Method and system for predicting risk of cardiovascular and cerebrovascular diseases

Abstract

The application provides a cardiovascular and cerebrovascular disease risk prediction method and system, relates to the technical field of data processing, and the method comprises the following steps: acquiring real-time physiological parameters of a target user, including dynamic blood pressure fluctuation rate, serum lipoprotein level, heart rate variability frequency domain index and sleep apnea hypopnea index; preprocessing the real-time physiological parameters, extracting the instantaneous waveform features of the blood pressure signal by using wavelet transform, and extracting the features related to the health condition, including blood pressure level, blood lipid level, blood glucose level, electrocardiogram abnormality index and cardiac ultrasound abnormality index, to form a feature vector set. Through the acquisition of the real-time physiological parameters, the preprocessing to form the feature vector set, the optimization by the phosphorus shrimp swarm algorithm, the construction of a three-level fusion model, the final accurate evaluation of the cardiovascular and cerebrovascular disease risk degree of the target user and the output of the early warning information, the accuracy and timeliness of the cardiovascular and cerebrovascular disease risk prediction are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for predicting the risk of cardiovascular and cerebrovascular diseases. Background Technology

[0002] Monitoring of cardiovascular and cerebrovascular diseases sometimes relies on limited physiological parameters for risk prediction, such as static indicators like blood pressure, blood lipids, and blood sugar. However, these indicators only reflect the body's condition at a certain point in time and are difficult to comprehensively and dynamically reflect the risk of cardiovascular and cerebrovascular diseases.

[0003] For example, focusing only on the absolute value of blood pressure while ignoring its dynamic fluctuations, such as the diurnal rhythm of blood pressure and the instantaneous fluctuation rate of blood pressure, is not an effective way to assess the risk of cardiovascular and cerebrovascular diseases.

[0004] Furthermore, traditional techniques primarily measure overall blood lipid levels such as total cholesterol and triglycerides, failing to accurately reflect the levels of different lipoprotein subclasses, such as low-density lipoprotein cholesterol (LDL-C) and high-density lipoprotein cholesterol (HDL-C), which are closely related to the risk of cardiovascular and cerebrovascular diseases. Focusing only on certain single characteristics, such as blood pressure and blood lipid levels, ignores the interactions and mutual influences between these characteristics. For example, there may be complex interactions between blood pressure and blood lipid levels, jointly affecting the risk of cardiovascular and cerebrovascular diseases. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for predicting the risk of cardiovascular and cerebrovascular diseases, so as to provide a basis for the early prediction of cardiovascular and cerebrovascular diseases.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] Firstly, a method for predicting the risk of cardiovascular and cerebrovascular diseases, the method comprising:

[0008] Acquire real-time physiological parameters of the target user, including dynamic blood pressure variability, serum lipoprotein levels, heart rate variability frequency domain index, and sleep apnea-hypopnea index;

[0009] Real-time physiological parameters are preprocessed, and wavelet transform is used to extract the instantaneous waveform features of the blood pressure signal. Features related to health status, including blood pressure level, blood lipid level, blood glucose level, abnormal electrocardiogram indicators and abnormal echocardiogram indicators, are extracted to form a feature vector set.

[0010] The krill swarm algorithm is used to select and optimize the feature vector set. By simulating the foraging behavior of krill swarms, the algorithm automatically searches for a subset of features that are key to assessing cardiovascular health status, so as to obtain the optimized feature vector set.

[0011] Based on the optimized feature vector set, a three-level fusion model is constructed using probabilistic time Petri nets, which includes spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning.

[0012] The feature vector of the target user is input into the three-level fusion model. According to the inference rules of Petri net, the health status risk level of the target user is obtained. Based on the health status risk level of the target user, the risk assessment result and early warning information are output.

[0013] Furthermore, real-time physiological parameters are preprocessed. Wavelet transform is used to extract instantaneous waveform features of the blood pressure signal, and features related to health status are extracted, including blood pressure level, blood lipid level, blood glucose level, abnormal electrocardiogram indicators, and abnormal echocardiogram indicators, forming a feature vector set, including:

[0014] Wavelet transform was used to extract instantaneous waveform features from blood pressure signals in order to obtain the dynamic characteristics of blood pressure changes.

[0015] Features related to health status are extracted from the data after wavelet transform processing, including blood pressure level, blood lipid level, blood sugar level, abnormal electrocardiogram indicators, and abnormal echocardiogram indicators, forming a feature vector set.

[0016] Furthermore, the krill swarm algorithm is used to select and optimize the feature vector set. By simulating the foraging behavior of krill swarms, a subset of features key to cardiovascular health assessment is automatically searched to obtain the optimized feature vector set, including:

[0017] Determine the size of the krill colony, i.e. the number of candidate features participating in feature selection, and assign an initial position to each krill;

[0018] Set the krill's sensing range, movement step size, and maximum number of iterations, and define a contribution function to evaluate the contribution of the krill's current location to the assessment of cardiovascular health status.

[0019] For each krill, extract the corresponding features based on its current location to build a prediction model;

[0020] The prediction model was trained using sample data containing cardiovascular health conditions, and the contribution value of each krill was calculated based on the contribution function.

[0021] The contribution value is assigned to the corresponding krill as a performance evaluation in the current iteration. For each krill, the set of neighbors within its perception range is calculated, and the learning object of the current krill is determined in the set of neighbors.

[0022] Update the current krill position based on the learning target, the current krill position difference, and the preset movement step size;

[0023] The process of feature extraction, model building, training, and contribution calculation is repeated until the preset maximum number of iterations is reached, ultimately resulting in a set of optimized feature combinations.

[0024] Based on the optimized feature combination, the corresponding features are extracted from the original feature vector set to form the optimized feature vector set.

[0025] Furthermore, based on the contribution function, the contribution of the krill's current location to the cardiovascular health assessment is calculated, including:

[0026] When constructing a prediction model based on the feature subset corresponding to the current location of the krill, the number of samples correctly predicted as positive and the number of samples correctly predicted as negative, as well as the number of samples incorrectly predicted as positive and the number of samples incorrectly predicted as negative are calculated; the accuracy index value is calculated based on the number of positive samples, the number of negative samples, the number of positive samples, and the number of negative samples.

[0027] Determine the size of the feature subset corresponding to the current location of the krill, that is, the number of features contained in the feature subset, and calculate the feature subset factor index value based on the number of features contained in the feature subset;

[0028] Calculate the entropy of the entire dataset and the entropy of the data subsets corresponding to different values ​​of each feature. For each feature, calculate the weighted average of the entropy of the data subsets corresponding to all its values. Calculate the difference between the entropy of the entire dataset and the weighted average, and calculate the ratio between the difference and the maximum information gain among all features. Based on the ratio, calculate the information gain impact value.

[0029] The accuracy index, feature subset factor index, and information gain influence value are fused to determine the contribution of the krill's current location to the assessment of cardiovascular health status.

[0030] Furthermore, based on the optimized feature vector set, a three-level fusion model is constructed using probabilistic temporal Petri nets, incorporating spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning. This model includes:

[0031] Initialize the construction environment for the probabilistic time Petri net, including defining the net's state, transitions, arcs, and associated probability and time parameters;

[0032] The blood pressure waveform features and heart rate variability physiological signal data in the optimized feature vector set are encoded and fused to form an intermediate feature representation that includes dynamic changes in physiological signals, spatial relationships and clinical knowledge;

[0033] By using a dynamic reinforcement learning algorithm, intermediate feature representations containing dynamic changes in physiological signals, spatial relationships, and clinical knowledge are taken as input data for learning and optimization. After multiple iterations of training, a three-level fusion model is formed, which includes spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning.

