A method and apparatus for diagnosing cardiac amyloidosis

Abstract

The application provides a prediction method and device for diagnosing cardiac amyloidosis, and applies to the technical field of data processing. The first training sample set is subjected to data cleaning processing to generate a second training sample set; the second training sample set is subjected to missing value supplement processing to generate a target training sample set; a preset cardiac amyloidosis prediction model is processed based on the target training sample set and a verification sample set to generate a target cardiac amyloidosis prediction model; clinical parameter information of a target user is processed to generate target physiological parameter information; life habit information of the target user is processed to generate a dynamic risk adjustment factor; the target physiological parameter information and the dynamic risk adjustment factor are processed based on the target cardiac amyloidosis prediction model to generate early warning information, wherein the early warning information is used to represent the relevance of the target user to cardiac amyloidosis, hypertrophic cardiomyopathy and hypertensive left ventricular hypertrophy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a prediction method and device for diagnosing cardiac amyloidosis. BACKGROUND

[0002] Cardiac amyloidosis (CA) is essentially the deposition of misfolded proteins in the heart, which induces infiltrative problems. Among all types of heart involvement, light chain cardiac amyloidosis (AL-CA) and transthyretin cardiac amyloidosis (ATTR-CA) are the main types. CA is classified as a rare disease, which is characterized by atypical symptoms and lack of specificity.

[0003] Currently, the diagnosis of CA usually relies on echocardiography screening, but the echocardiography score for judging cardiac amyloidosis often involves speckle tracking technology, which puts forward higher requirements for image quality. There are certain difficulties in actual measurement, and it is not simple and easy to operate. Moreover, the existing echocardiography score not only needs to use speckle tracking technology, but also requires high image quality, and needs to be measured after processing. The whole process is complex and cumbersome, and the use and promotion are faced with many obstacles. Therefore, how to better warn the user with cardiac amyloidosis is a technical problem that needs to be solved by the technical personnel in the field.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore includes information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide a prediction method and device for diagnosing cardiac amyloidosis, which at least partially overcomes the problems existing in the prior art. By collecting the information of the target user in multiple aspects and constructing the related data set and model, the foundation for analysis is laid. First, the training sample set is processed, including invalid value and abnormal value detection and correction deletion, and then the missing values are supplemented. The target training sample set is generated through threshold determination, interpolation prediction, backfilling test screening, etc. For the target user, on the one hand, the target physiological parameter information is obtained by processing the clinical parameters by random forest method, and on the other hand, the life habit information is preprocessed. The dynamic risk adjustment factor is generated through feature extraction and score calculation. Finally, the two are integrated to generate a feature vector, which is input into the model after processing. The warning information is generated according to the threshold value, so as to determine the correlation degree of the target user and cardiac amyloidosis and related problems.

[0006] Other characteristics and advantages of the present application will become apparent from the following detailed description, or will be learned by practice of the application.

[0007] According to an aspect of the present application, a method for predicting diagnosis of cardiac amyloidosis is provided, comprising: obtaining clinical parameter information of a target user, life habit information of the target user, a first training sample set, a validation sample set, and a preset cardiac amyloidosis prediction model, wherein the first training sample set comprises clinical parameter information and life habit information of a coronary artery user, clinical parameter information and life habit information of a hypertrophic cardiomyopathy user, and clinical parameter information and life habit information of a hypertensive left ventricular hypertrophy user, the life habit information of the target user comprises dietary intake data of the target user, exercise habit data of the target user, sleep quality data of the target user, drug use history data of the target user, and psychological state data of the target user; performing data cleaning processing on the first training sample set to generate a second training sample set; performing missing value supplement processing on the second training sample set to generate a target training sample set; processing the preset cardiac amyloidosis prediction model based on the target training sample set and the validation sample set to generate a target cardiac amyloidosis prediction model; processing the clinical parameter information of the target user to generate target physiological parameter information, wherein the target physiological parameter information comprises Sokolov-Leonid index, systolic blood pressure, interventricular septal thickness, left ventricular posterior wall thickness, tricuspid annulus systolic displacement value, and ratio of mitral valve blood flow E peak velocity to mitral annulus tissue Doppler e' peak velocity; processing the life habit information of the target user to generate a dynamic risk adjustment factor; processing the target physiological parameter information and the dynamic risk adjustment factor based on the target cardiac amyloidosis prediction model to generate early warning information, wherein the early warning information is used to represent the relevance of the target user to cardiac amyloidosis, hypertrophic cardiomyopathy, and hypertensive left ventricular hypertrophy.

[0008] In another aspect of the present application, a prediction device for diagnosing cardiac amyloidosis comprises: an acquisition module configured to acquire clinical parameter information of a target user, life habit information of the target user, a first training sample set, a verification sample set, and a preset cardiac amyloidosis prediction model, wherein the first training sample set comprises clinical parameter information and life habit information of a coronary artery user, clinical parameter information and life habit information of a hypertrophic cardiomyopathy user, and clinical parameter information and life habit information of a hypertensive left ventricular hypertrophy user, and the life habit information of the target user comprises dietary intake data of the target user, exercise habit data of the target user, sleep quality data of the target user, drug use history data of the target user, and psychological state data of the target user; a processing module configured to perform data cleaning processing on the first training sample set to generate a second training sample set, perform missing value supplement processing on the second training sample set to generate a target training sample set, process the target training sample set and the verification sample set to generate a target cardiac amyloidosis prediction model, process the clinical parameter information of the target user to generate target physiological parameter information, wherein the target physiological parameter information comprises Sokolow-Lyon index, systolic blood pressure, interventricular septum thickness, left ventricular posterior wall thickness, tricuspid annular systolic displacement value, and ratio of mitral valve blood flow E peak velocity to mitral annulus tissue Doppler e' peak velocity, process the life habit information of the target user to generate a dynamic risk adjustment factor, and process the target physiological parameter information and the dynamic risk adjustment factor based on the target cardiac amyloidosis prediction model to generate early warning information, wherein the early warning information is used to represent the relevance of the target user to cardiac amyloidosis, hypertrophic cardiomyopathy, and hypertensive left ventricular hypertrophy.

[0009] According to still another aspect of the present application, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a second processor to implement the prediction method for diagnosing cardiac amyloidosis.

[0010] The prediction method and device for diagnosing cardiac amyloidosis provided by the present application can lay a foundation for analysis by collecting multiple information of a target user and constructing related data sets and models. The training sample set is first processed, including invalid value and abnormal value detection and correction deletion, and then missing values are supplemented to generate a target training sample set through threshold determination, interpolation prediction, backfilling test screening, etc. For a target user, on the one hand, the clinical parameters are processed by a random forest method to obtain target physiological parameter information, and on the other hand, the life habit information is preprocessed to generate a dynamic risk adjustment factor through feature extraction and score calculation. Finally, the two are integrated to generate a feature vector, which is input into the model for processing to generate early warning information according to the threshold, so as to clearly determine the relevance of the target user to cardiac amyloidosis and related problems.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0012] Figure 1 A flowchart illustrating a predictive method for diagnosing myocardial amyloidosis provided in an embodiment of this application is shown.

[0013] Figure 2 A schematic diagram of a predictive device for diagnosing myocardial amyloidosis is shown in one embodiment of this application. Detailed Implementation

[0014] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0015] The following is combined Figure 1 This application describes a predictive method for diagnosing myocardial amyloidosis according to exemplary embodiments thereof. In one embodiment, this application also provides a predictive method and apparatus for diagnosing myocardial amyloidosis. Figure 1 A schematic flowchart illustrating a method for predicting the diagnosis of myocardial amyloidosis according to an embodiment of this application is shown. Figure 1 As shown, this method is applied to a server and includes:

[0016] S101, acquire the target user's clinical parameter information, target user's lifestyle information, first training sample set, validation sample set, and preset myocardial amyloidosis prediction model.

[0017] In one implementation, the clinical parameter information of the target user is as follows: accurately record the basic information of the target user such as age, gender, height, weight, etc. For example, the age is 55 years old, the gender is male, the height is 175 cm, and the weight is 80 kg. These basic information helps to preliminarily understand the physical characteristics of the target user, and provides basic data for subsequent analysis, such as age is one of the important risk factors for cardiovascular problems, and the risk of disease and problem performance is different in different age groups; gender also affects the incidence and symptom performance of some cardiovascular problems. Using professional blood pressure measuring equipment (such as electronic sphygmomanometer), according to the standard measurement method, the systolic pressure and diastolic pressure of the target user are measured in a quiet state, recorded as 130 / 85 mmHg, and the heart rate is measured as 75 times / minute. These data can directly reflect the instantaneous function state of the cardiovascular system, and high blood pressure is a common risk factor for myocardial amyloidosis, coronary artery problems and other cardiovascular problems, while abnormal heart rate (such as tachycardia or bradycardia) is also related to heart problems, such as myocardial lesions leading to arrhythmia. Get the cardiac function classification information of the target user, such as NYHA cardiac function classification II. This classification information can directly reflect the degree of impairment of its cardiac function, and also provides a reference for the overall function of the heart for subsequent analysis.

[0018] Through echocardiography, the left atrial size (LA) is accurately measured as 35 mm, the left ventricular size (LV) is 50 mm, the interventricular septal thickness (IVS) is 10 mm, the left ventricular posterior wall thickness (LVPW) is 10 mm, the right atrial size (RA) is 30 mm, the right ventricular size (RV) is 35 mm, and other parameters. These parameters can directly reflect the morphological structure of the heart, which is crucial for diagnosing myocardial amyloidosis, hypertrophic cardiomyopathy, and hypertensive left ventricular hypertrophy, etc. For example, myocardial amyloidosis leads to myocardial thickening, and the interventricular septum and left ventricular posterior wall thickness increase; hypertrophic cardiomyopathy is typically characterized by asymmetric hypertrophy of the ventricular wall. Get the left ventricular short-axis fractional shortening (FS%) of 30%, the left ventricular ejection fraction (EF%) of 55%, the degree of mitral regurgitation of mild, the degree of aortic valve regurgitation of no, the degree of tricuspid valve regurgitation of mild, the tricuspid valve regurgitation velocity of 2.5 m / s, the inferior vena cava width of 15 mm, the E / A of 1.2, the E / e' of 8, the tricuspid annular plane systolic excursion (TAPSE) of 18 mm, the pericardial effusion of no, the diastolic dysfunction classification of I, and the pulmonary artery systolic pressure of 30 mmHg. These parameters comprehensively evaluate the systolic and diastolic function of the heart, and impaired cardiac function is a core feature of cardiovascular problems, and different problems have different characteristics on the heart function. These parameters help to accurately distinguish the changes of heart function in different states, such as myocardial amyloidosis often accompanied by diastolic dysfunction, which is manifested by abnormal elevation of E / e' ratio.

