COPD-FE risk prediction method based on disease and symptom combination

Abstract

The invention discloses a COPD-FE risk prediction method based on disease and symptom combination, and is applied to the technical field of chronic obstructive pulmonary disease risk prediction. Comprising the following steps: acquiring chronic obstructive pulmonary frequent acute exacerbation influence factor data of a patient; the influence factors are screened through LASSO regression and an improved Boruta algorithm respectively; the LASSO independent influence factors and the Boruta independent influence factors are combined in different modes, a Logistic regression model and an XGBoost model are trained, and a plurality of COPD-FE risk prediction models are obtained; and evaluating the performance of all the COPD-FE risk prediction models, and selecting the COPD-FE risk prediction model meeting the requirement to predict the COPD-FE risk. According to the method, clinical data of patients are collected, a risk prediction model is constructed in combination with a feature selection method and machine learning, and an optimal model is screened out through comprehensive evaluation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chronic obstructive pulmonary disease risk prediction, and more particularly to a COPD-FE risk prediction method based on disease syndrome combination. BACKGROUND

[0002] Repeated acute exacerbations of chronic obstructive pulmonary disease can accelerate the decline in lung function and increase the risk of future mortality, regardless of the quality of life. In addition, repeated acute exacerbations of chronic obstructive pulmonary disease can also cause a series of complications, such as pulmonary heart disease caused by heart involvement, and increase the probability of other system-related complications. Therefore, early identification of chronic obstructive pulmonary disease is very important for treatment. The existing identification method is mainly realized by doctors according to symptoms and lung function evaluation reports. Although this method has high accuracy, it is too dependent on the personal experience of doctors. In addition to the judgment of doctors, the existing technology also uses machine learning algorithms to predict the risk of chronic obstructive pulmonary disease. However, since the risk of chronic obstructive pulmonary disease is related to many influencing factors and has many symptoms, if the conventional machine learning algorithm is directly used for chronic obstructive pulmonary disease risk prediction, the accuracy of the prediction result will be low. Therefore, how to provide a COPD-FE risk prediction method based on disease syndrome combination is a problem that those skilled in the art need to solve. SUMMARY

[0003] Therefore, the present application provides a COPD-FE risk prediction method based on disease syndrome combination, which collects the clinical data of patients, combines feature selection methods and machine learning to construct a risk prediction model, and screens the optimal model through comprehensive evaluation.

[0004] In order to achieve the above purpose, the present application provides the following technical scheme:

[0005] A COPD-FE risk prediction method based on disease syndrome combination comprises the following steps:

[0006] S1, obtaining the frequent acute exacerbation influencing factor data of chronic obstructive pulmonary disease of patients, pre-processing the influencing factor data and assigning variables, and constructing a training data set;

[0007] S2, respectively screening the influencing factors through LASSO regression and improved Boruta algorithm to obtain multiple groups of LASSO independent influencing factors and multiple groups of Boruta independent influencing factors;

[0008] S3, combining the LASSO independent influencing factors and the Boruta independent influencing factors in different ways to obtain multiple groups of sub-training sets, training the Logistic regression model and the XGBoost model based on each group of sub-training sets respectively, and obtaining multiple COPD-FE risk prediction models.

[0009] S4, verifying the performance indicators of all COPD-FE risk prediction models through the validation set;

[0010] S5, constructing an evaluation system of the COPD-FE risk prediction model, evaluating and screening the performance of all COPD-FE risk prediction models;

[0011] S6, selecting a COPD-FE risk prediction model meeting the requirements based on the screening results of the evaluation to predict the COPD-FE risk.

[0012] Optionally, S1 specifically comprises:

[0013] S11, determining the diagnostic criteria of patients, and including patients meeting the diagnostic criteria into the target group of the data set collection;

[0014] S12, determining the data source, the prediction time node and the true result label, and obtaining the time series data of the influencing factors of the COPD frequent acute exacerbation of the patients;

[0015] S13, performing the outlier rejection and the missing value supplement processing on the time series data of the influencing factors;

[0016] S14, assigning the variables to the processed influencing factors, constructing the input features containing the influencing factor data and the data acquisition time node, taking the input features and the corresponding prediction time node and true result label as the training sample, and obtaining the training data set.

[0017] Optionally, S13 specifically comprises:

[0018] For the time series data of an influencing factor, the Q L value, the Q U value and the Q R value are calculated, and according to the calculation results, it is judged whether all data are abnormal data, if the data q meets one of q < Q L -1.5Q R or q > Q U -1.5Q R , it is determined as abnormal data, and the data q is rejected; wherein Q L is the lower quartile, Q U is the upper quartile, and Q R is the difference between Q L and Q U .

[0019] For the time series data after the abnormal value is rejected, the prior distribution function p(x q of the missing value x q); calculate the joint distribution function p(X | x n ) based on the time series data X = {x1, x2,..., x q ) after removing outliers; calculate the posterior distribution p(x q | X) of the missing value x q ;

[0020]

[0021] Calculate the estimated value of the missing value x q :

[0022]

[0023] Fill the estimated value into the time series data as the missing value.

[0024] Optionally, the screening of the influencing factors in S2 through LASSO regression is specifically as follows: standardize the training data set, divide the training data set into multiple subsets, for each training data subset, use it as a validation set and use the remaining training data subsets as training sets, obtain multiple groups of LASSO training sets, for each candidate regularization parameter λ, train the LASSO model on all LASSO training sets and evaluate the model performance by the mean square error, and calculate the average mean square error of the LASSO model under all training sets as the performance index of the corresponding regularization parameter λ; the loss function of the LASSO model is:

[0025]

[0026] In the formula, N is the number of samples, y n is the true value corresponding to the sample n, x n is the nth sample, p is the model coefficient, ω n is the weight of the nth sample, s n is the quality score of the nth sample, the quality score is calculated according to the proportion of outliers and the proportion of missing values, X1 is the sample set with a quality score greater than a preset threshold, X2 is the sample set after removing outliers and supplementing missing values, X3 is the sample set with a quality score lower than a preset threshold, σ n1 , σ n2 , and σ n3 are the standard deviations of the quality scores of all samples in the sets X1, X2, and X3, respectively, α1, α2, and α3 are quality weight coefficients, and α1>α2>α3;

[0027] Select the regularization parameter λ with the optimal performance index to fit the LASSO model, train the entire training data set and extract the feature coefficients of each influencing factor, and retain all the influencing factors with non-zero feature coefficients as independent influencing factors; multiple groups of LASSO independent influencing factors are obtained through multiple screenings.

