Self-adaptive electrocardiosignal compression method based on meta-learning

Abstract

The invention discloses a meta-learning-based adaptive electrocardiosignal compression method, which comprises the following steps of: acquiring a target electrocardiosignal, and preprocessing the electrocardiosignal to obtain a preprocessed electrocardiosignal; inputting the preprocessed signal into a nested bottleneck auto-encoder, and selecting a compression ratio through a dynamic route; extracting meta-features of the signal segments, and selecting a compression model through a meta learner according to the meta-features and a preset reconstruction error threshold; when the prediction entropy exceeds a set threshold value, switching to a guarantee compression mode in which differential pulse code modulation and bzip2 are mixed; and carrying out compressed encoding on the electrocardiosignal by using the selected model. According to the method, the proper compression multiple model can be adaptively selected according to the complexity of the signal, so that better compression performance and higher compression efficiency are realized in a given reconstruction error range.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular relates to an adaptive electrocardiogram signal compression method based on meta-learning. Background Art

[0002] The electrocardiogram (ECG) is a crucial technology for detecting heart disease and monitoring treatment effectiveness. Thanks to the rapid development of the Internet of Things (IoT), doctors can now continuously monitor patients' vital signs to better assess their health remotely through wearable ECG devices. However, long-term ECG acquisition and processing often involve large amounts of data, requiring significant storage and transmission resources. This poses a significant challenge for resource-limited IoT devices. Therefore, there is an urgent need to design effective ECG signal compression methods to reduce resource consumption. To address this issue, cutting-edge technology research and development teams around the world have proposed ECG signal compression methods based on both lossless and lossy methods.

[0003] Lossless compression is generally based on lower-order linear prediction and entropy coding. It allows for rapid encoding and decoding, and high-precision restoration of the original data, making it useful for preventing diagnostic errors. However, the compression ratio is generally low. Common methods include delta encoding (DE), Huffman coding, and multi-level tree set splitting coding. If the compression ratio achieved using these methods directly is acceptable, lossless compression is the preferred technique. The pseudo-periodicity of ECGs and the similarity between heartbeats facilitate data compression. Therefore, based on lossless compression, researchers have explored ways to improve the compression ratio. Several lossy compression methods have been proposed, which lose some of the original signal information, improving the compression ratio while maintaining the quality of signal reconstruction. Lossy compression can be further categorized into signal processing methods and deep learning methods.

[0004] Using direct time-domain mathematical methods to reduce data redundancy can further improve the compression ratio, such as differential pulse code modulation, singular value decomposition, Tchebichef moments, and adaptive dictionary-based algorithms. However, the compression ratios of these methods are still relatively low. Signal processing methods in various transform domains are also commonly used for ECG compression, such as the fast Fourier transform, discrete cosine transform, and discrete wavelet transform, by adjusting the threshold to retain the required coefficients. Currently, the tunable-Q wavelet transform (TQWT) exhibits relatively good compression performance. In 2023, Hardev of the Indian Information Technology Center used optimization methods such as particle swarm optimization to adjust the Q value, thereby achieving adaptive compression. ECG compression is essentially a process of extracting concentrated feature expressions. Some research uses characteristic wavelet location to extract peak features for compression. For example, in 2024, Lin Yujie et al. from Southeast University located the QRS complex while retaining relevant features to achieve compression. Although signal processing methods offer high interpretability and low computational complexity, their compression performance is still very limited. Moreover, these methods often use empirical parameters, and the parameters need to be readjusted for signals with different degrees of regularity, while adaptive search will introduce additional computational complexity. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes an adaptive ECG signal compression method based on meta-learning, which makes full use of the shared neural network structure of each compression ratio model, while maintaining the ECG compression reconstruction accuracy, greatly reducing the model storage complexity.