[0019] The electrocardiogram records accurately measure the heart rate of 75 beats per minute, the PR interval of 0.18 seconds, the QRS interval of 0.08 seconds, and the QTc interval of 0.40 seconds. These basic interval parameters reflect the process of electrical physiology of the heart, and abnormal interval values suggest abnormalities in the cardiac conduction system or myocardial lesions, which are one of the important bases for diagnosing cardiovascular problems. For example, an extended QTc interval increases the risk of arrhythmia and is associated with cardiovascular problems such as myocardial lesions. The Sokolow-Lyons index in the electrocardiogram record is 3.5 mV, with low voltage in the limb lead and low voltage in the chest lead, false infarction, poor R wave progression, atrioventricular block, left bundle branch block (LBBB), right bundle branch block (RBBB), atrial fibrillation / flutter, and other indicators. These special electrocardiogram indicators reveal abnormal electrical activity of the heart from different angles and are closely related to changes in cardiac structure and function, and have important value in assisting the diagnosis of myocardial amyloidosis and other problems. For example, myocardial amyloidosis users exhibit QRS low voltage on electrocardiogram, and changes in the Sokolow-Lyons index are also related to myocardial lesions.

[0020] Blood samples were collected, and the level of troponin (TnI) was 0.05 ng / mL (the normal range is generally less than 0.1 ng / mL), which indicates that the myocardial cells are damaged, and the abnormal increase occurs in the process of myocardial amyloidosis, myocardial infarction, and other myocardial lesions, and is a sensitive indicator for early diagnosis of myocardial injury. The brain natriuretic peptide (BNP) was 100 pg / mL (the normal range varies depending on the detection method and reagent, generally less than 100-300 pg / mL), and the N-terminal pro-brain natriuretic peptide (NT-ProBNP) was 400 pg / mL (the normal range varies depending on factors such as age, generally slightly higher in the elderly, but this value reflects the cardiac function and load status to some extent), which significantly increases in the case of impaired cardiac function such as heart failure, and is helpful in assessing the severity and prognosis of cardiovascular problems.

[0021] Detecting alanine aminotransferase (ALT) of 40 U / L (normal range is generally 5-40 U / L), aspartate aminotransferase (AST) of 35 U / L (normal range is generally 8-40 U / L), urea nitrogen (BUN) of 5 mmol / L (normal range is generally 2.9-8.2 mmol / L), creatinine of 80 μmol / L (normal range is generally 53-106 μmol / L in men), uric acid (UA) of 400 μmol / L (normal range is generally 208-428 μmol / L in men), potassium (K+) of 4.0 mmol / L (normal range is generally 3.5-5.5 mmol / L), sodium (Na+) of 140 mmol / L (normal range is generally 135-145 mmol / L), chloride (Cl-) of 100 mmol / L (normal range is generally 96-108 mmol / L), blood glucose of 5.5 mmol / L (fasting normal range is generally 3.9-6.1 mmol / L), and other indicators reflect the overall metabolic status of the target user. Abnormal liver and kidney function affects drug metabolism and the internal environment stability of the cardiovascular system, for example, renal insufficiency leads to accumulation of toxins in the body, affecting heart function; electrolyte imbalance (such as hyperkalemia or hypokalemia) can directly affect cardiac electrical activity, causing arrhythmia; dyslipidemia (such as high cholesterol, high triglycerides, low-density lipoprotein elevation, and high-density lipoprotein reduction) is an important risk factor for atherosclerosis and is closely related to the occurrence and development of cardiovascular problems such as coronary artery problems. The detection of high-sensitivity C protein (hsCRP) of 3 mg / L (normal range is generally less than 3 mg / L) can be used as a marker of inflammation in the body, and chronic inflammation plays an important role in the occurrence and development of cardiovascular problems, and inflammatory response promotes the formation and instability of atherosclerotic plaques, increasing the risk of cardiovascular events.

[0022] In addition, at least one week of dietary information such as staple food, meat, vegetables, fruits, fat, and salt intake is obtained through questionnaires and mobile application records for evaluating nutritional intake and dietary patterns related to cardiovascular health. In combination with wearable device records and user reports, information such as exercise type, frequency, intensity, and duration, such as 4 times of exercise per week, including aerobic exercise and strength training, and running heart rate, which has an important impact on cardiovascular health, is obtained. Continuous sleep data for one week is obtained using a sleep monitoring device, including sleep duration, stage duration, number of interruptions, sleep onset and wake-up time, and poor sleep quality increases the risk of cardiovascular problems. All drug information is collected, including prescription drugs, non-prescription drugs, and health products, and the dosage form, usage, source, and drug use are recorded, which is of great significance to cardiovascular risk assessment. Regular psychological assessment is used to record scores and life events that affect mental state, and adverse psychological state is closely related to cardiovascular problems, which helps to comprehensively assess the risk.

[0023] The first training sample set is constructed by retrospectively selecting clinical data of coronary artery (CA) users, hypertrophic cardiomyopathy (HCM) users, and hypertension left ventricular hypertrophy (HTN-LVH) users from the medical record database of the cooperative hospital as the main source of the first training sample set. Ensure the integrity and accuracy of the data, including the user's clinical parameter information (such as various clinical indicators collected as described above) and lifestyle information (such as diet, exercise, sleep, drug use, psychological state, etc.). For example, data of 200 CA users, 150 HCM users, and 120 HTN-LVH users are obtained, covering information of users with different types of cardiovascular problems, providing diversified data support for model training. The sample inclusion criteria are strictly implemented, specifically:

[0024] CA users: the inclusion criteria are that the discharge diagnosis contains "myocardial amyloidosis" and is confirmed as AL-CA or ATTR-CA by tissue biopsy or genetic testing; the exclusion criteria are that the results of electrocardiogram or echocardiogram are missing and the type of amyloidosis cannot be determined. Ensure that the CA user samples included have clear diagnostic basis, complete and reliable data, and avoid affecting the model training effect due to data missing or unclear diagnosis.

[0025] HCM users: the inclusion criteria are that the left ventricular maximum wall thickness is ≥15mm, or the left ventricular maximum wall thickness of users with HCM family history is ≥13mm, or genetic testing confirms HCM; the exclusion criteria are ischemic heart disease, structural heart disease, heart valve disease, or other problems that can cause left ventricular hypertrophy (such as metabolic cardiomyopathy, etc.). By strictly selecting and excluding criteria, typical HCM user samples are selected, reducing the interference of other problem factors on model training, and improving the model's learning ability for HCM characteristics.

[0026] HTN-LVH users: the inclusion criteria are long-term history of hypertension with high blood pressure and poor blood pressure control; the exclusion criteria are severe heart valve disease (such as moderate to severe aortic valve stenosis) and severe renal dysfunction (eGFR <30ml·min-1·1.73m-2). This ensures that the HTN-LVH user samples included are mainly affected by hypertension leading to left ventricular hypertrophy, excluding the influence of other serious problems on heart structure and function, making the samples more representative and conducive to the model accurately learning the characteristics of HTN-LVH.

[0027] The selected user data is organized, and the clinical parameter information and lifestyle habit information of each user is organized in a unified format. At the same time, each sample is labeled with the corresponding problem category (CA, HCM or HTN-LVH) so as to be used for subsequent model training. For example, the user's age, gender, various clinical index values and lifestyle habit data are organized into a data record, and the problem category to which the user belongs is explicitly labeled in the record to form a structured first training sample set, providing accurate input data and classification labels for the model. The validation sample set is constructed by using stratified sampling method to extract a certain proportion (such as 20%-30%) of data from the original user data as the validation sample set, ensuring that the proportion of each problem (CA, HCM, HTN-LVH) in the validation sample set is similar to that in the first training sample set. In this way, the validation sample set can better represent the feature distribution of the entire user group, effectively evaluating the generalization ability of the model. For example, if the proportion of CA, HCM and HTN-LVH users in the first training sample set is 4:3:3, then when extracting the validation sample set, the corresponding number of users of each type is extracted according to approximately the same proportion to ensure the representativeness of the validation sample set. Ensure that the validation sample set and the first training sample set are independent of each other, that is, the user data in the validation sample set is not included in the first training sample set, avoiding data leakage and overfitting problems. Through the independent validation sample set, the performance of the model is objectively evaluated during the model training process, and whether the model has overfitting or underfitting phenomenon is found in time, so as to adjust the model parameters or improve the model structure, improve the accuracy and generalization ability of the model.

[0028] For the preset myocardial amyloidosis prediction model, considering that the myocardial amyloidosis prediction problem involves multi-dimensional clinical parameter information and lifestyle information, and needs to accurately distinguish different types of cardiovascular problems, a machine learning model with the ability to handle complex data relationships and multi-classification is selected as the preset model. For example, decision trees and their derivative models (such as random forests, gradient boosting trees, etc.) have good performance in handling non-linear relationships between features and classification problems, can classify and predict users according to different combinations of features, and are suitable for the situation where multiple clinical and lifestyle factors jointly affect the occurrence and development of the problem. The random forest model can effectively handle high-dimensional data, avoid overfitting, and provide feature importance evaluation, which is helpful to understand the contribution of different factors to problem prediction, so it is selected as one of the preset models. According to the characteristics of the selected model and the experience value, the key parameter range of the model is initially set. For example, for the random forest model, the number of decision trees is set to 100-500, the depth of the tree is set to 5-15 layers (to avoid overfitting caused by too deep, and too shallow cannot learn the data characteristics), the number of features selected when splitting the node is (n is the total number of features, that is, the total number of features of clinical parameter information and lifestyle information) and other parameter ranges. These initial parameter settings provide a basis for subsequent model training and optimization, and by adjusting these parameters, the model can better adapt to the data characteristics and improve the prediction performance.

