Cloud side end-based home heart real-time monitoring and risk prediction system and method

Abstract

The invention belongs to the field of cloud end collaboration, and relates to a home heart real-time monitoring and risk prediction system and method based on a cloud end. Multi-modal heart health data collected and uploaded by an end side device at an edge side is subjected to preprocessing and abnormal data monitoring, and then the data is transmitted to a cloud end for multi-modal fusion so as to predict the heart health risk. And real-time response of abnormal conditions and accurate prediction of heart health risks are realized based on nearby deployment of the edge server and strong computing power of the cloud. Meanwhile, the method can continuously optimize the data monitoring model and the risk prediction model according to the real-time feedback of the user, the prediction result and the actual monitoring data, provides a personalized adaptive adjustment mechanism for the user, and improves the prediction accuracy.

Description

Technical Field

[0001] This invention relates to the field of cloud-edge-device collaboration, and in particular to a home-based real-time cardiac monitoring and risk prediction system and method based on cloud-edge-device collaboration. Background Technology

[0002] Heart disease is one of the leading causes of death worldwide, with its high incidence and mortality rates posing a serious threat to human health and imposing a heavy economic and psychological burden on society and families. Health monitoring and risk prediction for heart disease patients can help detect cardiac health abnormalities early and intervene in a timely manner, effectively reducing the incidence and mortality of heart disease. Traditional monitoring methods rely on centralized hospital monitoring or periodic physical examinations, which cannot achieve 24 / 7 real-time monitoring. In a smart home environment, advanced monitoring and prediction technologies can be used to achieve real-time management of user health. This not only allows for the timely detection of potential health risks but also effectively alleviates the pressure on medical resources. Therefore, designing a home-based real-time heart health monitoring and risk prediction method based on a cloud-edge-device architecture is of great significance.

[0003] In recent years, some medical assistance systems based on sensors and cloud computing have emerged in the market. However, most existing systems rely on local terminal computing or are entirely dependent on cloud computing. Local terminal computing has limited capabilities, making it difficult to perform complex data analysis and resulting in low prediction accuracy. Cloud computing, on the other hand, suffers from high latency and bandwidth limitations, making real-time response difficult. Furthermore, most systems have significant limitations in real-time data monitoring and health prediction: the former relies on fixed thresholds to identify anomalies in collected data, lacking adaptive adjustment mechanisms for individual user differences, which easily leads to false alarms and reduces system reliability; the latter is based on single data analysis, which can only capture local features of the user's health, making it difficult to infer the overall condition, resulting in poor ability to identify potential health risks, high false negative rates, and missed opportunities for early intervention. Summary of the Invention

[0004] This invention provides a cloud-edge-device-based home-based real-time cardiac monitoring and risk prediction system and method. Employing a cloud-edge-device collaborative architecture, it leverages the proximity deployment of edge servers and the powerful computing capabilities of the cloud to achieve real-time response to abnormal situations and accurate prediction of cardiac health risks. Furthermore, the proposed real-time cardiac health detection method reduces false alarm rates and effectively improves system reliability by dynamically updating parameter thresholds and introducing duration thresholds. The proposed cardiac health risk prediction method overcomes the limitations of prediction based on single data by fusing and analyzing multimodal cardiac health data, thus reducing both false alarm and false negative rates.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A home-based real-time cardiac monitoring and risk prediction system based on cloud, edge, and device, including a device layer, an edge server layer, and a cloud server layer;

[0007] The device layer includes data acquisition devices and user terminal devices; the data acquisition devices collect and transmit multimodal cardiac health data from users; the user terminal devices display and broadcast cardiac health data and provide data query functions.

[0008] The edge server layer includes a data preprocessing module and an abnormal data monitoring module; the data preprocessing module receives and processes the collected multimodal cardiac health data, and sends it to the abnormal data monitoring module for real-time monitoring;

[0009] The cloud server layer includes a data storage module, a risk prediction module, and a model update module; the data storage module receives and stores cardiac health data, the risk prediction module predicts the user's cardiac health risk, and the model update module dynamically updates the abnormal data monitoring module and the risk prediction module.

[0010] The specific steps of the cloud-edge-device-based home-based real-time cardiac monitoring and risk prediction method using the above system are as follows:

[0011] S1 uses data acquisition equipment to collect multimodal cardiac health data, including digital data such as heart rate, blood pressure and body temperature, ECG signal waveform data and user image data, and sends them to the edge server layer.

[0012] S2, the collected multimodal cardiac health data is input into the corresponding modality data processing unit for data preprocessing and unified into a structured format. Then, the structured cardiac health data is sent to the abnormal data monitoring module for real-time monitoring.

[0013] S3 uses a threshold comparison method to monitor the preprocessed structured data in real time and upload it to the cloud server layer.

