Arrhythmia detection algorithm based on deep active learning and complex feature fusion

Abstract

The invention discloses an arrhythmia detection algorithm based on deep active learning and complex feature fusion, and the algorithm comprises the steps: building a new signal evaluation index through the weighted fusion of the statistical characteristics and morphological indexes of electrocardiosignals, and the complexity and dynamic indexes; precise screening of high-value samples is completed by means of a signal screening strategy; and a heart rhythm detection model fusing a convolutional neural network, a bidirectional long-short-term memory network, a channel attention mechanism and a Transform is constructed and is used for identifying 15 heart rhythm types. According to the algorithm, the marking cost can be remarkably reduced, the model detection performance is optimized, a stable technical scheme is provided for intelligent diagnosis of arrhythmia, the clinical application prospect is wide, and the market value is prominent; according to the method, efficient and accurate detection of arrhythmia can be realized, the industrial pain point of label data scarcity is effectively solved, the diagnosis efficiency is improved while the data labeling cost is remarkably reduced, and the method has relatively strong clinical application prospects and engineering popularization values.

Description

Technical Field

[0001] This invention belongs to the field of cardiac arrhythmia detection technology, and relates to a cardiac arrhythmia detection algorithm based on deep active learning and complex feature fusion. Background Technology

[0002] Cardiovascular disease (CVD) is one of the major diseases threatening human health worldwide. Electrocardiogram (ECG) signals, as the most commonly used preliminary examination method in clinical practice, consist of key waveforms such as the P wave, QRS complex, and T wave. Recording the changes in cardiac electrical activity over time using electrodes on the body surface, it can directly reflect the heart's rhythm status, and its morphological properties are core indicators for doctors and cardiologists in diagnosing heart diseases. However, the current model of relying entirely on manual diagnosis by doctors specializing in ECG is not only time-consuming and labor-intensive, but also requires extremely high levels of professional expertise. Furthermore, individual differences in doctors' clinical experience make it highly susceptible to missed diagnoses or misdiagnoses.

[0003] While computer-aided diagnostic systems offer a technological path to overcome this predicament, deep learning (DL), represented by convolutional neural networks, has been widely applied to various intelligent ECG signal recognition tasks due to its powerful feature extraction capabilities. However, existing computer-aided diagnostic models heavily rely on large amounts of labeled data. Although clinical practice has accumulated a vast amount of ECG data, much of it is unlabeled. Furthermore, labeling ECG data requires stringent expertise from domain specialists and involves significant costs, directly resulting in labeled ECG data becoming a scarce resource.

[0004] Active learning (AL) can effectively reduce data annotation costs to a certain extent. Its core logic is to actively select the most valuable data samples for annotation, maximizing model performance improvement while minimizing annotation costs. Therefore, to reduce ECG data annotation costs and improve the performance of computer-aided diagnostic models, it is crucial to combine the advantages of active learning to construct a detection scheme adapted to the characteristics of ECG signals, thus overcoming current technological bottlenecks. Existing solutions still fall short in terms of the comprehensiveness of feature fusion and the specificity of sample query strategies, making it difficult to fully explore the value of ECG signals and efficiently utilize limited annotation resources, thus limiting the accuracy and generalization of models in arrhythmia detection.

[0005] Chinese invention patent CN109620211A discloses an intelligent method for assisted diagnosis of abnormal electrocardiograms based on deep learning. This method includes: preprocessing electrocardiogram (ECG) signals to obtain binary images of each cardiac cycle in the ECG signal; extracting deep implicit features from the binary images using a deep convolutional neural network; and classifying the heartbeats in the ECG based on the features using a classifier.

[0006] However, the above-mentioned technical solutions are slow and have low accuracy when processing and analyzing electrocardiogram signals, which may affect the convenience for staff to use.

[0007] Therefore, this invention provides a cardiac arrhythmia detection algorithm based on deep active learning and complex feature fusion. This algorithm constructs a new signal evaluation index by weightedly fusing the statistical characteristics of electrocardiogram (ECG) signals with morphological indicators, and complexity with dynamic indicators. Based on this index, five signal query strategies (SCMCD-13, SCMCD-F, SCMCD-B, SCMCD-M, and SCMCD-FB) are designed to achieve accurate screening of high-value samples. Simultaneously, a cardiac rhythm detection model fused from a convolutional neural network, a bidirectional long short-term memory network, a channel attention mechanism, and a Transformer is designed for the detection of 15 cardiac rhythm types. This algorithm effectively reduces annotation costs and improves model detection performance, providing a reliable technical solution for intelligent diagnosis of cardiac arrhythmias, and has broad clinical application prospects and market value. Summary of the Invention

[0008] To address the problems of excessive reliance on large amounts of labeled data, scarcity of labeled samples, and high data labeling costs in existing electrocardiogram arrhythmia detection technologies, this invention proposes an arrhythmia detection algorithm based on deep active learning and complex feature fusion. This algorithm effectively solves the problems of excessive reliance on large amounts of labeled data, scarcity of labeled samples, and high data labeling costs in existing electrocardiogram arrhythmia detection technologies.

