Cerebrovascular disease risk assessment method

Abstract

The invention discloses a cerebrovascular disease risk assessment method, and relates to the technical field of risk prediction, and the method comprises the steps: determining a target population, and collecting first multi-dimensional data including demographic information, biochemical index information, lifestyle information and target risk type exclusive data; preprocessing the second multi-dimensional data to obtain second multi-dimensional data, and screening key feature variables by adopting a category-type feature improvement algorithm; constructing a machine learning model matched with the target risk, and outputting a risk assessment result or analyzing a variable risk association relationship after training; combining crowd characteristics or variable dependency relationship optimization model output, and verifying efficiency through decision curve analysis, an ROC curve and a confusion matrix; and finally integrating results to generate an evaluation report containing risk levels and key influence factors. According to the invention, through multi-dimensional data acquisition, scenarized model construction and two-dimensional verification, the assessment is more suitable for clinical requirements, and effective guidance can be provided for cerebrovascular disease risk prevention and control.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of risk prediction, in particular to a cerebral vascular disease risk assessment method. BACKGROUND

[0002] Cerebral vascular disease is an important disease that threatens the health of the population, especially in the middle-aged and elderly population, the incidence and mortality rate is high, not only seriously affects the quality of life of patients, but also brings heavy burden to the family and social medical system.

[0003] At present, the technical means for cerebral vascular disease risk assessment has gradually developed, for example, some studies use traditional statistical models to analyze the correlation between risk factors and disease incidence, and screen key risk variables through single factor or multi-factor analysis; Another study introduces machine learning algorithm to improve risk prediction accuracy, and optimizes the evaluation effect through feature selection or model construction.

[0004] However, the existing cerebral vascular disease risk assessment technology focuses on model construction itself, such as using random forest, extreme gradient boosting and other algorithms for risk prediction, but does not optimize and adjust the model according to the characteristics of different populations, resulting in prediction bias of the model in a specific group. Most machine learning models ignore clinical utility evaluation methods such as decision curve analysis, making it difficult to form an evaluation report that combines data support and clinical guidance. SUMMARY

[0005] In order to solve the technical problems existing in the prior art, the present application provides a cerebral vascular disease risk assessment method.

[0006] The technical scheme adopted by the present application is: a cerebral vascular disease risk assessment method, comprising the following steps:

[0007] S1, determine the target population of cerebral vascular disease risk assessment, collect the first multi-dimensional data of the target population; the first multi-dimensional data at least includes demographic information, biochemical index information and lifestyle information, and also includes exclusive data matched with the target risk type; the demographic information includes age, gender; the biochemical index information includes low density lipoprotein cholesterol, high density lipoprotein cholesterol, triglyceride-glucose index, apolipoprotein A and apolipoprotein B ratio, body mass index; the lifestyle information includes smoking status;

[0008] S2, pre-process the first multi-dimensional data to obtain second multi-dimensional data; the pre-processing includes rejecting samples with a missing proportion of key indicators exceeding the standard, and also includes exclusive pre-processing operations matched with the target risk type; using a category feature enhancement algorithm to analyze the feature importance of the second multi-dimensional data, and screening out key feature variables directly related to the target risk assessment;

[0009] S3, constructing a machine learning-based cerebral vascular disease risk assessment model taking the key feature variables as input; the type of the cerebral vascular disease risk assessment model matches the target risk type, the model is trained using a training set, and the cerebral vascular disease risk assessment model is used to output a risk assessment result or an association between an analytical variable and a target risk;

[0010] S4, based on the association between the variables and the target risk, the model output result is optimized and adjusted in combination with the population characteristics or variable dependence relationship of the assessment object; the performance of the optimized model is verified using a verification set, the clinical utility is verified through decision curve analysis, the prediction accuracy is verified through a receiver operating characteristic curve and a confusion matrix, and a verification result is obtained;

[0011] S5, based on the verification result, integrating the risk assessment result, the association between the key feature variables and the target risk, and generating a cerebral vascular disease risk assessment report containing risk levels and key influencing factors.

[0012] Preferably, the target risk type is the recurrence risk of cerebral vascular disease; and the target population is patients who have been diagnosed with cerebral vascular disease.

[0013] The exclusive data in step S1 includes disease-related data and treatment adherence data, the disease-related data includes first onset type, lesion site, interval time from onset to assessment, family history of cerebral vascular disease, and history of hypertension, diabetes, and dyslipidemia, the treatment adherence data includes anti-platelet drug use, lipid-lowering drug use, and blood pressure and blood glucose control compliance; the biochemical index information further includes plasma atherogenic index and qualitative results of urinary protein; and the lifestyle information further includes drinking status.

[0014] Preferably, the exclusive preprocessing operation in step S2 includes filling in missing continuous variable data and categorical variable data; the continuous variable data includes low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglyceride-glucose index, apolipoprotein A to apolipoprotein B ratio, body mass index, and plasma atherogenic index.

[0015] The categorical variable data includes smoking status, drinking status, family history of cerebral vascular disease, history of underlying diseases, anti-platelet drug use, lipid-lowering drug use, blood pressure and blood glucose control compliance, and qualitative results of urinary protein.

[0016] The key feature variables include disease-related features, physiological index features, and treatment-related features.

[0017] Preferably, the machine learning model in step S3 comprises an extreme gradient boosting main model and a Bayesian network auxiliary model; the extreme gradient boosting main model is used to output the recurrence probability of the evaluation object; and the Bayesian network auxiliary model is used to analyze the direct and indirect correlation between the key feature variable and the recurrence.

[0018] Preferably, the population feature in step S4 is gender; and the optimization adjustment mode is: for evaluation objects of different genders, the weight of the key feature variable in the extreme gradient boosting main model is adjusted respectively, so that the recurrence probability matches the actual risk characteristics of the evaluation object of the gender.

[0019] Preferably, the target risk type is the risk that the intervention response of cerebrovascular disease does not reach a preset effect; the target population is a population carrying known risk factors of cerebrovascular disease; the exclusive data in step S1 is intervention measure execution information, including the execution degree and execution persistence of the intervention measure; and the lifestyle information further comprises diet structure information.

[0020] Preferably, the exclusive preprocessing operation in step S2 comprises converting continuous information into categorical information according to medical specifications; the continuous information comprises body mass index, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, apolipoprotein A and apolipoprotein B ratio; the medical specifications comprise body mass index classification according to obesity diagnosis specifications, blood lipid related indicators classification according to blood lipid abnormality management specifications, and intervention measure execution degree classification according to clinical intervention effect evaluation specifications; and the preset effect comprises blood pressure control reaching the standard, blood glucose control reaching the standard, blood lipid index improvement and disease attack frequency reduction.

[0021] Preferably, the machine learning model in step S3 is a Bayesian network model; a directed acyclic graph structure is constructed based on variable dependency relationship; and the Bayesian network model is used to output the intervention response risk of the evaluation object.

