CN107680676A - A kind of gestational diabetes Forecasting Methodology based on electronic health record data-driven - Google Patents

Abstract

The invention discloses a kind of gestational diabetes Forecasting Methodology based on electronic health record data-driven, and the effect to become more and more important is played in intelligent medical treatment service, has invented the gestational diabetes based on machine learning herein（gestational diabetes mellitus,GDM）Gestational diabetes prediction framework, according to the time windows division methods of gathered data, construct universe data prediction model, by stages data prediction model and three groups of prediction frameworks of weekly data forecast model.After identification prediction problem, associated by input with ETL data cleansings, medical record coding with characteristic, the processing of electronic health record data prediction, secondary data, Feature Engineering, machine learning, predict and apply seven steps, realize the data mining of high-dimensional electronic health record.The flag data collection on making a definite diagnosis is constructed using clinical data, and is divided into 2 subsets for model training and test.It is predicted by SVMs, Bayesian network, decision tree and based on integrated mixed model, realizes GDM pattern classifications.

Description

A data-driven prediction method for gestational diabetes mellitus based on electronic medical records

technical field

The invention relates to the field of diabetes prediction, in particular to a method for predicting gestational diabetes driven by electronic medical record data.

Background technique

In disease prediction, take gestational diabetes mellitus (GDM) as an example. According to the International Diabetes Federation survey, although more and more women receive prenatal care, it is still the most common pregnancy complication, which is defined as Diabetes with normal glucose metabolism or potential impaired glucose tolerance before pregnancy, or diagnosed during pregnancy. The serious consequences of GDM have made the medical community attach great importance to its early diagnosis and prevention. Risks associated with gestational diabetes include type 2 diabetes in both mother and child, fetal overgrowth and related short-term adverse prognosis risks, and long-term obesity in offspring. GDM predictive diagnosis and prevention are important issues that are of great concern to the maternal and child health care group, making it an important field for the application of big data in health care. Both clinicians and pregnant women expect to be able to perceive the risk of GDM at an earlier stage of pregnancy, so as to prevent and intervene as early as possible. With the periodic collection of electronic medical records (EMR) and related clinical data and the accumulation of cross-border data, EMR reuse and big data analysis technologies provide prospective tools for the early diagnosis and prevention of GDM.

According to traditional diagnostic methods, such as the National Health Industry Standard Diagnosis Guidelines for Gestational Diabetes (2011), the screening and diagnosis of gestational diabetes are often performed by oral glucose tolerance test (OGTT) at 24-28 weeks of gestation. Before that, pregnant women will also undergo a large number of systematic prenatal examinations. The comprehensive and dynamic collection, utilization and transmission of pregnant women's health and medical data has accumulated a large amount of EMR data in the multi-source information system. Combined with massive health and medical big data, a large amount of fragmented information and tracking data on maternal and child health care services can also be collected through tools such as maternal and child health care intelligent applications and clinical decision support systems, providing a large amount of cross-border data for GDM forecasting. These applications can dynamically collect data in real time and provide personalized and precise health management services, which has become a new trend in the industry. Realizing GDM risk prediction and pattern identification through EMR, reducing the high risk of GDM disease to mothers and infants, has gradually become an important way to improve the health of women and children.

Contents of the invention

In view of the deficiencies of existing research, the present invention provides a method for predicting gestational diabetes driven by electronic medical records; it plays an increasingly important role in regional medical services; using clinical data, combined with methods such as artificial intelligence and machine learning, provides An intelligent decision support system for diseases, which helps to solve the problems of repeated inspections and repeated diagnosis and treatment, improves the work efficiency of doctors and reduces labor load, strengthens the control of medical errors, enriches the service methods of the regional medical information sharing platform and improves the application value.

