A self-supervised atrial fibrillation detection method based on cross-modal cross-reconstruction and related equipment

Abstract

This invention provides a self-supervised atrial fibrillation detection method and related equipment based on cross-modal cross-reconstruction. The method includes: acquiring a test electrocardiogram (ECG) signal; performing time-frequency conversion on the test ECG signal to obtain a frequency domain ECG signal; inputting the test ECG signal and the frequency domain ECG signal into a preset self-supervised learning model to obtain a reconstructed ECG signal and a reconstructed ECG frequency domain signal; the self-supervised learning model includes an encoder, a time-domain decoder, and a frequency-domain decoder; the encoder is used to extract time-domain low-dimensional embedding features and frequency-domain low-dimensional embedding features; the time-domain decoder is used to generate the reconstructed ECG signal based on the frequency-domain low-dimensional embedding features; the frequency-domain decoder is used to generate the reconstructed ECG frequency domain signal based on the time-domain low-dimensional embedding features; calculating the reconstruction error between the reconstructed ECG frequency domain signal and the original ECG frequency domain signal; if the reconstruction error is greater than a preset abnormality score, the test ECG signal is considered atrial fibrillation data.

Description

Technical Field

[0001] This invention relates to the field of atrial fibrillation detection technology, and in particular to a self-supervised atrial fibrillation detection method and related equipment based on cross-modal cross-reconstruction. Background Technology

[0002] Atrial fibrillation (AF) is a common cardiac arrhythmia. With the development of modern technology, AF detection has become an important topic in the medical and health field. Currently, AF detection methods mainly include temporal anomaly detection methods and deep learning-based methods. The former often ignores the unique medical characteristics of AF electrocardiogram data, and its accuracy needs improvement; the latter usually relies on a large number of labeled AF samples to train a deep learning model, which incurs high labeling costs. Summary of the Invention

[0003] To reduce the labeling cost of atrial fibrillation samples and improve the average accuracy of atrial fibrillation detection, this invention provides a self-supervised atrial fibrillation detection method and related equipment based on cross-modal cross-reconstruction.

[0004] In a first aspect, the present invention provides a self-supervised atrial fibrillation detection method based on cross-modal cross-reconstruction, comprising:

[0005] Acquire the electrocardiogram signal to be tested;

[0006] The electrocardiogram signal to be measured is converted from time to frequency to obtain the electrocardiogram frequency domain signal;

[0007] The ECG signal to be tested and the ECG frequency domain signal are input into a preset self-supervised learning model to obtain a reconstructed ECG signal and a reconstructed ECG frequency domain signal. The self-supervised learning model is trained based on normal ECG signals. The self-supervised learning model includes an encoder, a time-domain decoder, and a frequency-domain decoder. The encoder is used to extract time-domain low-dimensional embedding features and frequency-domain low-dimensional embedding features from the input ECG signal and ECG frequency domain signal, respectively. The time-domain decoder is used to generate the reconstructed ECG signal based on the frequency-domain low-dimensional embedding features. The frequency-domain decoder is used to generate the reconstructed ECG frequency domain signal based on the time-domain low-dimensional embedding features.

[0008] The reconstruction error between the reconstructed ECG frequency domain signal and the original ECG frequency domain signal is calculated. If the reconstruction error is greater than a preset abnormality score, the ECG signal to be tested is considered to be atrial fibrillation data.

[0009] Furthermore, the training process of the self-supervised learning model includes:

[0010] Construct a normal electrocardiogram signal dataset;

[0011] The normal electrocardiogram (ECG) signal dataset and its corresponding original ECG frequency domain signal dataset are input into the self-supervised learning model to generate a reconstructed ECG signal dataset and a reconstructed ECG frequency domain signal dataset.

[0012] Minimize the joint loss to iteratively optimize the parameters of the self-supervised learning model and save the self-supervised learning model with the optimal parameters; wherein, the joint loss consists of the reconstruction loss between the reconstructed ECG signal dataset and the original ECG signal dataset and the reconstruction loss between the reconstructed ECG frequency domain signal dataset and the original ECG frequency domain signal dataset.

