A Cardiovascular Disease Assessment and Management Method and System Based on Big Data

Abstract

The present invention discloses a method and system for cardiovascular disease assessment and management based on big data, relating to the field of big data, including: firstly, obtaining the original medical record data of the target patient and performing standardized processing to obtain the medical record data to be evaluated, comparing and analyzing it with the desensitized and archived medical record data to obtain an evaluation result, and then selecting the corresponding intervention strategy from a preset database accordingly, so as to realize the scientific assessment and effective management of cardiovascular diseases based on big data.

Description

Technical Field

[0001] The present invention relates to the field of big data, and in particular, to a method and system for cardiovascular disease assessment and management based on big data. Background Art

[0002] Cardiovascular diseases are one of the main diseases seriously threatening human health, and accurate assessment and effective management are crucial. Traditional assessments mostly rely on doctors' experience and limited case analysis, which have subjectivity and limitations. With the development of medical big data, using it for cardiovascular disease assessment and management has become a trend. However, existing related technologies still need to be improved in aspects such as data processing, assessment accuracy, and pertinence of intervention strategies. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for cardiovascular disease assessment and management based on big data.

[0004] In a first aspect, an embodiment of the present invention provides a method for cardiovascular disease assessment and management based on big data, including:

[0005] Obtaining the original medical record data of a target patient;

[0006] Performing data standardization processing on the original medical record data to obtain medical record data to be evaluated;

[0007] Comparing and analyzing the medical record data to be evaluated with at least two desensitized archived medical record data to obtain a cardiovascular disease assessment result corresponding to the medical record data to be evaluated;

[0008] According to the cardiovascular disease assessment result, selecting a corresponding cardiovascular disease intervention strategy in a preset cardiovascular disease intervention database.

[0009] In a second aspect, an embodiment of the present invention provides a server system, including a server, and the server is used to execute the method described in the first aspect.

[0010] Compared with the prior art, the beneficial effects provided by the present invention include: adopting a method and system for cardiovascular disease assessment and management based on big data disclosed by the present invention, by obtaining the original medical record data of a target patient and performing standardization processing to obtain medical record data to be evaluated, comparing and analyzing it with desensitized archived medical record data to obtain an assessment result, and then selecting a corresponding intervention strategy from a preset database accordingly, realizing scientific assessment and effective management of cardiovascular diseases based on big data. Brief Description of the Drawings

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0012] Figure 1 It is a schematic flowchart of the steps of the method for evaluating and managing cardiovascular diseases based on big data provided by the embodiments of the present invention;

[0013] Figure 2 It is a schematic block diagram of the structure of the computer device provided by the embodiments of the present invention. Detailed implementation manners

[0014] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. Usually, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0015] The following will describe the detailed implementation manners of the present invention in conjunction with the accompanying drawings.

[0016] To solve the technical problems in the foregoing background art, Figure 1 It is a schematic flowchart of the method for evaluating and managing cardiovascular diseases based on big data provided by the embodiments of the present disclosure. The method for evaluating and managing cardiovascular diseases based on big data will be introduced in detail below.

[0017] Step S201: Obtain the original medical record data of the target patient;

[0018] Step S202: Perform data standardization processing on the original medical record data to obtain the medical record data to be evaluated;

[0019] Step S203: Compare and analyze the medical record data to be evaluated with at least two desensitized and archived medical record data to obtain the cardiovascular disease evaluation result corresponding to the medical record data to be evaluated;

[0020] Step S204: Select the corresponding cardiovascular disease intervention strategy in the preset cardiovascular disease intervention database according to the cardiovascular disease evaluation result.

[0021] In an embodiment of the present invention, by way of example, there is a patient named Mr. Li who came to the doctor due to palpitations and chest tightness. After the doctor's preliminary examination of Mr. Li, the relevant information was entered into the hospital's electronic medical record system. At this time, the server would obtain Mr. Li's original medical record data from this electronic medical record system. These data include Mr. Li's basic personal information (such as age, gender, past medical history, etc.), the description of the symptoms during this visit (such as the frequency of palpitations, the degree of chest tightness, etc.), the results of various preliminary examinations (such as blood pressure, heart rate, etc.), and some relevant examination reports done before (such as the text descriptions corresponding to imaging data such as electrocardiograms and cardiac ultrasounds). After the server obtains Mr. Li's original medical record data, it begins to perform data standardization processing. Since there may be differences in formats, expressions, etc. when different departments and different doctors enter medical record data, the server needs to standardize these data. For example, for Mr. Li's blood pressure data, some doctors may record it as "120 / 80 mmHg", while some may write it as "12080", and the server will standardize it to "120 / 80 mmHg" according to the unified format standard. Another example is that for the symptom description part, some may be expressed in a more colloquial way, and the server will convert it into a standardized expression in medical terms. After such a series of processes, Mr. Li's medical record data becomes the medical record data to be evaluated that conforms to the standard format and standardized expression, so as to facilitate accurate analysis and comparison later. The server first extracts cardiovascular disease-related images from Mr. Li's medical record data to be evaluated, such as the electrocardiogram and cardiac ultrasound images done before. Suppose the server divides the electrocardiogram image into one data set and the cardiac ultrasound image into another data set. Then it obtains the target cardiovascular disease-related images from these data sets, such as selecting the electrocardiogram image during the current symptom attack as the target cardiovascular disease-related image, and extracts image features from it to obtain the first medical record image features, such as waveform features and heart rate change features on the electrocardiogram. At the same time, the server obtains the first medical record document data corresponding to Mr. Li's medical record data to be evaluated, that is, the text description part of the doctor's description of his condition. Semantic description features are extracted from these document data to obtain the first semantic description features, such as extracting semantic features regarding the severity of symptoms, the frequency of onset, etc. Then, the server obtains the first multi-dimensional features based on the first medical record image features and the first semantic description features, which include the target image features and the first semantic description features. Then, the server obtains the pre-established image feature matching pool and semantic description feature matching pool. The image feature matching pool records the mutual correlation relationships between each medical record data to be determined and its corresponding image features, and the semantic description feature matching pool is the same. The server matches these two matching pools respectively according to the target image features and also matches these two matching pools according to the first semantic description features, so as to obtain at least two target medical record data to be determined that match Mr. Li's medical record data to be evaluated.For example, the medical record data of several previous patients with the same symptoms of palpitation and chest tightness and similar relevant imaging features and semantic description features are matched as the target pending medical record data. For at least two target pending medical record data obtained by each matching method, the server calculates the matching coefficient between Mr. Li's pending medical record data to be evaluated and each target pending medical record data. For example, the matching coefficient is calculated by comparing aspects such as the similarity of imaging features and the fit of semantic description features. Then, for each target pending medical record data, the server calculates the arithmetic mean of its matching coefficients with the pending medical record data to be evaluated in each matching method. Based on this arithmetic mean, the de-identified archived medical record data is determined from at least two target pending medical record data. Suppose the de-identified archived medical record data of three patients is finally determined. These three patients have all undergone a clear diagnosis of cardiovascular disease and their medical record data is complete. The server constructs the disease correlation information based on the second multi-dimensional features corresponding to each de-identified archived medical record data, the cardiovascular disease target values configured for each de-identified archived medical record data, and the first multi-dimensional features of Mr. Li's pending medical record data to be evaluated. For example, the electrocardiogram imaging features, symptom description semantic features, etc. of Mr. Li are comprehensively correlated and analyzed with the corresponding features in the medical record data of the three determined de-identified archived medical record data to construct the disease correlation information. Then, based on this disease correlation information, the cardiovascular disease target value of Mr. Li's pending medical record data to be evaluated is identified to obtain the cardiovascular disease assessment result. If most of the patients in the medical record data of the three de-identified archived medical record data are diagnosed with coronary heart disease and the disease severity has different grades, through comparative analysis, the server may obtain the assessment result that Mr. Li has coronary heart disease and the disease is at a certain level. After the server obtains the assessment result that Mr. Li has coronary heart disease and the disease is at a certain level, it will select the corresponding intervention strategy in the preset cardiovascular disease intervention database. For example, for coronary heart disease patients with Mr. Li's disease severity, the intervention database may recommend first undergoing drug treatment, including taking a certain antiplatelet drug, lipid-regulating drug, etc., and cooperating with regular reexaminations (such as reexamining electrocardiogram, echocardiogram, etc. every month), and at the same time suggesting that Mr. Li adjust his lifestyle, such as reducing the intake of greasy food and increasing appropriate exercise. The server pushes these intervention strategies to the doctor so that the doctor can formulate a more appropriate treatment plan for Mr. Li.

[0022] In the embodiment of the present invention, the comparison and analysis of the pending medical record data to be evaluated with at least two de-identified archived medical record data to obtain the cardiovascular disease assessment result corresponding to the pending medical record data to be evaluated can be implemented through the following example.

[0023] Obtain the pending medical record data to be evaluated, and extract multi-dimensional features from the pending medical record data to be evaluated to obtain the first multi-dimensional features, where the first multi-dimensional features include at least two data dimension features corresponding to at least two data dimensions respectively;

[0024] Match at least two de-identified archived medical record data that match the to-be-evaluated medical record data according to the first multi-dimensional feature, and each of the de-identified archived medical record data is configured with a cardiovascular disease target value;

[0025] Construct disease condition association information according to the second multi-dimensional feature corresponding to each of the de-identified archived medical record data, the cardiovascular disease target value configured for each of the de-identified archived medical record data, and the first multi-dimensional feature of the to-be-evaluated medical record data;

[0026] Identify the cardiovascular disease target value of the to-be-evaluated medical record data according to the disease condition association information to obtain a cardiovascular disease evaluation result.

