Non-contact electrocardiogram generation method based on FMama neural network

Abstract

The invention relates to the technical field of vital sign monitoring, in particular to a non-contact electrocardiogram generation method based on an FMama neural network, and the method comprises the steps: S1, processing received radar data, and extracting heartbeat micro-motion signals of a heart from the radar data; s2, carrying out multi-level decomposition on the input heartbeat micro-motion signal by utilizing maximum overlapping discrete wavelet transform (MODWT); s3, the multiple levels are mapped to a high-dimensional feature space through an embedded network, and the high-dimensional feature space is input into an independent FMama neural network module for sequence feature extraction; step S4, performing self-adaptive variable-pitch scanning by a FreQUent SSM (FSSM) module in the FMama neural network module, and performing self-adaptive variable-pitch scanning for different frequency bands of each level to obtain Mo; the electrocardiogram can be generated in a non-contact mode, and the method is suitable for application scenes such as remote medical monitoring, vehicle-mounted health systems and home-based care.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vital sign monitoring, and more particularly to a non-contact electrocardiogram generation method based on FMamba neural network. BACKGROUND

[0002] According to statistics of the World Health Organization (WHO), cardiovascular diseases cause about 17.9 million deaths each year, accounting for about 32% of the total number of deaths worldwide. Arrhythmia, early myocardial ischemia and other heart diseases often develop gradually without obvious symptoms. By the time the patient feels unwell and seeks medical attention, the best intervention opportunity has usually been missed. Studies have shown that through high-precision, long-term electrocardiogram (ECG) monitoring, abnormalities such as atrial fibrillation and premature beats can be detected in the subclinical stage in a timely manner, which helps to significantly reduce the risk of major events such as myocardial infarction and stroke.

[0003] Traditional electrocardiogram signal acquisition mainly relies on attaching electrodes directly to the skin surface to detect the electrical activity of the heart. Although this method is widely used in clinical practice and has high measurement accuracy, it has significant limitations in actual use, such as being unsuitable for long-term continuous monitoring, causing skin discomfort due to adhesion, and being inconvenient for use in mobile scenarios.

[0004] Studies have shown that there is a coupling relationship between the electrical activity of the heart and its mechanical beating, and the electrical signal changes recorded by the electrocardiogram are closely related to the slight fluctuations of the chest caused by the contraction and relaxation of the heart. This provides a theoretical basis for indirectly inferring the electrical activity of the heart and reconstructing the ECG waveform by monitoring the subtle displacement of the chest caused by heartbeats. In recent years, non-contact vital sign detection technology has become a research hotspot, and ultra-wideband (UWB) radar has been widely used to monitor respiration and heart rate non-contactly due to its excellent penetration ability, time resolution and anti-interference characteristics. However, how to accurately capture the mechanical details corresponding to the electrical activity of the heart based on the chest micro-motion signals obtained by UWB radar, and to restore the key low-frequency electrocardiogram components (such as P-wave, QRS complex, T-wave) from them with high fidelity, still faces major challenges such as complex modeling, significant noise interference, and difficulty in effectively extracting long-term dependent features. This is also a key technical bottleneck that needs to be broken through in the current field of non-contact electrocardiogram reconstruction. SUMMARY

[0005] The purpose of the present application is to provide a non-contact electrocardiogram generation method based on FMamba neural network, which can generate electrocardiogram non-contactly and is suitable for remote medical monitoring, vehicle health systems and home care applications.

[0006] The purpose of the present application is achieved by the following technical solutions:

[0007] A non-contact electrocardiogram (ECG) generation method based on the FMamba neural network includes the following steps:

[0008] Step S1: Process the received radar data and extract the heartbeat micro-motion signal from the radar data;

[0009] Step S2: Perform multi-level decomposition on the input heartbeat micro-motion signal using the Maximum Overlap Discrete Wavelet Transform (MODWT);

[0010] Step S3: Each of the multi-level layers is mapped to a high-dimensional feature space using an embedding network, and then input into an independent FMamba neural network module for sequence feature extraction;

[0011] Step S4: The Frequent SSM (FSSM) module within the FMamba neural network performs adaptive variable spacing scanning, scanning different frequency bands at each level to obtain M. o ;

[0012] In step S2, the maximum overlap discrete wavelet transform (MODWT) decomposes each heartbeat micro-motion signal into multiple levels of detail signals and low-frequency signal components;

[0013] Right now: Where D j,t and S N,t These represent the reconstructed signal components obtained by inverse transformation of the detail signal obtained from layer j = 1, 2, 3, ..., N and the smooth signal from layer N, respectively. That is, the original heartbeat micro-motion signal is the sum of the detail components and the smooth component at all scales.