[0029] Random initialization or empirical distribution-based initialization methods are used to assign initial values to model parameters. For example, for the initial split point selection of decision trees, the threshold value of the feature is randomly selected for initialization, and then adjusted according to the data characteristics and optimization objectives during the model training process. Initial parameter setting speeds up the convergence speed of model training, reduces the training time, and also affects the final performance of the model. In order to comprehensively evaluate the performance of the preset model, multiple evaluation indicators are selected, including accuracy (the proportion of samples predicted correctly by the model to the total number of samples), recall rate (the proportion of true positive samples predicted as positive samples by the model), F1 score (an indicator that considers both accuracy and recall, which can balance the performance of the model on different classes), AUC-ROC curve (by calculating the True Positive Rate (TPR) and False Positive Rate (FPR) at different thresholds, the model's ability to distinguish different classes of samples is evaluated, and the AUC (Area Under Curve) value quantifies the overall classification performance of the model). The purpose of selecting these evaluation indicators is to comprehensively measure the accuracy, reliability and discrimination ability of the preset myocardial amyloidosis prediction model for different classes of problems (such as myocardial amyloidosis, hypertrophic cardiomyopathy, hypertensive left ventricular hypertrophy, etc.) from multiple angles during model training and optimization, so as to adjust and improve the model according to the evaluation results, improve the performance of the model, and make it more effective for actual problem prediction and diagnosis tasks. The complete expression is to comprehensively evaluate the performance of the preset model by selecting multiple evaluation indicators, including accuracy, recall rate, F1 score, AUC-ROC curve (by calculating the True Positive Rate and False Positive Rate at different thresholds, the model's ability to distinguish different classes of samples is evaluated, and the AUC value quantifies the overall classification performance of the model), etc., so as to measure the performance of the model from multiple angles during model training and optimization.

[0030] S102, performing data cleaning processing on the first training sample set to generate a second training sample set.

[0031] In one implementation, the first training sample set is subjected to invalid value detection processing to generate an initial training sample set. Each data record in the first training sample set is traversed row by row, and the data format of each field (corresponding to different clinical parameter information and lifestyle information) in each record is checked. For example, for a field recording a user's blood pressure value, it is checked whether it is in a numerical format, whether it contains a unit (such as mmHg) and whether the unit is correct. If data of an unstandardized format is found, such as "120 / ? mmHg" (containing an invalid character "?") or "150 kpa" (the unit is incorrect, which should be mmHg), the record is marked as an invalid value candidate. In order to more efficiently and accurately perform format checking, a format specification library is established in advance, which defines the format of each clinical parameter and lifestyle parameter in detail. For example, age should be a positive integer, height should be a numerical value with a unit of centimeters (cm), and drug names should follow the standard generic name format of drugs, etc. In the checking process, the data format of each field is compared with the definition in the format specification library, and any that do not meet the specification are invalid value candidates.

[0032] According to medical professional knowledge, a reasonable value range is set for each clinical parameter. For example, the normal range of systolic blood pressure is generally between 90-240 mmHg, the normal range of heart rate is 60-100 beats per minute, the normal range of tricuspid annular plane systolic excursion value for males is generally 53-106 pmol / L, and for females, it is 44-97 pmol / L, etc. For the exercise steps in lifestyle information, it is generally between 0-50,000 steps per day, and the salt intake should not exceed 6 grams per day, etc. For the fields in each data record, it is checked whether the value is within the corresponding reasonable value range. If data is found that is outside the reasonable range, such as a systolic blood pressure record of "300 mmHg" (significantly higher than the upper limit of the normal range) or an exercise step record of "100,000 steps", the record is marked as an invalid value candidate. At the same time, for data near the critical value (such as a systolic blood pressure of 85 mmHg, close to the lower limit of the normal range), special marking or further verification is performed to determine its accuracy.

[0033] For the records marked as invalid value candidates, direct deletion of these records is selected to ensure the data quality in the sample set. But before deletion, the impact of data volume needs to be considered carefully, if the number of invalid value candidate records is large (more than a certain proportion, such as 10%), further analysis is needed to find out the reason, whether there is a problem in the data collection process, or whether the data needs to be re-collected. In addition, some correctable invalid values are also corrected, for example, for data with format errors but the data itself can be recognized (such as "150kpa", which can be converted to the correct mmHg unit according to the conversion formula), format correction is performed and retained in the sample set; for data with obvious errors but with certain clues (such as age record "200 years old", which is an input error, if there is other information that can be referred to, such as birth date, etc., it can be verified and corrected), modification is performed under the premise of ensuring accuracy, otherwise it is deleted.

[0034] The initial training sample set is subjected to outlier detection processing to generate a second training sample set. For numerical data in the initial training sample set (such as blood pressure, heart rate, test index values in clinical parameters, and exercise data, dietary intake, etc. in lifestyle information), a visualization method is selected to observe the data distribution. Common methods include drawing histograms, box plots, scatter plots, etc. For example, a histogram of systolic blood pressure is drawn to observe whether its distribution pattern is approximately normally distributed, whether there are obvious skewness or outliers; a scatter plot of exercise steps and age is drawn to view the distribution of exercise steps in different age groups, whether there are data points that deviate significantly from the overall trend. Through the visualization results, it is preliminarily judged whether there are outliers in the data and the distribution trend of the outliers. For example, in the histogram of systolic blood pressure, if it is found that there are individual extremely high or extremely low columns far from the central area of most data, these data are outliers; in the scatter plot, if a point is far from the distribution range of other points, it is also an outlier. Visualization analysis helps to intuitively understand the overall characteristics of the data, and provides a reference basis for subsequent selection of appropriate outlier detection methods.

[0035] The descriptive statistics of each numerical variable in the initial training sample set are calculated, such as mean, median, standard deviation, interquartile range (IQR), etc. For example, the mean of tricuspid annular plane systolic excursion value is 80 μmol / L, the median is 75 μmol / L, the standard deviation is 10 μmol / L, and the interquartile range is 15 μmol / L (Q3-Q1, Q3 is the upper quartile, and Q1 is the lower quartile). These statistics will be used for subsequent outlier judgment. Normally distributed data (such as test indicators): For data that is approximately normally distributed, a method based on standard deviation is used to detect outliers. Data that is more than a certain multiple of the standard deviation (such as 3 times the standard deviation) away from the mean is generally considered an outlier. For example, assuming that the tricuspid annular plane systolic excursion value is approximately normally distributed, according to the mean and standard deviation calculation, if a user's tricuspid annular plane systolic excursion value is greater than the mean + 3 x standard deviation (i.e. 80 + 3 x 10 = 110 μmol / L) or less than the mean - 3 x standard deviation (i.e. 80 - 3 x 10 = 50 μmol / L), the value is considered an outlier. Non-normally distributed data (such as exercise steps, etc.): For non-normally distributed data, a method based on the interquartile range (IQR) can be used. In general, data less than Q1-1.5 x IQR or greater than Q3+1.5 x IQR is defined as an outlier. For example, for exercise step data, if Q1 = 5000 steps, Q3 = 10000 steps, and IQR = 5000 steps are calculated, then exercise step data less than 5000-1.5 x 5000 = -2500 steps or greater than 10000+1.5 x 5000 = 17500 steps can be considered an outlier.

[0036] For the detected outliers, first verify their authenticity. Check the original data record, ask the data collection personnel or communicate with the clinician to understand whether there is a data collection error, measurement error or special situation. If it is an outlier caused by data entry error or measurement error, correct it after verification to restore it to a reasonable value range. For example, if it is found that the systolic pressure of a certain user is recorded as "2000mmHg", which is obviously an entry error, and after verification it should be "120mmHg", it should be corrected to the correct value. If the reasonable reason of the outlier cannot be determined after verification, and the existence of the outlier has a great interference on the model training (such as deviating from the data distribution, affecting the learning of the model to the normal data pattern), delete the outlier. However, when deleting outliers, attention should be paid to the influence of data quantity to avoid excessive deletion leading to the loss of representativeness of the sample set. If the number of outliers is large (more than a certain proportion, such as 5%), the data collection and processing process needs to be reexamined to find the root cause, and additional data or improved data collection methods are needed. For some outliers with special clinical significance, they cannot be simply deleted or corrected, but need to be specially marked or analyzed separately. For example, for some rare disease users or users in special stages, their clinical parameters show outliers, but these outliers reflect the physiological process of special pathology. In this case, these outliers are marked as special categories, and are separately processed in the model training process, or the influence of these special data is studied separately when analyzing the model results.

[0037] S103, performing missing value supplement processing on the second training sample set to generate a target training sample set.

[0038] In one implementation, a missing value threshold for each clinical variable and a missing value threshold for each lifestyle variable are obtained. For clinical variables, such as age, gender, blood pressure, heart rate, various test indicators, etc., a preliminary analysis is performed on the data distribution thereof. Whether the data of each clinical variable is approximately normally distributed, skewed distributed, or other special distribution form is observed. For example, age is approximately normally distributed, while some related test indicators (such as the concentration of a specific protein in serum) exhibit skewed distribution. At the same time, considering the importance and clinical significance of the variable, some key clinical variables (such as heart function indicators) have a low tolerance for missing values, while some auxiliary indicators are relatively lenient. For lifestyle variables, such as dietary intake data, exercise habit data, sleep quality data (sleep duration, sleep interruption frequency, etc.), drug use history data, and psychological state data (psychological evaluation scale score, etc.), the data distribution characteristics thereof are also analyzed. Lifestyle data is more diverse and complex, with some data being categorical variables (such as exercise type) and some being numerical variables (such as exercise duration). For example, there are multiple categories of exercise types, and the distribution is uneven; while sleep duration fluctuates within a certain range, but there are also some abnormal values (such as extremely short or long sleep duration). Considering the potential impact of these variables on problem prediction, the relative importance of each variable in missing value processing is determined.

