A portable intelligent ECG monitoring system based on an edge-cloud architecture

Abstract

This invention discloses a portable intelligent electrocardiogram (ECG) monitoring system based on an edge-cloud architecture. The system consists of three parts: an end device (ECG monitor), an edge device (smartphone), and a cloud device (cloud server). The end device collects the user's real-time ECG signals and transmits them to the edge device via Bluetooth. The edge device receives the real-time ECG signals and uses a hexagonal recognition algorithm based on ECG morphology to detect any abnormal ECG signals. If any are found, the signals are transmitted to the cloud device via 5G. The cloud device uses a CNN-LSTM model to more accurately identify the user's health status and provides feedback to the edge device. When the edge device receives a result indicating illness, it sends an alert and the user's location to emergency contacts. This system provides users with a more portable and efficient intelligent ECG monitoring system.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and Internet of Things technology, and relates to a portable intelligent electrocardiogram monitoring system based on an edge-cloud architecture. Background Technology

[0002] Myocardial infarction (MI) is a serious cardiovascular disease primarily caused by blockage or occlusion of the coronary arteries, leading to myocardial ischemia and subsequent myocardial cell necrosis. Myocardial infarction has a rapid onset and a short effective treatment time. Typical symptoms include severe chest pain, chest tightness, and shortness of breath, and it has a high mortality rate. Furthermore, myocardial infarction can be classified according to its location: inferior myocardial infarction (IMI), anterior septal myocardial infarction (ASMI), lateral medial myocardial infarction (ILMI), anterior myocardial infarction (AMI), anterolateral myocardial infarction (ALMI), and posteromedial myocardial infarction (IPLMI). Treatment methods differ depending on the location of the myocardial infarction. Therefore, rapid identification of the location of the myocardial infarction lesion is of great importance.

[0003] There are many methods used in medicine to detect myocardial infarction, including electrocardiography (ECG), serum biochemical marker testing, cardiac magnetic resonance imaging (MRI), echocardiography, and radionuclide scanning. Among these, ECG is the most commonly used method because it is non-invasive and relatively easy to obtain. When a myocardial infarction occurs, the ST segment, S wave, and T wave on the ECG will change accordingly. Given these indicators, real-time monitoring of myocardial infarction patients is essential.

[0004] Currently, intelligent ECG monitoring systems can be broadly categorized into two types. The first type embeds deep learning models into edge devices; however, due to the limited computing and storage capabilities of edge devices, as well as energy consumption issues, the practical results are generally limited. The second type embeds deep learning models into cloud devices. This design solves the problems of computing power, storage capacity, and energy consumption, but sacrifices portability, preventing users from checking their health anytime, anywhere. Therefore, a portable intelligent ECG monitoring system with real-time monitoring remains in demand in the market. Summary of the Invention

[0005] The purpose of this invention is to provide a portable intelligent electrocardiogram monitoring system based on an edge-cloud architecture to overcome the shortcomings of existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A portable intelligent electrocardiogram monitoring system based on an edge-cloud architecture is characterized in that the system includes an end device, an edge device, and a cloud device.

[0008] The terminal device consists of a self-developed portable 12-lead holter;

[0009] The edge device consists of a smart device or a smartwatch;

[0010] The cloud device consists of cloud server A and cloud server B provided by the PC;

[0011] The terminal device is used to collect the user's electrocardiogram signal and send it to the edge device via Bluetooth;

[0012] The edge device is used to detect whether there are abnormal signals from the user in real time and send the abnormal signals to the cloud device. In addition, it can also feed back the identification results of the cloud device to the user.

[0013] The cloud device is used to more accurately identify abnormal signals and provide users with precise health reports;

[0014] Specifically, it involves the following steps:

[0015] Step 1: The terminal device first collects the user's electrocardiogram signal in real time and sends it to the edge device via Bluetooth;

[0016] Step 2: The edge device consists of two modules: an identification module and a display module. In the identification module, this invention proposes a hexagonal recognition algorithm based on electrocardiogram morphology. This algorithm has a simple structure, fast processing speed, and can identify whether electrocardiogram signals are abnormal. In the display module, the device can obtain the binary classification result of the hexagonal recognition method based on electrocardiogram morphology (displayed as a hexagonal image), as well as the multi-classification result fed back by the cloud device, and finally provide the user with a health report. If a danger occurs, it will send an alert to emergency contacts.