[0013] Furthermore, it also includes a time-domain predictor and a frequency-domain predictor; the time-domain predictor is used to predict time-domain ECG signal medical features based on the time-domain low-dimensional embedding features, and the frequency-domain predictor is used to predict frequency-domain ECG signal medical features based on the frequency-domain low-dimensional embedding features.

[0014] Correspondingly, the training process of the self-supervised learning model also includes:

[0015] The time-domain predictor and the frequency-domain predictor are used as regularizers to introduce a regularization term into the joint loss.

[0016] Furthermore, the time-domain electrocardiogram signal medical characteristics include heart rate and atrial fibrillation index.

[0017] Furthermore, the frequency domain electrocardiogram signal medical characteristics include power in the high-frequency band and power in the low-frequency band.

[0018] Secondly, the present invention provides a self-supervised atrial fibrillation detection device based on cross-modal cross-reconstruction, comprising:

[0019] The acquisition module is used to acquire the electrocardiogram signal to be tested;

[0020] The data preprocessing module is used to perform time-frequency conversion on the electrocardiogram signal to be tested to obtain the electrocardiogram frequency domain signal;

[0021] A cross-modal reconstruction module is used to input the ECG signal to be tested and the ECG frequency domain signal into a preset self-supervised learning model to obtain a reconstructed ECG signal and a reconstructed ECG frequency domain signal. The self-supervised learning model is trained based on normal ECG signals. The self-supervised learning model includes an encoder, a time-domain decoder, and a frequency-domain decoder. The encoder is used to extract time-domain low-dimensional embedding features and frequency-domain low-dimensional embedding features from the input ECG signal and ECG frequency domain signal, respectively. The time-domain decoder is used to generate the reconstructed ECG signal based on the frequency-domain low-dimensional embedding features. The frequency-domain decoder is used to generate the reconstructed ECG frequency domain signal based on the time-domain low-dimensional embedding features.

[0022] The detection module is used to calculate the reconstruction error between the reconstructed ECG frequency domain signal and the original ECG frequency domain signal. If the reconstruction error is greater than a preset abnormality score, the ECG signal to be tested is considered to be atrial fibrillation data.

[0023] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in the first aspect.

[0024] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0025] The beneficial effects of this invention are as follows:

[0026] (1) This invention models the atrial fibrillation detection task as an anomaly detection task and trains the self-supervised learning model based on normal electrocardiogram signal data. Therefore, it does not rely on a large number of labeled atrial fibrillation samples, thus reducing the labeling cost. Furthermore, by learning and reconstructing the global features of electrocardiogram signal data from both the time domain and frequency domain perspectives, the self-supervised learning model can learn more discriminative feature embeddings for atrial fibrillation detection, thereby improving the average accuracy of atrial fibrillation detection.

[0027] (2) When training the self-supervised learning model, the calculated time-frequency features are used as labels, and time-domain predictors and frequency-domain predictors are introduced as regularizers. Medical feature prediction is used to guide the model learning and extract effective features from the electrocardiogram data, so that the self-supervised learning model can learn more discriminative features for atrial fibrillation detection and further improve the accuracy of atrial fibrillation detection.

[0028] (3) Experimental results show that the method of the present invention outperforms the mainstream algorithms in AUC and AP of CPSC2021, Icentia11k and PTBXL by 6.29%, 9.22%, 3.57% and 9.37%, 11.94% and 2.84% respectively, proving the effectiveness of the model proposed in the present invention in the atrial fibrillation detection task. Attached Figure Description

[0029] Figure 1 A flowchart illustrating a self-supervised atrial fibrillation detection method based on cross-modal cross-reconstruction provided in an embodiment of the present invention;

[0030] Figure 2 The training process of a self-supervised learning model provided in an embodiment of the present invention;

[0031] Figure 3A framework diagram of a self-supervised atrial fibrillation detection method based on cross-modal cross-reconstruction provided in an embodiment of the present invention;

[0032] Figure 4 This is a schematic diagram of the structure of a self-supervised atrial fibrillation detection device based on cross-modal cross-reconstruction provided in an embodiment of the present invention;