[0009] To achieve the above objectives, the technical solution adopted by this invention is as follows: an arrhythmia detection algorithm based on deep active learning and complex feature fusion, comprising the following steps:

[0010] S1. Select the MIT-BIH arrhythmia database as the data source and extract the raw electrocardiogram signal according to a fixed time window.

[0011] S2. Denoise the original ECG signal and use the dual median filtering method. Set two median filter windows and filter the denoised signal in turn. Calculate the baseline and subtract the baseline from the original signal to obtain the corrected signal.

[0012] S3. Normalize the corrected signal to the [0,1] interval using the Min-Max normalization function to obtain the standardized signal segment, and store the standardized signal segment to construct an unlabeled arrhythmia sample set;

[0013] S4. Input the unlabeled sample set into the hybrid classification model to extract sample features. Then, extract the skewness and shape factor indices of the statistical characteristics and morphology of the electrocardiogram signal, the fractal dimension and sample entropy indices of complexity and dynamics. Calculate the fusion value of the above four indices with equal weights to form a new signal evaluation index. Then, based on the new signal evaluation index, select samples with high information gain and high labeling value to be labeled, and hand them over to electrocardiogram experts to complete the heart rhythm category labeling.

[0014] S5. The labeled samples are added to the labeled arrhythmia dataset, and the labeled samples are removed from the unlabeled arrhythmia dataset. Then, the deep learning model is trained and optimized using the labeled arrhythmia dataset as input, and the model parameters are updated.

[0015] S6. Iterate through steps S4 and S5, continuously filtering, labeling, and training the model until the performance of the hybrid classification model reaches the preset standard or there are no remaining samples in the unlabeled dataset. Stop the iteration and determine the final model.

[0016] Furthermore, the fixed time window includes at least one of 2 seconds, 5 seconds, 8 seconds, or 10 seconds.

[0017] Furthermore, the signal denoising adopts the wavelet transform denoising method, selects the 'db6' wavelet basis and 6-level decomposition, performs wavelet decomposition on the truncated electrocardiogram segment, and reconstructs the signal after processing the detail coefficients through a soft threshold of 0.008 to remove high-frequency noise.

[0018] Furthermore, when the labeled sample set is input into the hybrid classification model, the input is adjusted to the input_shape[0],1 format through the Reshape layer, and the channel dimension is added to adapt to subsequent convolution operations.

[0019] Furthermore, the hybrid classification model includes convolutional block design, bidirectional LSTM block design, Transformer block and subsequent processing, as well as fully connected and output layers.

[0020] Furthermore, the convolutional block design includes eight concatenated convolutional sub-blocks, each consisting of a Conv1D layer, a ChannelAttention layer with a compression ratio of 16, a BatchNormalization layer, an AveragePooling1D layer with a pooling kernel size of 2, and a Dropout layer with a ratio of 0.1. The convolutional kernel sizes of the eight sub-blocks are 9, 7, 5, 5, 5, 3, 3, and 3, respectively, and the number of output channels are 32, 64, 128, 512, 256, 128, 64, and 32, respectively. The activation function is ReLU, and L2 regularization with a coefficient of 0.001 is added to all of them.

[0021] Furthermore, the bidirectional LSTM block design includes 5 bidirectional LSTM sub-blocks, the output dimensions of the 5 bidirectional LSTM sub-blocks are 64, 128, 256, 64 and 32 respectively, the activation function is tanh, and each bidirectional LSTM sub-block is connected to a ChannelAttention layer with a compression ratio of 16, a BatchNormalization layer and a Dropout layer with a dropout rate of 0.1.

