Electrocardiogram feature point positioning method based on double u-net fusion of long and short range context information

Abstract

The application discloses a method for locating electrocardio feature points based on double U-Net fusion of long and short-range context information, comprising: preprocessing electrocardio signals, including resampling to a unified sampling rate and removing noise interference; segmenting the electrocardio signals, cutting into single heart beat segments and adjacent multi-heart beat segments based on R peak positions; applying data enhancement operations to the segmented heart beat segments; extracting features using double U-Net, including extracting local semantic features of single heart beat segments by a first U-Net and extracting global semantic features of multi-heart beat segments by a second U-Net; fusing the local semantic features and the global semantic features through a cross-level double gating module to generate purified fusion features; and locating feature points of the electrocardio signals based on the fusion features, including positions of P waves, QRS wave groups and T waves. The application improves the precision and robustness of feature point location.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of electrocardio feature point positioning, and particularly relates to an electrocardio feature point positioning method based on double U-Net fusion of long and short range context information. BACKGROUND

[0002] Electrocardiogram analysis can be used for screening and diagnosing cardiology diseases, and accurate feature point detection of electrocardio signals is the basis for diagnosis of cardiovascular diseases. The feature point detection algorithm of electrocardio signals faces several difficulties: firstly, electrocardio signals inevitably suffer from noise interference in the process of acquisition, and these interferences are difficult to be completely eliminated, thereby affecting feature extraction of electrocardio signals. In addition, for part of heartbeats, the individual's symptoms may cause the feature waveform to have distortion, so that it is particularly difficult to position feature points through the morphology of a single heartbeat. These challenges jointly constitute the complexity of feature point detection of electrocardio signals. Therefore, it is of great significance to study a flexible and accurate feature point positioning algorithm of electrocardio signals.

[0003] Significant progress has been made in the research of ECG signal feature point localization algorithms, demonstrating technological diversity and innovation. Jimenez et al. transferred the U-Net network from image segmentation to one-dimensional ECG signal processing, achieving efficient feature point localization by adjusting the convolution kernel dimension and introducing regularization techniques such as spatial dropout. Jimenez et al. combined multi-scale upsampling and ultra-dense connections to improve the structure, effectively alleviating the overfitting problem of the localization model through a two-layer U-Net structure and feature skip connections. Addressing the strong dependence of deep learning models on labeled data, Avetisyan et al. proposed synthetic data technology to expand the diversity of training samples. They extracted basic waveform segments such as P, QRS, and T from existing databases, combined probability rules and expert knowledge to construct synthetic ECG signals, and significantly improved the model's generalization ability to different morphological signals through pre-training with synthetic data. Chen et al. used a 1D U-Net and preprocessing and post-processing methods independent of R-peak position to achieve feature point localization for single heartbeats. The shortcomings of existing technologies are as follows: First, current ECG localization methods are limited to single-beat localization. Single-beat localization methods do not fully utilize the guidance of long-sequence context and the low-rank characteristics of ECG signals, lacking contextual dependency information between beats and failing to correct local noise interference through long-range context. This results in insufficient robustness of the model to noise interference and abnormal beats. Second, existing methods lack a guiding mechanism for multi-beat global high-level semantics to extract low-level features, leading to insufficient robustness in key feature point localization. On the one hand, the extracted multi-beat high-level semantic features cannot be used as guiding signals to directionally regulate the extraction process of local low-level features; on the other hand, there is a lack of a secondary purification mechanism for the fusion of long and short-range beat features, failing to effectively leverage the "purification-enhancement" effect of multi-beat global semantics on the local information of single beats, ultimately resulting in unsatisfactory feature point localization accuracy and generalization ability. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for locating ECG feature points based on dual U-Net fusion of long and short-range contextual information, thereby improving the accuracy and robustness of feature point localization.

[0005] To achieve the above objectives, this invention provides a method for locating ECG feature points based on dual U-Net fusion of long and short-range contextual information, comprising:

[0006] Preprocessing of ECG signals includes resampling to a uniform sampling rate and removing noise interference;

[0007] The electrocardiogram signal was segmented into single-heart beat segments and adjacent multi-heart beat segments based on the position of the R peak.

