Cardio-cerebrovascular disease risk prediction method and system based on multi-modal data

Abstract

The present application relates to the technical field of disease risk prediction, in particular to a cardiovascular and cerebrovascular disease risk prediction method based on multi-modal data, comprising the following steps: S1: multi-modal follow-up data acquisition and standardization; S2: construction of time-effective longitudinal follow-up sequence; S3: generation of longitudinal risk features, taking the effective longitudinal follow-up sequence as input to generate current indicators, time interval-aware change features, reference range deviation, abnormal duration, continuous abnormality number and medical concept evidence; S4: cardiovascular and cerebrovascular comprehensive risk and stroke special risk assessment; S5: risk result reliability judgment and key data item to be supplemented identification; S6: generation of minimum necessary data supplement task; S7: feedback data update and risk closed-loop reevaluation. The present application also provides a cardiovascular and cerebrovascular disease risk prediction system based on multi-modal data. Through the collection and processing of multi-modal data, the present application can ensure the accuracy of risk prediction.

Description

Technical Field

[0001] This invention relates to the field of disease risk prediction technology, specifically to a method and system for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data. Background Technology

[0002] In existing technologies, the focus for chronic diseases, especially cardiovascular and cerebrovascular diseases, is mainly on the statistical management of patients' basic information and conditions. Although some risk prediction tools for such diseases have emerged at home and abroad, they still have the following shortcomings: the data collected is limited and lacks reliable verification of the data's validity, as well as the reliability of the risk assessment results is not effectively evaluated and key missing data is not identified, resulting in low accuracy of risk prediction results. Summary of the Invention

[0003] The first aspect of this invention is intended to provide a method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data, which provides more accurate risk prediction results.

[0004] A method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data includes the following steps:

[0005] Step S1: Acquisition and standardization of multimodal follow-up data. Acquire multimodal follow-up data related to the comprehensive cardiovascular and cerebrovascular risk and stroke-specific risk of the target personnel. Unify personnel identification, fields, units of measurement, medical concepts, data sources and actual collection time to form standardized multimodal follow-up data for use in Step S2.

[0006] Step S2: Construction of time-valid longitudinal follow-up sequence. Using the data output from step S1 as input, establish the follow-up period according to the actual collection time, convert the continuous monitoring data into periodic features, and combine the validity period, collection status and monitoring coverage to generate a valid longitudinal follow-up sequence for use in step S3.

[0007] Step S3: Generation of longitudinal risk features. Step S3 uses the effective longitudinal follow-up sequences output in Step S2. As input, generate current indicators, time interval perceived change characteristics, reference range deviation, abnormal duration, number of consecutive abnormalities, and medical conceptual evidence to form longitudinal risk features for use in step S4;

[0008] Step S4: Comprehensive cardiovascular and cerebrovascular risk assessment and stroke-specific risk assessment; Step S4 uses the longitudinal risk characteristics output from Step S3. As input, the participation level of various features is determined based on data completeness, timeliness, and validity. A comprehensive risk status value is first generated, and then corrected using stroke-specific risk evidence.

[0009] Step S5: Risk Result Reliability Judgment and Key Data Items to be Supplemented. Step S5 receives the comprehensive cardiovascular and cerebrovascular risk results and their risk levels, stroke-specific risk results and their risk levels, main risk factors, risk contribution values, as well as data quality coefficients, fusion coefficients, feature missing identifiers, medical concept evidence conflict strength, feature validity coefficients, model participation paths, and traceability information. It then performs risk result reliability judgment and identifies key data items to be supplemented. If the above data is insufficient, step S5 is skipped, and the risk support, risk level stability, and risk result reliability of the two risk tasks are set to zero, and the overall risk result reliability status is set to unreliable. All missing or invalid core inputs are used as key data items to be supplemented, and the risk impact value is taken as one and sent to step S6, without outputting the two types of risk results.

[0010] Step S6: Minimum Necessary Data Supplementation Task Generation. Step S6 takes the overall risk result reliability status and key data items to be supplemented as inputs, and combines the risk impact value, reliability gap, data acquisition cost, expected completion time and collection conditions to determine the minimum necessary data supplementation task set.

[0011] Step S7: Feedback data update and risk closed-loop review. Step S7 receives the data returned by the task in Step S6 and updates the longitudinal follow-up sequence according to the task identifier and the actual collection time.

[0012] The second aspect of this invention is intended to provide a cardiovascular and cerebrovascular disease risk prediction system based on multimodal data, including a data acquisition module for acquiring multimodal follow-up data and standardizing it, acquiring multimodal follow-up data related to the comprehensive cardiovascular and cerebrovascular risk and stroke-specific risk of the target personnel, and unifying personnel identification, fields, units of measurement, medical concepts, data sources and actual collection time;

[0013] The longitudinal follow-up sequence construction module is used to establish follow-up cycles according to the actual collection time, convert continuous monitoring data into periodic features, and generate valid longitudinal follow-up sequences by combining the validity period, collection status and monitoring coverage.

[0014] The longitudinal risk feature generation module, with effective longitudinal follow-up sequences As input, it generates current indicators, time interval perceived change characteristics, reference range deviation, abnormal duration, number of consecutive abnormalities, and medical conceptual evidence;

[0015] The comprehensive cardiovascular and cerebrovascular risk assessment module, which includes stroke-specific risk assessment, uses longitudinal risk characteristics. As input, the participation level of various features is determined based on data completeness, timeliness, and validity to generate a comprehensive risk status value, which is then corrected using stroke-specific risk evidence.

[0016] The risk outcome reliability assessment and key supplementary data item identification module is used to receive the comprehensive cardiovascular and cerebrovascular risk results and their risk levels, the stroke-specific risk results and their risk levels, the main risk factors, the risk contribution value, as well as the data quality coefficient, fusion coefficient, feature missing identifier, medical concept evidence conflict intensity, feature validity coefficient, model participation path and traceability information, and to perform risk outcome reliability assessment and key supplementary data item identification.

[0017] The minimum necessary data supplementation task generation module is used to determine the minimum necessary data supplementation task set by taking the overall risk result reliability status and key data items to be supplemented as input, combined with risk impact value, reliability gap, data acquisition cost, expected completion time and collection conditions.

[0018] The feedback data update and risk closed-loop review module is used to receive data returned by the task and update the longitudinal follow-up sequence according to the task identifier and the actual collection time.

[0019] The beneficial effects of this invention are as follows: it collects multimodal source data, effectively judges the source data, and generates current indicators, time interval perception change characteristics, reference range deviation, abnormal duration, number of consecutive abnormalities, and medical conceptual evidence, thereby accurately assessing the comprehensive risk of cardiovascular and cerebrovascular diseases and the specific risk of stroke, and judging the reliability of the assessment results and accurately identifying key missing data, thereby ensuring the accuracy of risk prediction. Attached Figure Description

[0020] Figure 1 This is a flowchart of the cardiovascular and cerebrovascular disease risk prediction method based on multimodal data according to the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described below are only for explaining the present invention and do not limit the scope of protection of the present invention.

[0022] The cardiovascular and cerebrovascular disease risk prediction system based on multimodal data disclosed in this embodiment includes the following four subsystems: a report management subsystem, including: a case reporting management module for managing case report information; and a reporting management module for reporting case data to higher-level medical institutions or management departments.

[0023] The follow-up management subsystem includes: a follow-up reminder module, used to automatically remind medical staff of follow-up times; a follow-up management module, used to manage follow-up records and follow-up plans; a migration in / out module, used to manage patient migration information; and a statistical analysis module, used for statistical analysis of follow-up data.

[0024] The stroke screening and intervention management subsystem includes: a preliminary screening module for initial screening of stroke risk; and a statistical analysis module for statistical analysis of screening and intervention data.

[0025] The stroke screening, intervention, and management subsystem also includes:

[0026] The data acquisition module is used to acquire multimodal follow-up data and standardize it, to acquire multimodal follow-up data related to the comprehensive cardiovascular and cerebrovascular risk and stroke-specific risk of the target personnel, and to unify personnel identification, fields, units of measurement, medical concepts, data sources and actual collection time.

[0027] The longitudinal follow-up sequence construction module is used to establish follow-up cycles according to the actual collection time, convert continuous monitoring data into periodic features, and generate valid longitudinal follow-up sequences by combining the validity period, collection status and monitoring coverage.

[0028] The longitudinal risk feature generation module, with effective longitudinal follow-up sequences As input, it generates current indicators, time interval perceived change characteristics, reference range deviation, abnormal duration, number of consecutive abnormalities, and medical conceptual evidence;

[0029] The comprehensive cardiovascular and cerebrovascular risk assessment module, which includes stroke-specific risk assessment, uses longitudinal risk characteristics. As input, the participation level of various features is determined based on data completeness, timeliness, and validity to generate a comprehensive risk status value, which is then corrected using stroke-specific risk evidence.

[0030] The risk outcome reliability assessment and key supplementary data item identification module is used to receive the comprehensive cardiovascular and cerebrovascular risk results and their risk levels, the stroke-specific risk results and their risk levels, the main risk factors, the risk contribution value, as well as the data quality coefficient, fusion coefficient, feature missing identifier, medical concept evidence conflict intensity, feature validity coefficient, model participation path and traceability information, and to perform risk outcome reliability assessment and key supplementary data item identification.

[0031] The minimum necessary data supplementation task generation module is used to determine the minimum necessary data supplementation task set by taking the overall risk result reliability status and key data items to be supplemented as input, combined with risk impact value, reliability gap, data acquisition cost, expected completion time and collection conditions.

[0032] The feedback data update and risk closed-loop review module is used to receive data returned by the task and update the longitudinal follow-up sequence according to the task identifier and the actual collection time.

[0033] The health prescription management subsystem includes: a health prescription maintenance module for maintaining and managing health prescription content; and an intelligent health prescription push module for intelligently pushing personalized health prescriptions based on the patient's condition.

[0034] Existing technologies only focus on the statistical management of information on chronic patients. This solution innovatively integrates four subsystems: report management, follow-up management, stroke screening and intervention management, and health prescription management, forming a closed-loop monitoring and management system from data collection, risk assessment, follow-up intervention to health management, thus realizing full-process, intelligent monitoring and management of cardiovascular and cerebrovascular diseases.

[0035] This embodiment also discloses a method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data. By collecting follow-up data from follow-up personnel, the method predicts the risk of cardiovascular and cerebrovascular diseases for follow-up personnel and assesses the stroke risk level of patients in stroke screening.

[0036] As attached Figure 1 As shown, it includes the following steps:

[0037] Step S1: Acquisition and standardization of multimodal follow-up data;

[0038] Step S1: Obtain multimodal follow-up data related to the comprehensive cardiovascular and cerebrovascular risks and specific stroke risks of the target personnel, unify personnel identification, fields, units of measurement, medical concepts, data sources and actual collection time, and form standardized multimodal follow-up data for use in step S2.

[0039] S1 specifically includes the following steps:

[0040] S11: Unify the identification of target personnel and obtain source data. Obtain source data from the follow-up management subsystem, hospital information system, laboratory examination system, stroke screening subsystem and continuous monitoring equipment, and map the system number to the unique identifier of the target personnel. Data whose affiliation cannot be confirmed is marked as personnel affiliation to be verified data and is not directly merged. The original record index of matched records is retained for traceability.

[0041] S12: Obtain basic cardiovascular and cerebrovascular risk and dynamic indicator data, including basic risk data such as age, gender, family history, history of hypertension, history of diabetes, history of dyslipidemia, history of atrial fibrillation, history of coronary heart disease, history of stroke, history of transient ischemic attack, smoking status, and medication adherence, as well as dynamic indicators such as blood pressure, heart rate, weight, blood glucose, and blood lipids; simultaneously retain the actual collection time, data source, raw values, and units of measurement;

[0042] S13: Acquire stroke-specific data and supplementary modal data, acquire ECG, carotid ultrasound, stroke initial screening and high-risk survey results, acquire neurological symptoms in follow-up texts, and home blood pressure, continuous heart rate, activity level and sleep duration generated by continuous monitoring devices; retain the original content, actual collection time and collection status, and do not collect original medical images and large-volume physiological signals that are not used in subsequent risk assessments.

