A method and system for constructing a glomerular filtration rate estimation model

Abstract

A method and system for constructing a glomerular filtration rate estimation model, including: step 1: performing mRMR feature selection on the original data set, screening out the independent variable features, and using the transformed data set as the basic data set for subsequent modeling; step 2: constructing two linear regression models on the basic data set; step 3: performing sampling with replacement on the basic data set, and removing duplicate samples; step 4: training the set model on the N sampled data sets after removing duplicates in step 3, and obtaining its parameters: obtaining N temporary local models, and performing multivariate linear regression modeling on the N sampled data sets after removing duplicates according to the above two forms obtained in step 2; step 5: constructing a global model, using the average of the local model parameters obtained in step 4 as the parameters of the global model, and the obtained global model is the constructed glomerular filtration rate estimation model. The present invention models data for a specific population and obtains a GFR estimation model that is more suitable for the specific population.

Description

Technical Field

[0001] The present invention belongs to the technical field of glomerular filtration rate estimation models, and in particular relates to a method and system for constructing a glomerular filtration rate estimation model. Background Art

[0002] In the prior art, due to the poor operability of directly measuring the glomerular filtration rate (GFR), the GFR estimation method based on serum creatinine concentration (SCr) and serum cystatin C is widely used in clinical practice. The GFR estimated by the formula is called eGFR, which has become the main basis for diagnosing CKD stages 3 to 5. This method is low-cost and less harmful to the patient's body, but it is difficult to determine which of the existing multiple GFR estimation formulas is better, and the estimation accuracy is difficult to guarantee.

[0003] The reason why there are multiple GFR estimation formulas is that the GFR estimation formulas applicable to different races, ages, genders, and different types of kidney disease are different and cannot be generalized. Therefore, it is necessary to build a dedicated GFR estimation model for specific populations.

[0004] However, the training data available for model construction is very limited. The main reason is that the training data needs to include the GFR gold standard. However, due to the high cost and strong side effects of directly measuring GFR, the corresponding technology has been eliminated in the clinic, so the available qualified samples are extremely limited. In the field of machine learning, this is a small sample problem, and due to the small number of samples, the trained model is not reliable, and even the usual cross-validation technology cannot be used to evaluate the model performance. Summary of the invention

[0005] In order to overcome the deficiencies of the above-mentioned prior art, the purpose of the present invention is to provide a method and system for constructing a glomerular filtration rate estimation model, which uses small sample data to build a model for a specific population to obtain a more targeted GFR estimation model.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for constructing a glomerular filtration rate estimation model comprises the following steps:

[0008] Step 1: Data preprocessing, including feature selection and log transformation, performing mRMR feature selection on the original data set, screening out independent variable features, wherein the independent variable features include cystatin C and age, and the data set after log transformation of the independent variable features is used as the basic data set for subsequent modeling; illustratively, in the present invention, the independent variable features may also consist only of cystatin C and age, without any other features.

[0009] Step 2: Local model setting: construct two linear regression models based on the basic data set obtained in step 1;

[0010] Step 3: Remove over-resampled data, perform replacement sampling on the basic data set, and remove duplicate samples;

[0011] Step 4: Training of the local model: The set local model is trained on the N sample data sets after deduplication in step 3. The MSE is calculated for the training set samples according to formula (1) during each training.

[0012]

[0013] Where n is the number of samples, y i is the true value of the label of the i-th sample, is the predicted value of the i-th sample given by the model. When MSE no longer decreases with training iterations, the training ends. In this way, N temporary local models can be obtained. The two linear regression models obtained in step 2 are respectively used to perform multivariate linear regression modeling on the N sample data sets after deduplication; N 0+1 type local models and N 0+2 type local models are obtained;

[0014] Step 5: Global model construction: the local model trained in step 4 is used to take the average of the parameters of the local model as the parameters of the global model. The obtained global model is the constructed glomerular filtration rate estimation model.