[0028] Optionally, the screening of the influencing factors in S2 by improving the Boruta algorithm is specifically: for m*n original feature samples in the training set, copying the original feature samples of each influencing factor in the training set to obtain m*n copied samples, extracting samples from the copied samples according to a certain proportion p to obtain (m*p)*n extracted samples, performing random transformation on each column (m*p) extracted sample to obtain shadow feature samples, mixing the original feature samples and the shadow feature samples to obtain a new training set, training the XGBoost model through the new training set and calculating the importance scores of all features in the new training set, comparing the importance scores of each original influencing factor and the corresponding shadow feature, if the importance score of the original influencing factor is greater than the importance score of the shadow feature, the original influencing factor is retained, otherwise it is determined as pending; iteratively training, calculating the importance scores and comparing, gradually eliminating the original influencing factors, and retaining the influencing factors after stopping the iteration as independent influencing factors; and obtaining multiple groups of Boruta independent influencing factors through multiple screenings.

[0029] Optionally, S3 is specifically: taking the multiple groups of LASSO independent influencing factors and the multiple groups of Boruta independent influencing factors as sub-training sets respectively; taking the intersection between the multiple groups of LASSO independent influencing factors and the multiple groups of Boruta independent influencing factors as sub-training sets; taking the union between the multiple groups of LASSO independent influencing factors and the multiple groups of Boruta independent influencing factors as sub-training sets; training the Logistic regression model and the XGBoost model based on each group of sub-training sets respectively to obtain multiple LASSO-Logistic models, multiple Boruta-Logistic models, multiple LASSO-XGBoost models, multiple Boruta-XGBoost models, multiple LASSO∪Boruta-XGBoost models, and multiple LASSO∩Boruta-XGBoost models.

[0030] Optionally, S4 is specifically: inputting the validation set into all COPD-FE risk prediction models to obtain COPD-FE risk prediction results, and calculating the AUC index, accuracy, sensitivity, specificity, F1 score, Youden index, PPV index and NPV index of each COPD-FE risk prediction model according to the COPD-FE risk prediction results.

[0031] Optionally, S6 is specifically: obtaining the demand for COPD-FE risk prediction, when early screening is performed, selecting the COPD-FE risk prediction model with the optimal sensitivity and specificity from the evaluation and screening results; when clinical risk stratification decision is made, selecting the COPD-FE risk prediction model with the optimal interpretability from the evaluation and screening results.

[0032] Optionally, the evaluation system for the COPD-FE risk prediction model in S5 is as follows:

[0033] Model performance evaluation: The overall performance of the COPD-FE risk prediction model is evaluated based on five dimensions: core parameters, stability, interpretability, prediction efficiency, and prediction cost; the weight of each dimension is determined by the influence coefficient.

[0034] Core parameter assessment: The assessment score is calculated based on the risk prediction weight of each core parameter for COPD-FE. The weight of the core parameters is determined by the analytic hierarchy process.

[0035] Threshold screening mechanism: COPD-FE risk prediction models whose performance evaluation results are higher than the evaluation threshold are retained. Among the retained COPD-FE risk prediction models, a second screening is conducted based on the AUC index and accuracy. Among the remaining COPD-FE risk prediction models after the second screening, a third screening is conducted based on prediction cost and interpretability when the differences in AUC index, accuracy and stability are less than the allowable range.

[0036] Optionally, the weight of each dimension can be determined using the influence coefficient, specifically as follows:

[0037] Determine the impact of each dimension on patients, the model, and hospitals. Assign secondary impact coefficients to each impact based on its degree of influence. Assign primary impact coefficients to patients, the model, and hospitals based on prediction needs. Calculate the actual impact coefficient for each dimension based on the primary and secondary impact coefficients. Calculate the weights based on the actual impact coefficients for each dimension.

[0038]

[0039] In the formula, w k Let j be the weight of the k-th dimension. k Let j be the actual impact coefficient of the k-th dimension. i is the actual impact coefficient of the i-th dimension.

[0040] As can be seen from the above technical solution, compared with the prior art, the present invention provides a COPD-FE risk prediction method based on the combination of disease and symptoms, which has the following beneficial effects: The present invention collects patients' clinical data, combines feature selection methods and machine learning to construct multiple COPD-FE risk prediction models, and selects multiple optimal models by establishing a comprehensive evaluation system. The corresponding COPD-FE risk prediction can be selected according to the actual prediction needs, thereby improving the applicability of the model for predicting frequent acute exacerbations of COPD. Attached Figure Description

[0041] In order to make the technical solutions in the embodiments of the present application or the prior art clearer, the accompanying drawings needed in the embodiments or prior art description will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0042] Figure 1 Flow chart of COPD-FE risk prediction method of the present application;

[0043] Figure 2 LASSO regression coefficient path diagram in the embodiment of the present application;

[0044] Figure 3 LASSO cross-validation curve diagram in the embodiment of the present application;

[0045] Figure 4 LASSO regression coefficient visualization diagram in the embodiment of the present application;

[0046] Figure 5 Variable screening schematic diagram of improved Boruta algorithm in the embodiment of the present application;

[0047] Figure 6 Confusion matrix schematic diagram of LASSO-Logistic regression model in the training set in the embodiment of the present application;

[0048] Figure 7 ROC curve schematic diagram of LASSO-Logistic regression model in the training set in the embodiment of the present application;

[0049] Figure 8 Confusion matrix schematic diagram of LASSO-Logistic regression model in the validation set in the embodiment of the present application;

[0050] Figure 9 ROC curve schematic diagram of LASSO-Logistic regression model in the validation set in the embodiment of the present application;

[0051] Figure 10 Confusion matrix schematic diagram of Boruta-Logistic regression model in the training set in the embodiment of the present application;

[0052] Figure 11 ROC curve schematic diagram of Boruta-Logistic regression model in the training set in the embodiment of the present application;

[0053] Figure 12 Confusion matrix schematic diagram of Boruta-Logistic regression model in the validation set in the embodiment of the present application;

[0054] Figure 13 ROC curve of the Boruta-Logistic regression model in the embodiment of the application in the validation set;

[0055] Figure 14 Confusion matrix of the LASSO-XGBoost model in the embodiment of the application in the training set;

[0056] Figure 15 ROC curve of the LASSO-XGBoost model in the embodiment of the application in the training set;

[0057] Figure 16 Confusion matrix of the LASSO-XGBoost model in the embodiment of the application in the validation set;

