Atrial fibrillation classification prediction method based on millimeter wave radar signals

Abstract

The present invention provides an atrial fibrillation classification prediction method based on millimeter-wave radar signals, comprising: obtaining multiple reference signals for characterizing the heartbeat movement law of the heart part of the measured object, wherein the multiple reference signals are obtained by processing multiple original signals, and the multiple original signals are: multiple millimeter-wave radar echo signals reflected by the irradiated area of ​​the measured object; inputting the multiple reference signals into a trained student model, and outputting a prediction result of whether the measured object belongs to a predetermined category.

Description

Technical Field

[0001] The present invention relates to the field of intelligent sensing technology, and in particular to a method for classifying and predicting atrial fibrillation based on millimeter-wave radar signals. Background Art

[0002] Atrial fibrillation is a relatively common cardiovascular disease, which is associated with severe heart dysfunction and a significantly increased risk of stroke. With the trend of an aging population, the overall prevalence of atrial fibrillation is increasing year by year. Atrial fibrillation is a progressive disease characterized by changes in the structure of the heart over time, which can cause atrial fibrillation to develop from a paroxysmal state to a more stable and difficult-to-treat persistent atrial fibrillation state. Therefore, early detection of atrial fibrillation and appropriate intervention measures are crucial.

[0003] In the process of realizing the concept of the present invention, the inventors found that there are at least the following problems in the related technology: most of the commonly used atrial fibrillation monitoring devices are contact devices, so there is a risk of cross infection and inconvenience in use. In addition, the accuracy of the existing non-contact atrial fibrillation classification prediction method is not high enough, and there are very few other types in its non-atrial fibrillation samples, and some even only use normal sinus rhythm as non-atrial fibrillation category samples. Therefore, for non-contact atrial fibrillation detection, the current technology has certain limitations. Summary of the invention

[0004] In view of the above problems, the present invention provides an atrial fibrillation classification and prediction method based on millimeter wave radar signals.

[0005] One aspect of the present invention provides an atrial fibrillation classification prediction method based on millimeter-wave radar signals, comprising: obtaining multiple reference signals for characterizing the heartbeat movement pattern of the heart part of the object under test, wherein the multiple reference signals are obtained by processing multiple original signals, and the multiple original signals are: multiple millimeter-wave radar echo signals reflected by the irradiated area of ​​the object under test; inputting the multiple reference signals into a trained student model, and outputting a prediction result of whether the object under test belongs to a predetermined category.

[0006] According to an embodiment of the present invention, the student model is trained using the following method: obtaining multiple measured electrocardiogram data samples for multiple sample objects, multiple reference signal samples for characterizing the heartbeat movement patterns of the multiple sample objects, and multiple first sample labels for characterizing the true classification results of the multiple sample objects; using the multiple measured electrocardiogram data samples and the multiple first sample labels to train a first initial classification model to obtain a trained teacher model; using the teacher model as a guide, training a second initial classification model based on the multiple measured electrocardiogram data samples, the multiple reference signal samples, and the multiple first sample labels to obtain a trained student model.

[0007] According to an embodiment of the present invention, with the teacher model as a guide, the second initial classification model is trained based on multiple measured electrocardiogram data samples, multiple reference signal samples, and multiple first sample labels to obtain a trained student model, including: performing feature processing on multiple measured electrocardiogram data samples and multiple reference signal samples to determine common features between multiple measured electrocardiogram data samples and multiple reference signal samples; inputting multiple reference signal samples into the teacher model and the second initial classification model, while retaining the common feature channels corresponding to the common features in the teacher model, closing the remaining self-owned feature channels except the feature channels, using the teacher model to output a first prediction result of whether the sample object belongs to a predetermined category, and using the second initial classification model to output a second prediction result of whether the sample object belongs to a predetermined category; using the first prediction result as a guidance label, calculating a first loss between the guidance label and the second prediction result; calculating a second loss between the first sample label and the second prediction result; adjusting the network parameters of the second initial classification model based on the first loss and the second loss to obtain a trained student model.

[0008] According to an embodiment of the present invention, obtaining multiple reference signals for characterizing the law of heartbeat movement in the heart area of ​​the measured object includes: transmitting a millimeter-wave radar transmission signal to the measured object, and receiving multiple millimeter-wave radar echo signals reflected by the irradiated area of ​​the measured object; performing a first signal processing on the multiple millimeter-wave radar echo signals to obtain multiple first-stage signals for characterizing the law of mechanical vibration on the body surface of the irradiated area of ​​the measured object; performing a second signal processing on the multiple first-stage signals to eliminate the influence of the body surface vibration caused by the breathing of the measured object, and obtain multiple second-stage signals; performing a third signal processing on the multiple second-stage signals to eliminate the influence of the body surface vibration of the non-heart area in the irradiated area, and obtain multiple reference signals for characterizing the law of heartbeat movement in the heart area of ​​the measured object.

[0009] According to an embodiment of the present invention, performing first signal processing on multiple millimeter-wave radar echo signals includes: performing coherent accumulation operations on multiple millimeter-wave radar echo signals defined by multiple channels composed of N transmitting antennas and M receiving antennas.

