A coronary heart disease whole-cycle risk dynamic assessment method and system for urban and rural hierarchical diagnosis and treatment scenes

Abstract

The application discloses a kind of coronary heart disease whole cycle risk dynamic evaluation method and system for urban and rural hierarchical diagnosis and treatment scene, including through medical internet of things terminal and wearable device, multi-dimensional health data are automatically collected and standardization verification and quality control are completed;Standardized data are input into pre-trained GBDT risk assessment model to obtain static risk score;After the time series data of continuous physiological monitoring index collected by wearable device is treated with EWMA smoothing, CUSUM double-sided cumulative sum statistics is calculated, significant time series drift of each index is detected and weighted summation is obtained to obtain time series drift comprehensive score;Static risk score and time series drift comprehensive score are weighted fusion according to preset fusion coefficient to obtain dynamic comprehensive risk score;Risk grade is determined according to hierarchical threshold based on dynamic comprehensive risk score.The application realizes the whole cycle dynamic accurate quantification of coronary heart disease risk by the weighted fusion of GBDT static evaluation and CUSUM time series drift detection.

Description

Technical Field

[0001] This invention belongs to the field of medical and health information technology, and in particular relates to a method and system for dynamic risk assessment of coronary heart disease throughout its entire life cycle in urban and rural hierarchical medical treatment scenarios. Background Technology

[0002] Coronary heart disease (CHD) poses a significant threat to the health of urban and rural residents, characterized by high incidence, high disability rate, and high mortality rate. With the deepening of the tiered healthcare system, the uneven distribution of medical resources between urban and rural areas has become increasingly prominent. Primary healthcare institutions, due to limited equipment and personnel expertise, face significant challenges in CHD prevention and control, including inconsistent screening standards and low accuracy in identifying high-risk groups. Therefore, developing a dynamic CHD risk assessment scheme adapted to the urban and rural tiered healthcare system, and achieving a closed-loop technology process from standardized data collection to precise dynamic risk quantification, has become a core technical problem urgently needing to be solved in this field.

[0003] Currently, technical solutions for coronary heart disease risk assessment mainly focus on the following directions: First, risk assessment tools based on traditional statistical models, such as the Framingham Risk Score and the China-PAR model, utilize demographic information and limited structured risk factors for long-term risk prediction. These models rely on cross-sectional data at a single time point, cannot integrate continuously monitored time-series physiological indicators, and have fixed model parameters, making it difficult to capture dynamic changes in the patient's risk status. Second, risk prediction models based on machine learning integrate multi-dimensional clinical data from electronic health records for training and prediction. However, these models typically only output static risk scores and do not consider the time-series drift information contained in the continuous physiological monitoring data collected by wearable devices. When the patient's physiological indicators progressively deteriorate, the static model cannot respond in a timely manner. Third, health monitoring systems based on wearable devices can achieve continuous collection of physiological parameters such as heart rate and blood pressure and abnormal threshold alarms. However, these systems typically make simple judgments based on fixed thresholds, lack deep integration with machine learning risk assessment models, and are difficult to extract subtle drift trends with warning significance from time-series data.

[0004] However, the existing technical solutions mentioned above have the following shortcomings: First, the static risk assessment model only uses cross-sectional data collected in a single instance and cannot integrate time-series data from previous screenings and continuous monitoring by wearable devices to achieve dynamic updates of risk levels. When a patient's physiological indicators gradually shift over time, the static score cannot reflect the actual risk changes in a timely manner. Second, there is a lack of an effective fusion mechanism between the time-series monitoring data from wearable devices and the clinical assessment model. Subtle drift trends in continuous physiological signals are not quantified and incorporated into the comprehensive risk judgment, resulting in insufficient early warning capability for progressive disease deterioration in risk assessment. Third, primary healthcare institutions lack a standardized multi-source data collection system and unified data quality control methods. Inconsistent data formats and standards from different sources make it difficult to guarantee the applicability and accuracy of the risk assessment model in different urban and rural scenarios. Therefore, it is necessary to design a dynamic risk assessment system for the entire life cycle of coronary heart disease for urban and rural hierarchical medical treatment scenarios to solve the above problems. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a dynamic risk assessment system for the whole cycle of coronary heart disease in urban and rural hierarchical medical treatment scenarios. It aims to overcome the shortcomings of existing coronary heart disease risk assessment systems, such as the inability of static models to integrate continuous monitoring time-series information, the lack of an effective fusion mechanism between time-series physiological data and clinical assessment models, and the lack of uniformity in screening standards between urban and rural areas. It has the feature of achieving dynamic and accurate quantification of coronary heart disease risk throughout the whole cycle through weighted fusion of GBDT static risk score and CUSUM time-series drift comprehensive score.