[0033] Figure 5 This is a structural block diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0035] like Figure 1 As shown, this embodiment of the invention provides a self-supervised atrial fibrillation detection method based on cross-modal cross-reconstruction, comprising the following steps:

[0036] S101: Acquire the ECG signal to be measured;

[0037] Specifically, electrocardiogram (ECG) signals can be acquired using traditional ECG signal acquisition equipment (such as placing electrodes on specific parts of the user's body to capture ECG signals), or they can be acquired using portable and smart wearable devices (such as smartwatches or smartphones that acquire photoplethysmography (PPG) signals and then convert them into ECG signals), or they can be reconstructed using contactless ECG acquisition technology (such as millimeter-wave radar technology that reconstructs ECG signals by detecting minute vibrations in the chest wall).

[0038] S102: Perform time-frequency conversion on the electrocardiogram signal to be tested to obtain the electrocardiogram frequency domain signal;

[0039] Specifically, time-frequency conversion methods include, but are not limited to, Fourier transform, short-time Fourier transform, fast Fourier transform, and wavelet transform.

[0040] S103: Input the ECG signal to be tested and the ECG frequency domain signal into a preset self-supervised learning model to obtain the reconstructed ECG signal and the reconstructed ECG frequency domain signal;

[0041] Specifically, the self-supervised learning model is trained based on normal electrocardiogram (ECG) signals; the self-supervised learning model includes an encoder, a time-domain decoder, and a frequency-domain decoder; the encoder is used to extract time-domain low-dimensional embedding features and frequency-domain low-dimensional embedding features from the input ECG signal and ECG frequency-domain signal, respectively; the time-domain decoder is used to generate a reconstructed ECG signal based on the frequency-domain low-dimensional embedding features; the frequency-domain decoder is used to generate a reconstructed ECG frequency-domain signal based on the time-domain low-dimensional embedding features.

[0042] S104: Calculate the reconstruction error between the reconstructed ECG frequency domain signal and the original ECG frequency domain signal. If the reconstruction error is greater than the preset abnormality score, the ECG signal to be tested is considered to be atrial fibrillation data.

[0043] The self-supervised atrial fibrillation detection method based on cross-modal reconstruction provided in this invention modeles the atrial fibrillation detection task as an anomaly detection task and trains the self-supervised learning model based on normal electrocardiogram (ECG) signal data. Therefore, it does not rely on a large number of labeled atrial fibrillation samples, reducing labeling costs. Furthermore, by simultaneously learning and reconstructing the global features of ECG signal data from both the time and frequency domains, the self-supervised learning model can learn more discriminative feature embeddings for atrial fibrillation detection, thereby improving the average accuracy of atrial fibrillation detection.

[0044] In one embodiment, the present invention also provides a training process for a self-supervised learning model, including the following steps:

[0045] S201: Construct a normal electrocardiogram signal dataset;

[0046] S202: Input the normal electrocardiogram (ECG) signal dataset and its corresponding original ECG frequency domain signal dataset into the self-supervised learning model to generate a reconstructed ECG signal dataset and a reconstructed ECG frequency domain signal dataset;

[0047] Specifically, a time-frequency conversion is performed on a normal electrocardiogram (ECG) signal dataset to obtain a corresponding ECG frequency domain signal dataset. Meanwhile, to distinguish the ECG frequency domain signal dataset obtained through time-frequency conversion from the ECG frequency domain signal dataset generated by the self-supervised learning model, the former is referred to as the original ECG frequency domain signal dataset, and the latter as the reconstructed ECG frequency domain signal dataset.

[0048] S203: Minimize the joint loss to iteratively optimize the parameters of the self-supervised learning model and save the self-supervised learning model with the optimal parameters; wherein, the joint loss consists of the reconstruction loss between the reconstructed ECG signal dataset and the original ECG signal dataset and the reconstruction loss between the reconstructed ECG frequency domain signal dataset and the original ECG frequency domain signal dataset.

[0049] Specifically, the reconstruction loss can be, but is not limited to, mean squared error loss, mean absolute error loss, or a hybrid loss consisting of the two loss functions mentioned above.

