JPWO2020047750A5

JPWO2020047750A5 -

Info

Publication number: JPWO2020047750A5
Application number: JP2020568775A
Authority: JP
Publication date: 2023-06-15
Anticipated expiration: 2038-09-04

Description

本発明の実施例において、前記前処理は横方向引張処理及び／または横方向圧縮処理を含み、前記最初の心電図信号に対して前処理をすることによりトレーニング信号を形成するステップは、予め設定された比率により前記最初の心電図信号に対して横方向引張処理及び／または横方向圧縮処理をすることによりトレーニング信号を獲得するステップを含む。
In an embodiment of the present invention, said preprocessing comprises lateral tension processing and/or lateral compression processing, and the step of forming a training signal by preprocessing said initial electrocardiogram signal is preset to: obtaining a training signal by subjecting said initial electrocardiogram signal to lateral tension and/or lateral compression according to the ratio of said initial electrocardiogram signal .

（二）横方向引張処理及び／または横方向圧縮処理
具体的な実施例において、予め設定された比率によって最初の心電図信号に対して横方向引張処理及び／または横方向圧縮処理を行うことによりトレーニング信号を獲得する。
(2) Lateral tension processing and/or lateral compression processing In a specific embodiment, training by performing lateral tension processing and/or lateral compression processing on the initial electrocardiogram signal according to a preset ratio . get a signal.

トレーニング信号に対して横方向引張処理及び／または横方向圧縮処理を行う原因は、使用者の心拍数は一定の範囲内において変化し、その心拍数は使用者の現時点の健康状態によって決定されるからである。また、各使用者の心拍数が異なっているからである。一定の範囲の因子（すなわち所定の比率）で横方向引張処理／横方向圧縮処理、すなわち最初の心電図信号に対して引張処理／圧縮処理を行うことによりいろいろな心拍数に対応する心電図信号（すなわちトレーニング信号）を形成する。いろいろな心拍数に対応する心電図信号により不整脈検出モデルに対してトレーニングをする場合、不整脈検出モデルでいろいろな心拍数に対応する心電図信号を検出する効果を向上させ、不整脈検出モデルは心拍数の影響を受けない波形特徴と心拍数変異等の特徴を有効に検出することができる。
The reason for performing lateral tension processing and/or lateral compression processing on the training signal is that the user's heart rate varies within a certain range, and the heart rate is determined by the user's current health condition. It is from. Also, each user has a different heart rate. Electrocardiogram signals corresponding to different heart rates (i.e., training signal). When training an arrhythmia detection model with electrocardiogram signals corresponding to different heart rates, the arrhythmia detection model can improve the effect of detecting electrocardiogram signals corresponding to different heart rates, and the arrhythmia detection model can improve the effect of heart rate. It is possible to effectively detect features such as waveform features and heart rate variations that are not subject to

（一）正則化方法
ニューラルネットワークモデルのパラメーターが多いので、オーバーフィット現象（over-fit phenomena）が容易に生ずるおそれがある。本実施例のモデルでトレーニングをするとき、正則化方法によりオーバーフィット現象を防止することができる。正則化方法はＬ１正則化方法またはＬ２正則化方法等を採用することができる。モデル内のパラメーターの制限により一部分のパラメーターがモデル全体の変化方向を導くように更新されることを有効に制御することができる。実際の応用において、モデルトレーニングをするとき、損失函数に所定の比率の前記パラメーターの平方を添加することにより内部のパラメーターの柔軟性制限を形成し、モデルパラメーターが大きすぎる可能性を減少させ、モデルの安定なトレーニングを確保することができる。
(1) Regularization Method Since there are many parameters in the neural network model, over-fit phenomena may easily occur. The regularization method can prevent the overfitting phenomenon when training with the model of the present invention. As the regularization method, the L1 regularization method, the L2 regularization method, or the like can be adopted. Limiting the parameters in the model can effectively control that some parameters are updated to guide the direction of change for the entire model. In practical application, when training the model, a certain ratio of the square of the parameter is added to the loss function to form an internal parameter flexibility limit, which reduces the possibility that the model parameter is too large, and the model stable training can be ensured.

最初の心電図信号に対して横方向引張処理及び／または横方向圧縮処理をするとき、予め設定された比率によって前記最初の心電図信号に対して横方向引張処理及び／または横方向圧縮処理をすることによりトレーニング信号を獲得する。
performing lateral tension processing and/or lateral compression processing on the initial electrocardiogram signal according to a preset ratio when performing lateral tension processing and/or lateral compression processing on the initial electrocardiogram signal; to acquire the training signal.

Claims

A method of detecting arrhythmia, said method of detecting arrhythmia being implemented by a processor of an electronic device, said method of detecting arrhythmia comprising:
obtaining a detected waiting electrocardiogram signal;
inputting the electrocardiogram signal into a preset arrhythmia detection model;
obtaining a detection result corresponding to the electrocardiogram signal by detecting the electrocardiogram signal with the arrhythmia detection model;
The arrhythmia detection model includes a convolutional neural network and a recurrent neural network connected in sequence,
the detection result includes an arrhythmia type of the electrocardiogram signal;
The arrhythmia detection model training process includes:
obtaining a training signal;
inputting the training signal into an arrhythmia detection model awaiting training and calculating a loss function value according to the loss function of the arrhythmia detection model;
training on the parameters of the arrhythmia detection model using the loss function values and by a backpropagation algorithm method, stopping training when the loss function values reach a predetermined value;
Obtaining the training signal comprises:
obtaining an initial electrocardiogram signal;
forming a training signal by preprocessing the initial electrocardiogram signal;
The preprocessing includes local covering processing,
forming a training signal by preprocessing the initial electrocardiogram signal,
forming a signal corresponding to a disconnection by randomly detecting a signal of less than a predetermined duration in the initial electrocardiogram signal and zeroing the signal;
An arrhythmia detection method characterized by:

The step of obtaining a detection result corresponding to the electrocardiogram signal by detecting the electrocardiogram signal with the arrhythmia detection model includes obtaining a detection result corresponding to the electrocardiogram signal by detecting features of the input electrocardiogram signal with the convolutional neural network. obtaining feature information and inputting the feature information into the recurrent neural network; and determining an arrhythmia type of the electrocardiogram signal by performing a classification process on the input feature information in the recurrent neural network. 2. The arrhythmia detection method of claim 1, comprising the steps of:

The step of acquiring feature information of the electrocardiogram signal by detecting features of the electrocardiogram signal input by the convolutional neural network includes performing convolution processing and nonlinear transformation on the electrocardiogram signal by the convolutional neural network. 3. The arrhythmia detection of claim 2, comprising forming a feature time series by: Method.

The step of determining an arrhythmia type of the electrocardiogram signal by performing a classification process on the feature information input by the recurrent neural network includes: for each feature segment in the feature time series by the recurrent neural network: determining a final state of the characteristic time series by iterating through the characteristic time series; and determining an arrhythmia type of the electrocardiogram signal by performing a classification process on the final state of the characteristic time series. 4. The arrhythmia detection method according to claim 3.

The preprocessing further includes noise processing, and the step of forming a training signal by preprocessing the initial electrocardiogram signal generates random noise corresponding to a preset signal-to-noise ratio amplitude from the initial electrocardiogram signal. 2. The arrhythmia detection method of claim 1, comprising obtaining a training signal by inputting an electrocardiogram signal.

The preprocessing further includes lateral tension processing and/or lateral compression processing, and the step of forming a training signal by preprocessing the initial electrocardiogram signal comprises the 2. The arrhythmia detection method of claim 1, comprising obtaining the training signal by subjecting initial electrocardiogram signals to lateral tension and/or lateral compression.

The process of training the arrhythmia detection model according to claim 1, wherein one or more of a regularization method, a network connection structure random removal method, and a label interference method are used for stable training. Arrhythmia detection method.

The convolutional neural network comprises a plurality of convolution layers and a plurality of pooling layers, each convolution layer comprising a convolution unit, a nonlinear transformation unit and a batch normalization unit connected in sequence, and the recurrent neural network includes a plurality of long-term short-term memory units, each said long-term short-term memory unit including a forget gate, an update gate and an output gate. Method.

An arrhythmia detection device, wherein the arrhythmia detection device is provided on a processor side of an electronic device, the arrhythmia detection device comprises:
a signal acquisition module configured to acquire a detection waiting electrocardiogram signal;
a model input module configured to input the electrocardiogram signal into a preset arrhythmia detection model;
a model detection module configured to detect an electrocardiogram signal with the arrhythmia detection model to obtain a detection result corresponding to the electrocardiogram signal;
a training signal acquisition module configured to acquire a training signal;
a loss calculation module provided to input the training signal to an arrhythmia detection model awaiting training and to calculate a loss function value according to a preset loss function of the arrhythmia detection model;
a training module configured to train parameters of the arrhythmia detection model using the loss function value by a backpropagation algorithm method and to stop training when the loss function value reaches a predetermined value;
including
the arrhythmia detection model includes a convolutional neural network and a recurrent neural network connected in sequence;
the detection result includes an arrhythmia type of the electrocardiogram signal;
the training signal acquisition module is arranged to acquire an initial electrocardiogram signal and then preprocess the initial electrocardiogram signal to form a training signal, including local covering;
said training signal is formed by randomly detecting a signal of less than a predetermined duration in said initial electrocardiogram signal and zeroing said signal to form a signal corresponding to disconnection. An arrhythmia detection device characterized by:

the model detection module includes a feature detection unit and a type determination unit, the feature detection unit obtains feature information of the electrocardiogram signal by detecting features of the input electrocardiogram signal in the convolutional neural network; and inputting the feature information to the recurrent neural network, wherein the type determination unit determines the type of arrhythmia of the electrocardiogram signal by performing classification processing on the input feature information in the recurrent neural network. 10. The arrhythmia detection device of claim 9, wherein the arrhythmia detection device is provided to

The feature detection unit is arranged such that the convolutional neural network performs convolution processing and non-linear transformation on the electrocardiogram signal to form a feature time series, the feature time series arranged in chronological order, and 11. The arrhythmia detection device of claim 10, comprising feature segments extracted in a plurality of electrocardiogram signals.

The type determination unit determines a final state of the feature time series by causing the recurrent neural network to iterate over each feature segment in the feature time series, and for the final state of the feature time series: 12. Arrhythmia detection device according to claim 11, characterized in that it is arranged to determine the type of arrhythmia of said electrocardiogram signal by performing a classification process on it.

The convolutional neural network comprises a plurality of convolution layers and a plurality of pooling layers, each convolution layer comprising a convolution unit, a nonlinear transformation unit and a batch normalization unit connected in sequence, and the recurrent neural network includes a plurality of long-term short-term memory units, each said long-term short-term memory unit including a forget gate, an update gate and an output gate.

An electronic device, wherein the electronic device includes a storage device and a processing device, and the storage device controls the processing device to perform the arrhythmia detection method according to any one of claims 1 to 8. electronic device, characterized in that a program is stored thereon for executing a program, and the processing unit is arranged to execute the program stored in the storage device.

A program that is read into the arrhythmia detection device according to any one of claims 9 to 13 and causes the arrhythmia detection device to execute the arrhythmia detection method according to any one of claims 1 to 8 is stored. computer storage medium.