[0014] S4, the cloud server layer receives structured data uploaded by the edge server layer and stores it in the data storage module;

[0015] S5, the cloud server layer predicts users' heart health risks based on long-term stored structured heart health data, and sends early warning information in real time based on the prediction results, while also sending health reports to user terminals regularly.

[0016] S6, based on user feedback, prediction results and actual monitoring data, optimizes the abnormal data monitoring model and the heart health risk prediction model, and redeploys them to the abnormal data monitoring module of the edge server layer and the risk prediction module of the cloud server layer.

[0017] Compared with existing technologies, the cloud-edge-device architecture and prediction method for real-time monitoring of heart health in smart homes proposed in this invention have the following significant advantages:

[0018] (1) By introducing a cloud-edge-device collaborative architecture, this invention achieves efficient collaboration in data collection, processing, real-time monitoring and risk prediction. Compared to relying entirely on cloud computing, relying on edge servers deployed nearby to preprocess multimodal data and monitor abnormal data can reduce transmission latency and network bandwidth usage, and improve real-time performance. Compared to local terminal computing, relying on the stronger computing power of edge servers and the cloud can improve the accuracy of data monitoring and risk prediction, and reduce false alarm rate and false negative rate.

[0019] (2) This invention introduces a duration threshold combined with a parameter threshold to monitor abnormal data through threshold comparison, which can effectively reduce the false alarm rate and improve the reliability of the system. At the same time, it uses three devices—wearable sensors, portable devices with electrodes, and caustic cameras—to collect multimodal data and fuse the multimodal data to predict heart health risks, which solves the problem of high false alarm and missed alarm rates in the prior art due to ignoring the relationship between different health data or relying on only a single data for prediction.

[0020] (3) Based on real-time user feedback, prediction results and actual monitoring data, this invention dynamically optimizes the relevant parameters of the edge-side abnormal data monitoring model and the cloud-based prediction model, providing users with a personalized adaptive adjustment mechanism while improving prediction accuracy.

[0021] (4) This invention is designed for smart home environments. Compared with traditional centralized monitoring or periodic physical examinations, it can realize real-time monitoring of heart health data around the clock, effectively reduce the false alarm rate, avoid missing sudden situations, and alleviate the burden on medical institutions. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the home-based real-time heart monitoring and risk prediction system with a cloud-edge-device architecture according to a specific embodiment of the present invention.

[0023] Figure 2 This is a flowchart of a home-based real-time cardiac monitoring and risk prediction method according to a specific embodiment of the present invention.

[0024] Figure 3 This is a flowchart illustrating data preprocessing and real-time monitoring at the edge server layer in a specific embodiment of the present invention.

[0025] Figure 4 This is a flowchart illustrating the implementation of cardiac health risk prediction at the cloud server layer in a specific embodiment of the present invention. Detailed Implementation

[0026] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0027] Reference Figure 1 This embodiment provides a cloud-edge-device architecture for smart home environments that supports real-time monitoring and risk prediction of heart health, including a device layer 101 (device side), an edge server layer 102, and a cloud server layer 103.

[0028] The equipment layer 101 includes:

[0029] Data acquisition equipment: Specifically, it includes a data acquisition module and a data transmission module. The data acquisition module is used to collect the user's multimodal cardiac health data, including heart rate, blood pressure, body temperature, electrocardiogram (ECG) signals and monitoring images, and transmits the collected multimodal data to the data transmission module; the data transmission module is used to receive the multimodal data transmitted by the data acquisition module and send it to the edge server via a wireless network.

[0030] User terminal equipment: Specifically, it includes a visualization module, a voice broadcast module, a query request module, and a data receiving module. The visualization module displays alarm or warning information, health reports, and historical health data to the user on the terminal device, while providing a feedback window to collect user feedback on alarm or warning information and upload it to the cloud in real time. The voice broadcast module broadcasts received warning or alarm information, reminding users of any abnormalities in their current cardiac health or potential health risks. The query request module sends historical data query requests to the cloud storage via the cellular mobile network. The data receiving module receives historical health data, health reports, and warning information downloaded from the cloud, as well as alarm information transmitted from the edge, and then transmits it to the visualization module to visualize the received information or data for the user, and to the voice broadcast module to broadcast alarm or warning information.

[0031] Edge server layer 102 includes:

[0032] The data preprocessing module is divided into a digital data processing unit, a waveform data processing unit, and an image data processing unit. These units perform data preprocessing and feature extraction on digital data (heart rate, blood pressure, and body temperature), waveform data (ECG signals), and image data (monitoring images), respectively. The data is then uniformly formatted and sent to the abnormal data monitoring module for abnormal data monitoring.

