Coronary risk prediction method and system for patients with coronary heart disease

Abstract

The present application relates to a kind of coronary risk prediction method and system of coronary heart disease patient, to solve the limitation that traditional static evaluation model is difficult to integrate multi-source dynamic data. The multi-source heterogeneous data of patient is collected from hospital information system, gene database and wearable device;Key features are extracted by special processing module respectively: unstructured text and image plaque features are analyzed using deep learning model, multi-gene risk score is calculated, and time series analysis is used to capture physiological parameter attenuation trend, a three-level cascade integrated model is constructed, the first level filters high contribution degree features, the second level integrates space-time dynamic features through attention mechanism network, the last level integrates static and dynamic factors to generate dynamic risk score, and output visual risk trend chart and high-risk factor heat map, real-time early warning is generated through medical terminal and individualized intervention scheme is generated. The present application realizes the collaborative analysis of multidimensional data and dynamic risk tracking, significantly improves early warning capability and clinical decision efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical big data, and in particular to a coronary risk prediction method and system for coronary heart disease patients. BACKGROUND

[0002] There are significant limitations in the field of coronary heart disease risk prediction. Traditional clinical assessment mainly relies on static indicators (such as blood lipids, blood pressure) and single imaging examination, which is difficult to integrate multi-dimensional dynamic risk factors. Existing prediction models are usually constructed based on a single data source, lacking comprehensive analysis of genetic susceptibility, long-term physiological changes and plaque characteristic evolution, resulting in insufficient sensitivity in identifying high-risk patients, especially in ignoring the early warning of asymptomatic progressive lesions.

[0003] Existing technologies face bottlenecks in multi-source heterogeneous data processing. Unstructured text in electronic medical records, quantitative features of medical images, genomic data, and time-series physiological parameters generated by wearable devices have significant differences in data structure and spatio-temporal scale, and an effective collaborative analysis framework has not yet been established. Conventional prediction methods have limited depth in feature extraction of unstructured text, and the accuracy of automatic identification of image plaque characteristics is insufficient, and few models can simultaneously process the interactive effects of static clinical data and dynamic physiological monitoring signals.

[0004] Clinical practice urgently needs dynamic and personalized risk warning tools. Current risk scoring systems generally lack real-time updating mechanism, and cannot reflect the impact of short-term state changes of patients on prognosis; at the same time, the explainability of model decision-making process is insufficient, which makes it difficult for doctors to quickly locate the core risk driving factors. In addition, existing systems are mostly independent of diagnosis and treatment processes, and cannot directly translate risk prediction results into intervention strategies and coordinate medical resources, limiting the clinical transformation value. SUMMARY

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] According to the first aspect of the present application, a coronary risk prediction method for coronary heart disease patients is claimed, comprising:

[0007] S1, determining a target patient, and collecting and acquiring multi-source heterogeneous risk basic data of the target patient from a multi-source database;

[0008] S2, obtaining layered feature parameters of each multi-source heterogeneous risk basic data based on the multi-source heterogeneous risk basic data using different processing models;

[0009] S3, inputting the layered feature parameters into a three-level cascaded integrated model to output a dynamic risk prediction score of the target patient;

[0010] S4, generating a dynamic risk trend chart and a high-risk factor contribution heat map based on the dynamic risk prediction score, and pushing a warning signal in real time through a medical terminal.

[0011] Further, the S1 further comprises:

[0012] Real-time acquisition of clinical data, genomics data, and dynamic physiological monitoring data of the target patient from a hospital information system, a gene database, and a wearable device, wherein the clinical data at least includes electronic medical records EMR, coronary CT angiography CCTA images, and laboratory test reports;

[0013] The S2 further comprises:

[0014] Unstructured text in the clinical data is extracted for key pathological features using a BERT model, and CCTA images are automatically quantified for coronary artery calcification score CACS and plaque features using a 3D convolutional neural network;

[0015] Genomics data is processed through a polygenic risk score PRS model to calculate the genetic risk value of coronary heart disease;

[0016] Dynamic physiological data is processed through a time series anomaly detection algorithm to extract heart rate variability HRV attenuation indicators;

[0017] The S3 further comprises:

[0018] The hierarchical feature parameters output by S2 are input into a three-level cascade integrated model, wherein the first level uses XGBoost for feature importance screening, the second level uses a bidirectional LSTM network with attention mechanism to process time series features, and the third level integrates static and dynamic features through a random forest to generate a dynamic risk prediction score.