[0043] S14: Standardize fields, units, medical concepts, and time. Use standard field dictionaries and medical concept dictionaries to unify fields, diseases, symptoms, and abnormal examination items, and convert blood pressure, blood sugar, blood lipids, weight, heart rate, activity duration, and sleep duration into unified units of measurement.

[0044] The inspection, examination, follow-up, and equipment sampling times will be uniformly converted into actual collection times; a time missing indicator will be set when actual collection time is missing.

[0045] S15: Generate standardized multimodal follow-up data, identify and merge duplicate records with the same content, source and actual collection time; set collection status indicators for data with missing field values, units of measurement or time, or data collection failures; retain the original value and set an anomaly candidate indicator for data that exceeds the reference range.

[0046] Step S1 outputs standardized multimodal follow-up data. Each record includes the target person's unique identifier, data type, standard fields, standard data values, standard units of measurement, data source, actual collection time, collection status, and original record index, and is then input into step S2.

[0047] Step S2: Construction of time-efficient longitudinal follow-up sequences;

[0048] Step S2 takes the data output from step S1 as input, establishes a follow-up period according to the actual collection time, converts the continuous monitoring data into periodic features, and generates an effective longitudinal follow-up sequence for use in step S3 by combining the validity period, collection status and monitoring coverage.

[0049] S2 specifically includes the following steps:

[0050] S21: Determine follow-up nodes and divide follow-up periods;

[0051] The actual follow-up time, stroke screening time, and risk reassessment time triggered by feedback data for the target individuals are obtained and arranged in chronological order to form a follow-up node sequence for the target individuals. Follow-up, screening, and risk reassessment events with the same actual time are merged into the same follow-up node.

[0052] The output of step S1 Standardized multimodal follow-up data Based on actual collection time The first node to be included in the adjacent follow-up nodes One follow-up period, of which To record the serial number, This is the sequence number of the follow-up node.

[0053] Data with missing actual collection time, equipment collection failure, or personnel affiliation pending verification will not be included in the follow-up cycle; data later than the current follow-up node will not participate in the current risk handling, while data with abnormal candidate markers will still be retained.

[0054] S22: Convert continuous monitoring data into follow-up period-level features;

[0055] For continuous monitoring data such as home blood pressure, continuous heart rate, activity level and sleep duration, valid samples are screened according to the actual collection time, and the period mean, period fluctuation value, minimum value, maximum value, last valid sample value and monitoring coverage are calculated. The preset sampling interval is a configuration value greater than zero. The number of samples to be sampled is the period duration divided by the preset sampling interval and rounded up to at least one. The monitoring coverage is the ratio of the number of valid samples to the number of samples to be sampled. When the calculation result is greater than 1, it is taken as 1.

[0056] If no valid samples are available within a period, the monitoring coverage will be set to zero and a period data missing flag will be set.

[0057] S23: Determine the time validity factor of the data;

[0058] In the standard field dictionary, configure the corresponding validity period and stable validity ratio for dynamic indicators, test data, inspection results and screening results that require timeliness judgment.

[0059] The first Follow-up time points and multimodal follow-up data The difference between the actual collection times is denoted as the data time difference. Multimodal follow-up data The corresponding standard fields are standard fields. Standard fields The corresponding validity period is The stable and effective ratio is Then multimodal follow-up data In the The effective time coefficient for each follow-up node is:

[0060] (1)

[0061] in, Represents multimodal follow-up data In the The effective time coefficient of each follow-up node Indicates the time difference of the data. Represents standard fields The corresponding validity period, Represents standard fields The corresponding stable and effective ratio, and The validity period is greater than or equal to 0 and less than 1. The validity period is a configuration value greater than zero, and the data time difference and the validity period use the same time unit.

[0062] When the data is within a stable validity period, the time validity coefficient is 1. After entering the near-expiration stage, it continuously decreases and becomes 0 when the validity period is reached.

[0063] S24: Determine the overall effectiveness coefficient of the data;

[0064] The data availability coefficient is determined based on the standard data value, unit of measurement, actual collection time, and collection status: 1 is taken when the data is complete and in normal status or is an abnormal candidate, and 0 is taken when there are missing data, collection failure, or personnel attribution pending verification.

[0065] Based on the data availability coefficient, time validity coefficient, and monitoring coverage of continuous monitoring data, determine the comprehensive validity coefficient of the data at the current follow-up node:

[0066] (2)

[0067] in, Represents multimodal follow-up data In the The overall effective coefficient of each follow-up node Represents multimodal follow-up data Available data coefficients Represents multimodal follow-up data Time efficiency coefficient, This represents the coverage correction factor.

[0068] For the follow-up period-level features generated in step S22, the data availability coefficient is set to 1 when there is valid sampling within the period. The time of the last valid sampling within the period is taken as the actual collection time of the period-level feature, and the time validity coefficient is determined according to formula (1). The monitoring coverage of the period is used as the coverage correction coefficient, and the comprehensive validity coefficient is determined according to formula (2). When there is no valid sampling within the period, both the data availability coefficient and the comprehensive validity coefficient are set to 0. The coverage correction coefficient for other data is set to 1. Data with a comprehensive validity coefficient of zero do not participate in the generation of the current risk feature. Data with a comprehensive validity coefficient greater than 0 carry this coefficient into subsequent steps and do not directly change the original indicator value and its unit of measurement.

[0069] S25: Generate a valid longitudinal follow-up sequence;

[0070] For dynamic indicators such as systolic blood pressure, diastolic blood pressure, heart rate, weight, fasting blood glucose, glycated hemoglobin, and blood lipids, the data with the closest actual collection time to the current follow-up node and with a comprehensive effective coefficient greater than 0 were selected as the node representative data.

[0071] The continuous monitoring data uses the periodic features generated in step S22 as the node representative data; the basic risk data uses the most recently confirmed record.

[0072] For follow-up texts, electrocardiogram examination results, carotid ultrasound examination results, stroke screening results, and high-risk survey results, all event records with a comprehensive validity coefficient greater than 0 within the corresponding follow-up period are retained, and no single record is used to replace other valid records.

[0073] When a certain standard field does not have data with a comprehensive effective coefficient greater than 0 in the current follow-up node, a data missing flag is set for that standard field.

[0074] The node representative data, follow-up period-level features, comprehensive validity coefficient, data missing identifier, actual follow-up interval, and valid event records corresponding to each follow-up node are combined in chronological order to obtain the valid longitudinal follow-up sequence:

[0075] (3)

[0076] in, This represents the valid longitudinal follow-up sequence up to the current follow-up node. Indicates the current follow-up node number. Indicates the first The time corresponding to each follow-up node Indicates the first Each follow-up node corresponds to a data vector composed of node representative data and follow-up period-level features. This represents a vector composed of the comprehensive effective coefficients corresponding to each data point. This represents a vector composed of data missing identifiers corresponding to each standard field. Indicates the first The actual follow-up interval corresponding to each follow-up cycle. Indicates the first The collection of follow-up texts, examination conclusions, and stroke screening event records retained within each follow-up period.

[0077] The node-level comprehensive effective coefficient vector consists of the comprehensive effective coefficients of the data represented by the selected corresponding node and the periodic-level features. Each event record in the event record set also carries its comprehensive effective coefficient, actual collection time, data source, and original record index.

[0078] Step S2 outputs the valid longitudinal follow-up sequence. This sequence provides the current indicators, changes, duration of abnormalities, number of consecutive abnormalities, and medical conceptual evidence for step S3; it preserves the actual temporal relationship of irregular follow-ups and distinguishes between near-expiration, expired, and under-monitoring coverage data.

[0079] Step S3: Generation of longitudinal risk features. Step S3 uses the effective longitudinal follow-up sequences output in Step S2. As input, the system generates current indicators, time interval perceived change characteristics, reference range deviation, abnormal duration, number of consecutive abnormalities, and medical conceptual evidence to form longitudinal risk features for use in step S4.

[0080] S3 specifically includes the following steps:

[0081] S31: Determine the current indicator value and the previous valid indicator value;

[0082] Extract basic risk data such as age, gender, family history, medical history, smoking status, and medication adherence from the current node data, as well as the current values ​​of dynamic indicators such as blood pressure, heart rate, weight, blood sugar, and blood lipids.

[0083] Let the standard field number corresponding to the dynamic indicator be... ,but Indicates the first Standard fields in each follow-up node The corresponding current indicator value, This represents the overall effective coefficient corresponding to the current indicator value. This indicates a missing data flag corresponding to the current indicator value.

[0084] From the Each follow-up node begins a forward search to determine the nearest historical follow-up node that has a comprehensive validity coefficient greater than 0 and a data missing flag of 0. This historical follow-up node is then recorded as its sequence number. .

[0085] If a previous valid follow-up node exists, the current value and the previous valid value are retained; if no previous valid follow-up node exists, only the current value is retained and the change characteristics are marked as unavailable. Continuous monitoring data uses the periodic mean, fluctuation value, extreme value, last valid value and monitoring coverage as the basic data for change.

[0086] S32: Generate time interval-aware indicator change characteristics;

[0087] When both the current indicator value and the previous valid indicator value are available, Indicates the change in the two adjacent follow-up periods, in This represents the change per unit time calculated based on the actual follow-up interval. If any indicator value is missing, the corresponding change feature will be marked as unavailable. The change in adjacent follow-up is the current indicator value minus the previous valid indicator value, and the change per unit time is the change in adjacent follow-up divided by the actual time interval between the current node and the previous valid node that is greater than 0.

[0088] The time unit for the change per unit time is set to day by default according to the system. If the actual data collection time uses other time units, they will be converted to days before calculation.

[0089] S33: Deviation characteristics of the reference range of the generated indicator;

[0090] In the standard field dictionary, configure the lower reference limit, upper reference limit, and reference scale for dynamic indicators with medical reference ranges;

[0091] Set standard fields The lower reference limit is The upper limit of the reference is The reference scale is Then the standard field The deviation characteristics from the reference range at the current follow-up node are as follows:

[0092] (4)

[0093] in, Represents standard fields Deviation from the reference range at the current follow-up node; and These represent the lower reference limit and the upper reference limit, respectively. Indicates the reference scale, and The preset value is greater than zero, and its unit of measurement is consistent with the standard field. The corresponding standard units of measurement are consistent.

[0094] The reference range deviation characteristic is a dimensionless value. When the current indicator value is higher than the upper limit of the reference range, the reference range deviation characteristic is positive; when the current indicator value is lower than the lower limit of the reference range, the reference range deviation characteristic is negative; when the current indicator value is within the reference range, the reference range deviation characteristic is 0.

[0095] The reference scale can be determined based on the width of the medical reference range of the corresponding indicator, the robustness of dispersion in the training data, or the preset clinical change scale, and should be kept consistent during model training and actual risk assessment.

[0096] For standard fields with only a one-sided risk threshold, deviation calculations are performed using only the corresponding side. For basic risk data and event data that do not have a numerical medical reference range, no reference range deviation features are generated.

[0097] Based on the reference range deviation characteristics and the corresponding comprehensive effective coefficient, the abnormal status of the standard field at each follow-up node is determined. When the reference range deviation characteristics are not equal to 0 and the comprehensive effective coefficient is greater than 0, the corresponding standard field is marked as a valid abnormal status. When the reference range deviation characteristics are equal to 0 and the comprehensive effective coefficient is greater than 0, it is marked as a valid normal status. When the comprehensive effective coefficient is zero, the abnormal status is not determined based on the data.