[0008] Apply data augmentation operations to the segmented heartbeat fragments;

[0009] The features are extracted using a double U-Net, including a first U-Net extracting local semantic features of a single heartbeat segment and a second U-Net extracting global semantic features of a multi-heartbeat segment;

[0010] The local semantic features and the global semantic features are fused through a cross-layer double-gating module to generate purified fusion features;

[0011] Based on the fusion features, feature points of the electrocardiogram signal are located, including positions of P waves, QRS wave groups and T waves.

[0012] Optionally, removing noise interference includes:

[0013] A median filter is used to process baseline drift, with a window width of 200 milliseconds and 600 milliseconds, respectively;

[0014] An IIR digital band-stop filter is used to remove power frequency interference, with lower and upper limit cutoff frequencies of 49 Hz and 51 Hz, respectively;

[0015] A wavelet basis function is used for decomposition, and the signal is reconstructed by inverse wavelet transform to remove electromyographic interference.

[0016] Optionally, segmenting the electrocardiogram signal includes:

[0017] The R-peak position is located based on the Pan-Tompkins algorithm;

[0018] A single heartbeat segment is cut based on the R-peak position as the center and the RR interval as the cutting window width;

[0019] A multi-heartbeat segment is cut based on the R-peak position and its adjacent five heartbeats;

[0020] The single heartbeat segment and the multi-heartbeat segment are resampled to a uniform length according to the RR interval.

[0021] Optionally, applying data augmentation operations includes:

[0022] Time translation enhancement is performed with a probability, with a translation length of 0-10% of the total length of the signal;

[0023] Signal inversion enhancement is performed with a probability;

[0024] Gaussian noise is added to the signal with a probability;

[0025] Random occlusion enhancement is performed with a probability, with a randomly selected area filled;

[0026] The label position is adjusted synchronously to maintain alignment between the signal and the label.

[0027] Optionally, the features are extracted using a double U-Net, including:

[0028] The first U-Net and the second U-Net each comprise an encoder and a decoder;

[0029] The encoder receives a signal, and data compression and feature extraction are performed through a downsampling module, each downsampling module comprising a convolution layer, a batch normalization layer, an activation layer, and a pooling layer;

[0030] The decoder performs upsampling through an upsampling module, each upsampling module comprising a deconvolution layer, a batch normalization layer, and an activation layer;

[0031] The encoder and the decoder fuse multi-scale information through a skip connection.

[0032] Optionally, the fusion comprises:

[0033] Cross-attention calculation is performed on the input features to generate attention mapping features;

[0034] Multi-scale convolution operations are performed on the attention mapping features to extract multi-receptive field representations using different convolution kernel sizes;

[0035] The original input features are modulated by a gating weight;

[0036] The modulated features and the attention features are spliced to generate a gating weight through Sigmoid activation;

[0037] The fused features are output by weighted summation, and a residual connection is added.

[0038] Optionally, the feature points of the electrocardiogram signal include:

[0039] A task head is input based on the fused features to obtain a positioning prediction probability; the final positioning accurate position is obtained after post-processing; and a mean square loss function is used to calculate the positioning error.

[0040] Optionally, the mean square loss function comprises:

[0041] The mean square error of the predicted QRS complex position sequence and the real sequence is calculated;

[0042] The mean square error of the predicted P wave position sequence and the real sequence is calculated;

[0043] The mean square error of the predicted T wave position sequence and the real sequence is calculated;

[0044] The reconstruction loss of the single heart beat signal and the multi-heart beat signal is combined.

