Hierarchical diabetes risk assessment method based on big data and deep learning

Abstract

The invention discloses a hierarchical diabetes risk assessment method based on big data and deep learning, and belongs to the technical field of big data, and the method comprises the following steps: combining a one-dimensional convolutional neural network and an attention mechanism in deep learning, carrying out the hierarchical analysis of physical examination data in different scenes, and achieving the accuracy of diabetes risk assessment; a normal distribution mapping model is provided and is used for remapping a prediction result to standard normal distribution, so that risk grade division is realized; single sample feature importance sorting is carried out based on a data perturbation method, and personalized diabetes risk assessment is carried out for individual health conditions. A layered and scene-divided diabetes risk assessment system is constructed based on large-scale physical examination data in combination with artificial intelligence and a deep learning technology, and high-precision and personalized diabetes risk assessment can be provided according to different application scenes and individual health data, so that early warning and accurate health management are realized.

Description

Technical Field

[0001] The present invention belongs to the field of big data, and particularly relates to a hierarchical diabetes risk assessment method based on big data and deep learning. Background Art

[0002] As one of the four major chronic diseases in China, the problem of the increasing prevalence of diabetes in recent years has not been significantly improved. It is estimated that by 2045, the number of diabetic patients in China will increase to 174 million, which means that one out of every nine Chinese people will be a diabetic patient. Early diagnosis, early treatment, and timely adjustment of lifestyle are the most effective methods for preventing and controlling diabetes and its complications. However, due to the limitations of economic conditions and the understanding of their own health status, only half of the diabetic patients are currently diagnosed with diabetes, which poses a challenge to the early diagnosis and treatment of diabetes.

[0003] How can accurate and effective early diagnosis of diabetes be achieved? Domestic health care big data has become an important entry point. Health care big data is an important basic strategic resource in China, specifically including health big data centered on physical examination data, medical big data centered on electronic medical records, examinations and tests, prescriptions and doctor's orders, rapidly growing omics big data, and public health big data based on the detection of regional population diseases and health conditions. In recent years, physical examinations have received increasing attention, and physical examination institutions and data platforms have also developed rapidly. The functions of physical examination institutions have gradually shifted from simple health examinations to health management. For the potential diabetic population in China, health examinations have the advantages of convenience, universality, and comprehensive examination items. Therefore, physical examination big data shows an exponential growth trend. However, physical examination indicators, as an important tool for early detection of disease risk factors, have not been fully utilized to establish domestic chronic disease risk assessment tools. With the rapid development of artificial intelligence, deep learning algorithms can play performance advantages on big data, and machine learning can help understand and quantify the relationships between various features and diseases. By combining artificial intelligence to fully utilize the risk factors in physical examination data to construct a disease risk assessment tool, not only can accurate, convenient, and popular risk warnings be provided for potential patients, but it can also be directly applied to the physical examination population and health management centers. In addition, it can also be used as a tool for medical insurance systems and other entities to evaluate personal disease risks, which has great practical significance and broad application prospects and is of great significance for carrying out personalized health management and early disease prevention.

[0004] At present, a number of studies at home and abroad have constructed diabetes risk models. However, due to different living habits and environments in different countries, each country needs to build a diabetes risk scoring model suitable for its own population according to the actual situation of its people. China has also constructed some diabetes risk assessment tools based on the samples and physical characteristics of its own people. However, the application scenarios of existing risk prediction or risk scoring models are not clear enough.

[0005] Several major problems existing in the current diabetes risk assessment models: First, the accuracy is relatively low. Especially when using non-invasive physical examination data, the prediction effect of existing models is not sufficient to meet the needs of large-scale applications; Second, the application scenarios are limited. Many models are only applicable to specific data sets or specific populations and cannot be widely applied in different actual application scenarios; Third, there is a lack of personalization. Existing tools lack medical advice tailored according to individual differences and cannot meet the needs of personalized health management; Fourth, there is insufficient data integration and processing. Existing models fail to effectively integrate data sources from multiple aspects (such as living habits, past medical history, traditional Chinese medicine constitution, etc.), resulting in incomplete and inaccurate assessments.

[0006] Therefore, at the current stage, it is necessary to design a hierarchical diabetes risk assessment method, system and storage medium based on big data and deep learning to solve the above problems. Summary of the Invention

[0007] The purpose of the present invention is to provide a hierarchical diabetes risk assessment method, system and storage medium based on big data and deep learning, which are used to solve the technical problems existing in the above-mentioned prior art. By combining artificial intelligence and deep learning technologies based on large-scale physical examination data, a hierarchical and scenario-based diabetes risk assessment system is constructed, which can provide high-precision and personalized diabetes risk assessments according to different application scenarios and individual health data, and thus achieve early warning and precise health management.