[0022] Preferably, the optimization adjustment in step S4 comprises: if the evaluation object has missing key feature variable state, the influence of the known variable state and the variable dependency relationship on the missing variable is inferred based on probability, and the risk output result is optimized in combination with the probability parameters trained by the Bayesian network model.

[0023] Preferably, the category type feature promotion algorithm in step S2 is a CatBoost algorithm; the preset standard of the performance verification in step S4 further comprises that the net benefit of the decision curve is higher than a preset threshold; and the risk level in step S5 is divided based on the numerical range of the risk evaluation result.

[0024] The beneficial effects of the present application are at least one of the following:

[0025] The category feature importance algorithm is used for analyzing the feature importance of the preprocessed second multi-dimensional data, so that the key feature variables directly related to the target risk can be identified based on the association logic of the data itself, the deviation caused by artificial subjective judgment is reduced, and the possibility of missing key features or redundant irrelevant features is reduced.

[0026] Based on the association relationship between the variables and the target risk, the model output result is optimized and adjusted in combination with the population characteristics or variable dependency relationship of the evaluation object, so that the prediction deviation of the model for a specific evaluation object can be reduced; at the same time, through the double-dimension performance verification of the decision curve analysis verification of clinical utility, the receiver operating characteristic curve and the confusion matrix verification of prediction accuracy, the applicability of the model in actual application can be more comprehensively investigated, and the application limitation caused by the disconnection of the model from the clinical demand or the unstable prediction can be reduced. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The method flowchart of the embodiment one of the present application is shown. DETAILED DESCRIPTION

[0028] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0029] Embodiment one provides a cerebrovascular disease risk assessment method, as shown in the following formula (I), including the following steps: Figure 1

[0030] S1, determining the target population of cerebrovascular disease risk assessment, collecting the first multi-dimensional data of the target population; the first multi-dimensional data at least includes demographic information, biochemical index information and lifestyle information, and also includes exclusive data matched with the target risk type; the demographic information includes age, gender; the biochemical index information includes low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglyceride-glucose index, apolipoprotein A and apolipoprotein B ratio, body mass index; the lifestyle information includes smoking status.

[0031] It should be noted that the target population refers to a specific group meeting the demand of the cerebrovascular disease risk assessment scene; the first multi-dimensional data refers to the original data set collected directly; the demographic information is the information reflecting the basic population characteristics of the evaluation object, which is used to distinguish the risk difference of different groups. The biochemical index information is the physiological and biochemical parameters obtained through laboratory detection, which is the recognized associated index of cerebrovascular disease risk; the lifestyle information refers to the daily behavior related information of the evaluation object, which can be supplemented with other dimensions according to the target risk type; the exclusive data is the scene data collected additionally according to the target risk type.

[0032] ​Specifically, the risk assessment of cerebrovascular diseases needs to determine the assessment range to avoid the result deviation caused by the generalization of the assessment object; at the same time, the formation of the risk is related to the demographic characteristics, physiological indicators, living habits and scene factors, and it is easy to miss the key risk information by relying on single dimension data, so it is necessary to collect the first multi-dimensional data.

[0033] In the specific implementation process, the target population can be defined in combination with clinical standards or risk screening results, for example: if the subsequent assessment is the recurrence risk of cerebrovascular diseases, patients who have been diagnosed with cerebrovascular diseases need to be screened according to the ICD-10 diagnostic standard (such as ischemic stroke ICD-10 code I63); if the subsequent assessment is the intervention response risk, the population carrying known risk factors of cerebrovascular diseases such as hypertension and diabetes needs to be screened.

[0034] The first multi-dimensional data collection includes extracting age and gender demographic information from electronic medical records, obtaining biochemical indicator information from the laboratory information system (LIS), including low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), triglyceride-glucose index (TyG index), apolipoprotein A and apolipoprotein B ratio (ApoA / ApoB), body mass index (BMI).

[0035] Lifestyle information is collected through a structured questionnaire; special data collection is adapted according to the target risk type, such as collecting disease-related data from electronic medical record history when assessing the recurrence risk, and collecting treatment compliance data from medication follow-up records; the collected data are stored in the database according to the unique identifier (such as ID number) of the assessment object, forming a complete first multi-dimensional data set for each assessment object, avoiding data fragmentation.

[0036] S2, pre-processing the first multi-dimensional data to obtain second multi-dimensional data; the pre-processing includes rejecting samples with a missing proportion of key indicators exceeding the standard, and also includes special pre-processing operations matched with the target risk type; a category feature enhancement algorithm is used to analyze the feature importance of the second multi-dimensional data, and key feature variables directly related to the target risk assessment are screened out.

[0037] It should be noted that the pre-processing refers to a set of operations for quality optimization of the first multi-dimensional data, including rejecting samples with a missing proportion of key indicators exceeding the standard and personalized processing matched with the target risk type.

[0038] The proportion of missing key indicators exceeding the threshold means that the proportion of missing indicators that play a core role in risk assessment in the first multi-dimensional data exceeds the preset threshold. Such samples will interfere with the analysis results due to insufficient core information and need to be excluded. The exclusive preprocessing operation is a preprocessing operation customized according to the target risk type. For example, for recurrence risk, the missing data of medical history needs to be filled in, and for intervention response risk, the continuous intervention execution degree needs to be converted into a discrete type.

[0039] The category feature promotion algorithm refers to a machine learning algorithm that is good at processing category data and can output feature importance. In this scheme, the CatBoost algorithm is preferred, which can automatically process category features, reduce overfitting, and identify key variables through feature importance scores. The second multi-dimensional data is a high-quality data set obtained by preprocessing the first multi-dimensional data, which is the direct input for subsequent feature selection.

[0040] Key feature variables refer to variables that are directly related to the target risk assessment and are selected through feature importance analysis. For example, in the recurrence risk, the lesion site, and in the intervention response risk, the intervention execution persistence.

[0041] Considering that the first multi-dimensional data may have sample missing, data format not unified, variable redundancy, and other problems, directly using it for analysis will reduce the reliability of the results. At the same time, not all collected variables are directly related to the target risk, and redundant variables will increase the complexity of the model and reduce the operation efficiency. Therefore, it is necessary to improve the data quality through preprocessing and select key variables through a category feature promotion algorithm to provide more accurate input for subsequent modeling.

[0042] In the specific implementation process, samples with missing key indicators exceeding the threshold are excluded: according to the target risk type, key indicators are selected from the first multi-dimensional data, such as LDL-C, lesion site, and smoking status for recurrence risk, and BMI, intervention execution degree for intervention response risk. If the proportion of missing key indicators of a single sample is > 30%, the sample is excluded.

[0043] If the target risk is recurrence risk: for missing continuous variables (such as LDL-C, TyG index), the median is filled in, and for missing classification variables (such as smoking status, family history of disease), the mode is filled in (selecting the category with the highest frequency in the same population).