[0034] Furthermore, the three-level fusion model includes:

[0035] The first level is the physiological signal spatiotemporal coding module, which performs spatiotemporal coding on the physiological signals in the optimized feature vector set to obtain the temporal relationship and spatial features between the signals;

[0036] The second level is the clinical knowledge graph embedding module, which integrates clinical knowledge with feature vectors and transforms clinical knowledge into structured information.

[0037] The third level is the dynamic reinforcement learning module, which dynamically learns based on the outputs of the physiological signal spatiotemporal coding module and the clinical knowledge graph embedding module, ultimately outputting the target user's health status risk level.

[0038] Furthermore, the feature vector of the target user is input into the three-level fusion model. Based on the inference rules of the Petri net, the health risk level of the target user is obtained. Then, based on the health risk level of the target user, risk assessment results and early warning information are output, including:

[0039] The feature vector of the target user is used as input data and passed to the first level of the three-level fusion model, namely the physiological signal spatiotemporal coding module.

[0040] In the spatiotemporal coding module of physiological signals, the temporal relationship and spatial features between physiological signals are extracted to form intermediate feature representations;

[0041] The intermediate feature representation is passed to the second level, namely the clinical knowledge graph embedding module, where clinical knowledge is fused with the intermediate feature representation to generate a fused feature representation.

[0042] The fused feature representation is passed to the third level, namely the dynamic reinforcement learning module, which performs dynamic learning and optimization based on the fused feature representation. Through continuous adjustment, the target user's health status risk level is output.

[0043] Based on the target users' health status and risk level, generate corresponding risk assessment reports, including risk level, risk factor analysis, and recommended measures;

[0044] Based on the target user's health status and risk level, an early warning threshold is set. The target user's health status and risk level are compared with the early warning threshold. When the target user's health status and risk level are greater than or equal to the early warning threshold, an early warning message is automatically generated according to the preset early warning rules, including the risk level, main risk factors, and suggested emergency actions.

[0045] Furthermore, the fused feature representation is passed to the third level, namely the dynamic reinforcement learning module, which performs dynamic learning and optimization based on the fused feature representation. Through continuous adjustment, it outputs the target user's health status risk level, including:

[0046] Initialize the dynamic reinforcement learning model, including setting the structure, parameters, and learning rate of the dynamic reinforcement learning model;

[0047] The dynamic reinforcement learning model receives the fused feature representations and performs dynamic learning according to a preset learning algorithm;

[0048] At each stage of learning, the health risk level of a target user is predicted based on the current feature representation and compared with the actual health risk level, and the difference between the predicted result and the actual result is calculated.

[0049] The learning and optimization process is repeated until the preset learning stop condition is met, at which point the risk level of the target user's health status is output.

[0050] Secondly, the cardiovascular and cerebrovascular disease risk prediction system includes:

[0051] The data acquisition module is used to acquire physiological signal data of the target user;

[0052] The feature extraction module is used to extract features from the physiological signal data and form a feature vector set;

[0053] The feature selection and optimization module is used to select and optimize the feature vector set using the krill swarm algorithm, and automatically search for and determine the feature subset that is key to cardiovascular health assessment.

[0054] The three-level fusion module is used to perform deep, multi-dimensional fusion processing on the optimized feature vector set to obtain the fused feature representation.

[0055] The risk prediction module is used to predict the health status risk level of the target user based on the feature representation output by the three-level fusion module, obtain the risk prediction result, and automatically generate early warning information when the health status risk level of the target user exceeds the preset threshold.

[0056] Thirdly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0057] The above-described solution of the present invention has at least the following beneficial effects:

[0058] Wavelet transform is used to extract instantaneous waveform features from blood pressure signals, capturing subtle yet crucial changes. Simultaneously, it extracts various health-related features, such as blood pressure, blood lipid, blood glucose, ECG abnormalities, and echocardiogram abnormalities, forming a feature vector set. This preprocessing method not only preserves key physiological parameters but also removes noise and redundant data, improving data quality and usability. A krill swarm algorithm is then used to select and optimize features from the feature vector set. By simulating the foraging behavior of krill swarms, it automatically searches for a subset of features crucial for assessing cardiovascular health. This method avoids the subjectivity and limitations of manual feature selection, quickly and accurately finding the most predictive feature combinations, thus improving the efficiency and accuracy of feature selection.

[0059] The feature vector set optimized by the krill swarm algorithm removes features that contribute little to cardiovascular health assessment, retains key features, reduces the number and complexity of features, lowers the computational cost of the model, and improves the model's generalization ability and prediction accuracy. A three-level fusion model, incorporating spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning, is constructed using probabilistic temporal Petri nets. This organically combines the spatiotemporal information of physiological signals, clinical knowledge, and the optimization capabilities of reinforcement learning. Spatiotemporal encoding of physiological signals captures the dynamic changes of physiological signals in time and space, while embedding of clinical knowledge graphs introduces rich medical domain knowledge. Dynamic reinforcement learning, through continuous learning and optimization, improves the model's adaptability and accuracy. This three-level fusion model can fully utilize information from different sources to more comprehensively and accurately assess the cardiovascular and cerebrovascular disease risk level of target users. Compared with traditional models, it can better handle complex physiological data and medical knowledge, improving the accuracy and reliability of risk prediction.

[0060] The feature vectors of target users are input into a three-level fusion model, and the risk level of the target user's health status is obtained according to the inference rules of Petri nets. This personalized assessment method can fully consider the unique physiological characteristics and health status of each user, providing users with more accurate and targeted risk assessment results. Based on the target user's health status risk level, the risk assessment results and early warning information are output. When the risk level is high, a timely warning is issued to remind the user to take corresponding preventive and treatment measures, which helps to detect the risk of cardiovascular and cerebrovascular diseases early, reduce the incidence and mortality of diseases, and improve the user's health level and quality of life. This method provides doctors with a scientific and objective tool for predicting the risk of cardiovascular and cerebrovascular diseases, enabling them to more accurately assess patients' conditions and risks, formulate more reasonable treatment plans, and improve clinical treatment effects. The research and application of this method contributes to a deeper understanding of the pathogenesis and influencing factors of cardiovascular and cerebrovascular diseases, provides new ideas and methods for medical research, and promotes the development of the field of prevention and treatment of cardiovascular and cerebrovascular diseases. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating the cardiovascular and cerebrovascular disease risk prediction method provided in an embodiment of the present invention.

[0062] Figure 2 This is a schematic diagram of a cardiovascular and cerebrovascular disease risk prediction system provided in an embodiment of the present invention. Detailed Implementation

[0063] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0064] like Figure 1 As shown, embodiments of the present invention propose a method for predicting the risk of cardiovascular and cerebrovascular diseases, the method comprising the following steps:

[0065] Step 1: Obtain the target user's real-time physiological parameters, including dynamic blood pressure variability, serum lipoprotein levels, heart rate variability frequency domain index, and sleep apnea-hypopnea index.

[0066] Step 2: Preprocess the real-time physiological parameters, use wavelet transform to extract the instantaneous waveform features of the blood pressure signal, and extract features related to health status, including blood pressure level, blood lipid level, blood glucose level, abnormal electrocardiogram indicators and abnormal echocardiogram indicators, to form a feature vector set;

[0067] Step 3: Use the krill swarm algorithm to select and optimize the feature vector set. By simulating the foraging behavior of krill swarms, automatically search for the feature subset that is key to cardiovascular health assessment to obtain the optimized feature vector set.

[0068] Step 4: Based on the optimized feature vector set, construct a three-level fusion model that includes spatiotemporal encoding of physiological signals, embedding of clinical knowledge graph, and dynamic reinforcement learning through probabilistic time Petri net;

[0069] Step 5: Input the feature vector of the target user into the three-level fusion model, obtain the health status risk level of the target user according to the inference rules of Petri net, and output the risk assessment result and generate early warning information based on the health status risk level of the target user.

