WO2020181806A1 - 预测未来血糖值的方法、装置及计算机设备 - Google Patents

预测未来血糖值的方法、装置及计算机设备 Download PDF

Info

Publication number
WO2020181806A1
WO2020181806A1 PCT/CN2019/117222 CN2019117222W WO2020181806A1 WO 2020181806 A1 WO2020181806 A1 WO 2020181806A1 CN 2019117222 W CN2019117222 W CN 2019117222W WO 2020181806 A1 WO2020181806 A1 WO 2020181806A1
Authority
WO
WIPO (PCT)
Prior art keywords
blood glucose
data
sequence
historical
value
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2019/117222
Other languages
English (en)
French (fr)
Inventor
金晓辉
阮晓雯
徐亮
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2020181806A1 publication Critical patent/WO2020181806A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/04Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • GPHYSICS
    • G06COMPUTING OR CALCULATING; COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/06Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063Operations research, analysis or management
    • G06Q10/0639Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393Score-carding, benchmarking or key performance indicator [KPI] analysis
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H10/00ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/60ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients

Definitions

  • This application relates to the field of computer technology, in particular to a method, device and computer equipment for predicting future blood glucose levels.
  • Diabetes is a group of metabolic diseases characterized by high blood sugar, which can cause damage to large blood vessels and capillaries and endanger the heart, brain, kidneys, peripheral nerves, eyes, feet and other parts of the disease. There are also many complications. Therefore, it is absolutely necessary to strengthen the prediction of diabetes. However, with the advancement of technology, the diagnosis of disease types is no longer limited to the analysis of doctors. The use of artificial intelligence to predict diabetes is in line with today's development trend.
  • the existing diabetes prediction methods can only judge whether users without diabetes-related medical records are suffering from diabetes, but for users with existing data and long-term follow-up, they cannot predict future blood glucose levels, and cannot better realize the diagnosis of diabetes. Diagnosis and treatment.
  • this application provides a method, device, and computer equipment for predicting future blood glucose levels.
  • the main purpose is to solve the problem that at present, it can only determine whether a user has diabetes, but cannot predict the future blood glucose level, which cannot be better realized.
  • a method for predicting future blood glucose levels is provided.
  • the method includes: extracting sample user data corresponding to sample users whose diabetes follow-up records are greater than a preset duration threshold from original health files and electronic medical records;
  • the sample user data creates a prediction model of the blood glucose value time series;
  • the historical user data of the target user is input into the prediction model to predict the first prediction sequence of the fasting blood glucose value of the target user in a preset future time period, and
  • the historical duration corresponding to the historical user data is less than the preset duration threshold, and the target user and the sample user are both diabetic patients.
  • a device for predicting future blood glucose levels includes: an extracting module for extracting from the original health files and electronic medical records the samples corresponding to the users whose diabetes follow-up records are greater than a preset duration threshold Sample user data; a creation module, used to use the sample user data to create a prediction model of blood glucose level time series; a determination module, used to input historical user data of the target user into the prediction model to predict the target user’s presence
  • a first prediction sequence of fasting blood glucose values in a future time period and a second prediction sequence of blood glucose values two hours after a meal are preset, wherein the historical duration corresponding to the historical user data is less than the preset duration threshold, and the target Both the user and the sample user are diabetic patients.
  • a non-volatile readable storage medium having computer readable instructions stored thereon, and when the computer readable instructions are executed by a processor, the above method for predicting future blood glucose values is realized.
  • a computer device including a non-volatile readable storage medium, a processor, and a computer-readable storage medium that is stored on the non-volatile readable storage medium and can run on the processor. Instructions, when the processor executes the computer-readable instructions, the above method for predicting future blood glucose values is implemented.
  • this application provides a method, device and computer equipment for predicting future blood glucose levels. Compared with the current method of predicting diabetes using the constructed 0-1 classification model, this application is based on the original health files and The sample user data extracted from the electronic medical record is used to create a blood glucose time series prediction model, and the historical user data of the target user is input into the prediction model, and then the prediction model is used to obtain the target user’s fasting blood glucose value in the preset future time period. The first prediction sequence and the second prediction sequence for the blood glucose level of the preset duration after a meal. Since the sample user and the target user are both diabetic patients, the prediction model can be used to determine the sample users who have the same or similar user data as the target user.
  • FIG. 1 shows a schematic flow diagram of a method for predicting future blood glucose values provided by an embodiment of the present application
  • FIG. 2 shows a flow chart of another method for predicting future blood glucose values provided by an embodiment of the present application Schematic diagram
  • Figure 3 shows a schematic structural diagram of a device for predicting future blood glucose values provided by an embodiment of this application
  • Figure 4 shows a schematic structural diagram of another device for predicting future blood glucose values provided by an embodiment of this application.
  • this embodiment provides a method for predicting the future blood glucose level, as shown in Figure 1. As shown, the method includes:
  • sample user data corresponding to sample users whose diabetes follow-up records are greater than a preset duration threshold from original health files and electronic medical records.
  • the sample users are diabetic patients.
  • Sample user data can include patient visit data, physical examination index data, medication data, and health notification data, etc.
  • the visit data also includes outpatient, hospitalization, and medication data.
  • Patient medication, symptoms, signs, and lifestyle follow-up data for diabetes are mainly derived from various Large hospitals and physical examination centers.
  • the preset duration threshold can be set according to actual business needs. For example, the time threshold can be set to one year, that is, sample user data corresponding to sample users whose diabetes follow-up records are greater than one year are obtained.
  • the sample user data extracted from the original health files and electronic medical records generally contains a lot of useless data and abnormal data, so before using these data to create a predictive model, these sample user data need to be cleaned and integrated
  • the required relevant features are extracted from it, and the extraction method is mainly through regular expressions.
  • Regular expressions are logical formulas for character strings, which match the text through specific character combinations defined in advance to obtain the required text and numbers.
  • sample user data uses the sample user data to create a prediction model of the blood glucose value time series.
  • data such as fasting blood glucose value, blood glucose value two hours after a meal, medication, symptoms, signs, and lifestyle of the sample user during each diabetes follow-up of the sample user are obtained from the sample user data, and then the blood glucose value time is created based on these data
  • the prediction model of the sequence can be specifically obtained by training the relevant algorithm of the time sequence.
  • the prediction model of the blood glucose value time series obtained by creating can well reflect the fasting blood glucose value time series and the blood glucose value time series two hours after a meal corresponding to the sample users with different diabetes follow-up characteristics.
  • the historical user data is the historical user data of the target user for a recent period of time, for example, the fasting blood glucose value at each diabetes follow-up in the last six months, the blood glucose value two hours after a meal, medication, symptoms, signs, lifestyle, etc. data.
  • the preset future time period is set according to actual needs, so that the prediction model of the blood glucose value time series finally shows the blood glucose value in the future time period.
  • the preset future time period can be 30 days in the future;
  • the first prediction The sequence is a time sequence of the fasting blood glucose value of the target user in a preset future time period;
  • the second prediction sequence is a time sequence of the target user's blood glucose value two hours after a meal in the preset future time period.
  • a time series prediction model of blood glucose level is created, and the historical user data of the target user is input into the prediction model, and then the prediction is used
  • the model obtains the first prediction sequence of the fasting blood glucose level of the target user in the preset future time period and the second prediction sequence of the blood glucose level of the preset duration after the meal. Since the sample user and the target user are both diabetic patients, the prediction model can be used Determine the sample user who has the same or similar user data as the target user, and the corresponding fasting blood glucose value time series and the blood glucose value time series two hours after a meal.
  • the diabetes follow-up records of the sample users are greater than the preset duration threshold, and the historical duration corresponding to the historical user data of the target user is less than the threshold, it is equivalent to using a large-scale time series to determine the trend of the small-scale time series contained therein.
  • a good prediction of the future blood glucose level change law of the target user, and then real-time monitoring of the development of diabetes, and corresponding supporting treatments are of great significance in the treatment and control of diabetes.
  • the method includes : 201. Extract sample user data corresponding to sample users whose diabetes follow-up records are greater than a preset duration threshold from original health files and electronic medical records.
  • step 202 Preprocess the sample user data. Since there are usually useless data and abnormal data in the sample user data, the sample user data needs to be preprocessed to obtain the data required for the creation of the predictive model. To illustrate this process, as an optional method, step 202 may specifically include : Use regular expressions to extract the time series data of fasting blood glucose values and the time series data of blood glucose values two hours after a meal (that is, the time series data corresponding to each follow-up) from the sample user data; if the time series data of fasting blood glucose values or two hours after a meal Hourly blood glucose time series data.
  • the mode of the multiple blood glucose measurement values is used as the blood glucose measurement value corresponding to that day; and if fasting
  • the interpolation method is used to supplement the empty value corresponding to the date to obtain the blood glucose sequence of the daily granularity; finally the daily granularity
  • the blood glucose sequence is subjected to differential movement smoothing processing to obtain a stable sequence. Using this relatively stable and stable sequence training model can eliminate noise, and the obtained prediction model can predict more accurate results.
  • the sample user a corresponds to a fasting blood glucose value and a blood glucose value two hours after a meal every day in the past year, that is, two initial blood glucose time series; in order to make the two blood glucose time series trend
  • the two blood glucose value time series can be converted into a stationary series, specifically, the two blood glucose value time series can be differentiated and moved smoothly.
  • the difference between the two initial blood glucose value time series and the movement smoothing process are the same.
  • the time series transformation of fasting blood glucose value is taken as an example for illustration.
  • the time series obtained after the difference processing is moved and smoothed, that is, the series points that fluctuate greatly (the difference with the average value of each point is greater than a certain threshold) in the time series, and then use the average value of each point (here the average Arithmetic mean, geometric mean, harmonic mean, etc.) can be used to replace the value at this point.
  • the time series of fasting blood glucose values can be transformed from a non-stationary series to a stable series.
  • the description is divided into two prediction methods, one is to predict the time series of future fasting blood glucose values (that is, the process shown in steps 203a to 204a), and the other is Predict the time sequence of the blood glucose level two hours after a meal (that is, the process shown in steps 203b to 204b).
  • the process of using the time series data of the fasting blood glucose value in the sample user data after preprocessing as the label information Y1 may specifically include: using the stable sequence of the preprocessed fasting blood glucose value as Label information Y1.
  • the created first model training set contains each feature information X and each corresponding label information Y1. That is, the time series of fasting blood glucose values corresponding to sample users with different diabetes follow-up characteristics.
  • the time series algorithm can be obtained by the fusion of ARIMA and LSTM.
  • the evaluation of the first recognition model uses the average absolute percentage error (MAPE) index. When the MAPE index value corresponding to the first recognition model is less than the preset standard comparison threshold, it is determined The first recognition model meets the evaluation criteria.
  • the MAPE indicator is used to evaluate the error between the predicted value of the model and the true value. It also considers the ratio between the error and the true value.
  • the calculation formula is:
  • N is the total number of samples
  • X is the measured value
  • Y is the simulated value.
  • the standard comparison threshold can be set according to the actual situation. When the MAPE is less than the standard comparison threshold, it means that the first recognition model meets the evaluation standard. Prediction through the recognition model that meets the evaluation criteria can ensure the accuracy of the prediction results.
  • the first mapping relationship between the feature information X and the label information Y1 can be determined by the first recognition model that meets the evaluation standard.
  • the process specifically includes: (1) Using a random sampling method to obtain the first training sample set and the second training sample set from the first model training set. For example, n training samples are randomly selected from the first model training set, and two rounds of extraction are performed to obtain two training sets. (The two training sets are independent of each other and the elements can be repeated); (2) Based on the first training sample set, use the ARIMA algorithm to train the first classifier; based on the second training sample set use the LSTM algorithm to train the second Classifiers; where each training sample set contains different feature information X and their corresponding label information Y1.
  • the training process of these two classifiers can be obtained based on the training of their corresponding model training algorithms, and the obtained Both classifiers can independently predict the fasting blood glucose value time series, that is, input the characteristic data of the user to be tested (specific content corresponds to the characteristic information X), and find the corresponding label information Y1 through the classifier.
  • represents the coefficient of autoregressive AR
  • represents the coefficient of moving average MA
  • is the set offset value (such as taking the average of the difference between two adjacent points).
  • LSTM is equivalent to the threshold recurrent neural network RNN.
  • the ingenuity of LSTM is that by increasing the input threshold, the forgetting threshold and the output threshold, the weight of the self-loop is changed, so that in the model
  • the integration scale at different times can be dynamically changed, thereby avoiding the problem of gradient disappearance or gradient expansion.
  • the first classifier and the second classifier are fused by the method of calculating the average time series of the two as the first recognition model obtained after a round of training.
  • the results obtained by the two classifiers are the time series Ai ⁇ A1, A2...An ⁇ and the time series Bi ⁇ B1, B2....Bn ⁇ .
  • the prediction result of the recognition model is ⁇ (A1+B1)/2, (A2+B2)/2,...(An+Bn)/2 ⁇ .
  • the steps of dividing the first model training set and continuing to train the first recognition model that does not meet the standard are performed again until the first recognition model meets the standard. For example, re-divide the first model training set to obtain two new training sample sets, and use these two new training sample sets to continue training the two classifiers that have been trained, and then evaluate the two newly trained classes If the MAPE index value of the first recognition model obtained by the processor still does not meet the evaluation standard, then the first model training set is divided again, and the above process is repeated until the first recognition model meets the evaluation standard.
  • Step 203b which is parallel to step 203a, uses the time series data of the blood glucose level two hours after the meal in the preprocessed sample user data as label information Y2, and combines the feature information X to create a second model training set.
  • the time series data of the blood glucose level two hours after the meal in the preprocessed sample user data is used as the label information Y2, which specifically includes: stabilizing the preprocessed blood glucose level two hours after the meal The sequence is used as label information Y2.
  • step 204b Use the second model training set combined with the time sequence algorithm to train to obtain a second recognition model.
  • the evaluation of the second recognition model uses the MAPE index.
  • the MAPE index value corresponding to the second recognition model is less than the predetermined standard comparison threshold, it is determined that the second recognition model meets the evaluation standard, and the feature can be determined by the second recognition model that meets the evaluation standard.
  • step 204b may specifically include: obtaining the third training sample set and the fourth training sample set from the second model training set by random sampling; and using the ARIMA algorithm based on the third training sample set to obtain the first training sample set.
  • step 205 may specifically include: preprocessing the historical user data of the target user to obtain a stable sequence of historical fasting blood glucose values of the target user, a stable sequence of historical blood glucose values two hours after a meal, and historical medication data , Historical symptom data, historical physical sign data, historical lifestyle data; input historical medication data, historical symptom data, historical physical sign data, and historical lifestyle data into the first recognition model for similarity matching with feature information X, and use the first A mapping relationship, determining the target tag information Y1 corresponding to the feature information X with the similarity greater than a preset threshold and the highest similarity; comparing the first target time series corresponding to the target tag information Y1 with the historical fasting blood glucose value
  • the stable sequence performs approximate matching, and it is determined that the stable sequence of the historical fasting blood glucose value matches the first target position in the first target time series, so as to use the fasting blood glucose value after the first target position in the first target time series Time series, to determine the first prediction sequence of the fasting blood glucose level of the target user
