Diabetes collaborative prediction diagnosis model and method based on deep neural network

Abstract

The invention discloses a diabetes collaborative prediction diagnosis model and method based on a deep neural network. The diabetes collaborative prediction diagnosis model comprises a data preprocessing module, a diabetes prediction model based on different features and a custom risk decision strategy module. The data pre-processing module constructs a data sequence processing flow in a mode of series connection of abnormal value processing, data standardization and feature selection; the diabetes prediction model based on different features comprises an input layer, a hidden layer and an output layer; the input layer is constructed through an output construction prediction module of the preprocessing part; the hidden layer comprises a full connection layer and a Dropout layer; the output layer is used for outputting a prediction result; the custom risk decision strategy module comprises a diabetes prediction module based on high-dimensional nonlinear features, a prediction module based on potential data features and a prediction module based on patient similarity. Experimental results show that the sensitivity, specificity, error omission rate, diagnosis ratio and AUC of the prediction model provided by the invention are superior to those of a comparison algorithm.

Description

Technical Field

[0001] The present invention relates to a diabetes collaborative prediction diagnosis model and method based on a deep neural network, and belongs to the technical field of intelligent medical diagnosis. Background Art

[0002] Diabetes is a typical chronic disease with a long course and complex complications, which seriously affects the quality of life and economic burden of patients and their families. If early and accurate prediction of diabetes can be achieved based on patient health data, the development of diabetes can be effectively prevented or alleviated. Existing diabetes prediction algorithms have achieved good prediction results, but how to achieve noise filtering and high-dimensional nonlinear feature learning in diabetes data is still an important factor affecting the prediction performance of the algorithm. Therefore, how to improve the prediction performance of the algorithm through noise filtering and high-dimensional nonlinear feature learning deserves further study. In view of the above situation, this paper proposes an integrated diabetes prediction model based on a custom risk decision strategy. Summary of the invention

[0003] In view of the problems existing in the above-mentioned prior art, the present invention provides a collaborative predictive diagnosis model and method for diabetes based on deep neural networks. The model integrates modules such as data preprocessing, diabetes prediction models based on different features, and customized risk decision strategies in a serial manner, and uses each module to hierarchically solve the problems of data standardization, data set noise filtering, and high-dimensional nonlinear feature learning, thereby further improving the prediction performance of the integrated prediction model.

[0004] In order to achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a diabetes collaborative prediction diagnosis model and method based on deep neural network, characterized by comprising a data preprocessing module, a diabetes prediction model based on different characteristics, and a custom risk decision strategy module;

[0005] The data preprocessing module constructs a data sequential processing flow in a serial manner of outlier processing, data standardization and feature selection; the outliers include outliers and null values;

[0006] The characteristic vector of diabetes is defined as (1), where each element represents a specific clinical indicator;

[0007] r={f pregnant ,f glucose ,f pressure ,f triceps ,f insulin ,f bmi ,f pedigree ,f age} (1);

[0008] The outlier processing method is as follows:

[0009]

[0010] Among them, the clinical indicators in r are number of pregnancies, 2-hour blood glucose, blood pressure, skin thickness, 2-hour serum insulin, body mass index, diabetes spectrum function, and age; r i represents the i-th clinical indicator in r, and there exists r i ∈r; R i Represents clinical index r i The collection of R i,mean and R i,std Respectively represent R i The standard deviation and mean of i * Indicates that after processing i ;

[0011] The diabetes prediction model based on different features includes an input layer, a hidden layer and an output layer; the input layer is obtained by outputting the preprocessing part. select Construct a prediction module to construct the output of the preprocessing part; the hidden layer includes a fully connected layer and a Dropout layer; the introduction of the Dropout layer is used to reduce the computational complexity of the FDNN and avoid its overfitting; the output layer includes only one neuron for predicting the results Output:

[0012] The customized risk decision strategy module includes three base modules, namely, a diabetes prediction module based on high-dimensional nonlinear features, a prediction module based on potential data features, and a prediction module based on patient similarity; as shown in formula (11).