[0070] In this embodiment of the invention, by acquiring real-time physiological parameters such as dynamic blood pressure variability, serum lipoprotein levels, heart rate variability frequency domain index, and sleep apnea-hypopnea index, the physiological state of the target user can be comprehensively and continuously monitored. Real-time monitoring of these physiological parameters helps to promptly detect abnormal changes in physiological indicators, thereby enabling early warning of potential health risks and providing the possibility for timely intervention and treatment.

[0071] Step 2 involves preprocessing operations (such as denoising and normalization) to improve the quality of physiological parameter data and reduce the impact of noise and interference on subsequent analysis. Using wavelet transform and other methods to extract instantaneous waveform features of blood pressure signals, as well as health-related features (such as blood pressure levels and blood lipid levels), allows for a more accurate depiction of the target user's physiological state. The extracted features are then organized into a feature vector set, facilitating subsequent feature selection and optimization, as well as model construction.

[0072] Step 3: The krill swarm algorithm, by simulating the foraging behavior of krill swarms, can automatically search for a subset of features crucial for cardiovascular health assessment, reducing human intervention and subjective bias, and improving the objectivity and accuracy of feature selection. Through feature selection and optimization, a set of feature combinations that contributes most to cardiovascular health assessment can be obtained, forming an optimized feature vector set.

[0073] Step 4 involves constructing a three-level fusion model using probabilistic temporal Petri nets. This model integrates multi-source information, including spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning, to more comprehensively assess the target user's health status. The three-level fusion model comprehensively considers the dynamic changes in physiological signals, prior information from clinical knowledge, and the dynamic adjustment of individual health status, thereby improving the model's predictive performance and generalization ability.

[0074] Step 5: Input the feature vector of the target user into the three-level fusion model. According to the inference rules of Petri net, the health status risk level of the target user can be obtained, providing strong support for personalized medicine.

[0075] Based on the target user's health status and risk level, outputting risk assessment results and generating early warning information helps to promptly identify potential health risks and take corresponding intervention measures, reducing the incidence and mortality of diseases. By providing accurate risk assessment and early warning information, doctors can better understand patients' health status and develop more reasonable treatment plans, thereby improving the quality and efficiency of medical services.

[0076] In a preferred embodiment of the present invention, step 1 above, obtaining the real-time physiological parameters of the target user, including dynamic blood pressure variability, serum lipoprotein levels, heart rate variability frequency domain index, and sleep apnea-hypopnea index, may include:

[0077] In this embodiment of the invention, medical devices are selected to monitor the physiological parameters of the target user. For example, an ambulatory blood pressure monitor is used to measure blood pressure variability, a blood analyzer is used to detect serum lipoprotein levels, an electrocardiograph is used to record heart rate variability data, and a polysomnography system is used to assess the sleep apnea-hypopnea index. The devices are correctly configured and calibrated according to their operating manuals and clinical needs to ensure the accuracy and reliability of the collected data.

[0078] Connect the device to the target user and collect data according to the device's operating instructions. For example, attach the cuff to the blood pressure monitor and start the device to continuously monitor blood pressure; attach electrodes to the electrocardiograph and start recording ECG data; have the target user wear a polysomnography monitor for nighttime sleep monitoring. During data collection, ensure stable device operation and avoid interference and errors. Simultaneously, record key information such as collection time, location, and device model. Transmit the collected physiological data to the data processing center in real time for storage and processing. Wired or wireless connections such as USB, Bluetooth, and Wi-Fi can be used. During data transmission, ensure data security and privacy. Employ encryption technology to protect the data and prevent leakage and tampering.

[0079] In a preferred embodiment of the present invention, step 2 above involves preprocessing real-time physiological parameters, extracting instantaneous waveform features of the blood pressure signal using wavelet transform, and extracting features related to health status, including blood pressure level, blood lipid level, blood glucose level, abnormal electrocardiogram indicators, and abnormal echocardiogram indicators, forming a feature vector set, which may include:

[0080] Step 221: Wavelet transform is used to extract instantaneous waveform features from the blood pressure signal to obtain the dynamic features of blood pressure changes;

[0081] Step 222: Extract health-related features from the data after wavelet transform processing, including blood pressure level, blood lipid level, blood sugar level, abnormal electrocardiogram indicators, and abnormal echocardiogram indicators, to form a feature vector set.

[0082] In this embodiment of the invention, continuous blood pressure monitoring data of the target user is acquired, ensuring that the data time interval and sampling rate meet the requirements of wavelet transform. The data is then denoised to remove potential interference and noise, improving the signal-to-noise ratio. Based on the characteristics of the blood pressure signal, a suitable wavelet basis function, the Daubechies wavelet, is selected. The number of decomposition levels for the wavelet transform is determined, depending on the signal complexity and the required feature details. Wavelet transform is performed on the blood pressure signal to obtain approximation coefficients and detail coefficients at different scales. Instantaneous waveform features, such as peak values, trough values, and slopes, are extracted from the detail coefficients and organized into a feature vector.

[0083] Step 222: Calculate the statistical characteristics of blood pressure, such as the mean and standard deviation, to reflect the overall level and fluctuation of blood pressure. Based on clinical standards, blood pressure levels are categorized into normal, prehypertension, and hypertension. Extract indicators such as total cholesterol, triglycerides, high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C) from the blood analysis report. Assess whether blood lipid levels are normal and whether there is a risk of hyperlipidemia based on clinical standards. Extract indicators such as fasting blood glucose and postprandial blood glucose from blood glucose monitoring data. Assess whether blood glucose levels are normal and whether there is a risk of diabetes based on clinical standards. Automatically analyze electrocardiogram (ECG) data to identify abnormal waveforms such as arrhythmias and myocardial ischemia, and extract characteristic parameters such as the QRS complex and T wave from the ECG. Automatically analyze echocardiogram images to identify cardiac structural abnormalities and cardiac functional abnormalities. Extract indicators such as left ventricular ejection fraction (LVEF) and left ventricular end-diastolic diameter (LVEDD) from echocardiograms to assess cardiac health. Organize the extracted features into a feature vector set.

[0084] Assume you receive continuous blood pressure monitoring data from a target user, along with corresponding blood analysis reports, electrocardiogram (ECG), and echocardiogram results. Process the data as follows:

[0085] Wavelet transform is performed on blood pressure signals to extract instantaneous waveform features such as peak value and slope. Total cholesterol, triglycerides, HDL-C, and LDL-C are extracted from blood analysis reports, and blood lipid levels are assessed for normality. Fasting blood glucose and postprandial blood glucose are extracted from blood glucose monitoring data, and blood glucose levels are assessed for normality. Electrocardiogram (ECG) data is automatically analyzed to identify abnormal waveforms such as arrhythmias and myocardial ischemia, and characteristic parameters such as QRS complexes and T waves are extracted. Echocardiogram images are automatically analyzed to identify cardiac structural and functional abnormalities, and indicators such as LVEF and LVEDD are extracted. The extracted features are then organized into a feature vector set.

[0086] Employing methods such as wavelet transform to extract instantaneous waveform features from physiological signals allows for a more refined depiction of the dynamic changes in these signals, improving the accuracy of feature extraction. Automated extraction of health-related features, such as blood pressure, blood lipid levels, and blood glucose levels, enhances the efficiency of feature extraction. By extracting multiple health-related features to form a feature vector set, it is possible to more accurately assess the health status of target users and identify potential health risks. Health risk prediction based on feature vector sets can provide personalized medical advice and treatment plans for each target user, contributing to the development of personalized medicine and improving the targeting and effectiveness of medical services.

[0087] In a preferred embodiment of the present invention, step 3 above, which utilizes a krill swarm algorithm to select and optimize the feature vector set, automatically searches for a subset of features crucial to cardiovascular health assessment by simulating the foraging behavior of a krill swarm, to obtain an optimized feature vector set, may include:

[0088] Step 331: Determine the size of the krill group, i.e. the number of candidate features participating in feature selection, and assign an initial position to each krill.

