Cardiovascular disease multi-parameter joint monitoring method and system

Abstract

The invention discloses a cardiovascular disease multi-parameter joint monitoring method and system, and relates to the technical field of data processing. Comprising the following steps: monitoring multiple physiological parameters related to cardiovascular health of a patient by using wearable equipment to obtain monitoring data; preprocessing the monitoring data to obtain preprocessed monitoring data; respectively extracting time domain features, frequency domain features, nonlinear features and other statistical features from the preprocessed monitoring data to obtain multi-modal features; based on the multi-modal features, in combination with self-supervised learning of deep learning, a graph neural network and time sequence analysis, performing relation modeling, and performing data enhancement by using an adversarial generative network to obtain enhanced processing features; training a model by using marked normal and abnormal data by using feature importance scores and a random forest; and classifying real-time monitoring data by using the trained model, identifying normal and abnormal states, and obtaining an abnormal monitoring result.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a method and system for multi-parameter joint monitoring of cardiovascular diseases. Background Art

[0002] Cardiovascular disease (CVD) is one of the most lethal diseases worldwide, particularly among the elderly. CVD monitoring and data processing technologies face significant challenges worldwide. Currently, cardiovascular data analysis relies primarily on traditional methods such as time-domain and frequency-domain analysis. While these methods are effective to a certain extent, they exhibit numerous limitations in practical applications.

[0003] Cardiovascular data is characterized by its nonlinearity and non-stationarity, meaning that the data pattern is not a simple linear relationship, but rather changes over time and under different conditions. This complexity poses significant challenges to data analysis. First, the data is easily interfered with by various noises, such as motion artifacts of the patient during monitoring and electromagnetic interference in the environment. These noises can mask the true physiological signals and reduce data quality. Second, there are significant differences in physiological parameters between different individuals. These differences may lead to insufficient generalization capabilities of the analysis model, making it difficult to adapt to the characteristics and state changes of different individuals. In addition, in terms of feature extraction, traditional methods have difficulty extracting highly discriminative features from these complex physiological signals. Due to the dynamic nature and complexity of the cardiovascular system, simple statistical features often cannot accurately reflect the true state of the cardiovascular state, which requires more advanced feature extraction techniques to capture subtle changes in the signal.

[0004] Inadequate data integration is also a problem with traditional cardiovascular data processing methods. Cardiovascular health monitoring often involves multiple physiological parameters, such as blood pressure, electrocardiogram (ECG), and blood oxygen saturation (SpO2). However, these parameters are typically analyzed independently, lacking an effective integration strategy to comprehensively assess a patient's overall health. This fragmented analysis approach ignores the interactions and inherent connections between these parameters, resulting in insufficient accuracy and effectiveness of the analysis results. Summary of the Invention

[0005] In order to overcome the above-mentioned existing shortcomings, the main purpose of the present invention is to provide a multi-parameter joint monitoring method and system for cardiovascular disease, which can provide comprehensive support for the early diagnosis of cardiovascular disease, personalized adjustment of treatment plans and patient health management through real-time monitoring of multiple physiological parameters, combined with data analysis and intelligent early warning.

[0006] To achieve the above object, the present invention adopts the following technical solution, a method for multi-parameter joint monitoring of cardiovascular disease, comprising the following steps:

[0007] Acquire multiple physiological parameters related to cardiovascular health of the patient, perform synchronization preprocessing on the physiological parameters using dynamic time warping combined with Kalman filtering, and obtain preprocessed monitoring data;

[0008] Extracting time domain features, frequency domain features, nonlinear features and morphological features from the preprocessed monitoring data, and constructing a multimodal feature vector through feature concatenation and dimensionality reduction;

[0009] Physiological parameters are used as graph nodes, each corresponding to a multimodal feature vector. Edge weights are obtained using a preset sliding window size and step size. A masked self-supervised pre-training strategy is used to learn implicit associations between parameters. A combined temporal convolutional network is used to capture multi-scale temporal dependencies, and a spatiotemporal graph neural network dynamic relationship model is constructed. A generative adversarial network is embedded in the spatiotemporal graph neural network dynamic relationship model for data enhancement, thereby obtaining enhanced processing features.

[0010] Based on the SHAP value, combined with the enhanced processing features, the feature importance score is obtained, and combined with the stratified random forest, the trained two-layer random forest model is trained using the labeled normal and abnormal data to obtain a trained two-layer random forest model;

[0011] The trained two-layer random forest model is used to classify the real-time monitoring data, identify normal and abnormal states, and obtain abnormal monitoring results.

[0012] The method of obtaining the pre-processed monitoring data comprises the following steps:

[0013] Collect the patient's physiological parameter data through various sensors and monitoring devices to obtain the original physiological parameter time series data;

[0014] performing data cleaning and normalization on the original physiological parameter time series data to obtain normalized physiological parameter data;

[0015] Performing data alignment on the normalized physiological parameter data using dynamic time warping to obtain aligned physiological parameter time series data;

[0016] Performing denoising and smoothing processing on the aligned physiological parameter time series data using low-pass filtering and Kalman filtering to obtain denoised and smoothed physiological parameter data;

[0017] Missing value processing is performed on the denoised and smoothed physiological parameter data to obtain synchronized preprocessed physiological parameter data as preprocessed monitoring data.

[0018] The obtaining of the enhanced processing feature comprises the following steps:

[0019] Each physiological parameter is used as a node in the graph, and each node corresponds to a feature vector containing multimodal features. Through the preset sliding window size and step size, the mutual information between adjacent physiological parameter data points is obtained and used as the edge weight between nodes to construct a spatiotemporal graph neural network.

[0020] A masked self-supervised pre-training strategy is used to randomly mask 20% of the node features. The masked features are predicted by the graph autoencoder, the implicit connections between nodes are extracted, and the weights of the spatiotemporal graph neural network are optimized to obtain the optimized graph neural network. The graph autoencoder contains a two-layer graph convolutional network (GCN) with a hidden layer dimension of 64.