[0098] S34: Generate persistent anomaly features;

[0099] For dynamic indicators of systolic blood pressure, diastolic blood pressure, heart rate, weight, blood glucose and blood lipids, the effective duration of the current abnormal process is calculated based on the abnormal status at each follow-up node, the actual follow-up interval and the comprehensive effective coefficient.

[0100] The search starts backward from the current follow-up node; if the most recent valid normal state exists, that node is used as the starting point; otherwise, the earliest valid abnormal node is used as the starting point. A standard field is formed by composing the sequence numbers of adjacent time intervals from the starting point to the current node, where both adjacent endpoints have a composite validity coefficient greater than zero. The corresponding current abnormal range index set ;

[0101] Standard fields The duration of the abnormality at the current follow-up node is:

[0102] (5)

[0103] in, Represents standard fields The duration of the anomaly up to the current follow-up point; Represents standard fields The corresponding set of current abnormal interval indices; This indicates the first step determined in step S2. The actual follow-up interval corresponding to each follow-up cycle; This indicates that step S2 is a standard field. The validity period of the configuration; and Representing standard fields The combined effective coefficient between two adjacent follow-up nodes; and These represent the abnormal states of two adjacent follow-up nodes, respectively.

[0104] The abnormal status value is either 0 or 1. A valid abnormal status corresponds to an abnormal status value of 1, while a valid normal status corresponds to an abnormal status value of 0.

[0105] Equation (5) corrects the corresponding time interval by the average of the product of the combined effective coefficients of two adjacent nodes and the abnormal state; the contribution is greater when both adjacent nodes are abnormal states with high effective coefficients, and the contribution decreases accordingly when the effective coefficient of either node decreases.

[0106] Validity period Limit the maximum contribution of a single follow-up period to the duration of the abnormality, and avoid presuming the entire interval as a persistent abnormality due to excessively long intervals between two follow-up visits.

[0107] When the current follow-up node is in a valid normal state, the anomaly duration is set to 0. When the current follow-up node lacks valid data, the existing anomaly duration is not extended. When it is an empty set, Take 0.

[0108] S35: Generate a feature of consecutive anomaly counts;

[0109] by Represents standard fields As of the The number of consecutive valid abnormalities that occur at a follow-up node after its most recent valid normal state; the current valid normal state is reset to zero, and missing states are not counted as normal or abnormal.

[0110] When the standard field of the current follow-up node is in a valid normal state, the consecutive anomaly count is reset to zero. When the current follow-up node lacks valid data, the missing state is not counted as a normal state, nor is the consecutive anomaly count increased. Instead, the consecutive anomaly count corresponding to the most recent valid follow-up node is maintained, and the consecutive anomaly status is indicated by the data missing identifier of the current standard field, which is pending confirmation.

[0111] The duration of an anomaly characterizes the extent to which an anomaly persists over real time, while the number of consecutive anomalies characterizes the degree to which an anomaly recurs at different follow-up points. Both are retained to avoid treating multiple anomalies measured within a short period as the same as persistent anomalies over a longer period.

[0112] S36: Extract risk concepts from follow-up texts and examination results;

[0113] Obtain follow-up texts, medical histories, examination results, stroke screenings and high-risk investigation events from valid longitudinal follow-up sequences, and form unified medical concept evidence in the following manner;

[0114] For the follow-up text, medical named entity recognition, medical concept mapping, negation word recognition, degree word recognition and time expression recognition were used in sequence to extract symptoms such as dizziness, limb numbness, limb weakness, speech abnormality, facial drooping and visual abnormality, and to determine the positive state, negative state, degree of abnormality and actual occurrence time of each symptom.

[0115] For medical history information, extract the history of hypertension, diabetes, dyslipidemia, atrial fibrillation, coronary heart disease, stroke, and transient ischemic attack.

[0116] For the examination results, abnormal items such as atrial fibrillation, arrhythmia, carotid plaque and carotid stenosis are extracted, and the degree of abnormality is determined according to the structured abnormality level, degree words in the report or the preset examination result mapping table.

[0117] The extracted symptoms, medical history, and abnormal items are mapped to the unified medical concept established in step S1. When both positive and negative records exist for the same medical concept, positive and negative evidence are retained separately, without mutual cancellation in step S3.

[0118] Let the medical concept number be... The current follow-up point will be aligned with medical concepts. Events that are relevant and have a comprehensive validity coefficient greater than 0 are classified as the set of affirmative events. and the set of negative events And generate medical conceptual evidence features:

[0119] (6)

[0120] in, Indicates medical concept Characteristics of medical conceptual evidence at the current follow-up point; Indicates event logging; Represents event log The corresponding overall efficiency coefficient; Represents event log The corresponding event evidence strength ranges from 0 to 1. For affirmative events, the event evidence strength is determined by the degree of anomalousness or the preset mapping strength; for negative events, it is determined by the negative confirmation strength, which is 1 if no level is set.

[0121] The affirmative evidence component of a medical concept is denoted as: The weight of the negative evidence is denoted as The value is zero when the corresponding event set is empty.

[0122] Confirmed and uncorrected structured medical histories are considered positive evidence, while valid records clearly indicating the absence of the corresponding disease or symptoms are considered negative evidence; both types of evidence are retained for step S4 to identify source conflicts.

[0123] By retaining the strength of affirmative evidence and the strength of negative evidence separately, subsequent steps can identify potential inconsistencies between records from different sources, avoiding the loss of risky evidence information due to simple addition or mutual cancellation.

[0124] S37: Generate longitudinal risk characteristics;

[0125] The longitudinal risk characteristics are composed of current indicators, changes in adjacent follow-ups, changes per unit time, deviation from reference range, duration of abnormality, number of consecutive abnormalities, medical conceptual evidence, overall validity coefficient, data missing markers, and actual follow-up intervals. When the previous node or current data is missing, retain the corresponding unavailable identifier or data missing identifier, and output the feature to step S4.

[0126] For dynamic indicators that do not have a previous valid follow-up node, the corresponding unavailable change feature is retained in the longitudinal risk characteristics. For currently missing standard fields, their actual measured values ​​are not represented by zero values, but are distinguished by data missing indicators.

[0127] Through the above processing, the absolute level of the indicator, the change in adjacent follow-ups, the change per unit time, the duration of the abnormality, the number of consecutive abnormalities, and the medical conceptual evidence in the text and examination results are all included in the risk characteristics to distinguish between single abnormalities, short-term rapid deterioration, and long-term persistent abnormalities, and retain the data validity status and data missing status for subsequent risk assessment.

[0128] Step S4: Comprehensive cardiovascular and cerebrovascular risk assessment and stroke-specific risk assessment; Step S4 uses the longitudinal risk characteristics output from Step S3. As input, the participation level of various features is determined based on data completeness, timeliness, and validity. A comprehensive risk status value is first generated, and then corrected using stroke-specific risk evidence.

[0129] S4 specifically includes the following steps:

[0130] S41: Divide risk characteristic groups and transmit data status

[0131] First, distinguish between shared features and stroke-specific exclusive features according to the model participation path; then divide the shared features into structured longitudinal indicators, continuous monitoring and medical concept evidence feature groups according to the data formation method, so as to form a comprehensive risk status value for equations (9) and (10).

[0132] The structured longitudinal indicator feature group includes age, gender, family history, medical history, smoking status, medication adherence, actual follow-up interval, as well as current value, adjacent and unit time change, reference range deviation, abnormal duration and number of consecutive abnormalities formed by non-continuous monitoring data.

[0133] The continuous monitoring feature group includes the periodic mean, fluctuation, minimum, maximum, last valid value and coverage of home blood pressure, continuous heart rate, activity level and sleep duration, as well as the current value, adjacent and unit time changes, reference range deviation, abnormal duration and number of consecutive abnormalities formed by these values.

[0134] The medical concept evidence feature group includes medical concept evidence features corresponding to follow-up texts, electrocardiogram examination results, carotid ultrasound examination results, stroke initial screening results, and high-risk survey results.

[0135] Each risk feature is associated with source data, feature validity coefficient, feature missing identifier, and unique model participation path. Shared features are only assigned to one risk feature group; stroke-specific exclusive features are not included in the risk feature groups used in equations (9) and (10), but are only retained in the stroke-specific risk feature vector for use in equation (11).

[0136] The feature validity coefficient and feature missing identifier corresponding to the current indicator value are respectively the comprehensive validity coefficient and data missing identifier output in step S2.

[0137] The characteristic effectiveness coefficients for adjacent follow-up changes and unit time changes are taken as the smaller value between the comprehensive effectiveness coefficients corresponding to the current indicator value and the previous effective indicator value. When any indicator value is missing, the corresponding change characteristic is set to an unavailable state.

[0138] The effective coefficients for the anomaly duration and the number of consecutive anomalies are determined based on the combined effective coefficients of the follow-up nodes involved in their generation and the actual contribution of each node to the feature. When the current standard field is missing, the feature missing flag for the number of consecutive anomalies is set to 1 to retain the pending confirmation status.

[0139] The feature validity coefficient of a medical concept evidence feature is determined based on the comprehensive validity coefficient corresponding to the event record that generated the medical concept evidence feature. When there is neither affirmative nor negative evidence for the same medical concept, the feature missing flag corresponding to the medical concept is set to 1.

[0140] Through the above data status transmission, the timeliness of the data determined in step S2 and the usability of the features determined in step S3 can continue to apply to step S4, without being lost due to the risk feature conversion process.

[0141] S42: The medical concept of generating conflict suppression;

[0142] Step S3 retains the strength of affirmative evidence and the strength of negative evidence for each medical concept. To avoid directly adding the affirmative and negative evidence for the same medical concept in the risk assessment, the strength of evidence conflict for each medical concept is first calculated, and the directional features of the medical concept are generated accordingly.

[0143] (7)

[0144] in, Indicates medical concept The strength of conflict of medical conceptual evidence at the current follow-up point; This represents the medical concept generated in step S3. Corresponding strength of affirmative evidence; Indicates medical concept The corresponding strength of refutation evidence; Indicates medical concept Corresponding medical concept and directional characteristics; Indicates the sequence number of a medical concept.

[0145] The values ​​for the strength of conflict of evidence, strength of affirmative evidence, and strength of negative evidence for medical concepts all range from 0 to 1, while the values ​​for the directional features of medical concepts range from -1 to 1.

[0146] When a medical concept has only positive evidence, its directional characteristic is positive. When a medical concept has only negative evidence, its directional characteristic is negative. When both positive and negative evidence are strong, the intensity of the evidence conflict for the medical concept increases, and the absolute value of its directional characteristic decreases accordingly.

[0147] This step does not directly delete conflicting event records, but retains the strength of affirmative evidence, the strength of negative evidence, and the strength of conflict of medical conceptual evidence, for step S5 to determine whether the risk outcome is supported by sufficient and consistent data.

[0148] S43: Calculate the data quality coefficient for the risk characteristic group;

[0149] Let the risk characteristic group number be Risk Characteristic Group The corresponding feature index set is The number of features in the feature index set is .

[0150] Based on the feature missing identifiers and feature validity coefficients of each feature within the risk feature group, the feature completeness, average validity, and data quality coefficients are calculated respectively:

[0151] (8)

[0152] in, Indicates risk characteristic group Feature completeness at the current follow-up node; Indicates risk characteristic group Average validity at the current follow-up node; Indicates risk characteristic group Data quality coefficient at the current follow-up node; Indicates risk characteristics Corresponding feature missing identifier; Indicates risk characteristics The corresponding feature efficiency coefficient; Indicates the risk characteristic number; Indicates risk characteristic group The corresponding integrity adjustment coefficient, and Greater than or equal to 0 and less than or equal to 1.

[0153] The values ​​for feature completeness, average validity, and data quality coefficient all range from zero to one.

[0154] The comprehensive validity coefficient output in step S2 already includes the data acquisition status, time validity coefficient and continuous monitoring coverage. Therefore, the average validity in equation (8) can continue the judgment of data timeliness and validity made in step S2.