[0089] Step 332: Set the krill's sensing range, movement step size, and maximum number of iterations, and define a contribution function to evaluate the contribution of the krill's current position to the cardiovascular health assessment.

[0090] Step 333: For each krill, extract the corresponding features based on the current location to build a prediction model;

[0091] Step 334: Train the prediction model using sample data containing cardiovascular health status, and calculate the contribution value of each krill according to the contribution function.

[0092] Step 335: Assign the contribution value to the corresponding krill as a performance evaluation in the current iteration, and for each krill, calculate the neighbor set within its perception range, and determine the learning object of the current krill in the neighbor set.

[0093] Step 336: Update the current krill position based on the learning object, the current krill position difference, and the preset movement step size;

[0094] Step 337: Repeat the process of feature extraction, model building, training, and contribution calculation until the preset maximum number of iterations is reached, and finally obtain a set of optimized feature combinations.

[0095] Step 338: Based on the optimized feature combination, extract the corresponding features from the original feature vector set to form the optimized feature vector set.

[0096] In this embodiment of the invention, the dimension of the feature vector set is first read, which represents the number of candidate features. The size of the krill swarm is determined based on this number; for example, if the feature vector set contains 100 features, then 100 krill are initialized. Each krill is assigned an initial position in the feature space, which can be a random vector representing a point in the feature space. The initial positions can be selected through a uniform random distribution to ensure that the krill swarm extensively explores the feature space.

[0097] Step 332 sets the krill swarm's perception range (e.g., 5 units), movement step size (e.g., 0.1 units), and maximum number of iterations (e.g., 100). These parameters determine the krill's exploration ability and search efficiency in the feature space. A function is defined to evaluate the contribution of the krill's current location to the cardiovascular health assessment; this function can be constructed based on the accuracy of the prediction model.

[0098] Step 333: For each krill, extract a subset of features from the feature vector set based on its current position. For example, if the feature index corresponding to the krill's current position is [2, 5, 7], then extract these three features. Use the extracted feature subset to build a predictive model for predicting cardiovascular health status.

[0099] Step 334: Train the prediction model using sample data containing cardiovascular health status information. The training data should include feature vectors and corresponding health status labels. Based on the performance of the trained model, calculate the contribution value for each krill using a contribution function. This value reflects the importance of the krill's current location (i.e., the feature subset) to the cardiovascular health status assessment.

[0100] Step 335: Assign a contribution value to the corresponding krill as an evaluation of its performance in the current iteration. For each krill, calculate its neighbor set within its perceptual range. The krill in the neighbor set are the objects that the current krill can "learn" from. When calculating the neighbor set, the distance between krill (such as Euclidean distance) and the perceptual range can be considered.

[0101] Step 336: In the neighbor set, select the neighbor with the highest contribution as the learning target for the current krill. Update the current krill's position based on the learning target's position, the current krill's position difference, and the preset movement step size; the update formula can be expressed as: Repeat steps 333 to 336 until the preset maximum number of iterations is reached. In each iteration, the krill population gradually converges to a subset of features that contribute most to the assessment of cardiovascular health. After multiple iterations, an optimized set of features is obtained, which have the highest contribution to the assessment of cardiovascular health.

[0102] Step 338: Based on the optimized feature combination, extract the corresponding features from the original feature vector set, and organize the extracted features into an optimized feature vector set.

[0103] Suppose we have a feature vector set containing 100 features for assessing cardiovascular health. We initialize 100 krill, each krill representing a point in the feature space. We set the krill's perception range to 5 units, its step size to 0.1 units, and the maximum number of iterations to 100.

[0104] In the first iteration, each krill extracts a subset of features based on its initial position and builds a prediction model. After training the model using training data, the contribution value of each krill is calculated. Then, each krill calculates its neighbor set within its perceptual range and selects the neighbor with the highest contribution value as the learning target. The current krill's position is updated based on the difference between the learning target's position and the current krill's position.

[0105] After repeating the process 100 times, the krill population gradually converged to a subset of features that contributed most to the assessment of cardiovascular health. These features were then extracted from the original feature vector set to form an optimized feature vector set.

[0106] The krill swarm algorithm automatically searches for a subset of key features for cardiovascular health assessment, removing redundant and irrelevant features and improving the accuracy of the predictive model. The optimized feature vector set contains the most important features for cardiovascular health assessment, helping the model more accurately capture the relationship between physiological state and health status. The optimized feature vector set reduces the number and complexity of features, lowering the risk of overfitting and enhancing model stability. A stable model performs more consistently across different datasets, improving the reliability of prediction results. Reducing the number of features lowers model complexity and reduces computational resource and time consumption. Improved computational efficiency is particularly important in real-time or large-scale data processing scenarios. Cardiovascular health assessment based on the optimized feature vector set can provide personalized medical advice and treatment plans for each patient. Since feature vectors may differ between patients, the optimized feature vector set better reflects the individual physiological state of each patient. The krill swarm algorithm is an automated feature selection method, reducing human intervention and subjective bias. Automated feature selection improves the objectivity and accuracy of feature selection, making the feature selection process more scientific and reliable.

[0107] In a preferred embodiment of the present invention, the contribution of the krill's current location to the cardiovascular health assessment is calculated according to a contribution function, including:

[0108] When constructing a prediction model based on the feature subset corresponding to the current location of the krill, the number of samples correctly predicted as positive, the number of samples correctly predicted as negative, the number of samples incorrectly predicted as positive, and the number of samples incorrectly predicted as negative are calculated. The accuracy index value is calculated based on the number of positive samples, the number of negative samples, the number of positive samples, and the number of negative samples.

[0109] Determine the size of the feature subset corresponding to the current location of the krill, i.e. the number of features contained in the feature subset, and calculate the feature subset factor index value based on the number of features contained in the feature subset.

[0110] Calculate the entropy of the entire dataset and the entropy of the data subsets corresponding to different values ​​of each feature. For each feature, calculate the weighted average of the entropy of the data subsets corresponding to all values. Calculate the difference between the entropy of the entire dataset and the weighted average, and calculate the ratio between the difference and the maximum information gain among all features. Based on the ratio, calculate the information gain impact value.

[0111] The accuracy index, feature subset factor index, and information gain influence value are fused to determine the contribution of the krill's current location to the assessment of cardiovascular health status.

[0112] In this embodiment of the invention, the feature subset corresponding to the current position of the krill is used. Build a predictive model, use this model to make predictions, and count the number of samples for the following four scenarios:

[0113] The number of samples correctly predicted as positive ( );

[0114] The number of samples correctly predicted as negative ( );

[0115] The number of samples that were incorrectly predicted as positive ( );

[0116] The number of samples that were incorrectly predicted as negative ( ).

[0117] Calculate the accuracy index value based on the number of samples. Determine the size of the feature subset corresponding to the krill's current location, i.e., the number of features contained in the feature subset. ). Calculate the feature subset factor index values. ,here, It is a small constant used to avoid zero denominators when calculating the entire dataset. entropy For each feature in the feature subset Calculate all its possible values Corresponding data subset entropy ;in, It is a feature The value; It is a feature The set of all possible values ​​for each feature. Calculate the weighted average of the entropy of the data subsets corresponding to all its values. Calculate the difference between the entropy of the entire dataset and its weighted average. Calculate the difference and the maximum information gain among all features. The ratio between them. For each feature Calculate its information gain impact value. Perform the above calculations on all features in the feature subset and sum them to obtain the total information gain impact value. Then, perform a weighted sum of the accuracy index, feature subset factor index, and information gain impact value. ,in, , , These are weighting coefficients used to adjust the proportion of each component in the total contribution, outputting the contribution of the krill's current location to the cardiovascular health assessment. .