  • the time series data of fasting blood glucose and two-hour postprandial blood glucose in the target user data within the preset historical time period are cleaned Preprocessing and interpolation to obtain the target user's daily granular blood glucose sequence within the preset historical time period, and structure the target user data to facilitate the blood glucose prediction model combined with the target user's historical data to perform the blood glucose level Prediction, in order to make the historical blood glucose sequence trend of the target user more obvious, the historical blood glucose sequence is converted into a stable sequence for output, and the data is differentiated and moved smoothly to facilitate accurate analysis of the target user’s first prediction sequence and second prediction sequence .
  • the prediction model inputs the diabetes follow-up data such as medication, symptoms, signs, and lifestyle of the target user in the last 3 months, it is equivalent to inputting these data into the two classifiers in step 204a above, and each classifier
  • the corresponding feature information is matched by similarity, and the most similar feature information that is greater than a certain threshold is found, and then the corresponding fasting blood glucose value time series are obtained through the two classifiers, and then the final result is obtained by adding and averaging
  • the fasting blood glucose value time series B of the target user is equivalent to the fasting blood glucose level fluctuations of the sample user with similar characteristics as the target user in the last 1 year.
  • the three-month fasting blood glucose time series A of the target user is compared with the sample user Match the fasting blood glucose value time series B within this year to find the position of sequence A in sequence B, and then use the sequence data after this position as the time series of the target user’s future fasting blood sugar value, which represents the future fasting of the target user Changes in blood sugar levels.
  • the target user's blood glucose level change in the future two hours after a meal can be obtained. Time series of blood glucose values two hours after a meal.
  • further analysis can be made to obtain the influencing factors that affect the future blood glucose change, that is, the processes shown in steps 206 to 208 are performed.
  • step 207 Perform correlation analysis and calculation between the medication data, symptom data, physical sign data, and lifestyle data of each follow-up and the fasting blood glucose value of each follow-up, and determine the influence of the blood glucose in the first prediction sequence based on the obtained correlation analysis and calculation results Factors affecting indicator changes.
  • it further includes: obtaining a medication score value sequence composed of medications used in each follow-up, a symptom score value sequence composed of symptoms in each follow-up, a physical sign score value sequence composed of signs in each follow-up, and
  • Follow-up lifestyle consisted of a series of daily routine score values, a sequence of fasting blood glucose values measured at each follow-up, and a sequence of blood glucose values two hours after a meal.
  • the score can be based on the medication used at each diabetes follow-up (such as the greater the dose of the drug, the higher the score), and the score based on the symptoms that appear during each diabetes follow-up (such as the more symptoms, the greater the score) High, the score is higher when abnormal symptoms occur, etc.), the score is based on the signs that appear during each diabetes follow-up (such as the more abnormal the signs, the higher the score), and the score is based on the lifestyle conditions recorded during each diabetes follow-up (such as life The more abnormal the schedule, the higher the score).
  • step 207 may specifically include: calculating the first absolute value of the correlation coefficient between the medication score value sequence, the symptom score value sequence, the physical sign score value sequence, the daily routine score value sequence and the fasting blood glucose value sequence;
  • the value sequence, symptom score value sequence, physical sign score value sequence, and life schedule score value sequence obtain one or more of the highest first absolute values to determine the key influencing factors that affect the change of blood glucose indicators in the first prediction sequence.
  • the available calculation formula is:
  • ⁇ X, Y are the correlation coefficients between the X series and the Y series
  • ⁇ X is the standard deviation of the X series
  • ⁇ X is the average of the X series
  • ⁇ Y is the standard deviation of the Y series
  • ⁇ Y is The average of the Y series
  • E() is used to calculate expectations.
  • the absolute value of the correlation coefficient between the medication score value sequence and the fasting blood glucose value sequence is the highest, then the medication use of the target user is a key influencing factor that affects the future changes of the fasting blood glucose value.
  • step 208 may specifically include: calculating the medication score value sequence, the symptom score value sequence, the physical sign score value sequence, the daily routine score value sequence and the blood glucose value sequence two hours after a meal.
  • the second absolute value of the correlation coefficient between the two; the second absolute value is determined according to the medication score value sequence, the symptom score value sequence, the physical sign score value sequence, and the daily routine score value sequence to determine the influence of the second prediction
  • the key influencing factors of blood glucose index changes in the sequence.
  • the above calculation formula can be used to calculate the correlation coefficient between the medication score value sequence, the symptom score value sequence, the physical sign score value sequence, the daily routine score value sequence and the blood glucose value sequence two hours after a meal.
  • the absolute value of the correlation coefficient between the daily routine score value sequence and the blood glucose value sequence two hours after a meal is the highest, then the lifestyle of the target user is a key influencing factor that affects the change of blood glucose value two hours after a meal in the future.
  • the future blood glucose level change rule of the target user can be well predicted, and then the development of diabetes can be monitored in real time, and the corresponding supporting treatment can be carried out. It has great significance in the treatment and control of diabetes. Significance; and it can also determine the important factors that affect future changes in blood glucose indicators, which can make the diagnosis results more complete and cover more comprehensive types, which can provide patients and doctors with better treatment basis.
  • an embodiment of the present application provides a device for predicting future blood glucose levels. As shown in FIG. 3, the device includes: an extraction module 31, a creation module 32, and a determination Module 33.
  • the extraction module 31 can be used to extract sample user data corresponding to sample users whose diabetes follow-up records are greater than a preset duration threshold from the original health files and electronic medical records; the creation module 32 can be used to use the sample user data to create a time series of blood glucose values Prediction model; determination module 33, which can be used to input historical user data of the target user into the prediction model to predict the first prediction sequence of the fasting blood glucose level of the target user in a preset future time period, and two hours after a meal The second prediction sequence of the blood glucose level, wherein the historical duration corresponding to the historical user data is less than the preset duration threshold, and the target user and the sample user are both diabetic patients.
  • the creation module 32 may specifically include: a processing unit 321, a creation unit 322, and a training unit 323.
  • the processing unit 321 can be used to preprocess the sample user data; the creation unit 322 can be used to use the preprocessed time series data of fasting blood glucose values in the sample user data as label information Y1, and divide the sample user data Target feature data other than fasting blood glucose values and blood glucose values two hours after a meal are used as feature information X to create a first model training set, where the target feature data includes medication data, symptom data, physical sign data, and lifestyle data; training unit 323.
  • the first mapping relationship between the feature information X and the tag information Y1 can be determined; the creating unit 322 can also be used to use the preprocessed sample user data in the time series data of the blood glucose level two hours after a meal as tag information Y2, combined with the feature information X to create a second model training set; the training unit 323 can also be used to use the second model training set combined with the time series algorithm to train to obtain a second recognition model, where The evaluation of the second recognition model adopts the MAPE index.
  • the recognition model can determine the second mapping relationship between the feature information X and the tag information Y2.
  • the processing unit 321 can be specifically used to extract the time series data of fasting blood glucose values and the time series data of blood glucose values two hours after a meal from the sample user data using regular expressions; Time series data or the time series data of blood glucose two hours after a meal.
  • the mode of the multiple blood glucose measurement values is used as the Day’s corresponding blood glucose measurement value; and if the fasting blood glucose value or the two-hour postprandial blood glucose value on the date when the blood glucose measurement data is vacant within the preset historical time period, interpolation is used to supplement the date corresponding
  • the creation unit 322 can be specifically used to convert the preprocessed fasting blood glucose value
  • the stable sequence of is used as label information Y1; the creating unit 322 can be specifically used to use the preprocessed stable sequence of the blood glucose level two hours after a meal as label information Y2.
  • the determining module 33 can be specifically used to preprocess the historical user data of the target user to obtain the stable sequence of the historical fasting blood glucose level of the target user, and the stability of the historical blood glucose level two hours after a meal. Sequence, historical medication data, historical symptom data, historical physical sign data, and historical lifestyle data; input the historical medication data, the historical symptom data, the historical physical sign data, and the historical lifestyle data into the first Perform similarity matching with the feature information X in the recognition model, and use the first mapping relationship to determine the target tag information Y1 corresponding to the feature information X with the similarity greater than a preset threshold and the highest similarity; The first target time sequence corresponding to the target tag information Y1 is approximately matched with the stable sequence of the historical fasting blood glucose value, and it is determined that the stable sequence of the historical fasting blood glucose value matches the first target time sequence in the first target time sequence.
  • Target location so as to use the fasting blood glucose value time sequence after the first target location in the first target time sequence to determine the first prediction sequence of the fasting blood glucose value of the target user in a preset future time period;
  • the historical medication data, the historical symptom data, the historical physical sign data, and the historical lifestyle data are input into the second recognition model to perform similarity matching with the characteristic information X, and use the second Mapping relationship, determine the target tag information Y2 corresponding to the feature information X with the similarity greater than the preset threshold and the highest similarity; compare the second target time series corresponding to the target tag information Y2 with the historical post-dinner two Approximate matching is performed on the stable sequence of hourly blood glucose level, and it is determined that the stable sequence of blood glucose level two hours after the meal matches the second target position in the second target time series, so as to use the second target time series in the second target time series.
  • the two-hour postprandial blood glucose level time sequence after the second target location is used to determine the second predicted sequence of the target user’s two-hour postprandial blood glucose level within a preset future time period.
  • the training unit 323 can be specifically used to obtain a first training sample set and a second training sample set from the first model training set in a random sampling manner; based on the first training sample set, use ARIMA Algorithm, the first classifier is obtained by training; the second classifier is obtained by training using the LSTM algorithm based on the second training sample set; the first classifier and the second classifier are used to obtain the average time series of the two
  • the first recognition model obtained after one round of training is used as the first recognition model obtained after one round of training; if it is determined according to the MAPE index value that the first recognition model obtained after one round of training does not meet the evaluation criteria, the first model training is re-executed And continue to train the first recognition model that does not meet the standard until the first recognition model meets the standard; the training unit
  • Three training sample sets a fourth training sample set; based on the third training sample set using the ARIMA algorithm to train a third classifier; based on the fourth training sample set using the LSTM algorithm to train a fourth classifier;
  • the third classifier and the fourth classifier perform fusion processing by averaging the time series of the two as the second recognition model obtained after one round of training; if the MAPE index value is used to determine the after one round of training If the obtained second recognition model does not meet the evaluation standard, the steps of dividing the second model training set and continuing to train the second recognition model that does not meet the standard are executed again until the second recognition model meets the standard.
  • the device further includes: a calculation module 34; an extraction module 31, which can also be used to obtain medication data and symptom data of the target user during each follow-up visit from the historical user data , Physical sign data, lifestyle data, fasting blood glucose value, blood glucose value two hours after a meal; the calculation module 34 can be used to compare the medication data, symptom data, physical sign data, and lifestyle data of each follow-up Perform correlation analysis and calculation between fasting blood glucose values, and determine the factors that affect the changes in blood glucose indicators in the first prediction sequence based on the obtained correlation analysis and calculation results; and compare the medication data, symptom data, and physical sign data of each follow-up visit.
  • the extraction module 31 can also be used to obtain the medication scoring value sequence composed of medications used in each follow-up, the symptom score value sequence composed of symptoms in each follow-up, the physical sign score value sequence composed of signs in each follow-up, and The life schedule score value sequence composed of follow-up lifestyle, the fasting blood glucose value sequence measured at each follow-up and the blood glucose value sequence two hours after a meal; the calculation module 34 can be specifically used to calculate the medication score value sequence and the symptom score value sequence , The first absolute value of the correlation coefficient between the physical sign score value sequence, the daily routine score value sequence and the fasting blood glucose value sequence; according to the medication score value sequence, the symptom score value sequence, and the Obtain one or more of the highest first absolute value in the physical sign score value sequence and the life schedule score value sequence, and
  • the second absolute value according to the medication score value sequence, the symptom score value sequence, the physical sign score value sequence, and the daily routine score value sequence to obtain one or more of the second absolute value highest, Determine the key influencing factors that affect the change of the blood glucose index in the second prediction sequence.
  • the software product can be stored in a non-volatile storage medium (which can be a CD-ROM, U disk, mobile hard disk, etc.), including several
  • the instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute the methods in each implementation scenario of the present application.
  • an embodiment of the present application also provides a computer device, which may be a personal computer, Servers, network devices, etc.
  • the physical device includes a storage medium and a processor; the storage medium is used to store computer-readable instructions; the processor is used to execute the computer-readable instructions to realize the prediction shown in Figure 1 and Figure 2 above Methods of future blood sugar levels.
  • the computer device may also include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, and so on.
  • RF radio frequency
  • the user interface may include a display screen (Display), an input unit such as a keyboard (Keyboard), etc., and the optional user interface may also include a USB interface, a card reader interface, and the like.
  • the network interface can optionally include a standard wired interface, a wireless interface (such as a Bluetooth interface, a WI-FI interface), etc.
  • the non-volatile readable storage medium may also include an operating system and a network communication module.
  • the operating system is a program that manages the hardware and software resources of physical devices that predict future blood glucose levels, and supports the operation of information processing programs and other software and/or programs.
  • the network communication module is used to implement communication between various components in the non-volatile readable storage medium and communication with other hardware and software in the physical device.
  • this application can better predict the future blood glucose level change rule of the target user, and then can monitor the development of diabetes in real time, and carry out corresponding supporting treatments.
  • the treatment and control of diabetes is of great significance; and it can also determine important factors that affect future changes in blood glucose indicators, which can make the diagnosis results more complete and cover more comprehensive types, so as to provide patients and doctors with better treatment basis.
  • modules in the device in the implementation scenario can be distributed in the device in the implementation scenario according to the description of the implementation scenario, or can be changed to be located in one or more devices different from the implementation scenario.
  • the modules of the above implementation scenarios can be combined into one module or further divided into multiple sub-modules.
  • the above serial number of this application is only for description, and does not represent the merits of implementation scenarios.
  • the above disclosures are only a few specific implementation scenarios of the application, but the application is not limited to these, and any changes that can be thought of by those skilled in the art should fall into the protection scope of the application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Health & Medical Sciences (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Economics (AREA)
  • Strategic Management (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Epidemiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Data Mining & Analysis (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Quality & Reliability (AREA)
  • Databases & Information Systems (AREA)
  • Game Theory and Decision Science (AREA)
  • Biomedical Technology (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Pathology (AREA)
  • Tourism & Hospitality (AREA)
  • Physics & Mathematics (AREA)
  • General Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Investigating Or Analysing Biological Materials (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)

