A portable dynamic high-frequency electrocardiogram monitoring method and system

Abstract

The application provides a portable dynamic high-frequency electrocardiogram monitoring system and method, aiming at the problems of traditional high-frequency electrocardiogram in myocardial ischemia diagnosis, such as dependence on exercise load, diagnosis delay and low detection rate, and provides an efficient solution. The electrocardiogram acquisition hardware of the system adopts high signal-to-noise ratio and high fidelity design to enhance the detection capability of high-frequency electrocardiogram signal, and transmits the data to the mobile terminal in real time through the wireless module to realize continuous dynamic monitoring. The software algorithm preprocesses the collected dynamic resting high-frequency electrocardiogram data, removes abnormal waveforms, and extracts high-frequency signals in units of 3 seconds. Then, these data are input into a pre-trained electrocardiogram classification network, which integrates modules such as multi-scale convolution, feature splicing, long short-term memory network and adaptive convolution, realizes automatic feature extraction and time series modeling, and can efficiently identify early myocardial ischemia. The system provides an efficient and real-time solution for early diagnosis and intervention of myocardial ischemia.

Description

Technical Field

[0001] This invention relates to the field of medical instrument technology, and in particular to a portable dynamic high-frequency electrocardiogram monitoring method and system. Background Technology

[0002] Cardiovascular disease is the leading threat to human health, and electrocardiography (ECG), as a non-invasive and rapid diagnostic tool, has become the most common method for detecting cardiovascular diseases. High-frequency ECG is a technique developed based on conventional ECG, primarily analyzing ECG signals in the 150–250 Hz range. Studies have shown that high-frequency ECG has a higher detection rate for coronary heart disease compared to conventional ECG. Conventional high-frequency ECG monitoring methods rely on extracting indicators such as RMS for analysis, and require the patient to exercise to reach at least 80% of their age-appropriate maximum heart rate for it to be diagnostically valuable. This monitoring method has a long acquisition time, approximately 6 minutes or more, and if the heart rate does not reach the threshold, there is excessive noise, or too many noisy leads during the acquisition process, the analysis method may fail, rendering the acquired data meaningless. Furthermore, many patients may not be able to reach their heart rate threshold through exercise, making this data acquisition method unsuitable for them. First, high-frequency ECG information during exercise is acquired. A high-frequency QRS waveform curve within a preset time period is selected. The first reference point with the minimum root-mean-square voltage on the high-frequency QRS reference waveform curve is chosen, and first and second screening conditions are set to obtain a second and third reference point. The relative value of amplitude decrease is determined based on the first and second reference points. If the relative value of amplitude decrease and the relative value of amplitude increase meet preset conditions, reference information is determined based on the high-frequency QRS waveform curve. This method utilizes exercise-load high-frequency ECG to diagnose myocardial ischemia and myocardial infarction.

[0003] Current technologies rely on high-frequency ECG acquisition during exercise, requiring patients to undergo ECG recordings while exercising to elicit changes in the ECG signal for diagnosis. This presents limitations for patients with exercise difficulties (such as the elderly or seriously ill patients). Furthermore, current technologies based on exercise-based ECG recordings suffer from analysis lag, as data acquisition and analysis occur only after the patient has begun exercising. This means patients need to undergo testing under specific conditions, and analysis is typically performed offline, without real-time feedback. Summary of the Invention

[0004] This invention provides a portable dynamic high-frequency electrocardiogram monitoring method, which solves the problem in the prior art that requires patients to undergo exercise training before electrocardiogram recording and diagnosis, thereby improving diagnostic efficiency and accuracy.

[0005] In a first aspect, the present invention provides a portable dynamic high-frequency electrocardiogram monitoring method, the method comprising:

[0006] Acquire resting electrocardiogram (ECG) data, and perform data preprocessing on the resting ECG data to obtain resting high-frequency ECG data;

[0007] The resting high-frequency ECG data is input into a pre-trained ECG signal classification network to obtain the monitoring results of resting high-frequency ECG for myocardial ischemia; wherein, the ECG signal classification network includes: a multi-scale convolutional neural module, a feature splicing module, a long short-term memory module, an adaptive convolutional module, and a classification module;

[0008] The multi-scale convolutional neural module is used to extract multi-scale features from resting high-frequency electrocardiogram data to obtain local detail features corresponding to each convolutional layer.

[0009] The feature splicing module is used to splice the local detail features corresponding to each convolutional layer to obtain activation fusion spliced ​​features;

[0010] The long short-term memory module is used to perform temporal modeling on the activation fusion splicing features to obtain temporal related information;

[0011] The adaptive convolution module is used to randomly generate dynamic convolution kernels to extract features from the time-related information and obtain the time-related detailed features corresponding to each convolutional layer.

[0012] The classification module is used to classify the temporal detail features corresponding to each convolutional layer to obtain the classification result.

[0013] In conjunction with the first aspect, in one possible implementation, the data preprocessing of the resting electrocardiogram data to obtain resting high-frequency electrocardiogram data includes:

[0014] The resting ECG data is split to obtain multiple sets of split resting ECG data;

[0015] After cleaning the multiple sets of split resting ECG data, the correlation coefficient between the data within each set of split resting ECG data is calculated, and it is determined whether the correlation coefficient is less than a first preset value.

[0016] If the correlation coefficient between data within a group of split resting ECG data is less than a first preset value, then the split resting ECG data is discarded from the multiple groups of split resting ECG data.

[0017] For the resting ECG data after splitting the data within the group with a correlation coefficient not less than the first preset value, a bandpass filter is used to extract high-frequency ECG signals from the split resting ECG data to obtain resting high-frequency ECG data.

[0018] In conjunction with the first aspect, in one possible implementation, the multi-scale convolutional neural module includes a first convolutional layer, a second convolutional layer, and a third convolutional layer arranged in parallel.

[0019] The multi-scale convolutional neural module is used to extract multi-scale features from resting high-frequency electrocardiogram data to obtain local detail features corresponding to each convolutional layer, including:

[0020] Based on the first convolutional layer, rapid fluctuation detail features of the QRS band in the resting high-frequency ECG data are extracted to obtain the first feature;

[0021] The second feature is obtained by extracting the overall morphological features of the QRS band in the resting high-frequency ECG data based on the second convolutional layer.

[0022] Based on the third convolutional layer, the heart rate transformation and QRS intensity features of the QRS band in the resting high-frequency ECG data are extracted to obtain the third feature;

[0023] The first feature, the second feature, and the third feature are used as local detail features corresponding to each convolutional layer.

[0024] In conjunction with the first aspect, in one possible implementation, the feature splicing module includes: a filter splicing layer, a batch normalization layer, and a ReLU activation function layer;

[0025] The feature splicing module is used to splice the local detail features corresponding to each branch to obtain activated fusion spliced ​​features, including:

[0026] The first feature, the second feature, and the third feature are spliced ​​together according to the filter splicing layer to obtain the initial activation fusion splicing feature;

[0027] The initial activation fusion splicing features are normalized according to the batch normalization layer to obtain normalized splicing features;

[0028] The normalized splicing features are subjected to a nonlinear operation based on the ReLU activation function layer to obtain the activated fusion splicing features.

[0029] In conjunction with the first aspect, in one possible implementation, the long short-term memory module includes a TimeDistributed layer and an LSTM layer connected in sequence;

[0030] The long short-term memory module is used to perform temporal modeling on the activation fusion and splicing features to obtain temporal-related information, including:

[0031] The activation fusion and splicing features are flattened at each time step according to the TimeDistributed layer to obtain a one-dimensional feature vector corresponding to each time step;

[0032] The LSTM layer is used to perform time series modeling on the one-dimensional feature vector corresponding to each time step, and the long-short time dependency relationship is captured to obtain time series related information.

