A method for periodic assessment of cardiovascular disease risk in a patient
By constructing a multi-source longitudinal health data fusion framework and a time-aware deep state space model, the problem that existing cardiovascular disease risk assessment models cannot dynamically track the evolution of individual risk has been solved. This enables sensitive identification and precise intervention of risk inflection points, thereby improving the effectiveness of cardiovascular disease prevention.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE 900TH HOSPITAL OF THE CHINESE PEOPLES LIBERATION ARMY JOINT LOGISTICS SUPPORT FORCE
- Filing Date
- 2026-01-19
- Publication Date
- 2026-05-29
Smart Images
Figure CN122117360A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of biomedicine, specifically relating to a method for assessing the cardiovascular disease risk cycle in patients. Background Technology
[0002] With cardiovascular disease becoming the leading cause of death globally, personalized risk assessment is playing an increasingly important role in clinical preventive medicine. Current mainstream risk prediction models, such as the Framingham Risk Score and atherosclerotic cardiovascular disease (ASCVD) risk assessment tools, primarily rely on cross-sectional epidemiological data to construct static regression equations, outputting a 10-year risk probability at a fixed time point. These methods simplify complex physiological evolution into a linear combination of several baseline variables, neglecting the dynamic adaptation and decompensation mechanisms of the human cardiovascular system under long-term exposure to metabolic, behavioral, and environmental factors. Especially in chronic disease management, a patient's risk status is not static but fluctuates non-linearly with lifestyle interventions, drug treatments, or natural disease progression, and may even reach critical turning points—inflection points where risk accelerates or significantly reverses. However, existing models lack the ability to model the temporal dimension of continuous physiological signals, failing to capture such micro-dynamic changes, leading to delayed warnings and missed intervention opportunities.
[0003] Dynamic assessment of cardiovascular risk based on time-series physiological data has gradually become a research hotspot. This research aims to construct a longitudinal trajectory reflecting the evolution of an individual's health status by continuously collecting multi-source physiological indicators (such as ambulatory blood pressure, heart rate variability, continuous blood glucose, and activity levels) using wearable devices or remote monitoring systems. Its core objective is to go beyond static snapshot-based assessment and achieve real-time tracking and prospective extrapolation of risk evolution trends. However, existing attempts are mostly limited to simple sliding window statistics or traditional time series regression, which are difficult to effectively model long-term dependencies and sudden state transitions, and even more difficult to identify clinically significant risk inflection points.
[0004] When processing high-dimensional, heterogeneous, and non-stationary physiological time-series data, traditional methods generally suffer from insufficient model expressive power, sluggish response to nonlinear dynamics, and a lack of simulation and extrapolation mechanisms for intervention effects. Especially in real-world monitoring scenarios with significant individual differences and strong noise interference, traditional methods struggle to extract early warning signals from subtle but persistent shifts in physiological patterns. Furthermore, due to the lack of mature algorithmic frameworks similar to those used in the financial sector for quantifying market turning points, current clinical tools cannot predict substantial changes in risk status before routine biochemical indicators show abnormalities, nor can they simulate the impact of different intervention strategies on future risk trajectories like stress tests. Summary of the Invention
[0005] This invention provides a method for assessing the cyclical risk of cardiovascular disease in patients, aiming to address the technical problems of traditional risk assessment models that rely solely on cross-sectional data, cannot dynamically track the nonlinear evolution of individual cardiovascular risk, and lack the ability to sensitively identify key inflection points where risk states significantly accelerate or reverse. This method constructs an individual-centric multi-source longitudinal health data fusion framework and introduces a time-aware deep state-space modeling mechanism to achieve continuous, high-resolution, and personalized cyclical assessment of cardiovascular risk trajectories.
