Disease process twinborn construction method and system based on standard electronic medical record

Abstract

The invention discloses a disease process twinborn construction method and system based on a standard electronic medical record, and the method comprises the steps: obtaining electronic medical record data based on a standard FHIR electronic medical record of a patient; encoding and processing the electronic medical record data to obtain multivariable time sequence data; constructing a twinborn model in the disease process; acquiring a disease trajectory prediction report; a GPT model is introduced, numerical information in the disease trajectory prediction report and semantic information of the knowledge graph are converted into natural language description, and a report template is obtained; adjusting the report template; and generating a personalized medical report including the risk early warning information and the intervention suggestions. According to the method, the disease course of the patient with the diabetes complicated with the coronary heart disease can be subjected to individualized prediction by utilizing the constructed twinborn model in the disease process. By inputting personal information, medical history, laboratory examination and other data of a patient, the model can predict the disease development trend, the treatment effect and the occurrence risk of complications of the patient.

Description

Technical Field

[0001] The present application belongs to the field of medical information technology, and specifically relates to a method and system for constructing disease process twins based on standard electronic medical records. Background Art

[0002] With the rapid development of information technology, the medical industry has also ushered in a digital transformation. As an important part of medical informatization, the standard electronic medical record system has been widely used around the world. The medical record system not only records the patient's clinical information, diagnosis, treatment history, medical service data, but also contains a wealth of structured and unstructured information, providing important clinical reference materials for medical staff. However, how to effectively use these massive data to improve the scientificity and accuracy of clinical decision-making has always been a research hotspot in the field of medical informatics.

[0003] In this context, deep learning technologies such as recurrent neural network (RNN) and Transformer model have been proposed and applied to disease trajectory modeling and prediction. RNN can process sequence data and capture dynamic changes in time series, which is very suitable for processing medical time series data. The Transformer model can better handle long-distance dependencies through the self-attention mechanism and is suitable for processing complex medical text data.

[0004] Diabetes and coronary heart disease are two common chronic diseases that have a profound impact on patients' quality of life and health. As a metabolic disease, diabetes is mainly characterized by long-term elevated blood sugar levels, which leads to damage to multiple organs and systems, such as the cardiovascular, kidney, and nervous systems. Patients with diabetes and coronary heart disease will increase their risk of death and difficulty in treatment. Traditionally, doctors usually rely on clinical experience and expertise to develop treatment plans. However, the situation of diabetes and coronary heart disease is often very complicated, and there are huge individual differences between different patients. It is often difficult to develop the most optimized treatment strategy based on experience alone.

[0005] In order to meet this challenge, in recent years, the construction of disease process twins based on electronic medical record data has become a research direction that has attracted much attention. The disease process twin model is a model that uses machine learning and artificial intelligence technology to analyze a large amount of electronic medical record data to build a model of the patient's disease development, thereby simulating and predicting the patient's disease course. This model can provide doctors with more scientific and accurate clinical decision support, help medical staff better understand the development of the disease, and develop personalized treatment plans.

[0006] In the diagnosis and treatment of diabetic coronary heart disease, the application of the disease process twin model is of great significance. First, by analyzing a large amount of medical record data, the laws and characteristics of the course of patients with diabetic coronary heart disease can be revealed, and the key factors affecting the progression of the disease can be identified. Secondly, the establishment of a disease process twin model can simulate and predict the course of the patient's disease, providing doctors with scientific and accurate treatment recommendations. In addition, the disease process twin model can also realize personalized medical management, formulate the most suitable treatment plan according to the individual characteristics of the patient and the changes in the condition, and maximize the treatment effect and quality of life. Summary of the invention

[0007] The present application provides a disease process twin construction method and system based on standard electronic medical records to solve the above-mentioned technical problems.

[0008] In order to solve the above technical problems, a technical solution adopted in this application is: a disease process twin construction method based on standard electronic medical records, comprising the following steps:

[0009] Based on the patient's standard FHIR electronic medical record, obtain electronic medical record data;

[0010] Encode and process electronic medical record data to obtain multivariate time series data;

[0011] Construct a twin model of disease process based on multivariate time series data;

[0012] Obtain disease trajectory prediction reports based on the output results and prediction graphs of the disease process twin model;

[0013] The GPT model is introduced to convert the numerical information in the disease trajectory prediction report and the semantic information of the knowledge graph into natural language descriptions to obtain the report template;

[0014] Adjust report templates based on interactive interface validation and input of additional clinical parameter values;

[0015] Based on the report template and preset risk thresholds, a personalized medical report including risk warning information and intervention recommendations is generated.

