A risk prediction model training method, a disease risk prediction model and method

Abstract

The application provides a risk prediction model training method, which comprises the following steps: S1, acquiring risk values of a sample population suffering from a disease, wherein the risk values are obtained by detecting multiple physical indexes of the sample population; S2, acquiring fundus images of each person in the sample population; S3, labeling the fundus images of each person in the sample population by taking the risk value of each person suffering from the disease as a label to form samples and to constitute a training data set; and S4, training a neural network by using the training data set obtained in step S3 until convergence to obtain a risk prediction model. Through deep learning, the application can directly predict the disease risk of a person by using the fundus image of the person, is safe and non-invasive, can be remotely operated, does not need to collect blood in a way of invasiveness, skips the detection of physical indexes, saves waiting time, and makes the prediction faster.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to the application of artificial intelligence in the medical field, and more particularly to a risk prediction model training method, a disease risk prediction model and method. BACKGROUND

[0002] Risk models are commonly used in the medical field to predict the risk of diseases, such as cardiovascular disease, diabetes, dementia, etc. Existing risk models often use the Cox function or other exponential functions to calculate the risk, and the input indicators of these functions are generally one or more of age, gender, systolic blood pressure, diastolic blood pressure, family history of diabetes, high-density lipoprotein, triglycerides, BMI, waist circumference, fasting blood glucose, two-hour postprandial blood glucose (or using the oral glucose tolerance test (OGTT)), blood leukocyte count, years of education, active participation in physical activity, carrying the APOE ε4 gene, etc. As can be seen, many key indicators need to be obtained through invasive blood tests, which have a waiting period and are invasive to the body, and are not friendly to some people who cannot undergo blood tests, especially for remote prediction. SUMMARY

[0003] Therefore, the purpose of the present application is to overcome the defects of the prior art, to achieve non-invasive disease prediction by obtaining human fundus images, and to provide a risk prediction method based on neural networks, a risk prediction model, and a risk prediction model training method.

[0004] According to a first aspect of the present application, a risk prediction model training method is provided, the method comprising: S1, obtaining the risk value of a sample population suffering from a disease, wherein the risk value is calculated by detecting a plurality of physical indicators of the sample population; S2, obtaining the fundus image of each person in the sample population; S3, labeling the fundus image of each person in the sample population with the risk value of the disease of each person as a label to form a sample and to form a training data set; S4, training a neural network with the training data set obtained in step S3 until convergence to obtain a risk prediction model. Wherein the disease is one or more of cardiovascular disease, diabetes, and dementia.

[0005] Preferably, in some embodiments of the present application, in the step S1, the risk value of each person suffering from the disease is calculated according to the following formula:

[0006] Risk' = ln(Risk)

[0007] Wherein, Risk is the original risk value, which is obtained by calculating the physical indicators of the person through the cox function or other exponential functions.

[0008] In some embodiments of the present application, in the step S4, the neural network is trained by using a regression loss design and a gradient descent algorithm, and the training is ended when the training loss is not decreased compared with the minimum loss in all previous training for three times. Preferably, the loss is calculated by using the MSE or MAE loss function.

[0009] According to a second aspect of the present application, a disease risk prediction model is provided, which is obtained by training the neural network by using the method according to the first aspect of the present application. Wherein, the disease is one or more of cardiovascular disease, diabetes, dementia.

[0010] According to a third aspect of the present application, a disease risk prediction method is provided, which comprises: obtaining the fundus image of a person to be predicted; and obtaining the disease risk value based on the fundus image by using the disease risk prediction model trained by using the method according to the first aspect of the present application. Wherein, the disease is one or more of cardiovascular disease, diabetes, dementia.

[0011] Compared with the prior art, the present application has the following advantages: the present application can directly predict the disease risk of a person by using the fundus image of the person through deep learning, which is non-invasive and safe, and can be remotely operated, without the need for invasive blood collection and other physical indicator detection, saving waiting time and making the prediction faster; at the same time, the risk value of the population obtained by using the cox function or other exponential functions is reformed to form a data set to train the neural network risk model, which can converge faster. BRIEF DESCRIPTION OF DRAWINGS

[0012] The embodiments of the present application will be further described below with reference to the accompanying drawings, in which:

[0013] Figure 1 The risk prediction model training flowchart according to the embodiments of the present application;

[0014] Figure 2 The population distribution of dementia risk obtained by using the cox function or exponential function according to the embodiments of the present application;

[0015] Figure 3 The population distribution of dementia risk after the dementia risk value obtained by using the cox function or exponential function is reformed according to the embodiments of the present application. DETAILED DESCRIPTION

[0016] For the purposes of the present application, the technical solutions and advantages will be clearer. The present application will be further described in detail below with specific examples. It should be understood that the specific examples described herein are only used to explain the present application and not to limit the present application.