[0006] To achieve the above technical solution, the technical solution adopted by the present invention is as follows: A dynamic risk assessment method for coronary heart disease throughout its entire life cycle, oriented towards a tiered medical service model in urban and rural areas, includes the following steps: S1 automatically collects multi-dimensional health data of the target population through medical IoT terminal devices and wearable health monitoring devices. The collected data is then subjected to integrity verification, logical rationality verification and outlier identification in sequence. After standardization processing, the data is stored in the resident's full-cycle health database. S2, the standardized multi-dimensional health data is input into the pre-trained Gradient Boosting Tree (GBDT) risk assessment model to calculate the static risk score of coronary heart disease for the target population. ; S3. For the time-series data of continuous physiological monitoring indicators collected by wearable devices, after smoothing with exponentially weighted moving average (EWMA), the cumulative sum (CUSUM) two-sided cumulative sum statistic is calculated. When any statistic exceeds the judgment threshold, the corresponding indicator is judged to have significant time-series drift. A comprehensive time-series drift score is obtained by weighted summation based on the degree of drift of each indicator. ; S4, static risk scoring Overall score with time drift According to the preset fusion coefficient A weighted fusion method is used to calculate a dynamic comprehensive risk score. When no time-series drift signal is detected for any metric ; S5, based on dynamic comprehensive risk scoring The risk level of coronary heart disease in the target population is determined according to the preset grading thresholds, which are based on the 30th and 60th percentiles of the score distribution of the training set samples.

[0007] Preferably, the GBDT risk assessment model is trained and constructed according to the following method: S201, Training Dataset Construction: Collect clinical data of coronary heart disease patients and physical examination data of healthy individuals covering four levels of medical institutions: village, township, county and city, to form a training sample set; S202, Data Preprocessing: Clean, deduplicate, and impute missing values ​​of the multi-dimensional health data in the training sample set, and perform Z-score standardization on the imputed data; S203, Feature Selection and Core Feature Determination: The correlation between each feature and the risk of coronary heart disease is analyzed using the Pearson correlation coefficient on the standardized data; then, the retained features are ranked by importance using the gradient boosting tree (GBDT) model as the basis, and the core features that are strongly correlated with the risk of coronary heart disease are selected and the weight coefficient of each feature in the model is determined. S204, Model Training: Using core features as input and coronary heart disease risk score as output, the Gradient Boosting Tree (GBDT) algorithm is used for supervised training on the training sample set. During the training process, each new decision tree fits the prediction residual of the previous ensemble model, and the loss function is reduced iteratively round by round. Finally, the prediction results of all decision trees are weighted and summed to obtain the static risk score of coronary heart disease. S205, Hyperparameter Optimization and Performance Validation: Five-fold cross-validation is used to optimize the four hyperparameters of the GBDT algorithm—number of base learners, learning rate, maximum decision tree depth, and subsampling ratio—using grid search. The optimal combination of hyperparameters is determined with the goal of maximizing the average AUC value of cross-validation. After training, the model performance is validated on the test set using the confusion matrix, ROC curve, AUC value, precision, and recall. Once the AUC value reaches the preset acceptance standard, the model parameters are solidified.

[0008] Furthermore, the dynamic comprehensive risk score The score ranges from 0 to 100, with 0 to 30 indicating low risk, 31 to 60 indicating medium risk, and 61 to 100 indicating high risk. After each automatic synchronization of new screening or treatment data for the target population, the system automatically re-enters the new standardized data into the model for calculation. The risk level will be updated simultaneously.

[0009] Preferably, the EWMA smoothing and CUSUM two-sided cumulative sum statistic described in S3 are performed according to the following formula: ; ; ; in, For the first The monitoring indicators at any time The original collected values, EWMA smoothing value, The EWMA smoothing coefficient is set to 0.2 to 0.3. and These are the cumulative sums of CUSUM on the upper and lower sides, respectively. The baseline mean of health status. To allow for relaxation, half of the standard deviation of the health status of this indicator is taken; when or Time to determine the first The indicators showed significant time-series drift, among which This represents the standard deviation of the EWMA smoothed value when the indicator is in a healthy state. Drift rating: ; The overall time drift score is: ; in The number of continuous monitoring indicators included in drift monitoring, For the first The drift weight coefficient of an indicator is taken as the normalized importance score of that indicator in the feature importance ranking of the GBDT model.

[0010] Furthermore, the fusion coefficient described in S4 The candidate value that maximizes the classification AUC value is determined by searching the grid with a step size of 0.05 in the interval [0,1] on the validation set.

[0011] Preferably, it further includes a risk threshold adaptive update step: Using the initial training set's rating distribution as the baseline distribution and the rating distribution within the sliding time window as the current distribution, the rating interval is divided equally and the population stability index (PSI) is calculated. When the PSI exceeds the preset trigger threshold, the grading threshold is automatically re-estimated using the 30th and 60th percentiles of the current window's rating distribution, and a hysteresis band is set at the new threshold to prevent grade jumps caused by rating fluctuations.

[0012] Preferably, it also includes confidence quantification and automatic retesting steps: During the GBDT training phase, multiple Bagging sub-models are trained simultaneously, and the standard deviation of the scores of each sub-model is calculated during inference. As a confidence measure, when If the confidence level exceeds a preset threshold, the current prediction confidence is deemed insufficient. A retest command is automatically sent to the corresponding data acquisition terminal to re-collect the original indicator data and re-evaluate it. After the retest... If the value still exceeds the limit, the record is marked as an anomaly; the preset confidence threshold is taken as the percentage of correctly classified samples in the validation set. The 95th percentile of the distribution.