[0155] When all features within a risk feature group are missing, the data quality coefficient of that risk feature group is zero. Risk feature groups with a data quality coefficient of zero are not included in the current risk calculation. When the completeness adjustment coefficient is one, the data quality coefficient is the feature completeness; when the completeness adjustment coefficient is zero, the data quality coefficient is the average validity; in other cases, it is calculated according to equation (8).

[0156] The completeness adjustment coefficient is determined by the training and validation samples and remains fixed after model deployment. For datasets where missing data are common but still have auxiliary value, the completeness adjustment coefficient is decreased. For datasets requiring support from multiple fields, the completeness adjustment coefficient is increased.

[0157] S44: Generate risk evidence values ​​for risk feature groups;

[0158] Scale transformation or encoding is performed on continuous, enumerated, and medical concept-oriented features. Missing features are assigned neutral placeholder values ​​and a feature missingness indicator is input simultaneously. The risk feature group encoder outputs a risk evidence value. .

[0159] The risk feature group encoder consists of at least one fully connected layer and a nonlinear activation layer, and its output dimension is predetermined during the model training phase.

[0160] The risk evidence value ranges from -1 to 1. A positive risk evidence value indicates that the corresponding risk feature group supports an increase in risk. A negative risk evidence value indicates that the corresponding risk feature group supports a decrease in risk. A risk evidence value close to zero indicates that the corresponding risk feature group does not provide clear-cut risk evidence under the current data conditions.

[0161] S45: Fusion of evidence regarding the risks of enforcing data quality constraints;

[0162] Based on the data quality coefficient and risk evidence value of each risk feature group, the fusion coefficient for each risk feature group in the comprehensive cardiovascular and cerebrovascular risk calculation is calculated:

[0163] (9)

[0164] in, Indicates risk characteristic group The fusion coefficient at the current follow-up node; Indicates risk characteristic group Data quality coefficient; Indicates risk characteristic group Risk evidence value; This represents the fusion smoothing parameter, and The default value is greater than zero; Indicates the number of risk characteristic groups; The index represents the risk characteristic group used in the summation. The exponential term in equation (9) uses the natural exponential function, and the vertical line indicates taking the absolute value.

[0165] When the data quality coefficient of at least one risk feature group is greater than zero, the sum of the fusion coefficients of all risk feature groups is 1.

[0166] Equation (9) uses the absolute value of risk evidence to characterize the degree of correlation between the corresponding risk feature group and the current risk status of the target personnel, so that both obvious evidence of increased risk and obvious evidence of decreased risk can participate in the fusion. At the same time, the data quality coefficient independently constrains the participation degree of each risk feature group, avoiding the situation where low-completeness or low-effectiveness data obtains an excessively high fusion coefficient simply because the model output amplitude is large.

[0167] When the data quality coefficients of all risk feature groups are zero, skip the conventional risk calculations from equation (9) to equation (12) and do not generate external risk results; set the risk calculation status to the data insufficiency status and input all missing or invalid core data into step S5.

[0168] S46: Generate comprehensive cardiovascular and cerebrovascular risk results and stroke-specific risk results;

[0169] Based on the risk evidence value and fusion coefficient of each risk characteristic group, a comprehensive risk status value for the target individual at the current follow-up node is generated, and the comprehensive risk status value is converted into a comprehensive cardiovascular and cerebrovascular risk result:

[0170] (10)

[0171] in, This indicates the overall risk status value of the target individual at the current follow-up point; This indicates the overall cardiovascular and cerebrovascular risk outcome of the target individuals at the current follow-up point; This indicates a bias in the mapping of comprehensive cardiovascular and cerebrovascular risks; This represents the comprehensive risk mapping coefficient for cardiovascular and cerebrovascular diseases, and Positive numbers; superscript This indicates a comprehensive risk assessment task related to cardiovascular and cerebrovascular health. This represents a logical function that maps input values ​​to a range between zero and one.

[0172] The overall risk status value is a dimensionless numerical value, and the range of the overall cardiovascular and cerebrovascular risk result is 0 to 1.

[0173] From the perspective of vertical risk characteristics Stroke-specific risk features are extracted and scaled to form a stroke-specific risk feature vector. .

[0174] The stroke-specific risk characteristics include at least a history of atrial fibrillation, a history of stroke, a history of transient ischemic attack, evidence of atrial fibrillation, carotid artery plaque, carotid artery stenosis, limb numbness, limb weakness, speech abnormalities, facial drooping, visual abnormalities, initial stroke screening results and high-risk survey results, and also include the medical concept directional characteristics, characteristic validity coefficients and characteristic missing markers corresponding to the above risk factors.

[0175] The data quality coefficient for stroke-specific risk characteristics is determined according to formula (8). The stroke-specific risk feature vector, after scaling, is input into the encoder. This is then processed within the comprehensive risk state value. Based on this, stroke-specific risk evidence is introduced for correction, generating stroke-specific risk results:

[0176] (11)

[0177] in, This indicates the stroke-specific risk evidence value corresponding to the current follow-up node; Represents the stroke-specific risk evidence mapping vector; This represents a stroke-specific risk characteristic encoder; This represents the stroke-specific risk feature vector after scaling transformation. This indicates a bias in the mapping of evidence regarding the specific risks of stroke. This indicates the stroke-specific risk outcome corresponding to the current follow-up node; This indicates a bias in the risk mapping for stroke. This represents the overall risk status transmission coefficient; This indicates the correction factor for specific risk evidence related to stroke. Data quality coefficient representing specific risk characteristics of stroke; superscript This represents a stroke-specific risk task. The formula uses the hyperbolic tangent function, and the superscript transpose indicates vector transpose.

[0178] Both the overall risk status transmission coefficient and the stroke-specific risk evidence correction coefficient are positive.

[0179] The stroke-specific risk result in Formula (11) is based on the comprehensive risk status value, and then specifically modified according to the history of stroke, history of transient ischemic attack, atrial fibrillation, carotid artery abnormalities and related symptoms of the nervous system.

[0180] When the data quality coefficient of stroke-specific risk features is low, their risk evidence correction effect is correspondingly reduced. Features participating in the risk feature group coding and equations (9) and (10) are denoted as shared features; features only entering equation (11) are denoted as stroke-specific exclusive features. The data quality coefficient of stroke-specific risk features, feature missing identifiers of stroke-specific risk features, medical concept evidence conflict strength, and model participation path are continued to be output to step S5.

[0181] The comprehensive cardiovascular and cerebrovascular risk model and the stroke-specific risk model were trained using historical follow-up samples. First, the risk feature group encoder, risk evidence mapping parameters, and comprehensive cardiovascular and cerebrovascular risk mapping parameters were trained using the comprehensive cardiovascular and cerebrovascular risk labeled samples. Then, the trained comprehensive risk state values ​​were used as shared input to train the stroke-specific risk feature encoder and specific correction parameters using the stroke-specific risk labeled samples.

[0182] S47: Determine the risk level and key risk factors;

[0183] A first risk threshold and a second risk threshold are set for the comprehensive cardiovascular and cerebrovascular risk results and the stroke-specific risk results, respectively.

[0184] When the risk outcome is below the corresponding first risk threshold, the risk level is determined to be low risk. When the risk outcome is greater than or equal to the first risk threshold but lower than the second risk threshold, the risk level is determined to be medium risk. When the risk outcome is greater than or equal to the second risk threshold, the risk level is determined to be high risk.

[0185] Two risk tasks are configured with a first risk threshold and a second risk threshold, both ranging from zero to one, with the first threshold being strictly less than the second threshold. The thresholds are determined based on the risk outcome distribution, recognition sensitivity, and specificity of the training and validation samples, and are fixed after deployment.

[0186] To identify major risk factors, current indicator values, changes in adjacent follow-ups, changes per unit time, duration of abnormalities, number of consecutive abnormalities, and medical conceptual evidence belonging to the same medical risk factor are combined into a single risk factor unit.

[0187] Replace each risk factor unit with its corresponding reference neutral state, keeping other risk characteristics, data quality coefficients, and feature missing identifiers unchanged, and recalculate the risk results. Risk Factor Unit The risk contribution value for risk tasks is:

[0188] (12)

[0189] in, Representing risk factor units Risk tasks Risk contribution value; This indicates the risk result calculated using complete longitudinal risk characteristics; This indicates the risk factor unit. The risk result was recalculated after replacing it with a reference neutral state; Indicates the unit number of the risk factor; Pick The time indicates the comprehensive risk task of cardiovascular and cerebrovascular diseases, take The time indicates a stroke-specific risk task; subscript This indicates a reference to a neutral state.

[0190] For numerical risk factors, the reference neutrality status is determined based on the medical reference range of the corresponding standard field. For medical history and abnormal examination results, the reference neutrality status is set to the absence of corresponding positive risk evidence. For symptom-related risk factors, the reference neutrality status is set to the absence of corresponding positive symptom evidence, but existing negative evidence is not deleted.

[0191] A positive risk contribution value indicates that the risk factor increases the corresponding risk outcome. A negative risk contribution value indicates that the risk factor decreases the corresponding risk outcome.

[0192] A preset number of risk factor units are selected as the main risk factors for the corresponding risk tasks in descending order of positive risk contribution value. Their risk contribution value, standard fields, actual collection time, data source, comprehensive effectiveness coefficient, and original record index are associated with them so that they can be traced back to the original records obtained in step S1.

[0193] Step S4 ultimately outputs the comprehensive cardiovascular and cerebrovascular risk results. and its risk level, stroke-specific risk results Its risk level, main risk factors and their risk contribution value.

[0194] Step S4 simultaneously outputs the data quality coefficient and fusion coefficient of each risk feature group, the feature validity coefficient, feature missing identifier, the risk feature group to which each risk feature belongs, the model participation path and traceability information, as well as the data quality coefficient and medical concept evidence conflict strength of the stroke-specific risk feature, for step S5 to judge the reliability of the risk results and identify key data items to be supplemented.

[0195] Through the above processing, the risk assessment process calculates the risk direction of risk characteristics separately from data quality, avoiding directly equating the model's level of attention with the reliability of the data. Simultaneously, by integrating the risk status value with the hierarchical calculation relationship of stroke-specific risk evidence, the overall cardiovascular and cerebrovascular risk results and stroke-specific risk results share the overall risk basis of the target population, while retaining the corrective effect of stroke-specific risk factors on stroke-specific risk.

[0196] Step S5: Reliability assessment of risk outcomes and identification of key data items to be supplemented;

[0197] Step S5 receives the comprehensive cardiovascular and cerebrovascular risk results and their risk levels, stroke-specific risk results and their risk levels, main risk factors, risk contribution values, data quality coefficients, fusion coefficients, feature missing identifiers, medical concept evidence conflict strength, feature validity coefficients, model participation paths, and source information output from step S4. If data is insufficient, steps S51 to S56 are skipped, and the risk support, risk level stability, and risk result reliability of the two risk tasks are set to zero, with the overall risk result reliability status set to unreliable. All missing or invalid core inputs are treated as key supplementary data items, with a risk impact value of one, and are sent to step S6, without outputting either type of risk result. Core inputs are standard fields or medical concepts that are required by the model, and their absence or invalidity would prevent the corresponding risk feature group from forming a valid input. Shared features are associated with the two risk tasks, while stroke-specific exclusive features are only associated with the stroke-specific risk task. Source data entering both model participation paths are associated with both risk tasks. The risk result reliability of the affected risk task is set to zero, and the reliability gap is set to one. Otherwise, steps S51 to S56 are executed.

[0198] S5 specifically includes the following steps:

[0199] S51: Establish supporting relationships for risk outcomes and a set of candidate data items

[0200] Establish support relationships between the main risk factors and source risk characteristics of the comprehensive cardiovascular and cerebrovascular risk task and the stroke-specific risk task, respectively.