[0015] The feature selection in step 1 is specifically as follows:

[0016] The original data set contains 22 samples, each of which contains a patient's GFR measurement value and 231 features. First, mRMR feature selection is performed on these candidate features to screen out the top two features: cystatin C and age;

[0017] The log transformation is specifically:

[0018] The logarithm with base e was taken for the GFR measurement values ​​and the selected independent variable features, and the data set after feature selection and log transformation was used as the basic data set for subsequent modeling.

[0019] The local model in step 2 is a linear regression model, which is as follows:

[0020] y=w T x+b (2)

[0021] Where x is the input, y is the model output, w and b are the model parameters. After learning w and b through training, the model is determined. For the GFR measurement problem, the value y of the dependent variable GFR in step 1 and the value x1 of the independent variable characteristic age and the value x2 of cystatin C are taken with the base e as the logarithm, and x′1=log e x1,x′2=log e x2,y′=log e y, and use this data to build two linear regression models:

[0022] ①y′=a1x′1+a2x′2+a3 (0+1 type)

[0023] ②y′=c1(x′1) 2 +c2(x′2) 2 +c3x′1x′2+c4 (0+2 type)

[0024] Since formula ① only contains constant terms (zero-order terms) and linear terms, it is named 0+1 type; formula ② only contains constant terms and quadratic terms, it is named 0+2 type.

[0025] In step 3, the excess deduplication and replacement sampling is divided into two steps:

[0026] 1) Over-sampling with replacement: The basic data set is sampled with replacement. The number of samples in each sampling data set is the same as that of the basic data set. The sampling process is repeated N times to obtain N sampling data sets, so that the sampling data sets cover as comprehensive sample combinations as possible. The larger the number of sampling times N, the higher the accuracy of the final model. In the following modeling example, N is 1000, and each sampling data set contains 22 samples.

[0027] 2) Deduplication: Check whether there are duplicate samples in each sampling data set. If so, only one sample is retained from the duplicate samples.

[0028] The step 4 is specifically to train the set model on the N sample data sets after deduplication, and obtain its parameters: the model trained with the i-th data set is y=w i T x+b i , get N temporary local models

[0029] For the GFR estimation problem, multiple linear regression modeling is performed on the N sampling data sets of 0+1 type and 0+2 type set in step 2 above, and N 0+1 type local models and N 0+2 type local models (here N = 1000) are obtained. The 0+1 type local model constructed on the i-th data set is The 0+2 type local model is i=1,2,…,N.

[0030] The step 5 is specifically as follows:

[0031] Take the average of the parameters of N local models as the parameters of the global model, that is, in The corresponding global model is:

[0032]

[0033] For the GFR estimation problem, the average values ​​of the parameters of the 0+1 local model obtained in the previous step are calculated. As global model parameters, the 0+1 type global model is as follows:

[0034]

[0035] For the 0+2 type local model, the average value of its parameters is also calculated

[0036] As global model parameters, the 0+2 type global model is as follows:

[0037]

[0038] In the global model construction step, in addition to using the mean of the local model parameters to obtain the parameters of the global model, the median of the local model parameters can also be used to obtain the parameters of the global model.

[0039] For the GFR estimation problem, since the variables in the above formula have been log-transformed, it is also necessary to derive the reduced form of the formula;

[0040] For the 0+1 global model, log e x1=x′1,long e x2=x′2,long e Substituting y=y′ into equation (4) yields:

[0041]

[0042] Taking the exponential of both sides of equation (6) with base e at the same time, we can get the reduced form of the 0+1 type global model:

[0043]

[0044] Similarly, the reduction form for the 0+2 type global model is:

[0045]

[0046] The final 0+1 and 0+2 models are:

[0047] y=303.5359·x1 -0.3237 x2 -0.9369 (9)

[0048]

[0049] where x1 and x2 refer to age and cystatin C, respectively.

[0050] A method for evaluating a glomerular filtration rate estimation model, comprising the following steps:

[0051] Step 1: Add slight noise to the features of 22 samples in the original data set. Since only slight noise is added, the GFR measurement value of the sample is taken as the label value of the noisy sample to construct M noisy data sets.

