Non-contact electrocardiogram generation method and system based on millimeter wave radar

Abstract

The invention discloses a non-contact electrocardiogram generation method and system based on a millimeter wave radar, and relates to the field of electrocardiogram, and the method comprises the following steps: pairing a heartbeat signal RCG collected by the millimeter wave radar with a standard electrocardiogram signal ECG to construct a training data set; and carrying out time sequence feature extraction on the RCG, and learning forward and backward time sequence features through a bidirectional long short-term memory network Bi-LSTM. Signal importance weights at different time points are calculated by adopting an attention mechanism, and weighting processing is carried out to obtain electrocardio dynamic characteristics. And inputting the features into a generator, and performing signal conversion through a deep convolutional network. A double-discriminator structure is utilized to optimize and generate signals, the main discriminator verifies authenticity, and the auxiliary discriminator verifies consistency. And finally, carrying out processing after reverse normalization and smooth filtering to obtain a target electrocardiogram signal. By implementing the method, the accuracy of non-contact electrocardiogram generation can be improved.

Description

Technical Field

[0001] This application relates to the field of electrocardiogram, and particularly to a non-contact electrocardiogram generation method and system based on millimeter-wave radar. Background Art

[0002] Electrocardiogram detection, as an important means for diagnosing cardiovascular diseases, plays a crucial role in clinical medicine. Traditional electrocardiogram detection requires directly attaching electrode patches to the patient's body surface. This contact measurement method is prone to causing skin allergies in long-term monitoring scenarios and is also cumbersome to operate, which limits the wide application of electrocardiogram monitoring.

[0003] With the development of radar sensing technology, millimeter-wave radar has gradually been applied to the field of physiological signal detection due to its non-contact measurement advantage. Existing technologies capture the Doppler effect signals generated by heart movements through millimeter-wave radar, use simple signal processing methods for filtering and feature extraction, and then complete signal conversion using traditional machine learning algorithms.

[0004] However, due to the complex non-linear mapping relationship between heart movement signals and electrocardiac activity signals, existing signal processing and feature extraction methods are difficult to accurately capture the temporal dependence between signals, resulting in a large difference between the converted electrocardiogram signals and standard electrocardiograms in terms of waveform details and morphological features, which affects the clinical application value of the signals. Summary of the Invention

[0005] This application provides a non-contact electrocardiogram generation method and system based on millimeter-wave radar, which is used to improve the accuracy of non-contact electrocardiogram generation.

[0006] In a first aspect, the present application provides a non-contact electrocardiogram generation method based on a millimeter-wave radar, which is applied to a non-contact electrocardiogram generation system. The method includes: pairing the heartbeat signal RCG collected by the millimeter-wave radar with the simultaneously collected standard electrocardiogram signal ECG to obtain a training signal pair dataset. The heartbeat signal RCG reflects the physical characteristics of cardiac motion, and the standard electrocardiogram signal ECG represents the bioelectrical characteristics of cardiac electrical activity; extracting the temporal characteristics of the heartbeat signal RCG, and learning the forward and backward temporal characteristics of the heartbeat signal RCG through a bidirectional long short-term memory network Bi-LSTM to obtain a bidirectional temporal characteristic sequence; calculating the signal importance weights at different time points for the bidirectional temporal characteristic sequence through an attention mechanism and performing weighted processing to obtain an electrocardiodynamic weighted characteristic sequence; inputting the electrocardiodynamic weighted characteristic sequence into a generator, and performing signal conversion through a deep convolutional network to obtain an original electrocardiogram characteristic sequence; using a dual discriminator structure to optimize the original electrocardiogram characteristic sequence to obtain a corrected electrocardiogram characteristic sequence. The dual discriminator structure includes a main discriminator and an auxiliary discriminator. The main discriminator verifies the authenticity, and the auxiliary discriminator verifies the signal consistency; performing post-processing of inverse normalization and smoothing filtering on the corrected electrocardiogram characteristic sequence to obtain a target electrocardiogram signal.

[0007] In the above embodiment, first, the heartbeat signal RCG is collected by a millimeter-wave radar, and a training dataset is constructed with the standard electrocardiogram signal ECG. The temporal characteristics are extracted by a bidirectional long short-term memory network, and the important temporal information is highlighted by combining an attention mechanism. The signal conversion is realized through a deep convolutional network. The dual discriminator structure is innovatively introduced. The main discriminator ensures the waveform authenticity, and the auxiliary discriminator verifies the signal consistency. Finally, after post-processing, a target electrocardiogram signal with a complete shape and accurate details is obtained, realizing a high-quality conversion from the physical characteristics of cardiac motion to the bioelectrical characteristics.

[0008] Combined with some embodiments of the first aspect, in some embodiments, the step of extracting the temporal characteristics of the heartbeat signal RCG and learning the forward and backward temporal characteristics of the heartbeat signal RCG through a bidirectional long short-term memory network Bi-LSTM to obtain a bidirectional temporal characteristic sequence specifically includes: inputting the heartbeat signal RCG into two bidirectional long short-term memory network layers respectively for feature extraction to obtain an initial feature sequence, where each long short-term memory network includes an input gate, a forget gate, and an output gate; calculating the activation vectors of the input gate, the forget gate, and the output gate according to the input vector at the current moment and the state at the previous moment; using the activation vectors, combining the weight matrix and the bias matrix, to update the state information of the memory unit; based on the updated state of the memory unit and the activation vector of the output gate, generating the output feature at the current moment through an activation function; combining the output features of all time steps into a temporal characteristic sequence as the bidirectional temporal characteristic of the heartbeat signal RCG.

[0009] In the above embodiments, the bidirectional long short-term memory network performs bidirectional feature extraction on the RCG heartbeat signal in the forward and backward directions through the collaborative work of the input gate, forget gate, and output gate. The input gate and forget gate respectively control the input of new information and the forgetting of historical information, while the output gate is responsible for outputting the feature information in the current state. Through activation function processing and temporal combination, bidirectional temporal features with complete context dependencies are finally obtained, effectively capturing the long-term dependencies of the signal.

[0010] Combined with some embodiments of the first aspect, in some embodiments, the step of calculating the signal importance weights at different time points through the attention mechanism for weighted processing of the bidirectional temporal feature sequence to obtain the electrocardiodynamic weighted feature sequence specifically includes: learning global dependency information from the bidirectional temporal feature sequence through the attention layer; calculating the correlation degree scores between the output sequences at different time points and the output at the current time point using the softmax function; and performing weighted processing on the bidirectional temporal feature sequence according to the calculated correlation degree scores to obtain the electrocardiodynamic weighted feature sequence.

[0011] In the above embodiments, the attention mechanism first learns global dependency information from the bidirectional temporal feature sequence, and then calculates the correlation degree between the output sequences at different time points and the output at the current time point using the softmax function. Weighting is performed based on the calculated correlation degree scores, so that important temporal features obtain higher weights, and finally an electrocardiodynamic weighted feature sequence highlighting key information is obtained, enhancing the expression ability of important features.

[0012] Combined with some embodiments of the first aspect, in some embodiments, the step of optimizing the original electrocardiogram feature sequence using the dual discriminator structure specifically includes: using the main discriminator to receive the original electrocardiogram feature sequence and the real electrocardiogram signal as inputs, and calculating the authenticity probability through a four-layer one-dimensional convolutional network; using the auxiliary discriminator to receive the original electrocardiogram feature sequence and the RCG heartbeat signal as inputs, verifying the signal consistency and calculating the auxiliary loss; and optimizing the original electrocardiogram feature sequence according to the authenticity probability and the auxiliary loss to obtain the corrected electrocardiogram feature sequence.

[0013] In the above embodiments, the main discriminator receives the original electrocardiogram feature sequence and the real electrocardiogram signal as inputs, and calculates the authenticity probability through a four-layer one-dimensional convolutional network to ensure the waveform authenticity of the generated signal. The auxiliary discriminator verifies the consistency between the original electrocardiogram feature sequence and the RCG heartbeat signal and calculates the auxiliary loss. Under the combined optimization of the dual discriminator structure, the corrected electrocardiogram feature sequence is comprehensively improved in two dimensions: waveform authenticity and signal consistency.

[0014] In some embodiments in combination with some embodiments of the first aspect, the post - processing steps of performing denormalization and smoothing filtering on the pair of corrected electrocardiogram (ECG) feature sequences specifically include: performing denormalization processing on the corrected ECG feature sequences to restore the signal amplitude; performing smoothing filtering on the denormalized signal sequences to eliminate high - frequency noise and obtain the target ECG signal.

[0015] In the above - mentioned embodiments, first, denormalization processing is performed on the corrected ECG feature sequences to restore the signal amplitude to the original magnitude range. On this basis, the smoothing filtering operation effectively eliminates the high - frequency noise components, making the target ECG signal have a higher signal - to - noise ratio and a clearer waveform profile, ensuring the clinical application value and diagnostic reference significance of the signal.

