Multimodal assisted noise reduction system for electrocardiogram signals, electronic devices and computer-readable storage media

Abstract

This invention relates to a multimodal assisted denoising system for electrocardiogram (ECG) signals, an electronic device, and a computer-readable storage medium. The method includes: acquiring and preprocessing synchronously acquired ECG and photoplethysmography (PPG) pulse wave signals, extracting their respective modal-specific features and cross-modal shared features; inputting the ECG signal and its features into a diffuse noise prediction network, using modal-specific features as conditions and shared features to guide the output of predicted noise layer by layer; and finally reconstructing the output denoised ECG. By using PPG-assisted ECG denoising, with modal-specific features as constraints and shared features injected layer by layer into the network, key morphologies such as the QRS complex can be better preserved while suppressing motion artifacts. The waveform is restored using homologous physiological information, providing reliable input for subsequent diagnostic classification, and is suitable for long-term continuous health monitoring using wearable devices.

Description

Technical Field

[0001] This invention relates to the field of computer applications, specifically to a multimodal auxiliary noise reduction system for electrocardiogram signals, an electronic device, and a computer-readable storage medium. Background Technology

[0002] In recent years, wearable devices, represented by wristband devices and portable monitoring terminals, have become increasingly popular. Among them, electrocardiogram (ECG), as an important physiological signal reflecting cardiac electrical activity, is widely used in scenarios such as home health monitoring, chronic disease follow-up, and cardiovascular disease screening. Compared with the resting acquisition environment in hospitals, ECGs collected at home and in free-movement scenarios are more susceptible to motion artifacts, electromyography, and baseline drift. In particular, motion artifacts can lead to waveform drift, local distortion, and loss of key morphological details, thus affecting the results of subsequent diagnostic analysis.

[0003] Existing publicly available ECG denoising methods based on diffusion models are mostly based on single-modal inputs. They typically only perform filtering, reconstruction, or noise estimation based on the noisy ECG itself, making it difficult to fully utilize the physiological rhythm information carried by other synchronously acquired modalities. At the same time, existing multimodal methods usually focus on directly splicing different modalities or performing simple feature fusion, making it difficult to distinguish between modal shared information and modal-specific information, and easily introducing irrelevant components from auxiliary modalities into the main modal denoising process.

[0004] Furthermore, existing methods often struggle to simultaneously achieve both noise suppression and key ECG morphology fidelity in high-motion scenarios. They may weaken diagnostic-related structures such as the QRS complex while removing motion artifacts, thereby reducing the effectiveness of the denoised signal in downstream tasks. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide an ECG denoising method that can make full use of auxiliary modal physiological information in home scenarios, while taking into account noise suppression and waveform fidelity, so as to become a wearable ECG monitoring application; To achieve the above objectives, the present invention provides the following technical solution. A multimodal assisted denoising system for electrocardiogram signals based on a modal decoupling and conditional diffusion model includes: S1. Acquire synchronously acquired electrocardiogram signals and photoplethysmography (PPG) pulse wave signals, and perform preprocessing; S2. Input the preprocessed electrocardiogram (ECG) signal and photoplethysmography (PPG) signal into the modal decoupling feature extraction module to extract modal-specific features of the ECG signal, modal-specific features of the PPG signal, and cross-modal shared features. Among them, cross-modal shared features characterize the common physiological features between electrocardiogram (ECG) signals and photoplethysmography (PPG) pulse wave signals, while ECG signal modality-specific features characterize the waveform morphology features of ECG signals. S3. Input the ECG signal, ECG signal modal-specific features, and cross-modal shared features into the diffuse noise prediction network. Use the ECG signal modal-specific features as conditions and use the cross-modal shared features to guide the network layer by layer to output the predicted noise. S4. Perform reverse reconstruction based on the predicted noise, and restore and output the denoised electrocardiogram signal.

[0006] A further optimization of this invention is that the common physiological features represented by cross-modal shared features include heart rhythm, periodic changes, and cross-modal synchronization relationships; The waveform morphological features characterized by modal-specific features of electrocardiogram signals include at least one of the QRS complex, P wave, and T wave.

[0007] A further optimization of the present invention includes preprocessing that involves time alignment and pulse propagation delay compensation of the electrocardiogram signal and the photoplethysmography (PPG) signal, so that the hemodynamic changes reflected in the PPG signal are aligned with the corresponding cardiac events in the electrocardiogram signal.