[0033] In conjunction with the first aspect, in one possible implementation, the adaptive convolution module includes a fourth convolutional layer, a fifth convolutional layer, and a sixth convolutional layer arranged in parallel; wherein the number of convolutional kernels in the fourth, fifth, and sixth convolutional layers changes adaptively based on time-related information.

[0034] The adaptive convolution module is used to randomly generate dynamic convolution kernels to extract features from the temporal-related information, obtaining temporal detail features corresponding to each convolutional layer, including:

[0035] The first self-convolutional temporal-related information is obtained by extracting the small-scale high-frequency transient features of the time-related information from the fourth convolutional layer.

[0036] The second self-convolutional temporal related information is obtained by extracting the mesoscale mid-frequency morphological features of the time-related information from the fifth convolutional layer.

[0037] The third self-convolutional temporal related information is obtained by extracting the large-scale low-frequency rhythmic features of the time-related information from the sixth convolutional layer.

[0038] In conjunction with the first aspect, in one possible implementation, the classification module includes: a channel feature splicing layer, a batch normalized activation flattening layer, a regularized fully connected activation layer, and a fully connected discriminant layer;

[0039] The classification module is used to classify the temporal detail features corresponding to each convolutional layer to obtain classification results, including:

[0040] Based on the channel feature splicing layer, the temporal detail features corresponding to each convolutional layer are spliced ​​together to obtain the initial spliced ​​self-convolutional temporal related information.

[0041] Based on the batch normalization activation flattening layer, the initial spliced ​​self-convolutional temporal related information is subjected to batch normalization, activation and flattening operations to obtain one-dimensional self-convolutional temporal features.

[0042] Based on the regularized fully connected activation layer, the one-dimensional self-convolutional temporal features are regularized and compressed to obtain compressed self-convolutional temporal features.

[0043] The compressed self-convolutional temporal features are classified based on the fully connected discriminant layer to obtain the classification result.

[0044] In conjunction with the first aspect, in one possible implementation, the loss function used during the training of the ECG signal classification network is:

[0045]

[0046] Where N represents the total number of training samples; y i p represents the true label of the i-th training sample; i This represents the prediction result for the i-th training sample.

[0047] Secondly, the present invention provides a portable dynamic high-frequency electrocardiogram monitoring system, which includes: a front-end acquisition and processing module, a primary signal processing module, a wireless communication module, a result prediction module, and a power supply module;

[0048] The front-end acquisition and processing module is used to acquire multi-channel body surface potentials and obtain primary high-frequency electrocardiogram signals by differential amplification and AD conversion of the high-frequency electrocardiogram signals.

[0049] The primary signal processing module is used to perform primary processing on the acquired primary high-frequency ECG signals and obtain a noise-reduced standard high-frequency ECG signal through primary processing.

[0050] The wireless communication module is used to transmit the standard high-frequency electrocardiogram signal to the corresponding display module according to the wireless transmission function of the integrated Bluetooth module;

[0051] The result prediction module is used to predict the result based on the pre-trained ECG signal classification network and the pre-processed high-frequency ECG signal to obtain the monitoring result of resting high-frequency ECG for myocardial ischemia; wherein, the pre-trained ECG signal classification network is deployed on the dynamic high-frequency ECG monitoring system;

[0052] The power supply module is used to supply power to the front-end acquisition and processing module, the wireless communication module, and the primary signal processing module.

[0053] In conjunction with the second aspect, in one possible implementation, the front-end acquisition and processing module includes: an acquisition unit, a filtering unit, a buffer unit, a right leg drive noise reduction unit, a differential amplification unit, and an AD conversion unit;

[0054] The front-end acquisition and processing module is used to acquire multi-channel body surface potentials multiple times, and to obtain multiple sets of primary high-frequency electrocardiogram signals by differential amplification and AD conversion of the high-frequency electrocardiogram signals, including:

[0055] The acquisition unit is used to acquire multiple sets of multi-channel body surface potentials using an electrocardiogram sensor; each set of multi-channel body surface potentials includes the body surface potentials at electrode positions LA, LL, RA and V1 to V6.

[0056] The filtering unit is used to perform low-pass filtering on the multi-channel body surface potential to obtain the low-pass filtered multi-channel body surface potential.

[0057] The buffer unit is used to provide low output impedance to the multi-channel body surface potential after low-pass filtering, so as to obtain the buffered multi-channel body surface potential.

[0058] The right leg drive noise reduction unit is used to perform noise reduction processing on the buffered multi-channel body surface potential to obtain the noise-reduced multi-channel body surface potential.

[0059] The differential amplification unit is used to perform differential processing on the noise-reduced multi-channel body surface potential to obtain multiple sets of differential amplified signals.

[0060] The AD conversion unit is used to perform AD conversion on the multiple sets of differential amplified signals to obtain multiple sets of primary high-frequency electrocardiogram signals.

[0061] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention obtains resting electrocardiogram (ECG) data, performs data preprocessing on the resting ECG data, and obtains resting high-frequency ECG data; This invention uses resting high-frequency ECG data, eliminating the need for patients to undergo exercise load training, making it suitable for a wider range of people; The resting high-frequency ECG data is input into a pre-trained ECG signal classification network to obtain the monitoring results of resting high-frequency ECG for myocardial ischemia; wherein, the ECG signal classification network includes: a multi-scale convolutional neural module, a feature splicing module, a long short-term memory module, an adaptive convolutional module, and a classification module; The multi-scale convolutional neural module is used to perform multi-branch multi-scale feature extraction on the resting high-frequency ECG data to obtain the local detail features corresponding to each branch; The input resting high-frequency ECG data is processed by the pre-trained ECG signal classification network to obtain the monitoring results of resting high-frequency ECG for myocardial ischemia; Multi-scale feature extraction is performed on high-frequency ECG data, a process that captures local details in resting high-frequency ECG data, avoiding the limitations of traditional manual feature extraction. A Long Short-Term Memory (LSTM) module is used to perform temporal modeling of the local detail features corresponding to each convolution, obtaining temporal codes and hidden layer features. An LSTM module simultaneously focuses on the temporal relationships between the local detail features corresponding to each convolution, obtaining temporal-related information, which has a stronger ability to identify rapidly changing pathological signals in resting high-frequency ECG data. An adaptive convolution module is used to randomly generate dynamic convolution kernels to extract features from temporal-related information, obtaining temporal detail features corresponding to each convolutional layer. By highlighting key ECG segments through temporal detail features, the ability to capture pathological events such as myocardial ischemia and myocardial infarction is enhanced. Attached Figure Description

[0062] Figure 1 A flowchart illustrating the steps of a portable dynamic high-frequency electrocardiogram monitoring method provided in an embodiment of the present invention;

[0063] Figure 2 This is a schematic diagram of the electrocardiogram signal classification network structure provided in an embodiment of the present invention;

[0064] Figure 3 A schematic diagram of the electrode placement for bipolar limb leads and the limbs of the human body provided in an embodiment of the present invention;

[0065] Figure 4 This is a schematic diagram of a portable dynamic high-frequency electrocardiogram monitoring system provided in an embodiment of the present invention. Detailed Implementation

[0066] 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, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0067] In a first aspect, the present invention provides a portable dynamic high-frequency electrocardiogram monitoring method, see [link to previous article]. Figure 1 The method includes the following steps S101 to S102.