[0016] In some embodiments in combination with some embodiments of the first aspect, after the step of performing denormalization and smoothing filtering on the corrected ECG feature sequences to obtain the target ECG signal, the method further includes: segmenting the target ECG signal according to a preset time window to obtain multiple signal segments, and performing feature extraction on each signal segment, including calculating the heart rate, and extracting the waveform features of P - waves, QRS - waves, and T - waves; comparing the waveform features with a preset normal ECG feature range to determine whether there is an abnormality; when an abnormality is detected, grading the type of abnormality according to the degree of deviation of the waveform features to obtain an abnormality level; selecting a corresponding prompting method according to the abnormality level, where the prompting method includes display reminder, sound alarm, or remote notification; storing the type of abnormality, the abnormality level, and the corresponding target ECG signal in a database.

[0017] In the above - mentioned embodiments, segmenting the target ECG signal according to a preset time window and extracting waveform features realizes the feature quantification of key waveforms such as P - waves, QRS - waves, and T - waves. Abnormality grading is performed according to the degree of deviation of the waveform features from the preset range, and corresponding prompting methods are used for timely feedback. The complete abnormality detection and grading mechanism makes the ECG monitoring results more clinically instructive.

[0018] In some embodiments in combination with some embodiments of the first aspect, after the step of performing denormalization and smoothing filtering on the corrected ECG feature sequences to obtain the target ECG signal, the method further includes: statistically analyzing the time pattern of the occurrence of abnormalities in the database and performing a distribution analysis on the types of abnormalities; establishing an association model between user activities and the types of abnormalities according to the distribution analysis results; generating activity suggestions for the user's daily activities based on the association model, where the activity suggestions include adjusting the activity intensity and arranging the activity time; regularly generating an ECG monitoring analysis report, which includes statistical information on the types of abnormalities, the change trend of the abnormality levels, and preventive measures recommended to be taken.

[0019] In the above embodiments, the time pattern of abnormal occurrences in the database is statistically analyzed to establish an association model between user activities and abnormal types, thereby grasping the internal relationship between electrocardiogram abnormalities and user behavior patterns. Based on the activity suggestions and regular analysis reports generated by the association model, a complete feedback loop including abnormal type statistics, grade change trends, and preventive measures is formed, enabling the electrocardiogram monitoring system to have intelligent early warning and intervention capabilities.

[0020] In a second aspect, an embodiment of the present application provides a non-contact electrocardiogram generation system, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, and the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the non-contact electrocardiogram generation system to execute the method described in the first aspect and any possible implementation manner in the first aspect.

[0021] In a third aspect, an embodiment of the present application provides a computer program product containing instructions. When the computer program product runs on a non-contact electrocardiogram generation system, the non-contact electrocardiogram generation system is enabled to execute the method described in the first aspect and any possible implementation manner in the first aspect.

[0022] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, including instructions. When the instructions run on a non-contact electrocardiogram generation system, the non-contact electrocardiogram generation system is enabled to execute the method described in the first aspect and any possible implementation manner in the first aspect.

[0023] It can be understood that the non-contact electrocardiogram generation system provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the method provided in the embodiments of the present application. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method, and will not be elaborated here.

[0024] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. In the present application, through the collaborative work of the input gate, forget gate, and output gate of the bidirectional long short-term memory network, bidirectional feature extraction of the heartbeat signal RCG is performed forward and backward. The input gate and forget gate respectively control the input of new information and the forgetting of historical information, and the output gate is responsible for outputting the feature information in the current state. Through activation function processing and temporal combination, bidirectional temporal features including complete context dependencies are finally obtained, effectively capturing the long-term dependencies of the signal.

[0025] 2. By segmenting the target electrocardiogram signal according to a preset time window and extracting waveform features, this application achieves the quantitative characterization of key waveforms such as P waves, QRS waves, and T waves. Abnormality grading is performed based on the degree of deviation of the waveform features from the preset range, and corresponding prompting methods are used for timely feedback. The complete abnormality detection and grading mechanism makes the electrocardiogram monitoring results more clinically instructive.

[0026] 3. By statistically analyzing the time pattern of abnormality occurrences in the database, this application establishes an association model between user activities and abnormality types, thereby grasping the internal relationship between electrocardiogram abnormalities and user behavior patterns. The activity suggestions and regular analysis reports generated based on the association model form a complete feedback loop that includes abnormality type statistics, grade change trends, and preventive measures, enabling the electrocardiogram monitoring system to have intelligent early warning and intervention capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 is a schematic flowchart of a non-contact electrocardiogram generation method based on millimeter-wave radar in an embodiment of this application; Figure 2 is another schematic flowchart of a non-contact electrocardiogram generation method based on millimeter-wave radar in an embodiment of this application; Figure 3 is another schematic flowchart of a non-contact electrocardiogram generation method based on millimeter-wave radar in an embodiment of this application; Figure 4 is a system framework diagram of a non-contact electrocardiogram generation method based on millimeter-wave radar in an embodiment of this application; Figure 5 is a generator architecture diagram of a non-contact electrocardiogram generation method based on millimeter-wave radar in an embodiment of this application; Figure 6 is a dual discriminator structure diagram of a non-contact electrocardiogram generation method based on millimeter-wave radar in an embodiment of this application; Figure 7 is a schematic structural diagram of a physical device of a non-contact electrocardiogram generation system in an embodiment of this application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The terms used in the following embodiments of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. As used in the specification of this application, the singular forms "a", "an", "above-mentioned", "the", and "this" are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term " / and" used in this application refers to any or all possible combinations including one or more of the listed items.

[0029] Hereinafter, the terms "first" and "second" are for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "a plurality" is two or more.

[0030] For ease of understanding, the application scenarios of the embodiments of the present application are introduced below.

[0031] In the related art, cardiac health monitoring can be achieved by adopting the traditional contact electrocardiogram acquisition method. The scenario of using the non-contact electrocardiogram generation method based on millimeter-wave radar in the related art is introduced below. This method requires directly attaching electrode patches to the patient's body surface, which may cause skin allergies during long-term monitoring and is cumbersome to operate, limiting its application in scenarios such as home monitoring.

[0032] However, by adopting the non-contact electrocardiogram generation method based on millimeter-wave radar in the embodiments of the present application, the time-series features of the heartbeat signal are extracted through a bidirectional long short-term memory network, the important time-series information is highlighted by combining the attention mechanism, and the generated signal is optimized by using a dual discriminator structure, realizing a high-fidelity conversion from the radar heartbeat signal to the electrocardiogram signal. The scenario of using the non-contact electrocardiogram generation method based on millimeter-wave radar in the present application is introduced below. This method does not require any contact devices, and patients can complete long-term electrocardiogram monitoring in a natural state, which is especially suitable for continuous monitoring of special populations such as burn patients and premature infants.

[0033] For ease of understanding, the method provided in this embodiment is described in terms of its process in combination with the above scenarios. Please refer to Figure 1 , which is a schematic flowchart of a non-contact electrocardiogram generation method based on millimeter-wave radar in the embodiments of the present application.

[0034] S101. Pair the heartbeat signal RCG collected by the millimeter-wave radar with the simultaneously collected standard electrocardiogram signal ECG to obtain a training signal pair dataset.

[0035] Among them, the heartbeat signal RCG represents the heartbeat signal reflecting the physical characteristics of heart movement collected by the millimeter-wave radar, the ECG signal refers to the bioelectric characteristic signal representing the electrical activity of the heart collected through the standard lead method, and the training signal pair dataset is used to represent the data set composed of the paired RCG and ECG signals for model training.

[0036] This step is executed before starting to train the CardioGAN model and is used to prepare the data required for training. Specifically, first, a millimeter-wave radar is used to collect the subject's RCG (radial cardiogram) signal, and at the same time, a standard ECG (electrocardiogram) lead is used to collect the electrocardiogram signal. Then, the two signals are synchronously paired according to the acquisition timestamp to ensure that the RCG and ECG signals correspond one by one in time. Finally, the paired signal pairs are saved as the training dataset.

[0037] In some embodiments, the pairing of the RCG and ECG signals and the construction of the dataset can be achieved in various ways: Optionally, first, preprocess the original RCG and ECG signals, including operations such as filtering and denoising, normalization, etc.; then align the processed signals according to the sampling timestamp; finally, save the aligned signals as training sample pairs. Optionally, first segment the RCG and ECG signals by a fixed time window; then perform feature extraction and quality assessment on the segmented signals; finally, select the signal pairs with qualified quality as the training data. It can be understood that other signal pairing and dataset construction methods can also be adopted, which are not limited here.

[0038] S102. Extract the temporal features of the RCG (radial cardiogram) signal. The bidirectional long short-term memory network Bi-LSTM is used to learn the forward and backward temporal features of the RCG (radial cardiogram) signal, and a bidirectional temporal feature sequence is obtained.

