A multi-scale fusion optimization method for diabetic nephropathy classification based on routine test data

Abstract

The present invention discloses a multi-scale fusion optimization method for classifying diabetic nephropathy based on routine test data, and relates to the technical field of diabetic nephropathy classification. The method comprises inputting the dimension of blood routine test from the perspective of L1_1, inputting the dimension of urine routine test from the perspective of L1_2, fitting the blood routine test and urine routine test respectively using a double loss function, and extracting an abstract feature layer; the abstract feature layer is nonlinearly combined through L3; the three layers of neural networks L4, L5, and L6 are jointly learned, and finally the blood routine test dimension and urine routine test dimension information are output respectively through L7_1 and L7_2. The present invention discloses a multi-scale fusion optimization method for classifying diabetic nephropathy based on routine test data, which takes the neural network of a multilayer perceptron as the basic model, constructs a shared fusion structure, and adopts a shared fusion training method to make the blood routine test dimension richer than the information obtained by direct splicing, so that the model training effect is better.

Description

Technical Field

[0001] The present invention relates to the technical field of diabetic nephropathy classification, and in particular to a multi-scale fusion optimization diabetic nephropathy classification method based on conventional test data. Background Art

[0002] Diabetic nephropathy (DN) is one of the common microvascular complications of diabetes. The disease is very harmful and has a high incidence rate. Therefore, early detection of DN is of great significance for preventing the development of related diseases.

[0003] Most existing technologies are based on artificial intelligence algorithms to predict and identify diseases, but this method currently has many defects when applied to the medical industry:

[0004] First, standard medical data is difficult to obtain. Medical data is different from traditional data forms. Most of them are jointly produced by hospitals or published by hospitals for learning and discussion. They are difficult to obtain naturally and the amount of data is small.

[0005] Second, the current neural network training method is relatively simple, usually just taking the test dimension of uniform size as input, which makes it difficult to learn complex features in different dimensions;

[0006] Third, medical data requires the cooperation of doctors to organize and confirm the data, and doctors rely on years of clinical experience, which takes a long time, is costly, and has poor sample accuracy. Previous model learning only used direct linear splicing of data, ignoring the feature fusion information between data. Summary of the invention

[0007] To solve the above problems, the present invention provides a multi-scale fusion optimization diabetic nephropathy classification method based on routine test data. This method can learn more routine urine information from routine blood information, and can learn more nonlinear information than the previous direct splicing training. By constructing a multi-input parallel fusion model, the algorithm accuracy is improved, the workload of doctors is reduced, and the work efficiency of doctors is improved.

[0008] To achieve the above object, the present invention provides a multi-scale fusion optimization diabetic nephropathy classification method based on conventional test data, comprising the following steps:

[0009] (1) Establish blood routine data set S1 and urine routine data set S2;

[0010] (2) The blood routine dimension of S1 is input from the perspective of L1_1, and the urine routine dimension of S2 is input from the perspective of L1_2. The double loss function is used to fit the blood routine and urine routine respectively to extract the abstract feature layer;

[0011] (3) The abstract feature layer is nonlinearly combined through L3;

[0012] (4) After the three-layer neural network L4, L5, and L6 learn together, the blood routine dimension and urine routine dimension information are output through L7_1 and L7_2 respectively;

[0013] (5) The established multi-input multi-output shared parameter model is used to classify diabetic nephropathy.

[0014] Preferably, the input blood routine dimensions include: "gender", "age", "WBC", "RBC", "PDW", "MCV", "NEUT%", "LYMPH%", "EO%", "BASO%", "NEUT", "LYMPH", "BASO", "HGB", "HCT", "MCH", "MCHC", "R-CV", "PLT", "MPV", "PCT", "MONO", "MONO%", "EO", a total of 24 dimensions of features.

[0015] Preferably, the input urine routine dimensions include: 'urine α1 microspheres / creatinine', 'urine IGG / creatinine', 'urine TRU / creatinine', 'urine white microspheres / creatinine', a total of 4 dimensions.

[0016] Preferably, the data order should be disrupted when inputting the blood routine dimension and urine routine dimension to prevent the model from learning sequential features.

[0017] Preferably, both loss functions are cross entropy loss functions.

[0018] The present invention provides a multi-scale fusion optimization diabetic nephropathy classification method based on conventional test data, which has the following beneficial effects:

[0019] (1) A fusion training scheme using routine blood tests and routine urine tests is proposed. Routine blood tests and routine urine tests are used as routine test data. They are easy to obtain in a medical context, have high universality, are simple to classify, and have strong model interpretability.

[0020] (2) The characteristic information of urine routine was integrated into the dimension of blood routine through the fusion scheme, which improved the accuracy of the algorithm model. The data fusion method enabled the blood routine data to learn information from different data sources, thus avoiding the problem of data waste.

[0021] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a neural network model of multi-layer perceptron;

[0023] Figure 2 Fusion training process for multi-input and multi-output shared parameter models. DETAILED DESCRIPTION

[0024] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention clearer, the embodiments of the present invention are further described in detail in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals throughout represent the same or similar elements or elements with the same or similar functions.

[0025] It should be noted that the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or inherent to these processes, methods, products or devices.

[0026] Like reference numerals and letters denote similar items in the following drawings, and thus, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0027] Embodiment 1

[0028] A multi-layer perceptron neural network model is established by directly splicing and training single blood routine data or blood routine data set and urine routine data set. The model structure is as follows Figure 1 shown.

[0029] Figure 1 middle represents the input of the multilayer perceptron, represents the hidden layer, y Represents the model output category and the corresponding probability. The input and output formulas of the hidden layer are as follows:

[0030]

[0031] in, From 1 to n The order represents how many layers of connections there are; represents the weight of the fully connected layer.