[0118] The contribution function comprehensively considers the accuracy of the prediction model (through...). , , , (calculation), size of feature subset (through) The comprehensive consideration of both computational calculations and information gain of features (calculated through entropy and information gain) helps select the subset of features most predictive of cardiovascular health assessment, thereby improving the accuracy of the predictive model. By selecting the subset of features with the highest contribution, the number and complexity of features can be reduced, lowering the risk of model overfitting. Stable models perform more consistently across different datasets, improving the reliability of prediction results. Reducing the number of features reduces model complexity, decreasing computational resource and time consumption. The optimized feature vector set better reflects the individual physiological state of patients, providing strong support for personalized medicine. Different patients may have different feature vectors; cardiovascular health assessment based on the optimized feature vector set can provide personalized medical advice and treatment plans for each patient. The krill swarm algorithm, combined with a contribution function, implements an automated feature selection method, reducing human intervention and subjective bias. Automated feature selection improves the objectivity and accuracy of feature selection, making the feature selection process more scientific and reliable.

[0119] In a preferred embodiment of the present invention, step 4 above, which involves constructing a three-level fusion model comprising spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning using a probabilistic time Petri net based on the optimized feature vector set, may include:

[0120] Step 441: Initialize the construction environment for the probabilistic-time Petri net, including defining the net's state, transitions, arcs, and associated probability and time parameters;

[0121] Step 442: Encode and fuse the blood pressure waveform features and heart rate variability physiological signal data in the optimized feature vector set to form an intermediate feature representation that includes dynamic changes in physiological signals, spatial relationships and clinical knowledge.

[0122] Step 443: Using a dynamic reinforcement learning algorithm, intermediate feature representations containing dynamic changes in physiological signals, spatial relationships, and clinical knowledge are used as input data for learning and optimization. After multiple iterations of training, a three-level fusion model is formed, which includes spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning.

[0123] In this embodiment of the invention, a set of states in a probabilistic time Petri net is determined, representing possible configurations or states of the system at different points in time. Each state is assigned a unique identifier, and events or actions that can trigger state transitions are identified; these events or actions are called transitions. The triggering conditions and execution effects of each transition are defined, including the probability of the state transition and the time delay.

[0124] Define the connections between states and transitions; these connections are called arcs. Define a weight for each arc, representing the strength or probability of a state transition when a transition is triggered. Assign a trigger probability to each transition, representing the likelihood of the transition being triggered in a given state. Assign a time delay to each transition, representing the time required for the state transition after the transition is triggered.

[0125] Step 442: Extract blood pressure waveform features from the optimized feature vector set, such as peak value, slope, and duration. Convert these features into a processable format using a suitable encoding method (e.g., binary encoding, real number encoding). Extract heart rate variability (HRV) physiological signal data from the optimized feature vector set, such as RR interval and HRV index, and convert them into a processable format using the same encoding method. Fuse the encoded blood pressure waveform features and HRV physiological signal data to form an intermediate feature representation that includes dynamic changes in physiological signals, spatial relationships, and clinical knowledge. The fusion process can employ methods such as weighted averaging, concatenation, and convolution; the appropriate method should be selected based on the specific application scenario.

[0126] Step 333: Select a suitable dynamic reinforcement learning algorithm, such as Q-learning or Deep Q-Network (DQN). Initialize the parameters in the algorithm, such as the learning rate, discount factor, and exploration rate. Input the intermediate feature representation obtained in step 332 into the dynamic reinforcement learning algorithm. The dynamic reinforcement learning algorithm learns the optimal policy through multiple iterations of training based on the input data. In each iteration, the algorithm selects an action based on the current state, observes the new state and reward after executing the action, and updates the policy. Through continuous iterative training, the algorithm gradually converges to the final policy, forming a three-level fusion model that includes spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning.

[0127] Assuming we receive physiological signal data and a clinical knowledge graph of a target user, we process them according to the following steps:

[0128] The construction environment for the probabilistic temporal Petri net is initialized, defining states, transitions, arcs, and related probability and time parameters. Blood pressure waveform features and heart rate variability physiological signal data are encoded and fused to form intermediate feature representations. A dynamic reinforcement learning algorithm is used, employing these intermediate feature representations as input data for learning and optimization. Through multiple iterations of training, the algorithm gradually converges to the final policy, forming a three-level fusion model. When new physiological signal data is input, the model can make predictions and decisions based on the learned policy, providing doctors with accurate health risk assessments and early warning information.

[0129] By constructing probabilistic temporal Petri nets, the model can fully utilize the spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning mechanisms to improve its expressiveness and generalization ability. This enables the model to more accurately predict the health risks of target users, providing doctors with more reliable decision support. The three-level fusion model can perform personalized learning and optimization based on each target user's physiological signal data and clinical knowledge graph, which helps promote the development of personalized medicine and provides customized health risk assessment and early warning services for each target user. By automating the processing and analysis of physiological signal data and clinical knowledge graphs, the three-level fusion model can significantly improve the efficiency and quality of medical services. Doctors can obtain accurate health risk assessment results more quickly and take corresponding treatment measures, thereby improving patient treatment outcomes and satisfaction.

[0130] In another preferred embodiment of the present invention, the three-level fusion model may include:

[0131] The first level is the physiological signal spatiotemporal coding module, which performs spatiotemporal coding on the physiological signals in the optimized feature vector set to obtain the temporal relationship and spatial features between the signals;

[0132] The second level is the clinical knowledge graph embedding module, which integrates clinical knowledge with feature vectors and transforms clinical knowledge into structured information.

[0133] The third level is the dynamic reinforcement learning module, which dynamically learns based on the outputs of the physiological signal spatiotemporal coding module and the clinical knowledge graph embedding module, and finally outputs the health status risk level of the target user.

[0134] In this embodiment of the invention, physiological signal data, including blood pressure, heart rate, and electrocardiogram (ECG), are received from a feature vector set. The temporal relationships between these physiological signals are analyzed, such as the correlation between blood pressure fluctuations and heart rate changes. Time series analysis methods (such as LSTM) are used to encode the temporal relationships and extract temporal features. For physiological signals with spatial features (such as ECG lead signals), spatial encoding methods (such as convolutional neural networks (CNNs) are used to extract spatial features. The spatial features are then fused with the temporal features to form a comprehensive representation of the physiological signal features.

[0135] The second level is the clinical knowledge graph embedding module, which collects clinical knowledge related to cardiovascular and cerebrovascular diseases, including disease symptoms, risk factors, and treatment plans. It constructs a clinical knowledge graph, representing clinical knowledge in a structured form. Graph embedding methods (such as TransE and GraphSAGE) are used to transform the clinical knowledge graph into a low-dimensional vector space representation. Physiological signal feature vectors are fused with the clinical knowledge graph embedding vectors to form a feature representation containing rich clinical information. Appropriate fusion strategies (such as weighted averaging and concatenation) are designed to fuse physiological signal features with clinical knowledge graph embedding features, ensuring that the fused features fully reflect the correlation between physiological signals and clinical knowledge.

[0136] The third level: Dynamic reinforcement learning module. A suitable dynamic reinforcement learning algorithm (such as DQN, A3C, etc.) is selected, and model parameters are initialized, including learning rate, discount factor, and exploration rate. The outputs of the physiological signal spatiotemporal encoding module and the clinical knowledge graph embedding module are used as state representations. A state encoding method is designed to ensure that the state accurately reflects the current physiological state and clinical knowledge. Based on the current state, a policy network is used to select actions (such as predicting the health risk level of the target user). The action selection process needs to consider the balance between exploration and utilization. A suitable reward mechanism is designed to reward or penalize based on the difference between the predicted result and the actual result. The reward mechanism needs to guide the model towards more accurate risk prediction. The model is trained using a reinforcement learning algorithm, and the model parameters are optimized through multiple iterations. During training, parameters such as the learning rate and discount factor are continuously adjusted to improve the model's convergence speed and accuracy. After training, the model is used to predict new physiological signal data and output the health risk level of the target user.

