Millimeter wave electrocardiogram reconstruction system and method based on adaptive MODWT and CNN-BiLSTM-CA

Abstract

The invention discloses a millimeter wave electrocardiogram reconstruction system and method based on adaptive MODWT and CNN-BiLSTM-CA, and the method comprises the steps: collecting a tiny phase displacement signal of a target thoracic cavity region, and carrying out the multi-band decomposition and key component adaptive screening of an original signal through combining with a multi-scale stationary wavelet decomposition algorithm; and introducing a channel attention mechanism to enhance key information representation, inputting a reconstruction signal into a deep learning model combining a convolutional neural network and a bidirectional long-short term memory network, completing time sequence modeling and nonlinear mapping, and outputting a reconstruction waveform highly consistent with a standard electrocardiogram. According to the method, the signal reduction capacity under the complex interference condition is remarkably improved, high-precision and privacy-friendly remote physiological signal monitoring can be achieved under the condition of not depending on a lead electrode, and the method is superior to a traditional baseline model in the aspects of waveform reduction precision, signal time sequence consistency, model generalization capacity and the like; the method is suitable for various application scenes such as intelligent medical treatment.

Description

Technical Field

[0001] This invention relates to the fields of biomedical engineering, radar signal processing and artificial intelligence, and in particular to a millimeter-wave electrocardiogram reconstruction system and method based on adaptive MODWT and CNN-BiLSTM-CA. Background Technology

[0002] Electrocardiography (ECG), as a key vital sign reflecting the electrical activity of the heart, has wide applications in clinical cardiovascular disease diagnosis, rehabilitation monitoring, and telemedicine. However, traditional ECG signal acquisition methods rely on direct contact between patch electrodes and the skin, which not only affects the comfort and daily activities of the subject but is also prone to signal loss or distortion due to factors such as electrode loosening, skin interference, or poor contact.

[0003] With the development of non-contact sensing technology, millimeter-wave radar, with its advantages of high resolution, strong penetration, and strong resistance to environmental interference, has become a potential solution for non-contact reconstruction of electrocardiograms (ECGs). It indirectly infers cardiac electrical activity by monitoring minute movements of the chest cavity (such as micro-displacements on the body surface caused by heartbeats). However, millimeter-wave radar reflected signals suffer from strong noise backgrounds and weak features, making direct ECG reconstruction a significant challenge.

[0004] Against this backdrop, deep learning methods have proven powerful in complex pattern recognition and signal reconstruction tasks, especially the combination of convolutional neural networks (CNNs) and recurrent neural networks (RNNs), which can effectively extract latent ECG features from radar signals for accurate reconstruction. However, a systematic and integrated non-contact ECG reconstruction method is still lacking, particularly in signal preprocessing and deep model architecture design, where gaps urgently need to be addressed. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a millimeter-wave electrocardiogram reconstruction system and method based on adaptive MODWT and CNN-BiLSTM-CA, which is based on millimeter-wave radar for non-contact electrocardiogram reconstruction.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The millimeter-wave electrocardiogram reconstruction system based on adaptive MODWT and CNN-BiLSTM-CA provided by this invention includes a millimeter-wave radar acquisition module, a vital signs signal extraction module, a heartbeat component extraction module, an electrocardiogram reconstruction module, and an electrocardiogram output module; The millimeter-wave radar acquisition module is used to acquire the raw echo signal of the target's chest cavity region using millimeter-wave radar; The vital signs signal extraction module is used to preprocess the echo signal to obtain an effective physiological displacement signal sequence; it is used to perform signal preprocessing. The heartbeat component extraction module is used to perform multi-scale decomposition on the effective physiological displacement signal sequence using stationary wavelet transform MODWT, extract mid-to-high frequency coefficients containing electrocardiogram information, and reconstruct the coefficients of key scales to obtain a net signal with enhanced features. The electrocardiogram reconstruction module reconstructs the electrocardiogram by constructing a CNN-BiLSTM-CA deep neural network. It is used to input the net signal into a convolutional neural network (CNN) to extract spatial features, and then input it into a bidirectional long short-term memory network (BiLSTM) for time series modeling. The electrocardiogram output module is used to send the feature vector output by the BiLSTM into a fully connected layer for regression mapping and output the normalized reconstructed electrocardiogram signal.

[0007] Furthermore, the vital signs signal extraction module includes range-dimensional FFT, static clutter suppression, target range cell extraction, phase unwrapping, and differential calculus; The distance-dimensional FFT performs a fast Fourier transform on the echo signal, converting the time-domain signal into a distance-velocity spectrum; The static clutter suppression filters out interference signals generated by stationary objects in the environment; The target distance unit is extracted, and the distance unit corresponding to the chest cavity movement is locked to eliminate interference from non-target areas; The phase unwrapping and differential calculation are used to solve the radar phase change and extract the micro-displacement signal caused by the heartbeat.

[0008] Furthermore, the heartbeat component extraction module includes MODWT decomposition, determination of the effective decomposition layer number, dynamic screening of heartbeat components, and multi-resolution analysis; The adaptive MODWT decomposition performs multi-scale stationary wavelet transform on the preprocessed signal, decomposing it into components of different frequency bands. The effective number of decomposition layers is determined, the optimal number of decomposition layers is dynamically determined, and the heartbeat frequency band is determined. The dynamic screening of heartbeat components determines key scale components related to heartbeat based on energy contribution. The multi-resolution analysis reconstructs the selected scale coefficients to generate a net signal that enhances heartbeat characteristics and suppresses noise.

[0009] Furthermore, the electrocardiogram reconstruction module includes a channel attention module, a CNN module, and a BiLSTM module; The channel attention module (CA) weights the multiple components of the MODWT output to enhance key heartbeat features; it is used to assign different importance weights to the multiple heartbeat components extracted by multi-resolution analysis. The CNN module extracts spatial features through one-dimensional convolution: it achieves layer-by-layer abstraction through convolution kernels and improves generalization ability through BN+ReLU. The BiLSTM module employs bidirectional temporal modeling, including forward LSTM and backward LSTM. The forward LSTM simulates the direction of cardiac electrical signal conduction by learning the temporal evolution of the electrocardiogram in the forward direction; the backward LSTM captures the repolarization process and rhythm association by learning the electrocardiogram dependencies in the backward direction.