[0006] This invention provides a method for assessing the cardiovascular disease risk cycle in patients, comprising: acquiring a multi-source longitudinal health data sequence of a target patient, wherein the multi-source longitudinal health data sequence includes time-series data of physiological indicators from previous physical examination records, time-series data of laboratory test results, time-series data of medication history, time-series data of lifestyle records, and markers of past cardiovascular events; standardizing and imputing missing values in the multi-source longitudinal health data sequence to generate a structured longitudinal feature matrix with uniform time granularity; inputting the structured longitudinal feature matrix into a pre-trained time-aware deep state space model, wherein the time-aware deep state space model is configured to infer continuous latent variables of implicit physiological states in the input sequence and output a state trajectory vector sequence characterizing the intrinsic dynamic characteristics of the cardiovascular system; calculating a risk evolution curvature index based on the state trajectory vector sequence, wherein the risk evolution curvature index is used to quantify the local rate and direction of change of the current risk trajectory; comparing the risk evolution curvature index with a preset inflection point judgment threshold to identify a risk acceleration period, a risk plateau period, or a risk reversal period; generating corresponding risk level labels and clinical intervention suggestions based on the identified risk cycle stage, and outputting a visualized risk cycle map.
[0007] As one embodiment of the present invention, the acquisition of multi-source longitudinal health data sequences of the target patient specifically includes: extracting all outpatient and inpatient records of the target patient from the electronic health record system since the initial registration; acquiring data on the target patient's continuously monitored heart rate variability, resting heart rate, sleep quality, and daily activity levels from the wearable device data interface; synchronizing the target patient's lipid profile, glycated hemoglobin, high-sensitivity C-reactive protein, homocysteine, and glomerular filtration rate from the regional testing center database; acquiring the names, dosages, start and end dates, and durations of antihypertensive, lipid-lowering, antiplatelet, and hypoglycemic drugs used by the target patient from the pharmacy management system; and acquiring the target patient's periodically reported smoking status, drinking frequency, dietary pattern scores, and self-reported psychological stress scores from the structured questionnaire system.
[0008] As one embodiment of the present invention, the standardization and missing value imputation processing of the multi-source longitudinal health data sequence to generate a structured longitudinal feature matrix with unified time granularity specifically includes: performing Z-score standardization based on population distribution on all numerical features; converting categorical variables into binary vectors using one-hot encoding; aggregating heterogeneous data streams with inconsistent timestamps using 7 days as the basic time unit, and using a forward imputation combined with a linear interpolation strategy to handle missing values within the unit; marking feature dimensions without effective observations for more than 30 consecutive days as long-term missing and retaining the time mask; and finally forming a two-dimensional matrix with row indices as time units, column indices as standardized feature dimensions, and elements as real values.
[0009] In one embodiment of the present invention, the time-aware deep state-space model includes an input embedding layer, a continuous-time state evolution module, an observation decoder, and a state posterior inference network. The input embedding layer maps the feature vector of each time unit to a high-dimensional latent space representation. The continuous-time state evolution module uses neural ordinary differential equations to perform differential evolution of the latent state in the continuous time domain. Its state transition function is composed of a multilayer perceptron, and the input is the current latent state and the absolute timestamp. The observation decoder maps the latent state back to the original feature space to reconstruct the input, which is used as a supervision signal for model training. The state posterior inference network adopts a bidirectional recurrent neural network structure, encodes the observation sequence from both forward and backward directions, and generates the state posterior distribution parameters at each time point. In the inference phase, only the forward state evolution path is used for online risk assessment.
[0010] As one embodiment of the present invention, each state vector dimension in the state trajectory vector sequence characterizing the intrinsic dynamic characteristics of the cardiovascular system corresponds to the potential load level of a physiological subsystem, including the vascular elastic reserve dimension, metabolic homeostasis dimension, inflammation activation dimension, neural regulation dimension, and coagulation tendency dimension; the numerical range of each dimension is constrained to the interval between 0 and 1, and the higher the value, the greater the degree of deviation of the subsystem from healthy homeostasis.
[0011] As one embodiment of the present invention, the calculation of the risk evolution curvature index specifically includes: performing three spline smoothing fits on each dimension of the state trajectory vector sequence to obtain a smooth state trajectory function; obtaining the first derivative of the state trajectory function to obtain the instantaneous rate of change, and then obtaining the second derivative to obtain the instantaneous acceleration; weighting and fusing the instantaneous accelerations of each dimension, with the weighting coefficients pre-set according to the contribution of each physiological subsystem to the overall cardiovascular event; the weighted fusion result is the risk evolution curvature index.