[0013]

[0014] Among them, Output represents the prediction result of FDNN; m+1 Represents the diabetes prediction result based on high-dimensional nonlinear features; y output Represents the prediction result based on the underlying data features; y 0 / 1 Represents the prediction results based on patient similarity.

[0015] Furthermore, the clinical index r i The fluctuation and distribution difference of ∈r will affect the prediction performance of FDNN. i * Standardization is performed, that is, through the linear transformation formula (3), r i * Mapping to a uniform magnitude and interval;

[0016]

[0017] in, Represents r i * The value set of ; and Respectively The standard deviation and mean of i,norm represents the standardized clinical index r i * .

[0018] Furthermore, based on the principle of indicator-indicator independence and indicator-diagnosis correlation, feature selection is introduced to screen the indicators in r, that is, under the given hyperparameter Q condition, the correlation calculation function f is used Q (r norm ,r target ) Calculate the feature r i,norm With feature r j,norm , feature r i,norm and diagnosis target When the calculated correlation is greater than the threshold K, we consider the feature to be valid, otherwise we discard the feature; where r target Representation and feature vector r norm The corresponding diagnosis;

[0019] r select = {r i,select :f(r i,select ,r norm ,r target )≥K|Q} (4);

[0020] Among them, r select represents the filtered feature set, and r select ∈r; r i,select Represents r select The i-th clinical index in; the Pearson correlation coefficient is used as the specific correlation calculation function f Q (r norm ,r target ), the calculation formula is shown in formula (5);

[0021]

[0022] Among them, R i,norm,mean and R j,norm,mean Respectively represent r i,norm and r j,norm The set of corr (r i,norm ,r j,norm ) represents r i,norm and r j,norm The Pearson correlation coefficient of is in the range of [0,1]; when ρcorr (r i,norm ,r j,norm ) is closer to 1, r i,norm and r j,norm The more significant the correlation is, the lower it is; i,norm With r target The correlation is also calculated using formula (5). Based on formula (5), the characteristic heat map can also be obtained. In this paper, the larger the correlation coefficient, the lighter the color of the corresponding position, and vice versa.

[0023] Furthermore, the forward propagation function of the Dropout layer is shown in formula (6).

[0024]

[0025] Among them, i represents the output of the i-th layer. When i=1, o 1 That is the input layer; p i represents the probability vector, which obeys the Bernoulli distribution and determines the neurons discarded in the i-th hidden layer; W i represents the weight matrix of the i-th layer, W i T W i The transposed matrix of i represents the bias term of the i-th layer. m represents the number of hidden layers. When i=m+1, o m+1 It is the output layer of the diabetes prediction model based on deep neural network.

[0026] Furthermore, the difference between the output value and the true value is measured by the loss function. The loss function can be expressed as formula (7), that is, under a given weight parameter W i Under the conditions of other hyperparameters Q, determine whether the loss value is less than the given threshold K loss When loss>K loss When W i and Q are updated; and when loss ≤ K loss When the model converges, stop W i and the update of Q. Among them, loss represents the loss value, f loss represents the loss function, W i represents the weight of the i-th layer of the deep neural network, It means partial derivative.

[0027] loss = {min:f loss (o i ,o i,fact )|W i ,Q}(7);

[0028]

[0029] Furthermore, the gradient boosting tree GBDT is used to realize the diabetes prediction model based on potential data features. This module uses formula (7) as the loss function and takes the output of the preprocessing part {r select ,r target} is input, and diabetes prediction is achieved through formula (9);

[0030] Among them, y output represents the output of the prediction module based on the underlying data features; m represents the number of decision trees in the module; Represents the predicted value of the i-th decision tree; Represents the output of the i-th decision tree and satisfies

[0031] Furthermore, based on KNN, we explored the influence of expert experience on the prediction results and proposed a diabetes prediction model based on patient similarity. Specifically, we calculated the similarity between the target sample and the potential sample according to formula (10), and classified the target sample based on statistics according to the principle of "the minority obeys the majority".

[0032]

[0033] Among them, y 0 / 1 represents the output of the prediction module based on patient similarity; K indiv represents the number of potential samples; f distance Represents the similarity function, used to calculate K indiv The feature vector of the sample and the target feature vector r target The similarity between 0 / y 1 Optional output representing a sample.