[0050] In one embodiment, to further optimize the detection performance of the self-supervised learning model, a regularizer is introduced during training. The regularizer includes a time-domain predictor and a frequency-domain predictor. The time-domain predictor is used to predict time-domain ECG signal medical features based on the time-domain low-dimensional embedding features, and the frequency-domain predictor is used to predict frequency-domain ECG signal medical features based on the frequency-domain low-dimensional embedding features.

[0051] Specifically, based on normal electrocardiogram (ECG) signal data, time-domain ECG signal medical features (hereinafter referred to as first time-domain features) can be calculated, and based on the original ECG frequency-domain signal data, frequency-domain ECG signal medical features (hereinafter referred to as first frequency-domain features) can be calculated. The predicted time-domain ECG signal medical features are denoted as second time-domain features, and the predicted frequency-domain ECG signal medical features are denoted as second frequency-domain features. By calculating the distance between the first and second time-domain features, and the distance between the first and second frequency-domain features, regularization terms (such as L1 or L2 regularization) are generated based on these two distances. Finally, by introducing this regularization term into the original joint loss, the self-supervised learning model can learn more discriminative feature embeddings for atrial fibrillation detection, thereby improving its detection performance.

[0052] In one embodiment, the time-domain electrocardiogram (ECG) signal medical features include heart rate (HR) and the atrial fibrillation index (AFI). Specifically, for the first time-domain feature, the AFI is calculated as follows: first, the NN interval (NNI) is calculated based on the R peak of the ECG signal; then, the AFI is determined based on each obtained NNI, as shown in the following formula:

[0053]

[0054] Where n represents the number of NNIs. The R peak is an important characteristic peak in the QRS complex, usually representing the potential change during ventricular depolarization. The QRS complex reflects the entire process of ventricular depolarization, and the R peak is the first positive peak in the QRS complex, marking the beginning of ventricular depolarization.

[0055] In one embodiment, the frequency domain electrocardiogram (ECG) signal medical characteristics include high-frequency power and low-frequency power. Specifically, for the first frequency domain characteristic, a Fast Fourier Transform (FFT) is first applied to the ECG signal data, a time-series signal, to obtain the spectral feature distribution and the frequency domain signal X.f Next, the power in the high-frequency band (HF) and low-frequency band (LF) is extracted. The formulas for the above process are as follows:

[0056]

[0057] F(K)=|X(f)| 2 (3)

[0058]

[0059] HF=sum(F(K)[i]),i∈[0.15hz,0.4hz] (5)

[0060] LF=sum(F(K)[i]),i∈[0.04hz,0.15hz] (6)

[0061] Where X is the electrocardiogram signal, X f For the corresponding frequency domain signal, F(K) is the power spectral density in the frequency domain, which can be calculated using X(f) obtained from the FFT. HF represents high-frequency power, with a frequency range between 0.15 Hz and 0.4 Hz. LF represents low-frequency power, with a frequency range between 0.04 Hz and 0.15 Hz.

[0062] In one embodiment, the electrocardiogram signal is represented as X = {x} 1 ,x 2 ,x 3 …x n}, and define the encoder as E n Define the time-domain decoder as D t The frequency domain decoder is defined as D f The original electrocardiogram signal and its corresponding frequency domain signal were used as input, and the time-frequency signal was reconstructed by an autoencoder. The encoder used a full-scale convolutional neural network that dynamically adjusted the convolution kernel during feature representation, while the decoder consisted of multiple layers of transposed convolutional neural networks.

[0063] Specifically, the electrocardiogram signal X and the corresponding frequency domain signal X f Input to encoder E respectively n In this process, the time-domain low-dimensional embedded signal Z is obtained. t and frequency domain low-dimensional embedded signal Z f Then, a frequency domain decoder D is used. f Based on time-domain low-dimensional embedded signal Z t Frequency domain decoder reconstructs ECG frequency domain signal X f ′, and use time-domain decoder D t Based on the frequency domain low-dimensional embedded signal Z fThe electrocardiogram signal X′ is reconstructed using a time-domain decoder. The specific process and formula are shown below:

[0064] Z t =E n (X) (7)

[0065] Z f =E n (X f (8)

[0066] X′ f =D f (Z t (9)