[0013] Furthermore, it also includes a four-level hierarchical computing node architecture for collecting training data for the GBDT model; the responsibilities of the four-level hierarchical extreme node architecture are divided as follows: Village-level edge data collection nodes are deployed at village-level medical terminals and are responsible for collecting raw health data, caching and preprocessing it locally, and uploading it to township-level edge computing nodes. Township-level edge computing nodes are deployed on the servers of township-level medical institutions. They are responsible for data quality control and the calculation of routine risk scores for low- and medium-risk groups, and uploading the results to the county-level regional computing nodes. County-level regional computing nodes are deployed in the data centers of county-level medical institutions, and are responsible for deep inference computing and referral instruction generation for high-risk groups; The city-level cloud center computing nodes are deployed on the city-level cloud server and are responsible for the centralized training, hyperparameter optimization and version management of the GBDT model. After training is completed, the model is distributed to the lower-level nodes.

[0014] Preferably, a dynamic risk assessment system for the entire life cycle of coronary heart disease in a tiered urban and rural healthcare system is provided to implement the aforementioned dynamic risk assessment method for the entire life cycle of coronary heart disease in a tiered urban and rural healthcare system, comprising: The standardized screening module is used to automatically collect, standardize, verify, and control the multi-dimensional health data of the target population through medical IoT terminal devices and wearable health monitoring devices; The risk dynamic assessment model module is used to build the GBDT risk assessment model, integrates CUSUM time-series drift detection and EWMA smoothing, and weights and merges the GBDT static score and the time-series drift comprehensive score into a dynamic comprehensive risk score and completes the risk level determination. Preferably, the risk dynamic assessment model module includes a model building and training submodule, a CUSUM time-series drift detection and EWMA smoothing submodule, a risk level determination submodule, and a risk dynamic update submodule, wherein the output of the model building and training submodule is connected to the input of the risk level determination submodule, and the output of the risk dynamic update submodule is fed back to the input of the risk level determination submodule.

[0015] Furthermore, a computer device includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the aforementioned dynamic risk assessment method for the entire life cycle of coronary heart disease in urban and rural hierarchical medical treatment scenarios.

[0016] Furthermore, a computer-readable storage medium is provided, on which computer instructions are stored, for causing a computer to execute the aforementioned method for dynamic risk assessment of coronary heart disease throughout its entire life cycle in a tiered medical service scenario in urban and rural areas.

[0017] The beneficial effects of this invention are as follows: 1. This invention constructs a training sample set covering four levels of medical institutions: village, township, county, and city. It uses the GBDT algorithm to screen core risk features and determine weight coefficients through Pearson correlation coefficient analysis and feature importance ranking. Combined with five-fold cross-validation grid search to optimize hyperparameter combinations, a high-precision static risk assessment model for coronary heart disease is constructed. At the same time, through medical IoT terminal devices and wearable health monitoring devices, the automated collection of multi-dimensional health data and standardized quality control throughout the process are realized, solving the defects of existing technologies such as inconsistent screening standards in urban and rural areas and poor model adaptability to grassroots populations.

[0018] 2. This invention introduces EWMA smoothing and CUSUM two-sided cumulative sum statistics to detect drift in the time-series data of continuous physiological monitoring indicators collected by wearable devices. When the EWMA smoothed value of any monitoring indicator deviates from the health status benchmark and continues to accumulate beyond the judgment threshold, a significant time-series drift is determined to have occurred. The time-series drift comprehensive score is obtained by weighted summation according to the importance of the normalized features of each indicator. The static risk score and the time-series drift comprehensive score are then weighted and fused into a dynamic comprehensive risk score according to a preset fusion coefficient. This solves the shortcomings of existing static assessment models that cannot integrate continuous monitoring time-series information and have a lagging response to progressive physiological deterioration.

[0019] 3. This invention introduces the population stability index (PSI) to achieve adaptive updating of the classification threshold. When the score distribution shifts significantly, the classification threshold is automatically re-estimated and a hysteresis band is set to prevent level jumps. At the same time, by constructing multiple Bagging sub-models during the GBDT training phase and using the standard deviation of the sub-model scores as a confidence measure, a retest instruction is automatically triggered when the confidence is insufficient. This solves the defects of risk level classification standards becoming invalid as the population distribution changes and misjudgments caused by abnormal data quality in a single collection. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the workflow of the dynamic risk assessment model for coronary heart disease in this embodiment of the invention. Figure 2 This is a system structure block diagram of the present invention. Detailed Implementation

[0021] Example 1: like Figure 1 As shown, a method for dynamic risk assessment of coronary heart disease throughout its entire life cycle, oriented towards a tiered medical service model in urban and rural areas, includes the following steps: S1 automatically collects multi-dimensional health data of the target population through medical IoT terminal devices and wearable health monitoring devices. The collected data is then subjected to integrity verification, logical rationality verification and outlier identification in sequence. After standardization processing, the data is stored in the resident's full-cycle health database. S2, the standardized multi-dimensional health data is input into the pre-trained Gradient Boosting Tree (GBDT) risk assessment model to calculate the static risk score of coronary heart disease for the target population. ; S3. For the time-series data of continuous physiological monitoring indicators collected by wearable devices, after smoothing with exponentially weighted moving average (EWMA), the cumulative sum (CUSUM) two-sided cumulative sum statistic is calculated. When any statistic exceeds the judgment threshold, the corresponding indicator is judged to have significant time-series drift. A comprehensive time-series drift score is obtained by weighted summation based on the degree of drift of each indicator. ; S4, static risk scoring Overall score with time drift According to the preset fusion coefficient A weighted fusion method is used to calculate a dynamic comprehensive risk score. When no time-series drift signal is detected for any metric ; S5, based on dynamic comprehensive risk scoring The risk level of coronary heart disease in the target population is determined according to the preset grading thresholds, which are based on the 30th and 60th percentiles of the score distribution of the training set samples.