[0044] If the target risk is intervention response risk, continuous information (such as BMI, LDL-C) is converted into discrete types according to medical standards, for example, BMI is divided into <18.5 kg / m² (underweight), 18.5-24.0 kg / m² (normal), 24.0-28.0 kg / m² (overweight), and ≥28.0 kg / m² (obesity) according to the Guidelines for Prevention and Control of Overweight and Obesity in Chinese Adults; all variables after basic preprocessing and special preprocessing are integrated to form a second multi-dimensional data with unified structure and quality standards;

[0045] The second multi-dimensional data is input into the CatBoost algorithm, and parameters such as learning rate 0.1, tree depth 6, and iteration number 500 are set. The contribution of each variable to the target risk is calculated through algorithm iteration, that is, the feature importance score;

[0046] The top 80% of the variables in the feature importance score are selected as key feature variables (such as age, LDL-C, lesion site, and smoking status in the recurrence risk, and BMI, intervention execution persistence, and TyG index in the intervention response risk.

[0047] S3, the key feature variables are input to construct a machine learning-based cerebrovascular disease risk assessment model; the type of the cerebrovascular disease risk assessment model matches the type of the target risk, and the model is trained using the training set. The cerebrovascular disease risk assessment model is used to output risk assessment results or analyze the relationship between variables and target risks.

[0048] It should be noted that the machine learning-based cerebrovascular disease risk assessment model refers to a model constructed based on machine learning algorithms for assessing cerebrovascular disease risk, and the model type matches the target risk type, which means that the corresponding model is selected according to the characteristics of the target risk;

[0049] The training set refers to a subset of data from the second multi-dimensional data used for model training, which needs to cover different risk levels of evaluation objects to ensure the comprehensiveness of model training; the risk assessment result refers to the quantitative risk information output by the model, such as the recurrence risk probability and the risk level of intervention response not meeting the preset effect. The relationship between variables and target risks refers to the logic of how key feature variables affect target risks, which is different from a single risk result and can explain the causes of risk formation.

[0050] In the specific implementation process, if the target risk is the recurrence risk of cerebrovascular disease, a combined model of Extremely Gradient Boosting Main Model + Bayesian Network Auxiliary Model is constructed, the Extremely Gradient Boosting Main Model is used to quantify the risk probability, and the Bayesian Network Auxiliary Model is used to analyze the relationship between variables.

[0051] If the target risk is the risk that the intervention response for cerebrovascular disease does not achieve the preset effect, a Bayesian network model is constructed. The Bayesian network model presents the variable dependency relationship through a directed acyclic graph and can process the complex association between intervention execution and response effect.

[0052] S4, based on the association between the variable and the target risk, combining the population characteristics of the evaluation object or the variable dependency relationship, optimizing and adjusting the model output result; using the validation set to verify the performance of the optimized model, verifying the clinical utility through the decision curve analysis, verifying the prediction accuracy through the receiver operating characteristic curve and the confusion matrix, and obtaining the verification result;

[0053] It should be noted that the preset standard of the performance verification also includes that the net benefit of the decision curve is higher than the preset threshold; the population characteristics refer to the group attribution of the evaluation object, which will affect the association strength between the variable and the risk, such as different genders, different influence degrees of LDL-C on the recurrence risk;

[0054] The variable dependency relationship refers to the mutual influence relationship between the key characteristic variables, such as the mutual influence relationship between the intervention execution degree, blood pressure control and intervention response effect, and is mainly used to process variable missing or association analysis in the intervention response risk scenario.

[0055] Optimization and adjustment refers to the operation of modifying the model output result based on the variable association relationship and the population characteristics / variable dependency relationship, such as gender stratification adjustment of variable weight, missing variable inference through dependency relationship, and the purpose is to reduce the error caused by group bias or information missing.

[0056] The validation set is a data subset (20%-30% of the total data) for model performance verification divided from the second multi-dimensional data, which is independent of the training set and is used to objectively evaluate the actual performance of the model; the decision curve analysis (DCA) is a method for evaluating the clinical utility of the model, which calculates the net benefit (true positive benefit-false positive loss) to judge the clinical application value of the model at different thresholds, and the higher the net benefit, the stronger the guiding significance of the model to clinical decision-making; the receiver operating characteristic curve (ROC curve) is a method for evaluating the prediction accuracy of the model, which draws a curve with the false positive rate as the horizontal coordinate and the true positive rate as the vertical coordinate, and the larger the area under the curve (AUC), the stronger the ability of the model to distinguish risks; the confusion matrix quantifies the matching degree of the model prediction result and the actual result through four dimensions of true positive, false positive, true negative and false negative, and can calculate indicators such as accuracy and precision; the verification result refers to the evaluation conclusion of the integration of DCA, ROC curve and confusion matrix, which judges whether the model meets the preset standard.

[0057] Because the model may have group bias or information missing after training, it is necessary to optimize the results combined with the characteristics of the population or the dependence relationship of variables; At the same time, only through the training set validation cannot objectively reflect the actual application ability of the model, and the clinical utility and prediction accuracy need to be verified through the validation set to ensure that the model is not only accurate but also can provide effective guidance for clinical practice.

[0058] In the specific implementation process, if the target risk is recurrence risk: based on the correlation relationship of the Bayesian network auxiliary model analysis, it is found that "the influence weight of LDL-C on recurrence risk in men is 0.3, and that in women is 0.2"; the output results of the extreme gradient boosting main model are adjusted: the recurrence probability of male objects = original probability x (1+0.3x LDL-C over standard degree), and that of female objects = original probability x (1+0.2x LDL-C over standard degree);

[0059] If the target risk is intervention response risk: if the evaluation object is missing intervention execution continuity data, based on the dependence relationship of intervention execution degree→intervention execution continuity→response effect in the Bayesian network model; Through the known intervention execution degree = complete compliance, it is inferred that the probability of intervention execution continuity = high is 85%, and then the response risk is corrected combined with the model output;

[0060] The validation set sample is input into the optimized model, and the net income under different risk thresholds (0.1-0.9) is calculated; Draw the decision curve, if the curve is above the two reference lines of all samples predicted as positive and all samples predicted as negative, and the net income > 0, the model has clinical utility;

[0061] The prediction accuracy is verified by calculating the true positive rate and false positive rate of the validation set, drawing the ROC curve and calculating the AUC value. The confusion matrix is calculated by counting the true positive (model predicted positive and actual positive), false positive (model predicted positive but actual negative), true negative (model predicted negative and actual negative), and false negative (model predicted negative but actual positive), and calculating the accuracy and precision; Integrate the DCA net income, AUC value, accuracy and other indicators to determine whether it meets the preset standard, if not, return to step S2 to re-optimize feature selection or adjust the model parameters.

