Ischemic stroke prediction system based on big data

Abstract

The application discloses an ischemic stroke prediction system based on big data, which comprises a data acquisition module, a data set optimization module, an ischemic stroke prediction model construction module and an ischemic stroke prediction module. The application belongs to the field of data processing and specifically refers to an ischemic stroke prediction system based on big data. According to the scheme, samples are selected by combining active learning with weighted uncertainty, the weighted coefficient considers the category distribution of the samples, the prediction error of the Gaussian process model is combined, the uncertainty and rarity of each sample are used as the basis, and the patient data set is dynamically adjusted. The efficient selection of the patient data set is realized. A dynamic weight coefficient is used, the gradient density of the sample is automatically weighted, higher weight is given to the high-risk sample patients of ischemic stroke, a risk stratification loss is constructed, different risk levels of patients are accurately classified, and the accurate prediction of ischemic stroke is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of data processing, in particular to an ischemic stroke prediction system based on big data. BACKGROUND

[0002] The ischemic stroke prediction system is mainly based on a large amount of medical research and clinical data, and finds out the risk factors closely related to the occurrence of ischemic stroke by means of statistical methods and machine learning algorithms. Then the risk of individual ischemic stroke is evaluated and predicted. However, the general ischemic stroke prediction system has the problems of patient type imbalance, improper handling of noise and uncertainty often existing in clinical data, and poor final ischemic stroke prediction effect. The general ischemic stroke prediction system has the problems of patient risk category imbalance, not considering the dynamic change of patient risk, and not performing fine management on patients with different risk levels, which leads to the problem of poor accuracy of ischemic stroke prediction. SUMMARY

[0003] In view of the above problems, in order to overcome the defects of the prior art, the present application provides an ischemic stroke prediction system based on big data. For the problem of patient type imbalance in the general ischemic stroke prediction system, improper handling of noise and uncertainty often existing in clinical data, and poor final ischemic stroke prediction effect, the present application selects samples by combining active learning with weighted uncertainty, considers the class distribution of samples by weighted coefficient, combines the prediction error of Gaussian process model, and dynamically adjusts the patient data set based on the uncertainty and rarity of each sample. The efficient selection of patient data set is realized, and the final ischemic stroke prediction effect is improved. For the problem of patient risk category imbalance in the general ischemic stroke prediction system, not considering the dynamic change of patient risk, and not performing fine management on patients with different risk levels, which leads to the problem of poor accuracy of ischemic stroke prediction, the present application adopts dynamic weight coefficient, automatically weights the sample according to the gradient density, and gives higher weight to high-risk sample patients of ischemic stroke. The risk stratification loss is constructed to accurately classify patients with different risk levels, and the accurate prediction of ischemic stroke is realized.

[0004] The technical scheme adopted by the present application is as follows: the ischemic stroke prediction system based on big data provided by the present application comprises a data acquisition module, a data set optimization module, an ischemic stroke prediction model construction module and an ischemic stroke prediction module.

[0005] The data acquisition module acquires historical patient medical data, and obtains an original data set through feature engineering processing.

[0006] The data set optimization module optimizes the data set based on weighted uncertainty by using active learning technology.

[0007] The ischemic stroke prediction model construction module is based on a multilayer perceptron and designs dynamic weight coefficients and risk stratification loss functions to construct the ischemic stroke prediction model.

[0008] The ischemic stroke prediction module uses real-time patient medical data to predict ischemic stroke based on the established ischemic stroke prediction model.

[0009] Furthermore, in the data acquisition module, the historical patient medical data includes clinical signs, laboratory test indicators, lifestyle, and ischemic stroke status; the ischemic stroke status is used as a data label; the ischemic stroke status includes no ischemic stroke and ischemic stroke has occurred; the original dataset is obtained through feature engineering processing.