[0019] Further, the extraction of plaque features in S2 comprises:

[0020] Non-calcified plaques, mixed plaques, and low-density plaques in CCTA images are identified using a U-Net++ segmentation model;

[0021] The ratio of plaque volume to vessel cross-sectional area PV / VA is calculated as a high-risk plaque quantification indicator;

[0022] The identification of calcified regions in S2 comprises:

[0023] The inner membrane boundary of the blood vessel is automatically outlined on the CT image cross-section;

[0024] Pixel regions with CT values greater than 130 HU are detected and marked as calcified plaques;

[0025] The total volume of calcified plaques is calculated by continuously scanning along the vessel axis.

[0026] Further, the construction of the S2 multi-gene risk score PRS model also includes:

[0027] Extract coronary heart disease related SNP sites from the UK Biobank database, and use LASSO regression to screen significant sites;

[0028] Correct the weight based on the allele frequency of the target population, and generate patient-specific PRS values;

[0029] The S2 also includes the generation of a genetic risk index, and preferentially screens 20 gene sites that are strongly associated with coronary heart disease in the target population;

[0030] According to the fold increase in coronary heart disease risk of each site, set the weighting coefficient;

[0031] Add the weighting values of all high-risk sites carried by the patient to obtain the final index.

[0032] Further, the feature interaction module is introduced in the S3 three-level cascade integration model, and the interaction terms of clinical features and genomic features are generated through an explicit feature cross network; the interaction terms and original features are input into a bidirectional LSTM network;

[0033] The second level processing of the S3 includes arranging the heart rate variability data of the continuous 7 days in time sequence, automatically identifying the decay trend of the variability that lasts for more than 5 days, and inputting the decay slope as an independent risk factor into the third level model;

[0034] The third level processing of the S3 includes multiplying the genetic risk index and the calcification volume percentage to generate a composite index, and when the composite index is greater than a threshold value, automatically upgrading the final risk level.

[0035] Further, the attention mechanism of the S3 is specifically implemented as:

[0036] Add a time attention layer to the output layer of the bidirectional LSTM to dynamically weight the physiological monitoring data at different time points;

[0037] The correlation between attention weight and risk score is optimized through back propagation;

[0038] Before performing the S3, it also includes:

[0039] Smooth filtering is performed on abnormal peaks in ECG data caused by motion interference;

[0040] Delete blurred frames in CT images caused by respiratory motion artifacts.

[0041] Further, the method also includes:

[0042] The model continuously learns, and when the amount of new patient data reaches a threshold, an incremental learning process is started, and the elastic weight consolidation (EWC) algorithm is used to prevent catastrophic forgetting of the model.

[0043] The model is updated online, and every month, new coronary heart disease patient data from the hospital is received, and only the third-level decision tree model is retrained while the original feature screening rules are retained. When the accuracy of the validation set changes by more than 5%, the full model update is triggered.

[0044] Further, the generation of the high-risk factor contribution heat map of S4 includes:

[0045] The SHAP value is used to explain the model output, the marginal contribution of each feature to the risk score is calculated, and a feature importance map is generated according to the absolute value of the contribution.

[0046] The top 10% of calcified plaques in terms of volume proportion are highlighted in red, and a numerical dashboard of genetic risk index and dynamic physiological indicators is displayed beside the blood vessel model.

[0047] Further, the dynamic physiological monitoring data of S1 includes:

[0048] The continuous 7-day night resting heart rate, QT interval dispersion, and blood oxygen saturation fluctuation standard deviation are subjected to sensor noise suppression through Kalman filtering.

[0049] The processing of the dynamic physiological monitoring data further includes counting the number of blood oxygen abnormalities during night sleep, and activating an emergency data retransmission mechanism when the number of single-night abnormalities exceeds 10.