[0010] Furthermore, it also includes a model training module for training the model based on synchronously acquired real electrocardiogram data. The loss function is constructed by combining L1 loss and mean squared error, an early stop strategy is introduced, and the Adam optimizer is used for iterative model training.

[0011] Furthermore, it also includes an output evaluation module for evaluating the reconstructed electrocardiogram using the Pearson correlation coefficient (PCC) and mean squared error (RMSE).

[0012] The millimeter-wave electrocardiogram reconstruction method based on adaptive MODWT and CNN-BiLSTM-CA provided by this invention includes the following steps: S1: Use millimeter-wave radar to collect echo signals from the target's chest cavity region and extract phase information; S2: Preprocess the echo signal to obtain an effective physiological displacement signal sequence; S3: The effective physiological displacement signal sequence is decomposed into multiple scales using stationary wavelet transform MODWT, and mid-to-high frequency coefficients containing electrocardiogram information are extracted. The coefficients of key scales are reconstructed to obtain a net signal with enhanced features. S4: Input the net signal into a convolutional neural network (CNN) to extract spatial features, and then input it into a bidirectional long short-term memory (BiLSTM) network for time series modeling; S5: The feature vector output by the BiLSTM is fed into a fully connected layer for regression mapping, and the normalized reconstructed electrocardiogram signal is output.

[0013] Furthermore, the stationary wavelet transform MODWT decomposition process is an adaptive wavelet decomposition process, specifically including: S31: By setting energy convergence criteria, the minimum number of effective decomposition layers that satisfy the heartbeat frequency band coverage is dynamically determined; S32: Integrate the frequency band energy of each decomposition layer and select the three scale coefficient components that contribute the most to the energy in the target frequency band. S33: Use the multi-resolution analysis results composed of the three components as input to the deep neural network.

[0014] Furthermore, the Bidirectional Long Short-Term Memory (BiLSTM) network employs bidirectional temporal modeling, with the specific steps as follows: S41: Construct a forward LSTM to simulate the direction of cardiac electrical signal conduction by learning the temporal evolution of electrocardiograms in a forward manner; S42: Construct a backward LSTM to capture the association between repolarization processes and rhythms by learning ECG dependencies in reverse; S43: Bidirectional splicing, used to fuse the entire context of the depolarization-repolarization process; S44: Tanh normalization, used to constrain the output to the standard ECG voltage range.

[0015] The beneficial effects of this invention are as follows: This invention provides a millimeter-wave electrocardiogram (ECG) reconstruction system and method based on adaptive MODWT and CNN-BiLSTM-CA. This method falls within the technical application scope of ECG monitoring, remote health diagnosis, and medical artificial intelligence. It overcomes the limitations of traditional contact-based ECG acquisition methods and the problems of noise interference and feature extraction difficulties in the process of reconstructing ECGs from millimeter-wave radar data. By combining multi-scale wavelet decomposition (MODWT) preprocessing with a non-contact ECG reconstruction method based on CNN-BiLSTM deep neural network modeling, it ultimately achieves automatic reconstruction of radar signals into high-fidelity ECGs. This method first acquires signals using millimeter-wave radar: the target chest cavity region is illuminated by millimeter-wave radar to acquire echo signals reflecting the minute displacements caused by cardiac motion; signal preprocessing: the raw radar signal is subjected to multi-scale stationary wavelet decomposition (MODWT) to extract features of different frequency bands, followed by denoising and reconstruction to obtain a high-quality input signal sequence; feature modeling and reconstruction network construction: a CNN-BiLSTM deep neural network architecture is constructed, where CNN is used to extract local spatial features and bidirectional LSTM is used to learn temporal dependencies and contextual dynamic features; model training: synchronously acquired real ECG signals are used as supervision signals, and the network parameters are trained by minimizing the reconstruction error; ECG reconstruction and output: the preprocessed radar signal is input into the trained network, and the reconstructed ECG signal that is highly consistent with the real ECG waveform is output.

[0016] This method is non-contact: completely avoiding traditional electrode patches, improving comfort and continuous monitoring capabilities; strong multi-scale feature preservation: utilizing MODWT to decompose radar signals at multiple levels, preserving different frequency features and effectively suppressing noise interference; strong temporal modeling capability: bidirectional LSTM has a better ability to express the context of ECG signals, achieving higher fidelity reconstruction; strong scalability: the method can be widely applied to various scenarios such as remote monitoring, sleep detection, and smart healthcare; simple implementation and efficient training: based on a mature deep learning architecture, it is easy to port and deploy. This method and system are suitable for scenarios such as health monitoring, telemedicine, and intelligent care. To enhance the accuracy of reconstructed ECGs and the ability to restore waveform structure, this method combines a deep neural network architecture of convolutional neural network (CNN) and bidirectional long short-term memory network (BiLSTM). The CNN module is used to mine local spatial features and edge mutations, while the BiLSTM module is used to model the contextual temporal dependencies of ECG signals, thereby achieving higher fidelity ECG waveform reconstruction. Compared to traditional unidirectional models, bidirectional LSTM significantly enhances the ability to capture the temporal characteristics of key ECG components such as the P wave, QRS complex, and T wave.

[0017] This method employs end-to-end supervised learning, utilizing real electrocardiograms (ECGs) acquired synchronously with radar data as supervisory signals. By combining L1 loss and mean squared error for network training, it effectively guides the model to achieve accurate reconstruction in both time and amplitude dimensions. In the evaluation of the reconstruction results, this method demonstrates low error and high correlation performance across multiple test scenarios. By replacing electrodes with radar for non-contact signal acquisition, this method significantly improves user comfort and long-term continuous monitoring capabilities, making it particularly suitable for high-frequency, low-interference indoor health monitoring applications. Compared to traditional patch-based ECG acquisition methods, this system can operate stably without sensory input, avoiding issues such as electrode detachment and skin allergies. Furthermore, compared to image-based methods, this system requires no camera and is independent of lighting conditions, offering higher privacy protection and environmental adaptability. The overall solution boasts advantages such as high signal processing accuracy, strong modeling capabilities, and low deployment costs, providing a more efficient, reliable, and practical technical path for intelligent healthcare and home health monitoring.