[0022] Furthermore, the Transformer block and subsequent processing include a Transformer block with an embedding dimension of 32, an attention head of 4, and a feedforward network dimension of 128. The Transformer block is used to construct a TransformerBlock and connect it to a ChannelAttention layer with a compression ratio of 16, a BatchNormalization layer, and a Dropout layer with a dropout rate of 0.1. Global pooling of features is completed through the GlobalAveragePooling1D layer.

[0023] Furthermore, the fully connected and output layer includes two Dense layers arranged sequentially, with output dimensions of 80 and 50 respectively, and ReLU activation function. Both Dense layers are subjected to L2 regularization with a coefficient of 0.001. The Dense layers are connected to a BatchNormalization layer and a Dropout layer with a dropout rate of 0.2. Finally, the classification result is output through a Dense layer with an output dimension of the number of heart rhythm categories (num_classes) and an activation function of softmax.

[0024] Furthermore, the new signal evaluation index includes five exclusive signal query strategies: SCMCD-13, SCMCD-F, SCMCD-B, SCMCD-M, and SCMCD-FB.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. This invention solves the problem of insufficient representation by a single feature by fusing multi-dimensional electrocardiogram features and combining statistical characteristics with morphological, complexity and dynamic information, thus providing a reliable feature foundation for subsequent accurate classification;

[0027] 2. A signal query strategy was designed to screen high-value samples, significantly reducing the amount of labeled data and breaking through the technical bottleneck of scarce label data.

[0028] 3. The hybrid classification model combines the advantages of CNN, BiLSTM, ChannelAttention, and Transformer, and suppresses overfitting through strategies such as regularization, BatchNormalization, and Dropout. It achieves a detection accuracy of 99% for 15 heart rhythm types, and its generalization performance is significantly better than that of traditional models.

[0029] 4. The model is built on the TensorFlow framework, with a clear structure and strong portability. It can be quickly deployed to clinical electrocardiogram diagnostic equipment, providing an efficient and reliable technical solution for intelligent diagnosis of arrhythmias, and has significant clinical application value and promotion prospects. Attached Figure Description

[0030] Figure 1 This is a flowchart of the experimental process of the present invention;

[0031] Figure 2 This is a structural diagram of the model of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] See attached document Figure 1 , 2 The specific steps of the present invention will be described in further detail below.

[0034] Perform data preprocessing on electrocardiogram signals.

[0035] The MIT-BIH arrhythmia database was selected as the data source, and electrocardiogram signals were extracted using fixed time windows of 2 seconds, 5 seconds, 8 seconds, and 10 seconds.

[0036] Denoising with wavelet transform: Select the “db6” wavelet basis for 6-level decomposition, process the detail coefficients (including high-frequency noise) with a soft threshold of 0.008, and combine with the approximation coefficient reconstruction to remove high-frequency interference such as electromyography and power frequency.

[0037] Dual median filtering corrects baseline drift: First, a 0.2-second window is used to capture short-term baseline fluctuations, and then a 0.6-second window is used to smooth them. The difference between the resulting baseline estimation curve and the denoised signal is calculated to eliminate baseline drift.

[0038] After Min-Max normalization, the corrected signal is mapped to the [0,1] interval, stored and labeled according to "heart rhythm type-window duration-record number", and a complete and standardized arrhythmia dataset is constructed.

[0039] A hybrid classification model was built, consisting of a convolutional neural network, a bidirectional long short-term memory network, a channel attention mechanism, and a Transformer.

[0040] After receiving the electrocardiogram signal, it is converted into (input_shape[0],1) format by the Reshape layer to supplement the channel dimension to match the 1D convolution requirement;

[0041] Eight convolutional sub-blocks are set up in series. Each sub-block uses Conv1D (kernel 9→7→5→5→5→3→3→3, channels 32→64→128→512→256→128→64→32, ReLU+L2 regularization) to extract local features. ChannelAttention (compression ratio 16) is used to enhance key channels, BatchNormalization to accelerate convergence, AveragePooling1D (kernel 2) to reduce dimensionality, and Dropout (0.1) to prevent overfitting.

[0042] Five bidirectional LSTM sub-blocks (output dimensions 64→128→256→64→32, tanh activation) are set up to capture the timing dependency of the signal, followed by ChannelAttention, BatchNormalization and Dropout of the same specification.

[0043] The Transformer block (32 embedding dimensions, 4 heads, 128 feedforward dimensions) is set up to model long-distance feature associations. After channel attention, normalization, and dropout, the features are compressed by GlobalAveragePooling1D.