[0010] According to an embodiment of the present invention, performing coherent accumulation operation on a plurality of millimeter-wave radar echo signals defined by a plurality of channels includes adopting the following method:

[0011]

[0012] Where: S i , S(x,y,z,t) is the i-th signal corresponding to the spatial position (x,y,z) and signal acquisition period t among the multiple first-stage signals obtained after the coherent accumulation operation; N is the number of receiving antennas; M is the number of transmitting antennas; yn,m,t is the signal of the channel defined by the nth receiving antenna and the mth transmitting antenna in the acquisition period t among multiple millimeter-wave radar echo signals; k is the rate of change of the transmission frequency of the millimeter-wave radar transmitting signal; λ is the wavelength of the millimeter-wave radar transmitting signal; r(x, y, z, n, m) is the round-trip distance from the mth transmitting antenna to the target position (x, y, z) and back to the nth receiving antenna; c is the speed of light.

[0013] According to an embodiment of the present invention, performing second signal processing on the plurality of first stage signals includes: performing second-order difference processing on the plurality of first stage signals to eliminate the influence of body surface vibration caused by breathing of the measured object.

[0014] According to an embodiment of the present invention, performing second-order difference processing on a plurality of first-stage signals includes:

[0015]

[0016] Where: S i is the i-th signal among multiple first-stage signals; X i , S i ″ is the value for the i-th S i The i-th second-stage signal obtained by second-order difference processing; h is the period of each first-stage signal; S i,-n For S i The second-stage signal of the previous n-th sample, n = 1, 2, 3; S i,+n For S i The second stage signal of the nth sample is then obtained.

[0017] According to an embodiment of the present invention, performing third signal processing on multiple second-stage signals includes: calculating the correlation between each second-stage signal and each first-related signal group, wherein the first-related signal group is composed of multiple signals within a predetermined three-dimensional neighborhood range of the second-stage signal; determining the second-stage center signal from the multiple second-stage signals based on the correlation; and generating multiple reference signals for characterizing the heartbeat movement law of the heart part of the measured object based on the second-stage center signal and the second-related signal group, wherein the second-related signal group is composed of multiple signals within a predetermined two-dimensional neighborhood range of the second-stage center signal.

[0018] According to an embodiment of the present invention, calculating the correlation between each second stage signal and each first associated signal group includes:

[0019]

[0020] Where: X i is the i-th second-stage signal, m is the number of second-stage signals; Y iis any one of the multiple signals within the predetermined three-dimensional neighborhood of the i-th second stage signal; cov(X i ,Y i ) is X i and Y i The covariance between them; σ is the standard deviation.

[0021] According to the atrial fibrillation classification prediction method based on millimeter wave radar signals provided by the present invention, the mechanical motion information of the heart can be completely and accurately extracted, and the deep feature extraction of the mechanical motion signal of the heart can be realized. Classification and prediction of atrial fibrillation based on millimeter wave radar signals can avoid the discomfort caused by wearing electrodes and sensor equipment, and can avoid cross infection. Compared with the traditional contact measurement method, non-contact detection is realized, which greatly improves the user experience and scope of application (for example, due to poor physical health, the traditional measurement method cannot be implemented, and this method is not limited by this).

[0022] Furthermore, the student model is trained with the teacher model as a guide, so that the prediction accuracy of the trained student model is relatively high. Based on the prediction result of whether the object under test belongs to the predetermined category output by the trained student model, the purpose of using artificial intelligence methods to perform feature analysis on signal data to predict the results is achieved. The prediction results are reliable and have good reference value. The prediction results are provided for user reference, providing a new information dimension and improving user experience.

[0023] In addition, during the training process, by eliminating the influence of body surface vibration caused by breathing and body surface vibration of non-heart parts in the millimeter wave radar echo signal, the reference signal can more accurately represent the law of heartbeat movement of the heart part of the measured object; by calculating the correlation between each second stage signal and each first associated signal group, multiple reference signals that more accurately represent the law of heartbeat movement of the heart part of the measured object can be obtained. Thus, the prediction accuracy of the student model is further improved.

[0024] According to an embodiment of the present invention, heartbeat can cause mechanical vibration of the body surface in a certain area, and the millimeter wave radar echo signal in this area has a high correlation. Therefore, by calculating the correlation between each second-stage signal and each first associated signal group, multiple reference signals that more accurately characterize the heartbeat movement law of the heart part of the measured object can be obtained. This processing method can accurately eliminate the influence of the body surface vibration of the non-heart part in the irradiated area, and obtain a signal for characterizing the heartbeat movement law of the heart part of the measured object.

[0025] According to an embodiment of the present invention, since the electrocardiogram data and the heartbeat movement law signal data detected by the millimeter-wave radar signal have similar data distribution laws, and the measured electrocardiogram data and the millimeter-wave radar signal each have some own characteristics, if these own characteristics are involved in the training, it will affect the training effect of the student model. In the method of the present invention, the teacher network is guided based on the common features, which can improve the training effect of the student model. The common features of the teacher network can be used as knowledge and passed to the student network. When training the second initial classification model, in addition to extracting information about the heartbeat movement law of the heart part in the reference signal, the second initial classification model will also learn the knowledge of the teacher network, so as to achieve faster convergence speed and higher monitoring performance. According to an embodiment of the present invention, based on the common features between the measured electrocardiogram data and multiple reference signals, the process of training the second initial classification model is guided, which can improve the prediction accuracy of the student model. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The above contents and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:

[0027] Figure 1 A flowchart of a method for classifying and predicting atrial fibrillation based on millimeter-wave radar signals according to an embodiment of the present invention is schematically shown;

[0028] Figure 2 The appearance of a millimeter wave radar according to an embodiment of the present invention is schematically shown;

[0029] Figure 3 Schematically shows the appearance of an electrocardiogram detection device according to an embodiment of the present invention;

