Method and system for extracting fetal electrocardiogram based on multi-channel signal sensing

Abstract

The present application relates to the technical field of artificial intelligence, in particular to a fetal electrocardiogram extraction method and system based on multi-channel signal perception, the method comprising: acquiring a multi-channel abdominal signal and a trained multi-channel signal perception network; inputting the multi-channel abdominal signal into the trained multi-channel signal perception network, decomposing the multi-channel abdominal signal into multi-scale feature maps through a self-learning Q-guided wavelet module; encoding the multi-scale feature maps through an encoder module to obtain a feature mapping; processing the feature mapping through a tomographic attention transform module to generate enhanced output features; suppressing noise of the enhanced output features through an adaptive waveform contraction module to obtain de-noised sub-band feature maps processed by the adaptive waveform contraction module; and reconstructing a fetal electrocardiogram signal through a decoder module to reconstruct the fetal electrocardiogram signal from the de-noised sub-band feature maps.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a fetal electrocardiogram extraction method and system based on multi-channel signal perception. BACKGROUND

[0002] Non-invasive fetal electrocardiogram (fECG) extraction from maternal abdominal recordings is crucial for the diagnosis of prenatal arrhythmias and fetal distress. However, fetal electrocardiogram signals are weak and often masked by the dominant maternal electrocardiogram (mECG). In addition, interference factors such as electromyographic noise, baseline drift, and uterine contractions result in low signal-to-noise ratio (SNR).

[0003] To address these challenges, several traditional fetal extraction methods have been proposed, but they face some limitations. For example, template subtraction encounters difficulties in handling the non-stationary characteristics of maternal electrocardiogram signals. Blind source separation (BSS) requires multiple sensors and encounters difficulties when fetal and maternal signals overlap. Wavelet-based neural networks utilize multi-resolution analysis through time-frequency localization. Although they have advantages, these networks are often difficult to capture long-range contextual dependencies in sequences. SUMMARY

[0004] To solve the above problems, the present application provides a fetal electrocardiogram extraction method and system based on multi-channel signal perception, aiming to improve the accuracy of fetal electrocardiogram extraction.

[0005] To achieve the above purpose, the present application provides the following technical solutions:

[0006] On the one hand, the present application provides a fetal electrocardiogram extraction method based on multi-channel signal perception, which comprises the following steps:

[0007] S100, acquiring multi-channel abdominal signals and a trained multi-channel signal perception network; wherein the multi-channel signal perception network comprises a self-learning Q-guided wavelet module, an encoder module, a tomographic attention transform module, an adaptive waveform contraction module, and a decoder module;

[0008] S200, inputting the multi-channel abdominal signals into the trained multi-channel signal perception network, decomposing the multi-channel abdominal signals into multi-scale feature maps through the self-learning Q-guided wavelet module; encoding the multi-scale feature maps through the encoder module to obtain a feature mapping; processing the feature mapping through the tomographic attention transform module to generate enhanced output features; suppressing noise of the enhanced output features through the adaptive waveform contraction module to obtain denoised sub-band feature maps processed by the adaptive waveform contraction module; reconstructing the fetal electrocardiogram signal through the decoder module.

[0009] Optionally, in S200, the process of processing the feature map through the tomographic focus transformation module to generate enhanced output features includes:

[0010] S211, the feature map is divided into feature maps corresponding to the query channel group, key channel group and value channel group, and the interaction weight is calculated by using the linear transformation matrices of query, key and value respectively, combined with the weight parameters of the query channel group, key channel group and value channel group;

[0011] S212, the linear transformation matrix of the value channel group is adjusted using the interactive weights, and enhanced output features are generated by multiplying the interactive weights with the linear transformation matrix corresponding to the value channel group.

[0012] Optionally, in S211, the interaction weights are calculated by using the linear transformation matrices of the query, key, and value channels respectively, combined with the weight parameters of the query channel group, key channel group, and value channel group, including:

[0013] S2111, determine the linear transformation matrix corresponding to the query channel group and the key channel group, as well as the dimension of the key channel group;

[0014] S2112, perform a scaling dot product operation on the linear transformation matrices corresponding to the query channel group and the key channel group, and then scale by dividing by the square root of the dimension of the key channel group;

[0015] S2113, use the Softmax operation to normalize the scaled dot product result to obtain the interaction weights. The Softmax operation is used to convert the dot product result into a probability distribution so that the sum of all interaction weights is 1.

[0016] Optionally, in S200, the step of suppressing noise in the enhanced output feature using an adaptive waveform shrinkage module to obtain a denoised sub-band feature map processed by the adaptive waveform shrinkage module includes:

[0017] S221, Obtain the noise estimate for each sub-band. The noise estimate is obtained by calculating the median absolute deviation of the feature map after the sub-band is processed by the encoder module and dividing it by a constant factor. This constant factor is used to convert the median absolute deviation into a standard deviation estimate.

[0018] S222, input the noise estimate of each sub-band into the Sigmoid activation function to generate adaptive threshold coefficients, and use the product of the adaptive threshold coefficients and the noise estimate as the adaptive threshold;

[0019] S223, perform soft thresholding on each sub-band in the enhanced output feature according to the adaptive threshold to obtain the denoised sub-band feature map after processing by the adaptive waveform shrinkage module.

[0020] Optionally, in S223, the step of performing soft thresholding on each sub-band in the enhanced output features according to the adaptive threshold to obtain the denoised sub-band feature map after processing by the adaptive waveform shrinkage module includes:

[0021] When the absolute value of a feature is less than the adaptive threshold, it is set to zero;

[0022] When the absolute value of the feature value is greater than the adaptive threshold, the part exceeding the threshold is retained and a linear transformation is performed. At the same time, the sign information of the original feature is retained through the sign function, resulting in the denoised sub-band feature map after processing by the adaptive waveform shrinkage module.