[0032] The establishment process includes:

[0033] (1) A total of 74,982 blood routine and urine routine data sets were extracted from Beijing Tongren Hospital affiliated to Capital Medical University from 2013 to 2022. After data processing and feature engineering such as special value removal, the blood routine data set S1 and urine routine data set S2, as well as the data set S3 with complete dimensions of blood routine and urine routine data, were obtained;

[0034] Dataset S1 is the structured test data of routine blood tests, and dataset S2 is the structured test data of routine urine tests. The two datasets have the same dimension, the same behavior, and the same person's data; dataset S3 is the dataset after splicing under the second dimension. The S1, S2, and S3 datasets share the same test set Y.

[0035] (2) Construct a multi-layer neural network model_1 to train the data set S1 and obtain the model's indicators for the data set S1.

[0036] (2) Construct a multi-layer neural network model_2 to train the data set S3 and obtain the indicators for the test set Y.

[0037] Embodiment 2

[0038] Construct a multi-input fusion neural network model model_3 to perform fusion training on data set S1 and data set S2, and obtain the indicators of model_3 for the test set Y.

[0039] like Figure 2 As shown in the figure, routine blood data are loaded and input from the L1_1 and L2_2 model layers, and routine urine data are loaded and input from the L2_1 and L2_2 model layers. The first two layers of models compress the features of the test data respectively, extract them to the abstract feature layer, and then fuse them through L3 to nonlinearly combine the feature information of routine blood data with the feature information of routine urine data; then the feature extraction process is carried out through the three-layer neural network of L4, L5, and L6, and finally the dimension information of routine blood and routine urine is output through L7_1 and L7_2 respectively.

[0040] The specific steps include:

[0041] (1) A total of 22 items of blood routine test and 24 features of gender and age were put into the network for training, namely "gender", "age", "WBC", "RBC", "PDW", "MCV", "NEUT%", "LYMPH%", "EO%", "BASO%", "NEUT", "LYMPH", "BASO", "HGB", "HCT", "MCH", "MCHC", "R-CV", "PLT", "MPV", "PCT", "MONO", "MONO%", and "EO".

[0042] The dimensions of urine routine examination are 'urine α1 microspheres / creatinine', 'urine IGG / creatinine', 'urine TRU / creatinine', and 'urine white microspheres / creatinine', a total of 4 dimensions.

[0043] (2) The data order is disrupted to prevent the model from learning sequential features; the blood routine data set S1 and the urine routine data set S2 are input for fusion training.

[0044] L7_1 of the blood routine test dimension is used as the basis to verify the model performance in the test set Y.

[0045] Results Analysis

[0046] The evaluation indicators used for the validation set Y are: TPR , TNR , ACC , and the calculation formulas are:

[0047] ;

[0048] ;

[0049] ;

[0050] in, TP represents the number of correctly identified positives; TN represents the number of correctly identified negatives; FP represents the number of false positives; FN represents the number of falsely identified negatives; TPR The sensitivity represents the proportion of correctly identified positives to all positives that should have been identified; TNR It represents the proportion of correctly identified negatives to all those that should have been negatives, which is specificity; ACC Indicates the accuracy of the model's prediction of the classification results.

[0051] The classification accuracy of model fusion and simple data splicing is shown in Table 1:

[0052] Table 1 Algorithm results comparison table

[0053] ;

[0054] Data set S1 is single routine blood data; data set S3 is the result of concatenating routine blood data and routine urine data according to the second dimension; the fusion result of data sets S1 and S2 is obtained by using the fusion model of Example 2, the routine blood data and routine urine data are respectively input into the model layer, the input is collected into the test dimensions of routine blood data and routine urine data, extracted into the abstract feature layer, the features are combined and weights are shared in the third layer of the model, and finally different dimensional information is output respectively.

[0055] This fusion method allows blood routine to better learn the characteristic information of urine routine. The test results of the training model are shown in Table 1 above. The AUC value of the fusion model using data sets S1 and S2 is the largest in the table. The model correctly distinguishes positive and negative samples, and has better robustness and generalization ability. TPR and TNR Compared with the other two methods, it has decreased, but still remains at a reasonable level. It can be found that the overall effect of the fusion model is the best.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A multi-scale fusion optimization method for diabetic nephropathy classification based on conventional test data, characterized in that: The steps include: (1) Establish blood routine data set S1 and urine routine data set S2; (2) The blood routine dimension of S1 is input from the perspective of L1_1, and the urine routine dimension of S2 is input from the perspective of L1_2. The double loss function is used to fit the blood routine and urine routine respectively to extract the abstract feature layer; (3) The abstract feature layer is nonlinearly combined through L3; (4) After the three-layer neural network L4, L5, and L6 learn together, the blood routine dimension and urine routine dimension information are output through L7_1 and L7_2 respectively; (5) Apply the established multi-input multi-output shared parameter model to classify diabetic nephropathy; The input blood routine dimensions include: "gender", "age", "WBC", "RBC", "PDW", "MCV", "NEUT%", "LYMPH%", "EO%", "BASO%", "NEUT", "LYMPH", "BASO", "HGB", "HCT", "MCH", "MCHC", "R-CV", "PLT", "MPV", "PCT", "MONO", "MONO%", "EO", a total of 24 dimensions; The input urine routine dimensions include: ‘urine α1 microspheres / creatinine’, ‘urine IGG / creatinine’, ‘urine TRU / creatinine’, ‘urine IgG / creatinine’, a total of 4 dimensions; When inputting the dimensions of blood routine and urine routine, the data order should be disrupted to prevent the model from learning sequential features; Both loss functions are cross entropy loss functions.

Images