[0045] The technical effect of the present application: the present application discloses an electrocardio feature point positioning method based on double U-Net fusion of long and short range context information, which captures the correlation information between heartbeats through multi-heartbeat U-net, provides context information guidance for single-heartbeat U-Net, combines long and short range correlation semantic information, guides local feature extraction through high-level semantic information, and improves the precision and robustness of feature point positioning. BRIEF DESCRIPTION OF DRAWINGS

[0046] The drawings constituting a part of the present application are used to provide further understanding of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute improper limitation on the present application. In the drawings:

[0047] Figure 1 The flowchart of the electrocardio feature point positioning method based on double U-Net fusion of long and short range context information of the embodiment of the present application is shown in the figure;

[0048] Figure 2 The overall network framework diagram in the electrocardio feature point positioning method based on double U-Net fusion of long and short range context information of the embodiment of the present application is shown in the figure;

[0049] Figure 3 The cross-level double gate module diagram of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0050] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0051] It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that here.

[0052] As shown in the figure, the electrocardio feature point positioning method based on double U-Net fusion of long and short range context information is provided in the embodiment, which comprises: Figure 1

[0053] The electrocardio signal is preprocessed, including resampling to a unified sampling rate and removing noise interference;

[0054] The electrocardio signal is segmented, and based on the R peak position, it is cut into single heartbeat segments and adjacent multi-heartbeat segments;

[0055] The data enhancement operation is applied to the segmented heartbeat segment;

[0056] ​The features are extracted using a double U-Net, including a first U-Net for extracting local semantic features of a single heartbeat segment and a second U-Net for extracting global semantic features of a multi-heartbeat segment.

[0057] The local semantic features and the global semantic features are fused through a cross-layer double-gating module to generate purified fusion features.

[0058] Based on the fusion features, feature points of the electrocardiogram signal are located, including positions of P waves, QRS complexes, and T waves.

[0059] Specifically, the specific implementation process of the embodiment includes the following steps:

[0060] Step 1: Data preparation:

[0061] Step 1.1: Prepare the data set:

[0062] The electrocardiogram database contains 200 12-lead electrocardiogram records lasting 10 seconds. In addition, two certified cardiologists manually annotated the start and end points and peak values of P waves, T waves, and QRS complexes in the 200 records. The data is divided into training, validation, and test sets in a ratio of 6:2:2. To ensure data independence, ensure that the electrocardiogram segments in the training set, validation set, and test set do not overlap, and the data in each set comes from different subjects' electrocardiogram records.

[0063] Step 1.2: Resample:

[0064] Resample all electrocardiogram data to a sampling rate of 500 Hz.

[0065] Step 1.3: Electrocardiogram signal denoising:

[0066] Electrocardiogram signals generally contain three common noises: baseline drift, power frequency interference, and electromyographic interference. In order to ensure the performance of the electrocardiogram feature point detection algorithm, the original electrocardiogram signal must be denoised.

[0067] (1) Two median filters with window widths of 200 milliseconds and 600 milliseconds are used to handle baseline drift problems.(2) A bilinear transformation method is used to design an IIR digital band-stop filter with lower and upper cutoff frequencies of , to remove power frequency interference.(3) "bior4.4" is selected as the wavelet basis function, and a total of 5 layers of decomposition are performed. The processed signal is reconstructed through inverse wavelet transform, effectively removing electromyographic interference.

[0068] Step 2: Heartbeat segmentation:

[0069] The data processed in step 1 is preprocessed.

[0070] Step 2.1: ECG signal cutting:

[0071] After denoising the ECG signal, it is necessary to cut the signal according to the position of the R peak to cut the ECG signal into single heartbeat segments and adjacent heartbeat segments. The Pan-Tompkins algorithm is used to obtain the R peak position, and then the RR interval is calculated according to the label of the R peak in the data set. The R peak position obtained by the Pan-Tompkins method is taken as the center, and the RR interval is taken as the width of the cutting window, and the single heartbeat cutting is performed to obtain the single heartbeat segment. Similarly, the signal segment of five heartbeats including the R peak position and its adjacent two heartbeats is taken as the adjacent heartbeat segment. After cutting, the RR interval is uniformly resampled to a single heartbeat sequence of 128 lengths and a multi-adjacent heartbeat sequence of 512 lengths, and the sampling rate is adjusted synchronously.

