TWI701681B - Prediction model of atrial fibrillation and prediction system thereof - Google Patents

Abstract

A prediction system of atrial fibrillation is provided. The prediction system of atrial fibrillation includes an electrocardiogram obtaining unit and a non-transitory machine readable medium. The non-transitory machine readable medium storing a program which, when executed by a processing unit, obtains a prediction result. The program includes a reference database obtaining module, a reference feature selecting module, a training module, a target feature selecting module and a comparing module. Therefore, the prediction system of atrial fibrillation can be used to predict an incidence of stroke in a subject.

Description

Atrial fibrillation prediction model and its prediction system

The present invention relates to a medical information analysis model and system, especially an atrial fibrillation prediction model and atrial fibrillation prediction system.

Atrial fibrillation (atrial fibrillation) is a disorder in which the heartbeat is irregular and often too fast due to the abnormal function of the rhythm signal generated in the heart. The heartbeat can reach 350 beats per minute. Atrial fibrillation is the most common abnormal heart rhythm. An average of 1 in 100 people in the entire population suffers from atrial fibrillation. As the age increases, the proportion of atrial fibrillation is higher. Four out of every 100 people over the age of 60 suffer from atrial fibrillation, and one in ten people over the age of 80 suffers from atrial fibrillation. In 2010, an estimated 33.5 million people worldwide suffered from atrial fibrillation. In addition, there may be many potential patients who have not been diagnosed because they have no symptoms. It is estimated that the number of patients with atrial fibrillation in Asia will reach 72 million in 2050.

Patients with atrial fibrillation have 5 times the risk of developing thrombotic infarction diseases, including stroke, pulmonary embolism, and peripheral vascular embolism. Past studies have also shown that among patients suffering from atrial fibrillation, patients with paroxysmal atrial fibrillation have a lower incidence of stroke than patients with persistent atrial fibrillation. Patients with paroxysmal atrial fibrillation and persistent atrial fibrillation have a poorer prognosis after suffering from a stroke and also have a higher risk of second-degree stroke. Therefore, the pattern of atrial fibrillation and stroke are highly correlated. Therefore, in 2050, it is estimated that the number of stroke patients in Asia due to atrial fibrillation will reach 2.9 million. Patients with persistent atrial fibrillation have a higher incidence of stroke than patients with paroxysmal atrial fibrillation, and the prognosis after stroke is also worse. Clinically, CHA2DS2-VASc score is mainly used to evaluate the risk of stroke in patients with atrial fibrillation. The items evaluated by CHA2DS2IVASc score include age, gender, and comorbidities including infarct disease, hypertension, congestive heart failure, diabetes, and vascular disease. The CHA2DS2-VASc score increased, and the risk of vascular embolism also gradually increased, but so far, there has been no research on the correlation between the ECG characteristics of patients with atrial fibrillation and stroke.

The electrocardiogram provides information on atrial fibrillation, such as the frequency and pattern of atrial fibrillation during the day, but its data is huge and cannot be analyzed manually. Therefore, the conventional technology lacks the ability to efficiently analyze a large amount of data and can be applied to clinical assistant physicians to further judge the patient Whether the subject may be a patient with atrial fibrillation and stroke, in order to improve the accuracy of detection.

In view of this, one of the objectives of the present invention is to provide an atrial fibrillation prediction model and atrial fibrillation prediction system, which can objectively and accurately determine whether a subject has atrial fibrillation, and can further predict the probability of stroke. To assist doctors in clinical judgment.

One aspect of the present invention is to provide an atrial fibrillation prediction model, which includes the following establishment steps: obtaining a reference database, performing a feature selection step, and performing a training step. The reference database includes a plurality of reference twelve-lead ECG signal series. The feature selection step is to select at least one feature value based on a reference database, and the feature value includes using a calculation unit to calculate the ECG signal with the largest change in curvature obtained by referring to the peak-to-peak time difference in the twelve-lead ECG signal sequence Image interval. In the training step, a Long Short Term Memory (LSTM) is used to store a real-time value of an ECG signal, and the correlation between the characteristic value and the real-time value of the ECG signal is calculated. When the correlation exceeds A first preset threshold value updates the long-short memory unit, and obtains atrial fibrillation prediction model when the training reaches convergence, thereby obtaining a preset result.

According to the aforementioned atrial fibrillation prediction model, the long-short-term memory unit may be a Bi-directional Long Short Term Memory (Bi-directional LSTM).