[0039] According to the above analysis, for clinical variables with a large amount of data and relatively stable distribution, if the missing proportion is too high, it will seriously affect the model training effect, therefore, a relatively strict missing value threshold can be set. For example, for basic information such as age, gender, and key clinical indicators (such as systolic blood pressure, heart rate, etc.), if the missing proportion exceeds 10%, it is considered that the missing value of the variable is too much, and needs to be used carefully in subsequent analysis. For some clinical variables with small data volume or difficult to obtain, such as some special test indicators related to rare problems, the missing value threshold can be appropriately relaxed, such as setting it to 20%-30%, but the influence of missing value on the result needs to be paid attention to in subsequent analysis. The missing value threshold of lifestyle variables can be adjusted according to the nature and importance of the variable. For some lifestyle variables that are relatively easy to obtain and have an important influence on problem prediction, such as daily salt intake (which is closely related to cardiovascular problems such as hypertension), a lower missing value threshold can be set, such as 15%. While for some lifestyle variables that are difficult to accurately record or have a relatively small influence on the problem, such as some subtypes of exercise type (such as specific yoga poses, etc.), a higher missing value threshold can be set, such as 30%-40%. For missing value processing of categorical variables, the number of categories and distribution should be considered comprehensively. If the missing proportion of a certain category is too high, it will lead to insufficient representation of the category in model training, affecting the learning of the model on the features related to the category.

[0040] The second training sample set is processed based on a missing value threshold of each clinical variable and a missing value threshold of each lifestyle variable to generate missing sample information, where the missing sample information is used to represent samples with missing values and the location and variable name of the missing values. Each sample in the second training sample set is checked one by one, and for the clinical variables and lifestyle variables in each sample, a pre-set missing value threshold is used for judgment. For example, for a sample containing user age, systolic pressure, dietary intake, and the like, the clinical variables such as age and systolic pressure and the lifestyle variables such as various food intake amounts related to dietary intake are checked in turn to determine whether there are missing values. When a missing value is found in a certain variable, the identification (such as sample number) of the sample, the location (i.e., the position of the variable in the sample data structure, such as the element number in a data vector containing multiple clinical and lifestyle variables) of the missing value, and the variable name are recorded. For example, if it is found that the sample number 100 has missing systolic pressure data, it is recorded as (sample 100, position 3, variable name: systolic pressure); if it is found that the fruit intake amount of the same user's dietary intake is missing, it is recorded as (sample 100, position 15, variable name: fruit intake amount). All samples with missing values and their related missing information are arranged into a list or data structure to form missing sample information for subsequent targeted processing.

[0041] The missing sample information is processed to generate missing value imputation prediction information. A suitable imputation method is selected according to the data distribution characteristics of the clinical variables and the lifestyle variables. For clinical variables with approximately normal distribution, a mean-based imputation method (such as replacing the missing value with the mean of the variable) or a regression model-based imputation method (using other related variables as independent variables to establish a regression equation to predict the missing value) is used. For example, for the tricuspid annular plane systolic excursion value, which is a clinical variable with approximately normal distribution, if there is a missing value, the mean of the tricuspid annular plane systolic excursion values of other users is used for imputation, or a linear regression model is established with age, gender, weight, and other related variables as independent variables to predict the missing value. For lifestyle variables with non-normal distribution (such as exercise steps, which have a large number of zero values and a small number of high values, showing skew distribution), a median imputation or a quantile regression-based imputation method is used. For example, for missing values of exercise steps, the median of all sample exercise steps is used for imputation, or a quantile regression model is used to predict the missing value based on other related factors (such as age, exercise habit category, etc.).

[0042] Meanwhile, the correlation between variables is considered for the imputation. For example, there is a certain correlation between systolic blood pressure and age, weight, eating habits (such as salt intake), etc. If there is a missing value of systolic blood pressure, a multiple regression model or other suitable model (such as a decision tree model, a random forest model, etc., which can handle the nonlinear relationship between variables) is established using these related variables to predict the missing value. For lifestyle variables, such as exercise duration and exercise type, if the exercise duration is missing, the exercise duration distribution of users with similar exercise types and other users with similar exercise types is used for imputation. According to the selected imputation method, the effective data in the second training sample set is used to calculate the imputation prediction information of the missing value. Using the regression model-based imputation method, the form of the model (such as linear regression, logistic regression, decision tree regression, etc.) and the independent variables (i.e., other variables related to the missing variable) need to be determined first. For example, taking the missing value of tricuspid annular plane systolic excursion as an example, age, gender, weight, urea nitrogen, etc. are selected as independent variables to establish a linear regression model Y = β0+ β1X1+ β2X2+... + β n X n (Where Y is the tricuspid annular plane systolic excursion value, X1 is the age, X2 is the gender, X3 is the weight,..., X n is the urea nitrogen, β0, β1,..., β n are the regression coefficients), and the regression coefficients are estimated by least squares method, etc. Then, the known values of the independent variables in the missing value sample are substituted into the model to calculate the predicted value of the missing value. For lifestyle variables using median imputation, the median of all valid samples of the variable is directly calculated as the predicted value of the missing value.

[0043] The imputation prediction information of each missing value calculated is arranged into a set or data structure to form the missing value imputation prediction information. This information set contains the predicted value of each missing value and the sample identification, variable position, etc. related to it, so that the predicted value can be accurately backfilled into the original sample to generate complete target variable values. For example, for the missing value of tricuspid annular plane systolic excursion of sample No. 100, the imputation prediction information is (sample 100, position 5, predicted value: 85 μmol / L). All such imputation prediction information is summarized to form the missing value imputation prediction information set, which is prepared for the next step of generating target variable values.

[0044] The missing value imputation prediction information is processed to generate target variable values. According to the sample identifier and variable position information in the missing value imputation prediction information, the corresponding missing value position in the second training sample set is accurately found. For example, for the missing value mentioned above with the sample number 100 and the tricuspid annular plane systolic excursion value position 5, the 5th variable position of the sample is located in the original second training sample set. The predicted value in the imputation prediction information is backfilled to the corresponding missing value position to replace the original missing value. For example, the calculated tricuspid annular plane systolic excursion value prediction value 85 μmol / L is backfilled to the missing position of the tricuspid annular plane systolic excursion value of the sample with the sample number 100, so that the tricuspid annular plane systolic excursion value variable in the sample becomes complete. For all samples and variables with missing values, backfilling operation is performed in the same way, so as to obtain a complete data set without missing values (or with missing values replaced by predicted values), i.e. a target variable value set. Each sample in the set contains complete clinical variable values and lifestyle variable values, providing a complete data basis for subsequent statistical tests and model training.

[0045] The target variable value is processed based on a preset statistical test method to generate target variable parameter information. The preset statistical test method is selected according to the type (numerical type or category type) of the clinical variable and the lifestyle habit variable. For numerical variables such as age, blood pressure, and test index values, one-way ANOVA or independent sample t-test (when comparing two groups of data) is selected to test whether there is a significant difference in the mean value of the variable between different groups (such as different problem type user groups); for non-normal distribution numerical variables, Kruskal-Wallis H test is selected. For example, to compare whether there is a difference in the mean value of systolic blood pressure among three groups of people with cardiac amyloidosis, hypertrophic cardiomyopathy, and hypertensive left ventricular hypertrophy, if the systolic blood pressure is approximately normally distributed, one-way ANOVA can be used; if the distribution is not clear or does not conform to the normal distribution, Kruskal-Wallis H test is used. For categorical variables such as gender, drug use category, and exercise type, chi-square test is used to test whether there is a significant difference in the distribution of the categorical variable between different groups. For example, to test whether there is a difference in the proportion of men and women among different problem type users, chi-square test is used. The significance level of all statistical tests is set, and a value less than 0.01 is usually selected as the standard for judging whether the difference is significant. This means that when the calculated test statistic value is less than 0.01, it is considered that there is a significant difference in the variable between different groups, and the variable has an important influence on problem prediction, which needs to be focused on in subsequent analysis. For example, when performing one-way ANOVA to compare the tricuspid annular systolic displacement values of three groups of users, if the calculated value is less than 0.01, it means that there is a significant difference in the tricuspid annular systolic displacement values of the three groups of users, and the tricuspid annular systolic displacement value is an important prediction variable that should be considered when constructing a model or diagnosing a problem.

[0046] For numerical variables, the samples are grouped according to different symptom types (such as cardiac amyloidosis, hypertrophic cardiomyopathy, and hypertensive left ventricular hypertrophy) or other related grouping factors, and then the mean, standard deviation, and other descriptive statistics of each group are calculated, and the selected statistical test (such as one-way ANOVA or Kruskal-Wallis H test) is performed. For example, the collected user samples are divided into three groups, the mean and standard deviation of systolic blood pressure of each group are calculated, and one-way ANOVA is performed to obtain the test statistic value and value. By comparing the value with the set significance level (0.01), it is determined whether there is a significant difference in systolic blood pressure between different symptom groups. If the value is less than 0.01, it is considered that there is a significant difference in systolic blood pressure between different groups, and it is an important variable for distinguishing different symptoms or predicting symptom risk; if the value is greater than 0.01, the difference in systolic blood pressure between different groups is not significant, and its role in the current analysis of symptom prediction is relatively small.

[0047] For categorical variables, a contingency table is constructed, and the sample is cross-classified according to different groupings and categories of the categorical variable, and then the chi-square statistic and the corresponding value are calculated. For example, for the gender categorical variable, a 3x2 contingency table is constructed (3 rows represent three symptom types, and 2 columns represent men and women), the number of men and women in each symptom type is counted, the chi-square statistic is calculated, and the corresponding value is found according to the degrees of freedom and the chi-square distribution table. Similarly, the value is compared with 0.01 to determine whether there is a significant difference in the distribution of gender among different symptom groups. If the value is less than 0.01, it means that the distribution of gender in different symptom groups is different, and there is a correlation with the development of the symptom, which needs to be considered in the subsequent analysis; if the value is greater than 0.01, the difference in the distribution of gender in different symptom groups is not significant, and gender is not a key predictor variable in the current analysis.