[0008] A further optimization of this invention utilizes cross-modal shared features to guide the diffuse noise prediction network layer by layer, including: The quality of the photoplethysmography (PPG) signal is assessed, and a quality score is generated. The injection intensity of cross-modal shared features in each layer of the diffuse noise prediction network is dynamically adjusted based on the quality score.

[0009] In a further optimization of the present invention, the modal decoupling feature extraction module includes dual encoders, dual decoders, and a decoupling potential space; The modal decoupling feature extraction module is used to pre-train the modal feature extraction encoders, enabling each modal feature extraction encoder to extract modal-specific representations and cross-modal shared representations.

[0010] In a further optimization of the present invention, the diffuse noise prediction network adopts a three-branch structure, including: The main branches of the electrocardiogram signal are used to extract the timing representation from the raw electrocardiogram signal; The ECG signal modal-specific conditional branch is used to receive and introduce modal-specific features of noisy ECG signals as conditional constraints. The photoplethysmography (PPG) signal-assisted guiding branch is used to receive cross-modal shared features and, after combining them with diffusion time step information, injects them layer by layer into the main branch of the electrocardiogram signal.

[0011] A further optimization of the present invention is to use the modal-specific features of the input ECG signal as a conditional input, replacing the original waveform of the ECG signal.

[0012] A further optimization of the present invention is that the diffuse noise prediction network is trained through the following steps: The original ECG signal, the noisy ECG signal, and the original photoplethysmography pulse wave signal used for training are input into the diffuse noise prediction network to obtain the predicted noise. The predicted noise is used to estimate the fit of the diffuse noise prediction network to the input noise under the current parameters; Based on the difference between the predicted noise and the actual diffused noise used to generate the noisy ECG signal for training, a loss function is constructed. This loss function is then used to update the parameters of the diffused noise prediction network only through gradient backpropagation.

[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the ECG signal multimodal assisted denoising system based on the modal decoupling and conditional diffusion model as described above.

[0014] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described multimodal assisted denoising system for electrocardiogram signals based on a modal decoupling and conditional diffusion model.

[0015] This invention has the following characteristics and beneficial effects: 1. By using synchronously acquired PPG as an auxiliary mode, the shared physiological rhythm and target waveform structure are collaboratively modeled during the ECG denoising process. By using ECG modal-specific features as conditional constraints and combining modal-shared features with time step embedding, the spread noise prediction network (DNP) is injected layer by layer, which can better preserve key ECG morphological information such as QRS complexes while suppressing motion artifacts.

[0016] 2. By supplementing the original rhythm and heartbeat trend of ECG with homologous physiological information from other modalities, the ECG waveform can be better restored locally. At the same time, the output denoised ECG can provide a more reliable input signal for subsequent diagnostic, classification or estimation tasks, which is suitable for long-term continuous health monitoring scenarios supported by wearable devices.

[0017] 3. By using modal feature decoupling to filter information between multiple modalities and using multi-scale residual block extraction during model training, we can better obtain multimodal related information and provide more relevant contextual information for the prediction of ECG-specific noise during model training. Attached Figure Description

[0018] Figure 1 This is a flowchart of the ECG multimodal assisted denoising method of the present invention.

[0019] Figure 2This is a schematic diagram of the ECG multimodal assisted denoising model structure of the present invention.

[0020] Figure 3 This is a flowchart illustrating the ECG multimodal assisted denoising training construction process of the present invention.

[0021] Figure 4 This is a schematic diagram of the model structure of the modal decoupling feature extraction module of the present invention. Detailed Implementation

[0022] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0023] Example 1 Reference Figure 1 and Figure 2 This paper illustrates the overall process and model structure of an ECG multimodal assisted denoising method based on modal decoupling and conditional diffusion, provided by an embodiment of the present invention. This method is mainly used in home and free-movement scenarios, utilizing synchronously acquired PPGs to assist in denoising ECGs contaminated by motion artifacts. The method includes the following steps: Step 1: Obtain synchronized ECG and PPG fragments and complete the input sample construction.

[0024] In this embodiment, the input samples consist of synchronous ECG and PPG segments. Preferably, the original multimodal data is first divided into short time windows of 4 seconds in length, and a uniform sampling rate of 125Hz is applied, so that each segment contains 500 sampling points. The ECG and PPG segments are then normalized to reduce amplitude differences caused by different individuals, different devices, and different acquisition batches.

[0025] Step 2: Perform time alignment and pulse propagation delay compensation on ECG and PPG.