[0017] Step 3: The cloud device consists of server A and server B. The CNN-LSTM model proposed in this invention is deployed on this device. Server A continuously trains the CNN-LSTM model and optimizes its parameters using historically stored abnormal ECG segments. Server B receives real-time abnormal ECG segments, uses the CNN-LSTM model to detect the patient's condition, and provides feedback.

[0018] As a preferred embodiment of the present invention, the hexagonal recognition algorithm based on electrocardiogram morphology collects morphological features of electrocardiogram signals as samples, including the relative positions of the S wave and the baseline, the T wave and the baseline, and the T wave and the R wave. The algorithm judges whether the signal is abnormal by the discrimination criteria of electrocardiogram morphology and displays it to the user in the form of a hexagonal image, thereby enhancing the visualization of the algorithm.

[0019] As a preferred embodiment of the present invention, the CNN-LSTM model: This model integrates CNN and LSTM, effectively extracts local features from input data and captures temporal correlations, thereby accurately identifying MI and the onset location of MI, and is a seven-classification model.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] The system's hexagonal recognition algorithm based on electrocardiogram morphology, deployed on edge devices, has a time complexity of only O(m) (where m is the number of R waves appearing in a 2-minute electrocardiogram segment), significantly reducing computational complexity. This algorithm sends abnormal electrocardiogram signals to the cloud without uploading personal information, ensuring the privacy and security of patient data. Normal electrocardiogram segments are deleted from local storage, reducing storage space and energy consumption. The CNN-LSTM model deployed in the cloud can not only identify whether a user has myocardial infarction (MI), but also locate the MI position and report it back to the edge device. This system achieves both portability and accuracy, providing a more convenient and efficient intelligent electrocardiogram monitoring platform for MI patients. Furthermore, simulations of the two algorithms on the PTB diagnostic electrocardiogram database showed recognition rates of 83.2% and 99.3%, respectively, achieving satisfactory results. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a portable intelligent ECG monitoring system based on an edge-cloud architecture provided in an embodiment of the present invention;

[0023] Figure 2 The present invention provides a normal electrocardiogram signal segment and morphology.

[0024] Figure 3 Abnormal ECG signal segments and morphologies provided for embodiments of the present invention: (a) indicates ST segment depression, (b) indicates ST segment elevation, and (c) indicates abnormal T wave height;

[0025] Figure 4 A flowchart of a hexagon recognition algorithm based on electrocardiogram morphology provided in an embodiment of the present invention;

[0026] Figure 5 A hexagonal result diagram provided for an embodiment of the present invention;

[0027] Figure 6 A framework diagram of the CNN-LSTM model provided in the embodiments of the present invention; Detailed Implementation

[0028] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, the present invention provides an example of a portable intelligent electrocardiogram monitoring system based on an edge-cloud architecture, including: a portable 12-lead Holter electrocardiogram monitor, a smartphone running Android 12.0, and a PC equipped with a 12th generation i7 processor and an Intel 3060 graphics card.

[0030] The terminal device, consisting of a portable 12-lead Holter ECG monitor, has the main function of continuously transmitting the user's real-time ECG signal to the edge device via Bluetooth.

[0031] The edge devices consist of smartphones running Android 12.0: Edge devices are mostly composed of a single smartphone, but smartwatches and similar devices can also perform the same function. Figure 1 As can be seen, the edge device is divided into two modules: an identification module and a display module.

[0032] In the recognition module, the edge device first receives real-time raw ECG signals via Bluetooth and stores them locally. Then, it selects 2-minute segments of raw ECG signals for data analysis. The first step of data analysis is preprocessing and feature extraction. Preprocessing mainly includes filtering and QRS detection. Filtering is used to eliminate noise generated during ECG monitoring, and QRS detection is used to facilitate feature extraction. Once the R wave is located, the S and T waves are identified based on their position. The relative positions of the S and T waves with the baseline, as well as the relative positions of the T and R waves, are calculated to complete the feature extraction process. The second step of data analysis involves anomaly prediction. The extracted features are input into a hexagonal recognition algorithm based on ECG morphology. The algorithm outputs a hexagonal image with six types, only one of which is normal, and the rest are abnormal, allowing users to intuitively observe their health status. If the ECG signal is classified as abnormal, it is sent to the cloud platform for further analysis. Conversely, if it is classified as normal, the 2-minute ECG segment is deleted from local storage to prevent memory overflow.