[0012] As one embodiment of the present invention, the preset inflection point determination threshold includes a risk acceleration threshold and a risk reversal threshold; when the risk evolution curvature index is greater than the risk acceleration threshold for three consecutive time units, it is determined to enter the risk acceleration period; when the risk evolution curvature index is less than the risk reversal threshold for three consecutive time units and the current risk level is higher than the benchmark level, it is determined to enter the risk reversal period; the rest are determined to be the risk plateau period.
[0013] As one embodiment of the present invention, the generation of corresponding risk level labels and clinical intervention recommendations specifically includes: during the risk acceleration period, outputting a high-risk level label and recommending intensified drug treatment, increased follow-up frequency, and initiation of a lifestyle intervention program; during the risk plateau period, outputting a medium-risk or low-risk level label and recommending maintaining the existing treatment plan and routine annual assessment; during the risk reversal period, outputting an improvement trend label and recommending assessing the possibility of drug reduction and consolidating healthy behaviors.
[0014] As one embodiment of the present invention, the output visualization risk cycle map specifically includes: plotting a multi-curve time series graph with time as the horizontal axis and the comprehensive risk score and the load level of each physiological subsystem as the vertical axis; marking the boundaries of the identified risk cycle stages in the graph; distinguishing the acceleration phase, plateau phase and reversal phase with different colored background areas; and overlaying and displaying the time markers of key intervention events, including medication adjustments, major life events and clinical diagnoses.
[0015] The loss function of the time-aware deep state space model is composed of a weighted sum of a reconstruction error term, a state smoothing regularization term, and an event prediction cross-entropy term. The reconstruction error term uses the mean square error to measure the difference between the decoder output and the original input. The state smoothing regularization term applies an L2 norm penalty to the first derivative of the state trajectory; The event prediction cross-entropy term predicts the probability of major adverse cardiovascular events occurring within the next 5 years by attaching an event prediction header to the end of the state trajectory, and calculates the cross-entropy with the actual event labels.
[0016] This invention also provides a cardiovascular disease risk cycle assessment system for patients, comprising: a multi-source longitudinal health data acquisition unit for acquiring multi-source longitudinal health data sequences of a target patient; a data preprocessing unit for standardizing and imputing missing values in the multi-source longitudinal health data sequences to generate a structured longitudinal feature matrix with uniform time granularity; a deep state space modeling unit for inputting the structured longitudinal feature matrix into a pre-trained time-aware deep state space model and outputting a sequence of state trajectory vectors representing the intrinsic dynamic characteristics of the cardiovascular system; a risk evolution curvature calculation unit for calculating a risk evolution curvature index based on the state trajectory vector sequence; a risk cycle identification unit for comparing the risk evolution curvature index with a preset inflection point judgment threshold to identify a risk acceleration phase, a risk plateau phase, or a risk reversal phase; and a risk assessment output unit for generating corresponding risk level labels and clinical intervention recommendations based on the identified risk cycle stages, and outputting a visualized risk cycle map.
[0017] In one embodiment of the present invention, the multi-source longitudinal health data acquisition unit is connected to an electronic health record system, a wearable device data platform, a regional testing center database, a pharmacy management system, and a structured questionnaire system, and extracts data through an interface protocol that conforms to the medical information exchange standard.
[0018] In one embodiment of the present invention, the deep state space modeling unit is deployed on a server cluster with graphics processor acceleration capabilities. Its model parameters are trained end-to-end using a large-scale real-world cohort study dataset. The loss function is composed of a weighted sum of a reconstruction error term, a state smoothing regularization term, and an event prediction cross-entropy term.
[0019] As one embodiment of the present invention, the risk assessment output unit is integrated into the clinical decision support system, and its output results are pushed to the attending physician's workstation in a structured data format, and it supports linkage with the medical order module in the hospital information system.
[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention abandons the reliance of traditional static risk scoring models on single-point-of-time data and constructs a technical framework that can continuously track the dynamic evolution of an individual's cardiovascular health status.
[0021] 2. By fusing multi-source longitudinal health data and employing a time-aware deep state-space model, this invention can reconstruct the intrinsic physiological state trajectory of an individual's cardiovascular system with high precision, thereby capturing key inflection points of risk acceleration or reversal that cannot be identified by traditional methods.
[0022] 3. The introduction of the risk evolution curvature index provides clinical practice with an objective tool to quantify the rate of dynamic change in risk, enabling the selection of intervention timing to shift from experience-based judgment to data-driven decision-making.