[0039] Among them, the temporal feature refers to the change feature of the signal in the time dimension. Bi-LSTM represents a recurrent neural network structure that can simultaneously learn the forward and backward dependence relationships of the sequence. The bidirectional temporal feature sequence is used to represent the feature representation containing complete temporal information extracted by Bi-LSTM.

[0040] This step is executed after obtaining the paired dataset and is used to extract the temporal dependence features in the RCG (radial cardiogram) signal. Specifically, the RCG (radial cardiogram) signal is input into two LSTM networks with opposite directions. The forward LSTM processes the sequence from left to right, and the backward LSTM processes the sequence from right to left. The output features of the two LSTM networks are fused through a concat operation to obtain a feature sequence containing complete forward and backward temporal information.

[0041] In some embodiments, the extraction of temporal features can be achieved in various ways: Optionally, first, frame the input RCG (radial cardiogram) signal; then input each frame of the signal into the input gate, forget gate, and output gate of Bi-LSTM in turn; finally, update the cell state according to the gating state and output the features. Optionally, first convert the RCG (radial cardiogram) signal into a suitable input format; then perform feature extraction through multiple layers of Bi-LSTM; finally, perform weighted fusion on the features extracted by different layers. It can be understood that other temporal feature extraction methods can also be adopted, which are not limited here.

[0042] S103. Calculate the signal importance weights at different time points for the bidirectional time series feature sequence through the attention mechanism and perform weighted processing to obtain the electrocardiogram dynamics weighted feature sequence.

[0043] Among them, the attention mechanism refers to a calculation method that adaptively learns the importance of different time positions. The signal importance weight represents the contribution degree of each time point to the current output. The electrocardiogram dynamics weighted feature sequence is used to represent the sequence representation with dynamic time series features obtained after attention weighting. The weighted processing refers to the operation of weighted summation of features according to the calculated weights.

[0044] This step is executed after obtaining the bidirectional time series feature sequence and is used to highlight the feature information of important time points in the sequence. Specifically, first learn the global dependence information from the bidirectional time series feature sequence through the attention layer, and calculate the context vector as the weighted sum of the output states from the Pre - Attention layer. Then use the softmax function to calculate the correlation degree score between the output sequence at different time points and the output at the current time point. This score depends on the pre - attention output state and the hidden state at the previous time step in the generated sequence. Finally, perform weighted processing on the bidirectional time series feature sequence according to the calculated correlation degree score to obtain the electrocardiogram dynamics weighted feature sequence that integrates the time series importance information.

[0045] In some embodiments, the feature weighting of the attention mechanism can be implemented in multiple ways: Optionally, first calculate the similarity between the query vector and the key vector to obtain the original attention score; then normalize the attention score to obtain the weight coefficient; finally, perform weighted summation of the weight coefficient and the value vector to obtain the weighted feature. Optionally, first construct a multi - head attention structure to learn the feature relationships in different sub - spaces respectively; then splice the outputs of each attention head; finally, fuse through a linear transformation to obtain the final weighted feature sequence. It can be understood that other attention calculation and feature weighting methods can also be adopted, which are not limited here.

[0046] S104. Input the electrocardiogram dynamics weighted feature sequence into the generator, and perform signal conversion through a deep convolutional network to obtain the original electrocardiogram feature sequence.

[0047] Among them, the generator refers to a neural network module used to convert input features into target domain signals. The deep convolutional network represents a feature extraction and conversion network composed of multiple layers of convolutional operations. The original electrocardiogram feature sequence is used to represent the electrocardiogram signal feature representation obtained after preliminary conversion by the generator. The signal conversion refers to the process of converting one signal form into another signal form.

[0048] This step is executed after obtaining the weighted feature sequence and is used to convert the RCG features into ECG feature representations. Specifically, first, the electrocardiodynamic weighted feature sequence is input into the generator network. The generator adopts a multi-layer one-dimensional convolutional structure, and through layer-by-layer feature extraction and transformation, it realizes the mapping from the RCG feature space to the ECG feature space. Each layer of convolutional operation is equipped with batch normalization and non-linear activation functions to enhance the expressive power and training stability of the network. Finally, the converted original electrocardiogram feature sequence is output.

[0049] In some embodiments, feature conversion can be achieved in various ways: Optionally, first extract features through multi-layer downsampling convolution; then use dilated convolution to expand the receptive field to capture long-range dependencies; finally, restore the signal resolution through upsampling convolution. Optionally, first reshape the feature sequence into a shape suitable for convolution processing; then perform feature transformation through residual convolution blocks; finally, add skip connections to fuse multi-scale features. It can be understood that other network structures can also be used to achieve feature conversion, which is not limited here.

[0050] S105. Optimize the original electrocardiogram feature sequence using a dual discriminator structure to obtain a corrected electrocardiogram feature sequence.

[0051] Among them, the dual discriminator structure refers to a collaborative optimization architecture that includes a main discriminator and an auxiliary discriminator. The main discriminator represents a network module used to verify the authenticity of the generated signal, and the auxiliary discriminator is an auxiliary network module used to verify the signal consistency. The corrected electrocardiogram feature sequence is used to represent the high-quality electrocardiogram feature representation after being optimized by the dual discriminator. The authenticity probability represents the similarity between the generated signal and the real signal, and the auxiliary loss is a loss function used to measure the consistency between the generated signal and the input RCG heartbeat signal.

[0052] This step is executed after obtaining the original electrocardiogram feature sequence and is used to improve the quality and reliability of the generated signal. Specifically, first, the original electrocardiogram feature sequence and the real electrocardiogram signal are input into the main discriminator, and the authenticity probability is calculated through a four-layer one-dimensional convolutional network. The convolutional kernel sizes are 5×512, 5×256, 5×128, and 5×64 in sequence, which are used to evaluate the authenticity of the generated signal. At the same time, the original electrocardiogram feature sequence and the RCG heartbeat signal are input into the auxiliary discriminator to verify the signal consistency and calculate the auxiliary loss to ensure the correlation between the generated signal and the input signal. Finally, the original electrocardiogram feature sequence is optimized and adjusted according to the authenticity probability and the auxiliary loss to obtain a corrected electrocardiogram feature sequence with higher quality.

[0053] In some embodiments, the dual discriminator optimization can be achieved in various ways: Optionally, first calculate the adversarial loss using the main discriminator; then use the auxiliary discriminator to calculate the consistency loss; finally, guide the update of the generator parameters by weighted combination of the two loss functions. Optionally, first train the main discriminator and the auxiliary discriminator separately until convergence; then fix the discriminator parameters to optimize the generator; finally, use the alternating training strategy for fine-tuning. It can be understood that other optimization strategies can also be adopted to achieve the correction of the feature sequence, which is not limited here.

[0054] S106. Perform post-processing of inverse normalization and smoothing filtering on the corrected electrocardiogram feature sequence to obtain the target electrocardiogram signal.

[0055] Among them, inverse normalization refers to the operation of restoring the normalized eigenvalue to the original numerical range, and smoothing filtering represents a filtering processing method for eliminating high-frequency noise of the signal. The target electrocardiogram signal is used to represent the finally generated electrocardiogram signal with clinical application value. Post-processing refers to the signal optimization and adjustment operation performed after the main processing steps.

[0056] This step is executed after obtaining the corrected electrocardiogram feature sequence and is used to generate the finally available electrocardiogram signal. Specifically, first perform inverse normalization processing on the corrected electrocardiogram feature sequence, map the eigenvalue from the normalized range of [-1, 1] or [0, 1] back to the amplitude range of the original electrocardiogram signal to ensure the correctness of the signal amplitude. Then perform smoothing filtering on the signal sequence after inverse normalization, and use methods such as low-pass filters or moving averages to eliminate the high-frequency noise components introduced during the generation process, improve the smoothness and readability of the signal, and finally obtain the target electrocardiogram signal that can be used for clinical diagnosis.

[0057] In some embodiments, the signal post-processing can be achieved in various ways: Optionally, first perform inverse normalization according to the statistical parameters recorded during training; then use a Butterworth low-pass filter to remove high-frequency noise; finally, further improve the signal quality through wavelet threshold denoising method. Optionally, first perform segmentation processing on the signal; then perform inverse normalization and filtering on each signal segment independently; finally, merge the processed signal segments through a smoothing connection algorithm. It can be understood that other signal processing methods can also be adopted to achieve the final generation of the electrocardiogram signal, which is not limited here.

[0058] The following further describes the method provided in this embodiment in a more specific process. Please refer to Figure 2 , which is another process schematic diagram of the non-contact electrocardiogram generation method based on millimeter-wave radar in the embodiments of the present application.

[0059] S201. Pair the heartbeat signal RCG collected by the millimeter-wave radar with the simultaneously collected standard electrocardiogram signal ECG to obtain a training signal pair dataset.

[0060] Among them, the millimeter-wave radar heart rate signal RCG represents the cardiac motion displacement signal detected by the millimeter-wave radar through the Doppler effect. The standard electrocardiogram signal ECG refers to the bioelectrical signal reflecting cardiac electrical activity collected using the standard 12-lead. The training signal pair dataset is used to represent the set of data samples for model training composed of time-aligned RCG and ECG signals. Synchronous acquisition means obtaining the two signals simultaneously within the same time window.