Abstract

本申请公开了一种预测未来血糖值的方法、装置及计算机设备,涉及计算机技术领域,能有效解决现有技术中只能判断用户当前的血糖值,却无法对用户未来血糖值进行预测的问题。其中方法包括:从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据;利用所述样本用户数据创建血糖值时间序列的预测模型;将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,其中,所述历史用户数据对应的历史时长小于所述预设时长阈值,所述目标用户和所述样本用户均为糖尿病患者。本申请适用于对目标用户未来血糖值的预测。

Description

预测未来血糖值的方法、装置及计算机设备
本申请要求与2019年3月12日提交中国专利局、申请号为2019101850580、申请名称为“预测未来血糖值的方法、装置及计算机设备”的中国专利申请的优先权,其全部内容通过引用结合在申请中。
技术领域
本申请涉及计算机技术领域,尤其涉及到一种预测未来血糖值的方法、装置及计算机设备。
背景技术
糖尿病是一组以高血糖为特征的代谢性疾病,发病时会导致大血管、微血管受损并危及心、脑、肾、周围神经、眼睛、足等多个部位,还会伴有多种并发症,故加强糖尿病的预测工作是完全必要的。然而随着科技的进步,病种的诊断已经不局限于医生的分析,利用人工智能来预测糖尿病,才是符合如今的发展潮流。
发明人发现目前业内对于糖尿病预测的常见方法是通过收集糖尿病医案,将糖尿病患者数据与健康人群数据进行对比,构建0-1分类模型,通过患者的各类特征维度数据,判断用户是否患糖尿病。然而现有的糖尿病预测的方法只能判断没有糖尿病相关医案的用户是否患糖尿病,但针对已有数据且长期随访的用户,却不能进行未来血糖值的预测,不能更好的实现对糖尿病的诊断和治疗。
发明内容
有鉴于此,本申请提供了一种预测未来血糖值的方法、装置及计算机设备,主要目的在于解决当下只能判断出用户是否患糖尿病,却无法预测出未来血糖值,进而不能更好的实现对糖尿病跟踪治疗的问题。根据本申请的一个方面,提供了一种预测未来血糖值的方法,该方法包括:从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据;利用所述样本用户数据创建血糖值时间序列的预测模型;将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,其中,所述历史用户数据对应的历史时长小于所述预设时长阈值,所述目标用户和所述样本用户均为糖尿病患者。
根据本申请的另一个方面,提供了一种预测未来血糖值的装置,该装置包括:提取模块,用于从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据;创建模块,用于利用所述样本用户数据创建血糖值时间序列的预测模型;确定模块,用于将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,其中,所述历史用户数据对应的历史时长小于所述预设时长阈值,所述目标用户和所述样本用户均为糖尿病患者。
根据本申请的又一个方面,提供了一种非易失性可读存储介质,其上存储有计算机可读指令,所述计算机可读指令被处理器执行时实现上述预测未来血糖值的方法。根据本申请的再一个方面,提供了一种计算机设备,包括非易失性可读存储介质、处理器及存储在非易失性可读存储介质上并可在处理器上运行的计算机可读指令,所述处理器执行所述计算机可读指令时实现上述预测未来血糖值的方法。
借由上述技术方案,本申请提供的一种预测未来血糖值的方法、装置及计算机设备,与目前利用构建的0-1分类模型预测糖尿病的方法相比,本申请是依据从原始健康档案和电子病历中提取的样本用户数据,来创建血糖值时间序列的预测模型,将目标用户的历史用户数据输入到预测模型中,进而利用预测模型获取目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后预设时长血糖值的第二预测序列,由于样本用户和目标用户均为糖尿病患者,因此可利用预测模型确定与目标用户具有相同或类似用户数据的样本用户,其对应的空腹血糖值时间序列和餐后两小时血糖值时间序列。并且由于样本用户的糖尿病随访记录大于预设时长阈值,而目标用户的历史用户数据对应的历史时长小于该阈值,相当于利用大范围的时间序列判别其包含的小范围时间序列走向,进而可很好的预测出目标用户的未来血糖值变化规律,进而可实时监测糖尿病的发展情况,进行相应的配套治疗,在对糖尿病的治疗控制上具有很大的意义。上述说明仅是本申请技术方案的概述,为了能够更清楚了解本申请的技术手段,而可依照说明书的内容予以实施,并且为了本申请的上述和其他目的、特征和优点能够更明显易懂,以下特举本申请的具体实施方式。
附图说明
此处所说明的附图用来提供对本申请的进一步理解,构成本申请的一部分,本申请的示意性实施例及其说明用于解释本申请,并不构成对本地申请的不当限定。在附图中:图1示出了本申请实施例提供的一种预测未来血糖值的方法的流程示意图;图2示出了本申请实施例提供的另一种预测未来血糖值的方法的流程示意图;图3示出了本申请实施例提供的一种预测未来血糖值的装置的结构示意图;图4示出了本申请实施例提供的另一种预测未来血糖值的装置的结构示意图。
具体实施方式
下文中将参考实施例并结合附图来详细说明本申请。需要说明的是,在不冲突的情况下,本申请中的实施例及实施例中的特征可以相互结合。针对目前只能判断出用户是否患糖尿病,却无法预测出未来血糖值,进而不能更好的实现对糖尿病跟踪治疗的问题,本实施例提供了一种预测未来血糖值的方法,如图1所示,该方法包括:
101、从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据。其中,样本用户为糖尿病患者。样本用户数据可包括患者就诊数据、体检指标数据、用药数据和健康告知数据等,就诊数据又包括门诊、住院和用药数据,患者的用药、症状、体征和生活方式等糖尿病随访数据主要来源于各大医院和体检中心。预设时长阈值可根据实际业务需求设定,例如,可设定时间阈值为一年,即获取糖尿病随访记录大于一年的样本用户对应的样本用户数据。在具体的应用场景中,从原始健康档案和电子病历中提取的样本用户数据中一般包含许多无用数据以及异常数据等,故在利用这些数据创建预测模型之前,需要对这些样本用户数据进行清洗整合从中提取出需要的相关特征,提取方式主要通过正则表达式,正则表达式即对字符串的一种逻辑公式,通过事先定义好的特定字符组合去匹配文本,获取需要的文本及数字。
102、利用样本用户数据创建血糖值时间序列的预测模型。在本实施例中,从样本用户数据中获取样本用户每次糖尿病随访时的空腹血糖值、餐后两小时血糖值、用药、症状、体征和生活方式等数据,然后基于这些数据创建血糖值时间序列的预测模型,具体可通过时间序列的相关算法训练得到。通过创建得到的血糖值时间序列的预测模型可以很好的反应出具有不同糖尿病随访特征的样本用户分别对应的空腹血糖值时间序列和餐后两小时血糖值时间序列。
103、将目标用户的历史用户数据输入到预测模型中,预测目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列。其中,需要进行未来糖尿病预测的目标用户为糖尿病患者,历史用户数据对应的历史时长小于预设时长阈值。在本实施例中,历史用户数据为目标用户最近一段时间的历史用户数据,例如,最近半年每次糖尿病随访时的空腹血糖值、餐后两小时血糖值、用药、症状、体征和生活方式等数据。预设未来时间段为根据实际需要进行设定的,使血糖值时间序列的预测模型最终显示出未来该时间段内的血糖值,例如,预设未来时间段可为未来30日;第一预测序列为目标用户在预设未来时间段内空腹血糖值的时间序列;第二预测序列为目标用户在预设未来时间段内餐后两小时血糖值的时间序列。通过上述预测未来血糖值的方法,依据从原始健康档案和电子病历中提取的样本用户数据,来创建血糖值时间序列的预测模型,将目标用户的历史用户数据输入到预测模型中,进而利用预测模型获取目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后预设时长血糖值的第二预测序列,由于样本用户和目标用户均为糖尿病患者,因此可利用预测模型确定与目标用户具有相同或类似用户数据的样本用户,其对应的空腹血糖值时间序列和餐后两小时血糖值时间序列。并且由于样本用户的糖尿病随访记录大于预设时长阈值,而目标用户的历史用户数据对应的历史时长小于该阈值,相当于利用大范围的时间序列判别其包含的小范围时间序列走向,进而可很好的预测出目标用户的未来血糖值变化规律,进而可实时监测糖尿病的发展情况,进行相应的配套治疗,在对糖尿病的治疗控制上具有很大的意义。
进一步的,作为上述实施例具体实施方式的细化和扩展,为了完整说明本申请实施例中的具体实施过程,提供了另一种预测未来血糖值的方法,如图2所示,该方法包括:201、从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据。
202、对样本用户数据进行预处理。由于样本用户数据中通常会存在无用数据以及异常数据等,因此需要样本用户数据进行预处理,获取得到预测模型创建需要的数据,为了说明此过程,作为一种可选方式,步骤202具体可包括:利用正则表达式从样本用户数据中提取空腹血糖值的时序数据和餐后两小时血糖值的时序数据(即对应每次随访时的时序数据);若空腹血糖值的时序数据或餐后两小时血糖的时序数据,在预设历史时间段的部分日期存在一天内具有多条血糖测量数据的情况,则使用多条血糖测量值中的众数作为该天对应的血糖测量值;及若空腹血糖值或餐后两小时血糖值在预设历史时间段内出现血糖测量数据空缺的日期,则使用内插法补充日期对应的空值,得到日颗粒度的血糖序列;最后对日颗粒度的血糖序列,进行差分化移动平滑处理,得到稳定序列。利用这种相对平稳的稳定序列训练模型,可消除噪音,得到的预测模型可预测更加精确的结果。例如,获取过去一年中样本用户a每次随访时的空腹血糖值的时序数据和餐后两小时血糖值的时序数据,如果存在某一天内具有5条空腹血糖测量值的情况,分别为5.7mmol/L、5.7mmol/L、6.7mmol/L、5.9mmol/L、6.1mmol/L,因在这些空腹血糖测量值中5.7mmol/L为众数,故可用众数5.7mmol/L代替那一天的多次取值;如果存在某一天的餐后两小时血糖值空缺,则利用内插法补充空值,即使用相邻几天对应的餐后两小时血糖值,通过内插法公式计算得到(如已知(X1,Y1)、(X2,Y2),X为X1和X2中间点,那么X对应的Y=Y1+(Y2-Y1)*(X-X1)/ (X2-X1)。
通过上述预处理过程使得在过去一年中样本用户a每天都对应一个空腹血糖值和一个餐后两小时血糖值,即两个初始的血糖值时间序列;为了使这两个血糖值时间序列趋势性更加明显,可将这两个血糖值时间序列转化为平稳序列,具体可对这两个血糖值时间序列进行差分以及移动平滑处理。两个初始的血糖值时间序列的差分以及移动平滑处理过程相同,这里以空腹血糖值的时间序列变换为例进行说明,设空腹血糖值的时间序列为y t=f(t),t为每次糖尿病随访的时间,y t为随访时对应的空腹血糖值,一阶差分定义为Δy t=y t+1-y t=f(t+1)-f(t),k阶差分定义为
Figure PCTCN2019117222-appb-000001
求解带有k阶差分、且带有未知函数f(t)的差分方程,然后将求得的f(t)带入差分公式中,实现时间序列的差分处理过程。然后对差分处理后得到的时间序列进行移动平滑处理,即获取该时间序列中波动较大的(与各点平均值的差值大于一定阈值)序列点,然后利用各点平均值(这里平均值可利用算术平均数、几何平均数、调和平均数等)代替该点的值。最终通过这样的处理,可使得空腹血糖值的时间序列从非平稳序列转化为稳定序列。
在本实施例中,在对样本用户数据进行预处理后,分两种预测方式进行说明,一种是预测未来空腹血糖值的时间序列(即步骤203a至204a所示过程),另一种是预测餐后两小时血糖值的时间序列(即步骤203b至204b所示过程)。
