解本发明的技术方案，首先对人工智能模型尤其是卷积神经网络模型的基本原理进行介绍。
人工智能卷积神经网络(CNN)模型是深度学习中的监督学习方法，就是一个模拟神经网络的多层次网络(隐藏层hidden layer)连接结构，输入信号依次通过每个隐藏层，在其中进行一系列复杂的数学处理(Convolution卷积、Pooling池化、Regularization正则化、防止过拟合、Dropout暂时丢弃、Activation激活、一般使用Relu激活函数)，逐层自动地抽象出待识别物体的一些特征，然后把这些特征作为输入再传递到高一级的隐藏层进行计算，直到最后几层的全连接层(Full Connection)重构整个信号，使用Softmax函数进行逻辑(logistics回归)，达到多目标的分类。
CNN属于人工智能中的监督学习方法，在训练阶段，输入信号经过多个的隐藏层处理到达最后的全连接层，softmax逻辑回归得到的分类结果，与已知的分类结果(label标签)之间会有一个误差，深度学习的一个核心思想就是通过大量的样本迭代来不断地极小化这个误差，从而计算得到连接各隐藏层神经元的参数。这个过程一般需要构造一个特别的损失函数(cost function)，利用非线性优化的梯度下降算法和误差反向传播算法(backpropagation algorithm，BP)，快速有效地极小化整个深度(隐藏层的层数)和广度(特征的维数)都十分复杂的神经网络结构中所有连接参数。
深度学习把需要识别的图像输入到训练模型，经过第一隐藏层、第二隐藏层、第三隐藏层，最后是输出识别结果。每一层提取不同抽象程度的特征，最终识别出图像具体是什么类别，例如车、人或者动物。
深度学习的算法模型在数学上十分复杂。开发一个完整的算法程序需 要极强的专业背景知识和丰富的工作经验。近年来，Google、微软、百度、Facebook等公司和一些著名大学(比如加州伯克利大学、加拿大蒙特利尔大学)也都相继开发推出了各具特色的人工智能开发开源平台，帮助在深度学习领域的一些研发公司迅速掌握这个前沿技术。其中，伯克利的Caffe和Google的Tensorflow是目前使用最为广泛的两个框架工具。
由于深度学习模型极端复杂，而需要的训练数据从几十万、几百万到几千万，再加上反复循环迭代，导致非线性优化计算量非常庞大。对于实际的项目，使用普通计算机的中央处理器计算，耗时经常是从十几小时到数天，甚至更久。这种情况下，改用图形处理器(Graphics Processing Unit，GPU)计算可以极大地加快计算速度。目前，Nvidia公司提供的GPU卡，由于强大的图形与计算机视觉计算能力，大量的线性代数等计算库，支持并行处理，可以满足深度学习需要的各类方法计算，成为当前人工智能高性能训练与推理的基础硬件。
本发明的基于人工智能自学习的动态心电图分析方法，是基于CNN模型来实现的。
本发明实施例提供的基于人工智能自学习的动态心电图分析方法，基于图1所示的系统架构来实现。其系统架构包括心电监测设备1、数据存储传输装置2、分析系统硬件模块3、模型训练模块4和分析系统执行模块5，其中分析系统执行模块5具体包括：数据预处理模块501、心搏检测模块502、干扰识别模块503、导联心搏合并模块504、置信度计算模块505、心搏分类模块506、心搏审核模块507、P波和T波特征检测模块508、信号质量评估模块509、事件分类模块510、ST段和T波改变导联定位模块511和报告生成模块512。本发明通过系统架构中的分析系统硬件模块3、模型训练模块4和分析系统执行模块5实现人工智能自学习的动态心电图分析过程。
下面基于图1的系统架构，结合图2所示，对本发明的基于人工智能 自学习的动态心电图分析方法进行说明。
本方法主要包括如下步骤：
步骤100，接收动态心电监测设备输出的动态心电图数据；
具体的，动态心电监测设备将动态心电图电信号转换为数字信号输出，具体可以是单导联或多导联动态心电图设备测量采集输出的时间序列数据，也可以是通过数据库或者其他文件方式获得的单导联或多导联动态心电图数据，通过数据存储传输装置进行原始数据的存储，并可以通过WIFI、蓝牙，USB，3G/4G/5G移动通信网络，物联网等方式传输至分析系统硬件模块，并作为输入信号输入到分析系统执行模块。
作为记录动态心电图图形的动态心电图时间序列数据，由于不同设备厂家的动态心电图设备在采集模拟电路，滤波器和采样率诸方面的不同，产生的动态心电图时间序列数据在导联标记，数据编码格式，增益，精度，每秒的数据长度，基线位置等都有很大的差异，必须进行数据预处理，把所有输入的动态心电图时间序列数据根据本发明分析流程的要求，进行统一处理，按照预设标准数据格式进行存储。其中，预设标准数据格式，是指用于数据处理全流程的能够支持数据识别的数据格式。
步骤101，将动态心电图数据的数据格式经过重采样转换为预设标准数据格式，并对转换后的预设标准数据格式的心电图数据进行第一滤波处理；
具体的，通过数据预处理模块执行动态心电图数据的格式适配读取、重采样、滤波以及标准数据格式保存。动态心电图数据的格式适配读取，对不同的设备有不同的读取实现，读取后，需要调整基线、根据增益转换成毫伏数据。动态心电数据重采样，把数据转换成全流程能够处理的采样频率。然后通过滤波去除高频，低频的噪音干扰和基线漂移，提高人工智能分析准确率。将处理后的动态心电图数据以预设标准数据格式保存。
通过本步骤可以解决不同动态心电图设备在使用的导联，采样频率和 传输数据格式的差异，以及通过数字信号滤波去除高频，低频的噪音干扰和基线漂移。
在一个具体的例子中，重采样频率可以200Hz，保证每个固定时间范围的数据长度一致，进行计算、训练和推理，可以获得比较满意的分析结果而不会过分增加计算训练的复杂程度和时间。当然，随着心电图设备本身的采样率在逐渐的增大，GPU计算性能的不断跃升和成本下降，以及人工智能算法的不断创新优化，采样率还可以提高。
数字信号滤波可以分别采用高通滤波器，低通滤波器和中值滤波，把工频干扰、肌电干扰和基线漂移干扰消除，避免对后续分析的影响。
更具体的，可以采用低通、高通巴特沃斯滤波器进行零相移滤波，以去除基线漂移和高频干扰，保留有效的心电图信号；中值滤波则可以利用预设时长的滑动窗口内数据点电压幅值的中位数替代窗口中心序列的幅值。可以去除低频的基线漂移。
步骤102，对第一滤波处理后的动态心电图数据进行心搏检测处理，识别动态心电图数据包括的多个心搏数据；
每个心搏数据对应一个心搏周期，包括相应的P波、QRS波群、T波的幅值数据和起止时间数据；心搏检测模块执行的心搏检测包括QRS波群检测和P波，T波检测。QRS波群检测由两个过程构成，一是信号处理过程，从所述第一滤波处理后的心电图数据中提取QRS波群的特征频段；二是通过设置合理的阈值确定QRS波群的发生时间。在心电图中，一般会包含P波、QRS波群、T波成分以及噪声成分。一般QRS波群的频率范围在5到20Hz之间，可以通过一个在此范围内的带通滤波器提出QRS波群信号。然而P波、T波的频段以及噪声的频段和QRS波群频段有部分重叠，因此通过信号处理的方法并不能完全去除非QRS波群的信号。因此需要通过设置合理的阈值来从信号包络中提取QRS波群位置。具体的检测过程是一种基于峰值检测的过程。针对信号中每一个峰值顺序进行阈值判断，超过阈值 时进入QRS波群判断流程，进行更多特征的检测，比如RR间期、形态等。
由于动态心电图长时间佩戴记录过程中心搏信号的时域不平稳特性，心搏信号的幅度和频率时时刻刻都在变化，并且在疾病状态下，这种特性会表现的更强。在进行阈值设定时，需要根据数据特征在时域的变化情况动态的进行阈值调整。为了提高检测的准确率和阳性率，QRS波群检测大多采用双幅度阈值结合时间阈值的方式进行，高阈值具有更高的阳性率，低阈值具有更高的敏感率，在RR间期超过一定时间时(时间阈值)，使用低阈值进行检测，减少漏检情况。而低阈值由于阈值较低，容易受到T波、肌电噪声的影响，容易造成多检，因此优先使用高阈值进行检测。
对于不同导联的心搏数据，都具有导联参数，用以表征该心搏数据为哪个导联的心搏数据。因此在本步骤中，同时也确定了心搏数据的导联参数。
步骤103，根据心搏数据确定每个心搏的检测置信度；
具体的，置信度计算模块在心搏检测的过程中，根据QRS波群的幅度以及RR间期内噪声信号的幅度比例可以提供针对QRS波群检测置信度的估计值。
步骤104，根据训练得到的干扰识别二分类模型对所述心搏数据进行干扰识别，得到心搏数据是否存在干扰噪音，以及用于判断干扰噪音的一个概率值；