According to the aforementioned atrial fibrillation prediction model, the long-short memory unit may further include a Forget Gate, an Input Gate, and an Output Gate. The forgetting gate filters the real-time value of the ECG signal whose curvature changes too much to obtain an input value. The input gate system inputs the input value, and uses the Sigmoid function to calculate the correlation. The output gate system calculates the correlation using the Sigmoid function to obtain an output value, and when the output value exceeds a second preset threshold value, the output value is added to the long and short memory unit.

According to the aforementioned atrial fibrillation prediction model, the forgetting gate, the input gate and the output gate can be two-way serially connected.

According to the aforementioned atrial fibrillation prediction model, the first preset threshold and the second preset threshold can be determined by the tanh function.

Another aspect of the present invention is to provide an atrial fibrillation prediction system that includes an electrocardiogram capture unit and a non-transient machine-readable medium. The electrocardiogram capture unit is used to obtain a target twelve-lead electrocardiogram signal sequence. The non-transitory machine-readable medium is connected to the electrocardiogram capture unit by at least one signal, and the non-transitory machine-readable medium stores a program for obtaining a prediction result when the program is executed by at least one processing unit The program includes: a reference database acquisition module, a reference feature selection module, a training module, a target feature selection module, and a comparison module. The reference database acquisition module is used to acquire a reference database, and the reference database includes a plurality of reference twelve-lead ECG signal series. The reference feature selection module is used to select at least one reference feature value based on a reference database, and the reference feature value includes an ECG obtained by using a calculation unit to calculate a peak-to-peak time difference in a twelve-lead ECG signal sequence The image interval where the curvature changes the most. The training module includes a Long Short Term Memory (LSTM). The long-short memory unit is used to store a real-time value of an ECG signal, and calculate the correlation between the characteristic value and the real-time value of the ECG signal, and update the long-short memory unit when the correlation exceeds a first preset threshold , When the training reaches convergence, the atrial fibrillation prediction model is obtained. The target feature selection module is used to analyze the target twelve-lead ECG signal sequence to obtain a target feature value, and the target feature value includes calculating the target twelve-lead ECG signal using another calculation unit The image interval in which the curvature of the target ECG signal changes the most is obtained from the peak-to-peak time difference in the number sequence. The comparison module is used for analyzing and comparing the target feature value and the reference feature value with the atrial fibrillation prediction model, thereby obtaining a preset result.

According to the aforementioned atrial fibrillation prediction system, the long-short-term memory unit may be a Bi-directional Long Short Term Memory (Bi-directional LSTM).

According to the aforementioned atrial fibrillation prediction system, the long-short memory unit may further include a Forget Gate, an Input Gate, and an Output Gate. The forgetting gate is used to filter the real-time value of the ECG signal with excessive curvature change to obtain an input value. The input gate is used for inputting the input value, and calculating the correlation using a Sigmoid function. The output gate is used for calculating the correlation using a Sigmoid function to obtain an output value, and when the output value exceeds a second preset threshold value, the output value is added to the long and short memory unit.

According to the aforementioned atrial fibrillation prediction system, the forgetting gate, the input gate and the output gate can be bidirectionally connected in series.

According to the aforementioned atrial fibrillation prediction system, the first preset threshold and the second preset threshold can be determined by a tanh function.

The foregoing summary of the invention aims to provide a simplified summary of the disclosure so that readers have a basic understanding of the disclosure. This summary of the present invention is not a complete summary of the present disclosure, and its intention is not to point out important/key elements of the embodiments of the present invention or to define the scope of the present invention.