[0014] The N-level decomposition of the heartbeat micro-motion signal, wherein the decomposition formula for the coefficients corresponding to the heartbeat micro-motion signal in the j-th level is:

[0015]

[0016]

[0017] W j,t V represents the detail coefficients of the j-th layer, reflecting the rapid changes in the frequency band; j,t This represents the smoothing coefficient of the j-th layer, which preserves the low-frequency components of the signal; and These are the normalized high-pass and low-pass filter coefficients, respectively.

[0018] In step S3, the heartbeat micro-motion signal is decomposed into several components M at different scales using the maximum overlap discrete wavelet transform (MODWT). MODWT(k), k ∈ {1, 2, 3,..., N}, for each maximum overlap discrete wavelet transform (MODWT) decomposition level, an embedding network is used to map it to a high-dimensional feature space and input into an independent FMamba neural network module for sequence feature extraction;

[0019] The input cardiac micro-movement data M of each level MODWT (k) ∈ R B×L×D , B represents the batch size, L represents the sequence length, and D represents the sequence dimension. The input is input into the FMamba neural network module to obtain the final output M o ;

[0020] In the FMamba neural network module processing, the input number is first subjected to layer normalization processing to obtain LN(M MODWT (k)), which improves the stability of training. Then, the LN(M MODWT (k)) data is processed through two different branches.

[0021] Among them, the first branch obtains Linear(LN(M MODWT (k))) through a linear transformation layer, projects in the feature dimension, learns the combination between different features, and then obtains Conv(Linear(LN(M MODWT (k))) through a one-dimensional convolution layer to extract the local time context information. After processing by the SiLU activation function, the nonlinear relationship is introduced to enhance the ability of the FMamba neural network module to learn subtle patterns.

[0022] In the step S4, the Frequent SSM (FSSM) module introduces a variable interval scanning strategy based on frequency size. The variable interval scanning strategy based on frequency size performs detailed scanning with small intervals for useful low-frequency information in the signal and performs rough scanning with large intervals for more high-frequency useless information.

[0023] The variable interval scanning strategy based on frequency size is specifically: for the components M MODWT (k) of different scales after decomposition, which represent a time series [x1, x2, x3,..., x k ], the corresponding frequency band is [f k1 , f k2 ], and the discrete integral mean value is obtained with the function F(f) = a·(f-b) 2 +c to obtain the corresponding scanning interval n.

[0024] In the function, b represents the center frequency of 25Hz, c represents the minimum value of 1, and a is a learnable parameter. This function has the smallest QRS band scanning interval in the electrocardiogram, followed by the P band and T band scanning interval, and the largest high-frequency useless noise band scanning interval.

[0025] The formula for the scanning interval n is:

[0026]

[0027] For the initial time series [x1,x2,x3,...,x] k The forward interval scan time series obtained after forward scan:

[0028] X = [x1, x2] 1+n ,x 1+2n ,...,x 1+jn ]

[0029] Then, a reverse scan is performed to obtain the reverse interval scan time series:

[0030] X inverse =[x k ,x k-n ,x k-2n ,...,x k-jn ]

[0031] Finally, X and X inverse After inputting into the Frequent SSM (FSSM) module, the state space modeling and fusion are completed;

[0032] LN(M MODWT The second branch of the (k) data, input data LN(M) MODWT (k) is processed by linear transformation and activation function to obtain M. i , with M f Perform element-wise multiplication ( ) and then combine with the original input M MODWT (k) Add the residuals to obtain the final output M. o The overall mathematical process is as follows:

[0033] M f =FSSM(SiLU(Conv(Linear(LN(M) MODWT (k))))))

[0034] M i =SiLU(Linear(LN(M) MODWT (k))))