203a、将预处理后的样本用户数据中空腹血糖值的时序数据作为标签信息Y1,将样本用户数据中除空腹血糖值和餐后两小时血糖值以外的目标特征数据作为特征信息X,创建第一模型训练集。其中,目标特征数据包括用药数据、症状数据、体征数据、生活方式数据。结合步骤202中稳定序列转化的可选方式,将预处理后的样本用户数据中空腹血糖值的时序数据作为标签信息Y1的过程,具体可包括:将预处理后的空腹血糖值的稳定序列作为标签信息Y1。创建得到的第一模型训练集中包含各个特征信息X,以及各自对应的标签信息Y1。即不同糖尿病随访特征的样本用户分别对应的空腹血糖值时间序列。
204a、利用第一模型训练集并结合时间序列算法,训练得到第一识别模型。其中,时间序列算法可由ARIMA和LSTM两种算法融合得到,第一识别模型的评估采用平均绝对百分比误差(MAPE)指标,当第一识别模型对应的MAPE指标值小于预置标准比较阈值时,确定第一识别模型符合评估标准。MAPE指标用于评估模型预测值和真实值之间的误差,还考虑了误差与真实值之间的比例,其计算公式为:
Figure PCTCN2019117222-appb-000002
在上面公式中,N为样本总数,X为实测值,Y为模拟值。MAPE值越小,说明模型预测值和真实值之间的误差越小,在具体实施方式中,可根据实际情况设定标准比较阈值,当MAPE小于标准比较阈值时,说明第一识别模型符合评估标准。通过符合评估标准的识别模型进行预测,可保证预测结果的准确性。通过符合评估标准的第一识别模型可确定特征信息X和标签信息Y1之间的第一映射关系。为了说明上述第一识别模型的训练过程,作为一种可选方式,该过程具体包括:(1)采用随机采样方式从第一模型训练集中分别获取第一训练样本集、第二训练样本集,例如从第一模型训练集中随机抽取n个训练样本,共进行两轮抽取,得到两个训练集。(两个训练集之间相互独立,元素可以有重复);(2)基于第一训练样本集利用ARIMA算法,训练得到第一分类器;基于第二训练样本集利用LSTM算法,训练得到第二分类器;其中,每个训练样本集中都包含有不同的特征信息X,以及各自对应的标签信息Y1,这两种分类器的训练过程可基于各自对应的模型训练算法训练得到,且得到的这两种分类器都可单独进行空腹血糖值时间序列的预测,即输入待测用户的特征数据(具体内容对应特征信息X),通过分类器找到相对应的标签信息Y1。对于第一分类器的具体训练过程:1)设与特征信息Xi对应的空腹血糖值的时间序列Yi可用Yi={yi1,yi2,…,yik}来表示,其中1至k表示时间点,yit表示Yi第t个时间点的值。ARIMA(p,d,q)模型中有三个参数p,d,q,其中,p表示预测模型中采用的时序数据本身的滞后数;d表示时序数据需要进行几阶差分化才是稳定的;q表示预测模型中采用的预测误差的滞后数。2)通过步骤202中差分化处理的步骤确定d,而对空腹血糖值的时间序列分别求其自相关系数ACF和偏自相关系数PACF和偏自相关系数PACF;然后利用这两个系数分别生成自相关图和偏自相关图,然后从图中分析得到最佳的阶层p和阶数q。
3)基于上述p,d,q值,得到第一分类器的初始数学形式表示为:
Y i=μ+φ 1*f(X i)+...+φ p*f(X i-p)+θ 1*f(X i-1)+...+θ q*f(X i-q),其中,φ表示自回归AR的系数,θ表示移动平均MA的系数,μ为设定的偏移值(如取相邻两点差值的平均值)。
4)利用上述公式,带入第一训练样本集中的数据进行训练,求解f(Xi),作为第一分类器。
对于第二分类器的具体训练过程:LSTM相当于门限循环神经网络RNN,LSTM的巧妙之处在于通过增加输入门限,遗忘门限和输出门限,使得自循环的权重是变化的,这样一来在模型参数固定的情况下,不同时刻的积分尺度可以动态改变,从而避免了梯度消失或者梯度膨胀的问题。基于这一原理执行下列步骤:1)设与特征信息Xi对应的空腹血糖值的时间序列Yi可用Yi={yi1,yi2,…,yik}来表示,其中1至k表示时间点,yit表示Yi第t个时间点的值。2)利用公式Yi=A[(wf+wi)*f(Xi)/wo+(Bf+Bi)/Bo)],带入第二训练样本集中的数据进行训练,求解f(Xi),作为第二分类器。其中,A为设定的调整参数,wf为遗忘门限对应的权重,wi为输入门限对应的权重,wo为输出门限对应的权重,Bf为遗忘门限对应的补偿值,Bi为输入门限对应的补偿值,Bo为输出门限对应的补偿值。
(3)将第一分类器和第二分类器利用求二者平均时间序列的方式进行融合处理,作为经过一轮训练得到的第一识别模型。例如,两个分类器得到的结果分别为时间序列Ai{A1,A2….An},时间序列Bi{B1,B2….Bn},将Ai与Bi相加求平均,即最后得到的第一识别模型的预测结果为{(A1+B1)/2,(A2+B2)/2,…(An+Bn)/2}.进一步的,若根据MAPE指标值确定经过一轮训练得到的第一识别模型不符合评估标准,则重新执行划分第一模型训练集和继续训练不符合标准的第一识别模型的步骤,直至第一识别模型符合标准。例如重新划分第一模型训练集,得到两个新的训练样本集,并利用这两个新的训练样本集对已经训练得到的两个分类器继续进行训练,然后评估由两个新训练的分类器得到的第一识别模型的MAPE指标值,如果仍然不符合评估标准,则再重新执行划分第一模型训练集,重复上述过程,直至第一识别模型符合评估标准。
与步骤203a并列的步骤203b、将预处理后的样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,并结合特征信息X,创建第二模型训练集。结合步骤202中稳定序列转化的可选方式,将预处理后的样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,具体包括:将预处理后的餐后两小时血糖值的稳定序列作为标签信息Y2。
204b、利用第二模型训练集并结合时间序列算法,训练得到第二识别模型。其中,第二识别模型的评估采用MAPE指标,当第二识别模型对应的MAPE指标值小于预定标准比较阈值时,确定第二识别模型符合评估标准,通过符合评估标准的第二识别模型可确定特征信息X和标签信息Y2之间的第二映射关系。作为一种可选方式,步骤204b具体可包括:采用随机采样方式从第二模型训练集中分别获取第三训练样本集、第四训练样本集;基于第三训练样本集利用ARIMA算法,训练得到第三分类器;基于第四训练样本集利用LSTM算法,训练得到第四分类器;将第三分类器和第四分类器利用求二者平均时间序列的方式进行融合处理,作为经过一轮训练得到的第二识别模型;若根据MAPE指标值确定经过一轮训练得到的第二识别模型不符合评估标准,则重新执行划分第二模型训练集和继续训练不符合标准的第二识别模型的步骤,直至第二识别模型符合标准。上述可选方式的具体实现过程可参照步骤204a中的实现过程,在此不再赘述。
205、将目标用户的历史用户数据输入到预测模型中,预测目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列。
作为一种可选方式,步骤205具体可包括:将目标用户的历史用户数据经过预处理,得到目标用户的历史空腹血糖值的稳定序列、历史餐后两小时血糖值的稳定序列、历史用药数据、历史症状数据、历史体征数据、历史生活方式数据;将历史用药数据、历史症状数据、历史体征数据、历史生活方式数据输入到第一识别模型中与特征信息X进行相似度匹配,并利用第一映射关系,确定与相似度大于预设阈值、且相似度最高的所述特征信息X对应的目标标签信息Y1;将目标标签信息Y1相应的第一目标时间序列与所述历史空腹血糖值的稳定序列进行近似匹配,判定历史空腹血糖值的稳定序列匹配到第一目标时间序列中的第一目标位置,以便利用所述第一目标时间序列中在所述第一目标位置后的空腹血糖值时间序列,确定目标用户在预设未来时间段内空腹血糖值的第一预测序列;将历史用药数据、历史症状数据、历史体征数据、历史生活方式数据输入到第二识别模型中与特征信息X进行相似度匹配,并利用第二映射关系,确定与相似度大于预设阈值、且相似度最高的特征信息X对应的目标标签信息Y2;将目标标签信息Y2相应的第二目标时间序列与历史餐后两小时血糖值的稳定序列进行近似匹配,判定餐后两小时血糖值的稳定序列匹配到第二目标时间序列中的第二目标位置,以便利用第二目标时间序列中在第二目标位置后的餐后两小时血糖值时间序列,确定目标用户在预设未来时间段内餐后两小时血糖值的第二预测序列。其中,预定阈值可根据实际需求预先设置。例如,预定阈值设置得越大,相应的特征匹配精度越高,如果相似度为100%,则说明特征完全匹配。
通过上述获取空腹血糖的第一预测序列和餐后预设时长血糖的第二预测序列的方法,对预设历史时间段内的目标用户数据中空腹血糖和餐后两小时血糖的时序数据进行清洗和内插的预处理,以得到目标用户对应的预设历史时间段内日颗粒度的目标血糖序列,将目标用户数据结构化处理,方便血糖值预测模型结合目标用户的历史数据进行血糖值的预测,为了使目标用户的历史血糖序列趋势性更加明显,将历史血糖序列转化为平稳序列输出,对数据进行差分以及移动平滑处理,便于准确分析出目标用户的第一预测序列和第二预测序列。例如,在预测模型输入目标用户最近3个月的用药、症状、体征和生活方式等糖尿病随访数据后,相当于将这些数据分别输入到上述步骤204a的两个分类器中,并与分类器各自对应的特征信息进行相似度匹配,分别找到最相似且大于一定阈值的特征信息,进而通过这两个分类器分别求出各自对应的空腹血糖值时间序列,然后通过相加求平均的方式得到最终的空腹血糖值时间序列B,该序列B相当于与目标用户具有相似特征的样本用户在最近1年内的空腹血糖值波动情况,将目标用户这3个月的空腹血糖值时间序列A与样本用户这1年内的空腹血糖值时间序列B进行匹配,找到序列A在序列B中的位置,然后将该位置以后的序列数据作为目标用户的未来空腹血糖值的时间序列,即代表目标用户的未来空腹血糖值的变化情况。利用相同方法,将目标用户的用药、症状、体征和生活方式等糖尿病随访数据输入到上述步骤204b的两个分类器中,最后可求出代表目标用户的未来餐后两小时血糖值变化情况的餐后两小时血糖值时间序列。在本实施例中,除了计算出目标用户的未来血糖值变化情况以外,还可进一步分析得到影响未来血糖变化的影响因素,即执行步骤206至208所示过程。
206、从目标用户的历史用户数据中获取目标用户每次随访的用药数据、症状数据、体征数据、生活方式数据、空腹血糖值、餐后两小时血糖值。
207、将每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的空腹血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响第一预测序列中血糖指标变化的影响因素。作为一种可选方式,在步骤207之前还包括:获取每次随访用药组成的用药评分值序列、每次随访症状组成的症状评分值序列、每次随访体征组成的体征评分值序列、每次随访生活方式组成的生活作息评分值序列、每次随访测量的空腹血糖值序列和餐后两小时血糖值序列。
在本可选方式中,可依据每次糖尿病随访时的用药情况进行评分(如药剂量越大评分越高),依据每次糖尿病随访时出现的症状情况进行评分(如症状数越多评分越高、出现异常症状时评分也较高等),依据每次糖尿病随访时出现的体征情况进行评分(如体征越异常评分越高),依据每次糖尿病随访时记录的生活方式情况进行评分(如生活作息越异常评分越高)。相应的,步骤207具体可包括:计算用药评分值序列、症状评分值序列、体征评分值序列、生活作息评分值序列分别与空腹血糖值序列之间的相关系数的第一绝对值;依据用药评分值序列、症状评分值序列、体征评分值序列、生活作息评分值序列中获取第一绝对值最高的一项或多项,确定影响第一预测序列中血糖指标变化的关键影响因素。
例如,可利用计算公式为:
Figure PCTCN2019117222-appb-000003
计算相关系数。其中,ρ X,Y为X序列与Y序列二者之间的相关系数,σ X为X序列的标准差,μ X为X序列的平均值,σ Y为Y序列的标准差,μ Y为Y序列的平均值,E()用于计算期望。如用药评分值序列与空腹血糖值序列之间的相关系数的绝对值最高,那么目标用户的用药情况是影响未来空腹血糖值变化的关键影响因素。
208、将每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的餐后两小时血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第二预测序列中血糖指标变化的影响因素。