[0030] Figure 4 A schematic diagram schematically shows a measured electrocardiogram data sample and a reference signal sample according to an embodiment of the present invention;

[0031] Figure 5 A schematic diagram schematically illustrates common features between measured electrocardiogram data samples and reference signal samples according to an embodiment of the present invention;

[0032] Figure 6 A schematic diagram of the network structure of a teacher model and a student model according to an embodiment of the present invention is shown;

[0033] Figure 7 A schematic diagram schematically shows a result of spatial beamforming during non-cardiac micromotion sensing of a subject under test according to an embodiment of the present invention;

[0034] Figure 8 The test environment in Embodiment 1 of the present invention is schematically shown;

[0035] Fig. 9 A box plot schematically shows the distribution of prediction results obtained by using the prediction method of the present invention for multiple subjects. DETAILED DESCRIPTION

[0036] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.

[0037] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0038] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification, and should not be interpreted in an idealized or overly rigid manner.

[0039] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).

[0040] The present invention provides a method for classifying and predicting atrial fibrillation based on millimeter wave radar signals. In the process of implementing the present invention, it is found that most of the existing devices for monitoring atrial fibrillation are contact devices, such as smart watches, electrocardiographs, etc. Since electrodes, sensor devices, etc. need to be worn to monitor atrial fibrillation, users with allergies or skin damage may not be able to use such devices. In addition, for devices such as smart watches, the accuracy of atrial fibrillation monitoring is not high enough. For example, smart watches usually only judge whether it is atrial fibrillation based on the irregularity of heartbeats, ignoring other characteristics of atrial fibrillation, such as when users have arrhythmias caused by staying up late, they may also be misjudged as atrial fibrillation. In the process of implementing the present invention, it is found that millimeter wave radar has accurate spatial perception capabilities, thus providing the possibility for non-contact atrial fibrillation monitoring. Using millimeter wave radar for atrial fibrillation monitoring can solve the problem of user compliance, avoid the discomfort caused by wearing electrodes and sensor devices, and thus promote the integration of health-integrated life health monitoring into daily life. In addition, in the process of implementing the present invention, it is also found that the prediction accuracy of atrial fibrillation classification prediction based on a trained student model is higher. On this basis, the present invention provides a method for classifying and predicting atrial fibrillation based on millimeter wave radar signals.

[0041] Figure 1 The flowchart of the atrial fibrillation classification prediction method based on millimeter wave radar signals according to an embodiment of the present invention is schematically shown.

[0042] like Figure 1 As shown, the atrial fibrillation classification prediction method based on millimeter wave radar signals in this embodiment includes operations S110 to S120.

[0043] In operation S110, a plurality of reference signals for characterizing the heartbeat movement pattern of the measured object's heart are obtained, wherein the plurality of reference signals are obtained by processing a plurality of original signals, and the plurality of original signals are: a plurality of millimeter wave radar echo signals reflected by the irradiated area of ​​the measured object.

[0044] According to an embodiment of the present invention, a millimeter wave radar may be used to transmit a millimeter wave signal to an object under test.

[0045] Figure 2 The appearance of a millimeter wave radar according to an embodiment of the present invention is schematically shown. The millimeter wave radar may be a 3-transmitter 4-receiver device.

[0046] According to an embodiment of the present invention, after receiving a millimeter wave signal, the radiation area of ​​the measured object can reflect the received millimeter wave signal to obtain a plurality of original signals.

[0047] Specifically, the radiation area may be the radiation range of the millimeter wave radar, for example including but not limited to the chest area of ​​the object being measured.

[0048] In addition to the target signal required for detection, that is, the millimeter-wave radar echo signal generated by the body surface vibration caused by heartbeat movement, the original signal also implicitly contains a variety of interference signals, such as the millimeter-wave radar echo signal generated by the body surface vibration caused by breathing, and the millimeter-wave radar echo signal generated by other radiation areas except the heart. Therefore, it is necessary to process multiple original signals to obtain multiple reference signals for characterizing the heartbeat movement laws of the subject.

[0049] In operation S120, a plurality of reference signals are input into the trained student model, and a prediction result of whether the detected object belongs to a predetermined category is output.

[0050] According to an embodiment of the present invention, the prediction result may be: the object being tested belongs to a predetermined category, or does not belong to a predetermined category. The object belonging to the predetermined category is predicted to be likely to suffer from atrial fibrillation. It should be noted that the prediction result obtained here is only used as reference information for medical staff to make a diagnosis, and is not a direct diagnosis result.

[0051] According to an embodiment of the present invention, it is possible to determine whether atrial fibrillation occurs in the subject under test based on the student model. Specifically, after obtaining multiple reference signals, the multiple reference signals can be input into the trained student model to determine whether the subject under test belongs to a predetermined category.

[0052] According to an embodiment of the present invention, after a plurality of reference signals are input into a trained student model, the student model can output a prediction score. It can be determined whether atrial fibrillation occurs in the subject under test based on a predetermined threshold, for example: when the prediction score is greater than the predetermined threshold, the subject under test belongs to a predetermined category; when the prediction score is less than the predetermined threshold, the subject under test does not belong to the predetermined category.

[0053] According to the atrial fibrillation classification and prediction method based on millimeter-wave radar signals provided by the present invention, atrial fibrillation classification and prediction based on millimeter-wave radar signals can avoid the discomfort caused by wearing electrodes and sensor equipment, and can avoid cross-infection. Compared with traditional contact measurement methods, non-contact detection is achieved, which greatly improves user experience and scope of application.