[0022] Preferably, the GBDT risk assessment model is trained and constructed according to the following method: S201, Training Dataset Construction: Collect clinical data of coronary heart disease patients and physical examination data of healthy individuals covering four levels of medical institutions: village, township, county and city, to form a training sample set; S202, Data Preprocessing: Clean, deduplicate, and impute missing values ​​of the multi-dimensional health data in the training sample set, and perform Z-score standardization on the imputed data; S203, Feature Selection and Core Feature Determination: The correlation between each feature and the risk of coronary heart disease is analyzed using the Pearson correlation coefficient on the standardized data; then, the retained features are ranked by importance using the gradient boosting tree (GBDT) model as the basis, and the core features that are strongly correlated with the risk of coronary heart disease are selected and the weight coefficient of each feature in the model is determined. S204, Model Training: Using core features as input and coronary heart disease risk score as output, the Gradient Boosting Tree (GBDT) algorithm is used for supervised training on the training sample set. During the training process, each new decision tree fits the prediction residual of the previous ensemble model, and the loss function is reduced iteratively round by round. Finally, the prediction results of all decision trees are weighted and summed to obtain the static risk score of coronary heart disease. S205, Hyperparameter Optimization and Performance Validation: Five-fold cross-validation is used to optimize the four hyperparameters of the GBDT algorithm—number of base learners, learning rate, maximum decision tree depth, and subsampling ratio—using grid search. The optimal combination of hyperparameters is determined with the goal of maximizing the average AUC value of cross-validation. After training, the model performance is validated on the test set using the confusion matrix, ROC curve, AUC value, precision, and recall. Once the AUC value reaches the preset acceptance standard, the model parameters are solidified.

[0023] Furthermore, the dynamic comprehensive risk score The score ranges from 0 to 100, with 0 to 30 indicating low risk, 31 to 60 indicating medium risk, and 61 to 100 indicating high risk. After each automatic synchronization of new screening or treatment data for the target population, the system automatically re-enters the new standardized data into the model for calculation. The risk level will be updated simultaneously.

[0024] Preferably, the EWMA smoothing and CUSUM two-sided cumulative sum statistic described in S3 are performed according to the following formula: ; ; ; in, For the first The monitoring indicators at any time The original collected values, EWMA smoothing value, The EWMA smoothing coefficient is set to 0.2 to 0.3. and These are the cumulative sums of CUSUM on the upper and lower sides, respectively. The baseline mean of health status. To allow for relaxation, half of the standard deviation of the health status of this indicator is taken; when or Time to determine the first The indicators showed significant time-series drift, among which This represents the standard deviation of the EWMA smoothed value when the indicator is in a healthy state. Drift rating: ; The overall time drift score is: ; in The number of continuous monitoring indicators included in drift monitoring, For the first The drift weight coefficient of an indicator is taken as the normalized importance score of that indicator in the feature importance ranking of the GBDT model.

[0025] Furthermore, the fusion coefficient described in S4 The candidate value that maximizes the classification AUC value is determined by searching the grid with a step size of 0.05 in the interval [0,1] on the validation set.

[0026] Preferably, it further includes a risk threshold adaptive update step: Using the initial training set's rating distribution as the baseline distribution and the rating distribution within the sliding time window as the current distribution, the rating interval is divided equally and the population stability index (PSI) is calculated. When the PSI exceeds the preset trigger threshold, the grading threshold is automatically re-estimated using the 30th and 60th percentiles of the current window's rating distribution, and a hysteresis band is set at the new threshold to prevent grade jumps caused by rating fluctuations.

[0027] Preferably, it also includes confidence quantification and automatic retesting steps: During the GBDT training phase, multiple Bagging sub-models are trained simultaneously, and the standard deviation of the scores of each sub-model is calculated during inference. As a confidence measure, when If the confidence level exceeds a preset threshold, the current prediction confidence is deemed insufficient. A retest command is automatically sent to the corresponding data acquisition terminal to re-collect the original indicator data and re-evaluate it. After the retest... If the value still exceeds the limit, the record is marked as an anomaly; the preset confidence threshold is taken as the percentage of correctly classified samples in the validation set. The 95th percentile of the distribution.

[0028] like Figure 2 As shown, a dynamic risk assessment system for the entire life cycle of coronary heart disease in a tiered medical service model for urban and rural areas is used to implement the aforementioned dynamic risk assessment method for the entire life cycle of coronary heart disease in a tiered medical service model for urban and rural areas, including: The standardized screening module is used to automatically collect, standardize, verify, and control the multi-dimensional health data of the target population through medical IoT terminal devices and wearable health monitoring devices; The risk dynamic assessment model module is used to build the GBDT risk assessment model. It integrates CUSUM time-series drift detection and EWMA smoothing processing, and weights and merges the GBDT static score and the time-series drift comprehensive score into a dynamic comprehensive risk score to complete the risk level determination.

[0029] Preferably, the risk dynamic assessment model module includes a model building and training submodule, a CUSUM time-series drift detection and EWMA smoothing submodule, a risk level determination submodule, and a risk dynamic update submodule, wherein the output of the model building and training submodule is connected to the input of the risk level determination submodule, and the output of the risk dynamic update submodule is fed back to the input of the risk level determination submodule.