[0027] In an embodiment of the present invention, exemplarily, the server has completed the standardization process of Mr. Li's original medical record data and obtained the medical record data to be evaluated. Now, the server needs to perform multi-dimensional feature extraction on this medical record data to be evaluated. From the data dimension perspective, on the one hand, it is the basic information dimension of Mr. Li. The server extracts features such as age and gender from it. For example, Mr. Li is 55 years old this year and his gender is male. These features have certain reference value for the evaluation of cardiovascular diseases because people of different ages and genders may have different risks and manifestations of cardiovascular diseases. On the other hand, it is the symptom description dimension. The server carefully analyzes the specific conditions of Mr. Li's palpitations and chest tightness recorded by the doctor, including features such as the attack frequency, the duration of each attack, and whether there are other accompanying symptoms (such as chest pain, difficulty breathing, etc.). For example, Mr. Li's palpitations symptom occur 2-3 times a week on average, each lasting about 10-15 minutes, and occasionally accompanied by mild chest pain. There is also the examination result dimension. The server focuses on various cardiovascular-related examination data of Mr. Li. Taking blood pressure as an example, Mr. Li's blood pressure value is 130 / 85 mmHg, which reflects the pressure state of his cardiovascular system. Looking at the electrocardiogram examination results again, the server extracts waveform features on the electrocardiogram, such as whether there are ST-segment changes, T-wave inversion, etc. These are all important clues for judging cardiovascular diseases. Through the analysis and extraction of these different data dimensions, the server obtains the first multi-dimensional feature, which covers data dimension features in multiple aspects and lays a foundation for subsequent matching and analysis with the de-identified archived medical record data. After the server has the first multi-dimensional feature of Mr. Li's medical record data to be evaluated, it begins to perform matching searches in the huge de-identified archived medical record database. This database stores a large amount of medical record data of past patients after de-identification processing, and each medical record data is configured with corresponding cardiovascular disease target values, that is, information such as the disease condition and degree after a clear diagnosis. The server first compares each data dimension feature in Mr. Li's first multi-dimensional feature with the medical record data in the database one by one. For example, in the age dimension, the server will screen out the medical records of patients with similar ages because people with similar ages may have more similarities in cardiovascular health conditions. For Mr. Li's situation of being 55 years old, the server may first lock in the medical record data of patients between 45 and 65 years old as the preliminary matching range. In the symptom description dimension, the server will search for medical records that also have palpitations and chest tightness symptoms, and whose attack frequency, duration, and accompanying symptoms are similar. Suppose there is a medical record of Ms. Wang in the database. She also often has palpitations, with a similar attack frequency per week as Mr. Li, a similar duration of each attack, and occasionally accompanied by mild chest pain. Then Ms. Wang's this medical record data will be preliminarily matched in this dimension. In the examination result dimension, the server compares examination data such as blood pressure and electrocardiogram.If in the medical record of another Mr. Zhang, the blood pressure value is relatively close to that of Mr. Li, and there is also a similar ST-segment change on the electrocardiogram, then the medical record data of Mr. Zhang will also fall into the scope of further matching. Through the comprehensive comparison and screening of the characteristics of multiple data dimensions, the server finally determines at least two desensitized archived medical record data with a relatively high matching degree to the medical record data of Mr. Li to be evaluated. For example, in addition to the medical records of Ms. Wang and Mr. Zhang mentioned above, the medical record of a Mr. Zhao is also matched. Their medical record data are somewhat similar to Mr. Li's in each dimension, and these desensitized archived medical record data are clearly marked with their corresponding cardiovascular disease target values. For example, Ms. Wang is diagnosed with mild coronary heart disease, Mr. Zhang with moderate coronary heart disease, and Mr. Zhao with early myocardial infarction, etc. After the server determines the desensitized archived medical record data of Ms. Wang, Mr. Zhang, Mr. Zhao, etc. that match the medical record data of Mr. Li to be evaluated, the next step is to construct disease-related information. For the desensitized archived medical record data of Ms. Wang, the server extracts its corresponding second multi-dimensional characteristics. This includes the dimensional characteristics of Ms. Wang's basic information, such as her age is 52 years old and her gender is female; the dimensional characteristics of symptom description, such as the specific feelings and attack frequencies when her palpitation symptoms occur; the dimensional characteristics of examination results, such as her blood pressure value and electrocardiogram characteristics. At the same time, combining the cardiovascular disease target value of Ms. Wang being diagnosed with mild coronary heart disease and the first multi-dimensional characteristics of Mr. Li's medical record data to be evaluated (such as Mr. Li's age, symptoms, examination results, etc.), the server begins to construct the associated information. The server will analyze that although there are differences in age between Mr. Li and Ms. Wang, they are also within a certain range, and how this age difference may affect the development of the disease. In terms of symptoms, compare the similarities and differences in symptoms such as palpitation and chest tightness between the two, and the relationship between these symptoms and the diagnosis of coronary heart disease. For the examination results, compare the similarities and differences in data such as blood pressure and electrocardiogram between the two, and how these data reflect the condition of cardiovascular diseases. Through such a comprehensive analysis and synthesis, the server constructs a part of the disease-related information for the medical record data of Mr. Li and Ms. Wang. Similarly, for the desensitized archived medical record data of Mr. Zhang and Mr. Zhao, the server also extracts their second multi-dimensional characteristics respectively, combines their respective cardiovascular disease target values and Mr. Li's first multi-dimensional characteristics, and conducts similar comprehensive analysis and synthesis. For example, for Mr. Zhang's moderate coronary heart disease situation, the server will focus on analyzing the association between the degree of his disease and Mr. Li's current condition, and the impact of the differences in examination results on the disease judgment, etc. Finally, the server integrates the disease-related information constructed for the medical record data of different patients such as Ms. Wang, Mr. Zhang, and Mr. Zhao to form a complete disease-related information, which comprehensively covers the association situation between Mr. Li and these matching desensitized archived medical record data in each dimension and the corresponding cardiovascular disease target values, etc.After the server constructs the complete disease-related information, it uses this as a basis to identify the cardiovascular disease target values of Mr. Li's medical record data to be evaluated, thereby obtaining the cardiovascular disease assessment result. The server will carefully analyze the comparison of Mr. Li with other patients such as Ms. Wang, Mr. Zhang, and Mr. Zhao in various dimensions in the disease-related information and the corresponding cardiovascular disease target values. For example, from the age dimension, Mr. Li is slightly older than Ms. Wang, which may mean that he has a certain potential risk increase in the development of cardiovascular diseases; from the symptom dimension, the attack frequency and duration of Mr. Li's symptoms of palpitation and chest tightness may be within a certain specific range compared with those of Ms. Wang, Mr. Zhang, etc., which has certain reference value for judging the degree of the disease; from the examination result dimension, the similarity and difference between Mr. Li's blood pressure value, electrocardiogram characteristics, etc. and those of other patients can also reflect his cardiovascular disease status. Considering these factors comprehensively, the server combines the cardiovascular disease target values such as Ms. Wang being diagnosed with mild coronary heart disease, Mr. Zhang with moderate coronary heart disease, and Mr. Zhao with early myocardial infarction, as well as their various associations with Mr. Li in the disease-related information. After complex data analysis and judgment, the server finally obtains the cardiovascular disease assessment result of Mr. Li. Suppose after analysis, the server determines that Mr. Li is more likely to have moderate coronary heart disease. Then this "having moderate coronary heart disease" is the cardiovascular disease assessment result corresponding to Mr. Li's medical record data to be evaluated. In this way, through a series of detailed comparisons, analyses, and comprehensive judgments, the server has completed the transformation process from the medical record data to be evaluated to the cardiovascular disease assessment result in a scientific and accurate manner, providing an important basis for subsequent treatment and intervention.

[0028] In an embodiment of the present invention, the first multi-dimensional feature includes a first medical record image feature, and the second multi-dimensional feature includes a second medical record image feature; the construction of the disease-related information based on the second multi-dimensional feature corresponding to each of the desensitized and archived medical record data, the cardiovascular disease target values configured for each of the desensitized and archived medical record data, and the first multi-dimensional feature of the medical record data to be evaluated can be implemented through the following examples.

[0029] Perform feature mapping processing on the first medical record image feature to obtain the first disease semantic description information, and perform feature mapping processing on the second medical record image feature corresponding to each of the desensitized and archived medical record data to obtain the second disease semantic description information corresponding to each of the desensitized and archived medical record data;

[0030] Obtain the first medical record document data corresponding to the medical record data to be evaluated and the second medical record document data corresponding to each of the desensitized and archived medical record data;

[0031] Construct the disease association information based on the second disease semantic description information, the second medical record document data, and the cardiovascular disease target value corresponding to each of the desensitized archived medical record data, as well as the first disease semantic description information and the first medical record document data.

[0032] In an embodiment of the present invention, exemplarily, for the medical record data to be evaluated of Mr. Li, which includes cardiovascular disease-related images such as electrocardiograms and cardiac ultrasounds, the first medical record image features extracted by the server are very crucial. Taking Mr. Li's electrocardiogram image as an example, the server first needs to perform feature mapping processing on these first medical record image features. Through the built-in professional image analysis algorithms and models, it converts various waveform features, time interval features, etc. on the electrocardiogram. For example, the morphological and amplitude features of waveforms such as the P wave, QRS complex, and T wave on a normal electrocardiogram, as well as the time intervals between them, after feature mapping processing, will be transformed into descriptive information with semantic meanings. Suppose after processing, the server maps the image feature that the P wave morphology in Mr. Li's electrocardiogram is slightly wider to the first disease semantic description information such as "there may be a slight abnormality in the atrial depolarization process", and maps the feature that the QRS complex duration is normal but the voltage is slightly lower to semantic descriptions similar to "the overall ventricular depolarization is normal but the intensity of myocardial electrical activity may be slightly weaker". Similarly, for the de-identified archived medical record data of patients such as Ms. Wang, Mr. Zhang, and Mr. Zhao, the server will also perform similar feature mapping processing on their respective second medical record image features. Taking Ms. Wang's cardiac ultrasound image as an example, the server analyzes second medical record image features such as the size of the heart chambers, the thickness of the ventricular walls, and the blood flow conditions in the image. For example, if Ms. Wang's cardiac ultrasound shows that the thickness of the left ventricular wall is slightly thicker than the normal range, after feature mapping processing, the server will obtain the second disease semantic description information such as "the myocardium of the left ventricle may have compensatory thickening, indicating a possible potential increase in cardiovascular burden". For Mr. Zhang's electrocardiogram image, if there is a feature of mild ST segment depression, after processing, it will be mapped to the second disease semantic description information of "there may be local myocardial ischemia". Through such feature mapping processing, the server obtains corresponding disease semantic description information for each medical record data, and these information can more intuitively reflect the cardiovascular disease-related conditions implied by the image features. After the server completes the processing of the above image features, it will then obtain relevant medical record document data. For the medical record data to be evaluated of Mr. Li, the server accurately extracts the corresponding first medical record document data from the electronic medical record system. These document data are the text contents of the doctor's detailed records of his condition during the consultation of Mr. Li. It includes Mr. Li's chief complaint, that is, the specific feelings, attack frequencies, durations, etc. of the symptoms he described himself, such as palpitations and chest tightness; the doctor's physical examination results, such as whether there are murmurs when auscultating the heart and whether the heart rate is regular; and the doctor's preliminary diagnostic impressions based on the initial examination and clinical experience, etc. Similarly, the server will also obtain the second medical record document data corresponding to the de-identified archived medical record data of patients such as Ms. Wang, Mr. Zhang, and Mr. Zhao from the storage system.Taking Ms. Wang as an example, the data in her second medical record document records similar information when she visited the doctor. For example, when she described the symptom of palpitation, there was a slight sense of dizziness, and the attack frequency was slightly lower than that of Mr. Li. The doctor's physical examination found that there was a slight murmur during her heart auscultation, and other situations, as well as the detailed diagnosis process and basis for the final diagnosis of mild coronary heart disease. By obtaining this detailed medical record document data, the server is fully prepared for the subsequent construction of comprehensive and accurate disease condition correlation information. After the server has all the above necessary data, it begins to construct the disease condition correlation information. Taking Ms. Wang's desensitized and archived medical record data as an example, the server first combines her second disease condition semantic description information "There may be compensatory thickening of the left ventricular myocardium, indicating a possible potential increase in cardiovascular burden", the second medical record document data (including symptom description, physical examination results, diagnosis process, etc.) and her cardiovascular disease target value (mild coronary heart disease), and then comprehensively compares and analyzes it with Mr. Li's first disease condition semantic description information (such as the semantic description after the mapping of the electrocardiogram image features mentioned above) and the first medical record document data (Mr. Li's symptoms, physical examination results, etc.). The server will analyze the similarities and differences between Ms. Wang and Mr. Li in terms of symptoms. For example, Ms. Wang has palpitation accompanied by slight dizziness, while Mr. Li has palpitation occasionally accompanied by slight chest pain. These two different accompanying symptoms may imply different development directions of cardiovascular diseases. In terms of physical examination results, Ms. Wang has a slight murmur during heart auscultation, and Mr. Li may have a slightly irregular heart rate. These differences and similarities are also important for judging the cardiovascular disease condition. At the same time, by comparing the disease condition semantic description information of the two, the compensatory thickening of the left ventricular myocardium in Ms. Wang and the myocardial electrical activity intensity situation implied by Mr. Li's electrocardiogram image features reflect the state of the cardiovascular system from different angles. By analyzing these associations, the similarities and differences in the disease conditions of the two can be understood more deeply. For the desensitized and archived medical record data of Mr. Zhang and Mr. Zhao, the server will also follow the same method to comprehensively compare and analyze their respective second disease condition semantic description information, second medical record document data and cardiovascular disease target values with Mr. Li's first disease condition semantic description information and first medical record document data. For example, for Mr. Zhang, combining his second disease condition semantic description information "There may be local ischemia in the myocardium", the second medical record document data (such as the symptom attack situation, physical examination results, etc.) and the cardiovascular disease target value (moderate coronary heart disease), and comparing it with Mr. Li's relevant data, analyzing the differences between the two in terms of myocardial ischemia and the impact of this difference on the disease condition judgment, etc. Finally, the server will integrate all the association information obtained from the comprehensive comparison and analysis of the medical record data of different patients such as Ms. Wang, Mr. Zhang and Mr. Zhao with Mr. Li's medical record data to construct a complete disease condition correlation information.This disease-related information covers the similarities and differences among patients in terms of symptoms, physical examination results, semantic descriptions of the diseases implied by imaging features, etc., as well as the corresponding cardiovascular disease target values, providing an important basis for accurately identifying the cardiovascular disease target values of the medical record data to be evaluated subsequently.