[0054] Furthermore, based on the trained student model output, a prediction result of whether the tested object belongs to a predetermined category is achieved, thereby achieving the purpose of using artificial intelligence methods to perform feature analysis on signal data and predict results. The prediction result is reliable and has good reference value. The prediction result is provided for user reference, providing a new information dimension and improving user experience.

[0055] According to the embodiment of the present invention, the student model is obtained through operations 11 to 13.

[0056] Operation 11: Acquire multiple measured electrocardiogram data samples for multiple sample objects, multiple reference signal samples for characterizing the heartbeat movement patterns of the multiple sample objects, and multiple first sample labels for characterizing the true classification results of the multiple sample objects.

[0057] According to an embodiment of the present invention, electrocardiogram (ECG) detection may be performed on a plurality of sample objects in advance to obtain a plurality of actually measured ECG data samples for the plurality of sample objects.

[0058] Figure 3 The external appearance of an electrocardiogram detection device according to an embodiment of the present invention is schematically shown. The electrocardiogram detection device is used to obtain a plurality of measured electrocardiogram data samples of a plurality of sample objects.

[0059] Figure 4 The schematic diagram schematically shows a measured electrocardiogram data sample and a reference signal sample according to an embodiment of the present invention.

[0060] Depend on Figure 4 It can be seen that the reference signals of non-atrial fibrillation patients are significantly different from those of atrial fibrillation patients. Similarly, the measured electrocardiogram data of non-atrial fibrillation patients are also significantly different from those of atrial fibrillation patients. Therefore, by obtaining multiple reference signals for characterizing the heartbeat movement law of the subject under test and inputting the multiple reference signals into the trained student model, the atrial fibrillation classification prediction of the subject under test can be achieved.

[0061] According to an embodiment of the present invention, the actual classification result may include whether the sample object belongs to a predetermined category.

[0062] Operation 12: Use multiple measured electrocardiogram data samples and multiple first sample labels to train a first initial classification model to obtain a trained teacher model.

[0063] According to an embodiment of the present invention, the accuracy of using an electrocardiogram to determine whether atrial fibrillation occurs is relatively high, so a first initial classification model can be trained using multiple measured electrocardiogram data samples and multiple first sample labels to obtain a trained teacher model. The trained teacher model has a high classification accuracy.

[0064] Operation 13: Guided by the teacher model, a second initial classification model is trained based on multiple measured electrocardiogram data samples, multiple reference signal samples, and multiple first sample labels to obtain a trained student model.

[0065] According to an embodiment of the present invention, since the teacher model has a high classification accuracy, in order to improve the classification accuracy of the student model, the teacher model can be used as a guide to train a second initial classification model based on multiple measured electrocardiogram data samples, multiple reference signal samples, and multiple first sample labels to obtain a trained student model.

[0066] According to an embodiment of the present invention, considering the lack of sufficient samples for training based on the millimeter wave radar signal method, the teacher model in the present invention is first trained with real electrocardiogram samples and real labels, and then used to guide the training of the student model, taking into account the problem of lack of samples and ensuring the accuracy of model training. Specifically, the teacher model is trained with real electrocardiogram samples and real labels, and has good prediction accuracy. Since the classification accuracy of the teacher model is relatively high, the present invention uses the teacher model as a guide to train the second initial classification model based on multiple measured electrocardiogram data samples, multiple reference signal samples, and multiple first sample labels, so that the prediction accuracy of the trained student model can be higher.

[0067] According to an embodiment of the present invention, specifically, guided by the teacher model, a second initial classification model is trained based on multiple measured electrocardiogram data samples, multiple reference signal samples, and multiple first sample labels to obtain a trained student model, including operations 21 to 25.

[0068] Operation 21: perform feature processing on the multiple measured electrocardiogram data samples and the multiple reference signal samples to determine common features between the multiple measured electrocardiogram data samples and the multiple reference signal samples.

[0069] Operation 22: input multiple reference signal samples into the teacher model and the second initial classification model, retain the general feature channels corresponding to the general features in the teacher model, close the remaining self-owned feature channels except the feature channels, use the teacher model to output a first prediction result of whether the sample object belongs to a predetermined category, and use the second initial classification model to output a second prediction result of whether the sample object belongs to a predetermined category.

[0070] Figure 5 A schematic diagram schematically illustrates common features between measured electrocardiogram data samples and reference signal samples according to an embodiment of the present invention.

[0071] like Figure 5 As shown, the features shared by the measured electrocardiogram data and the reference signal can be common features; the features possessed by the measured electrocardiogram data samples and the reference signal samples respectively can be used as their own features.

[0072] According to an embodiment of the present invention, the teacher network has a common feature channel corresponding to the common features, and has an own feature channel corresponding to the own features.

[0073] The teacher network is pre-trained together with the second initial classification model. Three kinds of losses will be generated at this time: the loss between the prediction result of the teacher network and the true classification result; the loss between the prediction result of the second initial classification model and the true classification result; the loss between the prediction result of the teacher network and the prediction result of the second initial classification model. Through joint pre-training, the teacher network and the second initial classification model can reach the "feature alignment" state. It is possible to identify which feature dimensions have a greater impact on the "feature alignment" state through arrangement. For example, feature channels that have a greater impact on the "feature alignment" state can be assigned a larger significance value. The significance values ​​of all feature channels of the teacher network are ranked. The feature channels with higher rankings are closer to the common feature channels of the measured electrocardiogram data and the reference signal. Therefore, the features corresponding to the feature channels with higher rankings can be used as common features.