[0030] Example 2: This embodiment also discloses a method for further applying the assessment method to subsequent hierarchical management and early warning, the specific method being as follows: Based on the determined risk level, a pre-set three-level differentiated management path is automatically matched. Through a four-level hierarchical computing node architecture deployed at village-level edge collection nodes, township-level edge computing nodes, county-level regional computing nodes, and city-level cloud center computing nodes, management tasks are automatically dispatched and collaboratively executed. Personalized health intervention content is automatically pushed to the target population based on the risk factors associated with the risk level.

[0031] Furthermore, the division of responsibilities in the four-level hierarchical computing node architecture is as follows: Village-level edge data collection nodes are deployed at village-level medical terminals and are responsible for collecting raw health data, caching and preprocessing it locally, and uploading it to township-level edge computing nodes. Township-level edge computing nodes are deployed on the servers of township-level medical institutions. They are responsible for data quality control and the calculation of routine risk scores for low- and medium-risk groups, and uploading the results to the county-level regional computing nodes. County-level regional computing nodes are deployed in the data centers of county-level medical institutions, and are responsible for deep inference computing and referral instruction generation for high-risk groups; The city-level cloud center computing nodes are deployed on the city-level cloud server and are responsible for the centralized training, hyperparameter optimization and version management of the GBDT model. After training is completed, the model is distributed to the lower-level nodes.

[0032] Furthermore, the three-tiered differentiated management path is as follows: The management path for low-risk levels is to automatically generate annual routine screenings and automatically push personalized health intervention content; the management path for medium-risk levels is to automatically generate risk factor intervention plans and increase the frequency of monitoring data collection; and the management path for high-risk levels is to enhance continuous monitoring with wearable devices and automatically refer patients.

[0033] Furthermore, it also includes a hierarchical management module, which is used to match differentiated standardized management paths based on risk levels, and automatically dispatch management tasks and push intervention content through a four-level hierarchical computing node architecture; The monitoring and early warning module is used to automatically trigger tiered early warnings based on changes in the risk of the target population and track the handling status until the closed loop is completed. The output of the standardized screening module is connected to the input of the risk dynamic assessment model module, the output of the risk dynamic assessment model module is connected to the input of the hierarchical management module, the output of the hierarchical management module is connected to the input of the monitoring and early warning module, and the output of the monitoring and early warning module is fed back to the input of the risk dynamic assessment model module.

[0034] Preferably, the monitoring and early warning module includes a tiered early warning preset submodule and an early warning triggering and closed-loop handling submodule; the output of the tiered early warning preset submodule is connected to the input of the early warning triggering and closed-loop handling submodule, and the output of the early warning triggering and closed-loop handling submodule is connected to the input of the risk dynamic assessment model module, so as to send back the early warning handling status to trigger the recalculation of the risk level.

[0035] Example 3: This embodiment provides a specific example of a training method for a GBDT risk assessment model, including the following steps: S201, Training dataset construction: Collect clinical data of coronary heart disease patients and physical examination data of healthy people covering four levels of medical institutions: village, township, county and city, to form a training sample set. Each sample contains multidimensional health indicators and their corresponding coronary heart disease incidence labels. The label value of 1 indicates a confirmed coronary heart disease patient and a value of 0 indicates a healthy person. S202, Data Preprocessing and Z-score Standardization: The multi-dimensional health data in the training sample set is cleaned, deduplicated, and imputed for missing values. Missing value imputation uses the median of the feature across all samples in the training set. The imputed data is then Z-score standardized item by item according to each feature dimension. ; in, Indicates the first The feature in the first The original collected values ​​in each sample Indicates the first The mean of each feature across all samples in the training set. Indicates the first The standard deviation of each feature across all samples in the training set This represents the standardized feature value; after standardization, each feature has a mean of 0 and a standard deviation of 1, so that indicators such as blood pressure, blood sugar, blood lipids, and heart rate of different dimensions can participate in model training on the same scale; S203, Feature Selection and Core Feature Determination: For the standardized data, first calculate the linear correlation between each feature and the coronary heart disease incidence label using the Pearson correlation coefficient: ; in, Indicates the first The Pearson correlation coefficient between these characteristics and the incidence of coronary heart disease. This represents the total number of training samples. For the first The feature in the first Standardized values ​​in each sample For the first The incidence label of coronary heart disease in a single sample, and The first The mean of each feature value and the disease label in all samples; Features with extremely low correlation to the incidence of coronary heart disease were initially screened out using a threshold, while features with a linear association with the risk of developing the disease were retained. The retained features are then trained using Gradient Boosting Tree (GBDT) as the base model. After training, the weighted number of times each feature is selected as a splitting node across all decision trees is extracted from the model as the feature importance score. After normalization, the weight coefficients of each feature in the model are obtained: ; in, This represents the total number of features retained. Indicates the first Feature importance score of each feature Indicates the first The weight coefficients after normalization of each feature; according to Sort by value from largest to smallest, and select the 20 items with the highest weight coefficients as core features. These 20 core features include age, gender, smoking history, alcohol consumption history, family history of coronary heart disease, history of hypertension, history of diabetes, history of hyperlipidemia, systolic blood pressure, diastolic blood pressure, fasting blood glucose, 2-hour postprandial blood glucose, total cholesterol, triglycerides, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, heart rate, PR interval, QRS duration, and ST segment deviation. The weight coefficients of these 20 core features are as follows: Simultaneously, it serves as the comprehensive score for calculating time-series drift in claim 4. Drift weighting coefficients of each indicator The basis for its value; S204, Model Training: Standardized feature vectors composed of the aforementioned 20 core features. Using the coronary heart disease risk score as input and the Gradient Boosting Tree Derivative (GBDT) algorithm as output, supervised training is performed on the training sample set. The GBDT algorithm uses a regression decision tree as the base learner and the mean squared error as the loss function, and constructs the ensemble model in the following recursive manner: Initialize the model to predict constant values: ; in, This represents the predicted output of the initial model. For the first Coronary heart disease incidence labels for each training sample; From round 1 to round 2 In each round, the prediction residual of the current model for each training sample is calculated: ; in, Indicates the first After rounds of iteration, the ensemble model is applied to the first... The predicted value for each sample, Indicates the first The new decision tree needs to fit the residuals. , The total number of base learners is between 100 and 500; No. Decision Tree by To fit the residuals to the training data, update the ensemble model: ; in, The learning rate is set to 0.01 to 0.1 to control the contribution weight of each new decision tree to the final model; For the first The output of each decision tree is set to a maximum depth of 3 to 8 and a subsampling ratio of 0.6 to 0.9. Iteration After rounds of testing, the final integrated model, namely the GBDT risk assessment model, is obtained, which includes the feature vectors of the target population. Output static risk score for coronary heart disease : ; in, The value ranges from 0 to 100, with higher values ​​indicating a greater risk of developing coronary heart disease. S205, Hyperparameter Optimization and Performance Verification: Five-fold cross-validation is used to optimize the number of base learners. Learning rate The four hyperparameters—maximum depth of the decision tree, subsampling ratio, etc.—are optimized using a grid search. The training sample set is randomly divided into five equal parts, with each part serving as the validation set in turn, and the remaining four parts serving as the training set. The optimal combination of hyperparameters is determined by maximizing the mean AUC value on the five-part validation set. ; in, Indicates the first The AUC value on the validation set. This represents the average AUC value from the five-fold cross-validation. After training, model performance is validated on independent test sets using the confusion matrix, ROC curve, AUC value, precision, and recall. Once the preset acceptance criteria are met, the model parameters are solidified.