[0058] Figure 17 ROC curve of the LASSO-XGBoost model in the embodiment of the application in the validation set;

[0059] Figure 18 Confusion matrix of the Boruta-XGBoost model in the embodiment of the application in the training set;

[0060] Figure 19 ROC curve of the Boruta-XGBoost model in the embodiment of the application in the training set;

[0061] Figure 20 Confusion matrix of the Boruta-XGBoost model in the embodiment of the application in the validation set;

[0062] Figure 21 ROC curve of the Boruta-XGBoost model in the embodiment of the application in the validation set;

[0063] Figure 22 Confusion matrix of the LASSO∪Boruta-XGBoost model in the embodiment of the application in the training set;

[0064] Figure 23 ROC curve of the LASSO∪Boruta-XGBoost model in the embodiment of the application in the training set;

[0065] Figure 24 Confusion matrix of the LASSO∪Boruta-XGBoost model in the embodiment of the application in the validation set;

[0066] Figure 25A ROC curve diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the validation set;

[0067] Figure 26 A confusion matrix diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the training set;

[0068] Figure 27 A ROC curve diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the training set;

[0069] Figure 28 A confusion matrix diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the validation set;

[0070] Figure 29 A ROC curve diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the validation set;

[0071] Figure 30 A calibration curve diagram of the Boruta-XGBoost model in the embodiment of the application in the training set;

[0072] Figure 31 A calibration curve diagram of the Boruta-XGBoost model in the embodiment of the application in the test set

[0073] Figure 32 A calibration curve diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the training set;

[0074] Figure 33 A calibration curve diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the test set;

[0075] Figure 34 A decision curve diagram of the Boruta-XGBoost model in the embodiment of the application in the training set;

[0076] Figure 35 A decision curve diagram of the Boruta-XGBoost model in the embodiment of the application in the test set;

[0077] Figure 36 A decision curve diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the training set;

[0078] Figure 37 A decision curve diagram of the LASSO∩Boruta-XGBoost model in the embodiment of the application in the test set;

[0079] Figure 38 A bee chart for the Boruta-XGBoost model in the embodiment of the present application;

[0080] Figure 39 A waterfall chart for the Boruta-XGBoost model in the embodiment of the present application;

[0081] Figure 40 A bee chart for the LASSO∩Boruta-XGBoost model in the embodiment of the present application;

[0082] Figure 41 A waterfall chart for the LASSO∩Boruta-XGBoost model in the embodiment of the present application. DETAILED DESCRIPTION

[0083] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0084] The embodiments of the present application disclose a COPD-FE risk prediction method based on disease syndrome combination, as shown in the following steps: Figure 1

[0085] S1, acquiring the data of factors affecting frequent acute exacerbation of chronic obstructive pulmonary disease of a patient, pre-processing the data of the factors and assigning variables, and constructing a training data set;

[0086] S2, screening the factors by LASSO regression and improved Boruta algorithm respectively, and obtaining multiple groups of LASSO independent factors and multiple groups of Boruta independent factors;

[0087] S3, combining the LASSO independent factors and the Boruta independent factors in different ways to obtain multiple groups of sub-training sets, training a Logistic regression model and an XGBoost model based on each group of sub-training sets respectively, and obtaining multiple COPD-FE risk prediction models;

[0088] S4, verifying the performance indicators of all COPD-FE risk prediction models through a verification set;

[0089] S5, constructing an evaluation system of the COPD-FE risk prediction model, and evaluating and screening the performance of all COPD-FE risk prediction models;

[0090] ​S6, selecting a COPD-FE risk prediction model meeting the requirements based on the evaluation screening result to predict the COPD-FE risk.

[0091] Further, S1 is specifically:

[0092] S11, determining a diagnosis standard of the patient, and including the patient meeting the diagnosis standard into a target group of the data set collection;

[0093] S12, determining a data source, a prediction time node and a real result label, and acquiring time series data of the influence factors of the COPD-FE of the patient;

[0094] S13, performing outlier rejection and missing value supplement processing on the time series data of the influence factors;

[0095] S14, performing variable assignment on the processed influence factors, constructing input features containing the influence factor data and the data acquisition time node, taking the input features and the corresponding prediction time node and real result label as training samples, and obtaining a training data set.

[0096] In the embodiment of the present application, the diagnosis standard is specifically:

[0097] Western medicine standard: ① having symptoms of dyspnea, chronic cough and sputum; ② having a history of smoking, occupational or environmental exposure to harmful substances, and being prone to attack in cold seasons, often accompanied by a history of repeated respiratory tract infections and acute exacerbations, and showing a tendency of family aggregation; ③ having persistent airflow limitation: lung function is the gold standard for diagnosing COPD, that is, FEV1 / FVC < 70% after using bronchodilators; ④ excluding other known causes or specific pathological manifestations of airflow limitation diseases, including but not limited to pulmonary tuberculosis, bronchiectasis, bronchial asthma, diffuse panbronchiolitis, lung cancer and congestive heart failure, etc. Stable period diagnosis standard: the patient's clinical symptoms such as shortness of breath, cough and sputum are relatively stable or mild, or return to the symptoms and disease manifestations before the acute exacerbation.

[0098] Traditional Chinese medicine diagnosis standard:

[0099] 1) Lung Qi deficiency syndrome

[0100] Diagnosis standard: ① cough or wheezing, shortness of breath, and aggravation with exercise; ② fatigue, weakness, or spontaneous sweating; ③ aversion to wind and easy to catch a cold; ④ pale tongue, white fur, and fine or weak pulse. Having 3 of ①, ②, ③ and ④.

[0101] 2) Lung and spleen Qi deficiency syndrome

[0102] Diagnostic criteria: ① cough or wheezing, shortness of breath, and aggravation of exercise; ② fatigue, lack of energy or spontaneous sweating, and aggravation of exercise; ③ aversion to wind and easy to catch a cold; ④ poor appetite or less food; ⑤ fullness or abdominal distension or loose stools; ⑥ tongue body is big or has teeth marks, tongue coating is thin white or white greasy, pulse is deep and fine or slow and fine or weak. 2 of ①, ②, ③, plus 2 of ④, ⑤, ⑥.