[0018] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0019] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following drawings are provided for illustration.

[0020] Figure 1 This is the overall system architecture diagram.

[0021] Figure 2 This is the phase signal of the target range unit.

[0022] Figure 3 This is a graph showing the decomposition coefficients of MODWT at various scales.

[0023] Figure 4 The image shows the result after MODWT signal reconstruction.

[0024] Figure 5 This is a diagram of the overall network structure of CNN-BiLSTM.

[0025] Figure 6 This is a diagram of the BiLSTM network structure.

[0026] Figure 7 This is a loss curve for the training process.

[0027] Figure 8 This is a comparison chart of the output electrocardiogram and the true electrocardiogram.

[0028] Figure 9 Error / evaluation metric graphs for different test data. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0030] Example 1 like Figure 1 As shown, Figure 1 The system overall structure diagram provided in this embodiment is a millimeter-wave electrocardiogram reconstruction system based on adaptive MODWT and CNN-BiLSTM-CA. It is characterized by including a millimeter-wave radar acquisition module, a vital sign signal extraction module, a heartbeat component extraction module, an electrocardiogram reconstruction module, an electrocardiogram output module, a model training module, and an output evaluation module. The millimeter-wave radar acquisition module is used to acquire the raw echo signal of the target's chest cavity region using millimeter-wave radar; The vital signs signal extraction module is used to preprocess the echo signal, including setting the distance from doors and windows, removing background interference, and obtaining an effective physiological displacement signal sequence; it is used to implement signal preprocessing. The heartbeat component extraction module is used to perform multi-scale decomposition of the effective physiological displacement signal sequence using stationary wavelet transform (MODWT), extract mid-to-high frequency coefficients containing electrocardiogram information, and reconstruct the coefficients of key scales to obtain a net signal with enhanced features. The MODWT preprocessing supports custom selection of mother wavelet (such as sym4 or db4) and can adjust the number of decomposition layers to adapt to different frequency band distribution characteristics.

[0031] The electrocardiogram reconstruction module reconstructs the electrocardiogram by constructing a CNN-BiLSTM-CA deep neural network module. It is used to input the net signal into a convolutional neural network (CNN) to extract spatial features, and then input it into a bidirectional long short-term memory network (BiLSTM) for time series modeling. The CNN part in this embodiment includes a one-dimensional convolution module and a batch normalization module. The BiLSTM part includes two LSTM units in opposite directions, and their outputs are fused by a fully connected layer to generate the final prediction signal.

[0032] The electrocardiogram output module is used to send the feature vector output by the BiLSTM into a fully connected layer for regression mapping and output the normalized reconstructed electrocardiogram signal.

[0033] The vital signs signal extraction module described in this embodiment includes range-dimensional FFT, static clutter suppression, target range cell extraction, phase unwrapping, and differential calculus. The distance-dimensional FFT performs a fast Fourier transform on the echo signal, converting the time-domain signal into a distance-velocity spectrum; The static clutter suppression filters out interference signals generated by stationary objects in the environment (such as furniture and walls); The target distance unit is extracted, and the distance unit corresponding to the chest cavity movement is locked to eliminate interference from non-target areas; The phase unwrapping and differential calculation are used to solve the radar phase change and extract the micro-displacement signal caused by heartbeat (differential processing eliminates the influence of respiratory motion). The heartbeat component extraction module described in this embodiment is used to extract heartbeat components based on adaptive MODWT, including MODWT decomposition, determination of effective decomposition layer number, dynamic screening of heartbeat components, and multi-resolution analysis. The adaptive MODWT decomposition performs multi-scale stationary wavelet transform on the preprocessed signal, decomposing it into components of different frequency bands. The determination of the effective number of decomposition layers and the dynamic determination of the optimal number of decomposition layers can avoid redundant calculations and focus on the heartbeat frequency band of 0.8-5Hz. The dynamic screening of heartbeat components is based on energy contribution analysis, selecting three key scale components related to heartbeat. The multi-resolution analysis reconstructs the selected scale coefficients to generate a net signal that enhances heartbeat characteristics and suppresses noise. The electrocardiogram reconstruction module described in this embodiment is used to reconstruct the electrocardiogram based on CNN-BiLSTM-CA, and includes a channel attention module, a CNN module, and a BiLSTM module. The Channel Attention Module (CA) weights the multiple components of the MODWT output to enhance key heartbeat features (such as QRS waves); it is used to assign different importance weights to multiple heartbeat components extracted by multi-resolution analysis. Key cardiac features include typical electrophysiological waveforms such as the P wave, QRS complex, and T wave, whose energy distribution varies across different decomposition layers. To achieve weighted processing of different key cardiac features, power spectral density analysis is first used to assess the energy proportion of each component within the cardiac frequency band, and the component with higher energy contribution is selected as the dominant feature channel. Subsequently, the channel attention module dynamically generates an importance score for each cardiac feature channel based on the statistical characteristics of each channel, using global average pooling and learnable weight parameters. This achieves enhanced expression of key features and suppression of redundant features, thereby improving the accuracy and stability of ECG reconstruction.

[0034] In this embodiment, the wavelet transform denoising module dynamically determines the number of decomposition layers based on the energy convergence criterion and performs integral evaluation on the frequency band energy of each decomposition layer.

[0035] The CNN module extracts spatial features through one-dimensional convolution: it can achieve layer-by-layer abstraction through convolution kernels and improve generalization ability through BN+ReLU. The CNN module described in this embodiment consists of three one-dimensional convolutional sub-modules and three one-dimensional transposed convolutional sub-modules, used to implement downsampling and inverse upsampling.