100‧‧‧Building steps of atrial fibrillation prediction model

110, 120, 130‧‧‧step

200‧‧‧Atrial Fibrillation Prediction System

300‧‧‧ECG capture unit

400‧‧‧Non-transitory machine-readable media

410‧‧‧Refer to the database to obtain the module

420‧‧‧Reference feature selection module

421, 441‧‧‧computing unit

430‧‧‧Training Module

432、600‧‧‧Long and short memory unit

440‧‧‧Target feature selection module

450‧‧‧Comparison Module

610‧‧‧Input layer

620‧‧‧1st level long and short memory unit

630‧‧‧2nd level long and short memory unit

640‧‧‧Level 3 Long Short Memory Unit

650‧‧‧Level 4 Long Short Memory Unit

660‧‧‧Maximum pooling layer

670‧‧‧All company class

In order to make the above and other objects, features, advantages and embodiments of the present invention more comprehensible, the description of the accompanying drawings is as follows: Figure 1 shows the establishment of an atrial fibrillation prediction model according to an embodiment of the present invention Step flowchart; Figure 2 shows a block diagram of an atrial fibrillation prediction system according to another embodiment of the present invention; Figure 3 shows a schematic diagram of the data labeling platform of the reference database of the atrial fibrillation prediction model of the present invention; Figure 4 is a schematic diagram showing the structure of the long and short memory unit of the atrial fibrillation prediction model of the present invention; Figure 5 is a diagram showing the structure of the long and short memory unit of the atrial fibrillation prediction model of the present invention; and Figure 6 is the atrial fibrillation prediction of the present invention The system is used to predict the receiver operating characteristic curve of the fan rate in the brain of the subject.

The various embodiments of the present invention will be discussed in more detail below. However, this embodiment can be an application of various inventive concepts and can be implemented in various specific ranges. The specific implementation is for illustrative purposes only, and is not limited to the scope of disclosure.

Please refer to Fig. 1, which shows a flowchart of steps 100 for establishing an atrial fibrillation prediction model according to an embodiment of the present invention. The step 100 of establishing the atrial fibrillation prediction model of the present invention includes step 110, step 120 and In step 130, the established atrial fibrillation prediction model can be used to predict the incidence of stroke in the subject.

Step 110 is to obtain a reference database, which includes a plurality of reference twelve-lead ECG signal series. Furthermore, referring to the twelve-lead ECG signal sequence, preliminary classification can be carried out first, which is divided into an abnormal data and a non-abnormal data and marked, so as to divide the reference database into two categories.

Step 120 is to perform a feature selection step, which is to select at least one feature value based on the reference database, the feature value including the ECG signal curvature obtained by referring to the peak-to-peak time difference in the twelve-lead ECG signal sequence calculated by a calculation unit The most changed image interval.

Step 130 is a training step, which uses a Long Short Term Memory (LSTM) to store an ECG signal real-time value, and calculates the correlation between the characteristic value and the ECG signal real-time value, when the correlation When the sex exceeds a first preset threshold, the long-short memory unit is updated, and when the training reaches convergence, the atrial fibrillation prediction model is obtained, thereby obtaining a preset result. The long and short memory unit may further include a Forget Gate, an Input Gate, and an Output Gate. The forgetting gate filters the real-time value of the ECG signal whose curvature changes too much to obtain an input value. The input gate system inputs the input value, and uses the Sigmoid function to calculate the correlation. The output gate system calculates the correlation using the Sigmoid function to obtain an output value, and when the output value exceeds a second preset threshold value, the output value is added to the long and short memory unit. Preferably, the forgotten gate, the input gate and the output gate may be two-way serially connected, and the first preset threshold and the second preset threshold may be determined by a tanh function. among them The long and short term memory unit may be a Bi-directional Long Short Term Memory (Bi-directional LSTM).

The first preset threshold and the second preset threshold are determined by the tanh function, and the output value of the tanh function is between -1 and 1, which is to throw a large number of twelve-lead ECG signal series into The preset value calculated by the mathematical formula of machine learning. During the training of the atrial fibrillation prediction model, when the correlation between the characteristic value and the real-time value of the ECG signal exceeds the first preset threshold, the long and short memory units are updated to achieve convergence to obtain the atrial fibrillation prediction model. Among them, when the correlation is closer to -1, the probability that the subject does not have atrial fibrillation is higher, and when the correlation is close to 1, the probability that the subject has atrial fibrillation is higher. When the atrial fibrillation prediction model is used to predict whether the subject has atrial fibrillation, the forgetting gate will first filter the real-time value of the ECG signal to obtain an input value, and by inputting the correlation calculated by the Sigmoid function by the input gate, the output gate will be The correlation uses a Sigmoid function to calculate the output value, and when the output value exceeds a second preset threshold, the output value is added to the long-short memory unit. Among them, when the output value is closer to -1, it means that the subject has a higher probability of not having atrial fibrillation. Conversely, when the output value is closer to 1, it means that the subject is more likely to have atrial fibrillation.