因为动态心电图检测长时间记录过程中易受多种影响出现干扰现象，导致获取数据无效或不准确，不能正确反映受测者的状况，同时也增加医生诊断难度及工作量；而且干扰数据也是导致智能分析工具无法有效工作的主要因素。因此，将外界信号干扰降到最低显得尤为重要。一直以来，在动态心电图的计算机自动分析方法中，对干扰信号进行准确识别，没有得到比较满意的解决，从而导致动态心电图计算机分析算法的总体准确率不高。
本步骤基于以深度学习算法为核心的端到端二分类识别模型，具有精度高，泛化性能强的特点，可有效地解决电极片脱落、运动干扰和静电干扰等主要干扰来源产生的扰动问题，克服了传统算法因干扰数据变化多样无规律而导致的识别效果差的问题。
具体的，干扰识别过程由干扰识别模块来完成，主要步骤可以包括：
步骤A，对心搏数据使用干扰识别二分类模型进行干扰识别；
步骤B，识别心搏数据中，心搏间期大于等于预设间期判定阈值的数据片段；
步骤C，对心搏间期大于等于预设间期判定阈值的数据片段进行信号异常判断，确定是否为异常信号；
其中，异常信号的识别主要包括是否为电极片脱落、低电压等情况。
步骤D，如果不是异常信号，则以预设时间宽度，根据设定时值确定数据片段中滑动取样的起始数据点和终止数据点，并由起始数据点开始对数据片段进行滑动取样，至终止数据点为止，得到多个取样数据段；
步骤E，对每个取样数据段进行干扰识别。
以一个具体的例子对上述步骤A-E进行说明。对每个导联的心搏数据以第一数据量进行切割采样，然后分别输入到干扰识别二分类模型进行分类，获得干扰识别结果和对应结果的一个概率值；对心搏间期大于等于2秒的心搏数据，先判断是否是信号溢出，低电压，电极脱落；如果不是上述情况，就按照第一数据量，从左边心搏开始，向右连续以第一数据量不重叠滑动取样，进行识别。
输入可以是任一导联的第一数据量心搏数据，然后采用干扰识别二分类模型进行分类，直接输出是否为干扰的分类结果，获得结果快，精确度高，稳定性好，可为后续分析提供更有效优质的数据。
上述干扰识别二分类模型的结构，是基于人工智能深度学习卷积神经网络LeNet-5和AlexNet等模型启发构建的一个端到端二分类识别系统。
对于该模型的训练，我们采用了来源于30万病人近400万精确标注的数据片段。标注分为两类：正常心电图信号或者是有明显干扰的心电图信号片段。我们通过定制开发的工具进行片段标注，然后以自定义标准数据格式保存干扰片段信息。
在训练过程，使用两台GPU服务器进行几十次轮循训练。在一个具体的例子中，采样率是200Hz，数据长度是300个心电图电压值(毫伏)的一个片段D[300]，输入数据是：InputData(i，j),其中，i是第i个导联，j是导联i第j个片段D。输入数据全部经过随机打散才开始训练，保证了训练过程收敛，同时，控制从同一个病人的心电图数据中收集太多的样本，提高模型的泛化能力，既真实场景下的准确率。训练收敛后，使用100万独立的测试数据进行测试，准确率可以到达99.3％。另有具体测试数据如下表1。
干扰 正常
敏感率(Sensitivity) 99.14％ 99.32％
阳性预测率(Positive Predicitivity) 96.44％ 99.84％
表1
因为干扰数据往往是由外界扰动因素的作用而引起的，主要有电极片脱落、低电压、静电干扰和运动干扰等情况，不但不同扰动源产生的干扰数据不同，而且相同扰动源产生的干扰数据也是多种多样；同时考虑到干扰数据虽然多样性布较广，但与正常数据的差异很大，所以在收集干扰的训练数据时也是尽可能的保证多样性，同时采取移动窗口滑动采样，尽可能增加干扰数据的多样性，以使模型对干扰数据更加鲁棒，即使未来的干扰数据不同于以往任何的干扰，但相比于正常数据，其与干扰的相似度也会大于正常数据，从而使模型识别干扰数据的能力增强。
本步骤中采用的干扰识别二分类模型可以如图3所示，网络首先使用2层卷积层，卷积核大小是1x5,每层后加上一个最大值池化。卷积核数目从128开始，每经过一次最大池化层，卷积核数目翻倍。卷积层之后是两个 全连接层和一个softmax分类器。由于该模型的分类数为2，所以softmax有两个输出单元，依次对应相应类别，采用交叉熵做为损失函数。
步骤105，根据检测置信度确定心搏数据的有效性，并且，根据确定有效的心搏数据以及其导联参数，基于所述干扰识别的结果和时间规则合并生成心搏时间序列数据，并根据心搏时间序列数据生成心搏分析数据；
由于心电图信号的复杂性以及每个导联可能受到不同程度的干扰影响，依靠单个导联检测心搏会存在多检和漏检的情况，不同导联检测到心搏结果的时间表征数据没有对齐，所以需要对所有导联的心搏数据根据干扰识别结果和时间规则进行合并，生成一个完整的心搏时间序列数据，统一所有导联心搏数据的时间表征数据。其中，时间表征数据用于表示每个数据点在心电图数据信号时间轴上的时间信息。根据这个统一的心搏时间序列数据，在后续的分析计算时，可以使用预先设置好的阀值，对各导联心搏数据进行切割，从而生成具体分析需要的各导联的心搏分析数据。
上述每个导联的心搏数据在合并前，需要根据步骤102中获得的检测置信度确定心搏数据的有效性。
具体的，导联心搏合并模块执行的心搏数据合并过程如下：根据心电图基本规律参考数据的不应期获取不同导联心搏数据的时间表征数据组合，丢弃其中偏差较大的心搏数据，对上述时间表征数据组合投票产生合并心搏位置，将合并心搏位置加入合并心搏时间序列，移动到下一组待处理的心搏数据，循环执行直至完成所有心搏数据的合并。
其中，心电图活动不应期可以优选在200毫秒至280毫秒之间。获取的不同导联心搏数据的时间表征数据组合应满足以下条件：心搏数据的时间表征数据组合中每个导联最多包含一个心搏数据的时间表征数据。在对心搏数据的时间表征数据组合进行投票时，使用检出心搏数据的导联数占有效导联数的百分比来决定；若心搏数据的时间表征数据对应导联的位置为低电压段、干扰段以及电极脱落时认为该导联对此心搏数据为无效导联。 在计算合并心搏具体位置时，可以采用心搏数据的时间表征数据平均值得到。在合并过程中，本方法设置了一个不应期来避免错误合并。
在本步骤中，通过合并操作输出一个统一的心搏时间序列数据。该步骤同时能够降低心搏的多检率和漏检率，有效的提高心搏检测的敏感度和阳性预测率。
步骤106，根据训练得到的心搏分类模型对心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到心搏分析数据的一次分类信息；
具体的，不同动态心电图设备在信号测量、采集或者输出的导联数据等方面存在的差异，以及实际应用场景的不同，本发明心搏分类模块中执行的心搏的分类方法，可以根据具体情况，使用简单的单导联分类方法，或者是多导联分类方法。多导联分类方法又包括导联投票决策分类方法和导联同步关联分类方法两种。导联投票决策分类方法是基于各导联的心搏分析数据进行导联独立分类，再把结果投票融合确定分类结果的投票决策方法；导联同步关联分类方法则采用对各导联的心搏分析数据进行同步关联分析的方法。单导联分类方法就是对单导联设备的心搏分析数据，直接使用对应导联模型进行分类，没有投票决策过程。下面对以上所述几种分类方法分别进行说明。
单导联分类方法包括：
根据心搏时间序列数据，以第二数据量将单导联心搏数据进行切割生成单导联的心搏分析数据，并输入到训练得到的对应该导联的心搏分类模型进行幅值和时间表征数据的特征提取和分析，得到单导联的所述一次分类信息。
导联投票决策分类方法可以具体包括：
第一步、根据心搏时间序列数据，以第三数据量对各导联心搏数据进行切割，从而生成各导联的心搏分析数据；
第二步、根据训练得到的各导联对应的心搏分类模型对各导联的心搏分 析数据进行幅值和时间表征数据的特征提取和分析，得到各导联的分类信息；
第三步、根据各导联的分类信息和导联权重值参考系数进行分类投票决策计算，得到所述一次分类信息。具体的，导联权重值参考系数是基于心电数据贝叶斯统计分析得到各导联对不同心搏分类的投票权重系数。
导联同步关联分类方法可以具体包括：
根据心搏时间序列数据，以第四数据量对各导联心搏数据进行切割，从而生成各导联的心搏分析数据；然后根据训练得到的多导联同步关联分类模型对各导联的心搏分析数据进行同步幅值和时间表征数据的特征提取和分析，得到心搏分析数据的一次分类信息。
心搏数据的同步关联分类方法输入是动态心电图设备所有导联数据，按照心搏分析数据统一的心搏位点，截取各导联上相同位置和一定长度的数据点，同步输送给经过训练的人工智能深度学习模型进行计算分析，输出是每个心搏位置点综合考虑了所有导联心电图信号特征，以及心搏在时间上前后关联的心律特征的准确心搏分类。