[0033] The abnormal data monitoring module uses a threshold comparison method to monitor preprocessed formatted data in real time. The formatted data is then uploaded to the cloud via a fiber optic network for further analysis. The threshold comparison method involves the edge server promptly transmitting an alarm to the user terminal device via a wireless network when one or more cardiac health data points exceed their corresponding parameter thresholds for a set duration threshold. Here, the parameter thresholds refer to the upper and lower limits of various data points for maintaining a healthy heart state, while the duration threshold is the maximum allowed duration for a particular cardiac health data point to exceed its corresponding parameter threshold.

[0034] Cloud server layer 103 includes:

[0035] Data storage module: Used to receive and store formatted data preprocessed by the edge server. It is divided into a historical data area and a real-time data area, which store historical data and real-time data respectively.

[0036] Risk prediction module: Used to predict the user's heart health risk and periodically (weekly or daily, which can be set according to user needs) send health reports to the user terminal and send timely warning information to the user terminal based on the predicted heart health risk index.

[0037] Model update module: This module is used to dynamically update and optimize the abnormal data monitoring model and the cardiac health risk prediction model based on user feedback, prediction results, and actual monitoring data, and then redeploy them to the edge server and cloud server, respectively.

[0038] like Figure 2 As shown, this embodiment provides a method for real-time home-based cardiac health monitoring and risk prediction based on a cloud-edge-device architecture. The specific steps are as follows:

[0039] S1 uses data acquisition equipment to collect multimodal cardiac health data and sends it to an edge server via a wireless network.

[0040] The data acquisition equipment includes wearable sensors, portable devices with electrodes, and a caustic camera. The sensors collect digital data such as heart rate, blood pressure, and body temperature; the electrodes collect ECG signal waveform data; and the caustic camera collects user image data to prepare for subsequent extraction of the user's blood oxygen saturation and respiratory rate. Furthermore, it utilizes facial information from the optically blurred images of the caustic camera to protect user privacy.

[0041] S2 inputs the collected multimodal cardiac health data into the corresponding modality's data processing unit for data preprocessing and unification into a structured format. Then, the structured data is sent to the abnormal data monitoring module for real-time monitoring.

[0042] like Figure 3As shown, the data preprocessing module includes a digital data processing unit, a waveform data processing unit, and an image data processing unit, with the corresponding processing flows as follows:

[0043] Digital data processing unit: Heart rate, blood pressure, and body temperature data directly collected by sensors are cleaned through filtering and interpolation techniques, relevant features are extracted, and the data is collected in the form of time series data, with readings taken periodically. Specifically, a fixed window of 5 sampling points is set, the mean and standard deviation of the data within the window are calculated, outliers deviating from the mean by more than 3 times the standard deviation are removed, and the filtered data gaps are filled by constructing a linear relationship between the valid data points before and after the missing segment. After data cleaning, relevant features are extracted, the data is collected in the form of time series data, and readings are taken periodically.

[0044] Waveform data processing unit: The continuous cardiac activity electrical signals collected by the electrodes are formed into waveform signals. The electrical signals are then processed by signal filtering, baseline correction and QRS complex monitoring to extract features, thereby obtaining data on respiratory rate interval, QRS complex width and heart rate variability.

[0045] Specifically, continuous cardiac activity electrical signals collected by electrodes form waveform signals. These signals are filtered using a 0.5–35 Hz bandpass filter to select the main frequency components of the ECG signal, eliminating 50 Hz power line interference and electromyographic noise. A sliding window of 50 ms is then set, and the filtered signal is processed by moving average. The baseline is corrected by the mean signal within the window to eliminate baseline shift caused by low-frequency drift. Subsequently, a threshold of 0.4 times the peak signal value is set, and the signal is scanned to identify waveform segments exceeding the threshold. The start point (the position where the rising edge of the signal first exceeds the threshold) and end point (the position where the falling edge of the signal falls back below the threshold) of the QRS complex are located, thus completing accurate monitoring of the QRS complex. The respiratory rate interval can be calculated based on the time interval between adjacent QRS complexes, and the rhythm variability index is extracted through the interval sequence. The QRS complex width can be obtained from the start and end times of a single QRS complex.

[0046] Image data processing unit: The image uploaded after being captured by the caustic camera is divided into image blocks of different sizes using multi-scale image segmentation. To improve efficiency, blood oxygen saturation extraction selects images from fingertips, earlobes, and wrists—areas with good blood circulation and thin skin suitable for precise measurement—using smaller-scale blocks of 6×6, 12×12, 24×24, and 48×48 pixels. The pulsatile region of interest is located using live skin monitoring technology. The average pixel value of each small block is calculated frame-by-frame to obtain the time series x[n]. This time series x[n] is multiplied by the window function ω[n] within a sliding window of length L, and a power spectrum is generated using Fast Fourier Transform (FFT).