[0021] The optimized graph neural network performs spatial convolution on the node features to obtain a spatial feature sequence;

[0022] Design a temporal convolutional network, input the spatial feature sequence into the designed temporal convolutional network, combine the temporal convolutional network to capture multi-scale temporal dependencies, build a spatiotemporal graph neural network dynamic relationship model, and obtain temporal and spatial features by stacking graph convolution layers and temporal convolutional network layers in the spatiotemporal graph neural network dynamic relationship model;

[0023] Using conditional Wasserstein GAN as the adversarial generative network, the generator and discriminator are designed to be embedded in the dynamic relationship model of the spatiotemporal graph neural network. The obtained temporal and spatial features are enhanced in feature space. The discriminator constrains the generator to generate feature data that conforms to the clinical distribution, and then fuses it with the original feature data to obtain enhanced processing features.

[0024] The method of training a machine learning model using feature importance scoring and a two-layer random forest includes the following steps:

[0025] Based on the SHAP value, the enhanced processing features are ranked by importance using a feature importance scoring method, and features whose importance is greater than a set threshold are selected as the filtered feature subset;

[0026] The filtered feature subset is used as the input feature of the first layer of the double-layer random forest to obtain an initial value of the abnormal probability. The initial value of the abnormal probability is used as the sample weight of the second layer of the forest. The weighted Gini coefficient is used and an adaptive weighting mechanism is introduced to adjust the tree node splitting of the second layer of the random forest to obtain a trained double-layer random forest model.

[0027] The trained two-layer random forest model is trained using the labeled normal and abnormal data to obtain a trained two-layer random forest model.

[0028] The trained model is used to classify real-time monitoring data and identify normal and abnormal states, including:

[0029] Input the real-time monitoring data into the trained two-layer random forest model;

[0030] According to the output results of the trained two-layer random forest model, the status of the real-time monitoring data is judged and abnormal monitoring results are generated.

[0031] A detection device for a multi-parameter combined monitoring method for cardiovascular disease, comprising:

[0032] A data acquisition module is used to collect physiological parameter data related to cardiovascular health, obtain multiple physiological parameters related to cardiovascular health of patients, synchronize and pre-process the physiological parameters using dynamic time warping combined with Kalman filtering, and obtain pre-processed monitoring data;

[0033] The data processing module is used to preprocess, extract features, model relationships, and enhance data for the collected data; extract time domain features, frequency domain features, nonlinear features, and morphological features from the preprocessed monitoring data, and construct a multimodal feature vector through feature splicing and dimensionality reduction; use physiological parameters as graph nodes, nodes correspond to multimodal feature vectors, edge weights are obtained through preset sliding window size and step size, and a masked self-supervised pre-training strategy is used to learn implicit associations between parameters. A joint time convolutional network is used to capture multi-scale temporal dependencies, and a spatiotemporal graph neural network dynamic relationship model is constructed. The adversarial generative network is embedded in the spatiotemporal graph neural network dynamic relationship model for data enhancement to obtain enhanced processing features;

[0034] The model training module is used to train a two-layer random forest model. Based on the SHAP value, it combines enhanced processing features to obtain feature importance scores. In addition, it combines the stratified random forest and uses the labeled normal and abnormal data to train the trained two-layer random forest model to obtain a trained two-layer random forest model.

[0035] The anomaly detection module is used to classify real-time monitoring data using a trained two-layer random forest model, identify normal and abnormal states, and obtain abnormal monitoring results;

[0036] The early warning module is used to judge the abnormal monitoring results obtained according to the set threshold. When the judgment result is not satisfied with the set threshold, the early warning mechanism is automatically triggered and a reminder message is sent to the patient, doctor and family members.

[0037] Compared with existing technologies, the present invention offers the following advantages: By utilizing wearable devices to monitor multiple physiological parameters related to cardiovascular health in real time, combined with the extraction of time-domain features, frequency-domain features, nonlinear features, and other statistical features, the present invention can comprehensively capture multidimensional information about a patient's cardiovascular health. This multi-parameter monitoring helps improve diagnostic accuracy, provides more detailed and comprehensive health data support, and enables earlier detection of potential abnormalities and timely warnings. Furthermore, through denoising methods such as low-pass filtering and Kalman filtering, as well as data preprocessing methods such as normalization and standardization, noise can be effectively removed from the data, reducing external interference with data collection and ensuring the accuracy and consistency of the final monitored data. These processing measures help improve the quality of subsequent analysis and modeling, reducing misjudgments caused by data quality issues. Data augmentation, combining self-supervised learning, graph neural networks, and generative adversarial networks in deep learning, can effectively increase the diversity of model training, making the model more robust and generalizable. Enhanced data not only helps improve model recognition accuracy but also prevents overfitting caused by data scarcity or imbalance. The self-supervised learning model extracts features and reduces dimensionality from monitoring data, reducing manual intervention and reliance on empirical knowledge, thereby improving the model's automation and efficiency. Furthermore, using feature importance scoring and random forest training methods, it can identify the most informative features from a large number of features, further optimizing model performance and recognition accuracy. By inputting and classifying real-time data and utilizing a trained two-layer random forest model to promptly determine the status of monitoring data, real-time monitoring of patient health is possible.

[0038] The real-time nature of the detection method proposed in the present invention is of great significance for early detection of abnormalities and intervention, especially in cardiovascular diseases, where early identification of abnormal conditions is crucial for reducing risks and improving treatment outcomes. The method can perform personalized health status assessments and abnormality warnings for patients based on real-time analysis of multi-parameter data. This has great practical value for patients with chronic diseases, especially those at high risk of cardiovascular disease. Through personalized monitoring and real-time feedback, patients can better adjust their lifestyles and avoid worsening of health problems. Furthermore, through automated data collection and analysis, combined with deep learning and machine learning technologies, the method can reduce the workload of doctors, improve diagnostic efficiency, avoid repeated and unnecessary examinations, and ultimately help reduce medical costs. In summary, the cardiovascular disease monitoring system based on the multi-parameter joint monitoring method can effectively improve the quality of monitoring data, enhance the reliability of monitoring results, and provide more accurate disease warning and intervention strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings described herein are used to provide further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute improper limitations on the present application.