本方法充分考虑了动态心电图不同导联数据实际上就是测量了心脏电信号在不同的心电轴向量方向传递的信息流，把动态心电图信号在时间和空间上传递的多维度数字特征进行综合分析，极大地改进了传统方法仅仅依靠单个导联独立分析，然后把结果汇总进行一些统计学的投票方式而比较容易得出的分类错误的缺陷，极大地提高了心搏分类的准确率。
本步骤中的心搏分类模型，我们采用了训练集包含30万病人的1700万数据样本进行训练。这些样本是根据动态心电图分析诊断的要求对数据进行准确的标注产生的，标注主要是针对常见心律失常，传导阻滞以及ST段和T波改变，可满足不同应用场景的模型训练。具体以预设标准数据格式保存标注的信息。在训练数据的预处理上，为增加模型的泛化能力，对于样本量较少的分类做了小幅的滑动来扩增数据，具体的说，就是以每个 心搏为基础，按照一定步长(比如10-50个数据点)移动2次，这样就可以增加2倍的数据，提高了对这些数据量比较少的分类样本的识别准确率。经过实际结果验证，泛化能力也得到了改善。
在一个实际训练过程使用了两台GPU服务器进行几十次轮循训练，训练收敛后，使用500万独立的测试数据进行测试，准确率可以到达91.92％。
其中，训练数据的截取的长度，可以是1秒到10秒。比如采样率是200Hz，以2.5s为采样长度，取得的数据长度是500个心电图电压值(毫伏)的一个片段D[500]，输入数据是：InputData(i，j),其中，i是第i个导联，j是导联i第j个片段D。输入数据全部经过随机打散才开始训练，保证了训练过程收敛，同时，控制从同一个病人的心电图数据中收集太多的样本，提高模型的泛化能力，既真实场景下的准确率。训练时候，同步输入所有导联对应的片段数据D，按照图像分析的多通道分析方法，对每个时间位置的多个空间维度(不同心电轴向量)的导联数据进行同步学习，从而得到一个比常规算法更准确的分类结果。
本步骤中采用的心搏分类模型可以如图4所示，具体可以为基于人工智能深度学习的卷积神经网络AlexNet，VGG16，Inception等模型启发的端对端多标签分类模型。具体的讲，该模型的网络是一个7层的卷积网络，每个卷积之后紧跟一个激活函数。第一层是两个不同尺度的卷积层，之后是六个卷积层。七层卷积的卷积核分别是96,256,256,384,384,384,256。除第一层卷积核有两个尺度分别是5和11外，其他层卷积核尺度为5。第三、五、六、七层卷积层后是池化层。最后跟着两个全连接层。
步骤107，对一次分类信息结果中的特定心搏的心搏分析数据输入到训练好的ST段和T波改变模型进行识别，确定ST段和T波评价信息；
其中，ST段和T波评价信息具体为心搏分析数据对应的ST段和T波发生改变的导联位置信息。因为临床诊断要求对于ST段和T波的改变定位到具体的导联。
其中，一次分类信息的特定心搏数据是指包含窦性心搏(N)和其它可能包含ST改变的心搏类型的心搏分析数据。
ST段和T波改变导联定位模块将一次分类信息的特定心搏数据，按照每个导联依次输入到一个为识别ST段和T波改变的人工智能深度学习训练模型，进行计算分析，输出的结果说明导联片段的特征是否符合ST段和T波改变的结论，这样就可以确定ST段和T波改变发生的在具体那些导联的信息，即ST段和T波评价信息。具体方法可以是：把一次分类信息中结果是窦性心搏的各导联心搏分析数据，输入给ST段和T波改变模型，对窦性心搏分析数据进行逐一识别判断，以确定窦性心搏分析数据是否存在ST段和T波改变特征以及发生的具体导联位置信息，确定ST段和T波评价信息。具体执行过程示意图可以如图6所示。
带有ST段和T波改变的心搏在所有心搏中的占比较低，为了兼顾训练数据的多样性及各个类别数据量的均衡性，选取无ST段和T波改变以及有ST段和T波改变的训练数据比例约为2:1，保证了模型在分类过程中良好的泛化能力且不出现对训练数据占比较多一类的倾向性。由于心搏的形态多种多样，不同个体表现的形态不尽相同，因此，为了模型更好估计各分类的分布，能有效提取特征，训练样本从不同年龄，体重，性别和居住地区的个体收集；另外，因为单个个体在同一时间段内的心电图数据往往是高度相似的，所以为了避免过度学习，在获取单个个体的数据时，从所有数据中随机选取不同时间段的少量样本；最后，由于患者的心搏形态存在个体间差异大，而个体内相似度高的特点，因而在划分训练、测试集时，把不同的患者分到不同的数据集，避免同一个体的数据同时出现在训练集与测试集中，由此，所得模型测试结果最接近真实应用场景，保证了模型的可靠性和普适性。
本步骤中采用的ST段和T波改变模型可以如图5所示，具体可以为基于人工智能深度学习的卷积神经网络AlexNet和VGG16等模型启发的端对 端分类模型。具体的讲，该模型是一个7层的网络，模型包含了7个卷积，5个池化和2个全连接。卷积使用的卷积核均为1x5，每层卷积的滤波器个数各不相同。第1层卷积滤波器个数为96；第2层卷积和第3层卷积连用，滤波器个数为256；第4层卷积和第5层卷积连用，滤波器个数为384；第6层卷积滤波器个数为384；第7层卷积滤波器个数为256；第1、3、5、6、7层卷积层后是池化。随后是两个全连接，最后还采用Softmax分类器将结果分为两类。为了增加模型的非线性，提取数据更高维度的特征，故采用两个卷积连用的模式。
步骤108，根据心搏时间序列数据，对心搏分析数据进行P波和T波特征检测，确定每个心搏中P波和T波的详细特征信息；
其中，详细特征信息包括幅值、方向、形态和起止时间等数据；在对心搏信号的分析中，P波、T波以及QRS波中的各项特征也是心电图分析中的重要依据。
在P波和T波特征检测模块中，通过计算QRS波群中切分点位置，以及P波和T波的切分点位置，来提取P波、T波以及QRS波群中的各项特征。可以分别通过QRS波群切分点检测、单导联PT检测算法和多导联PT检测算法来实现。
QRS波群切分点检测：根据QRS波群检测算法提供的QRS波群段功率最大点以及起止点，寻找单个导联中QRS波群的R点，R’点，S点以及S’点。在存在多导联数据时，计算各个切分点的中位数作为最后的切分点位置。
单导联P波、T波检测算法：P波和T波相对QRS波群幅度低、信号平缓，容易淹没在低频噪声中，是检测中的难点。本方法依据QRS波群检测的结果，在消除QRS波群对低频频段的影响后，使用低通滤波器对信号进行第三滤波，使PT波相对幅度增高。之后通过峰值检测的方法在两个QRS波群之间寻找T波。因为T波是心室复极产生的波群，因此T波和QRS波 群之间有明确的锁时关系。以检测到的QRS波群为基准，在每个QRS波群到下一个QRS波群间期取中点(比如限制在第一个QRS波群后400ms到600ms之间的范围)作为T波检测结束点，在此区间内选取最大的峰作为T波。再在剩余的峰值内选择幅度最大的峰为P波。同时也根据P波和T波的峰值与位置数据，确定P波和T波的方向和形态特征。优选的，低通滤波的截止频率设置为10-30Hz之间。
多导联P波、T波检测算法：在多导联的情况中，由于心搏中各个波的产生时间相同，空间分布不同，而噪声的时间空间分布不同，可以通过溯源算法来进行P、T波的检测。首先对信号进行QRS波群消除处理并使用低通滤波器对信号进行第三滤波以去除干扰。之后通过独立成分分析算法计算原始波形中的各个独立成分。在分离出的各个独立成分中，依据其峰值的分布特征以及QRS波群位置，选取相应的成分作为P波和T波信号，同时确定P波和T波的方向和形态特征。
步骤109，对心搏分析数据在一次分类信息下根据心电图基本规律参考数据、P波和T波的详细特征信息以及ST段和T波评价信息进行二次分类处理，最终得到心搏分类信息；
具体的，心电图基本规律参考数据是遵循权威心电图教科书中对心肌细胞电生理活动和心电图临床诊断的基本规则描述生成的，比如两个心搏之间最小的时间间隔，P波与R波的最小间隔等等，用于将心搏分类后的一次分类信息再进行细分；主要根据是心搏间RR间期以及不同心搏信号在各导联上的医学显著性；心搏审核模块依据心电图基本规律参考数据结合一定连续多个心搏分析数据的分类识别，以及P波和T波的详细特征信息将室性心搏分类拆分更细的心搏分类，包括：室性早搏(V)、室性逸搏(VE)、加速性室性早搏(VT)，将室上性类心搏细分为室上性早搏(S)、房性逸搏(SE)、交界性逸搏(JE)和房性加速性早搏(AT)等等。
此外，通过二次分类处理，还可以纠正一次分类中发生的不符合心电图 基本规律参考数据的错误分类识别。将细分后的心搏分类按照心电图基本规律参考数据进行模式匹配，找到不符合心电图基本规律参考数据的分类识别，根据RR间期及前后分类标识纠正为合理的分类。