[0023] Optionally, in S200, the step of reconstructing the denoised sub-band feature map through the decoder module to reconstruct the fetal electrocardiogram signal includes:

[0024] S231, input the denoised subband feature map processed by the adaptive waveform shrinkage module into the decoder module;

[0025] S232, the decoder module reconstructs the signal through stacked 1D convolutional layers, instance normalization layers and sinusoidal activation functions, and recombines the denoised sub-band feature maps of multiple scales into a time-domain signal to obtain a denoised and enhanced fetal electrocardiogram signal.

[0026] Optionally, the composite loss function for training the multi-channel signal sensing network is constructed in the following manner:

[0027] S101, acquire the estimated fetal electrocardiogram signal and the actual fetal electrocardiogram signal, wherein the estimated fetal electrocardiogram signal is the output obtained by processing multi-channel abdominal signals through a multi-channel signal sensing network;

[0028] S102, the L1 loss is obtained by calculating the sum of the absolute differences between the estimated fetal electrocardiogram (ECG) signal and the actual fetal ECG signal; the predicted morphological feature vector and the target morphological feature vector are obtained, and the morphological loss is obtained by calculating the similarity measure between the predicted morphological feature vector and the target morphological feature vector; the predicted morphological feature vector and the target morphological feature vector respectively represent the feature mapping maps of the estimated fetal ECG signal and the actual fetal ECG signal on the waveform structure; after performing a Fourier transform operation on the estimated fetal ECG signal and the actual fetal ECG signal, the corresponding spectral representation is obtained, and the frequency domain L2 loss is obtained by calculating the Euclidean distance between the two spectral representations;

[0029] S103, the L1 loss, morphological loss and frequency domain L2 loss are weighted and summed according to the set hyperparameters to construct the composite loss function.

[0030] On the other hand, embodiments of the present invention provide a fetal electrocardiogram extraction system based on multi-channel signal sensing, comprising:

[0031] At least one processor;

[0032] At least one memory for storing at least one program;

[0033] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0034] On the other hand, embodiments of the present invention provide a computer-readable storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the above-described method.

[0035] The beneficial effects of this invention are as follows: This invention discloses a method and system for fetal electrocardiogram (ECG) extraction based on multi-channel signal sensing. This invention decomposes multi-channel abdominal signals into multi-scale feature maps using a self-learning Q-guided wavelet module. By utilizing the synergistic effects of an encoder module, a tomographic focus transformation module, an adaptive waveform contraction module, and a decoder module, it effectively extracts and enhances fetal ECG signals. This method overcomes the limitations of traditional methods in handling non-stationary signals, multi-source interference, and long-distance contextual dependencies, significantly improving the signal-to-noise ratio and the accuracy of signal extraction, providing more reliable technical support for prenatal diagnosis. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic flowchart of a fetal electrocardiogram extraction method based on multi-channel signal sensing in an embodiment of the present invention;

[0038] Figure 2 This is an overall architecture diagram of the multi-channel signal sensing network in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram showing the results of the multi-channel signal sensing network proposed in this embodiment of the invention on the ADFECGDB dataset. Detailed Implementation

[0040] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0042] It is understood that the terms “first,” “second,” etc., used in this invention may be used herein to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are used only to distinguish one concept from another. For example, first information may also be referred to as second information without departing from the scope of embodiments of the invention, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to determination” as used herein may be interpreted as “when…” or “when…” or “in response to determination.”

[0043] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0044] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0045] Advances in signal processing technologies aim to combine signal processing priors with deep learning to overcome existing limitations. Deep learning techniques, such as convolutional neural networks (CNNs), often fail to provide the explicit temporal features required for fetal electrocardiogram (ECG) extraction. In contrast, Transformer architectures excel at modeling global dependencies through self-attention mechanisms. However, their fixed word segmentation strategies may overlook subtle temporal details crucial for accurate waveform reconstruction. To address this, hybrid approaches have emerged, combining temporal and spectral inputs. While effective, these methods typically rely on late-stage fusion, limiting interaction between different modalities during processing. Adaptive template subtraction with echo-state networks dynamically updates the maternal template. However, this approach faces challenges related to error accumulation, particularly during fetal-maternal QRS overlap. Similarly, the VMD-CNN hybrid model decomposes the signal into intrinsic modes for CNN processing, but it requires manual mode selection, potentially leading to distortion of fetal P / T wave features. Finally, CycleGAN-based modality transformation attempts to convert abdominal signals into chest signals. However, it ignores electrophysiological correlations, leading to motion artifacts.

[0046] Inspired by these observations, this invention proposes a multi-channel signal sensing network designed to perform robust fetal electrocardiogram (fECG) estimation from abdominal signals by introducing a tomographic attention transformation module and an adaptive waveform contraction module. This approach suppresses maternal interference while preserving fetal morphological components. The key contributions of the proposed multi-channel signal sensing network are as follows:

[0047] To separate fetal electrocardiograms, we propose a tomographic focus transformation module to model the maternal-fetal interaction of decomposed abdominal signals at multiple scales.

[0048] To enhance fetal features in the noisy subband, we introduced an adaptive waveform shrinkage module to mitigate the effects of noise while preserving key fetal ECG features.

[0049] Experiments on the ADFECGDB and B2LABOUR datasets demonstrate that the proposed tomographic focus transformation module and adaptive waveform shrinkage module are effective in multi-channel signal sensing networks.