[0047]

[0048] n represents the sampling point number (the nth sampling point within the sliding window), n = 0, 1, 2, ..., L-1; k is the discrete frequency index, k = 0, 1, 2, ..., L-1, corresponding to the kth frequency component. f s This indicates the sampling frequency; it can be understood as one being the sequence number in the timing sequence and the other being the sequence number in the frequency sequence.

[0049] The power spectrum is normalized and sorted to construct a one-dimensional feature vector that reflects the stability of the pulse peak. Based on this value, it is determined whether a pulse is present. The small blocks that are determined to contain a pulse are merged in space to obtain the final region of interest for pulsatility.

[0050] A spatial redundancy signal extraction architecture is used to extract photoplethysmography (PPG) signals, and Fourier transform (FFT) is performed on the PPG signals in blocks to generate power spectra. The pixel average value is calculated in the time dimension for each pulsatile region of interest (ROI) to form a region-level time-series signal; DC normalization is then performed to eliminate illumination offset; a sliding window FFT is then used to obtain the power spectrum of each window, and the signal-to-noise ratio (SNR) is calculated. The image blocks with the highest SNR from all blocks are selected, and their signals are synthesized into the final global PPG signal.

[0051] Finally, blood oxygen saturation was extracted using a bandpass filter at [0.6, 2] Hz; respiratory rate extraction used larger-scale blocks of 48×48, 72×72, 96×96 and 120×120 pixels to cover the whole-body image. The initial operation steps were the same as for blood oxygen saturation extraction. After extracting the PPG signal and fusing the high-quality image block signal, the respiratory rate was obtained using a bandpass filter at [0.15, 1] ​​Hz.

[0052] The feature data, after being preprocessed by the corresponding modality processing unit, is standardized into a structured format by adding a unified timestamp, marking the data source, and performing unit internationalization operations, and is then transmitted to the abnormal data monitoring module for real-time monitoring.

[0053] S3 uses a threshold comparison method to monitor the preprocessed structured data in real time and uploads it to the cloud server via a fiber optic network.

[0054] like Figure 3As shown, the threshold comparison method compares various cardiac health data with corresponding thresholds and determines whether to send an alarm message to the user terminal based on the comparison results: if all cardiac health data do not exceed the set parameter thresholds or the duration of exceeding the parameter thresholds is less than T, no alarm is triggered; otherwise, an alarm message is sent to the user terminal via the wireless network. The duration threshold T is related to the acquisition period t, T = αt, where α represents a constant coefficient. For example, if t = 15s, the corresponding α for body temperature data is 40, and the upper limit of the parameter threshold is 37.5℃ and the lower limit is 36℃, then T = 15s × 40 = 10min, indicating that the upper limit of the allowed duration for body temperature data to exceed 37.5℃ or fall below 36℃ is 10min. Both the parameter thresholds and T can be personalized according to the characteristics of different physiological parameters and can be optimized and updated by the cloud model update module.

[0055] The specific thresholds for various cardiac health data are as follows: Body temperature: upper limit 37.5℃, lower limit 36℃, α=40; Heart rate: lower limit 50bpM, upper limit 120bpM, α=1; Blood pressure: lower limit diastolic blood pressure <90mmHg, systolic blood pressure <60mmHg, upper limit diastolic blood pressure >160mmHg, systolic blood pressure >100mmHg, α=2; Respiratory rate interval: lower limit 400ms, upper limit 1200ms, α=1; QRS complex width: lower limit none, upper limit 120ms, α=1; Heart rate variability: lower limit 50ms, upper limit none, α=240; Blood oxygen saturation: lower limit <90%, upper limit none, α=1; Respiratory rate: lower limit 8 breaths / min, upper limit >30 breaths / min, α=1.

[0056] S4 receives structured data uploaded from the edge server in the cloud and stores it in the data storage module.

[0057] The data storage module deploys a data receiving interface to receive structured data uploaded from edge servers and store it within the module. The data storage module employs a distributed storage architecture, divided into a real-time data area and a historical data area. The real-time data area uses the in-memory database Redis to store the structured data for the most recent 72 hours, facilitating retrieval during cloud-based predictive model training and inference. The historical data area utilizes the distributed file system HDFS combined with the relational database MySQL for long-term storage, and optimizes retrieval efficiency for large-scale historical data through data sharding and indexing.

[0058] The S5 cloud server predicts users' heart health risks based on long-term stored structured heart health data and sends early warning information in a timely manner based on the prediction results, while also sending health reports to user terminals on a regular basis.