[0103] 3) Lung and kidney qi deficiency syndrome

[0104] Diagnostic criteria: ① wheezing, shortness of breath, and aggravation of exercise; ② lack of energy, or spontaneous sweating, and aggravation of exercise; ③ easy to catch a cold, aversion to wind; ④ soreness of waist and knees; ⑤ tinnitus, dizziness or facial pallor; ⑥ frequent urination, nocturia, or cough and incontinence; ⑦ pale tongue, white tongue coating, and deep and fine pulse or weak pulse. 2 of ①, ②, ③, plus 2 of ④, ⑤, ⑥, ⑦.

[0105] In the embodiment of the present application, the patients are excluded by exclusion criteria, specifically: ① participated in an intervention clinical experiment within the last three months; ② pregnant and lactating women; ③ patients with abnormal emotions and mental illness; ④ patients with cardiac function reaching grade IV; ⑤ patients with atrial fibrillation and hemodynamic instability; ⑥ patients who have been in bed for a long time; ⑦ patients with severe hypoxemia; ⑧ patients with severe liver and kidney diseases, specifically, severe liver disease refers to patients with cirrhosis, portal hypertension and variceal bleeding, and severe kidney disease includes patients who are undergoing dialysis or have received kidney transplantation; ⑨ patients with bronchiectasis, bronchial asthma, active tuberculosis, pulmonary embolism, pneumothorax, pleural effusion, malignant tumor and other neuromuscular diseases that affect respiratory function; ⑩ patients with lung and kidney qi and yin deficiency syndrome are mostly in the severe stage and cannot safely cooperate with 6MWT examination.

[0106] Further, S13 specifically is:

[0107] For a time series data of an influencing factor, Q L , Q U and Q R values are calculated, and according to the calculation results, it is respectively judged whether all data are abnormal data, if the data q meets one of q L -1.5Q R or q U -1.5Q R , it is determined as abnormal data, and the data q is excluded; wherein Q L is the lower quartile, Q U is the upper quartile, and Q R is the difference between Q L and Q U .

[0108] For the time series data after excluding abnormal values, the prior distribution function p(x q of the missing value x q is calculated.q );calculate joint distribution function p(X|x n ) according to time series data X={x1,x2,...,x q ) after removing outliers; q calculate posterior distribution p(x q |X) of missing value x

[0109]

[0110] calculate the estimate of missing value x q ;

[0111]

[0112] fill the estimate into time series data as missing value.

[0113] In the embodiments of the present application, the collected information includes: gender, age, body mass index, smoking index, medication history, TCM syndrome type, TCM syndrome score, forced expiratory volume in one second, predicted forced expiratory volume in one second, predicted forced expiratory volume in one second percentage, forced vital capacity, predicted forced vital capacity, predicted forced vital capacity percentage, forced vital capacity to forced expiratory volume in one second ratio percentage, number of colds, number of acute exacerbations, acute exacerbation degree, 6-minute walk distance, chronic obstructive pulmonary disease score, dyspnea grading score, chronic obstructive pulmonary disease patient-reported outcomes, self-rating anxiety scale, self-rating depression scale, physical function score, physical role score, bodily pain score, general health score, vitality score, social function score, emotional role score, mental health score, health change score, white blood cells, red blood cells, hemoglobin, platelets, neutrophils, neutrophil percentage, lymphocytes, lymphocyte percentage, monocytes, monocyte percentage, eosinophils, eosinophil percentage, basophils, basophil percentage, alanine aminotransferase, aspartate aminotransferase, blood creatinine, urea nitrogen, and uric acid.

[0114] In the embodiments of the present application, after the missing values are processed, the neutrophil / lymphocyte ratio, platelet / lymphocyte ratio, and lymphocyte / monocyte ratio are calculated according to the existing blood routine data, and the patients are divided into GOLD 1, 2, 3, and 4 levels according to FEV1pred%. The medication history of the patients is tracked, and the use of hormones during the treatment course is assigned a value of “0”, and the use of hormones is assigned a value of “1”. The smoking index = 0, i.e., it is determined that the patient has never smoked, and is assigned a value of “0”; ≤200 is determined to be mild smoking, and is assigned a value of “1”; >200 and <400 is determined to be moderate smoking, and is assigned a value of “2”; and ≥400 is determined to be heavy smoking, and is assigned a value of “3”. The specific variable assignment table is shown in Table 1:

[0115] Table 1 variable assignment

[0116]

[0117] Further, the screening of the influencing factors in S2 by LASSO regression is specifically as follows: the training data set is standardized, the training data set is divided into multiple subsets, for each training data subset, the training data subset is taken as a validation set, and the remaining training data subsets are taken as training sets, multiple groups of LASSO training sets are obtained, for each candidate regularization parameter λ, the LASSO model is trained on all LASSO training sets respectively, and the model performance is evaluated by the mean square error, and the average mean square error of the LASSO model under all training sets is calculated as the performance index corresponding to the regularization parameter λ; the loss function of the LASSO model is:

[0118]

[0119] In the formula, N is the number of samples, y n is the true value corresponding to the sample n, x n is the nth sample, p is the model coefficient, ω n is the weight of the nth sample, s n is the quality score of the nth sample, the quality score is calculated according to the proportion of outliers and the proportion of missing values, X1 is the sample set with a quality score greater than a preset threshold, X2 is the sample set after removing outliers and supplementing missing values, X3 is the sample set with a quality score less than a preset threshold, σ n1 , σ n2 , σ n3 are the quality score standard deviations of all samples in the sets X1, X2 and X3 respectively, α1, α2 and α3 are all quality weight coefficients, and α1>α2>α3.

[0120] The regularization parameter λ with the optimal performance index is selected to fit the LASSO model, the entire training data set is used for training, and the feature coefficients of each influencing factor are extracted, and all the influencing factors with non-zero feature coefficients are retained as independent influencing factors; multiple groups of LASSO independent influencing factors are obtained through multiple screenings. In this embodiment, the quality weight of the sample is introduced into the LASSO model, so that the influence of the samples with too many outliers and missing values in the original samples on the model is reduced.

[0121] In the embodiment of the application, the quality score calculation directly takes 1-missing value / outlier ratio as the corresponding missing value / outlier quality score, and the weights of the missing values and the outliers are both 0.5.