[0035] M g =M f M i

[0036] M o =M MODWT (k)+Linear(M g )

[0037] After that, in order to fully fuse the information of different time scales and frequency scales, the output M o (N) of the Nth layer is spliced with the maximum overlap discrete wavelet transform (MODWT) transformed component M MODWT (N-1) of the (N-1)th layer, and then input into the FMamba neural network module of the (N-1)th layer:

[0038] M input (k-1)=Concat(M o (k),M MODWT (k-1)),k∈{2,3,...,N}

[0039] By analogy, the output M o (2) of the second layer is spliced with the maximum overlap discrete wavelet transform (MODWT) transformed component M MODWT (1) of the first layer to form a multi-scale feature fusion representation M input (1), which is input into the FMamba neural network module to obtain the output M o (1) of the first layer; then, the outputs M o (1) to M o (N) of each layer are subjected to inverse wavelet transform operation, finally, the fused features are input into the FMamba neural network module of the last layer for further global sequence modeling, and output as a waveform sequence M output : M output =Mamba(InverseMODWT(M o (1),M o (2),...,M o (N))) of the simulated standard lead electrocardiogram.

[0040] The present application has the following advantages:

[0041] The present application is a method for generating electrocardiogram based on FMamba neural network module deep learning, which combines the maximum overlap discrete wavelet transform (MODWT) FMamba neural network module, effectively separates different frequency components of the cardiac micro-motion signal through multi-scale wavelet decomposition, and designs a frequency-based adaptive scanning interval strategy, which scans the low-frequency information in detail and sparsely scans the high-frequency noise, effectively preserves the integrity of the useful low-frequency part of the signal, and improves the fidelity and calculation efficiency of the main low-frequency signals P, QRS and T wave reconstruction in the electrocardiogram.

[0042] Further, the dynamic dependence coupling across time scales and frequency scales is realized by layer-by-layer feature splicing and cascading FMamba neural network modules, which fully utilizes the linear complexity advantage of the FMamba neural network module in long sequence modeling, not only enhances the ability to capture the overall waveform structure of the electrocardiogram, but also facilitates real-time deployment on resource-constrained edge devices and wearable devices. BRIEF DESCRIPTION OF DRAWINGS

[0043] The application will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0044] Figure 1 is a schematic diagram of the FMamba neural network-based non-contact electrocardiogram generation method of the application;

[0045] Figure 2 is a schematic diagram of the FMamba neural network module of the application;

[0046] Figure 3 is a schematic diagram of the variable interval scanning strategy based on frequency size. DETAILED DESCRIPTION

[0047] The application will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0048] As Figures 1 to 3 shown below, the steps and functions of a FMamba neural network-based non-contact electrocardiogram generation method are described in detail;

[0049] A FMamba neural network-based non-contact electrocardiogram generation method, the method comprising the following steps:

[0050] Step S1: processing the received radar data to extract the heartbeat micro-motion signals of the heart from the radar data;

[0051] Step S2: using maximum overlap discrete wavelet transform (MODWT) to perform multi-level decomposition on the input heartbeat micro-motion signals, as Figure 1 shown;

[0052] To overcome the shortcomings of maximum overlap discrete wavelet transform (MODWT) in time positioning and translation invariance, the maximum overlap discrete wavelet transform (MODWT) omits the downsampling operation in the filtering process, and in the step S2, the maximum overlap discrete wavelet transform (MODWT) decomposes each heartbeat micro-motion signal into a plurality of degree detail signals and low-frequency signal parts;

[0053] That is, where D j,t and S N,tThe reconstructed signal component, i.e. the original heartbeat micro-motion signal, is the sum of the detail components at all scales and the smooth component at the bottom layer.

[0054] The N-level decomposition of the heartbeat micro-motion signal, wherein the decomposition formula of the heartbeat micro-motion signal corresponding to the coefficient of the jth layer is:

[0055]

[0056] W j,t represents the detail coefficient of the jth layer, reflecting the rapid change of the frequency band; V j,t represents the smooth coefficient of the jth layer, retaining the low-frequency component of the signal. and are the normalized high-pass and low-pass filter coefficients, respectively.