[0006] To achieve the above objectives, the present invention provides an adaptive ECG signal compression method based on meta-learning, comprising:

[0007] Acquiring a target electrocardiogram signal, and preprocessing the electrocardiogram signal to obtain a preprocessed electrocardiogram signal;

[0008] Inputting the preprocessed signal into a nested bottleneck autoencoder and selecting a compression ratio through dynamic routing;

[0009] Extract the meta-features of the signal segments and select a compression model based on the meta-features and a preset reconstruction error threshold through a meta-learner;

[0010] When the predicted entropy exceeds the set threshold, it switches to the guaranteed compression mode that combines differential pulse code modulation and bzip2;

[0011] Use the selected model to compress and encode the ECG signal.

[0012] Optionally, the preprocessing process includes: using a 4th-order Butterworth bandpass filter to eliminate baseline drift; performing sliding segmentation on the signal sampled at 360 Hz using a 1024-sampling point window; and performing range normalization on the segmented signal segments to obtain a preprocessed ECG signal.

[0013] Optionally, the construction process of the nested bottleneck autoencoder includes: a fully shared encoder extracts primary features; four branches respectively achieve different compression ratios through partially shared encoder layers and branch-specific encoder layers; and the decoder includes an inverse operation block corresponding to the encoder.

[0014] Optionally, the training process of the branch-specific encoder layer includes: initially training the branch with the highest compression ratio until convergence; unfreezing and training the specific parameters of each branch in stages in order from low to high compression ratio; and adopting dual reconstruction loss constraints in the incremental training stage.

[0015] Optionally, the meta-feature extraction process includes: calculating the signal mean and standard deviation; detecting the number of intersections between the signal and three quantile lines; calculating the correlation coefficient between the four equal sub-segments; and obtaining the maximum amplitude value of each sub-segment.

[0016] Optionally, the application process of the meta-learner includes: using an XGBoost classifier to establish a mapping relationship between meta-features and the optimal model; calculating a predicted entropy value to evaluate selection reliability; and activating a guaranteed compression mode when the entropy value exceeds 0.8.

[0017] Optionally, the implementation process of the guaranteed compression mode includes: performing second-order differential processing on the signal; quantizing the differential result into 8-bit data; and using the bzip2 algorithm for lossless compression.

[0018] Optionally, the branch-specific encoder layer includes: a hybrid local channel attention module and a multi-scale convolution module;

[0019] The hybrid local channel attention module is used for feature weight allocation;

[0020] The multi-scale convolution module is used to support structural reparameterization.

[0021] Technical effect of the invention: The present invention discloses an adaptive ECG signal compression method based on meta-learning, which can adaptively select a suitable compression multiple model according to the complexity of the signal, thereby achieving better compression performance and higher compression efficiency within a given reconstruction error range. The design of NBAE reduces the consumption of storage and computing resources of candidate models. The uncertainty estimation of the meta-learner can avoid unnecessary decompression operations in the compression stage, and the time occupied by the model selection strategy using meta-learning is negligible. The present invention proposes an efficient module that combines lightweight channel attention and multi-scale deep convolution, which can capture richer features and highlight the importance, thereby improving the performance of the neural network. At the same time, the most reliable compression model is introduced into the compression strategy based on meta-learning. On the one hand, it serves as one of the predictable models of the meta-learner. On the other hand, the uncertainty of the meta-learner is first estimated to judge the reliability of the model selection in advance. If the predicted entropy is higher than the set critical value, this model can ensure high-quality compression and reconstruction of the signal, thereby improving the robustness and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0023] Figure 1 Schematic diagram of a process of an adaptive electrocardiogram signal compression method based on meta-learning according to an embodiment of the present invention;

[0024] Figure 2 Schematic diagram of the model structure of the nested bottleneck autoencoder according to an embodiment of the present invention;

[0025] Figure 3 Schematic diagram of the NBAE staged training procedure shown in Algorithm 1 of an embodiment of the present invention. DETAILED DESCRIPTION