[0074] According to an embodiment of the present invention, since the electrocardiogram data and the heartbeat movement law signal data detected by the millimeter-wave radar signal have similar data distribution laws, and the measured electrocardiogram data and the millimeter-wave radar signal each have some inherent characteristics, if these inherent characteristics are involved in the training, it will affect the training effect of the student model. In the method of the present invention, in operation 22, the teacher network provides guidance based on the general features, which can improve the training effect of the student model. The general features of the teacher network can be used as knowledge and passed to the student network. When training the second initial classification model, in addition to extracting information about the heartbeat movement law of the heart part in the reference signal, the second initial classification model will also learn the knowledge of the teacher network, so as to achieve faster convergence speed and higher monitoring performance.

[0075] According to an embodiment of the present invention, the prediction accuracy of the student model can be improved by guiding the process of training the second initial classification model based on the common features between the measured electrocardiogram data and multiple reference signals.

[0076] Specifically, multiple reference signal samples can be input into the teacher model and the second initial classification model respectively, while retaining the general feature channel corresponding to the general feature in the teacher model, and closing the remaining self-owned feature channels except the feature channel. The teacher model can output a first prediction result of whether the sample object belongs to a predetermined category. The second initial classification model can output a second prediction result of whether the sample object belongs to a predetermined category.

[0077] According to an embodiment of the present invention, in the process of training the second initial classification model, two kinds of losses will be generated: the loss between the prediction results output by the teacher network and the prediction results output by the second initial classification model; the loss between the prediction results output by the second initial classification model and the true classification results of multiple sample objects. Therefore, the loss between the prediction results output by the teacher network and the prediction results output by the second initial classification model, and the loss between the prediction results output by the second initial classification model and the true classification results of multiple sample objects can be calculated by the following operations 23 to 25.

[0078] Operation 23: Using the first prediction result as a guidance label, calculate a first loss between the guidance label and the second prediction result.

[0079] According to an embodiment of the present invention, in operation 23, the first prediction result may be used as a guidance label to calculate a first loss between the guidance label and the second prediction result. The first loss may represent the loss between the prediction result output by the teacher network and the prediction result output by the second initial classification model.

[0080] Operation 24: Calculate a second loss between the first sample label and the second prediction result.

[0081] According to an embodiment of the present invention, a second loss between the first sample label and the second prediction result may be calculated in operation 24. The second loss may represent the loss between the prediction result output by the second initial classification model and the actual classification results of the plurality of sample objects.

[0082] Operation 25: Adjust the network parameters of the second initial classification model based on the first loss and the second loss to obtain a trained student model.

[0083] According to an embodiment of the present invention, when the number of reference signal samples collected is small, sample expansion can be performed by enhancing the signal. For example, the following two methods can be used.

[0084] First, the reference signal is converted into a frequency domain signal through discrete cosine transform (DCT), and then a certain number of frequency components are randomly set to zero, that is, through random dropout in the frequency domain. After that, the signal is converted into the time domain through inverse discrete cosine transform (IDCT) to construct new samples and realize the expansion construction of samples.

[0085] Second, by randomly cutting a section of the reference signal and resampling the cut signal to the original length through some data sampling methods (such as linear interpolation and filling), a new sample is constructed to achieve sample expansion construction.

[0086] According to an embodiment of the present invention, specifically, the step of constructing a teacher network includes: inputting a plurality of measured electrocardiogram data samples and a plurality of first sample labels into a first initial classification model, firstly performing feature extraction and expansion of the receptive field through a multi-scale convolution layer. Then, extracting features layer by layer through convolution layers, pooling layers, etc. Obtaining a high-dimensional feature representation in the last hidden layer of the teacher network, and using the extracted high-dimensional features for classification.

[0087] Figure 6 A schematic diagram of the network structure of a teacher model and a student model according to an embodiment of the present invention is shown schematically.

[0088] Among them, both the teacher model and the student model include Conv convolution layer, BN (Batch Normallization) standardization layer, ReLU activation function layer, Max pooling layer, etc. In addition, each also includes the Concatenation feature fusion layer, SE compression and excitation network layer, Sum Linear fully connected layer, etc. shown in the figure.

[0089] The biggest difference between the teacher model and the student model is that the number of convolution kernels in the convolution layer of the teacher model is greater than that of the student model. Since the teacher model has more convolution kernels, feature extraction is more accurate and the prediction accuracy of the teacher model is better, but it also reduces the computing efficiency to a certain extent. The student model is just the opposite. The prediction accuracy of the student model before training is worse than that of the teacher model. However, with the teacher model as a guide and the knowledge distillation method used to guide the student model, the trained student model has better prediction accuracy and higher computing efficiency.

[0090] According to an embodiment of the present invention, obtaining a plurality of reference signals for characterizing the heartbeat movement regularity of the subject includes operations 31 to 34 .

[0091] Operation 31: transmitting a millimeter-wave radar transmission signal to the object under test, and receiving a plurality of millimeter-wave radar echo signals reflected by the irradiated area of ​​the object under test.

[0092] Operation 32: performing first signal processing on the plurality of millimeter-wave radar echo signals to obtain a plurality of first-stage signals for characterizing the mechanical vibration law of the body surface of the irradiated area of ​​the measured object.

[0093] According to the embodiment of the present invention, a plurality of first-stage signals for characterizing the mechanical vibration law of the body surface of the irradiated area of ​​the measured object can be obtained through the first signal processing.

