Non-contact arrhythmia detection method and system and storage medium

Abstract

The invention provides a non-contact arrhythmia detection method and system and a storage medium, and the method comprises the steps: obtaining a heart micro-motion signal of a patient obtained through the detection of a radar system, and carrying out the time-domain phase feature, the symbolic dynamics feature and the motion vector feature of the heart micro-motion signal based on the time-domain phase feature, the symbolic dynamics feature and the motion vector feature of the heart micro-motion signal; and obtaining input of a pre-trained disease diagnosis model based on the time domain phase feature, the symbolic dynamics feature and the motion vector feature, inputting the input into the disease diagnosis model, and outputting to obtain an arrhythmia detection result of the patient. According to the invention, high-frequency heart motion characteristics can be captured, a non-contact detection mode is adopted, and high detection precision and user comfort can be achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular, to a non-contact arrhythmia detection method and system. Background Art

[0002] Arrhythmia phenomena such as atrial fibrillation and ventricular premature contractions seriously affect the health of patients, and timely detection helps to maintain heart health. Existing arrhythmia detection means are mainly limited to hospital examinations, such as electrocardiogram (ECG) and echocardiogram, as well as wearable devices such as smart watches.

[0003] However, the existing detection schemes have the following defects: (1) The detection frequency of signal waves is relatively low, and the traditional detection wave frequency (0.5 - 40 Hz) cannot capture the characteristics of high-frequency mechanical movements. For example, the current detection method is difficult to detect paroxysmal atrial fibrillation, which easily leads to low detection accuracy; (2) The current detection method mostly adopts a contact detection method. For example, when a patient undergoes an ECG test, electrodes need to be attached, resulting in poor comfort and compliance, and being easily affected by skin impedance.

[0004] Therefore, there is an urgent need for a brand-new arrhythmia detection method that can not only capture the high-frequency cardiac movement characteristics to improve the detection accuracy, but also improve the user comfort. Summary of the Invention

[0005] In view of this, the embodiments of the present invention provide a non-contact arrhythmia detection method and system, which can not only capture the high-frequency cardiac movement characteristics, but also adopt a non-contact detection method, with high detection accuracy and user comfort.

[0006] One aspect of the present invention provides a non-contact arrhythmia detection method, which includes the following steps:

[0007] Obtain the cardiac micro-motion signal of a patient detected by a radar system, and obtain the time-domain phase feature, symbol dynamics feature, and motion vector feature based on the cardiac micro-motion signal;

[0008] Based on the time-domain phase feature, symbol dynamics feature, and motion vector feature, obtain the input of a pre-trained disease diagnosis model, input it into the disease diagnosis model, and output the arrhythmia detection result of the patient;

[0009] Wherein, the time-domain phase feature and the motion vector feature are respectively obtained through the following methods:

[0010] Divide the cardiac micro-motion signal into two orthogonal signals, obtain a phase signal based on the phase information of the two orthogonal signals, and perform Fourier transform on the phase signal to obtain the time-domain phase feature; and

[0011] Based on a biomechanical non - linear deformation model, a non - rigid displacement component is decomposed from the cardiac micro - motion signal, and a non - rigid stress tensor is calculated based on the non - rigid displacement component and the visco - elastic constitutive equation of myocardial tissue, so as to obtain a motion vector feature based on the non - rigid stress tensor.

[0012] In some embodiments of the present invention, after acquiring the cardiac micro - motion signal, the method further includes: constructing a cardiac dynamics model based on the cardiac micro - motion signal;

[0013] Constructing a cardiac dynamics model based on the cardiac micro - motion signal includes:

[0014] By finite - element mesh generation, the cardiac surface is discretized into multiple nodes, and the phase change of each node is calculated using the transmitted signal of the radar system and the cardiac micro - motion signal;

[0015] Based on the phase change of each node and the extended Kalman filter, the displacement vector of each node is estimated, so as to construct a cardiac dynamics model.

[0016] In some embodiments of the present invention, calculating the non - rigid stress tensor based on the non - rigid displacement component and the visco - elastic constitutive equation of myocardial tissue includes:

[0017] In the cardiac dynamics model, the difference between the non - rigid displacement components of adjacent nodes on the cardiac surface is calculated to obtain a local strain tensor, and the time derivative of the local strain tensor is taken to obtain a strain rate tensor;

[0018] Determine the relaxation modulus based on biomechanical experiments or inverse - problem solving methods;

[0019] Using the strain rate tensor, the relaxation modulus, and the visco - elastic constitutive equation of myocardial tissue, determine the stress tensor in the time domain.

[0020] In some embodiments of the present invention, the time - domain phase feature is obtained by the following method:

[0021] The cardiac micro - motion signal is divided into two orthogonal signals, and the arctangent function of the ratio of the two orthogonal signals is calculated to obtain a phase signal;

[0022] Take the first - order derivative of the phase signal, and obtain the time - domain phase feature based on the result of the first - order derivative and the signal sampling rate.

[0023] In some embodiments of the present invention, obtaining the input of the pre - trained disease diagnosis model based on the time - domain phase feature, symbolic dynamics feature, and motion vector feature includes:

[0024] The local waveform features are obtained from the time-domain phase features using a one-dimensional convolutional neural network model, the temporal dependence features of the symbol sequence are obtained from the symbolic dynamics features using a long short-term memory network model, and the multi-scale deformation features are obtained from the motion vector features using a graph diffusion network model;

[0025] Fuse the local waveform features, temporal dependence features and multi-scale deformation features, and obtain the input of the pre-trained disease diagnosis model based on the fused mechanical features.