[0048] The results of the statistical test are organized and recorded to generate target variable parameter information. For each variable, record its test method, test statistic value (such as F value, X 2 value, p value, mean (for numerical variables), standard deviation (for numerical variables), and distribution in different groups (for categorical variables), etc. For example, for the systolic blood pressure variable, it is recorded as (test method: one-way ANOVA, F value: 5.23, p value: 0.005, mean of cardiac amyloidosis group: 130 mmHg, standard deviation: 15 mmHg; hypertrophic cardiomyopathy group mean: 135 mmHg, standard deviation: 18 mmHg; left ventricular hypertrophy group mean: 140 mmHg, standard deviation: 20 mmHg). These target variable parameter information comprehensively reflects the difference and distribution characteristics of each variable between different groups, providing an important basis for subsequent model construction and variable selection, and ignoring or further analyzing the relationship between variables to some extent.

[0049] Based on the target variable parameter information, a target training sample set is generated. According to the test results in the target variable parameter information, the importance of each variable is evaluated. Generally, variables with p values less than 0.01 are considered to have significant differences and are given priority; while variables with p values greater than 0.01 are relatively unimportant in the current analysis, but are not absolutely excluded and need to be further considered. For example, if the p value of a certain clinical test indicator is greater than 0.01, and the mean difference in different symptom groups is small and the standard deviation is large, it means that the variable has weak discrimination ability for the symptom, and it is considered to be deleted from the sample set to reduce the data dimension and improve the model training efficiency. However, when deleting variables, the correlation between variables needs to be considered to avoid deleting variables that are highly correlated with important variables, resulting in loss of information.

[0050] In addition to the p-value, the clinical significance of the variable, the completeness and availability of the data, and other factors need to be considered. For example, although the p-value of some variables is slightly larger than 0.01, it has important clinical significance (such as an index related to a rare symptom), or the data completeness of the variable is better (with less missing values), the variable is considered to be retained. At the same time, for some highly correlated variables, dimensionality reduction processing is performed through principal component analysis (PCA) and other methods to convert multiple correlated variables into a few uncorrelated principal components, reducing the number of variables while retaining most of the information. For example, if it is found that systolic blood pressure and diastolic blood pressure are highly correlated, they are converted into one or two principal components through PCA, which are used as new variables in the target training sample set, reducing the dimensionality of the variables while retaining the contribution of blood pressure information to symptom prediction. The sample data after variable screening is arranged into a structure suitable for model training. Usually, the clinical variable values and lifestyle variable values of each sample are arranged in a certain order to form a vector or data matrix, where each row represents a sample and each column represents a variable. For example, for a sample set containing age, gender, systolic blood pressure, exercise frequency, and salt intake in diet, the data of each sample is arranged into a vector [age, gender, systolic blood pressure, exercise frequency, salt intake,...], and all samples form a matrix as the target training sample set. Ensure that the data format and order in the sample set are consistent with the requirements of the subsequent model training algorithm, so that the data can be smoothly input into the model for training. After constructing the target training sample set, standardization processing is also considered to make different variables have the same scale and order of magnitude.

[0051] In another embodiment, the application also includes a calculation formula for obtaining the missing value interpolation prediction information, and the calculation formula is: wherein, represents the prediction of the missing value at the mth interpolation, Y i represents the prediction value based on the regression model, and ε (m) represents a random error term. Suppose there are missing values when processing the systolic blood pressure data of a user. These missing values are predicted by establishing a regression model, and the independent variables of the regression model include the user's age, gender, weight, and other factors. i is a preliminary systolic blood pressure prediction value based on the regression model according to these independent variables. For example, for a 50-year-old male user weighing 80 kg, the preliminary predicted systolic blood pressure Y i is 130 mmHg. In each interpolation process, ε (m) will fluctuate randomly. For example, the first interpolation ε (1) is +5 mmHg, so the predicted systolic blood pressure Y after the first interpolation is 130+5=135 mmHg; the second interpolation ε(2) is -3 mmHg, then is 130-3=127 mmHg. By interpolating multiple times and considering different random situations, the prediction of the missing value of systolic pressure is more accurate and robust.

[0052] In another embodiment, the application further includes a calculation formula for obtaining the target variable value, and the calculation formula is: wherein x i represents the original variable value, μ i represents the mean of the variable x i , and σ i represents the standard deviation of the variable x i . Suppose there is a set of original values x i of left ventricular posterior wall thickness of users. First, the mean μ i and the standard deviation σ i of the set of left ventricular posterior wall thickness values are calculated. For example, the mean μ i of the left ventricular posterior wall thickness of the users is 12 mm, and the standard deviation σ i is 2 mm. For an original value x i of left ventricular posterior wall thickness of a user, which is 14 mm, the standardized value is calculated by the formula as (14-12) / 2=1. The purpose of standardization is to make the variable of left ventricular posterior wall thickness comparable in value with other clinical variables (such as other structural sizes of heart, functional indicators, etc.), convert them to a standard scale, and avoid bias in subsequent data analysis, model construction, etc. due to the difference in the original numerical magnitude.

[0053] S104, processing the preset myocardial amyloidosis prediction model based on the target training sample set and the verification sample set to generate a target myocardial amyloidosis prediction model.

[0054] In an embodiment, any number of data features in the target training sample set is obtained, and a sampling ratio is generated based on the number of each data feature in the target training sample set. A certain number of data features are randomly selected from the target training sample set. Assuming that there are 100 data features (features include various variables related to clinical parameter information and lifestyle habit information, such as age, systolic blood pressure, exercise frequency in exercise habit, salt intake in diet, etc.) in the target training sample set, 20 data features are randomly selected as the basis for subsequent processing. This random selection method avoids the bias of feature selection to some extent, and ensures that the model can learn the relationship between different types of features. By selecting part of the data features, the data dimension is reduced, the calculation amount is reduced, the model training efficiency is improved, and the overfitting problem caused by too many irrelevant or redundant features is avoided. For example, if feature selection is not performed and all 100 features are directly used for model training, the model may be too complex to learn the real pattern in the data, and is easily disturbed by noise features. By selecting 20 representative data features, the model focuses on the factors that have the greatest impact on symptom prediction, making the model more concise and effective.

[0055] The number of each of the 20 selected data features in the target training sample set is counted (for example, there are 200 sample values for the age feature and 180 sample values for the systolic blood pressure feature, etc.). The proportion of the number of each feature in the total number of samples (assuming that the total number of samples is 500) is calculated as the sampling ratio of the feature. For example, the sampling ratio of the age feature is 200÷500=0.4, the sampling ratio of the systolic blood pressure feature is 180÷500=0.36, etc. These sampling ratios will be used in the subsequent sampling process to ensure that the sampled samples can retain the feature distribution of the original sample set. The determination of the sampling ratio makes the sampling process more scientific and reasonable, and ensures that different features are extracted according to their importance (reflected by the proportion of the number) in the original sample set when generating the sampling features. The sampling features generated in this way can better reflect the feature distribution of the original data, avoid over-sampling or under-sampling of certain features in the sampling set due to improper sampling, and thus improve the accuracy and stability of model training. For example, if a certain feature is very important (has a large number) in the original sample set, but the sampling ratio is calculated incorrectly, resulting in over-sampling or under-sampling of the feature in the sampling set, the model will make a wrong judgment about the importance of the feature during the learning process, affecting the prediction performance of the model.

[0056] The target training sample set is sampled based on a sampling ratio to generate a preset number of sampling features. Based on processing any data feature and each sampling feature, multiple groups of data groups are generated, wherein each group of data groups includes a preset number of data samples, and at least one data sample includes identification information. The target training sample set is sampled according to the calculated sampling ratio. For each data feature, a corresponding number of sample values are randomly extracted from the sample set according to its sampling ratio to generate a preset number (assuming 30) of sampling features. For example, for the age feature, 40 sample values are randomly extracted from 200 age sample values as the sampling results of the age feature according to the sampling ratio of 0.4; for the systolic pressure feature, 180 sample values (rounded down) are randomly extracted from 180 systolic pressure sample values as the sampling results of the systolic pressure feature according to the sampling ratio of 0.36. Such sampling operations are performed on all 20 data features, and finally 30 sampling features are obtained, each of which contains part of the sample values of different data features. These sampling features will serve as the basis for subsequent data group construction. Through this sampling method based on the sampling ratio, the sampling features generated not only retain the feature distribution information of the original sample set, but also reduce the data volume, making subsequent data processing and model training more efficient. At the same time, since the sampling is random, each sampling feature has a certain randomness, which to some extent avoids the over-reliance of the model on specific sample patterns and improves the generalization ability of the model.

[0057] The 30 generated sampling features are combined to construct multiple groups of data groups. The specific operation is to select one data feature from the 30 sampling features as the basis each time, and then combine it with the other 29 sampling features to form a group of data groups. In this way, 30 groups of data groups (each group of data groups contains 30 data features) are obtained. In the process of constructing data groups, it is ensured that at least one data sample contains identification information (such as the symptom category to which the sample belongs, such as myocardial amyloidosis, hypertrophic cardiomyopathy, or hypertensive left ventricular hypertrophy, etc.), so that the sample category can be clearly defined during model training and verification, and accurate classification and prediction can be performed. For example, the first group of data groups is (age sampling feature 1, systolic pressure sampling feature 2, exercise frequency sampling feature 3, …, identification information sample 1), the second group of data groups is (age sampling feature 2, systolic pressure sampling feature 3, exercise frequency sampling feature 4, …, identification information sample 2), and so on.

[0058] The purpose of constructing the data set is to organize the sampling features into a format suitable for model training, and at the same time, through the samples containing identification information, the model can learn the relationship between different feature combinations and symptom categories. Each data set is regarded as a training sample, which contains information of multiple data features that work together to predict the symptom category to which the sample belongs. By constructing multiple data sets, the data diversity of model training is increased, and the learning ability and generalization ability of the model are improved. For example, if only a single data set is used for model training, the model may only learn the relationship between a specific feature combination and the symptom category, and cannot cope with other different feature combination cases. By constructing multiple data sets, the model is exposed to more different feature combination patterns, thereby better learning the rules in the data and improving the prediction accuracy of the model in various situations.