[0050] refer to Figure 1 ,like Figure 1 The image shows a method for extracting fetal electrocardiograms based on multi-channel signal sensing, according to an embodiment of the present invention. The method includes the following steps:

[0051] S100, acquire multi-channel abdominal signals and a trained multi-channel signal sensing network; wherein, the multi-channel signal sensing network includes a self-learning Q-guided wavelet module, an encoder module, a tomographic focus transformation module, an adaptive waveform shrinkage module, and a decoder module;

[0052] S200, the multi-channel abdominal signal is input into a trained multi-channel signal sensing network. The multi-channel abdominal signal is decomposed into multi-scale feature maps by a self-learning Q-guided wavelet module. The multi-scale feature maps are encoded by an encoder module to obtain a feature map. The feature map is processed by a tomographic attention transformation module to generate enhanced output features. The enhanced output features are noise-suppressed by an adaptive waveform contraction module to obtain a denoised sub-band feature map processed by the adaptive waveform contraction module. The denoised sub-band feature map is reconstructed by a decoder module to reconstruct the fetal electrocardiogram signal.

[0053] In the embodiments provided by this invention, the self-learning Q-guided wavelet module can adaptively adjust the decomposition scale, thereby more effectively capturing multi-scale features in the abdominal signal. This is crucial for accurately separating the fetal electrocardiogram (ECG) signal in subsequent processing. The encoder module employs an efficient neural network structure to deeply encode the multi-scale feature maps, extracting more representative feature maps. The tomographic attention transformation module, by introducing a self-attention mechanism, can establish effective correlations between features at different scales, enhancing the ability to capture relevant features of the fetal ECG. The adaptive waveform contraction module, by adaptively adjusting the noise threshold, effectively suppresses interference from maternal signals while preserving key features of the fetal ECG. Finally, the decoder module, through inverse wavelet transform, recombines the processed, denoised sub-band feature maps into a time-domain signal, obtaining a denoised and enhanced fetal ECG signal. This series of processing steps not only improves the accuracy of fetal ECG extraction but also significantly enhances the robustness of the system, providing a more reliable basis for clinical diagnosis.

[0054] In some embodiments, in S200, the process of processing the feature map through the tomographic focus transformation module to generate enhanced output features includes:

[0055] S211, the feature map is divided into feature maps corresponding to the query channel group, key channel group and value channel group, and the interaction weight is calculated by using the linear transformation matrices of query, key and value respectively, combined with the weight parameters of the query channel group, key channel group and value channel group;

[0056] S212, the linear transformation matrix of the value channel group is adjusted using the interactive weights, and enhanced output features are generated by multiplying the interactive weights with the linear transformation matrix corresponding to the value channel group.

[0057] In this embodiment, by dividing the feature map into query, key, and value channel groups and calculating interaction weights, the correlation between features in different channels can be dynamically captured. This design allows the tomographic focus transformation module to adaptively focus on feature regions related to the fetal electrocardiogram while suppressing irrelevant interference.

[0058] In some embodiments, S211, the step of calculating the interaction weight using the linear transformation matrices of the query, key, and value channels, respectively, combined with the weight parameters of the query channel group, key channel group, and value channel group, includes:

[0059] S2111, determine the linear transformation matrix corresponding to the query channel group and the key channel group, as well as the dimension of the key channel group;

[0060] S2112, perform a scaling dot product operation on the linear transformation matrices corresponding to the query channel group and the key channel group, and then scale by dividing by the square root of the dimension of the key channel group;

[0061] S2113, use the Softmax operation to normalize the scaled dot product result to obtain the interaction weights. The Softmax operation is used to convert the dot product result into a probability distribution so that the sum of all interaction weights is 1.

[0062] In this embodiment, the numerical stability and effectiveness of the interaction weights are ensured through scaling the dot product operation and normalization. The scaling operation prevents the gradient vanishing problem caused by excessively large dot product results, while Softmax normalization ensures that the interaction weights reflect the relative importance of features between different channels. This calculation method enables the tomographic focus transformation module to more accurately capture key features in the fetal electrocardiogram signal.

[0063] In some embodiments, in S200, the step of performing noise suppression on the enhanced output features through an adaptive waveform shrinkage module to obtain a denoised sub-band feature map processed by the adaptive waveform shrinkage module includes:

[0064] S221, Obtain the noise estimate for each sub-band. The noise estimate is obtained by calculating the median absolute deviation of the feature map after the sub-band is processed by the encoder module and dividing it by a constant factor. This constant factor is used to convert the median absolute deviation into a standard deviation estimate.

[0065] S222, input the noise estimate of each sub-band into the Sigmoid activation function to generate adaptive threshold coefficients, and use the product of the adaptive threshold coefficients and the noise estimate as the adaptive threshold;

[0066] S223, perform soft thresholding on each sub-band in the enhanced output feature according to the adaptive threshold to obtain the denoised sub-band feature map after processing by the adaptive waveform shrinkage module.

[0067] In this embodiment, by introducing noise estimation and an adaptive threshold mechanism, the adaptive waveform shrinkage module can dynamically adjust the noise suppression intensity of each sub-band. The dynamic threshold adjustment strategy significantly improves the model's adaptability to signals with different signal-to-noise ratios, especially maintaining stable feature extraction capabilities even in low-quality abdominal signals.

[0068] In some embodiments, S223, the step of performing soft thresholding on each sub-band in the enhanced output features according to the adaptive threshold to obtain the denoised sub-band feature map after processing by the adaptive waveform shrinkage module includes:

[0069] When the absolute value of a feature is less than the adaptive threshold, it is set to zero;

[0070] When the absolute value of the feature value is greater than the adaptive threshold, the part exceeding the threshold is retained and a linear transformation is performed. At the same time, the sign information of the original feature is retained through the sign function, resulting in the denoised sub-band feature map after processing by the adaptive waveform shrinkage module.