[0026] In some embodiments of the present invention, before fusing the local waveform features and the temporal dependence features, the method further includes: using the JS divergence constraint to align the feature spaces of the local waveform features and the temporal dependence features;

[0027] Fusing the local waveform features, temporal dependence features and multi-scale deformation features includes:

[0028] Adaptively adjust the contribution weights of the local waveform features, temporal dependence features and multi-scale deformation features to the detection results according to the signal-to-noise ratio, and fuse the local waveform features, temporal dependence features and multi-scale deformation features in a weighted fusion manner.

[0029] In some embodiments of the present invention, obtaining the input of the pre-trained disease diagnosis model based on the fused mechanical features includes:

[0030] Based on the pre-trained cross-modal diffusion model or the myocardial displacement-electrical conduction coupling equation, map the fused mechanical features to electrophysiological signals, and use the electrophysiological signals as the input of the pre-trained disease diagnosis model.

[0031] In some embodiments of the present invention, the cardiac micro-motion signal is obtained by signal preprocessing of the original signal collected by the radar system, and the signal preprocessing process includes:

[0032] Input the original signal collected by the radar system into a high-order anti-aliasing filter and a clutter suppressor for filtering;

[0033] Perform dynamic beamforming processing and phase compensation on the filtered signal to obtain a noise-reduced signal;

[0034] Optimize the noise-reduced signal through the beamforming algorithm that maximizes the autocorrelation coefficient and the compressive sensing technology to obtain the cardiac micro-motion signal.

[0035] Another aspect of the present invention provides a non-contact arrhythmia detection system, including a processor, a memory, and a computer program / instruction stored on the memory, and the processor is used to execute the computer program / instruction, and when the computer program / instruction is executed, the system implements the steps of the method described in any one of the above embodiments.

[0036] Another aspect of the present invention provides a computer-readable storage medium, on which computer programs / instructions are stored, and when the computer programs / instructions are executed by a processor, the steps of the method described in any of the above embodiments are implemented.

[0037] The non-contact arrhythmia detection method and system proposed by the present invention can utilize the non-contact sensing ability of the radar system to sense weak human motion signals, and extract multi-modal features of cardiac activity by analyzing the motion signals, so as to realize the detection and identification of arrhythmias. This application can not only capture high-frequency cardiac motion features, but also adopt a non-contact detection method, with high detection accuracy and user comfort.

[0038] Additional advantages, objects, and features of the present invention will be partially described below, and will become partially apparent to those of ordinary skill in the art after studying the following text, or can be learned from the practice of the present invention. The objects and other advantages of the present invention can be achieved and obtained by the structure specifically pointed out in the specification and the drawings.

[0039] Those skilled in the art will understand that the objects and advantages that can be achieved by the present invention are not limited to the above specifically described, and the above and other objects that the present invention can achieve will be more clearly understood according to the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not limit the present invention. In the drawings:

[0041] Figure 1 It is a schematic flowchart of a non-contact arrhythmia detection method in an embodiment of the present invention.

[0042] Figure 2 It is a schematic flowchart of a non-contact arrhythmia detection method in another embodiment of the present invention.

[0043] Figure 3 It is a schematic flowchart of constructing a cardiac dynamics model in an embodiment of the present invention.

[0044] Figure 4 It is a schematic flowchart of fusing features and obtaining an arrhythmia detection result in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] To make the objects, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the embodiments and the drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but do not limit the present invention.

[0046] Here, it should also be noted that in order to avoid obscuring the present invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the present invention are shown in the drawings, while other details less relevant to the present invention are omitted.

[0047] It should be emphasized that the term "comprising / including" as used herein refers to the presence of features, elements, steps, or components, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0048] Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0049] Most existing solutions for detecting a patient's arrhythmia adopt contact detection, which has the disadvantages of poor comfort and being easily affected by skin impedance. Moreover, a detection wave with a relatively low frequency is used to capture the heart movement to generate an electrocardiogram signal, and it is easy to miss a diagnosis or make a misdiagnosis. That is, the current arrhythmia detection method not only has a relatively low detection accuracy but also poor comfort.

[0050] Based on this, the present application proposes a method for detecting arrhythmia based on a radar system. As Figure 1 shown, this method can utilize the non-contact sensing ability of the radar system to sense weak human movement signals (such as minute deformations of the chest caused by heartbeats). By analyzing these signals, multi-modal features of heart activities can be extracted, and an artificial intelligence algorithm is used to achieve automatic detection and identification of arrhythmia.

[0051] Figure 2 It is a schematic flowchart of a non-contact arrhythmia detection method according to an embodiment of the present application. As Figure 2 shown, this method includes steps S110 to S120.

[0052] Step S110: Obtain a heart micro-motion signal and perform feature extraction. Step S110 can be divided into the following steps: Step S111: Obtain the patient's heart micro-motion signal detected by the radar system; Step S112: Extract features in the time domain, frequency domain, and spatial domain from the heart micro-motion signal to obtain time-domain phase features, symbolic dynamics features, and motion vector features respectively.