[0036] The BiLSTM module employs bidirectional temporal modeling, including forward LSTM and backward LSTM. The forward LSTM simulates the direction of cardiac electrical signal conduction by learning the temporal evolution of the electrocardiogram in the forward direction; the backward LSTM captures the repolarization process and rhythm association by learning the electrocardiogram dependencies in the backward direction. The simulated cardiac electrical signal conduction direction is P wave → QRS wave → T wave; the capture of the repolarization process and its correlation with rhythm is T wave → QRS wave → P wave; the physiological significance of the bidirectional timing model is to capture the dominant timing of ventricular depolarization (rapid rising edge of QRS wave) through forward flow, and to strengthen the correlation between repolarization and rhythm (such as the influence of T wave morphology and QT interval on P wave) through backward flow.

[0037] In this embodiment, forward propagation captures the electrical signal conduction timing, backward propagation models the physiological rhythm closed loop, and feature splicing is combined to achieve high-fidelity reconstruction of the electrocardiogram waveform. Its bidirectional design significantly outperforms the unidirectional model in restoring asymmetric ECG features (such as T-wave inversion and prolonged PR interval). In this embodiment, the number of hidden units in both the forward and backward LSTM can be set to 64. The bidirectional temporal output is spliced ​​along the feature dimensions (Concat), and the spliced ​​output is then input into a fully connected layer for regression prediction, which can fuse bidirectional contextual information: the output dimension can be set to 128, equivalent to the sum of 64 forward and 64 backward dimensions.

[0038] The ECG feature enhancement strategy in this embodiment divides ECG components into forward LSTM and backward LSTM for analysis. For the P wave, forward LSTM detects the initial atrial excitation point, while backward LSTM correlates with the atrial recovery process during the TP interval. For the QRS complex, forward LSTM identifies the R wave peak value and duration width, while backward LSTM enhances the ST segment transition features. For the T wave, forward LSTM predicts the ventricular repolarization endpoint, while backward LSTM correlates with the QT interval and the start point of the P wave in the next cycle. This bidirectional modeling can reduce the QRS width prediction error by 37%. In this embodiment, the BiLSTM input receiving channel is attention-weighted by CA to the MODWT component, and the weight of key heartbeat features (such as the R wave) is increased by 3-5 times. The BiLSTM output is compressed by a fully connected layer, and the ECG voltage value at each time point is predicted by regression. Tanh activation is used to accurately constrain the output range to the standard medical ECG value range [-1,1]mV.

[0039] The model training module is used to train the model based on synchronously acquired real electrocardiogram data. It uses L1 loss and mean squared error to jointly construct the loss function, and uses the Adam optimizer for iterative model training. At the same time, it introduces an early stop strategy to improve training efficiency.

[0040] The output evaluation module is used to evaluate the accuracy and robustness of the reconstructed electrocardiogram using Pearson correlation coefficient (PCC) and mean squared error (RMSE).

[0041] The millimeter-wave radar acquisition module described in this embodiment uses a 77GHz millimeter-wave radar with a bandwidth of 4GHz, a sampling frequency of 20 frames / second, and a distance resolution of 4cm.

[0042] The preprocessing module described in this embodiment includes a distance threshold setting unit and a background interference removal unit, which are used to remove background interference and filter distance unit data with a high signal-to-noise ratio.

[0043] The system described in this embodiment is deployed on an edge computing device, supporting real-time data collection and electrocardiogram output in contactless scenarios such as smart mattresses, office chairs, and nursing chairs, with an inference latency of less than 70ms.

[0044] The millimeter-wave electrocardiogram reconstruction method based on adaptive MODWT and CNN-BiLSTM-CA provided in this embodiment includes the following steps: S1: Use millimeter-wave radar to collect echo signals from the target's chest cavity region and extract phase information; S2: Preprocess the echo signal, including setting the distance to doors and windows, removing background interference, and filtering the distance cell data with a high signal-to-noise ratio through amplitude analysis to obtain an effective physiological displacement signal sequence; S3: The effective physiological displacement signal sequence is decomposed into multiple scales using stationary wavelet transform (MODWT), mid-to-high frequency coefficients containing electrocardiogram information are extracted, and the coefficients of key scales are reconstructed to obtain a net signal with enhanced features. S4: Input the net signal into a convolutional neural network (CNN) to extract spatial features, and then input it into a bidirectional long short-term memory network (BiLSTM) for time series modeling; In this embodiment, the convolutional neural network has a channel attention module (CA) after the input layer, which is used to assign different importance weights to the multiple heartbeat components extracted by multi-resolution analysis. S5: The feature vector output by the BiLSTM is fed into a fully connected layer for regression mapping, and the normalized reconstructed electrocardiogram signal is output.

[0045] The stationary wavelet transform MODWT decomposition process described in this embodiment is an adaptive wavelet decomposition process, specifically including: S31: By setting energy convergence criteria, the minimum number of effective decomposition layers that satisfy the heartbeat frequency band coverage is dynamically determined; S32: Integrate the frequency band energy of each decomposition layer and select the three scale coefficient components that contribute the most to the energy in the target frequency band. S33: Use the multi-resolution analysis results composed of the three components as input to the deep neural network.

[0046] The Bidirectional Long Short-Term Memory (BiLSTM) network described in this embodiment employs bidirectional temporal modeling, and the specific steps are as follows: S41: Construct a forward LSTM to simulate the direction of cardiac electrical signal conduction by learning the temporal evolution of electrocardiograms in a forward manner; S42: Construct a backward LSTM to capture the association between repolarization processes and rhythms by learning ECG dependencies in reverse.

[0047] S43: Bidirectional splicing, used to fuse the entire context of the depolarization-repolarization process; S44: Tanh normalization, used to constrain the output to the standard ECG voltage range.

[0048] In the BiLSTM network described in this embodiment, the hidden state dimension of both the forward and backward LSTM units is 64, and the outputs are concatenated and input into the fully connected layer for regression prediction.

[0049] The convolutional neural network described in this embodiment consists of three one-dimensional convolutional sub-modules and three one-dimensional transposed convolutional sub-modules. The one-dimensional convolutional sub-modules are used to perform downsampling and extract local features, while the one-dimensional transposed convolutional sub-modules are used to perform inverse upsampling and gradually restore the time dimension.