[0068] S101, acquire resting electrocardiogram data, perform data preprocessing on the resting electrocardiogram data, and obtain resting high-frequency electrocardiogram data;

[0069] Specifically, in step S101, resting electrocardiogram (ECG) data is acquired, and the resting ECG data is preprocessed to obtain resting high-frequency ECG data, including the following steps S1011 to S1012.

[0070] S1011, split the resting ECG data to obtain multiple sets of split resting ECG data;

[0071] S1012, after cleaning the multiple sets of split resting ECG data, calculate the correlation coefficient between the data within each set of split resting ECG data, and determine whether the correlation coefficient is less than the first preset value.

[0072] If the correlation coefficient between data within a group of split resting ECG data is less than the first preset value, then the split resting ECG data will be discarded from multiple groups of split resting ECG data.

[0073] For the resting ECG data after splitting the data within the group with a correlation coefficient not less than the first preset value, a bandpass filter is used to extract high-frequency ECG signals from the split resting ECG data to obtain resting high-frequency ECG data.

[0074] S102, input the resting high-frequency ECG data into the pre-trained ECG signal classification network to obtain the monitoring results of resting high-frequency ECG for myocardial ischemia; see [link / reference]. Figure 2 The ECG signal classification network includes: a multi-scale convolutional neural module, a feature splicing module, a long short-term memory module, an adaptive convolutional module, and a classification module.

[0075] The multi-scale convolutional neural module is used to extract multi-scale features from resting high-frequency ECG data to obtain the local detail features corresponding to each convolutional layer.

[0076] Specifically, the multi-scale convolutional neural module includes a first convolutional layer, a second convolutional layer, and a third convolutional layer arranged in parallel;

[0077] A multi-scale convolutional neural module is used to extract multi-scale features from resting high-frequency ECG data, obtaining local detail features corresponding to each convolutional layer, including:

[0078] Based on the first convolutional layer, rapid fluctuation detail features of the QRS band in the resting high-frequency ECG data are extracted to obtain the first feature;

[0079] The second feature is obtained by extracting the overall morphological features of the QRS band in the resting high-frequency ECG data based on the second convolutional layer.

[0080] Based on the third convolutional layer, the heart rate transformation and QRS intensity features of the QRS band in the resting high-frequency ECG data are extracted to obtain the third feature;

[0081] The first, second, and third features are used as the local detail features corresponding to each convolutional layer.

[0082] Understandably, the first convolutional layer uses a 1*31 convolutional kernel to extract features from resting high-frequency ECG data, focusing on capturing the rapid fluctuation details of the QRS band in the signal. It mainly extracts the steep edges and drastic changes in the QRS band to obtain the first feature.

[0083] The second convolutional layer is used to extract the overall morphological features of the QRS band in the resting high-frequency ECG data using a 1*36 convolutional kernel to obtain the second feature;

[0084] Understandably, the second convolutional layer uses a 1*36 convolutional kernel to extract features from resting high-frequency ECG data, focusing on capturing the overall morphological features of the QRS complex and extracting features with a certain time dependence to obtain the second feature.

[0085] The third convolutional layer is used to extract heart rate changes and QRS intensity features from the QRS band in the resting high-frequency ECG data using a 1*41 convolutional kernel, thus obtaining the third feature.

[0086] Understandably, the third convolutional layer uses a 1*41 convolutional kernel to extract features from resting high-frequency ECG data, focusing on capturing the rhythm and trend information of the QRS complex, extracting features including heart rate changes and QRS intensity, and obtaining the third feature.

[0087] For example, the multi-scale convolutional neural module uses a 12x3000 input size and a three-layer one-dimensional convolutional (Conv1D) structure, containing 31, 36, and 41 filters respectively, with a kernel size of 3 and a stride of 5. The last layer, Conv1D-41, has an output dimension of 3x3x600, indicating that local high-frequency features are extracted through multi-channel convolution, and the receptive field is gradually expanded by stacking layers, providing hierarchical temporal feature representation for subsequent modules. The addition of Dropout and BatchNormalization further enhances the model's generalization ability. Information at different time scales in high-frequency ECG signals is extracted using convolutional kernels of different sizes, and the features extracted at each scale are concatenated into a fused feature vector. Subsequently, the fused feature vector is batch normalized to stabilize the input of each layer, accelerate network convergence, and introduce nonlinearity through the ReLU activation function to enhance feature representation ability.

[0088] The feature stitching module is used to stitch together the local detail features corresponding to each convolutional layer to obtain activation fusion stitched features;

[0089] Specifically, the feature splicing module includes: a filter splicing layer, a batch normalization layer, and a ReLU activation function layer;

[0090] The feature concatenation module is used to concatenate the local detail features corresponding to each branch to obtain activation fusion concatenation features, including:

[0091] The first feature, the second feature, and the third feature are concatenated according to the filter concatenation layer to obtain the initial activation fusion concatenation feature;

[0092] The initial activation fusion and splicing features are normalized according to the batch normalization layer to obtain normalized splicing features;

[0093] The normalized splicing features are subjected to nonlinear operations based on the ReLU activation function layer to obtain the activated fusion splicing features.

[0094] For example, the local detail features corresponding to each convolutional layer are concatenated to form a 9×600 fused feature matrix, and then further subjected to batch normalization and ReLU activation to obtain activated fused concatenated features; this ensures the stability of the features, speeds up the model convergence, and improves the feature representation capability.

[0095] The Long Short-Term Memory (LSTM) module is used to perform temporal modeling on the activation fusion splicing features to obtain temporal-related information;

[0096] Specifically, this includes the TimeDistributed layer and the LSTM layer connected in sequence;

[0097] The Long Short-Term Memory (LSTM) module is used to perform temporal modeling on activation fusion and splicing features to obtain temporal-related information, including:

[0098] Based on the TimeDistributed layer, the activated fusion and splicing features are flattened at each time step to obtain a one-dimensional feature vector corresponding to each time step;

[0099] The LSTM layer is used to perform time series modeling on the one-dimensional feature vector corresponding to each time step, and the long-short time dependency relationship is captured to obtain time series related information.

[0100] For example, in the temporal modeling stage, the LSTM layer processes the flattened temporal features. The LSTM layer has 200 hidden units to capture the temporal dependencies of ECG signals, extract global temporal features, and enable the network to recognize long-term signal pattern changes, improving its ability to perceive dynamic features. It captures the sequential dependencies of signals and extracts time-related information. The LSTM layer uses a 200-hidden-unit Long Short-Term Memory module to perform deep modeling of the temporal features extracted by the multi-scale convolutional neural module. Its core function is to capture the long-range dependencies of ECG signals, such as the periodic features of arrhythmia or the temporal correlation of abnormal waveforms. The Transformation layer, combined with feature dimension transformation, adapts the LSTM layer output to the input dimension of the adaptive convolution module Involution, forming an end-to-end joint optimization.

[0101] The adaptive convolution module is used to randomly generate dynamic convolution kernels to extract features of time-related information and obtain the time-related detailed features corresponding to each convolutional layer;

[0102] Specifically, the adaptive convolution module includes a fourth, fifth, and sixth convolutional layer in parallel; wherein the number of convolutional kernels in the fourth, fifth, and sixth convolutional layers changes adaptively based on time-related information.

[0103] The adaptive convolution module is used to randomly generate dynamic convolution kernels to extract features related to temporal information, obtaining temporal detail features corresponding to each convolutional layer, including:

[0104] Based on the small-scale high-frequency transient features of time-related information extracted from the fourth convolutional layer, the first self-convolutional time-related information is obtained.