Please refer to FIG. 2, which shows a block diagram of an atrial fibrillation prediction system 200 according to another embodiment of the present invention. The atrial fibrillation prediction system 200 of the present invention includes an electrocardiogram capture unit 300 and a non-transitory machine-readable medium 400. The atrial fibrillation prediction system 200 can be used to predict the occurrence probability of a subject's stroke.

The ECG acquisition unit 300 is used to obtain the target twelve-lead ECG signal sequence of the subject and obtain the reference twelve-lead ECG signal sequence. The electrocardiogram capturing unit 300 may be an electrocardiograph. Preferably, the electrocardiogram capture unit 300 can be a twelve-lead electrocardiograph, which includes 10 electrode patches, and more than 2 electrode patches are placed on the limbs, which are formed in pairs for measurement, and record 12 groups of body surfaces Lead potential changes, and draw 12 sets of lead signals on the ECG paper to obtain the 12-lead ECG signal series.

The non-transitory machine-readable medium 400 is connected to the electrocardiogram capture unit 300 by at least one signal, and the non-transitory machine-readable medium 400 stores a program, wherein when the program is executed by at least one processing unit, the program is used A prediction result is obtained, and the prediction result is the probability of occurrence of stroke of the subject. The program includes a reference database acquisition module 410, a reference feature selection module 420, a training module 430, a target feature selection module 440, and a comparison module 450.

The reference database obtaining module 410 is used to obtain a reference database, which includes a plurality of reference twelve-lead ECG signal series. Furthermore, referring to the twelve-lead ECG signal sequence, preliminary classification can be carried out first, which is divided into an abnormal data and a non-abnormal data and marked, so as to divide the reference database into two categories.

The reference feature selection module 420 is used for selecting at least one reference feature value based on a reference database, the reference feature value including an ECG obtained by using a calculation unit 421 to calculate a peak-to-peak time difference in a twelve-lead ECG signal sequence The image interval where the curvature changes the most.

The training module 430 includes a long and short memory unit 432. The long and short memory unit 432 is used to store a real-time value of an ECG signal, and calculate the correlation between the characteristic value and the real-time value of the ECG signal, and update the long and short memory unit when the correlation exceeds a first preset threshold. 432. Obtain the atrial fibrillation prediction model when the training reaches convergence. The long and short memory unit 432 may further include a forget gate, an input gate, and an output gate. The forgetting gate system filters the real-time value of the ECG signal with excessive curvature change to obtain the input value. The input gate system inputs the input value, and uses the Sigmoid function to calculate the correlation. The output gate system calculates the correlation using a Sigmoid function to obtain an output value, and when the output value exceeds a second preset threshold value, the output value is added to the long and short memory unit 432. Preferably, the forgotten gate, the input gate and the output gate may be two-way serially connected, and the first preset threshold and the second preset threshold may be determined by a tanh function. In addition, the long and short memory unit 432 may be a bidirectional long and short memory unit.

The target characteristic selection module 440 is used to analyze the target twelve-lead ECG signal series to obtain a target characteristic value. The target characteristic value includes using another calculation unit 441 to calculate the target twelve-lead ECG signal series An image interval in which the curvature of a target ECG signal changes the most, obtained by a peak-to-peak time difference.

The comparison module 450 is used for analyzing and comparing the target feature value and the reference feature value with the atrial fibrillation prediction model, thereby obtaining a preset result. The preset result is the probability of the subject having a stroke within 3-6 months, and the probability value is 0%-100%, which is used as an auxiliary reference for the doctor's diagnosis.

<Test Example>

1. Reference database

The reference database used in the present invention is China Medical University and affiliated hospitals retrospectively collecting in-hospital de-linked clinical content of examinees from 2009/01/01 to 2018/12/31, which is approved by China Medical University and affiliated hospitals. The clinical trial plan approved by the Hospital Research Ethics Committee (China Medical University & Hospital Research Ethics Committee) is numbered: CMUH107-REC2-134 (AR-1). The data is collected through the GE Healthcare MUSE system with keyword parameter search method to collect the patient's electrocardiography (ECG/EKG) waveform data including atrial fibrillation and myocardial infarction, including a series of twelve-lead ECG signals. The original data is Extensible Markup Language (XML) format. There is no special restriction on the gender of the subject to whom the images are collected, and there is no special age range. The reference subjects included 5,000 reference subjects without atrial fibrillation, and 10,012 reference subjects with atrial fibrillation, for a total of 15,012 reference subjects. The above data is the actual "number of data" used. The possibility of "examination of the same patient" at different time points/dates is not excluded.

