An adaptive electrocardiosignal compression method based on meta learning

Abstract

The application discloses an adaptive electrocardiosignal compression method based on meta learning, comprising the following steps: obtaining a target electrocardiosignal, preprocessing the electrocardiosignal to obtain a preprocessed electrocardiosignal; inputting the preprocessed signal into a nested bottleneck self-encoder to select a compression ratio through dynamic routing; extracting the meta features of the signal segment, and selecting a compression model according to the meta features and a preset reconstruction error threshold through a meta learner; when the predicted entropy exceeds a set threshold, switching to a differential pulse code modulation and bzip2 hybrid guarantee compression mode; and using the selected model to compress and encode the electrocardiosignal. The application can adaptively select a suitable compression multiple model according to the complexity of the signal, so that better compression performance and higher compression efficiency can be realized within a given reconstruction error range.

Description

Technical Field

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

[0002] Electrocardiography (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 using wearable ECG devices. However, the long-term acquisition and processing of ECG data often involves massive amounts of data, requiring significant storage and transmission resources. This poses a serious challenge to IoT devices with limited resources. Therefore, there is an urgent need to design an effective ECG signal compression method to reduce resource consumption. To address this issue, leading technology research teams around the world have proposed lossless or lossy ECG signal compression methods.

[0003] Lossless compression is generally based on lower-order linear prediction and entropy coding methods, enabling fast encoding and decoding and high-precision reconstruction of original data. It can be used to prevent errors in diagnosis, but the compression ratio is generally low. Common methods include differential coding (delta encoding, DE), Huffman coding, and multilevel tree set splitting coding. If the compression ratio of these methods is acceptable, lossless compression is the preferred technique. The pseudo-periodicity of ECG and the similarity between heartbeats facilitate data compression. Therefore, based on lossless compression, researchers have explored ways to improve the compression ratio, proposing various lossy compression methods that lose some original signal information, thereby improving the compression ratio while ensuring the quality of signal reconstruction. Lossy compression is further divided 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. Examples include differential pulse code modulation (DCM), singular value decomposition (SVD), Tchebichef moments, and adaptive dictionary-based algorithms. However, the compression ratios of these methods are still relatively low. Signal processing methods in different transform domains are also commonly used for ECG compression, such as Fast Fourier Transform (FFT), Discrete Cosine Transform (DCT), and Discrete Wavelet Transform (DWT), which retain the required coefficients by adjusting the threshold. Currently, the tunable-Q wavelet transform (TQWT) shows relatively good compression performance. In 2023, Hardev from the Indian Centre for Information Technology 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 and condensing feature representations. Some studies use feature wave localization to extract peak features for compression. For example, in 2024, Lin Yujie et al. from Southeast University localized QRS waves while retaining relevant features to achieve compression. Although signal processing methods are highly interpretable and have low computational complexity, their compression performance remains 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 introduces additional computational complexity. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes an adaptive ECG signal compression method based on meta-learning. This method fully utilizes the shared neural network structure of various compression ratio models, significantly reducing model storage complexity while maintaining ECG compression and reconstruction accuracy.

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

[0007] Acquire the target electrocardiogram (ECG) signal, preprocess the ECG signal, and obtain the preprocessed ECG signal;

[0008] The preprocessed signal is input into a nested bottleneck autoencoder, and the compression ratio is selected through dynamic routing.

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

[0010] When the prediction entropy exceeds the set threshold, switch to a guaranteed compression mode that combines differential pulse code modulation and bzip2.

[0011] The selected model is used to compress and encode the electrocardiogram signal.

[0012] Optionally, 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 360Hz; and performing range normalization on the segmented signal segments to obtain the preprocessed electrocardiogram signal.

[0013] Optionally, the construction process of the nested bottleneck autoencoder includes: a fully shared encoder extracting primary features; four branches achieving different compression ratios through a partially shared encoder layer and a branch-specific encoder layer, respectively; and a decoder containing inverse operation blocks 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 according to the compression ratio from low to high; and using double 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 the three quantile lines; calculating the correlation coefficients between the four sub-segments; and obtaining the maximum amplitude value of each sub-segment.

[0016] Optionally, the application process of the meta-learner includes: establishing a mapping relationship between meta-features and the optimal model using an XGBoost classifier; calculating the prediction entropy value to evaluate the reliability of the selection; and activating the 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 performing lossless compression using the bzip2 algorithm.