[0026] In this embodiment, coarse synchronization is performed on the ECG and PPG based on timestamps, and cardiac feature points in the ECG and pulse feature points in the PPG are extracted to complete peak alignment. Considering the propagation delay between cardiac electrical activity and peripheral pulse arrival, pulse propagation delay compensation is further performed to align the hemodynamic changes reflected in the PPG with the corresponding cardiac events in the target ECG.

[0027] Step 3: Construct a noisy input for ECG and keep PPG as the auxiliary mode.

[0028] After preprocessing, a noisy input is constructed for the ECG. Specifically, motion artifact segments are extracted from the NSTDB noise library and superimposed onto the original ECG segment according to a preset signal-to-noise ratio to generate a noisy ECG. The PPG signal is not used for noise addition and remains as an auxiliary modal input to provide supplementary physiological information related to cardiac activity during the denoising process.

[0029] Step 4: Pre-training of the modal decoupling feature extraction module.

[0030] The pre-training objective of the modal decoupling encoder is to constrain shared features to represent common physiological information across modalities, and to constrain specific features to retain unique information of their respective modalities, so as to improve the quality of conditional representation in the subsequent diffusion denoising stage.

[0031] In this embodiment, the Modal Decoupling Feature Extraction (MOD) module adopts a pre-trained structure composed of dual encoders, dual decoders, and a decoupled latent space. Each modal encoder adds two MLP classification heads to the existing pre-trained encoder to output modality-specific and shared representations, thereby generating shared candidate representations and modality-specific candidate representations, respectively. Correspondingly, the ECG reconstruction decoder and PPG reconstruction decoder are used to reconstruct the input samples based on the shared and modality-specific representations, so that different latent subspaces can assume different information expression functions through reconstruction consistency constraints, and are only used for encoder tuning.

[0032] Furthermore, the pre-training of the Modal Decoupling Feature Extraction (MOD) module is performed using a three-step iterative method: First, learn the modality-specific representations of the two modalities and use the complementary information provided by the paired modalities to complete sample reconstruction. At the same time, train the mask parameters corresponding to the shared representations. The loss function is as follows: in, This represents the specific characterization of the m-th mode. This represents the corresponding modality-specific characterization target. This represents the reconstruction decoder for the m-th mode. This represents complementary information provided by the paired modes. This represents the mask parameters corresponding to the shared representation. This represents the weight of the reconstruction term; a value of 0.5 is used during training. This represents the m-th modal input sample.

[0033] Second, based on the learned modality-specific representations, shared representations are further learned. Shared features from another modality are used to continuously train the decoder of the current modality. At the same time, the mask parameters corresponding to each modality-specific representation are trained. The loss function is as follows: in, This represents the shared candidate representation generated by the m-th modality. Indicates a shared representational objective. This represents the mask parameter corresponding to the m-th modal specific characterization. This represents the reconstruction decoder for the m-th mode. Indicates the weight of the reconstruction item. This represents the m-th modal input sample.

[0034] Third, fix the shared representation and modality-specific representation targets obtained in the first two steps, and fine-tune the ECG encoder and PPG encoder so that their shared prediction head and modality-specific prediction head regress to their corresponding latent representations, in order to obtain the Modality Decoupling Feature Extraction Module (MOD) that can be directly called in the subsequent diffusion denoising stage. The loss function is as follows: in, Indicates shared representation, Represents modal-specific characterization. and Let these represent the shared representation prediction head and the specific representation prediction head of the m-th modal encoder, respectively. This represents the m-th modal input.

[0035] Preferably, the shared characterization subspace in the MOD is used to characterize the cardiac rhythm, periodic changes, and cross-modal synchronization relationships that coexist in ECG and PPG. The ECG modality-specific characterization subspace is used to retain structural information such as QRS complex, P wave, and T wave that are directly related to ECG electrical activity and cannot be fully replaced by PPG. The PPG modality-specific characterization subspace is used to retain modality-specific information related to pulse rise edge, amplitude changes, and peripheral hemodynamics.

[0036] After the MOD completes pre-training, the ECG modal-specific features output by the ECG encoder are fed into the conditional branch of the diffuse noise prediction network DNP, the shared representation is fed into the layer-by-layer injection branch of the diffuse backbone, and the PPG modal-specific representation is retained for use in modal decoupling training and reconstruction constraints, rather than as the final denoising output, thereby ensuring that the present invention still aims at pure ECG reconstruction.