[0010] Furthermore, the dataset optimization module specifically includes the following:

[0011] Initialize the unit; randomly select nl samples from the original dataset as the initial training set; the Gaussian process model used for initial training is represented as: Assume the noise follows a set order. The distribution of the true labels is represented as follows: ;in, These are the input features of each sample in the initial training set; These are the corresponding real tags; Let the implicit function corresponding to the sample follow a mean of . A Gaussian process with covariance function K; The noise is assumed to have a mean of 0 and a variance of . The normal distribution; Given input features Below, real tags The probability distribution of ; I is the identity matrix;

[0012] Actively learn to select units; predict the remaining dataset, and calculate the uncertainty of each sample; represented as: ; ; Calculate the rarity , is represented as: Constructing weighted uncertainty , is represented as: ;in, It is the output predicted by the Gaussian process model of the remaining data; It is the covariance vector between the new sample and the initial training sample; T is the matrix transpose; It is uncertain; It is the number of samples in the category corresponding to the i-th sample; is the total number of samples in the dataset;

[0013] a data set construction unit; selecting the first z samples with lower weighted uncertainty from small to large according to the weighted uncertainty; repeating the active learning selection step until the maximum iteration number is reached; obtaining the final dataset.

[0014] Further, the ischemic stroke prediction model construction module is based on a multi-layer perception model and a final dataset to implement the construction of an ischemic stroke prediction model; specifically including the following contents:

[0015] a feature extraction unit; using a multi-layer perception to perform a nonlinear transformation on the input features; for input data X, after the first linear transformation, it is represented as: ; the second full connection transformation further linearly transforms the output of the first layer and applies ReLU activation again, represented as: ; residual connection is introduced; when x and dimension match, the output is represented as: ; when x and dimension do not match, the output is represented as: ; wherein, and are the outputs of the first layer transformation and the second layer transformation, respectively; and are the weight matrix and bias of the first layer; and are the weight matrix and bias of the second layer; ReLU(·) is the ReLU activation function; y is the feature extraction output; x is the input feature vector of the patient; and are the weight and bias used to adjust the input dimension matching;

[0016] a prediction output unit; using a fully connected layer to output the final prediction result, represented as: is ; wherein, is the ischemic stroke risk probability predicted by the model; sigmoid(·) is the sigmoid function;

[0017] a loss function design unit; specifically including:

[0018] constructing an initial cross-entropy loss function, represented as: ; calculating the gradient, let the predicted probability ; the gradient is represented as: ; let ; introduce the gradient density , represented as: ; ; define the dynamic weight coefficient , is expressed as: ; wherein, is an initial cross-entropy loss function; Y is a real label; is a gradient of the initial cross-entropy loss function; is a prediction error; and are prediction errors of the i-th sample and the k-th sample, respectively; n is a total number of samples; k is a sample index; is a local kernel function; u is a window width parameter;

[0019] A risk stratification loss is constructed; the data of the occurred ischemic stroke is re-labeled, and the labels include very high risk, high risk, medium risk and low risk; the probability output by the model for each patient is converted into a risk score; is expressed as: ; the risk stratification loss is expressed as: ; wherein, is a risk score of the i-th sample; is a prediction probability of the i-th sample; is a real label of the i-th sample; and are balance factors of the medium risk interval; the data of is output as a very high risk; the data of is output as a high risk; the data of is output as a medium risk; the data of is output as a high risk;

[0020] The loss function L of the final ischemic stroke prediction model is expressed as: .

[0021] Further, the ischemic stroke prediction module acquires patient medical data in real time based on the established ischemic stroke prediction model, inputs the patient medical data processed by feature engineering into the ischemic stroke prediction model, and takes the model output as an ischemic stroke prediction result.

[0022] The present application has the following beneficial effects by adopting the above-mentioned scheme:

[0023] (1) In view of the problems that the general ischemic stroke prediction system has unbalanced patient types, does not properly handle the noise and uncertainty often existing in clinical data, and thus leads to poor final ischemic stroke prediction effect, the present scheme selects samples by active learning combined with weighted uncertainty, considers the class distribution of samples in the weighted coefficient, combines the prediction error of the Gaussian process model, and dynamically adjusts the patient data set based on the uncertainty and rarity of each sample; the efficient selection of the patient data set is realized; and the final ischemic stroke prediction effect is improved.

[0024] (2) For the general ischemic stroke prediction system, there are problems of imbalance of patient risk categories, not considering the dynamic changes of patient risks, not performing fine management on patients of different risk levels, and thus leading to the problem of inaccuracy of ischemic stroke prediction. The scheme adopts a dynamic weight coefficient, automatically weights the sample gradient density, and gives higher weight to high-risk sample patients of ischemic stroke; a risk stratification loss is constructed to accurately classify patients of different risk levels; and thus the accurate prediction of ischemic stroke is realized. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 A flowchart of the ischemic stroke prediction system based on big data provided by the present application is shown.