[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 Effects of this Invention: This invention discloses an adaptive ECG signal compression method based on meta-learning. It adaptively selects 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 design of NBAE reduces the consumption of storage and computational resources for candidate models. The uncertainty estimation of the meta-learner avoids unnecessary decompression operations during the compression stage, and the time occupied by the model selection strategy using meta-learning is negligible. This invention proposes an efficient module combining lightweight channel attention and multi-scale deep convolution, which can capture richer features and highlight their importance, thereby improving the performance of the neural network. Simultaneously, the safest compression model is introduced into the meta-learning-based compression strategy. On the one hand, it serves as one of the predictable models for the meta-learner; on the other hand, the uncertainty estimation of the meta-learner is performed first to pre-judge the reliability of the model selection. If the prediction entropy is higher than a set threshold, this model can ensure high-quality signal compression and reconstruction, thereby improving the robustness and reliability of the system. Attached Figure Description

[0022] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0023] Figure 1 This is a flowchart illustrating an adaptive electrocardiogram signal compression method based on meta-learning according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the model structure of the nested bottleneck autoencoder in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the phased training procedure of NBAE as shown in Algorithm 1 of this embodiment of the invention. Detailed Implementation

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

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

[0028] With the advancements in artificial intelligence and deep learning, an increasing number of studies are focusing on compression of electrocardiogram (ECG) signals using neural network-based autoencoders. In 2024, Yang Minxi et al. from Xi'an University of Electronic Science and Technology used a superimposed semantic encoder to compress heartbeats and transmit them to the cloud. They then simultaneously performed classification through superimposed semantic decoding, enabling deployment on IoT devices and saving communication resources. Also in 2024, He Peng et al. from Chongqing University of Posts and Telecommunications used a residual convolutional autoencoder to simultaneously compress ECG and photoplethysmography (PPG) signals, and then combined it with a generative adversarial network for encryption tasks, proposing the HCEN model, which achieved good performance in both tasks. Some works also combine deep neural networks with lossless coding. For example, Yangyang Chang from the University of Minnesota used a network with a relatively low compression ratio and then used Huffman coding to compensate for some of the compression ratio, achieving lightweight compression of 12 leads. Although neural network compression methods can automatically learn key features and achieve better performance than mathematical methods, the above works all use models with a fixed compression ratio, which cannot guarantee quality for specific signals. The reasons are as follows: ECG signals have varying low-rank levels; irregular signals require more information to store, and neural networks may fail to generalize to unfamiliar samples. Compression quality varies depending on the signal; otherwise, useful clinical information may be lost, leading to misdiagnosis. In response, in 2024, Tahir Bekiryaz1c1 et al. from Bursa Technical University in Turkey proposed an ECG compression method with constant error and adjustable compression ratio using a convolutional autoencoder. This method uses a shared encoder and independent decoders to achieve different compression ratios, then uses a heuristic search method to compare with a set threshold to ensure an upper limit on the compression error. This method relies solely on neural networks, and is not optimal when encountering signals that all networks cannot reconstruct well. Furthermore, using multiple models introduces additional storage space and computational complexity, especially when the number of models is large. There is still room for improvement in how to efficiently deploy and select appropriate models.

[0029] Existing ECG compression technologies can be broadly categorized into two main types: signal processing and deep learning. On the one hand, signal processing methods rely on prior signal decomposition, such as wavelet transform, which cannot effectively represent the distorted ECG waveforms commonly found in dynamic electrocardiograms. While ensuring signal reconstruction accuracy, their compression ratios are generally low, and their compression performance heavily depends on the manual selection of empirical parameters, making it difficult to automate large-scale ECG data processing. On the other hand, most deep learning methods use fixed compression ratio models, which cannot flexibly adapt to varying ECG waveforms, making it difficult to balance signal reconstruction accuracy and compression ratio. The few deep learning methods that use multi-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 electrocardiogram signal compression method based on meta-learning, including:

[0031] The dataset is preprocessed, including removing baseline drift noise, splitting the data by a sliding window, and normalizing the range of the split segments to eliminate differences in the model input.

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

[0033] In the online application phase, the model selection and uncertainty estimation results are used through a meta-learner to achieve 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 electrocardiogram records sampled at 360 Hz and stored in 2^12 format (two 12-bit sample points in every two bytes).

[0037] Step 1.2: Data Preprocessing:

[0038] Baseline drift was removed using a 4th-order Butterworth bandpass filter (0.5-150Hz). The signal was divided into segments containing multiple heartbeats by 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 the compression model:

[0040] Step 2.1: Design of a nested bottleneck autoencoder:

[0041] NBAE structure such as Figure 2 As shown, there are four information bottlenecks, each operating independently. By partially sharing parameters between the encoder and decoder, the high compression ratio model parameters can be utilized as much as possible. Bottleneck sections with a small number of parameters supporting different compression ratios can be introduced, thus enabling models with multiple compression ratios without significantly increasing storage space. During model inference, NBAE is a dynamic hard routing method, meaning only one path is activated and executed. The encoder E is fully shared. fulExtracting local features from multiple channels, expanding the channels, and reducing the spatial dimension can yield a preliminary potential representation of the ECG signal. (Partially shared encoder layer) Further reducing the spatial dimension, where j∈[1,2,3,4]. The intermediate features of NBAE preserve the uncompressed channels, avoiding premature information loss and ensuring the integrity of shared parts. Specific encoder layers in the bottleneck branch. Further reducing the number of channels to 1 allows for different compression ratios. The decoder contains branch-specific blocks. Partial shared blocks and fully shared block D ful This is the inverse operation 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 multi-scale convolution block (MSCB). Mixed local channel attention 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 can be reparameterized structurally.