基于步骤207中的可选方式过程,相应的,步骤208具体可包括:计算用药评分值序列、症状评分值序列、体征评分值序列、生活作息评分值序列分别与餐后两小时血糖值序列之间的相关系数的第二绝对值;依据用药评分值序列、症状评分值序列、体征评分值序列、生活作息评分值序列中获取第二绝对值最高的一项或多项,确定影响第二预测序列中血糖指标变化的关键影响因素。例如,可利用上述计算公式,计算用药评分值序列、症状评分值序列、体征评分值序列、生活作息评分值序列分别与餐后两小时血糖值序列之间的相关系数。如用生活作息评分值序列与餐后两小时血糖值序列之间的相关系数的绝对值最高,那么目标用户的生活方式情况是影响未来餐后两小时血糖值变化的关键影响因素。通过上述预测未来血糖值的方法,可很好的预测出目标用户的未来血糖值变化规律,进而可实时监测糖尿病的发展情况,进行相应的配套治疗,在对糖尿病的治疗控制上具有很大的意义;并且还可判断出影响未来血糖指标变化的重要因素,可使诊断结果更加完善,覆盖类型更全面,从而能为患者及医生提供更好的治疗 依据。进一步的,作为图1和图2所示方法的具体体现,本申请实施例提供了一种预测未来血糖值的装置,如图3所示,该装置包括:提取模块31、创建模块32、确定模块33。
提取模块31,可用于从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据;创建模块32,可用于利用所述样本用户数据创建血糖值时间序列的预测模型;确定模块33,可用于将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,其中,所述历史用户数据对应的历史时长小于所述预设时长阈值,所述目标用户和所述样本用户均为糖尿病患者。
在具体的应用场景中,为了创建出血糖值时间序列的预测模型,如图4所示,创建模块32,具体可包括:处理单元321、创建单元322、训练单元323。
处理单元321,可用于对样本用户数据进行预处理;创建单元322,可用于将预处理后的所述样本用户数据中空腹血糖值的时序数据作为标签信息Y1,将所述样本用户数据中除空腹血糖值和餐后两小时血糖值以外的目标特征数据作为特征信息X,创建第一模型训练集,其中,所述目标特征数据包括用药数据、症状数据、体征数据、生活方式数据;训练单元323,可用于利用所述第一模型训练集并结合时间序列算法,训练得到第一识别模型,其中,所述时间序列算法由ARIMA和LSTM两种算法融合得到,所述第一识别模型的评估采用平均绝对百分比误差MAPE指标,当所述第一识别模型对应的MAPE指标值小于预置标准比较阈值时,确定所述第一识别模型符合评估标准,通过符合评估标准的所述第一识别模型可确定所述特征信息X和所述标签信息Y1之间的第一映射关系;创建单元322,还可用于将预处理后的所述样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,并结合所述特征信息X,创建第二模型训练集;训练单元323,还可用于利用所述第二模型训练集并结合所述时间序列算法,训练得到第二识别模型,其中,所述第二识别模型的评估采用MAPE指标,当所述第二识别模型对应的MAPE指标值小于预定标准比较阈值时,确定所述第二识别模型符合评估标准,通过符合评估标准的所述第二识别模型可确定所述特征信息X和所述标签信息Y2之间的第二映射关系。
在具体的应用场景中,处理单元321,具体可用于利用正则表达式从所述样本用户数据中提取空腹血糖值的时序数据和餐后两小时血糖值的时序数据;若所述空腹血糖值的时序数据或所述餐后两小时血糖的时序数据,在预设历史时间段的部分日期存在一天内具有多条血糖测量数据的情况,则使用所述多条血糖测量值中的众数作为该天对应的血糖测量值;及若所述空腹血糖值或所述餐后两小时血糖值在所述预设历史时间段内出现血糖测量数据空缺的日期,则使用内插法补充所述日期对应的空值,得到日颗粒度的血糖序列;对所述日颗粒度的血糖序列,进行差分化移动平滑处理,得到稳定序列;创建单元322,具体可用于将预处理后的所述空腹血糖值的稳定序列作为标签信息Y1;创建单元322,具体还可用于将预处理后的所述餐后两小时血糖值的稳定序列作为标签信息Y2。在具体的应用场景中,确定模块33,具体可用于将所述目标用户的历史用户数据经过预处理,得到所述目标用户的历史空腹血糖值的稳定序列、历史餐后两小时血糖值的稳定序列、历史用药数据、历史症状数据、历史体征数据、历史生活方式数据;将所述历史用药数据、所述历史症状数据、所述历史体征数据、所述历史生活方式数据输入到所述第一识别模型中与所述特征信息X进行相似度匹配,并利用所述第一映射关系,确定与相似度大于预设阈值、且相似度最高的所述特征信息X对应的目标标签信息Y1;将所述目标标签信息Y1相应的第一目标时间序列与所述历史空腹血糖值的稳定序列进行近似匹配,判定所述历史空腹血糖值的稳定序列匹配到所述第一目标时间序列中的第一目标位置,以便利用所述第一目标时间序列中在所述第一目标位置后的空腹血糖值时间序列,确定所述目标用户在预设未来时间段内空腹血糖值的第一预测序列;将所述历史用药数据、所述历史症状数据、所述历史体征数据、所述历史生活方式数据输入到所述第二识别模型中与所述特征信息X进行相似度匹配,并利用所述第二映射关系,确定与相似度大于预设阈值、且相似度最高的所述特征信息X对应的目标标签信息Y2;将所述目标标签信息Y2相应的第二目标时间序列与所述历史餐后两小时血糖值的稳定序列进行近似匹配,判定所述餐后两小时血糖值的稳定序列匹配到所述第二目标时间序列中的第二目标位置,以便利用所述第二目标时间序列中在所述第二目标位置后的餐后两小时血糖值时间序列,确定所述目标用户在预设未来时间段内餐后两小时血糖值的第二预测序列。在具体的应用场景中,训练单元323,具体可用于采用随机采样方式从所述第一模型训练集中分别获取第一训练样本集、第二训练样本集;基于所述第一训练样本集利用ARIMA算法,训练得到第一分类器;基于所述第二训练样本集利用LSTM算法,训练得到第二分类器;将所述第一分类器和所述第二分类器利用求二者平均时间序列的方式进行融合处理,作为经过一轮训练得到的第一识别模型;若根据MAPE指标值确定所述经过一轮训练得到的第一识别模型不符合评估标准,则重新执行划分所述第一模型训练集和继续训练不符合标准的所述第一识别模型的步骤,直至所述第一识别模型符合标准;训练单元323,具体还可用于采用随机采样方式从所述第二模型训练集中分别获取第三训练样本集、第四训练样本集;基于所述第三训练样本集利用ARIMA算法,训练得到第三分类器;基于所述第四训练样本集 利用LSTM算法,训练得到第四分类器;将所述第三分类器和所述第四分类器利用求二者平均时间序列的方式进行融合处理,作为经过一轮训练得到的第二识别模型;若根据MAPE指标值确定所述经过一轮训练得到的第二识别模型不符合评估标准,则重新执行划分所述第二模型训练集和继续训练不符合标准的所述第二识别模型的步骤,直至所述第二识别模型符合标准。
在具体的应用场景中,如图4所示,本装置还包括:计算模块34;提取模块31,还可用于从所述历史用户数据中获取所述目标用户每次随访的用药数据、症状数据、体征数据、生活方式数据、空腹血糖值、餐后两小时血糖值;计算模块34,可用于将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的空腹血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第一预测序列中血糖指标变化的影响因素;及将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的餐后两小时血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第二预测序列中血糖指标变化的影响因素。在具体的应用场景中,提取模块31,还可用于获取每次随访用药组成的用药评分值序列、每次随访症状组成的症状评分值序列、每次随访体征组成的体征评分值序列、每次随访生活方式组成的生活作息评分值序列、每次随访测量的空腹血糖值序列和餐后两小时血糖值序列;计算模块34,具体可用于计算所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列分别与所述空腹血糖值序列之间的相关系数的第一绝对值;依据所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列中获取所述第一绝对值最高的一项或多项,确定影响所述第一预测序列中血糖指标变化的关键影响因素;计算模块34,具体还可用于计算所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列分别与所述餐后两小时血糖值序列之间的相关系数的第二绝对值;依据所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列中获取所述第二绝对值最高的一项或多项,确定影响所述第二预测序列中血糖指标变化的关键影响因素。
需要说明的是,本实施例提供的一种预测预测未来血糖值的装置所涉及各功能模块的其它相应描述,可以参考图1至图2中的对应描述,在此不再赘述。基于上述如图1和图2所示方法,相应的,本申请实施例还提供了一种存储介质,其上存储有计算机可读指令,该计算机可读指令被处理器执行时实现上述如图1和图2所示的预测未来血糖值的方法。基于这样的理解,本申请的技术方案可以以软件产品的形式体现出来,该软件产品可以存储在一个非易失性存储介质(可以是CD-ROM,U盘,移动硬盘等)中,包括若干指令用以使得一台计算机设备(可以是个人计算机,服务器,或者网络设备等)执行本申请各个实施场景的方法。基于上述如图1、图2所示的方法,以及图3、图4所示的虚拟装置实施例,为了实现上述目的,本申请实施例还提供了一种计算机设备,具体可以为个人计算机、服务器、网络设备等,该实体设备包括存储介质和处理器;存储介质,用于存储计算机可读指令;处理器,用于执行计算机可读指令以实现上述如图1和图2所示的预测未来血糖值的方法。可选地,该计算机设备还可以包括用户接口、网络接口、摄像头、射频(Radio Frequency,RF)电路,传感器、音频电路、WI-FI模块等等。用户接口可以包括显示屏(Display)、输入单元比如键盘(Keyboard)等,可选用户接口还可以包括USB接口、读卡器接口等。网络接口可选的可以包括标准的有线接口、无线接口(如蓝牙接口、WI-FI接口)等。
本领域技术人员可以理解,本实施例提供的计算机设备结构并不构成对该实体设备的限定,可以包括更多或更少的部件,或者组合某些部件,或者不同的部件布置。非易失性可读存储介质中还可以包括操作系统、网络通信模块。操作系统是管理预测未来血糖值的实体设备硬件和软件资源的程序,支持信息处理程序以及其它软件和/或程序的运行。网络通信模块用于实现非易失性可读存储介质内部各组件之间的通信,以及与该实体设备中其它硬件和软件之间通信。通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到本申请可以借助软件加必要的通用硬件平台的方式来实现,也可以通过硬件实现。通过应用本申请的技术方案,与目前现有技术相比,本申请可很好的预测出目标用户的未来血糖值变化规律,进而可实时监测糖尿病的发展情况,进行相应的配套治疗,在对糖尿病的治疗控制上具有很大的意义;并且还可判断出影响未来血糖指标变化的重要因素,可使诊断结果更加完善,覆盖类型更全面,从而能为患者及医生提供更好的治疗依据。本领域技术人员可以理解附图只是一个优选实施场景的示意图,附图中的模块或流程并不一定是实施本申请所必须的。本领域技术人员可以理解实施场景中的装置中的模块可以按照实施场景描述进行分布于实施场景的装置中,也可以进行相应变化位于不同于本实施场景的一个或多个装置中。上述实施场景的模块可以合并为一个模块,也可以进一步拆分成多个子模块。上述本申请序号仅仅为了描述,不代表实施场景的优劣。以上公开的仅为本申请的几个具体实施场景,但是,本申请并非局限于此,任何本领域的技术人员能思之的变化都应落入本申请的保护范围。