[0050] Further, the dynamic risk prediction score is used for:

[0051] A personalized intervention strategy recommendation is generated, including drug adjustment suggestions and review cycle planning, which is linked to the hospital resource scheduling system to automatically schedule a heart specialist follow-up;

[0052] An intervention priority list is popped up on the doctor's workstation, and the sorting is based on the descending order of the risk score, and a report draft containing the review CT suggestion date and statin dose adjustment scheme is automatically generated.

[0053] According to the second aspect of the present application, the present application claims to protect a coronary risk prediction system for coronary heart disease patients, comprising:

[0054] One or more processors;

[0055] A memory having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the coronary risk prediction method for coronary heart disease patients.

[0056] The present application relates to a coronary risk prediction method and system for coronary heart disease patients, aiming to solve the limitation of traditional static evaluation model which is difficult to integrate multi-source dynamic data. The multi-source heterogeneous data of patients is collected from hospital information system, gene database and wearable devices; key features are extracted by special processing modules: deep learning model is used to analyze unstructured text and image plaque features, calculate multi-gene risk score, and capture physiological parameter decay trend by time series analysis, to build a three-level cascade integrated model, the first level screens high contribution features, the second level integrates space-time dynamic features through attention mechanism network, and the last level integrates static and dynamic factors to generate dynamic risk score, outputs visual risk trend chart and high-risk factor heat map, and generates personalized intervention scheme through real-time warning and association of medical terminal. The present application realizes collaborative analysis of multi-dimensional data and dynamic risk tracking, significantly improves early warning ability and clinical decision-making efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The workflow diagram of a coronary risk prediction method for coronary heart disease patients claimed by the embodiments of the present application;

[0058] Figure 2 The second workflow diagram of a coronary risk prediction method for coronary heart disease patients claimed by the embodiments of the present application;

[0059] Figure 3 The structure module diagram of a coronary risk prediction system for coronary heart disease patients claimed by the embodiments of the present application. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0061] The terms "first", "second", "third", etc. in the present application are only for descriptive purpose and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second", "third" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "multiple" is at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present application are only used to explain the relative position relationship, movement condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directional indications also change accordingly. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0062] Reference herein to "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment can be included in at least one embodiment of the application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily independent or alternative embodiments to other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0063] According to the first embodiment of the present application, the present application claims a coronary risk prediction method for coronary heart disease patients, referring to Figure 1 , comprising:

[0064] S1, determining a target patient, collecting and acquiring multi-source heterogeneous risk basic data of the target patient from a multi-source database;

[0065] S2, obtaining layered feature parameters of each multi-source heterogeneous risk basic data based on the multi-source heterogeneous risk basic data using different processing models;

[0066] S3, inputting the layered feature parameters into a three-level cascaded integrated model to output a dynamic risk prediction score of the target patient;

[0067] S4, generating a dynamic risk trend chart and a high-risk factor contribution degree heat map based on the dynamic risk prediction score, and pushing an early warning signal in real time through a medical terminal.

[0068] In this embodiment, the structured database of the hospital information system is extracted from the structured database of the hospital information system, including the patient electronic medical record, including the medical history record, the medication record and the operation record; the original DICOM file of the coronary CT angiography is obtained from the medical image archiving system; the continuous heart rate and blood oxygen data collected by the medical-grade electrocardiogram monitor worn by the patient are received in real time through the Bluetooth transmission protocol; the genetic detection report containing single nucleotide polymorphism sites is obtained from the genetic sequencing agency;

[0069] For clinical data processing, the unstructured text in the electronic medical record is input into the pre-trained medical language model, and the key pathological feature vector containing myocardial ischemia and angina symptoms is output; the coronary CT image is three-dimensionally reconstructed, the calcification area on the blood vessel wall is automatically identified, and the percentage of calcification volume to total blood vessel volume is calculated;

[0070] For genomic data processing, the patient gene site is compared with the risk site in the coronary heart disease genetic database, and the frequency of high-risk sites is counted; the genetic risk index is generated based on the clinical weight coefficient of different sites;