具体的，经过二次分类处理，可以输出多种心搏分类，比如：正常窦性心搏(N)、完全性右束支阻滞(N_CRB)、完全性左束支阻滞(N_CLB)、室内阻滞(N_VB)、一度房室传导阻滞(N_B1)、预激(N_PS)、室性早搏(V)、室性逸搏(VE)、加速性室性早搏(VT)、室上性早搏(S)、房性逸搏(SE)、交界性逸搏(JE)、加速性房性早搏(AT)、房扑房颤(AF)、伪差(A)等分类结果。
通过本步骤，还可以完成基础心率参数的计算。其中基础计算的心率参数包括：RR间期、心率、QT时间、QTc时间等参数。
步骤110，对心搏分析数据进行信号质量分析，评估心搏分析数据的信号质量，得到心搏分析数据的信号质量评价参数；
具体的，信号质量评估模块通过提取心搏分析数据中的RR间期信号，对RR间期信号进行第二滤波处理和包络计算，确定噪声信号强度，并计算相对于相应的心搏时间序列数据的最大振幅的信号噪声比，确定信号质量评价参数。其具体步骤可以包括：
步骤A，提取心搏分析数据中QRS波群的位置信息和宽度信息；
步骤B，提取相邻两个QRS波群信号之间的RR间期的信号；
步骤C，对RR间期的信号进行滤波处理，并对滤波处理后的信号进行包络计算，得到RR间期的噪声信号的平均功率；其中，RR间期的噪声信号的平均功率为RR间期内包络幅度的均值。
步骤D，根据噪声信号的平均功率和QRS波群信号的功率得到信号质量评价参数；
进一步具体的，信号质量评价参数可以由RR间期内相对于QRS波群的噪声水平来表征。具体通过根据QRS波群的功率和噪声信号的平均功率计算得到。
以公式表示为：
Figure PCTCN2018072359-appb-000001
其中，S i为第i个QRS波群信号的幅度，N i,t为第i个RR间期区间内第t个样本点的幅值，T为RR间期长度。
步骤111，根据心电图基本规律参考数据对心搏分类信息进行分析匹配，根据信号质量评价参数对分析匹配的结果进行处理，生成心电图事件数据；
具体的，事件分类模块根据心搏二次分类结果，按照心电图基本规律参考数据进行模式匹配，并结合信号质量评价参数得到心搏数据中的心电图事件数据。
具体的，事件分类模块可以分类对应于心电图事件数据的以下这些典型的心电图事件，包括但不限于：
室上性早搏
室上性早搏成对
室上性早搏二联律
室上性早搏三联律
房性逸搏
房性逸搏心律
交界性逸搏
交界性逸搏心律
非阵发性室上性心动过速
最快室上性心动过速
最长室上性心动过速
室上性心动过速
短阵室上性心动过速
心房扑动-心房颤动
室性早搏
室性早搏成对
室性早搏二联律
室性早搏三联律
室性逸搏
室性逸搏心律
加速性室性自主心律
最快室性心动过速
最长室性心动过速
室性心动过速
短阵室性心动过速
二度I型窦房传导阻滞
二度II型窦房传导阻滞
一度房室传导阻滞
二度I型房室传导阻滞
二度II型房室传导阻滞
二度II型(2：1)房室传导阻滞
高度房室传导阻滞
完全性左束支阻滞
完全性右束支阻滞
室内阻滞
预激综合症
ST段和T波改变
最长RR间期
步骤112，对心电图事件数据进行统计分析，得到相应的报告结论数据和报告表项数据；根据信号质量评价参数对心电图事件数据进行评估，生 成报告图形数据；并且输出报告表项数据、报告图形数据和报告结论数据。
具体的，报告生成模块根据参考信号质量评价参数，剔除心电图事件数据中不可信事件，计算各种心率参数，统计事件的数量和发生时间等，得到报告结论数据和报告表项数据。
具体的心率参数计算包括计算平均心率、及最大、最小心率等。在计算最大和最小心率时，以固定长度片段为统计单位，逐个心搏进行全程扫描和统计比较。片段的长度一般为8-10秒，也可以根据需要自由设定。计算心率时，针对窦性心率占主体和异位心率占主体采用不同的心搏类型统计计算方法。在计算最大最小心率时：针对窦性心率占主体的心电图，仅计算窦性分类心搏；针对房扑房颤类占主体的心电图，仅计算房扑房颤类心搏及窦性心搏；针对非房扑房颤的其他异位心搏占主体的心电图，除伪差外所有类型心搏均参与计算。
具体的，报告生成模块可以计算下述的心率参数：
时间：
监测总时长
干扰总时长
有效时长
开始结束时间
心率：
平均心率
最大心率，及发生时间
最小心率，及发生时间
总心搏数量
异常心搏占总心搏百分比
房扑房颤占总心搏百分比
心率变异性：
窦性心搏间标准差
RR间期平均值标准差
RR间期：
长RR间期(1.5-2.0s)发生次数
3s>长RR间期>2.0s发生次数
5s>长RR间期>3s发生次数
长RR间期>5s发生次数
最长RR间期(s)及发生时间
以及各种心电图事件的数量，或者持续时间，发生时间等等。
对心电图事件数据，根据信号质量评价参数进行质量评估，选取数据信号质量最高的事件片段，同时分析片段中的包含的事件种类数量，优先选取最具代表性的片段，优选片段的开始位置，尽量保证事件心搏位于所选取片段的中部，生成报告图形数据。
在优选的方案中，心电图事件的选择规则可以具体如下例子所述：
对于一般心电图事件，选取单个片段。当片段第一个心搏心率大于等于100时，片段起始点位置距离第一个心搏距离为0.3秒。当片段第一个心搏心率小于等于45时，片段起始点位置距离第一个心搏距离为0.37秒。
对于有多个片段的心电图事件，需要对多个片段进行质量评估，依据信号质量评价参数计算截取片段的无干扰信号占比，当无干扰信号占比达到设定的阈值(优选的在60％-95％范围中确定阈值)，片段符合筛选条件，从中获得最优典型数据段。
最后，输出报告表项数据、报告图形数据和报告结论数据。
在上述过程之外，本发明还可以基于得到的分类结果，进行人工检验，并将检验出与自动分析输出的分类结果不符的心搏分类数据再进行修正，作为训练样本反馈输入给训练好的模型，用于人工智能自学习。深度学习的模型可以通过输入新的训练数据进行不断的迭代循环训练，从而持续改进本发明所使用的3个分类模型(干扰二分类模型，心搏分类模型，ST段和T波改变模型)的准确率。
需要说明，在上述执行步骤的过程中虽然对3个分类模型的具体实现方式进行了具体说明，但并不限定其具体的实现方式只能采用所例举的方式实现。三个模型均可采用LeNet-5模型，AlexNet模型、VGG16模型、Inception模型中的一个或多个来实现，也可以使用其它一些模型，比如，ResNet模型，RNN相关的模型，Auto-Encoding模型或者SeqToSeq等模型进行训练和分类推理，同时，也不限定模型分类的类别数量，不限定识别的心电图事件种类和统计分析的心率参数。
图7为本发明实施例提供的一种设备结构示意图，该设备包括：处理器和存储器。存储器可通过总线与处理器连接。存储器可以是非易失存储器，例如硬盘驱动器和闪存，存储器中存储有软件程序和设备驱动程序。软件程序能够执行本发明实施例提供的上述方法的各种功能；设备驱动程序可以是网络和接口驱动程序。处理器用于执行软件程序，该软件程序被执行时，能够实现本发明实施例提供的方法。
需要说明的是，本发明实施例还提供了一种计算机可读存储介质。该计算机可读存储介质上存储有计算机程序，该计算机程序被处理器执行时，能够实现本发明实施例提供的方法。
本发明实施例还提供了一种包含指令的计算机程序产品。当该计算机程序产品在计算机上运行时，使得处理器执行上述方法。
本发明实施例提供的基于人工智能自学习的动态心电图分析方法及设备，通过数据的预处理，心搏特征检测，使用深度学习方法进行的干扰信 号检测，信号质量评估与导联合并，采用深度学习方法进行心搏分类，心搏的审核，心电事件和参数的分析计算，最终自动输出报告数据的一个完整快速流程的自动化分析方法。本发明的自动分析方法，还可以记录对自动分析结果的修改信息，并收集这些修改的数据反馈给深度学习模型继续训练，不断地改进和提升自动分析方法的准确率。
专业人员应该还可以进一步意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、计算机软件或者二者的结合来实现，为了清楚地说明硬件和软件的可互换性，在上述说明中已经按照功能一般性地描述了各示例的组成及步骤。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本发明的范围。
结合本文中所公开的实施例描述的方法或算法的步骤可以用硬件、处理器执行的软件模块，或者二者的结合来实施。软件模块可以置于随机存储器(RAM)、内存、只读存储器(ROM)、电可编程ROM、电可擦除可编程ROM、寄存器、硬盘、可移动磁盘、CD-ROM、或技术领域内所公知的任意其它形式的存储介质中。
以上所述的具体实施方式，对本发明的目的、技术方案和有益效果进行了进一步详细说明，所应理解的是，以上所述仅为本发明的具体实施方式而已，并不用于限定本发明的保护范围，凡在本发明的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。