[0061] In the data acquisition stage, first, a millimeter-wave radar system with a sampling frequency of 1000 Hz is used to collect the heart rate signal of the subject. The radar transmission frequency is 60 GHz, and the bandwidth is 7 GHz. At the same time, a standard electrocardiograph is used to record the ECG signal at the same sampling frequency. Preprocessing is performed on the collected original signals, including band-pass filtering (0.5 - 40 Hz) to remove baseline drift and high-frequency noise, signal segmentation (each segment is 10 s), amplitude normalization, etc. Then, based on the time stamp, the processed RCG and ECG signals are aligned to ensure that the two signals are precisely corresponding in time. Finally, the paired signals are saved as training samples, and each pair of samples contains a segment of the heart rate signal RCG and the corresponding ECG signal. Repeat this process to obtain a sufficient number of training samples and construct a complete training dataset.

[0062] S202: Input the heart rate signal RCG into two bidirectional long short-term memory network layers respectively for feature extraction to obtain an initial feature sequence, where each long short-term memory network includes an input gate, a forget gate, and an output gate.

[0063] Among them, the bidirectional long short-term memory network is a neural network structure that can process forward and backward temporal information simultaneously. The input gate controls the degree to which the current input information enters the cell state. The forget gate determines how much of the previous cell state information is retained. The output gate controls how much of the current cell state information is output to the hidden state. The initial feature sequence represents the temporal feature representation preliminarily extracted by the bidirectional LSTM.

[0064] The preprocessed RCG heartbeat signal is segmented according to the time step T, and the signal of each time step is input into two LSTM networks with opposite directions. The forward LSTM processes the signal sequentially from the start to the end of the sequence, and the backward LSTM processes the signal from the end to the start of the sequence. Both LSTM networks contain 200 hidden units, and the cell state and hidden state are updated at each time step. For each direction, the LSTM network controls the information flow through three gate structures: the input gate uses the sigmoid function to judge the importance of the new input information, the forget gate determines the retention ratio of the historical information, and the output gate controls the information output. The outputs of the LSTM in both directions are fused through a concatenation operation to obtain a feature sequence containing bidirectional temporal information. The sequence length is the same as the input signal, and the feature dimension at each time step is 400.

[0065] S203. Calculate the activation vectors of the input gate, forget gate, and output gate according to the input vector at the current moment and the state at the previous moment.

[0066] Among them, the input vector xt represents the RCG value of the heartbeat signal at the current time step. The state at the previous moment includes the hidden state ht-1 and the cell state ct-1. The activation vector refers to the output value of each gate unit calculated through the activation function. The weight matrices W and U are used to process the current input and the hidden state at the previous moment respectively, and the bias vector b is used to adjust the fitting ability of the network.

[0067] At each time step of the LSTM network, the states of each gate unit are updated through the following specific calculation process: First, calculate the input gate activation vector it = σ(Wixt + Uiht-1 + bi), where σ is the sigmoid activation function, and the output is limited to the range [0, 1], indicating the degree of allowing new information to enter. Then calculate the output gate activation vector ot = σ(Woxt + Uoht-1 + bo) to control the proportion of information output. At the same time, calculate the forget gate activation vector ft = σ(Wfxt + Ufht-1 + bf) to determine how much historical information to retain. Based on these gate values, update the memory unit state ct = ft * ct-1 + it * tanh(Wcxt + Ucht-1 + bc), where * represents element-wise multiplication, and the tanh activation function compresses the newly generated candidate values to the range [-1, 1]. Finally, calculate the output vector ht = ot * tanh(ct) at the current moment as the feature representation of this time step.

[0068] S204. Use the activation vector, combined with the weight matrix and bias matrix, to update the state information of the memory unit.

[0069] Among them, the activation vector includes the activation values of the input gate \(i_t\), the forget gate \(f_t\), and the output gate \(o_t\). The weight matrices \(W_c\) and \(U_c\) are used to process the current input and historical information respectively, and the bias matrix \(b_c\) is used to adjust the overall offset of the network. The memory cell state \(c_t\) stores the long-term memory information. The update of the state information refers to the process of selectively updating the memory cell state through the gating mechanism.

[0070] The update calculation process of the memory cell state is: \(c_t = f_t * c_{t - 1}+i_t * \tanh(W_cx_t + U_ch_{t - 1}+b_c)\). The specific implementation is as follows: First, multiply the current input \(x_t\) by the weight matrix \(W_c\), and at the same time multiply the previous hidden state \(h_{t - 1}\) by the weight matrix \(U_c\). After adding the two and adding the bias vector \(b_c\), the candidate memory value is obtained. The candidate memory value is compressed to the range of \([-1, 1]\) through the \(\tanh\) activation function, and then element-wise multiplied by the input gate activation vector \(i_t\) to control the input ratio of new information. At the same time, the forget gate activation vector \(f_t\) is element-wise multiplied by the memory cell state \(c_{t - 1}\) at the previous moment to control the retention ratio of historical information. Finally, these two parts are added together to obtain the updated memory cell state \(c_t\), realizing the selective memory and forgetting of long-term dependence information.

[0071] S205. Based on the updated memory cell state and the activation vector of the output gate, generate the output feature at the current moment through the activation function.

[0072] Among them, the updated memory cell state \(c_t\) contains all the memory information retained at the current moment. The output gate activation vector \(o_t\) controls the output ratio of information. The output feature \(h_t\) represents the feature representation extracted at the current time step. The activation function \(\tanh\) is used to map the feature values to the standard range.

[0073] The calculation expression of the output feature is \(h_t = o_t * \tanh(c_t)\). The specific implementation process is as follows: First, apply the \(\tanh\) activation function to the updated memory cell state \(c_t\) to compress the value range to the interval of \([-1, 1]\) to make the feature distribution more regular. Then, multiply the activated state value element-wise by the activation vector of the output gate \(o_t\). The activation value of the output gate is between \([0, 1]\). Through this multiplicative gating mechanism, the network can adaptively control the amount of information output in each dimension. The finally obtained output feature \(h_t\) integrates the current input information and historical memory information, and through selective filtering, retains the most valuable feature representation.

[0074] S206. Combine the output features of all time steps into a temporal feature sequence as the bidirectional temporal features of the heartbeat signal RCG.

[0075] Among them, the updated memory cell state $c_t$ contains all the memory information retained at the current moment, the output gate activation vector $o_t$ controls the proportion of information output, and the output feature $h_t$ represents the feature representation extracted at the current time step. The activation function tanh is used to map the feature values to the standard range.

[0076] The calculation of the output feature is achieved through the following process: First, apply the tanh activation function to the updated memory cell state $c_t$ to compress the value range to the interval [-1, 1], making the feature distribution more regular. Then perform an element-wise multiplication operation on the activated state value and the activation vector of the output gate $o_t$ ($h_t = o_t * tanh(c_t)$). The activation value of the output gate is between [0, 1]. Through this multiplicative gating mechanism, the network can adaptively control the amount of information output in each dimension. The finally obtained output feature $h_t$ fuses the current input information and historical memory information, and through selective filtering, retains the most valuable feature representation.

[0077] S207. Calculate the signal importance weights at different time points for the bidirectional time series feature sequence through the attention mechanism and perform weighted processing to obtain the electrocardiodynamic weighted feature sequence.

[0078] Among them, the bidirectional time series feature sequence represents a feature representation containing the complete time series information of the bidirectional LSTM output. The attention mechanism is used to adaptively calculate the importance degree at different time positions. The signal importance weight represents the contribution size of each time point to the current output. The electrocardiodynamic weighted feature sequence refers to the sequence representation with dynamic time series features obtained after attention weighting.

[0079] The specific steps of processing the bidirectional time series features through the attention mechanism are as follows: First, input the bidirectional time series feature sequence $H = [h_1, h_2,..., h_T]$ into the attention layer, where $T$ is the sequence length. For each time step $t$, calculate the similarity between the query vector $q_t$ and the key vectors $k_t$ of all time steps to obtain the attention score $e_t$. Normalize the attention score through the softmax function to obtain the weight coefficient $\alpha_t$. The weight value represents the importance degree of different time positions to the current output. Then perform a weighted sum $c_t=\sum\alpha_t * v_t$ of the weight coefficient and the corresponding value vector $v_t$ to obtain a feature representation containing global context information. Repeat this process for all time steps, and finally obtain the electrocardiodynamic weighted feature sequence that fuses the time series importance information.

[0080] S208. Learn the global dependence information from the bidirectional time series feature sequence through the attention layer.

[0081] Among them, the attention layer refers to the network layer structure used to calculate the relationship between features. The global dependence information represents the long-range correlation between different positions in the sequence. The context vector is used to represent the global information representation concerned at the current time step.