[0033] In an embodiment of the present invention, the number of the first medical record imaging features is at least two; obtaining the first disease semantic description information by performing feature mapping processing on the first medical record imaging features can be implemented through the following examples.

[0034] Performing a feature integration operation on at least two of the first medical record imaging features to obtain a target imaging feature;

[0035] Converting the target imaging feature to the semantic feature domain through a preset feature conversion component to obtain the first disease semantic description information.

[0036] In an embodiment of the present invention, exemplarily, multiple imaging features such as P wave morphology, QRS complex duration and voltage, and T wave morphology are extracted from Mr. Li's electrocardiogram image, and features such as cardiac chamber size and ventricular wall thickness are obtained from the echocardiogram image. The server performs a feature integration operation on these different first medical record imaging features. Through a specific algorithm, it comprehensively processes these scattered imaging features and integrates the information about the cardiovascular system state contained in each of them. Just like combining the features of each wave of the electrocardiogram with the cardiac structure features shown in the echocardiogram, after complex calculations and integrations, a target imaging feature that can comprehensively reflect Mr. Li's cardiovascular condition is obtained. This target imaging feature is no longer an isolated individual imaging feature but an overall feature representation that integrates multiple imaging information. After obtaining the target imaging feature, the server uses the preset feature conversion component for the next step of processing. This feature conversion component is like a translator that can convert the information in the form of data and graphics of the target imaging feature into the semantic feature domain, that is, into a form that can accurately describe the cardiovascular condition in literal language. For example, after being processed by the feature conversion component, the originally complex target imaging feature may be converted into a first disease semantic description information such as "the overall electrical activity and structure of the heart present a certain specific state, and there may be a certain degree of myocardial functional abnormality", thus providing a clearer and more understandable basis for subsequent operations such as constructing disease-related information.

[0037] In an embodiment of the present invention, constructing the disease-related information according to the second disease semantic description information, the second medical record document data, and the cardiovascular disease target value corresponding to each desensitized archived medical record data, as well as the first disease semantic description information and the first medical record document data, can be implemented through the following examples.

[0038] Construct context deduction information based on the first identifier for distinguishing different contents included in the same medical record data, and the second disease semantic description information, the second medical record document data, and the cardiovascular disease target value corresponding to the same de-identified archived medical record data, so as to obtain at least two context deduction information corresponding to at least two de-identified archived medical record data;

[0039] Integrate the at least two context deduction information to obtain context integration information, wherein the at least two context deduction information in the context integration information are divided by a second identifier for distinguishing the contents of different medical record data;

[0040] Construct association information based on the first identifier, the first disease semantic description information, and the first medical record document data;

[0041] Generate the disease association information according to the context integration information and the association information.

[0042] In an embodiment of the present invention, by way of example, in a hospital information management system, the server is responsible for processing a large amount of medical record data for accurate cardiovascular disease assessment. Taking the medical record data of Mr. Li to be evaluated mentioned above and the de-identified archived medical record data of Ms. Wang, Mr. Zhang, Mr. Zhao, etc. that match it as an example to illustrate this step. For each de-identified archived medical record data, there are different types of information. In order to better sort out and utilize this information to construct disease-related information, the server will use a pre-set first identifier. This first identifier is like a label attached to different contents within the medical record data to clearly distinguish them. Taking the de-identified archived medical record data of Ms. Wang as an example, the server first extracts the second disease semantic description information therein, such as "there may be compensatory thickening of the left ventricular myocardium, indicating a possible potential increase in cardiovascular burden", which is obtained by mapping the cardiac ultrasound image features. Then there is the second medical record document data, which details the symptom description (palpitations accompanied by slight dizziness, attack frequency, etc.) when Ms. Wang visited the doctor, the physical examination results (slight murmur during cardiac auscultation, etc.) and the detailed process of finally being diagnosed with mild coronary heart disease. And her cardiovascular disease target value is mild coronary heart disease. The server uses the first identifier to distinguish these different contents. For example, attach a "symptom identifier" to the content of the symptom description part, a "physical examination identifier" to the physical examination result part, a "diagnosis identifier" to the cardiovascular disease target value, and an "image semantic identifier" to the second disease semantic description information. Then, based on these different contents with identifiers, the server starts to construct context deduction information. It will analyze that in Ms. Wang's case, the symptom of palpitations accompanied by slight dizziness with the "symptom identifier", combined with the situation of slight murmur during cardiac auscultation with the "physical examination identifier", and then consider the semantic description of compensatory thickening of the left ventricular myocardium with the "image semantic identifier" and the cardiovascular disease target value of mild coronary heart disease with the "diagnosis identifier", and through logical analysis and medical knowledge, deduce the development of Ms. Wang's condition at different stages, possible influencing factors, etc., so as to construct a context deduction information about Ms. Wang. Similarly, for the de-identified archived medical record data of Mr. Zhang and Mr. Zhao, the server will also follow the above method, use their respective second disease semantic description information, second medical record document data and cardiovascular disease target values, and use the first identifier to distinguish different contents, and then construct context deduction information about Mr. Zhang and Mr. Zhao respectively. In this way, the server obtains at least two context deduction information corresponding to at least two de-identified archived medical record data. After the server obtains the context deduction information about Ms. Wang, Mr. Zhang, Mr. Zhao, etc., next it needs to integrate this information. During the integration process, in order to be able to clearly distinguish the context deduction information of different patients, the server will use a second identifier for distinguishing the content of different medical record data.For example, attach the "Ms. Wang identifier" to the overall context deduction information of Ms. Wang, the "Mr. Zhang identifier" to that of Mr. Zhang, and the "Mr. Zhao identifier" to that of Mr. Zhao. Then, the server combines these context deduction information with different identifiers. Just like putting together different jigsaw puzzle pieces, it arranges and combines this information in a certain order and logic so that they can jointly form a more comprehensive and macroscopic information set, which is the context integration information. In this context integration information, through the second identifiers such as the "Ms. Wang identifier", the "Mr. Zhang identifier", and the "Mr. Zhao identifier", it is very clear which part of the information is about which patient, thus facilitating subsequent analysis and processing. At the same time, the server also needs to construct associated information based on the medical record data to be evaluated for Mr. Li. The server first extracts the first disease semantic description information of Mr. Li, assuming it is "The overall electrical activity and structure of the heart present a certain specific state, and there may be a certain degree of functional abnormality in the myocardium" obtained by mapping the electrocardiogram image features. Then there is the first medical record document data, which contains Mr. Li's symptom descriptions (palpitations, chest tightness, etc.), physical examination results (the heart rate may be slightly irregular, etc.). Similarly, the server uses the first identifier to distinguish these different contents. For example, attach the "symptom identifier" to the symptom description part of Mr. Li, attach the "regularity identifier" to the physical examination result part (here it is assumed to be used to distinguish the physical examination result part to differentiate from the "physical examination identifier" used for other patients' medical records), and attach the "image semantic identifier" to the first disease semantic description information. Then, based on these different contents with identifiers, the server constructs a piece of associated information about Mr. Li by analyzing the relationships between them, such as the association between the symptom description and the physical examination result, the association between the symptom description and the first disease semantic description information, and the association between the physical examination result and the first disease semantic description information. Finally, the server needs to generate disease-associated information based on the already constructed context integration information and associated information. The server will organically combine the context integration information and the associated information. It will compare, analyze, and integrate the information about the disease development, influencing factors, etc. of patients such as Ms. Wang, Mr. Zhang, and Mr. Zhao in the context integration information with the information about Mr. Li's symptoms, physical examination results, disease semantic descriptions, etc. in the associated information.For example, by comparing Ms. Wang's symptoms of palpitation accompanied by mild dizziness with Mr. Li's symptoms of palpitation and chest tightness, analyze the similarities and differences in their symptom manifestations and the possible development directions of cardiovascular diseases they may imply; by comparing the physical examination result of a slight murmur during Ms. Wang's cardiac auscultation with the physical examination result of a slightly irregular heart rate that may exist in Mr. Li, explore the similarities and differences in their physical states of the cardiovascular system; by comparing the semantic description of the compensatory thickening of the left ventricular myocardium in Ms. Wang with the semantic description of a specific state presented by the overall electrical activity and structure of Mr. Li's heart, study the similarities and differences in their internal states of the cardiovascular system. Then, the server will organize the results of these comparisons, analyses, and fusions to form a comprehensive and integrated disease condition correlation information. This disease condition correlation information covers various correlation situations between Mr. Li and other matching patients in terms of symptoms, physical examination results, image semantic descriptions, cardiovascular disease target values, etc., providing a solid foundation for accurately identifying the cardiovascular disease target values of the medical record data to be evaluated in the future.