[0053] The cardiac micro-motion signal can be used to characterize the minute mechanical motion in the patient's cardiac region, and the waveform of the signal can be presented in the form of time-cardiac displacement as the horizontal and vertical coordinates. Since the cardiac motion is relatively weak, to improve the resolution, this application can transmit a signal for detecting the patient's cardiac motion through a radar system and receive the cardiac micro-motion signal carrying the motion characteristics of the human cardiac region. By using a radar system to detect the cardiac micro-motion signal, this application can not only detect using high-frequency signals, but also the non-contact detection can provide a more comfortable user experience.

[0054] As an example, traditional ECG relies on contact electrodes to capture myocardial electrical conduction signals (0.05 - 100 Hz), and there is a problem of signal attenuation caused by changes in skin impedance; while this application can use a frequency-modulated continuous wave (FMCW) millimeter-wave radar (76 - 81 GHz band), by measuring the micro-Doppler effect caused by myocardial mechanical motion (for example, the cardiac displacement accuracy determined by the millimeter-wave signal can reach 50 μm), the cardiac dynamics model can be reconstructed and arrhythmia detection can be performed, breaking through the physical limitations of traditional electrophysiological detection. The above-mentioned selection of millimeter-wave radar to collect cardiac micro-motion signals is only an example, and this application can also use other radar systems capable of receiving high-frequency signal waves to collect human motion signals, such as lidar or infrared radar.

[0055] Furthermore, in step S111, the cardiac micro-motion signal can be the original signal detected by the radar system, or the cardiac micro-motion signal can be the signal after the original signal detected by the radar system undergoes signal preprocessing. Considering that there are various interferences in the process of the radar system transmitting and receiving signals, in addition to common electromagnetic interference, there may also be a problem of inability to focus on the patient's cardiac region (that is, in addition to the micro-motion signals in the cardiac region, there may also be micro-motion signals in other chest regions, causing interference). Therefore, this application usually needs to perform filtering, noise reduction, and optimization processing on the original signal detected by the radar system, and further perform feature extraction steps on the cardiac micro-motion signal obtained after signal preprocessing. Moreover, in the process of signal preprocessing, the execution order of the filtering, noise reduction, and optimization processing steps is not specifically limited in this application. For example, after noise reduction processing, filtering and optimization processing can be performed on the signal.

[0056] In some embodiments of the present invention, the signal preprocessing process includes: inputting the raw signal collected by the radar system (i.e., the raw cardiac micro-motion signal of the patient) into a high-order anti-aliasing filter and a clutter suppressor for filtering; wherein, the clutter suppressor can effectively separate the cardiac micro-motion signal of 0.1 - 5 Hz from the respiratory harmonics of 30 - 60 times per minute; performing dynamic beamforming processing (the signal-to-noise ratio can be increased by about 15 dB) and phase compensation (used to eliminate the carrier frequency offset caused by radar local oscillator leakage and reduce phase distortion) on the filtered signal to obtain a noise-reduced signal; optimizing the noise-reduced signal through the beamforming algorithm that maximizes the autocorrelation coefficient and the compressive sensing technology to obtain the cardiac micro-motion signal.

[0057] Specifically, the present application can use a high-order anti-aliasing filter and a clutter suppressor (the clutter suppressor can be constructed based on a cross-domain diffusion model or empirical mode decomposition) for filtering, combine the dynamic beamforming algorithm and phase compensation technology for noise reduction, and use the beamforming algorithm that maximizes the autocorrelation coefficient and the compressive sensing technology for optimization. That is, the present application can perform preprocessing of the raw signal from multiple dimensions, specifically as follows: Compared with the traditional ECG signal preprocessing method of suppressing baseline drift and electromyographic noise, the filtering process of the present application needs to match the characteristics of high-frequency signals and enhance the anti-interference ability. Therefore, for the high-frequency characteristics of radar signals, the present application uses a high-order anti-aliasing filter to eliminate signal aliasing and uses a clutter suppression algorithm to suppress environmental electromagnetic interference to reduce the influence of noise in the Wi-Fi or Bluetooth frequency bands, etc. During the noise reduction process, the dynamic beamforming algorithm is used in the spatial dimension to construct an adaptive cardiac motion tracking beam to achieve spatial filtering, so as to suppress the multipath effect caused by thoracic breathing or limb movement, thereby focusing on the signal in the cardiac region, and in the time dimension, the phase unwrapping technology can be used to eliminate signal phase ambiguity. In the signal optimization stage, the beamforming algorithm that maximizes the autocorrelation coefficient can be used to accurately locate the cardiac motion region, and the compressive sensing technology is applied to optimize the signal reconstruction efficiency.