[0017] According to one embodiment of the present application, a risk prediction model training method is provided, which is used to predict the risk of a disease according to an eye fundus image, such as diabetic retinopathy. Figure 1 As shown, the method comprises steps S1, S2, S3 and S4. Each step will be described in detail below.

[0018] In step S1, the risk value of the sample population suffering from the disease is obtained, wherein the risk value is calculated by detecting a plurality of physical indicators of the sample population. As described in the background art, the physical indicators of a person are obtained by invasive blood detection in the prior art, and the risk value of the related disease is obtained by a cox function or an exponential function.

[0019] Wherein, the classical risk model using cox function is calculated as follows:

[0020]

[0021] Wherein β is the weighted value of each indicator, X is the indicator value, is the mean value of each indicator used for model establishment. According to one example of the present application, it is assumed that the classical risk model has two inputs: age (ln(age) is commonly used to represent the age indicator) and gender (1 and 0 represent male and female respectively), and the corresponding weights β1 and β2 are 24.87 and 0.36 respectively, and the average age of the population used for modeling is and the gender is 39.5 and 0.49. Then for the input and the risk can be written as:

[0022]

[0023] Wherein S0(t) is a constant value, which is 0.9707 here, and the output Risk value is generally between 0 and 1.

[0024] The calculation of the risk value of the logistic regression model using the exponential function is as follows:

[0025]

[0026] The calculation method is similar to cox, β is the weighted value of each indicator, X is the indicator value, and β0 is a constant value, generally the intersection value of the function curve and the coordinate.

[0027] In step S2, the eye fundus image of each person in the sample population is obtained.

[0028] In step S3, the risk value of each person in the sample population suffering from the disease is labeled as a label to form a sample and constitute a training data set.

[0029] In step S4, the training data set obtained in step S3 is used to train the neural network to convergence to obtain a risk prediction model.

[0030] As can be seen from the description of step S1 above, the risk value calculated by the Cox function corresponding to the above risk model and the logistic regression model can be used as a label for model training in the framework of deep learning. The risk value expressed by the function is predicted and used as the label of the fundus image for the training of the neural network. The risk model obtained can directly predict the risk value through the fundus image. For example, the cardiovascular risk value is used as a label to label the fundus image, and then the labeled fundus image is used as a data set to train the neural network, and the obtained model can predict the cardiovascular risk of the corresponding patient through the fundus image. The model obtained by the above embodiment can directly predict the risk of a person suffering from a disease (such as cardiovascular disease, diabetes, cerebral palsy, etc.) through the fundus image of the person, without the need for invasive blood collection, which is safe and non-invasive, and can be remotely operated.

[0031] Further, the inventors have found that the risk value samples calculated by the Cox function corresponding to the above risk model and the logistic regression model have the following two problems: 1. Highly uneven distribution of numerical values; 2. Large difference in order of magnitude between high risk and low risk.

[0032] For the above two problems, the inventors realize that if the neural network is directly trained with the risk value calculated by the Cox function corresponding to the above risk model and the logistic regression model, the model convergence speed is slow, and the accuracy of the trained model is not high. Therefore, preferably, the present application proposes to modify the risk value samples calculated by the Cox function corresponding to the existing risk model and the logistic regression model, so that they are more conducive to deep learning while ensuring equivalence and reasonable distribution. The equivalence refers to maintaining the correspondence of the original sample risk ordering, and the reasonable distribution refers to the modified sample value distribution being conducive to feature learning using a deep learning scheme.

[0033] The risk value samples calculated by the Cox function corresponding to the existing risk model and the logistic regression model have the characteristic that the distribution of numerical values is highly uneven, and there is a significant difference in order of magnitude between the numerical values. The present application modifies the risk value to obtain a uniform risk value distribution. According to one embodiment of the present application, the risk value samples calculated by the Cox function corresponding to the existing risk model and the logistic regression model are modified, and the original risk value is modified as follows:

[0034] Risk' = ln(Risk)

[0035] The formula is the logarithm of the Risk value with base e, where Risk' is the modified risk value, and Risk is the original risk value. Using the improved risk value to train the model makes the overall distribution of the risk more uniform and the value more linear, and the maximum and minimum values generally do not differ by more than an order of magnitude.