[0071] For dynamic physiological data processing, the electrocardiogram data is segmented and aggregated by hour, the standard deviation of adjacent heartbeat intervals is calculated as the heart rate variability index, and the abnormal period and its duration when the oxygen saturation is less than 92% are marked;

[0072] When constructing the dynamic risk prediction model, the hierarchical feature parameters generated by S2 are arranged in time sequence, and input into the cascade prediction system:

[0073] The first level uses gradient boosting decision tree to select the top 30% important features;

[0074] The second level uses recurrent neural network with time series memory function to analyze the trend of dynamic physiological characteristics over time;

[0075] The third level fuses the static features and trend analysis results, and generates a risk score of 0-100 by voting through multiple decision trees;

[0076] When the risk visualization output is generated, the risk score is converted into red (high risk), yellow (medium risk) and green (low risk) three-color warning levels; the coronary three-dimensional model and the visualization atlas superimposed with the plaque position are generated, and the calcification area and position coordinates are marked.

[0077] Further, the S1 further comprises:

[0078] The clinical data, genomic data and dynamic physiological monitoring data of the target patient are obtained from the hospital information system, genetic database and wearable device in real time, wherein the above-mentioned clinical data at least includes electronic medical record EMR, coronary CT angiography CCTA image and laboratory test report;

[0079] Referring to Figure 2 , the S2 further comprises:

[0080] The BERT model is used to extract key pathological features from unstructured text in clinical data, and a 3D convolutional neural network is used to automatically quantify coronary artery calcification score (CACS) and plaque features from CCTA images.

[0081] The PRS model is used to calculate the genetic risk value of coronary heart disease from genomics data.

[0082] The time series anomaly detection algorithm is used to extract heart rate variability (HRV) attenuation indicators from dynamic physiological data.

[0083] The S3 further comprises:

[0084] The hierarchical feature parameters output by the S2 are input into a three-level cascade integrated model, wherein the first level uses XGBoost to perform feature importance screening, the second level uses a bidirectional LSTM network with attention mechanism to process time series features, and the third level integrates static and dynamic features through a random forest to generate a dynamic risk prediction score.

[0085] Further, the extraction of plaque features in the S2 comprises:

[0086] The U-Net++ segmentation model is used to identify non-calcified plaques, mixed plaques and low-density plaques in CCTA images.

[0087] The ratio of plaque volume to vessel area (PV / VA) is calculated as a high-risk plaque quantification indicator.

[0088] The calcification region identification in the S2 comprises:

[0089] The intima boundary in the CT image cross-section is automatically outlined.

[0090] The pixel region with CT value greater than 130HU is detected and marked as calcified plaque.

[0091] The total volume of calcified plaque is calculated by continuously scanning along the vessel axis.

[0092] Further, the construction of the PRS model in the S2 further comprises:

[0093] The SNP sites related to coronary heart disease are extracted from the UK Biobank database, and LASSO regression is used to screen significant sites.

[0094] Based on the allele frequency correction weight of the target population, the patient-specific PRS value is generated.

[0095] The S2 further comprises the generation of a genetic risk index, which preferentially screens 20 gene loci strongly associated with coronary heart disease in the target population;

[0096] The weighting coefficient is set according to the multiple of the increase in the risk of coronary heart disease of each locus;

[0097] The weighted values of all high-risk loci carried by the patient are accumulated as the final index.

[0098] Further, the feature interaction module is introduced in the three-level cascade integrated model of the S3, and the interaction terms between the clinical features and the genomic features are generated through the explicit feature cross network, and the interaction terms and the original features are input into the bidirectional LSTM network;

[0099] The second-level processing of the S3 includes arranging the heart rate variability data of the continuous 7 days in time sequence, automatically identifying the decay trend of the variability degree lasting more than 5 days, and inputting the decay slope as an independent risk factor into the third-level model;

[0100] The third-level processing of the S3 includes multiplying the genetic risk index and the calcification volume percentage to generate a composite index, and automatically upgrading the final risk level when the composite index is greater than a threshold.