[0058] As an example, during the filtering process, the high-order anti-aliasing filter can be an existing Butterworth filter, Chebyshev filter or elliptic filter, or a self-designed filter. For example, during the noise reduction process, the long short-term memory network (LSTM) can be used to learn the cardiac cycle pattern and predict the cardiac position θ at the next moment. t, thereby dynamically adjusting the beam pointing angle to achieve a beam tracking accuracy at the 0.1° level. For example, the present application can adopt a double-loop phase-locked structure to construct a phase compensator to achieve more precise frequency and phase control. Among them, the inner loop uses a Costas loop to eliminate the carrier frequency offset, and the outer loop compensates for the phase ambiguity caused by movement through the cross-correlation of reference signals, and can greatly reduce the phase noise by correcting the phase error (for example, the phase noise can be reduced to 0.05 rad RMS). For another example, when optimizing the signal in the present application, the method of optimizing the signal using the beamforming algorithm that maximizes the autocorrelation coefficient can be the same as the prior art, that is, it includes steps such as constructing a signal model, calculating the beamforming output, determining the optimization objective, and obtaining the optimal weight vector by solving the optimization problem, or it can also be different from the prior art. For example, in the optimization objective design and solution stage, a cost function can be constructed. (where W is the weight vector and R is the autocorrelation matrix of the desired signal), and the optimal weight vector is solved by maximizing the generalized Rayleigh quotient, so that the autocorrelation coefficient of the optimized signal reaches the peak within the cardiac rhythm period (0.5 - 5 Hz).

[0059] In the signal preprocessing process, the above-mentioned filters or algorithms are only examples, and the present application can also adopt other signal processing devices or algorithms, and the present invention is not limited thereto.

[0060] In some embodiments of the present invention, after step S111, the method further includes step S113: constructing a cardiac dynamics model based on the cardiac micro-motion signal. The process of constructing the cardiac dynamics model in step S113 can be realized by a finite element model. As Figure 3 shown, the specific steps are as follows:

[0061] Step S1131: Discretize the cardiac surface into multiple nodes through finite element mesh generation.

[0062] Step S1132: Calculate the phase change of each node using the transmitted signal of the radar system and the cardiac micro-motion signal. In this step, the phase change of each node is the difference between the phase of the transmitted signal of the radar system and the phase of the cardiac micro-motion signal for each node on the cardiac surface, which can be expressed by the formula △φ i = φ i,发射 - φ i,微动信号 , where △φ i represents the phase change of node i, and φ i,发射 and φ i,微动信号 represent the transmitted signal phase of node i and the phase of the cardiac micro-motion signal of node i, respectively.

[0063] Step S1133: Based on the phase changes of each node and the Extended Kalman Filter (EKF), estimate the displacement vector di=(dix, diy, diz) of each node, thereby constructing a cardiac dynamics model.

[0064] As an example, after step S1133, the method further includes step S1134: iteratively update the cardiac dynamics model, specifically: use the Extended Kalman Filter to fuse the displacement observation values (determine the displacement values of each node on the cardiac surface by the signals received during continuous detection by the radar) and the model prediction values (the displacement prediction values of the EKF) in real time, iteratively optimize the displacement vector of each node, realize real-time tracking of the deformation process, and improve the dynamic response accuracy.

[0065] In the in vitro porcine heart experiment, the cardiac dynamics model constructed in this application can detect local myocardial abnormal bulges (simulating the mechanical activity of atrial fibrillation) at the 0.2 mm level, which is better than the minimum detection threshold (0.5 mm) of traditional methods. And the clinical test data shows that for atrial fibrillation, the capture success rate of the cardiac micro-motion signal is increased from "68% of the traditional radar" to 93%, and the false alarm rate is decreased to 2.1%.

[0066] Step S112: Extract the mechanical features of cardiac activity from the cardiac micro-motion signal, including time-domain phase features, symbolic dynamics features, and motion vector features.

[0067] Traditional methods mostly use time-frequency analysis (such as wavelet transform) or morphological features (such as the standard deviation of the RR interval), and have the following limitations: limited time-frequency resolution, difficult to capture electrophysiological abnormalities at the microsecond level; sensitive to noise (such as electromyogram interference); ignoring the phase information of the original signal in the IQ domain (such as non-linear changes in the quadrature phase); and mostly using single features (such as the RR interval), without realizing multi-modal fusion. In view of this, this application not only proposes to fuse multi-modal features, but also proposes new feature extraction methods in the time-frequency dimension and the space dimension to improve the resolution. In addition, this application also introduces a deep learning model to extract weak cardiac mechanical motion features from complex signals, breaking through the limitations of traditional ECG relying on wavelet transform or empirical mode decomposition to separate motion artifacts.

[0068] (1) For the time-domain dimension: This application can decompose the cardiac micro-motion signal received by the radar system into the I channel and the Q channel (the I channel and the Q channel signals are orthogonal signals), and use dual-channel differential phase analysis and Short-Time Fourier Transform (STFRFT) to realize time-frequency analysis, thereby extracting the chaotic phase features caused by arrhythmia. Specifically, the formula for extracting the time-domain phase features from the cardiac micro-motion signal can be expressed as:

[0069]

[0070] where, φ d (t) represents the time-domain phase feature at time t, and f s is the signal sampling rate.

[0071] The extraction formula of the above time-domain phase feature is proposed based on the complex signal (I + jQ) and the Fourier transform. Through the Fourier transform, the time-domain I / Q signal can be converted into the frequency domain to analyze the spectral characteristics of the signal (such as Doppler frequency shift). Therefore, this application can also be regarded as extracting the time-domain phase feature through the following process: dividing the cardiac micro-motion signal into two orthogonal signals, obtaining the phase signal based on the phase information of the two orthogonal signals, and performing the Fourier transform on the phase signal to obtain the time-domain phase feature. Among them, obtaining the phase signal based on the phase information of the two orthogonal signals and performing the Fourier transform on the phase signal to obtain the time-domain phase feature includes: dividing the cardiac micro-motion signal into two orthogonal signals, calculating the arctangent function of the ratio of the two orthogonal signals to obtain the phase signal; performing the short-time fractional Fourier transform on the phase signal, and adjusting the fractional-order parameter α to obtain the time-domain representation (for example, by rotating the analysis signal in the time-frequency plane and separating different motion modes); extracting the short-time phase change feature from the time-domain representation (for example, by peak detection or energy focusing to quantify the instantaneous frequency offset of cardiac motion), and using it as the time-domain phase feature.