[0026] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0027] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0028] With the advancement of artificial intelligence and deep learning, a growing number of studies are focusing on the compression of ECG signals using neural network-based autoencoders. In 2024, Yang Minxi et al. from Xidian University used a stacked semantic encoder to compress heartbeats and transmit them to the cloud. They then used stacked semantic decoding for simultaneous classification, enabling deployment on IoT devices and saving communication resources. In 2024, He Peng et al. from Chongqing University of Posts and Telecommunications used a residual convolutional autoencoder to simultaneously compress ECG and photoplethysmography signals. They then combined this with a generative adversarial network for encryption, proposing the HCEN model that achieved good performance on both tasks. Some work has also combined deep neural networks with lossless coding. For example, Yangyang Chang from the University of Minnesota used a network with a smaller compression factor and then used Huffman coding to compensate for some of the compression factor, achieving lightweight compression of 12 leads. Although neural network compression methods can automatically learn key features and achieve better performance than mathematical methods, these studies all use fixed compression ratio models, which cannot guarantee quality for specific signals. This is because ECG signals vary in low rank. Irregular signals require more information to store, and neural networks often fail to generalize to unseen samples. Compression quality varies depending on the signal; failure to do so can result in the loss of useful clinical information and lead to misdiagnosis. In response, in 2024, Tahir Bekiryaz et al. from Bursa Technical University in Turkey proposed an ECG compression method using convolutional autoencoders with constant error and adjustable compression ratio. This method uses a shared encoder and independent decoders to achieve different compression ratios, then uses a heuristic search method to compare the compression error with a set threshold to ensure an upper bound on the compression error. This method relies solely on neural networks and is still suboptimal when encountering signals that all networks cannot reconstruct well. Furthermore, using multiple models requires additional storage space and computational complexity, a problem that is particularly pronounced when the number of models is large. There is still room for improvement in the efficient deployment and selection of appropriate models.

[0029] Existing ECG compression technologies can be roughly divided into two categories: signal processing and deep learning. On the one hand, signal processing methods rely on prior signal decomposition such as wavelet transform, which cannot effectively characterize the distorted ECG waveforms commonly found in dynamic electrocardiograms. While ensuring signal reconstruction accuracy, their compression ratios are generally low, and the compression performance is heavily dependent on the manual selection of empirical parameters, making it difficult to automatically process ECG data in large quantities. On the other hand, deep learning methods mostly use fixed compression ratio models, which cannot flexibly adapt to changing ECG waveforms, making it difficult to strike a balance between signal reconstruction accuracy and compression ratio. The few deep learning methods that use multiple compression ratio models have high storage complexity and inefficient model selection strategies, making them unsuitable for deployment on terminals with limited computing resources.

[0030] like Figure 1 As shown, this embodiment provides an adaptive ECG signal compression method based on meta-learning, including:

[0031] Preprocess the dataset, including removing baseline drift noise, segmenting the data by sliding window, and performing range normalization on the segmented segments to eliminate model input differences;

[0032] In the offline training stage, the proposed nested bottleneck autoencoder (NBAE) needs to be trained to provide candidate models with different compression ratios, and then the meta-features of the training set and the model with the best reconstruction effect are used as labels to train the meta-learner.

[0033] In the online application phase, the meta-learner is used to select the model and estimate the uncertainty, thereby achieving efficient, flexible, and reliable ECG data compression. The specific steps are as follows:

[0034] Step 1: Data Preparation:

[0035] Step 1.1: Prepare the dataset:

[0036] The MIT-BIH Arrhythmia Database was used as the experimental dataset. This dataset contains 48 two-lead 30-minute ECG records sampled at 360 Hz and stored in 212 format (two 12-bit sample points per two bytes).

[0037] Step 1.2: Data preprocessing:

[0038] A 4th-order Butterworth bandpass filter (0.5-150 Hz) was used to remove baseline drift. The signal was segmented into segments containing multiple heartbeats using a sliding window of 1024 sampling points (approximately 2.84 seconds). These segments were then randomly divided into training, validation, and test sets in an 8:1:1 ratio.