[0037] Preferably, the modal decoupling feature can be expressed as: in, Indicates ECG modal-specific features, This indicates PPG modal-specific features. Indicates cross-modal shared features. and These represent the specific coding mappings of ECG and PPG, respectively. This represents a shared feature fusion mapping.

[0038] Step 5: Training the conditional diffusion-based noise prediction network DNP.

[0039] like Figure 2 As shown, the diffusion noise prediction network (DNP) adopts a three-branch hierarchical structure, including a noisy ECG backbone branch, an ECG modal-specific condition branch, and a PPG auxiliary guidance branch. The noisy ECG backbone branch takes the noisy ECG in the diffusion state at step t as input. After channel enhancement by the bottom one-dimensional convolutional layer, it sequentially passes through a multi-level feature extraction unit (Block) to extract temporal representations related to noise estimation, such as local waveform texture, QRS group edges, and long-term rhythm changes.

[0040] In this embodiment, the feature extraction unit Block is a residual module consisting of four 20-dimensional convolutional layers connected in series. After the module input is extracted by multiple convolutional layers, the residual is added to the shortcut branch, and the output dimension is 80.

[0041] The forward noise addition process of diffusion can be represented as: in, This indicates a pure, unprocessed ECG. Indicates the first Noisy ECG in step-diffusion state This represents the cumulative product of the diffusion scheduling coefficients. This represents Gaussian noise.

[0042] In the current diffusion step, the conditional noise prediction process can be expressed as: in, This represents the DNP (Diffusion Noise Prediction Network). Represents a linear mapping of shared features. Indicates time step embedding, This indicates feature splicing.

[0043] Furthermore, the modality-shared features are transformed into feature representations consistent with the channel dimensions of the backbone network after linear mapping, then concatenated with the diffusion time-step embedding, and injected layer by layer into the original ECG feature extraction module Block. Through this layer-by-layer injection method, shared physiological information can continuously participate in the estimation of motion artifacts at different depths and scales. The layer-by-layer injection process can be represented as follows: in, Indicates the first The original characteristics of the layer residual block, This represents the output feature after injecting shared features. Indicates the first Layer linear mapping matrix, This represents the injection transformation function for the corresponding layer.

[0044] At the top level of the ECG modality-specific conditional branch, the fused high-level features are mapped by the output convolutional layer to a predicted noise sequence with the same length as the input ECG, serving as the noise estimate in the DDPM inverse process. Through this structure, DNP can simultaneously utilize the current diffusion state of the noisy ECG, the stable structural representation of the ECG itself, and the synchronous rhythm information in the PPG auxiliary mode, thereby improving the ability to identify and remove motion artifacts.

[0045] During the diffusion denoising stage, the parameters of the ECG encoder and PPG encoder are frozen and used only as stable characterization extractors. The current diffusion step... The original noisy ECG is fed into the diffuse noise prediction network DNP as the main input, while the ECG modal-specific features are input into the conditional branch, instead of directly using the original ECG waveform or its shallow features as the conditional input.

[0046] Step 6: Perform reverse reconstruction based on the predicted noise to restore a clean ECG.

[0047] At each diffusion time step, the diffusion noise prediction network (DNP) outputs the noise prediction result corresponding to the current noisy ECG. Because the conditional branch highlights the structural features of the ECG itself, while the layer-by-layer injection branch continuously introduces cross-modal shared information, the network can more accurately distinguish between the true ECG morphology and motion artifacts.

[0048] The pure ECG estimate can be recovered from the predicted noise: Furthermore, the posterior mean of a single-step backward update can be written as: in, This represents the pure ECG estimate at step t. This represents the posterior mean during back diffusion. and This represents the time correlation coefficient determined by the diffusion scheduling parameters.

[0049] Subsequently, based on the predicted noise results, a reverse reconstruction process of the diffusion model is performed to gradually recover a clean ECG waveform from the high-noise state. The final output is the denoised ECG signal, without outputting a denoised PPG, thus maintaining the technical boundary of this invention, which aims at ECG denoising.

[0050] Preferably, an auxiliary modal quality assessment module can be set before modal sharing feature injection. This quality assessment module calculates the quality score of the current PPG segment based on PPG waveform stability, spectral energy concentration, consistency of adjacent pulse intervals, and cross-modal correlation between the PPG and ECG after delay compensation. The quality assessment and shared feature gating process can be represented as follows: in, Indicators representing waveform morphology stability Indicates the index of spectral concentration. This indicates the consistency index of adjacent pulse intervals. This represents the cross-modal correlation index after delay compensation. Indicates the weights injected for shared features. This represents the shared features after quality gating.