[0034] Furthermore, sensitivity, specificity and accuracy are used as evaluation indicators of the model prediction performance, and the calculation methods of each indicator are shown in formulas (12)-(14).

[0035]

[0036] Among them, TP is the number of correctly predicted positive samples; TN is the number of correctly predicted negative samples; FP is the number of incorrectly predicted positive samples; FN is the number of incorrectly predicted negative samples; in addition, ROC-AUC is also a comparative evaluation indicator, which is calculated by sklearn's roc_auc_score; in order to reduce the error, all experiments are repeated multiple times, and then the average of their prediction results is taken as the experimental result.

[0037] The beneficial effect of the present invention is that the model integrates data preprocessing, diabetes prediction model based on different features, customized risk decision strategy and other modules in a serial manner, hierarchically solves the data standardization, data set noise filtering and high-dimensional nonlinear feature learning problems in the prediction model, thereby further improving the prediction performance of our diabetes prediction model. A series of experiments were conducted based on the PID data set to compare this model with other commonly used models. The experimental results show that the sensitivity, specificity, false omission rate, diagnostic odds ratio and AUC of the prediction model proposed in this paper are better than the comparison algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic diagram of an integrated diabetes prediction model based on a customized risk decision-making strategy of the present invention;

[0039] Figure 2 It is a schematic diagram of the diabetes prediction model framework based on high-dimensional nonlinear features of the present invention;

[0040] Figure 3 It is a schematic diagram of the diabetes prediction model framework based on potential data features of the present invention;

[0041] Figure 4 (a) is a schematic diagram of data distribution before data preprocessing of the present invention;

[0042] Figure 4 (b) is a schematic diagram of data distribution after data preprocessing of the present invention;

[0043] Figure 5 It is a schematic diagram of the correlation between the processed corresponding clinical indicator attributes and the diagnosis of the present invention;

[0044] Figure 6 (a) is a schematic diagram of feature selection of the PIMA dataset of the present invention;

[0045] Figure 6 (b) is a schematic diagram of feature selection of the noisy PIMA data set of the present invention;

[0046] Figure 7 (a) is a schematic diagram of the sensitivity, specificity, accuracy and AUC-ROC of the FDNN on the PIMA dataset of the present invention;

[0047] Figure 7 (b) is a schematic diagram of the sensitivity, specificity, accuracy and AUC-ROC of the FDNN on the noisy PIMA dataset of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail below through the accompanying drawings and embodiments. However, it should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by technicians in the technical field to which the present invention belongs. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0050] This paper focuses on the noise filtering and high-dimensional nonlinear feature learning problems of diabetes datasets, and proposes a diabetes integrated prediction model FDNN based on a custom risk decision strategy. The model mainly includes data preprocessing, diabetes prediction models based on different features, and custom risk decision strategies. Its framework is as follows Figure 1 shown.

[0051] As an indispensable module in FDNN, data preprocessing constructs a data sequential processing flow in the form of outlier processing, data standardization and feature selection in series to improve the quality of input data for the prediction model. Outliers include outliers and null values. Among them, outliers are abnormal data that obviously deviate from the normal value range, which will significantly affect the fit and performance of the prediction model; null values ​​are not the same as 0, and null values ​​cannot be used as input data for the model. Therefore, data preprocessing first filters outliers in the data set. In this paper, the feature vector of diabetes is defined as formula (1), where each element represents a specific clinical indicator. The outlier processing method is shown in formula (2).

[0052] r={f pregnant ,f glucose ,f pressure ,f triceps ,f insulin ,f bmi ,f pedigree ,f age}(1)

[0053]

[0054] Among them, the clinical indicators in r are number of pregnancies, 2-hour blood glucose, blood pressure, skin thickness, 2-hour serum insulin, body mass index, diabetes spectrum function, and age; r i represents the i-th clinical indicator in r, and there exists r i ∈r; R i Represents clinical index r i The collection of R i,mean and R i,std Respectively represent Ri The standard deviation and mean of i * Indicates that after processing i .