[0105] Based on the mid-scale mid-frequency morphological features of the time-related information extracted from the fifth convolutional layer, the second self-convolutional time-related information is obtained.

[0106] Based on the large-scale low-frequency rhythmic features of time-related information extracted from the sixth convolutional layer, the third self-convolutional time-related information is obtained.

[0107] For example, the network introduces an adaptive convolution module called Involution. Unlike traditional convolution, this module dynamically generates convolution kernels based on the input, thus extracting local features more flexibly. By setting different kernel sizes, the Involution module further extracts features at different scales and concatenates and fuses them to obtain richer local spatiotemporal information. Involution employs a channel-adaptive deconvolution structure; Involution-31 uses 5 channels with a stride of 4.5, while Involution-41 compresses to 3 channels and outputs a 3x16x120 feature map. Compared to traditional convolution, Involution, by dynamically generating convolution kernel parameters, reduces computational cost while enhancing the ability to model the spatial-channel interactions of signals from different leads, making it suitable for processing multi-dimensional feature fusion of 12-lead ECG signals.

[0108] The classification module is used to classify the temporal detail features corresponding to each convolutional layer and obtain the classification results.

[0109] Specifically, the classification module includes: a channel feature splicing layer, a batch normalized activation flattening layer, a regularized fully connected activation layer, and a fully connected discriminant layer;

[0110] The classification module is used to classify the temporal detail features corresponding to each convolutional layer, and obtain the classification results, including:

[0111] Based on the channel feature splicing layer, the temporal detail features corresponding to each convolutional layer are spliced ​​together to obtain the initial spliced ​​self-convolutional temporal related information;

[0112] Based on the batch normalization activation flattening layer, the initial spliced ​​autoconvolutional temporal information is subjected to batch normalization, activation and flattening operations to obtain one-dimensional autoconvolutional temporal features.

[0113] Based on the regularized fully connected activation layer, the one-dimensional self-convolutional temporal features are regularized and compressed to obtain compressed self-convolutional temporal features.

[0114] Based on the fully connected discriminant layer, the compressed self-convolutional temporal features are classified to obtain the classification results.

[0115] The loss function used in training the ECG signal classification network is:

[0116]

[0117] Where N represents the total number of training samples; y i p represents the true label of the i-th training sample; i This represents the prediction result for the i-th training sample.

[0118] This invention uses resting high-frequency electrocardiogram (ECG) data for training, completely independent of exercise load. Once a set number of ECG data points are collected, they can be directly input into a pre-trained ECG signal classification network, providing real-time monitoring results of resting high-frequency ECG data for myocardial ischemia, eliminating the lag in the analysis process of traditional methods.

[0119] This invention utilizes an automatic feature extraction network based on electrocardiogram (ECG) signals, avoiding the shortcomings of traditional methods that rely on manually designed feature extraction and classification algorithms, thus improving diagnostic efficiency and accuracy. After deep learning optimization, the diagnostic accuracy of ECG can reach over 90%, significantly improving the early diagnostic capability for myocardial ischemia and myocardial infarction.

[0120] When training an ECG signal classification network: first clean abnormal ECG signals, then perform signal filtering and noise removal.

[0121] Abnormal ECG signal cleaning: To construct a high-quality ECG signal classification network, training was first performed based on resting high-frequency ECG data from 56 groups of patients, with each group's data acquisition time being approximately 3 minutes. First, the QRS waveform was extracted from each group of high-frequency ECG data, and each group was divided into sets of 3000 points (i.e., 3 seconds).

[0122] To remove noise and abnormal waveforms, correlation coefficient analysis was performed on each heartbeat beat within each group of 3000 points. If the correlation coefficient was less than 0.8, the data set was considered to contain abnormal waveforms or noise and was discarded. Synchronous processing of 12-lead signals was used for QRS wave extraction and correlation coefficient analysis. If the QRS waveform of a lead could not be extracted, or the correlation coefficient was less than 0.8, all lead data in that group were discarded. This process ensured the quality and consistency of the data used.

[0123] Signal filtering and noise removal: After completing the correlation coefficient analysis, we performed a bandpass filter (150–250 Hz) to extract resting high-frequency ECG data. Subsequently, the root mean square (RMS) value was calculated for the signal intensity outside the QRS complex, serving as a measure of noise level. If the noise level exceeded 1 μV, the data set was considered too noisy and was discarded.

[0124] After completing the above data preprocessing, the resting high-frequency electrocardiogram data of 56 patients were divided into 2022 valid data sets, each with a data length of 3 seconds (i.e., 3000 data points). This amount of data is sufficient for deep learning training, and each data set includes labels from doctors indicating the presence or absence of myocardial ischemia.

[0125] In a specific embodiment of the present invention, the deep learning method used in the present invention is based on multi-scale convolutional neural modules and long short-term memory modules, and is designed to better adapt to the multi-scale features and temporal dynamics of high-frequency ECG signals for resting high-frequency ECG data.

[0126] Specifically, the first step is to extract multi-scale features from the input resting high-frequency ECG data in the multi-scale convolutional neural module. This process can capture the local details of the QRS waveform in the resting high-frequency ECG data, avoiding the limitations of traditional manual feature extraction.

[0127] Subsequently, multi-scale features are spliced ​​and activated to perform temporal modeling, obtaining time-related information. The LSTM layer also focuses on the temporal relationship between local detail features, which has a stronger ability to identify rapidly changing pathological signals in resting high-frequency ECG data.

[0128] This invention acquires and preprocesses resting electrocardiogram (ECG) data to obtain high-frequency resting ECG data, eliminating the need for patients to undergo exercise training and making it applicable to a wider range of people. On the hardware side, a portable high-frequency ECG acquisition system was designed, focusing on solving the technical challenges of portability and high-frequency signal acquisition. It employs a multi-channel signal acquisition circuit, filtering and input protection circuit, buffer circuit, differential amplification and AD conversion circuit, right leg drive circuit, ECG digital signal processor, and wireless data transmission module. These modules work together efficiently to achieve high-quality ECG signal acquisition, real-time processing, and wireless transmission, ensuring the device is lightweight and compact while possessing high-performance monitoring capabilities. On the software side, the high-frequency ECG data processing flow was optimized. A multi-scale convolutional neural module was used to extract multi-scale features, and batch normalization and ReLU activation were combined to enhance feature representation. Subsequently, time-distributed layers were used to flatten time-series features, and LSTM was used for temporal modeling to extract the sequential dependencies of the signal. Simultaneously, Involution adaptive convolution was introduced to dynamically generate convolution kernels to extract local spatiotemporal features and fuse information from different scales. Finally, the features are batch normalized, activated, Flattened, and dimensionality reduced by a fully connected layer to output class predictions, and the model is optimized using cross-entropy loss. This invention achieves high-quality ECG monitoring while significantly improving the portability of the device.

[0129] Secondly, this invention provides a portable dynamic high-frequency electrocardiogram (ECG) monitoring system and a portable dynamic high-frequency ECG monitoring method, see [link to relevant documentation]. Figure 4 The system includes: a front-end acquisition and processing module, a primary signal processing module, a wireless communication module, a result prediction module, and a power supply module;

[0130] Front-end acquisition and processing module: used to acquire multiple channels of body surface potentials multiple times, and to obtain multiple sets of primary high-frequency ECG signals by differential amplification and AD conversion of high-frequency ECG signals;

[0131] Here, multiple sets of primary high-frequency ECG signals represent the collection of multiple sampled data continuously acquired by the system at the original sampling frequency (16kSPS) after AD conversion. Each set represents a complete signal sample captured in a very short time.