Claims (16)

  1. 一种基于人工智能自学习的动态心电图分析方法，其特征在于，所述方法包括：
    接收动态心电监测设备输出的动态心电图数据；将所述动态心电图数据的数据格式经过重采样转换为预设标准数据格式，并对转换后的预设标准数据格式的动态心电图数据进行第一滤波处理；
    对所述第一滤波处理后的动态心电图数据进行心搏检测处理，识别所述动态心电图数据包括的多个心搏数据，每个所述心搏数据对应一个心搏周期，包括相应的P波、QRS波群、T波的幅值和起止时间数据；
    根据所述心搏数据确定每个心搏的检测置信度；
    根据训练得到的干扰识别二分类模型对所述心搏数据进行干扰识别，得到心搏数据是否存在干扰噪音，以及用于判断干扰噪音的一个概率值；
    根据所述检测置信度确定心搏数据的有效性，并且，根据确定有效的心搏数据的导联参数和心搏数据，基于所述干扰识别的结果和时间规则合并生成心搏时间序列数据；根据所述心搏时间序列数据生成心搏分析数据；
    根据训练得到的心搏分类模型对所述心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到所述心搏分析数据的一次分类信息；
    对所述一次分类信息结果中的特定心搏的心搏分析数据输入到训练好的ST段和T波改变模型进行识别，确定ST段和T波评价信息；
    根据所述心搏时间序列数据，对所述心搏分析数据进行P波和T波特征检测，确定每个心搏中P波和T波的详细特征信息，详细特征信息包括幅值、方向、形态和起止时间的数据；
    对所述心搏分析数据在所述一次分类信息下根据所述心电图基本规律参考数据、所述P波和T波的详细特征信息以及所述ST段和T波评价信息进行二次分类处理，最终得到心搏分类信息；
    对所述心搏分析数据进行信号质量分析，评估所述心搏分析数据的信号质量，得到所述心搏分析数据的信号质量评价参数；
    根据心电图基本规律参考数据对所述心搏分类信息进行分析匹配，根据信号质量评价参数对所述分析匹配的结果进行处理，生成心电图事件数据；
    对所述心电图事件数据进行统计分析，得到相应的报告结论数据和报告表项数据；根据信号质量评价参数对所述心电图事件数据进行评估，生成报告图形数据；并且输出所述报告表项数据、报告图形数据和报告结论数据。
  2. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述确定检测置信度具体包括：
    根据所述心搏数据确定RR间期并计算所述RR间期内噪声的估计值；
    根据所述噪声的估计值和每个心搏数据中的最大幅值确定每个心搏数据的检测置信度。
  3. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述根据训练得到的干扰识别二分类模型对所述心搏数据进行干扰识别具体包括：
    以第一数据量对所述心搏数据进行切割采样，并将切割采样得到的数据输入到干扰识别二分类模型进行干扰识别；
    识别心搏数据中，心搏间期大于等于预设间期判定阈值的数据片段；
    对所述心搏间期大于等于预设间期判定阈值的数据片段进行信号异常判断，确定是否为异常信号；
    如果不是异常信号，则以预设时间宽度，根据设定时值确定所述数据片段中滑动取样的起始数据点和终止数据点，并由所述起始数据点开始对所述数据片段进行滑动取样，至所述终止数据点为止，得到多个取样数据段；
    对每个所述取样数据段进行所述干扰识别。
  4. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述根据所述心搏时间序列数据生成心搏分析数据具体为：
    根据预设阈值，对所述心搏时间序列数据中各导联心搏数据进行切割，从而生成各导联的心搏分析数据。
  5. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述根据训练得到的心搏分类模型对所述心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到所述心搏分析数据的一次分类信息具体包括：
    对于单导联的心搏分析数据，将所述心搏分析数据输入到训练得到的对应所述单导联心搏分类模型，以第二数据量进行幅值和时间表征数据的特征提取和分析，得到所述单导联的一次分类信息。
  6. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述根据训练得到的心搏分类模型对所述心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到所述心搏分析数据的一次分类信息具体包括：
    根据训练得到的各导联对应的心搏分类模型对各导联的心搏分析数据以第三数据量进行幅值和时间表征数据的特征提取和分析，得到各导联的分类信息；
    根据各导联的分类信息和导联权重值参考系数进行分类投票决策计算，得到所述一次分类信息。
  7. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述根据训练得到的心搏分类模型对所述心搏分析数据进行幅值和时间表征数据的特征提取和分析，得到所述心搏分析数据的一次分类信息具体包括：
    根据训练得到的多导联同步关联分类模型对各导联的心搏分析数据以第四数据量进行同步幅值和时间表征数据的特征提取和分析，得到心搏分析数据的一次分类信息。
  8. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述对 所述心搏分析数据进行信号质量分析，评估所述心搏分析数据的信号质量，得到所述心搏分析数据的信号质量评价参数具体为：
    提取所述心搏分析数据中的RR间期信号，对所述RR间期信号进行第二滤波处理和包络计算，确定噪声信号强度，并计算相对于相应的心搏时间序列数据的最大振幅的信号噪声比，确定所述信号质量评价参数。
  9. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述对所述一次分类信息结果中的特定心搏的心搏分析数据输入到训练好的ST段和T波改变模型进行识别，确定ST段和T波评价信息，具体包括：
    将所述一次分类信息中的特定心搏数据，按导联依次输入到根据训练得到的ST段和T波改变模型，对各导联的特定心搏数据进行幅值和时间表征数据的特征提取和分析，得到各导联的ST段和T波改变信息，确定ST段和T波评价信息；所述ST段和T波评价信息具体为所述心搏片段数据对应的ST段和T波发生改变的导联位置信息。
  10. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述对所述心搏分析数据进行P波和T波特征检测，得到详细特征信息具体包括：
    对所述心搏分析数据进行QRS波群信号消除处理，并对所述QRS波群信号消除处理后的心搏分析数据进行第三滤波，通过独立成分分析算法对心搏分析数据进行数据分离，得到各独立成分序列数据，依据所述独立成分序列数据的峰值的分布特征以及QRS波群的位置，选取概率最大的独立成分作为相应的P波、T波成分，同时确定P波和T波的方向和形态特征。
  11. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述对所述心搏分析数据进行P波和T波特征检测，得到详细特征信息具体包括：
    对所述心搏分析数据进行P波、T波信号特征提取处理；
    对所述信号特征数据进行峰值检测，确定相邻两个QRS波群之间的T 波检测区间，并在所述T波检测区间内确定最大幅值的数据点为T波；
    确定相邻两个QRS波群之间除T波检测区间之外的区间内最大幅值的数据点为P波；
    根据P波和T波的峰值与位置数据，确定P波和T波的方向和形态特征。
  12. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述根据信号质量评价参数对所述心电图事件数据进行评估，生成报告图形数据，具体包括：
    根据所述信号质量评价参数，对每种心电图事件的数据片段进行评估，选择信号质量评价参数最高的数据片段为所述心电图事件中的典型数据片段，生成报告图形数据。
  13. 根据权利要求1所述的动态心电图分析方法，其特征在于，所述方法还包括：
    接收对所述心搏分类信息的修改信息；
    将修改后的数据作为训练样本数据，用于所述基于人工智能自学习的心电图自动分析方法中的模型训练。
  14. 一种设备，包括存储器和处理器，其特征在于，所述存储器用于存储程序，所述处理器用于执行如权利要求1至13任一项所述的方法。
  15. 一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行如权利要求1至13任一项所述的方法。
  16. 一种计算机可读存储介质，包括指令，当所述指令在计算机上运行时，使所述计算机执行根据权利要求1至13任一项所述的方法。
PCT/CN2018/072359 2017-11-27 2018-01-12 基于人工智能自学习的动态心电图分析方法和装置 WO2019100565A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
EP18881678.9A EP3692904A4 (en) 2017-11-27 2018-01-12 METHOD AND DEVICE FOR DYNAMIC SELF-LEARNING ELECTROCARDIOGRAPHY ANALYSIS USING ARTIFICIAL INTELLIGENCE
US16/651,912 US11234629B2 (en) 2017-11-27 2018-01-12 Method and device for self-learning dynamic electrocardiography analysis employing artificial intelligence