[0051] Shared feature injection weights are generated based on the quality scores, and the injection branches are dynamically adjusted layer by layer. When the PPG quality is high, the modulation effect of shared features on each residual block is enhanced; when the PPG quality is low, the influence of shared features on noise prediction is weakened to avoid introducing additional interference when the auxiliary mode quality is poor.

[0052] Step 7: Output the denoised ECG and use it for subsequent tasks.

[0053] Through the above structural design, this embodiment can more effectively suppress motion artifacts common in home exercise scenarios while preserving key ECG morphological information, thereby obtaining high-quality ECG signals that are more suitable for subsequent tasks.

[0054] Example 2 Reference Figure 3 This paper demonstrates a training and construction process for an ECG multimodal assisted denoising model based on modal decoupling and conditional diffusion, provided by an embodiment of the present invention. The training process employs a phased construction approach, including a modal decoupling encoder pre-training phase and a diffusion denoising backbone training phase.

[0055] Step 1: Obtain synchronized data and complete sample partitioning.

[0056] In this embodiment, the training data can be derived from publicly available synchronous datasets that contain both ECG and PPG, such as PulseDB and BIDMC. First, the original multimodal data is cleaned, segmented, and has its sampling rate standardized to obtain paired ECG-PPG segments of 4 seconds in length and a sampling rate of 125Hz. Then, the dataset is divided into training, validation, and test sets according to a preset ratio.

[0057] Step 2: Construct motion artifact training samples specifically for ECG.

[0058] To construct the training samples required for the ECG denoising task, motion artifact segments are extracted from the NSTDB noise database. These motion artifacts are then superimposed onto the original ECG segments at random or preset signal-to-noise ratios to generate noisy ECGs. The corresponding original ECGs serve as reference targets during model training. PPG segments retain their original auxiliary modality roles during training and are not subject to noise superposition.

[0059] Step 3: Pre-trained modal decoupling encoder.

[0060] During the pre-training phase of the modal decoupled encoder, synchronous ECG and PPG segments are input into the ECG encoder and PPG encoder. Through shared representation constraints, specific representation constraints, and reconstruction constraints, the two encoders can learn cross-modal shared physiological information and their respective modal-specific information. After pre-training, the ECG encoder and PPG encoder possess stable decoupled representation capabilities.

[0061] Step 4: Freeze the encoder and train the diffusion denoising backbone.

[0062] After the encoder pre-training is complete, the parameters of the ECG encoder and PPG encoder are fixed and no longer jointly updated during subsequent diffusion training phases. This avoids the diffusion backbone training process from damaging the learned decoupled representations and helps maintain the stability of conditional feature and shared feature representations.

[0063] During the training phase of the diffusion denoising backbone, a forward diffusion process is first applied to the original ECG to obtain the noisy ECG state at different time steps. Then, the noisy ECG at the current time step is fed into the diffusion noise prediction network DNP, the ECG modality-specific features are fed into the conditional branch, and the modality-shared features are linearly mapped and concatenated with the time step embedding, and then fed into the layer-by-layer injection branch.

[0064] In this embodiment, the diffusion noise prediction network (DNP) outputs the corresponding noise estimation results at each diffusion time step. During training, the diffusion noise prediction loss is constructed solely based on the difference between the network's predicted noise and the actual diffusion noise, and this difference is used to update the diffusion backbone's relevant parameters. The loss function for the diffusion stage can be expressed as: in, The loss for predicting diffuse noise is represented by the MES loss function. This represents the actual diffused noise. This represents network prediction noise. This indicates shared features after quality gating.

[0065] With encoder parameters frozen, the optimization focus during the diffusion backbone training phase is on how to more fully utilize the decoupled ECG specific features and modal-shared features to improve noise modeling capabilities under motion artifact conditions. Compared to schemes that directly use single-modal noisy ECGs for noise prediction, this embodiment can learn more stable and physiologically constrained noise estimation strategies during training.

[0066] Preferably, an auxiliary modal quality assessment mechanism can be introduced simultaneously during the training phase to dynamically adjust the intensity of shared feature injection based on the quality of PPG fragments, so that the model maintains good robustness under different wearing states and different motion intensities.

[0067] Step 5: After training, obtain the denoised model and use it for downstream tasks.

