A disease risk level prediction model and its construction method and application

Abstract

The present invention provides a disease risk level prediction model and its construction method and application. The prediction model is constructed by the following method: obtaining the physical examination report data of the physical examination personnel as a data set, dividing the people with diseases and the people without diseases, and then converting the multivariate data in the data set into application-level structured data and filling it, using a machine learning method for learning, and then expanding the positive sample training set, using the optimal indicator decreasing method for modeling, and then performing model fusion to obtain the disease risk level prediction model. The obtained disease risk level prediction model can efficiently, quickly and accurately predict the risk level of the patient's disease, providing the possibility of early prevention and treatment of the disease.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical data processing, and specifically relates to a disease risk level prediction model and a construction method and application thereof. Background Art

[0002] With the continuous progress and development of society, people's work and life pressures have increased invisibly. In addition, a series of external factors such as food safety, radiation, and environmental pollution have dramatically accelerated the incidence of hidden major diseases at a younger age. These major diseases are generally characterized by strong concealment, high incidence, and great harm. Since the early symptoms of the disease are not obvious, once it reaches the later stage, the harm is very great, which will have a significant impact on people's quality of life, work, family, and even life. Therefore, early detection and prevention of major diseases is particularly important.

[0003] Regular physical examinations are a convenient and effective way to understand one's own health status. Generally speaking, physical examination items include: clinical physical examination department physical examination items, instrument physical examination department physical examination items and laboratory physical examination department physical examination items. Combining the doctor's rich clinical experience, the auxiliary detection of the equipment and the laboratory chemical test data of blood, urine and stool, it can comprehensively evaluate the individual's physical condition and play the role of early detection, early diagnosis and early treatment of major diseases.

[0004] CN112863659A discloses a physical examination service management method, device, electronic terminal and storage medium, the method comprising: responding to a reservation request from a user terminal, determining physical examination reservation information, wherein the physical examination reservation information includes physical examination items; obtaining medical history data corresponding to the user terminal from a preset database, generating physical examination precautions according to the medical history data and physical examination items; and sending the physical examination precautions to the user terminal. The invention realizes intelligent physical examination service management, achieves the technical effect of saving resource consumption, and facilitates people to make physical examination reservations and tests.

[0005] CN112349415A discloses a method, device and storage medium for constructing a resident health index based on physical examination big data. The method includes: obtaining resident physical examination data, and constructing a first physical examination data set containing the examination results of each physical examination item of the residents; preprocessing the first physical examination data to construct a second physical examination data set; normalizing the second physical examination data set to obtain a third physical examination data set, and obtaining the index weight of each physical examination item; isotropically processing the second physical examination data set to obtain a fourth physical examination data set, and constructing the physical examination data of the best healthy person and the physical examination data of the worst healthy person; obtaining a first healthy distance and a second healthy distance; calculating an individual health index based on the first healthy distance and the second healthy distance. Based on physical examination big data, the invention comprehensively evaluates the health index of the examinee for the physical examination items and unexamined items of the examinee, thereby providing a data basis for the comprehensive evaluation of the health of the residents.

[0006] However, currently, there is no method or model in the art that can accurately predict and determine the risk of an individual suffering from a certain disease based on the individual's physical examination data. Summary of the invention

[0007] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a disease risk level prediction model and its construction method and application. The present invention combines physical examination report data and clinical data for analysis and develops a new artificial intelligence fusion model for accurately predicting and judging the risk of a certain disease. It can more efficiently and quickly predict the risk level of patients suffering from diseases to determine whether they have asymptomatic or temporarily asymptomatic diseases, improve the accuracy and precision of the prediction effect, and provide the possibility for early screening, early prevention and early treatment of diseases.

[0008] To achieve this object, the present invention adopts the following technical solutions:

[0009] In a first aspect, the present invention provides a method for constructing a disease risk level prediction model, characterized in that the construction method comprises:

[0010] S1. Obtaining the physical examination report data of physical examination personnel with outpatient or hospitalization records as a data set, and dividing the population suffering from a specific disease and the population not suffering from the specific disease in the data set;

[0011] S2, converting the unstructured multivariate data in the data set into application-level structured data, and filling the obtained application-level structured data;

[0012] S3. Using the filled data as a training set, adopting a machine learning method to learn, screen and obtain important characteristic indicators of the specific disease;

[0013] S4, training the model with the positive and negative sample sets divided by the physical examination report data of the physical examination personnel with outpatient records or hospitalization records, and then screening the positive samples in the physical examination report data of the physical examination personnel without outpatient records or hospitalization records, expanding the positive sample training set, and again distinguishing the people suffering from the specific disease from the people not suffering from the specific disease;

[0014] S5. For the expanded training set data with determined characteristic indicators, a model is built using the optimal indicator reduction method, and then the model is fused to obtain the disease risk level prediction model.