Claims (20)

  1. 一种预测未来血糖值的方法,包括:
    从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据;
    利用所述样本用户数据创建血糖值时间序列的预测模型;
    将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,其中,所述历史用户数据对应的历史时长小于所述预设时长阈值,所述目标用户和所述样本用户均为糖尿病患者。
  2. 根据权利要求1所述的方法,所述利用所述样本用户数据创建血糖值时间序列的预测模型,具体包括:对样本用户数据进行预处理;将预处理后的所述样本用户数据中空腹血糖值的时序数据作为标签信息Y1,将所述样本用户数据中除空腹血糖值和餐后两小时血糖值以外的目标特征数据作为特征信息X,创建第一模型训练集,其中,所述目标特征数据包括用药数据、症状数据、体征数据、生活方式数据;利用所述第一模型训练集并结合时间序列算法,训练得到第一识别模型,其中,所述时间序列算法由ARIMA和LSTM两种算法融合得到,所述第一识别模型的评估采用平均绝对百分比误差MAPE指标,当所述第一识别模型对应的MAPE指标值小于预置标准比较阈值时,确定所述第一识别模型符合评估标准,通过符合评估标准的所述第一识别模型可确定所述特征信息X和所述标签信息Y1之间的第一映射关系;将预处理后的所述样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,并结合所述特征信息X,创建第二模型训练集;利用所述第二模型训练集并结合所述时间序列算法,训练得到第二识别模型,其中,所述第二识别模型的评估采用MAPE指标,当所述第二识别模型对应的MAPE指标值小于预定标准比较阈值时,确定所述第二识别模型符合评估标准,通过符合评估标准的所述第二识别模型可确定所述特征信息X和所述标签信息Y2之间的第二映射关系。
  3. 根据权利要求2所述的方法,所述对样本用户数据进行预处理,具体包括:利用正则表达式从所述样本用户数据中提取空腹血糖值的时序数据和餐后两小时血糖值的时序数据;若所述空腹血糖值的时序数据或所述餐后两小时血糖的时序数据,在预设历史时间段的部分日期存在一天内具有多条血糖测量数据的情况,则使用所述多条血糖测量值中的众数作为该天对应的血糖测量值;及若所述空腹血糖值或所述餐后两小时血糖值在所述预设历史时间段内出现血糖测量数据空缺的日期,则使用内插法补充所述日期对应的空值,得到日颗粒度的血糖序列;对所述日颗粒度的血糖序列,进行差分化移动平滑处理,得到稳定序列;
    所述将预处理后的所述样本用户数据中空腹血糖值的时序数据作为标签信息Y1,具体包括:将预处理后的所述空腹血糖值的稳定序列作为标签信息Y1;
    所述将预处理后的所述样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,具体包括:将预处理后的所述餐后两小时血糖值的稳定序列作为标签信息Y2。
  4. 根据权利要求3所述的方法,所述将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,具体包括:将所述目标用户的历史用户数据经过预处理,得到所述目标用户的历史空腹血糖值的稳定序列、历史餐后两小时血糖值的稳定序列、历史用药数据、历史症状数据、历史体征数据、历史生活方式数据;将所述历史用药数据、所述历史症状数据、所述历史体征数据、所述历史生活方式数据输入到所述第一识别模型中与所述特征信息X进行相似度匹配,并利用所述第一映射关系,确定与相似度大于预设阈值、且相似度最高的所述特征信息X对应的目标标签信息Y1;将所述目标标签信息Y1相应的第一目标时间序列与所述历史空腹血糖值的稳定序列进行近似匹配,判定所述历史空腹血糖值的稳定序列匹配到所述第一目标时间序列中的第一目标位置,以便利用所述第一目标时间序列中在所述第一目标位置后的空腹血糖值时间序列,确定所述目标用户在预设未来时间段内空腹血糖值的第一预测序列;将所述历史用药数据、所述历史症状数据、所述历史体征数据、所述历史生活方式数据输入到所述第二识别模型中与所述特征信息X进行相似度匹配,并利用所述第二映射关系,确定与相似度大于预设阈值、且相似度最高的所述特征信息X对应的目标标签信息Y2;将所述目标标签信息Y2相应的第二目标时间序列与所述历史餐后两小时血糖值的稳定序列进行近似匹配,判定所述餐后两小时血糖值的稳定序列匹配到所述第二目标时间序列中的第二目标位置,以便利用所述第二目标时间序列中在所述第二目标位置后的餐后两小时血糖值时间序列,确定所述目标用户在预设未来时间段内餐后两小时血糖值的第二预测序列。
  5. 根据权利要求2所述的方法,所述利用所述第一模型训练集并结合时间序列算法,训练得到第一识别模型,具体包括:采用随机采样方式从所述第一模型训练集中分别获取第一训练样本集、第二训练样本集;基于所述第一训练样本集利用ARIMA算法,训练得到第一分类器;基于所述第二训练样本集利用LSTM算法,训练得到第二分类器;将所述第一分类器和所述第二分类器利用求二者平均时 间序列的方式进行融合处理,作为经过一轮训练得到的第一识别模型;若根据MAPE指标值确定所述经过一轮训练得到的第一识别模型不符合评估标准,则重新执行划分所述第一模型训练集和继续训练不符合标准的所述第一识别模型的步骤,直至所述第一识别模型符合标准;
    所述利用所述第二模型训练集并结合所述时间序列算法,训练得到第二识别模型,具体包括:采用随机采样方式从所述第二模型训练集中分别获取第三训练样本集、第四训练样本集;基于所述第三训练样本集利用ARIMA算法,训练得到第三分类器;基于所述第四训练样本集利用LSTM算法,训练得到第四分类器;将所述第三分类器和所述第四分类器利用求二者平均时间序列的方式进行融合处理,作为经过一轮训练得到的第二识别模型;若根据MAPE指标值确定所述经过一轮训练得到的第二识别模型不符合评估标准,则重新执行划分所述第二模型训练集和继续训练不符合标准的所述第二识别模型的步骤,直至所述第二识别模型符合标准。
  6. 根据权利要求1所述的方法,在所述将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列之后,所述方法还包括:从所述历史用户数据中获取所述目标用户每次随访的用药数据、症状数据、体征数据、生活方式数据、空腹血糖值、餐后两小时血糖值;将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的空腹血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第一预测序列中血糖指标变化的影响因素;及将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的餐后两小时血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第二预测序列中血糖指标变化的影响因素。
  7. 根据权利要求6所述的方法,在所述将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的空腹血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第一预测序列中血糖指标变化的影响因素之前,所述方法还包括:获取每次随访用药组成的用药评分值序列、每次随访症状组成的症状评分值序列、每次随访体征组成的体征评分值序列、每次随访生活方式组成的生活作息评分值序列、每次随访测量的空腹血糖值序列和餐后两小时血糖值序列;
    所述将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的空腹血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第一预测序列中血糖指标变化的影响因素,具体包括:计算所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列分别与所述空腹血糖值序列之间的相关系数的第一绝对值;依据所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列中获取所述第一绝对值最高的一项或多项,确定影响所述第一预测序列中血糖指标变化的关键影响因素;
    所述将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的餐后两小时血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第二预测序列中血糖指标变化的影响因素,具体包括:
    计算所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列分别与所述餐后两小时血糖值序列之间的相关系数的第二绝对值;依据所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列中获取所述第二绝对值最高的一项或多项,确定影响所述第二预测序列中血糖指标变化的关键影响因素。
  8. 一种预测未来血糖值的装置,包括:
    提取模块,用于从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据;
    创建模块,用于利用所述样本用户数据创建血糖值时间序列的预测模型;
    确定模块,用于将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,其中,所述历史用户数据对应的历史时长小于所述预设时长阈值,所述目标用户和所述样本用户均为糖尿病患者。
  9. 根据权利要求8所述的装置,所述创建模块,具体包括:处理单元、创建单元、训练单元;
    所述处理单元,用于对样本用户数据进行预处理;所述创建单元,用于将预处理后的所述样本用户数据中空腹血糖值的时序数据作为标签信息Y1,将所述样本用户数据中除空腹血糖值和餐后两小时血糖值以外的目标特征数据作为特征信息X,创建第一模型训练集,其中,所述目标特征数据包括用药数据、症状数据、体征数据、生活方式数据;
    所述训练单元,用于利用所述第一模型训练集并结合时间序列算法,训练得到第一识别模型,其中,所述时间序列算法由ARIMA和LSTM两种算法融合得到,所述第一识别模型的评估采用平均绝对百分比误差MAPE指标,当所述第一识别模型对应的MAPE指标值小于预置标准比较阈值时,确定所述第一识别模型符合评估标准,通过符合评估标准的所述第一识别模型可确定所述特征信息X和所述标签信息Y1之间的第一映射关系;
    所述创建单元,还用于将预处理后的所述样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,并结合所述特征信息X,创建第二模型训练集;
    所述训练单元,还用于利用所述第二模型训练集并结合所述时间序列算法,训练得到第二识别模型,其中,所述第二识别模型的评估采用MAPE指标,当所述第二识别模型对应的MAPE指标值小于预定标准比较阈值时,确定所述第二识别模型符合评估标准,通过符合评估标准的所述第二识别模型可确定所述特征信息X和所述标签信息Y2之间的第二映射关系。
  10. 根据权利要求9所述的方法,所述处理单元,具体用于利用正则表达式从所述样本用户数据中提取空腹血糖值的时序数据和餐后两小时血糖值的时序数据;若所述空腹血糖值的时序数据或所述餐后两小时血糖的时序数据,在预设历史时间段的部分日期存在一天内具有多条血糖测量数据的情况,则使用所述多条血糖测量值中的众数作为该天对应的血糖测量值;及若所述空腹血糖值或所述餐后两小时血糖值在所述预设历史时间段内出现血糖测量数据空缺的日期,则使用内插法补充所述日期对应的空值,得到日颗粒度的血糖序列;对所述日颗粒度的血糖序列,进行差分化移动平滑处理,得到稳定序列;