[0033] In the display module, the edge device ultimately obtains the hexagonal recognition result of the ECG segment from the hexagonal recognition algorithm based on ECG morphology, while the cloud platform provides a seven-classification result based on the CNN-LSTM model. If the hexagonal image shows an anomaly, and the feedback result of the CNN-LSTM model is MI (Miscarriage of Injury), an alert will be immediately sent to emergency contacts, along with the location of the MI.

[0034] The cloud device, comprised of PCs equipped with 12th-generation i7 processors and Intel 3060 graphics cards, consists of two servers, Cloud Server A and Cloud Server B, each powered by a single PC. The cloud device employs a CNN-LSTM-based MI recognition model pre-trained using the PTB ECG diagnostic database. Cloud Server A's primary function is to store historical ECG signals deemed abnormal by the edge device and utilize these signals to continuously train the model, update its parameters, and optimize the CNN-LSTM-based multi-class MI recognition model. Cloud Server B's primary function is to receive real-time ECG signals deemed abnormal by the edge device via 5G and analyze the data. First, a two-minute ECG segment is segmented into 1024-bit samples for feature extraction. Then, these samples are fed into the trained CNN-LSTM model for recognition. If more than three ECG samples are identified as MI within two minutes, the patient is diagnosed with an MI attack, and the recognition result and MI location are fed back to the edge device.

[0035] like Figure 2 and Figure 3 As shown, both the hexagon recognition algorithm based on electrocardiogram morphology and the recognition model based on CNN-LSTM use electrocardiogram signals as training samples. Therefore, the electrocardiogram signal is described as follows: a complete heartbeat includes a P wave, QRS complex, T wave, and U wave, as shown in the diagram. Figure 2 As shown. The morphology, amplitude, and duration of the P wave provide information about the functional state of the atria. The QRS complex is the largest band on the electrocardiogram, representing the current of ventricular depolarization. The T wave and U wave represent ventricular repolarization, i.e., the electrical signals that restore the ventricular muscle to its initial state. When myocardial infarction (MI) occurs, there will be some obvious changes on the electrocardiogram, such as ST segment depression, ST segment elevation, abnormal T wave height, and T wave inversion, as shown. Figure 3 As shown in the figure. Therefore, these ECG regions can be extracted as features for the identification of MI.

[0036] Preprocessing, including filtering and QRS detection, is often required before using ECG signals. Filtering is crucial because ECG signals acquired by portable ECG monitors often contain significant noise, such as baseline drift, power line interference, and electromyographic interference. Baseline drift is caused by poor electrode contact and low-frequency interference signals due to respiration, which can lead to abnormal amplitude calculations for various waves. Power line interference, primarily generated by the power supply equipment, significantly impacts the signal, causing overall waveform blurring. Electromyographic interference originates from muscle contractions and tremors, manifesting as irregularly rapidly changing waveforms. The presence of this noise severely affects the accuracy of the analysis; therefore, filtering is essential before using ECG signals. Since ECG signals are typically concentrated in the low-frequency range, this invention employs a fourth-order Butterworth bandpass filter to effectively filter out ECG signal noise below 250Hz and above 2000Hz, improving data quality and the reliability of experimental results.

[0037] The QRS detection method utilizes the morphology, amplitude, and location of the R-peak to provide crucial information about cardiac health. For example, the height and shape of the R-peak can be influenced by the health of the ventricular muscle, myocardial hypertrophy, or dilation. Furthermore, the occurrence and interval of the R-peak can be used to calculate heart rate and assess the normality or abnormality of heart rhythm. Therefore, locating the R-peak and performing feature selection based on its location is beneficial for MI identification. This invention employs the Pan-Tompkins algorithm for R-peak location. The two main operations of the Pan-Tompkins algorithm are R-peak enhancement and R-peak decision. R-peak enhancement strengthens the features of the R-peak, while R-peak decision selects the R-peak from noise.