[0008] To achieve the above purpose, the technical solution of the present invention is:

[0009] The hierarchical diabetes risk assessment method based on big data and deep learning includes the following steps:

[0010] S1. Combine the one-dimensional convolutional neural network and attention mechanism in deep learning, and achieve the accuracy of diabetes risk assessment by performing hierarchical analysis on physical examination data in different scenarios;

[0011] S2. Propose a normal distribution mapping model for remapping the prediction results onto the standard normal distribution, so as to achieve risk level classification;

[0012] S3. Adopt a data perturbation method for single-sample feature importance ranking for personalized diabetes risk assessment of individual health conditions.

[0013] Furthermore, in step S1,

[0014] A feature extraction framework based on a one-dimensional convolutional neural network is proposed. All convolutional layers use a kernel size of 3, and a ReLU activation function is added after the last three convolutional operations. The extracted features are flattened into one-dimensional features and then passed to a fully connected neural network for final classification. An attention mechanism is introduced to assign specific weights to the samples.

[0015] First, for the input data Four one-dimensional convolutional operations are performed, where L is the length of the input sequence and C is the number of channels of the input signal. The convolutional output of each layer is calculated by the following formula:

[0016] X (l+1) = ReLU(W (l) * X (l) + b (l) )

[0017] where X (l) is the input of the l-th layer, is the convolutional kernel of the l-th layer, b (l) is the bias term, k is the size of the convolutional kernel, the symbol * represents the convolutional operation, and ReLU(·) is the ReLU activation function. Through four one-dimensional convolutions, the extracted features are denoted as X (4) ;

[0018] The extracted features X (4) are first flattened into one-dimensional features and then input into a fully connected neural network, and an attention mechanism is introduced. First, the features are mapped to a one-dimensional space through a fully connected layer:

[0019] a = Sigmoid(W a X (4) + b a )

[0020] where W a is the weight matrix of the attention layer, and b a is the bias term;

[0021] Through the mapping, the weight a of the sample ∈ [0, 1] is obtained and assigned to the input feature X':

[0022] X'' = a ⊙ X'

[0023] Then, after being processed by the fully connected layer, the Dropout and ReLU activation functions are added during this process. Finally, the Sigmoid function is used in the output layer to map the result to between 0 and 1 to obtain the final classification result: y = σ(W o X' + b o)。

[0024] Further, in step S2,

[0025] First, generate a set of data that follows the standard normal distribution and has the same number of samples as the training set as the target variable Y; let the number of training set samples be N, and draw N samples from the standard normal distribution to form Y, that is

[0026]

[0027] Next, sort Y in ascending order to obtain the sorted target variable Y sorted ; at the same time, sort the predicted values X = {x1, x2,..., x N} of the training set in ascending order to obtain X sorted ;

[0028] Next, use the sorted X sorted and Y sorted data to train a fully connected neural network; the input is the sorted predicted value X sorted , and the output is the sorted target value Y sorted ; use the mean squared error loss function MSELoss) to optimize the model parameters, and the loss function is defined as follows:

[0029]

[0030] where f(x i ) represents the predicted output of the neural network for the input x i , and y i is the corresponding target value;

[0031] Based on the model output results of the normal distribution mapping, determine the thresholds of five risk levels through quantiles; select the 20%, 40%, 60%, and 80% quantiles as the thresholds for dividing different risk levels; divide the samples into five levels: extremely low risk, low risk, medium risk, high risk, and extremely high risk;

[0032] Let the results of the model prediction output be X = {x1, x2,..., x N}, where N is the number of samples; first sort the predicted values and calculate the 20%, 40%, 60%, and 80% quantile values, which are denoted as P 20 , P 40 , P 60 , P 80 ; based on these quantiles, the risk level division rules are as follows: extremely low risk: when x < P 20 , that is, the predicted value of the sample is less than the 20% quantile, low risk: when P 20 ≤ x < P40 , that is, the predicted value of the sample is between the 20%-40% quantiles, Medium risk: When P 40 ≤x<P 60 , that is, the predicted value of the sample is between the 40%-60% quantiles, High risk: When P 60 ≤x<P 80 , that is, the predicted value of the sample is between the 60%-80% quantiles, Extremely high risk: When x≥P 80 , that is, the predicted value of the sample is greater than the 80% quantile; The division of risk levels is expressed as:

[0033]

[0034] Furthermore, in step S3,

[0035] Use a single-sample-based feature contribution ranking method to evaluate the relative contribution of each feature to the model prediction result; Assume the predicted value of the given sample sample1 is X; By randomly perturbing each feature, calculate the change in the predicted value after perturbation to estimate the contribution value of the feature to the prediction result;

[0036] For each feature f in sample 1 i , perform n random perturbation experiments; Each time of perturbation, change the feature value f i to a random value where:

[0037]

[0038] where, min i and max i are the minimum and maximum values of the feature in the entire dataset respectively;

[0039] Each time after perturbing the feature f i , record the new predicted value x i，j output by the model, where j = 1, 2,..., n represents the jth perturbation experiment, and then calculate the average predicted value of the feature f i

[0040]

[0041] The contribution value C i of each feature f i is calculated by the absolute value of the difference between the initial predicted value X and the average predicted value :

[0042]

[0043] Finally, according to C i ​Sort them in descending order of size to obtain the contribution value ranking of each feature.

[0044] A hierarchical diabetes risk assessment system based on big data and deep learning uses the hierarchical diabetes risk assessment method based on big data and deep learning as described above for hierarchical diabetes risk assessment.

[0045] A storage medium stores a computer program thereon, and when the computer program is run, it executes the hierarchical diabetes risk assessment method based on big data and deep learning as described above.

[0046] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0047] By building a hierarchical diabetes risk assessment system based on large-scale physical examination data, combining artificial intelligence and deep learning technologies, the present invention can provide high-precision and personalized diabetes risk assessment according to different application scenarios and individual health data, and then achieve early warning and precise health management. This technical solution provides a more scientific and practical auxiliary solution for the early diagnosis and prevention and control of diabetes. Description of the Drawings

[0048] Figure 1 It is a schematic flow chart of the steps of a hierarchical diabetes risk assessment method based on big data and deep learning. Detailed Embodiments

[0049] The hierarchical diabetes risk assessment method based on big data and deep learning includes the following steps:

[0050] S1. Combine the one-dimensional convolutional neural network and attention mechanism in deep learning, and achieve the accuracy of diabetes risk assessment through hierarchical analysis of physical examination data in different scenarios;

[0051] S2. Propose a normal distribution mapping model for remapping the prediction results to the standard normal distribution, so as to achieve risk level division;

[0052] S3. Use the data perturbation method to rank the importance of single-sample features for personalized diabetes risk assessment of individual health conditions.

[0053] (1) Data

[0054] First, we excluded data columns with more than 50% missing values and removed features that had no significant discriminatory effect in the diagnosis of coronary heart disease, such as cervical smear and self-assessment of the health status of the elderly. In addition, we cleaned the irregular items in the data, for example, corrected the non-standard suffixes in the occupation column and converted some medical terms into numerical forms so that the model could better process this information. We also converted multi-select features such as cerebrovascular diseases into one-hot encoding. After these steps, we finally obtained 569,634 healthy person data (negative samples) with 131 features and 157,453 coronary heart disease patient data (positive samples) with 131 features, and the ratio of positive to negative samples was approximately 1:3. Then, we divided the data into a training set and a test set at a ratio of 9:1, and then further divided the training set into a final training set and a validation set at a ratio of 9:1.

[0055] (2) Method

[0056] We first divided the physical examination data into five versions: the first version included resident information, routine physical examination information, blood routine information, and biochemical index information; the second version added lifestyle information and other examination results on the basis of the first version; the third version further added other disease information on the basis of the second version; the fourth version added traditional Chinese medicine constitution examination information on the basis of the third version; the fifth version was a special version, namely the non-invasive self-test version. We targeted the home test group, that is, those who did not need to go to the hospital to obtain new data and only needed old disease and routine physical information to complete the diagnosis and prediction. Our model was based on a one-dimensional convolutional architecture and was optimized by combining an attention mechanism. In addition, we also constructed a normal distribution mapping model to remap the prediction results of the integrated model onto a normal distribution. According to the remapped data, we used the data at the 20%, 40%, 60%, and 80% positions as thresholds to divide the prediction results into five risk levels: below 20% was low risk, 20%-40% was relatively low risk, 40%-60% was medium risk, 60%-80% was relatively high risk, and over 80% was high risk. Finally, we performed single-sample feature importance ranking based on the method of data perturbation.