[0062] S5, based on the verification result, integrating the risk assessment result, the correlation between the key feature variables and the target risk, generating a cerebrovascular disease risk assessment report containing risk level and key influencing factors.

[0063] The risk level is divided based on the numerical range of the risk assessment result. Integration refers to integrating the risk assessment result output by the model, the association between the key feature variable and the target risk, and the verification result according to logical association to avoid scattered information. The risk level refers to the qualitative risk level divided based on the risk assessment result (such as the probability value), such as low risk, medium risk, and high risk, which is different from the quantitative risk assessment result and is more convenient for clinical understanding and application. The key influencing factor refers to the key feature variable with the highest contribution degree to the target risk and the corresponding association. The cerebrovascular disease risk assessment report refers to a structured document containing core assessment information, which is different from single risk value output and simultaneously presents risk level, risk reason, and reference basis to provide a clear direction for subsequent intervention.

[0064] If the risk assessment result (such as probability) output by the model and the variable association are presented separately, it is difficult for clinical personnel to quickly understand the risk profile and core reasons of the assessment object. At the same time, the verification result can prove the reliability of the assessment and needs to be included in the report to enhance the credibility. Therefore, a structured report needs to be generated through integration to convert quantitative results into "understandable and applicable" clinical information and promote the transformation of risk assessment results into actual intervention measures.

[0065] In the specific implementation process, the risk level is divided based on the numerical range of the risk assessment result to set the grading standard, for example: recurrence risk: risk probability < 30% is low risk, risk probability 30%-60% is medium risk, and risk probability > 60% is high risk; intervention response risk: risk probability < 20% is low risk, 20%-50% is medium risk, and > 50% is high risk.

[0066] Key influencing factor extraction is to select the key feature variables with a feature importance score of 3-5 and the corresponding association rules from the association between the variables and the target risk, for example:

[0067] Recurrence risk: extract LDL-C: contribution degree 28%, LDL-C > 3.4 mmol / L increases risk by 20%, smoking contribution degree 22%, current smoking increases risk by 18%, age contribution degree 18%, ≥ 65 years old increases risk by 15%;

[0068] Intervention response risk: extract intervention execution persistence contribution degree 32%, low persistence increases the risk of substandard response by 25%, BMI contribution degree 25%, ≥ 28 kg / m² increases risk by 20%;

[0069] Report structure design: includes the basic information of the evaluation object: the attribution of the target population, age, gender; risk assessment results: quantitative risk value such as recurrence probability 89.28% and risk level high risk; key influencing factors ranked by contribution degree and association, such as LDL-C exceeding the standard being the primary influencing factor, increasing the recurrence risk by 20%; evaluation reliability statement: verification result summary (such as AUC=0.76, accuracy rate 92.28%); preliminary suggestion direction: intervention direction based on key influencing factors (such as controlling LDL-C level, quitting smoking); the above information is integrated into a structured document through an automated tool (such as Python report generation library).

[0070] Embodiment two

[0071] On the basis of embodiment one, the target risk type is recurrence risk of cerebrovascular disease; the target population is patients who have been diagnosed with cerebrovascular disease;

[0072] The exclusive data in step S1 includes disease-related data and treatment adherence data, the disease-related data includes first onset type, lesion site, interval time from onset to evaluation, family history of cerebrovascular disease, history of hypertension, history of diabetes, and history of dyslipidemia, the treatment adherence data includes anti-platelet drug taking situation, lipid-lowering drug taking situation, and blood pressure and blood glucose control standard reaching situation; the biochemical index information further includes plasma atherosclerosis index and qualitative results of urine protein; the lifestyle information further includes drinking status.

[0073] It should be noted that the disease-related data refers to the medical history data specifically used for recurrence risk assessment, which is different from the intervention measure implementation information for intervention response risk, including first onset type, lesion site, interval time from onset to evaluation, family history of cerebrovascular disease, and past medical history; the treatment adherence data refers to the medication and control data specifically used for recurrence risk assessment, including anti-platelet drug taking situation, lipid-lowering drug taking situation, and blood pressure and blood glucose control standard reaching situation.

[0074] In the specific implementation process, the disease-related data is extracted from the electronic medical record discharge summary and medical history collection module, for example, the first onset type is based on ICD-10 coding (ischemic stroke I63), and the lesion site is based on the head CT / MRI report;

[0075] The treatment adherence data is extracted from the medication follow-up record and outpatient review record, for example, the blood pressure reaching standard situation is counted in the last 3 months of 12 measurements, and the number of times of systolic pressure <140 mmHg and diastolic pressure <90 mmHg accounts for the proportion.

[0076] In a possible implementation, the exclusive pre-processing operation in step S2 includes filling in missing continuous variable data and categorical variable data; the continuous variable data includes low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglyceride-glucose index, apolipoprotein A to apolipoprotein B ratio, body mass index, and plasma atherogenic index;

[0077] The categorical variable data includes smoking status, drinking status, family history of cerebrovascular disease, history of underlying diseases, anti-platelet drug use, lipid-lowering drug use, blood pressure and blood glucose control, and urine protein qualitative results.

[0078] The key feature variable includes disease-related features, physiological indicator features, and treatment-related features.

[0079] It should be noted that the continuous variable data in the recurrence risk scenario refers to continuous numerical variables in disease-related data and biochemical indicators, including low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglyceride-glucose index, apolipoprotein A to apolipoprotein B ratio, body mass index, and plasma atherogenic index (AIP, calculation method: log10(TG / HDL-C);

[0080] The categorical variable data in the recurrence risk scenario refers to category variables in lifestyle, medical history, and treatment compliance, including smoking status, drinking status, family history of cerebrovascular disease, history of underlying diseases, anti-platelet drug use, lipid-lowering drug use, blood pressure and blood glucose control, and urine protein qualitative results (negative / positive).

[0081] The key feature variable classification refers to disease-related features, physiological indicator features, and treatment-related features according to data sources.

[0082] The exclusive pre-processing operation refers to filling in missing continuous variable plasma atherogenic index (AIP) samples with the median of the population, and filling in missing categorical variable urine protein qualitative results samples with the mode (negative).

[0083] The key feature variable screening refers to screening out disease-related features such as first onset type, lesion site, family history of cerebrovascular disease, and hypertension history, physiological indicator features such as age, gender, LDL-C, AIP, and urine protein qualitative results, and treatment-related features such as anti-platelet drug use and blood pressure control.

[0084] In a possible implementation, the machine learning model in step S3 includes an extreme gradient boosting main model and a Bayesian network auxiliary model; the extreme gradient boosting main model is used to output the recurrence probability of the evaluation object; and the Bayesian network auxiliary model is used to analyze the direct and indirect correlation between the key feature variable and the recurrence. In step S4, the population characteristics are gender, and the optimization adjustment mode is: for evaluation objects of different genders, the weights of the key feature variables in the extreme gradient boosting main model are adjusted respectively, so that the recurrence probability matches the actual risk characteristics of the evaluation object of the gender.