[0038] like Figure 4 and Figure 5 As shown, the hexagonal algorithm based on electrocardiogram morphology is an electrocardiogram (ECG) recognition algorithm proposed in this invention, involving ECG morphological features and image analysis. Its main advantages are simple structure and fast computation speed. The algorithm mainly includes three parts: preprocessing, feature extraction, and classification prediction, as illustrated in the flowchart below. Figure 4 As shown.

[0039] Preprocessing: Preprocessing consists of three parts: filtering, QRS detection, and S-wave and T-wave localization. The specific processes of filtering and QRS detection have been discussed above. The following describes how to use the R-wave to locate the S-wave and T-wave. First, the ECG dataset is normalized to the first quadrant, then:

[0040] (1)

[0041] (2)

[0042] Where S i T represents the position of the i-th S-wave.i R represents the position of the i-th T wave. i Let represent the position of the i-th R-wave, j represent the j-th position after the R-wave (j=1,2,3,...,60), k represent the k-th position after the R-wave (k=101,102,103,...,300), ReP(p,q) represent the relative position of p and q, and BL represent the baseline.

[0043] Feature extraction: When myocardial infarction (MI) occurs, corresponding changes will appear on the electrocardiogram (ECG). The ST segment may be elevated or depressed, and the T wave height may be abnormal or inverted; therefore, studying these locations has important medical value. Based on the positions of the R wave, S wave, and T wave, their relative positions to the baseline are calculated.

[0044] (3)

[0045] (4)

[0046] (5)

[0047] Where ReP(BL,S) i ReP(T) represents the relative position of the baseline and the i-th S-wave. i ReP(T,BL) represents the relative position of the i-th T wave with respect to the baseline. i , R i ) represents the relative position of the i-th T wave and the i-th R wave, and Voltage() represents the voltage at that position. The other symbols are referenced in formula (1-2).

[0048] Classification prediction:

[0049] Algorithm 1: Electrocardiographic Morphological Judgment Criteria <![CDATA[Input: R i , S i , T i , m (m represents the number of R waves, S waves, and T waves that appear within 2 minutes).]]> for i in m-1 do: <![CDATA[if ReP(T i , R i )>0:]]> An abnormal ECG segment was sent to the cloud, and image 7(a) was returned. <![CDATA[if ReP(T i ,BL)<0:<!-- 4 --> ]]> An abnormal ECG segment was sent to the cloud, and image 7(b) was returned. <![CDATA[if ReP(BL,S i )<0.3:]]> The abnormal ECG segment was sent to the cloud and image 7(c) was returned. <![CDATA[if ReP(T i , R i )>0 and ReP(BL,S i )<0.3:]]> An abnormal ECG segment was sent to the cloud, and image 7(d) was returned. <![CDATA[if ReP(T i ,BL)<0 and ReP(BL,S i )<0.3:]]> The abnormal ECG segment was sent to the cloud, and image 7(e) was returned. None of the above conditions are met: The ECG segment appeared normal; see image 7(f).

[0050] Depend on Figure 5 It can be seen that S' and T' represent abnormal S and T waves, while S and T represent normal S and T waves. Among them, 5(a) indicates a highly abnormal T wave, corresponding to... Figure 3 (c); Figure 5 (b) indicates T-wave inversion, corresponding to Figure 3 (d); Figure 5 (c) indicates an ST segment anomaly, corresponding to Figure 3 (a) or Figure 3 (b); Figure 5 (d) indicates ST segment abnormalities and T wave height abnormalities; Figure 5 (e) indicates ST segment abnormality and T wave inversion; Figure 5 (f) indicates normal operation. When the display module displays... Figure 5 When (ae) is displayed, it indicates that the edge device has determined that the ECG segment is abnormal. Figure 5(f) indicates that the condition is normal.

[0051] like Figure 6 As shown, electrocardiogram (ECG) signals are one-dimensional data that gradually change over time. This invention utilizes a CNN-LSTM model to capture local features and long-term dependencies in the ECG signal sequence. By leveraging both local and global feature information, it improves the model's ability to distinguish between seven types of ECG signals: normal, IMI, ASMI, ILMI, AMI, ALMI, and IPLMI.