[0050] The model described in this embodiment is deployed on an edge computing device, supporting real-time data collection and electrocardiogram output in contactless scenarios such as smart mattresses, office chairs, and nursing chairs, with an inference latency of less than 70ms.

[0051] In this embodiment, the Pearson correlation coefficient (PCC) and mean squared error (RMSE) are used as performance indicators to evaluate the accuracy and robustness of the reconstructed electrocardiogram during the output effect evaluation process.

[0052] In the model training process described in this embodiment, real electrocardiograms collected synchronously are used as label data. A loss function is constructed based on L1 loss and mean squared error. The Adam optimizer is used for iterative model training, and an early stopping strategy is introduced to improve training efficiency.

[0053] Example 2 This embodiment details the construction of a millimeter-wave electrocardiogram (ECG) reconstruction system using MODWT and CNN-BiLSTM-CA, and illustrates the method of ECG reconstruction. This method is a non-contact ECG reconstruction approach, demonstrating the complete processing flow from millimeter-wave radar signal acquisition to ECG output. The system mainly includes a millimeter-wave radar acquisition module, a vital sign signal extraction module for wavelet transform denoising preprocessing of the radar signal, a heartbeat component extraction module for constructing a CNN-BiLSTM-CA deep neural network, and a signal reconstruction module. This method is suitable for human physiological monitoring and remote ECG sensing scenarios. The method first acquires micro-displacement signals of the target chest cavity region using millimeter-wave radar, then performs denoising processing on the signal using multi-scale stationary wavelet decomposition, and then uses a convolutional neural network (CNN) and a bidirectional long short-term memory network (BiLSTM) to construct a deep model, achieving temporal modeling and feature extraction of the original radar signal, and finally outputting a reconstructed signal that is highly consistent with the actual ECG waveform.

[0054] The technical details of this embodiment will be described in detail below, including system structure, signal acquisition, preprocessing, modeling method, and reconstruction evaluation.

[0055] S1. Construct and initialize each functional module. The functional modules include a millimeter-wave radar acquisition module, a vital sign signal extraction module (preprocessing module, multi-scale wavelet decomposition module), a heartbeat component extraction module (deep neural network modeling module), and an electrocardiogram reconstruction module (signal reconstruction module) for a non-contact electrocardiogram reconstruction system. The modules are connected in the order of data flow. It can acquire minute surface displacement signals in the chest cavity region under non-contact conditions and restore high-fidelity electrocardiogram waveforms by combining deep learning methods, realizing accurate mapping from radar signals to medical signals.

[0056] Compared to traditional adhesive electrode acquisition systems, this system enables continuous, high-fidelity ECG monitoring without relying on electrode contact with the skin. This avoids interference caused by electrode detachment or skin sensitivity, making it particularly suitable for special groups such as the elderly, pregnant women, and comatose patients. The entire system consists of multiple functional modules working together, forming a complete non-contact physiological signal reconstruction path from signal acquisition, noise removal, feature extraction, model prediction to final ECG output.

[0057] S2. Millimeter-wave radar signal acquisition and preprocessing: Millimeter-wave radar is used to acquire echo signals from the target's thoracic cavity region, extracting phase information. Signal filtering and preprocessing are then performed using methods such as setting distance thresholds and removing background interference to obtain effective physiological displacement signal sequences, as detailed below: S201 This embodiment uses an AWR1642 millimeter-wave radar module, operating at a frequency of 77 GHz, with a bandwidth of 4 GHz, a range resolution of 4 cm, and a sampling frequency of 20 frames / second. The radar is installed within 30 to 60 cm of the subject's chest, requiring no fixed contact device, and can be flexibly deployed on medical beds, walls, or wearable devices; To enhance signal stability, the acquisition environment should be relatively quiet, and the subject should remain in a resting seated position. By setting an appropriate range gate, the target cell with the strongest echo energy is selected, and the phase change sequence is extracted from it as an indirect reflection of the chest wall micro-vibration displacement caused by cardiac activity. S203 phase signal as Figure 2 As shown, it exhibits certain periodic changes, but is also affected by factors such as respiratory movements, posture changes, and electromagnetic interference, including issues such as high-frequency noise and baseline drift. Because it is taken from non-contact radar and is not a specific lead type, its structural characteristics differ from those of a standard electrocardiogram, requiring further extraction of effective components. To further improve signal accuracy, the S204 integrates a target tracking algorithm to dynamically lock onto the chest cavity region, eliminate non-target reflection interference, and improve acquisition stability.

[0058] S3. MODWT Wavelet Decomposition and Signal Reconstruction: The preprocessed signal is decomposed into multiple scales using stationary wavelet transform (MODWT) to extract mid-to-high frequency coefficients containing ECG information. The coefficients at key scales are reconstructed to obtain a net signal with enhanced features. Preprocessing of the phase signal includes setting a distance threshold to remove background interference and filtering distance cell data with high signal-to-noise ratios through amplitude analysis. The MODWT wavelet decomposition process is an adaptive wavelet decomposition process, specifically including: dynamically determining the minimum effective decomposition layer to satisfy the heartbeat frequency band coverage by setting an energy convergence criterion, avoiding redundant calculations; integrating and evaluating the frequency band energy of each decomposition layer to select the three scale coefficient components with the highest energy contribution in the target frequency band; and using the multi-resolution analysis results composed of these three components as input to a deep neural network to achieve adaptive extraction of the heartbeat principal component. To achieve signal denoising and feature enhancement, S301 employs the Multi-Scale Stationary Wavelet Decomposition (MODWT) method to process the original phase signal. Compared to traditional DWT, MODWT possesses translation invariance and redundancy, exhibiting stronger robustness in processing non-stationary biological signals. S302 In this embodiment, sym4 is selected as the wavelet mother function to perform a 6-level decomposition on the original signal, obtaining six sets of detail coefficients and scaling coefficients, such as Figure 3 As shown. Each layer corresponds to different frequency band characteristics, with lower layers focusing on high-frequency changes (such as instantaneous spikes) and higher layers focusing on low-frequency trends (such as periodic rhythms). S303's power spectral density analysis at various scales revealed that the coefficients in layers 3 to 5 exhibited concentrated energy and a clear periodicity, corresponding to the QRS complex and T wave information in the electrocardiogram. The coefficients in other layers were primarily affected by noise or respiration. S304 selects the above three-layer coefficients for weighted reconstruction, effectively filtering out irrelevant frequency band interference. The reconstructed signal is as follows: Figure 4 As shown, the waveform is continuous and the rhythm is clear, which greatly improves the quality of subsequent model input data; The S305 reconstructed signal still retains some non-ideal factors (such as slight jitter), but compared with the original signal, it significantly reduces high-frequency clutter while preserving the target characteristics, thus enhancing learnability. S306 In the data post-processing stage, the present invention can further introduce operations such as normalization and bandpass filtering to further improve signal smoothness and boundary feature expression, and enhance the modeling effect.