[0085] Based on example one, this example is a brain vascular disease recurrence risk scenario, and the specific construction process of the extreme gradient boosting main model and the Bayesian network auxiliary model is described in detail.

[0086] In this example, the target population is 950 patients who have been diagnosed with ischemic cerebral stroke (ICD-10 code I63), which is divided into a training set (665 cases) and a validation set (285 cases) according to a 7:3 stratified sampling method;

[0087] The key feature variable refers to a disease-related feature (first onset type = basal ganglia infarction, family history of cerebrovascular disease = yes / no, history of hypertension = yes / no) selected by the CatBoost algorithm, a physiological index feature (age, gender, LDL-C, AIP, qualitative result of urine protein), and a treatment-related feature (anti-platelet drug taking situation, blood pressure control compliance); and the output label is whether the training set sample will relapse within 2 years of follow-up, with 199 relapses and 466 non-relapses.

[0088] The extreme gradient boosting main model adopts a multi-decision tree serial integration architecture, and the core level includes: an input layer that receives standardized data of 10 key feature variables (continuous variables are normalized to the [0, 1] interval, and classification variables are one-hot encoded);

[0089] The basic decision tree layer is composed of 800 regression decision trees, and the parameters are: tree depth = 6, maximum leaf node number of each tree = 32, and each tree fits the prediction residual of the previous model; the weight distribution layer distributes weights to each decision tree, and dynamically adjusts the weights based on the prediction error of the tree; and the output layer accumulates the weighted prediction results of all decision trees, and converts them into recurrence probability values through an activation function.

[0090] The gradient boosting main model includes a loss function (log loss function) for measuring the deviation between the predicted value of the gradient boosting main model and the actual label, and the formula is as follows:

[0091]

[0092] wherein, Total number of training set samples, corresponding to 665 training set samples in the embodiment; Actual recurrence label of the i-th sample, taking values {0, 1} (1 represents that the sample is diagnosed with recurrence within 2 years of follow-up, corresponding to 199 recurrence samples in the embodiment; 0 represents non-recurrence, corresponding to 466 non-recurrence samples); The recurrence probability prediction value of the i-th sample, taking values in the range [0, 1] (for example, the recurrence probability of a 65-year-old male patient in the embodiment is 89.28%). When or , the logarithmic term will be infinite, and the prediction value needs to be processed at the boundary: , wherein ; in the embodiment, the loss value of each iteration is calculated by the function, and the training is terminated when the loss fluctuation is less than 0.001 for 50 consecutive iterations.

[0093] In each iteration, the negative gradient of the loss function with respect to the current model prediction value is calculated as the pseudo-residual, and the formula is as follows:

[0094]

[0095] , wherein t is the current iteration round, corresponding to the t-th decision tree, and in the embodiment, t takes values in the range of 1~800; The pseudo-residual of the i-th sample in the t-th round, for example, the prediction probability of a recurrence sample in the embodiment is for the first t-1 rounds, then the pseudo-residual is ; ; The recurrence probability prediction value integrated by the first t-1 decision trees, for example, when the 100th iteration is performed, the prediction result of the first 99 trees for the i-th sample is obtained.

[0096] Boundary processing is performed on to avoid a denominator of 0; in the embodiment, when the pseudo-residual is positive, it indicates that the prediction probability of the current model for the recurrence sample is low, and needs to be corrected by a new decision tree.

[0097] The contribution weight of each decision tree to the final prediction result is determined, and optimization is performed by minimizing the loss function:

[0098]

[0099] The weight of the t-th decision tree, for example, the weight of the 50th decision tree in the embodiment is 0.08 after optimization; The learning rate, which is fixed at 0.05 in the embodiment, is used to control the contribution strength of a single tree to avoid overfitting; The prediction output of the i-th sample for the t-th decision tree, the maximum number of leaf nodes of each decision tree in this embodiment is 32, and the output range is [-1, 1]; The key feature variable set of the i-th sample, in the corresponding embodiment, 10 key features such as age, gender, LDL-C, AIP, etc.

[0100] Fuse the prediction results of all decision trees, output the final recurrence probability, formula content:

[0101]

[0102] Sigmoid activation function, used to map the weighted sum of decision trees to the [0, 1] interval to get the final recurrence probability; The total number of decision trees, which is fixed at 800 in this embodiment, and can also be terminated early according to the loss convergence; The key feature variables of the i-th sample, such as the feature combination of "65-year-old male, LDL-C = 4.2 mmol / L, and smoking = yes" in the embodiment.

[0103] In this embodiment, the above formula is used to calculate the initial recurrence probability, for example, the initial prediction probability of a male patient after 800 tree integration is 72%, and the optimized probability is 89.28% combined with the gender weight analysis of the Bayesian network; If the loss fluctuation is less than 0.001 for 50 consecutive iterations, the iteration can be stopped early, and there is no need to train to 800 trees.

[0104] The Bayesian network auxiliary model adopts a directed acyclic graph (DAG) architecture, and the core components include: node layer: 11 nodes, corresponding to key feature variables and target variables, specifically:

[0105] Feature nodes: age, gender, LDL-C, AIP, smoking status, family history of cerebrovascular disease, history of hypertension, anti-platelet drug use, blood pressure control compliance, and urine protein qualitative results;

[0106] Target node: recurrence risk (binary: yes / no);

[0107] Directed edge layer: based on variable association relationship to build directed edges, such as gender→LDL-C, LDL-C→AIP, AIP→recurrence risk, family history of cerebrovascular disease→recurrence risk, etc., a total of 13 directed edges;

[0108] Conditional probability table (CPT) layer: each node corresponds to a conditional probability table, which stores the conditional probability of the node under all value combinations of the parent nodes (such as the probability of LDL-C > 3.4 mmol / L under gender = male and gender = female). ​

[0109] The network structure is constructed using the Minimum Maximal Hill Climbing (MMHC) algorithm, and the conditional probabilities are calculated using maximum likelihood estimation, as follows:

[0110] The scoring function BIC is used to evaluate the goodness of fit of the network structure, and the formula is as follows:

[0111]

[0112] in, The likelihood function is used to measure the model parameters. For training set data The degree of fit, Key features and recurrence label data of the 665 training sets in the corresponding embodiment; These are model parameters, i.e., the conditional probabilities of each node. For example, in the embodiment, the probability of "LDL-C > 3.4 mmol / L" is 68% when "gender = male". The total number of model parameters is 42 independent probability values ​​in the conditional probability table of the 11 nodes in this example. This represents the total number of training set samples, corresponding to 665 samples in this example.