[0052] Preprocessing and Feature Extraction: The ECG segments received by this model have already been filtered by the edge device, so no further preprocessing is required. The 12-lead ECG signal is segmented with a sample length of 1024 to form (12, 1024) sample points, thereby completing feature extraction.

[0053] CNN-LSTM Model: The structure of the CNN-LSTM model is as follows Figure 6 As shown, the CNN part consists of one-dimensional convolutional layers, max pooling layers, and dropout layers. The convolutional layers are used to extract local features from the ECG signal, the max pooling layers are used for feature enhancement, and the dropout layers are used to prevent overfitting. The LSTM part consists of three bidirectional LSTM layers with 256, 128, and 64 nodes respectively. Finally, a Flatten layer is used for dimensionality reduction, and a fully connected layer is used for classification.

[0054] In summary, this invention proposes a portable intelligent electrocardiogram (ECG) monitoring system based on an edge-cloud architecture, aiming to achieve long-term monitoring of microcardiac dysregulation (MI) using edge devices and cloud servers. The system employs a hexagonal recognition algorithm based on ECG morphology on the edge device, which is simpler in model and faster in computation, alleviating the computational limitations and energy consumption issues of the edge device. Furthermore, a more accurate and powerful deep learning model is deployed on the cloud server, capable of accurately identifying abnormal information and locating the occurrence of MI. The effectiveness and reliability of both algorithms are verified on the publicly available PTB ECG diagnostic dataset, demonstrating that the system is fully functional and performs well.

[0055] The portable intelligent ECG monitoring system based on an edge-cloud architecture provided by this invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. It should be noted that those skilled in the art can make several improvements and modifications to this invention without departing from the principles of this invention, and these improvements and modifications also fall within the protection scope of the claims of this invention.

Claims

1. A portable intelligent electrocardiogram monitoring system based on an edge-cloud architecture, characterized in that, The system includes terminal devices, edge devices, and cloud devices: The terminal device consists of a self-developed portable 12-lead holter; The edge device consists of a smartphone or smartwatch; The cloud device consists of cloud server A and cloud server B provided by the PC; The terminal device is used to collect the user's electrocardiogram signal and send it to the edge device via Bluetooth; The edge device is used to detect whether there are abnormal signals from the user in real time and send the abnormal signals to the cloud device, which then feeds back the identification results to the user. The cloud device is used to more accurately identify abnormal signals and provide users with precise health reports; Specifically, it involves the following steps: Step 1: The terminal device first collects the user's electrocardiogram signal in real time and sends it to the edge device via Bluetooth; Step 2: The edge device consists of two modules: an identification module and a display module. In the identification module, a hexagonal recognition algorithm based on electrocardiogram (ECG) morphology is proposed, involving ECG morphological features and image analysis, including preprocessing, feature extraction, and classification prediction. This algorithm collects morphological features of ECG signals as samples, including the relative positions of the S wave and baseline, the T wave and baseline, and the T wave and R wave. It determines whether the signal is abnormal based on ECG morphological criteria and displays it to the user in the form of a hexagonal image, including abnormal T wave height, T wave inversion, ST segment abnormality, ST segment abnormality and T wave height abnormality, ST segment abnormality and T wave inversion, and normal. In the display module, the device can obtain binary classification results from the ECG morphology-based hexagonal recognition method, as well as multi-classification results from cloud device feedback, ultimately providing the user with a health report. If danger occurs, it will send an alert to emergency contacts. Step 3: The cloud device consists of server A and server B. A CNN-LSTM fusion model is deployed in the device. Server A continuously trains the CNN-LSTM model and optimizes the parameters using historically stored abnormal ECG segments. Server B receives real-time abnormal ECG segments and uses the CNN-LSTM model to detect the patient's condition and provide feedback.

2. The portable intelligent ECG monitoring system based on an edge-cloud architecture according to claim 1, characterized in that, The CNN-LSTM model: This model combines CNN and LSTM to effectively extract local features from input data and capture temporal correlations, thereby accurately identifying myocardial infarction and its location. It is a seven-classification model.

Images