[0071] In this embodiment, the adaptive waveform shrinkage module achieves precise noise suppression through soft thresholding while preserving key features of the fetal electrocardiogram (ECG) signal to the maximum extent. This processing method effectively solves the problem of feature loss when processing low signal-to-noise ratio signals using traditional methods, significantly improving the accuracy and reliability of fetal ECG extraction.

[0072] In some embodiments, S200, the reconstruction of the denoised sub-band feature map by the decoder module to reconstruct the fetal electrocardiogram signal includes:

[0073] S231, input the denoised subband feature map processed by the adaptive waveform shrinkage module into the decoder module;

[0074] S232, the decoder module reconstructs the signal through stacked 1D convolutional layers, instance normalization layers and sinusoidal activation functions, and recombines the denoised sub-band feature maps of multiple scales into a time-domain signal to obtain a denoised and enhanced fetal electrocardiogram signal.

[0075] In this embodiment, an efficient inverse wavelet transform process is achieved through a combination of stacked 1D convolutional layers, instance normalization layers, and a sinusoidal activation function in the decoder module. This structural design not only ensures the accuracy of signal reconstruction but also introduces nonlinear characteristics through the sinusoidal activation function, enhancing the model's adaptability to complex signal patterns.

[0076] In some embodiments, the composite loss function for training the multi-channel signal sensing network is constructed in the following manner:

[0077] S101, acquire the estimated fetal electrocardiogram signal and the actual fetal electrocardiogram signal, wherein the estimated fetal electrocardiogram signal is the output obtained by processing multi-channel abdominal signals through a multi-channel signal sensing network;

[0078] S102, the L1 loss is obtained by calculating the sum of the absolute differences between the estimated fetal electrocardiogram (ECG) signal and the actual fetal ECG signal; the predicted morphological feature vector and the target morphological feature vector are obtained, and the morphological loss is obtained by calculating the similarity measure between the predicted morphological feature vector and the target morphological feature vector; the predicted morphological feature vector and the target morphological feature vector respectively represent the feature mapping maps of the estimated fetal ECG signal and the actual fetal ECG signal on the waveform structure; after performing a Fourier transform operation on the estimated fetal ECG signal and the actual fetal ECG signal, the corresponding spectral representation is obtained, and the frequency domain L2 loss is obtained by calculating the Euclidean distance between the two spectral representations;

[0079] S103, the L1 loss, morphological loss and frequency domain L2 loss are weighted and summed according to the set hyperparameters to construct the composite loss function.

[0080] In this embodiment, a composite loss function is constructed by weighting and summing L1 loss, morphological loss, and frequency domain L2 loss. This comprehensively considers the characteristics of the signal in the time domain, waveform structure, and frequency domain. L1 loss directly measures the numerical difference between the estimated signal and the true signal, ensuring the overall accuracy of the signal values. Morphological loss focuses on the waveform structure features of the signal, ensuring that the extracted fetal electrocardiogram (ECG) signal has the correct morphology. Frequency domain L2 loss, from a frequency domain perspective, makes the estimated signal as similar to the true signal as possible in spectral distribution. This composite loss function design allows the multi-channel signal sensing network to optimize its parameters from multiple dimensions during training, thereby extracting fetal ECG signals more comprehensively and accurately. Through this comprehensive loss function, the network not only focuses on the precise matching of signal values ​​but also emphasizes the similarity of waveform morphology and the consistency of frequency domain characteristics, effectively improving the model's ability to capture complex and variable fetal ECG signals and its resistance to interference.

[0081] The following are specific embodiments provided by the present invention:

[0082] Extracting fetal electrocardiograms from abdominal signals presents significant challenges due to the overlap of maternal and fetal cardiac activity, non-stationary noise, and low signal-to-noise ratio. Therefore, abdominal signals... a (t) can be expressed as:

[0083] (1);

[0084] Where t represents the time variable, and These represent the maternal and fetal electrocardiogram sources, respectively. and , representing the abdominal signal transmission path, i.e. the channel response of maternal and fetal electrocardiogram signals propagating in the abdomen, respectively, and n(t) represents additive noise.

[0085] Therefore, the objective of this invention is to obtain from the observed Accurately restore fetal components .

[0086] Deep learning methods in related technologies often struggle to simultaneously achieve accurate time-frequency localization and capture fine-grained temporal details in signal extraction. To address these challenges, a multi-channel signal sensing network (MCN) for fetal electrocardiogram estimation is proposed. This network comprises five modules: a self-learning Q-guided wavelet module (TQWT), an encoder module, a tomographic attention transformation module (CAT), an adaptive waveform contraction module (RS), and a decoder module. Figure 2 As shown, TQWT is first applied to decompose the multi-channel abdominal signal into multi-scale feature maps. These feature maps are then encoded by an encoder module to capture a hierarchical representation. Subsequently, a CAT module suppresses maternal ECG interference by modeling cross-modal interactions, while an RS module further eliminates residual noise components through adaptive thresholding. Finally, a decoder reconstructs the fetal electrocardiogram (fECG) signal from the processed feature maps.

[0087] Self-learning Q-guided wavelet module for multi-scale decomposition:

[0088] Transient fetal electrocardiographic features (e.g., QRS complexes) are often masked by broadband noise in the raw signal. To preserve these key dynamics, we first segment the abdominal signal into non-overlapping windows:

[0089] (2);

[0090] Subsequently, we applied a self-learning Q-guided wavelet module (TQWT), setting J=6 scale and Q=3.0, to decompose the two signals:

[0091] (3);

[0092] in, This represents the j-th scale feature map after the i-th channel is decomposed. This represents the operation of the self-learning Q-guided wavelet module.