[0043] In an embodiment of the present invention, the first multi-dimensional feature includes a target image feature and a first semantic description feature; at least two de-identified archived medical record data that match the medical record data to be evaluated can be matched according to the first multi-dimensional feature, and the implementation can be carried out through the following examples.

[0044] Obtain a pre-established image feature matching pool and a semantic description feature matching pool. The image feature matching pool includes the mutual correlation relationship between each pending medical record data and the image feature corresponding to the pending medical record data, and the semantic description feature matching pool includes the mutual correlation relationship between each pending medical record data and the semantic description feature corresponding to the pending medical record data;

[0045] Match the image feature matching pool and the semantic description feature matching pool respectively according to the target image feature, and match the image feature matching pool and the semantic description feature matching pool respectively according to the first semantic description feature to obtain at least two target pending medical record data that match the medical record data to be evaluated;

[0046] Determine the de-identified archived medical record data from the at least two target pending medical record data according to the medical record matching coefficient between the at least two target pending medical record data and the medical record data to be evaluated.

[0047] In an embodiment of the present invention, exemplarily, in the hospital information management system, the server undertakes the tasks of processing and analyzing a large amount of medical record data to accurately evaluate cardiovascular diseases. To achieve efficient and accurate matching of medical record data, the server has pre-established an image feature matching pool and a semantic description feature matching pool. The construction of the image feature matching pool is a long-term and complex process. The server collects a large amount of pending medical record data of past patients, and these pending medical record data cover various examination images related to cardiovascular diseases, such as electrocardiogram, echocardiogram, coronary angiography and other image materials. For each piece of pending medical record data, the server extracts its corresponding image features through professional image analysis algorithms. For example, from a piece of pending electrocardiogram medical record data, the server will extract features such as the morphology, duration, and voltage of the P wave, QRS complex, and T wave; from the pending echocardiogram medical record data, image features such as the size of the heart cavity, the thickness of the ventricular wall, and the blood flow velocity will be extracted. Then, an association relationship is established between each piece of pending medical record data and its corresponding image features and stored in the image feature matching pool. The establishment of the semantic description feature matching pool is the same. The server analyzes and processes the text description part of a large amount of pending medical record data and extracts the semantic description features therein. These text descriptions include the patient's self-reported symptoms (such as the specific feelings, attack frequency, and duration of palpitation and chest tightness), the doctor's physical examination records (such as auscultation results, heart rate, blood pressure, etc.) and the preliminary diagnosis impression and other contents. Through natural language processing technology, the server can extract key semantic description features from them, such as "frequent palpitation accompanied by mild chest pain" and "slightly fast and irregular heart rate". Similarly, after establishing an association relationship between each piece of pending medical record data and its corresponding semantic description features, it is stored in the semantic description feature matching pool. Taking the pending medical record data of Mr. Li mentioned above as an example, the server has extracted the target image features and the first semantic description features from this medical record data. For the target image features, assume that the target image features in Mr. Li's pending medical record data are from a recent electrocardiogram image, and its features include that the P wave morphology is slightly wide, the QRS complex duration is normal but the voltage is slightly low, etc. The server will match these target image features with the data in the image feature matching pool. It will traverse each piece of pending medical record data and its corresponding image features in the image feature matching pool and judge the matching degree by calculating the similarity between the features. For example, for the electrocardiogram image features in a piece of pending medical record data, if its P wave morphology also has a similar slightly wide situation, and the QRS complex duration and voltage are close to Mr. Li's target image features, then this piece of pending medical record data has a high matching degree in terms of image feature matching. At the same time, the server will also match Mr. Li's target image features with the semantic description feature matching pool.Although the semantic description feature matching pool mainly stores semantic description features, due to the possible association between the imaging features and semantic description features of different medical record data, matching can also be performed in this pool through a certain mapping relationship. For example, if the semantic description features of a pending medical record data include a description of "possible abnormalities in cardiac electrical activity", and the corresponding electrocardiogram imaging features are similar to the target imaging features of Mr. Li, then this pending medical record data will also be considered to have a certain degree of matching in this regard. For the first semantic description feature, assume that the first semantic description feature in Mr. Li's medical record data to be evaluated includes content such as "palpitations occur 2-3 times a week, each episode lasts about 10-15 minutes, and is occasionally accompanied by mild chest pain". The server will also match these first semantic description features with the imaging feature matching pool and the semantic description feature matching pool respectively. When matching with the imaging feature matching pool, it will look for pending medical record data in which the symptom descriptions of the patients in the medical record data corresponding to the imaging features are similar to Mr. Li's first semantic description feature. For example, if a patient in a pending medical record data also has palpitations, and the attack frequency, duration, and accompanying symptoms are similar to Mr. Li's, then this pending medical record data will be considered to have a match in this regard. When matching with the semantic description feature matching pool, it will directly compare the similarity of the semantic description features through natural language processing technology to find the pending medical record data that is closest to Mr. Li's first semantic description feature. Through the above two-way matching process in the two matching pools, the server finally obtains at least two target pending medical record data that match Mr. Li's medical record data to be evaluated. For example, it matches the pending medical record data of patients such as Ms. Wang, Mr. Zhang, and Mr. Zhao, and their medical record data have a certain similarity to Mr. Li's medical record data to be evaluated in terms of imaging features and semantic description features. After the server obtains the target pending medical record data of Ms. Wang, Mr. Zhang, and Mr. Zhao that match Mr. Li's medical record data to be evaluated, it will next calculate the medical record matching coefficient to further determine the de-identified archived medical record data. For each target pending medical record data, the server will calculate the medical record matching coefficient between it and Mr. Li's medical record data to be evaluated. The calculation of this matching coefficient is a comprehensive consideration process. Taking the target pending medical record data of Ms. Wang as an example, the server will calculate the matching coefficient from multiple aspects. In terms of imaging features, it will compare the imaging features such as electrocardiogram and echocardiogram in Ms. Wang's pending medical record data with the corresponding imaging features in Mr. Li's medical record data to be evaluated. For example, compare the morphology, duration, voltage, etc. of the P wave, QRS complex, and T wave of the electrocardiogram, and calculate the similarity of the imaging features through a certain mathematical formula (such as the cosine similarity formula, etc.), and use this similarity as part of the matching coefficient in terms of imaging features. In terms of semantic description features, it will compare the symptom descriptions, physical examination records, etc. in Ms. Wang's pending medical record data with the first semantic description feature in Mr. Li's medical record data to be evaluated.For example, by comparing the attack frequency, duration, accompanying symptoms, etc. of symptoms such as palpitation and chest tightness, the similarity of semantic description features is calculated through natural language processing technology, and this similarity is used as part of the matching coefficient for semantic description features. Then, the matching coefficient for image features and the matching coefficient for semantic description features are combined according to a certain weight (assuming the weight of the matching coefficient for image features is 0.6 and the weight of the matching coefficient for semantic description features is 0.4) to obtain the medical record matching coefficient between the target pending medical record data of Ms. Wang and the medical record data to be evaluated of Mr. Li. Similarly, the server will perform similar calculations on other target pending medical record data such as Mr. Zhang and Mr. Zhao to obtain the medical record matching coefficients between their respective data and the medical record data to be evaluated of Mr. Li. Finally, based on these medical record matching coefficients, the de-identified archived medical record data is determined from at least two target pending medical record data such as Ms. Wang, Mr. Zhang, and Mr. Zhao. For example, if the medical record matching coefficient of Ms. Wang is the highest, then the pending medical record data of Ms. Wang will be determined as the de-identified archived medical record data for subsequent operations such as constructing disease-related information and determining the evaluation results of cardiovascular diseases.

[0048] In the embodiment of the present invention, the at least two target pending medical record data include at least two target pending medical record data obtained by each matching method; the determining of the de-identified archived medical record data from the at least two target pending medical record data according to the medical record matching coefficient between the at least two target pending medical record data and the medical record data to be evaluated can be implemented through the following examples.

[0049] For each of the at least two target pending medical record data obtained by each matching method, calculate the matching coefficient between the medical record data to be evaluated and each target pending medical record data;

[0050] For each target pending medical record data, calculate the arithmetic mean of the matching coefficients between the target pending medical record data and the medical record data to be evaluated in each matching method;

[0051] Determine the de-identified archived medical record data from the at least two target pending medical record data according to the arithmetic mean of the matching coefficients.

[0052] In an embodiment of the present invention, by way of example, in a hospital information management system, the server continuously processes medical record data related to cardiovascular diseases. Taking the medical record data to be evaluated of Mr. Li mentioned above as an example, the server has obtained multiple target pending medical record data through different matching methods. For example, through the image feature matching pool, the target pending medical record data of patients such as Ms. Wang and Mr. Zhang are obtained, and through the semantic description feature matching pool, the target pending medical record data of patients such as Ms. Zhao and Mr. Liu are obtained. For the target pending medical record data of Ms. Wang obtained through the image feature matching pool, the server needs to calculate the matching coefficient between the medical record data to be evaluated of Mr. Li and it. The server will compare the image features of the two in detail, such as the P-wave morphology, QRS complex duration and voltage of Mr. Li's electrocardiogram image, with the corresponding electrocardiogram image features in Ms. Wang's pending medical record data. Through a professional algorithm (such as a calculation method based on the distance of feature vectors), the similarity in image features between the two is accurately calculated, and this is used as part of the matching coefficient in the image feature matching method. At the same time, for the semantic description features of Ms. Wang's target pending medical record data, the server compares the first semantic description features such as the attack frequency and duration of symptoms of palpitation and chest tightness in Mr. Li's medical record data to be evaluated with the similar symptom descriptions, physical examination records and other semantic description features in Ms. Wang's pending medical record data. The similarity of semantic description features is calculated using natural language processing technology and used as part of the matching coefficient in the semantic description feature matching method. Then, by combining the partial matching coefficients of these two aspects (image feature and semantic description feature matching methods) and according to a certain weight (assuming the image feature weight is 0.6 and the semantic description feature weight is 0.4), the complete matching coefficient between Mr. Li's medical record data to be evaluated and Ms. Wang's target pending medical record data in this matching method is obtained. Similarly, the server will also perform the above similar matching coefficient calculation process on other target pending medical record data obtained through different matching methods, such as the target pending medical record data of Mr. Zhang, Ms. Zhao, Mr. Liu, etc. Still taking Ms. Wang's target pending medical record data as an example, assuming that the matching coefficient with Mr. Li's medical record data to be evaluated in the image feature matching method calculated above is 0.7, and the matching coefficient in the semantic description feature matching method is 0.6. Then the server will perform an arithmetic mean calculation on these two matching coefficients (0.7 and 0.6), that is, (0.7 + 0.6) ÷ 2 = 0.65, to obtain the arithmetic mean of the matching coefficients between Ms. Wang's target pending medical record data and Mr. Li's medical record data to be evaluated in these two matching methods. For other target pending medical record data such as Mr. Zhang, Ms. Zhao, Mr. Liu, etc., the server will also perform the same arithmetic mean calculation on their respective matching coefficients with Mr. Li's medical record data to be evaluated in different matching methods to obtain their respective arithmetic means of the matching coefficients. After the server calculates the arithmetic mean of the matching coefficients of each target pending medical record data, it will determine the desensitized archived medical record data accordingly.For example, the arithmetic mean of the matching coefficients of Ms. Wang's target-undetermined medical record data is 0.65, that of Mr. Zhang is 0.58, that of Ms. Zhao is 0.62, and that of Mr. Liu is 0.55. The server will compare these arithmetic means and find that the arithmetic mean of the matching coefficients of Ms. Wang is the highest. Then, the server will determine that Ms. Wang's target-undetermined medical record data is desensitized archived medical record data for subsequent operations such as constructing disease-related information and determining the final cardiovascular disease assessment results using this desensitized archived medical record data.