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201711203048.2 2017-11-27
CN201711203048.2A CN107951485B (zh) 2017-11-27 2017-11-27 基于人工智能自学习的动态心电图分析方法和装置

Publications (1)

Publication Number Publication Date
WO2019100565A1 true WO2019100565A1 (zh) 2019-05-31

Family

ID=61961871

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/072359 WO2019100565A1 (zh) 2017-11-27 2018-01-12 基于人工智能自学习的动态心电图分析方法和装置

Country Status (4)

Country Link
US (1) US11234629B2 (zh)
EP (1) EP3692904A4 (zh)
CN (1) CN107951485B (zh)
WO (1) WO2019100565A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111177280A (zh) * 2019-12-26 2020-05-19 北京亚信数据有限公司 一种数据权威性评价方法和装置
CN111493858A (zh) * 2020-03-16 2020-08-07 浙江好络维医疗技术有限公司 一种基于聚类分析的单导联特异主波识别定位方法
CN111904410A (zh) * 2020-06-19 2020-11-10 陕西省医疗器械质量监督检验院 一种动态心电图准确性的检测系统及检测方法
CN113080995A (zh) * 2021-03-17 2021-07-09 深圳邦健生物医疗设备股份有限公司 心动过速属性的识别方法、装置、设备和介质
CN113796873A (zh) * 2021-10-26 2021-12-17 山东建筑大学 穿戴式动态心电信号分类方法及系统
CN114219028A (zh) * 2021-12-15 2022-03-22 东软集团股份有限公司 心电st段抬高识别模型的训练、st段抬高识别方法及装置
US11568991B1 (en) 2020-07-23 2023-01-31 Heart Input Output, Inc. Medical diagnostic tool with neural model trained through machine learning for predicting coronary disease from ECG signals