[0016] Further, EHR data include basic demographic information; medical history and diagnosis records; laboratory test results; treatment and medication records; lifestyle and behavioral data; and hospitalization and surgical records.

[0017] Furthermore, the method of encoding and processing the electronic medical record data to obtain multivariate time series data includes:

[0018] Parse and structure text information extracted from FHIR electronic medical records based on the GPT model to obtain key information including medical history, diagnosis records, laboratory test results, treatment and medication records, lifestyle and behavior data;

[0019] Arrange the key information in chronological order and convert the medical data in the text information into time series data to form multivariate time series data.

[0020] Furthermore, the method of constructing a twin model of a disease process includes:

[0021] Combine the RNN model with the NBM model to generate an RNN-NBM model for sequence modeling and prediction of multivariate time series data;

[0022] Based on the RNN-NBM model and the AFT model for temporal prediction of specific clinical events, a disease process twin model (RNN-NBM-AFT) was obtained.

[0023] Another technical solution adopted by this application is: a disease process twin construction system based on standard electronic medical records, including:

[0024] The electronic medical record coding module is used to parse the text information in the FHIR electronic medical record through the GP model T, convert the text information into multivariate time series data, and perform feature extraction and structural processing;

[0025] The sequence modeling and prediction module is used to compress and decompress the input data through an autoencoder to fill in missing data, and combine the RNN and NBM models to make high-precision predictions of disease trajectories;

[0026] The knowledge graph generation module is used to combine the model prediction results and the medical knowledge graph to generate a disease trajectory prediction report;

[0027] The medical report generation module converts the predicted numerical information and the semantic information of the knowledge graph into natural language descriptions by introducing the GPT model, and generates a complete personalized medical report;

[0028] The risk warning module is used to generate risk warning information and provide personalized treatment plan adjustments and health management suggestions.

[0029] The beneficial effect of this application is that this application uses the constructed disease process twin model to make individualized predictions of the course of disease in patients with diabetes and coronary heart disease. By inputting the patient's personal information, medical history, laboratory test data and other data, the model can predict the patient's disease development trend, treatment effect and risk of complications. Based on the results of the course prediction, doctors can develop more personalized treatment plans for patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a flowchart of an embodiment of a method for constructing a disease process twin based on a standard electronic medical record of the present application;

[0031] Figure 2 yes Figure 1 A schematic flow chart of an embodiment of step S2 in FIG.

[0032] Figure 3 yes Figure 1 A schematic flow chart of an embodiment of step S3 in FIG.

[0033] Figure 4 It is a structural block diagram of an embodiment of the disease process twin construction system based on standard electronic medical records of the present application. DETAILED DESCRIPTION

[0034] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with specific embodiments.

[0035] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways than those described herein. Therefore, the present invention is not limited to the specific embodiments of the following disclosure.

[0036] See also Figure 1 , Figure 1 : is a flow chart of an embodiment of a method for constructing a disease process twin based on a standard electronic medical record of the present application. The method comprises:

[0037] Step S1. Acquire electronic medical record data based on the patient's standard FHIR electronic medical record.

[0038] Specifically, this application extracts key information (such as age, gender, BMI, blood sugar level, blood lipid level, blood pressure, diagnosis records of diabetes and coronary heart disease, etc.) from the FHIR medical records of diabetic patients. Specifically including: basic demographic information: such as age, gender, weight, height, BMI. Medical history and diagnosis records: diabetes diagnosis date and type (type 1 or type 2), coronary heart disease diagnosis date and type (such as angina, myocardial infarction, etc.), and other related chronic diseases (such as hypertension, kidney disease, etc.). Laboratory test results: blood sugar level (fasting blood sugar, postprandial blood sugar, HbA1c), blood lipid levels (total cholesterol, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglycerides), blood pressure (systolic and diastolic blood pressure), serum creatinine and urine protein (to assess renal function), electrocardiogram (ECG) results, cardiac ultrasound (Echo) results. Treatment and medication records: diabetes medications (such as insulin, oral hypoglycemic drugs), coronary heart disease medications (such as statins, beta-blockers, antiplatelet drugs), other related medications (such as antihypertensive drugs, diuretics). Lifestyle and behavior data: eating habits, exercise frequency and type, smoking and drinking. Hospitalization and surgery records: number and duration of hospitalization, surgical history (such as coronary artery bypass grafting, heart stenting). Pre-process the above information, exclude irrelevant information, ensure data quality and information integrity, and prepare for data analysis.