[0055] Clinical indicators i The fluctuation and distribution difference of ∈r may also affect the prediction performance of FDNN. Here we i * Standardization is performed, that is, ri is mapped to a uniform magnitude and interval through the linear transformation formula (3).

[0056]

[0057] in, Represents r i * The value set of ; and Respectively The standard deviation and mean of i,norm represents the standardized clinical index r i * .

[0058] In addition, the dimension of r may also be an important factor affecting the prediction performance of FDNN. When the dimension of r is much larger than the number of samples, it may lead to the dimensionality disaster; and when there are redundant indicators in r, it not only directly increases the computational complexity of FDNN but also may lead to overfitting of FDNN. For this reason, we introduce feature selection to screen the indicators in r based on the principle of indicator-indicator independence and indicator-diagnosis correlation. The feature selection used in this paper is shown in formula (4), that is, under the given hyperparameter Q condition, the correlation calculation function f is used to calculate the feature. Q (r norm ,r target ) Calculate the feature r i,norm With feature r j,norm , feature r i,norm and diagnosis target When the calculated correlation is greater than the threshold K, we consider the feature to be valid, otherwise we discard the feature. targe t represents the characteristic vector r norm The corresponding diagnosis.

[0059] r select = {r i,select :f(r i,select ,r norm ,r target )≥K|Q} (4)

[0060] Among them, r selectrepresents the filtered feature set, and r selec t∈r.r i,select Represents r selec The Pearson correlation coefficient is used as the specific correlation calculation function f Q (r norm ,r target ), the calculation formula is shown in formula (5).

[0061]

[0062] Among them, R i,norm,mean and R j,norm,mean Respectively represent r i,norm and r j,norm The collection of ρ corr (r i,norm ,r j,norm ) represents r i,norm and r j,norm The Pearson correlation coefficient of is in the range [0,1]. corr (r i,norm ,r j,norm ) is closer to 1, r i,norm and r j,norm The more significant the correlation is, the lower it is. ,norm With r target The correlation is also calculated using formula (5). Based on formula (5), the characteristic heat map can also be obtained, and in this paper, the larger the correlation coefficient, the lighter the color of the corresponding position, and vice versa.

[0063] Deep neural networks have significant advantages in learning high-dimensional nonlinear features of data. Therefore, we introduce deep neural networks into FDNN to build a diabetes prediction model based on high-dimensional nonlinear features. This module is mainly composed of input layer, hidden layer and output layer. The hidden layer includes fully connected layer and Dropout layer. Its framework is as follows: Figure 2 shown.

[0064] like Figure 2 As shown in the figure, in the diabetes prediction model based on high-dimensional nonlinear features, we first use the output r of the preprocessing part select The input layer of the prediction module is constructed. The hidden layer is connected in an alternating manner with fully connected layers and Dropout layers. The introduction of the Dropout layer can not only reduce the computational complexity of the FDNN, but also avoid its overfitting. The output layer includes only one neuron, which is used to predict the results. Output of . In this paper, the forward propagation function of the Dropout layer is shown in formula (6).

[0065]

[0066] Among them, i represents the output of the i-th layer. When i=1, o 1 That is the input layer; p i represents the probability vector, which obeys the Bernoulli distribution and determines the neurons discarded in the i-th hidden layer; W i represents the weight matrix of the i-th layer, W i T W i The transposed matrix of i represents the bias term of the i-th layer. m represents the number of hidden layers. When i=m+1, o m+1 is the output layer of the diabetes prediction model based on deep neural network. In addition, we use the loss function to measure the difference between the output value and the true value. The loss function can be expressed by formula (7), that is, under a given weight parameter W i Under the conditions of other hyperparameters Q, determine whether the loss value is less than the given threshold K loss When loss>K loss When W i and Q are updated; and when loss ≤ K loss When the model converges, stop W i and the update of Q. Among them, loss represents the loss value, f loss represents the loss function, W i represents the weight of the i-th layer of the deep neural network, It means partial derivative.