[0101] Further, the attention mechanism of the S3 is specifically implemented as:

[0102] The time attention layer is stacked on the output layer of the bidirectional LSTM, and the physiological monitoring data at different time points is dynamically weighted;

[0103] The correlation between the attention weight and the risk score is optimized through back propagation;

[0104] Before executing the S3, it further includes:

[0105] The abnormal peak values in the electrocardiogram data caused by motion interference are smoothed and filtered;

[0106] The blurred frames caused by respiratory motion artifacts in the CT image are deleted.

[0107] Further, the method further includes:

[0108] The model continuous learning step, when the amount of newly added patient data reaches a threshold, an incremental learning process is started, and the elastic weight consolidation EWC algorithm is used to prevent catastrophic forgetting of the model;

[0109] The model online updating step, receiving the data of the newly diagnosed coronary heart disease patients in the hospital every month, only retraining the third-level decision tree model and retaining the original feature screening rule, and triggering the full model update when the accuracy rate change of the validation set exceeds 5%.

[0110] Further, the generation of the high-risk factor contribution degree heat map of the S4 includes:

[0111] The SHAP value is used to explain the model output, and the marginal contribution of each feature to the risk score is calculated to generate a feature importance map according to the absolute value of the contribution degree;

[0112] The top 10% of the volume proportion of calcified plaques is highlighted in red, and the numerical dashboard of the genetic risk index and the dynamic physiological indicators is displayed beside the blood vessel model.

[0113] Further, the dynamic physiological monitoring data of S1 includes:

[0114] The night resting heart rate, QT interval dispersion, and blood oxygen saturation fluctuation standard deviation for 7 consecutive days are subjected to sensor noise suppression through Kalman filtering.

[0115] The processing of the dynamic physiological monitoring data further includes counting the number of blood oxygen abnormalities during night sleep, and activating an emergency data retransmission mechanism when the number of single-night abnormalities exceeds 10.

[0116] Further, the dynamic risk prediction score is used for:

[0117] A personalized intervention strategy recommendation is generated, including drug adjustment suggestions and review cycle planning, which is linked to a hospital resource scheduling system to automatically schedule a heart specialist follow-up;

[0118] An intervention priority list is popped up on the doctor's workstation, sorted in descending order of risk score, and a report draft containing review CT suggestion dates and statin dose adjustment schemes is automatically generated.

[0119] According to the second embodiment of the present application, the present application claims a coronary risk prediction system for coronary heart disease patients, referring to Figure 3 , comprising:

[0120] One or more processors;

[0121] A memory having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, so that the one or more processors implement the coronary risk prediction method for coronary heart disease patients.

[0122] In this embodiment, the target patient multi-source heterogeneous data acquisition is aimed at clinical data acquisition, and the patient's electronic health record is extracted from the structured database of the hospital information system, including:

[0123] Disease diagnosis items and disease course description text in medical history records;

[0124] Drug categories and medication cycles in long-term medication lists;

[0125] Intervention treatment methods and postoperative follow-up summaries in surgical records.

[0126] Accessing coronary CT angiography raw image files via a picture archiving and communication system, including a complete heart scan sequence, and integrating biomarker detection results from laboratory test reports;

[0127] Genomics data access obtains patient genetic sequencing reports from genetic testing institutions through encrypted transmission protocols, and parses single nucleotide polymorphism site information and gene annotations in the reports;

[0128] During dynamic physiological monitoring, a Bluetooth low-energy communication link is established with a medical-grade electrocardiogram monitor, and real-time continuous electrocardiogram waveforms and blood oxygen saturation raw signals are received and cached. An automatic data stream recovery mechanism is started for interrupted data streams.

[0129] For hierarchical feature parameter extraction, including clinical data processing and text feature extraction, unstructured text in electronic medical records is input into a pre-trained medical semantic understanding model:

[0130] Identify and encode myocardial ischemia-related symptom description entities, extract semantic features of angina attack frequency and duration, and output pathological feature vector representation

[0131] Coronary image analysis includes vessel reconstruction and calcification recognition. The CT image sequence is isotropically reconstructed, and the coronary vessel tree is automatically segmented in three-dimensional space. The continuous region of the vessel wall CT value exceeding the calcification threshold is detected, and the proportion of the calcification region volume to the total vessel segment volume is calculated.