[0059] S4. CNN-BiLSTM Deep Neural Network Modeling: The reconstructed signal is sequentially input into a convolutional neural network (CNN) to extract spatial features, and then input into a bidirectional long short-term memory network (BiLSTM) for time series modeling to capture the rhythmicity and waveform changes in the signal. The CNN has a channel attention module (CA) after the input layer to assign different importance weights to multiple heartbeat components extracted by multi-resolution analysis, thereby enhancing key features and suppressing redundant components. The CNN consists of three one-dimensional convolutional submodules and three one-dimensional transposed convolutional submodules. The one-dimensional convolutional submodules are used for downsampling, extracting local features, and progressively enhancing cardiac electrophysiological features. The one-dimensional transposed convolutional submodules are used for inverse upsampling, gradually restoring the time dimension, and achieving feature reconstruction.

[0060] S401 This embodiment designs a deep model combining a convolutional neural network (CNN) and a bidirectional long short-term memory network (BiLSTM) for electrocardiogram reconstruction of processed radar signal sequences. The overall structure is as follows: Figure 5 As shown. Figure 5 The diagram shows the ECG reconstruction network structure based on CNN-BiLSTM-CA proposed in this embodiment. The overall structure includes a channel attention module (CA), a convolutional neural network module (CNN), a deconvolution module (TransCNN), a bidirectional long short-term memory network module (BiLSTM), and a fully connected output module (FC). The CA module weights the multiple heartbeat components extracted after MODWT decomposition, dynamically allocating weights based on the contribution of each channel to enhance the expressive power of key heartbeat features (such as QRS waves and T waves) and suppress redundant interference. The weighted features are first encoded by the CNN module, which consists of one-dimensional convolution (Conv1D), batch normalization, Tanh activation function, and max pooling, used to extract multi-scale local heartbeat temporal features layer by layer. Subsequently, the features are restored by the TransCNN module, gradually recovering the original feature dimensions. Next, the features are fed into the BiLSTM module, which utilizes bidirectional temporal modeling capabilities to simultaneously capture the positive and negative physiological dependence features of the ECG, improving the reconstruction integrity of each stage from the P wave to the T wave. Finally, the final electrocardiogram sequence is output through multiple fully connected layers (FC) and Dropout regularization, achieving accurate reconstruction of key electrocardiographic waveforms; The S402 CNN module consists of two one-dimensional convolutional layers. The first layer has a kernel size of 5 and 16 channels, while the second layer has a kernel size of 3 and 32 channels. Each convolutional layer is followed by Batch Normalization (BN) and ReLU activation functions to extract local mutations and edge features, and to suppress overfitting. The S403 CNN output feature map is fed into the BiLSTM module. The BiLSTM structure is as follows: Figure 6 As shown, it contains two LSTM units with opposite time directions, and the number of hidden units is set to 64 in each case. BiLSTM can learn forward and backward temporal dependencies simultaneously, which helps to preserve symmetrical structures and rhythm information such as P waves, QRS waves, and T waves in ECG signals; Figure 6 The diagram shows a Bidirectional Long Short-Term Memory (BiLSTM) network structure, used for bidirectional time-series modeling of time series signals. In the diagram, x... t Let h represent the input features at time t. t This represents the hidden state output at that moment; the LSTM (Long Short-Term Memory) can effectively capture long-term dependency information in the sequence. Unlike traditional LSTM, which is based solely on forward propagation, BiLSTM consists of a set of forward LSTMs and a set of backward LSTMs running in parallel. These processes the input sequence along the forward and backward time axes, respectively, and then fuses the hidden states from both directions at each time step. This allows for the simultaneous use of past and future contextual information, achieving a more comprehensive representation of temporal features. In this embodiment, the BiLSTM module is used to capture the bidirectional physiological dependencies between key waveforms such as the P wave, QRS complex, and T wave in the electrocardiogram (ECG), overcoming the difficulty of unidirectional modeling in depicting issues such as the impact of ventricular repolarization on the preceding depolarization phase, and improving the continuity and accuracy of ECG waveform reconstruction.

[0061] The output sequence of S404 BiLSTM is spliced ​​and then fed into a fully connected layer for regression prediction. Finally, the output is normalized to the [-1,1] interval by the Tanh activation function to match the numerical range of the standardized ECG signal. If the application scenario is sensitive to computing resources, the CNN module can be simplified to a single layer or replaced with a lightweight convolutional block (such as the MobileNet structure) to achieve low-power deployment.

[0062] S5. Network Training and Deployment: The network includes output layer mapping and ECG reconstruction. Specifically, the feature vector output by the BiLSTM is fed into a fully connected layer for regression mapping, outputting a normalized reconstructed ECG signal. In the BiLSTM network, the hidden state dimension of both the forward and backward LSTM units is 64. The outputs are concatenated and then input into the fully connected layer for regression prediction. The BiLSTM module includes one Bidirectional Long Short-Term Memory (BiLSTM) layer and three fully connected layers. The BiLSTM layer extracts time-dependent features from the ECG sequence simultaneously through forward and backward information flows to enhance rhythm modeling capabilities.