[0057] Step S3: The multi-level is respectively mapped to a high-dimensional feature space by an embedding network, and input into an independent FMamba neural network module for sequence feature extraction.

[0058] The heartbeat micro-motion signal is decomposed into components M MODWT (k), k∈{1,2,3,...,N} by maximum overlap discrete wavelet transform (MODWT), and position embedding is performed. MODWT For each maximum overlap discrete wavelet transform (MODWT) decomposition level, an embedding network is used to map it to a high-dimensional feature space, and the input heart micro-motion data M B×L×D (k)∈R o is input into the FMamba neural network module to obtain the final output M MODWT The structure of the FMamba neural network module is shown in Figure 2 .

[0059] The input heart micro-motion data M B×L×D (k)∈R o , B represents batch size, L represents sequence length, and D represents sequence dimension. MODWT MODWT In the FMamba neural network module processing, the input number is first subjected to layer normalization to obtain LN(M MODWT (k)), which improves the stability of training, and then the LN(M MODWT (k)) data is processed through two different branches.

[0061] The first branch obtains Linear(LN(M)) through a linear transformation layer. MODWT (k))), projecting onto the feature dimension, learns the combinations between different features, and then passes through a one-dimensional convolutional layer to obtain Conv(Linear(LN(M)). MODWT (k) Extract local time context information, process it with the SiLU activation function to introduce nonlinear relationships, and enhance the ability of the FMamba neural network module to learn subtle patterns;

[0062] Step S4: The Frequent SSM (FSSM) module within the FMamba neural network performs adaptive variable spacing scanning, scanning different frequency bands at each level to obtain M. o ;

[0063] The Frequent SSM (FSSM) module introduces a frequency-based variable-interval scanning strategy. This strategy performs detailed scanning with small intervals on useful low-frequency information in the signal, while performing coarse scanning with large intervals on useless high-frequency information. The implementation process of the frequency-based variable-interval scanning strategy is as follows: Figure 3 As shown, this strategy can better restore the frequency characteristics of low-frequency P, QRS and T waves in the electrocardiogram. While ensuring the integrity of effective low-frequency information, it effectively avoids redundant sampling caused by invalid high-frequency information and improves the feature acquisition rate.

[0064] The strategy of variable interval scanning based on frequency magnitude is specifically as follows: for components M of different scales after decomposition... MODWT (k) represents a time series [x1,x2,x3,...,x k The corresponding frequency band is [f] k1 ,f k2 ], and the function F(f) = a·(fb) 2 +c is used to perform discrete integration and calculate the mean value to obtain the corresponding scanning interval n;

[0065] In the function, b represents the center frequency of 25Hz, c represents the minimum value of 1, and a is a learnable parameter. This function has the smallest QRS band scanning interval in the electrocardiogram, followed by the P band and T band scanning interval, and the largest high-frequency useless noise band scanning interval.

[0066] The formula for the scanning interval n is:

[0067]

[0068] For the initial time series [x1,x2,x3,...,x] kThe forward interval scanning time sequence obtained after forward scanning:

[0069] X = [x1, x 1+n ,x 1+2n ,...,x 1+jn ]

[0070] The reverse interval scanning time sequence obtained after reverse scanning:

[0071] X inverse = [x k ,x k-n ,x k-2n ,...,x k-jn ]

[0072] Finally, X and X inverse are input into the Frequent SSM (FSSM) module to complete the modeling and fusion of the state space;

[0073] The second branch of the LN(M MODWT (k)) data, the input data LN(M MODWT (k)) is processed through linear transformation and an activation function to obtain M i , which is multiplied element by element with M f , and then the original input M MODWT (k) is added to obtain the final output M o ; the overall mathematical process is as follows:

[0074] M f = FSSM(SiLU(Conv(Linear(LN(M MODWT (k))))))

[0075] M i = SiLU(Linear(LN(M MODWT (k))))

[0076] M g = M f M i

[0077] M o = M MODWT (k) + Linear(M g )

[0078] In order to fully fuse the information of different time scales and frequency scales, the output M o (N) of the Nth layer is combined with the maximum overlap discrete wavelet transform (MODWT) transform component M MODWT of the (N-1)th layer.(N-1) After concatenation (Concat), the input is input into the FMamba neural network module of the N-1 layer:

[0079] M input (k-1) = Concat (M o (k), M MODWT (k-1)), k ∈ {2, 3,..., N}

[0080] By analogy, the output M o (2) of the second layer is concatenated with the maximum overlap discrete wavelet transform (MODWT) transform component M MODWT (1) of the first layer to form a multi-scale feature fusion representation M input (1) and input into the FMamba neural network module to obtain the output M o (1) of the first layer; then the outputs M o (1) to M o (N) of each layer are subjected to inverse wavelet transform operation, finally, the fused features are input into the FMamba neural network module of the last layer for further global sequence modeling, and output as a waveform sequence M output : M output = Mamba (InverseMODWT (M o (1), M o (2),..., M o (N));

[0081] In the training process of the generative ECG model of the application, the Adam optimizer is used, the initial learning rate is set to 0.001, and the dynamic adjustment strategy learning rate decay is used to ensure the stability and convergence of the training. The loss function uses the L2 loss function, which aims to measure the difference between the predicted results and the actual results, and promote the model to be effectively optimized. In addition, Dropout is applied in the training process to prevent overfitting and improve the generalization ability of the model. The training batch size is set to 32 and adjusted according to the memory usage during training, and each training round is 100 rounds. The dataset uses Nature public dataset A dataset of clinically recorded radar vital signs with synchronized reference sensor signals, selects the ECG and radar data measured by all subjects in the resting state, and uses R-R interval error, QRS interval error, P-R interval error, Q-T interval error to measure the learning effect of the model, and the specific results are shown in Table 1.

[0082] Table 1 Training results

[0083] R-R interval QRS interval P-R interval Q-T interval Error (ms) 2.8 ms 5.4 ms 16.6 ms 16.4 ms Error percentage (%) 94％ 95％ 52％ 70％

Claims

1. A non-contact electrocardiogram generation method based on FMamba neural network, characterized in that: The method includes the following steps: Step S1: Process the received radar data and extract the heartbeat micro-motion signal from the radar data; Step S2: Perform multi-level decomposition on the input heartbeat micro-motion signal using the Maximum Overlap Discrete Wavelet Transform (MODWT); Step S3: Each of the multi-level layers is mapped to a high-dimensional feature space using an embedding network, and then input into an independent FMamba neural network module for sequence feature extraction; Step S4: The Frequent SSM (FSSM) module within the FMamba neural network performs adaptive variable spacing scanning, scanning different frequency bands at each level to obtain M. o。 2. The non-contact electrocardiogram generation method based on Mamba neural network according to claim 1, characterized in that: In step S2, the maximum overlap discrete wavelet transform (MODWT) decomposes each heartbeat micro-motion signal into multiple levels of detail signals and low-frequency signal components; Right now: Where D j,t and S N,t These represent the reconstructed signal components obtained by inverse transformation of the detail signal obtained from layer j = 1, 2, 3, ..., N and the smooth signal from layer N, respectively. In other words, the original heartbeat micro-motion signal is the sum of the detail components and the smooth component at all scales.

3. The non-contact electrocardiogram generation method based on Mamba neural network according to claim 2, characterized in that: The N-level decomposition of the heartbeat micro-motion signal, wherein the decomposition formula for the coefficients corresponding to the heartbeat micro-motion signal in the j-th level is: W j,t V represents the detail coefficients of the j-th layer, reflecting the rapid changes in the frequency band; j,t This represents the smoothing coefficient of the j-th layer, which preserves the low-frequency components of the signal; and These are the normalized high-pass and low-pass filter coefficients, respectively.

4. The non-contact electrocardiogram generation method based on Mamba neural network according to claim 1, characterized in that: In step S3, the heartbeat micro-motion signal is decomposed into several components M at different scales using the maximum overlap discrete wavelet transform (MODWT). MODWT For each Maximum Overlap Discrete Wavelet Transform (MODWT) decomposition level (k), k∈{1,2,3,…,N}, an embedding network is used to map it to a high-dimensional feature space and input it into an independent FMamba neural network module for sequence feature extraction.