[0036] According to an example of the present application, a risk distribution diagram of dementia based on a logistic regression statistic is shown as Figure 2 , where most of the values are concentrated between 0 and 0.01, and the discrimination of these samples is poor. The dementia risk regression based on the original distribution has poor discrimination of risk, and if there is a high-risk sample, such as a risk of 0.8, the loss of a sample using MSE (mean square error) can be more than 6000 times that of an ordinary sample, which causes an imbalance in the contribution of data to the loss in the model training process of deep learning. By modifying the original risk value as described above, as shown in Figure 3 , the overall distribution of the modified risk value is more uniform, and the value is more linear, and the maximum and minimum values generally do not differ by more than an order of magnitude. When using the risk value as a label for deep learning, the label can be more conducive to model convergence and data balance in model training, and using this method can directly improve the performance of the determination coefficient of the prediction target of the model. Wherein, using the modified risk value as the label of the fundus image for neural network training, a regression loss design can be used, for example, using MSE and MAE. The corresponding loss function can be set as Mean Square Error (MSE), Mean Absolute Error and Mean Absolute Percentage Error, which is used to quantify the risk value index calculated by the above existing risk model based on the Cox function and the logistic regression model:

[0037] The MSE formula is as follows:

[0038]

[0039] Where Y i is the true measurement value, is the model prediction index, and n is the number of samples.

[0040] Or use MAE, whose formula is:

[0041]

[0042] wherein Y i is the real measurement value, is the model predicted index, n is the number of samples. On the risk value without modification, MAE and MSE are difficult to use, because MAE and MSE are more biased in the numerical distribution of the labels generated by the Cox and logistic regression formula. After the logarithmic conversion of the e-based risk value, it can better adapt to the training requirements of deep learning.

[0043] According to one embodiment of the present application, the training of the risk prediction model by MAE includes the following steps:

[0044] T1, obtaining a series of fundus images of a sample population and the real risk values of the sample population corresponding to the calculation by the Cox index function, taking the logarithm of the real risk value of each sample, and taking the logarithm value as the label of the corresponding fundus image, for example, the label of the i-th fundus image is given as ln(Y i ), wherein Y i is the real risk value of the i-th sample calculated by the Cox model, and all the labeled fundus images form a training data set;

[0045] T2, inputting each image in the training data set into the neural network in turn for training to obtain a predicted risk value, for example, inputting the i-th labeled fundus image into the neural network and obtaining the predicted value of the model

[0046] T3, MAE loss function calculates the loss:

[0047]

[0048] T4, using the Adam optimizer to update the model parameters based on the loss

[0049] T5, repeating steps T2-T4, and performing multiple rounds of iterative training on the neural network until convergence.

[0050] Inputting any fundus image into the trained risk prediction model can obtain the corresponding output fitted risk, for example, inputting the k-th fundus image into the trained risk prediction model can directly obtain the output of the risk This output fits ln(Y k , by calculating the Cox function risk value corresponding to the k-th fundus image can be obtained without clinical data calculation.

[0051] The following experimental data is used to illustrate the effects achieved by the present application. The present application uses the clinical data in the following literature as a verification data set to verify the present application:

[0052] Wu YF, Liu XQ, Li X, Li Y, Zhao LC, Chen Z, Li YH, Rao XX, Zhou BF, Detrano R, USA-PRC and MUCA C. Prediction of 10-year risk of fatal and nonfatal ischemic cardiovascular diseases in Chinese adults. Circulation. 2006; 114: 2217-2225.

[0053] In the document, the clinical data information is based on the indicators of systolic blood pressure, total cholesterol, diabetes history, age, gender, BMI, and smoking history collected in the clinic, and the validation data set information based on the clinical data is shown in Table 1:

[0054] Table 1

[0055]

[0056] The clinical trial in the document uses a cardio-cerebrovascular risk model based on a cox function to predict the cardio-cerebrovascular risk of the population, and the theoretical risk value based on the document can be calculated based on the indicators of systolic blood pressure, total cholesterol, diabetes history, age, gender, BMI, and smoking history collected in the clinic. Among them, the top 5% of the theoretical risk value is positive, and the rest is negative. A binary classification test is performed on the validation data set, and the effectiveness of the model of the application is verified by comparing the AUC values of the neural network risk models trained respectively by directly using the theoretical risk value and the modified risk value. It is found through experiments that the AUC value of the neural network risk model trained by directly using the theoretical risk value of the document on the validation data set is 0.771, and the AUC value of the neural network risk model trained by using the modified risk value on the validation data set is 0.859. For ordinary skilled persons in the art, if the AUC value of a neural network risk model is greater than 0.5, it means that the neural network risk model trained based on the theoretical risk value corresponding to the clinical data is feasible. Moreover, the greater the AUC value, the better the effect. As can be seen, the neural network risk model trained by modifying the risk value in the application has better effect.

