Diabetes risk prediction model based on multi-mode large model

Abstract

The invention relates to the technical field of diabetes, in particular to a diabetes risk prediction model based on a multi-mode large model. The system comprises a multi-data fusion acquisition unit, a model construction selection unit and a user classification unit. According to gene detection and family history data, different models are allocated to predict the diabetes risk, it is ensured that the prediction result is closer to the actual situation of an individual, the family history data are combined, the individual risk can be evaluated more comprehensively, targeted health management suggestions are provided for family members, the family data are integrated into the models, and the risk of diabetes is predicted. The method is beneficial for discovering a family genetic mode and improving the prediction accuracy, and can improve the prediction accuracy through family history data even if no gene signal exists, so that the existing information is fully utilized.

Description

Technical Field

[0001] The present invention relates to the technical field of diabetes, and in particular to a diabetes risk prediction model based on a multimodal large model. Background Art

[0002] Diabetes is a group of clinical syndromes characterized by chronic hyperglycemia caused by the combined effects of genetic and environmental factors. It is a common metabolic disease and has become a global health problem. Diabetes is highly harmful and measures must be taken to prevent and manage it. In the prior art, a system for predicting diabetes risk, such as publication number CN111968748A, provides a modeling method for a diabetes complication prediction model, including the following steps: S10 data acquisition, collecting hospital diabetes complication data; S20 data processing, preprocessing the complication data to obtain modeling data; S30 using SPSS software to discretize the modeling data to obtain feature data; and S40 using 70% of the feature data to train the machine learning model, and using 30% of the feature data to test the machine learning model to obtain the diabetes complication prediction model. The modeling method of a diabetes complication prediction model of the present invention uses a large amount of existing electronic medical record data of domestic diabetic patients to establish a diabetes complication prediction model, which can improve the prediction effect of complications of domestic diabetic patients. The application of the diabetes complication prediction model to diabetes complication prediction can take appropriate measures for high-risk groups and assist doctors in diagnosis and treatment;

[0003] Currently, users are divided into two types: those who carry diabetes genes (with a family history) and those who do not carry diabetes genes. The content that needs to be predicted for different users is also different. However, the existing technology is not convenient for allocating prediction models according to the user's genetic conditions, resulting in the inability to carry out targeted risk prediction, affecting subsequent diagnosis and treatment. In view of this, we propose a diabetes risk prediction model based on a multimodal large model, which can predict disease progression and complications for users who carry diabetes genes, and predict future disease risks for users who do not carry diabetes genes. Summary of the invention

[0004] The purpose of the present invention is to provide a diabetes risk prediction model based on a multimodal large model. Different models are assigned to predict diabetes risk according to genetic testing and family history data. For users with a family history, family history data is combined to more comprehensively assess individual risks and provide targeted health management recommendations for family members.

[0005] To achieve the above-mentioned object, the present invention provides a diabetes risk prediction model based on a multimodal large model, comprising a multi-data fusion acquisition unit, a model construction selection unit and a user classification unit;

[0006] The multi-data fusion acquisition unit is used to mine historical data of diabetic patients, including historical data of diabetic patients with family history and historical data of diabetic patients without family history;

[0007] The model construction selection unit uses the historical data of the diabetic patients with a family history in the multi-data fusion acquisition unit as the initial training set to train the carrying prediction model for analyzing time series data and predicting blood sugar fluctuations and complication risks; and uses the historical data of the diabetic patients without a family history as the training set to train the normal prediction model for predicting future disease risks;

[0008] The user classification unit is used to collect current user data, including gene data, family history data, clinical data and lifestyle data, and obtain gene chips from the gene data to analyze the current user's diabetes status;

[0009] If the user carries the diabetes gene and the family history data reflects no family history, the current user data will be input into the carrier prediction model to predict blood sugar fluctuations and complication risks;

[0010] If the user carries a diabetes gene and the family history data reflects a family history, the family data corresponding to the current user data will be used as a training set to retrain the actual gene training model to predict blood sugar fluctuations and complication risks;

[0011] If the user does not carry the diabetes gene, and the family history data reflects a family history, the current user data is input into the actual carrier prediction model to predict blood sugar fluctuations and complication risks;

[0012] If the user does not carry the diabetes gene and the family history data reflects no family history, the current user data is input into the normal prediction model to predict future disease risk.