[0132] Here, the front-end acquisition and processing module includes: an acquisition unit, a filtering unit, a buffer unit, a right leg drive noise reduction unit, a differential amplification unit, and an AD conversion unit;

[0133] Front-end acquisition and processing module: used to acquire multiple channels of body surface potentials, and to obtain multiple sets of primary high-frequency ECG signals by differential amplification and AD conversion of the high-frequency ECG signals, including:

[0134] Acquisition Unit: Used to acquire multiple sets of multi-channel body surface potentials using an electrocardiogram sensor; each set of multi-channel body surface potentials includes: body surface potentials at electrode positions LA, LL, RA and V1 to V6;

[0135] Filtering unit: used to perform low-pass filtering on multi-channel body surface potentials to obtain low-pass filtered multi-channel body surface potentials;

[0136] Buffer unit: Used to provide low output impedance for the multi-channel body surface potential after low-pass filtering, so as to obtain the buffered multi-channel body surface potential;

[0137] Right leg drive noise reduction unit: used to perform noise reduction processing on the buffered multi-channel body surface potential to obtain the noise-reduced multi-channel body surface potential;

[0138] Differential amplifier unit: used to perform differential processing on the noise-reduced multi-channel body surface potential to obtain multiple sets of differential amplified signals;

[0139] AD conversion unit: used to convert multiple sets of differential amplified signals into multiple sets of primary high-frequency electrocardiogram signals.

[0140] Primary signal processing module: used to sum and downsample multiple sets of primary high-frequency ECG signals to obtain a noise-reduced primary high-frequency ECG signal; then, the noise-reduced primary high-frequency ECG signal is calculated to obtain a standard high-frequency ECG signal;

[0141] The primary signal processing module here includes: a signal noise reduction module and a standard high-frequency ECG signal conversion module;

[0142] Signal noise reduction module: used to sum and downsample multiple sets of data from the acquired primary high-frequency ECG signal to obtain the noise-reduced primary high-frequency ECG signal;

[0143] Standard high-frequency ECG signal conversion module: used to convert the noise-reduced primary high-frequency ECG signal using standard ECG signal conversion formulas to obtain a standard high-frequency ECG signal, formulas (1.2) to (1.12).

[0144] Wireless communication module: used to transmit standard high-frequency ECG signals to the corresponding display module via the wireless transmission function of the integrated Bluetooth module;

[0145] The result prediction module is used to predict the results based on the pre-trained ECG signal classification network and standard high-frequency ECG signals, and obtain the monitoring results of resting high-frequency ECG for myocardial ischemia; wherein, the pre-trained ECG signal classification network is deployed on the dynamic high-frequency ECG monitoring system;

[0146] Power supply module: Used to supply power to the front-end acquisition and processing module, wireless communication module and primary signal processing module.

[0147] For example, the principle of electrocardiogram (ECG) signal detection is essentially based on the weak periodic currents generated by the heart during depolarization and repolarization. These currents are conducted throughout the body via tissues and fluids, causing periodically changing potentials in various parts of the body surface. By placing electrodes at different locations on the skin, corresponding potential signals can be collected, and the potential differences between different locations constitute the common electrocardiogram (ECG). Figure 3 As shown, the electrodes of the bipolar limb leads are placed on the four limbs of the human body. The measuring electrodes are located in three positions: the right arm (RA), the left arm (LA), and the left leg (LL). Different combinations of these electrodes form the signals of the three leads I, II, and III.

[0148] like Figure 3 As shown, this system comprehensively reflects the electrical activity of the heart in the coronary and transverse planes through the electrode layout at different locations on the limbs and chest.

[0149] The following are the formulas for calculating the WCT at the Wilson center of a 12-lead ECG:

[0150] The calculation formulas for bipolar limb leads I, II, and III are as follows:

[0151] Lead I = V LA -V RA (1.2)

[0152] Lead II = V LL -V RA (1.3)

[0153] Lead III = V LL -V LA =Lead II - Lead I (1.4)

[0154] The formulas for calculating aVR, aVL, and aVF in compression limb leads are as follows:

[0155] Lead III = V LL -V LA =Lead II - Lead I (1.5)

[0156] aVL=V LA -(V RA +V LL ) / 2=LeadI-0.5*LeadII (1.6)

[0157] aVF=V LL -(V RA +V LL ) / 2=LeadII-0.5*LeadI (1.7)

[0158] The formula for calculating the Wilson central terminal potential (WCT) is:

[0159] WCT = (V RA +V LA +V LL ) / 3(1.8)

[0160] The formulas for calculating leads V1 to V6 in the chest wall are as follows:

[0161] V i =V cheat,i -WCT,i∈{1,2,3,4,5,6}(1.9)

[0162] The above eight formulas can be used to calculate the 12-lead ECG signal, including three bipolar limb leads, three compression limb leads, and six chest wall leads. Of these 12 leads, Lead III, aVR, aVL, and aVF can all be calculated using Lead I and Lead II. Therefore, in actual data acquisition, only V needs to be acquired. LA V RA V LL The potentials of the three electrodes can be used to deduce the electrocardiogram signals of the bipolar limb leads and the compression limb leads.

[0163] Similarly, for chest wall leads, after obtaining V LA V RA V LL After obtaining the potentials at the three electrodes, the reference potential of the WCT can be calculated. Therefore, only the potentials at the additional six chest wall lead locations need to be acquired to obtain the complete ECG signals for the six chest wall leads. Thus, in the 12-lead ECG signal, a total of 10 electrodes are used, of which 9 electrodes are used to record the signal, obtaining the ECG signals of 8 leads (Lead I, Lead II, V1-V6). The ECG signals of the other 4 leads (Lead III, aVR, aVL, aVF) are obtained through calculation, thereby generating 12 ECG leads. The RL (right leg) electrode serves as the driving reference electrode and does not participate in the formation of the ECG signal, but is used to reduce common-mode noise and improve signal quality.

[0164] Furthermore, for the chest wall lead signals of Lead I, Lead II, and 6 that need to be measured, there is a simpler method that plays a crucial role in circuit optimization. The Lead I, Lead II, and V signals are transformed using formulas (1.2), (1.3), and (1.8). RA V LA V LL The relationship can be modified by changing WCT, which can be transformed into the following formula:

[0165] WCT = (V RA +Lead I+V RA +Lead II+V RA ) / 3(1.10)

[0166] Finally, WCT can be transformed into:

[0167] WCT = V RA +(Lead I+Lead II) / 3(1.11)

[0168] At this point, the calculation formula for chest leads V1 to V6 can be derived from formula (1.9) and become:

[0169] V i =V cheat,i -V RA -(Lead I+Lead II) / 3,i∈{1,2,3,4,5,6}(1.12)

[0170] Combining equations (1.2), (1.3), and (1.21), among the eight ECG signals (Lead I, Lead II, V1-V6) that need to be measured, V can be obtained. RA As the negative terminal, this transformation is applied to the front-end acquisition circuit of ECG, which can effectively reduce the complexity of the circuit. It eliminates the need to first construct a Wilson central potential (WCT) circuit; all differential signals can be directly connected to V. RA Then, through simple mathematical calculations in the microcontroller, the electrocardiogram signals of each lead in the standard formula can be obtained.