[0059] The preset myocardial amyloidosis prediction model is trained based on the data samples in the multiple data sets to generate a trained myocardial amyloidosis prediction model. The data samples in the 30 constructed data sets are used to train the preset myocardial amyloidosis prediction model. The preset model here is a random forest model. In the training process, the data features in each data set are taken as input, and the corresponding identification information (symptom category) is taken as output, so that the model learns the mapping relationship between the data features and the symptom category. For example, for the random forest model, the model will divide different sample spaces by constructing decision trees according to the values of the age, systolic pressure, exercise habit and other features in the data set, so that the leaf nodes of each decision tree belong to the same symptom category as much as possible, thereby realizing the prediction of the symptom. In the training process, the model will continuously adjust the structure and parameters (such as the selection of split nodes, the depth of the tree, etc.) of the decision tree to minimize the prediction error. When training the random forest model, the number of decision trees, the depth of the tree, the selection criteria of the split nodes and other hyperparameters are adjusted. The optimal combination of hyperparameters is selected through cross-validation and other methods to improve the performance of the model. At the same time, the idea of ensemble learning is adopted to combine multiple trained models (such as multiple random forest models with different parameter settings), and the final prediction result is obtained through voting or weighted averaging, further improving the accuracy and stability of the model. In addition, regularization techniques such as L1 and L2 regularization are used in the training process to prevent overfitting of the model and improve the generalization ability of the model.

[0060] The trained myocardial amyloidosis prediction model is processed based on a verification sample set to generate a verification result. If the data sample containing the identification information in the verification result represents a risk factor affecting myocardial amyloidosis, the trained myocardial amyloidosis prediction model is used as a target myocardial amyloidosis prediction model. The trained myocardial amyloidosis prediction model is verified using an independent verification sample set. The verification sample set is constructed in a similar manner to the training sample set, but it is ensured that the verification sample set is independent of the training sample set, i.e., the samples in the verification sample set cannot be included in the training sample set, to avoid data leakage and overfitting problems. The samples in the verification sample set are input into the trained model to obtain the prediction results of the model. Then, the prediction results are compared with the true identification information (symptom category) of the samples in the verification sample set, and the verification indicators of the model are calculated, such as accuracy, recall rate, F1 score, etc. For example, the accuracy is obtained by calculating the proportion of the number of correctly predicted samples to the total number of verification samples; the recall rate is obtained by calculating the proportion of the true positive samples that are predicted as positive samples by the model; and the F1 score is an indicator that considers the accuracy and recall rate, balancing the performance of the model on different categories.

[0061] If the data sample containing the identification information in the verification result is accurately predicted as a risk factor representing myocardial amyloidosis (i.e., the model can accurately identify samples with myocardial amyloidosis or related risks), and the model performs well in overall performance (such as accuracy, recall rate, F1 score, etc. reaching a certain threshold, for example, accuracy greater than 80%, F1 score greater than 0.7, etc.), the trained myocardial amyloidosis prediction model is used as the target myocardial amyloidosis prediction model. This target model will be used to process the clinical parameter information and dynamic risk adjustment factor of a new target user to generate early warning information, providing support for clinical diagnosis and problem prediction.

[0062] S105, processing the clinical parameter information of the target user to generate target physiological parameter information.

[0063] In one implementation, the clinical parameter information of the target user is processed based on a random forest method to generate an importance score value of each clinical parameter information; assuming that there is a data set containing 1000 target users, the clinical parameter information of each user includes age, gender, blood pressure (systolic pressure and diastolic pressure), heart rate, electrocardiogram parameters (such as PR interval, QRS interval, QTc interval, Sokolow-Lyons index, etc.), echocardiogram parameters (such as left atrial size, left ventricular size, interventricular septal thickness, left ventricular posterior wall thickness, left ventricular short axis shortening rate, left ventricular ejection fraction, mitral regurgitation degree, tricuspid regurgitation velocity, tricuspid annulus systolic displacement value, E / A, E / e', etc.), and test information (such as troponin, brain natriuretic peptide, liver and kidney function indicators, etc.). The random forest model is composed of multiple decision trees, and it is assumed that the random forest model is initialized to contain 100 decision trees. When constructing each decision tree, a part of the features (for example, n features are selected, n is the total number of features, and it is assumed that the total number of features is 30, so that n features are randomly selected) is randomly selected from all the clinical parameter information as the candidate split feature of the decision tree.

[0064] For each decision tree, a sample is randomly drawn with replacement from the training data (i.e., the clinical parameter information of 1000 target users) (for example, 632 samples are drawn, which is calculated according to the empirical formula n , where n sub is the number of sub-samples, and n is the total number of samples, so that the training data of each tree has a certain randomness and diversity). Then, according to the impurity index such as Gini impurity or information entropy, the feature with the maximum impurity reduction is selected as the split node, and the decision tree is gradually constructed until the stopping condition is met (such as the depth of the tree reaches 10 layers, or the number of samples in the node is less than 5, or the impurity reduction is less than a certain threshold, such as 0.01, etc.). In this process, each decision tree learns different decision rules according to the data features and sample distribution.

[0065] After the random forest model is constructed, the importance score value of each clinical parameter information is calculated. The method also includes a calculation formula for obtaining the importance score value, and the calculation formula is: where I j represents the importance score value of the jth clinical parameter information, N represents the total number of decision trees in the random forest, T represents the total number of nodes of each decision tree, n t represents the number of samples contained in node t, n represents the total number of samples in the training set, v it represents the split feature of node t in decision tree i, represents the indicator function, and v it ​​when v it = j (i.e., the split feature of node t is the jth clinical parameter information), which is the impurity reduction of node t before and after splitting (impurity is measured by Gini impurity or information entropy, here we assume Gini impurity is used, the formula is where I(t) represents the Gini impurity of node t before splitting, n left and n right are the sample sizes of the left and right child nodes of node t after splitting, I(t left ) and I(t right ) are the Gini impurities of the left and right child nodes, respectively); when v it ≠ j,

[0066] For each decision tree, when a feature is selected as the split node, the impurity reduction (here we take Gini impurity as an example) of this split point is calculated. For example, for a certain node in a decision tree, suppose the Gini impurity before splitting is I(t) = 0.5, the sample sizes of the left and right child nodes after splitting are n left = 300, n right = 332, and the Gini impurities are I(t left ) = 0.4, I(t right ) = 0.35, respectively. According to the Gini impurity reduction formula (where v it is the split feature of the node in the decision tree, here we assume the split feature is systolic blood pressure, j represents the systolic blood pressure clinical parameter information, n t = n left + n right = 632 is the total sample size of node t), the Gini impurity reduction of systolic blood pressure when this node is split is For each node split in each decision tree, if the split feature is a certain clinical parameter information (such as systolic blood pressure), the impurity reduction of each split is accumulated. Suppose in these 100 decision trees, systolic blood pressure appears as a split feature 200 times, and the impurity reduction of each split is accumulated to 15, then according to the importance score value formula (where I j is the importance score value of the jth clinical parameter information (such as systolic blood pressure), N = 100 is the total number of decision trees, T is the total number of nodes in each decision tree, and n = 10000 is the total sample size of the training set), the importance score value of systolic blood pressure is In the same way, the importance score values of other clinical parameter information (such as age, gender, other electrocardiogram and echocardiogram parameters, etc.) are calculated.

[0067] The importance score value of each clinical parameter information is processed based on the preset feature elimination method to generate target clinical parameter combination information. According to the importance score value (or normalized importance proportion) of each clinical parameter information calculated, the relative importance ranking thereof is determined. For example, the clinical parameter information is sorted according to the importance score value from large to small, the importance score value of the systolic blood pressure is the highest, and the ranking is 1; the age is the second, and the ranking is 2; the gender is ranked 15th, and so on. Assuming that there are 30 clinical parameter information, each parameter has a corresponding ranking. A minimum relative importance ranking threshold r min is preset. For example, through multiple tests and analysis, it is found that when the top 8 clinical parameter information is retained, the performance (such as accuracy, recall rate and the like) of the model is good on the training set and the validation set, and at the same time, the influence of too many irrelevant or redundant features on the model can be avoided, so r min = 8. The setting of this threshold needs to balance between the number of features and the performance of the model, and if the threshold is too large, some important features may be deleted, resulting in loss of model information and performance decline; if the threshold is too small, too many less important features may be retained, increasing the complexity of the model and easily leading to overfitting.

[0068] The target clinical parameter combination information is processed to generate target physiological parameter information, wherein the target physiological parameter information includes Sokolow-Lyons index, systolic blood pressure, interventricular septal thickness, left ventricular posterior wall thickness, tricuspid annular systolic displacement value, and ratio of E peak velocity of mitral blood flow to e' peak velocity of mitral annular tissue Doppler. The corresponding clinical parameter values are extracted from the target clinical parameter combination information as the basis of the target physiological parameter information. For example, for the target user 1, the systolic blood pressure is 130 mmHg, the age is 55 years old, the Sokolow-Lyons index is 3.0 mV, the left ventricular posterior wall thickness is 10 mm, the interventricular septal thickness is 12 mm, the tricuspid annular systolic displacement value is 18 mm, the ratio of E peak velocity of mitral blood flow to e' peak velocity of mitral annular tissue Doppler is 8, and the heart rate is 75 beats per minute, and so on. These parameter values will be used for subsequent analysis and symptom prediction.

[0069] The preset feature elimination method includes a calculation formula for obtaining the target clinical parameter combination information, and the calculation formula is: C = {c j |j∈{1,2,...,m},r j ≥r min}; wherein C represents the target clinical parameter combination information, which is a set, c j represents the jth clinical parameter information, m represents the total number of initial clinical parameter information, r j represents the importance score value I j of the jth clinical parameter information, and r minThe extracted target physiological parameter information is arranged in a certain order to form a vector or matrix for input into the subsequent prediction model for processing. For example, a vector X = [130, 55, 3.0, 10, 12, 18, 8, 75] is constructed (the order here is consistent with the order of the parameters in the target clinical parameter combination information determined above), where each element corresponds to the systolic blood pressure, age, Sokolow-Lyons index, left ventricular posterior wall thickness, interventricular septal thickness, tricuspid annular systolic displacement value, ratio of mitral valve flow E peak velocity to mitral annulus tissue Doppler e' peak velocity, and heart rate, respectively. If there are multiple target users, the target physiological parameter information vectors of each user are combined into a matrix, with each row representing the target physiological parameter information of a user, which facilitates batch processing of data for multiple users. The process of constructing the target physiological parameter information vector or matrix needs to ensure the accuracy and consistency of the data, and match the input requirements of the subsequent model, so that the model can correctly learn and predict the problem state of the target user.