[0036] Example 4: This embodiment provides a specific example of applying the method to a certain location. The system hardware layer includes a primary mobile terminal, a PC terminal, physical examination equipment, an electrocardiograph, and a server; the software layer includes a standardized screening module, a dynamic risk assessment model module, a hierarchical management module, and a follow-up early warning module. Each module is interconnected through a data interface and deployed on a cloud server. Medical institutions at all levels in urban and rural areas access the system through terminals to achieve data sharing and collaborative operation.

[0037] The standardized screening module incorporates a unified screening indicator system, covering five dimensions and 28 core indicators, specifically including: Basic information: age, gender, height, weight; Underlying medical conditions: hypertension, diabetes, hyperlipidemia, etc.; Physiological indicators: blood pressure, fasting blood glucose, 2-hour postprandial blood glucose, total cholesterol, triglycerides, and low-density lipoprotein cholesterol; Characteristics of chest pain: location, nature, duration, and triggering factors; Core ECG parameters: heart rate, PR interval, QRS duration, and ST segment deviation.

[0038] The screening guidelines are pre-defined as follows: the screening population consists of urban and rural residents aged 35 and above; the general population is screened once a year, while high-risk groups such as those with hypertension and diabetes are screened every six months. Primary healthcare workers can quickly input data via mobile devices or connect to medical examination equipment to automatically synchronize blood pressure, blood sugar, and other data. The system automatically verifies the completeness and rationality of the data, and provides real-time alerts for corrections of abnormal data to ensure the quality of the screening data.

[0039] The risk dynamic assessment model employs the Gradient Boosting Tree (GBDT) algorithm. The training dataset comprises clinical data from 50,000 coronary heart disease patients and physical examination data from 10 tertiary hospitals, 20 county-level hospitals, and 50 township health centers in a certain province. The data dimensions include: Traditional risk factors include: age, sex, smoking history, family history, etc. Real-time clinical indicators: blood pressure, blood glucose, blood lipids, and electrocardiogram parameters; Dynamic follow-up data: changes in indicators during each follow-up period and the effectiveness of intervention measures.

[0040] During the model training process, the multi-source data is first cleaned, standardized, and features are extracted to select 20 core risk features. Then, the GBDT algorithm is used to build a risk assessment model, and the model parameters are optimized through cross-validation to determine the weight of each feature. Finally, the model outputs a risk score of 0-100, with 0-30 being low risk, 31-60 being medium risk, and 61-100 being high risk.

[0041] After each screening, diagnosis, and follow-up data entry is completed, the system automatically inputs the new data into the model and recalculates the risk score and level. For example, if a resident's initial screening risk score is 45 points (medium risk), and a significant increase in low-density lipoprotein cholesterol is found during the 3-month follow-up, the model automatically updates the risk score to 68 points (high risk) and pushes it to the high-risk management path.

[0042] The tiered management module develops standardized management pathways for three types of at-risk individuals, clearly defining intervention measures, follow-up frequency, and division of responsibilities.