[0023] 4. The clear division of risk cycles and corresponding intervention recommendations significantly improve the accuracy and timeliness of cardiovascular disease prevention, helping to effectively block risks before they actually turn into clinical events, thereby reducing overall morbidity and medical costs. Attached Figure Description
[0024] Figure 1 This is a schematic diagram of the overall technical solution architecture of a method for assessing the cardiovascular disease risk cycle in patients, as proposed in this invention. Figure 2 This is a schematic diagram of the core principle framework of the time-aware depth state space model in this invention; Figure 3 This is a logical flowchart of the acquisition of multi-source longitudinal health data and the construction of structured feature matrices in this invention; Figure 4 This is a flowchart illustrating the logical process framework for calculating the risk evolution curvature index and identifying the risk cycle in this invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the load dimensions of each physiological subsystem and the comprehensive risk score in this invention; Figure 6 This is a logical flowchart of the process for generating a visualized risk cycle map and outputting clinical intervention recommendations in this invention. Detailed Implementation
[0025] Please refer to the attached document. Figure 1 Hertz Figure 6 This invention provides a method for assessing the cardiovascular disease risk cycle in patients. Its core lies in constructing an individual-centered multi-source longitudinal health data fusion framework and introducing a time-aware deep state-space modeling mechanism to achieve continuous, high-resolution, and personalized periodic assessment of cardiovascular risk trajectories. The method acquires multi-source longitudinal health data sequences from target patients, performs standardization and missing value imputation to generate a structured longitudinal feature matrix with uniform time granularity, inputs this matrix into a pre-trained time-aware deep state-space model, and outputs a sequence of state trajectory vectors representing the intrinsic dynamic characteristics of the cardiovascular system. Based on this, a risk evolution curvature index is calculated, and compared with a preset inflection point threshold to identify risk acceleration phases, risk plateau phases, or risk reversal phases. Finally, based on the identified risk cycle stages, corresponding risk level labels and clinical intervention recommendations are generated, and a visualized risk cycle map is output.
[0026] The acquisition of multi-source longitudinal health data sequences for the target patients includes: extracting all outpatient and inpatient records from the initial registration of the target patients from the electronic health record system; acquiring continuously monitored heart rate variability, resting heart rate, sleep quality, and daily activity data from wearable device data interfaces; synchronizing the target patients' lipid profile, glycated hemoglobin, high-sensitivity C-reactive protein, homocysteine, and glomerular filtration rate test results from the regional testing center database; acquiring the names, dosages, start and end dates, and durations of antihypertensive, lipid-lowering, antiplatelet, and hypoglycemic drugs used by the target patients from the pharmacy management system; and acquiring the target patients' periodically submitted smoking status, alcohol consumption frequency, dietary pattern scores, and self-reported psychological stress scores from the structured questionnaire system. All data are accurately timestamped, with a timestamp accuracy of at least daily levels, and some physiological signal data, such as heart rate variability, are sampled at minute-level frequencies. Each data source is extracted through an interface protocol conforming to medical information exchange standards to ensure data integrity and consistency. The data extraction process includes identity desensitization, permission verification, and audit log recording to meet medical data security compliance requirements.
[0027] After acquiring the aforementioned multi-source longitudinal health data sequences, data preprocessing steps were performed. Specifically, this included Z-score standardization based on population distribution for all numerical features; that is, for each numerical indicator, the original observations were converted into Z-scores under a standard normal distribution based on the mean and standard deviation of the large-scale reference population; categorical variables were converted into binary vectors using one-hot encoding, for example, drug categories such as statins, ACEIs, and ARBs were mapped to independent binary dimensions; and heterogeneous data streams with inconsistent timestamps were aggregated using 7-day periods as the basic time unit. If multiple observations existed within each time unit, their arithmetic mean was taken as the unit value. Representative values; for missing values within a cell, a forward imputation combined with linear interpolation strategy is used for imputation, that is, the previous valid observation value is used first for imputation, if there is no valid value in the forward direction, the valid value in the backward direction is used for reverse imputation, and if there are valid values in both directions, linear interpolation is used; for feature dimensions without valid observations for more than 30 consecutive days, they are marked as long-term missing and a time mask is retained. This mask is used for explicit processing of missing information in subsequent model inference; finally, a two-dimensional matrix is formed with row indices representing time cells, column indices representing standardized feature dimensions, and elements representing real values. This matrix is the structured vertical feature matrix, which has fixed dimensions and covers all physiological, biochemical, behavioral, and treatment-related variables.