[0072] Step 2.2: Heartbeat data enhancement:

[0073] In order to simulate the complex conditions such as noise interference, signal shape variation and position offset in the actual acquisition scene, and improve the robustness and generalization ability of the model to non-ideal signals. For the ECG signal feature point positioning task, multiple data augmentations are performed on the training set data.

[0074] (1) Time translation enhancement. With a probability of 50%, the single heartbeat signal is translated left and right: the translation length is 0-10% of the total length of the signal, simulating the position fluctuation of the heartbeat on the time axis (such as signal offset caused by breathing and movement). After translation, the position of the label is adjusted synchronously to ensure that the signal and the label are aligned.

[0075] (2) Signal inversion enhancement. With a probability of 20%, the single heartbeat signal is inverted, and the same operation is performed on the adjacent heartbeat signal. This operation simulates the signal polarity inversion caused by different lead acquisition or device differences (such as the switching of positive and negative waves), and enhances the adaptability of the model to signal direction changes.

[0076] (3) Noise injection enhancement. With a probability of 20%, Gaussian noise is added to the single heartbeat signal, and the noise intensity is controlled by the signal-to-noise ratio. The noise is generated by a function that matches the signal and is superimposed. This operation simulates the power frequency interference, electromyographic noise and other noises in the actual acquisition environment, and improves the anti-interference ability of the model in the noise environment.

[0077] (4) Random occlusion augmentation. With 10% probability, randomly select two regions, each with a length of 0.8%-1.0% of the total length of the signal, and fill the occluded regions with a constant. This operation simulates the local missing or severe interference (such as poor electrode contact) in the signal acquisition process, forcing the model to learn the context information of the unoccluded regions, enhancing the fault tolerance to local anomalies.

[0078] Step 3: ECG localization framework based on double U-Net and long-short beat information:

[0079] The proposed network architecture is shown in Figure 2 , which mainly consists of two parts: long-short beat information feature extraction stage and context global-local semantic information purification stage.

[0080] In the long-short beat information feature extraction stage, two 1D U-net networks of different sizes are used to extract features from the near-neighbor beat information and short-time beat information. The model inputs different time long-short beat data into the encoder, which uses a 1D-Unet network to extract ECG features while outputting semantic information for different long-short inputs. The output of the encoder-decoder network after the decoder is used for the reconstruction of the original signal auxiliary task.

[0081] In the context global-local information purification stage, first, the single beat and multi-beat semantic features output by the bottom layer of the encoder-decoder network are deconvolved and upsampled, and the dimensions are aligned. Then, a cross-level double gated attention module is used to obtain the weighted fusion features, which are then input into the task head to obtain the localization prediction probability. Finally, the final localization accurate position is obtained after post-processing.

[0082] Step 4: Double U-net long-short ECG feature extraction network construction:

[0083] Step 4.1: Construct long-short beat feature extraction network:

[0084] The feature extraction network consists of two U-net networks of different levels, each containing an encoder and a decoder. The encoder extracts multi-scale ECG signal features from the ECG signal, and then the decoding process splices different scale information for decoding and upsampling. The encoder receives the original signal, which is compressed and feature-extracted by the down-sampling module. Each down-sampling module contains two 3x3 convolution layers, a batch normalization layer, a Relu activation layer, and a 2x2 max-pooling layer. In the decoding stage, the decoded feature map is spliced with the corresponding layer feature map of the encoder and sent to the decoder through the up-sampling module for up-sampling. Each down-sampling module contains two 3x3 deconvolution layers, a batch normalization layer, and a Relu activation layer. Finally, the network restores the input signal to the original signal while outputting the fused features at each level.