[0113] Hill Climbing Search: Generate candidate structures (add, delete, or reverse directed edges); calculate the BIC value of the candidate structures and retain the structure with the best BIC value; repeat the search until the BIC value no longer increases, and obtain the final directed acyclic graph; Parent Node Limitation: Set the maximum number of parent nodes for each node to 3 to avoid excessive complexity in the network structure.

[0114] Conditional probability calculation (maximum likelihood estimation): Calculate the conditional probability of each node under the combination of values ​​of its parent node. The formula is as follows:

[0115]

[0116] Where X is the current node, which can be a feature node or a target node, x is a specific value of node X, Pa(X) is the set of parent nodes of node X, pa is a combination of values ​​of the set of parent nodes Pa(X), and I(·) is an indicator function that takes the value 1 if the condition in parentheses is met, otherwise takes the value 0. , For the first The values ​​of node X and its parent node in each sample.

[0117] In this embodiment, the continuous features are first discretized and then the conditional probability is calculated. If the number of samples corresponding to a certain combination of parent node values ​​is 0 (e.g., "gender = female and AIP = exceeding the standard and family history = yes" is only 1 case), Laplace smoothing (adding 1 smoothing) is used to avoid the probability being 0.

[0118] The strength formula of direct association relationship is used to quantify the association degree of the direct directed edge between two nodes, and analyze the direct influence of characteristic variable on the recurrence risk. The formula is as follows:

[0119]

[0120] Wherein, represents the direct directed edge from the characteristic node X to the target node Y (recurrence risk), is the fixed value combination of X. is the “recurrence = yes” value of the target node Y; represents the parent node set of the target node Y except X, is the fixed value combination of X.

[0121] The direct association strength of LDL-C to the recurrence risk in male is 0.3, and that in female is 0.2, which provides a basis for subsequent weight adjustment of the main model. The association strength is in the range of [0, 1], and the greater the value is, the more significant the direct influence of X on Y is.

[0122] The strength formula of indirect association relationship is used to quantify the indirect influence of characteristic variable on the recurrence risk through the intermediate node. The formula is as follows:

[0123]

[0124] Wherein, represents the indirect association path from the characteristic node X to the target node Y through the intermediate node Z, which is smoking→LDL-C→recurrence risk in this embodiment; : the direct association strength of X to Z, which is 0.22 in this embodiment; : the direct association strength of Z to Y, which is 0.28 in this embodiment; and the final indirect association strength is 0.22*0.28=0.0616.

[0125] In this embodiment, the strength is calculated only for the indirect path with a length of 2. The longer path such as age→AIP→LDL-C→recurrence risk can be expanded through the product of multiple direct strengths. The indirect association strength needs to be analyzed in combination with the direct association strength. For example, the direct strength of smoking to the recurrence risk is 0.22, and the indirect strength is 0.0616 in this embodiment, and the total influence is mainly the direct association.

[0126] Embodiment Three

[0127] ​​​On the basis of the embodiment one, the target risk type is a risk that an intervention response for a cerebrovascular disease does not reach a preset effect; the target population is a population carrying known risk factors of the cerebrovascular disease; the exclusive data in the step S1 is intervention measure execution information, including an execution degree and an execution persistence of the intervention measure; and the lifestyle information further includes diet structure information.

[0128] It should be noted that the intervention measure execution information refers to operation type data that is used for intervention response risk assessment, including an execution degree of the intervention measure, such as whether a dose of antihypertensive drugs meets a standard or whether a frequency of exercise meets a requirement; and an execution persistence, such as a proportion of days of continuous execution of the intervention measure in a recent one month.

[0129] The population carrying known risk factors of the cerebrovascular disease refers to a population carrying known risk factors of the cerebrovascular disease in the Chinese Guidelines for Prevention and Treatment of Cerebrovascular Disease, and the known risk factors include a history of hypertension, a history of diabetes, a history of dyslipidemia, a family history of cerebrovascular disease, a smoking status, and a body mass index (BMI) exceeding a normal range (≥24 kg / m²).

[0130] In a specific implementation process, the target population is determined through risk factor screening questionnaires and laboratory detection screening, for example, patients with hypertension (systolic blood pressure ≥140 mmHg) combined with dyslipidemia (LDL-C ≥3.4 mmol / L) are included.

[0131] The intervention measure execution information is collected by recording the frequency of exercise through an intelligent wearable device and recording the dose and days of medication through a medication APP, for example, an execution degree of antihypertensive drugs = dose meeting a standard (daily use of 5 mg of amlodipine, in line with a medical order), and an execution persistence = 27 days of continuous medication in a recent 30 days (a proportion of 90%).

[0132] In a possible implementation, the exclusive preprocessing operation in the step S2 includes converting continuous type information into discrete type information according to medical specifications; the continuous type information includes a body mass index, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, a ratio of apolipoprotein A to apolipoprotein B; the medical specifications include classification of the body mass index according to a diagnosis specification of obesity, classification of blood lipid related indexes according to a management specification of dyslipidemia, and classification of an execution degree of the intervention measure according to a clinical intervention effect evaluation specification; and the preset effect includes standard blood pressure control, standard blood glucose control, improvement of blood lipid indexes, and reduction of disease attack frequency.

[0133] Medical norms refer to authoritative standards for continuous information classification in intervention response scenarios, including obesity diagnosis norms (WHO or Chinese guidelines), blood lipid abnormality management norms (2023 ESC / EAS blood lipid abnormality management guidelines), and clinical intervention effect evaluation norms (China Chronic Disease Prevention and Control Work Norms); preset effects refer to health goals expected to be achieved by intervention measures, including blood pressure control compliance (systolic pressure < 140 mmHg and diastolic pressure < 90 mmHg), blood glucose control compliance (fasting blood glucose < 7.0 mmol / L), blood lipid index improvement (LDL-C < 2.6 mmol / L), and disease onset frequency reduction (stroke onset frequency reduction ≥ 50% compared with before intervention).

[0134] In the specific implementation process, BMI is divided into < 18.5 (slim), 18.5-24.0 (normal), 24.0-28.0 (overweight), and ≥ 28.0 (obesity) according to the China Adult Overweight and Obesity Prevention and Control Guidelines.

[0135] LDL-C is divided into < 1.8 mmol / L (extremely high-risk group target), 1.8-2.6 mmol / L (high-risk group target), 2.6-3.4 mmol / L (medium-risk group target), and ≥ 3.4 mmol / L (low-risk group target) according to the 2023 ESC / EAS guidelines; intervention implementation degree is divided into complete compliance (implementation degree ≥ 90%), partial compliance (60%-89%), and non-compliance (< 60%) according to the China Chronic Disease Prevention and Control Work Norms.

[0136] In one possible implementation, the machine learning model in step S3 is a Bayesian network model; a directed acyclic graph structure is constructed based on variable dependency relationships, and the Bayesian network model is used to output the intervention response risk of the evaluation object. The optimization adjustment in step S4 includes: if there is a missing state of a key feature variable of the evaluation object, the influence of the missing variable on the known variable state and the variable dependency relationship is probabilistically inferred, and the risk output result is optimized in combination with the probability parameters trained by the Bayesian network model.