[0201] Let the risk task identifier be .when Pick When, it indicates a comprehensive cardiovascular and cerebrovascular risk task; when Pick At that time, it indicated a special risk task for stroke.

[0202] For each major risk factor, associate its current indicators, changes, deviations from reference ranges, duration of anomalies, number of consecutive anomalies, and medical conceptual evidence, and obtain the corresponding feature validity coefficients, feature missing indicators, data quality coefficients of the risk feature group to which it belongs, and the intensity of conflict of medical conceptual evidence.

[0203] Candidate data items include missing, expired, data exceeding the validity period, data below the preset lower limit of the validity coefficient, data with conflicting positive and negative evidence, data with only abnormal candidate records, or data that supports the main risk factors but currently has no valid records. Each candidate data item is associated with a standard field or medical concept, current status, most recent collection time, source, original record index, and affected risk task. The preset lower limit of the validity coefficient is greater than 0 and less than or equal to 1.

[0204] S52: Assess the extent of effective data support for the main risk factors;

[0205] Set risk tasks The corresponding set of main risk factors is as follows Key risk factors The corresponding set of supporting risk feature indexes is The set of supporting risk feature indices is a non-empty set, and its number of elements is greater than or equal to 1.

[0206] Based on the missing data identifiers, feature validity coefficients, data quality coefficients of the respective risk feature groups, and the strength of conflict of medical conceptual evidence for each supporting risk feature, the effective support of the main risk factors is calculated, and the risk support of the corresponding risk tasks is further calculated:

[0207] (13)

[0208] in, Indicates key risk factors Effective support at the current follow-up node; Indicates the main risk factors The strength of conflict of relevant medical conceptual evidence is zero when no medical conceptual evidence is involved. This indicates support for the main risk factors. A set of risk characteristic indexes Indicates the number of elements in the set; Indicates the risk characteristic number; and These represent the risk feature missing identifier and the risk feature validity coefficient determined in step S4, respectively. Indicates risk characteristics The risk characteristic group to which it belongs; This represents the data quality coefficient for the corresponding risk characteristic group. When a major risk factor is associated with multiple medical concepts, the strength of the conflict of evidence for those medical concepts is taken as the maximum value of the conflict of evidence for each related medical concept.

[0209] Indicates risk task Risk support level at the current follow-up node; This indicates the main risk factors calculated in step S4. Risk tasks Risk contribution value; This indicates a preset positive number to prevent the denominator from being zero.

[0210] The effective support and risk support of the main risk factors both range from 0 to 1.

[0211] When the supporting data for a major risk factor is complete, within its validity period, and does not present significant conflict of evidence, its effective support level is close to 1. When the major risk factor is supported only by outdated data, low-coverage continuous monitoring data, or conflicting event records, its effective support level decreases.

[0212] Equation (13) uses the risk contribution value to weight the effective support of different major risk factors, so that the major risk factors that have a greater impact on the current risk outcome have a higher proportion in the risk support calculation, and avoids treating low-impact auxiliary data as the same as the key data that determines the risk level.

[0213] When the corresponding risk task does not output the main risk factors with a positive risk contribution value greater than the preset lower limit of the positive risk contribution value, the conventional calculation of risk support in formula (13) is not performed. Instead, the fusion coefficient and data quality coefficient output in step S4 are used to calculate the risk support, and the fusion coefficient of each risk feature group is used as the weight of the corresponding data quality coefficient.

[0214] S53: Construct candidate data to complete scenarios and determine risk outcome ranges;

[0215] For candidate data item set For each candidate data item, low-risk completion scenarios and high-risk completion scenarios are established respectively.

[0216] For missing numerical dynamic indicators, low-risk and high-risk completion values ​​are determined based on the most recent valid indicator value of the target person, the medical reference range of the corresponding standard field, and the reasonable range of variation in the training sample.

[0217] For numerical dynamic indicators that have expired or have a low overall effectiveness coefficient, their original values ​​are retained, and two completion scenarios are generated using stable and abnormal continuation states, respectively.

[0218] For discrete data such as medical history, symptoms, abnormal examination results, and stroke screening results, the state without corresponding positive risk evidence is identified as a low-risk completion scenario, and the state with corresponding positive risk evidence and assigned a value according to a preset degree level is identified as a high-risk completion scenario.

[0219] For medical concepts where there is a conflict between positive and negative evidence, positive evidence and negative evidence are respectively used as two complementary scenarios.

[0220] For standard fields with anomaly candidate identifiers but lacking verification data, the two completion scenarios are: the anomaly record failing verification and the record succeeding after verification.

[0221] When constructing candidate data completion scenarios, only the risk characteristics and data status corresponding to the current candidate data item are replaced, while other longitudinal risk characteristics, comprehensive effectiveness coefficients, and data missing indicators remain unchanged.

[0222] Input the low-risk and high-risk completion scenarios into the risk assessment model that has been trained in step S4, respectively, to obtain candidate data items. The two corresponding risk recalculation results are denoted as follows: and .

[0223] Based on the current risk results and the recalculated risk results of each candidate data item, the risk tasks are determined. Corresponding risk outcome range:

[0224] (14)

[0225] in, Indicates risk task The lower bound of risk outcomes under the current data conditions; This indicates the upper bound of the corresponding risk outcome; This indicates the current risk result output in step S4; Indicates candidate data items; Represents a set of candidate data items; Representing candidate data items Recalculation of risk results when using low-risk completion scenarios; Representing candidate data items Recalculate the risk results when using high-risk completion scenarios.

[0226] The lower bound of the risk outcome, the upper bound of the risk outcome, and the current risk outcome all range from 0 to 1.

[0227] When the candidate data item set is empty, both the lower bound and the upper bound of the risk result are set to the current risk result output in step S4.

[0228] Equation (14) uses a single candidate data completion method to determine the risk outcome range, without assuming that all missing data are in the highest or lowest risk state at the same time, thereby avoiding the formation of an overly wide risk range that is detached from the actual data conditions.

[0229] S54: Determine the stability of the risk level and calculate the reliability of the risk outcome;

[0230] Obtain step S4 for risky tasks A first risk threshold and a second risk threshold are set, and the two risk thresholds are combined to form a risk level boundary set. ;

[0231] The stability of the risk level is calculated based on the width of the risk outcome interval and the number of risk level boundaries that the risk outcome interval crosses:

[0232] (15)

[0233] in, Indicates risk task Corresponding risk level stability; Indicates the risk level boundary; Indicates risk task The corresponding risk level boundary set; This indicates a conditional indicator function that returns 1 if the condition within the parentheses is true, and zero otherwise.

[0234] The value of the risk level stability ranges from 0 to 1.

[0235] When the risk outcome range is narrow and does not cross any risk level boundaries, the risk level stability is high. When the risk outcome range crosses one risk level boundary, it indicates that at least one candidate data item may cause the current risk level to change to an adjacent level. When the risk outcome range crosses two risk level boundaries simultaneously, it indicates that the current data state is insufficient to stably determine the risk level.

[0236] By combining risk support and risk level stability, we obtain the reliability of risk outcomes corresponding to risk tasks:

[0237] (16)

[0238] in, Indicates risk task Reliability of risk outcomes at the current follow-up node; Indicates risk tolerance; Indicates the stability of the risk level; Indicates risk task The corresponding support adjustment coefficient, and It is greater than or equal to zero and less than or equal to one.

[0239] The reliability of the risk outcome ranges from 0 to 1. The support adjustment coefficient is determined by the training and validation samples and remains fixed after the model is deployed. When the support adjustment coefficient is one, the reliability of the risk outcome is the risk support; when the support adjustment coefficient is zero, the reliability of the risk outcome is the risk level stability; otherwise, it is calculated according to equation (16).

[0240] Equation (16) calculates whether key risk factors have valid data support and whether missing data may change the risk level separately, avoiding judging the reliability of the result solely based on whether the risk prediction probability is close to zero or one.

[0241] Even if the current risk outcome is high, the reliability of the corresponding risk outcome will still decrease if the main risk factors lack effective data support or if the risk level may decrease after candidate data is supplemented. Even if the current risk outcome is near the risk level boundary, the corresponding risk outcome can still have high reliability if there is sufficient effective data and supplementing candidate data will not change the risk level.

[0242] S55: Calculate the risk impact value of candidate data items;

[0243] For each candidate data item, calculate the difference in risk outcomes between its low-risk completion scenario and high-risk completion scenario, and determine whether the two completion scenarios correspond to different risk levels.

[0244] Calculate the risk impact value of the candidate data items based on the degree of data gap, differences in risk outcomes, and changes in risk levels:

[0245] (17)

[0246] in, Representing candidate data items Risk impact value at the current follow-up node; Representing candidate data items The extent of the data gap; Indicates risk task The corresponding risk difference adjustment coefficient, and Greater than or equal to zero and less than or equal to one; This represents the risk level mapping function that maps the risk outcome to low, medium, or high risk levels based on the risk threshold in step S4. The conditional indicator function follows the aforementioned definition. The data gap level is set to 0 to 1. The risk difference adjustment coefficient is determined by the training and validation samples and remains fixed after model deployment.

[0247] For data items that are completely missing or have become invalid, set the data gap level to one.

[0248] For data items with a comprehensive effective coefficient greater than zero but lower than the preset lower limit of the effective coefficient, the degree of data gap is determined by subtracting the ratio of the comprehensive effective coefficient to the preset lower limit of the effective coefficient from 1.

[0249] For medical concepts where there is a conflict between affirmative and negative evidence, the strength of the conflict of medical concept evidence output in step S4 is directly determined as the degree of data gap.

[0250] For data items with only a single candidate record of anomaly and lacking support from other sources, the ratio of the sum of the comprehensive effective coefficients of existing supporting records to a preset minimum total effective support quantity greater than zero is limited to no more than 1, and the degree of data gap is determined by subtracting this ratio from 1. When the same candidate data item simultaneously meets multiple of the aforementioned data gap rules, the degree of data gap for completely missing or invalid data items is taken as one; in other cases, the degree of data gap corresponding to each applicable rule is calculated separately, and the maximum value is taken as the unique degree of data gap for that candidate data item; when no applicable rule exists, it is taken as zero.

[0251] Equation (17) considers both the numerical change in risk outcome and the change in risk level that may be caused by the candidate data item. When two completion scenarios for a candidate data item can change the risk level, the candidate data item can still increase the risk impact value even if the difference in the numerical difference in risk outcome is relatively small.

[0252] For risk tasks unaffected by candidate data items, both completion scenarios use the current risk outcome, making the risk difference term and risk level change term zero; therefore, taking the maximum value in the two risk tasks is equivalent to taking the maximum risk impact value in the risk task affected by the project.

[0253] S56: Determine the overall risk outcome reliability status and key data items to be supplemented;

[0254] The first reliability threshold is lower than the second reliability threshold. An unreliable state is first determined when the reliability of the risk outcome of any risk task is lower than the first reliability threshold or the risk outcome range crosses two risk level boundaries. Otherwise, a conditionally reliable state is defined when the reliability of the risk outcome of any risk task is lower than the second reliability threshold or the risk outcome range crosses one risk level boundary. A reliable state is defined when the reliability of the risk outcomes of both risk tasks reaches the second reliability threshold and neither risk outcome range crosses a risk level boundary. If the states of two risk tasks are different, the state with the lower reliability is used as the overall risk outcome reliability state. The preset effective coefficient lower limit, preset positive risk contribution value lower limit, first reliability threshold, second reliability threshold, and preset risk impact value threshold are all determined based on the risk identification performance, risk level stability, and supplementary collection benefits in the training and validation samples, and remain fixed after model deployment. The preset effective coefficient lower limit, preset positive risk contribution value lower limit, and preset risk impact value threshold are all greater than zero and less than or equal to one; the first reliability threshold and the second reliability threshold are both between zero and one, and the first reliability threshold is lower than the second reliability threshold.