[0093] By using TQWT multi-scale decomposition and acquiring signal features at different frequency bands, the encoder module receives the multi-scale feature maps after TQWT decomposition as input. Through a series of convolution and pooling operations, it downsamples and extracts features from these feature maps, gradually generating hierarchical feature maps.

[0094] The encoder module is used for feature modeling:

[0095] Each sub-band The features are extracted by the encoder module, which consists of five one-dimensional convolutional blocks:

[0096] (4);

[0097] in, This represents the feature map of the j-th subband after processing by the encoder module. InstanceNorm1d represents the one-dimensional instance normalization operation, and 1DConv represents the one-dimensional convolution operation.

[0098] The activations in each input feature map are normalized using InstanceNorm1d, and the sinusoidal activation function Sin is applied to introduce nonlinearity, which further helps to distinguish relevant features from noise components.

[0099] By applying these convolutional layers and normalization, the encoder module captures the basic temporal and spectral features of each subband, progressively filtering out irrelevant noise while maintaining the integrity of the fetal electrocardiogram features.

[0100] Tomographic Focus Transformation Module:

[0101] Fetal electrocardiogram (fECG) signals are consistently masked by maternal electrocardiogram (mECG) interference, especially in overlapping spectral bands. To address this issue, a Tomographic Attention Transformation (CAT) module was designed to suppress mECG and enhance the fetal component by modeling the complex interactions between multiple channels, defined as:

[0102] (5);

[0103] Wherein, ICG and CWI represent intra-channel grouping and inter-channel interaction, respectively, and X represents the input feature matrix.

[0104] The CAT module first performs fine-grained feature division of each subband through intra-channel grouping (ICG), aggregating similar features to enhance local representation capabilities. Then, it calculates cross-channel attention weights through inter-channel interaction (CWI), dynamically adjusting the feature importance between different channels to effectively suppress maternal ECG interference and enhance fetal ECG features. This mechanism enables the model to more accurately capture key information in the fetal ECG, maintaining high recognition accuracy even under strong maternal ECG signals.

[0105] 1) Intra-channel grouping: The key breakthrough lies in the strategic separation and reorganization of feature maps to establish clear interaction paths. Specifically, the feature tensors are divided into three distinct groups, each playing a different role in the interaction process:

[0106] (6);

[0107] in, , , These represent the weight parameters for the query channel group (Q), key channel group (K), and value channel group (V), respectively. , , Let Q, K, and V represent the feature maps corresponding to the three channel groups Q, K, and V after the feature tensor is partitioned. Q, K, and V represent the linear transformation matrices of query, key, and value, respectively.

[0108] 2) Inter-channel interaction mechanism: The inter-channel interaction mechanism establishes explicit relationships between different channel groups through attention-based computation. Specifically, the interaction weight (IW) is first calculated in the following way:

[0109] (7);

[0110] in, The interaction weights are represented by Q and K, which represent the linear transformation matrices corresponding to the query channel group and the key channel group, respectively, and T represents the matrix transpose. It is a dimension of the key channel group; The operation normalizes the scaled dot product between the query and the key, generating interaction weights that quantify the mutual information and dependency strength between channel groups.

[0111] Next, these interactive weights are used to modulate the value channels, producing an enhanced output, as shown below:

[0112] (8);

[0113] in, The value represents the linear transformation matrix corresponding to the channel group, and Output represents the enhanced output features generated after adjustment by the inter-channel interaction mechanism.

[0114] Compared to the standard transformer, which uses the same inputs for query (Q), key (K), and value (V), the proposed tomographic focus transformation module offers the following significant advantages:

[0115] Explicit interaction modeling: This mechanism forces the network to learn and model the relationships between different subsets of abdominal channels, thereby achieving more accurate interaction capture.

[0116] Information bottleneck avoidance: By forcing cross-channel processing, the model avoids the risk of simply copying the input to the output, ensuring a more informative transformation.

[0117] Reinforced representation learning: Controlled information flow between channel groups leads to richer and more diverse feature maps, improving the model's ability to capture complex patterns.

[0118] Adaptive waveform shrinkage module and reconstruction:

[0119] Wavelet subbands often retain residual noise, which can interfere with subtle fetal morphological features. To address this issue, the Adaptive Waveform Shrinkage (RS) module aims to suppress noise components through adaptive thresholding, defined as follows:

[0120] (9);

[0121] Where T represents the adaptive thresholding operation, S represents the shrinkage function applied to the wavelet subband, and ∘ represents element-wise multiplication. This represents the feature map after processing by the encoder module. This indicates the features after processing by the adaptive waveform shrinkage module.

[0122] 1) Adaptive Threshold Strategy: The core innovation lies in determining an adaptive threshold through a data-driven approach, achieving frequency-specific denoising. Unlike traditional wavelet denoising methods with fixed thresholds, our method uses learnable parameters to dynamically adapt to the noise characteristics of each sub-band.

[0123] , , (10);

[0124] in, This represents the adaptive threshold for the feature map Z, where j represents the subband index. Represents the dynamic scaling factor of the j-th sub-band. Here, Median represents the noise estimate for the j-th subband, and Median denotes the median calculation operation. This represents the j-th sub-band. Let represent the feature map of the j-th subband after processing by the encoder module, |·| represents absolute value operation, Median represents median calculation, and Sigmoid represents the Sigmoid activation function.