[0057] (2.1) Data version division

[0058] The data features of the first version include resident information, routine physical examination information, blood routine information, and biochemical index information. Among them, resident information includes age and gender, routine physical examination information includes body temperature, pulse rate, heart rate, respiratory rate, height, weight, waist circumference, body mass index, average systolic blood pressure, average diastolic blood pressure, pulse pressure, waist-to-weight ratio, waist-to-height ratio, and waist-to-body mass ratio. Blood routine information includes hemoglobin, white blood cells, and platelets. Biochemical index information includes fasting blood glucose, serum alanine aminotransferase, serum aspartate aminotransferase, total bilirubin, serum creatinine, blood urea nitrogen, total cholesterol, triglycerides, serum low-density lipoprotein cholesterol, serum high-density lipoprotein cholesterol, urine protein, urine glucose, urine ketone body, and urine occult blood.

[0059] The second version adds lifestyle information and other examination information on the basis of the first version. Lifestyle information includes exercise frequency, eating habits, smoking status, drinking frequency, and whether one has been drunk in the past year. Other examination information includes electrocardiogram, other B-ultrasound, B-ultrasound, and chest X-ray.

[0060] The third version adds other disease information on the basis of the second version. Other disease information includes dorsalis pedis artery pulsation, cerebrovascular diseases, kidney diseases, heart diseases, vascular diseases, eye diseases, nervous system diseases, other system diseases, and symptoms.

[0061] The fourth version adds traditional Chinese medicine constitution examination information on the basis of the third version. Traditional Chinese medicine constitution examination information includes peaceful constitution, qi deficiency constitution, yang deficiency constitution, yin deficiency constitution, phlegm-dampness constitution, damp-heat constitution, blood stasis constitution, qi stagnation constitution, and special endowment constitution.

[0062] The fifth version is a non-invasive self-test version, which includes resident information, general physical examination information, lifestyle-related information, and other disease information. Resident information includes age, gender, ethnicity, blood type, marital status, permanent residence type, education level, whether one is a floating population, whether one is a poor person, occupation code, and household registration category. General physical examination information includes height, weight, waist circumference, body mass index, waist-to-weight ratio, waist-to-height ratio, and waist-to-body mass ratio. Lifestyle-related information includes exercise frequency, eating habits, smoking status, drinking frequency, and whether one has been drunk in the past year. Other disease information includes cerebrovascular diseases, kidney diseases, heart diseases, vascular diseases, eye diseases, nervous system diseases, other system diseases, and symptoms.

[0063] (2.2) Diabetes prediction model

[0064] In this study, we propose a feature extraction framework based on one-dimensional convolutional neural network (CNN) for diabetes diagnosis. All convolutional layers use a kernel size of 3, and the ReLU activation function is added after the last three convolutional operations. The extracted features are flattened into one-dimensional features and then passed to a fully connected neural network for final classification, which introduces an attention mechanism to assign specific weights to samples.

[0065] First, perform four one-dimensional convolutional operations on the input data , where L is the length of the input sequence and C is the number of channels of the input signal. The convolutional output of each layer is calculated by the following formula:

[0066] X (l+1) = ReLU(W (l) * X (l) + b (l) )

[0067] where X (l) is the input of the l-th layer, is the convolutional kernel of the l-th layer, b (l) is the bias term, k is the size of the convolutional kernel (k = 3 in this model), the symbol * represents the convolutional operation, and ReLU(·) is the ReLU activation function. Through four one-dimensional convolutions, the extracted features are denoted as X (4) .

[0068] The extracted features X (4) are first flattened into one-dimensional features and then input into a fully connected neural network that introduces an attention mechanism. First, map the features to a one-dimensional space through a fully connected layer:

[0069] a = Sigmoid(W a X (4) + b a )

[0070] where W a is the weight matrix of the attention layer, and b a is the bias term.