[0171] The filtering unit is a protection circuit that limits the energy and performs low-pass filtering on the input high-frequency ECG signal. The entire circuit design consists of a gas discharge tube (2RK075M-4), a current-limiting resistor, a filter capacitor, and a rectifier diode (BAV199). The two ends of the gas discharge tube (2RK075M-4) are connected to the lead input terminal and the floating ground terminal, respectively, to limit the voltage amplitude of the input signal. When the amplitude of the input signal exceeds the set threshold, the gas discharge tube will conduct, transferring the excess energy of the signal to the ground wire, thereby preventing excessive signal from damaging subsequent circuits.

[0172] One end of the current-limiting resistor is connected to the input terminal, and the other end is connected to the rectifier diode. This limits the input signal current, preventing excessive current from affecting subsequent circuits. One end of the filter capacitor is connected to the non-inverting input of the input buffer circuit, and the other end is grounded, forming an RC low-pass filter circuit. This circuit performs low-pass filtering on the input signal, achieving high-frequency noise suppression. The low-pass RC filter parameters set here are resistor R = 44kΩ and capacitor C = 100pF. This circuit can effectively attenuate high-frequency signals above 36kHz. The rectifier diode (BAV199) adopts a center-tapped structure, with the center end connected to the non-inverting input of the input buffer circuit, and the other two ends connected to the positive and negative power supplies, respectively. The main function of this circuit is to rectify the input signal and provide a stable voltage reference for differential signal processing in subsequent circuits.

[0173] The buffer unit employs a unity-gain (Voltage Follower) architecture, with its core component being the low-noise, low-offset-voltage TLV2221IDBVR operational amplifier. This design aims to improve the input impedance of high-frequency ECG signals, reduce signal attenuation caused by skin-electrode contact impedance, and provide low output impedance to ensure stable signal transmission to subsequent circuits. The circuit utilizes a ±2.5V symmetrical dual power supply to ensure signal integrity and is equipped with an R25 (18kΩ) current-limiting resistor and C108 and C109 (100nF) decoupling capacitors to suppress noise and improve power supply stability. The overall design effectively reduces signal distortion, enhances the accuracy of high-frequency ECG signal acquisition, and provides high-quality input for subsequent processing.

[0174] The Right Leg Drive (RLD) circuit is a negative feedback technique used to reduce common-mode noise and improve the signal-to-noise ratio (SNR) of electrocardiogram (ECG) signals. Because human electrode signals are highly susceptible to power line interference (50Hz / 60Hz) and other common-mode noise, a right leg drive circuit is typically used to feed the common-mode signal back to the body to actively cancel out this noise and improve the quality of the ECG signal.

[0175] Common-mode signal extraction in the right leg drive noise reduction unit typically involves extracting the common-mode signal from the left and right hand electrodes (RA and LA) or all limb leads (RA, LA, LL). The extracted common-mode signal is then amplified in reverse by an inverting amplifier. To avoid the introduction of high-frequency noise after inversion, an RC low-pass filter is usually added to the output of the right leg drive circuit to filter out high-frequency noise. The filtered signal is finally fed back to the human body through the right leg electrode (RL), thereby effectively suppressing common-mode interference and improving the accuracy and stability of ECG signal acquisition.

[0176] The electrocardiogram sensor of this invention is the ADS131E08S chip. With its advantages of high precision, low noise and multi-channel synchronous sampling, the ADS131E08S chip can effectively improve signal quality and provide high-quality input for subsequent data processing, making it very suitable for the needs of this system.

[0177] The ADS131E08S is a highly integrated multi-channel synchronous sampling ADC chip that supports up to 24-bit Δ-Σ (ΔΣ) analog-to-digital conversion and is designed for high-precision signal acquisition. This chip integrates a programmable gain amplifier (PGA), a high-precision reference source, and a low-noise oscillator, reducing reliance on external components and thus optimizing overall system performance and power consumption. Its sampling rate can be set from 1 to 64 kSPS via a configuration register, and the PGA gain is adjustable to 1, 2, 4, 8, or 12 to accommodate input signals of different amplitudes. Furthermore, the chip boasts a dynamic range of 118 dB at a 1 kSPS sampling rate, a crosstalk of -125 dB, and a total harmonic distortion (THD) of -100 dB at both 50 Hz and 60 Hz, with an internal reference voltage drift of only 8 ppm / ℃. Its operating temperature range covers -40℃ to +105℃, meeting the environmental requirements for routine ECG signal acquisition.

[0178] After passing through the differential amplifier unit, the analog signal is transmitted to the ADS131E08S and amplified by the integrated programmable gain amplifier (PGA). This PGA module supports adjustable gains of 1, 2, 4, 8, and 12 times, with varying system noise levels at different gain ratios. Generally, a higher gain ratio results in lower system noise, but the noise level is also affected by the sampling frequency.

[0179] Furthermore, the ADS131E08S supports both 3V and 5V power supplies. Considering the portability requirements, this solution chooses a 3V power supply to reduce power consumption. At this point, the reference voltage is 2.4V, while the amplitude range of the ECG signal is 0.5–4mV. Table 2.1 shows the system noise levels at different sampling frequencies and gains. To minimize noise, a higher gain factor should be selected. If a gain of 12 is set, the maximum amplitude of the ECG signal can be amplified to 48mV, far below the reference voltage of 2.4V, ensuring the signal remains within a safe range. Therefore, this invention selects a gain factor of 12 to optimize signal quality and improve acquisition accuracy.

[0180] The amplified analog signal then needs to be processed by a ΔΣ analog-to-digital converter (ADC). The ADS31E08S offers seven selectable sampling frequencies: 1, 2, 4, 8, 16, 32, and 64 kSPS. Sampling frequencies of 16 kSPS and below can support a maximum resolution of 24 bits, while 32 kSPS and 64 kSPS can only support a maximum resolution of 16 bits.

[0181] Monitoring high-frequency ECG signals requires capturing subtle changes in electrical signals; therefore, the sampling bit depth should be maximized. At 24-bit resolution, the maximum supported sampling frequency is 16 kSPS, corresponding to a system -3dB cutoff frequency of 4192 Hz. Typically, the frequency range required for high-frequency ECG signal analysis is 150–250 Hz, and to ensure complete signal acquisition, the -3dB cutoff frequency generally needs to be at least 10 times the maximum analysis frequency. Therefore, 4192 Hz ≥ 10 × 250 Hz meets this requirement. Thus, this invention employs a 16 kSPS sampling rate to achieve the highest sampling efficiency at 24-bit precision, ensuring the integrity and accuracy of the high-frequency ECG signal.

[0182] Eight differential signals can be obtained by acquiring data through the eight differential channels of the AD conversion unit, where LA, LL, and V1 to V6 are positive, and RA is negative. According to formulas (1.2) and (1.3) in Section 2.2, the Lead I and Lead II signals correspond to the digital signals acquired by channel 1 and channel 2, respectively.

[0183] According to formula (1.21), the differential signals acquired by channels 3 to 8 still have a certain deviation from the standard ECG chest lead signals. Since Lead I and Lead II are acquired by channels 1 and 2 respectively, the differential signals need to be converted at the microcontroller end to obtain the standard chest lead signals.

[0184] In addition, the Lead III, aVR, aVL and aVF lead signals can be obtained by performing mathematical operations on Lead I and Lead II (as shown in formulas (1.4) to (1.7)).

[0185] The ADS131E08S has a sampling frequency set to 16kSPS and supports eight-channel simultaneous acquisition, resulting in a total data transmission volume of 16kSPS × 8. Each channel's sampled data has a bit width of 24 bits (3 bytes), typically transmitted in a 32-bit format to preserve the sign bit and optimize data processing. Due to the large overall data volume, direct transmission would significantly increase the microcontroller's computational and storage burden. Furthermore, while a higher ADC sampling rate improves time-domain resolution, it also increases the computational overhead of data processing. Therefore, averaging and downsampling of the raw ECG signal can be performed to reduce the data volume and improve signal quality.