[0042] Step 2.2: Training the nested bottleneck autoencoder

[0043] The NBAE model has multiple shared encoder-decoder layers. Training a model with only one compression factor without considering other compression factors will result in a loss of overall performance. However, training multiple compression factor models together makes it difficult to adjust the weights of their respective loss functions. Essentially, specific compression factors are applied to different specific tasks. Therefore, a heuristic training strategy was 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 phase, the branch with the highest compression factor is trained until convergence 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) specific encoder and specific decoder (phase 1); (2) adjacent shared encoder and shared decoder (phase 2), with backward compatibility ensured through selective parameter updates. In the initial training phase and phase 1, the mean squared error loss L1 is used as the objective function, and its calculation formula is as follows:

[0044]

[0045] Among them, s o The input signal representing the model, s rThe reconstructed signal represents the model, and n is the length of the signal. In Phase 2, to maintain the performance of the branch with the highest compression ratio 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 a bi-objective loss L2, calculated as follows:

[0046]

[0047] in, This represents the reconstruction signal of the j-th branch of the currently trained model. This represents the reconstructed signal from the fourth model branch used during initial training. The Adam optimizer was used during training, trained via backpropagation, and hyperparameters such as the adjustment factor, learning rate, and number of training epochs were determined using labeled samples from the validation set. The optimal model was then selected for testing. Since the mean squared error (MSE) is relatively small and difficult to observe, the percentage root mean squared error (PRD) was used as one of the evaluation metrics 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 may fail to generalize to unfamiliar data, a rapidly deployable compression model is used when other models fail to reconstruct the data effectively. This method is an improvement on DE encoding: the signal is first differentially divided to increase redundancy, then quantized to 8 bits, and finally compressed losslessly using the efficient bzip2 encoding. Although the DE-bzip2 hybrid encoding has a relatively low compression ratio, it ensures reliable compression.

[0051] Step 3: Training and preparation of meta-learners:

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

[0053] The compression effect of the model is related to the morphological characteristics of the waveform. For a long ECG signal, the main influencing factors include the regularity and variability of heartbeats, and the noise level of the signal. A normal heartbeat lasts 0.75 to 0.8 seconds, so the ECG segment divided according to step 1 contains approximately 3 to 4 heartbeats. 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 level. Zero-crossing points reflect the presence or absence of lead drops, high-frequency interference, or heartbeat distortion. The correlation coefficient and range of the four-segment segments reflect the different types of heartbeats within a signal. If all heartbeats are morphologically similar, less information is needed to store compared to heartbeats with diverse categories. The Pearson correlation coefficient is calculated pairwise between each of the four-segment segments and the other three four-segment segments, and the average value is used as a correlation coefficient index for that segment. The maximum allowed PRD (Percentage Reduction) can be used as a compression condition or as a feature, without requiring one-hot encoding.

[0054] Table 1

[0055]

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

[0057] For a given ECG signal, the task of the meta-learner is to map the extracted meta-features to the most suitable compressed model labels, which is essentially a classification problem. Extreme gradient boosting (XGBoost) is chosen as the meta-learner because of its interpretability and low computational cost. XGBoost effectively achieves gradient boosting through regularized learning and caching of the perceptual tree structure, and improves accuracy through 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 have five models with different compression ratios. These models are used to compress and reconstruct all signal segments in the training set, and the models that meet the PRD threshold condition set in the meta-features are selected as labels. Then, we use these labels and their corresponding meta-features to train a meta-learner, using cross-entropy as the loss function for XGBoost.

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

[0060] In the application phase, for the preprocessed data that needs compression, meta-feature extraction is first performed according to Table 1, followed by prediction by a meta-learner and uncertainty estimation. This is because the meta-learner cannot always predict the optimal result. A common practice is to use the safest method as a safeguard when the reconstruction error of the prediction model does not meet requirements, but this requires decompression during compression, increasing runtime. To address this issue, the need for a safeguard method is determined by measuring the prediction uncertainty of the meta-learner. Higher prediction entropy indicates higher uncertainty, with predictions distributed across more classes. Furthermore, even if the prediction is incorrect, the reconstruction error is sometimes acceptable, thus uncertainty tolerance needs to be considered. Therefore, it is stipulated that if the total prediction entropy of a meta-learner under multiple PRD conditions exceeds a set boundary value, the meta-learner's prediction of this signal is unreliable, and the compression method in step 2.3 should be switched to. The prediction entropy H of the i-th segment... i The calculation formula is as follows:

[0061]

[0062] Where f is the prediction probability of the meta-learner. i θ represents the meta-feature of this segment, and θ is the set PRD condition. This is the model number predicted by the meta-learner. Based on the preparation in step 2, there are a total of M = 5 numbers.