[0125] Specifically, These are learnable parameters; the reduction intensity is adjusted based on the feature map of the denoised sub-band, and the noise estimate is... This provides a statistically robust estimate of the noise variance within each subband, reflecting the noise level of the j-th subband. 0.6745 is a constant factor used to convert the median to a standard deviation estimate, and is used to convert the absolute deviation of the median to a standard deviation estimate. The Sigmoid function converts the noise estimates... Mapping to the (0,1) interval generates adaptive threshold coefficients. This dynamic thresholding mechanism enables the model to differentiate noise characteristics across different frequency bands, effectively suppressing residual noise while preserving key fetal electrocardiogram features.

[0126] 2) Soft Thresholding Mechanism: The soft thresholding mechanism applies a calculated threshold to suppress noise components while preserving signal integrity. The reduction operation is defined as:

[0127] (11);

[0128] in, This represents the sub-band after thresholding. ReLU implements soft thresholding behavior, while the sign function is used to preserve the sign information of the original features. This indicates an adaptive threshold.

[0129] This operation effectively suppresses noise components below the adaptive threshold while preserving the phase information of the original signal by maintaining the sign. Specifically, when the absolute value of an eigenvalue is less than the threshold, the ReLU function sets it to zero, thus eliminating noise; when the absolute value of an eigenvalue is greater than the threshold, the ReLU function retains the portion exceeding the threshold and performs only a linear transformation, thus removing noise while preserving the useful components of the signal. The sign function ensures that the polarity (positive or negative) of the signal is maintained during thresholding, which is crucial for accurately reconstructing the fetal electrocardiogram. Through this soft thresholding mechanism, the RS module can effectively reduce noise interference and improve signal quality without losing key morphological features of the fetal electrocardiogram.

[0130] Signal reconstruction decoder:

[0131] The reconstruction phase synthesizes denoised subbands using a 1D convolutional block that learns the optimal ensemble weights. The reconstruction process is defined as follows:

[0132] (12);

[0133] in, Indicates the number of sub-bands. This represents the inverse wavelet transform operation. This represents the feature map of the j-th denoised sub-band after processing by the adaptive waveform shrinkage module. This represents the reconstructed fetal electrocardiogram signal.

[0134] Specifically, The inverse wavelet transform is achieved through stacked 1D convolutional layers, instance normalization layers, and a sinusoidal activation function. This design enables the model to capture complex temporal dependencies while preserving the morphological integrity of the reconstructed fetal signal. The processed multi-scale denoised sub-band feature maps are recombined into a time-domain signal via inverse wavelet transform, resulting in a denoised and enhanced fetal electrocardiogram (ECG). During this process, the 1D convolutional blocks learn and apply optimal weights to ensure that the sub-band signals are optimally fused, preserving the detailed features of the fetal ECG while further eliminating residual noise. The final output sf signal is closer to the real fetal ECG in both time and amplitude.

[0135] Loss function:

[0136] Accurate fetal electrocardiogram (fECG) estimation requires joint optimization of temporal accuracy and spectral fidelity. To this end, we propose a composite loss function that combines three complementary objectives, formulated as follows:

[0137] (13);

[0138] in, These represent the estimated fetal electrocardiogram (ECG) signal and the actual fetal ECG signal, respectively. Represents morphological loss, used to measure the similarity of signals in the spatial domain. This indicates the Fourier transform operation. and λ1, λ2, and λ3 represent the spectral representations of the estimated fetal electrocardiogram (ECG) signal and the actual fetal ECG signal, respectively, and are hyperparameters for adjusting the loss weights.

[0139] Specifically, the first L1 loss By calculating the sum of the absolute differences between the estimated and true signals, the model is forced to precisely align temporally with key features of the fetal electrocardiogram, such as the QRS complex. The second term is morphological loss. This compares the signal's brightness, contrast, and structural information at different scales, ensuring that the model captures subtle morphological changes while preserving the overall signal shape. The third term is the frequency domain L2 loss. By comparing the energy distribution of the estimated signal with that of the real signal in the frequency domain, the model's ability to model spectral characteristics is optimized, ensuring accurate recovery of the spectral components of the fetal electrocardiogram. This composite loss function design allows the model to be optimized simultaneously in the time, space, and frequency domains, thereby significantly improving the accuracy and robustness of fetal electrocardiogram estimation.

[0140] For example, , , By setting these hyperparameters appropriately, the impact of different loss terms on model training can be balanced.

[0141] Time loss: Minimize amplitude deviation and maintain reference point alignment (e.g., R-wave peak timing accuracy).

[0142] Shape loss Used to maintain waveform structural similarity, defined as:

[0143] (14);

[0144] Where P and T represent the predicted morphological feature vector and the target morphological feature vector, respectively.

[0145] Spectral loss: Forced frequency distribution uniformity through Fourier transform ℱ(·).

[0146] Experimental verification:

[0147] A. Experimental setup:

[0148] 1) Datasets: To evaluate the performance of the proposed method, we used two publicly available datasets: the Abdominal and Direct Fetal Electrocardiogram Database (ADFECGDB) and the B2LABOUR dataset. Specifically, the ADFECGDB (AD) dataset contains five fetal electrocardiogram (FECG) records: r01, r04, r07, r08, and r10. Each record contains four abdominal MECG channels and one FECG record obtained directly from the fetal scalp. In our experiments, a five-fold cross-validation scheme was applied to this dataset.

[0149] In addition, the B2LABOUR (B2) dataset contains 12 recorded sessions, each containing four abdominal channels sampled at 500Hz, one direct FECG ground truth sampled at 1kHz, and four additional channels for MECG suppression to enhance FECG visibility. For standardized evaluation, a six-fold cross-validation scheme is used on this dataset, with two records used for testing and the remaining ten for training each fold.