[0094] Operation 33: Perform second signal processing on the plurality of first-stage signals to eliminate the influence of body surface vibration caused by breathing of the measured object, and obtain a plurality of second-stage signals.

[0095] According to an embodiment of the present invention, in addition to the heartbeat which can cause mechanical vibrations on the body surface in the radiation area, breathing can also cause vibrations on the body surface in the radiation area. Therefore, it is necessary to perform second signal processing on multiple first-stage signals to eliminate the influence of body surface vibrations caused by the breathing of the subject.

[0096] Operation 34 : performing third signal processing on the plurality of second-stage signals to eliminate the influence of body surface vibration of non-heart parts in the irradiated area, and obtain a plurality of reference signals for characterizing the heartbeat movement law of the heart part of the measured object.

[0097] According to an embodiment of the present invention, the non-cardiac part in the irradiated area also has body surface vibration, so it is necessary to perform third signal processing on multiple second stage signals to eliminate the influence of the body surface vibration of the non-cardiac part in the irradiated area.

[0098] According to an embodiment of the present invention, by eliminating the influence of body surface vibration caused by breathing and body surface vibration in non-heart parts in the millimeter wave radar echo signal, the reference signal can more accurately characterize the law of heartbeat movement in the heart part of the measured object.

[0099] According to an embodiment of the present invention, specifically, performing the first signal processing on multiple millimeter-wave radar echo signals includes: performing a coherent accumulation operation on multiple millimeter-wave radar echo signals defined by multiple channels composed of N transmitting antennas and M receiving antennas.

[0100] According to an embodiment of the present invention, a virtual antenna array surface can be constructed based on the physical arrangement of the millimeter-wave radar antenna, and a phase shift vector can be constructed according to the antenna spacing and signal bandwidth in the virtual antenna array surface, and the spatial beamforming can be calculated to complete the spatial domain filtering of the millimeter-wave radar echo signal.

[0101] Figure 7 The schematic diagram schematically shows the spatial beamforming result during the non-cardiac micro-motion sensing of the measured object according to an embodiment of the present invention.

[0102] like Figure 7 As shown, the signal representation of the millimeter-wave radar signal reflected by the chest area of ​​the object under test is obviously different from that of the surrounding area.

[0103] According to an embodiment of the present invention, a coherent accumulation operation is performed on a plurality of millimeter-wave radar echo signals defined by a plurality of channels as shown in the following formula (1):

[0104]

[0105] in:

[0106] S i, S(x,y,z,t) is the i-th signal corresponding to the spatial position (x,y,z) and signal acquisition period t among the multiple first-stage signals obtained after the coherent accumulation operation;

[0107] N is the number of receiving antennas;

[0108] M is the number of transmitting antennas;

[0109] y n,m,t is a signal corresponding to a channel defined by an nth receiving antenna and an mth transmitting antenna in a collection period t among multiple millimeter-wave radar echo signals;

[0110] k is the rate of change of the transmission frequency of the millimeter wave radar transmission signal;

[0111] λ is the wavelength of the signal transmitted by the millimeter-wave radar;

[0112] r(x,y,z,n,m) is the round-trip distance from the mth transmitting antenna to the target location (x,y,z) and back to the nth receiving antenna;

[0113] c is the speed of light.

[0114] According to an embodiment of the present invention, specifically, performing second signal processing on a plurality of first-stage signals includes: performing second-order difference processing on the plurality of first-stage signals to eliminate the influence of body surface vibration caused by breathing of the measured object.

[0115] According to an embodiment of the present invention, by performing second-order differential processing on multiple first-stage signals, the influence of body surface vibration caused by the breathing of the subject can be eliminated, thereby amplifying multiple reference signals used to characterize the heartbeat movement pattern of the heart area of ​​the subject.

[0116] According to an embodiment of the present invention, the second-order difference processing is performed on multiple first-stage signals as shown in the following formula (2):

[0117]

[0118] in:

[0119] S i is the i-th of multiple first-stage signals;

[0120] X i , S i ″ is the value for the i-th S i The i-th second-stage signal obtained by performing second-order difference processing;

[0121] h is the period of each first-stage signal;

[0122] S i,-n For S iThe second stage signal of the previous n-th sample, n = 1, 2, 3;

[0123] S i,+n For S i The second stage signal of the nth sample is then obtained.

[0124] According to the embodiment of the present invention, specifically, performing the third signal processing on the plurality of second-stage signals includes operations 41 to 43 .

[0125] Operation 41 : Calculate the correlation between each second stage signal and each first correlation signal group, wherein the first correlation signal group is composed of a plurality of signals within a predetermined three-dimensional neighborhood range of the second stage signal.

[0126] According to an embodiment of the present invention, in order to improve the signal-to-noise ratio of the reference signal and eliminate the influence of the body surface vibration of the non-heart part in the irradiated area, the third signal processing can be performed on multiple second-stage signals. Specifically, the heartbeat can cause mechanical vibration of the body surface in a certain area, and the millimeter-wave radar echo signal in this area has a high correlation. Therefore, in operation 41, all the second-stage signals in the millimeter-wave radar scanning space can be traversed, and the correlation between each second-stage signal in the three-dimensional space and all the second-stage signals in the neighborhood can be calculated.

[0127] Operation 42 : Determine a second stage center signal from a plurality of second stage signals according to the correlation.

[0128] According to an embodiment of the present invention, the second stage signal with the highest correlation with all second stage signals in the neighborhood can be selected as the second stage center signal. The second stage center signal can be regarded as the center point of the heartbeat movement.