Families Citing this family (75)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11013471B2 (en) * 2018-04-26 2021-05-25 Vektor Medical, Inc. Display of an electromagnetic source based on a patient-specific model
US11259871B2 (en) 2018-04-26 2022-03-01 Vektor Medical, Inc. Identify ablation pattern for use in an ablation
CN108875575B (zh) * 2018-05-12 2021-05-11 鲁东大学 基于多通道信号融合深度神经网络自动心律失常分析方法
CN108764457B (zh) * 2018-05-12 2019-10-18 鲁东大学 基于采样通道融合深度神经网络的自动心律失常分析方法
CN110610183A (zh) * 2018-06-15 2019-12-24 佛山市顺德区美的电热电器制造有限公司 谷物评估方法、装置和存储介质
CN109117729A (zh) * 2018-07-11 2019-01-01 上海夏先机电科技发展有限公司 心电图室性逸搏实时判断方法、装置、系统及存储介质
CN109117730B (zh) * 2018-07-11 2021-05-11 上海夏先机电科技发展有限公司 心电图心房颤动实时判断方法、装置、系统及存储介质
CN109044347B (zh) * 2018-07-11 2021-08-27 上海夏先机电科技发展有限公司 识别心电波图像交界性逸搏方法、装置、系统和电子设备
CN109009143B (zh) * 2018-07-12 2021-01-29 杭州电子科技大学 一种通过人体步态预测心电信息的方法
CN108836314A (zh) * 2018-07-13 2018-11-20 希蓝科技（北京）有限公司 一种基于网络和人工智能的动态心电图分析方法及系统
CN109091139B (zh) * 2018-07-13 2021-04-30 希蓝科技（北京）有限公司 一种修正阵发性心律失常事件的方法及装置
CN108937915B (zh) * 2018-07-24 2021-02-19 杭州质子科技有限公司 一种动态心电图中早搏的识别方法
CN110786847B (zh) * 2018-08-02 2022-11-04 深圳市理邦精密仪器股份有限公司 心电信号的建库方法和分析方法
CN109276242A (zh) * 2018-08-02 2019-01-29 深圳市理邦精密仪器股份有限公司 心电信号类型识别的方法和装置
CN109214318B (zh) * 2018-08-22 2021-10-22 北京天泽智云科技有限公司 一种寻找非稳态时间序列微弱尖峰的方法
CN110897626A (zh) * 2018-09-14 2020-03-24 杭州脉流科技有限公司 基于深度神经网络的心电图分析方法、装置、计算机设备和系统
CN110897628A (zh) * 2018-09-14 2020-03-24 杭州脉流科技有限公司 基于深度神经网络的心电图信号特征提取方法、装置、系统、设备和分类方法
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
CN109157200B (zh) * 2018-10-31 2021-03-23 安徽华米信息科技有限公司 噪声检测方法及装置
CN109637610A (zh) * 2018-11-19 2019-04-16 深圳市理邦精密仪器股份有限公司 心电报告的配置方法、终端设备及介质
CN109350050B (zh) * 2018-11-20 2021-08-17 锐捷网络股份有限公司 心跳间期检测方法及装置
CN111195124A (zh) * 2018-11-20 2020-05-26 上海翰临电子科技有限公司 Ecg信号噪声处理方法及其装置、设备和存储介质
CN109620214B (zh) * 2018-12-07 2020-09-25 上海数创医疗科技有限公司 基于人工智能技术的心电信号st段自动判断方法及装置
CN111317463B (zh) * 2018-12-13 2022-12-06 深圳市理邦精密仪器股份有限公司 心搏的插入方法、心搏自动插入装置及存储介质
CN109589110B (zh) * 2018-12-28 2021-09-21 上海联影智能医疗科技有限公司 心电导联检测方法、装置、设备和存储介质
CN109887595A (zh) * 2019-01-16 2019-06-14 成都蓝景信息技术有限公司 基于深度学习技术的心搏异常识别算法
CN109875521A (zh) * 2019-04-18 2019-06-14 厦门纳龙科技有限公司 一种心电图数据分析以及系统
CN110179456B (zh) * 2019-05-23 2021-11-02 中国航天员科研训练中心 心电噪声识别模型训练及心电噪声检测方法、装置
US10709347B1 (en) 2019-06-10 2020-07-14 Vektor Medical, Inc. Heart graphic display system
JP6865329B1 (ja) * 2019-07-29 2021-04-28 株式会社カルディオインテリジェンス 表示装置、表示方法及びプログラム
CN110693483A (zh) * 2019-09-02 2020-01-17 乐普智芯(天津)医疗器械有限公司 一种动态心电图自动分析的方法
CN110693485A (zh) * 2019-09-18 2020-01-17 乐普智芯(天津)医疗器械有限公司 一种可穿戴式的心电信息监测系统及方法
CN110664394A (zh) * 2019-09-18 2020-01-10 乐普智芯(天津)医疗器械有限公司 一种实时心电监测系统和监测方法
CN110916646A (zh) * 2019-12-10 2020-03-27 电子科技大学 一种基于faster-rcnn的qrs波检测方法
CN110916645A (zh) * 2019-12-10 2020-03-27 电子科技大学 一种结合小波变换和图像分割网络的qrs波识别方法
CN111053551B (zh) * 2019-12-27 2021-09-03 深圳邦健生物医疗设备股份有限公司 Rr间期心电数据分布显示方法、装置、计算机设备和介质
CN113116357A (zh) * 2019-12-30 2021-07-16 华为技术有限公司 一种数据采集方法、终端设备及存储介质
CN111310572B (zh) * 2020-01-17 2023-05-05 上海乐普云智科技股份有限公司 利用心搏时间序列生成心搏标签序列的处理方法和装置
CN111242098B (zh) * 2020-02-27 2022-04-22 西安交通大学 一种结合特征提取和inception网络的心电数据分类方法及系统
CN111419220A (zh) * 2020-03-27 2020-07-17 山东省人工智能研究院 一种基于共振稀疏分解的十二导联心电信号房颤检测方法
CN111612042B (zh) * 2020-04-24 2023-09-08 广州视源电子科技股份有限公司 左前分支传导阻滞的检测装置、存储介质及终端
CN113712566B (zh) * 2020-05-12 2024-02-06 深圳市科瑞康实业有限公司 一种生成心搏间期差值数据序列的方法和装置
CN113712563A (zh) * 2020-05-12 2021-11-30 深圳市理邦精密仪器股份有限公司 动态心电图分析方法、心电监测设备以及存储介质
CN113712525A (zh) * 2020-05-21 2021-11-30 深圳市理邦精密仪器股份有限公司 一种生理参数处理方法、装置及医疗设备
CN113768511B (zh) * 2020-06-04 2023-09-22 深圳市理邦精密仪器股份有限公司 生理参数检测方法及电子设备
CN111739631A (zh) * 2020-06-19 2020-10-02 首都医科大学附属北京潞河医院 基于心电图的处理方法及处理系统
CN111759303B (zh) * 2020-07-21 2023-08-04 迪姆软件(北京)有限公司 数据处理方法及装置
CN111956201B (zh) * 2020-07-22 2022-09-06 上海数创医疗科技有限公司 基于卷积神经网络的心拍类型识别方法和装置
CN114224355A (zh) * 2020-09-07 2022-03-25 中国移动通信有限公司研究院 心电信号分类训练方法、分类方法、装置及存储介质
CN112244860B (zh) * 2020-09-27 2023-06-20 山东平伟医疗技术有限公司 一种动态心电图数据分析方法
CN112244855B (zh) * 2020-11-13 2022-08-26 上海乐普云智科技股份有限公司 一种面向心电数据的数据处理系统
CN112463334B (zh) * 2020-12-04 2023-08-18 苏州浪潮智能科技有限公司 一种训练任务排队原因分析方法、系统、设备以及介质
TWI758039B (zh) * 2020-12-29 2022-03-11 財團法人國家衛生研究院 選擇心電圖的特徵的電子裝置和方法
CN113017613B (zh) * 2021-03-03 2022-05-06 四川大学华西医院 基于人工智能的心冲击波信号处理方法及计算机设备
CN112949525B (zh) * 2021-03-12 2022-07-12 咸宁职业技术学院 一种基于向量心电的心拍分类方法及装置
CN113080996B (zh) * 2021-04-08 2022-11-18 大同千烯科技有限公司 一种基于目标检测的心电图分析方法及装置
KR102441309B1 (ko) * 2021-05-17 2022-09-08 주식회사 웰리시스 심전도 해석 매칭지원 서비스 시스템
CN113598788B (zh) * 2021-06-24 2024-02-13 上海数创医疗器械有限公司 一种心跳识别预判断模型
CN113349753A (zh) * 2021-07-19 2021-09-07 成都芯跳医疗科技有限责任公司 一种基于便携式动态心电监护仪的心律失常检测方法
CN113768514B (zh) * 2021-08-09 2024-03-22 西安理工大学 基于卷积神经网络与门控循环单元的心律失常分类方法
US20230038493A1 (en) 2021-08-09 2023-02-09 Vektor Medical, Inc. Tissue state graphic display system
US20230112597A1 (en) 2021-10-11 2023-04-13 Biosense Webster (Israel) Ltd. Suppressing interference in electrocardiogram signals using a trained neural network
CN113892955B (zh) * 2021-10-18 2024-03-29 上海乐普云智科技股份有限公司 基于人工智能模型定位qrs波并进行分类的方法和装置
CN113974644B (zh) * 2021-11-26 2023-03-24 苏州大学 一种房颤信号生成方法及系统
CN113951893B (zh) * 2021-12-02 2023-02-03 清华大学 一种结合深度学习和电生理知识的多导联心电信号特征点提取方法
CN114176519A (zh) * 2021-12-09 2022-03-15 复旦大学 一种非接触式心电信号质量分类方法
CN114668401B (zh) * 2022-03-11 2023-04-18 肇庆星网医疗科技有限公司 一种ai心电训练数据标注方法、装置、电子设备及介质
CN114451879B (zh) * 2022-03-15 2024-04-09 武汉中旗生物医疗电子有限公司 一种智能型心率变异性分析系统
CN115486856A (zh) * 2022-09-20 2022-12-20 江宁 消除信号相移的表面电生理信号处理方法和系统
CN115423379B (zh) * 2022-10-21 2023-03-24 北京汇达城数科技发展有限公司 基于溯源信息的置信度评估方法、系统、终端及存储介质
TWI833462B (zh) * 2022-11-22 2024-02-21 國立勤益科技大學 一種動態心電圖st改變事件自動偵測方法
CN116570295B (zh) * 2023-07-14 2024-04-30 浙江好络维医疗技术有限公司 一种心电图低电压t波终点定位方法
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
CN117462141B (zh) * 2023-12-25 2024-03-26 深圳市先健心康医疗电子有限公司 心电信号检测方法、装置、计算机设备和存储介质
CN117547283B (zh) * 2024-01-11 2024-05-14 湘江实验室 心电信号数据分类方法、装置、计算机设备及介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6370423B1 (en) * 1998-10-05 2002-04-09 Juan R. Guerrero Method for analysis of biological voltage signals
EP2033574A1 (en) * 2005-03-31 2009-03-11 Q & S S.P.A Apparatus for detecting cardiac signals
CN102138789A (zh) * 2011-01-24 2011-08-03 无锡微感科技有限公司 一种动态心电和运动记录与分析系统
CN106108889A (zh) * 2016-07-20 2016-11-16 杨平 基于深度学习算法的心电图分类方法
CN107358196A (zh) * 2017-07-12 2017-11-17 北京卫嘉高科信息技术有限公司 一种心搏类型的分类方法、装置及心电仪