[0186] This invention employs downsampling by averaging 16 sets of data to reduce the data volume, lessen the data transmission burden on the microcontroller, and smooth random noise, thereby improving the signal-to-noise ratio. After downsampling, the sampling frequency is reduced from 16 kSPS to 1 kSPS. According to the Nyquist sampling theorem, the sampling frequency must be at least twice the highest frequency of the signal to avoid aliasing (F_s ≥ 2F_max). The frequency range after downsampling is still four times that of the 150–250 Hz band of high-frequency ECG signals (1 kSPS ≥ 2 × 250 Hz), which can completely preserve the key information of high-frequency ECG signals.

[0187] In addition, to filter out baseline drift interference, this invention employs a Butterworth digital high-pass filter to filter the data after conversion of the standard ECG signal, and deploys the designed digital filter into the MCU. A second-order IIR high-pass Butterworth filter was designed using the filterDesigner toolkit in Matlab, with a sampling frequency of 1kHz and a cutoff frequency of 0.67Hz.

[0188] After the STM32F405, acting as the main control chip, processes and packages the data, it needs to be transmitted. To meet portability requirements, wireless transmission is necessary. In this invention, the STM32F072 is used as the wireless communication control unit, paired with the HC-04 wireless communication module for wireless data transmission, thus meeting the real-time transmission requirements of portable high-frequency electrocardiogram signals.

[0189] The STM32F072 is a low-power microcontroller based on the ARM Cortex-M0 core, with a main frequency of up to 48MHz, and integrates USART, SPI, and I2C interfaces. 2 It features multiple communication interfaces such as USB and has powerful peripheral control capabilities. In the system of this invention, the STM32F072 was chosen as the wireless transmission control unit primarily based on the following considerations: First, this chip can independently handle wireless transmission tasks, decoupling data communication from the main control chip STM32F405, reducing the computational burden on the main control unit, and improving system stability; second, the STM32F072 is compatible with various wireless communication modules, supports flexible communication protocols, meets the portability requirements of this system, and optimizes the efficiency and power consumption of wireless data transmission.

[0190] The HC-04 is a wireless serial communication module based on the Bluetooth 2.0 protocol, supporting a UART interface, and featuring low power consumption, high stability, and ease of integration. In the system of this invention, the HC-04 acts as a wireless communication module, responsible for transmitting data processed by the STM32F072 to a host computer (such as a PC or mobile device) via Bluetooth. Its low power consumption makes it suitable for portable devices, while the simple configuration of the UART interface facilitates efficient interfacing with the STM32F072, enabling stable data transmission.

[0191] The HC-04, as a wireless communication module, is mainly responsible for receiving and sending data. The receiving part is used to parse the instructions sent by the host computer, which include two types of instructions: start (“START”) and end (“CLOSE”), and then pass the instructions to the STM32F072 for execution of the corresponding tasks. In the STM32F072, the main functions are data acquisition and data packet integrity judgment, specifically including: (1) Data verification: performing checksum verification and data length checks on the data acquired from the main control chip STM32F405 to ensure the integrity and accuracy of the data. (2) Anomaly handling: if the data is complete, it is sent to the host computer through the HC-04; if the data is missing or abnormal, it is discarded to avoid error propagation.

[0192] The front-end acquisition and processing module uses a central electrocardiogram (ECG) sensor to collect high-frequency ECG signals (150Hz-250Hz) from the human body surface, capturing subtle changes in myocardial activity. The acquired weak high-frequency ECG signals are amplified to ensure a high signal-to-noise ratio. The amplified analog signal is then converted into a digital signal by a sampling rate AD converter to meet the high-frequency characteristic analysis requirements of the ECG signal, providing high-quality signal input for subsequent data processing and analysis. The processing includes averaging and noise reduction, baseline drift filtering, and low-pass filtering to reduce noise interference and ensure the extracted high-frequency ECG signal is as pure as possible. Furthermore, the microcontroller within the processing module is responsible for overall system control, ensuring the coordinated operation of all modules.

[0193] The wireless communication module, integrating a Bluetooth module, enables the device to transmit data and communicate in real time with mobile devices (such as smartphones and tablets) or cloud platforms. This function facilitates remote monitoring and storage of data, ensuring that medical personnel can monitor the patient's ECG status at any time and provide timely diagnosis and intervention.

[0194] This invention combines LSTM in a pre-trained ECG signal classification network to capture the time-series features of signals, enabling real-time analysis of pathological states such as myocardial ischemia and myocardial infarction. After the ECG signal classification network is trained, it undergoes quantization and compression processing, and is successfully deployed in a portable device, achieving efficient and low-power dynamic high-frequency ECG monitoring. Compared with traditional methods, this invention significantly improves the accuracy of high-frequency ECG signal analysis, and its portability allows the device to provide early warning of myocardial infarction in daily life. This device does not rely on exercise load as an excitation condition, making it more suitable for people with exercise difficulties, contributing to the prevention of myocardial infarction, and providing important support for the early diagnosis of cardiovascular diseases.

[0195] The power module utilizes a high-efficiency battery module, supporting extended portable monitoring. This battery module is optimized to ensure the device's stability and reliability during prolonged use, meeting patients' needs for continuous wear in daily life and ensuring sufficient power endurance for early warning of diseases such as myocardial ischemia and myocardial infarction.

[0196] This invention utilizes deep learning technology to train high-frequency electrocardiogram (ECG) signals and deploys the trained model onto a portable device. Unlike traditional methods, this invention uses resting high-frequency ECG data for training, completely independent of exercise load. Once a set number of ECG data points are collected, they can be directly input into the trained network, providing real-time diagnostic results and eliminating the lag in the analysis process of traditional methods. Through automatic feature extraction by the deep learning model, the shortcomings of traditional methods relying on manually designed feature extraction and classification algorithms are avoided, improving diagnostic efficiency and accuracy. After deep learning optimization, the diagnostic accuracy of ECG can reach over 90%, significantly improving the early diagnostic capability of myocardial ischemia and myocardial infarction.

[0197] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0198] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A portable dynamic high-frequency electrocardiogram monitoring method, characterized in that, include: Acquire resting electrocardiogram (ECG) data, and perform data preprocessing on the resting ECG data to obtain resting high-frequency ECG data; The resting high-frequency ECG data is input into a pre-trained ECG signal classification network to obtain the monitoring results of resting high-frequency ECG for myocardial ischemia; wherein, the ECG signal classification network includes: a multi-scale convolutional neural module, a feature splicing module, a long short-term memory module, an adaptive convolutional module, and a classification module; The multi-scale convolutional neural module is used to extract multi-scale features from resting high-frequency electrocardiogram data to obtain local detail features corresponding to each convolutional layer. The feature splicing module is used to splice the local detail features corresponding to each convolutional layer to obtain activation fusion spliced ​​features; The long short-term memory module is used to perform temporal modeling on the activation fusion splicing features to obtain temporal related information; The adaptive convolution module is used to randomly generate dynamic convolution kernels to extract features from the time-related information and obtain the time-related detailed features corresponding to each convolutional layer. The classification module is used to classify the temporal detail features corresponding to each convolutional layer to obtain the classification result.