The implementation process of the present invention is to construct a method for predicting gestational diabetes based on electronic medical record data, which is characterized in that: comprising the following steps:

(1) Input and ETL data cleaning module. Obtain historical archived pregnant women data corresponding to EMR, and complete preliminary data cleaning through steps such as extraction, conversion, and loading, and complete deprivation and data quality management;

(2), medical record coding and feature data association module. Through the patient identification code ID of the EMR system, the time-space desensitization data association is carried out, and the characteristic data is screened by combining clinical knowledge and experience to generate a GDM data warehouse;

(3), EMR data preprocessing module. Perform missing value, discretization and normalization processing on input data;

(4), secondary data processing module. Perform classification label calibration and complete inclusion and exclusion criteria checks;

(5), feature engineering module. Divide the data into two categories: GDM and non-GDM, use the clinical data associated with the disease as the conditional attribute, mark the category as the decision attribute, and perform embedded feature selection;

(6), machine learning module. According to the selected input features, the whole data is divided into training samples and test samples, the time window and machine learning model are selected, and the cross method training is carried out to obtain the prediction algorithm;

(7) Forecast application module. Input the electronic medical record data of undiagnosed pregnant women into the machine learning model in step (6), and infer the occurrence value (or risk rate) of GDM in these undiagnosed pregnant women.

Specifically, first start the processing process of the EMR data-driven gestational diabetes prediction method, access the EMR data storage, input the data into the process, and complete the data cleaning work through the ETL module. Then, combined with the electronic medical record data collection process, the prediction problem is identified, and the medical record coding and characteristic data association are completed. Through de-privacy-related methods, the privacy information in patient data is eliminated, the data quality is checked, and the query associated desensitized data module is realized to complete the construction of the GDM data warehouse. In the EMR data preprocessing stage, the missing value processing is completed, and the data discretization and normalization processing are realized. Then, realize classification label analysis and calibration, complete the electronic medical record data inclusion and exclusion criteria check, and complete the secondary data processing. The experimental data set is divided into samples, including two data sets for GDM diagnosis or not, and then the feature engineering module is realized. Then, through the time window division and model selection module, enter: the global data prediction model GDPM, the staged data prediction model SDPM or the weekly data prediction model WDPM, and then enter the GDM prediction model based on machine learning, and then realize the prediction application. Use the application results as feedback control to optimize the model and enable data reuse in subsequent stages of feature engineering. Finally, end the data processing process.

According to a method for predicting gestational diabetes driven by electronic medical record data according to the present invention, it is characterized in that: different time window division methods are collected to construct a prediction model:

(1) Global data prediction model, using EMR data earlier than OGTT, to infer classification values;

(2) Stage data prediction model, using EMR in the first trimester or 13 to 23 weeks, to infer classification values;

(3) Weekly data prediction model, using weekly EMRs starting from week 12, to infer categorical values.

According to a kind of gestational diabetes prediction method driven by electronic medical record data according to the present invention, it is characterized in that: the GDM data warehouse is constructed by associating the medical record code with the feature data:

(1) File creation data set, including file ID, file doctor, file date, husband's age, husband's physical condition, husband's hobbies, BMI, etc.;

(2) Obstetrical inspection data set, including obstetrical inspection ID, weight, obstetric inspection time, gestational week, diastolic blood pressure, etc.;

(3), LIS data, including blood routine, liver function, kidney function, electrolytes, glycosylated hemoglobin, blood lipids, ferritin, etc.;

(4) The first page of the medical record, including the consultation number, medical record number, diagnosis number, diagnosis type, diagnostic ICD code (doctor), diagnostic ICD code (medical record), etc.

In view of the increasingly important role of prediction systems based on intelligent reasoning models in smart medical services, the present invention provides a prediction framework for gestational diabetes based on machine learning, and constructs a global data prediction model, There are three sets of forecasting frameworks: periodical data forecasting model and weekly data forecasting model. After identifying the prediction problem, through seven steps of input and ETL data cleaning, medical record coding and feature data association, EMR data preprocessing, secondary data processing, feature engineering, machine learning, and predictive application, the data of high-dimensional electronic medical records is realized. dig. A labeled dataset about diagnosis was constructed using clinical data and divided into 2 subsets for model training and testing. GDM pattern classification is achieved through support vector machines, Bayesian networks, CHAID decision trees, and predictions based on ensemble-based hybrid models. The regional medical application system was developed based on the prediction model, and the results showed that the GDM prediction model based on machine learning provided an effective application tool for early prediction of gestational diabetes in pregnant women in regional medical care.