[0072] Considering that the cardiac micro-motion signal has non-stationary characteristics (such as transient abnormalities during atrial fibrillation), and STFRFT can more flexibly match the time-frequency distribution of non-stationary signals and improve the detection accuracy of weak phase changes. Therefore, the scheme for processing signals in the time-domain dimension proposed in this application can directly extract the phase change rate feature, significantly improving the processing accuracy of complex non-stationary signals, and its technical complexity far exceeds the linear filtering process of traditional ECG.

[0073] (2) For the frequency-domain dimension: Using multi-scale symbolic dynamics operations, extract the symbolic dynamics features from the cardiac micro-motion signal to capture the non-linear dynamics characteristics. The process of performing symbolic dynamics operations in this application can be the same as that of the prior art, that is, it includes three steps: symbolization processing, multi-scale decomposition, and feature extraction. Specifically, the cardiac micro-motion signal can be decomposed into orthogonal I-channel and Q-channel signals; the amplitude of the I-channel or Q-channel signal is quantized into multiple symbols to form a symbol sequence; the transfer law of the symbol sequence is statistically analyzed to generate a transition probability matrix. For example, quantizing the amplitude of the I-channel or Q-channel sub-signal to obtain 4 symbols {A, B, C, D}, statistically analyzing the transfer law of the symbol sequence with a length of 3 (such as the symbol sequence "A→B→C") and forming a 16×4-dimensional transition probability matrix.

[0074] (3) For the airspace dimension: Based on the non - linear deformation model of biomechanics, the non - rigid displacement component is decomposed from the cardiac micro - motion signal, and the non - rigid stress tensor (also known as the stress distribution tensor) is calculated based on the non - rigid displacement component and the visco - elastic constitutive equation of myocardial tissue, so as to obtain the motion vector feature based on the stress tensor.

[0075] More specifically, traditional ECG relies on 12 - lead potential difference analysis and cannot detect mechanical asynchrony (such as left bundle branch block accompanied by ventricular asynchrony); while this application can quantify the mechanical delay between ventricles through three - dimensional motion vector field analysis. For example, local systolic abnormalities of premature ventricular contractions can be identified through the analysis of the curvature of the motion trajectory (usually, the curvature mutation of cardiac motion > 35%). The process of extracting features in the airspace dimension of this application is as follows: Based on the non - linear deformation model of biomechanics, cardiac motion can be decomposed into a rigid displacement component and a non - rigid deformation component. Among them, the rigid displacement component can be used to represent the overall position change of the patient's body, such as the thoracic motion caused by breathing, and the non - rigid displacement component can be used to represent the position change caused by local contraction or dilation of the myocardium. Filtering out the rigid displacement component can improve the signal - to - noise ratio. After calculating the non - rigid stress tensor using the non - rigid displacement component, the non - rigid stress tensor can be used as the motion vector feature, or a deep learning network can be used to extract the motion vector feature from the non - rigid stress tensor for subsequent feature joint analysis. In addition, the non - linear deformation model of biomechanics mentioned in this application can be a non - linear anisotropic material model, a non - linear finite - element model, or a non - linear elastic model, etc., and the visco - elastic constitutive equation of myocardial tissue can also be constructed based on existing non - linear visco - elastic models such as the generalized Maxwell model or the distributed visco - elastic model. The present invention does not specifically limit the non - linear deformation model of biomechanics and the visco - elastic constitutive equation of myocardial tissue, and can be selected according to requirements.

[0076] As an example, this application establishes the visco - elastic constitutive equation of myocardial tissue with the following formula to describe the dynamic deformation response of the myocardium under the action of millimeter - wave:

[0077]

[0078] Among them, ∈(τ) is the strain - rate tensor, σ(t) is the stress tensor, and G(t) is the relaxation modulus, which is used to describe the visco - elastic properties of myocardial materials.

[0079] In some embodiments of the present invention, calculating the non-rigid stress tensor based on the non-rigid displacement components and the viscoelastic constitutive equation of myocardial tissue includes: in the cardiac dynamics model, calculating the difference in non-rigid displacement components between adjacent nodes on the cardiac surface to obtain the local strain tensor, and performing time differentiation on the local strain tensor to obtain the strain rate tensor; determining the relaxation modulus based on biomechanical experiments or inverse problem solving methods; using the strain rate tensor, the relaxation modulus, and the viscoelastic constitutive equation of myocardial tissue to determine the stress tensor in the time domain. Among them, the process of determining G(t) through experimental calibration is: based on biomechanical experimental data (such as in vitro myocardial stretching tests), preset the mathematical form of G(t) (such as an exponential decay model), and the process of determining G(t) through inverse problem solving is: input the cardiac micro-motion signal containing displacement data measured by the radar system into the constitutive equation, and combine an optimization algorithm (such as the least squares method) to inversely deduce the parameters of G(t). That is, after calculating the strain rate tensor and the relaxation modulus according to the cardiac micro-motion signal, the non-rigid stress tensor can be calculated using the above constitutive equation. The above-mentioned methods for calculating the strain rate tensor and the relaxation modulus are only examples, and other existing calculation methods can also be used to calculate the strain rate tensor and the relaxation modulus, and the present invention is not limited thereto.