[0122] In an embodiment of the present application, 54 factors influencing the TCM syndrome type, the number of colds, age, etc. are included, including the basic conditions of the research subjects, lung function, psychological quality, blood routine, liver and kidney function, etc. LASSO regression is used in the training set, and the regression coefficient path diagram is shown in Figure 2 It can be seen that, as the regularization parameter Log(λ) increases, the regression coefficient gradually approaches 0, Figure 2 reflects the process of screening the influencing factors of the model, and then 10-fold cross-validation is performed, and the cross-validation curve is shown in Figure 3 The horizontal axis represents the regularization parameter Log(λ), and the regularization parameter Log(λ) gradually increases from left to right. The vertical axis represents the binomial deviation, and the lower the deviation, the better the model fitting effect. The point represents the average value of the binomial deviation under different regularization parameter Log(λ) values, and as the regularization parameter Log(λ) increases, the binomial deviation first decreases and then increases. In this embodiment, the regularization parameter corresponding to the minimum mean square error is 0.055, at which time 5 influencing factors are selected: the number of colds, the acute exacerbation degree, 6MWD, COPD-PRO and UA, and the coefficient visualization diagram is shown in Figure 4 .

[0123] Further, in S2, the Boruta algorithm is improved to screen the influencing factors, specifically: for the m×n original feature samples in the training set, the original feature samples of each influencing factor in the training set are copied to obtain m×n copied samples, a certain proportion p of the copied samples is extracted to obtain (m×p)×n extracted samples, each column (m×p) extracted sample is randomly transformed to obtain shadow feature samples, the original feature samples and the shadow feature samples are mixed to obtain a new training set, the XGBoost model is trained through the new training set and the importance score of all features in the new training set is calculated, the importance score of each original influencing factor and the corresponding shadow feature is compared, if the importance score of the original influencing factor is greater than the importance score of the shadow feature, the original influencing factor is retained, otherwise it is determined to be pending; the importance score is calculated and compared by iteration to gradually eliminate the original influencing factors, and the remaining influencing factors after stopping the iteration are used as independent influencing factors; a plurality of groups of Boruta independent influencing factors are obtained through multiple screenings.

[0124] In an embodiment of the present application, the Boruta package of R software is used for independent influencing factor screening, and in the feature selection process of the improved Boruta algorithm, the maximum number of iterations is set to 500 to obtain the feature importance score. Figure 5The feature selection results of the improved Boruta algorithm are shown, wherein the blue box plot represents three key statistics (minimum value, median value, maximum value) of the shadow variable importance score, the red box plot represents the irrelevant features rejected (P>0.05), and the green box plot identifies the significant correlation features passing the test (P≤0.05). In this embodiment, one of the groups of independent influencing factors screened by the improved Boruta algorithm includes 6 influencing factors, which are: acute exacerbation degree, cold frequency, CAT, COPD-PRO, SAS and MON%.

[0125] Further, S3 is specifically: taking the plurality of groups of LASSO independent influencing factors and the plurality of groups of Boruta independent influencing factors as sub-training sets respectively; taking the intersection between the plurality of groups of LASSO independent influencing factors and the plurality of groups of Boruta independent influencing factors as a sub-training set; taking the union between the plurality of groups of LASSO independent influencing factors and the plurality of groups of Boruta independent influencing factors as a sub-training set; training a Logistic regression model and an XGBoost model based on each group of sub-training sets respectively to obtain a plurality of LASSO-Logistic models, a plurality of Boruta-Logistic models, a plurality of LASSO-XGBoost models, a plurality of Boruta-XGBoost models, a plurality of LASSO∪Boruta-XGBoost models, and a plurality of LASSO∩Boruta-XGBoost models.

[0126] In the embodiments of the present application, different kinds of COPD-FE risk prediction models are illustrated respectively, and specifically:

[0127] One of the groups of 5 independent influencing factors screened by the LASSO regression is incorporated into the Logistic regression model, and the results show that the cold frequency, acute exacerbation degree, 6MWD and UA are risk factors for frequent acute exacerbation of chronic obstructive pulmonary disease (P<0.05), and the specific parameters of each factor are shown in Table 2; the regression equation is: Logit(P)=-1.1909+0.3206×cold frequency+3.2033×mild acute exacerbation+3.5753×moderate acute exacerbation+4.008×severe acute exacerbation-0.0040×6MWD-0.0052×UA.

[0128] Table 2 LASSO-Logistic regression analysis results

[0129]

[0130] The LASSO-Logistic regression model, on the training set, achieved an accuracy of 0.756, a sensitivity of 0.479, a specificity of 0.897, a PPV of 0.703, an NPV of 0.772, an F1 score of 0.567, a Youden index of 0.376, and an AUC of 0.857. A visualization of the confusion matrix can be found here. Figure 6 ROC curves are shown below. Figure 7 The LASSO-Logistic regression model, in the validation set, achieved an accuracy of 0.771, a sensitivity of 0.412, a specificity of 0.887, a PPV of 0.538, an NPV of 0.825, an F1 score of 0.467, a Youden index of 0.299, and an AUC of 0.847. A visualization of the confusion matrix can be found here. Figure 8 ROC curves are shown below. Figure 9 .

[0131] Six independent influencing factors selected by Boruta features were included in a logistic regression model. The results indicated that the number of colds and the severity of acute exacerbations were risk factors for frequent acute exacerbations of chronic obstructive pulmonary disease (COPD) (P < 0.05). The specific parameters of each factor are shown in Table 3. The regression equation was: Logit(P) = -4.4030 + 0.332 × number of colds + 2.795 × mild acute exacerbation + 3.543 × moderate acute exacerbation + 3.953 × severe acute exacerbation;

[0132] Table 3. Results of Boruta-Logistic Regression Analysis

[0133]

[0134] The Boruta-Logistic regression model, on the training set, achieved an accuracy of 0.717, a sensitivity of 0.319, a specificity of 0.919, a PPV of 0.667, an NPV of 0.726, an F1 score of 0.432, a Youden index of 0.238, and an AUC of 0.835. A visualization of the confusion matrix can be found here. Figure 10 ROC curves are shown below. Figure 11 The Boruta-Logistic regression model, on the validation set, achieved an accuracy of 0.857, a sensitivity of 0.588, a specificity of 0.943, a PPV of 0.769, an NPV of 0.877, an F1 score of 0.667, a Youden index of 0.531, and an AUC of 0.892. A visualization of the confusion matrix can be found here. Figure 12 ROC curves are shown below. Figure 13 .

[0135] The factors of LASSO regression and Boruta feature screening were taken as the union set, and a total of 8 factors were included in the Logistic regression model. The results showed that the frequency of cold, the severity of acute exacerbation, 6MWD and UA were the risk factors for frequent acute exacerbation of chronic obstructive pulmonary disease (P<0.05), and the model was equivalent to the LASSO-Logistic regression model.