2. Used to determine the brain fan rate of subjects

In this experimental example, first establish an optimized atrial fibrillation prediction model. First obtain the reference database. The reference database contains a plurality of reference twelve-lead ECG signal series, and the reference twelve-lead ECG signal series will be preliminarily classified into one abnormal data and one non-abnormal data and mark . Please refer to Figure 3, which shows a schematic diagram of the data labeling platform of the reference database of the atrial fibrillation prediction model of the present invention, in order to enable the subsequent establishment of atrial fibrillation The predictive model correctly learns the disease problem corresponding to the twelve-lead ECG signal sequence. First, a data marking platform will be established without providing any patient-related personal information and restricting specific connections. The physician will use this platform to refer to the database Perform a variety of markers as a reference basis for the learning of atrial fibrillation prediction model.

The reference feature selection module is then used to select at least one feature value according to the reference database, the feature value including the ECG signal with the largest change in curvature obtained by calculating the peak-to-peak value time difference in the twelve-lead ECG signal sequence by the calculation unit Image interval.

Then proceed to the training step, using the two-way long and short memory network architecture for neural network learning, and let the machine learn time series signals from different neural directions. When optimizing parameters, the traditional Recurent Neural Network (RNN) uses gradient descent (Gradient Descent) to optimize the method of updating parameters, and the method of seeking parameter changes is Backward Propagation (Backward). Propagation) algorithm to achieve, but this algorithm will cause gradient explosion (Gradient Explosion) and gradient vanish (Gradient Vanish) due to the parameters taken. The atrial fibrillation prediction model of the present invention adds a forgetting gate during training, so that if a gradient explosion is encountered during the backpropagation algorithm, the forgetting gate can be used to block it, but the input value is mathematically calculated. Approaching 0 (that is, the value after a dozen digits below the decimal point) will cause the computer to ignore the gradient disappearance. The auxiliary input gate (Pass Gate) can be used to transmit the message to avoid the gradient disappear.

In detail, the training step is to use the long-short memory unit to store the real-time value of the ECG signal, and calculate the correlation between the characteristic value and the real-time value of the ECG signal. When the correlation exceeds a first preset threshold, all the values are updated. The long and short memory unit. Please refer to FIG. 4, which is a schematic diagram of the structure of the long and short memory unit of the atrial fibrillation prediction model of the present invention. The long and short memory unit uses a memory branch that is updated over time to enhance the current decision result, and the long and short memory unit includes forgetting gate, input gate and output gate to determine whether the memory is updated, and forgetting gate, input gate and The output gate is two-way serial connection. The forgetting gate is used to filter the real-time value of the ECG signal with excessive curvature change to obtain the input value. In detail, the forget gate uses the calculated zf (f means forget) as the forget gate to control which ct-1 in the previous state needs to be left or forgotten, usually a Sigmoid function. The input gate is used to input the input value and calculate the correlation using the Sigmoid function. In detail, the input gate determines whether the current input (Input) and the newly generated memory cell (Memory Cell Candidate) are added to the long term memory (Long Term Memory). The input gate also uses the Sigmoid function to indicate whether to add. Specifically, it is to select and memorize the input xt. Record what is important and write less what is not. The current input content is represented by the previously calculated z. The selected gating signal is controlled by zi (i stands for information). The output gate is used to calculate the correlation using the Sigmoid function to obtain the output value, and when the output value exceeds the second preset threshold value, the output value is added to the long and short memory unit. In detail, the output gate determines which output will be regarded as the current state. It is mainly controlled by zo. And also change the co obtained in the previous stage through a tanh activation function. Forget Please refer to formula (I), formula (II) and formula (III) for detailed calculation methods of gate, input gate and output gate.

c ^t = z ^f ⊙ c ^{t -1} + z ⁱ ⊙ z ………………………… formula (I); h ^t = z ^o ⊙ tanh ( c ^t )…………………… …Formula (II); y ^t = σ ( W'h ^t )……………………………………Formula (III). The first preset threshold and the second preset threshold are determined by the tanh function. The output value of the tanh function is between -1 and 1, which is to throw a large number of twelve-lead ECG signal series into the machine The preset value calculated by the learned mathematical formula. When the training reaches convergence, the atrial fibrillation prediction model is obtained, thereby obtaining a preset result, which is the subject's brain fan rate.