[0059] like Figure 4As shown, cardiac health risk prediction includes cardiac health data prediction based on CNN and cardiac health risk index prediction based on multimodal data fusion analysis. The specific steps are as follows:

[0060] Step S51, the process of establishing a CNN-based cardiac health data prediction model is as follows: historical cardiac health data stored in the cloud is imported into the CNN for training. The input consists of 8 feature sequences composed of time-series cardiac health data collected at all hourly times over 3 consecutive days. in ,b in ,c in ,d in ,e in ,f in ,g in ,h in > These correspond to heart rate, blood pressure, body temperature, respiratory rate interval, QRS complex width, heart rate variability, blood oxygen saturation, and respiratory rate, respectively. Convolutional operations are performed on these feature sequences in a feature fusion layer to extract the linkage features of different parameters along the time axis. Through linear transformation and nonlinear activation, the multi-parameter fusion features extracted at each time step are mapped to specific parameter prediction values. Finally, the predicted sequences for all hourly times corresponding to the eight cardiac health data points on day 4 are output. out ,b out ,c out ,d out ,e out ,f out ,g out ,h out >.

[0061] The constructed CNN network structure consists of an input layer, two one-dimensional convolutional layers followed by pooling layers, two fully connected layers, and an output layer. The input layer receives an input tensor composed of eight sets of temporal cardiac health feature sequences. The two convolutional layers have kernels of size 3 and stride of 1, using 32 and 64 kernels respectively. Each convolutional layer is followed by a max pooling layer with a window size of 2 and a stride of 2 to downsample the features along the time axis. The two fully connected layers use the ReLU activation function to perform nonlinear mapping on the high-dimensional features after convolution and pooling. The output layer consists of eight neurons, each corresponding to a predicted value of one of the eight cardiac health data points.

[0062] In step S52, the cloud server layer continuously supplements the latest structured cardiac health data uploaded by the edge server layer and stored in the data storage module. Referring to step S51, the prediction model is continuously trained using the latest data to ensure the accuracy of the prediction results. The CNN prediction model, after long-term training, can output predictions of the user's cardiac health data for the next 24 hours when inputting cardiac health data from approximately 72 hourly intervals.

[0063] Step S53: Set the prediction period to 6 hours, meaning that heart health data prediction is performed every 6 hours. The data storage module inputs 8 sets of structured data at the hour within the past 72 hours into the CNN prediction model to form a time-series feature sequence. After passing through the CNN network, it outputs the prediction sequence H of the 8 data items at the hour within the next 24 hours. i ={h i,1 ,h i,2 ,...h i,24}(i = 1, 2, ..., 8, corresponding to the above 8 health data items, h i,j (This represents the predicted data for the i-th type of health data at the j-th integer time).

[0064] Step S54: Based on the user's personalized characteristics, customize the initial weights x of each data point in the prediction of heart health risks. i (i = 1, 2, ..., 8, ...) (Corresponding to the above 8 types of health data respectively). If the user suffers from palpitations, then x1 = 0.22, x2 = 0.1, x3 = 0.05, x4 = 0.15, x5 = 0.08, x6 = 0.25, x7 = 0.08, x8 = 0.07 to reflect the importance of heart rate and heart rate variability to the user's cardiac health. The 8 sets of H data output in step S53... i Normalization And according to x i Perform fusion and splicing to build a unified high-level fusion representation. Simultaneously, an attention mechanism is introduced to adaptively adjust the weights of various data points in risk prediction based on their correlation with different characteristics and cardiac health risks. Specifically, H... i ' As the "key" and "value" in the attention mechanism, a learnable query vector q is set, and the correlation between sample features and cardiac health risk is learned, and the correlation score s between the two is calculated. i =q T ·H i ′, dynamically adjust the weight of various data in risk prediction To enhance the contribution of key features to the prediction of cardiac health risks.

[0065] Step S55: The fused feature K is input into the output layer consisting of a fully connected network and a sigmoid activation function to predict heart health risk. The fully connected network performs a mapping from "high-dimensional feature to risk index" on K, converting the complex feature vector into an initial predicted value z. The sigmoid activation function then normalizes z to [0,1], outputting the user's heart health risk index. Determine the heart health risk level based on y and perform the corresponding action: When When it is determined to be low risk, no warning is sent; when When the risk level is determined to be medium, an orange alert is sent to the user's terminal; when When the risk is deemed high, a red alert is sent to the user's terminal.

[0066] S6 cloud servers optimize the abnormal data monitoring model and the heart health risk prediction model based on user feedback, prediction results and actual monitoring data, and redeploy them to edge servers and cloud servers.

[0067] Model optimization refers to comparing real-time collected data, user feedback, and prediction results uploaded from the cloud and edge servers, continuously optimizing and adjusting the relevant parameters of the abnormal data monitoring model and the heart health risk prediction model based on the error, and then redeploying them to the edge servers and the cloud.