[0039] For non-standard electronic medical records, you need to convert them into FHIR standard electronic medical records. Conduct a comprehensive analysis of non-standard electronic medical records, including their data structure, field definition, data format, etc. According to the content and structure of non-standard electronic medical records, select appropriate FHIR resources for mapping. For example, patient information can be mapped to FHIR's Patient resource, medical information can be mapped to Encounter resource, examination results can be mapped to Observation resource, etc. Design detailed data mapping rules to map each field or data block in the non-standard electronic medical record to the corresponding FHIR resource attribute. This step requires special attention to the conversion of data types, formats, and units to ensure that the mapped data meets the requirements of the FHIR standard. According to the designed data mapping rules, write code or configure tools to implement the actual mapping of data.

[0040] Step S2: Encode and process the electronic medical record data to obtain multivariate time series data.

[0041] For details, see Figure 2 , Figure 2 for Figure 1 Schematic diagram of the process of step S2 in an embodiment. Step S2 includes:

[0042] Step S21. Based on the GPT model, the text information extracted from the FHIR electronic medical record is parsed and structured to obtain key information including medical history, diagnosis records, laboratory test results, treatment and medication records, lifestyle and behavior data.

[0043] Specifically, large language models (such as GPT) are used to parse and structure text information extracted from FHIR electronic medical records. Key information extracted from electronic medical records includes medical history, diagnosis records, laboratory test results, treatment and medication records, lifestyle and behavior data, etc.

[0044] Step S22. Arrange the key information in chronological order and convert the medical data in the text information into time series data to form multivariate time series data.

[0045] Specifically, the above key information will be arranged in chronological order to form multivariate time series data. According to the parsed text information, the relevant medical data is converted into time series data. Each time point contains multiple variables, such as the patient's physiological indicators (such as blood sugar level, blood pressure, etc.), drug dosage, diagnosis results, laboratory test results, etc. The construction of time series data needs to ensure the continuity and integrity of the data, and exclude irrelevant information and noise data through data cleaning and preprocessing steps. In the process of constructing time series data, data from different sources are aligned and synchronized. For example, laboratory test results and diagnosis records may be recorded at different time points, so data alignment is required through timestamps and event matching. Through the semantic understanding ability of the GPT model, related events and data are associated to ensure the accuracy and consistency of time series data. The processed time series data will be stored in a structured data storage system for subsequent model training and prediction. The data storage system needs to have efficient data retrieval and management capabilities to support a variety of query and analysis needs.

[0046] Step S3. Construct a disease process twin model based on multivariate time series data.

[0047] For details, see Figure 3 , Figure 3 for Figure 1 Schematic diagram of the process of step S3 of an embodiment. Step S3 includes:

[0048] Step S31. Combine the RNN model and the NBM model to generate an RNN-NBM model for sequence modeling and prediction of multivariate time series data.

[0049] Step S32. Based on the RNN-NBM model and the AFT model for time prediction of specific clinical events, a disease process twin model (RNN-NBM-AFT) is obtained.

[0050] See also Figure 4 , the point predictor takes the patient’s data at baseline, background information, and two times, current time and future time. The autoencoder compresses the input vector into a lower dimensional embedding, then decompresses the embedding to reconstruct the input. Missing data is filled in by replacing any missing variables with their reconstructed values. And by adding the reconstruction loss of the autoencoder to the negative log-likelihood of the NBM, the network can be trained to fill in missing data in a way that improves prediction performance.