During the training of the atrial fibrillation prediction model, when the correlation between the characteristic value and the real-time value of the ECG signal exceeds the first preset threshold, the long and short memory unit is updated to achieve convergence to obtain the atrial fibrillation prediction model. When it approaches -1, it means that the subject has a higher probability of not having atrial fibrillation. Conversely, when the correlation is closer to 1, it means that the subject is more likely to have atrial fibrillation. When the atrial fibrillation prediction model is used to predict and predict atrial fibrillation, the forgetting gate will first filter the real-time value of the ECG signal to obtain the input value and input the correlation calculated by the Sigmoid function through the input gate, and the output gate will use the correlation The sigmoid function calculates the output value. When the output value exceeds the second preset threshold, the output value is added to the long and short memory unit. When the output value approaches -1, it means that the subject does not have atrial fibrillation The higher the probability of, the higher the probability that the subject has atrial fibrillation when it approaches 1, on the contrary.

In addition, please refer to FIG. 5 again, which shows the structure diagram of the long and short memory unit 600 of the atrial fibrillation prediction model of the present invention. The long and short memory unit 600 of the atrial fibrillation prediction model of the present invention is a 128*4 long and short memory unit in the fourth-level long and short memory group, which includes an input layer 610, a first-level long and short memory unit 620, a second-level long and short memory unit 630, The third-level long and short memory unit 640, the fourth-level long and short memory unit 650, the maximum pooling layer 660, and the fully connected layer 670. The first-level long and short memory unit 620, the second-level long and short memory unit 630, the third-level long and short memory unit 640, and the fourth-level long and short memory unit 650 each have 128 long and short memory units. The first-level long-short memory unit 620 can handle low-complexity feature values, the second-level long-short memory unit 630 can handle slightly more complex feature values, and the third-level long-short memory unit 640 can handle more complex features. Value, the fourth-level long-short memory unit 650 can handle the most complex feature value. The maximum pooling layer will be collected according to the characteristics of fourth-order long and short memory learning, and the fully connected layer (Sigmod function/tanh function) will output the final result according to the part of the feature learning.

In this test example, the atrial fibrillation prediction system including the established atrial fibrillation prediction model is further used to predict the subject's stroke. The steps are as follows: Provide the atrial fibrillation prediction model established above. Provide the subject's target 12-lead ECG signal sequence. The target characteristic value obtained by analyzing the target 12-lead ECG signal sequence with the target characteristic selection module. Finally, a comparison module is used to analyze and compare the target feature value and the reference feature value with the atrial fibrillation prediction model, thereby obtaining a preset result to predict the subject's brain fan rate.

Please refer to Figure 6, which is a receiver operating characteristic curve (ROC) diagram used by the atrial fibrillation prediction system of the present invention to predict the fan rate in the brain of a subject. The results show that when the atrial fibrillation prediction model of the present invention predicts the subject's brain fan rate, the area under the curve (AUC) of the test is 0.996, and the ROC value is 99.6%. It is shown that the atrial fibrillation prediction model and the atrial fibrillation prediction system of the present invention can accurately predict the brain fan rate of the subject with the twelve-lead ECG signal sequence.

In this way, the present invention provides an atrial fibrillation prediction model and an atrial fibrillation prediction system. The neural network learning is performed through the structure of the long and short memory network, and the neural direction of different items allows the machine to learn time series signals, which can be objective and Accurately use the 12-lead ECG signal sequence to determine whether the subject has atrial fibrillation, and further predict the probability of stroke, and provide a second opinion to the specialist to assist the doctor in clinical judgment. From the original image input to the interpretation result, it only takes 0.1-1 seconds on average to complete, and the accuracy rate can be as high as 0.996. Based on the atrial fibrillation prediction model and atrial fibrillation prediction system of the present invention, the twelve-lead ECG signal sequence of a case can be used for automated and rapid data analysis, assisting medical staff in interpretation and immediate diagnosis, and improving early stroke The discovery rate allows the physician to plan the patient’s follow-up course of treatment.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Anyone familiar with the art can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection of the present invention The scope shall be subject to those defined in the attached patent scope.