[0068] The specific steps of the model optimization process are as follows:

[0069] S61, the cloud records the predicted values ​​output by the model at each time point. (For the detection model, abnormal data is detected.) on the contrary For prediction models, (The predicted risk index) and the corresponding actual label value y obtained based on user feedback. t (For the detection model, in the event of a cardiac abnormality, y) t =1, otherwise y t =0; For the predictive model, the user has cardiac risk, y t =1, otherwise y t =0), and calculate the error. The error dataset D for the anomaly monitoring model and risk prediction model is formed by arranging the data in chronological order. e and D c .

[0070] S62, based on D e and D c and the corresponding parameter vector θ e ={alarm thresholds for each parameter} and θ c = {CNN convolutional layer weights, convolutional layer bias parameters, pooling stride, fully connected layer weights, fully connected layer bias parameters, and weight matrices of the query vector, key vector, and value vector in the attention mechanism} are retrained separately, and a loss function with user weight factors is constructed. And through optimization algorithms, the updated parameter vector θ is made... e and θ c It minimizes errors in user data.

[0071] S63, will update θe and θ c The new parameters are distributed to the edge server and the user terminal respectively, so that the user's anomaly monitoring model and heart health risk prediction model can automatically adopt the new parameter configuration to achieve model optimization.

[0072] User feedback is obtained after the user selects an option through a feedback pop-up window: After the user terminal receives and visualizes the alarm or warning information, it will ask the user about the accuracy of the monitoring or prediction results through a feedback pop-up window containing "accurate" and "inaccurate" options. Alarm and warning information feedback pop-ups will appear 1 hour and 24 hours after the user terminal receives the alarm information, respectively. The user can then select based on the actual results. Figure 1 As shown, the terminal device transmits this feedback information to the cloud in real time via the cellular mobile network.

[0073] The above description is only a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. Any equivalent modifications or changes made by those skilled in the art based on the content disclosed in the present invention should be included within the scope of protection set forth in the claims.

Claims

1. A home-based real-time cardiac monitoring and risk prediction system based on cloud-edge computing, characterized in that: The system includes a device layer, an edge server layer, and a cloud server layer; The device layer includes data acquisition devices and user terminal devices; the data acquisition devices collect and transmit multimodal cardiac health data from users; the user terminal devices display and broadcast cardiac health data and provide data query functions. The edge server layer includes a data preprocessing module and an abnormal data monitoring module; the data preprocessing module receives and processes the collected multimodal cardiac health data, and sends it to the abnormal data monitoring module for real-time monitoring; The cloud server layer includes a data storage module, a risk prediction module, and a model update module; the data storage module receives and stores cardiac health data, the risk prediction module predicts the user's cardiac health risk, and the model update module dynamically updates the abnormal data monitoring module and the risk prediction module.

2. The home-based real-time cardiac monitoring and risk prediction system based on cloud-edge computing as described in claim 1, characterized in that: In the device layer The data acquisition device includes a data acquisition module and a data transmission module; The data acquisition module is used to collect the user's multimodal cardiac health data, including heart rate, blood pressure, body temperature, ECG signals and monitoring images, and transmit the collected multimodal data to the data transmission module; the data transmission module is used to receive the multimodal data transmitted by the data acquisition module and send it to the edge server layer via a wireless network; The user terminal device includes a visualization module, a voice broadcasting module, a query request module, and a data receiving module; The visualization module is used to display alarm or warning information, health reports and historical health data to the user on the user terminal device, and provides a feedback window to collect user feedback on alarm or warning information and upload it to the cloud in real time; the voice broadcast module is used to broadcast the received warning or alarm information by voice, reminding the user that there is an abnormality in the current heart health status or potential health risks. The query request module is used to send historical data query requests to the cloud storage via the cellular mobile network. The data receiving module is used to receive historical health data, health reports and early warning information from the cloud, as well as alarm information transmitted from the side, and then transmit them to the visualization module to visualize the received information or data and the voice broadcast module to broadcast the alarm or early warning information.

3. The home-based real-time cardiac monitoring and risk prediction system based on cloud-edge computing as described in claim 1, characterized in that: In the edge server layer The data preprocessing module includes a digital data processing unit, a waveform data processing unit, and an image data processing unit. These units perform data preprocessing and feature extraction on the digital data, waveform data, and image data, respectively, and then send the data to the abnormal data monitoring module in a structured format for abnormal data monitoring. Subsequently, the preprocessed cardiac health formatted data is uploaded to the cloud server layer for storage and further analysis via a fiber optic network. The abnormal data monitoring module monitors the preprocessed structured data in real time using a threshold comparison method. The threshold comparison method refers to the method whereby when the monitoring detects that the duration of cardiac health data exceeding the corresponding parameter threshold reaches a set duration threshold, the edge server promptly transmits alarm information to the user terminal device via a wireless network. Here, the parameter threshold refers to the upper and lower limits of various data when the heart maintains a healthy state, and the duration threshold refers to the maximum time limit for cardiac health data to exceed the corresponding parameter threshold.