[0080] Step S120: This application can analyze by combining multi-modal information such as time-domain phase change and frequency-domain Doppler characteristics to achieve precise detection of the patient's cardiac motion. Specifically, based on the time-domain phase characteristics, symbolic dynamics characteristics, and motion vector characteristics, the input of the pre-trained disease diagnosis model is obtained; inputting it into the disease diagnosis model can output the detection result of the patient's arrhythmia.

[0081] Obtaining the input of the pre-trained disease diagnosis model based on the time-domain phase characteristics, symbolic dynamics characteristics, and motion vector characteristics includes: using the time-domain phase characteristics, symbolic dynamics characteristics, and motion vector characteristics as the input data set to input into the pre-trained disease diagnosis model; or, fusing the time-domain phase characteristics, symbolic dynamics characteristics, and motion vector characteristics, and using the fused characteristics as the input of the pre-trained disease diagnosis model.

[0082] In some embodiments of the present invention, since the features in three different dimensions extracted from the cardiac micro-motion signals have relatively low information abundance, this application can utilize a multi-branch encoder to achieve the condensation and fusion of multi-modal features. In the time-domain branch, symbolic dynamics branch, and motion branch, the time-domain phase features, symbolic dynamics features, and motion vector features can be encoded respectively. The fused encoded features can obtain the fused encoded mechanical features. The process of encoding and fusing multi-modalities using a multi-branch encoder is as follows: Use a one-dimensional convolutional neural network model (i.e., 1D CNN model) to obtain local waveform features (also referred to as local waveform encoded features) from the time-domain phase features, use a long short-term memory network model (i.e., LSTM model) to obtain the temporal dependence features of the symbolic sequence (also referred to as temporal dependence encoded features) from the symbolic dynamics features, and use a graph diffusion network model to obtain multi-scale deformation features (also referred to as multi-scale deformation encoded features) from the motion vector features; fuse the local waveform features, temporal dependence features, and multi-scale deformation features, and obtain the input of the pre-trained disease diagnosis model based on the fused mechanical features.

[0083] As an example, the layer structure of the 1D CNN model can be a one-dimensional convolutional layer (64, k = 7) - max pooling layer - batch normalization layer, and the number of hidden units of the LSTM model can be 128. The structures and parameters of the one-dimensional convolutional neural network model, long short-term memory network model, and graph diffusion network model mentioned in this application can be set according to requirements, and the present invention is not limited thereto. In addition, the motion vector features can also not be encoded, but directly fused with the local waveform features and temporal dependence features to obtain the fused mechanical features.

[0084] Furthermore, fusing the local waveform features, temporal dependence features, and multi-scale deformation features includes: adaptively adjusting the contribution weights of the local waveform features, temporal dependence features, and multi-scale deformation features to the detection results according to the signal-to-noise ratio, and fusing the local waveform features, temporal dependence features, and multi-scale deformation features in a weighted fusion manner.

[0085] According to the signal-to-noise ratio (SNR), the contribution degrees of the features of each branch to the diagnosis result can be adaptively adjusted through dynamic weights. For example, the weights of each branch can be dynamically adjusted through learnable parameters α ∈ [0, 1] and β ∈ [0, 1], where α and β can be generated by the sigmoid function from the SNR value of the current signal, so as to achieve noise adaptive fusion. The formula for fusing features by weight adjustment is as follows:

[0086]

[0087] where, represents the local waveform features encoded from the time-domain phase features, represents the temporal dependence features encoded from the symbolic dynamics features, The motion vector features are encoded to obtain multi-scale deformation features, and the mechanical features encoded by fusion are obtained.

[0088] As an example, as Figure 4 shown, since the time-domain phase features and symbol dynamics features are obtained based on the IQ path signals, the time-domain and frequency-domain features can be coupled first based on an in-phase encoder or a quadrature encoder, and then the motion vector features are coupled to obtain the mechanical features encoded by fusion. Input into a decoder (the decoder is used to convert into information that can be recognized by a pre-trained disease diagnosis model), and finally the arrhythmia detection result of the patient can be obtained through the disease diagnosis model.

[0089] In addition, there may be a problem of misalignment of the feature spaces between the time-domain and frequency-domain features. Therefore, before fusing the local waveform features and the temporal dependence features, the method further includes: using the JS divergence constraint to align the feature spaces of the local waveform features and the temporal dependence features. The JS divergence (Jensen-Shannon Divergence) is an index used to measure the similarity of two probability distributions. Adding the JS divergence constraint to the loss function means constructing a loss function based on the JS divergence. Based on the original formula of the JS divergence and the branch features in this application, the following formula of the loss function can be obtained:

[0090]

[0091] where KL represents the KL divergence,

[0092] In the MIT-BIH AF database test, the fused multi-modal encoded features make the F1-score reach 94.7%, which is significantly improved compared with the single time-domain features (89.2%) and symbol features (91.1%). Moreover, for the specific f wave of atrial fibrillation, the detection sensitivity can reach 98.3%, and it can identify the microvolt-level (<50 μV) low-frequency oscillation components missed by traditional methods.