[0013] As a further improvement of the technical solution, the multi-data fusion acquisition unit includes a crawler mining module and a family history classification module;

[0014] The crawler mining module is used to crawl the disease data of all diabetic patients from medical records, and the disease data includes family history data, clinical data and lifestyle data;

[0015] The family history classification module is used to call out disease data, analyze whether there are disease records in the corresponding family history data for classified storage, and if there are disease records, the diabetic patients marked as having a family history are stored in the first database; if there are no disease records, the diabetic patients marked as having no family history are stored in the second database.

[0016] As a further improvement of the technical solution, the training carrying prediction model in the model construction selection unit is in the first database, and a recurrent neural network is used to train the historical data of diabetic patients with a family history, including the following steps:

[0017] Select the disease data corresponding to the diabetic patients marked as having a family history in the first database, clean the data, process missing values, normalize the data, and convert it into a format suitable for recurrent neural network input;

[0018] Extract periodic time features as the input layer, use RNN units as hidden layers to process sequence data, capture time dependencies, and design the output layer based on predicting blood sugar fluctuations and complication risks. The expression is:

[0019] h t =f(h t-1 ,x t ,W h ,b h )

[0020] y t =g(h t ,W y ,b y )

[0021] Among them, h t is the hidden state at time step t, and the previous hidden state h t-1 , x t is the input at time step t, W h and b h are the weights and biases of the hidden layer, f is the activation function, usually tanh or ReLU, y t is the output at time step t, W y and b y are the weight and bias of the output layer, and g is the activation function of the output layer.

[0022] As a further improvement of the technical solution, the normal prediction model is trained in the model construction selection unit in the second database, and a deep learning algorithm is used to train the historical data of diabetic patients without a family history, including the following steps:

[0023] Select the disease data corresponding to the diabetic patients marked as having no family history in the second database, clean the data, process the missing values, and normalize them, so that all numerical features are standardized into feature vectors;

[0024] A multi-layer perceptron model is constructed. The input layer accepts the patient's feature vector, and the hidden layer uses a nonlinear activation function to extract and transform the input data. Finally, the output layer uses a sigmoid activation function and outputs the disease risk probability, with a value between 0 and 1.

[0025] As a further improvement of the technical solution, the user classification unit includes a data collection module, a gene detection module, a model selection module and an update adjustment module;

[0026] The data collection module is used to collect the user's family history data through family medical history records, output family history conditions, including family history and no family history, obtain clinical data from the electronic health record system, and collect lifestyle data through a questionnaire filled out by the user;

[0027] The gene detection module is used to obtain the user's gene information through gene detection, and use known diabetes-related gene chips to perform targeted detection. If the diabetes gene exists, a gene signal is output; if the diabetes gene does not exist, a no gene signal is output;

[0028] The model selection module is used to simultaneously receive the family history and targeted detection results of the data collection module and the gene detection module, including the following gestures:

[0029] Posture 1: receiving the gene signal and no family history, inputting the clinical data and lifestyle data in the data collection module to carry the gene training model, and outputting the predicted blood sugar fluctuation and complication risk results;

[0030] Posture 2: receiving a gene signal and having a family history, transmitting a signal to an update adjustment module, for collecting family data of the current user's family members and inputting it into a first database, wherein the family data includes family clinical data and family lifestyle data, and using a recurrent neural network to train the family clinical data and lifestyle data again to form an actual gene-carrying training model, inputting the clinical data and family lifestyle data in the data collection module, and outputting predicted blood sugar fluctuations and complication risks;

[0031] Posture 3: receiving no gene signal and family history, inputting clinical data and lifestyle data in the data acquisition module into the actual gene-carrying training model, and outputting predicted blood sugar fluctuations and complication risks;

[0032] Posture 4: receiving no gene signal and no family history, inputting the clinical data and lifestyle data in the data collection module into the normal prediction model to predict future disease risks.