[0040] Figure 1 It is a schematic diagram of the process structure of the present invention;

[0041] Figure 2 is a schematic diagram of a process for obtaining enhanced features according to the present invention;

[0042] Figure 3 It is a schematic diagram of the pretreatment process of the present invention. DETAILED DESCRIPTION

[0043] Cardiovascular data is characterized by its nonlinearity and non-stationarity, meaning that the data pattern is not a simple linear relationship, but rather changes over time and under different conditions. This complexity poses significant challenges to data analysis. First, the data is easily interfered with by various noises, such as motion artifacts of the patient during monitoring and electromagnetic interference in the environment. These noises can mask the true physiological signals and reduce data quality. Second, there are significant differences in physiological parameters between different individuals. These differences may lead to insufficient generalization capabilities of the analysis model, making it difficult to adapt to the characteristics and state changes of different individuals. In addition, in terms of feature extraction, traditional methods have difficulty extracting highly discriminative features from these complex physiological signals. Due to the dynamic nature and complexity of the cardiovascular system, simple statistical features often cannot accurately reflect the true state of the cardiovascular state, which requires more advanced feature extraction techniques to capture subtle changes in the signal.

[0044] Inadequate data integration is also a problem with traditional cardiovascular data processing methods. Cardiovascular health monitoring often involves multiple physiological parameters, such as blood pressure, electrocardiogram (ECG), and blood oxygen saturation (SpO2). However, these parameters are typically analyzed independently, lacking an effective integration strategy to comprehensively assess a patient's overall health. This fragmented analysis approach ignores the interactions and inherent connections between these parameters, resulting in insufficient accuracy and effectiveness of the analysis results.

[0045] In order to overcome the above-mentioned existing shortcomings, the main purpose of the present invention is to provide a multi-parameter joint monitoring method and system for cardiovascular disease, which can provide comprehensive support for the early diagnosis of cardiovascular disease, personalized adjustment of treatment plans and patient health management through real-time monitoring of multiple physiological parameters, combined with data analysis and intelligent early warning.

[0046] To achieve the above object, the present invention adopts the following technical solutions: a multi-parameter combined monitoring method for cardiovascular disease, see Figure 1 , including the following steps:

[0047] Acquire multiple physiological parameters related to cardiovascular health of the patient, perform synchronization preprocessing on the physiological parameters using dynamic time warping combined with Kalman filtering, and obtain preprocessed monitoring data;

[0048] Extracting time domain features, frequency domain features, nonlinear features and morphological features from the preprocessed monitoring data, and constructing a multimodal feature vector through feature concatenation and dimensionality reduction;

[0049] Physiological parameters are used as graph nodes, each corresponding to a multimodal feature vector. Edge weights are obtained using a preset sliding window size and step size. A masked self-supervised pre-training strategy is used to learn implicit associations between parameters. A combined temporal convolutional network is used to capture multi-scale temporal dependencies, and a spatiotemporal graph neural network dynamic relationship model is constructed. A generative adversarial network is embedded in the spatiotemporal graph neural network dynamic relationship model for data enhancement, thereby obtaining enhanced processing features.

[0050] Based on the SHAP value, combined with the enhanced processing features, the feature importance score is obtained, and combined with the stratified random forest, the trained two-layer random forest model is trained using the labeled normal and abnormal data to obtain a trained two-layer random forest model;

[0051] The trained two-layer random forest model is used to classify the real-time monitoring data, identify normal and abnormal states, and obtain abnormal monitoring results.

[0052] The technical solution of the present invention includes the following key points:

[0053] Multi-parameter real-time monitoring: Real-time monitoring of physiological parameters related to cardiovascular health through multiple sensors, including but not limited to electrocardiogram (ECG), blood pressure, blood oxygen saturation (SpO2), heart rate variability (HRV), body temperature, etc.

[0054] Data processing and comprehensive analysis: upload monitoring data to the data analysis platform via wireless transmission, conduct comprehensive analysis of multiple physiological parameters, and perform intelligent evaluation of abnormal data.

[0055] Early warning mechanism: Based on data analysis results, when a parameter is detected to be abnormal, the system automatically triggers the early warning mechanism and pushes reminder information to patients, doctors or family members.

[0056] Personalized health management: Provide users with personalized health management plans and treatment recommendations based on analysis results, and interact with doctors remotely through the platform.

[0057] Remote monitoring and diagnosis and treatment support: Doctors can remotely view patients' monitoring data, track their health status in a timely manner, and adjust treatment plans based on data changes.

[0058] This method is applicable to hospitals, telemedicine monitoring platforms, health management systems, and other fields, and is designed to accurately monitor and identify the health status of patients with cardiovascular disease in real time. Patients may experience fluctuations in various physiological parameters (such as heart rate, blood pressure, and blood oxygen saturation) in different health management scenarios. Timely diagnosis and prediction of abnormal events are crucial for preventing cardiovascular events.

[0059] The cardiovascular disease multi-parameter joint monitoring method and system of the present invention includes the following main steps and modules:

[0060] Use multiple physiological parameter monitoring devices such as smart watches, wristbands, chest patches, etc. to simultaneously monitor multiple physiological parameters related to cardiovascular health. Physiological parameters include electrocardiogram (ECG), blood pressure, blood oxygen saturation (SpO2), heart rate variability (HRV), body temperature, etc. All monitoring devices transmit data to the data management platform in real time via wireless network or Bluetooth. All physiological data is uploaded to the cloud data storage platform through smart terminal devices (such as smartphones, tablets, etc.). The data platform stores and manages data based on different user accounts, and users can view historical data and health reports at any time.

[0061] The data analysis platform preprocesses real-time data, including noise reduction and standardization, to ensure data accuracy. Combining machine learning with big data analysis techniques, it analyzes multiple parameters to assess the user's cardiovascular health. Based on the data analysis results, a health report is generated, providing a comprehensive assessment of the patient's current health status.