[0255] Candidate data items whose risk impact value reaches the preset risk impact value threshold or may change the adjacent risk level are identified as key data items to be supplemented, and sorted by risk impact value; each item retains the standard field or medical concept, current status, gap type, most recent collection time, main risk factors, affected risk tasks, two types of risk completion results, risk level change identifier, risk impact value, data source and original record index.

[0256] Based on the type of data gap, key data items to be supplemented are marked as candidate items for data collection, information verification, repeated testing, or further screening.

[0257] If the overall risk result reliability status is reliable, the current risk result is retained, the set of key data items to be supplemented and the set of minimum necessary supplementary data collection tasks are set to empty, and the process terminates according to step S74 without proceeding to step S6. If the overall risk result reliability status is conditionally reliable or unreliable, the risk result output in step S4 is retained, but not replaced with a supplementary scenario, and the risk support, risk result range, risk level stability, risk result reliability, overall risk result reliability status, and set of key data items to be supplemented are output to step S6. The data-insufficient branch does not output the risk result and risk result range, but only outputs the zeroed risk result reliability, the unreliable overall risk result reliability status, and the set of key data items to be supplemented to step S6.

[0258] Step S6: Generation of Minimum Necessary Data Acquisition Task

[0259] Step S6 takes the overall risk result reliability status and key data items to be supplemented as input, and combines the risk impact value, reliability gap, data acquisition cost, expected completion time and collection conditions to determine the minimum necessary supplementary data collection task set.

[0260] S6 specifically includes the following steps:

[0261] S61: Generate executable candidate tasks

[0262] Step S6 only handles situations where the conditions are reliable, unreliable, or insufficient; each time, it regenerates four types of candidate tasks and judges their executability based on the current key data items to be supplemented, without using the results from the previous round.

[0263] Candidate tasks are configured with terminals, subjects, devices, return formats, and estimated completion times. Path collection condition coefficients are calculated based on device, interface, subject, and personnel status. A coefficient of zero is assigned if necessary conditions are missing, and a coefficient of one is assigned if all conditions are met. The project collection condition coefficient is the maximum value of its candidate paths. Task units only enter the executable index set if their own path coefficient is greater than zero. Step S71 records the failure reason and path for use in step S74 when deciding to re-enter step S6. Failed paths with unchanged conditions are no longer entered into the executable set; alternative paths are selected if available, and manual review is required if all paths are unexecutable.

[0264] S62: Burden of Quantitative Data Acquisition

[0265] The project-level standardized data acquisition cost and estimated completion time are respectively taken as the original value divided by the positive reference upper limit of the corresponding task type and capped at one. The reference upper limit is fixed after deployment; both are used in Equation (18). The actual cost and estimated completion time of the candidate task unit are determined according to its configuration and are only used for parallel selection in Equation 19.

[0266] S63: Calculate the task priority of key data items to be supplemented.

[0267] by Indicates key data items that need to be supplemented. The gap is calculated by subtracting the maximum reliability of the risk outcome from the risk task, with a value ranging from zero to one. The lower the reliability of the risk outcome of the risk task, the larger the gap.

[0268] Calculate the task priority of key data items to be supplemented based on risk impact value, reliability gap, data acquisition burden, and collection conditions:

[0269] (18)

[0270] in, Indicates key data items that need to be supplemented. Task priority; Indicates the data collection condition coefficient; This represents the risk impact value output in step S5; This indicates a reliability gap; Indicates the cost of standardized data acquisition; Indicates the estimated completion time for standardization; , and These represent the adjustment coefficients corresponding to the reliability gap, data acquisition cost, and expected completion time, respectively, and all are non-negative preset values.

[0271] Task priority is set from zero to one, with higher priority given when the risk impact value is greater, the reliability of the risk result is lower, and the collection conditions are more sufficient, and lower priority given when the cost of standardized data acquisition and the expected completion time are greater; under the same task priority, key data items that may change the risk level are given priority.

[0272] S64: Determine the minimum necessary set of replenishment tasks

[0273] When multiple key data items that need to be supplemented can be obtained through the same data collection, information verification, repeated testing, or further screening, they are merged into a single candidate task unit. The same key data item can be retained in multiple candidate task units that can obtain the item, forming a many-to-many coverage relationship between key data items and candidate task units. They are not uniquely assigned before the minimum necessary supplementary data collection task set is determined.

[0274] Configure a unique index for each candidate task unit and the set of key supplementary data items it can acquire. For any subset of candidate task unit indices, the sum of task priorities within the union of the key supplementary data items covered by each candidate task unit is used as the coverage value of that subset; the same key supplementary data item is calculated only once. Determine the minimum necessary supplementary data collection task set according to the principle of minimizing the number of tasks required to cover the preset proportion of task value.

[0275]

[0276] (19)

[0277] in, Represents a subset of the index The coverage value; Indicates the current node; and These represent the candidate task unit index and the key data item index to be supplemented, respectively. Represents candidate task units A set of key data items that need to be supplemented is available; Indicates task priority; This represents the set of indexes for executable candidate task units, with superscripts indicating executable. This represents the unique candidate task unit index subset that satisfies the coverage constraint, has the fewest number of tasks, and is determined by the parallel rules. Represents candidate task units ; This represents the minimum set of necessary re-collection tasks; This represents the task value coverage threshold greater than zero and not exceeding one. The summation operation accumulates different task priorities, and the union eliminates duplicate key data items to be supplemented. Paired vertical lines indicate the number of elements; a single vertical line within curly braces indicates that elements satisfying the subsequent condition are selected; the summation subset relationship limits the index range; and English abbreviations indicate that the subsequent constraints are satisfied.

[0278] When the number of tasks is equal, the following criteria are selected in order: a large number of key data items to be supplemented that may change the risk level after deduplication; low total cost; short critical path; and candidate task unit unique index that appears earlier. The total cost is the sum of the actual costs of the candidate task units; when the critical path is executed in parallel, the longest estimated completion time is used; when executed sequentially, the sum of the estimated completion times is used. If the set of executable candidate task unit indexes is empty or the total coverage value is zero, the minimum necessary supplementary data collection task set is set to empty, and step S65 is not executed; only one manual review task and manual review identifier are generated before proceeding to step S74. The manual review task adopts the record fields of step S65 and the return structure of step S1, including task and personnel identifiers, standard fields or medical concepts to be reviewed, and return format.

[0279] S65: Generate and issue supplementary mining tasks

[0280] Four types of tasks are issued only when the minimum necessary supplementary data collection task set is not empty. The task and personnel identifiers, task type, standard fields or medical concepts corresponding to key data items to be supplemented, main risk factors, affected risk tasks, task priority, execution conditions, estimated completion time, return format, and original record index are recorded. The supplementary data collection round is initialized to zero before the first entry into step S6 and is retained across rounds. Each time a non-empty minimum necessary supplementary data collection task set is issued, it increments by one, and is not reset to zero upon subsequent entries into step S6.

[0281] The returned data uses the standardized data structure from step S1, and carries the task identifier and actual collection time, for step S7 to update the longitudinal follow-up sequence.

[0282] Step S6 outputs the minimum necessary supplementary data collection task set, the key data items to be supplemented set, task priority, estimated completion time, execution status, and supplementary data collection round; only when the minimum necessary supplementary data collection task set is not empty, the corresponding automatic supplementary data collection task proceeds to step S7, and empty sets are processed according to step S64.

[0283] Step S7: Feedback Data Update and Risk Closed-Loop Review

[0284] Step S7 receives the data returned by the task in step S6 and updates the longitudinal follow-up sequence according to the task identifier and the actual collection time.

[0285] S7 specifically includes the following steps:

[0286] S71: Receive and verify task return data

[0287] Data is returned based on task and personnel identification, verifying task type, standard fields or medical concepts corresponding to key data items to be supplemented, actual collection time, data source, and execution status. Supplemented, tested, and screened data are standardized according to step S1, and information verification generates traceable records. Each key data item to be supplemented is recorded as in progress, pending verification, completed, failed, or terminated: passing verification indicates completion; pending verification indicates attribution or format is pending confirmation; missing necessary identifiers or times indicates pending verification and the sequence is not updated; execution failure indicates failure. After pending verification, it is converted to completed or failed; if the execution exceeds the expected completion time or the pending verification exceeds the positive fixed verification time limit, it is converted to failure; key data items to be supplemented that are in progress or pending verification are not reissued. Tasks containing multiple key data items to be supplemented are allowed to be partially completed, with only completed data proceeding to step S72. The failure reason and path are recorded in the current supplementary collection round and transmitted together when all key data items to be supplemented reach the final state and step S74 decides to proceed to step S6.

[0288] S72: Update the effective longitudinal follow-up sequence

[0289] Completed and verified data are sent to step S2; supplementary collection, testing, and screening results are entered according to the actual collection time. Information verification is linked to the original record through the original record index, and the original collection time and verification time are retained. After standardization, the follow-up period and validity coefficient are re-determined; when the time is later than the current node, a new review node is created; if it belongs to an existing period, it is included in that period without replacing it with the receiving time.

[0290] Numerical data supplementation or testing generates new standardized records; review of medical history, symptoms, or examination conclusions generates confirmation or correction records related to the original records and updates positive and negative evidence; examination and screening results generate corresponding event and medical concept evidence. Updates do not overwrite historical records; node representative data, comprehensive validity coefficients, and data missing indicators are recalculated in step S2 based on all records.

[0291] S73: Re-perform risk assessment and reliability determination

[0292] After all key data items to be supplemented in this round of data collection have been completed or failed, steps S3 to S5 are re-executed using the sequence updated by the completed key data items. The comprehensive effectiveness coefficients and derived characteristics affected by time or historical dependencies are recalculated from the affected follow-up period to the current node, reusing only uncorrected static fields unaffected by time. Step S74 is not performed if there are key data items to be supplemented that are in progress or awaiting review; if there are no completed key data items in this round of data collection, step S74 is performed after maintaining the reliability status of the risk results and overall risk results.

[0293] S74: Perform the closed-loop termination check and output the result.

[0294] When the overall risk result reliability status is reliable, the supplementary data collection is terminated, and two types of risk results, their risk levels, main risk factors, and risk result reliability are output. If the overall risk result reliability status is still conditionally reliable, unreliable, or insufficient data, and the maximum supplementary data collection round has not been reached and there is no manual review indicator, the process returns to step S6, where step S61 reassesses the executability based on the latest key data items to be supplemented. Automatic supplementary data collection stops when a manual review indicator exists or the maximum supplementary data collection round has been reached; when the maximum supplementary data collection round has been reached and there is no manual review task, only one manual review task is generated, using the record fields and return structure from step S64. Unfinished automatic tasks are marked as terminated, while manual review tasks remain pending or in progress; the currently available risk results, overall risk result reliability status, unresolved key data items to be supplemented, final task status, manual review task, and manual review indicator are output. The data-insufficient branch does not output risk results. When valid data is generated by manual review, the manual review task is marked as completed, the manual review indicator for this task is cleared, and the data is entered into step S71 as information review data. The maximum number of supplementary sampling rounds is a positive configuration value. The final risk results are used only for risk assessment and follow-up control; only data that may affect the reliability or risk level of the risk results and are actionable are supplemented to reduce duplicate sampling and unnecessary screening.

[0295] The above methods can be used to uniformly process multimodal follow-up data collected from different sources and at different times, and conduct comprehensive cardiovascular and cerebrovascular risk assessments and stroke-specific risk assessments based on the completeness, timeliness, and effectiveness of the data, while also judging the reliability of the risk results.

[0296] When critical data that may affect the reliability or risk level of risk results is found to be missing, invalid, or conflicting, priority is given to identifying the data that needs to be supplemented and determining the minimum necessary data collection set. After the supplemented data is returned, the longitudinal follow-up sequence is updated and a new risk assessment and reliability judgment are performed, thus forming a risk assessment, necessary supplementation, and closed-loop review process. This reduces duplicate data collection and unnecessary screening, making the risk assessment results more consistent with the current actual data status of the target personnel.