[0051] A residual network (Resnet) is used to predict the baseline variable values ​​at the current time point and future time points. These predictions are based on the patient's historical data and background information. The residuals between the predicted values ​​and the observed values ​​at the current time point are then calculated. These residuals are then adjusted using a function that decays with the difference between the current time and the future time (such as exponential decay). The adjusted residuals are added to the predictions for future time points to incorporate autocorrelation. The RNN receives the baseline prediction, contextual information, and the autoencoder outputs for the current and future time points as input and is trained to predict the standard deviation of each longitudinal variable at future time points. For indications that require prediction of specific clinical events, an accelerated failure time (AFT) model is used. The AFT model uses features extracted from the baseline prediction and context vectors and uses the RNN to calculate the time to event. The AFT model is able to capture changes in risk over time and predict acceleration or deceleration relative to the average event time. The above three components (baseline prediction, RNN variance prediction, and AFT model) are integrated into a unified digital twin prediction model. The entire model is trained using a composite loss function that combines the reconstruction error of the autoencoder, the negative log-likelihood of the NBM, the loss of the RNN variance prediction, and the negative log-likelihood of the AFT model.

[0052] In other embodiments, the integrated RNN-NBM-AFT model is trained by using a training set. During the training process, appropriate optimization algorithms (such as Adam, RMSprop, etc.) and loss functions (such as mean square error, cross entropy loss, etc.) are used to update the model parameters through the back propagation algorithm to minimize the prediction error. The model performance is evaluated on an independent test set, and the prediction effect is measured using indicators such as accuracy, precision, recall, F1 score, ROC curve, etc. The prediction performance of the model in different time periods and under different conditions is analyzed to ensure stability and accuracy. Hyperparameter tuning is performed based on the performance of the model on the test set, which may include changing the number of LSTM layers, the number of neurons, the learning rate, the batch size, etc. Use techniques such as cross-validation to avoid overfitting and ensure that the model has good generalization ability.

[0053] Step S4. Obtain a disease trajectory prediction report based on the output results and prediction graph of the disease process twin model.

[0054] Specifically, the processed data is input into the trained disease process twin model, and the disease trajectory prediction report is obtained by outputting the results and prediction graphs.

[0055] Step S5. Introduce the GPT model, convert the numerical information in the disease trajectory prediction report and the semantic information of the knowledge graph into natural language description, and obtain the report template.

[0056] Specifically, through the GPT model, the predicted numerical information and semantic information of the knowledge graph are converted into natural language descriptions, or the prediction results are converted into interpretable formats, such as tables, charts, or text descriptions, so that doctors and patients can understand and apply them and generate complete report templates.

[0057] Step S6. Adjust the report template based on the interactive interface verification and input of additional clinical parameter values.

[0058] Step S7. Generate a personalized medical report including risk warning information and intervention suggestions based on the report template and the preset risk threshold.

[0059] Specifically, based on the trained RNN-NBM-AFT model, we first need to set reasonable risk thresholds. These thresholds are determined based on the model's performance on an independent test set to ensure that alarms are effectively triggered without generating too many false positives. By analyzing the model's prediction performance in different time periods and under different conditions, these thresholds can be adjusted to optimize the model's accuracy and sensitivity.

[0060] When the model's prediction results exceed the preset risk threshold, the model will automatically generate risk warning information. The risk warning information will be promptly communicated to relevant medical staff and patients through appropriate means (such as email, text messages, or notifications in the medical system). According to the patient's specific situation and prediction results, the system will provide personalized intervention suggestions. Intervention suggestions include adjustments to treatment plans, optimization of drug dosages, and other health management suggestions. In addition, targeted health education and management suggestions, such as diet control and exercise plans, can be provided based on the prediction results to help patients better manage their condition. As the patient's condition develops and treatment proceeds, the model needs to continuously receive new data and feedback to update risk assessments and intervention recommendations. In addition, doctors and patients can also adjust their responses to strategies based on actual results to form a dynamically adjusted closed-loop management system.

[0061] See also Figure 4 , Figure 4This is a structural diagram of an embodiment of a disease process twin construction system based on standard electronic medical records of the present application. The system includes: an electronic medical record encoding module 1, a sequence modeling and prediction module 2, a knowledge graph generation module 3, a medical report generation module 4 and a risk warning module 5.