[0150] 2) Implementation Details: The multi-channel signal sensing network was implemented using PyTorch 1.8.0. During training, we used a batch size of 16 and trained the model for 200 epochs using the LambdaLR optimizer, with an initial learning rate of... Training and testing were conducted on workstations equipped with NVIDIA GeForce A100 GPUs. To improve the quality of the input signal, a digital comb filter was used to remove power frequency interference and baseline drift, while a low-pass digital filter eliminated white noise with frequencies above 200Hz.

[0151] 3) Evaluation Metrics: To evaluate the detection performance of the F-QRS complex, we used the following metrics: F1-score (F1), sensitivity (Sen), and positive predictive value (PPV). Additionally, we used mean squared error (MSE) and R-squared goodness of fit (R²). Waveform quality of the FECG signal was predicted using wavelet energy-based diagnostic distortion (WEDD). For fetal QRS complex detection, the Pan-Tompkins method was applied to the predicted FECG signal.

[0152] B. Comparison with the state-of-the-art (SOTA) method on the ADFECGDB dataset:

[0153] In this section, we compare the proposed multichannel signal sensing network with several state-of-the-art (SOTA) methods, including Varanini et al., Behar et al., encoder-decoder, CycleGAN, PA-Net, and CSGSA-Net. These methods demonstrate varying degrees of success in fetal R-peak detection and waveform reconstruction quality.

[0154] For fetal R-wave detection, as shown in Table I, the proposed multi-channel signal sensing network outperforms all previous methods. Our method achieves the highest positive predictive value (PPV) of 99.82%, sensitivity of 99.77%, and F1-score of 99.83%. Specifically, the multi-channel signal sensing network improves upon CycleGAN, PA-Net, and CSGSA-Net, which have PPVs of 99.23%, 99.54%, and 99.59%, respectively. These significant improvements demonstrate that the multi-channel signal sensing network provides superior performance in fetal R-wave detection, consistently delivering higher detection accuracy with lower variance.

[0155] For waveform reconstruction quality, we use mean square error (MSE) and R-squared (R²). The method was compared with wavelet energy-based diagnostic distortion (WEDD). As shown in Table II, the multichannel signal sensing network significantly outperformed all other methods in these metrics. It achieved the lowest MSE of 0.034, the highest R-squared of 93.65%, and the lowest WEDD of 6.81%. This marks a significant improvement over CycleGAN, which has an MSE of 0.042, an R-squared of 92.71%, and a WEDD of 7.08%.

[0156] The reduction in mean square error (MSE) and weighted root mean square error (WEDD) highlights the higher quality reconstruction provided by the multi-channel signal sensing network, ensuring better preservation of waveform features compared to other methods. Finally, the proposed multi-channel signal sensing network significantly improves waveform reconstruction quality, performing excellently on all evaluation metrics. These improvements, coupled with minimal increase in computational cost compared to previous methods, underscore the effectiveness of the multi-channel signal sensing network approach. Visual comparison and detailed analysis of the reconstructed waveforms can be performed... Figure 3 Found it. Figure 3 In the diagram, (a) represents the input abdominal electrocardiogram (AECGs), (b) represents the reconstructed fetal electrocardiogram signal, and (c) represents the actual fetal electrocardiogram signal.

[0157] Table I: Performance of fetal R-peak detection on ADFECGDB (AD) and B2_LABOUR (B2) datasets;

[0158]

[0159] Table II: Waveform reconstruction quality on the ADFECGDB (AD) and B2_LABOUR (B2) datasets;

[0160]

[0161] C. Comparison with state-of-the-art (SOTA) methods on the B2LABOUR dataset:

[0162] In this section, we evaluate the performance of the proposed multichannel signal sensing network on the B2LABOUR dataset and compare it with similar methods. The results in Tables I and II demonstrate the robustness and versatility of the multichannel signal sensing network in fetal R-wave peak detection and waveform reconstruction.

[0163] For fetal R-wave peak detection, as quantified in Table I, the proposed multichannel signal sensing network achieved the highest scores across all three metrics: PPV of 99.87%, sensitivity of 99.85%, and F1 score of 99.86%. This represents a consistent improvement over the strongest benchmarks, surpassing PA²Net's 99.17% and CycleGAN's 98.95% in detection accuracy.

[0164] In waveform reconstruction, the multichannel signal sensing network achieved the lowest mean squared error (MSE) of 0.038 and the lowest waveform distortion (WEDD) of 97.17%, significantly outperforming competing methods. The significantly lower WEDD confirms the multichannel signal sensing network's superior ability to preserve diagnostically important waveform features. In summary, the experimental results consistently validate the superior performance and strong generalization ability of the multichannel signal sensing network on this comprehensive dataset, highlighting its effectiveness in non-invasive fetal ECG analysis.

[0165] In summary, this invention proposes a wavelet transform attention network (multi-channel signal sensing network) for extracting fetal electrocardiograms (fECG) from abdominal recordings, addressing challenges such as maternal ECG interference and low signal-to-noise ratio (SNR). The tomographic attention transform module aims to focus on the fetal component by adaptively suppressing maternal interference, while the adaptive waveform contraction module preserves key fetal ECG features. Experimental results demonstrate that the multi-channel signal sensing network outperforms state-of-the-art methods, improving the reliability of prenatal monitoring.

[0166] This invention also provides a fetal electrocardiogram extraction system based on multi-channel signal sensing, comprising:

[0167] At least one processor;

[0168] At least one memory for storing at least one program;

[0169] When the at least one program is executed by the at least one processor, the at least one processor performs the method described above.

[0170] The content of the above method embodiments is applicable to this embodiment. The specific functions implemented in this embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments. Therefore, they will not be repeated here.

[0171] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0172] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0173] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0174] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0175] This invention also provides a computer program product, including a computer program or computer instructions, which are stored in a memory. A processor of a computer device reads the computer program or computer instructions from the memory and executes the computer program or computer instructions, causing the computer device to perform the above-described method.