    所述创建单元,具体用于将预处理后的所述空腹血糖值的稳定序列作为标签信息Y1;所述创建单元,具体还用于将预处理后的所述餐后两小时血糖值的稳定序列作为标签信息Y2。
  11. 根据权利要求10所述的装置,所述确定模块,具体用于将所述目标用户的历史用户数据经过预处理,得到所述目标用户的历史空腹血糖值的稳定序列、历史餐后两小时血糖值的稳定序列、历史用药数据、历史症状数据、历史体征数据、历史生活方式数据;将所述历史用药数据、所述历史症状数据、所述历史体征数据、所述历史生活方式数据输入到所述第一识别模型中与所述特征信息X进行相似度匹配,并利用所述第一映射关系,确定与相似度大于预设阈值、且相似度最高的所述特征信息X对应的目标标签信息Y1;将所述目标标签信息Y1相应的第一目标时间序列与所述历史空腹血糖值的稳定序列进行近似匹配,判定所述历史空腹血糖值的稳定序列匹配到所述第一目标时间序列中的第一目标位置,以便利用所述第一目标时间序列中在所述第一目标位置后的空腹血糖值时间序列,确定所述目标用户在预设未来时间段内空腹血糖值的第一预测序列;将所述历史用药数据、所述历史症状数据、所述历史体征数据、所述历史生活方式数据输入到所述第二识别模型中与所述特征信息X进行相似度匹配,并利用所述第二映射关系,确定与相似度大于预设阈值、且相似度最高的所述特征信息X对应的目标标签信息Y2;将所述目标标签信息Y2相应的第二目标时间序列与所述历史餐后两小时血糖值的稳定序列进行近似匹配,判定所述餐后两小时血糖值的稳定序列匹配到所述第二目标时间序列中的第二目标位置,以便利用所述第二目标时间序列中在所述第二目标位置后的餐后两小时血糖值时间序列,确定所述目标用户在预设未来时间段内餐后两小时血糖值的第二预测序列。
  12. 根据权利要求9所述的装置,所述训练单元,具体用于采用随机采样方式从所述第一模型训练集中分别获取第一训练样本集、第二训练样本集;基于所述第一训练样本集利用ARIMA算法,训练得到第一分类器;基于所述第二训练样本集利用LSTM算法,训练得到第二分类器;将所述第一分类器和所述第二分类器利用求二者平均时间序列的方式进行融合处理,作为经过一轮训练得到的第一识别模型;若根据MAPE指标值确定所述经过一轮训练得到的第一识别模型不符合评估标准,则重新执行划分所述第一模型训练集和继续训练不符合标准的所述第一识别模型的步骤,直至所述第一识别模型符合标准;
    所述训练单元,具体还用于采用随机采样方式从所述第二模型训练集中分别获取第三训练样本集、第四训练样本集;基于所述第三训练样本集利用ARIMA算法,训练得到第三分类器;基于所述第四训练样本集利用LSTM算法,训练得到第四分类器;将所述第三分类器和所述第四分类器利用求二者平均时间序列的方式进行融合处理,作为经过一轮训练得到的第二识别模型;若根据MAPE指标值确定所述经过一轮训练得到的第二识别模型不符合评估标准,则重新执行划分所述第二模型训练集和继续训练不符合标准的所述第二识别模型的步骤,直至所述第二识别模型符合标准。
  13. 根据权利要求8所述的装置,所述装置还包括:计算模块;
    所述提取模块,还用于从所述历史用户数据中获取所述目标用户每次随访的用药数据、症状数据、体征数据、生活方式数据、空腹血糖值、餐后两小时血糖值;
    所述计算模块,用于将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的空腹血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第一预测序列中血糖指标变化的影响因素;及将所述每次随访的用药数据、症状数据、体征数据、生活方式数据分别与每次随访的餐后两小时血糖值之间进行相关分析计算,依据得到的相关分析计算结果判定影响所述第二预测序列中血糖指标变化的影响因素。
  14. 根据权利要求13所述的装置,所述提取模块,还用于获取每次随访用药组成的用药评分值序列、每次随访症状组成的症状评分值序列、每次随访体征组成的体征评分值序列、每次随访生活方式组成的生活作息评分值序列、每次随访测量的空腹血糖值序列和餐后两小时血糖值序列;
    所述计算模块,具体用于计算所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列分别与所述空腹血糖值序列之间的相关系数的第一绝对值;依据所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列中获取所述第一绝对值最高的一项或多项,确定影响所述第一预测序列中血糖指标变化的关键影响因素;
    所述计算模块,具体还用于计算所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列分别与所述餐后两小时血糖值序列之间的相关系数的第二绝对值;依据所述用药评分值序列、所述症状评分值序列、所述体征评分值序列、所述生活作息评分值序列中获取所述第二绝对值最高的一项或多项,确定影响所述第二预测序列中血糖指标变化的关键影响因素。
  15. 一种非易失性可读存储介质,其上存储有计算机可读指令,所述计算机可读指令被处理器执行时实现预测未来血糖值的方法,包括:从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据;利用所述样本用户数据创建血糖值时间序列的预测模型;将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,其中,所述历史用户数据对应的历史时长小于所述预设时长阈值,所述目标用户和所述样本用户均为糖尿病患者。
  16. 根据权利要求15所述的非易失性可读存储介质,所述计算机可读指令被所述处理器执行时实现所述利用所述样本用户数据创建血糖值时间序列的预测模型,具体包括:对样本用户数据进行预处理;将预处理后的所述样本用户数据中空腹血糖值的时序数据作为标签信息Y1,将所述样本用户数据中除空腹血糖值和餐后两小时血糖值以外的目标特征数据作为特征信息X,创建第一模型训练集,其中,所述目标特征数据包括用药数据、症状数据、体征数据、生活方式数据;利用所述第一模型训练集并结合时间序列算法,训练得到第一识别模型,其中,所述时间序列算法由ARIMA和LSTM两种算法融合得到,所述第一识别模型的评估采用平均绝对百分比误差MAPE指标,当所述第一识别模型对应的MAPE指标值小于预置标准比较阈值时,确定所述第一识别模型符合评估标准,通过符合评估标准的所述第一识别模型可确定所述特征信息X和所述标签信息Y1之间的第一映射关系;将预处理后的所述样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,并结合所述特征信息X,创建第二模型训练集;利用所述第二模型训练集并结合所述时间序列算法,训练得到第二识别模型,其中,所述第二识别模型的评估采用MAPE指标,当所述第二识别模型对应的MAPE指标值小于预定标准比较阈值时,确定所述第二识别模型符合评估标准,通过符合评估标准的所述第二识别模型可确定所述特征信息X和所述标签信息Y2之间的第二映射关系。
  17. 根据权利要求16所述的非易失性可读存储介质,所述计算机可读指令被所述处理器执行时实现所述对样本用户数据进行预处理,具体包括:利用正则表达式从所述样本用户数据中提取空腹血糖值的时序数据和餐后两小时血糖值的时序数据;若所述空腹血糖值的时序数据或所述餐后两小时血糖的时序数据,在预设历史时间段的部分日期存在一天内具有多条血糖测量数据的情况,则使用所述多条血糖测量值中的众数作为该天对应的血糖测量值;及若所述空腹血糖值或所述餐后两小时血糖值在所述预设历史时间段内出现血糖测量数据空缺的日期,则使用内插法补充所述日期对应的空值,得到日颗粒度的血糖序列;对所述日颗粒度的血糖序列,进行差分化移动平滑处理,得到稳定序列;
    所述计算机可读指令被所述处理器执行时实现所述将预处理后的所述样本用户数据中空腹血糖值的时序数据作为标签信息Y1,具体包括:将预处理后的所述空腹血糖值的稳定序列作为标签信息Y1;所述计算机可读指令被所述处理器执行时实现所述将预处理后的所述样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,具体包括:将预处理后的所述餐后两小时血糖值的稳定序列作为标签信息Y2。
  18. 一种计算机设备,包括非易失性可读存储介质、处理器及存储在非易失性可读存储介质上并可在处理器上运行的计算机可读指令,所述处理器执行所述计算机可读指令时实现预测未来血糖值的方法,包括:从原始健康档案和电子病历中提取糖尿病随访记录大于预设时长阈值的样本用户对应的样本用户数据;利用所述样本用户数据创建血糖值时间序列的预测模型;将目标用户的历史用户数据输入到所述预测模型中,预测所述目标用户在预设未来时间段内空腹血糖值的第一预测序列,和餐后两小时血糖值的第二预测序列,其中,所述历史用户数据对应的历史时长小于所述预设时长阈值,所述目标用户和所述样本用户均为糖尿病患者。
  19. 根据权利要求18所述的计算机设备,所述计算机可读指令被所述处理器执行时实现所述利用所述样本用户数据创建血糖值时间序列的预测模型,具体包括:对样本用户数据进行预处理;将预处理后的所述样本用户数据中空腹血糖值的时序数据作为标签信息Y1,将所述样本用户数据中除空腹血糖值和餐后两小时血糖值以外的目标特征数据作为特征信息X,创建第一模型训练集,其中,所述目标特征数据包括用药数据、症状数据、体征数据、生活方式数据;利用所述第一模型训练集并结合时间序列算法,训练得到第一识别模型,其中,所述时间序列算法由ARIMA和LSTM两种算法融合得到,所 述第一识别模型的评估采用平均绝对百分比误差MAPE指标,当所述第一识别模型对应的MAPE指标值小于预置标准比较阈值时,确定所述第一识别模型符合评估标准,通过符合评估标准的所述第一识别模型可确定所述特征信息X和所述标签信息Y1之间的第一映射关系;将预处理后的所述样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,并结合所述特征信息X,创建第二模型训练集;利用所述第二模型训练集并结合所述时间序列算法,训练得到第二识别模型,其中,所述第二识别模型的评估采用MAPE指标,当所述第二识别模型对应的MAPE指标值小于预定标准比较阈值时,确定所述第二识别模型符合评估标准,通过符合评估标准的所述第二识别模型可确定所述特征信息X和所述标签信息Y2之间的第二映射关系。
  20. 根据权利要求19所述的计算机设备,所述计算机可读指令被所述处理器执行时实现所述对样本用户数据进行预处理,具体包括:利用正则表达式从所述样本用户数据中提取空腹血糖值的时序数据和餐后两小时血糖值的时序数据;若所述空腹血糖值的时序数据或所述餐后两小时血糖的时序数据,在预设历史时间段的部分日期存在一天内具有多条血糖测量数据的情况,则使用所述多条血糖测量值中的众数作为该天对应的血糖测量值;及若所述空腹血糖值或所述餐后两小时血糖值在所述预设历史时间段内出现血糖测量数据空缺的日期,则使用内插法补充所述日期对应的空值,得到日颗粒度的血糖序列;对所述日颗粒度的血糖序列,进行差分化移动平滑处理,得到稳定序列;
    所述计算机可读指令被所述处理器执行时实现所述将预处理后的所述样本用户数据中空腹血糖值的时序数据作为标签信息Y1,具体包括:将预处理后的所述空腹血糖值的稳定序列作为标签信息Y1;所述计算机可读指令被所述处理器执行时实现所述将预处理后的所述样本用户数据中餐后两小时血糖值的时序数据作为标签信息Y2,具体包括:将预处理后的所述餐后两小时血糖值的稳定序列作为标签信息Y2。
PCT/CN2019/117222 2019-03-12 2019-11-11 预测未来血糖值的方法、装置及计算机设备 Ceased WO2020181806A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910185058.0 2019-03-12
CN201910185058.0A CN110085318A (zh) 2019-03-12 2019-03-12 预测未来血糖值的方法、装置及计算机设备