[0028] The structured longitudinal feature matrix is input into a pre-trained time-aware deep state-space model. This model includes an input embedding layer, a continuous-time state evolution module, an observation decoder, and a state posterior inference network. The input embedding layer maps the feature vector of each time unit to a high-dimensional latent space representation. This mapping is achieved through a fully connected neural network, and its weights are determined through end-to-end optimization during model training. The continuous-time state evolution module uses ordinary differential equations to perform differential evolution of the latent state in the continuous time domain. Its state transition function is composed of a multilayer perceptron. The input is the current latent state and the absolute timestamp, and the output is the state derivative, which is used to advance the latent state through an ordinary differential equation solver such as the Runge-Kutta method. The observation decoder maps the latent state back to the original feature space to reconstruct the input. Its structure is another fully connected network, which is used to provide reconstruction error as a supervision signal during the training phase. The state posterior inference network adopts a bidirectional recurrent neural network structure, which encodes the observation sequence in both forward and backward directions to generate state posterior distribution parameters at each time point, including the mean and diagonal covariance. During the inference phase, only the forward state evolution path is used for online risk assessment to avoid dependence on future data and ensure that the method is applicable to real-time clinical scenarios.
[0029] The output of the time-aware deep state-space model is a sequence of state trajectory vectors. Each dimension of this sequence corresponds to a potential load level of a physiological subsystem, including dimensions of vascular elastic reserve, metabolic homeostasis, inflammation activation, neural regulation, and coagulation tendency. The numerical range of each dimension is constrained to the interval between 0 and 1; higher values indicate a greater deviation of the subsystem from healthy homeostasis. This constraint is achieved by applying a sigmoid activation function to the output layer of the state evolution module. The length of the state trajectory vector sequence is equal to the number of time units in the structured longitudinal feature matrix, and the state vector at each time point reflects the overall intrinsic physiological state of the cardiovascular system at that moment.
[0030] Based on the state trajectory vector sequence, the risk evolution curvature index is calculated. This calculation process first performs a third-order spline smoothing fit on each dimension of the state trajectory vector sequence to obtain a smooth state trajectory function. Natural boundary conditions are used in this fitting process to avoid endpoint oscillations. Then, the first derivative of the state trajectory function is calculated to obtain the instantaneous rate of change, and the second derivative is calculated to obtain the instantaneous acceleration. The instantaneous accelerations of each dimension are weighted and fused using pre-set weighting coefficients. These weighting coefficients are determined based on the contribution of each physiological subsystem to the overall cardiovascular event. This contribution is obtained through multivariate Cox regression analysis in large-scale cohort studies. For example, the weight for the vascular elastic reserve dimension is 0.35, the weight for the metabolic homeostasis dimension is 0.25, the weight for the inflammation activation dimension is 0.20, the weight for the neural regulation dimension is 0.10, and the weight for the coagulation tendency dimension is 0.10. The weighted fusion result is the risk evolution curvature index, which physically represents the degree of local curvature of the comprehensive risk trajectory at the current time point. A positive value indicates risk acceleration, and a negative value indicates risk deceleration or reversal.
[0031] The risk evolution curvature index is used for risk cycle identification. Preset inflection point thresholds include a risk acceleration threshold and a risk reversal threshold, both empirically set values: the risk acceleration threshold is 0.08, and the risk reversal threshold is -0.06. When the risk evolution curvature index is greater than the risk acceleration threshold for three consecutive time units, the risk is considered to be in an acceleration phase; when the risk evolution curvature index is less than the risk reversal threshold for three consecutive time units and the current risk level is higher than a baseline level, the risk is considered to be in a reversal phase; all other cases are considered risk plateau phases. The risk level baseline is defined as a comprehensive state where the mean of each dimension of the state trajectory vector is less than 0.4. This judgment logic is implemented using a sliding window mechanism, with a fixed window length of three time units. The judgment result is updated every time a time unit is advanced.