[0132] When quantifying plaque features, use a multi-layer segmentation network to distinguish plaque subtypes, including non-calcified / mixed / low-density. Calculate the relative area ratio of the plaque in the blood vessel cross-section based on its three-dimensional morphology, and mark the morphological features of vulnerable plaques, such as surface irregularity.

[0133] Process genomic data to model genetic risk, match patient gene sites with coronary heart disease genetic databases, aggregate multi-site risk effect values based on population frequency corrected weight coefficients, and output standardized genetic risk index;

[0134] For dynamic physiological data processing, analyze electrocardiogram signals, segment continuous electrocardiogram data by fixed time, calculate the fluctuation dispersion degree of adjacent heartbeat intervals as an autonomic nervous function indicator, identify motion artifacts in the signal and perform data repair;

[0135] Abnormal monitoring of blood oxygen, detecting the duration of blood oxygen saturation below the clinical warning line, counting the frequency of abnormal events per unit time, and calculating the coefficient of variation of blood oxygen fluctuation amplitude.

[0136] For three-stage cascading dynamic risk assessment, the first stage is feature importance screening, which receives all hierarchical feature parameters and classifies them by source, evaluates the correlation strength of the features and coronary events through a tree structure model, and retains a subset of features with high correlation strength;

[0137] The second stage is time trend modeling, which constructs a time series matrix for dynamic physiological features, uses a recurrent network structure with memory units to capture the continuous change pattern of indicators such as heart rate variability, automatically identifies decaying trajectories with continuously decreasing variability, and outputs mathematical representations of decaying trends with dynamic weighting of key time nodes through attention mechanisms;

[0138] The third stage is multi-modal feature fusion for feature interaction processing, generating cross-combination items of clinical features and genetic features, and calculating a multiplicative compound factor of calcification indicators and genetic risk;

[0139] An integrated decision mechanism is used to input static features, time trends, and interaction items into multiple decision trees, and a 0-100 score is generated based on tree model voting. When the compound factor exceeds the pre-set threshold, the risk level promotion rule is activated.

[0140] For dynamic risk atlas generation for risk visualization and early warning, risk scores are mapped to three-color warning levels (red / yellow / green), and by drawing a trend curve of risk scores over time, labeling the time nodes corresponding to clinical intervention events, and highlighting the critical period of risk level transition;

[0141] For the construction of three-dimensional coronary models, CT image reconstruction results and plaque localization information are fused, and the spatial distribution of different plaque types is rendered with a color gradient. The coronary anatomical structure atlas can be interactively rotated by superimposing the outline of calcified areas on the surface of the vascular model;

[0142] For high-risk factor contribution analysis, model interpretability techniques are used to calculate feature contribution values, and a heat map is generated by sorting the contribution values. For plaque characteristics and genetic factors, the contribution is highlighted, and the following is displayed simultaneously on the side of the three-dimensional model: real-time data dashboard of key physiological indicators, genetic risk index level identification, early warning push mechanism, pop-up warning sent to the attending physician terminal when the risk level enters the high-risk interval, and generation of a summary report containing core risk factors, and personalized intervention recommendations are pushed;

[0143] For drug adjustment schemes and dose optimization suggestions, including image review time window planning and specialist clinic appointment recommendation.

[0144] The incremental learning process of the system's continuous optimization mechanism of the present embodiment includes regularly receiving new patient data and initiating model updates, using an anti-forgetting learning strategy to retain historical knowledge, and triggering full model retraining when model performance fluctuations exceed the tolerance threshold;

[0145] The quality control module is used to perform integrity check on the original data, discard physiological signal segments with serious motion artifacts, exclude CT image blur frames caused by respiratory motion, and establish a failure retry mechanism for the feature extraction process.

[0146] The above design ensures clinical practicability:

[0147] The fusion depth integrates multi-source genomic static risk and physiological dynamic changes to capture multi-dimensional representations of coronary lesions.

[0148] The progressive attenuation pattern of time-sensitive indicators such as heart rate variability is used to warn of potential functional deterioration.

[0149] The interpretable output performs three-dimensional visualization positioning of high-risk plaques, and a heat map quantifies the contribution of each factor to assist clinical decision-making.