The present invention has the following advantages: a gestational diabetes mellitus (GDM) prediction model set based on machine learning is constructed, and according to different time window division methods of collected data, it covers the main needs of pregnant women to predict gestational diabetes mellitus in advance, and improves clinical decision support The intelligent level of the system plays an increasingly important role in regional medical services; through the association of medical record coding and clinical feature data, a GDM data warehouse based on core data is constructed to provide technical support for knowledge mining and data management of high-dimensional EMR data, which can Improve doctors' work efficiency and reduce workload, enhance medical error control, enrich the service methods of the regional medical information sharing platform and improve the application value.

Description of drawings

Figure 1 Prediction of gestational diabetes driven by electronic medical record data;

Figure 2 GDM prediction framework based on secondary cleaning and machine learning;

Figure 3 Global data prediction model;

Figure 4 staging data prediction model;

Figure 5 Weekly data prediction model;

Figure 6 Data collection process before filing in the first trimester;

Fig. 7 GDM prediction module based on machine learning;

Fig. 8 is a schematic flowchart of a method for predicting gestational diabetes driven by electronic medical record data.

detailed description

The present invention will be described in detail below in conjunction with the accompanying drawings, and the technical solutions in the embodiments of the present invention will be clearly and completely described. The described embodiments are some but not all embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

To explore whether electronic medical record data covering clinical experience is helpful for the prediction of GDM status in the first-trimester group of pregnant women. Combining structured and unstructured data in electronic medical records and other external network data can not only conduct clinical data association analysis on complications of gestational diabetes, but also predict the risk of pregnant women suffering from gestational diabetes mellitus (GDM) in advance. This paper provides a GDM prediction framework earlier than OGTT diagnosis through the analysis method of health care big data, as shown in Figure 1, that is, the prediction of gestational diabetes driven by electronic medical record data;

A kind of gestational diabetes prediction method driven by electronic medical record data according to the present invention comprises the following steps:

(1) Obtain the historical archived pregnant women data corresponding to the electronic medical record data;

(2) Obtain the relevant GDM information determined by the OGTT, and divide the data of pregnant women into confirmed GDM and non-GDM for reference;

(3) The samples obtained from step (2) are associated with spatiotemporal data through the patient identification code ID of the electronic medical record system, and the feature data of the marked electronic medical records associated with the GDM disease are screened through feature selection;

(4) Obtain the clinical data (archive data, obstetric examination data, LIS data, medical record home page) and its characteristic attributes of each sample. The clinical data associated with the disease is used as a conditional attribute, and the data of the marked category is a classification decision attribute ; The conditional attribute of each sample electronic case is used as the input vector, and the GDM occurrence value of the disease is output to establish a machine learning model;

(5) Obtain the characteristic data of the electronic medical records of undiagnosed pregnant women, input the machine learning model in step (4), and reason to predict the occurrence value (or risk rate) of GDM in pregnant women to be investigated.

The process of gestational diabetes prediction driven by the above electronic medical record data is refined, and a GDM prediction framework based on secondary cleaning and machine learning is invented, as shown in Figure 2, that is, a GDM prediction framework based on secondary cleaning and machine learning:

In the model design phase, the first step is to identify the forecasting problem. On this basis, various medical data of patients with gestational diabetes are collected through various channels and stored in the computer information system. According to the system hypothesis analysis based on the time window division, the invented scheme mainly includes three groups of schemes. Using the LIS system, analyze the inspection data sent by the experimental instruments, generate inspection reports, and store them in the database through the network. The basis for clinical data extraction mainly includes the following three points: ①Qualitative-quantitative data regularization; ②Pre-gestational diabetes mellitus (PGDM) with pre-gestational diabetes; ③Clinical knowledge judgment, activated partial thromboplastin time (Activated Partial ThromboplastinTime, APTT) interval; ④Other data calibration and preprocessing processes.