[0093] In some embodiments of the present invention, the disease diagnosis model can perform multi-task diagnosis to determine whether a patient has an arrhythmia problem and output the classification result of the arrhythmia, realizing the identification and classification of the arrhythmia. For example, the output results of the disease diagnosis model include: normal heart rhythm, or types of arrhythmias such as atrial fibrillation, ventricular premature beat grading, or conduction block localization. According to the output results of the disease diagnosis model, an alarm can be triggered or abnormal event data can be stored simultaneously. According to the function of the disease diagnosis model, the disease diagnosis model can be constructed based on a classification model, and the classification model used to construct the disease diagnosis model can be a support vector machine, a decision tree, or a CardioNet neural network, etc. The disease diagnosis model mentioned in this application can adopt an existing model or a self-designed disease diagnosis model, and the present invention does not specifically limit it.

[0094] As an example, the CardioNet neural network can be used for the diagnosis of cardiovascular diseases. For example, based on deep learning models (such as CNN and LSTM) and the CardioNet neural network, feature extraction and classification can be performed on electrocardiogram signals. However, the input data of existing disease diagnosis models are mostly electrocardiogram signals. To be applicable to existing disease diagnosis models, after obtaining the fused encoded mechanical features, based on a pre-trained cross-modal diffusion model or a myocardial displacement-electrical conduction coupling equation, the fused mechanical features can be mapped into electrophysiological signals, and the electrophysiological signals can be used as the input of the pre-trained disease diagnosis model. That is, the fused mechanical features can be used as the input of the disease diagnosis model, or after converting the fused mechanical features into electrophysiological signals, the electrophysiological signals can be used as the input of the disease diagnosis model. For example, this application constructs a myocardial displacement-electrical conduction coupling equation for converting the mechanical displacement of the heart into the phase change of myocardial electrical activity, and its formula is as follows:

[0095]

[0096] where λ is the wavelength of the electromagnetic wave emitted by the radar system, v myocardium(t) is the ventricular wall movement speed at time t, is the electrocardiogram signal corresponding to the heart micro-motion signal.

[0097] The above-mentioned pre-trained cross-modal diffusion model or myocardial displacement-electrical conduction coupling equation is only an example, and this application does not limit the specific method of mapping mechanical features into electrophysiological signals.

[0098] In addition, considering the individual differences among patients, a disease diagnosis model can be constructed by combining a classification model and an adaptive learning mechanism. For example, an individual physiological feature library can be constructed, and real-time calibration of individual physiological features can be achieved through an online adaptive learning mechanism (such as establishing a user baseline through the cardiac micro-motion signals collected in the first 30 seconds), and an environmental interference feature library can be constructed to preprocess the original signals through online transfer learning to identify environmental electromagnetic interference features in real time. Moreover, since the disease diagnosis model is used to identify arrhythmias caused by differences such as heart rate variability, atrial activity, and ventricular activity, prior knowledge of myocardial kinematics can be embedded in the disease diagnosis model as a constraint on the diagnosis results, such as displacement continuity constraint or periodic symmetry constraint.

[0099] As an example, a radar sign imaging system with high resolution can be developed based on the above-mentioned arrhythmia detection method to achieve non-contact cardiac function assessment and break through the functional limitations of traditional ECG. For example, the spatial resolution of the radar sign imaging system can reach 3mm (axial) × 8° (azimuthal), and the time resolution can reach a frame interval of 50ms to achieve non-contact cardiac function assessment (the calculation error of ejection fraction can be less than 5%), breaking through the functional limitations of traditional ECG.

[0100] The arrhythmia detection method proposed in this application can be applied to various scenarios such as homes, hospitals, and nursing homes, and has the following significant advantages:

[0101] (1) Non-contact and non-intrusive detection: There is no need to wear any sensors, avoiding the problems of poor comfort and compliance of traditional detection methods, and improving the user experience. And continuous 24-hour monitoring can be achieved to improve the detection rate of atrial fibrillation.

[0102] (2) Brand-new data processing method: High-frequency detection waves are used, and a brand-new signal processing method is proposed for the high-frequency characteristics of the signals; a new method for extracting time-domain phase features is proposed, and on the basis of currently considering time-domain or frequency-domain analysis, time-domain phase, frequency-domain Doppler, and spatial-domain motion vector field features are combined to overcome the problem of feature singularity; and mechanical motion features are mapped to electrocardiogram signals to break through the limitations of ECG electrical signals.

[0103] (3) Adoption of a dynamic disease diagnosis model: An online learning mechanism is used to calibrate individual physiological features in real time, improving the individual's adaptability and reducing the false alarm rate.

[0104] (4) High detection accuracy: Verified through clinical trials, the atrial fibrillation recognition accuracy reaches 97.32%.

[0105] Correspondingly, the present invention also provides a non-contact arrhythmia detection system, which includes a computer device. The computer device includes a processor and a memory. A computer program / instructions is stored in the memory. The processor is configured to execute the computer program / instructions stored in the memory. When the computer program / instructions are executed by the processor, the system implements the steps of the method described above.

[0106] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program / instructions is stored. When the computer program / instructions are executed by a processor, the steps of the foregoing edge computing server deployment method are implemented. The computer-readable storage medium may be a tangible storage medium, such as a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0107] Those of ordinary skill in the art should understand that the various exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Specifically, whether to implement in hardware or software depends on the specific application and design constraints of the technical solution. A professional technician can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium or transmitted through a data signal carried in a carrier wave on a transmission medium or a communication link.