[0026] Figure 2 A flowchart of the ischemic stroke prediction model construction module is shown.

[0027] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, and are used to explain the present application together with embodiments of the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0029] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0030] Embodiment one, refer to Figure 1 The ischemic stroke prediction system based on big data provided by the present application comprises a data acquisition module, a data set optimization module, an ischemic stroke prediction model construction module and an ischemic stroke prediction module.

[0031] The data acquisition module acquires historical patient medical data, and obtains an original data set through feature engineering processing;

[0032] The data set optimization module optimizes the data set based on weighted uncertainty by using an active learning technique;

[0033] The ischemic stroke prediction model construction module designs a dynamic weight coefficient and a risk stratification loss function based on a multi-layer perception to construct an ischemic stroke prediction model.

[0034] The ischemic stroke prediction module realizes ischemic stroke prediction for real-time patient medical data based on the established ischemic stroke prediction model.

[0035] Embodiment two, see Figure 1 Based on the above embodiment, in the data acquisition module, the historical patient medical data includes clinical signs, laboratory examination indexes, lifestyle and ischemic stroke status; the ischemic stroke status is taken as a data label; the ischemic stroke status includes no ischemic stroke and has occurred ischemic stroke; the clinical signs include blood pressure, heart rate, body weight, height, BMI, waist circumference and body fat rate; the laboratory examination indexes include blood glucose, glycosylated hemoglobin, blood lipid, inflammation index, renal function index and liver function index; the lifestyle includes smoking and drinking, exercise habit, diet structure and sleep quality; the original data set is obtained by feature engineering processing.

[0036] Embodiment three, see Figure 1 Based on the above embodiment, in the data set optimization module, the proportions of ischemic stroke patients who have not occurred ischemic stroke and have occurred ischemic stroke are different in the whole data, and the data set is optimized by using active learning; specifically including the following contents:

[0037] An initialization unit; nl sample data are randomly selected from the original data set as an initial training set; an initial Gaussian process model is used for training and is expressed as: Let the noise obey The distribution of the true label is expressed as: Wherein, is the input feature of each sample in the initial training set; is the corresponding true label; is the hidden function corresponding to the sample, which obeys a Gaussian process with a mean of and a covariance function K; is a noise assumption obeying a normal distribution with a mean of 0 and a variance of ; is the probability distribution of the true label under the given input feature ; I is the unit matrix;

[0038] Actively learn to select units; predict the remaining dataset, and calculate the uncertainty of each sample; represented as: ; ; Calculate the rarity , represented as: Constructing weighted uncertainty , represented as: ;in, It is the output predicted by the Gaussian process model of the remaining data; It is the covariance vector between the new sample and the initial training sample; T is the matrix transpose; It is uncertain; It is the number of samples in the category corresponding to the i-th sample; It is the total number of samples in the dataset;

[0039] Dataset construction unit; sort the samples according to their weighted uncertainty from smallest to largest, select the top z samples with the lowest weighted uncertainty and add them to the training set; repeat the active learning selection step until the maximum number of iterations is reached; obtain the final dataset.

[0040] By performing the above operations, this solution addresses the problems of imbalanced patient types and inadequate handling of noise and uncertainty often present in clinical data in general ischemic stroke prediction systems, which often lead to poor prediction results. This solution utilizes active learning combined with weighted uncertainty selection of samples. The weighting coefficients consider the class distribution of the samples and incorporate the prediction error of the Gaussian process model. Based on the uncertainty and rarity of each sample, the patient dataset is dynamically adjusted, achieving efficient selection of the patient dataset and thus improving the final ischemic stroke prediction effect.