[0082] The calculation process of the attention layer is as follows: First, perform a linear transformation on the input bidirectional temporal feature sequence H to obtain the query matrix Q, the key matrix K, and the value matrix V. Perform matrix multiplication on the query matrix Q and the key matrix K to obtain the attention score matrix E = QK^T. The attention scores reflect the strength of the correlation between different time positions. For time step t, perform a dot product operation between its query vector qt and the key vectors of all time steps to measure the degree of association between the current time step and other time steps. Through this calculation mechanism, the attention layer can capture the long-range dependencies in the sequence and overcome the problem that traditional recurrent networks are difficult to handle long sequences.

[0083] S209. Calculate the correlation degree score between the output sequences at different time points and the output at the current time point using the softmax function.

[0084] Among them, the softmax function is an activation function that maps a real-valued vector to a probability distribution. The correlation degree score represents the contribution weight of different time positions to the current output, and the normalization process ensures that the sum of all weights is 1.

[0085] The correlation degree score is calculated in the following way: For each row (corresponding to a time step) in the attention score matrix E, apply the softmax function for normalization: αt,t' = exp(et,t') / Σexp(et,i). Where et,t' represents the attention score between time step t and time step t', and the denominator is the sum of the scores for all time steps i. Through the softmax function, the original attention scores are converted into weight values in the range [0, 1], and the sum of all weights is 1. These normalized weights directly reflect the importance of different time positions to the current output. The larger the value, the greater the contribution of the information contained in that time position to the current output.

[0086] Use a bidirectional LSTM and an attention mechanism to extract the temporal dependence features of the RCG signal. The pre-attention layer includes two bidirectional long short-term memory network layers (Bi-LSTM Layer), which extract hidden and sequential information from the sequence. The Recurrent Neural Network (RNN) is often used to handle short-term temporal problems but cannot handle long-term dependence problems, which will result in the vanishing gradient and inability to perform gradient updates. The emergence of LSTM solves this problem. Microscopically, LSTM introduces a cell state and retains the useful information of the time series by controlling the input gate, forget gate, and output gate, thus avoiding the vanishing gradient. Bi-LSTM is an improvement based on LSTM, mainly learning the time series in both forward and backward directions, strengthening the connection of time information before and after, and capturing richer features. Specifically, the forward propagation calculation formula of LSTM is as follows: i t = σ(W i x t + U i h t-1 + b i )(1) o t = σ(W o x t + U o h t-1 + b o )(2) f t = s(W f x t + U f h t-1 + b f )(3) c t = f t * c t-1 + i t * tanh(W c x t + U c h t-1 + b c )(4) h t = o t * tanh(c t )(5) where x t is the input vector at the current time step t, i t , o t and f t are the activation vectors of the input gate, output gate, and forget gate at time step t respectively, c t is the memory cell vector at time step t, h t is the output vector at time step t, W i , W o , W f and W c are the weight matrices of the input gate, output gate, forget gate, and memory cell respectively, b i , b o , b f and b c are the bias matrices of the input gate, output gate, forget gate, and memory cell respectively, and σ is the activation function.

[0087] Calculate the output for T time steps to obtain a vector of shape (T×1).

[0088] The role of the B2 attention layer is to learn precise global dependency information from the entire real sample sequence, regardless of the sequence length, and it helps to model the RCG signal generated from the radar-synchronized heartbeat signal. This layer greatly improves the accuracy of the generated sequence. Equation (6) describes the attention layer as a feed-forward neural network model. This layer learns the sequence at a given time step based on the entire input sequence.

[0089] Context vector c t It is calculated as a weighted sum of the output states from the Pre-Attention layer: The learnable parameter α(t, t′) is calculated as the softmax function: e (t,t′) = s(a t-1 , y t ″)(8) Sofamax calculates a score that represents the degree of correlation between the pre-attention output sequence around position t′ and the output at position t. As is obvious from the previous discussion, this score depends on the pre-attention output state y t ″ and the hidden state of the previous time step a t-1 in the generated sequence.

[0090] B3: The Post-Attention layer is responsible for generating the electrocardiogram signal based on the latent information obtained from the hidden state, the corresponding output of the previous time step in the Post-Attention layer, and the context vector from the Attention layer. The conditional probability of each unit can be defined as: p(y t |y 1 , y 2 ,..., y t-1 , c t ) = g(y t-1 , a t , c t )(9) a t = f(a t-1 , y t-1 , c t )(10) S210. Weight the bidirectional time-series feature sequence according to the calculated degree of correlation score to obtain the electrocardiodynamic weighted feature sequence.

[0091] Among them, the correlation degree score refers to the normalized weight value αt,t' calculated by the softmax function. The bidirectional time series feature sequence is the feature sequence H output by the LSTM network. The weighting process refers to the weighted sum operation of the weight value and the feature vector. The electrocardiodynamic weighted feature sequence represents a feature representation that integrates the time series importance information.

[0092] The generation process of the weighted feature sequence is implemented by the context vector calculation method: for each time step t, multiply all the attention weights αt,t' corresponding to this time step by the feature vector y't' at the corresponding time step and sum them, that is, ct = Σαt,t'y't'. The specific steps are as follows: First, obtain the attention weight vector [αt,1, αt,2,..., αt,T] of time step t for all other time steps, and these weight values have been normalized by softmax. Then multiply each weight value by the feature vector at the corresponding time step to obtain the weighted feature vector. Finally, sum all the weighted feature vectors to obtain the context vector ct at time step t. Repeat this process for each time step in the sequence, and finally obtain the electrocardiodynamic weighted feature sequence containing global context information.

[0093] S211. Use the main discriminator to receive the original electrocardiogram feature sequence and the real electrocardiogram signal as inputs, and calculate the authenticity probability through a four-layer one-dimensional convolutional network.

[0094] Among them, the main discriminator is a neural network used to distinguish real and generated data. The original electrocardiogram feature sequence refers to the electrocardiogram signal features that have not been processed. The real electrocardiogram signal is the standard ECG data collected clinically. The one-dimensional convolutional network consists of four convolutional layers. The authenticity probability represents the confidence that the input data is real data.

[0095] The specific implementation of the main discriminator adopts a four-layer one-dimensional convolutional structure: the first layer uses 64 convolutional kernels with a kernel size of 3 and a stride of 2; the second layer uses 128 convolutional kernels with a kernel size of 3 and a stride of 2; the third layer uses 256 convolutional kernels with a kernel size of 3 and a stride of 2; the fourth layer uses 512 convolutional kernels with a kernel size of 3 and a stride of 2. After each layer of convolution, a batch normalization layer and a LeakyReLU activation function are connected. Input the original electrocardiogram feature sequence and the real electrocardiogram signal into this network respectively. After four layers of convolution processing, a scalar value is output through a fully connected layer, and then mapped to the [0, 1] interval through the sigmoid function to obtain the probability value representing the authenticity of the input data.

[0096] S212. Use the auxiliary discriminator to receive the original electrocardiogram feature sequence and the RCG heartbeat signal as inputs, verify the signal consistency and calculate the auxiliary loss.

[0097] Among them, the auxiliary discriminator is a neural network used to evaluate the similarity between the generated data and the source data. The original ECG feature sequence refers to the features of the unprocessed ECG signal. The RCG of the heartbeat signal is the ECG signal collected by radar. The signal consistency represents the matching degree of the two signals in the temporal characteristics. The auxiliary loss is used to measure the difference between the generated data and the source data.

[0098] The calculation process of the auxiliary discriminator is similar to that of the main discriminator, adopting the same four-layer one-dimensional convolutional structure. The original ECG feature sequence and the RCG of the heartbeat signal are used as inputs, and features are extracted through the convolutional network respectively. The auxiliary loss is obtained by calculating the mean square error between the two feature sequences. The smaller the loss value, the better the consistency between the generated ECG feature sequence and the RCG of the heartbeat signal. The introduction of the auxiliary discriminator provides additional supervision information, which helps the generator generate an ECG signal that better matches the source signal in temporal characteristics.

[0099] S213. Optimize the original ECG feature sequence according to the authenticity probability and the auxiliary loss to obtain the corrected ECG feature sequence.

[0100] Among them, the authenticity probability refers to the score of the data authenticity output by the main discriminator. The auxiliary loss represents the difference measure between the original ECG feature sequence and the RCG of the heartbeat signal. The optimization process refers to adjusting the feature sequence by the gradient descent method. The corrected ECG feature sequence refers to the feature representation of the ECG signal after optimization.

[0101] The feature sequence optimization is carried out using the combined loss function: the total loss L = λ1Ladv + λ2Laux, where Ladv = -log(D(G(x))) represents the adversarial loss, D(G(x)) is the authenticity probability output by the discriminator, Laux represents the mean square error loss calculated by the auxiliary discriminator, and λ1 and λ2 are the weight coefficients for balancing the two losses. For each training batch, first calculate the discriminator score and the auxiliary loss of the feature sequence output by the generator, and then calculate the gradient of the combined loss with respect to the generator parameters. Update the generator parameters through backpropagation to make the generated feature sequence improve the discriminator score while maintaining consistency with the RCG of the heartbeat signal. After multiple rounds of iterative optimization, finally obtain the corrected ECG feature sequence that can pass the discriminator verification and match the RCG of the heartbeat signal in features.