[0136] The factors of LASSO regression and Boruta feature screening were taken as the intersection set, and a total of 3 factors were included in the Logistic regression model. The results showed that the frequency of cold and the severity of acute exacerbation were the risk factors for frequent acute exacerbation of chronic obstructive pulmonary disease (P<0.05), and the model was equivalent to the Boruta-Logistic regression model.

[0137] The 5 factors screened by LASSO regression were included in the XGBoost model. After 5-fold cross-validation, the best parameters of the model were nrounds=100, max_depth=3, learning_rate (eta)=0.1, gamma=0.1, colsample_bytree=0.8, min_child_weight=1, and subsample=0.8. The LASSO-XGBoost model in the training set had an accuracy of 0.943, a sensitivity of 0.936, a specificity of 0.946, a PPV of 0.898, a NPV of 0.967, an F1 score of 0.917, a Youden index of 0.882, and an AUC value of 0.984. The visualization of the confusion matrix is shown in Figure 14 , and the ROC curve is shown in Figure 15 . The LASSO-XGBoost model in the validation set had an accuracy of 0.771, a sensitivity of 0.647, a specificity of 0.811, a PPV of 0.524, a NPV of 0.878, an F1 score of 0.579, a Youden index of 0.458, and an AUC value of 0.811. The visualization of the confusion matrix is shown in Figure 16 , and the ROC curve is shown in Figure 17 .

[0138] The 6 factors screened by Boruta feature selection were included in the XGBoost model, and after 5-fold cross-validation, the best parameters of the model were nrounds = 100, max_depth = 3, learning_rate (eta) = 0.1, gamma = 0, colsample_bytree = 0.8, min_child_weight = 3, and subsample = 0.8. The Boruta-XGBoost model in the training set had an accuracy of 0.867, a sensitivity of 0.787, a specificity of 0.908, a PPV of 0.813, a NPV of 0.894, an F1 score of 0.8, a Youden index of 0.695, and a visualization of the confusion matrix was shown in Figure 18 . The AUC value was 0.952, and the ROC curve was shown in Figure 19 . The Boruta-XGBoost model in the validation set had an accuracy of 0.814, a sensitivity of 0.824, a specificity of 0.811, a PPV of 0.583, a NPV of 0.935, an F1 score of 0.683, a Youden index of 0.635, and a visualization of the confusion matrix was shown in Figure 20 . The AUC value was 0.811, and the ROC curve was shown in Figure 21 .

[0139] The union of LASSO regression and Boruta feature selection factors was taken, and a total of 8 factors were included in the XGBoost model. After 5-fold cross-validation, the best parameters of the model were nrounds = 100, max_depth = 3, learning_rate (eta) = 0.1, gamma = 0, colsample_bytree = 0.8, min_child_weight = 1, and subsample = 0.8. The LASSO∪Boruta-XGBoost model in the training set had an accuracy of 0.95, a sensitivity of 0.947, a specificity of 0.951, a PPV of 0.908, a NPV of 0.972, an F1 score of 0.927, a Youden index of 0.898, and an AUC value of 0.99. The visualization of the confusion matrix was shown in Figure 22 , and the ROC curve was shown in Figure 23 . The LASSO∪Boruta-XGBoost model in the validation set had an accuracy of 0.757, a sensitivity of 0.706, a specificity of 0.774, a PPV of 0.5, a NPV of 0.891, an F1 score of 0.585, a Youden index of 0.48, and an AUC value of 0.814. The visualization of the confusion matrix was shown in Figure 24 , and the ROC curve was shown in Figure 25 .

[0140] The intersection of the LASSO regression and the Boruta feature screening is taken as the influencing factor, and 3 factors are taken into the XGBoost model. After 5-fold cross-validation, the best parameters of the model are nrounds = 100, max_depth = 3, learning_rate (eta) = 0.1, gamma = 0.1, colsample_bytree = 0.8, min_child_weight = 1, and subsample = 0.8. The LASSO intersection Boruta-XGBoost model has an accuracy of 0.828, a sensitivity of 0.787, a specificity of 0.849, a PPV of 0.725, a NPV of 0.887, an F1 score of 0.755, a Youden index of 0.636, and an AUC value of 0.91 in the training set. The visualization of the confusion matrix is shown in Figure 26 , and the ROC curve is shown in Figure 27 . The LASSO intersection Boruta-XGBoost model has an accuracy of 0.829, a sensitivity of 0.706, a specificity of 0.868, a PPV of 0.632, a NPV of 0.902, an F1 score of 0.667, a Youden index of 0.574, and an AUC value of 0.877 in the validation set. The visualization of the confusion matrix is shown in Figure 28 , and the ROC curve is shown in Figure 29 .

[0141] Further, S4 is specifically: inputting the validation set into all COPD-FE risk prediction models respectively to obtain COPD-FE risk prediction results, and calculating the AUC index, accuracy, sensitivity, specificity, F1 score, Youden index, PPV index, and NPV index of each COPD-FE risk prediction model according to the COPD-FE risk prediction results.

[0142] Further, S6 is specifically: obtaining the demand for COPD-FE risk prediction, and selecting the COPD-FE risk prediction model with the optimal sensitivity and specificity from the evaluation and screening results when early screening is performed; and selecting the COPD-FE risk prediction model with the optimal interpretability from the evaluation and screening results when clinical risk stratification decision is made.

[0143] In an embodiment of the present application, the remaining models after screening are 1 Boruta-XGBoost model and 1 LASSO intersection Boruta-XGBoost model, and the re-evaluation and visualization of the models are performed, specifically:

[0144] The calibration curves are established for the training set and the validation set of the Boruta-XGBoost model and the LASSO intersection Boruta-XGBoost model. Figures 30-31Calibration curve of Boruta-XGBoost model, Figure 30 Training set, Figure 31 Validation set. It can be seen from the figure that the calibration curves of the two are close to the ideal diagonal line as a whole, and the trend is the same, which means that the predicted probability of the model and the actual occurrence probability are well matched.

[0145] Figures 32-33 Calibration curve of LASSO∩Boruta-XGBoost model, Figure 32 Training set, Figure 33 Validation set. It can be seen from the figure that the calibration curves of the two are close to the ideal diagonal line as a whole, and the trend is the same, which means that the predicted probability of the model and the actual occurrence probability are well matched.