[0041] Example 4, see Figure 1 and Figure 2 This embodiment, based on the above embodiment, involves a predictive model for ischemic stroke. The prediction of ischemic stroke involves not only simple clinical information but also complex nonlinear relationships. It uses a multilayer perceptron model from deep learning to capture these nonlinear features; a loss function is designed based on dynamic weight coefficients and risk stratification loss; and the ischemic stroke prediction model is constructed based on the final dataset. Specifically, it includes the following:

[0042] Feature extraction unit; uses a multilayer perceptron to perform a nonlinear transformation on the input features. For input data X, after the first linear transformation, it is represented as: The second fully connected layer performs a further linear transformation on the output of the first layer and applies ReLU activation again, as follows: Introduce residual join; when x and When performing dimension matching, the output is represented as follows: When x and When the dimensions do not match, the output is represented as: ; wherein, and are the outputs of the first layer transformation and the second layer transformation, respectively; and are the weight matrix and bias of the first layer; and are the weight matrix and bias of the second layer; ReLU(·) is the ReLU activation function; y is the feature extraction output; x is the input feature vector of the patient; and are the weight and bias for adjusting the input dimension matching;

[0043] The prediction output unit outputs the final prediction result using a fully connected layer, represented as: ; wherein, is the ischemic stroke risk probability predicted by the model; sigmoid(·) is the sigmoid function;

[0044] The loss function design unit; since the data of ischemic stroke is usually highly unbalanced, high-risk samples are usually less; in order to solve the problem of sample imbalance, a dynamic weight coefficient is designed, so that the model gives greater weight to high-risk samples during training; specifically including:

[0045] An initial cross-entropy loss function is constructed, represented as: ; the gradient is calculated, and let the predicted probability ; the gradient is represented as: ; let ; in order to measure the distribution of sample gradient in the whole sample, the gradient density is introduced, represented as: ; ; the dynamic weight coefficient is defined, represented as: ; wherein, is the initial cross-entropy loss function; Y is the real label; is the gradient of the initial cross-entropy loss function; is the prediction error; and are the prediction errors of the i-th sample and the k-th sample, respectively; n is the total number of samples; k is the sample index; is the local kernel function; u is the window width parameter;

[0046] A risk stratification loss is constructed; data on ischemic stroke that has occurred are relabeled with labels including very high risk, high risk, intermediate risk, and low risk; since the risk of ischemic stroke is continuously distributed, in order to accurately manage different risk levels, the probability output by each patient through the model is converted into a risk score; represented as: Risk stratification loss Represented as: ;in, It is the risk score of the i-th sample; It is the predicted probability of the i-th sample; It is the true label of the i-th sample; and It is a balancing factor for the intermediate risk range; The data label output is extremely high risk; The data is labeled as high risk; The data label output is medium risk; The data is labeled as high risk.

[0047] The loss function L of the final ischemic stroke prediction model is expressed as: .

[0048] By performing the above operations, this solution addresses the issue of inaccurate prediction of ischemic stroke in general ischemic stroke prediction systems. This is due to imbalances in patient risk categories, failure to consider dynamic changes in patient risk, and lack of refined management for patients at different risk levels. Instead, this solution employs dynamic weighting coefficients, automatically assigning weights based on the gradient density of samples, giving higher weights to high-risk patients with ischemic stroke. A risk stratification loss function is constructed to accurately classify patients at different risk levels, thereby achieving precise prediction of ischemic stroke.

[0049] Example 5, see Figure 1 This embodiment is based on the above embodiment. The ischemic stroke prediction module collects the patient's medical data in real time based on the established ischemic stroke prediction model, and inputs it into the ischemic stroke prediction model after feature engineering processing. The model output is used as the ischemic stroke prediction result. If the model output indicates that an ischemic stroke has occurred and is not low risk, an early warning is issued to relevant personnel.

[0050] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and implementations, it is to be understood that the terminology used is for the purpose of descriptive clarity and that it is intended to be limited only by the claims.

[0051] While the embodiments of the application have been shown and described, it is to be understood that the embodiments described are only by way of example and that modifications, changes, substitutions and variations can be made by those skilled in the art without departing from the spirit and scope of the application.

[0052] The above description of the present application has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the application to the precise form described. Many modifications, variations, alterations, and equivalents can be suggested to one skilled in the art in light of this disclosure and it is intended that the protected application embrace all such alternatives, modifications, and variations.