[0043] For low-risk individuals with scores of 0-30: Intervention measures include health education, promoting knowledge about low-salt and low-fat diets, regular exercise, smoking cessation and alcohol limitation, and annual routine screening; follow-up frequency is once a year; the responsible entity is the village-level medical institution, with the township-level medical institution responsible for technical guidance.

[0044] For individuals with scores of 31-60 at medium risk: the intervention measures include quarterly follow-up to monitor blood pressure, blood sugar, and blood lipids; risk factor intervention, including developing personalized diet and exercise plans and medication guidance; follow-up frequency is once every 3 months; the responsible entity is the township-level medical institution, while the county-level medical institution is responsible for technical support and referral coordination.

[0045] For high-risk individuals scoring 61-100: Intervention measures include monthly follow-ups, comprehensive monitoring of the condition, medication adherence management, specialist outreach, automatic referral to county-level or higher specialist medical institutions, enhanced intervention, regular electrocardiogram re-examinations, and adjustments to treatment plans; follow-up frequency is once a month; the responsible entity is the county-level medical institution, while the municipal-level medical institution is responsible for specialist diagnosis and treatment and guidance on difficult cases.

[0046] The follow-up early warning module is based on a hierarchical management path and automatically generates personalized follow-up plans. For example, the follow-up plan for residents at medium risk is "to complete the quarterly follow-up on May 10, 2026". The system will remind medical staff and residents 7 days in advance via SMS and system messages.

[0047] The tiered early warning rules are as follows: Level 1 Warning for Low-Risk Abnormalities: If one of the follow-up indicators of a low-risk resident shows a slight abnormality, the warning will be sent to the village-level medical staff, requiring a re-examination within one month; Level II Warning: If two or more follow-up indicators of medium-risk residents are abnormal or the risk level increases, the warning will be sent to township-level medical staff, requiring intervention and re-examination within two weeks. Level 3 warning for high-risk abnormalities: If a high-risk resident's follow-up indicators are significantly abnormal, symptoms such as chest pain appear, or the risk level continues to rise, an alert will be sent to county-level medical staff and grid administrators, requiring them to immediately connect with a specialist doctor and complete the intervention assessment within 24 hours.

[0048] Taking a 58-year-old rural male resident as an example, the full-cycle management process is as follows: First, standardized screening is conducted. Village-level medical staff complete the annual screening via mobile devices, entering data such as blood pressure 150 / 95 mmHg, fasting blood glucose 7.2 mmol / L, and smoking history of 20 years. The system automatically synchronizes the data to the risk assessment model. Secondly, a dynamic risk assessment is conducted. The model calculates a risk score of 52 points, indicating medium risk, and automatically matches a medium-risk management path. During the subsequent tiered management intervention, township-level medical staff will receive management tasks, conduct follow-up visits every 3 months, guide patients to quit smoking and follow a low-salt diet, and prescribe antihypertensive and hypoglycemic drugs. A follow-up warning was issued after 6 months of follow-up. The resident's blood pressure was poorly controlled, and a repeat blood pressure reading of 160 / 100 mmHg triggered a level 3 warning. The system automatically updated the risk score to 65 points. Then, specialists connect with county-level medical staff who receive the early warning and immediately connect with municipal-level specialist medical institutions to arrange two-way referrals for residents, complete examinations such as electrocardiograms and coronary CT scans, and develop intensive treatment plans. Finally, continuous follow-up management is required. Residents are transferred to the high-risk management path, and county-level medical staff follow up once a month to monitor changes in their condition until the risk level is reduced, thus achieving closed-loop management throughout the entire cycle.

Claims

1. A method for dynamic risk assessment of coronary heart disease throughout its entire life cycle, oriented towards a tiered medical service model in urban and rural areas, characterized in that... Includes the following steps: S1 automatically collects multi-dimensional health data of the target population through medical IoT terminal devices and wearable health monitoring devices. The collected data is then subjected to integrity verification, logical rationality verification and outlier identification in sequence. After standardization processing, the data is stored in the resident's full-cycle health database. S2, the standardized multi-dimensional health data is input into the pre-trained Gradient Boosting Tree (GBDT) risk assessment model to calculate the static risk score of coronary heart disease for the target population. ; S3. For the time-series data of continuous physiological monitoring indicators collected by wearable devices, after smoothing with exponentially weighted moving average (EWMA), the cumulative sum (CUSUM) two-sided cumulative sum statistic is calculated. When any statistic exceeds the judgment threshold, the corresponding indicator is judged to have significant time-series drift. A comprehensive time-series drift score is obtained by weighted summation based on the degree of drift of each indicator. ; S4, static risk scoring Overall score with time drift According to the preset fusion coefficient A weighted fusion method is used to calculate a dynamic comprehensive risk score. When no time-series drift signal is detected for any metric ; S5, based on dynamic comprehensive risk scoring The risk level of coronary heart disease in the target population is determined according to the preset grading threshold.