[0129] Operation 43: Based on the second-stage central signal and the second associated signal group, generate multiple reference signals for characterizing the heartbeat movement pattern of the subject's heart, wherein the second associated signal group consists of multiple signals within a predetermined two-dimensional neighborhood range of the second-stage central signal.

[0130] According to an embodiment of the present invention, since heartbeat motion can cause body surface motion within a certain area, the signals of all neighborhoods within a certain distance unit of the center point of the heartbeat motion can be used as reference signals. Specifically, multiple second-stage signals within a two-dimensional neighborhood (within a two-dimensional plane parallel to the human chest cavity) of the second-stage center signal can be used as a second associated signal group. The second-stage center signal and the second associated signal group can be used as multiple reference signals to characterize the law of heartbeat motion in the heart of the subject being measured.

[0131] According to an embodiment of the present invention, heartbeats can cause mechanical vibrations on the body surface in a certain area, and the millimeter-wave radar echo signals in this area have a high correlation. Therefore, by calculating the correlation between each second-stage signal and each first-related signal group, multiple reference signals that more accurately characterize the heartbeat movement law of the heart part of the measured object can be obtained. This processing method can accurately eliminate the influence of body surface vibrations in non-heart parts in the irradiated area, and obtain a signal for characterizing the heartbeat movement law of the heart part of the measured object.

[0132] According to an embodiment of the present invention, the correlation between each second stage signal and each first associated signal group is calculated as shown in the following formula (3):

[0133]

[0134] in:

[0135] X i is the i-th second-stage signal, and m is the number of second-stage signals;

[0136] Y i is any one of a plurality of signals within a predetermined three-dimensional neighborhood of the i-th second stage signal;

[0137] cov(X i ,Y i ) is X i and Y i The covariance between

[0138] σ is the standard deviation.

[0139] The present invention is further described below by embodiments and related test experiments. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present invention. However, it is obvious that one or more embodiments may also be implemented without these specific details. And, in the absence of conflict, the details in the following embodiments may be combined into other feasible embodiments at will.

[0140] Example 1

[0141] The atrial fibrillation classification prediction method based on millimeter wave radar signals of the present invention was used to classify and predict atrial fibrillation in 882 subjects in a clinical environment (851 of whom belonged to the predetermined category and 31 did not belong to the predetermined category). The specific test conditions are: the subject was in a sleeping position and the distance from the millimeter wave radar was 0.5 meters. The millimeter wave radar was in a 3-transmit 4-receive working state. The starting frequency of the millimeter wave radar signal was 77 GHz, the bandwidth was 4 GHz, and the frame rate was 100 Hz.

[0142] Figure 8The test environment in Embodiment 1 of the present invention is schematically shown.

[0143] The prediction effect of the atrial fibrillation classification prediction method based on millimeter-wave radar signals of the present invention can be evaluated based on accuracy, F1 score, sensitivity, and specificity.

[0144] The calculation method of F1 score is shown in the following formula (4):

[0145]

[0146] Among them, TP is the true positive sample, that is, the number of positive samples predicted by the student model and the actual classification result of the subject is also the positive sample; FP is the false positive sample, that is, the number of positive samples predicted by the student model and the actual classification result of the subject is the negative sample, also known as the number of false positive negative samples; TN is the true negative sample, that is, the number of negative samples predicted by the student model and the actual classification result of the subject is also the negative sample; FN is the false negative sample, that is, the number of negative samples predicted by the student model and the actual classification result of the subject is the positive sample, also known as the number of false negative positive samples. Among them, the positive sample indicates that the subject belongs to the predetermined category, and the negative sample indicates that the subject does not belong to the predetermined category.

[0147] The calculation method of sensitivity is shown in the following formula (5):

[0148]

[0149] Wherein, Sensitivity is sensitivity, and the definitions of TP and FN are as described above and will not be repeated here.

[0150] The specificity is calculated as shown in the following formula (6):

[0151]

[0152] Among them, Specificity is specificity, and the definitions of TP and FP are as described above and will not be repeated here.

[0153] The results of AF classification prediction for 882 subjects in a clinical setting are shown in Table 1 .

[0154] Table 1

[0155] Accuracy F1 score Sensitivity Specificity True negative example False Positive False Negative True example result 0.9909 0.8788 0.9355 0.9929 845 6 2 29

[0156] As can be seen from Table 1, the main evaluation index F1 score is 0.8788, indicating that the atrial fibrillation classification prediction method based on millimeter wave radar signals of the present invention has a good classification prediction performance.

[0157] Fig. 9A box plot schematically shows the distribution of prediction results obtained by using the prediction method of the present invention for multiple subjects.

[0158] The above prediction results are obtained based on reference signals corresponding to 2646 10-second cardiac mechanical movements of 882 subjects.

[0159] like Fig. 9 As shown, the dotted line represents the threshold for object category judgment. The prediction score of patients with atrial fibrillation is basically higher than the threshold, and the prediction score of patients without atrial fibrillation is lower than the threshold. There is a significant difference in the output prediction scores of patients with atrial fibrillation and those without atrial fibrillation, which further shows that the classification prediction performance of the atrial fibrillation classification prediction method based on millimeter wave radar signals of the present invention is good.

[0160] In addition, experimental verification has shown that the method of the embodiment of the present invention has for the first time achieved high-precision identification of a predetermined category (atrial fibrillation) from 34 common categories, including arrhythmias such as atrial premature beats and ventricular premature beats, cardiac abnormalities such as bundle branch block and atrioventricular block, and normal sinus rhythm, with high recognition accuracy.