[0137] Assuming we receive physiological signal data and clinical medical records from a target user, we process them according to the following steps:

[0138] The received physiological signal data is cleaned and preprocessed. Time series analysis and spatial coding methods are used to extract the temporal relationships and spatial features of the physiological signals. The extracted features are then fused to form a comprehensive representation of the physiological signal features. A clinical knowledge graph related to cardiovascular and cerebrovascular diseases is constructed based on clinical medical records. The graph embedding method is used to transform the clinical knowledge graph into a representation in a low-dimensional vector space.

[0139] Physiological signal feature vectors are fused with clinical knowledge graph embedding vectors to form a feature representation containing rich clinical information. A dynamic reinforcement learning model is initialized, with appropriate parameters and state representation settings.

[0140] Using the fused features as state input, the policy network selects actions (predicting the health risk level of the target user), and rewards or penalties are given based on the difference between the prediction result and the actual result. The model is then trained and optimized. After training, the model is used to predict new physiological signal data and output the health risk level of the target user.

[0141] By using spatiotemporal encoding of physiological signals and embedding clinical knowledge graphs, the model can fully leverage the correlation between physiological signals and clinical knowledge to improve the accuracy of risk prediction. The dynamic reinforcement learning module can dynamically learn and optimize based on new physiological signal data and clinical knowledge, enhancing the model's generalization ability. This allows the model to adapt to the physiological states and clinical situations of different target users, improving its applicability.

[0142] The three-level fusion model can perform personalized learning and prediction based on each target user's physiological signal data and clinical medical records. This helps to promote the development of personalized medicine and provide customized cardiovascular and cerebrovascular disease risk assessment and early warning services for each target user.

[0143] In a preferred embodiment of the present invention, step 5 above, which involves inputting the feature vector of the target user into the three-level fusion model, obtaining the health status risk level of the target user according to the inference rules of the Petri net, and outputting the risk assessment result and generating early warning information based on the health status risk level of the target user, may include:

[0144] Step 552: The feature vector of the target user is used as input data and passed to the first level of the three-level fusion model, namely the physiological signal spatiotemporal coding module.

[0145] Step 553: In the physiological signal spatiotemporal coding module, extract the temporal relationship and spatial features between physiological signals to form an intermediate feature representation;

[0146] Step 554: The intermediate feature representation is passed to the second level, namely the clinical knowledge graph embedding module, where clinical knowledge is fused with the intermediate feature representation to generate the fused feature representation.

[0147] Step 555: The fused feature representation is passed to the third level, namely the dynamic reinforcement learning module, which performs dynamic learning and optimization based on the fused feature representation. Through continuous adjustment, the target user's health status risk level is output.

[0148] Step 556: Based on the target user's health status and risk level, generate a corresponding risk assessment report, including risk level, risk factor analysis, and recommended measures;

[0149] Step 557: Based on the target user's health status and risk level, set an early warning threshold, compare the target user's health status and risk level with the early warning threshold, and when the target user's health status and risk level are greater than or equal to the early warning threshold, automatically generate early warning information according to the preset early warning rules, including risk level, main risk factors, and suggested emergency actions.

[0150] In this embodiment of the invention, physiological signal data is acquired from the target user through medical devices (such as electrocardiographs, blood pressure monitors, etc.) to ensure the accuracy and continuity of data acquisition. Based on the characteristics of the physiological signals, relevant features are extracted, such as time series features (such as mean, standard deviation, peak value, etc.) and spatial features (such as the relationship between leads in an electrocardiogram, etc.). The extracted features are then encoded to form a feature vector set.

[0151] Step 552: Design the input interface of the three-level fusion model to ensure that it can receive the feature vector set as input data and transmit the feature vector set to the physiological signal spatiotemporal coding module through the input interface.

[0152] Step 553: Use time series analysis methods (such as LSTM, GRU, etc.) to extract the temporal relationship between physiological signals, analyze the changing trend of physiological signals over time, and extract temporal features. For physiological signals with spatial features (such as electrocardiogram), use spatial coding methods (such as CNN) to extract spatial features, analyze the spatial distribution and relationship of physiological signals, extract spatial features, and fuse the extracted temporal relationship and spatial features to form a comprehensive intermediate feature representation, ensuring that the fused features can fully reflect the temporal and spatial characteristics of physiological signals.

[0153] Step 554: Construct a clinical knowledge graph related to cardiovascular and cerebrovascular diseases, including disease symptoms, risk factors, and treatment plans, ensuring the accuracy and completeness of the knowledge graph. Use graph embedding methods to transform the clinical knowledge graph into a low-dimensional vector space representation, ensuring that the embedded vectors accurately reflect the semantic and structural information of clinical knowledge. Fuse the intermediate feature representations with the embedded vectors of the clinical knowledge graph, designing appropriate fusion strategies (such as weighted averaging, concatenation, etc.) to ensure that the fused features fully reflect the correlation between physiological signals and clinical knowledge.

[0154] Step 555: Initialize the dynamic reinforcement learning model, setting appropriate parameters and state representation to ensure the model can receive the fused feature representation as input data. Use the fused feature representation as state input and select actions (such as predicting the degree of health risk) according to the policy network. During the learning process, reward or penalize based on the difference between the predicted result and the actual result, adjusting the model parameters accordingly. Through multiple iterations of training, continuously optimize the model parameters to improve prediction accuracy. Adjust parameters such as the learning rate and discount factor according to actual conditions to accelerate convergence and improve model performance. After training, use the model to predict new physiological signal data and output the degree of health risk for the target user.

[0155] Step 556: Based on the preset risk level classification standards (e.g., low, medium, high, very high), map the target user's health risk level to the corresponding risk level to ensure that the risk level classification is scientific and reasonable and accurately reflects the target user's health risk level. Conduct in-depth analysis of the fused feature representation to identify the main factors affecting the target user's health status. Combine clinical knowledge to explain and illustrate the risk factors, providing strong support for the risk assessment report. Based on the risk level and risk factor analysis results, develop personalized recommendations for the target user. Integrate the risk level, risk factor analysis, and recommendations into a complete risk assessment report.

[0156] Step 557: Based on actual needs and safety considerations, set a reasonable early warning threshold to ensure that the threshold accurately reflects the critical level of health risk and avoids false alarms and missed alarms. Compare the health risk level of the target user with the early warning threshold. When the risk level is greater than or equal to the early warning threshold, trigger the early warning mechanism and automatically generate early warning information according to preset early warning rules. The early warning information includes the risk level, main risk factors, and suggested emergency actions, ensuring that the information is comprehensive and accurate.

[0157] Assuming you receive electrocardiogram data and clinical medical records from a target user, process them according to the following steps:

[0158] ECG data is acquired from the target user using an electrocardiograph (ECG) machine, and then cleaned and standardized to extract ECG features such as R-wave peak value and QT interval, forming a feature vector set. This feature vector set is then passed to a physiological signal spatiotemporal coding module, which uses LSTM to extract the temporal relationships of the ECG data and CNN to extract spatial features. The extracted temporal relationships and spatial features are fused to form an intermediate feature representation, constructing a clinical knowledge graph related to cardiovascular and cerebrovascular diseases, including disease symptoms and risk factors. A graph embedding method is used to transform the knowledge graph into a low-dimensional vector space representation. The intermediate feature representation is then fused with the embedded vectors of the clinical knowledge graph to generate a fused feature representation. This fused feature representation is passed to a dynamic reinforcement learning module, which dynamically learns and optimizes based on the fused feature representation. By continuously adjusting model parameters, the prediction accuracy is improved. After training, the model is used to predict new ECG data and output the target user's health risk level.