2. The method for dynamic risk assessment of coronary heart disease throughout its entire life cycle in a hierarchical medical system in urban and rural areas, as described in claim 1, is characterized in that... The GBDT risk assessment model was trained and constructed using the following method: S201, Training Dataset Construction: Collect clinical data of coronary heart disease patients and physical examination data of healthy individuals covering four levels of medical institutions: village, township, county and city, to form the training sample set; S202, Data Preprocessing: Clean, deduplicate, and impute missing values ​​of the multi-dimensional health data in the training sample set, and perform Z-score standardization on the imputed data; S203, Feature Selection and Core Feature Determination: The correlation between each feature and the risk of coronary heart disease is analyzed using the Pearson correlation coefficient on the standardized data; then, the retained features are ranked by importance using the gradient boosting tree (GBDT) model as the basis, and the core features that are strongly correlated with the risk of coronary heart disease are selected and the weight coefficient of each feature in the model is determined. S204, Model Training: Using core features as input and coronary heart disease risk score as output, the gradient boosting tree (GBDT) algorithm is used for supervised training on the training sample set. During the training process, each new decision tree fits the prediction residual of the previous ensemble model, and the loss function is reduced iteratively in each round. Finally, the prediction results of all decision trees are weighted and summed to obtain the static risk score of coronary heart disease. S205, Hyperparameter Optimization and Performance Verification: Five-fold cross-validation is used to perform grid search optimization on the four hyperparameters of the GBDT algorithm: number of base learners, learning rate, maximum depth of decision tree, and subsampling ratio. The optimal combination of hyperparameters is determined with the goal of maximizing the average AUC value of cross-validation. After training, the model performance is validated on the test set using confusion matrix, ROC curve, AUC value, precision and recall. Once the AUC value reaches the preset acceptance standard, the model parameters are solidified.

3. The method for dynamic risk assessment of coronary heart disease throughout its entire life cycle in a hierarchical medical system in urban and rural areas, as described in claim 1, is characterized in that... The EWMA smoothing and CUSUM two-sided cumulative sum statistic described in S3 are performed according to the following formula: ； ； ； in, For the first The monitoring indicators at any time The original collected values, EWMA smoothing value, EWMA smoothing coefficient, and These are the cumulative sums of CUSUM on the upper and lower sides, respectively. The baseline mean of health status. To allow for relaxation, half of the standard deviation of the health status of this indicator is taken.

4. The method for dynamic risk assessment of coronary heart disease throughout its entire life cycle in a hierarchical medical system in urban and rural areas, as described in claim 3, is characterized in that... when or Time to determine the first The indicators showed significant time-series drift, among which This represents the standard deviation of the EWMA smoothed value when the indicator is in a healthy state. Drift rating: ； The overall time drift score is: ； in The number of continuous monitoring indicators included in drift monitoring, For the first The drift weight coefficient of an indicator is taken as the normalized importance score of that indicator in the feature importance ranking of the GBDT model.

5. The method for dynamic risk assessment of coronary heart disease throughout its entire life cycle in a hierarchical medical system in urban and rural areas, as described in claim 1, is characterized in that... It also includes a risk threshold adaptive update step: Using the rating distribution of the initial training set as the baseline distribution and the rating distribution within the sliding time window as the current distribution, the rating interval is divided equally and the population stability index (PSI) is calculated. When the PSI exceeds the preset trigger threshold, the grading threshold is automatically re-estimated, and a hysteresis band is set at the new threshold to prevent grade jumps caused by rating fluctuations.

6. The method for dynamic risk assessment of coronary heart disease throughout its entire life cycle in a hierarchical medical system in urban and rural areas, as described in claim 1, is characterized in that... It also includes confidence quantification and automated retesting steps: During the GBDT training phase, multiple Bagging sub-models are trained simultaneously, and the standard deviation of the scores of each sub-model is calculated during inference. As a confidence measure, when If the confidence level exceeds a preset threshold, the current prediction confidence is deemed insufficient. A retest command is automatically sent to the corresponding data acquisition terminal to re-collect the original indicator data and re-evaluate it. After the retest... If the data still exceeds the limit, the record will be marked as an anomaly.

7. A dynamic risk assessment system for the entire life cycle of coronary heart disease in a hierarchical medical system in urban and rural areas, used to implement the dynamic risk assessment method for the entire life cycle of coronary heart disease in a hierarchical medical system in urban and rural areas as described in any one of claims 1 to 6, characterized in that, include: The standardized screening module is used to automatically collect, standardize, verify, and control the multi-dimensional health data of the target population through medical IoT terminal devices and wearable health monitoring devices; The risk dynamic assessment model module is used to build the GBDT risk assessment model. It integrates CUSUM time-series drift detection and EWMA smoothing processing, and weights and merges the GBDT static score and the time-series drift comprehensive score into a dynamic comprehensive risk score to complete the risk level determination.

8. A dynamic risk assessment system for the entire life cycle of coronary heart disease in a hierarchical medical system in urban and rural areas, as described in claim 7, is characterized in that... The risk dynamic assessment model module includes a model building and training submodule, a CUSUM time-series drift detection and EWMA smoothing submodule, a risk level determination submodule, and a risk dynamic update submodule. The output of the model building and training submodule is connected to the input of the risk level determination submodule, and the output of the risk dynamic update submodule is fed back to the input of the risk level determination submodule.

9. A computer device, characterized in that, It includes a memory and a processor, which are interconnected and communicate with each other. The memory stores computer instructions, and the processor executes the computer instructions to perform the dynamic risk assessment method for the whole cycle of coronary heart disease in urban and rural hierarchical medical treatment scenarios as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the method for dynamic risk assessment of coronary heart disease throughout the entire life cycle for urban and rural hierarchical medical treatment scenarios as described in any one of claims 1 to 6.

Images