[0039] Step 2: Training and preparation of compression model:

[0040] Step 2.1: Design of Nested Bottleneck Autoencoder:

[0041] NBAE architecture Figure 2 As shown in Figure 1, there are four information bottlenecks, each of which works independently. By partially sharing the encoder and decoder, we can maximize the use of model parameters with high compression ratios and introduce bottleneck parts with a small number of parameters that support different compression ratios, thereby implementing models with multiple compression ratios without increasing too much storage space. During model inference, NBAE is a dynamic hard routing method, that is, only one path is activated and executed. Fully shared encoder E fulExtracting local features from multiple channels, expanding channels and reducing spatial dimensions can obtain a preliminary potential representation of ECG signals. Partially shared encoder layer Further reduce the spatial dimension, where j∈[1,2,3,4]. NBAE’s intermediate features retain the uncompressed channels, which can avoid premature information loss and ensure the integrity of the shared parts. Specific encoder layers in the bottleneck branch Further reducing the number of channels to 1, achieving different compression ratios. The decoder contains branch specific blocks , some shared blocks and fully shared blocks D ful , which is the inverse of the corresponding encoder, thereby recovering the reconstructed signal. The efficient multi-scale convolution block (EMCB) used in the model combines mixed local channel attention (MLCA) and a multi-scale convolution block (MSCB). MLCA highlights important features by fusing global and local contextual information, while the multi-scale convolution block enriches the extracted features through different receptive fields and is structurally reparameterizable.

[0042] Step 2.2: Training of the Nested Bottleneck Autoencoder

[0043] The NBAE model has multiple shared encoding and decoding layers. Training a model with only one compression factor without considering other compression factors will result in a loss of balance. However, it is difficult to adjust the weights of the respective loss functions when jointly training the model parts with multiple factors. In essence, a specific compression factor faces different specific tasks, so a heuristic training strategy is designed to optimize the branch-specific parameters of NBAE to effectively adapt to complex tasks, as shown in Algorithm 1. First, in the initial training stage, the branch with the highest compression factor is trained to converge to maintain the integrity of the model. Subsequent incremental training starts from the branch with the lowest compression factor, and each branch is gradually unfrozen and trained: (1) a specific encoder and a specific decoder (stage 1); (2) adjacent shared encoders and shared decoders (stage 2), ensuring backward compatibility through selective parameter updates. In the initial training stage and stage 1, the mean square error loss L1 is used as the objective function, which is calculated as follows:

[0044]

[0045] Among them, s o Represents the input signal of the model, s rRepresents the reconstructed signal of the model, and n is the length of the signal. In stage 2, in order to maintain the performance of the branch with the highest compression factor during incremental training, the reconstruction error of the current branch is added to the reconstruction error of the initial training branch constrained by λ using the dual-objective loss L2, calculated as follows:

[0046]

[0047] in, represents the reconstruction signal of the j-th model branch currently being trained, This represents the reconstructed signal of the fourth model branch used in initial training. Training was performed using the Adam optimizer, with back-gradient propagation. Labeled samples from the validation set were used to determine hyperparameters such as the control factor, learning rate, and number of training rounds, and the optimal model was selected for testing. Because the mean square error (MSE) is small and difficult to observe, the root mean square error (PRD) percentage was used as an evaluation metric for the quality of the reconstructed signal. Its calculation formula is as follows:

[0048]

[0049] Step 2.3: Design the safest compression model:

[0050] Because deep learning can fail to generalize when presented with unfamiliar data, a rapidly deployable compression model is employed when other models fail to provide adequate reconstruction results. This method is an improvement on DE encoding: the signal is first subjected to second-order differencing to increase redundancy, then quantized to 8 bits, and further losslessly compressed using the efficient bzip2 encoding. While the DE-bzip2 hybrid encoding yields lower compression, it ensures reliable compression.