[0062] When a monitored parameter changes beyond a set threshold, the system automatically issues a warning signal. This warning signal is sent to the patient, doctor, or family member via push notification, text message, or email. The system can also provide targeted emergency advice based on the type of abnormal data, such as abnormal blood pressure or heart rate fluctuations.

[0063] Based on comprehensive analysis results, the system provides each user with a personalized health management plan, including advice on diet, exercise, medication, etc. Users can remotely interact with doctors through smart terminals to receive personalized treatment plan adjustments.

[0064] Doctors can view patients' monitoring data in real time through a dedicated platform and adjust treatment plans based on data changes. Doctors can also provide remote medical support through video conferencing, voice calls, and other means.

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] Example:

[0067] A patient is at high risk for cardiovascular disease. The system uses devices such as smart watches and blood pressure monitors to monitor multiple physiological parameters in real time, including electrocardiogram (ECG), blood pressure, and blood oxygen levels. All patient monitoring data is uploaded to the cloud platform via wireless network, where it is stored and processed. Figure 1-Figure 3 Doctors can monitor patients' health status in real time through the data analysis platform. If abnormal blood pressure fluctuations are detected, the system immediately sends an alert, prompting both the patient and the doctor to take appropriate emergency measures. The system also provides personalized health management plans, advising patients to reduce salt intake, increase aerobic exercise, and seek regular follow-up visits.

[0068] To use wearable devices to simultaneously monitor multiple physiological parameters related to cardiovascular health of patients, the first step is to select wearable devices: wearable devices with high sensitivity, stability and reliability, such as smart watches, health monitoring bracelets, etc.

[0069] Then determine the monitoring parameters. Wearable devices should at least include the following physiological parameters monitoring functions:

[0070] a) Electrocardiogram (ECG): monitors the heart's electrical activity;

[0071] b) Blood pressure: monitor systolic and diastolic blood pressure;

[0072] c) Heart rate: monitor the number of heart beats per minute;

[0073] d) Blood oxygen saturation (SpO2): monitors the oxygen saturation in the blood;

[0074] e) Respiratory rate: monitor the number of breaths per minute;

[0075] f) Temperature: monitor body surface temperature;

[0076] g) Activity level: monitor daily activities and exercise levels.

[0077] Wearable devices continuously monitor the above physiological parameters during patients' daily activities to ensure the real-time and continuity of data.

[0078] The monitoring data is transmitted to the data management platform in real time. Before data transmission, the monitoring data is encrypted to ensure patient privacy and security. The encrypted data is transmitted to the data management platform in real time via wireless transmission technologies such as Bluetooth, Wi-Fi, and mobile networks. This ensures data synchronization between the wearable device and the data management platform to avoid data loss or delay.

[0079] Comprehensively analyze the monitoring data to generate multi-parameter detection results, including the following:

[0080] The collected physiological data needs to be cleaned through denoising and normalization steps. First, Kalman filtering and low-pass filtering are applied to remove device noise and high-frequency interference. Then, dynamic time warping (DTW) is used to align the data collected by different sensors. This ensures that all physiological parameters are synchronized on the time axis. Missing data is interpolated to fill in the gaps, resulting in preprocessed monitoring data.

[0081] From the cleaned and preprocessed data, extract time domain features such as mean, standard deviation, maximum value, etc., frequency domain features such as frequency components, nonlinear features such as Lyapunov exponents, and morphological features such as electrocardiogram waveform features.

[0082] By concatenating the eigenvectors of each physiological parameter and combining them with PCA (Principal Component Analysis) for dimensionality reduction, a single eigenvector containing multimodal information is constructed. Each physiological parameter is represented as a node, and the connections (edges) between nodes are defined by the mutual information between adjacent data points.

[0083] A spatiotemporal graph neural network was constructed to capture the dynamic spatiotemporal relationships between physiological parameters. Using a masked self-supervised pre-training strategy, 20% of node features were randomly masked. A graph autoencoder was then used to predict the masked features and learn the implicit associations between nodes. The graph autoencoder used a two-layer graph convolutional network (GCN) with a hidden layer dimension of 64.

[0084] The output of the spatiotemporal graph neural network is enhanced using a conditional Wasserstein GAN (generative adversarial network). The generator generates feature data that conforms to the clinical distribution, and the discriminator fuses the data generated by the constrained generator with the original feature data to enhance feature diversity.

[0085] The importance of the enhanced features is evaluated based on the Shapley weighted approach (SHAP). Each feature is scored based on its contribution to the model decision, and a subset of features with higher importance is selected.

[0086] After selecting the most important features, we used them as input features to train a two-layer random forest model. This model uses a weighted mechanism to optimize the classification process. The first layer calculates an initial value for anomaly probability, and the second layer adjusts the sample weights based on this initial value, ultimately training a highly accurate classification model.

[0087] Real-time monitoring data is collected by sensors and uploaded to the cloud platform. This data is then inferred using a trained two-layer random forest model to generate monitoring status. This two-layer random forest model can identify abnormal patient conditions such as arrhythmias and hypertension in real time.

[0088] When a patient's real-time heart rate fluctuates significantly and the blood oxygen level is lower than the critical value, the model will judge it as abnormal and trigger an early warning notification to alert medical staff.

[0089] The monitoring data of patient A in this embodiment are as follows, with time points from T0 to T3:

[0090] Heart rate (bpm): 60, 72, 65, 85;

[0091] Blood pressure (mmHg): Systolic pressure: 120, 125, 130, 135; Diastolic pressure: 80, 82, 84, 88; Blood oxygen saturation (%): 98, 97, 95, 94;

[0092] ECG: waveform data is uploaded in real time;

[0093] Through spatiotemporal graph neural network and enhanced generative feature processing, the final feature data includes: time series feature vector of heart rate, time domain feature vector of blood pressure, frequency domain feature of blood oxygen saturation and nonlinear feature of ECG waveform;

[0094] The abnormal prediction results output after processing by the two-layer random forest model: Patient A had an abnormal decrease in blood oxygen saturation at T3, and the model predicted an abnormal state.