[0085] Step 4.2: Construct a cross-layer dual-gating module:

[0086] This module first extracts relevant information at different time scales using a standard cross-attention mechanism. Then, it uses this information to refine the original input features. The refined main path and branch path features are added together, and then gated and weighted with high-level global semantic information. These are the input features of the main road and the branch road, respectively. The number of feature channels, Let be the feature dimension. First, use... As a query, Cross-attention is performed using the key and value to obtain attention-mapped features:

[0087] (1);

[0088] in, This is the cross-attention output without residual connections. Then, Inputting into a feedforward neural network (FFN) yields a representation that has undergone linear transformation and activation:

[0089] (2);

[0090] in, for Feature representation after passing through a feedforward network; For feedforward network ; This is the cross-attention output without residual connections; For element-wise Sigmoid function operation; These are the linear weights of the feedforward network; This is the nonlinear bias of the feedforward network;

[0091] Then consider the above attention results Feature representation with multiple receptive fields is performed using multi-scale convolutional kernels, where Indicates different convolution kernel sizes:

[0092] (3);

[0093] Then, the gate weights are obtained through two different 1×3 convolutional layers. and Used to modulate the original input features respectively and :

[0094] (4);

[0095] where denotes element-wise multiplication. Then, the modulated inputs are concatenated with and :

[0096] (5);

[0097] where denotes concatenation along the channel dimension. Subsequently, the and are concatenated along the channel dimension to obtain the fused feature and a linear transformation with channel compression and a sigmoid activation function to generate the gating weight:

[0098] (6);

[0099] (7);

[0100] where, denotes element-wise sigmoid function operation, is a weight matrix, is a gating weight, is a fused feature, is a weight matrix of the channel compression linear transformation layer, is a bias of the channel compression linear transformation layer, is the number of fused feature channels, is the dimension of the fused feature; the weight is used to weight the sum of and :

[0101] (8);

[0102] where, denotes element-wise multiplication, after a layer of 1x1 convolution compression channel, the final output is a fused feature representation.

[0103] Finally, the output result is the sum of the multi-layer information fusion feature and the residual connection cross attention output:

[0104] (9);

[0105] Step 4.3: Design and optimization of positioning network loss function

[0106] The loss function is a mean square loss function, and the calculation method is as follows:

[0107] (10);

[0108] wherein, are trainable network parameters, is the total number of training samples, , , are the predicted position sequences of QRS complex, P wave and T wave respectively, , , denote the real sequences of the corresponding positions. and denote the reconstructed single heartbeat signal and the original single heartbeat signal respectively, and denote the reconstructed multi- heartbeat signal and the original multi- heartbeat signal respectively. The present application adopts AdamW optimizer to propagate parameters through backpropagation gradient. The long-short heartbeat feature extraction network based on double U-net includes step 4.1, as shown in

[0109] , through the U-net encoding-decoding network of different sizes, combined with the heartbeat reconstruction auxiliary task, the effective long-short sequence heartbeat feature extraction is realized. Figure 2 The cross-module gating fusion module includes step 4.2, and the module is as shown in

[0110] . The module first obtains the correlation information of long-short range electrocardiogram deep semantics through cross attention, generates fusion features; then the fusion features are used as gating signals, and sigmoid activation is used to purify the original semantic features of long-short heartbeats; then, combined with high-level global semantics, the purified semantic features are subjected to secondary gating, and the purified features fused with local semantic information and global high-level semantic information are output. Figure 3 Single heartbeat signal is easy to be disturbed by noise, resulting in feature distortion and poor positioning accuracy. The existing single heartbeat positioning method only relies on single heartbeat local waveform features, and cannot utilize the low-rank regularity of electrocardiogram signals for noise correction, resulting in significant increase of feature point positioning error in noise scene. The present application proposes a long-short heartbeat feature extraction architecture based on double U-net, which captures high-level semantic information through long sequence multi- heartbeat input, retains local semantic information through single heartbeat input, and combines heartbeat reconstruction auxiliary task to constrain feature extraction, thereby improving the positioning robustness of single heartbeat noise and the generalization ability of feature extraction. In order to effectively utilize high-level global semantic information to guide low-level feature extraction, the present application proposes a double gating purification module, which first purifies primary semantic information by using fusion semantic information, and then combines global fusion semantic information to perform secondary gating on the purified primary semantic information, realizing "global semantic guidance-local detail purification-cross-level fusion" feature purification.