[0067] loss = {min:f loss (o i ,o i,fact )|W i ,Q}(7)

[0068]

[0069] In addition to high-dimensional nonlinear features, there may be other potential data features in the data set, which may also affect the prediction performance of FDNN. Gradient boosting tree GBDT is an integrated algorithm based on decision trees, which consists of multiple decision trees and iterates the model through gradient descent. Therefore, we introduced the gradient boosting tree GBDT (GBDT) to implement a diabetes prediction model based on potential data features. This module uses formula (7) as the loss function and takes the output of the preprocessing part {r select ,r target} is used as input, and diabetes prediction is achieved through formula (9).

[0070]

[0071] Among them, y output represents the output of the prediction module based on the underlying data features; m represents the number of decision trees in the module. Represents the predicted value of the i-th decision tree. Represents the output of the i-th decision tree and satisfies

[0072] In the medical field, machine learning models based on data features can effectively assist clinical diagnosis, but the experience of clinical experts is equally important. Therefore, based on KNN, we explore the impact of expert experience on prediction results and propose a diabetes prediction model based on patient similarity. Specifically, the similarity between the target sample and the potential sample is calculated according to formula (10), and the target sample is classified based on statistics according to the principle of "the minority obeys the majority".

[0073]

[0074] Among them, y 0 / 1 represents the output of the prediction module based on patient similarity; K indiv represents the number of potential samples; f distance Represents the similarity function, used to calculate K indiv The feature vector of the sample and the target feature vector r target The similarity between 0 / y 1 Optional output representing a sample.

[0075] There are three main base modules in FDNN, namely diabetes prediction module based on high-dimensional nonlinear features, prediction module based on potential data features, and prediction module based on patient similarity. In order to fully consider the impact of high-dimensional nonlinear features, potential features and expert experience on the prediction results, we build a custom risk decision strategy based on different prediction modules, as shown in formula (11).

[0076]

[0077] Among them, Output represents the prediction result of FDNN; m+1 Represents the diabetes prediction result based on high-dimensional nonlinear features; y output Represents the prediction result based on the underlying data features; y 0 / 1 Represents the prediction results based on patient similarity.

[0078] In this paper, sensitivity, specificity and accuracy are used as evaluation indicators of model prediction performance. The calculation methods of each indicator are shown in formulas (12)-(14).

[0079]

[0080] Among them, TP is the number of correctly predicted positive samples; TN is the number of correctly predicted negative samples; FP is the number of incorrectly predicted positive samples; FN is the number of incorrectly predicted negative samples. In addition, ROC-AUC is also an evaluation indicator for our comparison, which is calculated by sklearn's roc_auc_score. In order to reduce the error, we repeated all experiments 20 times and took the average of the prediction results as the experimental results.

[0081] From the above discussion, it can be seen that variable parameters directly affect the prediction performance of FDNN. Therefore, this paper adjusts and determines the specific values ​​of each variable parameter through experiments. The initial values ​​of each variable parameter are shown in Table 2. First, the central value of the variable parameter is determined by the grid search algorithm, and then the central value is fine-tuned by interval search to avoid the fitting difference problem caused by outliers, so as to determine the optimal value of the variable parameter. At the same time, considering the sensitivity of the L2 paradigm to outliers, the L1 paradigm is introduced in the loss function of FDNN.

[0082] like Figure 4 As shown in (a), a scatter plot is used to visualize the data distribution of each clinical indicator (element in the diabetes feature vector) in the PIMA dataset. As can be seen from the figure, there are significant outliers in the distribution of each clinical indicator, and there is no obvious distinction between negative and positive samples, which may be an important factor affecting the performance of diabetes prediction. In this case, we preprocess the PIMA dataset according to equations (2) and (3). After preprocessing, the distribution of each clinical indicator is as follows Figure 4 (b) is shown. Figure 4 (b) It can be seen that the distribution of each clinical indicator after filtering tends to be concentrated, and the fluctuation tends to be stable. At this time, we calculate the correlation matrix between each clinical indicator and disease diagnosis before and after data preprocessing, as shown in Figure 5 Compared with the correlation matrix before data preprocessing, the correlation between Glucose, Blood Pressure and BMI and disease diagnosis has been improved after data preprocessing, making them more relevant to disease labels. Figure 5 Among them, 0-7 correspond to clinical indicators Pregnancies, Glucose, Blood Pressure, Skin Thickness, 2-hour Serum Insulin (short for Insulin), BMI, Diabetes Pedigree Function, Age; 8 corresponds to disease diagnosis.