[0067] X′=D t (Z f (10)

[0068] In one embodiment, the prediction module enhances the model's ECG signal representation capability by predicting time-domain and frequency-domain ECG signal medical features using low-dimensional features. The original ECG signal X is used to obtain the time-domain AFI and HR values, as well as the frequency-domain HF and LF values. The calculated AFI value is defined as N. a The calculated HR value is defined as N. hr The high-frequency characteristic HF value is defined as N. hf The low-frequency characteristic LF value is defined as N. lf Subsequently, the raw electrocardiogram signal X and the frequency domain signal are input into encoder E. n To obtain the time-domain low-dimensional embedding Z t and frequency domain low-dimensional embedding Z f Subsequently, a multilayer perceptron (MLP) prediction head was used to process Z. t Generate predicted values ​​of time-domain feature values ​​AFI and HR, and process Z. f The predicted values ​​of the frequency domain eigenvalues ​​HF and LF are generated. The above prediction process can be expressed by the following formula:

[0069] N a ′,N hr ′=P t (Z t (11)

[0070] N hf ′,N lf ′=P f (Z f (12)

[0071] Where, N a ′、N hr ′、N hf ′ and Nlf ′ represents the predicted result of ECG medical characteristics, while P t P f It is a time-domain and frequency-domain predictor that obtains ECG feature prediction values ​​from low-dimensional feature embedding.

[0072] To verify the effectiveness of the present invention, the present invention also quantitatively evaluated the present invention's solution (model framework as follows). Figure 3 The anomaly detection performance (shown in Table 1) is presented below. Some experimental details are given below. The results of all comparison algorithms on each dataset are shown in Table 1. The effectiveness of the proposed method is illustrated by showing the model's results on time series anomaly detection tasks. On the CPSC2021 dataset, compared with other anomaly detection algorithms with fixed network parameters, the proposed method has the highest AUC and AP scores, exceeding the best comparison algorithm by 6.29% and 9.37%, respectively. On the Icentia11k dataset, when the network parameters are fixed, the proposed method outperforms other anomaly detection algorithms by 9.22% and 11.94% in AUC and AP, respectively. On the PTBXL dataset, when the network parameters are fixed, the proposed method outperforms other anomaly detection algorithms by 3.57% and 2.28% in AUC and AP, respectively.

[0073] Table 1 Comparison of Algorithm Performance

[0074]

[0075] The results show that the proposed solution is superior to traditional anomaly detection algorithms, demonstrating its potential for practical application in clinical settings.

[0076] Based on the same inventive concept, such as Figure 4 As shown, this embodiment of the invention also provides a self-supervised atrial fibrillation detection device based on cross-modal cross-reconstruction, including: an acquisition module, a data preprocessing module, a cross-modal cross-reconstruction module, and a detection module.

[0077] The acquisition module acquires the ECG signal to be tested; the data preprocessing module performs time-frequency conversion on the ECG signal to obtain the ECG frequency domain signal; the cross-modal reconstruction module inputs the ECG signal to be tested and the ECG frequency domain signal into a preset self-supervised learning model to obtain a reconstructed ECG signal and a reconstructed ECG frequency domain signal; the self-supervised learning model is trained based on normal ECG signals; the self-supervised learning model includes an encoder, a time-domain decoder, and a frequency-domain decoder; the encoder extracts time-domain low-dimensional embedding features and frequency-domain low-dimensional embedding features from the input ECG signal and ECG frequency domain signal, respectively; the time-domain decoder generates the reconstructed ECG signal based on the frequency-domain low-dimensional embedding features; the frequency-domain decoder generates the reconstructed ECG frequency domain signal based on the time-domain low-dimensional embedding features; the detection module calculates the reconstruction error between the reconstructed ECG frequency domain signal and the original ECG frequency domain signal; if the reconstruction error is greater than a preset abnormality score, the ECG signal to be tested is considered to be atrial fibrillation data.