[0176] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0177] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0178] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically include computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0179] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method for extracting fetal electrocardiogram based on multi-channel signal sensing, characterized in that, The method comprises the following steps: S100, acquiring a multi-channel abdominal signal and a trained multi-channel signal perception network; wherein the multi-channel signal perception network comprises a self-learning Q-guided wavelet module, an encoder module, a tomographic attention transformation module, an adaptive waveform shrinkage module, and a decoder module; S200, inputting the multi-channel abdominal signal into the trained multi-channel signal perception network, decomposing the multi-channel abdominal signal into multi-scale feature maps through the self-learning Q-guided wavelet module; encoding the multi-scale feature maps through the encoder module to obtain a feature mapping graph; processing the feature mapping graph through the tomographic attention transformation module to generate enhanced output features; performing noise suppression on the enhanced output features through the adaptive waveform shrinkage module to obtain denoised sub-band feature maps processed by the adaptive waveform shrinkage module; and reconstructing the fetal electrocardiogram signal by reconstructing the denoised sub-band feature maps through the decoder module; In S200, the processing of the feature mapping graph through the tomographic attention transformation module to generate enhanced output features comprises: S211, dividing the feature mapping graph into feature mapping graphs corresponding to the query channel group, the key channel group, and the value channel group, respectively using linear transformation matrices of the query, the key, and the value, and combining weight parameters of the query channel group, the key channel group, and the value channel group to calculate interaction weights; S212, adjusting the linear transformation matrix of the value channel group using the interaction weights, and multiplying the interaction weights and the linear transformation matrix corresponding to the value channel group to generate enhanced output features; In S200, the noise suppression on the enhanced output features through the adaptive waveform shrinkage module to obtain denoised sub-band feature maps processed by the adaptive waveform shrinkage module comprises: S221, obtaining a noise estimate value of each sub-band, wherein the noise estimate value is obtained by calculating the median absolute deviation of the feature mapping graph processed by the encoder module and dividing the median absolute deviation by a constant factor, and the constant factor is used to convert the median absolute deviation into a standard deviation estimate; S222, inputting the noise estimate value of each sub-band into a Sigmoid activation function to generate an adaptive threshold coefficient, and taking the product of the adaptive threshold coefficient and the noise estimate value as an adaptive threshold; S223, performing soft threshold processing on each sub-band in the enhanced output features according to the adaptive threshold to obtain denoised sub-band feature maps processed by the adaptive waveform shrinkage module; In S200, the reconstruction of the fetal electrocardiogram signal by reconstructing the denoised sub-band feature maps through the decoder module comprises: S231, inputting the denoised sub-band feature maps processed by the adaptive waveform shrinkage module into the decoder module; S232, the decoder module realizes signal reconstruction through a stacked 1D convolution layer, an instance normalization layer, and a sine activation function, recombines the multi-scale denoised sub-band feature maps into a time domain signal to obtain a denoised and enhanced fetal electrocardiogram signal.

2. The method of claim 1, wherein, In S211, the calculation of the interaction weights by respectively using linear transformation matrices of the query, the key, and the value, and combining weight parameters of the query channel group, the key channel group, and the value channel group comprises: S2111, determine the linear transformation matrix corresponding to the query channel group and the key channel group, and the dimension of the key channel group; S2112, perform a scaled dot product operation on the linear transformation matrix corresponding to the query channel group and the key channel group, and divide by the square root of the dimension of the key channel group for scaling; S2113, normalize the scaled dot product result using the Softmax operation to obtain the interaction weight, wherein the Softmax operation is used to convert the dot product result into a probability distribution, so that the sum of all interaction weights is 1.

3. The method of claim 1, wherein, In S223, the adaptive threshold is used to process each sub-band in the enhanced output feature to obtain a denoised sub-band feature map processed by the adaptive waveform shrinkage module, including: When the absolute value of the feature value is less than the adaptive threshold, it is set to zero; When the absolute value of the feature value is greater than the adaptive threshold, the part exceeding the threshold is reserved and linearly transformed, and the sign function is used to retain the sign information of the original feature to obtain a denoised sub-band feature map processed by the adaptive waveform shrinkage module.

4. The method of claim 1, wherein, The composite loss function of the multi-channel signal perception network is constructed by the following method: S101, obtaining an estimated fetal electrocardiogram signal and a true fetal electrocardiogram signal, wherein the estimated fetal electrocardiogram signal is obtained by processing a multi-channel abdominal signal through a multi-channel signal perception network; S102, obtaining an L1 loss by calculating the absolute difference between the estimated fetal electrocardiogram signal and the true fetal electrocardiogram signal; obtaining a predicted morphology feature vector and a target morphology feature vector, and obtaining a morphology loss by calculating the similarity between the predicted morphology feature vector and the target morphology feature vector; the predicted morphology feature vector and the target morphology feature vector respectively represent the feature mapping of the estimated fetal electrocardiogram signal and the true fetal electrocardiogram signal in the waveform structure; after performing a Fourier transform operation on the estimated fetal electrocardiogram signal and the true fetal electrocardiogram signal, corresponding frequency spectrum representations are obtained, and a frequency domain L2 loss is obtained by calculating the Euclidean distance between the two frequency spectrum representations; S103, weighting and summing the L1 loss, the morphology loss and the frequency domain L2 loss according to the set hyperparameters to construct a composite loss function.

5. A multi-channel signal based fetal electrocardiogram sensing extraction system, characterized by, Comprise: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method of any one of claims 1 to 4.

6. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 5. The computer program is executed by the processor to implement the method of any one of claims 1 to 4.

Images