[0150] A closed-loop intervention program is formed, and the risk score is directly linked to treatment recommendation generation and medical resource scheduling to form a prevention and control closed loop.

[0151] This embodiment demonstrates that through the landing of each technology, a closed loop from data acquisition → feature processing → model decision → clinical output is realized. Under the premise of excluding diagnostic methods, quantitative features (such as 130HU calcification threshold), dynamic mechanisms (such as blood oxygen retransmission), and clinical embedding (automatic medical order generation) are used to construct a complete technical solution with patent eligibility.

[0152] In the several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. The division of units is only a logical function division. In actual implementation, another division mode can be used, for example, a plurality of units or components can be combined or integrated into another system, or some features can be omitted or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, which can be electrical, mechanical or other forms.

[0153] In addition, each functional unit in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically, or two or more units can be integrated into one unit. The above integrated unit can be realized in the form of hardware or in the form of a software functional unit. The above is only an embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structure or equivalent process conversion using the contents of the specification and drawings, or direct or indirect application in other related technical fields, is also included in the patent protection scope of the present application.

[0154] The specific embodiments of the present application have been described above in detail, but the application is not limited to the above-described specific embodiments. Any equivalent modifications or substitutions made to the application by those skilled in the art are also within the scope of the present application, and thus, equivalent transformations, modifications, improvements, etc. made without departing from the spirit and principle of the present application should be encompassed within the scope of the present application.

Claims

1. A method of predicting coronary risk in a patient with coronary heart disease, characterized in that, The method comprises the following steps: S1, determining a target patient, collecting multi-source heterogeneous risk basic data of the target patient from a multi-source database; S2, obtaining layered feature parameters of each multi-source heterogeneous risk basic data based on different processing models of the multi-source heterogeneous risk basic data; S3, inputting the layered feature parameters into a three-level cascade integrated model to output a dynamic risk prediction score of the target patient; S4, generating a dynamic risk trend graph and a high-risk factor contribution degree heat map based on the dynamic risk prediction score, and pushing an early warning signal in real time through a medical terminal; The S3 further comprises: inputting the layered feature parameters output by S2 into the three-level cascade integrated model, wherein the first level uses XGBoost to perform feature importance screening, the second level uses a bidirectional LSTM network with a fusion attention mechanism to process time sequence features, and the third level generates a dynamic risk prediction score by integrating static and dynamic features through a random forest; In the three-level cascade integrated model of S3, a feature interaction module is introduced, and an interaction term of clinical features and genomic features is generated through an explicit feature cross network, and the interaction term and the original features are input into a bidirectional LSTM network; The second level processing of S3 comprises arranging heart rate variability data of 7 consecutive days in time sequence, automatically identifying a decay trend of the variability degree lasting more than 5 days, and inputting the decay slope as an independent risk factor into the third level model; The third level processing of S3 comprises multiplying a genetic risk index and a calcification volume percentage to generate a composite index, and automatically upgrading the final risk level when the composite index is greater than a threshold value; The first level uses gradient boosting decision trees to screen the top 30% of important features; The second level uses a recurrent neural network with time sequence memory function to analyze the trend of dynamic physiological characteristics over time; The third level fuses static features and trend analysis results to generate a risk score of 0-100 points through multiple decision trees voting; When the risk is visualized and output, the risk score is converted into a high-risk, medium-risk and low-risk three-color early warning level; A three-dimensional model of the coronary artery and a visual atlas of the superposition of the plaque position are generated, and the area and position coordinates of the calcified region are labeled.

2. The method of predicting coronary risk in a patient with coronary heart disease according to claim 1, wherein, The S1 further comprises: real-time acquisition of clinical data, genomics data and dynamic physiological monitoring data of the target patient from a hospital information system, a genetic database and a wearable device, wherein the above-mentioned clinical data at least includes electronic medical records EMR, coronary CT angiography CCTA images and laboratory test reports; The S2 further comprises: extracting key pathological features from unstructured text in the clinical data using a BERT model, and automatically quantifying coronary calcification score CACS and plaque characteristics from the CCTA image through a 3D convolutional neural network; calculating the genetic risk value of coronary heart disease from the genomics data through a polygenic risk score PRS model; extracting heart rate variability HRV attenuation indicators from the dynamic physiological data through a time sequence anomaly detection algorithm.