In the primary processing stage, data cleaning tools such as Extract-Transform-Load (ETL) are mainly used. After the data is processed, time window iterative reasoning is performed. After the data warehouse is determined, clinical data extraction and secondary processing are carried out. In ICD-10, the diagnosis code for gestational diabetes is O24.4. Use ICD-10 to divide the data into the first page of medical records, inspection items, inspection results, etc., and establish electronic medical records. The data after secondary processing is subjected to feature engineering processing. It mainly includes two tasks: algorithm automatic selection; expert participation in selection of common feature set. The last is model mining, through processes such as knowledge discovery, optimization prediction and data analysis, to establish an appropriate model and store the collected information.

In the forecasting framework construction stage, three sets of forecasting models are formed based on the multi-scale time window division method. The choice of time scale determines the granularity of data used for machine learning. According to the multi-scale time window division method, according to the three grades of large, medium and small according to the time scale, the invented data mining scheme mainly includes three groups of models: global data prediction model, staged data prediction model and weekly data prediction model.

Table 1 Three groups of prediction models based on multi-scale time window division method

Note: Training data labels: positive and negative values are derived from the discrete value mapping of OGTT's diagnostic ICD.

In addition, you can choose a wider time window according to the scale, such as days, etc., and I will not introduce more here due to the space.

The following three groups of models are introduced respectively, namely, the global data forecasting model, the phased data forecasting model and the weekly data forecasting model.

Global Data Prediction Model GDPM: Global Data Prediction Model (Global Data Prediction Model, GDPM) extracts maternal electronic medical record data earlier than OGTT as a whole data set for machine learning and prediction. The global data prediction model is shown in Figure 3. In this line of thinking, the data is electronic case data collected throughout pregnancy, a sparse dataset. Through the prediction model to realize the identification of GDM in the first trimester and predict the occurrence of GDM in the later stage, the classification value (and its probability) of the global data is obtained, that is, the prediction value of gestational diabetes in the whole pregnancy cycle.

Staged data prediction model SDPM: The staged data prediction model (Staged Data Prediction Model, SDPM) divides the maternal electronic medical record data earlier than OGTT into the first trimester data set (before 13 weeks of pregnancy), the second trimester data set (13 weeks - Before OGTT), two data sets are formed for prediction. The staging data forecasting model is shown in Figure 4. It includes two sub-modes: first, input the data set of the first trimester (before 13 weeks of pregnancy), including the filing data of the first trimester, obstetric inspection data, LIS data, and the front page of medical records, and correspondingly obtain the predicted classification value (and its probability) of the first trimester . Second, input the second-trimester data set (13 weeks-before OGTT), including second-trimester filing data, obstetric inspection data, LIS data, medical record homepage, etc., and obtain the second-trimester prediction classification value (and its probability) accordingly.

Weekly data prediction model: The weekly data prediction model (Weekly Data Prediction Model, WDPM) divides the maternal electronic medical record data earlier than OGTT into multiple data sets according to a certain time window (such as gestational weeks) for prediction. The weekly data forecasting model is shown in Figure 5. Under this set of models, the input data are the 12th, 13th, 14th, ..., 23rd gestational week data sets, including the obstetric inspection data, LIS data, medical record homepage, etc. of each gestational week. Although the time interval of each gestational week is the same, due to the actual needs and behaviors of pregnant women, the statistical frequency distribution of each gestational week is determined to be uneven. In some gestational weeks, the data scale is relatively large, such as the 12th week. Some data on gestational weeks are relatively small, because the number of pregnant women who go to check-ups at these gestational weeks is not large.

Combining clinical knowledge and expert opinion in the field, considering the program GDPM, the statistical distribution of data is uneven, which is not suitable for subsequent data processing; in the program SDPM, the data time span is too long, and the signs of pregnant women vary greatly, so it is not suitable for further data processing. Only the relatively concentrated data collected in the first trimester meet the requirements and can be used for further analysis, that is, the implementation plan WDPM.