[0146] Figures 34-37 Decision curve results of Boruta-XGBoost model and LASSO∩Boruta-XGBoost model. It can be seen that the performance of Boruta-XGBoost model on the training set is relatively stable, and the net benefit is higher than the "TreatAll" and "Treat None" strategies in most threshold ranges( Figure 34 ); the threshold probability on the validation set has a greater net benefit in the range of about 6.25%~56.25%, 78.13%~81.25%( Figure 35 ). The performance of LASSO∩Boruta-XGBoost model on the training set is also relatively stable, but the net benefit area is not as good as Boruta-XGBoost model( Figure 36 ); the threshold probability on the validation set has a greater net benefit in the range of about 6.25%~81.25%( Figure 37 ). This shows that both have good clinical practicability.

[0147] SHAP method is used to analyze the explainability of Boruta-XGBoost model and LASSO∩Boruta-XGBoost model. In the bee swarm chart, each row corresponds to a variable, and the horizontal axis represents the SHAP value. Each point represents a sample, and its position reflects the SHAP value of the corresponding variable. Each feature is arranged from top to bottom according to importance. SHAP value is used to measure the contribution of a variable to the model output. Generally, red in the bee swarm chart represents a higher SHAP value, indicating that the variable has a greater positive impact on the model output; while blue represents a lower SHAP value, indicating that the variable has a smaller or negative impact on the model output. The bee swarm chart of Boruta-XGBoost model is as follows: Figure 38As shown, the ranking of the features in the model from important to unimportant is: acute exacerbation degree, frequency of cold, monocyte percentage, SAS, CAT and COPD-PRO. From the swarm plot, it can be preliminarily judged that, except for monocyte percentage, the samples with high SHAP values of the remaining variables fall in the positive quadrant, which are positive factors affecting the model, i.e. risk factors. Further verified by the waterfall plot, Figure 39 The waterfall plot of the Boruta-XGBoost model for a single sample, the dashed line in the figure points to E[f(x)] = -0.767, which represents the expected prediction value when the contribution of all features is zero; f(x) = 1.8 is the final prediction value. After adding the expected prediction value and the contribution of each feature, the final prediction value is obtained. Figure 40 The swarm plot of the LASSO∩Boruta-XGBoost model, Figure 41 The waterfall plot of the LASSO∩Boruta-XGBoost model for a single sample. Similarly, acute exacerbation degree, frequency of cold and COPD-PRO are risk factors of the model. The common feature conclusion of the two is the same.

[0148] Further, the evaluation system of the COPD-FE risk prediction model in S5 is specifically:

[0149] Model performance evaluation: the overall performance of the COPD-FE risk prediction model is evaluated based on five dimensions of core parameters, stability, interpretability, prediction efficiency and prediction cost; the weight of each dimension is determined by the influence coefficient;

[0150] Core parameter evaluation: the evaluation score is calculated based on the weight of each core parameter in the COPD-FE risk prediction; the weight of the core parameter is determined by the analytic hierarchy process;

[0151] Threshold screening mechanism: retain the COPD-FE risk prediction model whose performance evaluation result is higher than the evaluation threshold, in the retained COPD-FE risk prediction model, based on AUC index and accuracy for secondary screening, in the remaining COPD-FE risk prediction model of secondary screening, when the difference of AUC index, accuracy and stability is less than the allowed range value, based on prediction cost and interpretability for tertiary screening.

[0152] Further, the weight of each dimension is determined by the influence coefficient:

[0153] Determine the influence result of each dimension on patients, models and hospitals respectively, assign a secondary influence coefficient to each influence result according to different influence degrees, assign a primary influence coefficient to patients, models and hospitals according to prediction requirements, calculate the actual influence coefficient of each dimension based on the primary influence coefficient and the secondary influence coefficient, and calculate the weight according to the actual influence coefficient of each dimension:

[0154]

[0155] wherein w k is the weight of the kth dimension, j k is the actual influence coefficient of the kth dimension, j i is the actual influence coefficient of the ith dimension.

[0156] In one embodiment of the present application, the first influence coefficient assignment for the patient, the model and the hospital are s1, s2 and s3 respectively, wherein the influence of one dimension on the patient's result can have two kinds of results, and the influence degree of the two kinds of results is selected as: greater influence, general influence, lighter influence and basically no influence, and the influence coefficient assignment is s 11 , s 12 , s 13 , s 14 , and the influence on the model and the hospital is determined according to the actual situation analysis, and so on, and then the actual influence coefficient is:

[0157] j = s1*s 13 + s1*s 14 + s2*s 24 + s3*s 33 .

[0158] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same and similar parts between the various embodiments can be referred to each other.

[0159] Various modifications to these embodiments will be apparent to those of ordinary skill in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the present application will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A COPD-FE risk prediction method based on disease syndrome combination, characterized in that, The method comprises the following steps: S1, acquiring the influencing factor data of frequent acute exacerbation of chronic obstructive pulmonary disease of a patient, pre-processing the influencing factor data and assigning variables, and constructing a training data set; S2, screening the influencing factors by LASSO regression and improved Boruta algorithm respectively, to obtain multiple groups of LASSO independent influencing factors and multiple groups of Boruta independent influencing factors; S3, combining the LASSO independent influencing factors and the Boruta independent influencing factors in different ways to obtain multiple groups of sub-training sets, training a Logistic regression model and an XGBoost model based on each group of sub-training sets respectively, and obtaining multiple COPD-FE risk prediction models; S4, verifying the performance indicators of all COPD-FE risk prediction models through a validation set; S5, constructing an evaluation system of the COPD-FE risk prediction model, evaluating and screening the performance of all COPD-FE risk prediction models; S6, selecting a COPD-FE risk prediction model meeting the requirements based on the evaluation and screening results to predict the COPD-FE risk.

2. The COPD-FE risk prediction method based on syndrome combination according to claim 1, characterized in that, S1 specifically comprises: S11, determining the diagnostic criteria of the patient, and including the patients meeting the diagnostic criteria into the target population of the data set collection; S12, determining the data source, the prediction time node and the true result label, and acquiring the time series data of the influencing factors of the patient's frequent acute exacerbation of chronic obstructive pulmonary disease; S13, performing outlier rejection and missing value supplement processing on the time series data of the influencing factors; S14, assigning variables to the processed influencing factors, constructing input features containing the influencing factor data and the data acquisition time node, taking the input features and the corresponding prediction time node and true result label as training samples, and obtaining the training data set.