[0015] In the present invention, for the training set data of the determined indicators, the optimal indicator reduction method is used to model the model. The optimal indicator reduction means that after modeling with all indicators, the model is saved, the most important indicator is deleted, and the model is retrained, and so on, until the effect of the model drops below the set threshold. Its advantage is to reduce the influence of strong indicators on the model and increase the robustness and generalization ability of the model.

[0016] The disease risk level prediction model can be used to predict a variety of diseases, such as coronary heart disease, stroke, cerebral hemorrhage, cerebral infarction, cerebral infarction, gastric cancer, etc.

[0017] As a preferred technical solution of the present invention, the data set includes personal information data, inspection data and examination data.

[0018] The personal information data includes physical sign data, medical history, family history or surgical history data.

[0019] The test data include laboratory test data, such as routine blood test, routine urine test, blood lipid test and other data.

[0020] The inspection data is text description data, including conclusion data of inspections such as X-ray, CT, MRI, electrocardiogram, abdominal color ultrasound, and pathology.

[0021] As a preferred technical solution of the present invention, the filling method described in step S2 includes:

[0022] For personal body shape indicators (such as height, weight, etc.), use the average value of the individual in recent years to fill in, or use the average value of people of the same age and gender to fill in;

[0023] For other indicators except the missing personal body shape, the average value of people of the same age and gender is used to fill in, the difference method is used to fill in, or the machine learning method is used to fill in.

[0024] As a preferred technical solution of the present invention, the method for obtaining the important characteristic indicators of the specific disease in step S3 includes:

[0025] Using the xgboost machine learning method to learn, based on feature recursive elimination and / or genetic algorithm, to screen and obtain the feature indicators;

[0026] After obtaining the important characteristic indicators of the specific disease, step S3 also includes the step of obtaining derived characteristics;

[0027] The derived features include performing square, square root, exponential or exponential power operations on a single feature index, or performing addition, subtraction, multiplication and division operations on at least two feature indexes.

[0028] As a preferred technical solution of the present invention, after obtaining the important characteristic indicators of the specific disease in step S3, it also includes the steps of discovering abnormal samples and removing abnormal samples;

[0029] The method for discovering abnormal samples includes: learning structured data of people suffering from the specific disease and people not suffering from the specific disease, using at least two machine learning methods to build models, fusing at least two models, and using the fused model to predict the risk level of the specific disease of the corresponding physical examination personnel, comparing the predicted results, and screening out abnormal samples.

[0030] The method for removing abnormal samples includes: marking or modifying data in the abnormal samples that affect important characteristic indicators, or directly deleting the abnormal samples.

[0031] In the method of the present invention, the data affecting the characteristic index in the abnormal sample will be marked or modified, or the abnormal sample will be directly deleted. The removal of abnormal samples will be adjusted according to the actual situation so that it does not affect the accuracy of model construction.

[0032] As a preferred technical solution of the present invention, the model fusion method in step S5 includes:

[0033] The model modeled with all indicators is model_0. After removing the most important indicators, the model modeled is model_1. And so on, we get model_0, model_1, ..., model_n, a total of n+1 models. We use the linear machine learning model to model the model and learn the corresponding coefficients.

[0034] model=a1*(model_0)+a2*(model_1)+∧+a n *(model_n)

[0035] In the present invention, the model fusion method also includes using other models for filling, including:

[0036] Taking model_0 as the main model, recall the samples whose prediction value of model_0 is less than M but whose prediction value of model_i is greater than N, where M and N are thresholds of different sizes and M<N;

[0037] The predicted values ​​are:

[0038] y pred =max(model_i), i∈(1≤i≤n).

[0039] In the present invention, the model fusion step can also be replaced by a voting method, that is, a majority vote, in which a sample is scored by multiple models and graded according to the scoring results, such as: high risk, medium-high risk, low risk or medium-low risk, etc.; the risk level is determined by voting.

[0040] Alternatively, the model fusion step may also be replaced by a mean method.

[0041] In the present invention, the disease risk level prediction model divides the prediction results into four levels;

[0042] Among them, the real sick population has a high risk level, the non-sick population has a low risk level, and the risk levels of the intermediate transition population other than the real sick population and the non-sick population are divided into medium-low risk level and medium-high risk level.

[0043] For people whose predicted results are medium-low risk or medium-high risk, it can remind people to take early prevention and treatment as early as possible to avoid the occurrence of the disease.