[0053] In the embodiment of the present invention, the obtaining of the first multi-dimensional feature by performing multi-dimensional feature extraction on the medical record data to be evaluated may be implemented through the following examples.

[0054] Extract at least two cardiovascular disease-related images from the medical record data to be evaluated, and divide the at least two cardiovascular disease-related images into at least two data sets;

[0055] Obtain a target cardiovascular disease-related image from the at least two data sets, and perform image feature extraction on the target cardiovascular disease-related image to obtain a first medical record image feature;

[0056] Obtain the first medical record document data corresponding to the medical record data to be evaluated, and perform semantic description feature extraction on the first medical record document data to obtain a first semantic description feature;

[0057] Obtain the first multi-dimensional feature according to the first medical record image feature and the first semantic description feature.

[0058] In an embodiment of the present invention, by way of example, in a hospital information management system, the server is responsible for processing a large amount of medical record data for cardiovascular disease assessment. Taking the medical record data to be evaluated of Mr. Li as an example, this medical record data contains rich information, including various cardiovascular disease-related images. The server will extract at least two such images from it, such as an electrocardiogram (ECG) image and an echocardiogram image. The ECG image can reflect the electrical activity of the heart, and many cardiovascular problems can be insight through the waveform changes on it; the echocardiogram image can clearly show the structure, size and blood flow conditions of the heart, etc. After extraction, the server will reasonably divide these images to form at least two data sets. For example, the ECG image is divided into one data set and managed separately for subsequent targeted analysis; the echocardiogram image is divided into another data set, so that different types of image data can be processed more orderly. After the server divides the data sets, it will select the target cardiovascular disease-related images from them for further analysis. Suppose the server selects the ECG image during a recent episode of palpitations of Mr. Li from the ECG image data set as the target cardiovascular disease-related image. Then, the server uses a professional image analysis algorithm to extract features from the target image. For the ECG image, it will extract features such as the morphology of the P wave (whether it is tall or flat), the duration of the QRS complex (whether it is within the normal range), and the inversion of the T wave. These features combined form the first medical record image features, which can reflect the specific conditions of Mr. Li's heart electrical activity from different angles and provide important basis for subsequent disease assessment. At the same time, the server will also obtain the first medical record document data corresponding to the medical record data to be evaluated of Mr. Li. These document data are detailed records made by the doctor during the consultation of Mr. Li, including the symptoms described by Mr. Li himself (such as the feeling of palpitations, the degree of chest tightness, the attack frequency, etc.), the doctor's physical examination results (such as whether there are murmurs in the heart auscultation, whether the heart rate is regular, etc.) and the preliminary diagnosis impression and other contents. The server then uses natural language processing technology to extract semantic description features from these first medical record document data. For example, from Mr. Li's description of the palpitations symptoms "occurring two or three times a week, each lasting about ten minutes, and feeling a very strong heartbeat during the attack", the server can extract semantic description features such as "medium palpitations attack frequency, medium duration, and high heartbeat intensity during the attack". These first semantic description features can reflect the characteristics of Mr. Li's condition in a more intuitive way. Finally, the server will obtain the first multi-dimensional features based on the first medical record image features and the first semantic description features extracted previously. For example, combine the first medical record image features such as the morphology of the P wave and the duration of the QRS complex extracted from the ECG image with the first semantic description features such as "medium palpitations attack frequency, medium duration, and high heartbeat intensity during the attack" extracted from the medical record document data.In this way, a first multi-dimensional feature covering both image and semantic information is formed, which can more comprehensively and accurately describe Mr. Li's cardiovascular disease-related conditions, laying a foundation for subsequent operations such as comparison and analysis with de-identified archived medical record data.

[0059] In an embodiment of the present invention, the identification of the cardiovascular disease target value of the to-be-evaluated medical record data according to the disease correlation information can be implemented through the following examples.

[0060] Perform a temporal feature mapping on the disease correlation information to obtain disease correlation temporal features, and the temporal features are input parameters corresponding to the cardiovascular disease recognition model;

[0061] Place the disease correlation temporal features into the cardiovascular disease recognition model, which is based on maintaining the original initial network parameters of a preset classification model and updating the iterative network parameters of the classification model according to the sample disease correlation temporal features, and the iterative network parameters are related to the PEFT fine-tuning component added to the classification model;

[0062] Obtain the target cardiovascular disease target value of the to-be-evaluated medical record data output by the cardiovascular disease recognition model.

[0063] In an embodiment of the present invention, exemplarily, in the hospital information management system, the server has completed the construction of disease-related information. For example, for the medical record data to be evaluated of Mr. Li, through multi-faceted correlation analysis with multiple desensitized archived medical record data, a detailed disease-related information has been obtained. This disease-related information includes the comparison of Mr. Li with other similar cases in terms of symptom manifestations, examination results, disease progression, and other aspects. Now, the server needs to perform temporal feature mapping on this disease-related information. The server will sort and analyze the contents of the disease-related information in chronological order according to the preset algorithms and rules. For example, for the attack frequency of Mr. Li's palpitation symptoms, the time change from the initial onset to the most recent attack; and the change trends of his various examination results (such as electrocardiogram, cardiac ultrasound, etc.) at different time points. Through such processing, the originally relatively static disease-related information is transformed into disease-related temporal features with temporal characteristics. These temporal features can more clearly reflect the development dynamics of the disease over time, which is exactly the input parameter required by the cardiovascular disease recognition model, preparing for accurately identifying the cardiovascular disease target value in the follow-up. After obtaining the disease-related temporal features, the server will put them into the cardiovascular disease recognition model for analysis. This cardiovascular disease recognition model is obtained through special training. Initially, it is based on a preset classification model, such as a commonly used neural network classification model. During the training process, the server keeps the initial network parameters of this preset classification model unchanged to retain some basic characteristics and structures of the model. Then, the server will use a large number of sample disease-related temporal features to update the iterative network parameters of the classification model. These sample disease-related temporal features are obtained by performing similar processing (such as constructing disease-related information and performing temporal feature mapping) on the medical record data of many previously diagnosed cardiovascular disease patients. When updating the iterative network parameters, the key lies in adding a PEFT fine-tuning component. This component adopts specific technologies (such as LoRA technology), which can fine-tune the iterative network parameters of the classification model according to the error between the sample disease-related temporal features and the actual cardiovascular disease target value, enabling the model to more accurately adapt to the cardiovascular disease recognition task. When the server puts the disease-related temporal features of Mr. Li into this specially trained cardiovascular disease recognition model, the model can analyze Mr. Li's condition based on the optimized network parameters and algorithms inside it. Finally, the server waits for the output result of the cardiovascular disease recognition model. After complex calculations and analyses inside the model, based on Mr. Li's disease-related temporal features and the optimized network parameters from the previous training, the cardiovascular disease recognition model will output a target cardiovascular disease target value for Mr. Li's medical record data to be evaluated.For example, if the result output by the model is "moderate coronary heart disease", then this is the target value of cardiovascular disease determined by the server for Mr. Li's medical record data to be evaluated after a series of complex processes. This value will provide an important basis for formulating targeted treatment plans and intervention measures in the follow-up.

[0064] In the embodiment of the present invention, the cardiovascular disease recognition model can be obtained in the following manner, and the implementation can be performed through the following examples.

[0065] Obtain the first sample medical record data image features, the first sample medical record document data, and the first sample cardiovascular disease target value corresponding to the first sample medical record data, as well as the second sample medical record data image features, the second sample medical record document data, and the second sample cardiovascular disease target value corresponding to the second sample medical record data;

[0066] Perform feature transformation operations on the first sample medical record data image features and the second sample medical record data image features respectively according to a preset initial feature transformation component to obtain the first sample disease semantic description information and the second sample disease semantic description information, and construct sample disease association information according to the first sample disease semantic description information, the first sample medical record document data, and the first sample cardiovascular disease target value, as well as the second sample disease semantic description information and the second sample medical record document data;

[0067] Perform a temporal feature mapping on the sample disease association information to obtain the sample disease association temporal features;

[0068] Add the PEFT fine-tuning component to the classification model to add the iterative network parameters to the classification model through the PEFT fine-tuning component;

[0069] Keep the initial network parameters of the classification model unchanged, and update the iterative network parameters of the classification model according to the sample disease association temporal features and the second sample cardiovascular disease target value to obtain the cardiovascular disease recognition model.