[0052] Step 2: For the glomerular filtration rate estimation model to be evaluated, the model input is the sample feature value, and the output is the GFR estimation value. The samples in the noisy data set obtained in step 1 are used as the model input to obtain the GFR estimation value of the model for these samples. Then, the label value of the sample is used to calculate the MSE of the model to be evaluated (which can be any glomerular filtration rate estimation model) on each noisy data set. The calculation formula is as shown in formula (1); M MSEs can be obtained on M noisy data sets;

[0053] Step 3: Let the MSE on the i-th data set be MSE i , take the average As an evaluation indicator;

[0054] MSE i It reflects the matching ability of the model with the i-th noisy data set. The closer the two are, the better the MSE is. i The lower, It reflects the average matching ability of the model with various noisy data. This shows that on average, the model matches all kinds of noisy data well, and thus also matches unseen data well, that is, the model has good generalization ability.

[0055] A device for estimating glomerular filtration rate, comprising:

[0056] A collection unit, the collection unit comprising a cystatin C collection device and an age collection device;

[0057] A processing unit, running the glomerular filtration rate estimation model, receiving the data from the cystatin C collection device and the age collection device in a wired or wireless manner, and estimating the glomerular filtration rate;

[0058] The display unit receives and displays the estimated glomerular filtration rate via wired or wireless means.

[0059] The hardware form of the cystatin C collection device is a smart phone or a computer, and the data transmission mode with the processing unit is a Socket mode.

[0060] The hardware form of the age collection device is a smart phone or a computer, and the data transmission mode with the processing unit is Socket mode.

[0061] The hardware form of the processing unit is a computer.

[0062] Beneficial effects of the present invention:

[0063] (1) The technology involved in the present invention can be used for, but not limited to, GFR estimation for kidney disease. For GFR estimation for kidney disease, a more targeted model can be developed, for example, a GFR estimation model that is more targeted than the existing model can be constructed based on data from a specific population.

[0064] (2) Low sample bias and high model accuracy. During sampling, duplicate samples are removed so that each local model avoids sample bias. The global model established through multiple different local models avoids the problem of each local model preferring to respond to local samples, thus achieving high accuracy.

[0065] (3) Low data dependence, especially suitable for small sample situations. By sampling with replacement multiple times to obtain a sampling data set containing a variety of sample combinations, the dependence on data and its scale is reduced, solving the problem of poor modeling results due to insufficient training data.

[0066] (4) Small model scale. The global model finally established is a single linear regression model with small model scale, small storage volume and high response speed.

[0067] (5) A measure of the model’s generalization ability is proposed, which is applicable to small sample situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a schematic diagram of the technical route of the present invention.

[0069] Figure 2 For various models Schematic diagram. DETAILED DESCRIPTION

[0070] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0071] like Figure 1 Shown: A method for constructing a glomerular filtration rate estimation model

[0072] (1) Data preprocessing: A method for constructing a glomerular filtration rate estimation model includes the following steps;

[0073] Data preprocessing is divided into two steps:

[0074] Feature selection: The original data set contains 22 samples, each of which contains a patient's GFR measurement value and 231 features. First, mRMR feature selection was performed on these candidate features, and the top two features were screened out: cystatin C and age.

[0075] Log transformation: Take the logarithm with base e for the GFR measurement value and the selected feature independent variables. The dataset after feature selection and log transformation is used as the basic dataset for subsequent modeling.

[0076] (2) Local model setting

[0077] The local model is set as a linear regression model, as follows:

[0078] y=w T x+b (1)

[0079] Where x is the input, y is the model output, and w and b are model parameters. After learning w and b through training, the model is determined.

[0080] For the GFR measurement problem, we take the logarithm with base e for the dependent variable GFR value y and the characteristic independent variable age value x1 and cystatin C value x2, and we get x′1=log e x1,x′2=log e x2,y′=log e y, and use this data to build two linear regression models:

[0081] y′=a1x′1+a2x′2+a3 (type 0+1)

[0082] y′=c1(x′1) 2 +c2(x′2) 2 +c3x′1x′2+c4 (0+2 type)

[0083] If it only contains constant terms (zero-order terms) and linear terms, it is named 0+1 type; if it only contains constant terms and quadratic terms, it is named 0+2 type.