[0108] It should be clear that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order between steps after understanding the spirit of the present invention.

[0109] In the present invention, the features described and / or illustrated for one embodiment can be used in the same or a similar manner in one or more other embodiments, and / or combined with the features of other embodiments or replace the features of other embodiments.

[0110] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the embodiments of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A non-contact arrhythmia detection method, characterized in that, The method includes the following steps: Obtain the micro-motion signal of the patient's heart detected by a radar system, and obtain the time-domain phase feature, symbolic dynamics feature, and motion vector feature based on the micro-motion signal of the heart; Obtain the input of a pre-trained disease diagnosis model based on the time-domain phase feature, symbolic dynamics feature, and motion vector feature, and input it into the disease diagnosis model to output the arrhythmia detection result of the patient; Among them, the time-domain phase feature and the motion vector feature are obtained respectively in the following ways: Divide the micro-motion signal of the heart into two orthogonal signals, obtain a phase signal based on the phase information of the two orthogonal signals, and perform a Fourier transform on the phase signal to obtain the time-domain phase feature; and Based on the non-linear deformation model of biomechanics, decompose the non-rigid displacement component from the micro-motion signal of the heart, and calculate the non-rigid stress tensor based on the non-rigid displacement component and the viscoelastic constitutive equation of myocardial tissue, so as to obtain the motion vector feature based on the non-rigid stress tensor.

2. The method according to claim 1, wherein After obtaining the micro-motion signal of the heart, the method further includes: constructing a cardiac dynamics model based on the micro-motion signal of the heart; The constructing a cardiac dynamics model based on the micro-motion signal of the heart includes: By finite element mesh division, discretize the heart surface into multiple nodes, and calculate the phase change of each node using the transmitted signal of the radar system and the micro-motion signal of the heart; Estimate the displacement vector of each node based on the phase change of each node and the extended Kalman filter, so as to construct a cardiac dynamics model.

3. The method according to claim 2, characterized in that, The calculating the non-rigid stress tensor based on the non-rigid displacement component and the viscoelastic constitutive equation of myocardial tissue includes: In the cardiac dynamics model, calculate the difference between the non-rigid displacement components of adjacent nodes on the heart surface to obtain a local strain tensor, and perform a time derivative on the local strain tensor to obtain a strain rate tensor; Determine the relaxation modulus based on biomechanical experiments or inverse problem solving methods; Use the strain rate tensor, the relaxation modulus, and the viscoelastic constitutive equation of myocardial tissue to determine the stress tensor in the time domain.

4. The method according to claim 1, wherein The time-domain phase feature is obtained in the following way: Divide the micro-motion signal of the heart into two orthogonal signals, calculate the arctangent function of the ratio of the two orthogonal signals to obtain a phase signal; Take the first derivative of the phase signal, and obtain the time-domain phase feature based on the result of the first derivative and the signal sampling rate.

5. The method according to claim 1, characterized in that The obtaining the input of a pre-trained disease diagnosis model based on the time-domain phase feature, symbolic dynamics feature, and motion vector feature includes: Use a one-dimensional convolutional neural network model to obtain local waveform features from the time-domain phase feature, use a long short-term memory network model to obtain the temporal dependence features of the symbolic sequence from the symbolic dynamics feature, and use a graph diffusion network model to obtain multi-scale deformation features from the motion vector feature; Fuse the local waveform features, temporal dependence features, and multi-scale deformation features, and obtain the input of the pre-trained disease diagnosis model based on the fused mechanical features.

6. The method according to claim 5, characterized in that, Before fusing the local waveform features and the temporal dependence features, the method further includes: using the JS divergence constraint to align the feature spaces of the local waveform features and the temporal dependence features; The fusing of the local waveform features, the temporal dependence features, and the multi-scale deformation features includes: Adapting the contribution weights of the local waveform features, the temporal dependence features, and the multi-scale deformation features to the detection result according to the signal-to-noise ratio, and fusing the local waveform features, the temporal dependence features, and the multi-scale deformation features in a weighted fusion manner.

7. The method according to claim 5, wherein The obtaining of the input of the pre-trained disease diagnosis model based on the fused mechanical features includes: Based on a pre-trained cross-modal diffusion model or a myocardial displacement-electrical conduction coupling equation, mapping the fused mechanical features into electrophysiological signals, and using the electrophysiological signals as the input of the pre-trained disease diagnosis model.

8. The method according to claim 1, characterized in that, The cardiac micro-motion signal is obtained by performing signal preprocessing on the original signal collected by the radar system, and the signal preprocessing process includes: Inputting the original signal collected by the radar system into a high-order anti-aliasing filter and a clutter suppressor for filtering; Performing dynamic beamforming processing and phase compensation on the filtered signal to obtain a noise-reduced signal; Optimizing the noise-reduced signal through a beamforming algorithm that maximizes the autocorrelation coefficient and compressive sensing technology to obtain the cardiac micro-motion signal.

9. A non-contact arrhythmia detection system, comprising a processor, a memory, and computer programs / instructions stored on the memory, characterized in that, The processor is configured to execute the computer program / instructions, and when the computer program / instructions are executed, the system implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having computer programs / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images