[0044] Two Dense layers (80→50 dimensions, ReLU+L2 regularization) are used to further process the features. Each layer is normalized and Dropout (0.2). Finally, the heart rhythm category is output through the Dense layer activated by softmax.

[0045] Construct a multi-dimensional ECG feature fusion system and signal query strategy.

[0046] The statistical characteristics and morphological indicators (skewness, shape factor), complexity and dynamic indicators (fractal dimension, sample entropy) of electrocardiogram (ECG) signals are extracted. The four types of indicators are weighted equally (0.25 each) to calculate the fusion value, thereby constructing a new evaluation index that can comprehensively reflect the value of ECG signals.

[0047] Based on the above-mentioned integrated evaluation indicators, five query strategies—SCMCD-13, SCMCD-F, SCMCD-B, SCMCD-M, and SCMCD-FB—were innovatively designed. Through differentiated sample sorting and selection logic, high-value data (such as samples containing key cardiac rhythm features and high information gain) in unlabeled ECG samples were accurately identified.

[0048] SCMCD-13: This strategy first sorts the unlabeled datasets in ascending order according to the fusion evaluation metric, then divides all unlabeled data into 13 equal parts, and selects the top n / / 13 samples from each part as the samples to be labeled.

[0049] SCMCD-F: This strategy first sorts the data in the unlabeled dataset according to the fusion evaluation metric from smallest to largest, and then takes the top results starting from the smallest. One sample is selected as the sample to be labeled.

[0050] SCMCD-B: This strategy first sorts the data in the unlabeled dataset according to the fusion evaluation metric from smallest to largest, and then takes the last data from the largest. One sample is selected as the sample to be labeled.

[0051] SCMCD-M: This strategy first sorts the data in the unlabeled dataset according to the fusion evaluation metric from smallest to largest, and then takes the middle value starting from the middle. One sample is selected as the sample to be labeled.

[0052] SCMCD-FB: This strategy first sorts the data in the unlabeled dataset from smallest to largest according to the fusion evaluation metric, and then takes data from both the minimum and maximum values. One sample is selected as the sample to be labeled.

[0053] The entire algorithm flow of this invention is as follows:

[0054] S1. Select the MIT-BIH arrhythmia database as the data source and extract the raw electrocardiogram signal according to a fixed time window.

[0055] S2. Denoise the original ECG signal and use the dual median filtering method. Set two median filter windows and filter the denoised signal in turn. Calculate the baseline and subtract the baseline from the original signal to obtain the corrected signal.

[0056] S3. Normalize the corrected signal to the [0,1] interval using the Min-Max normalization function to obtain the standardized signal segment, and store the standardized signal segment to construct an unlabeled arrhythmia sample set;

[0057] S4. Input the unlabeled sample set into the hybrid classification model to extract sample features. Then, extract the statistical characteristics and morphological indicators (skewness, shape factor), complexity and dynamic indicators (fractal dimension, sample entropy) of the electrocardiogram signal. Calculate the fusion value of the above four types of indicators with equal weights to form a new signal evaluation index. Then, based on the new signal evaluation index, select samples with high information gain and high labeling value to be labeled, and hand them over to electrocardiogram experts to complete the heart rhythm category labeling.

[0058] S5. The labeled samples are added to the labeled arrhythmia dataset, and the labeled samples are removed from the unlabeled arrhythmia dataset. Then, the deep learning model is trained and optimized using the labeled arrhythmia dataset as input, and the model parameters are updated.

[0059] S6. Iterate through steps S4 and S5, continuously filtering, labeling, and training the model until the performance of the hybrid classification model reaches the preset standard or there are no remaining samples in the unlabeled dataset. Stop the iteration and determine the final model.

[0060] This invention presents a cardiac arrhythmia detection algorithm based on deep active learning and complex feature fusion. It creates new signal evaluation metrics by weighted fusion of statistical and morphological indicators, complexity and dynamic indicators of electrocardiogram (ECG) signals; it accurately selects high-value samples using a signal screening strategy; and it constructs a cardiac rhythm detection model integrating convolutional neural networks, bidirectional long short-term memory networks, channel attention mechanisms, and Transformers for the identification of 15 cardiac rhythm types. This algorithm significantly reduces annotation costs and optimizes model detection performance, providing a stable technical solution for intelligent diagnosis of cardiac arrhythmias, with broad clinical application prospects and significant market value.