[0033] As a further improvement of the present technical solution, the update adjustment module collects family data of the current user's family members, adopts a pre-designed questionnaire, uses a text classification algorithm to classify the text data, and identifies clinical data and lifestyle data.

[0034] As a further improvement of the technical solution, the questionnaire is divided into two categories: clinical data and lifestyle data;

[0035] Clinical data: including but not limited to disease diagnosis and medical records;

[0036] Lifestyle data: including but not limited to eating habits, exercise habits, sleeping habits and emotional state.

[0037] As a further improvement of the present technical solution, the user classification unit also includes a node division module, which is used to define a window size, such as data from the past 3 months, and use the user data collected by the data acquisition module in the sliding window to input the model selection module for prediction. Whenever new blood glucose monitoring data is received, the window is moved forward one time step and the latest data is added to the window. Whenever the window is updated, the prediction result is updated based on the user's latest data and time changes.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] In this diabetes risk prediction model based on a multimodal large model, a deep learning algorithm is used to train the historical data of diabetic patients without a family history, and a recurrent neural network is used to train the historical data of diabetic patients with a family history, forming multiple prediction models. It is convenient to assign corresponding prediction models according to whether the user carries the diabetes gene and family history, and can provide individuals with more accurate health assessments and prevention suggestions. The prediction results can help doctors make decisions faster, which is conducive to subsequent early prevention and intervention.

[0040] In addition, a recurrent neural network is used to train the family clinical data and lifestyle data again to form an actual gene carrier training model. The clinical data and family lifestyle data in the data acquisition module are input to output predicted blood sugar fluctuations and complication risks to ensure that the prediction results are closer to the individual's actual situation. Combined with family history data, individual risks can be assessed more comprehensively, and targeted health management suggestions can be provided to family members. Integrating family data into the model will help discover family inheritance patterns and improve the accuracy of predictions. Even if there is no genetic signal, the accuracy of predictions can be improved through family history data, making full use of existing information. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is the overall structural principle block diagram of the present invention;

[0042] Figure 2 This is a principle block diagram of the multi-data fusion acquisition unit and model construction selection unit of the present invention.

[0043] The meaning of each number in the figure is:

[0044] 100. Multi-data fusion acquisition unit; 110. Crawler mining module; 120. Family history classification module;

[0045] 200, Model construction selection unit;

[0046] 300, user classification unit; 310, data collection module; 320, gene detection module; 330, model selection module; 340, update adjustment module. DETAILED DESCRIPTION

[0047] The following will be combined with the accompanying drawings in the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Example 1

[0049] See also Figure 1-Figure 2 As shown, this embodiment provides a diabetes risk prediction model based on a multimodal large model, including a multi-data fusion acquisition unit 100, a model construction selection unit 200 and a user classification unit 300;

[0050] The multi-data fusion acquisition unit 100 is used to mine historical data of diabetic patients, including historical data of diabetic patients with family history and historical data of diabetic patients without family history;

[0051] However, the multi-data fusion acquisition unit 100 includes a crawler mining module 110 and a family history classification module 120;

[0052] The crawler mining module 110 is used to crawl the disease data of all diabetic patients from the medical records. The disease data includes family history data, clinical data and lifestyle data. The disease data can be obtained by using web crawler technology, which is an automated tool for automatically crawling information on the Internet. It is used to extract data from websites. First, determine the source of the data to ensure that the data source is legal and has obtained the patient's authorization and consent. The data source may include but is not limited to hospital websites, electronic health record systems, research databases, etc. When accessing these data sources, it is necessary to comply with relevant laws and regulations to ensure the legal acquisition of data, set the basic parameters of the crawler, such as the start URL of the crawler, the target data type diabetic patient information, etc., generate or obtain the URLs of all relevant pages according to the website structure, send HTTP requests to each URL, obtain the web page content, use a parsing library such as BeautifulSoup or lxml to parse the web page content, extract the target data, and store the extracted data in a database or CSV file;