[0297] The scope of protection claimed in this application shall be determined by the contents of its claims, and the contents described in the invention description, specific embodiments and drawings shall be used to interpret the claims.

[0298] Within the scope of the technical concept of this application, several modifications can be made to the specific implementation of this application, and these modified implementations should also be considered within the protection scope of this application.

Claims

1. A method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data, characterized in that, Includes the following steps: Step S1: Acquisition and standardization of multimodal follow-up data. Acquire multimodal follow-up data related to the comprehensive cardiovascular and cerebrovascular risk and stroke-specific risk of the target personnel. Unify personnel identification, fields, units of measurement, medical concepts, data sources and actual collection time to form standardized multimodal follow-up data for use in Step S2. Step S2: Construction of time-valid longitudinal follow-up sequence. Using the data output from step S1 as input, establish follow-up period according to actual collection time, convert continuous monitoring data into periodic features, and combine validity period, collection status and monitoring coverage to generate a valid longitudinal follow-up sequence for use in step S3. Step S3: Generation of longitudinal risk features. Step S3 uses the effective longitudinal follow-up sequences output in Step S2. As input, generate current indicators, time interval perceived change characteristics, reference range deviation, abnormal duration, number of consecutive abnormalities, and medical conceptual evidence to form longitudinal risk features for use in step S4; Step S4: Comprehensive cardiovascular and cerebrovascular risk assessment and stroke-specific risk assessment; Step S4 uses the longitudinal risk characteristics output from Step S3. As input, the participation level of various features is determined based on data completeness, timeliness, and validity. A comprehensive risk status value is first generated, and then corrected using stroke-specific risk evidence. Step S5: Reliability assessment of risk results and identification of key data items to be supplemented. Step S5 receives the comprehensive risk results of cardiovascular and cerebrovascular diseases and their risk levels, the specific risk results of stroke and their risk levels, the main risk factors, the risk contribution value, as well as the data quality coefficient, fusion coefficient, feature missing identifier, medical concept evidence conflict intensity, feature validity coefficient, model participation path and traceability information, and performs reliability assessment of risk results and identification of key data items to be supplemented. Step S6: Minimum Necessary Data Supplementation Task Generation. Step S6 takes the overall risk result reliability status and key data items to be supplemented as inputs, and combines the risk impact value, reliability gap, data acquisition cost, expected completion time and collection conditions to determine the minimum necessary data supplementation task set. Step S7: Feedback data update and risk closed-loop review. Step S7 receives the data returned by the task in Step S6 and updates the longitudinal follow-up sequence according to the task identifier and the actual collection time.

2. The method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data according to claim 1, characterized in that, S2 specifically includes the following steps: S21: Determine follow-up nodes and divide follow-up periods; S22: Convert continuous monitoring data into follow-up period-level features. For continuous monitoring data such as family blood pressure, continuous heart rate, activity level and sleep duration, filter valid samples according to the actual collection time, and calculate the period mean, period fluctuation value, minimum value, maximum value, last valid sample value and monitoring coverage. S23: Determine the time validity coefficient of the data. In the standard field dictionary, configure the corresponding validity period and stable validity ratio for dynamic indicators, test data, inspection results and screening results that need to be judged for timeliness. The first Follow-up time points and multimodal follow-up data The difference between the actual collection times is denoted as the data time difference. Multimodal follow-up data The corresponding standard fields are standard fields. Standard fields The corresponding validity period is The stable and effective ratio is Then multimodal follow-up data In the The effective time coefficient for each follow-up node is: （1） in, Represents multimodal follow-up data In the The time effectiveness coefficient of each follow-up node Indicates the time difference of the data. Represents standard fields The corresponding validity period, Represents standard fields The corresponding stable and effective ratio; S24: Determine the overall validity coefficient of the data. Based on the standard data value, unit of measurement, actual collection time, and collection status, determine the data usability coefficient. Based on the data usability coefficient, time validity coefficient, and monitoring coverage of continuous monitoring data, determine the overall validity coefficient of the data at the current follow-up node. （2） in, Represents multimodal follow-up data In the The overall effective coefficient of each follow-up node Represents multimodal follow-up data Available data coefficients Represents multimodal follow-up data Time efficiency coefficient, Indicates the coverage correction factor; S25: Generate a valid longitudinal follow-up sequence by combining the node representative data, follow-up period-level features, comprehensive validity coefficient, data missing identifier, actual follow-up interval, and valid event records corresponding to each follow-up node in chronological order to obtain the valid longitudinal follow-up sequence. （3） in, This represents the valid longitudinal follow-up sequence up to the current follow-up node. Indicates the current follow-up node number. Indicates the first The time corresponding to each follow-up node Indicates the first Each follow-up node corresponds to a data vector composed of node representative data and follow-up period-level features. This represents a vector composed of the comprehensive effective coefficients corresponding to each data point. This represents a vector composed of data missing identifiers corresponding to each standard field. Indicates the first The actual follow-up interval corresponding to each follow-up cycle. Indicates the first The collection of follow-up texts, examination conclusions, and stroke screening event records retained within each follow-up period.

3. The method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data according to claim 2, characterized in that, S3 specifically includes the following steps: S31: Determine the current indicator value and the previous valid indicator value; S32: Generate time-interval-aware indicator change characteristics, when both the current indicator value and the previous valid indicator value are available, This indicates the change in the two values ​​during adjacent follow-ups, expressed as... This represents the change per unit time calculated based on the actual follow-up interval; where the change between adjacent follow-ups is the current indicator value minus the previous valid indicator value, and the change per unit time is the change between adjacent follow-ups divided by the actual time interval between the current node and the previous valid node that is greater than 0. S33: Deviation characteristics of the reference range of the generated indicator; In the standard field dictionary, configure the lower reference limit, upper reference limit, and reference scale for dynamic indicators with medical reference ranges; set the standard fields. The lower reference limit is The upper limit of the reference is The reference scale is Then the standard field The deviation characteristics from the reference range at the current follow-up node are as follows: （4） in, Represents standard fields Deviation from the reference range at the current follow-up node; and These represent the lower reference limit and the upper reference limit, respectively. Indicates the reference scale, and The default value is greater than zero.

4. The method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data according to claim 3, characterized in that, S3 also includes the following steps: S34: Generate persistent anomaly features; For dynamic indicators of systolic blood pressure, diastolic blood pressure, heart rate, weight, blood glucose and blood lipids, the effective duration of the current abnormal process is calculated based on the abnormal status at each follow-up node, the actual follow-up interval and the comprehensive effective coefficient. Standard fields The duration of the abnormality at the current follow-up node is: （5） in, Represents standard fields The duration of the anomaly up to the current follow-up point; Represents standard fields The corresponding set of current abnormal interval indices; This indicates the first step determined in step S2. The actual follow-up interval corresponding to each follow-up cycle; This indicates that step S2 is a standard field. The validity period of the configuration; and Representing standard fields The combined effective coefficient between two adjacent follow-up nodes; and These represent the abnormal states of two adjacent follow-up nodes, respectively. S35: Generate a feature of consecutive anomaly counts; by Represents standard fields up to the The number of consecutive valid anomalies that occur after the most recent valid normal state for each follow-up node; when the standard field of the current follow-up node is in a valid normal state, the consecutive anomaly count is cleared to zero; when the current follow-up node lacks valid data, the missing state is not counted as a normal state, nor is the consecutive anomaly count increased. Instead, the consecutive anomaly count corresponding to the most recent valid follow-up node is maintained, and the consecutive anomaly state is represented by the data missing identifier of the current standard field to indicate that it is pending confirmation. The duration of an anomaly is used to characterize the extent to which an anomaly persists in the real time dimension, while the number of consecutive anomalies is used to characterize the extent to which an anomaly recurs at different follow-up nodes. S36: Extract risk concepts from follow-up texts and examination results; Obtain follow-up texts, medical history, examination results, stroke screening and high-risk investigation events from valid longitudinal follow-up sequences, and form unified medical concept evidence; The extracted symptoms, medical history, and abnormal items are mapped to the unified medical concept established in step S1; for cases where both positive and negative records exist for the same medical concept, positive evidence and negative evidence are retained respectively. Let the medical concept number be... The current follow-up point will be aligned with medical concepts. Events that are relevant and have a comprehensive validity coefficient greater than 0 are classified as the set of affirmative events. and the set of negative events And generate medical conceptual evidence features: （6） in, Indicates medical concept Characteristics of medical conceptual evidence at the current follow-up point; Indicates event logging; Indicates event logging The corresponding overall efficiency coefficient; Indicates event logging The corresponding event evidence strength is determined by the degree of anomalousness or the preset mapping strength for affirmative events and by the negative confirmation strength for negative events, with a value of 1 when no level is set. The affirmative evidence component of a medical concept is denoted as: The weight of the negative evidence is denoted as The value is zero when the corresponding event set is empty. A confirmed and uncorrected structured medical history serves as positive evidence, while a valid record clearly indicating the absence of the corresponding disease or symptoms serves as negative evidence. S37: Generate longitudinal risk characteristics; The longitudinal risk characteristics are composed of current indicators, changes in adjacent follow-ups, changes per unit time, deviation from reference range, duration of abnormality, number of consecutive abnormalities, medical conceptual evidence, overall validity coefficient, data missing markers, and actual follow-up intervals. When the previous node or current data is missing, retain the corresponding unavailable flag or data missing flag, and output the feature to step S4.

5. The method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data according to claim 4, characterized in that, S4 specifically includes the following steps: S41: Divide risk feature groups and transmit data status. First, distinguish shared features and stroke-specific exclusive features according to the model participation path. Shared features are then divided into structured longitudinal indicators, continuous monitoring and medical concept evidence feature groups according to the data formation method. S42: Generate conflict-suppressing medical concept directional features; for each medical concept, retain the strength of affirmative evidence and the strength of negative evidence separately. To avoid the direct addition of affirmative and negative evidence for the same medical concept in risk assessment, first calculate the conflict strength of medical concept evidence, and then generate medical concept directional features based on this: （7） in, Indicates medical concept The strength of conflict of medical conceptual evidence at the current follow-up point; This represents the medical concept generated in step S3. Corresponding strength of affirmative evidence; Indicates medical concept The corresponding strength of refutation evidence; Indicates medical concept Corresponding medical concept and directional characteristics; Indicates the sequence number of a medical concept; S43: Calculate the data quality coefficient for the risk characteristic group; Let the risk characteristic group number be Risk Characteristic Group The corresponding feature index set is The number of features in the feature index set is ; Based on the feature missing identifiers and feature validity coefficients of each feature within the risk feature group, the feature completeness, average validity, and data quality coefficients are calculated respectively: （8） in, Indicates risk characteristic group Feature completeness at the current follow-up node; Indicates risk characteristic group Average validity at the current follow-up node; Indicates risk characteristic group Data quality coefficient at the current follow-up node; Indicates risk characteristics Corresponding feature missing identifier; Indicates risk characteristics The corresponding feature efficiency coefficient; Indicates the risk characteristic number; Indicates risk characteristic group The corresponding integrity adjustment coefficient; S44: Generate risk evidence values ​​for risk feature groups, perform scaling or encoding on continuous, enumerated, and medical concept-oriented features, use neutral placeholder values ​​for missing features and simultaneously input feature missing labels; the risk feature group encoder outputs risk evidence values. .