[0063] To train the aforementioned neural network model, S501 uses standard lead electrocardiograms acquired synchronously with radar signals as supervisory labels. The acquisition frequency is 500Hz to ensure sufficient temporal resolution. All training data for S502 are Z-score normalized to eliminate the impact of amplitude differences on model training. The S503 loss function uses a weighted combination, consisting of L1 Loss (mean absolute error) and RMSE (mean squared error), which respectively measure the overall fitting accuracy and the degree of local distortion. The weight parameters can be dynamically adjusted according to the distribution of the dataset; The S504 optimizer uses the Adam algorithm with an initial learning rate of 0.001, a batch size of 32, and 300 training rounds. An early stopping strategy is introduced to prevent overfitting. The S505 model training is based on the PyTorch platform, with a single training session taking approximately 45 minutes on an NVIDIA RTX3060 GPU. After the final model is exported, it is deployed on an NVIDIA Jetson Nano edge device, which can process 15 frames per second in inference mode to meet the needs of real-time applications. After being lightweighted, the S506 model has an inference latency of less than 70ms, an overall power consumption of less than 6W, supports long-term remote continuous operation, and has good deployability.

[0064] S6. Data Acquisition Scenarios and Platform Adaptation: The model is deployed on edge computing devices to support real-time data acquisition and ECG output in contactless scenarios such as smart mattresses, office chairs, and nursing chairs, meeting the requirements of low power consumption and long-term stable operation. The non-contact electrocardiogram (ECG) system of the present invention, S601, possesses excellent deployment flexibility and can be widely applied in various daily life and medical monitoring scenarios, including integrated deployment in locations such as smart mattresses, office chairs, nursing chairs, and ward walls. During operation, the system requires no direct contact with the human body; it achieves continuous reconstruction of high-fidelity ECGs solely through remote detection of the chest cavity area using millimeter-wave radar. During deployment, the radar's deployment angle and distance can be flexibly selected according to usage requirements, and it supports integration with synchronization devices, remote monitoring platforms, etc., meeting the real-time physiological signal monitoring needs of diverse scenarios such as hospitals, homes, and elderly care institutions.

[0065] This embodiment of S602 also supports Bluetooth / 5G wireless data transmission, realizing a front-end acquisition + back-end modeling mode, which is suitable for distributed deployment applications such as smart elderly care beds and remote hospital monitoring.

[0066] S7. Model Training and Optimization: The training process and results were analyzed. Standard electrocardiograms collected synchronously were used as label data. A loss function was constructed based on L1 loss and mean squared error. The Adam optimizer was used for iterative model training, and an early stopping strategy was introduced to improve training efficiency. The training process of the S701 network model is as follows: Figure 7 The figure shows the trend of the loss value with epochs during training. It shows a stable decreasing curve overall, indicating that the model is well trained and has strong convergence. S702 uses a validation set to evaluate the model's fitting ability. In most test samples, the model's predictions are highly consistent with the actual ECG, with waveforms that largely overlap and minimal error. S703 can compare and analyze the results of different training rounds and different parameter combinations, and optimize the network structure and learning strategy that are most suitable for specific collection environments and target groups.

[0067] S8. Training process evaluation: ECG reconstruction output, record the trend of model loss curve during training, monitor convergence speed and training stability through validation set to ensure excellent model fitting ability. S801 Comparison of the network output results of the present invention with a real electrocardiogram (ECG) Figure 8 The figure shows the waveform alignment between the reconstructed ECG sequence output by the model and the original real lead ECG. The main features (P wave, QRS wave, T wave) are accurately reproduced. Compared with the original signal, the reconstructed S802 signal exhibits stable rhythm and significantly reduced noise, demonstrating the advantages of this method in morphological restoration and feature reconstruction.

[0068] The S803 output can also be converted and mapped to different lead formats, adapting to the interface of standard hospital ECG analysis systems.

[0069] S9. Error assessment and generalization performance test: Output effect analysis and generalization evaluation: Compare and analyze the reconstruction results with the real electrocardiogram, and use indicators such as Pearson correlation coefficient, mean square error, and dynamic time regularization distance to evaluate the reconstruction accuracy and cross-sample generalization ability of the model. S901 This embodiment uses multiple test samples to evaluate the model's generalization ability, and the error statistics are as follows: Figure 9 The figure shows the mean squared error (RMSE) and Pearson correlation coefficient (PCC) of the model under different test data.

[0070] The S902 model has an average RMSE of 0.047 and a PCC greater than 0.92, indicating that the model has good overall predictive ability and time series fitting ability. The S903 model maintains stable output under different subjects and different data collection scenarios, indicating that it has good versatility and robustness, and is suitable for scenarios such as remote health monitoring, elderly care, and mobile healthcare. Compared to the comparison models (such as those without MODWT processing or using a unidirectional LSTM network), S904 is shown.

[0071] The system and method provided in this embodiment outperform those in all evaluation metrics. Employing strategies such as adaptive wavelet layer determination, weighted important feature components, feature reconstruction, and rhythm extraction, it significantly improves signal restoration capabilities under complex interference conditions. Furthermore, the system offers advantages such as low power consumption, wearability, and edge computing, enabling high-precision, privacy-friendly remote physiological signal monitoring without relying on lead electrodes. Experimental verification shows that this method surpasses traditional baseline models in waveform restoration accuracy, signal timing consistency, and model generalization ability, making it suitable for various application scenarios such as smart healthcare, elderly health management, and long-term remote monitoring.