[0070] In an embodiment of the present invention, exemplarily, in the huge medical record database of a hospital, the server undertakes the important task of collecting and organizing data to build a cardiovascular disease recognition model. Taking the medical record data of a large number of past cardiovascular disease patients as an example, the server first selects a part of it as the first sample medical record data, and then selects another part as the second sample medical record data. For the first sample medical record data, assume it is the medical record of a patient named Mr. Wang. The server extracts relevant information from Mr. Wang's medical record: First sample medical record data imaging features: From the electrocardiogram image taken by Mr. Wang, the server extracts imaging features such as the P wave morphology is normal but slightly wide, and the QRS complex duration is normal but the voltage is slightly low; from the echocardiogram image, features such as the left ventricular wall thickness is slightly thicker than the normal range are extracted. These imaging features can reflect the electrical activity and structural conditions of Mr. Wang's heart. First sample medical record document data: This part includes the symptoms described by Mr. Wang himself, such as the palpitation symptom occurs two or three times a week, each time lasting about ten minutes, and he feels a little stuffy in the chest during the attack; the doctor's physical examination results, such as a slight murmur is heard during heart auscultation, the heart rate is slightly fast but basically regular; and the doctor's preliminary diagnosis impression and other contents, which detail the situation when Mr. Wang visited the doctor. First sample cardiovascular disease target value: After a series of detailed examination and diagnosis processes, Mr. Wang was finally diagnosed with mild coronary heart disease, which is his first sample cardiovascular disease target value. Similarly, for the second sample medical record data, such as the medical record of Ms. Li. The server also performs similar extractions: Second sample medical record data imaging features: From Ms. Li's electrocardiogram image, features such as the T wave has a slight inversion and the QRS complex duration is slightly prolonged are extracted; from the echocardiogram image, it is found that the right atrium is slightly enlarged and other features. Second sample medical record document data: Ms. Li describes that the palpitation symptom occurs several times a day, each time lasting for a few minutes, accompanied by slight dizziness; the doctor's physical examination finds that the blood pressure is slightly high and there is no obvious murmur during heart auscultation; and the doctor's preliminary diagnosis and other situations. Second sample cardiovascular disease target value: Ms. Li was finally diagnosed with early myocardial ischemia, which is her second sample cardiovascular disease target value. After the server obtains the various features of the above two groups of sample medical record data, it begins to process them using a preset initial feature conversion component. For the imaging features of the first sample medical record data of Mr. Wang, the server inputs them into the initial feature conversion component. Taking the electrocardiogram imaging features as an example, through the built-in algorithms and rules, the initial feature conversion component will convert the feature that the P wave morphology is normal but slightly wide into a first sample disease semantic description information similar to "There may be slight changes in the atrial depolarization process. Although the P wave morphology is normal, there is a trend of being slightly wide, which may imply subtle abnormalities in the initial stage of cardiac electrical activity"; for the echocardiogram imaging feature that the left ventricular wall thickness is slightly thicker than the normal range, it is converted into a semantic description such as "There may be a certain degree of compensatory thickening of the left ventricular myocardium, indicating that the cardiovascular system may be under certain pressure".Similarly, for the image feature of Ms. Li's second sample medical record data, after being processed by the initial feature conversion component, the feature of slight T wave inversion may be converted into semantic description information of the second sample condition such as "there may be local slight ischemia in the myocardium, which is reflected as T wave inversion on the electrocardiogram"; the feature of slightly enlarged right atrium is converted into semantic descriptions such as "the structure of the right atrium is slightly enlarged, which may affect cardiac hemodynamics, and it is necessary to pay attention to whether it is related to cardiovascular diseases". Then, the server constructs sample condition association information based on these converted semantic description information of the first sample condition, the first sample medical record document data and the first sample cardiovascular disease target value, as well as the semantic description information of the second sample condition, the second sample medical record document data and the second sample cardiovascular disease target value. Taking Mr. Wang and Ms. Li as examples, the server will analyze the relationship between the first sample medical record document data such as the onset of Mr. Wang's palpitation symptoms and the slight murmur heard during heart auscultation and his semantic description information of the first sample condition (such as the subtle abnormalities in the initial stage of cardiac electrical activity and the possible pressure on the cardiovascular system), and his first sample cardiovascular disease target value of mild coronary heart disease; at the same time, compare the relationship between the second sample medical record document data such as the number of times Ms. Li's palpitation symptoms occur every day and the accompanied slight dizziness and her semantic description information of the second sample condition (such as there may be local slight ischemia in the myocardium and the slightly enlarged right atrium structure may affect cardiac hemodynamics), and her second sample cardiovascular disease target value of early myocardial ischemia. By comprehensively considering these factors, the server constructs a sample condition association information that includes the association situation in terms of symptoms, image semantic description, cardiovascular disease diagnosis, etc. among different sample medical record data such as Mr. Wang and Ms. Li. After constructing the sample condition association information, the server then needs to perform a temporal feature mapping on it. For the part of Mr. Wang's sample condition association information, the server will sort out the change situation of the attack frequency of his palpitation symptoms in chronological order, such as the change process from once a week at the initial onset to two or three times a week later; and the change trend of his various examination results (such as electrocardiogram, echocardiogram, etc.) at different time points, such as the change of features such as the P wave morphology and QRS complex voltage on the electrocardiogram over time. Similarly, for the part of Ms. Li's sample condition association information, the server will also sort out the change process of the attack frequency of her palpitation symptoms from several times a day to a decrease after treatment; and the change trend of her various examination results at different time points, such as whether the T wave inversion on the electrocardiogram has improved. By performing a comprehensive temporal feature mapping on the sample condition association information of different sample medical record data in this way, the server obtains the sample condition association temporal features. These temporal features can more clearly reflect the development dynamics of the condition over time in the sample medical record data and provide an important basis for subsequent model training. After preparing the sample condition association temporal features, the server needs to adjust the preset classification model.Assume that the preset classification model is a common neural network model, such as a multi-layer perceptron (MLP) model. The server adds the PEFT fine-tuning component to this classification model. The PEFT fine-tuning component is implemented using specific techniques, such as LoRA technology. After adding the PEFT fine-tuning component to the classification model, iterative network parameters can be added to the classification model through this component. These iterative network parameters are the parameters required when updating the classification model based on the sample disease condition-related time series features and the corresponding cardiovascular disease target values in the subsequent process. They can continuously optimize the classification model to adapt to the task of cardiovascular disease recognition. After the server adds the PEFT fine-tuning component to the classification model, it starts to update and train the classification model to obtain a cardiovascular disease recognition model. In this process, the server keeps the initial network parameters of the classification model unchanged. This is to retain some original basic characteristics and structures of the classification model so that it will not lose its original advantages due to excessive adjustment. Then, the server updates the iterative network parameters of the classification model according to the obtained sample disease condition-related time series features and the second sample cardiovascular disease target values. Taking the sample medical record data of Mr. Wang and Ms. Li as an example, the server inputs the sample disease condition-related time series features obtained from the sample medical record data of Mr. Wang, Ms. Li, and many others after the previous processing into the classification model added with the PEFT fine-tuning component, and at the same time takes the second sample cardiovascular disease target values of Ms. Li, etc. as the target output. By comparing the error between the predicted value output by the model and the actual second sample cardiovascular disease target value, and using a specific algorithm (such as the stochastic gradient descent algorithm), the server adjusts and updates the iterative network parameters of the classification model according to this error. After multiple such iterative update processes, with the continuous increase and processing of the sample medical record data, the iterative network parameters of the classification model are gradually optimized, and finally a cardiovascular disease recognition model that can accurately identify cardiovascular diseases is obtained. This model can be used for subsequent operations such as identifying the cardiovascular disease target values of new medical record data to be evaluated.

[0071] In the embodiment of the present invention, before performing the feature transformation operation on the image features according to the preset initial feature transformation component, the following implementation manners are also provided.

[0072] Obtain sample cardiovascular disease-related images and sample medical record document data corresponding to the sample cardiovascular disease-related images;

[0073] Extract features from the sample cardiovascular disease-related images to obtain sample image features, and place the sample image features into a preliminary feature transformation component, so that the preliminary feature transformation component transforms the sample image features into the input semantic feature domain of the classification model;

[0074] Obtain the sample semantic description features corresponding to the feature conversion request, and place the sample semantic description features and the target sample disease condition semantic description features output by the feature conversion component into the classification model, where the network parameters of the classification model remain unchanged;

[0075] Obtain the sample inferred medical record document data output by the classification model, and train the preliminary feature conversion component according to the sample medical record document data and the sample inferred medical record document data to obtain the initial feature conversion component.

[0076] In an embodiment of the present invention, by way of example, in a hospital information management system, the server is responsible for preparing data for constructing an accurate cardiovascular disease recognition model. First, the server selects a large number of sample cardiovascular disease-related images and their corresponding sample medical record document data from a vast medical record database. For example, the case of Mr. Chen is selected, and the sample cardiovascular disease-related images include electrocardiogram images and echocardiogram images taken multiple times. The corresponding sample medical record document data details Mr. Chen's symptoms, such as the irregular onset of palpitation symptoms, with varying durations for each onset, sometimes accompanied by mild chest pain; the doctor's physical examination results, such as occasional mild murmurs detected during heart auscultation and a generally normal heart rate; and the doctor's preliminary diagnostic impressions based on these conditions. After obtaining Mr. Chen's electrocardiogram and echocardiogram and other sample cardiovascular disease-related images, the server uses professional image analysis algorithms to extract features from them. For electrocardiogram images, the server extracts sample image features such as the morphology of the P wave, the duration and voltage of the QRS complex, and the inversion of the T wave. From echocardiogram images, sample image features such as the size of the cardiac chambers, the thickness of the ventricular walls, and the blood flow velocity are extracted. Then, these extracted sample image features are placed into a preliminary feature conversion component. The role of this preliminary feature conversion component is to convert the sample image features from their original image data format to the input semantic feature domain that the classification model can understand, just like translating one "language" into another "language", so as to enable accurate analysis and processing in the classification model subsequently. During this process, the server also obtains sample semantic description features corresponding to the sample cardiovascular disease-related images. For example, from Mr. Chen's medical record document data, sample semantic description features such as "the onset of palpitation symptoms is unstable, and there may be abnormal myocardial electrical activity" are extracted through natural language processing technology, which is a feature refinement of Mr. Chen's condition based on text description. At the same time, the preliminary feature conversion component will convert the input sample image features and output target sample condition semantic description features. For example, it will convert electrocardiogram image features into a more semantic description of the heart's electrical activity state. Then, the obtained sample semantic description features and the target sample condition semantic description features output by the preliminary feature conversion component are placed into the classification model together, and during this process, the network parameters of the classification model are kept unchanged, that is, the original structure and parameter settings of the classification model are not changed, so as to solely examine the effect of the preliminary feature conversion component. After receiving the sample semantic description features and the target sample condition semantic description features, the classification model will perform operations based on its internal algorithms and logic and output sample inferred medical record document data. For example, for Mr. Chen's case, the classification model may output an inferred description of Mr. Chen's condition, such as predictions of the attack frequency and duration of palpitation symptoms and speculations on the heart auscultation results.Then, the server will train the preliminary feature transformation component based on the difference between Mr. Chen's original sample medical record document data and the sample inferred medical record document data output by the classification model. If the prediction of the palpitation attack frequency in the sample inferred medical record document data does not match the actual situation in the original sample medical record document data, the server will adjust the preliminary feature transformation component according to this difference through a specific training algorithm (such as the backpropagation algorithm, etc.) to continuously optimize it. After a large number of such training processes, using the data of numerous different sample cases, an initial feature transformation component that can accurately perform feature transformation is finally trained, laying a foundation for the subsequent construction of a cardiovascular disease recognition model.

[0077] In the embodiment of the present invention, the iterative network parameters of the classification model are updated according to the sample disease-related temporal features to obtain the cardiovascular disease recognition model, which can be implemented through the following examples.

[0078] Place the sample disease-related temporal features into the classification model with the PEFT fine-tuning component added, and obtain the sample inferred cardiovascular disease target value output by the classification model;

[0079] Update the iterative network parameters of the classification model according to the error between the sample inferred cardiovascular disease target value and the second sample cardiovascular disease target value to obtain the cardiovascular disease recognition model;

[0080] The embodiment of the present invention also provides the following implementation manners.

[0081] Update the component parameters of the initial feature transformation component according to the error between the sample inferred cardiovascular disease target value and the second sample cardiovascular disease target value.