[0071] Through this mapping, we obtain the weight a ∈ [0, 1] of the sample and assign it to the input feature X':

[0072] X'' = a ⊙ X'

[0073] Then, after further processing by the fully connected layer, we add Dropout (weight 0.1) and the ReLU activation function in this process. Finally, use the Sigmoid function in the output layer to map the result to between 0 and 1 to obtain the final classification result:

[0074] y = σ(Wo X′ + b o )

[0075] (2.3) Normal distribution mapping model

[0076] First, we generated a set of data that follows the standard normal distribution and has the same number of samples as the training set as the target variable Y. Let the number of samples in the training set be N. We drew N samples from the standard normal distribution to form Y, that is

[0077]

[0078] Next, we sorted Y in ascending order to obtain the sorted target variable Y sorted . At the same time, the predicted values X = {x1, x2,..., x N} of the training set were also sorted in ascending order to obtain X sorted。

[0079] Next, we used the sorted X sorted and Y sorted data to train a fully connected neural network. Specifically, the input is the sorted predicted value X sorted , and the output is the sorted target value Y sorted . We used the mean squared error loss function (MSELoss) to optimize the model parameters. The loss function is defined as follows:

[0080]

[0081] where f(x i ) represents the predicted output of the neural network for the input x i , and y i is the corresponding target value.

[0082] (2.4) Risk division

[0083] In this study, based on the output results of the normal distribution mapping model, we determined the thresholds for five risk levels through quantiles. Specifically, we selected the 20%, 40%, 60%, and 80% quantiles as the thresholds for dividing different risk levels. Through these thresholds, the samples were divided into five levels: extremely low risk, low risk, medium risk, high risk, and extremely high risk.

[0084] Let the results of the model prediction output be X = {x1, x2,..., x N}, where N is the number of samples. To determine the thresholds for different risk levels, we first sorted the predicted values and calculated the 20%, 40%, 60%, and 80% quantile values, denoted as P 20 , P 40 , P60 , P 80 . Based on these quantiles, the rules for dividing risk levels are as follows: Extremely low risk: when x < P 20 , that is, the predicted value of the sample is less than the 20% quantile. Low risk: when P 20 ≤ x < P 40 , that is, the predicted value of the sample is between the 20% - 40% quantiles. Medium risk: when P 40 ≤ x < P 60 , that is, the predicted value of the sample is between the 40% - 60% quantiles. High risk: when P 60 ≤ x < P 80 , that is, the predicted value of the sample is between the 60% - 80% quantiles. Extremely high risk: when x ≥ P 80 , that is, the predicted value of the sample is greater than the 80% quantile. Therefore, the division of risk levels can be expressed as:

[0085]

[0086] (2.5) Single-sample feature ranking

[0087] In this study, we used a single-sample-based feature contribution ranking method to evaluate the relative contribution of each feature to the model prediction result. Assume that the predicted value of the given sample sample1 is X. By randomly perturbing each feature, the change in the predicted value after perturbation is calculated to estimate the contribution value of the feature to the prediction result.

[0088] For each feature f in sample 1 i , perform n random perturbation experiments. Each time of perturbation, change the feature value f i to a random value where:

[0089]

[0090] where, min i and max i are the minimum and maximum values of the feature in the entire dataset respectively.

[0091] Each time after perturbing the feature f i , record the new predicted value x i，j (where j = 1, 2,..., n represents the jth perturbation experiment), and then calculate the average predicted value of the feature f i

[0092]

[0093] The contribution value C of each feature f i ​​i Calculated from the absolute value of the difference between the initial predicted value X and the average predicted value as follows:

[0094]

[0095] Finally, sort in descending order according to the magnitude of C i to obtain the contribution value ranking of each feature. The larger the contribution value, the more significant the impact of the feature on the model predicted value.

Claims

1. A hierarchical diabetes risk assessment method based on big data and deep learning, characterized in that, It includes the following steps: S1. Combine the one-dimensional convolutional neural network and the attention mechanism in deep learning, and achieve the accuracy of diabetes risk assessment by performing hierarchical analysis on physical examination data in different scenarios; S2. Propose a normal distribution mapping model to remap the prediction results to the standard normal distribution, thereby realizing the risk level division; S3. Adopt a data perturbation method to rank the importance of single-sample features for personalized diabetes risk assessment of individual health conditions.