[0159] Based on the target users' health status and risk level, they were categorized into "high-risk" levels. Analysis of the fused feature representations identified hypertension and unhealthy lifestyle habits as the main risk factors. Recommendations were developed for the target users, including improving lifestyle habits and regularly monitoring blood pressure. This information was integrated into a risk assessment report and sent to the target users. A warning threshold was set for the "medium-risk" level. Comparison of the target users' health status risk level with the warning threshold revealed that the risk level was higher than the warning threshold.

[0160] Based on preset warning rules, the system automatically generates warning information, including risk level, main risk factors (hypertension and unhealthy lifestyle habits), and recommended emergency actions (seek immediate medical attention and adjust lifestyle habits). The warning information is promptly released to target users and medical teams to enable emergency action to be taken.

[0161] By embedding physiological signals into spatiotemporal encoding and clinical knowledge graphs, the model can fully utilize the correlation between physiological signals and clinical knowledge, improving the accuracy and reliability of risk prediction. The dynamic reinforcement learning module can dynamically learn and optimize based on new physiological signal data and clinical knowledge, enhancing the model's adaptability and generalization ability. This allows the model to adapt to the physiological states and clinical conditions of different target users, improving its practicality and value. The three-level fusion model can perform personalized learning and prediction based on each target user's physiological signal data and clinical medical records. This helps promote the development of personalized medicine and health management, providing customized health risk assessment and early warning services for each target user, improving the efficiency and quality of medical services. The generation of risk assessment reports and early warning information enhances the target user's awareness and attention to health risks, helping them take proactive preventative measures to reduce these risks. When the risk level reaches or exceeds the early warning threshold, the timely release of early warning information prompts medical teams or relevant personnel to take swift emergency action, helping to prevent health deterioration and ensuring the safety of the target user's life. The generation of risk assessment reports and early warning information provides strong support for the rational allocation of medical resources, helping to achieve precise delivery of medical resources and improve the quality and efficiency of medical services.

[0162] In another preferred embodiment of the present invention, step 555 above, which involves dynamic learning and optimization based on the fused feature representation, and continuously adjusting to output the health status risk level of the target user, may include:

[0163] Step 5551: Initialize the dynamic reinforcement learning model, including setting the structure, parameters, and learning rate of the dynamic reinforcement learning model;

[0164] Step 5552: The dynamic reinforcement learning model receives the fused feature representation and performs dynamic learning according to the preset learning algorithm;

[0165] Step 5553: In each stage of learning, predict the health status risk level of a target user based on the current feature representation, compare it with the actual health status risk level, and calculate the difference between the prediction result and the actual result.

[0166] Step 5554: Repeat the learning and optimization process. When the preset learning stop condition is reached, output the health status risk level of the target user.

[0167] In this embodiment of the invention, a suitable dynamic reinforcement learning algorithm (such as DQN, DDPG, etc.) is selected according to the task requirements, and the network structure of the model is designed, including an input layer, a hidden layer, and an output layer. The input layer is used to receive the fused feature representation, the hidden layer is used to process and learn the features, and the output layer is used to output the predicted health status risk level. The number of neurons, activation function, and other parameters of each layer are determined. The model's weights and biases are initialized to ensure that the initialized parameters enable the model to have reasonable performance in the early stages of training. An appropriate learning rate is set to control the step size of parameter updates during training. The selection of the learning rate needs to be adjusted according to the task complexity and data scale to ensure that the model can converge within a reasonable time.

[0168] Step 5552 involves passing the fused feature representation as input data to the dynamic reinforcement learning model, ensuring that the format and dimensions of the input data match the requirements of the model's input layer. Based on a pre-defined learning algorithm (such as Q-learning, policy gradient, etc.), the model processes and learns from the input data. During the learning process, the model selects a policy based on the current state and action, and updates the model parameters to optimize prediction performance.

[0169] Step 5553: Use the current model parameters to predict the input data to obtain the target user's health status risk level, ensuring the reasonableness and accuracy of the prediction results and reflecting the target user's actual health status. Compare the prediction results with the actual health status risk level, calculate the difference between the prediction results and the actual results, and use this as the basis for model performance evaluation. Calculate the loss function (such as mean squared error) based on the difference to measure the model's predictive performance.

[0170] Step 5554: Based on the loss function and learning algorithm, continuously update the model parameters and repeat the learning process. During the learning process, gradually decrease the loss function value to improve the model's predictive performance. Set an appropriate learning stopping condition (such as reaching the maximum number of iterations). When the learning stopping condition is met, stop model training to avoid overfitting or excessively long training time. Use the trained model to predict new input data and output the health risk level of the target user, ensuring the rationality and accuracy of the output results and providing support for medical decision-making.

[0171] Assuming we have already received the fused feature representations, we now need to use a dynamic reinforcement learning model to predict the level of health risk. The following is the execution process:

[0172] The DDPG algorithm was selected as the dynamic reinforcement learning model, and a corresponding network structure was designed. The model's weights and biases were initialized, and the learning rate was set to 0.001. The received fused feature representations, including physiological signal features and clinical knowledge graph embedding vectors, were ensured to match the format and dimensionality of the input data with the requirements of the model's input layer. The input data was processed and learned using the DDPG algorithm. During the learning process, a strategy was selected based on the current state and actions, and the model parameters were updated to optimize prediction performance.

[0173] The model uses the current parameters to predict the health risk level of the target user from the input data. The predicted result is compared with the actual health risk level, and a loss function value is calculated. The model parameters are continuously updated based on the loss function and the learning algorithm, and training is repeated until the loss function converges or the maximum number of iterations is reached. The trained model is then used to predict new input data, outputting the health risk level of the target user. The prediction results are fed back to the healthcare system or the user to support medical decision-making.

[0174] Dynamic reinforcement learning models can dynamically learn and optimize based on the fused feature representations, improving prediction accuracy. Through continuous iterative training, the model can gradually approximate the real situation, providing a more reliable basis for medical decision-making. Dynamic reinforcement learning models can adapt to different physiological signal data and clinical knowledge graph embedding vectors, enhancing the model's adaptability. This allows the model to be applied in different scenarios, providing more users with personalized health risk assessment services. Accurate prediction of health risk levels can provide strong support for medical decision-making, enabling doctors to formulate more scientific and reasonable treatment plans based on the prediction results, improving treatment effectiveness and patient satisfaction. The application of dynamic reinforcement learning models in predicting health risk levels promotes the development of the healthcare field, contributing to the rational allocation and efficient utilization of medical resources and improving the quality and efficiency of medical services.

[0175] like Figure 2 As shown, embodiments of the present invention also provide a cardiovascular and cerebrovascular disease risk prediction system, comprising:

[0176] The data acquisition module is used to acquire physiological signal data of the target user;

[0177] The feature extraction module is used to extract features from the physiological signal data and form a feature vector set;

[0178] The feature selection and optimization module is used to select and optimize the feature vector set using the krill swarm algorithm, and automatically search for and determine the feature subset that is key to cardiovascular health assessment.

[0179] The three-level fusion module is used to perform deep, multi-dimensional fusion processing on the optimized feature vector set to obtain the fused feature representation.

[0180] The risk prediction module is used to predict the health status risk level of the target user based on the feature representation output by the three-level fusion module, obtain the risk prediction result, and automatically generate early warning information when the health status risk level of the target user exceeds the preset threshold.