[0044] In a second aspect, the present invention provides a disease risk level prediction model constructed using the construction method described in the first aspect.

[0045] In a third aspect, the present invention further provides a disease risk level prediction device, the device comprising:

[0046] The data collection module is used to collect the physical examination report data, outpatient data and hospitalization data of the physical examination personnel;

[0047] A data conversion and filling module, used to convert the obtained data into application-level structured data and fill the obtained application-level structured data;

[0048] A prediction module predicts the disease risk level of the physical examination person based on the disease risk level prediction model described in the second aspect.

[0049] In the present invention, the device may further include a front-end rendering and display module to render and display the predicted data of the disease risk level on the front end, so as to facilitate viewing and processing by users.

[0050] In a fourth aspect, the present invention provides a computer device, the device comprising:

[0051] one or more processors;

[0052] A storage device for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method for constructing a disease risk level prediction model as described in any one of the first aspects.

[0053] In addition, the present invention also provides a computer storage medium on which a computer program is stored, and when the program is executed by a processor, the method for constructing a disease risk level prediction model as described in the first aspect is implemented.

[0054] All or part of the steps of the above method can be completed by hardware related to the computer program. The above computer program can be stored in a computer-readable storage medium.

[0055] When the program is executed, it executes the steps of the above-mentioned method embodiments; the computer storage medium may include but is not limited to floppy disks, optical disks, CD-ROMs (read-only disk memories), magneto-optical disks, ROMs (read-only memory), RAMs (random access memory), EPROMs (erasable programmable read-only memory), EEPROMs (electrically erasable programmable read-only memory), magnetic cards or optical cards, flash memory, or other types of media / machine-readable media suitable for storing machine-executable instructions.

[0056] The computer storage medium may be a product that is not connected to a computer device, or a component that is connected to a computer device for use.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The method for constructing a disease risk level prediction model provided by the present invention is based on the physical examination data and clinical data of a large number of physical examination personnel, and goes through the steps of collection, conversion, learning and optimization. Among them, a multivariate data training model is used, combined with the optimal indicator reduction method, to construct a disease risk level prediction model; the model divides disease risk into multiple levels such as low risk, medium-low risk, medium-high risk and high risk. It is an effective assessment tool that can provide reliable evidence-based medicine risk level recommendations, help patients understand their own health status, and thus effectively prevent and intervene in early stages to reduce the incidence of the disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 Schematic diagram of the process of constructing the disease risk level prediction model provided in the present invention.

[0060] Figure 2It is a schematic diagram of the structure of the disease risk level prediction system provided in the present invention. DETAILED DESCRIPTION

[0061] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and through specific implementation methods. However, the following examples are only simple examples of the present invention and do not represent or limit the scope of protection of the present invention. The scope of protection of the present invention shall be subject to the claims.

[0062] The disease risk level prediction model provided by the present invention is Figure 1 The steps shown are as follows:

[0063] S1. Prepare the data set: obtain the medical examination report data and outpatient diagnosis data of medical examinees for many years;

[0064] It should be noted here that the physical examination report data and outpatient diagnosis data of the physical examination personnel are obtained after the physical examination personnel know and agree, and are only used to construct the disease risk level prediction model described in the present invention, and are properly managed to prevent personal information leakage.

[0065] Then, the data set is divided: according to the diagnosis information of the medical examinees when they visit the hospital and the medical history, surgical history and other information in the medical examination report, the medical examinees are divided into a group of people suffering from a specific disease and a group of people without a disease.

[0066] S2. Convert unstructured multivariate data into application-level structured data: parse the physical examination report data obtained in the previous step, and parse the physical examination report data into multiple data including: physical sign data, past history, family history, surgical history and other personal information data, laboratory data, electrocardiogram, bilateral carotid artery color Doppler ultrasound conclusions and other text-based examination data.

[0067] Furthermore, data integrity processing is performed: for application-level structured data, features are selected based on the missing rate; for personal body shape indicators with little change in height, weight, and waist circumference, the average values ​​of the individual in recent years can be used to fill in. If there are none, the average values ​​of the same age and gender can be used to fill in;

[0068] For other missing indicators, the mean value of people of the same age and gender is used to fill in the missing indicators, or the difference method or machine learning method is used to fill in the missing indicators.

[0069] S3. Screening representative characteristic indicators: using machine learning methods such as xgboost to learn and obtain the main indicators; selecting the main indicators based on the random sorting method;

[0070] In some specific embodiments, the main index may also be selected by a genetic algorithm;

[0071] After the characteristic index is obtained, in some specific embodiments, the step of obtaining a derived characteristic may also be included.