4. The home-based real-time cardiac monitoring and risk prediction system based on cloud-edge computing as described in claim 1, characterized in that: In the cloud server layer The data storage module is used to receive and store formatted data processed by the data preprocessing module in the edge server layer; The risk prediction module is used to predict the user's heart health risk and periodically send health reports to the user terminal, as well as send timely warning information to the user terminal based on the predicted heart health risk index. The model update module is used to dynamically update the abnormal data monitoring model and the heart health risk prediction model based on user feedback, prediction results and actual monitoring data, and deploy the updated models to the abnormal data monitoring module at the edge server layer and the risk prediction module at the cloud server layer, respectively.

5. The method for a home-based real-time cardiac monitoring and risk prediction system based on cloud-edge computing according to any one of claims 1-4, characterized in that: The specific steps of the method are as follows: S1 uses data acquisition equipment to collect multimodal cardiac health data, including digital data such as heart rate, blood pressure and body temperature, ECG signal waveform data and user image data, and sends them to the edge server layer. S2, the collected multimodal cardiac health data is input into the corresponding modality data processing unit for data preprocessing and unified into a structured format. Then, the structured cardiac health data is sent to the abnormal data monitoring module for real-time monitoring. S3 uses a threshold comparison method to monitor the preprocessed structured data in real time and upload it to the cloud server layer. S4, the cloud server layer receives structured data uploaded by the edge server layer and stores it in the data storage module; S5, the cloud server layer predicts users' heart health risks based on long-term stored structured heart health data, and sends early warning information in real time based on the prediction results, while also sending health reports to user terminals regularly. S6, based on user feedback, prediction results and actual monitoring data, optimizes the abnormal data monitoring model and the heart health risk prediction model, and redeploys them to the abnormal data monitoring module of the edge server layer and the risk prediction module of the cloud server layer.

6. The method for real-time home cardiac monitoring and risk prediction based on cloud-edge computing as described in claim 5, characterized in that: In step S2, the computational resources and processing latency required for preprocessing data of different modalities vary. Image data requires more computational resources and has a longer processing latency compared to digital data. Depending on the acquisition device, data of different modalities are input into the corresponding data preprocessing module for preprocessing. The specific steps are as follows: S21, the data preprocessing module includes a digital data processing unit, a waveform data processing unit, and an image data processing unit, with the corresponding processing flows as follows: Digital data processing unit: Heart rate, blood pressure and body temperature data directly collected by the sensors are cleaned by interpolation and filtering technology, relevant features are extracted, collected in the form of time series data, and readings are read periodically; Waveform data processing unit: The continuous cardiac activity electrical signals collected by the electrodes are used to form waveform signals. The electrical signals are then processed by signal filtering, baseline correction and QRS complex monitoring to extract features, thereby obtaining data on respiratory rate interval, QRS complex width and heart rate variability. Image data processing unit: The images uploaded after being captured by the caustic camera are divided into image blocks of different sizes through multi-scale image segmentation. For blood oxygen saturation extraction, images of fingertips, earlobes, and wrists with good blood circulation and thin skin, suitable for precise measurement, are selected. The scale blocks are divided into 6×6, 12×12, 24×24, and 48×48 pixels. The pulsatile region of interest is located using live skin monitoring technology. The photoplethysmography (PPG) signal is extracted using a spatial redundancy signal extraction architecture. The PPG signal is divided into blocks and Fourier transform is performed to generate the power spectrum. The optimal image block is selected for signal fusion based on the signal-to-noise ratio. Finally, blood oxygen saturation is extracted by bandpass filtering at [0.6, 2] Hz. For respiratory rate extraction, scale blocks of 48×48, 72×72, 96×96, and 120×120 pixels are used to cover the whole body image. After extracting the PPG signal and fusing the high-quality image block signal, the respiratory rate is obtained by bandpass filtering at [0.15, 1] ​​Hz. The feature data, after being preprocessed by the corresponding modality processing unit, is standardized into a structured format by adding a unified timestamp, marking the data source, and performing unit internationalization operations. The structured data is then transmitted to the abnormal data monitoring module for real-time monitoring.

7. The method for real-time home cardiac monitoring and risk prediction based on cloud-edge computing as described in claim 5, characterized in that: In step S3, the threshold comparison method compares each cardiac health data point with its corresponding parameter threshold and determines whether to send an alarm message to the user terminal based on the comparison result: if all cardiac health data points do not exceed the set parameter threshold, no alarm is triggered; otherwise, an alarm message is sent to the user terminal via the wireless network. A duration threshold T is introduced, representing the upper limit of the allowed duration for cardiac health data to exceed the corresponding parameter threshold. The alarm mechanism is triggered only when the duration for which the health data exceeds the corresponding parameter threshold reaches the preset T. The duration threshold is related to the acquisition period t, T = αt, where α represents a constant coefficient. It is set individually according to the characteristics of different physiological parameters and can be optimized and updated by the model update module in the cloud.