[0032] Based on the identified risk cycle stages, corresponding risk level labels and clinical intervention recommendations are generated. During the risk acceleration phase, a high-risk label is output, and recommendations include strengthening medication, increasing follow-up frequency, and initiating a lifestyle intervention program. Specific recommendations include adjusting the dosage of antihypertensive or lipid-lowering medications, scheduling follow-up blood lipid and blood pressure checks within 3 months, and providing personalized diet and exercise plans. During the risk plateau phase, a medium-risk or low-risk label is output, and recommendations include maintaining the existing treatment plan and routine annual assessments. The mean value of the status trajectory for medium-risk is between 0.4 and 0.6, while the mean value for low-risk is below 0.4. During the risk reversal phase, an improvement trend label is output, and recommendations include assessing the possibility of medication reduction and reinforcing healthy behaviors. Specific recommendations include gradually reducing the types or dosages of medications under the guidance of a physician, extending the follow-up interval to one year, and strengthening positive feedback mechanisms for healthy behaviors.
[0033] The method also includes outputting a visualized risk cycle map. This map plots a multi-curve time series graph with time on the horizontal axis and the overall risk score and the load level of each physiological subsystem on the vertical axis. The boundaries of identified risk cycle stages are marked on the graph, with the boundary positions determined by inflection point judgment logic. Different colored background areas distinguish acceleration, plateau, and reversal phases; for example, red represents acceleration, gray represents plateau, and green represents reversal. Time stamps of key intervention events are overlaid, including medication adjustments, major life events, and clinical diagnoses, which are automatically extracted from medical records and diagnostic codes in the electronic health record system. The map generation process is completed by a front-end rendering engine, supporting interactive zooming and data exploration, facilitating clinicians' intuitive understanding of patient risk dynamics.
[0034] The training process of the time-aware deep state-space model is based on a large-scale real-world cohort study dataset containing over 100,000 person-years of cardiovascular health follow-up data, covering complete multi-source longitudinal observations and endpoint event records. The model loss function consists of three weighted terms: a reconstruction error term using mean squared error to measure the difference between the observed decoder output and the original input; a state smoothing regularization term applying an L2 norm penalty to the first derivative of the state trajectory to encourage smoothness in the state evolution path; and an event prediction cross-entropy term appending an event prediction header to the end of the state trajectory to predict the probability of major adverse cardiovascular events occurring within the next 5 years, and calculating the cross-entropy with the real event labels. The weights of the three terms are 1.0, 0.1, and 0.5, respectively, and were optimized on the validation set using grid search. Model training was performed on a server cluster with GPU acceleration capabilities, using the Adam optimizer, with an initial learning rate of 0.001, a batch size of 128, and 200 training epochs.
[0035] During the model deployment phase, the deep state-space modeling unit is integrated into the clinical decision support system as a microservice. Its input is a structured longitudinal feature matrix processed by the preprocessing unit, and its output is a sequence of state trajectory vectors. This microservice receives requests via an application programming interface (API) and returns results in JSON format, including state vectors at each time point, risk evolution curvature indices, and risk cycle stage labels. The risk assessment output unit receives these results, combines them with a pre-defined intervention rule base, generates structured clinical recommendations, and pushes them to the attending physician's workstation in a standardized message format. This output also supports integration with the medical order module in the hospital information system; for example, it can automatically trigger medication adjustment reminders when a risk acceleration period is identified.
[0036] In practical application, the method first involves the system administrator configuring patient identifiers to trigger a multi-source data extraction process. The data acquisition unit then accesses electronic health records, wearable device platforms, laboratory databases, pharmacy systems, and questionnaire systems in parallel, aggregating all historical records. The preprocessing unit cleans, standardizes, aligns, and imputs the raw data, generating a structured longitudinal feature matrix. The modeling unit loads pre-trained model parameters, performs forward inference, and outputs a state trajectory. The curvature calculation unit performs differential operations on the state trajectory to obtain a risk evolution curvature index. The cycle identification unit applies sliding window judgment logic to determine the current risk cycle. The output unit integrates all information to generate risk level labels, intervention suggestions, and a visualization map. The entire process is automatically triggered each time a patient adds new health data, achieving dynamic and continuous risk assessment.