[0072] The derived features include: performing square, square root, exponential or exponential power operations on a single feature index, or performing addition, subtraction, multiplication and division operations on at least two feature indexes.

[0073] In addition, in some specific embodiments, after the characteristic indicators are obtained, the steps of discovering abnormal samples and removing abnormal samples are also included.

[0074] For the people who have been divided into those with specific diseases and those who do not, we learn from structured data, use a variety of machine learning methods to build models, fuse multiple models, and comprehensively consider and evaluate the physical examination reports. We compile statistics on the predictions of all physical examination reports of each person, screen out those whose physical examination reports vary greatly over the years, and conduct review and cleaning.

[0075] In certain specific embodiments, the abnormal samples found during the screening process include substitute samples and invalid samples;

[0076] The substitute samples include: elderly substitute (old for young), young substitute (young for old), male substitute (male for female), female substitute (female for male);

[0077] Invalid samples include: early physical examination reports and physical examination reports of potential patients.

[0078] Cleaning of test substitution samples: For the four situations of old replacing young, young replacing old, male replacing female, and female replacing male, it is necessary to consider the accuracy of test substitution reports affecting classification or marking. Generally, the items that affect classification or marking in test substitution reports are marked or modified, or all test substitution reports of the person who replaced the test are deleted.

[0079] S4. Expand the positive sample training set: For potential illness, it is generally manifested as actual illness, but the steps recorded in S1 to distinguish between the sick and the healthy based on the physical examination information cannot correctly classify them; through training the model, multiple iterations, the sick and the healthy people are distinguished again.

[0080] S5. Modeling by using the optimal index decreasing method: For the training set data of the determined index, the present invention adopts the optimal index decreasing method to perform modeling;

[0081] Among them, optimal index reduction means that after using all indicators to build a model, the model is saved, the most important indicator is deleted, and the model is retrained, and so on, until the effect of the model drops below the set threshold. This method can reduce the impact of strong indicators on the model and increase the robustness and generalization ability of the model.

[0082] Furthermore, in certain specific embodiments, the method provided by the present invention further comprises a step of model fusion. For the model established by the optimal index reduction method, model fusion is performed;

[0083] The specific operations of the model fusion method are:

[0084] Assume that the model for modeling all indicators is model_0. After removing the most important indicator, the model for modeling is model_1. And so on, we get model_0, model_1, ..., model_n, a total of n+1 models. We use linear machine learning models such as LR to model the models and learn the corresponding coefficients. The fused model is as follows:

[0085] model=a1*(model_0)+a2*(model_1)+∧+a n *(model_n)

[0086] In the present invention, a model with a higher accuracy rate can be used as the main model, and other models can be used to supplement it; for example:

[0087] Taking model_0 as the main model, the samples with low prediction of model_0 but prediction of model_i (n>=i>=1) greater than a certain threshold are recalled, and the prediction value is:

[0088] y pred =max(model_i),i∈(1≤i≤n)

[0089] In some specific embodiments, the model fusion step can also be replaced by a voting method, that is, a majority vote, in which a sample is scored by multiple models and graded according to the scoring results, such as: high risk, medium-high risk, low risk or medium-low risk, etc.; the risk level is determined by voting.

[0090] In some specific embodiments, the model fusion step can also be performed by the mean method, that is, averaging the scoring results of each model to the model, and then grading the average scoring results.

[0091] The present invention also provides a disease risk level prediction device, the device is as follows Figure 2 As shown, including:

[0092] (1) Data collection module: The data collection module is connected to the hospital's physical examination system, outpatient system, and inpatient system to collect physical examination report data, outpatient data, and inpatient data of the examinees;

[0093] (2) a data conversion and filling module, used to convert the obtained data into application-level structured data and fill the obtained application-level structured data;

[0094] (3) A prediction module, which predicts the disease risk level of the person under physical examination based on a disease risk level prediction model.

[0095] In some specific embodiments, the system further includes a front-end rendering and display module for rendering and displaying the predicted data of the disease risk level on the front-end.

[0096] In some specific embodiments, the system also includes a user management module for managing information of medium-low risk, medium-high risk, and high-risk users. For example, when new user data has not been reviewed by a doctor, it will be displayed as unreviewed information on the front end; when the doctor clicks to review the user information, the user will enter the history library and archive.

[0097] The disease risk level prediction model or prediction method constructed by the above method can grade the disease risk of the physical examination personnel.

[0098] At present, it is too one-sided to simply classify disease levels into two types: low risk or high risk. In the classification process, the real sick people and the non-sick people are relatively concentrated, but the potential sick people or the sick people who have been controlled still exist in the intermediate transition part.