When using EMR data or conducting clinical data mining, in addition to mastering clinical guidelines and clinical pathways, the following important sub-processes need to be completed:

Input and ETL data cleaning: The input data (or data collection process) takes the traditional outpatient service process as a clue to analyze the process of data collection and information aggregation during the filing period. The process of data collection before filing in the first trimester is shown in Figure 5. In the first trimester, that is, within 0 to 12 weeks of pregnancy, no filing data has been formed. When pregnant women confirm pregnancy at 0-5 weeks, they will choose to have a gynecological examination or not to do a gynecological examination. If there is no abnormality, the obstetrician and gynecologist will file when the pregnant woman is 12 weeks pregnant. For pregnant women who do not have gynecological examinations, obstetricians and gynecologists will also file when they are 12 weeks pregnant. The next steps include: First, for the pregnant women undergoing gynecological examinations, obstetricians and gynecologists will perform B-ultrasound examinations for them. If the test result is an ectopic pregnancy, the pregnancy must be terminated; if not, a gynecological triad examination should be performed. Secondly, check whether the uterine fibroids affect the fetal development. If so, the surgeon needs to operate on the pregnant woman to remove the uterine fibroids; if there is no effect, perform routine gynecological examinations, including leucorrhea examination, cervical smear examination and gynecological speculum examine. Then, determine whether the pregnant woman is infected by bacteria. If the pregnant woman is infected by bacteria, a physician should be arranged for treatment immediately; if not infected, chorionic villus sampling can be performed, but this test is risky. Then, check whether the fetus is seriously abnormal. If the test results show that the fetus is seriously abnormal, the pregnancy must be terminated; if the fetus is not seriously abnormal, a physician will check the heart, liver, and kidney functions of the pregnant woman to determine whether other functions of the pregnant woman are abnormal. . If there is any abnormality, the pregnant woman should see a doctor for treatment; if there is no abnormality, the obstetrician and gynecologist will file when the pregnant woman is 12 weeks pregnant. For pregnant women who do not have gynecological examinations, obstetricians and gynecologists will also file when they are 12 weeks pregnant. In addition, in the 24-28 weeks of gestational glucose tolerance test data flow, pregnant women first receive the checklist, and after paying the fee, they arrive at the laboratory for blood routine, blood type and various antibody tests. Urine is then prepared for routine urinalysis. A vaginal discharge test is then performed. Finally, the 24-28 week glucose tolerance test is carried out. The medical staff of the endocrinology department first draw blood for the pregnant woman on an empty stomach; then the pregnant woman drinks the glucose water prescribed by the hospital, which takes about 5 minutes; one hour later, the medical staff draws blood for the pregnant woman; An hour later, the medical staff drew blood for the third time. Later, the medical staff tested the blood sugar concentration, which took a day, and finally the pregnant woman got the test result.

In the input and ETL data cleaning phase, data cleaning tools such as Extract-Transform-Load (ETL) are mainly used. After the data is processed, time window iterative reasoning is performed. The type of data was identified, followed by clinical data extraction. In data processing, the data cleaning methods used mainly include: data standardization processing, lookup table processing (removing invalid data), fuzzy dynamic matching, basic common sense logic judgment, domain common sense judgment, and data correlation verification. Through these methods, missing data processing, similar duplicate object detection, abnormal data processing, logical error detection, inconsistent data processing, etc. are realized.

Medical record coding and characteristic data association: through the patient identification code ID of the EMR system, the temporal and spatial desensitization data association is carried out, and the characteristic data is screened in combination with clinical knowledge and experience to generate a GDM data warehouse. The data sources of the data warehouse used for GDM machine learning are mainly inpatients (such as pregnant women) and outpatient data, including: archive data, obstetric inspection details, LIS data, and medical record homepages. The filing data selects the filing ID, filing doctor, filing time, husband's age, husband's physical condition, husband's hobbies, BMI, etc. Obstetric inspection details select registration number, obstetric inspection ID, weight, obstetric inspection time, gestational week, diastolic blood pressure, etc. Blood routine, liver function, kidney function, electrolytes, glycosylated hemoglobin, blood lipids, ferritin, etc. were selected for LIS data. On the first page of the medical record, visit number, medical record number, diagnosis number, diagnosis type, diagnosis ICD code (doctor), diagnosis ICD code (medical record).