[0053] In addition, after obtaining disease data, you can also extract test reports and medical records from CV image recognition users, and use text matching algorithms to extract disease data of diabetic patients from test reports and medical records, including the following steps: locate and identify text areas from images, use deep learning models such as YOLO or Mask R-CNN to locate text areas, convert the detected text areas into editable text formats, such as Tesseract OCR, and extract key information from the recognized text. Suppose you have a database of test reports and want to extract specific disease information from it. You can first perform image preprocessing and OCR recognition on the test reports, and then use the above text matching algorithm to compare the similarity between each test report and keywords (such as "hypertension", "diabetes", etc.) to determine whether the test report contains information about related diseases;

[0054] The family history classification module 120 is used to call out disease data and analyze whether there are disease records in the corresponding family history data for classified storage. If there are disease records, the diabetic patients marked as having a family history are stored in the first database. If there are no disease records, the diabetic patients marked as having no family history are stored in the second database. During the analysis, it is necessary to determine which fields in the family history data can be used to determine whether the patient has a family history. For example, the fields may include the diabetes status of relatives such as "parents", "siblings", and "children". A standard is defined to judge the family history. For example: if the family history data contains at least one person with diabetes among "parents", "siblings", or "children", the patient is marked as having a family history. According to the classification logic, the data of each patient is classified and the results are stored in a database or CSV file.

[0055] The model building selection unit 200 uses the historical data of the diabetic patients with a family history in the multi-data fusion acquisition unit 100 as the initial training set to train the carrying prediction model. The carrying prediction model is conducive to predicting patients with a higher probability of suffering from diabetes, improving the risk prediction effect, and is used to analyze time series data and predict blood sugar fluctuations and complications. The historical data of the diabetic patients without a family history are used as the training set to train the normal prediction model for predicting future disease risks. The normal prediction model is conducive to predicting patients with a low probability of suffering from diabetes.

[0056] Further, the training carrying prediction model in the model construction selection unit 200 is in the first database, and a recurrent neural network is used to train the historical data of patients with diabetes with a family history, including the following steps:

[0057] Select the disease data corresponding to the diabetic patients marked as having a family history in the first database, clean the data, process missing values, normalize the data, and convert it into a format suitable for recurrent neural network input;

[0058] Extract periodic time features as the input layer, use RNN units as hidden layers to process sequence data, capture time dependencies, and design output layers based on predicted blood sugar fluctuations and complication risks. For example, blood sugar fluctuation prediction can be a continuous value output, and complication risk prediction is a classification output. The expressions are:

[0059] h t =f(h t-1 ,x t ,W h ,b h )

[0060] y t =g(h t ,W y ,b y )

[0061] Among them, h t is the hidden state at time step t, and the previous hidden state h t-1 , x t is the input at time step t, W h and b h are the weights and biases of the hidden layer, f is the activation function, usually tanh or ReLU, y t is the output at time step t, W y and b yare the weight and bias of the output layer, respectively, and g is the activation function of the output layer. When predicting blood sugar fluctuations: the input data includes historical blood sugar levels, timestamps, possible genetic factors, etc. RNN can learn the change pattern of blood sugar levels over time, including trends, periodicity and seasonality, through its cyclic structure, and the updated hidden state h t contains information about the history of blood glucose fluctuations, which is used to generate the blood glucose prediction y for the next time step t ;

[0062] When predicting the risk of complications: the input data may include blood sugar level, blood pressure, weight, family history, genotype, etc. RNN can analyze the relationship between these data and identify patterns related to the risk of diabetes complications. By learning these patterns, RNN can predict the likelihood of an individual developing complications in the future and output the corresponding risk score y t .