2. The portable dynamic high-frequency electrocardiogram monitoring method according to claim 1, characterized in that, The step of preprocessing the resting electrocardiogram (ECG) data to obtain resting high-frequency ECG data includes: The resting ECG data is split to obtain multiple sets of split resting ECG data; After cleaning the multiple sets of split resting ECG data, the correlation coefficient between the data within each set of split resting ECG data is calculated, and it is determined whether the correlation coefficient is less than a first preset value. If the correlation coefficient between data within a group of split resting ECG data is less than a first preset value, then the split resting ECG data is discarded from the multiple groups of split resting ECG data. For the resting ECG data after splitting the data within the group with a correlation coefficient not less than the first preset value, a bandpass filter is used to extract high-frequency ECG signals from the split resting ECG data to obtain resting high-frequency ECG data.

3. The portable dynamic high-frequency electrocardiogram monitoring method according to claim 1, characterized in that, The multi-scale convolutional neural module includes a first convolutional layer, a second convolutional layer, and a third convolutional layer arranged in parallel; The multi-scale convolutional neural module is used to extract multi-scale features from resting high-frequency electrocardiogram data to obtain local detail features corresponding to each convolutional layer, including: Based on the first convolutional layer, rapid fluctuation detail features of the QRS band in the resting high-frequency ECG data are extracted to obtain the first feature; The second feature is obtained by extracting the overall morphological features of the QRS band in the resting high-frequency ECG data based on the second convolutional layer. Based on the third convolutional layer, the heart rate transformation and QRS intensity features of the QRS band in the resting high-frequency ECG data are extracted to obtain the third feature; The first feature, the second feature, and the third feature are used as local detail features corresponding to each convolutional layer.

4. The portable dynamic high-frequency electrocardiogram monitoring method according to claim 3, characterized in that, The feature concatenation module includes: a filter concatenation layer, a batch normalization layer, and a ReLU activation function layer; the feature concatenation module is used to concatenate the local detail features corresponding to each branch to obtain activation fusion concatenation features, including: The first feature, the second feature, and the third feature are spliced ​​together according to the filter splicing layer to obtain the initial activation fusion splicing feature; The initial activation fusion splicing features are normalized according to the batch normalization layer to obtain normalized splicing features; The normalized splicing features are subjected to a nonlinear operation based on the ReLU activation function layer to obtain the activated fusion splicing features.

5. The portable dynamic high-frequency electrocardiogram monitoring method according to claim 1, characterized in that, The long short-term memory module includes a TimeDistributed layer and an LSTM layer connected in sequence; The long short-term memory module is used to perform temporal modeling on the activation fusion and splicing features to obtain temporal-related information, including: The activation fusion and splicing features are flattened at each time step according to the TimeDistributed layer to obtain a one-dimensional feature vector corresponding to each time step; The LSTM layer is used to perform time series modeling on the one-dimensional feature vector corresponding to each time step, and the long-short time dependency relationship is captured to obtain time series related information.

6. The portable dynamic high-frequency electrocardiogram monitoring method according to claim 1, characterized in that, The adaptive convolution module includes a fourth, fifth, and sixth convolutional layer arranged in parallel; wherein the number of convolutional kernels in the fourth, fifth, and sixth convolutional layers changes adaptively based on time-related information. The adaptive convolution module is used to randomly generate dynamic convolution kernels to extract features from the temporal-related information, obtaining temporal detail features corresponding to each convolutional layer, including: The first self-convolutional temporal-related information is obtained by extracting the small-scale high-frequency transient features of the time-related information from the fourth convolutional layer. The second self-convolutional temporal related information is obtained by extracting the mesoscale mid-frequency morphological features of the time-related information from the fifth convolutional layer. The third self-convolutional temporal related information is obtained by extracting the large-scale low-frequency rhythmic features of the time-related information from the sixth convolutional layer.

7. The portable dynamic high-frequency electrocardiogram monitoring method according to claim 1, characterized in that, The classification module includes: a channel feature splicing layer, a batch normalized activation flattening layer, a regularized fully connected activation layer, and a fully connected discriminant layer; The classification module is used to classify the temporal detail features corresponding to each convolutional layer to obtain classification results, including: Based on the channel feature splicing layer, the temporal detail features corresponding to each convolutional layer are spliced ​​together to obtain the initial spliced ​​self-convolutional temporal related information. Based on the batch normalization activation flattening layer, the initial spliced ​​self-convolutional temporal related information is subjected to batch normalization, activation and flattening operations to obtain one-dimensional self-convolutional temporal features. Based on the regularized fully connected activation layer, the one-dimensional self-convolutional temporal features are regularized and compressed to obtain compressed self-convolutional temporal features. The compressed self-convolutional temporal features are classified based on the fully connected discriminant layer to obtain the classification result.

8. The portable dynamic high-frequency electrocardiogram monitoring method according to claim 1, characterized in that, The loss function used in training the ECG signal classification network is: Where N represents the total number of training samples; y i p represents the true label of the i-th training sample; i This represents the prediction result for the i-th training sample.

9. A portable dynamic high-frequency electrocardiogram monitoring system, characterized in that, For implementing the portable dynamic high-frequency electrocardiogram monitoring method as described in any one of claims 1 to 8, the system comprises: a front-end acquisition and processing module, a primary signal processing module, a wireless communication module, a result prediction module, and a power supply module; The front-end acquisition and processing module is used to acquire multi-channel body surface potentials multiple times, and to obtain multiple sets of primary high-frequency electrocardiogram signals by differential amplification and AD conversion of the high-frequency electrocardiogram signals. The primary signal processing module is used to sum and downsample the multiple sets of primary high-frequency ECG signals to obtain a denoised primary high-frequency ECG signal; then, it calculates the denoised primary high-frequency ECG signal to obtain a standard high-frequency ECG signal. The wireless communication module is used to transmit the standard high-frequency electrocardiogram signal to the corresponding display module according to the wireless transmission function of the integrated Bluetooth module; The result prediction module is used to predict the result based on the pre-trained ECG signal classification network and the standard high-frequency ECG signal to obtain the monitoring result of resting high-frequency ECG for myocardial ischemia; wherein, the pre-trained ECG signal classification network is deployed on the dynamic high-frequency ECG monitoring system; The power supply module is used to supply power to the front-end acquisition and processing module, the wireless communication module, and the primary signal processing module.

10. The portable dynamic high-frequency electrocardiogram monitoring system according to claim 9, characterized in that, include: The front-end acquisition and processing module includes: an acquisition unit, a filtering unit, a buffer unit, a right leg drive noise reduction unit, a differential amplification unit, and an AD conversion unit; The front-end acquisition and processing module is used to acquire multi-channel body surface potentials multiple times, and to obtain multiple sets of primary high-frequency electrocardiogram signals by differential amplification and AD conversion of the high-frequency electrocardiogram signals, including: The acquisition unit is used to acquire multiple sets of multi-channel body surface potentials using an electrocardiogram sensor; each set of multi-channel body surface potentials includes the body surface potentials at electrode positions LA, LL, RA and V1 to V6. The filtering unit is used to perform low-pass filtering on the multi-channel body surface potential to obtain the low-pass filtered multi-channel body surface potential. The buffer unit is used to provide low output impedance to the multi-channel body surface potential after low-pass filtering, so as to obtain the buffered multi-channel body surface potential. The right leg drive noise reduction unit is used to perform noise reduction processing on the buffered multi-channel body surface potential to obtain the noise-reduced multi-channel body surface potential. The differential amplification unit is used to perform differential processing on the noise-reduced multi-channel body surface potential to obtain multiple sets of differential amplified signals. The AD conversion unit is used to perform AD conversion on the multiple sets of differential amplified signals to obtain multiple sets of primary high-frequency electrocardiogram signals.

Images