[0095] Based on the monitoring data and the trained two-layer random forest model, the system identified abnormal conditions in real time during the monitoring of patient A, generated an alarm, and promptly notified medical staff for intervention.

[0096] The advantage of applying the method of the present invention is that it can integrate multiple physiological parameters and provide accurate health status monitoring and early warning through the combination of deep learning and machine learning.

[0097] Through this method, hospitals, telemedicine platforms, etc. can achieve real-time and efficient monitoring of patients with cardiovascular diseases, provide patients with personalized and accurate health management services, prevent acute events caused by cardiovascular diseases, and improve patients' quality of life and treatment effects.

[0098] Example 2

[0099] This example describes a specific application of a multi-parameter joint monitoring method for cardiovascular disease. It uses wearable devices to monitor multiple cardiovascular-related physiological parameters and uses a series of advanced technical means to process, extract features, model, and classify the data, achieving accurate monitoring of cardiovascular health status and abnormality identification. The following is an analysis of the technical means and technical effects:

[0100] Various sensors and monitoring devices collect data on patients' physiological parameters, including electrocardiogram (ECG), heart rate (HR), blood pressure (BP), and blood oxygen saturation (SpO2). These parameters are widely used in cardiovascular disease monitoring and can reflect different aspects of heart function.

[0101] The original physiological parameter time series data is cleaned to remove outliers and interference signals. The data is then normalized to ensure that different parameters are within a similar numerical range to facilitate subsequent analysis. Because the sampling frequencies and timestamps of different parameters may differ, dynamic time warping (DTW) technology is used to align the data. The aligned physiological parameter time series data are denoised and smoothed using low-pass filtering and Kalman filtering. Low-pass filtering can remove high-frequency noise, while Kalman filtering can effectively estimate the system state, filter interference signals from the noise, and improve data quality. Missing value processing is performed on the denoised and smoothed physiological parameter data, and appropriate interpolation methods (such as linear interpolation and spline interpolation) are used to fill in the missing data to obtain synchronized preprocessed physiological parameter data, which are used as preprocessed monitoring data.

[0102] Multimodal features are extracted from the preprocessed monitoring data, including time domain features, frequency domain features, nonlinear features, and morphological features. Time domain features analyze physiological signals on the time axis. Frequency domain features analyze physiological signals from a frequency domain perspective, reflecting the frequency components and energy distribution of the signal. Nonlinear features analyze physiological signals from a nonlinear perspective. Morphological features analyze the shape and structure of the signal waveform, reflecting the morphological properties of the signal. Morphological features are particularly useful for analyzing signals with distinct waveform characteristics, such as electrocardiograms (ECGs) and blood oxygen saturation (SpO2), and can help identify abnormal waveform patterns. These features reflect the complexity and dynamic nature of the signal, revealing its overall level, degree of fluctuation, and trend. Multimodal feature vectors are then constructed through feature concatenation and dimensionality reduction.

[0103] Physiological parameters are used as graph nodes, each corresponding to a multimodal feature vector. Edge weights are obtained using a preset sliding window size and step size. A masked self-supervised pre-training strategy is used to learn the implicit associations between parameters. A joint temporal convolutional network is used to capture multi-scale temporal dependencies. A dynamic relationship model of a spatiotemporal graph neural network is constructed, and a generative adversarial network is embedded in the model for data augmentation to obtain enhanced processing features. In the graph structure, each physiological parameter (such as electrocardiogram, heart rate, blood pressure, blood oxygen saturation, etc.) is represented as a node, and the node is characterized by its multimodal feature vector. The edge weights between nodes are obtained using a preset sliding window size and step size. Mutual information is usually used to measure the correlation between two physiological parameters. The graph structure can effectively capture the correlation and complementarity between different physiological parameters, providing structured information for subsequent feature learning and pattern recognition.

[0104] Masked self-supervised pre-training is an emerging graph representation learning method. It learns implicit connections between nodes by randomly masking some node features and then using a graph autoencoder to predict the masked features. This method employs a masked self-supervised pre-training strategy, randomly masking 20% of the node features and using a graph autoencoder to predict the masked features. The graph autoencoder consists of a two-layer graph convolutional network (GCN) with a hidden layer dimension of 64. This pre-training process learns implicit connections between nodes and optimizes the weights of the spatiotemporal graph neural network (GNN), resulting in an optimized GNN. The optimized GNN then performs spatial convolution on the node features to obtain a spatial feature sequence. Spatial convolution captures local structural information and feature associations between nodes, helping to extract meaningful spatial patterns. A temporal convolutional network (TCN) is designed to feed the spatial feature sequence into the TCN, which is then combined to capture multi-scale temporal dependencies. By stacking TCN layers in the dynamic relationship model of the spatiotemporal GNN, temporal and spatial features are obtained. Temporal convolutional networks can capture long-term dependencies and multi-scale features in time series data, helping to identify temporal patterns and trends. Using a conditional Wasserstein GAN as a generative adversarial network, the generator and discriminator are designed and embedded in a spatiotemporal graph neural network dynamic relationship model to perform feature space enhancement on the acquired temporal and spatial features. The discriminator constrains the generator to generate feature data that conforms to the clinical distribution and fuses it with the original feature data to obtain enhanced features. This data augmentation method can effectively increase the diversity of training data and improve the generalization and robustness of the model.

[0105] Atrial fibrillation (AFib) is a common cardiac arrhythmia characterized by irregular tremors of the atria, leading to irregular heartbeats. By monitoring parameters such as ECG, HR, BP, and SpO2, typical features of atrial fibrillation, such as irregular heart rhythms and increased heart rate variability, can be captured.

[0106] Early diagnosis and effective monitoring of atrial fibrillation are crucial for preventing serious complications such as stroke. Traditional monitoring methods often rely on single-parameter analysis, which struggles to capture the multidimensional characteristics of the disease. Multi-parameter combined monitoring can provide a more comprehensive assessment of patient health status, improving diagnostic accuracy and monitoring effectiveness.