[0072] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A millimeter-wave electrocardiogram reconstruction system based on adaptive MODWT and CNN-BiLSTM-CA, characterized in that, It includes a millimeter-wave radar acquisition module, a vital signs signal extraction module, a heartbeat component extraction module, an electrocardiogram reconstruction module, and an electrocardiogram output module; The millimeter-wave radar acquisition module is used to acquire the raw echo signal of the target's chest cavity region using millimeter-wave radar; The vital signs signal extraction module is used to preprocess the echo signal to obtain an effective physiological displacement signal sequence. Used to perform signal preprocessing; The heartbeat component extraction module is used to perform multi-scale decomposition on the effective physiological displacement signal sequence using stationary wavelet transform MODWT, extract mid-to-high frequency coefficients containing electrocardiogram information, and reconstruct the coefficients of key scales to obtain a net signal with enhanced features. The electrocardiogram reconstruction module reconstructs the electrocardiogram by constructing a CNN-BiLSTM-CA deep neural network. It is used to input the net signal into a convolutional neural network (CNN) to extract spatial features, and then input it into a bidirectional long short-term memory network (BiLSTM) for time series modeling. The electrocardiogram output module is used to send the feature vector output by the BiLSTM into a fully connected layer for regression mapping and output the normalized reconstructed electrocardiogram signal.

2. The millimeter-wave electrocardiogram reconstruction system based on adaptive MODWT and CNN-BiLSTM-CA according to claim 1, characterized in that, The vital signs signal extraction module includes range-dimensional FFT, static clutter suppression, target range cell extraction, phase unwrapping, and differential calculus. The distance-dimensional FFT performs a fast Fourier transform on the echo signal, converting the time-domain signal into a distance-velocity spectrum; The static clutter suppression filters out interference signals generated by stationary objects in the environment; The target distance unit is extracted, and the distance unit corresponding to the chest cavity movement is locked to eliminate interference from non-target areas; The phase unwrapping and differential calculation are used to solve the radar phase change and extract the micro-displacement signal caused by the heartbeat.

3. The millimeter-wave electrocardiogram reconstruction system based on adaptive MODWT and CNN-BiLSTM-CA according to claim 1, characterized in that, The heartbeat component extraction module includes MODWT decomposition, determination of the effective decomposition layer number, dynamic screening of heartbeat components, and multi-resolution analysis. The adaptive MODWT decomposition performs multi-scale stationary wavelet transform on the preprocessed signal, decomposing it into components of different frequency bands. The effective number of decomposition layers is determined, the optimal number of decomposition layers is dynamically determined, and the heartbeat frequency band is determined. The dynamic screening of heartbeat components determines key scale components related to heartbeat based on energy contribution. The multi-resolution analysis reconstructs the selected scale coefficients to generate a net signal that enhances heartbeat characteristics and suppresses noise.

4. The millimeter-wave electrocardiogram reconstruction system based on adaptive MODWT and CNN-BiLSTM-CA according to claim 1, characterized in that, The electrocardiogram reconstruction module includes a channel attention module, a CNN module, and a BiLSTM module; The channel attention module (CA) weights the multiple components of the MODWT output to enhance key heartbeat features; it is used to assign different importance weights to the multiple heartbeat components extracted by multi-resolution analysis. The CNN module extracts spatial features through one-dimensional convolution: it achieves layer-by-layer abstraction through convolution kernels and improves generalization ability through BN+ReLU. The BiLSTM module employs bidirectional temporal modeling, including forward LSTM and backward LSTM. The forward LSTM simulates the direction of cardiac electrical signal conduction by learning the temporal evolution of the electrocardiogram in the forward direction; the backward LSTM captures the repolarization process and rhythm association by learning the electrocardiogram dependencies in the backward direction.

5. The millimeter-wave electrocardiogram reconstruction system based on adaptive MODWT and CNN-BiLSTM-CA according to claim 1, characterized in that, It also includes a model training module for training models based on synchronously acquired real electrocardiogram data. The module uses L1 loss and mean squared error to construct a loss function, introduces an early stop strategy, and uses the Adam optimizer for iterative model training.

6. The millimeter-wave electrocardiogram reconstruction system based on adaptive MODWT and CNN-BiLSTM-CA according to claim 1, characterized in that, It also includes an output evaluation module for evaluating the reconstructed electrocardiogram using the Pearson correlation coefficient (PCC) and mean squared error (RMSE).

7. A millimeter-wave electrocardiogram reconstruction method based on adaptive MODWT and CNN-BiLSTM-CA, characterized in that, Includes the following steps: S1: Use millimeter-wave radar to collect echo signals from the target's chest cavity region and extract phase information; S2: Preprocess the echo signal to obtain an effective physiological displacement signal sequence; S3: The effective physiological displacement signal sequence is decomposed into multiple scales using stationary wavelet transform MODWT, and mid-to-high frequency coefficients containing electrocardiogram information are extracted. The coefficients of key scales are reconstructed to obtain a net signal with enhanced features. S4: Input the net signal into a convolutional neural network (CNN) to extract spatial features, and then input it into a bidirectional long short-term memory (BiLSTM) network for time series modeling; S5: The feature vector output by the BiLSTM is fed into a fully connected layer for regression mapping, and the normalized reconstructed electrocardiogram signal is output.

8. The millimeter-wave electrocardiogram reconstruction method based on adaptive MODWT and CNN-BiLSTM-CA according to claim 7, characterized in that, The stationary wavelet transform (MODWT) decomposition process is an adaptive wavelet decomposition process, specifically including: S31: By setting energy convergence criteria, the minimum number of effective decomposition layers that satisfy the heartbeat frequency band coverage is dynamically determined; S32: Integrate the frequency band energy of each decomposition layer and select the three scale coefficient components that contribute the most to the energy in the target frequency band. S33: Use the multi-resolution analysis results composed of the three components as input to the deep neural network.

9. The millimeter-wave electrocardiogram reconstruction method based on adaptive MODWT and CNN-BiLSTM-CA according to claim 7, characterized in that, The bidirectional long short-term memory network BiLSTM employs bidirectional temporal modeling, and the specific steps are as follows: S41: Construct a forward LSTM to simulate the direction of cardiac electrical signal conduction by learning the temporal evolution of electrocardiograms in a forward manner; S42: Construct a backward LSTM to capture the association between repolarization processes and rhythms by learning ECG dependencies in reverse; S43: Bidirectional splicing, used to fuse the entire context of the depolarization-repolarization process; S44: Tanh normalization, used to constrain the output to the standard ECG voltage range.

Images