6. The method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data according to claim 5, characterized in that, S4 also includes the following steps: S45: Fusion of evidence regarding the risks of enforcing data quality constraints; Based on the data quality coefficient and risk evidence value of each risk feature group, the fusion coefficient for each risk feature group in the comprehensive cardiovascular and cerebrovascular risk calculation is calculated: （9） in, Indicates risk characteristic group The fusion coefficient at the current follow-up node; Indicates risk characteristic group Data quality coefficient; Indicates risk characteristic group Risk evidence value; This represents the fusion smoothing parameter, and The default value is greater than zero; Indicates the number of risk characteristic groups; This indicates the risk characteristic group number used for summation; S46: Generate comprehensive cardiovascular and cerebrovascular risk results and stroke-specific risk results; Based on the risk evidence value and fusion coefficient of each risk characteristic group, a comprehensive risk status value for the target individual at the current follow-up node is generated, and the comprehensive risk status value is converted into a comprehensive cardiovascular and cerebrovascular risk result: （10） in, This indicates the overall risk status value of the target individual at the current follow-up point; This indicates the overall cardiovascular and cerebrovascular risk outcome of the target individuals at the current follow-up point; This indicates a bias in the comprehensive risk mapping of cardiovascular and cerebrovascular diseases; This represents the comprehensive risk mapping coefficient for cardiovascular and cerebrovascular diseases, and Positive numbers; superscript This indicates a comprehensive risk assessment task related to cardiovascular and cerebrovascular health. This represents a logical function that maps input values ​​to a range between zero and one. In the comprehensive risk status value Based on this, stroke-specific risk evidence is introduced for correction, generating stroke-specific risk results: （11） in, This indicates the stroke-specific risk evidence value corresponding to the current follow-up node; Represents the stroke-specific risk evidence mapping vector; This represents a stroke-specific risk characteristic encoder; This represents the stroke-specific risk feature vector after scaling transformation. This indicates a bias in the mapping of evidence regarding the specific risks of stroke. This indicates the stroke-specific risk outcome corresponding to the current follow-up node; This indicates a bias in the risk mapping for stroke. This represents the overall risk status transmission coefficient; This indicates the correction factor for specific risk evidence related to stroke. Data quality coefficient representing specific risk characteristics of stroke; superscript This indicates a specific risk assessment task related to stroke. S47: Determine the risk level and main risk factors; set a first risk threshold and a second risk threshold for the comprehensive cardiovascular and cerebrovascular risk results and the stroke-specific risk results respectively, and divide the risk level into three levels: low, medium and high. In order to determine the main risk factors, the current indicator value, adjacent follow-up change, unit time change, abnormal duration, number of consecutive abnormalities and medical conceptual evidence belonging to the same medical risk factor are combined into a risk factor unit. Replace the risk factor units sequentially with the corresponding reference neutral states, while keeping other risk characteristics, data quality coefficients, and feature missing identifiers unchanged. Recalculate the risk results. The risk contribution value for risk tasks is: （12） in, Representing risk factor units Risk tasks Risk contribution value; This indicates the risk result calculated using complete longitudinal risk characteristics; This indicates the risk factor unit. The risk result was recalculated after replacing it with a reference neutral state; Indicates the unit number of the risk factor; Pick The time indicates the comprehensive risk task of cardiovascular and cerebrovascular diseases, take The time indicates a specific risk task related to stroke; subscript This indicates a reference to a neutral state.

7. The method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data according to claim 6, characterized in that, S5 specifically includes the following steps: S51: Establish risk outcome support relationships and a candidate data item set; S52: Assess the extent of effective data support for the main risk factors; Set risk tasks The corresponding set of main risk factors is as follows Key risk factors The corresponding set of supporting risk feature indices is Based on the missing data identifiers, feature validity coefficients, data quality coefficients of their respective risk feature groups, and the strength of conflict of medical conceptual evidence for each supporting risk feature, the effective support of the main risk factors is calculated, and the risk support of the corresponding risk tasks is further calculated. （13） in, Indicates key risk factors Effective support at the current follow-up node; Indicates the main risk factors The strength of conflict of relevant medical conceptual evidence is zero when no medical conceptual evidence is involved. This indicates support for the main risk factors. A set of risk characteristic indexes Indicates the number of elements in the set; Indicates the risk characteristic number; and These represent the risk feature missing identifier and the risk feature validity coefficient determined in step S4, respectively. Indicates risk characteristics The risk characteristic group to which it belongs; This represents the data quality coefficient for the corresponding risk characteristic group. Indicates risk task Risk support level at the current follow-up node; This indicates the main risk factors calculated in step S4. Risk tasks Risk contribution value; This indicates a pre-defined positive number to prevent the denominator from being zero; S53: Construct candidate data completion scenarios and determine risk outcome ranges for the candidate data item set. For each candidate data item, establish low-risk completion scenarios and high-risk completion scenarios respectively; Input the low-risk and high-risk completion scenarios into the risk assessment model that has been trained in step S4, respectively, to obtain candidate data items. The two corresponding risk recalculation results are denoted as follows: and ; Based on the current risk results and the recalculated risk results of each candidate data item, the risk tasks are determined. Corresponding risk outcome range: （14） in, Indicates risk task The lower bound of risk outcomes under the current data conditions; This indicates the upper bound of the corresponding risk outcome; This indicates the current risk result output in step S4; Indicates candidate data items; Represents a set of candidate data items; Representing candidate data items Recalculation of risk results when using low-risk completion scenarios; Representing candidate data items Recalculation of risk results when using high-risk completion scenarios; S54: Determine the stability of the risk level and calculate the reliability of the risk outcome; Obtain step S4 for risky tasks A first risk threshold and a second risk threshold are set, and the two risk thresholds are combined to form a risk level boundary set. ; The stability of the risk level is calculated based on the width of the risk outcome interval and the number of risk level boundaries that the risk outcome interval crosses: （15） in, Indicates risk task Corresponding risk level stability; Indicates the risk level boundary; Indicates risk task The corresponding risk level boundary set; This indicates a conditional indicator function that returns 1 if the condition within the parentheses is true, and zero otherwise. By combining risk support and risk level stability, we obtain the reliability of risk outcomes corresponding to risk tasks: （16） in, Indicates risk task Reliability of risk outcomes at the current follow-up node; Indicates risk tolerance; Indicates the stability of the risk level; Indicates risk task The corresponding support adjustment coefficient; S55: Calculate the risk impact value of candidate data items; For each candidate data item, calculate the difference in risk outcomes between its low-risk completion scenario and high-risk completion scenario, and determine whether the two completion scenarios correspond to different risk levels; Calculate the risk impact value of the candidate data items based on the degree of data gap, differences in risk outcomes, and changes in risk levels: （17） in, Representing candidate data items Risk impact value at the current follow-up node; Representing candidate data items The extent of the data gap; Indicates risk task The corresponding risk difference adjustment coefficient; This represents the risk level mapping function that maps the risk result to a low-risk, medium-risk, or high-risk level based on the risk threshold in step S4. The condition indicator function follows the aforementioned definition. S56: Determine the overall risk outcome reliability status and key data items to be supplemented; first determine the unreliable state, i.e., the reliability of the risk outcome of any risk task is lower than the first reliability threshold or the risk outcome interval crosses two risk level boundaries; in other cases, when the reliability of the risk outcome of any risk task is lower than the second reliability threshold or the risk outcome interval crosses one risk level boundary, it is judged as conditionally reliable; when the reliability of the risk outcome of two risk tasks both reach the second reliability threshold and the risk outcome intervals do not cross the risk level boundary, it is judged as reliable. When the states of two risk tasks are different, the one with the lower reliability is used as the overall risk outcome reliability status; Candidate data items whose risk impact value reaches the preset risk impact value threshold or may change the adjacent risk level are identified as key data items to be supplemented, and sorted by risk impact value; according to the data gap type, key data items to be supplemented are marked as candidate items for data supplementation, information verification, repeated testing, or further screening.

8. The method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data according to claim 7, characterized in that, S6 specifically includes the following steps: S61: Generate executable candidate tasks. Each time, four types of candidate tasks are regenerated and their executableness is judged based on the current key data items to be supplemented. S62: Quantifying the burden of data acquisition; S63: Calculate the task priority of key data items to be supplemented, based on risk impact value, reliability gap, data acquisition burden, and collection conditions. （18） in, Indicates key data items that need to be supplemented. Task priority; Indicates the data collection condition coefficient; This represents the risk impact value output in step S5; This indicates a reliability gap; Indicates the cost of standardized data acquisition; Indicates the estimated completion time for standardization; , and These represent adjustment coefficients corresponding to reliability gap, data acquisition cost, and estimated completion time, respectively. S64: Determine the minimum necessary set of replenishment tasks When multiple key data items to be supplemented can be obtained through the same data collection, information verification, duplicate detection, or further screening, they are merged into a candidate task unit; each candidate task unit is configured with a unique index and the set of key data items to be supplemented that it can obtain. Based on the principle of minimizing the number of tasks required to cover the preset proportion of task value, determine the minimum necessary set of supplementary collection tasks: ， （19） in, Represents a subset of the index The coverage value; Indicates the current node; and These represent the candidate task unit index and the key data item index to be supplemented, respectively. Represents candidate task units A set of key data items that need to be supplemented is available; Indicates task priority; This represents the set of indexes for executable candidate task units, with superscripts indicating executable. This represents the unique candidate task unit index subset that satisfies the coverage constraint, has the fewest number of tasks, and is determined by the parallel rules. Represents candidate task units ; This represents the minimum set of necessary re-collection tasks; This represents the task value coverage threshold that is greater than zero and does not exceed one. S65: Generate and issue supplementary mining tasks.

9. The method for predicting the risk of cardiovascular and cerebrovascular diseases based on multimodal data according to claim 8, characterized in that, S7 specifically includes the following steps: S71: Receive and verify the data returned by the task; S72: Update the valid longitudinal follow-up sequence. Completed and verified data are sent to step S2. Supplementary collection, testing and screening results are entered according to the actual collection time. Information verification is linked to the original record through the original record index, and the original collection time and verification time are retained. After standardization, the follow-up period and validity coefficient are re-determined. Numerical supplementary collection or testing generates new standardized records. S73: Re-perform the risk assessment and reliability determination; S74: Perform the closed-loop termination judgment and output the result.

10. A cardiovascular and cerebrovascular disease risk prediction system based on multimodal data, characterized in that: It includes a data acquisition module, which is used to acquire multimodal follow-up data and standardize it, to acquire multimodal follow-up data related to the comprehensive cardiovascular and cerebrovascular risk and stroke-specific risk of the target personnel, and to unify personnel identification, fields, units of measurement, medical concepts, data sources and actual collection time; The longitudinal follow-up sequence construction module is used to establish follow-up cycles according to the actual collection time, convert continuous monitoring data into periodic features, and generate valid longitudinal follow-up sequences by combining the validity period, collection status and monitoring coverage. The longitudinal risk feature generation module, with effective longitudinal follow-up sequences As input, it generates current indicators, time interval perceived change characteristics, reference range deviation, abnormal duration, number of consecutive abnormalities, and medical conceptual evidence; The comprehensive cardiovascular and cerebrovascular risk assessment module, which includes stroke-specific risk assessment, uses longitudinal risk characteristics. As input, the participation level of various features is determined based on data completeness, timeliness, and validity to generate a comprehensive risk status value, which is then corrected using stroke-specific risk evidence. The risk outcome reliability assessment and key supplementary data item identification module is used to receive the comprehensive cardiovascular and cerebrovascular risk results and their risk levels, the stroke-specific risk results and their risk levels, the main risk factors, the risk contribution value, as well as the data quality coefficient, fusion coefficient, feature missing identifier, medical concept evidence conflict intensity, feature validity coefficient, model participation path and traceability information, and to perform risk outcome reliability assessment and key supplementary data item identification. The minimum necessary data supplementation task generation module is used to determine the minimum necessary data supplementation task set by taking the overall risk result reliability status and key data items to be supplemented as input, combined with risk impact value, reliability gap, data acquisition cost, expected completion time and collection conditions. The feedback data update and risk closed-loop review module is used to receive data returned by the task and update the longitudinal follow-up sequence according to the task identifier and the actual collection time.

Images