[0099] Therefore, the present invention defines the risk level into four levels: low risk, medium-low risk, medium-high risk, and high risk. Low risk and high risk are in a relatively objective state, and reasonable reminders and controls are provided to the medical examinees with medium-low risk and medium-high risk, which is of great significance for early prevention and treatment of diseases.

[0100] The applicant declares that the above is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention shall fall within the protection scope and disclosure scope of the present invention.

Claims

1. A method for constructing a disease risk level prediction model, characterized in that: The construction method comprises: S1. Obtaining the physical examination report data of physical examination personnel with outpatient or hospitalization records as a data set, and dividing the population suffering from a specific disease and the population not suffering from the specific disease in the data set; S2, converting the unstructured multivariate data in the data set into application-level structured data, and filling the obtained application-level structured data; S3. Using the filled data as a training set, adopting a machine learning method to learn, screen and obtain important characteristic indicators of the specific disease; S4, training the model with the positive and negative sample sets divided by the physical examination report data of the physical examination personnel with outpatient records or hospitalization records, and then screening the positive samples in the physical examination report data of the physical examination personnel without outpatient records or hospitalization records, expanding the positive sample training set, and again distinguishing the people suffering from the specific disease from the people not suffering from the specific disease; S5. For the expanded training set data with the determined important characteristic indicators, a model is built by using the optimal indicator decreasing method, and then the model is integrated to obtain the disease risk level prediction model; The data set in step S1 includes personal information data, inspection data and examination data; The personal information data includes physical sign data, medical history, family history or surgical history data; The test data include laboratory test data, including blood routine test, urine routine test and blood lipid data; The examination data is text description data, including electrocardiogram data and / or bilateral carotid artery color Doppler ultrasound conclusion data; The filling method in step S2 includes: For personal body shape indicators, use the average value of the individual in recent years to fill in, or use the average value of people of the same age and gender to fill in; For other missing indicators except for personal body shape, the average value of people of the same age and gender is used to fill in, the difference method is used to fill in, or the machine learning method is used to learn and fill in; The method for obtaining the important characteristic indicators of the specific disease in step S3 includes: The xgboost machine learning method is used for learning, and the characteristic indicators are screened based on feature recursive elimination and / or genetic algorithm; after obtaining the important characteristic indicators of the specific disease, step S3 also includes the step of obtaining derived characteristics; the derived characteristics include performing square, square root, exponential or exponential power operations on a single characteristic indicator, or performing addition, subtraction, multiplication and division operations on at least two characteristic indicators; The model fusion method in step S5 includes: The model modeled with all indicators is model_0. After removing the most important indicators, the model modeled is model_1. Then, by analogy, we get n+1 models including model_0, model_1 and model_n. We use the linear machine learning model to model the model and learn the corresponding coefficients, which are as follows: model=a1*(model_0)+a2*(model_1)+Λ+a n *(model_n); The model fusion method also includes using other models for filling, including: Model_0 is used as the main model. Samples whose prediction values ​​of model_0 are less than M but whose prediction values ​​of model_i are greater than N are recalled. M and N are thresholds of different sizes and M <N; The predicted values ​​are: y pred =max(model_i),i∈(1≤i≤n); The disease risk level prediction model divides the prediction results into four levels; Among them, the real sick population has a high risk level, the non-sick population has a low risk level, and the risk levels of the intermediate transition population other than the real sick population and the non-sick population are divided into medium-low risk level and medium-high risk level.

2. The construction method according to claim 1, characterized in that: After obtaining the important characteristic indicators of the specific disease in step S3, the steps of discovering abnormal samples and removing abnormal samples are also included; The method for discovering abnormal samples includes: learning structured data of a population suffering from the specific disease and a population not suffering from the specific disease, using at least two machine learning methods to build models, fusing at least two models, and using the fused model to predict the specific disease risk level of the corresponding physical examination personnel, comparing the predicted results, and screening to obtain abnormal samples; The method for removing abnormal samples includes: marking or modifying data in the abnormal samples that affect important characteristic indicators, or directly deleting the abnormal samples.

3. A disease risk level prediction model constructed using the construction method as described in claim 1 or 2.

4. A disease risk level prediction device, characterized in that: The disease risk level prediction device comprises: The data collection module is used to collect the physical examination report data, outpatient data and hospitalization data of the physical examination personnel; A data conversion and filling module, used to convert the obtained data into application-level structured data and fill the obtained application-level structured data; A prediction module predicts the disease risk level of the physical examination person based on the disease risk level prediction model according to claim 3.

5. A computer device, characterized in that: The device comprises: one or more processors; A storage system for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method for constructing a disease risk level prediction model as described in claim 1 or 2.

Images