[0082] In an embodiment of the present invention, exemplarily, the server has prepared sample disease-related time-series features through a series of previous steps, and these features clearly reflect the development dynamics of the disease over time in the sample medical record data. The server places these sample disease-related time-series features into a classification model that has incorporated a PEFT fine-tuning component. Suppose this classification model is a carefully designed neural network model, and the PEFT fine-tuning component adopts the advanced LoRA technology, which can effectively fine-tune and optimize the model. When the sample disease-related time-series features are input into this model, the model starts to perform arithmetic processing based on its internal complex algorithms and structures, as well as the fine-tuning ability brought by the PEFT fine-tuning component. After a series of calculations, the model finally outputs a sample inferred cardiovascular disease target value. For example, for the sample disease-related time-series features corresponding to a sample medical record data, the sample inferred cardiovascular disease target value output by the model may be "moderate coronary heart disease", which is the cardiovascular disease condition corresponding to the sample medical record data inferred by the model based on the input time-series features and its own parameters and algorithms. After the server obtains the sample inferred cardiovascular disease target value output by the model, it is necessary to compare it with the known second sample cardiovascular disease target value. The second sample cardiovascular disease target value was clearly determined when the sample medical record data was initially collected, and it is the true cardiovascular disease condition obtained after a detailed diagnosis of the sample medical record data. For example, the second sample cardiovascular disease target value corresponding to the just-mentioned sample medical record data is "mild coronary heart disease", while the sample inferred cardiovascular disease target value output by the model is "moderate coronary heart disease", which means there is a certain error. The server will update the iterative network parameters of the classification model through a specific algorithm (such as the stochastic gradient descent algorithm, etc.) according to this error. These iterative network parameters are related to the PEFT fine-tuning component. By continuously adjusting these parameters, the model can more accurately identify cardiovascular diseases. After each update of the iterative network parameters, the server will input the sample disease-related time-series features into the updated model again to obtain a new sample inferred cardiovascular disease target value, and then continue to compare it with the second sample cardiovascular disease target value, and repeat this iterative update process multiple times. Through the repeated training and iterative update process of a large number of sample medical record data, the iterative network parameters of the classification model are gradually optimized, and finally a cardiovascular disease recognition model that can accurately identify cardiovascular diseases is obtained. At the same time, the server will also update the component parameters of the initial feature conversion component according to the error between the sample inferred cardiovascular disease target value and the second sample cardiovascular disease target value. The initial feature conversion component played an important role in the previous step of converting sample image features into semantic features that are more easily understood by the classification model. When it is found that there is an error between the sample inferred cardiovascular disease target value and the second sample cardiovascular disease target value, it indicates that there may also be certain problems in the feature conversion link.For example, if the error is large, it may be because the initial feature conversion component is not accurate enough when converting the sample image features into semantic features, resulting in the subsequent classification model making incorrect inferences based on these inaccurate semantic features. Therefore, the server will adjust and update the component parameters of the initial feature conversion component according to this error, also through specific algorithms (such as backpropagation algorithm, etc.). By continuously optimizing the parameters of the initial feature conversion component, it can perform feature conversion more accurately, thereby providing more accurate inputs for the cardiovascular disease recognition model and further improving the accuracy of the entire cardiovascular disease recognition system.

[0083] In addition, in the cardiovascular disease assessment and management method based on big data, there are many detailed and interrelated steps, which will be integrated and elaborated below. First, in the link of obtaining the original medical record data of the target patient, the server plays an important role. It establishes a data interface connection with the electronic medical record system of the medical institution. This data interface uses a secure encryption protocol to ensure the security and confidentiality of data transmission, so as to obtain the original medical record data generated by the target patient in this medical institution in real time. At the same time, in order to more comprehensively evaluate the cardiovascular disease status of the patient, the server will also obtain the real-time physiological data of the target patient from wearable devices, such as heart rate, blood pressure and other information, and integrate it into the original medical record data as a supplementary data source. Then, the obtained original medical record data is processed by data standardization to obtain the medical record data to be evaluated. According to the preset data standard specifications related to cardiovascular diseases, format unification processing is carried out on each index data in the original medical record data. For example, information such as dates and times recorded in different formats are converted into a unified standard format. For different types of data, the processing methods are also different. For image data, operations such as resolution adjustment and contrast enhancement are carried out; for text data, vocabulary standardization and grammar specification are implemented. Through these operations, it is ensured that the obtained medical record data to be evaluated has consistency and comparability, so as to smoothly carry out comparison and analysis operations subsequently. In the process of comparing and analyzing the medical record data to be evaluated with at least two desensitized archived medical record data to obtain the cardiovascular disease assessment result, there are a series of key steps. Before the matching operation, it is necessary to first process the first multi-dimensional feature of the medical record data to be evaluated. Specifically, data cleaning operations need to be carried out on it to remove the outliers, duplicate data and data features irrelevant to cardiovascular disease assessment in the first multi-dimensional feature. Subsequently, a data normalization method is used to process the cleaned first multi-dimensional feature, and the values of the data dimension features are mapped into a preset standard interval, so as to improve the accuracy and efficiency of subsequent matching with the desensitized archived medical record data. When matching, it involves multiple complex sub-steps. First, at least two cardiovascular disease-related images are extracted from the medical record data to be evaluated and divided into at least two data sets. Then, target cardiovascular disease-related images are obtained from these data sets, and image feature extraction is carried out on them to obtain the first medical record image features. At the same time, the first medical record document data corresponding to the medical record data to be evaluated is obtained, and semantic description feature extraction is carried out on it to obtain the first semantic description features. Furthermore, based on these two, the first multi-dimensional feature is obtained. After that, a pre-established image feature matching pool and a semantic description feature matching pool are obtained, and matching operations are carried out according to specific rules, that is, the target image features and the first semantic description features in the first multi-dimensional feature are respectively used to match these two matching pools to obtain at least two target pending medical record data that match the medical record data to be evaluated.For these medical record data with undetermined targets, it is necessary to calculate the matching coefficients between the medical record data to be evaluated and each piece of medical record data with undetermined targets, and further calculate the arithmetic mean of the matching coefficients between each piece of medical record data with undetermined targets and the medical record data to be evaluated in each matching method. Finally, determine the de-identified archived medical record data from at least two pieces of medical record data with undetermined targets. After determining the de-identified archived medical record data, construct the disease condition association information based on the second multi-dimensional features corresponding to each de-identified archived medical record data, the cardiovascular disease target values configured for each de-identified archived medical record data, and the first multi-dimensional features of the medical record data to be evaluated. In this process, if the number of the first medical record image features is at least two, it is necessary to perform a feature integration operation on them to obtain the target image features, and then convert the target image features to the semantic feature domain through a preset feature conversion component to obtain the first disease condition semantic description information. At the same time, perform feature mapping processing on the second medical record image features corresponding to each de-identified archived medical record data respectively to obtain the second disease condition semantic description information corresponding to each de-identified archived medical record data. Then, obtain the first medical record document data corresponding to the medical record data to be evaluated and the second medical record document data corresponding to each de-identified archived medical record data respectively, and construct the context deduction information based on the first identifier used to distinguish different contents included in the same medical record data, the second disease condition semantic description information, the second medical record document data, and the cardiovascular disease target values corresponding to the same de-identified archived medical record data, so as to obtain at least two context deduction information corresponding to at least two de-identified archived medical record data. Integrate these context deduction information to obtain the context integration information, and then construct the association information based on the first identifier, the first disease condition semantic description information, and the first medical record document data. Finally, generate the disease condition association information based on the context integration information and the association information. After obtaining the disease condition association information, it is necessary to perform a temporal feature mapping on it to obtain the disease condition association temporal features, which serve as the input parameters for the cardiovascular disease recognition model. The specific operation is to use the time window sliding technology to segment the disease condition association information, divide it into data of multiple time window segments. For the data of each time window segment, use a preset temporal feature extraction algorithm (which can be implemented based on a recurrent neural network or a long short-term memory network) to extract its corresponding temporal features, and then integrate and splice the temporal features extracted from each time window segment to form the disease condition association temporal features, which fully reflect the feature changes of the disease condition association information in the time dimension. There is also a rigorous process for obtaining the cardiovascular disease recognition model. First, obtain the first sample medical record data image features, the first sample medical record document data, and the first sample cardiovascular disease target values corresponding to the first sample medical record data, as well as the second sample medical record data image features, the second sample medical record document data, and the second sample cardiovascular disease target values corresponding to the second sample medical record data.Then, perform feature transformation operations on the first sample medical record data image features and the second sample medical record data image features respectively according to the preset initial feature transformation component, obtain the first sample disease semantic description information and the second sample disease semantic description information, construct sample disease association information based on this information, and perform temporal feature mapping on it to obtain sample disease association temporal features. Next, add the PEFT fine-tuning component (implemented using the LoRA technique) to the classification model to add iterative network parameters to the classification model through this component. During the process of updating the iterative network parameters using the PEFT fine-tuning component of the LoRA technique, set different rank values for experiments, and determine the optimal rank value by comparing the error between the sample inferred cardiovascular disease target value output by the classification model under different rank values and the second sample cardiovascular disease target value. Configure the PEFT fine-tuning component according to the optimal rank value to complete the update and optimization of the cardiovascular disease recognition model, and improve its accuracy and generalization ability for different disease association temporal feature inputs. At the same time, during the process of updating the iterative network parameters of the classification model according to the sample disease association temporal features to obtain the cardiovascular disease recognition model, it is also necessary to update the component parameters of the initial feature transformation component according to the error between the sample inferred cardiovascular disease target value and the second sample cardiovascular disease target value. Specifically, the stochastic gradient descent algorithm is used to calculate the gradient values of the component parameters of the initial feature transformation component according to this error, and then adjust and update the component parameters of the initial feature transformation component according to the preset learning rate based on the gradient values, so that the initial feature transformation component can more accurately convert the image features into the features of the semantic feature domain suitable for the input of the classification model in subsequent processing. Finally, according to the cardiovascular disease assessment results, select the corresponding cardiovascular disease intervention strategies in the preset cardiovascular disease intervention database. First, analyze the key information such as the disease severity and development trend in the cardiovascular disease assessment results, and perform hierarchical screening in the preset cardiovascular disease intervention database based on these key information. First, screen out all intervention strategies applicable to this disease type, and then further narrow down the scope according to the disease severity to select the most suitable cardiovascular disease intervention strategy for the current assessment results. And during the selection process, fully consider the individual difference factors of the patient, such as age, gender, underlying diseases, etc., and perform personalized adjustment and optimization on the selected intervention strategies. For example, if the patient is an elderly person and the assessment results show that the cardiovascular disease is in the early stage, on the basis of the selected intervention strategy, appropriately increase the frequency of regular physical examinations and adjust the diet recommendations to increase the intake proportion of foods rich in dietary fiber and unsaturated fatty acids; for female patients, according to the characteristics of their physiological cycle, reasonably arrange the exercise time and intensity in the intervention strategy, and avoid high-intensity exercise during the physiological period; for patients with other underlying diseases such as diabetes, combine the disease control requirements of diabetes and make coordinated adjustments to aspects such as drug use and diet control in the cardiovascular disease intervention strategy to achieve the effect of comprehensive treatment and prevention.Through this series of complete and elaborate steps, the big data-based cardiovascular disease assessment and management method can more accurately and comprehensively evaluate the cardiovascular disease status of patients and formulate more targeted and effective intervention strategies.