[0137] Based on example one, this embodiment is a risk scenario of brain vascular disease intervention response not reaching the preset effect, and the construction process of the Bayesian network model is described in detail.

[0138] The target population in this embodiment is 800 patients with hypertension (systolic pressure ≥ 140 mmHg) combined with blood lipid abnormalities (LDL-C ≥ 3.4 mmol / L); they are divided into a training set (denoted as , 560 cases) and a validation set (denoted as , 240 cases) according to 7:3 stratified sampling, and the proportion of intervention response non-compliance samples in both groups is 28%, ensuring consistent data distribution.

[0139] The key variables in this embodiment include 6 feature variables (denoted as...). ) and 1 target variable (denoted as These are all categorical variables used to construct the nodes of the Bayesian network. The Bayesian network model outputs the risk of intervention response failure, while also addressing the sustainability of intervention implementation. Risk inference problem when missing.

[0140] The Bayesian network model adopts a three-layer architecture: a node layer, a directed edge layer, and a conditional probability table (CPT) layer, as follows: The node layer has a total of 7 nodes. .in, (Age): Values ​​can be... <55 years old , ( ) ( ) Par( )= .

[0141] ( , Overweight, 24.0-28.0 ) ( No parent node, i.e. (LDL-C): Values ​​are... <2.6mmol / L , ( ) ( ) Par( )={ , }

[0142] (Smoking status): Values ​​are... (yes), (No); No parent node, i.e. . (Intervention Implementation Level): Values ​​are... (Not up to standard, <60%) (Partially met the standard, 60%-89%) (Fully compliant, ≥90%); No parent node, i.e. . (Intervention Implementation Continuity): Values ​​are... (Low, <60%) (In China, 60%-89%) (high, ≥ 90%); parent node is and , i.e. . (intervention response substandard): takes value (yes), (no); parent node is and , i.e. .

[0143] There are 6 directed edges, represented by parent node→child node, reflecting the dependency relationship between nodes, determined based on clinical logic and training set data association analysis, specifically: , , , , , . For example indicates that the higher the intervention implementation degree, the easier it is to maintain the intervention implementation persistence, which conforms to the clinical intervention law.

[0144] Conditional probability table (CPT) layer The CPT layer stores model parameters , i.e. the conditional probability of each node under all value combinations of its parent nodes. For example stores the CPT as:

[0145] (total of 3×3×3=27 probability values), and the CPT of G stores (total of 2×3×3=18 probability values). All probability values are calculated through subsequent parameter learning.

[0146] BIC score function (select the optimal directed edge) Structure learning is to select the network structure with the highest goodness of fit and the lowest complexity through the BIC score function, the formula is as follows:

[0147]

[0148] : likelihood function, measures the fitting degree of parameters to the data set , the calculation method is the logarithm of the product of the joint probability of all samples under the current structure, i.e. ; where is the number of samples in the data set , represents the value of in the th sample, represents the value of in the th sample. Total number of model parameters, total number of independent probability values of all CPTs in this scenario is 48 (the sum of the probability values of each node CPT is 1, so the repeated count needs to be deducted). The number of samples of the data set The number of samples of the training set .

[0149] Conditional probability maximum likelihood estimation (calculate ) Parameter learning calculates the conditional probability of each node through the sample frequency statistics of the training set , The formula is as follows:

[0150]

[0151] The number of samples of the training set 560. Indicates the indicator function, if the i-th sample satisfies and the parent node set , then the value is 1, otherwise 0. Indicates a specific value of feature node M, for example Indicates that the intervention execution persistence is high. Indicates a specific value combination of the parent node set of feature node M, for example Indicates that the intervention execution degree is fully met and the BMI is normal.

[0152] Exemplary:

[0153]

[0154] In the training set , there are 125 samples that satisfy and , among which there are 82 samples of , and the formula is:

[0155]

[0156] That is, when the intervention execution degree = fully met and the BMI = normal, the conditional probability of the intervention execution persistence = high is 65.6%.

[0157] Bayes theorem (for missing) When the evaluation object is missing (intervention execution persistence), the value probability of is inferred based on the known variables (such as ), The formula is as follows:

[0158] ​​

[0159] where, represents a certain value of missing variable Z, e.g. represents high intervention adherence persistence. represents the value of known variable set K, e.g. represents complete compliance of known intervention adherence and normal BMI. represents all possible values of missing variable Z, in this scenario represents the value of (low, medium, high). represents the marginal probability of missing variable Z, obtained by counting the proportion of samples with each value of Z in the training set , e.g. represents that 42% of samples in the training set have high intervention adherence persistence.

[0160] Because is the parent node of G, according to the local Markov property of Bayesian networks, (i.e. the conditional probability obtained by parameter learning), the formula can be simplified as: ; for example, given , , which is consistent with the parameter learning result.

[0161] Combined with the inference result of missing variables, the probability of target variable G = g1 (intervention non-compliance) is calculated by weighting, and the formula is as follows:

[0162]

[0163] represents the test sample to be evaluated, i.e. the data of a patient who needs to calculate the intervention response risk. represents the known parent node value of target variable G, e.g. represents that LDL-C is known to be 2.6-3.4 mmol / L. represents the conditional probability obtained from the CPT of G, e.g. represents that when LDL-C = 2.6-3.4 mmol / L and intervention adherence persistence = high, the probability of intervention non-compliance is 15%. represents the inference probability of missing variable Z, e.g. .

[0164] Application example: given the known information of a test sample: missing, substitute into the formula to calculate:

[0165]

[0166] The risk of intervention response failure of the test sample is 23.8%, which is classified as medium risk according to the risk classification standard of 20%-50% in Example One, and is determined as medium risk.