Publications (1)

Publication Number Publication Date
WO2020181806A1 true WO2020181806A1 (zh) 2020-09-17

Family

ID=67413150

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/117222 Ceased WO2020181806A1 (zh) 2019-03-12 2019-11-11 预测未来血糖值的方法、装置及计算机设备

Country Status (2)

Country Link
CN (1) CN110085318A (zh)
WO (1) WO2020181806A1 (zh)

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110085318A (zh) * 2019-03-12 2019-08-02 平安科技(深圳)有限公司 预测未来血糖值的方法、装置及计算机设备
CN112489798A (zh) * 2019-09-12 2021-03-12 金安国科智能科技(北京)有限公司 一种人体状态评价、人体数据可视化方法、装置、设备及介质
CN110782989B (zh) * 2019-09-18 2022-06-17 平安科技(深圳)有限公司 数据分析方法、装置、设备及计算机可读存储介质
JP6790207B1 (ja) * 2019-09-25 2020-11-25 株式会社東芝 検査値予測装置、検査値予測システム、検査値予測方法、およびプログラム
CN111383758A (zh) * 2020-03-06 2020-07-07 三七二二(北京)健康咨询有限公司 基于多维数据预测餐后血糖的方法和装置
CN112102953B (zh) * 2020-10-22 2023-06-16 平安科技(深圳)有限公司 个性化糖尿病健康管理系统、设备及存储介质
CN113113142A (zh) * 2021-04-09 2021-07-13 长沙理工大学 一种利用智能分析技术对糖尿病患病风险的预测方法
CN113628755B (zh) * 2021-08-20 2024-03-12 东南大学附属中大医院 患者的血糖控制方法、装置、设备以及存储介质
CN113951879B (zh) * 2021-12-21 2022-04-05 苏州百孝医疗科技有限公司 血糖预测方法和装置、监测血糖水平的系统
CN114021672B (zh) * 2022-01-06 2022-04-22 苏州百孝医疗科技有限公司 标签生成方法和装置、监控血糖水平的系统
CN114530250B (zh) * 2022-04-24 2022-08-02 广东工业大学 基于数据增强的可穿戴血糖检测方法、系统及存储介质
CN118039139A (zh) * 2022-11-14 2024-05-14 华为技术有限公司 一种健康管理的方法、装置、系统、电子设备及存储介质
CN117612737B (zh) * 2024-01-24 2024-05-03 胜利油田中心医院 一种糖尿病护理数据智能优化方法
CN120527040B (zh) * 2025-07-25 2025-11-21 佛山市南海区大沥镇社区卫生服务中心(佛山市南海区大沥镇疾病预防控制中心) 一种慢性病监测方法、系统、设备及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103310113A (zh) * 2013-06-24 2013-09-18 浙江大学 一种基于频带分离和数据建模的通用血糖预测方法
CN106384013A (zh) * 2016-09-28 2017-02-08 湖南老码信息科技有限责任公司 一种基于增量式神经网络模型的ii型糖尿病预测方法和预测系统
US20180150609A1 (en) * 2016-11-29 2018-05-31 Electronics And Telecommunications Research Institute Server and method for predicting future health trends through similar case cluster based prediction models
CN110085318A (zh) * 2019-03-12 2019-08-02 平安科技(深圳)有限公司 预测未来血糖值的方法、装置及计算机设备

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150347707A1 (en) * 2014-05-30 2015-12-03 Anthony Michael Albisser Computer-Implemented System And Method For Improving Glucose Management Through Cloud-Based Modeling Of Circadian Profiles
CN108766578A (zh) * 2018-05-16 2018-11-06 清华大学深圳研究生院 一种血糖预测方法及装置

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103310113A (zh) * 2013-06-24 2013-09-18 浙江大学 一种基于频带分离和数据建模的通用血糖预测方法
CN106384013A (zh) * 2016-09-28 2017-02-08 湖南老码信息科技有限责任公司 一种基于增量式神经网络模型的ii型糖尿病预测方法和预测系统
US20180150609A1 (en) * 2016-11-29 2018-05-31 Electronics And Telecommunications Research Institute Server and method for predicting future health trends through similar case cluster based prediction models
CN110085318A (zh) * 2019-03-12 2019-08-02 平安科技(深圳)有限公司 预测未来血糖值的方法、装置及计算机设备

Also Published As

Publication number Publication date
CN110085318A (zh) 2019-08-02

Similar Documents

Publication Publication Date Title
WO2020181806A1 (zh) 预测未来血糖值的方法、装置及计算机设备
Al-Zaiti et al. Machine learning-based prediction of acute coronary syndrome using only the pre-hospital 12-lead electrocardiogram
Gao et al. An effective LSTM recurrent network to detect arrhythmia on imbalanced ECG dataset
US11600390B2 (en) Machine learning clinical decision support system for risk categorization
CN108461110B (zh) 医疗信息处理方法、装置及设备
US20210035693A1 (en) Methods, systems, and apparatuses for predicting the risk of hospitalization
Zhang et al. Nonlaboratory-based risk assessment model for type 2 diabetes mellitus screening in Chinese rural population: a joint bagging-boosting model
CN113345577A (zh) 诊疗辅助信息的生成方法、模型训练方法、装置、设备以及存储介质
US20180210925A1 (en) Reliability measurement in data analysis of altered data sets
Ieva et al. Risk prediction for myocardial infarction via generalized functional regression models
Morris et al. A new method to assess asymmetry in fingerprints could be used as an early indicator of type 2 diabetes mellitus
Liu et al. Beatquency domain and machine learning improve prediction of cardiovascular death after acute coronary syndrome
Obaido et al. An improved ensemble method for predicting hyperchloremia in adults with diabetic ketoacidosis
CN120513491A (zh) 用于妊娠病症的早期识别的预测模型
Nikolaou et al. The cardiovascular phenotype of Chronic Obstructive Pulmonary Disease (COPD): Applying machine learning to the prediction of cardiovascular comorbidities
JP6177609B2 (ja) 医療カルテシステム及び医療カルテ検索方法
LK et al. COVID-19 outbreak based coronary heart diseases (CHD) prediction using SVM and risk factor validation
Saputra et al. Hyperparameter optimization for cardiovascular disease data-driven prognostic system
Uddin et al. A Voice assistive mobile application tool to detect cardiovascular disease using machine learning approach
Patnaik et al. Intelligent decision support system in healthcare using machine learning models
Praveen et al. Diabetes prediction with ensemble learning techniques in machine learning
Naz et al. Meta-ensemble learning for heart disease prediction: a stacking-based approach with explainable AI
Sinha et al. CARDPSoML: Comparative approach to analyze and predict cardiovascular disease based on medical report data and feature fusion approach
Liao et al. Application of machine learning to identify clustering of cardiometabolic risk factors in US adults
Sinha et al. Automated detection of coronary artery disease using machine learning algorithm

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19918959

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19918959

Country of ref document: EP

Kind code of ref document: A1

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 25/01/2022)