[0061] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An arrhythmia detection algorithm based on deep active learning and complex feature fusion, characterized in that: Includes the following steps: S1. Select the MIT-BIH arrhythmia database as the data source and extract the raw electrocardiogram signal according to a fixed time window. S2. Denoise the original ECG signal and use the dual median filtering method. Set two median filter windows and filter the denoised signal in turn. Calculate the baseline and subtract the baseline from the original signal to obtain the corrected signal. S3. Normalize the corrected signal to the [0,1] interval using the Min-Max normalization function to obtain the standardized signal segment, and store the standardized signal segment to construct an unlabeled arrhythmia sample set; S4. Input the unlabeled sample set into the hybrid classification model to extract sample features. Then, extract the skewness and shape factor indices of the statistical characteristics and morphology of the electrocardiogram signal, the fractal dimension and sample entropy indices of complexity and dynamics. Calculate the fusion value of the above four indices with equal weights to form a new signal evaluation index. Then, based on the new signal evaluation index, select samples with high information gain and high labeling value to be labeled, and hand them over to electrocardiogram experts to complete the heart rhythm category labeling. S5. The labeled samples are added to the labeled arrhythmia dataset, and the labeled samples are removed from the unlabeled arrhythmia dataset. Then, the deep learning model is trained and optimized using the labeled arrhythmia dataset as input, and the model parameters are updated. S6. Iterate through steps S4 and S5, continuously filtering, labeling, and training the model until the performance of the hybrid classification model reaches the preset standard or there are no remaining samples in the unlabeled dataset. Stop the iteration and determine the final model.

2. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 1, characterized in that: The fixed time window includes at least one of 2 seconds, 5 seconds, 8 seconds, or 10 seconds.

3. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 1, characterized in that: The signal denoising method employs wavelet transform, selecting the 'db6' wavelet basis and level 6 decomposition. Wavelet decomposition is performed on the truncated ECG segments, and the signal is reconstructed after processing the detail coefficients with a soft threshold of 0.008 to remove high-frequency noise.

4. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 1, characterized in that: When the labeled sample set is input into the hybrid classification model, the input is adjusted to the input_shape[0],1 format through the Reshape layer, and the channel dimension is added to adapt to the subsequent convolution operation.

5. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 1, characterized in that: The hybrid classification model includes convolutional block design, bidirectional LSTM block design, Transformer block and subsequent processing, as well as fully connected and output layers.

6. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 5, characterized in that: The convolutional block design includes eight concatenated convolutional sub-blocks, each consisting of a Conv1D layer, a ChannelAttention layer with a compression ratio of 16, a BatchNormalization layer, an AveragePooling1D layer with a pooling kernel size of 2, and a Dropout layer with a ratio of 0.

1. The kernel sizes of the eight convolutional sub-blocks are 9, 7, 5, 5, 5, 3, 3, and 3, respectively, and the number of output channels are 32, 64, 128, 512, 256, 128, 64, and 32, respectively. The activation function is ReLU, and all are subject to L2 regularization with a coefficient of 0.

001.

7. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 5, characterized in that: The bidirectional LSTM block design includes 5 bidirectional LSTM sub-blocks, with output dimensions of 64, 128, 256, 64, and 32 respectively, and activation function of tanh. Each bidirectional LSTM sub-block is connected to a ChannelAttention layer with a compression ratio of 16, a BatchNormalization layer, and a Dropout layer with a dropout rate of 0.

1.

8. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 5, characterized in that: The Transformer block and subsequent processing include a Transformer block with an embedding dimension of 32, an attention head of 4, and a feedforward network dimension of 128. The Transformer block is used to construct a TransformerBlock and connect it to a ChannelAttention layer with a compression ratio of 16, a BatchNormalization layer, and a Dropout layer with a dropout rate of 0.

1. Global pooling of features is completed through the GlobalAveragePooling1D layer.

9. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 5, characterized in that: The fully connected and output layers consist of two Dense layers arranged sequentially, with output dimensions of 80 and 50 respectively. The activation function is ReLU, and both Dense layers are subjected to L2 regularization with a coefficient of 0.

001. The Dense layers are connected to a BatchNormalization layer and a Dropout layer with a dropout rate of 0.

2. Finally, the classification results are output through a Dense layer with an output dimension of the number of heart rhythm categories (num_classes) and an activation function of softmax.

10. The arrhythmia detection algorithm based on deep active learning and complex feature fusion according to claim 1, characterized in that: The new signal evaluation index includes five exclusive signal query strategies: SCMCD-13, SCMCD-F, SCMCD-B, SCMCD-M, and SCMCD-FB.

Images