[0084] (3) Over-sampling of data:

[0085] Excessive deduplication and replacement sampling is divided into two steps:

[0086] Excess sampling with replacement: The basic data set is sampled with replacement. The number of samples in each sampling data set is the same as that of the basic data set. The sampling process is repeated N times to obtain N sampling data sets. Here, the number of sampling times N is required to be as large as possible so that the sampling data set can cover as comprehensive sample combinations as possible. The larger the number of sampling times N, the higher the accuracy of the final model. In the following modeling example, N is 1000 and each sampling data set contains 22 samples.

[0087] Deduplication: Check whether there are duplicate samples in each sampling data set. If so, delete the duplicate samples.

[0088] (4) Training of local models:

[0089] The set model is trained on the N sample data sets after deduplication to obtain its parameters: Let the model trained with the i-th data set be y=w i T x+b i , get N temporary local models It is worth noting that the local models here are only used as an intermediate process to build the global model and will not be retained after the global model is obtained. This is different from the base learners in ensemble learning, which are components of the final classifier.

[0090] For the GFR estimation problem, multiple linear regression modeling is performed on the N sample data sets after deduplication according to the above two forms, and N 0+1 type local models and N 0+2 type local models are obtained (here N = 1000). Suppose the 0+1 type local model constructed on the i-th data set is The 0+2 type local model is i=1,2,…,N.

[0091] (5) Global model construction

[0092] Take the average of the parameters of N local models as the parameters of the global model, that is, in The corresponding global model is:

[0093]

[0094] The global model obtained by this method is a single regression model, which is different from the model obtained by ensemble learning which is a fusion of multiple models.

[0095] For the GFR estimation problem, the average values ​​of the parameters of the 0+1 local model obtained in the previous step are calculated. As global model parameters, the 0+1 type global model is as follows:

[0096]

[0097] For the 0+2 type local model, the average value of its parameters is also calculated

[0098] As global model parameters, the 0+2 type global model is as follows:

[0099]

[0100] In particular, in the global model construction step, in addition to using the mean of the local model parameters to obtain the parameters of the global model, the median of the local model parameters can also be used to obtain the parameters of the global model.

[0101] Regarding the GFR estimation problem, since the variables in the above formula have undergone log transformation, it is also necessary to derive the reduced form of the formula.

[0102] For the 0+1 global model, log e x1=x′1,log e x2=x′2,log e Substituting y=y′ into equation (3) yields:

[0103]

[0104] Taking the exponential of both sides of equation (5) with base e at the same time, we can get the reduced form of the 0+1 type global model:

[0105]

[0106] Similarly, the reduction form for the 0+2 type global model is:

[0107]

[0108] The final 0+1 and 0+2 models are:

[0109] y=303.5359·x1 -0.3237 x2 -0.9369 (8)

[0110]

[0111] where x1 and x2 refer to age and cystatin C, respectively.

[0112] (6) Global model performance evaluation

[0113] In view of the small sample size, it is impossible to use cross-validation for model training and testing. The present invention proposes to use all samples for training and a new evaluation measure: by adding small noise to the samples as input, the average value of the MSE of the model output is obtained. As an evaluation indicator of the generalization ability of the global model. The smaller it is, the stronger the model's generalization ability is.

[0114] The evaluation method consists of three steps:

[0115] 1) Add noise to the samples and construct M noisy data sets. Noise is added to simulate data that the model has never seen. The ability of a model to predict data it has never seen is its generalization ability.

[0116] 2) Calculate the model's predicted value for each sample in a noisy data set, and then use the sample's true label value to calculate the model's MSE on this data set. The calculation formula is as follows:

[0117]

[0118] Where n is the number of samples, y i is the true value of the label of the i-th sample, is the predicted value of the i-th sample given by the model.

[0119] In this way, M MSEs can be obtained on M noisy data sets.

[0120] 3) Let the MSE on the i-th data set be MSE i , take the average As an evaluation indicator.

[0121] MSE i It reflects the matching ability of the model with the i-th noisy data set. The closer the two are, the better the MSE is. i The lower. It reflects the average matching ability of the model with various noisy data. This shows that on average, the model matches all kinds of noisy data well, and thus also matches unseen data well, that is, the model has good generalization ability.