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8632465B1 (en) 2009-11-03 2014-01-21 Vivaquant Llc Physiological signal denoising
EP2744402B1 (en) 2011-10-12 2020-12-09 Koninklijke Philips N.V. System and method for determining the q and j points of an electrocardiogram
EP3878364A1 (en) * 2015-10-27 2021-09-15 Cardiologs Technologies An automatic method to delineate or categorize an electrocardiogram

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6370423B1 (en) * 1998-10-05 2002-04-09 Juan R. Guerrero Method for analysis of biological voltage signals
EP2033574A1 (en) * 2005-03-31 2009-03-11 Q & S S.P.A Apparatus for detecting cardiac signals
CN102138789A (zh) * 2011-01-24 2011-08-03 无锡微感科技有限公司 一种动态心电和运动记录与分析系统
CN106108889A (zh) * 2016-07-20 2016-11-16 杨平 基于深度学习算法的心电图分类方法
CN107358196A (zh) * 2017-07-12 2017-11-17 北京卫嘉高科信息技术有限公司 一种心搏类型的分类方法、装置及心电仪

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3692904A4 *

Cited By (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111177280A (zh) * 2019-12-26 2020-05-19 北京亚信数据有限公司 一种数据权威性评价方法和装置
CN111493858A (zh) * 2020-03-16 2020-08-07 浙江好络维医疗技术有限公司 一种基于聚类分析的单导联特异主波识别定位方法
CN111493858B (zh) * 2020-03-16 2022-12-09 浙江好络维医疗技术有限公司 一种基于聚类分析的单导联特异主波识别定位方法
CN111904410A (zh) * 2020-06-19 2020-11-10 陕西省医疗器械质量监督检验院 一种动态心电图准确性的检测系统及检测方法
CN111904410B (zh) * 2020-06-19 2023-11-03 陕西省医疗器械质量监督检验院 一种动态心电图准确性的检测系统及检测方法
US11568991B1 (en) 2020-07-23 2023-01-31 Heart Input Output, Inc. Medical diagnostic tool with neural model trained through machine learning for predicting coronary disease from ECG signals
CN113080995A (zh) * 2021-03-17 2021-07-09 深圳邦健生物医疗设备股份有限公司 心动过速属性的识别方法、装置、设备和介质
CN113080995B (zh) * 2021-03-17 2024-02-02 深圳邦健生物医疗设备股份有限公司 心动过速属性的识别方法、装置、设备和介质
CN113796873A (zh) * 2021-10-26 2021-12-17 山东建筑大学 穿戴式动态心电信号分类方法及系统
CN113796873B (zh) * 2021-10-26 2024-05-03 山东建筑大学 穿戴式动态心电信号分类方法及系统
CN114219028A (zh) * 2021-12-15 2022-03-22 东软集团股份有限公司 心电st段抬高识别模型的训练、st段抬高识别方法及装置
CN114219028B (zh) * 2021-12-15 2024-04-12 东软集团股份有限公司 心电st段抬高识别模型的训练、st段抬高识别方法及装置

Also Published As

Publication number Publication date
US11234629B2 (en) 2022-02-01
CN107951485B (zh) 2019-06-11
CN107951485A (zh) 2018-04-24
EP3692904A1 (en) 2020-08-12
US20200260980A1 (en) 2020-08-20
EP3692904A4 (en) 2021-07-14

Similar Documents

Publication Publication Date Title
WO2019100565A1 (zh) 基于人工智能自学习的动态心电图分析方法和装置
US11529103B2 (en) Artificial intelligence self-learning-based automatic electrocardiography analysis method and apparatus
US11344243B2 (en) Artificial intelligence self-learning-based static electrocardiography analysis method and apparatus
US11517212B2 (en) Electrocardiogram information dynamic monitoring method and dynamic monitoring system
US11350868B2 (en) Electrocardiogram information processing method and electrocardiogram workstation system
CN109411042B (zh) 心电信息处理方法和心电工作站
US11324455B2 (en) Artificial intelligence-based interference recognition method for electrocardiogram
CN106815570B (zh) 一种基于动态模式识别的心电信号st-t段识别方法
WO2022135449A1 (zh) 癫痫患者发作间期痫样电活动检测装置和方法
Zhang et al. Detection of atrial fibrillation from variable-duration ECG signal based on time-adaptive densely network and feature enhancement strategy
Wang et al. Multiscale residual network based on channel spatial attention mechanism for multilabel ECG classification
Zhao et al. PVC recognition for wearable ECGs using modified frequency slice wavelet transform and convolutional neural network
CN110960207A (zh) 一种基于树模型的房颤检测方法、装置、设备及存储介质
Jokić et al. An efficient ECG modeling for heartbeat classification
Zhang et al. Evaluation of single-lead ECG signal quality with different states of motion
Al-Ani ECG Signal Recognition Based on Lookup Table and Neural Networks

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: 18881678

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2018881678

Country of ref document: EP

Effective date: 20200504

NENP Non-entry into the national phase

Ref country code: DE