[0107] In this example, the following data was collected from a 50-year-old male patient who was diagnosed with atrial fibrillation:

[0108] Table 1 Physiological data of fifty-year-old men

[0109]

[0110]

[0111] The data has been cleaned and normalized, and then dynamic time warping (DTW) combined with Kalman filtering is used to denoise and smooth the data. The processed data is as follows:

[0112] Table 2 Physiological parameters after pretreatment

[0113]

[0114]

[0115] The following features are extracted from each physiological parameter, including:

[0116] Time domain features: ECG features: average value: 0.2, standard deviation: 0.1, maximum value: 0.3, minimum value: 0.1; HR features: average value: 0.3, standard deviation: 0.1, maximum value: 0.4, minimum value: 0.2; BP features: average value: 0.4, standard deviation: 0.1, maximum value: 0.5, minimum value: 0.3; SpO2 features: average value: 0.5, standard deviation: 0.1, maximum value: 0.6, minimum value: 0.4;

[0117] Frequency domain feature extraction: ECG feature: main frequency component: 0.15, frequency band energy: 0.25; HR feature: main frequency component: 0.2, frequency band energy: 0.3; BP feature: main frequency component: 0.25, frequency band energy: 0.35; SpO2 feature: main frequency component: 0.3, frequency band energy: 0.4,

[0118] Nonlinear feature extraction, ECG feature: sample entropy: 0.4, approximate entropy: 0.5; HR feature: sample entropy: 0.5, approximate entropy: 0.6; BP feature: sample entropy: 0.55, approximate entropy: 0.65, SpO2 feature: sample entropy: 0.6, approximate entropy: 0.7;

[0119] Morphological feature extraction, ECG features: number of peaks: 3, number of troughs: 2, waveform area: 0.15; HR features: number of peaks: 2, number of troughs: 1, waveform area: 0.2; BP features: number of peaks: 2, number of troughs: 1, waveform area: 0.25; SpO2 features: number of peaks: 1, number of troughs: 0, waveform area: 0.3;

[0120] All extracted features are concatenated into a multimodal feature vector, which is then reduced using techniques such as principal component analysis (PCA) or t-SNE to obtain the reduced feature vector. The reduced feature vector is: [0.25, 0.3, 0.35, 0.4, 0.45, 0.5, 0.55, 0.6].

[0121] Physiological parameters are used as graph nodes, and each node corresponds to a feature vector containing multimodal features. Using a preset sliding window size and step size, the mutual information between adjacent physiological parameter data points is obtained and used as the edge weight between nodes to construct a spatiotemporal graph neural network.

[0122] The graph structure is as follows:

[0123] Node 1: ECG

[0124] Node 2: HR

[0125] Node 3: BP

[0126] Node 4: SpO2

[0127] The edge weights are as follows:

[0128] Node 1-Node 2: 0.6

[0129] Node 1-Node 3: 0.5

[0130] Node 1 - Node 4: 0.4

[0131] Node 2-Node 3: 0.7

[0132] Node 2 - Node 4: 0.5

[0133] Node 3-Node 4: 0.6

[0134] A masked self-supervised pre-training strategy was used, randomly masking 20% of the node features. The masked features were predicted by a graph autoencoder, extracting implicit connections between nodes. The weights of the spatiotemporal graph neural network were optimized to obtain the optimized graph neural network. In this example, the masked node features were the feature vectors of the HR. The graph autoencoder successfully predicted the masked features and optimized the weights of the graph neural network. The optimized graph neural network then performed spatial convolution on the node features to obtain a spatial feature sequence. The spatial feature sequence obtained in this example is: [0.28, 0.32, 0.36, 0.4].

[0135] A temporal convolutional network is designed, and the spatial feature sequence is input into the temporal convolutional network. Combined with the temporal convolutional network, it captures multi-scale temporal dependencies and constructs a spatiotemporal graph neural network dynamic relationship model. By stacking graph convolutional layers and temporal convolutional network layers in the spatiotemporal graph neural network dynamic relationship model, temporal and spatial features are obtained. In this example, the temporal and spatial features obtained are: [0.3, 0.35, 0.4, 0.45].

[0136] Using a conditional Wasserstein GAN as a generative adversarial network, we designed a generator and discriminator embedded in a spatiotemporal graph neural network dynamic relationship model. We then enhanced the temporal and spatial features obtained. The discriminator constrained the generator to generate feature data that matched the clinical distribution and fused it with the original feature data to obtain enhanced features. The enhanced features obtained in this example were: [0.32, 0.37, 0.42, 0.47].

[0137] Based on the SHAP value, the feature importance scoring method is used to rank the importance of the enhanced features, and the features whose importance is greater than the set threshold are selected as the filtered feature subset.

[0138] The feature importance scores are: feature 1: 0.25, feature 2: 0.3, feature 3: 0.28, feature 4: 0.32. The threshold is set to 0.3, and the filtered feature subsets are feature 2 and feature 4.

[0139] The filtered feature subset is used as the input features for the first layer of a two-layer random forest to obtain the initial value of the anomaly probability. The initial value of the anomaly probability is used as the sample weight for the second layer of the forest. The weighted Gini coefficient is used and an adaptive weighting mechanism is introduced to adjust the node splitting of the second layer of the random forest to obtain the trained two-layer random forest model.

[0140] The trained two-layer random forest model is trained using the labeled normal and abnormal data to obtain a trained two-layer random forest model.

[0141] The training data includes 100 normal samples and 50 abnormal samples. The model achieves 95% accuracy, 90% precision and 95% recall on the validation set.

[0142] The real-time monitoring data is input into the trained two-layer random forest model. According to the output of the model, the status of the real-time monitoring data is judged and abnormal monitoring results are generated.

[0143] The output is:

[0144] Time point 55: Normal

[0145] Time point 60: Normal

[0146] Time Point 65: Anomaly

[0147] Time Point 70: Abnormal

[0148] Time Point 75: Abnormal

[0149] Therefore, the system will issue an alarm at time points 65, 70 and 75 to remind medical personnel that the patient may be in an abnormal state.