[0122] Model comparison experiment

[0123] Next, we will conduct a comparative experiment between our established model and the existing model to verify the advantages of our established model. The experimental platform is MATLAB.

[0124] All the models used in the experiment are shown in Table 1, of which the first eight are models reported in the literature, and the last two are models established by us.

[0125] Table 1 All models used in the experiment

[0126]

[0127] The experiment is to test the model's ability to predict and generalize data that it has never seen by adding noise to the samples. The experimental plan is as follows:

[0128] There are 22 samples in total for the comparative experiment. The sample labels are GFR measurement values, and the features involved include age, gender, cystatin C, and serum creatinine (different models use different features, see Table 1 for details).

[0129] (1) Adding noise to the sample

[0130] The noise adding method is to add a small random noise to the value of each feature of the existing sample to form a new sample. Let max be one-fiftieth of the feature mean, and the value range of the random noise added to this feature is (0, max), as shown in Table 2. Each sample is noised in this way to obtain a noised data set. Repeat this method M times to obtain M noised data sets, where M=1000.

[0131] Table 2 Noise values ​​of input samples

[0132]

[0133] (2) MSE calculation

[0134] For a model, find its predicted value for each sample in a noisy data set, and then use the true value of the sample label to find its MSE on this data set. In this way, M MSEs can be obtained on M noisy data sets. Let the MSE on the i-th data set be MSE i , take the average As an evaluation indicator.

[0135] (3) Experimental results

[0136] Various models like Figure 2 shown.

[0137] Depend on Figure 2 It can be seen that Model 9 and Model 10 The two models are the 0+1 model and the 0+2 model established by the method of this patent. This shows that the generalization ability of the model established by the method of this patent is stronger than that of the existing models, and the 0+2 model has a better effect.

[0138] Purpose of the present invention:

[0139] (1) Modeling is performed based on data from a specific population, so that a GFR estimation model that is more suitable for that specific population can be obtained.

[0140] (2) A sampling data set containing a variety of sample combinations is obtained through multiple sampling with replacement. This can reduce the dependence on the data size and solve the problem of difficulty in dividing the training / test set due to a small data set. It is more suitable for small sample situations.

[0141] (3) A sampling data set without duplicate samples is obtained through deduplication operation, and multiple temporary local models are established on different sampling data sets to construct the final global model, which better avoids sample bias.

[0142] (4) Construct a single regression model, which can make the model small in size, small in storage volume, and high in response speed.

[0143] (5) Propose a model generalization ability measure suitable for small sample situations.

[0144] The key points and protection points of the present invention are:

[0145] (1) Modeling process: By building a temporary local model as an intermediate process, taking the average parameter values ​​of all local models as the parameters of the global model, the global model is used for GFR assessment. This model is only a single regression model, unlike the Bagging algorithm which is a fusion of multiple regression models, so the model size is small, the storage volume is small, and the response speed is high.

[0146] (2) Oversampling process: The number of bootstraps N is required to be large, and a local model is trained using the samples from each bootstrap. Since N is large, the sampled data set can basically cover all combinations of samples, so that the final global model has high accuracy.

[0147] (3) The model has no sample bias: The sample data set is deduplicated, so that each local model avoids sample bias. The global model established by the local models obtained through each sampling avoids the problem that each local model prefers the corresponding local samples. The final model has no sample bias and is highly accurate.

[0148] (4) Evaluation index of generalization ability: The present invention proposes a new evaluation method, which adds small noise to the sample as input, obtains the average value of the MSE of the model output, and uses it as the evaluation index of the global model. This index can measure the generalization ability of the model. The generalization ability of a model refers to the model's ability to predict data that it has never seen. As we all know, the MSE of a model reflects the matching ability between the model and the data. The better the match, the lower the MSE. Therefore, the MSE may be the result of the model overfitting the data. The evaluation index proposed in this patent - a model's ——Reflects the average matching ability of the model with various noisy data, thus This shows that on average, the model matches all kinds of noisy data well, and thus also matches unseen data well, that is, the model has good generalization ability.