[0051] Step 3: Meta-learner training and preparation:

[0052] Step 3.1: Design and extraction of meta-features:

[0053] The compression effectiveness of the model depends on the waveform's morphological characteristics. For longer ECG signals, the main influencing factors include the regularity and variability of the heartbeats, and the noise content of the signal. A normal heartbeat lasts 0.75 to 0.8 seconds, so the ECG segments segmented in step 1 contain approximately three to four beats. Based on the characteristics of the ECG signal, 14 meta-features were designed for each signal segment, as shown in Table 1. The mean and variance of the normalized signal reflect the noise content. Zero crossings indicate lead dropout, high-frequency interference, or heartbeat distortion. The correlation coefficient and range of the four-part segments reflect the different types of heartbeats within a signal segment. If all heartbeats have similar morphology, less information is preserved than if beats have diverse categories. The Pearson correlation coefficient is calculated for each four-part segment with the other three four-part segments, and the average is taken as the correlation coefficient indicator for the segment. The maximum PRD can be used as a compression condition and can also be used as a feature, eliminating the need for one-hot encoding.

[0054] Table 1

[0055]

[0056] Step 3.3: Meta-learner design and training

[0057] Given an ECG signal, the meta-learner's task is to map the extracted meta-features to the most appropriate compression model labels, which is essentially a classification problem. Extreme gradient boosting (XGBoost) was chosen as the meta-learner due to its interpretability and low computational overhead. XGBoost efficiently implements gradient boosting through regularized learning and a cache-aware tree structure, and improves accuracy through a second-order Taylor expansion. XGBoost uses decision trees as weak learners, and its performance is highly dependent on hyperparameters such as tree depth, number of trees, and learning rate.

[0058] After step 2, we now have five models with different compression factors. We compress and reconstruct all signal segments in the training set using these models, selecting the models that meet the PRD threshold set in the meta-features as labels. We then use these labels and the corresponding meta-features to train a meta-learner, using cross-entropy as the XGBoost loss function.

[0059] Step 4: Application of adaptive compression method:

[0060] In the application stage, for the pre-processed data that needs to be compressed, meta-features are first extracted according to Table 1, and then it needs to be predicted by the meta-learner and uncertainty estimation is performed. Because the meta-learner cannot always predict the best results. A common practice is to use the safest method as a guarantee when the reconstruction error of the prediction model does not meet the requirements, but this requires decompression during compression, which adds additional running time. To solve this problem, the need for a guarantee method is determined by measuring the prediction uncertainty of the meta-learner. The higher the prediction entropy, the higher the uncertainty, and the prediction is distributed in more classes. Even if the prediction is incorrect, the reconstruction error is sometimes still acceptable, so uncertainty tolerance needs to be considered. Therefore, it is stipulated that: if the total prediction entropy of a meta-learner under multiple PRD conditions is higher than the set boundary value, then the meta-learner's prediction of this signal is unreliable and should switch to the compression method of step 2.3. The prediction entropy H of the i-th segment i The calculation formula is as follows:

[0061]

[0062] Where, is the predicted probability of the meta-learner, f i is the meta-feature of the segment, θ is the set PRD condition, is the model number predicted by the meta-learner. According to the preparation in step 2, there are M = 5 numbers in total.

[0063] After the meta-learner predicts and estimates uncertainty, the selected candidate model can be used for compression. The model selection step was verified to be time-efficient during actual C++ program deployment. The experiment used eight levels of PRD, which not only improved uncertainty tolerance but also enabled more comprehensive training of the meta-learner, enabling it to select the best candidate model for most scenarios and maximize the combined power of the candidate models.

[0064] The hybrid channel attention module in the EMCB module in step 2.1 can be replaced with a convolutional block attention module or a more advanced channel attention module. Reparameterizable multi-scale depthwise convolution can be replaced with a single multi-scale convolution. The most reliable compression model in step 2.3 can be replaced with a higher-multiplier lossless compression model, such as LZMA.