[0102] S214. Perform denormalization processing on the corrected ECG feature sequence to restore the signal amplitude.

[0103] Among them, the corrected ECG feature sequence is the optimized normalized feature representation. The denormalization processing refers to restoring the normalized data to the original data range. The signal amplitude represents the actual voltage value size of the ECG signal.

[0104] Inverse normalization processing uses a calculation process opposite to that of data preprocessing: for the normalized eigenvalue x', the inverse transformation is performed using x = x' * σ + μ, where σ is the standard deviation of the training data and μ is the mean of the training data. The specific implementation steps are as follows: First, obtain the mean μ and standard deviation σ saved during the preprocessing of the training data. Then, multiply each data point in the corrected feature sequence by the standard deviation σ and add the mean μ. In this way, the data normalized to the interval [-1, 1] or [0, 1] is restored to the voltage value range of the original electrocardiogram signal, giving the signal practical physical significance.

[0105] S215. Perform smoothing filtering on the signal sequence after inverse normalization to eliminate high-frequency noise and obtain the target electrocardiogram signal.

[0106] Among them, the signal sequence after inverse normalization refers to the electrocardiogram signal after restoring the amplitude. Smoothing filtering is a processing method for removing high-frequency interference in the signal. High-frequency noise represents the fast-fluctuating components in the signal, and the target electrocardiogram signal is the final output standard electrocardiogram signal. The filtering process uses a low-pass filter in digital filtering technology. This filter allows low-frequency signals to pass through while attenuating high-frequency components.

[0107] The generation of the target electrocardiogram signal adopts the following specific steps: First step, design a 4th-order Butterworth low-pass filter, set the sampling frequency to 500 Hz (that is, sample 500 points per second), and the cut-off frequency to 40 Hz to ensure that the main frequency components (0.05 - 35 Hz) of the electrocardiogram signal can pass through the filter. Second step, perform zero-phase filtering on the signal sequence after inverse normalization, and eliminate phase distortion by applying the filter in the forward and reverse directions respectively. The specific implementation method is: first filter the signal from front to back once, then flip the filtered signal, perform the second filtering, and finally flip the signal again. Third step, calculate the output signal through the difference equation of the Butterworth filter: y(n) = b0x(n) + b1x(n - 1) + b2x(n - 2) +... - a1y(n - 1) - a2y(n - 2) -..., where x(n) is the input signal, y(n) is the output signal, and bi and ai are the filter coefficients. This filtering method can effectively remove the 50 Hz power frequency interference and other high-frequency noises in the signal, while maintaining the steep edge characteristics of the QRS complex, making the waveforms of slow changes such as P waves and T waves smoother. The finally obtained target electrocardiogram signal has a clear waveform contour, accurately reflects the characteristics of cardiac electrical activities, and is suitable for clinical diagnosis and analysis.

[0108] The following further describes the method provided in this embodiment in a more specific process. Please refer to Figure 2 , which is another process schematic diagram of the non-contact electrocardiogram generation method based on millimeter-wave radar in the embodiment of the present application.

[0109] After step S215, the following steps are further included: S301. Segment the target electrocardiogram (ECG) signal according to a preset time window to obtain multiple signal segments, and perform feature extraction on each signal segment, including calculating the heart rate, and extracting waveform features of P wave, QRS complex, and T wave.

[0110] Among them, the preset time window refers to a time interval with a fixed length (usually set to 10 seconds). The signal segment refers to the ECG signal sequence intercepted within the time window. The heart rate represents the number of heartbeats per minute. The P wave represents the potential change generated during atrial depolarization. The QRS complex represents the potential change generated during ventricular depolarization. The T wave represents the potential change generated during ventricular repolarization. The waveform features include quantitative indexes such as the amplitude, duration, and interval of the wave.

[0111] The specific steps of signal segmentation and feature extraction are as follows: First, segment the ECG signal into equal-length segments according to a 10-second time window. For each segment, locate the peak point of the R wave through the differential threshold method. The R wave with an amplitude exceeding the threshold and an interval greater than 200 ms is marked as a valid R wave. Calculate the instantaneous heart rate based on the interval between adjacent R waves: Heart rate = 60 / RR interval (seconds). After determining the position of the R wave, search forward for the Q wave (the first negative deflection point) and backward for the S wave (the lowest point) to obtain the start point, end point, and duration of the QRS complex. Search for the P wave before the QRS complex: The peak point above the baseline is the vertex of the P wave, and determine the start point and end point of the P wave. Search for the T wave after the QRS complex: The peak point above the baseline is the vertex of the T wave, and determine the start point and end point of the T wave. Extract the characteristic parameters of each wave: the amplitude and duration of the P wave, the amplitude and duration of the QRS complex, the amplitude and duration of the T wave, the PR interval (from the start point of the P wave to the start point of the QRS complex), and the QT interval (from the start point of the QRS complex to the end point of the T wave).

[0112] S302. Compare the waveform features with the preset normal ECG feature range to determine whether there is an abnormality.

[0113] Among them, the waveform features refer to the quantitative indexes extracted from the signal segments. The normal ECG feature range refers to the normal value interval of each index specified by the clinical standard. Abnormality refers to the situation where the characteristic parameters exceed the normal range.

[0114] Feature comparison uses a threshold judgment method: Heart rate is judged. The normal range is 60 - 100 beats per minute. A heart rate lower than 60 beats per minute is bradycardia, and a heart rate higher than 100 beats per minute is tachycardia. The characteristics of the P wave are judged. The normal P wave amplitude is 0.05 - 0.25 mV, and the duration is less than 0.12 seconds. Exceeding the range indicates atrial abnormality. The characteristics of the QRS complex are judged. The normal QRS duration is 0.06 - 0.10 seconds, and the amplitude is 0.5 - 2.5 mV. Exceeding the range indicates ventricular conduction abnormality. The characteristics of the T wave are judged. The normal T wave amplitude is 1 / 8 to 2 / 3 of the Q wave amplitude, and the duration is 0.10 - 0.25 seconds. An abnormal morphology indicates myocardial repolarization abnormality. The PR interval is judged. The normal range is 0.12 - 0.20 seconds. A too long PR interval indicates atrioventricular block, and a too short PR interval indicates preexcitation syndrome. The QT interval is judged. The normal QTc (corrected QT interval) is 0.35 - 0.44 seconds. A too long QT interval indicates repolarization abnormality.

[0115] S303. When an abnormality is detected, the type of abnormality is classified according to the degree of deviation of the waveform characteristics to obtain an abnormality level.

[0116] Among them, the type of abnormality refers to the specific manifestation forms of various electrocardiogram abnormalities (such as tachycardia, conduction block, ST change, etc.). The degree of deviation refers to the specific value by which the characteristic parameter exceeds the normal range. The abnormality level uses a three - level classification standard (mild, moderate, severe) to quantitatively classify the severity of the abnormality.

[0117] The abnormality classification process uses a multi - parameter quantitative evaluation method: First, set the classification criteria for various abnormalities. Heart rate abnormality classification (mild bradycardia 50 - 59 beats per minute, moderate 40 - 49 beats per minute, severe <40 beats per minute; mild tachycardia 100 - 120 beats per minute, moderate 121 - 150 beats per minute, severe >150 beats per minute), QRS complex abnormality classification (mild prolongation 0.10 - 0.12 seconds, moderate 0.12 - 0.15 seconds, severe >0.15 seconds), ST - segment change classification (mild elevation 0.1 - 0.2 mV, moderate 0.2 - 0.3 mV, severe >0.3 mV), T - wave abnormality classification (mild: T - wave inversion depth <0.2 mV, moderate: 0.2 - 0.5 mV, severe: >0.5 mV), QT - interval abnormality classification (mild prolongation 0.44 - 0.48 seconds, moderate 0.48 - 0.52 seconds, severe >0.52 seconds). For each detected abnormality, determine the level according to the specific value of its characteristic parameter with reference to the classification criteria. When multiple abnormalities exist simultaneously, use the highest level as the final abnormality level.

[0118] S304. Select the corresponding prompt method according to the abnormality level. The prompt method includes display reminder, sound alarm, or remote notification.

[0119] Among them, display reminder means marking abnormal information in text or graphic form on the display interface, sound alarm means prompting abnormalities through sound signals with different frequencies and rhythms, and remote notification means sending abnormal information to the designated recipient through the network.