[0149] One application form of the present invention provides a glomerular filtration rate estimation device, including a collection unit, a processing unit and a display unit. Among them, the collection unit is a collection device for cystatin C and age. Cystatin C is data separated from a blood sample. To obtain cystatin C and age data, the collection device can be a data upload directly from the detection end, or it can be manually entered. Obviously, in terms of hardware, the collection device can be a device with communication function such as a smart phone, a computer, a touch screen, etc. In this embodiment, the data transmission method selects the Socket method. The processing unit runs the aforementioned glomerular filtration rate estimation model, which is a device with computing and communication functions, such as a computer, a cloud, etc., which receives cystatin C and age data by wire or wirelessly, and estimates the glomerular filtration rate using the estimation model. The display unit is used to display the results, which should have a display interface and communication functions, for example, it can be integrated with the collection device, it can also be integrated with the processing unit, or it can be a separate display device.

[0150] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When the use is implemented in whole or in part in the form of a computer program product, the computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, the process or function described in the embodiment of the present invention is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website site, computer, server or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL) or wireless (e.g., infrared, wireless, microwave, etc.) mode) to another website site, computer, server or data center. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state hard disk SolidState Disk (SSD)), etc.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a glomerular filtration rate estimation model, characterized in that: The steps include: Step 1: Data preprocessing, Including feature selection and log transformation, mRMR feature selection is performed on the original data set to screen out independent variable features, the independent variable features include cystatin C and age, and the data set after the independent variable features are log-transformed is used as the basic data set for subsequent modeling; Step 2: Local model setting, construct two linear regression models for the basic data set obtained in step 1; Step 3: Remove over-resampled data, perform replacement sampling on the basic data set, and remove duplicate samples; Step 4: Training of the local model: training the set model on the N sample data sets after deduplication in step 3. For each training, the MSE is calculated according to formula (1) for the training set samples. Where n is the number of samples, y i is the true value of the label of the i-th sample, is the predicted value of the i-th sample given by the model. When the MSE no longer decreases with the training iterations, the training is terminated. In this way, N temporary local models can be obtained. The two linear regression models obtained in step 2 are used to perform multivariate linear regression modeling on the N sample data sets after deduplication. N 0+1 type local models and N 0+2 type local models are obtained. Step 5: constructing a global model, using the local model trained in step 4, taking the average of the parameters of the local model as the parameters of the global model, and the obtained global model is the constructed glomerular filtration rate estimation model; The local model in step 2 is a linear regression model, which is as follows: y=w T x+b (2) Where x is the input, y is the model output, w and b are the model parameters. After learning w and b through training, the model is determined. For the GFR measurement problem, the value y of the dependent variable GFR in step 1 and the value x1 of the independent variable characteristic age and the value x2 of cystatin C are taken with the base e as the logarithm, and x′1=log2x1,x′2=lgo e x2,y′=log e y, and use this data to build two linear regression models: ①y′=a1x1′+a2x2′+a3(0+1 type) ②y′=c1(x1′) 2 +c2(x2′) 2 +c3x1′x2′+c4(0+2 type) Since formula ① only contains constant terms (zero-order terms) and linear terms, it is named 0+1 type; formula ② only contains constant terms and quadratic terms, it is named 0+2 type.

2. The method for constructing a glomerular filtration rate estimation model according to claim 1, characterized in that: The feature selection in step 1 is specifically as follows: The original data set contains 22 samples, each of which contains a patient's GFR measurement value and 231 features. First, mRMR feature selection is performed on these candidate features to screen out the top two features: cystatin C and age; The log transformation is specifically: The logarithm with base e was taken for the GFR measurement values ​​and the selected independent variable features, and the data set after feature selection and log transformation was used as the basic data set for subsequent modeling.

3. The method for constructing a glomerular filtration rate estimation model according to claim 1, characterized in that: In step 3, the excess deduplication and replacement sampling is divided into two steps: 1) Oversampling with replacement: The basic data set is sampled with replacement. The number of samples in each sampling data set is the same as that of the basic data set. The sampling process is repeated N times to obtain N sampling data sets, so that the sampling data set N covers as comprehensive sample combinations as possible. The larger the sampling number N, the higher the accuracy of the final model. Each sampling data set contains 22 samples. 2) Deduplication: Check whether there are duplicate samples in each sampling data set. If so, only one sample is retained from the duplicate samples.