[0037] The mathematical expression of the risk evolution curvature index is as follows. Let the first... The state trajectory function of each physiological subsystem is: Its second derivative is The weighting coefficient is The risk evolution curvature index Defined as:
[0038] in, The values range from 1 to 5, corresponding to five dimensions: vascular elastic reserve, metabolic homeostasis, inflammation activation, neural regulation, and coagulation tendency.
[0039] The state trajectory function The node is obtained by cubic spline interpolation. discrete time points State observations at the location The spline coefficients are determined by minimizing the following objective function:
[0040] in, This is a smoothing parameter with a value of 0.1, used to balance fitting accuracy and curve smoothness.
[0041] The state evolution process of the aforementioned constant differential equation is described by the following differential equation:
[0042] in, for The hidden state vector at time t. Let be the state transition function composed of a multilayer sensing mechanism. These are its learnable parameters.
[0043] The output of the state posterior inference network is at each time point. State posterior distribution parameters and The calculation formula is as follows:
[0044] in, For the entire observation sequence, It is a bidirectional recurrent neural network. Its parameters.
[0045] In summary, this invention, by fusing multi-source longitudinal health data to construct a time-aware deep state-space model, achieves high-precision reconstruction and dynamic analysis of individual cardiovascular risk trajectories. The introduction of the risk evolution curvature index transforms the identification of risk inflection points from qualitative judgment to quantitative analysis, significantly improving the accuracy and timeliness of cardiovascular disease prevention. This method is entirely data-driven, requiring no reliance on expert experience or rules, and is suitable for large-scale population health management and personalized clinical decision support.
Claims
1. A method for periodic assessment of cardiovascular disease risk in patients, characterized in that, include: Acquire multi-source longitudinal health data sequences of the target patient, including time-series data of physiological indicators, laboratory test results, medication history, lifestyle records, and markers of past cardiovascular events from previous physical examination records. The multi-source longitudinal health data sequence is standardized and missing value imputation is performed to generate a structured longitudinal feature matrix with a unified time granularity; The structured longitudinal feature matrix is input into a pre-trained time-aware deep state space model, which is configured to perform continuous latent variable inference on the implicit physiological states in the input sequence and output a sequence of state trajectory vectors characterizing the intrinsic dynamic characteristics of the cardiovascular system. Based on the state trajectory vector sequence, a risk evolution curvature index is calculated, which is used to quantify the local rate and direction of change of the current risk trajectory. The risk evolution curvature index is compared with a preset inflection point judgment threshold to identify the risk acceleration period, risk plateau period, or risk reversal period. Based on the identified risk cycle stages, corresponding risk level labels and clinical intervention recommendations are generated, and a visualized risk cycle map is output.
2. The method for periodic assessment of cardiovascular disease risk in patients according to claim 1, characterized in that, The acquisition of the multi-source longitudinal health data sequence of the target patient includes: Extract all outpatient and inpatient records of the target patient from the time of initial registration from the electronic health record system; Data on heart rate variability, resting heart rate, sleep quality, and daily activity levels of the target patient are obtained from the wearable device data interface. Synchronize the target patient's lipid profile, glycated hemoglobin, high-sensitivity C-reactive protein, homocysteine, and glomerular filtration rate from the regional testing center database. Obtain the names, dosages, start and end dates, and durations of antihypertensive, lipid-lowering, antiplatelet, and hypoglycemic drugs used by the target patient from the pharmacy management system; The target patients' smoking status, drinking frequency, dietary pattern scores, and self-reported psychological stress scores were obtained from a structured questionnaire system.
3. The method for periodic assessment of cardiovascular disease risk in patients according to claim 2, characterized in that, The standardization and missing value imputation processing of the multi-source longitudinal health data sequences to generate a structured longitudinal feature matrix with uniform temporal granularity includes: Perform Z-score standardization based on population distribution on all numerical features; Categorical variables are converted into binary vectors using one-hot encoding; For heterogeneous data streams with inconsistent timestamps, aggregation is performed using 7 days as the basic time unit, and a forward filling combined with linear interpolation strategy is used to handle missing values within the unit. For feature dimensions that have no valid observations for more than 30 consecutive days, they are marked as long-term missing and the time mask is retained. The final result is a two-dimensional matrix with row indices representing time units, column indices representing standardized feature dimensions, and elements representing real values.