[0120] The selection and execution of the prompting method adopt a hierarchical response mechanism: for mild abnormalities, the abnormal waveform area is highlighted with a yellow mark on the display interface, and the abnormal type and specific parameters are displayed at the bottom of the interface. For moderate abnormalities, in addition to the display reminder, an intermittent sound alarm is triggered (ringing once every 30 seconds, lasting for 1 second, with a tone frequency of 1000 Hz), and at the same time, a text message notification containing the abnormal type and parameters is sent to the mobile phones of the preset medical staff. For severe abnormalities, the abnormal area is displayed with a red mark, a continuous sound alarm is triggered (ringing once every 5 seconds, lasting for 2 seconds, with a tone frequency of 2000 Hz), text message and email notifications are sent to the medical staff, and the preset emergency contact number is automatically dialed.

[0121] S305. Store the abnormal type, abnormal level, and the corresponding target electrocardiogram signal in the database.

[0122] Among them, the abnormal type refers to the specific electrocardiogram abnormal manifestation, the abnormal level refers to the determined severity classification, the target electrocardiogram signal refers to the original waveform data when the abnormality occurs, and the database is a structured data storage system for storing and managing this information.

[0123] The data storage process is implemented using a relational database: create a database structure including an abnormal record table and a waveform data table. The abnormal record table contains fields: record ID (primary key), timestamp, patient ID, abnormal type, abnormal level, abnormal parameter value, processing status, etc. The waveform data table contains fields: waveform ID (primary key), record ID (foreign key), sampling time, signal data points, sampling rate, etc. For each detected abnormality, first insert a record into the abnormal record table, including information such as the time, type, and level of the abnormality occurrence, and then store the corresponding electrocardiogram signal data (waveform data from 10 seconds before to 10 seconds after the abnormality occurs) into the waveform data table, and establish the corresponding relationship between the abnormal record and the waveform data through the foreign key association. This storage structure facilitates subsequent data query, statistical analysis, and medical record management.

[0124] S306. Statistically analyze the time pattern of abnormal occurrences in the database and analyze the distribution of abnormal types.

[0125] Among them, the time pattern of abnormal occurrences refers to the time characteristics of electrocardiogram abnormalities (such as occurrence time, duration, repetition period, etc.), the abnormal type distribution refers to the proportion and distribution characteristics of different types of abnormalities in the overall abnormalities, and the distribution analysis refers to the regular results obtained by statistically processing the abnormal data.

[0126] The time pattern statistics and distribution analysis adopt the following methods: First, divide a day into 48 half-hour time periods with a 24-hour cycle, and count the occurrence times and durations of various types of abnormalities in each time period. Calculate the frequency distribution of each type of abnormality in each time period to obtain the time density curve. For each type of abnormality, calculate its average daily occurrence times, average duration, and occurrence interval. Statistically analyze the peak and trough periods of the occurrence of abnormalities and analyze the day-night distribution characteristics. Calculate the percentage of each type of abnormality in the total number of abnormalities and draw a pie chart distribution of the abnormality types. Conduct time series analysis on the data of consecutive days to identify the periodic patterns of the occurrence of abnormalities.

[0127] S307. According to the distribution analysis results, establish an association model between user activities and abnormality types.

[0128] Among them, user activities refer to the daily behaviors of the wearer (such as sleeping, exercising, working, etc.), and the association model refers to a mathematical model that describes the corresponding relationship between activity types, intensities, times, and electrocardiogram abnormalities.

[0129] The process of establishing the association model adopts a multivariate analysis method: Collect user activity logs and record activity types, start times, durations, and activity intensities. Align the activity data and electrocardiogram abnormality data on the time axis and establish a sliding observation window with a time window of 30 minutes. Calculate the correlation coefficients between the activity characteristics (types, intensities) and the occurrence of abnormalities in each observation window. Use the Logistic regression model to construct the mapping relationship from activity characteristics to the probability of abnormality occurrence. The input variables of the model include activity types (encoded as discrete variables), activity intensities (normalized continuous variables), activity durations, the types of the previous activity, and the interval time, and the output variable is the probability of abnormality occurrence.

[0130] S308. Generate activity suggestions for the user's daily activities based on the association model. The activity suggestions include activity intensity adjustment and activity time arrangement.

[0131] Among them, the activity suggestions refer to the behavioral guidance opinions given based on the association analysis results. The activity intensity adjustment refers to the modification suggestions for the amount of exercise and intensity of different activities, and the activity time arrangement refers to the rational planning of the start times of various activities.

[0132] The activity suggestion generation adopts a rule-based reasoning method: First, set the activity intensity threshold, and classify activities into low-intensity (such as walking, office work), medium-intensity (such as fast walking, cycling), and high-intensity (such as running, ball games). According to the relationship between activity intensity and the probability of abnormality in the association model, determine the safe intensity range for each type of activity. Put forward adjustment suggestions for activities beyond the safe range, such as adjusting strenuous exercise to medium-intensity exercise, or segmenting activities with too long a duration. Based on the time distribution of abnormality occurrence, identify the high-risk periods of the user, and suggest avoiding high-intensity activities during these periods. Generate a specific activity schedule, including the suggested activity type, start time, duration, and activity intensity.

[0133] S309. Regularly generate an electrocardiogram monitoring analysis report, which includes statistical information on the types of abnormalities, the changing trend of the abnormality levels, and preventive measures recommended to be taken.

[0134] Among them, the analysis report refers to the summary document of electrocardiogram monitoring regularly generated by the system. The statistical information includes quantitative indicators such as the number of abnormality occurrences and type distributions. The changing trend refers to the evolving characteristics of abnormal conditions over time. The preventive measures refer to targeted risk control suggestions.

[0135] The report generation process adopts a templatized method: Generate analysis reports on a weekly, monthly, and quarterly basis. The report content includes: (1) An overview of abnormality statistics, listing the number of occurrences, proportion, and average duration of various types of abnormalities, and presenting them in the form of charts; (2) Abnormality level analysis, counting the changes in the number of different-level abnormalities, and drawing a trend chart to show the changes in severity; (3) Time distribution analysis, showing the time pattern of abnormality occurrences, including daily distribution and weekly distribution; (4) Activity correlation analysis, explaining which activities are likely to induce abnormalities; (5) Risk assessment, assessing the overall risk level based on the frequency and severity of abnormalities; (6) Preventive suggestions, providing specific preventive measures for the high-incidence abnormality types, including activity adjustment plans, suggestions for improving living habits, and precautions. The report adopts a hierarchical display method, including both professional medical terms and detailed data, as well as easy-to-understand explanations and suggestions.

[0136] To more clearly illustrate the technical solution of this application, the following will describe in detail the specific implementation manners of this application with reference to the accompanying drawings. First, explain the overall system architecture, and then introduce in detail the specific implementation of the core modules.

[0137] Please refer to Figure 4 , which is the system framework diagram of the non-contact electrocardiogram generation method based on millimeter-wave radar in the embodiment of this application.

[0138] Figure 4It fully shows the overall process from signal input to the generation of the target electrocardiogram signal: First, the heart rate signal RCG collected by the millimeter-wave radar and the synchronously collected standard electrocardiogram signal ECG are reshaped into a (1, T) dimension; then, they are processed by the generator module. The generator includes an input layer (Input), a bidirectional long short-term memory network (Bilstm) for extracting temporal features, an attention mechanism (Attention) for calculating the importance weights of signals, an LSTM layer, and a Dense layer for feature transformation; then, a dual discriminator structure is adopted for optimization. Among them, the main discriminator verifies the authenticity of the generated signal through a one-dimensional convolutional network, and the auxiliary discriminator verifies the signal consistency through a one-dimensional convolutional network; finally, the optimized target electrocardiogram signal is output.

[0139] To further illustrate the specific structural design of the generator, please refer to Figure 5 , which is the architecture diagram of the generator for the non-contact electrocardiogram generation method based on millimeter-wave radar in the embodiment of this application, showing three key layer structures for feature extraction and processing. The pre-attention layer segments the heart rate signal RCG by time steps (x1 to xT) and processes it through a bidirectional LSTM network. Each LSTM includes an input gate, a forget gate, and an output gate, generating a bidirectional temporal feature sequence (h1 to hT). The attention layer calculates the importance weights of signals at different time points and generates a feature representation (c1 to cT) containing global dependency information. The post-attention layer further processes the features through an LSTM, transforms the features using a Dense layer, and finally outputs an electrocardiodynamic weighted feature sequence (y0 to yT).

[0140] To illustrate in detail the specific implementation manner of the dual discriminator in this application, please refer to Figure 6 , which is the structure diagram of the dual discriminator for the non-contact electrocardiogram generation method based on millimeter-wave radar in the embodiment of this application. Figure 6 It includes a main discriminator (Figure a) and an auxiliary discriminator (Figure b), both of which adopt a similar four-layer one-dimensional convolutional network structure. Among them, the main discriminator takes the original electrocardiogram feature sequence and the real electrocardiogram signal as inputs, includes four Conv1D convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function, and outputs the authenticity probability through a Dense layer; the auxiliary discriminator takes the original electrocardiogram feature sequence and the heart rate signal RCG as inputs, adopts the same network structure as the main discriminator, and obtains the auxiliary loss by calculating the mean square error to verify the signal consistency. This dual discriminator structure can optimize the authenticity of the generated signal and its consistency with the input signal simultaneously, and finally obtain a high-quality corrected electrocardiogram feature sequence.