3. The method of claim 2, wherein the method is for predicting the coronary risk of a patient with coronary heart disease. The extraction of the plaque characteristics in S2 comprises: using a U-Net++ segmentation model to identify non-calcified plaques, mixed plaques and low-density plaques in the CCTA image; A ratio of plaque volume to vessel area, PV / VA, is calculated as a high-risk plaque quantification index; The calcium region identification of S2 includes: The intima boundary of the vessel is automatically outlined on the CT image cross-section; A pixel region with a CT value greater than 130 HU is detected and marked as a calcified plaque; The total volume of the calcified plaque is calculated by continuously scanning along the vessel axis.

4. The method of claim 2, wherein the method is for predicting the coronary risk of a patient with coronary heart disease. The PRS model of S2 is constructed, further comprising: SNP sites related to coronary heart disease are extracted from the UK Biobank database, and LASSO regression is used to screen significant sites; Based on the allele frequency correction weight of the target population, a patient-specific PRS value is generated; The S2 further includes the generation of a genetic risk index, and 20 gene sites strongly related to coronary heart disease in the target population are preferentially screened; A weighting coefficient is set according to the multiple of the risk of coronary heart disease increased by each site; The final index is obtained by accumulating the weighting values of all high-risk sites carried by the patient.

5. The method of claim 2, wherein the method is for predicting the coronary risk of a patient with coronary heart disease. The attention mechanism of S3 is specifically implemented as: A time attention layer is added to the output layer of the bidirectional LSTM to dynamically weight the physiological monitoring data at different time points; The correlation between attention weight and risk score is optimized through back propagation; Before executing S3, it further includes: Abnormal peaks in ECG data caused by motion interference are smoothed and filtered; Fuzzy frames caused by respiratory motion artifacts in CT images are deleted.

6. The method of claim 2, wherein the method is for predicting the coronary risk of a patient with coronary heart disease. The method further includes: A model continuous learning step, when the amount of new patient data reaches a threshold, an incremental learning process is started, and an elastic weight consolidation (EWC) algorithm is used to prevent model catastrophic forgetting; A model online updating step, every month, the data of newly diagnosed coronary heart disease patients in the hospital are received, only the third level decision tree model is retrained and the original feature screening rule is retained, and when the accuracy of the validation set changes by more than 5%, the whole model is updated.

7. The method of claim 1, wherein the method is for predicting coronary risk in a patient with coronary heart disease. The generation of the high-risk factor contribution degree heat map of S4 includes: SHAP value is used to explain the model output, calculate the marginal contribution of each feature to the risk score, and sort the absolute values to generate a feature importance map; The top 10% of calcified plaques in volume are highlighted in red, and a numerical dashboard of genetic risk index and dynamic physiological indicators is displayed beside the vessel model.

8. The method of claim 1, wherein the method is for predicting coronary risk in a patient with coronary heart disease. The dynamic physiological monitoring data of S1 includes: Nighttime resting heart rate, QT interval dispersion, and blood oxygen saturation fluctuation standard deviation for 7 consecutive nights are obtained through Kalman filtering for sensor noise suppression; The processing of the dynamic physiological monitoring data further includes counting the number of blood oxygen abnormalities during nighttime sleep, and activating an emergency data retransmission mechanism when the number of single-night abnormalities exceeds 10 times; The dynamic risk prediction score is used to: Generate personalized intervention strategy recommendations, including drug adjustment suggestions and review cycle planning, and automatically book a follow-up visit to a heart specialist by linking with the hospital resource scheduling system; Pop up an intervention priority list on the doctor's workstation, sort it in descending order of risk score, and automatically generate a report draft containing review CT suggestion dates and statin dose adjustment schemes.

9. A coronary risk prediction system for a patient with coronary heart disease, characterized in that, It includes: One or more processors; A memory having one or more programs stored thereon, wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to carry out the method of predicting coronary risk of a coronary heart disease patient according to any one of claims 1 to 8.

Images