[0065] The advantages of the present invention are that: existing technologies cannot effectively balance compression factor and reconstruction error, and cannot quickly and adaptively adjust empirical parameters for signals with different degrees of regularity; the method of using multiple models for mixed compression increases the time of model selection and introduces additional storage space and computational complexity. The present invention proposes a meta-learning-based adaptive ECG signal compression method that can adaptively select an appropriate compression factor model based on the complexity of the signal, thereby achieving better compression performance and higher compression efficiency within a given reconstruction error range. The NBAE design reduces the consumption of storage and computing resources for candidate models, the uncertainty estimation of the meta-learner can avoid unnecessary decompression operations during the compression stage, and the time taken by the meta-learning model selection strategy is negligible. The existing methods that only use deep learning models may fail to generalize when faced with unfamiliar data, resulting in a decrease in the quality of the reconstructed signal. The present invention proposes an efficient module that combines lightweight channel attention and multi-scale deep convolution to capture richer features and highlight importance, thereby improving the performance of neural networks. At the same time, the most reliable compression model is introduced into the meta-learning-based compression strategy. On the one hand, it serves as one of the predictive models of the meta-learner. On the other hand, it first estimates the uncertainty of the meta-learner to judge the reliability of the model selection in advance. If the predicted entropy is above a set threshold, this model can ensure high-quality compression and reconstruction of the signal. This improves the robustness and reliability of the system.

[0066] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. An adaptive electrocardiogram signal compression method based on meta-learning, characterized in that: include: Acquiring a target electrocardiogram signal, and preprocessing the electrocardiogram signal to obtain a preprocessed electrocardiogram signal; Inputting the preprocessed signal into a nested bottleneck autoencoder and selecting a compression ratio through dynamic routing; Extract the meta-features of the signal segments and select a compression model based on the meta-features and a preset reconstruction error threshold through a meta-learner; When the predicted entropy exceeds the set threshold, it switches to the guaranteed compression mode that combines differential pulse code modulation and bzip2; Use the selected model to compress and encode the ECG signal.

2. The adaptive electrocardiogram signal compression method based on meta-learning according to claim 1, characterized in that: The preprocessing process includes: using a 4th-order Butterworth bandpass filter to eliminate baseline drift; using a 1024-sampling point window to perform sliding segmentation on the signal sampled at 360 Hz; and performing range normalization processing on the segmented signal segments to obtain preprocessed electrocardiogram signals.

3. The adaptive electrocardiogram signal compression method based on meta-learning according to claim 1, characterized in that: The construction process of the nested bottleneck autoencoder includes: a fully shared encoder extracts primary features; four branches achieve different compression ratios through partially shared encoder layers and branch-specific encoder layers respectively; and a decoder contains an inverse operation block corresponding to the encoder.

4. The adaptive electrocardiogram signal compression method based on meta-learning according to claim 3, characterized in that: The training process of the branch-specific encoder layer includes: initially training the branch with the highest compression ratio until convergence; unfreezing and training the specific parameters of each branch in stages according to the compression ratio from low to high; and adopting a dual reconstruction loss constraint in the incremental training stage.

5. The adaptive electrocardiogram signal compression method based on meta-learning according to claim 1, characterized in that: The meta-feature extraction process includes: calculating the signal mean and standard deviation; detecting the number of intersections between the signal and three quantile lines; calculating the correlation coefficient between the four equal sub-segments; and obtaining the maximum amplitude value of each sub-segment.

6. The adaptive electrocardiogram signal compression method based on meta-learning according to claim 1, characterized in that: The application process of the meta-learner includes: using the XGBoost classifier to establish a mapping relationship between meta-features and the optimal model; calculating the predicted entropy value to evaluate the selection reliability; and activating the guaranteed compression mode when the entropy value exceeds 0.

8.

7. The adaptive electrocardiogram signal compression method based on meta-learning according to claim 1, characterized in that: The implementation process of the guaranteed compression mode includes: performing second-order differential processing on the signal; quantizing the differential result into 8-bit data; and using the bzip2 algorithm for lossless compression.

8. The adaptive electrocardiogram signal compression method based on meta-learning according to claim 3, characterized in that: The branch-specific encoder layer includes: a hybrid local channel attention module and a multi-scale convolution module; The hybrid local channel attention module is used for feature weight allocation; The multi-scale convolution module is used to support structural reparameterization.

Images