8. The method for real-time home cardiac monitoring and risk prediction based on cloud-edge computing as described in claim 5, characterized in that: In step S4, a data receiving interface is deployed in the cloud to receive structured data uploaded by the edge server and store it in the data storage module. The risk prediction module and the user terminal obtain historical health data by accessing the data storage module, specifically including: The data storage module adopts a distributed storage architecture, divided into a real-time data area and a historical data area. The real-time data area uses the in-memory database Redis to store nearly 72 hours of structured data. The historical data area is based on the distributed file system HDFS combined with the relational database MySQL to achieve long-term storage, and establishes a data sharding and index optimization mechanism.

9. The method for real-time home cardiac monitoring and risk prediction based on cloud-edge computing as described in claim 5, characterized in that: In step S5, cardiac health risk prediction includes cardiac health data prediction based on convolutional neural networks (CNNs) and cardiac health risk index prediction based on multimodal data fusion analysis. The specific steps are as follows: Step S51, the process of establishing a CNN-based cardiac health data prediction model is as follows: historical cardiac health data stored in the cloud is imported into the CNN for training. The input consists of 8 sets of feature sequences , which are time-series cardiac health data collected at all times over multiple consecutive days. in ,b in ,c in ,d in ,e in ,f in ,g in ,h in > corresponds to heart rate, blood pressure, body temperature, respiratory rate interval, QRS complex width, heart rate variability, blood oxygen saturation, and respiratory rate, respectively. These feature sequences are convolved in a feature fusion layer to extract the linkage features of different parameters along the time axis. Through linear transformation and nonlinear activation, the multi-parameter fusion features extracted at each time step are mapped to specific parameter prediction values. The final output is a predicted sequence corresponding to eight cardiac health data points. out ,b out ,c out ,d out ,e out ,f out ,g out ,h out >; In step S52, the cloud server layer continuously receives the latest structured heart health data uploaded by the edge server layer and stored in the data storage module. Referring to step S51, the prediction model is continuously trained in combination with the latest data to ensure the accuracy of the prediction results. Step S53: Set the prediction period, that is, perform heart health data prediction once every prediction period. The data storage module inputs 8 sets of structured data into the CNN prediction model to form a time-series feature sequence, and outputs 8 data future prediction sequences H through the CNN network. i ={h i,1 ,h i,2 ,...h i,24 }, i = 1, 2, ..., 8, corresponding to 8 health data items, h i,j This represents the predicted data for the i-th type of health data at the j-th integer time. Step S54: Based on user-personalized characteristics, customize the initial weights x of various data points in the prediction of heart health risks. i i = 1, 2, ..., 8 Corresponding to 8 different health data types, the 8 sets of H output in step S53 are... i Normalization It also performs fusion and splicing to build a unified high-level fusion representation. Simultaneously, an attention mechanism and a learnable query vector q are introduced. By learning the correlation between sample features and cardiac health risks, the correlation score s between the two is calculated. i =q T ·H i ′, dynamically adjust the weight of each data point in risk prediction Strengthen the contribution of key features to the prediction of cardiac health risks; Step S55: The fused feature K is input to the output layer consisting of a fully connected network and a sigmoid activation function to output the heart health risk index. The fully connected network performs a high-dimensional feature mapping to the risk index on K, converting the complex feature vector into an initial predicted value z, and then normalizing z to [0,1] using the sigmoid activation function, outputting the user's heart health risk index y, where 0 < y ≤ 10. The heart health risk level is determined based on y and the corresponding operation is performed: when 0 < y ≤ 2, it is determined to be low risk and no warning is issued; when 2 < y ≤ 6, it is determined to be medium risk and an orange warning is sent to the user terminal; when y > 6, it is determined to be high risk and a red warning is sent to the user terminal.

10. The method for real-time home cardiac monitoring and risk prediction based on cloud-edge computing as described in claim 5, characterized in that: In step S6, the user terminal receives and visualizes the alarm or warning information to the user, and provides a feedback pop-up window with two options: accurate and inaccurate, asking the user about the accuracy of the monitoring or prediction results. The alarm and warning information feedback pop-ups will appear 1 hour and 24 hours after the user terminal receives the information, respectively. After the user selects according to the actual results, the terminal immediately transmits the feedback information to the cloud. After receiving the user feedback, the cloud compares the real-time collected data uploaded by the edge server with the prediction results, continuously optimizes and adjusts the relevant parameters of the abnormal data monitoring model and the heart health risk prediction model based on the error, and redeploys them to the edge server layer and the cloud server layer.

Images