Claims

1. A big data based ischemic stroke prediction system characterized in that: The system comprises a data collection module, a data set optimization module, an ischemic stroke prediction model construction module and an ischemic stroke prediction module. The data collection module collects historical patient medical data and obtains an original data set through feature engineering processing. The data set optimization module optimizes the data set based on weighted uncertainty using active learning technology. The ischemic stroke prediction model construction module designs a dynamic weight coefficient and a risk stratification loss function based on a multilayer perception machine to construct an ischemic stroke prediction model. The ischemic stroke prediction module realizes ischemic stroke prediction for real-time patient medical data based on the established ischemic stroke prediction model. The data set optimization module specifically comprises the following contents: An initialization unit; Randomly select nl samples from the original dataset as the initial training set; the Gaussian process model used for initial training is represented as: Assume the noise follows a set order. The distribution of the true labels is represented as follows: ;in, These are the input features of each sample in the initial training set; These are the corresponding real tags; Let the implicit function corresponding to the sample follow a mean of . A Gaussian process with covariance function K; The noise is assumed to have a mean of 0 and a variance of . The normal distribution; Given input features Below, real tags The probability distribution of ; I is the identity matrix; active learning selection unit; predicting the remaining dataset, calculating the uncertainty of each sample; denoted as: ; ; calculating the rarity , denoted as: ; constructing the weighted uncertainty , denoted as: ; wherein, is the Gaussian process model prediction output of the remaining data; is the covariance vector between the new sample and the initial training sample; T is the matrix transpose; is the uncertainty; is the number of samples corresponding to the category of the i-th sample; is the total number of samples in the dataset; A data set construction unit; the z samples with lower weighted uncertainty are selected to join the training set according to the ascending order of the weighted uncertainty; the active learning selection step is repeated until the maximum iteration number is reached; and a final data set is obtained. The ischemic stroke prediction model construction module realizes the construction of the ischemic stroke prediction model based on the multilayer perception machine model and the final data set; specifically comprising the following contents: Feature extraction unit; use multi-layer perception to carry out nonlinear transformation to input feature, for input data X, after first layer linear transformation, it is expressed as: ; Second layer full connection transformation carries out further linear transformation to the output of first layer, and applies ReLU activation again, and is expressed as: ; Introduce residual connection; when x and Dimension match, the output is expressed as: ; When x and Dimension does not match, the output is expressed as: ; Wherein, And First layer transformation and second layer transformation output respectively; And First layer weight matrix and bias; And Second layer weight matrix and bias; ReLU (·) is ReLU activation function; y is feature extraction output; x is the input feature vector of patient; And Weight and bias for adjusting input dimension matching; The prediction output unit outputs a final prediction result using a fully connected layer, represented as: ; wherein, is the ischemic stroke risk probability predicted by the model; sigmoid(·) is a sigmoid function. A loss function design unit; specifically comprising: The initial cross-entropy loss function is constructed and denoted as: The gradient is calculated, and let the prediction probability The gradient is denoted as: Let The gradient density is introduced and denoted as: ; The dynamic weight coefficient is defined and denoted as: wherein, is the initial cross-entropy loss function; Y is the true label; is the gradient of the initial cross-entropy loss function; is the prediction error; and are the prediction errors of the i-th sample and the k-th sample, respectively; n is the total number of samples; k is the sample index; is the local kernel function; u is the window width parameter; Constructing risk stratification loss; re-labeling the data of the occurred ischemic stroke, the labels include very high risk, high risk, medium risk and low risk; converting the probability output by the model for each patient into a risk score; denoted as: ; risk stratification loss denoted as: ; wherein, is the risk score of the ith sample; is the predicted probability of the ith sample; is the true label of the ith sample; and is the balance factor of the medium risk interval; the data label of is output as very high risk; the data label of is output as high risk; the data label of is output as medium risk; the data label of is output as low risk; The loss function L of the final ischemic stroke prediction model is represented as: .

2. The big data based ischemic stroke prediction system as claimed in claim 1, wherein: In the data collection module, the historical patient medical data comprises clinical signs, laboratory examination indexes, lifestyle and ischemic stroke status; the ischemic stroke status is taken as a data label; the ischemic stroke status comprises no ischemic stroke and ischemic stroke; The original data set is obtained through feature engineering processing. 3.The big data based ischemic stroke prediction system according to claim 2, wherein: The ischemic stroke prediction module collects patient medical data in real time based on the established ischemic stroke prediction model, inputs the processed data into the ischemic stroke prediction model, and takes the model output as the ischemic stroke prediction result.

Images