[0181] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0182] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0183] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the risk of cardiovascular and cerebrovascular diseases, characterized in that, The method includes: Acquire real-time physiological parameters of the target user, including dynamic blood pressure variability, serum lipoprotein levels, heart rate variability frequency domain index, and sleep apnea-hypopnea index; Real-time physiological parameters are preprocessed, and wavelet transform is used to extract the instantaneous waveform features of the blood pressure signal. Features related to health status, including blood pressure level, blood lipid level, blood glucose level, abnormal electrocardiogram indicators and abnormal echocardiogram indicators, are extracted to form a feature vector set. The krill swarm algorithm is used to select and optimize the feature vector set. By simulating the foraging behavior of krill swarms, a subset of features key to cardiovascular health assessment is automatically searched to obtain the optimized feature vector set, including: Determine the size of the krill colony, i.e. the number of candidate features participating in feature selection, and assign an initial position to each krill; Set the krill's sensing range, movement step size, and maximum number of iterations, and define a contribution function. Based on the contribution function, calculate the contribution of the krill's current position to the cardiovascular health assessment. For each krill, extract the corresponding features based on its current location to build a prediction model; The prediction model was trained using sample data containing cardiovascular health conditions, and the contribution value of each krill was calculated based on the contribution function. The contribution value is assigned to the corresponding krill as a performance evaluation in the current iteration. For each krill, the set of neighbors within its perception range is calculated, and the learning object of the current krill is determined in the set of neighbors. Update the current krill position based on the learning target, the current krill position difference, and the preset movement step size; The process of feature extraction, model building, training, and contribution calculation is repeated until the preset maximum number of iterations is reached, ultimately resulting in a set of optimized feature combinations. Based on the optimized feature combination, the corresponding features are extracted from the original feature vector set to form the optimized feature vector set. Based on the optimized feature vector set, a three-level fusion model is constructed using probabilistic temporal Petri nets, incorporating spatiotemporal encoding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning. This model includes: Initialize the construction environment for the probabilistic time Petri net, including defining the net's state, transitions, arcs, and associated probability and time parameters; The blood pressure waveform features and heart rate variability physiological signal data in the optimized feature vector set are encoded and fused to form an intermediate feature representation that includes dynamic changes in physiological signals, spatial relationships and clinical knowledge; Through dynamic reinforcement learning algorithms, intermediate feature representations containing dynamic changes in physiological signals, spatial relationships, and clinical knowledge are used as input data for learning and optimization. After multiple iterations of training, a three-level fusion model is formed, which includes spatiotemporal coding of physiological signals, embedding of clinical knowledge graphs, and dynamic reinforcement learning. The feature vector of the target user is input into the three-level fusion model. According to the inference rules of Petri net, the health status risk level of the target user is obtained. Based on the health status risk level of the target user, the risk assessment result and early warning information are output.

2. The method for predicting the risk of cardiovascular and cerebrovascular diseases according to claim 1, characterized in that, Real-time physiological parameters are preprocessed, and wavelet transform is used to extract the instantaneous waveform features of the blood pressure signal. Features related to health status are also extracted, including blood pressure level, blood lipid level, blood glucose level, abnormal electrocardiogram indicators, and abnormal echocardiogram indicators, forming a feature vector set, including: Wavelet transform was used to extract instantaneous waveform features from blood pressure signals in order to obtain the dynamic characteristics of blood pressure changes. Features related to health status are extracted from the data after wavelet transform processing, including blood pressure level, blood lipid level, blood glucose level, abnormal electrocardiogram indicators, and abnormal echocardiogram indicators, forming a feature vector set.

3. The method for predicting the risk of cardiovascular and cerebrovascular diseases according to claim 2, characterized in that, The contribution of the krill's current location to the cardiovascular health assessment is calculated based on the contribution function, including: When constructing a prediction model based on the feature subset corresponding to the current location of the krill, the number of samples correctly predicted as positive, the number of samples correctly predicted as negative, the number of samples incorrectly predicted as positive, and the number of samples incorrectly predicted as negative are calculated; and the accuracy index value is calculated based on the number of samples correctly predicted as positive, the number of samples correctly predicted as negative, the number of samples incorrectly predicted as positive, and the number of samples incorrectly predicted as negative. Determine the size of the feature subset corresponding to the current location of the krill, i.e. the number of features contained in the feature subset, and calculate the feature subset factor index value based on the number of features contained in the feature subset. Calculate the entropy of the entire dataset and the entropy of the data subsets corresponding to different values ​​of each feature. For each feature, calculate the weighted average of the entropy of the data subsets corresponding to all values. Calculate the difference between the entropy of the entire dataset and the weighted average, and calculate the ratio between the difference and the maximum information gain among all features. Based on the ratio, calculate the information gain impact value. The accuracy index, feature subset factor index, and information gain influence value are fused to determine the contribution of the krill's current location to the assessment of cardiovascular health status.

4. The method for predicting the risk of cardiovascular and cerebrovascular diseases according to claim 3, characterized in that, The three-level fusion model includes: The first level is the physiological signal spatiotemporal coding module, which performs spatiotemporal coding on the physiological signals in the optimized feature vector set to obtain the temporal relationship and spatial features between the signals; The second level is the clinical knowledge graph embedding module, which integrates clinical knowledge with feature vectors and transforms clinical knowledge into structured information. The third level is the dynamic reinforcement learning module, which dynamically learns based on the outputs of the physiological signal spatiotemporal coding module and the clinical knowledge graph embedding module, and finally outputs the health status risk level of the target user.

5. The method for predicting the risk of cardiovascular and cerebrovascular diseases according to claim 4, characterized in that, The feature vectors of the target users are input into the three-level fusion model. Based on the inference rules of the Petri net, the health risk level of the target users is obtained. Then, based on the health risk level, the risk assessment results and early warning information are output, including: The feature vector of the target user is used as input data and passed to the first level of the three-level fusion model, namely the physiological signal spatiotemporal coding module. In the spatiotemporal coding module of physiological signals, the temporal relationship and spatial features between physiological signals are extracted to form intermediate feature representations; The intermediate feature representation is passed to the second level, namely the clinical knowledge graph embedding module, where clinical knowledge is fused with the intermediate feature representation to generate a fused feature representation. The fused feature representation is passed to the third level, namely the dynamic reinforcement learning module, which performs dynamic learning and optimization based on the fused feature representation. Through continuous adjustment, the target user's health status risk level is output. Based on the target users' health status and risk level, generate corresponding risk assessment reports, including risk level, risk factor analysis, and recommended measures; Based on the target user's health status and risk level, an early warning threshold is set. The target user's health status and risk level are compared with the early warning threshold. When the target user's health status and risk level are greater than or equal to the early warning threshold, an early warning message is automatically generated according to the preset early warning rules, including the risk level, main risk factors, and suggested emergency actions.

6. The method for predicting the risk of cardiovascular and cerebrovascular diseases according to claim 5, characterized in that, Dynamic learning is performed based on the outputs of the physiological signal spatiotemporal coding module and the clinical knowledge graph embedding module to ultimately output the target user's health status risk level, including: Initialize the dynamic reinforcement learning model, including setting the structure, parameters, and learning rate of the dynamic reinforcement learning model; The dynamic reinforcement learning model receives the fused feature representations and performs dynamic learning according to a preset learning algorithm; At each stage of learning, the health risk level of a target user is predicted based on the current feature representation and compared with the actual health risk level, and the difference between the predicted result and the actual result is calculated. The learning and optimization process is repeated until the preset learning stop condition is met, at which point the risk level of the target user's health status is output.

7. A cardiovascular and cerebrovascular disease risk prediction system, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The data acquisition module is used to acquire physiological signal data of the target user; The feature extraction module is used to extract features from the physiological signal data and form a feature vector set; The feature selection and optimization module is used to select and optimize the feature vector set using the krill swarm algorithm, and automatically search for and determine the feature subset that is key to cardiovascular health assessment. The three-level fusion module is used to perform deep, multi-dimensional fusion processing on the optimized feature vector set to obtain the fused feature representation. The risk prediction module is used to predict the health status risk level of the target user based on the feature representation output by the three-level fusion module, obtain the risk prediction result, and automatically generate early warning information when the health status risk level of the target user exceeds the preset threshold.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images