[0070] S106, processing the life habit information of the target user to generate a dynamic risk adjustment factor.

[0071] In an embodiment, the data type is ensured to meet the requirements, and the data that does not meet the requirements (such as the exercise frequency being a decimal number) is verified, corrected or marked as an abnormal value. The numerical data in the life habit information (such as exercise duration, sleep duration, etc.) is subjected to abnormal value detection, the mean value and standard deviation are calculated, and the data with a distance from the mean value exceeding 3 times the standard deviation is regarded as an abnormal value, which is verified and corrected or deleted. The missing value situation of each life habit variable is analyzed to determine a reasonable threshold (such as a missing proportion of dietary intake data exceeding 15% being regarded as more, a missing proportion of exercise type subdivision reaching 30% being regarded as more, and a missing proportion of psychological state data exceeding 10% being treated with caution). For approximately normally distributed data (such as exercise duration), a mean value or regression model based interpolation method is used; for non-normally distributed data (such as the number of sleep interruptions), a median interpolation or quantile regression interpolation method is used, and the data accuracy and reliability are verified after the interpolation operation.

[0072] The risk of nutritional imbalance is assessed and scored based on dietary intake characteristics (e.g., excessive carbohydrate intake and insufficient dietary fiber, inadequate protein intake, etc.); the risk of insufficient exercise is assessed and scored based on exercise habits (considering exercise frequency, duration, and intensity); the risk is assessed and scored based on sleep quality characteristics (sleep duration and stages); the risk of adverse reactions is assessed and scored based on medication use history characteristics (drug classification, frequency, dosage, and interactions); and the risk of psychological problems is assessed and scored based on psychological state characteristics (assessment scale scores), with scores adjusted for changes in psychological state. The mean and standard deviation of each characteristic of the lifestyle information for different target users are calculated, and weights are determined based on the relationship between user data and the mean and standard deviation (e.g., higher than the mean, plus one standard deviation, high weight; close to the mean, medium weight; lower than the mean, minus one standard deviation, low weight). The scores for diet, exercise, sleep, medication, and psychological risk are multiplied by their corresponding dynamic characteristic weights and summed to obtain a dynamic risk adjustment factor. This factor comprehensively reflects the impact of lifestyle habits on the risk of problems such as myocardial amyloidosis.

[0073] S107, Based on the target myocardial amyloidosis prediction model, the target physiological parameter information and the dynamic risk adjustment factor are processed to generate early warning information.

[0074] In one implementation, the target physiological parameter information and dynamic risk adjustment factor are processed to generate a multi-dimensional feature vector, which is used to characterize the dimensional information of the target physiological parameter information and the dynamic risk adjustment factor. Assume the target physiological parameter information includes the Sokolov-Lyons index (3.5mV), systolic blood pressure (130mmHg), interventricular septal thickness (10mm), left ventricular posterior wall thickness (10mm), tricuspid annular systolic displacement (18mm), and the ratio of mitral valve blood flow E-peak velocity to mitral annular tissue Doppler e'-peak velocity (8); the dynamic risk adjustment factor is 19.65 (calculated previously). These parameters and factors are combined into a vector, for example, 3.5, 130, 10, 10, 18, 8, 19.65. This vector is the multi-dimensional feature vector, which integrates the target physiological parameter information reflecting the cardiovascular physiological state of the target user and the dimensional information of the dynamic risk adjustment factor reflecting lifestyle risks, providing a comprehensive data foundation for subsequent model processing.

[0075] Data normalization and standardization are performed on the multi-dimensional feature vector to generate the target feature vector. For each dimension (feature) in the multi-dimensional feature vector, its minimum and maximum values ​​are found across all samples. For example, for the systolic blood pressure dimension, the minimum value is 90 mmHg and the maximum value is 200 mmHg; for the dynamic risk adjustment factor dimension, the minimum value is 10 and the maximum value is 30. The minimum-maximum normalization formula is then used. The values of each dimension are normalized. Taking systolic blood pressure as an example, for a sample with a systolic blood pressure value of 130 mmHg, the normalized systolic blood pressure value is For the dynamic risk adjustment factor 19.65, the normalized value is All dimensions in the multi-dimensional feature vector are normalized in this way to obtain a normalized feature vector, for example [0.18, 0.36, 0.2, 0.2, 0.3, 0.4, 0.4].

[0076] The mean and standard deviation of each dimension in the multi-dimensional feature vector are calculated. For example, for the above feature vector containing seven dimensions, the mean of the systolic blood pressure dimension is calculated to be 120 mmHg and the standard deviation is 15 mmHg; the mean of the tricuspid annular systolic displacement value dimension is 15 mm and the standard deviation is 3 mm, etc. The standardization formula is The values of each dimension are standardized. Taking systolic blood pressure as an example, for a sample with a systolic blood pressure value of 130 mmHg, the standardized systolic blood pressure value is For a tricuspid annular systolic displacement value of 18 mm, the standardized value is All dimensions in the multi-dimensional feature vector are standardized in this way to obtain a standardized target feature vector, for example 0.33, 0.67, 0.5, 0.5, 1, 1.33, 1.2. Standardization makes the mean of each dimension 0 and the standard deviation 1, further converting the data to a standard scale, eliminating the influence of different dimensions due to the difference in dimension and data distribution, thereby improving the training effect and generalization ability of the model.

[0077] The target feature vector is processed based on the target myocardial amyloidosis prediction model to generate a target prediction value, and the target prediction value is processed based on a preset threshold to generate a warning information, which is used to represent the relevance of the target user to myocardial amyloidosis, hypertrophic cardiomyopathy and hypertensive left ventricular hypertrophy. The target feature vector 0.33, 0.67, 0.5, 0.5, 1, 1.33, 1.2 after normalization and standardization is input into the target myocardial amyloidosis prediction model, which has the ability to predict myocardial amyloidosis, hypertrophic cardiomyopathy and hypertensive left ventricular hypertrophy. The model is constructed based on the random forest algorithm, which calculates the probability or score of the target user suffering from different symptoms according to the dimensional information in the feature vector (i.e. the normalized and standardized target physiological parameter information and dynamic risk adjustment factor) through voting or weighted averaging of multiple decision trees. Assuming that the model output result is that the target user has a 0.3 probability of suffering from myocardial amyloidosis, a 0.2 probability of suffering from hypertrophic cardiomyopathy, a 0.4 probability of suffering from hypertensive left ventricular hypertrophy, and a comprehensive symptom risk score (for example, a score obtained by weighting and summing the probabilities of the three symptoms according to certain rules, assuming 0.3x3+0.2x2+0.4x1=1.7), the comprehensive score is the target prediction value. The output result of the model reflects the risk degree of the target user in cardiovascular problems, which provides a basis for subsequent generation of warning information.

[0078] The setting of the preset threshold is based on a large amount of clinical data and the evaluation of the performance of the model. For myocardial amyloidosis, it is assumed that after research, when the prediction probability is greater than 0.25, the patient has a high probability of suffering from the disease, and therefore the warning threshold of myocardial amyloidosis is set to 0.25; for hypertrophic cardiomyopathy, the warning threshold is set to 0.2; for hypertensive left ventricular hypertrophy, the warning threshold is set to 0.3. At the same time, for the comprehensive risk score, a corresponding threshold is also set, such as when the comprehensive risk score is greater than 1.5, a high-risk warning is issued, prompting close attention and further evaluation.

[0079] The server collects various types of information of the target user, including clinical parameters and details of living habits, and constructs a training sample set, a verification sample set and a preset prediction model containing information related to the user with specific cardiovascular problems, thereby providing sufficient data basis for subsequent analysis. Subsequently, detailed data processing is carried out on the training sample set. In the data cleaning stage, invalid value detection is carried out, and the data format and value range are carefully checked. The data that does not meet the requirements is modified or deleted; the abnormal value detection is accurately identified by combining visualization and statistical calculation. After verification, the abnormal values are reasonably modified or deleted to ensure the data quality. After cleaning, the missing value supplement processing is carried out. First, the missing value threshold of the clinical and living habit variables is determined, and the missing sample information is generated according to the threshold. The missing value is supplemented and predicted, and the target variable value is backfilled. The target variable parameter information is generated by statistical test, and the target training sample set is finally generated based on the screened variables.

[0080] For the target user, on the one hand, the clinical parameter information is processed, the random forest method is used to calculate the importance score of each parameter, the key clinical parameter combination is determined according to the preset feature elimination method, and the target physiological parameter information is generated; on the other hand, the living habit information is processed. First, the data is preprocessed to ensure the integrity and accuracy, and then various features are extracted, the risk score and dynamic feature weight value of each aspect are calculated, and finally the dynamic risk adjustment factor is generated. Finally, the target physiological parameter information and the dynamic risk adjustment factor are integrated into a multi-dimensional feature vector, which is input into the target model after normalization and standardization processing. According to the preset threshold, the target prediction value output by the model is judged, and the warning information is generated, so as to clearly determine the correlation degree of the target user and the myocardial amyloidosis and related problems.