4. The method for periodic assessment of cardiovascular disease risk in patients according to claim 3, characterized in that, The time-aware deep state space model includes an input embedding layer, a continuous-time state evolution module, an observation decoder, and a state posterior inference network. The input embedding layer maps the feature vector of each time unit to a high-dimensional latent space representation. The continuous-time state evolution module uses a neural ordinary differential equation to perform differential evolution of the hidden state in the continuous-time domain. Its state transition function is composed of a multilayer sensing mechanism, and the input is the current hidden state and the absolute timestamp. The observation decoder maps the hidden states back to the original feature space to reconstruct the input, which serves as a supervision signal for model training. The state posterior inference network adopts a bidirectional recurrent neural network structure, which encodes the observation sequence from both forward and backward directions to generate state posterior distribution parameters at each time point; During the reasoning phase, online risk assessment is performed using only the forward state evolution path.
5. The method for periodic assessment of cardiovascular disease risk in patients according to claim 4, characterized in that, Each state vector dimension in the state trajectory vector sequence characterizing the intrinsic dynamic characteristics of the cardiovascular system corresponds to the potential load level of a physiological subsystem, including the vascular elastic reserve dimension, metabolic homeostasis dimension, inflammation activation dimension, neural regulation dimension, and coagulation tendency dimension; the numerical range of each dimension is constrained to the interval between 0 and 1, and the higher the value, the greater the degree to which the subsystem deviates from healthy homeostasis.
6. The method for periodic assessment of cardiovascular disease risk in patients according to claim 5, characterized in that, The calculated risk evolution curvature index includes: Perform a 3rd-order spline smoothing fit on each dimension of the state trajectory vector sequence to obtain a smooth state trajectory function; The first derivative of the state trajectory function is obtained to obtain the instantaneous rate of change, and the second derivative is obtained to obtain the instantaneous acceleration. The instantaneous accelerations of each dimension are weighted and fused, and the weighting coefficients are pre-set according to the contribution of each physiological subsystem to the overall cardiovascular event. The weighted fusion result is the risk evolution curvature index.
7. The method for periodic assessment of cardiovascular disease risk in patients according to claim 6, characterized in that, The preset inflection point determination thresholds include a risk acceleration threshold and a risk reversal threshold; When the risk evolution curvature index is greater than the risk acceleration threshold for three consecutive time units, it is determined that the risk acceleration period has begun. When the risk evolution curvature index is less than the risk reversal threshold for three consecutive time units and the current risk level is higher than the benchmark level, it is determined that the risk reversal period has begun. All other situations are considered to be in a risk plateau period.
8. The method for periodic assessment of cardiovascular disease risk in patients according to claim 7, characterized in that, The generation of corresponding risk level labels and clinical intervention recommendations includes: During periods of accelerated risk, a high-risk label is issued, and recommendations are made to intensify drug treatment, increase follow-up frequency, and initiate lifestyle intervention programs. During the risk plateau period, a medium- or low-risk label is issued, and it is recommended to maintain the existing treatment plan and routine annual assessments; During the risk reversal period, output improvement trend labels and recommend assessing the possibility of medication reduction and consolidating healthy behaviors.
9. The method for periodic assessment of cardiovascular disease risk in patients according to claim 8, characterized in that, The output visualized risk cycle map includes: Plot a multi-curve time series diagram with time as the horizontal axis and the comprehensive risk score and the load level of each physiological subsystem as the vertical axis. Mark the boundaries of the identified risk cycle stages on the diagram; Different colored background areas are used to distinguish the acceleration phase, plateau phase, and reversal phase; The time stamps of key intervention events are overlaid, including medication adjustments, major life events, and clinical diagnoses.
10. The method for periodic assessment of cardiovascular disease risk in patients according to claim 4, characterized in that, The loss function of the time-aware deep state space model is composed of a weighted sum of a reconstruction error term, a state smoothing regularization term, and an event prediction cross-entropy term. The reconstruction error term uses the mean square error to measure the difference between the decoder output and the original input. The state smoothing regularization term applies an L2 norm penalty to the first derivative of the state trajectory; The event prediction cross-entropy term predicts the probability of major adverse cardiovascular events occurring within the next 5 years by attaching an event prediction header to the end of the state trajectory, and calculates the cross-entropy with the actual event labels.