EMR prediction data preprocessing: missing value, discretization and normalization processing of input data; data quality problems and examples in data cleaning, as shown in Table 2.

Table 2 Data quality problems and examples in data cleaning

Secondary data processing: In ICD-10, the diagnosis code for gestational diabetes mellitus is O24.4. Classification label calibration was performed and inclusion and exclusion criteria checks were completed. Use ICD-10 to divide the data into the first page of medical records, inspection items, inspection results, etc., and establish electronic medical records.

Feature engineering: In the statistical information table of clinical data sets in data cleaning and secondary processing, data sources are divided into archive data, obstetric inspection details, LIS data, medical record home pages, etc. The data after secondary processing is subjected to feature engineering processing. Divide the data into two types, GDM and non-GDM, regard the clinical data associated with the disease as the conditional attribute, mark the category as the decision attribute, and perform embedded feature selection. It mainly includes two tasks: automatic algorithm selection; experts participate in the selection of common feature sets and store the extracted feature information.

Machine learning: Predictive model selection: Combining clinical knowledge and expert opinions in the field, fully considering the main features of the scheme GDPM, scheme SDPM, and scheme WDPM, such as data time span, differences in pregnant women's signs, etc., to select a GDM prediction model based on machine learning. Through the processes of knowledge discovery, optimization prediction and data analysis, an appropriate data mining model is established. According to the selected input features, the whole data is divided into training samples and test samples, the time window and machine learning model are selected, and the cross method training is carried out to obtain the prediction algorithm. Using Bayesian network, support vector machine, CHAID decision tree and integrated mixed model, the prediction experiment of gestational diabetes is carried out, as shown in Figure 7, which is the GDM prediction module based on machine learning. The trained learning algorithm can provide feedback information to improve the parameter learning and model tuning process. And use receiver operating characteristic curve (receiver operating characteristic curve, ROC) and area under the curve (Area Underroc Curve, AUC) to provide evaluation indicators, comprehensively evaluate the performance of the model, and then feed back the control prediction model and algorithm.

Prediction application: input the electronic medical record data of undiagnosed pregnant women into the machine learning model, and deduce the GDM occurrence value (or risk rate) of these undiagnosed pregnant women. This method realizes the GDM data mining of the unbalanced positive and negative samples of high-dimensional EHRs, constructs the feature dimension of the archived data and the nuclear data set of OGTT diagnostic ICD markers, and realizes the identification of GDM in the first trimester through the prediction model. Divide the archival sample dataset into GDM and non-GDM (nGDM) subsets for model training and testing. Predict the occurrence of GDM in the later period through the prenatal inspection data filed in the first trimester, which is earlier than the OGTT diagnosis result under normal circumstances. Among them, the global data prediction model is slightly earlier than the OGTT diagnosis result, and the staging data prediction model will advance the prediction status to the first trimester or the second trimester before OGTT; the weekly data prediction model will advance the risk of GDM prediction to each gestational week.

This method provides an important prospective tool for the early diagnosis and prevention of GDM, which is based on the process module of the GDM prediction method driven by electronic medical record data, as shown in Figure 8. Realize GDM risk prediction and pattern recognition through HER. Using the model with the highest accuracy for prediction, the results are significantly earlier than the GDM screening and diagnosis time at 24 to 28 weeks of gestation, which can prompt pregnant women to participate in exercise and healthy diet plans in time.