2. The hierarchical diabetes risk assessment method based on big data and deep learning according to claim 1, wherein In step S1, Propose a feature extraction framework based on the one-dimensional convolutional neural network. All convolutional layers use a kernel size of 3, and the ReLU activation function is added after the last three convolutional operations; the extracted features are flattened into one-dimensional features and then passed to the fully connected neural network for final classification, and the attention mechanism is introduced to assign specific weights to the samples; First, the input data is subjected to four layers of one-dimensional convolution operations, where L is the length of the input sequence and C is the number of channels of the input signal; the convolution output of each layer is calculated by the following formula: X (l+1) = ReLU(W (l) * X (l) + b (l) ) Among them, X (l) is the input of the l-th layer, is the convolution kernel of the l-th layer, b (l) is the bias term, k is the size of the convolution kernel, the symbol * represents the convolution operation, ReLU(·) is the ReLU activation function; through four layers of one-dimensional convolution, the extracted feature is denoted as X (4) ; Extracted feature X (4) First, it is flattened into a one-dimensional feature and then input into a fully connected neural network, introducing an attention mechanism; first, the feature is mapped to a one-dimensional space through a fully connected layer: a = Sigmoid(W a X (4) + b a ) Among them, W a is the weight matrix of the attention layer, and b a is the bias term; Through mapping, the weight a of the sample is obtained, where a ∈ [0, 1], and it is assigned to the input feature X ′ : X ′‘ = a ⊙ X ′ Then, after being processed by the fully connected layer, the Dropout and ReLU activation functions are added during this process; finally, the Sigmoid function is used in the output layer to map the result to between 0 and 1 to obtain the final classification result: y = σ(W o X ′ + b o )。 3. The hierarchical diabetes risk assessment method based on big data and deep learning according to claim 2, wherein In step S2, First, generate a set of data that follows the standard normal distribution and has the same number of samples as the training set as the target variable Y; assume the number of training set samples is N, and draw N samples from the standard normal distribution to form Y, that is Next, sort Y from smallest to largest to obtain the sorted target variable Y sorted ; At the same time, sort the predicted values X = {x1, x2,..., x N} of the training set in ascending order to obtain X sorted ; Next, use the sorted X sorted and Y sorted data to train a fully connected neural network; the input is the sorted predicted value X sorted , and the output is the sorted target value Y sorted ; use the mean squared error loss function MSELoss) to optimize the model parameters, and the loss function is defined as follows: Among them, f(x i ) represents the predicted output of the neural network for the input x i , and y i is the corresponding target value; Based on the output results of the normal distribution mapping model, determine the thresholds of five risk levels through quantiles; select the 20%, 40%, 60%, and 80% quantiles as the thresholds for dividing different risk levels; divide the samples into five levels: extremely low risk, low risk, medium risk, high risk, and extremely high risk; Let the result of the model prediction output be X = {x1, x2,..., x N}, where N is the number of samples; First, sort the predicted values and calculate the quantile values of 20%, 40%, 60%, and 80%, denoted as P 20 , P 40 , P 60 , P 80 respectively; Based on these quantiles, the risk level classification rules are as follows: Very low risk: when x < P 20 , that is, the predicted value of the sample is less than the 20% quantile, Low risk: when P 20 ≤ x < P 40 , that is, the predicted value of the sample is between the 20% - 40% quantiles, Medium risk: when P 40 ≤ x < P 60 , that is, the predicted value of the sample is between the 40% - 60% quantiles, High risk: when P 60 ≤ x < P 80 , that is, the predicted value of the sample is between the 60% - 80% quantiles, Extremely high risk: when x ≥ P 80 , that is, the predicted value of the sample is greater than the 80% quantile; The classification of the risk level is expressed as:

4. The hierarchical diabetes risk assessment method based on big data and deep learning according to claim 3, wherein In step S3, Use the feature contribution ranking method based on single samples to evaluate the relative contribution of each feature to the model prediction results; assume that the predicted value of the given sample sample1 is X; by randomly perturbing each feature, calculate the change in the predicted value after perturbation to estimate the contribution value of this feature to the prediction result; For each feature f in sample 1 i , perform n times of random perturbation experiments; each time of perturbation, change the feature value f i to a random value where: where min i and max i are the minimum and maximum values of the feature in the entire dataset, respectively; Each perturbation feature f i After that, record the new predicted value x output by the model i，j , where j = 1, 2,..., n represents the j-th perturbation experiment, and then calculate the feature f i Average predicted value of Each feature f i has a contribution value C i calculated by the absolute value of the difference between the initial prediction value X and the average prediction value as follows: Finally, sort in descending order according to the size of C i to obtain the contribution value ranking of each feature.

5. A hierarchical diabetes risk assessment system based on big data and deep learning, characterized in that, Adopt the hierarchical diabetes risk assessment method based on big data and deep learning as described in any one of claims 1-4 for hierarchical diabetes risk assessment.

6. A storage medium, characterized in that, A computer program is stored on the storage medium, and when the computer program is run, it executes the hierarchical diabetes risk assessment method based on big data and deep learning as described in any one of claims 1-4.

Images