[0063] After the meta-learner makes predictions and performs uncertainty estimation, 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 set the PRD conditions to eight levels, which not only improves the tolerance for uncertainty but also allows for more comprehensive training of the meta-learner, enabling it to select the best candidate model for most scenarios and maximize the combined strength of the candidate models.

[0064] In step 2.1, the hybrid channel attention module of the EMCB module can be replaced with a convolutional block attention module or a more advanced channel attention module. The reparameterizable multi-scale depthwise convolution can use only a single multi-scale convolution. The safest compression model in step 2.3 can be replaced with a lossless compression model with a higher compression ratio, such as LZMA.

[0065] The advantages of this invention are as follows: Existing technologies cannot effectively balance compression ratio and reconstruction error, and cannot quickly and adaptively adjust empirical parameters for signals with different levels of regularity; the method of using multiple models for hybrid compression increases model selection time and introduces additional storage space and computational complexity. This invention proposes an adaptive ECG signal compression method based on meta-learning, which can adaptively select an appropriate compression ratio 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 computational resources for candidate models, and the uncertainty estimation of the meta-learner can avoid unnecessary decompression operations during the compression stage, and the time occupied by the model selection strategy using meta-learning is negligible. Existing methods that only use deep learning models may experience generalization failure when faced with unfamiliar data, leading to a decrease in the quality of the reconstructed signal. This invention proposes an efficient module that combines lightweight channel attention and multi-scale deep convolution, which can capture richer features and highlight their importance, thereby improving the performance of the neural network. Meanwhile, the most reliable compression model is introduced into the meta-learning-based compression strategy. On the one hand, it serves as one of the predictable models for the meta-learner; on the other hand, the uncertainty of the meta-learner is estimated first to determine the reliability of the model selection in advance. If the prediction entropy is higher than a set critical value, 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 this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for adaptive electrocardiogram signal compression based on meta-learning, characterized in that, The application relates to a method for compressing and encoding electrocardio signals. The method comprises the following steps: acquiring a target electrocardio signal, preprocessing the electrocardio signal to obtain a preprocessed electrocardio signal; inputting the preprocessed signal into a nested bottleneck self-encoder, and selecting a compression ratio through dynamic routing; extracting meta features of the signal segment, and selecting a compression model according to the meta features and a preset reconstruction error threshold through a meta learner; when the predicted entropy exceeds a set threshold, switching to a differential pulse code modulation and bzip2 mixed guarantee compression mode; using the selected model to compress and encode the electrocardio signal; the construction process of the nested bottleneck self-encoder comprises the following steps: a fully shared encoder extracts primary features; four branches respectively realize different compression ratios through a partially shared encoder layer and a branch-specific encoder layer; and a decoder comprises inverse operation blocks corresponding to the encoder; the training process of the branch-specific encoder layer comprises the following steps: initially training a highest compression ratio branch to convergence; in order from low to high compression ratio, unfreezing and training branch-specific parameters in stages; and in the incremental training stage, adopting double reconstruction loss constraints; the meta feature extraction process comprises the following steps: calculating signal mean and standard deviation; detecting the number of intersection points of the signal and three quantile lines; calculating the correlation coefficient between four equal segments; and obtaining the maximum amplitude value of each subsegment; the application process of the meta learner comprises the following steps: using an XGBoost classifier to establish a mapping relationship between meta features and optimal models; calculating a predicted entropy value to evaluate the reliability of selection; and when the entropy value exceeds 0.8, activating a guarantee compression mode; the implementation process of the guarantee compression mode comprises the following steps: performing second-order differential processing on the signal; quantizing the differential result to 8-bit data; and adopting a bzip2 algorithm for lossless compression.

2. The meta-learning based adaptive electrocardiosignal compression method of claim 1, wherein, The preprocessing process comprises the following steps: adopting a fourth-order Butterworth band-pass filter to eliminate baseline drift; using a 1024-sample-point window to perform sliding segmentation on the signal sampled at 360 Hz; and performing range normalization processing on the segmented signal segment to obtain the preprocessed electrocardio signal.

3. The meta-learning based adaptive electrocardiosignal compression method of claim 1, wherein, The branch-specific encoder layer comprises a hybrid local channel attention module and a multi-scale convolution module; The hybrid local channel attention module is used for feature weight distribution; The multi-scale convolution module is used for supporting structure reparameterization.

Images