[0111]

[0112] ​The feature extractor of single heart beat and multi heart beat can use other forms of coding and decoding model, such as Transformer, Resnet, nn U-net, etc. The fusion mode is not limited, and any feature fusion means can be used, such as multiple attention mechanisms or convolution modulation.

[0113] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for locating electrocardio feature points based on a dual-U-Net fusion of long and short-range context information, characterized in that, The method comprises the following steps: Preprocessing the electrocardiogram signal, including resampling to a unified sampling rate and removing noise interference; Segmenting the electrocardiogram signal, cutting into single heartbeat segments and adjacent multi-heartbeat segments based on the R-peak position; Applying data enhancement operations to the segmented heartbeat segments; Extracting features using a double U-Net, including a first U-Net extracting local semantic features of single heartbeat segments and a second U-Net extracting global semantic features of multi-heartbeat segments; Fusing the local semantic features and the global semantic features through a cross-level double gating module to generate purified fusion features; Locating feature points of the electrocardiogram signal based on the fusion features, including the positions of P waves, QRS complexes, and T waves; Using a double U-Net to extract features includes: Both the first U-Net and the second U-Net include an encoder and a decoder; The encoder receives the signal, performs data compression and feature extraction through a downsampling module, and each downsampling module includes a convolution layer, a batch normalization layer, an activation layer, and a pooling layer; The decoder performs upsampling through an upsampling module, and each upsampling module includes an inverse convolution layer, a batch normalization layer, and an activation layer; The encoder and the decoder fuse multi-scale information through a skip connection; Fusing through a cross-level double gating module includes: Performing cross-attention calculation on the input features to generate attention mapping features; Performing multi-scale convolution operations on the attention mapping features to extract multi-receptive field representations using different convolution kernel sizes; Modulate the original input features through gating weights; Concatenate the modulated features and the attention features, and generate gating weights through Sigmoid activation; Weighted sum to output fusion features and add residual connection; Locating feature points of the electrocardiogram signal includes: Input the fusion features into the task head to obtain the positioning prediction probability; After post-processing, the final positioning accurate position is obtained; The positioning error is calculated using the mean square loss function; Using the mean square loss function includes: Calculate the mean square error of the predicted QRS complex position sequence and the true sequence; Calculate the mean square error of the predicted P wave position sequence and the true sequence; Calculate the mean square error of the predicted T wave position sequence and the true sequence; Joint single heartbeat signal and multi-heartbeat signal reconstruction loss.

2. The method of claim 1, wherein the method is based on a dual-U-Net fusion of long and short-range context information. Removing noise interference includes: Using a median filter to process baseline drift with window widths of 200 milliseconds and 600 milliseconds; Using an IIR digital band-stop filter to remove power frequency interference, with lower and upper cutoff frequencies of the stop band being 49 Hz and 51 Hz, respectively; Using a wavelet basis function for decomposition, and reconstructing the signal to remove electromyographic interference through inverse wavelet transform.

3. The method of claim 1, wherein the method is based on a dual-U-Net fusion of long and short-range context information. Segmenting the electrocardiogram signal includes: Locating the R-peak position based on the Pan-Tompkins algorithm; Cutting single heartbeat segments with the R-peak position as the center and the RR interval as the cutting window width; Cutting multi-heartbeat segments with the R-peak position and its adjacent five heartbeats; Resample the single heartbeat segments and the multi-heartbeat segments to a unified length according to the RR interval.

4. The method of claim 1, wherein the method is based on a dual-U-Net fusion of long and short-range context information. Applying data enhancement operations includes: Performing time shift enhancement with a probability, with a shift length of 0-10% of the total signal length; Performing signal inversion enhancement with a probability; Adding Gaussian noise to the signal with a probability; Performing random occlusion enhancement with a probability, randomly selecting an area to fill; Synchronously adjust the tag position to keep the signal aligned with the tag.

Images