[0167] Based on the validation set Model validation: Through ROC curve evaluation, the area under the curve (AUC) is calculated to be 0.78, indicating that the model has good ability to distinguish between intervention compliance and non-compliance; the formula is as follows:

[0168] The true positive rate (TPR) represents the proportion of actual non-compliance and model prediction of non-compliance TPR= ; Where: TP=42 true positives, actual non-compliance and prediction of non-compliance, FN=35 false negatives, actual non-compliance but prediction of compliance Calculation: TPR=

[0169] The false positive rate (FPR) represents the proportion of actual compliance but model prediction of non-compliance FPR= Where: FP=15 false positives, actual compliance but prediction of non-compliance, TN=148 true negatives, actual compliance and prediction of compliance Calculation: FPR=

[0170] AUC is the area of the ROC curve (FPR as the horizontal axis, TPR as the vertical axis) and the horizontal axis, which is calculated by traversing all risk thresholds and integrating. The calculation result of this validation set is 0.78, indicating that the model has good ability to distinguish between intervention compliance and non-compliance

[0171] The accuracy rate calculated by the confusion matrix is 81% (42 true positives, 15 false positives, 148 true negatives, and 35 false negatives), which meets the accuracy requirements of clinical application. The formula is as follows:

[0172] Accuracy rate=

[0173] Substitute the validation set data to calculate: TP=42, TN=148, FP=15, FN=35, total sample size=240 accuracy rate=

[0174] The clinical utility is evaluated by the decision curve (DCA), and within the risk threshold range of 0.2-0.7, the net benefit of the model is 0.15-0.22, which is higher than the reference lines of predicting all samples as positive and predicting all samples as negative, proving that the model can provide effective guidance for clinical intervention decision-making, the formula is as follows:

[0175] Net benefit=

[0176] Where: (total number of samples in the validation set), For risk threshold (0.2~0.7), TP, FP are true positive, false positive number under corresponding threshold. (corresponding TP=60, FP=30):

[0177] Net benefit When risk threshold (corresponding TP=48, FP=20): net benefit

[0178] In 0.2~0.7 threshold range, model net benefit is 0.15~0.22, higher than all sample prediction positive net benefit 0.08 and all sample prediction negative net benefit 0, proving its clinical decision guidance value.

[0179] The above-described embodiments only express the specific implementation of the present application, which is described in detail and specifically, but cannot be understood as the limitation of the patent scope of the present application. It should be noted that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which belong to the protection scope of the present application.

Claims

1. A method for assessing the risk of cerebrovascular disease, characterized by, The method comprises the following steps: S1, determining a target population for cerebrovascular disease risk assessment, collecting first multi-dimensional data of the target population; the first multi-dimensional data at least includes demographic information, biochemical index information and lifestyle information, and further includes exclusive data matched with a target risk type; the demographic information includes age and gender; the biochemical index information includes low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglyceride-glucose index, apolipoprotein A and apolipoprotein B ratio, and body mass index; the lifestyle information includes smoking status; S2, pre-processing the first multi-dimensional data to obtain second multi-dimensional data; the pre-processing includes eliminating samples with a missing proportion of key indicators exceeding a standard, and further includes exclusive pre-processing operations matched with the target risk type; performing feature importance analysis on the second multi-dimensional data by using a category feature promotion algorithm, and screening out key feature variables directly related to the target risk assessment; S3, taking the key feature variables as inputs to construct a cerebrovascular disease risk assessment model based on machine learning; the type of the cerebrovascular disease risk assessment model is matched with the target risk type, the model is trained by using a training set, and the cerebrovascular disease risk assessment model is used to output a risk assessment result or an association between variables and the target risk; S4, based on the association between the variables and the target risk, the model output result is optimized and adjusted in combination with the population characteristics of an assessment object or the variable dependency relationship; the performance of the optimized model is verified by using a verification set, the clinical utility is verified by using a decision curve analysis, the prediction accuracy is verified by using a receiver operating characteristic curve and a confusion matrix, and a verification result is obtained; S5, based on the verification result, integrating the risk assessment result, the association between the key feature variables and the target risk, and generating a cerebrovascular disease risk assessment report including a risk level and key influencing factors.

2. The method of claim 1, wherein, The target risk type is a recurrence risk of cerebrovascular disease; and the target population is a patient who has been diagnosed with cerebrovascular disease. In step S1, the exclusive data includes disease-related data and treatment compliance data; the disease-related data includes a first onset type, a lesion site, an interval time from onset to assessment, a family history of cerebrovascular disease, a history of hypertension, a history of diabetes, and a history of dyslipidemia; the treatment compliance data includes anti-platelet drug taking, lipid-lowering drug taking, and blood pressure and blood glucose control; the biochemical index information further includes a plasma atherogenic index and a qualitative result of urinary protein; and the lifestyle information further includes drinking status.

3. The method of claim 2, wherein the method is for assessing the risk of a cerebrovascular disease. In step S2, the exclusive pre-processing operation includes filling in missing continuous variable data and classification variable data; the continuous variable data includes low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglyceride-glucose index, apolipoprotein A and apolipoprotein B ratio, body mass index, and plasma atherogenic index. The categorical variable data includes smoking status, drinking status, family history of cerebrovascular disease, history of underlying diseases, anti-platelet drug use, lipid-lowering drug use, blood pressure and blood glucose control, and qualitative results of urinary protein; The key feature variables include disease-related features, physiological indicator features, and treatment-related features.

4. The method of claim 3, wherein the method is for assessing the risk of a cerebrovascular disease. The machine learning model in step S3 includes an extreme gradient boosting main model and a Bayesian network auxiliary model; the extreme gradient boosting main model is used to output the recurrence probability of the evaluation object; and the Bayesian network auxiliary model is used to analyze the direct and indirect association between the key feature variables and recurrence.

5. The method of claim 4, wherein the method further comprises: The population feature in step S4 is gender; and the optimization adjustment mode is: for evaluation objects of different genders, the weights of the key feature variables in the extreme gradient boosting main model are adjusted respectively, so that the recurrence probability matches the actual risk characteristics of the evaluation object.

6. The method of claim 1, wherein the method further comprises: The target risk type is the risk of cerebrovascular disease intervention response not reaching a preset effect; the target population is a population carrying known risk factors of cerebrovascular disease; the exclusive data in step S1 is intervention measure execution information, including execution degree and execution persistence of the intervention measure; and the lifestyle information further includes diet structure information.

7. The method of claim 6, wherein the method further comprises: The exclusive preprocessing operation in step S2 includes converting continuous information into discrete information according to medical specifications; the continuous information includes body mass index, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, and the ratio of apolipoprotein A to apolipoprotein B; the medical specifications include classification of body mass index according to obesity diagnosis specifications, classification of blood lipid-related indicators according to blood lipid abnormality management specifications, and classification of intervention measure execution degree according to clinical intervention effect evaluation specifications; and the preset effect includes blood pressure control, blood glucose control, blood lipid indicator improvement, and reduction of disease attack frequency.

8. The method of claim 7, wherein, The machine learning model in step S3 is a Bayesian network model; a directed acyclic graph structure is constructed based on variable dependency relationships, and the Bayesian network model is used to output the intervention response risk of the evaluation object.

9. The method of claim 8, wherein the method further comprises: The optimization adjustment in step S4 includes: if the evaluation object has missing key feature variable states, the influence of the missing variables on the known variable states is inferred based on probability, and the risk output result is optimized in combination with the probability parameters trained by the Bayesian network model.

10. The method of claim 1, wherein the method further comprises: The category feature improvement algorithm in step S2 is a CatBoost algorithm; the preset standard of the performance verification in step S4 further includes that the net benefit of the decision curve is higher than a preset threshold; and the risk level in step S5 is divided based on the numerical range of the risk assessment result.

Images