Application of GDM prediction in maternal and child medical services:

The GDM forecasting system provides an important platform for the storage and analysis of regional medical maternal and child health big data. It provides highly integrated system functions with an open architecture. It can meet the needs of different levels through flexible configuration of regulatory topics. Provide strong regulatory business support. The intelligent application of maternal and child health care based on the GDM prediction function of machine learning can enable pregnant women to understand the dangers of high-risk pregnancy, so as to consciously participate in the systemic health care during pregnancy and childbirth, and specifically remind pregnant women to actively participate in physical exercise and health care. Diet plan, etc., to reduce the incidence of abnormal perinatal pregnancy and perinatal disease and mortality of pregnant women, and ensure the safety of women and children. Maternal and child health care smart applications can improve individual health management capabilities for women and children by providing health information services.

Conclusion: Prediction of gestational diabetes becomes the key to early diagnosis and prevention of related diseases in pregnant women. The data-driven GDM prediction model invented in this paper provides an effective method for predicting pregnant women in advance. According to the different time window division methods of the collected data, three groups of prediction frameworks were constructed. Predict the incidence of GDM in the later period by collecting the data of pregnant women's obstetric examination, and predict the risk of GDM in pregnant women in advance to varying degrees compared with the OGTT diagnosis time under normal circumstances. This article provides an important prospective tool for early diagnosis and prevention of GDM, and follow-up work will improve the accuracy of prediction methods and provide individualized medical services.

The above description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (1)

1. A kind of 1. gestational diabetes Forecasting Methodology based on electronic health record data-driven, it is characterised in that：In such a way Realize；Structure is with lower module；

   （1）, input with ETL data cleansing modules：History corresponding to obtaining EMR is filed pregnant woman's data, and by extracting, change and The steps such as loading complete preliminary data cleaning, and privacy and data quality management are gone in completion；

   （2）, medical record coding with characteristic relating module：Space-time desensitization data are carried out by the patient identification code ID of emr system Association, with reference to clinical knowledge and experience screening characteristic, generate GDM data warehouses；

   （3）, EMR data pretreatment module：Missing values, discretization and normalized are carried out to input data；

   （4）, secondary data processing module：Tag along sort calibration is carried out, exclusion standard inspection is included in completion；

   （5）, Feature Engineering module：Data are divided into GDM and the classes of non-GDM two, using with the clinical data of disease association as bar Part attribute, mark classification are decision attribute, carry out embedded feature selecting；

   （6）, machine learning module：According to the input feature vector of selection, total evidence is divided into training sample and test sample, selected Time window and machine learning model, crossing method training is carried out, obtains prediction algorithm；

   （7）, prediction application module：By not yet diagnosed pregnant woman's electronic health record data, input step（6）In machine learning model, The GDM occurrence values of these pregnant woman to be diagnosed of reasoning（Or relative risk）；

   During implementation, start the processing procedure of the gestational diabetes Forecasting Methodology of electronic health record data-driven first, access electronics Medical record data memory, and flow is entered data into, completing data cleansing with ETL data cleansing modules by input works；With Electronic health record Data Collection flow is combined afterwards, identification prediction problem, is completed medical record coding and is associated with characteristic；By going privacy Correlation technique, the privacy information in patient data is eliminated, check the quality of data, realize the desensitization data module of inquiry association, it is complete Built into GDM data warehouses；

   Handled by EMR data pretreatment module in EMR data pretreatment stage, completion missing values, realize Data Discretization with returning One changes processing work；

   Then, analyzed and calibrated by secondary data processing modules implement tag along sort, completed electronic health record data and include exclusion mark Standard checks, completes secondary data processing work；Sample division is carried out to experimental data set, including GDM whether two numbers made a definite diagnosis According to collection, and then realize Feature Engineering module；

   Then, the division of passage time window and Model selection module, enter：Universe data prediction model GDPM, by stages data prediction Model SDPM or weekly data forecast model WDPM, subsequently into the GDM forecast models based on machine learning, and then realize that prediction should With；Optimized model will be carried out as feedback control using result, data-reusing is realized in the Feature Engineering of follow-up phase；Finally tie Beam data handling process.