[0078] In one embodiment, the self-supervised atrial fibrillation detection device based on cross-modal cross-reconstruction provided by this invention further includes a training module; the training module is used to construct a normal electrocardiogram (ECG) signal dataset; input the normal ECG signal dataset and its corresponding original ECG frequency domain signal dataset into the self-supervised learning model to generate a reconstructed ECG signal dataset and a reconstructed ECG frequency domain signal dataset; minimize the joint loss to iteratively optimize the parameters of the self-supervised learning model, and save the self-supervised learning model with the optimal parameters; wherein, the joint loss consists of the reconstruction loss between the reconstructed ECG signal dataset and the original ECG signal dataset and the reconstruction loss between the reconstructed ECG frequency domain signal dataset and the original ECG frequency domain signal dataset.

[0079] In one embodiment, the self-supervised atrial fibrillation detection device based on cross-modal cross-reconstruction provided by this invention further includes a time-domain predictor and a frequency-domain predictor. The time-domain predictor is used to predict time-domain ECG signal medical features based on the time-domain low-dimensional embedding features, and the frequency-domain predictor is used to predict frequency-domain ECG signal medical features based on the frequency-domain low-dimensional embedding features. Correspondingly, the training module is used to use the time-domain predictor and the frequency-domain predictor as regularizers to introduce a regularization term into the joint loss.

[0080] It should be noted that the self-supervised atrial fibrillation detection device based on cross-modal cross-reconstruction provided in the various embodiments of the present invention is for the purpose of implementing the above methods. Its specific functions can be referred to in the above method embodiments, and will not be repeated here.

[0081] Figure 5 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 5 As shown, the electronic device may include: a processor 501, a communication interface 502, a memory 503, and a communication bus 504, wherein the processor 501, the communication interface 502, and the memory 503 communicate with each other through the communication bus 504. The processor 501 can call logic instructions in the memory 503 to execute a self-supervised atrial fibrillation detection method. This method includes: acquiring a test electrocardiogram (ECG) signal; performing time-frequency conversion on the test ECG signal to obtain a frequency domain ECG signal; inputting the test ECG signal and the frequency domain ECG signal into a preset self-supervised learning model to obtain a reconstructed ECG signal and a reconstructed ECG frequency domain signal; the self-supervised learning model is trained based on normal ECG signals; the self-supervised learning model includes an encoder, a time-domain decoder, and a frequency-domain decoder; the encoder is used to extract time-domain low-dimensional embedding features and frequency-domain low-dimensional embedding features from the input ECG signal and the ECG frequency domain signal, respectively; the time-domain decoder is used to generate a reconstructed ECG signal based on the frequency-domain low-dimensional embedding features; the frequency-domain decoder is used to generate a reconstructed ECG frequency domain signal based on the time-domain low-dimensional embedding features; calculating the reconstruction error between the reconstructed ECG frequency domain signal and the original ECG frequency domain signal; if the reconstruction error is greater than a preset abnormality score, the test ECG signal is considered atrial fibrillation data.

[0082] Furthermore, when the logical instructions in the aforementioned memory 503 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0083] This invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute the self-supervised atrial fibrillation detection method provided in the above-described method embodiments.

[0084] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the self-supervised atrial fibrillation detection method provided in the above-described method embodiments.