[0057] By using the method described in the above embodiment, a disease risk prediction model with pertinence can be obtained, and when it is necessary to predict the disease risk of a patient, the fundus image of the patient can be directly collected, and the disease risk of the patient can be obtained based on the fundus image by using the corresponding risk prediction model. According to one embodiment of the present application, a disease risk prediction method is provided, and the risk prediction model of a cardiovascular disease is obtained by using the above embodiment, and when it is necessary to predict the cardiovascular disease risk of a patient, the fundus image of the patient is collected, and the disease risk value of the patient is obtained based on the fundus image by using the above risk prediction model.

[0058] The present application can directly predict the disease risk (including but not limited to cardiovascular disease, diabetes, cerebral palsy, etc.) of a person by using the fundus image of the person through deep learning, is non-invasive and safe, and can be remotely operated without invasive blood collection; meanwhile, the neural network risk model trained by using the data set formed by modifying the risk value of the population predicted by using the Cox function or other exponential functions can converge faster, and the trained neural network risk model can be used for prediction, and since the detection of the physical indicators is skipped, the waiting time is saved, and the prediction speed is faster.

[0059] It should be noted that although the above describes the steps in sequence, it does not mean that the steps must be performed in the above sequence, and in fact, some of the steps can be performed concurrently or even in a changed sequence, as long as the required functions can be achieved.

[0060] The present application can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions embodied therewith to implement various aspects of the present application.

[0061] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch cards or punched tape, and any suitable combination of the foregoing. A non-transitory, computer-readable storage medium does not include a signal.

[0062] While various embodiments of the present invention have been described above, the above descriptions are intended to be illustrative, non-exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or technological improvements in the marketplace, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A risk prediction model training method, characterized in that: The method comprises: S1. Obtain a disease risk value for the sample population, wherein the risk value is calculated by testing multiple physical indicators of the sample population; wherein the disease risk value for each person is obtained by the following method: Risk′=ln(Risk) Among them, Risk′ is the risk value of each person suffering from the disease, and Risk is the original risk value of each person. The original risk value is obtained by calculating the person's physical indicators using the cox function or exponential function, where: When using the cox function, the original risk value of a person is calculated as follows: Among them, β is the weighted value of each indicator, X is the indicator value, is the mean of each indicator, S0(t) is a constant; When using the exponential function, the original risk value of a person is calculated as follows: Among them, β is the weighted value of each indicator, X is the indicator value, and β0 is a constant value; S2. Acquire fundus images of each person in the sample population; S3, using the risk value of each person in the sample population to develop a disease as a label to annotate their fundus images to form samples, and to form a training data set; S4. Using the training data set obtained in step S3, the neural network is trained until convergence to obtain a risk prediction model.

2. The risk prediction model training method according to claim 1, characterized in that: The disease is one or more of cardiovascular disease, diabetes, and dementia.

3. The risk prediction model training method according to claim 1, characterized in that: In step S4, the neural network is trained using a regression loss design and a gradient descent algorithm. The training is terminated when the loss of three consecutive trainings does not decrease compared to the minimum loss of all previous trainings.

4. The risk prediction model training method according to claim 3, characterized in that: In step S4, the loss is calculated using the MSE or MAE loss function.

5. A disease risk prediction model, characterized in that: The risk prediction model is obtained by training a neural network using the method described in any one of claims 1 to 4.

6. The disease risk prediction model according to claim 5, characterized in that: The disease is one or more of cardiovascular disease, diabetes, and dementia.

7. A computer-readable storage medium, characterized in that A computer program is included thereon, and the computer program can be executed by a processor to implement the steps of any one of the methods of claims 1 to 4.

8. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs, which, when executed by the one or more processors, enables the electronic device to implement the steps of the method according to any one of claims 1 to 4.

Images