[0063] In addition, the training normal prediction model in the model construction selection unit 200 is in the second database, and the deep learning algorithm is used to train the historical data of diabetic patients without family history, and the historical data of diabetic patients without family history includes demographic information (age, gender, race, etc.), lifestyle (diet, exercise, etc.), medical history (blood pressure, cholesterol, etc.) and genetic information (such as single nucleotide polymorphism SNP), including the following steps:

[0064] Select the disease data corresponding to the diabetic patients marked as having no family history in the second database, clean the data, process the missing values, and normalize them, so that all numerical features are standardized into feature vectors;

[0065] Construct a multi-layer perceptron model, which accepts the patient's feature vector in the input layer, extracts and transforms the input data through nonlinear activation functions (such as ReLU) in the hidden layer. Multiple hidden layers can learn complex nonlinear relationships. Finally, the output layer uses the sigmoid activation function, and the output is the disease risk probability, which is between 0 and 1.

[0066] Multilayer Perceptron (MLP) is a feedforward neural network consisting of multiple layers, including input layer, hidden layer and output layer. Each layer consists of multiple neurons, which are connected by weights. The information transmission direction is unidirectional, starting from the input layer, passing through the hidden layer, and finally reaching the output layer. It is applied to the risk prediction of diabetes without family history.

[0067] The user classification unit 300 is used to collect current user data, which includes gene data, family history data, clinical data and lifestyle data, and obtain gene chips from the gene data to analyze the current user's diabetes status;

[0068] If the user carries a diabetes gene and the family history data shows that there is no family history (because random mutations may occur in the gene transmission process, causing the offspring to carry genes that have never appeared in the family, including diabetes-related genes, and the parental genes will recombine when passed to the children, the gene combination obtained by the children may be a completely new combination of the parents' genes. Even if the parents do not carry the diabetes gene, the children may obtain the gene through gene recombination), the current user data will be input into the carrier prediction model to predict blood sugar fluctuations and complications.

[0069] If the user carries a diabetes gene and the family history data reflects a family history (since family members share the same genes, if a parent or grandparent has diabetes, the possibility of the child carrying the diabetes gene will increase significantly), the family data corresponding to the current user data will be used as a training set to retrain the actual gene training model to predict blood sugar fluctuations and complication risks;

[0070] If the user does not carry the diabetes gene, and the family history data shows that there is a family history (even if the user does not carry the diabetes gene, the presence of a family history may increase the risk of the disease, for example, family members may carry other genetic variants related to diabetes, or the lifestyle of family members may affect the individual's risk of disease), the current user data will be input into the actual carrying prediction model to predict blood sugar fluctuations and complications.

[0071] If the user does not carry the diabetes gene and the family history data reflects no family history (the individual may not carry the gene related to diabetes and there is no diabetes patient in the family history, so the risk of disease is low, but if the individual does not have good living habits, such as a balanced diet, regular exercise, weight control, etc., the risk of disease will also increase), the current user data will be input into the normal prediction model to predict future disease risks.

[0072] It is worth noting that the user classification unit 300 includes a data collection module 310, a gene detection module 320, a model selection module 330 and an update adjustment module 340;

[0073] The data collection module 310 is used to collect the user's family history data through family medical history records, output family history conditions, including whether there is a family history or not, obtain clinical data from the electronic health record system, including medical records such as diagnosis, treatment, and examination results, etc., and collect lifestyle data through questionnaires filled out by users, including eating, exercise, sleep, smoking, drinking and other habits;

[0074] The gene detection module 320 is used to obtain the user's gene information through gene detection, and use a known diabetes-related gene chip for targeted detection. If the diabetes gene exists, a gene signal is output, and if the diabetes gene does not exist, a no gene signal is output. The gene chip is a high-throughput gene detection technology. Its principle is to fix a large number of known gene fragment probes related to a specific disease on the chip, and then hybridize the user's DNA sample with the chip. If there is a gene sequence in the sample that matches the probe on the chip, a fluorescent signal will be generated, thereby determining whether the user carries the relevant gene. The diabetes-related gene chip contains a large number of genes related to diabetes. Multiple gene fragments related to the onset of diabetes, such as: genes related to insulin secretion: TCF7L2, KCNJ11, SLC30A8, INS, etc.; genes related to insulin resistance: PPARG, IRS1, LEP, LEPR, etc.; genes related to glucose metabolism: GCK, HNF1A, HNF4A, SLC2A2, etc.; during the test, the user's blood or saliva sample is collected, DNA is extracted from the sample, the user's DNA sample is hybridized with the diabetes-related gene chip, a fluorescence scanner is used to detect the fluorescence signal on the chip, the fluorescence signal intensity is analyzed, and it is determined whether the user carries the diabetes-related gene;