5. The non-contact electrocardiogram generation method based on Mamba neural network according to claim 4, characterized in that: The input cardiac micromotion data M at each level MODWT (k)∈R B×L×D B represents the batch size, L represents the sequence length, and D represents the sequence dimension. The input is fed into the FMamba neural network module to obtain the final output M. o .

6. The non-contact electrocardiogram generation method based on Mamba neural network according to claim 5, characterized in that: In the FMamba neural network module, the input numbers are first normalized to obtain LN(M) MODWT (k)) to improve training stability, then LN(M) MODWT (k) The data will be processed through two different branches; The first branch obtains Linear(LN(M)) through a linear transformation layer. MODWT (k))), projecting onto the feature dimension, learns the combinations between different features, and then passes through a one-dimensional convolutional layer to obtain Conv(Linear(LN(M)). MODWT (k) Extracts local time context information, which is then processed by the SiLU activation function to introduce nonlinear relationships, thereby enhancing the ability of the FMamba neural network module to learn subtle patterns.

7. The non-contact electrocardiogram generation method based on Mamba neural network according to claim 6, characterized in that: In step S4, the Frequent SSM (FSSM) module introduces a strategy of variable interval scanning based on frequency magnitude. This strategy performs detailed scanning of useful low-frequency information in the signal with small intervals, and performs coarse scanning of useless high-frequency information with large intervals.

8. The non-contact electrocardiogram generation method based on Mamba neural network according to claim 7, characterized in that: The strategy of variable interval scanning based on frequency magnitude is specifically as follows: for components M of different scales after decomposition... MODWT (k) represents a time series [x1,x2,x3,...,x k The corresponding frequency band is [f] k1 ,f k2 ], and the function F(f) = a·(fb) 2 +c is used to perform discrete integration and calculate the mean value to obtain the corresponding scanning interval n; In the function, b represents the center frequency of 25Hz, c represents the minimum value of 1, and a is a learnable parameter. This function has the smallest QRS band scanning interval in electrocardiogram, followed by the P band and T band scanning interval, and the largest high-frequency useless noise band scanning interval.

9. The non-contact electrocardiogram generation method based on Mamba neural network according to claim 8, characterized in that: The formula for the scanning interval n is: For the initial time series [x1,x2,x3,...,x] k The forward interval scan time series obtained after forward scan: X=[x1,x 1+n ,x 1+2n ,...,x 1+jn ] Then, a reverse scan is performed to obtain the reverse interval scan time series: X inverse =[x k ,x k-n ,x k-2n ,...,x k-jn ] Finally, X and X inverse After inputting into the Frequent SSM (FSSM) module, the state space modeling and fusion are completed.

10. A non-contact electrocardiogram generation method based on a Mamba neural network according to claim 9, characterized in that: LN(M MODWT The second branch of the (k) data, input data LN(M) MODWT (k) is processed by linear transformation and activation function to obtain M. i , with M f Perform element-wise multiplication () and then combine with the original input M. MODWT (k) Add the residuals to obtain the final output M. o The overall mathematical process is as follows: M f =FSSM(SiLU(Conv(Linear(LN(M MODWT (to)))))) M i =SiLU(Linear(LN(M MODWT (to)))) M g =M f M i M o =M MODWT (k)+Linear(M g ) To fully integrate information from different time and frequency scales, the output M of the Nth layer is then... o (N) and the maximum overlap discrete wavelet transform (MODWT) component M of the (N-1)th layer MODWT After concatenating (N-1), the data is then input into the FMamba neural network module of layer N-1: M input (k-1)=Concat(M o (k),M MODWT (k-1)),k∈{2,3,...,N} And so on, finally the output M of the second layer o (2) Maximum overlap with the first layer Discrete Wavelet Transform (MODWT) component M MODWT (1) After splicing, a multi-scale feature fusion representation M is formed. input (1) Input the FMamba neural network module to obtain the output M of the first layer. o (1); then the output M of each layer o (1)~M o (N) Perform inverse wavelet transform. Finally, the fused features are input into the last layer of the FMamba neural network module for further global sequence modeling, and output as a waveform sequence M simulating a standard lead electrocardiogram. output M output =Mamba(InverseMODWT(M o (1),M o (2),...,M o (N))).

Images