200‧‧‧Atrial Fibrillation Prediction System

300‧‧‧ECG capture unit

400‧‧‧Non-transitory machine-readable media

410‧‧‧Refer to the database to obtain the module

420‧‧‧Reference feature selection module

421, 441‧‧‧computing unit

430‧‧‧Training Module

432‧‧‧Long and short memory unit

440‧‧‧Target feature selection module

450‧‧‧Comparison Module

Claims (8)

An atrial fibrillation prediction model includes the following steps: obtaining a reference database, where the reference database includes a plurality of reference twelve-lead ECG signal sequences; performing a feature selection step, which selects at least A feature value, the feature value comprising a calculation unit calculating the time difference of a peak-to-peak value in the reference twelve-lead ECG signal sequence to obtain an image interval where the curvature of the ECG signal changes the most; and performing a training step using four Long and short memory groups are trained. Each of the long and short memory groups has 128 Long Short Term Memory (LSTM) units, and each of the long and short memory units is a bi-directional Long Short Term Memory (Bi-directional Long Short Term Memory, Bi-directional). LSTM), each of the long and short memory units stores a real-time value of the ECG signal, and calculates the correlation between the characteristic value and the real-time value of the ECG signal. When the correlation exceeds a first preset threshold, the long and short memory unit is updated, When the training reaches convergence, the atrial fibrillation prediction model is obtained, thereby obtaining a preset result, wherein the training step does not include using a convolutional neural network to train the feature value. According to the atrial fibrillation prediction model described in claim 1, wherein the long and short memory unit further includes: A Forget Gate is used to filter the real-time value of the ECG signal whose curvature changes too much to obtain an input value; an Input Gate is used to input the input value and use the Sigmoid function to calculate the correlation; And an output gate. The correlation is calculated using the Sigmoid function to obtain an output value. When the output value exceeds a second preset threshold, the output value is added to the long-short memory unit. The atrial fibrillation prediction model described in item 2 of the scope of patent application, wherein the forgetting gate, the input gate and the output gate are bidirectionally connected in series. In the atrial fibrillation prediction model described in item 2 of the scope of patent application, the first preset threshold and the second preset threshold are determined by a tanh function. An atrial fibrillation prediction system, comprising: an electrocardiogram capture unit for obtaining a target twelve-lead ECG signal sequence; and a non-transient machine-readable medium connected to the electrocardiogram capture unit by at least one signal, wherein The non-transitory machine-readable medium is used to store a program that is used to obtain a prediction result when the program is executed by a processing unit, and the program includes: A reference database acquisition module is used to acquire a reference database, and the reference database includes a plurality of reference twelve-lead ECG signal series; a reference feature selection module is used to select at least A reference eigenvalue, the reference eigenvalue includes an image interval in which the curvature of the ECG signal changes the most, which is obtained by calculating a peak-to-peak time difference in the reference twelve-lead ECG signal sequence by a calculation unit; a training module includes: Four long and short memory groups, each of which has 128 Long Short Term Memory (LSTM), and each of the long and short memory units is a Bi-directional Long Short Term Memory (Bi-directional LSTM) ), each of the long and short memory units is used to store a real-time value of an ECG signal, and calculate the correlation between the characteristic value and the real-time value of the ECG signal, and update the long and short memory unit when the correlation exceeds a first preset threshold , When the training reaches convergence, the atrial fibrillation prediction model is obtained, wherein the training module does not include a convolutional neural network; a target feature selection module is used to analyze the target twelve-lead ECG signal sequence to obtain a target The characteristic value, the target characteristic value includes calculating the target twelve-lead ECG signal number using another calculation unit A target ECG signal curvature change image interval obtained by a peak-to-peak time difference in the column; and a comparison module for analyzing and comparing the target characteristic value and the reference characteristic value with the atrial fibrillation prediction model , Thereby obtaining a preset result. For example, the atrial fibrillation prediction system described in item 5 of the scope of patent application, wherein the long and short memory unit further includes: a Forget Gate for filtering the real-time value of the ECG signal with excessive curvature change to obtain an input value ; An input gate (Input Gate) to input the input value, and use the Sigmoid function to calculate the correlation; and an output gate (Output Gate) to use the Sigmoid function to calculate the correlation to obtain an output value When the output value exceeds a second preset threshold value, the output value is added to the long-short memory unit. The atrial fibrillation prediction system described in item 6 of the scope of patent application, wherein the forgetting gate, the input gate and the output gate are two-way serially connected. For the atrial fibrillation prediction system described in item 6 of the scope of patent application, the first preset threshold and the second preset threshold are determined by a tanh function.

Images