[0150] Example 3:

[0151] Apply the multi-parameter combined monitoring method of cardiovascular disease to conduct real-time monitoring in patients with acute myocardial infarction.

[0152] A 65-year-old man with a history of coronary artery disease was admitted to the hospital with sudden chest pain. Monitoring equipment included a 12-lead electrocardiograph with a 500Hz sampling rate, a noninvasive blood pressure monitor with a 100Hz sampling rate, and a pulse oximeter with a 30Hz sampling rate.

[0153] The monitoring data was preprocessed, starting with signal alignment: cubic spline interpolation was performed on the blood oxygen signal to 500Hz. Dynamic time warping was used to align the ECG R wave peak with the rising edge of the blood pressure. The time deviation correction results are as follows:

[0154] Table 3 Correction results

[0155] Signal pair Maximum phase difference Post-alignment error ECG-blood pressure 120ms 28ms ECG-Blood Oxygen 300ms 42ms

[0156] Then Kalman filter denoising is performed, with process noise covariance Q = 1e-4I and observation noise covariance R = 0.01I. The denoising effect, taking the ST segment as an example:

[0157] Table 4 Denoising results

[0158] index Before filtering After filtering Signal-to-noise ratio (SNR) 32.7dB 47.2dB Standard deviation 0.15mV 0.06mV

[0159] For multimodal feature extraction, the first step is feature calculation. In this embodiment, the preset sliding window length is 60 seconds. ST segment features include: slope 60 milliseconds after the J point: -0.12 mV / s (normal range: ±0.05 mV / s). ST segment depression amplitude: 0.22 mV (threshold > 0.1 mV is abnormal). Nonlinear features include: sample entropy: 1.32 (normal reference value 0-0.8). Hurst exponent: 0.87 (> 0.75 indicates abnormal long-range correlation).

[0160] Original feature dimension: 512. Compression to 64 dimensions using a variational autoencoder, with a reconstruction error MSE of 0.037 (<0.05 meets the standard).

[0161] Next, the spatiotemporal graph neural network model was built to detect the blood pressure-ECG time lag correlation. The correlation results are as follows:

[0162] Table 5 Correlation results

[0163] Parameter pair Optimal time delay τ Mutual Information Systolic blood pressure-heart rate +2.3s 0.68 Blood oxygen-ST slope -1.8s 0.51

[0164] Data enhancement effect: Generate 2000 sets of synthetic feature data.

[0165] The SHAP value distribution verification results are as follows:

[0166] Table 6 SHAP value distribution verification

[0167] Key Features Real data SHAP mean Generate data SHAP mean deviation ST segment slope 0.34 0.31 8.8％ Systolic blood pressure variance 0.28 0.26 7.1％

[0168] The SHAP threshold is 0.1, and 32 key features are screened out.

[0169] The top 3 important features are as follows:

[0170] Table 7 Top 3 important features

[0171] feature SHAP value Medical significance ST segment slope 0.41 Quantification of myocardial ischemia Systolic blood pressure variance 0.38 Autonomic nervous system dysregulation QRS complex area 0.35 Abnormal ventricular depolarization

[0172] The classification performance obtained is as follows:

[0173] Table 8 Classification performance

[0174] index First layer of forest Second layer of forest Sensitivity 89.2％ 94.7％ Specificity 85.6％ 91.3％ AUC 0.88 0.96

[0175] The monitoring timeline is as follows:

[0176] Table 9 Monitoring timeline

[0177] Time Node event System response 09:25:00 ST segment slope <-0.1 mV / s for the first time Yellow warning (P=0.73) 09:27:15 Systolic blood pressure drops by > 20 mmHg / minute Red alarm (P=0.93) 09:28:00 Doctors confirmed acute anterior myocardial infarction System records diagnostic results

[0178] System performance: Feature calculation latency: 128ms. Classification inference latency: 43ms. Total response time: 171ms (<200ms to meet the standard).

[0179] Test dataset: MIT-BIH arrhythmia database n = 48 cases; MIMIC-III acute coronary syndrome cases n = 127 cases.

[0180] Performance comparison is as follows:

[0181] Table 10 Performance comparison

[0182] method Sensitivity Specificity AUC False alarm rate / 24h Traditional LSTM 82.3％ 79.1％ 0.85 18.7％ Method of the present invention 93.7％ 90.3％ 0.96 6.8％ Improvement +11.4％ +11.2％ +0.11 -63.6％

[0183] Summary of the technical effects of the method of the present invention:

[0184] Early warning capability: In cases of ST-segment elevation myocardial infarction, the system issues an early warning an average of 4.2 minutes earlier than traditional methods.

[0185] Diagnostic accuracy: The detection sensitivity for NSTEMI (non-ST-segment elevation myocardial infarction) reaches 91.4%, an increase of 23.6% over existing equipment.

[0186] Clinical practicality: The interpretable reports generated were adopted by 92.3% of cardiologists. During a six-month clinical trial at a tertiary hospital, the missed report rate was only 0.8%.

[0187] This example verifies the significant advantages of this patented technical solution in terms of real-time performance, accuracy, and clinical applicability through specific data, and provides an innovative solution for intelligent monitoring of cardiovascular diseases.

[0188] It should be noted that, in the present invention, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0189] The above embodiments are merely examples of the present invention and do not limit the scope of protection of the present invention. Any designs that are identical or similar to the present invention fall within the scope of protection of the present invention.