[0084] An embodiment of the present invention provides a computer device 100. The computer device 100 includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned big data-based cardiovascular disease assessment and management method. As Figure 2 shown, Figure 2 is a structural block diagram of the computer device 100 provided by an embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112, and a communication unit 113. To achieve data transmission or interaction, the elements of the memory 111, the processor 112, and the communication unit 113 are electrically connected to each other directly or indirectly. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines.

[0085] For illustrative purposes, the foregoing description has been made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the disclosure to the precise forms disclosed. Numerous modifications and variations are possible in light of the above teachings. The embodiments were chosen and described in order to best illustrate the principles of the disclosure and its practical application, thereby enabling those skilled in the art to best utilize the disclosure and to adapt various embodiments with different modifications to the particular applications contemplated.

Claims

1. A cardiovascular disease assessment and management method based on big data, characterized in that: include: Obtain original medical records of target patients; Performing data standardization processing on the original medical record data to obtain medical record data to be evaluated; Compare and analyze the medical record data to be evaluated with at least two desensitized archived medical record data to obtain the cardiovascular disease evaluation results corresponding to the medical record data to be evaluated; According to the cardiovascular disease assessment results, selecting a corresponding cardiovascular disease intervention strategy from a preset cardiovascular disease intervention database; The step of comparing and analyzing the medical record data to be evaluated with at least two desensitized archived medical record data to obtain a cardiovascular disease evaluation result corresponding to the medical record data to be evaluated includes: Constructing disease association information according to the second multidimensional features corresponding to each of the desensitized archived medical record data, the cardiovascular disease target value configured for each of the desensitized archived medical record data, and the first multidimensional features of the medical record data to be evaluated; According to the disease association information, the cardiovascular disease target value of the medical record data to be evaluated is identified to obtain a cardiovascular disease evaluation result; The step of identifying the cardiovascular disease target value of the medical record data to be evaluated according to the disease association information includes: Performing time series feature mapping on the disease condition related information to obtain disease condition related time series features; Placing the disease-associated time series features into the cardiovascular disease recognition model; Obtaining a target cardiovascular disease target value of the medical record data to be evaluated output by the cardiovascular disease recognition model; The cardiovascular disease recognition model is obtained by the following methods, including: Obtaining first sample medical record data image features, first sample medical record document data, and first sample cardiovascular disease target values ​​corresponding to the first sample medical record data, and second sample medical record data image features, second sample medical record document data, and second sample cardiovascular disease target values ​​corresponding to the second sample medical record data; According to a preset initial feature conversion component, feature conversion operations are performed on the first sample medical record data image feature and the second sample medical record data image feature respectively to obtain first sample condition semantic description information and second sample condition semantic description information, and sample condition association information is constructed according to the first sample condition semantic description information, the first sample medical record document data and the first sample cardiovascular disease target value, and the second sample condition semantic description information and the second sample medical record document data; Performing time series feature mapping on the sample disease condition related information to obtain the sample disease condition related time series feature; Adding a PEFT fine-tuning component to a classification model, so as to add iterative network parameters to the classification model through the PEFT fine-tuning component; The initial network parameters of the classification model are maintained as they are, and the iterative network parameters of the classification model are updated according to the sample condition-related time series characteristics and the second sample cardiovascular disease target value to obtain the cardiovascular disease recognition model.

2. The method according to claim 1, characterized in that The step of comparing and analyzing the medical record data to be evaluated with at least two desensitized archived medical record data to obtain a cardiovascular disease evaluation result corresponding to the medical record data to be evaluated further includes: Acquire medical record data to be evaluated, extract at least two cardiovascular disease-related images from the medical record data to be evaluated, and divide the at least two cardiovascular disease-related images into at least two data sets; Acquire target cardiovascular disease related images from the at least two data sets, and extract image features from the target cardiovascular disease related images to obtain first medical record image features; Acquire first medical record document data corresponding to the medical record data to be evaluated, and extract semantic description features from the first medical record document data to obtain first semantic description features; Obtaining a first multidimensional feature according to the first medical record image feature and the first semantic description feature, wherein the first multidimensional feature includes a target image feature and a first semantic description feature; Acquire a pre-established image feature matching pool and a semantic description feature matching pool, wherein the image feature matching pool includes a correlation relationship between each pending medical record data and the image feature corresponding to the pending medical record data, and the semantic description feature matching pool includes a correlation relationship between each pending medical record data and the semantic description feature corresponding to the pending medical record data; According to the target image features, the image feature matching pool and the semantic description feature matching pool are matched respectively, and according to the first semantic description features, the image feature matching pool and the semantic description feature matching pool are matched respectively to obtain at least two target pending medical record data that match the medical record data to be evaluated; the at least two target pending medical record data include at least two target pending medical record data obtained by matching each matching method; For the at least two target pending medical record data obtained by matching in each matching method, a matching coefficient between the medical record data to be evaluated and each target pending medical record data is calculated; For each target pending medical record data, calculating the arithmetic mean of the matching coefficients of the target pending medical record data in each matching mode with the medical record data to be evaluated; Determine the desensitized archived medical record data from the at least two target pending medical record data according to the arithmetic mean of the matching coefficients, each of the desensitized archived medical record data being configured with a cardiovascular disease target value; Constructing disease association information according to the second multidimensional features corresponding to each of the desensitized archived medical record data, the cardiovascular disease target value configured for each of the desensitized archived medical record data, and the first multidimensional features of the medical record data to be evaluated; According to the disease association information, the cardiovascular disease target value of the medical record data to be evaluated is identified to obtain a cardiovascular disease evaluation result.

3. The method according to claim 2, characterized in that The first multidimensional feature includes a first medical record image feature, the second multidimensional feature includes a second medical record image feature, and the number of the first medical record image features is at least two; the condition association information is constructed according to the second multidimensional feature corresponding to each of the desensitized archived medical record data, the cardiovascular disease target value configured for each of the desensitized archived medical record data, and the first multidimensional feature of the medical record data to be evaluated, including: Performing a feature integration operation on at least two of the first medical record image features to obtain a target image feature; The target image feature is converted into a semantic feature domain through a preset feature conversion component to obtain first semantic description information of the condition, and the second medical record image feature corresponding to each of the desensitized archived medical record data is subjected to feature mapping processing to obtain second semantic description information of the condition corresponding to each of the desensitized archived medical record data; Obtaining first medical record document data corresponding to the medical record data to be evaluated and second medical record document data corresponding to each of the desensitized archived medical record data; Constructing context deduction information according to a first identifier for distinguishing different contents included in the same medical record data, and the second condition semantic description information, the second medical record document data, and the cardiovascular disease target value corresponding to the same desensitized archived medical record data, so as to obtain at least two context deduction information corresponding to at least two desensitized archived medical record data; Integrating the at least two pieces of context deduction information to obtain context integration information, wherein at least two pieces of context deduction information in the context integration information are divided by a second identifier for distinguishing contents of different medical record data; Constructing association information according to the first identifier, the first condition semantic description information and the first medical record document data; The disease condition related information is generated according to the context integration information and the related information.

4. The method according to claim 2, characterized in that: The step of identifying the cardiovascular disease target value of the medical record data to be evaluated according to the disease association information further includes: Performing time series feature mapping on the disease-related information to obtain disease-related time series features, wherein the time series features are input parameters corresponding to a cardiovascular disease recognition model; Placing the disease-associated time series features into the cardiovascular disease recognition model, the cardiovascular disease recognition model is obtained by maintaining the initial network parameters of a preset classification model, and updating the iterative network parameters of the classification model according to the sample disease-associated time series features, the iterative network parameters being related to the PEFT fine-tuning component added to the classification model; The target cardiovascular disease target value of the medical record data to be evaluated output by the cardiovascular disease recognition model is obtained.

5. The method according to claim 4, characterized in that Before performing a feature conversion operation on the image feature according to the preset initial feature conversion component, the method further includes: Acquire sample cardiovascular disease-related images and sample medical record document data corresponding to the sample cardiovascular disease-related images; Extracting features from the sample cardiovascular disease-related images to obtain sample image features, and placing the sample image features into a preparatory feature conversion component, so that the preparatory feature conversion component converts the sample image features into an input semantic feature domain of the classification model; Obtaining a sample semantic description feature corresponding to a feature conversion request, and placing the sample semantic description feature and the target sample condition semantic description feature output by the feature conversion component into the classification model, wherein the network parameters of the classification model remain unchanged; The sample inferred medical record document data output by the classification model is obtained, and the preliminary feature conversion component is trained according to the sample medical record document data and the sample inferred medical record document data to obtain the initial feature conversion component.

6. The method according to claim 4, characterized in that The updating of the iterative network parameters of the classification model according to the sample disease-associated time series characteristics to obtain the cardiovascular disease recognition model includes: Placing the sample disease-associated time series features into the classification model to which the PEFT fine-tuning component is added, and obtaining the sample inferred cardiovascular disease target value output by the classification model; According to the error between the sample inferred cardiovascular disease target value and the second sample cardiovascular disease target value, updating the iterative network parameters of the classification model to obtain the cardiovascular disease recognition model; The method further comprises: The component parameters of the initial feature conversion component are updated according to the error between the sample inferred cardiovascular disease target value and the second sample cardiovascular disease target value.

7. The method according to claim 6, characterized in that The updating of the component parameters of the initial feature conversion component includes: Using a stochastic gradient descent algorithm, according to the error between the sample inferred cardiovascular disease target value and the second sample cardiovascular disease target value, the gradient value of each component parameter of the initial feature conversion component is calculated; According to the preset learning rate, the component parameters of the initial feature conversion component are adjusted and updated according to the gradient value.

8. The method according to claim 6, characterized in that The classification model is a neural network model, the PEFT fine-tuning component is implemented using LoRA technology, the iterative network parameters of the classification model are updated according to the sample disease-related time series characteristics and the second sample cardiovascular disease target value to obtain the cardiovascular disease recognition model, and further includes: In the process of updating the iterative network parameters by using the PEFT fine-tuning component of the LoRA technology, different rank values ​​are set for testing, and the optimal rank value is determined by comparing the error between the sample inferred cardiovascular disease target value output by the classification model under different rank values ​​and the second sample cardiovascular disease target value; The PEFT fine-tuning component is configured according to the optimal rank value to complete the update optimization of the cardiovascular disease recognition model.

9. The method according to claim 2, characterized in that: Before matching at least two desensitized archived medical record data that match the medical record data to be evaluated according to the first multidimensional feature, the method further includes: Performing a data cleaning operation on the first multidimensional feature of the medical record data to be evaluated, removing abnormal values, duplicate data, and data features irrelevant to cardiovascular disease evaluation in the first multidimensional feature; The data normalization method is used to process the cleaned first multidimensional features, and the value of each data dimension feature is mapped to a preset standard range.

10. A server system, characterized in that: The method comprises a server, wherein the server is used to execute the method described in any one of claims 1 to 9.

Images