[0081] In one embodiment, as shown in Figure 2 The application also provides a prediction device for diagnosing myocardial amyloidosis, comprising: an acquisition module 201 for acquiring clinical parameter information of a target user, living habit information of the target user, a first training sample set, a verification sample set and a preset myocardial amyloidosis prediction model, wherein the first training sample set comprises clinical parameter information and living habit information of a coronary artery user, clinical parameter information and living habit information of a hypertrophic cardiomyopathy user, and clinical parameter information and living habit information of a hypertensive left ventricular hypertrophy user, and the living habit information of the target user comprises dietary intake data of the target user, exercise habit data of the target user, sleep quality data of the target user, drug use history data of the target user and psychological state data of the target user;

[0082] The processing module 202 is configured to perform data cleaning processing on the first training sample set to generate a second training sample set, perform missing value supplement processing on the second training sample set to generate a target training sample set, process the preset myocardial amyloidosis prediction model based on the target training sample set and the verification sample set to generate a target myocardial amyloidosis prediction model, process the clinical parameter information of the target user to generate target physiological parameter information, wherein the target physiological parameter information includes Sokolow-Lyon index, systolic blood pressure, interventricular septal thickness, left ventricular posterior wall thickness, tricuspid annular plane systolic excursion value, and ratio of mitral valve flow E peak velocity to mitral annulus tissue Doppler e' peak velocity, process the life habit information of the target user to generate a dynamic risk adjustment factor, and process the target physiological parameter information and the dynamic risk adjustment factor based on the target myocardial amyloidosis prediction model to generate early warning information, wherein the early warning information is used to represent the relevance of the target user to myocardial amyloidosis, hypertrophic cardiomyopathy, and hypertensive left ventricular hypertrophy.

[0083] Each of the embodiments in the present application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other. Each of the embodiments focuses on the difference from other embodiments. In particular, for the prediction method, the electronic device, the electronic equipment, and the readable storage medium embodiments for evaluating and diagnosing myocardial amyloidosis, since they are basically similar to the above-mentioned prediction method embodiments for diagnosing myocardial amyloidosis, the description is relatively simple, and the relevant parts can be referred to the above-mentioned description of the prediction method embodiments for diagnosing myocardial amyloidosis.

Claims

1. A method of predicting diagnosis of cardiac amyloidosis, characterized by, The method comprises the following steps: obtaining clinical parameter information of a target user, life habit information of the target user, a first training sample set, a verification sample set, and a preset myocardial amyloidosis prediction model, wherein the first training sample set comprises clinical parameter information and life habit information of a coronary artery user, clinical parameter information and life habit information of a hypertrophic cardiomyopathy user, and clinical parameter information and life habit information of a hypertensive left ventricular hypertrophy user, and the life habit information of the target user comprises dietary intake data of the target user, exercise habit data of the target user, sleep quality data of the target user, drug use history data of the target user, and psychological state data of the target user; performing data cleaning processing on the first training sample set to generate a second training sample set; performing missing value supplement processing on the second training sample set to generate a target training sample set; processing the target training sample set and the verification sample set to generate a target myocardial amyloidosis prediction model; processing the clinical parameter information of the target user to generate target physiological parameter information, wherein the target physiological parameter information comprises Sokolov-Leonidze index, systolic pressure, interventricular septum thickness, left ventricular posterior wall thickness, tricuspid annulus systolic displacement value, and ratio of mitral valve blood flow E peak velocity to mitral annulus tissue Doppler e' peak velocity; processing the life habit information of the target user to generate a dynamic risk adjustment factor; processing the target physiological parameter information and the dynamic risk adjustment factor based on the target myocardial amyloidosis prediction model to generate early warning information, wherein the early warning information is used to represent the relevance of the target user to myocardial amyloidosis, hypertrophic cardiomyopathy, and hypertensive left ventricular hypertrophy.

2. The method of claim 1, wherein, The method comprises the following steps: performing invalid value detection processing on the first training sample set to generate an initial training sample set; performing abnormal value detection processing on the initial training sample set to generate a second training sample set.

3. The method of claim 2, wherein, The method comprises the following steps: obtaining a missing value threshold of each clinical variable and a missing value threshold of each life habit variable; processing the second training sample set based on the missing value threshold of each clinical variable and the missing value threshold of each life habit variable to generate missing sample information, wherein the missing sample information is used to represent samples with missing values, and the position and variable name of the missing values; processing the missing sample information to generate missing value imputation prediction information; processing the missing value imputation prediction information to generate a target variable value; processing the target variable value based on a preset statistical test method to generate target variable parameter information; generating a target training sample set based on the target variable parameter information; The method comprises a calculation formula for obtaining missing value imputation prediction information, and the calculation formula is: wherein, represents the prediction of the missing value at the mth imputation, Y i represents the predicted value based on the regression model, ε (m) represents the random error term; The method comprises a calculation formula for obtaining a target variable value, and the calculation formula is: where x i represents the mean of the original variable values, μ i represents the mean of the variable x i , σ i represents the standard deviation of the variable x i .

4. The method of claim 3, wherein, The preset myocardial amyloidosis prediction model is processed based on the target training sample set and the verification sample set to generate a target myocardial amyloidosis prediction model, including: Obtaining any number of data features in the target training sample set; Based on the number of data features in the target training sample set, a sampling ratio is generated; Based on the sampling ratio, the target training sample set is sampled to generate a preset number of sampling features; Based on any data feature and each sampling feature, a plurality of data groups are generated, wherein each data group contains a preset number of data samples, and at least one data sample includes identification information; The preset myocardial amyloidosis prediction model is trained based on the data samples in the plurality of data groups to generate a trained myocardial amyloidosis prediction model; The trained myocardial amyloidosis prediction model is processed based on the verification sample set to generate a verification result; If the data sample containing the identification information in the verification result is a risk factor affecting myocardial amyloidosis, the trained myocardial amyloidosis prediction model is used as the target myocardial amyloidosis prediction model.

5. The method of claim 1, wherein, The target user's clinical parameter information is processed to generate target physiological parameter information, including: The target user's clinical parameter information is processed based on a random forest method to generate an importance score value for each clinical parameter information; The importance score value of each clinical parameter information is processed based on a preset feature elimination method to generate target clinical parameter combination information; The target clinical parameter combination information is processed to generate target physiological parameter information, wherein the target physiological parameter information includes Sokolow-Lyon index, systolic blood pressure, interventricular septal thickness, left ventricular posterior wall thickness, tricuspid annular systolic displacement value, and ratio of mitral valve flow E peak velocity to mitral annular tissue Doppler e' peak velocity; The method further includes a calculation formula for obtaining the importance score value, and the calculation formula is: wherein I j represents the importance score value of the jth clinical parameter information, N represents the total number of decision trees in the random forest, T represents the total number of nodes of each decision tree, n t represents the number of samples contained in node t, n represents the total number of samples in the training set, v it represents the split feature of node t in decision tree i, represents the indicator function; The preset feature elimination method includes a calculation formula for obtaining target clinical parameter combination information, the calculation formula being: C={c j |j∈{1,2,...,m},r j ≥r min}; wherein C represents target clinical parameter combination information, which is a set, c j represents the jth clinical parameter information, m represents the total number of initial clinical parameter information, r j represents the importance score value I j of the jth clinical parameter information, and r min represents a preset minimum relative importance ranking threshold.

6. The method of claim 5, wherein, The target user's lifestyle information is processed to generate a dynamic risk adjustment factor, including: The target user's lifestyle information is preprocessed to generate lifestyle information after data cleaning and missing value supplementation; The data cleaning and missing value supplemented lifestyle information is processed for feature extraction to generate dietary intake features, exercise habit features, sleep quality features, drug use history features, and psychological state features; The dietary intake features, exercise habit features, sleep quality features, drug use history features, and psychological state features are processed to generate dietary risk scores, exercise risk scores, sleep quality scores, drug risk comprehensive scores, psychological risk comprehensive scores, and dynamic feature weight values, wherein the dynamic feature weight values are generated based on the lifestyle information of different target users; The dietary risk scores, exercise risk scores, sleep quality scores, drug risk comprehensive scores, psychological risk comprehensive scores, and dynamic feature weight values are processed to generate a dynamic risk adjustment factor.

7. The method of claim 6, wherein, processing the target physiological parameter information and the dynamic risk adjustment factor based on the target cardiac amyloidosis prediction model to generate early warning information, including: processing the target physiological parameter information and the dynamic risk adjustment factor to generate a multi-dimensional feature vector, wherein the multi-dimensional feature vector is used to represent the dimension information of the target physiological parameter information and the dynamic risk adjustment factor; performing data normalization and standardization processing on the multi-dimensional feature vector to generate a target feature vector; processing the target feature vector based on the target cardiac amyloidosis prediction model to generate a target prediction value; processing the target prediction value based on a preset threshold to generate early warning information.

8. A prognostic device for diagnosing cardiac amyloidosis, characterized by comprising: The device includes: an acquisition module configured to acquire clinical parameter information of a target user, life habit information of the target user, a first training sample set, a validation sample set, and a preset cardiac amyloidosis prediction model, wherein the first training sample set includes clinical parameter information and life habit information of a coronary artery user, clinical parameter information and life habit information of a hypertrophic cardiomyopathy user, and clinical parameter information and life habit information of a hypertensive left ventricular hypertrophy user, and the life habit information of the target user includes dietary intake data of the target user, exercise habit data of the target user, sleep quality data of the target user, drug use history data of the target user, and psychological state data of the target user; a processing module configured to perform data cleaning processing on the first training sample set to generate a second training sample set, perform missing value supplement processing on the second training sample set to generate a target training sample set, process the target training sample set and the validation sample set based on the preset cardiac amyloidosis prediction model to generate a target cardiac amyloidosis prediction model, process the clinical parameter information of the target user to generate target physiological parameter information, wherein the target physiological parameter information includes Sokolow-Lyon index, systolic blood pressure, interventricular septal thickness, left ventricular posterior wall thickness, tricuspid annular systolic displacement value, and ratio of mitral flow E peak velocity to mitral annular tissue Doppler e' peak velocity, process the life habit information of the target user to generate a dynamic risk adjustment factor, and process the target physiological parameter information and the dynamic risk adjustment factor based on the target cardiac amyloidosis prediction model to generate early warning information, wherein the early warning information is used to represent the relevance of the target user to cardiac amyloidosis, hypertrophic cardiomyopathy, and hypertensive left ventricular hypertrophy.

9. An electronic device, comprising: includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to perform the prediction method for diagnosing cardiac amyloidosis according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a second processor to implement the prediction method for diagnosing cardiac amyloidosis according to any one of claims 1-7.

Images