[0068] After training, a multimodal denoising model for ECG based on modal decoupling and conditional diffusion was obtained. This model can reconstruct noisy ECGs collected in home scenarios with high quality and provide more reliable input signals for subsequent diagnostic, classification, or estimation tasks, thereby effectively improving task performance.

[0069] Example 3 This embodiment provides a device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the ECG multimodal assisted denoising method based on modal decoupling and conditional diffusion described in Embodiments 1 and 2.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A multimodal assisted noise reduction system for electrocardiogram (ECG) signals, characterized in that, include: S1. Acquire synchronously acquired electrocardiogram signals and photoplethysmography (PPG) pulse wave signals, and perform preprocessing; S2. Input the preprocessed electrocardiogram (ECG) signal and photoplethysmography (PPG) signal into the modal decoupling feature extraction module to extract modal-specific features of the ECG signal, modal-specific features of the PPG signal, and cross-modal shared features. Among them, cross-modal shared features characterize the common physiological features between electrocardiogram (ECG) signals and photoplethysmography (PPG) pulse wave signals, while ECG signal modality-specific features characterize the waveform morphology features of ECG signals. S3. Input the ECG signal, ECG signal modal-specific features, and cross-modal shared features into the diffuse noise prediction network. Use the ECG signal modal-specific features as conditions and use the cross-modal shared features to guide the network layer by layer to output the predicted noise. S4. Perform reverse reconstruction based on the predicted noise, and restore and output the denoised electrocardiogram signal.

2. The ECG signal multimodal assisted noise reduction system according to claim 1, characterized in that, The common physiological features represented by cross-modal shared features include cardiac rhythm, periodic variations, and cross-modal synchronization relationships; The waveform morphological features characterized by modal-specific features of electrocardiogram signals include at least one of the QRS complex, P wave, and T wave.

3. The ECG signal multimodal assisted noise reduction system according to claim 1, characterized in that, Preprocessing includes time alignment and pulse propagation delay compensation of electrocardiogram (ECG) signals and photoplethysmography (PPG) signals, so that the hemodynamic changes reflected in the PPG signals are aligned with the corresponding cardiac events in the ECG signals.

4. The ECG signal multimodal assisted noise reduction system according to claim 1, characterized in that, Layer-by-layer guidance of the diffuse noise prediction network is achieved by leveraging cross-modal shared features, including: The quality of the photoplethysmography (PPG) signal is assessed, and a quality score is generated. The injection intensity of cross-modal shared features in each layer of the diffuse noise prediction network is dynamically adjusted based on the quality score.

5. The ECG signal multimodal assisted noise reduction system according to claim 1, characterized in that, The modal decoupling feature extraction module includes dual encoders, dual decoders, and a decoupling potential space; The modal decoupling feature extraction module is used to pre-train the modal feature extraction encoders, enabling each modal feature extraction encoder to extract modal-specific representations and cross-modal shared representations.

6. The ECG signal multimodal assisted noise reduction system according to claim 1, characterized in that, The diffuse noise prediction network adopts a three-branch structure, including: The main branches of the electrocardiogram signal are used to extract the timing representation from the raw electrocardiogram signal; The ECG signal modal-specific conditional branch is used to receive and introduce modal-specific features of noisy ECG signals as conditional constraints. The photoplethysmography (PPG) signal-assisted guiding branch is used to receive cross-modal shared features and, after combining them with diffusion time step information, injects them layer by layer into the main branch of the electrocardiogram signal.

7. The ECG signal multimodal assisted noise reduction system according to claim 6, characterized in that, The modal-specific features of the input ECG signal are used as conditional inputs to replace the original waveform of the ECG signal.

8. The ECG signal multimodal assisted noise reduction system according to claim 7, characterized in that, The diffuse noise prediction network is trained through the following steps: The original ECG signal, the noisy ECG signal, and the original photoplethysmography pulse wave signal used for training are input into the diffuse noise prediction network to obtain the predicted noise. The predicted noise is used to estimate the fit of the diffuse noise prediction network to the input noise under the current parameters; Based on the difference between the predicted noise and the actual diffused noise used to generate the noisy ECG signal for training, a loss function is constructed. This loss function is then used to update the parameters of the diffused noise prediction network only through gradient backpropagation.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes a computer program, it implements the ECG signal multimodal assisted noise reduction system as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the ECG signal multimodal assisted noise reduction system as described in any one of claims 1-8.

Images