[0161] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for classifying and predicting atrial fibrillation based on millimeter wave radar signals, comprising: Acquire multiple reference signals for characterizing the heartbeat movement law of the heart of the measured object, wherein the multiple reference signals are obtained by processing multiple original signals, and the multiple original signals are: multiple millimeter wave radar echo signals reflected by the irradiated area of ​​the measured object; Inputting the multiple reference signals into a trained student model, and outputting a prediction result of whether the object under test belongs to a predetermined category; The student model is trained using the following method: Acquire a plurality of measured electrocardiogram data samples for a plurality of sample objects, a plurality of reference signal samples for characterizing the heartbeat movement regularity of the heart parts of the plurality of sample objects, and a plurality of first sample labels for characterizing the true classification results of the plurality of sample objects; Using a plurality of measured electrocardiogram data samples and a plurality of first sample labels to train a first initial classification model to obtain a trained teacher model; Using the teacher model as a guide, training a second initial classification model based on a plurality of measured electrocardiogram data samples, a plurality of reference signal samples, and a plurality of first sample labels to obtain a trained student model; Wherein, guided by the teacher model, a second initial classification model is trained based on multiple measured electrocardiogram data samples, multiple reference signal samples, and multiple first sample labels, and the trained student model is obtained, including: Performing feature processing on a plurality of measured electrocardiogram data samples and a plurality of reference signal samples to determine common features between the plurality of measured electrocardiogram data samples and the plurality of reference signal samples; Input the plurality of reference signal samples into the teacher model and the second initial classification model, retain the common feature channel corresponding to the common feature in the teacher model, close the remaining self-owned feature channels except the feature channel, use the teacher model to output a first prediction result of whether the sample object belongs to a predetermined category, and use the second initial classification model to output a second prediction result of whether the sample object belongs to a predetermined category; Using the first prediction result as a guidance label, calculating a first loss between the guidance label and the second prediction result; Calculating a second loss between the first sample label and the second prediction result; The network parameters of the second initial classification model are adjusted based on the first loss and the second loss to obtain a trained student model.

2. The method according to claim 1, wherein: Acquiring multiple reference signals for characterizing the heartbeat movement law of the subject includes: Transmitting a millimeter-wave radar transmission signal to the object under test, and receiving a plurality of millimeter-wave radar echo signals reflected by the irradiated area of ​​the object under test; Performing first signal processing on the multiple millimeter wave radar echo signals to obtain multiple first stage signals for characterizing the mechanical vibration law of the body surface of the irradiated area of ​​the measured object; Performing second signal processing on the plurality of first-stage signals to eliminate the influence of body surface vibration caused by breathing of the measured object, and obtaining a plurality of second-stage signals; The third signal processing is performed on the multiple second stage signals to eliminate the influence of the body surface vibration of the non-heart part in the irradiated area, and obtain multiple reference signals for characterizing the heartbeat movement law of the heart part of the measured object.

3. The method according to claim 2, wherein: Performing a first signal processing on the multiple millimeter-wave radar echo signals includes: A coherent accumulation operation is performed on the multiple millimeter-wave radar echo signals defined by multiple channels composed of M transmitting antennas and N receiving antennas.

4. The method according to claim 3, wherein: The coherent accumulation operation on the multiple millimeter-wave radar echo signals defined by the multiple channels includes adopting the following method: in: , The i-th signal corresponding to the spatial position (x, y, z) and the signal acquisition period t among the multiple first-stage signals obtained after the coherent accumulation operation; N is the number of receiving antennas; M is the number of transmitting antennas; is a signal corresponding to a channel defined by an nth receiving antenna and an mth transmitting antenna in a collection period t among multiple millimeter-wave radar echo signals; k is the rate of change of the transmission frequency of the millimeter wave radar transmission signal; is the wavelength of the signal transmitted by the millimeter wave radar; is the round-trip distance from the mth transmitting antenna to the target location (x, y, z) and back to the nth receiving antenna; The speed of light.

5. The method according to claim 2, wherein: Performing second signal processing on the plurality of first-stage signals includes: The plurality of first-stage signals are subjected to second-order difference processing to eliminate the influence of body surface vibration caused by the breathing of the measured object.

6. The method according to claim 5, wherein: Performing second-order difference processing on the plurality of first-stage signals includes: = in: is the i-th of multiple first-stage signals; , For the i-th The i-th second-stage signal obtained by performing second-order difference processing; h is the period of each first-stage signal; for The second-stage signal of the previous n-th sample, n=1, 2, 3; for The second stage signal of the nth sample after that.

7. The method according to claim 2, wherein: Performing third signal processing on the plurality of second-stage signals includes: Calculating the correlation between each second-stage signal and each first associated signal group, wherein the first associated signal group is composed of a plurality of signals within a predetermined three-dimensional neighborhood of the second-stage signal; determining a second-stage center signal from the plurality of second-stage signals according to the correlation; Based on the second-stage central signal and the second associated signal group, multiple reference signals are generated for characterizing the heartbeat movement pattern of the subject's heart, wherein the second associated signal group consists of multiple signals within a predetermined two-dimensional neighborhood range of the second-stage central signal.

8. The method according to claim 7, wherein: Calculating the correlation between each second stage signal and each first correlation signal group includes: in: is the i-th second-stage signal, L is the number of second-stage signals; is any one of a plurality of signals within a predetermined three-dimensional neighborhood of the i-th second stage signal; )for and The covariance between is the standard deviation.

Images