4. The method for constructing a glomerular filtration rate estimation model according to claim 1, characterized in that: The step 4 is specifically to train the set model on the N sample data sets after deduplication, and obtain its parameters: the model trained with the i-th data set is y=w i T x+b i , get N temporary local models For the GFR estimation problem, multiple linear regression modeling is performed on the N sampling data sets of 0+1 type and 0+2 type set in step 2 above, and N 0+1 type local models and N 0+2 type local models are obtained. The 0+1 type local model constructed on the i-th data set is The 0+2 type local model is 5. The method for constructing a glomerular filtration rate estimation model according to claim 1, characterized in that: The step 5 is specifically as follows: Take the average of the parameters of N local models as the parameters of the global model, that is, in The corresponding global model is: For the GFR estimation problem, the average values ​​of the parameters of the 0+1 local model obtained in the previous step are calculated. N = 1000 is used as the global model parameter, and the 0+1 type global model is as follows: For the 0+2 type local model, the average value of its parameters is also calculated N = 1000 is used as the global model parameter, and the 0+2 type global model is as follows: In the global model construction step, in addition to using the mean of the local model parameters to obtain the parameters of the global model, the median of the local model parameters can also be used to obtain the parameters of the global model.

6. The method for constructing a glomerular filtration rate estimation model according to claim 5, characterized in that: For the GFR estimation problem, since the variables in the above formula have been log-transformed, it is also necessary to derive the reduced form of the formula; For the 0+1 global model, log e x1=x′1,log e x2=x′2,log e Substituting y=y′ into equation (4) yields: Taking the exponential of both sides of equation (6) with base e at the same time, we can get the reduced form of the 0+1 type global model: Similarly, the reduction form for the 0+2 type global model is: The final 0+1 and 0+2 models are: y=303.5359·x1 -0.3237 ·x2 -0.9369 (9) where x1 and x2 refer to age and cystatin C, respectively.

7. A method for evaluating a glomerular filtration rate estimation model obtained based on the method according to any one of claims 1 to 6, characterized in that: A method for evaluating a glomerular filtration rate estimation model to be evaluated comprises the following steps; Step 1: Add slight noise to the features of 22 samples in the original data set. Since only slight noise is added, the GFR measurement value of the sample is taken as the label value of the noisy sample to construct M noisy data sets. Step 2: For the glomerular filtration rate estimation model to be evaluated, the model input is the sample feature value, and the output is the GFR estimation value. The samples in the noisy data set obtained in step 1 are used as the model input to obtain the GFR estimation value of the model for these samples. The label value of the sample is then used to calculate the MSE of the model to be evaluated, which can be any glomerular filtration rate estimation model on each noisy data set; M MSEs can be calculated on M noisy data sets; Step 3: Let the MSE on the i-th data set be MSE i , take the average As an evaluation indicator; MSE i It reflects the matching ability of the model with the i-th noisy data set. The closer the two are, the better the MSE is. i The lower, It reflects the average matching ability of the model with various noisy data. This shows that on average, the model matches all kinds of noisy data well, and thus also matches unseen data well, that is, the model has good generalization ability.

8. A glomerular filtration rate estimation device constructed based on the method according to any one of claims 1 to 6, characterized in that: include: A collection unit, the collection unit comprising a cystatin C collection device and an age collection device; A processing unit, running the glomerular filtration rate estimation model, receiving the data from the cystatin C collection device and the age collection device in a wired or wireless manner, and estimating the glomerular filtration rate; A display unit receives and displays the estimated glomerular filtration rate via a wired or wireless method; The hardware form of the cystatin C collection device is a smart phone or a computer, and the data transmission method with the processing unit is a Socket method; The hardware form of the age collection device is a smart phone or a computer, and the data transmission method with the processing unit is Socket; The hardware form of the processing unit is a computer.

Images