Claims

1. A multi-parameter joint monitoring method for cardiovascular disease, characterized in that: The following steps are involved: Acquire multiple physiological parameters related to cardiovascular health of the patient, perform synchronization preprocessing on the physiological parameters using dynamic time warping combined with Kalman filtering, and obtain preprocessed monitoring data; Extracting time domain features, frequency domain features, nonlinear features and morphological features from the preprocessed monitoring data, and constructing a multimodal feature vector through feature concatenation and dimensionality reduction; Physiological parameters are used as nodes in the graph neural network, and the nodes correspond to multimodal feature vectors. The edge weights are obtained through the preset sliding window size and step size. A masked self-supervised pre-training strategy is used to learn the implicit associations between parameters. A combined temporal convolutional network is used to capture multi-scale temporal dependencies. A dynamic relationship model of the spatiotemporal graph neural network is constructed. The adversarial generative network is embedded in the dynamic relationship model of the spatiotemporal graph neural network for data enhancement and to obtain enhanced processing features. Based on the SHAP value, combined with the enhanced processing features, the feature importance score is obtained, and combined with the stratified random forest, the trained two-layer random forest model is trained using the labeled normal and abnormal data to obtain a trained two-layer random forest model; The trained two-layer random forest model is used to classify the real-time monitoring data, identify normal and abnormal states, and obtain abnormal monitoring results.

2. The multi-parameter combined monitoring method for cardiovascular disease according to claim 1, characterized in that: The method of obtaining the pre-processed monitoring data comprises the following steps: Collect the patient's physiological parameter data through various sensors and monitoring devices to obtain the original physiological parameter time series data; performing data cleaning and normalization on the original physiological parameter time series data to obtain normalized physiological parameter data; Performing data alignment on the normalized physiological parameter data using dynamic time warping to obtain aligned physiological parameter time series data; Performing denoising and smoothing processing on the aligned physiological parameter time series data using low-pass filtering and Kalman filtering to obtain denoised and smoothed physiological parameter data; Missing value processing is performed on the denoised and smoothed physiological parameter data to obtain synchronized preprocessed physiological parameter data as preprocessed monitoring data.

3. The multi-parameter combined monitoring method for cardiovascular disease according to claim 1, wherein: The obtaining of the enhanced processing feature comprises the following steps: Each physiological parameter is used as a node in the graph neural network. Each node corresponds to a feature vector containing multimodal features. Through the preset sliding window size and step size, the mutual information between adjacent physiological parameter data points is obtained and used as the edge weight between nodes to construct a spatiotemporal graph neural network. A masked self-supervised pre-training strategy is used to randomly mask 20% of the node features. The masked features are predicted by the graph autoencoder, the implicit connections between nodes are extracted, and the weights of the spatiotemporal graph neural network are optimized to obtain the optimized graph neural network. The graph autoencoder contains a two-layer graph convolutional network (GCN) with a hidden layer dimension of 64. The optimized graph neural network performs spatial convolution on the node features to obtain a spatial feature sequence; Design a temporal convolutional network, input the spatial feature sequence into the designed temporal convolutional network, combine the temporal convolutional network to capture multi-scale temporal dependencies, build a spatiotemporal graph neural network dynamic relationship model, and obtain temporal and spatial features by stacking graph convolution layers and temporal convolutional network layers in the spatiotemporal graph neural network dynamic relationship model; Using conditional Wasserstein GAN as the adversarial generative network, the generator and discriminator are designed to be embedded in the dynamic relationship model of the spatiotemporal graph neural network. The obtained temporal and spatial features are enhanced in feature space. The discriminator constrains the generator to generate feature data that conforms to the clinical distribution, and then fuses it with the original feature data to obtain enhanced processing features.

4. The multi-parameter combined monitoring method for cardiovascular disease according to claim 1, wherein: The method of obtaining a trained double-layer random forest model comprises the following steps: Based on the SHAP value, the enhanced processing features are ranked by importance using a feature importance scoring method, and features whose importance is greater than a set threshold are selected as the filtered feature subset; The filtered feature subset is used as the input feature of the first layer of the double-layer random forest to obtain an initial value of the abnormal probability. The initial value of the abnormal probability is used as the sample weight of the second layer of the forest. The weighted Gini coefficient is used and an adaptive weighting mechanism is introduced to adjust the tree node splitting of the second layer of the random forest to obtain a trained double-layer random forest model. The trained two-layer random forest model is trained using the labeled normal and abnormal data to obtain a trained two-layer random forest model.

5. The multi-parameter combined monitoring method for cardiovascular disease according to claim 1, wherein: The trained model is used to classify real-time monitoring data and identify normal and abnormal states, including: Input the real-time monitoring data into the trained two-layer random forest model; According to the output results of the trained two-layer random forest model, the status of the real-time monitoring data is judged and abnormal monitoring results are generated.

6. A detection system for a multi-parameter joint monitoring method for cardiovascular disease, characterized in that: include: A data acquisition module is used to collect physiological parameter data related to cardiovascular health, obtain multiple physiological parameters related to cardiovascular health of patients, synchronize and pre-process the physiological parameters using dynamic time warping combined with Kalman filtering, and obtain pre-processed monitoring data; Data processing module, used for preprocessing, feature extraction, relationship modeling and data enhancement of collected data; Time domain features, frequency domain features, nonlinear features, and morphological features are extracted from the preprocessed monitoring data, and a multimodal feature vector is constructed through feature splicing and dimensionality reduction. Physiological parameters are used as graph nodes, and the nodes correspond to multimodal feature vectors. Edge weights are obtained through a preset sliding window size and step size. A masked self-supervised pre-training strategy is used to learn implicit associations between parameters. A joint temporal convolutional network is used to capture multi-scale temporal dependencies, and a spatiotemporal graph neural network dynamic relationship model is constructed. A generative adversarial network is embedded in the spatiotemporal graph neural network dynamic relationship model for data enhancement to obtain enhanced processing features. The model training module is used to train a two-layer random forest model. Based on the SHAP value, it combines enhanced processing features to obtain feature importance scores. In addition, it combines the stratified random forest and uses the labeled normal and abnormal data to train the trained two-layer random forest model to obtain a trained two-layer random forest model. The anomaly detection module is used to classify real-time monitoring data using a trained two-layer random forest model, identify normal and abnormal states, and obtain abnormal monitoring results; The early warning module is used to judge the abnormal monitoring results obtained according to the set threshold. When the judgment result is not satisfied with the set threshold, the early warning mechanism is automatically triggered and a reminder message is sent to the patient, doctor and family members.

7. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

8. A computer device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

Images