[0083] In order to verify the prediction performance of FDNN, we compared it with base algorithms such as SVM, Logistic, RF, GBDT, etc. through experiments. The experimental results are shown in Table 3 and Figure 7 shown.

[0084] Table 3

[0085]

[0086] According to Table 3 and Figure 7 , we get the following results:

[0087] (1) The sensitivity, specificity, accuracy and AUC-ROC of FDNN are basically not affected by data noise. FDNN also has stable prediction performance on noisy data sets. It should be noted that the AUC-ROC of FDNN is 0.8463 in the absence of noise, which is lower than the 0.8467 of Logistic. In addition, the sensitivity, specificity, accuracy and AUC-ROC of FDNN are always the highest under different conditions, which shows that FDNN has better prediction performance than the comparison algorithm. For example, on the PID data set and the noisy data set, the prediction accuracy of FDNN is 10.15% and 5.79% higher than that of GBDT, respectively.

[0088] (2) The evaluation indicators of FDNN on the PIMA dataset and the noisy dataset are basically the same, while the evaluation indicators of the comparison algorithm fluctuate. This not only verifies the importance of feature selection, but also shows that FDNN with feature selection function has better anti-interference ability and is suitable for application scenarios with noise.

[0089] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent substitution or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A diabetes collaborative prediction and diagnosis model and method based on deep neural network, characterized in that: It includes data preprocessing module, diabetes prediction model based on different features, and customized risk decision strategy module; The data preprocessing module constructs a data sequential processing flow in a serial manner of outlier processing, data standardization and feature selection; the outliers include outliers and null values; The characteristic vector of diabetes is defined as (1), where each element represents a specific clinical indicator; r={f pregnant ,f glucose ,f pressure ,f triceps ,f insulin ,f bmi ,f pedigree ,f age }(1); The outlier processing method is as follows: Among them, the clinical indicators in r are number of pregnancies, 2-hour blood glucose, blood pressure, skin thickness, 2-hour serum insulin, body mass index, diabetes spectrum function, and age; r i represents the i-th clinical indicator in r, and there exists r i ∈r; R i Represents clinical index r i The collection of R i,mean and R i,std Respectively represent R i The standard deviation and mean of Indicates that after processing i ; The diabetes prediction model based on different features includes an input layer, a hidden layer and an output layer; the input layer is obtained by outputting the preprocessing part. select Construct a prediction module to construct the output of the preprocessing part; the hidden layer includes a fully connected layer and a Dropout layer; the introduction of the Dropout layer is used to reduce the computational complexity of the FDNN and avoid its overfitting; the output layer includes only one neuron for predicting the results Output: The customized risk decision strategy module includes three base modules, namely, a diabetes prediction module based on high-dimensional nonlinear features, a prediction module based on potential data features, and a prediction module based on patient similarity; as shown in formula (11). Among them, Output represents the prediction result of FDNN; m+1 Represents the diabetes prediction result based on high-dimensional nonlinear features; y output Represents the prediction result based on the underlying data features; y 0 / 1 Represents the prediction results based on patient similarity.