[0085] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A self-supervised atrial fibrillation detection method based on cross-modal cross-reconstruction, characterized in that, include: Acquire the electrocardiogram signal to be tested; The electrocardiogram signal to be measured is converted from time to frequency to obtain the electrocardiogram frequency domain signal; The ECG signal to be tested and the ECG frequency domain signal are input into a preset self-supervised learning model to obtain a reconstructed ECG signal and a reconstructed ECG frequency domain signal; the self-supervised learning model is trained based on normal ECG signals; the self-supervised learning model includes a shared weight encoder. Time-domain decoder and frequency domain decoder The shared weight encoder Used to extract the input electrocardiogram signals respectively and electrocardiogram frequency domain signal Extracting low-dimensional embedding features in the temporal domain and frequency domain low-dimensional embedding features The time-domain decoder For use based on the frequency domain low-dimensional embedding features Generate reconstructed electrocardiogram signals The frequency domain decoder For use based on the time-domain low-dimensional embedding features Generate reconstructed electrocardiogram frequency domain signal The corresponding formula is as follows: During the training of the self-supervised learning model, a regularizer is introduced, which includes a time-domain predictor and a frequency-domain predictor. The time-domain predictor is used to predict time-domain ECG signal medical features based on the time-domain low-dimensional embedding features, and the frequency-domain predictor is used to predict frequency-domain ECG signal medical features based on the frequency-domain low-dimensional embedding features. The time-domain predictor and the frequency-domain predictor are used as regularizers to introduce a regularization term into the original joint loss, minimizing the joint loss after introducing the regularization term, so as to iteratively optimize the parameters of the self-supervised learning model and save the self-supervised learning model with the optimal parameters; wherein, the original joint loss consists of the reconstruction loss between the reconstructed ECG signal dataset and the original ECG signal dataset and the reconstruction loss between the reconstructed ECG frequency domain signal dataset and the original ECG frequency domain signal dataset; the time-domain ECG signal medical features include heart rate and atrial fibrillation index, and the frequency-domain ECG signal medical features include high-frequency power and low-frequency power; The reconstruction error between the reconstructed ECG frequency domain signal and the original ECG frequency domain signal is calculated. If the reconstruction error is greater than a preset abnormality score, the ECG signal to be tested is considered to be atrial fibrillation data.

2. The self-supervised atrial fibrillation detection method based on cross-modal cross-reconstruction according to claim 1, characterized in that, The training process of the self-supervised learning model also includes: Construct a normal electrocardiogram signal dataset; The normal electrocardiogram (ECG) signal dataset and its corresponding original ECG frequency domain signal dataset are input into the self-supervised learning model to generate a reconstructed ECG signal dataset and a reconstructed ECG frequency domain signal dataset.

3. A self-supervised atrial fibrillation detection device based on cross-modal cross-reconstruction, characterized in that, include: The acquisition module is used to acquire the electrocardiogram signal to be tested; The data preprocessing module is used to perform time-frequency conversion on the electrocardiogram signal to be tested to obtain the electrocardiogram frequency domain signal; A cross-modal reconstruction module is used to input the ECG signal to be tested and the ECG frequency domain signal into a preset self-supervised learning model to obtain a reconstructed ECG signal and a reconstructed ECG frequency domain signal; the self-supervised learning model is trained based on normal ECG signals; the self-supervised learning model includes a shared weight encoder. Time-domain decoder and frequency domain decoder The shared weight encoder Used to extract the input electrocardiogram signals respectively and electrocardiogram frequency domain signal Extracting low-dimensional embedding features in the temporal domain and frequency domain low-dimensional embedding features The time-domain decoder For use based on the frequency domain low-dimensional embedding features Generate reconstructed electrocardiogram signals The frequency domain decoder For use based on the time-domain low-dimensional embedding features Generate reconstructed electrocardiogram frequency domain signal The corresponding formula is as follows: During the training of the self-supervised learning model, a regularizer is introduced, which includes a time-domain predictor and a frequency-domain predictor. The time-domain predictor is used to predict time-domain ECG signal medical features based on the time-domain low-dimensional embedding features, and the frequency-domain predictor is used to predict frequency-domain ECG signal medical features based on the frequency-domain low-dimensional embedding features. The time-domain predictor and the frequency-domain predictor are used as regularizers to introduce a regularization term into the original joint loss, minimizing the joint loss after introducing the regularization term, so as to iteratively optimize the parameters of the self-supervised learning model and save the self-supervised learning model with the optimal parameters; wherein, the original joint loss consists of the reconstruction loss between the reconstructed ECG signal dataset and the original ECG signal dataset and the reconstruction loss between the reconstructed ECG frequency domain signal dataset and the original ECG frequency domain signal dataset; the time-domain ECG signal medical features include heart rate and atrial fibrillation index, and the frequency-domain ECG signal medical features include high-frequency power and low-frequency power; The detection module is used to calculate the reconstruction error between the reconstructed ECG frequency domain signal and the original ECG frequency domain signal. If the reconstruction error is greater than a preset abnormality score, the ECG signal to be tested is considered to be atrial fibrillation data.

4. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in claim 1 or 2.

5. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in claim 1 or 2.

Images