[0062] The electronic medical record encoding module 1 is used to parse the text information in the FHIR electronic medical record through the GP model T, convert the text information into multivariate time series data, and perform feature extraction and structural processing. The sequence modeling and prediction module 2 is used to compress and decompress the input data through the autoencoder to fill in the missing data, and combine the RNN and NBM models to make high-precision predictions of disease trajectories; the knowledge graph generation module 3 is used to combine the model prediction results and the medical knowledge graph to generate a disease trajectory prediction report. The medical report generation module 4 converts the predicted numerical information and the semantic information of the knowledge graph into natural language descriptions by introducing the GPT model, and generates a complete personalized medical report. The risk warning module 5 can set a reasonable risk threshold, generate risk warning information, and provide personalized treatment plan adjustments and health management suggestions.

[0063] The system of this application forms a dynamically adjusted closed-loop management system through comprehensive model prediction results and professional input from doctors to continuously optimize patients' treatment plans and health management strategies. The system of this application forms a dynamically adjusted closed-loop management system through comprehensive model prediction results and professional input from doctors to continuously optimize patients' treatment plans and health management strategies. The system is highly scalable and can be applied to the course prediction and management of other chronic diseases. By adjusting model parameters and feature extraction methods, accurate prediction of different diseases can be achieved.

[0064] This application uses the constructed twin model of the disease process to make individualized predictions on the course of disease in patients with diabetes and coronary heart disease. By inputting the patient's personal information, medical history, laboratory test data and other data, the RNN-NBM-AFT model can predict the patient's disease development trend, treatment effect and the risk of complications. Based on the results of the course prediction, doctors can develop more personalized treatment plans for patients. For example, adjust the dosage and type of hypoglycemic drugs or anti-coronary heart disease drugs according to the results of the disease prediction to reduce the risk of complications. Through the prediction of the course of disease, more accurate health education and management suggestions can be provided to patients. For example, in response to the risk of complications that may appear in the predicted results, popularize relevant knowledge to patients in advance, strengthen self-management, and reduce the incidence of complications.

[0065] The above description is only an embodiment of the present application and does not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly used in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A method for constructing disease process twins based on standard electronic medical records, characterized in that: include: Based on the patient's standard FHIR electronic medical record, obtain electronic medical record data; Encoding and processing the electronic medical record data to obtain multivariate time series data; Based on the multivariate time series data, construct a disease process twin model; Obtaining a disease trajectory prediction report based on the output results and prediction graph of the disease process twin model; The GPT model is introduced to convert the numerical information in the disease trajectory prediction report and the semantic information of the knowledge graph into a natural language description to obtain a report template; adjusting the report template based on interactive interface verification and input of additional clinical parameter values; Based on the report template and the preset risk threshold, a personalized medical report including risk warning information and intervention suggestions is generated.

2. The method according to claim 1, characterized in that The electronic medical record data includes basic demographic information; medical history and diagnosis records; laboratory test results; treatment and medication records; lifestyle and behavior data; and hospitalization and surgery records.

3. The method according to claim 1, characterized in that The method of encoding and processing the electronic medical record data to obtain multivariate time series data includes: Parsing and structuring the text information extracted from the FHIR electronic medical record based on the GPT model to obtain key information including medical history, diagnosis records, laboratory test results, treatment and medication records, lifestyle and behavior data; The key information is arranged in chronological order and the medical data in the text information is converted into time series data to form multivariate time series data.

4. The method according to claim 3, characterized in that The method for constructing a disease process twin model comprises: Combining the RNN model with the NBM model to generate an RNN-NBM model for sequence modeling and prediction of the multivariate time series data; Based on the RNN-NBM model and the AFT model for time prediction of specific clinical events, a disease process twin model is obtained.

5. A disease process twin construction system based on standard electronic medical records, characterized in that: include: An electronic medical record encoding module, used to parse the text information in the FHIR electronic medical record through the GP model T, convert the text information into multivariate time series data, and perform feature extraction and structural processing; The sequence modeling and prediction module is used to compress and decompress the input data through an autoencoder to fill in missing data, and combine the RNN and NBM models to make high-precision predictions of disease trajectories; The knowledge graph generation module is used to combine the model prediction results and the medical knowledge graph to generate a disease trajectory prediction report; The medical report generation module converts the predicted numerical information and the semantic information of the knowledge graph into natural language descriptions by introducing the GPT model, and generates a complete personalized medical report; The risk warning module is used to generate risk warning information and provide personalized treatment plan adjustments and health management suggestions.

Images