2. A diabetes collaborative prediction and diagnosis model and method based on deep neural network according to claim 1, characterized in that: The clinical index i The fluctuation and distribution difference of ∈r will affect the prediction performance of FDNN. Standardization is performed, that is, through the linear transformation formula (3), Mapping to a uniform magnitude and interval; in, express The value set of ; and Respectively The standard deviation and mean of i,norm Indicates standardized clinical indicators 3. A diabetes collaborative prediction and diagnosis model and method based on deep neural network according to claim 2, characterized in that: Based on the principle of indicator-indicator independence and indicator-diagnosis correlation, feature selection is introduced to screen the indicators in r, that is, under the given hyperparameter Q condition, the correlation calculation function f is used Q (r norm ,r target ) Calculate the feature r i,norm With feature r j,norm , feature r i,norm and diagnosis target When the calculated correlation is greater than the threshold K, we consider the feature to be valid, otherwise we discard the feature; where r target Representation and feature vector r norm The corresponding diagnosis; r select ={r i,select :f(r i,select ,r norm ,r target )≥K|Q}(4); Among them, r select represents the filtered feature set, and r select ∈r; r i,select Represents r select The i-th clinical index in; the Pearson correlation coefficient is used as the specific correlation calculation function f Q (r norm ,r target ), the calculation formula is shown in formula (5); Among them, R i,norm,mean and R j,norm,mean Respectively represent r i,norm and r j,norm The set of corr (r i,norm ,r j,norm ) represents r i,norm and r j,norm The Pearson correlation coefficient of is in the range of [0,1]; when ρ corr (r i,norm ,r j,norm ) is closer to 1, r i,norm and r j,norm The more significant the correlation is, the lower it is; i,norm With r target The correlation is also calculated using formula (5). Based on formula (5), the characteristic heat map can also be obtained. In this paper, the larger the correlation coefficient, the lighter the color of the corresponding position, and vice versa.

4. A diabetes collaborative prediction and diagnosis model and method based on deep neural network according to claim 3, characterized in that: The forward propagation function of the Dropout layer is shown in formula (6). Among them, i represents the output of the i-th layer. When i=1, o1 is the input layer; p i represents the probability vector, which obeys the Bernoulli distribution and determines the neurons discarded in the i-th hidden layer; W i represents the weight matrix of the i-th layer, W i T W i The transposed matrix of i represents the bias term of the i-th layer. m represents the number of hidden layers. When i=m+1, o m+1 It is the output layer of the diabetes prediction model based on deep neural network.

5. A diabetes collaborative prediction and diagnosis model and method based on deep neural network according to claim 4, characterized in that: The difference between the output value and the true value is measured by the loss function. The loss function can be expressed as formula (7), that is, under a given weight parameter W i Under the conditions of other hyperparameters Q, determine whether the loss value is less than the given threshold K loss When loss>K loss When W i and Q are updated; and when loss ≤ K loss When the model converges, stop W i and the update of Q. Among them, loss represents the loss value, f loss represents the loss function, W i represents the weight of the i-th layer of the deep neural network, It means partial derivative. loss={min:f loss (o i ,o i,fact )|W i ,Q}(7); 6. A diabetes collaborative prediction and diagnosis model and method based on deep neural network according to claim 5, characterized in that: The diabetes prediction model based on potential data features is implemented by gradient boosting tree GBDT. This module uses formula (7) as the loss function and takes the output of the preprocessing part {r select ,r target } is input, and diabetes prediction is achieved through formula (9); Among them, y output represents the output of the prediction module based on the underlying data features; m represents the number of decision trees in the module; Represents the predicted value of the i-th decision tree; Represents the output of the i-th decision tree and satisfies 7. A diabetes collaborative prediction and diagnosis model and method based on deep neural network according to claim 6, characterized in that: Based on KNN, we explore the influence of expert experience on prediction results and propose a diabetes prediction model based on patient similarity. Specifically, we calculate the similarity between the target sample and the potential sample according to formula (10), and classify the target sample based on statistics according to the principle of "the minority obeys the majority". Among them, y 0 / 1 represents the output of the prediction module based on patient similarity; K indiv represents the number of potential samples; f distance Represents the similarity function, used to calculate K indiv The feature vector of the sample and the target feature vector r target The similarity between them; y0 / y1 represents the optional output of the sample.

8. A diabetes collaborative prediction and diagnosis model and method based on deep neural network according to claim 1, characterized in that: Sensitivity, specificity and accuracy are used as evaluation indicators of the model prediction performance, and the calculation methods of each indicator are shown in formulas (12)-(14). Among them, TP is the number of correctly predicted positive samples; TN is the number of correctly predicted negative samples; FP is the number of incorrectly predicted positive samples; FN is the number of incorrectly predicted negative samples; in addition, ROC-AUC is also a comparative evaluation indicator, which is calculated by sklearn's roc_auc_score; in order to reduce the error, all experiments are repeated multiple times, and then the average of their prediction results is taken as the experimental result.

Images