3. The COPD-FE risk prediction method based on syndrome combination according to claim 2, characterized in that, S13 specifically comprises: For time series data of an influencing factor, calculate its Q L value, Q U value and Q R value, and determine whether all data are abnormal data according to the calculation results, if the data q meets one of q < Q L -1.5Q R or q > Q U -1.5Q R , it is determined as abnormal data, and the data q is excluded; wherein Q L is the lower quartile, Q U is the upper quartile, Q R is the difference between Q L and Q U . For the time series data after removing outliers, the prior distribution function p(x q ) of the missing value x q is calculated; according to the time series data X={x1,x2,...,x n} after removing outliers, the joint distribution function p(X|x q ) is calculated; the posterior distribution p(x q |X) of the missing value x q is calculated; Compute the estimate of the missing value x q : The estimated value is supplemented into the time series data as a missing value.

4. The COPD-FE risk prediction method based on syndrome combination according to claim 2, characterized in that, In S2, the LASSO regression is used to screen the influencing factors, specifically: the training data set is standardized, the training data set is divided into multiple subsets, for each training data subset, it is taken as a validation set, and the remaining training data subsets are taken as training sets, to obtain multiple LASSO training sets, for each candidate regularization parameter λ, the LASSO model is trained on all LASSO training sets and the model performance is evaluated by mean square error, and the average mean square error of the LASSO model under all training sets is calculated as the performance indicator of the corresponding regularization parameter λ; the loss function of the LASSO model is: In the formula, N is the number of samples, y n is the true value corresponding to the sample n, x n is the nth sample, p is the model coefficient, ω n is the weight of the nth sample, s n is the quality score of the nth sample, the quality score is calculated according to the proportion of abnormal values and the proportion of missing values, X1 is the sample set with a quality score greater than a preset threshold, X2 is the sample set after abnormal value elimination and missing value supplementation, X3 is the sample set with a quality score lower than the preset threshold, σ n1 , σ n2 , σ n3 are the standard deviations of the quality scores of all samples in the sets X1, X2 and X3 respectively, and α1, α2 and α3 are all quality weight coefficients, and α1>α2>α3. The regularization parameter λ with the optimal performance indicator is selected to fit the LASSO model, the entire training data set is used for training, and the feature coefficients of each influencing factor are extracted, and all influencing factors with non-zero feature coefficients are retained as independent influencing factors; multiple groups of LASSO independent influencing factors are obtained through multiple screenings.

5. The COPD-FE risk prediction method based on syndrome combination according to claim 1, characterized in that, In S2, the influencing factors are screened by improving the Boruta algorithm. Specifically, for the m*n original feature samples in the training set, the original feature samples of each influencing factor in the training set are copied to obtain m*n copied samples. A certain proportion p of the copied samples is extracted to obtain (m*p) n extracted samples. Each column (m*p) extracted sample is randomly transformed to obtain shadow feature samples. The original feature samples and the shadow feature samples are mixed to obtain a new training set. The XGBoost model is trained based on the new training set, and the importance scores of all features in the new training set are calculated. The importance scores of each original influencing factor and the corresponding shadow feature are compared. If the importance score of the original influencing factor is greater than the importance score of the shadow feature, the original influencing factor is retained, otherwise it is determined as pending. The original influencing factors are gradually eliminated by iterative training, importance score calculation and comparison. The remaining influencing factors after the iteration are stopped are used as independent influencing factors. A plurality of groups of Boruta independent influencing factors are obtained through multiple screenings.

6. The COPD-FE risk prediction method based on syndrome combination according to claim 1, characterized in that, In S3, a plurality of groups of LASSO independent influencing factors and a plurality of groups of Boruta independent influencing factors are used as sub-training sets. The intersection between the plurality of groups of LASSO independent influencing factors and the plurality of groups of Boruta independent influencing factors is used as a sub-training set. The union between the plurality of groups of LASSO independent influencing factors and the plurality of groups of Boruta independent influencing factors is used as a sub-training set. Logistic regression models and XGBoost models are trained based on each group of sub-training sets to obtain a plurality of LASSO-Logistic models, a plurality of Boruta-Logistic models, a plurality of LASSO-XGBoost models, a plurality of Boruta-XGBoost models, a plurality of LASSO∪Boruta-XGBoost models, and a plurality of LASSO∩Boruta-XGBoost models.

7. The COPD-FE risk prediction method based on syndrome combination according to claim 1, characterized in that, In S4, the validation set is input into all COPD-FE risk prediction models to obtain COPD-FE risk prediction results. The AUC index, accuracy, sensitivity, specificity, F1 score, Youden index, PPV index and NPV index of each COPD-FE risk prediction model are calculated based on the COPD-FE risk prediction results.

8. The COPD-FE risk prediction method based on syndrome combination according to claim 7, characterized in that, In S6, the demand for COPD-FE risk prediction is obtained. When early screening is performed, the COPD-FE risk prediction model with the optimal sensitivity and specificity is selected from the evaluation and screening results. When clinical risk stratification decision is made, the COPD-FE risk prediction model with the optimal interpretability is selected from the evaluation and screening results.

9. The COPD-FE risk prediction method based on syndrome combination according to claim 7, characterized in that, The evaluation system of the COPD-FE risk prediction model in S5 is as follows: Model performance evaluation: The overall performance of the COPD-FE risk prediction model is evaluated based on five dimensions of core parameters, stability, interpretability, prediction efficiency and prediction cost. The weight of each dimension is determined by an influence coefficient. Core parameter evaluation: based on each core parameter, the evaluation score is calculated for the COPD-FE risk prediction weight, and the weight of the core parameter is determined by the analytic hierarchy process; Threshold screening mechanism: the COPD-FE risk prediction model with a performance evaluation result higher than the evaluation threshold is retained, and in the retained COPD-FE risk prediction model, secondary screening is carried out based on the AUC index and accuracy, and in the remaining COPD-FE risk prediction model after secondary screening, when the differences in AUC index, accuracy and stability are less than the allowed range value, tertiary screening is carried out based on prediction cost and interpretability.

10. The COPD-FE risk prediction method based on syndrome combination according to claim 9, characterized in that, The weight of each dimension is determined by the influence coefficient, which is specifically: Determine the influence result of each dimension on the three aspects of patients, models and hospitals respectively, assign a secondary influence coefficient to each influence result according to the different influence degrees, assign a primary influence coefficient to patients, models and hospitals according to the prediction requirements, calculate the actual influence coefficient of each dimension based on the primary influence coefficient and the secondary influence coefficient, and calculate the weight according to the actual influence coefficient of each dimension: In the formula, w k is the weight of the kth dimension, j k is the actual influence coefficient of the kth dimension, j i is the actual influence coefficient of the ith dimension.

Images