[0141] Next, the non-contact electrocardiogram generation system in the embodiment of this invention application will be described from the perspective of hardware processing. Please refer to Figure 7, which is a schematic structural diagram of an entity device of the non-contact electrocardiogram generation system in the embodiment of the present application.

[0142] It should be noted that Figure 7 The structure of the non-contact electrocardiogram generation system shown is only an example, and should not bring any restrictions to the functions and usage scope of the embodiments of the present invention.

[0143] As Figure 7 shown, the non-contact electrocardiogram generation system includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 702 or the program loaded from the storage section 708 into the random access memory (RAM) 703, such as executing the method described in the above embodiments. In the RAM 703, various programs and data required for system operation are also stored. The CPU 701, ROM 702, and RAM 303 are connected to each other via a bus 704. The input / output (I / O) interface 705 is also connected to the bus 704.

[0144] The following components are connected to the I / O interface 305: an input section 306 including an audio input device, a button switch, etc.; an output section 707 including a liquid crystal display (LCD), an audio output device, an indicator light, etc.; a storage section 708 including a hard disk, etc.; and a communication section 709 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 709 performs communication processing via a network such as the Internet. The drive 710 is also connected to the I / O interface 705 as needed. A removable medium 711, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 710 as needed, so that the computer program read from it can be installed into the storage section 708 as needed.

[0145] Particularly, according to the embodiments of the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments of the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from the network through the communication section 709, and / or installed from the removable medium 311. When the computer program is executed by the central processing unit (CPU) 301, various functions defined in the present invention are executed.

[0146] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fibers, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above. In the present invention, a computer-readable storage medium may be any tangible medium that contains or stores a program, and the program can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0147] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. Among them, each block in the flowchart or block diagram may represent a module, a program segment, or a part of code, and the above module, program segment, or part of code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order from that marked in the accompanying drawings.

[0148] Specifically, the non-contact electrocardiogram generation system of this embodiment includes a processor and a memory, and a computer program is stored on the memory. When the computer program is executed by the processor, the non-contact electrocardiogram generation method based on millimeter-wave radar provided in the above embodiment is implemented.

[0149] On the other hand, the present invention also provides a computer-readable storage medium, which may be included in the non-contact electrocardiogram generation system described in the above embodiment; or it may exist alone and not be assembled into the non-contact electrocardiogram generation system. The above storage medium carries one or more computer programs. When the one or more computer programs are executed by a processor of the non-contact electrocardiogram generation system, the non-contact electrocardiogram generation system implements the non-contact electrocardiogram generation method based on millimeter-wave radar provided in the above embodiment.

[0150] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

[0151] As used in the foregoing embodiments, depending on the context, the term "when" can be interpreted to mean "if" or "after" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "when determining" or "if (the stated condition or event) is detected" can be interpreted to mean "if determined" or "in response to determining" or "when (the stated condition or event) is detected" or "in response to detecting (the stated condition or event)".

[0152] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the foregoing embodiments can be implemented, and the processes can be completed by relevant hardware instructed by a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the foregoing method embodiments. The foregoing storage media include: various media such as ROM or random access memory RAM, magnetic disks, or optical discs that can store program codes.

Claims

1. A non-contact electrocardiogram generation method based on millimeter wave radar, characterized in that: Applied to a non-contact electrocardiogram generation system, the method comprises: The heartbeat signal RCG collected by the millimeter wave radar is paired with the standard electrocardiogram signal ECG collected synchronously to obtain a training signal pair data set, wherein the heartbeat signal RCG reflects the physical characteristics of the heart movement, and the standard electrocardiogram signal ECG represents the bioelectric characteristics of the heart's electrical activity; Extracting time series features from the heartbeat signal RCG, learning forward and backward time series features of the heartbeat signal RCG through a bidirectional long short-term memory network Bi-LSTM to obtain a bidirectional time series feature sequence; The signal importance weights at different time points are calculated for the bidirectional time series feature sequence through an attention mechanism and weighted processing is performed to obtain a cardiodynamic weighted feature sequence; Inputting the electrocardiodynamic weighted feature sequence into a generator, performing signal conversion through a deep convolutional network, and obtaining an original electrocardiodynamic feature sequence; The original ECG feature sequence is optimized by using a dual discriminator structure to obtain a corrected ECG feature sequence, wherein the dual discriminator structure comprises a main discriminator and an auxiliary discriminator, wherein the main discriminator verifies the authenticity, and the auxiliary discriminator verifies the signal consistency; The corrected electrocardiogram feature sequence is post-processed by inverse normalization and smoothing filtering to obtain a target electrocardiogram signal.

2. The method according to claim 1, characterized in that The step of extracting the time series features of the heartbeat signal RCG, learning the forward and backward time series features of the heartbeat signal RCG through a bidirectional long short-term memory network Bi-LSTM, and obtaining a bidirectional time series feature sequence specifically includes: Inputting the heartbeat signal RCG into two bidirectional long short-term memory network layers for feature extraction to obtain an initial feature sequence, wherein each long short-term memory network includes an input gate, a forget gate and an output gate; According to the input vector at the current moment and the state at the previous moment, the activation vectors of the input gate, forget gate and output gate are calculated; Using the activation vector in combination with the weight matrix and the bias matrix, updating the state information of the memory unit; Based on the updated memory unit state and the activation vector of the output gate, generating the output feature at the current moment through an activation function; The output features of all the time steps are combined into a time series feature sequence as the bidirectional time series feature of the heartbeat signal RCG.

3. The method according to claim 1, characterized in that The step of calculating the signal importance weights at different time points of the bidirectional time series feature sequence through an attention mechanism and performing weighted processing to obtain a cardiodynamic weighted feature sequence specifically includes: learning global dependency information from the bidirectional temporal feature sequence through an attention layer; The softmax function is used to calculate the correlation score between the output sequence at different time points and the output at the current time point; The bidirectional time series feature sequence is weighted according to the calculated correlation degree score to obtain the electrocardiodynamic weighted feature sequence.

4. The method according to claim 1, characterized in that: The step of optimizing the original ECG feature sequence using a dual discriminator structure specifically includes: Using the main discriminator to receive the original ECG feature sequence and the real ECG signal as input, and calculating the authenticity probability through a four-layer one-dimensional convolutional network; Using the auxiliary discriminator to receive the original ECG feature sequence and the heartbeat signal RCG as input, verify signal consistency and calculate auxiliary loss; The original ECG feature sequence is optimized according to the authenticity probability and the auxiliary loss to obtain the corrected ECG feature sequence.

5. The method according to claim 4, characterized in that The post-processing step of performing denormalization and smoothing filtering on the corrected ECG feature sequence specifically includes: Performing a denormalization process on the corrected ECG characteristic sequence to restore the signal amplitude; The denormalized signal sequence is smoothed and filtered to eliminate high-frequency noise to obtain the target electrocardiogram signal.

6. The method according to claim 1, characterized in that After the step of performing post-processing of denormalization and smoothing filtering on the corrected ECG feature sequence to obtain a target ECG signal, the method further includes: Segmenting the target electrocardiogram signal according to a preset time window to obtain a plurality of signal segments, and performing feature extraction on each of the signal segments, including calculating the heart rate, and extracting waveform features of the P wave, QRS wave, and T wave; Comparing the waveform characteristics with a preset normal electrocardiogram characteristic range to determine whether there is an abnormality; When an abnormality is detected, the abnormality type is graded according to the degree of deviation of the waveform characteristics to obtain an abnormality level; Selecting a corresponding prompting method according to the abnormality level, the prompting method including display reminder, sound alarm or remote notification; The abnormality type, the abnormality level and the corresponding target electrocardiogram signal are stored in a database.

7. The method according to claim 6, characterized in that After the step of performing post-processing of denormalization and smoothing filtering on the corrected ECG feature sequence to obtain a target ECG signal, the method further includes: Counting the time regularity of anomalies in the database, and performing distribution analysis on the anomaly types; and establishing a correlation model between user activities and the anomaly types based on the distribution analysis results; Generate activity suggestions for the user's daily activities based on the association model, the activity suggestions including activity intensity adjustment and activity time arrangement; Generate an ECG monitoring analysis report regularly, the report including statistical information of the abnormality type, the changing trend of the abnormality level and recommended preventive measures.

8. A non-contact electrocardiogram generation system, characterized in that: The non-contact electrocardiogram generating system includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to enable the non-contact electrocardiogram generating system to perform the method described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that: When the instructions are executed on a non-contact electrocardiogram generating system, the non-contact electrocardiogram generating system is caused to execute the method according to any one of claims 1 to 7.

10. A computer program product, characterized in that When the computer program product runs on a non-contact electrocardiogram generating system, the non-contact electrocardiogram generating system is caused to perform the method according to any one of claims 1 to 7.

Images