[0075] The model selection module 330 is used to simultaneously receive the family history and targeted detection results from the data collection module 310 and the gene detection module 320, including the following gestures:

[0076] Posture 1: receiving gene signals and no family history, inputting clinical data and lifestyle data in the data acquisition module 310 to carry the gene training model, outputting the predicted blood sugar fluctuation and complication risk results, and by analyzing the gene signals of individuals without family history, more accurately predicting their blood sugar fluctuation and complication risks, and realizing personalized prevention;

[0077] Posture 2: receiving a gene signal and having a family history, transmitting a signal to the update adjustment module 340, for collecting family data of the current user's family members and inputting it into the first database, wherein the family data includes family clinical data and family lifestyle data, and using a recurrent neural network to train the family clinical data and lifestyle data again to form an actual gene-carrying training model, inputting the clinical data and family lifestyle data in the data collection module 310, and outputting predicted blood sugar fluctuations and complication risks, using the actual gene-carrying training model to ensure that the prediction results are closer to the actual individual situation;

[0078] Posture 3: receiving no gene signal and family history, inputting clinical data and lifestyle data in the data acquisition module 310 into the actual gene-carrying training model, outputting predicted blood sugar fluctuation and complication risk, combining with family history data, more comprehensive assessment of individual risk, providing targeted health management advice for family members, integrating family data into the model, helping to discover family inheritance patterns and improve the accuracy of prediction, even without gene signals, the accuracy of prediction can be improved through family history data, making full use of existing information;

[0079] Posture 4: receiving no gene signal and no family history, inputting clinical data and lifestyle data in the data collection module 310 into the normal prediction model to predict future disease risk. For individuals with no family history and no gene signal, future disease risk can be predicted and health intervention can be carried out in advance;

[0080] Therefore, by integrating genetic testing and family history data and combining them with advanced machine learning models, it is possible to provide individuals with more accurate health assessments and prevention recommendations, which will help improve medical quality, reduce medical costs, and enhance patients' health awareness. Predictive results can help doctors make decisions faster and reduce unnecessary examinations and treatments. Early prevention and intervention can reduce the occurrence of complications, thereby reducing long-term medical costs.

[0081] Furthermore, the update adjustment module 340 collects family data of the current user's family members, uses a pre-designed questionnaire, and uses a text classification algorithm to classify the text data to identify clinical data and lifestyle data, wherein the text classification algorithm includes the following steps:

[0082] Convert the text into a bag-of-words model, use the frequency of each word as a feature, use naive Bayes to divide the data set into training set, validation set, and test set, use the training set data to train the selected model, and use the relevant functions in the scikit-learn library, which is suitable for text classification, especially when the number of categories is large and the feature dimension is high, use the trained model to classify new text data.

[0083] Specifically, the questionnaire was divided into two categories: clinical data and lifestyle data, and contained multiple multiple-choice questions, fill-in-the-blank questions, and rating questions;

[0084] Clinical data: including but not limited to disease diagnosis and medical records. These data are usually given in text form and may require text analysis using natural language processing (NLP) technology to extract key information such as disease name, diagnosis date, treatment plan, etc.

[0085] Lifestyle data: including but not limited to eating habits, exercise habits, sleeping habits and emotional states. These data are usually given in the form of multiple-choice questions, fill-in-the-blank questions, and scoring questions, and can be directly converted into numerical data for analysis.

[0086] Embodiment 2:

[0087] By integrating genetic testing and family history data, the model can provide individuals with more accurate diabetes risk assessment, including blood sugar fluctuation prediction and complication risk prediction, as well as future disease risk prediction. This helps doctors and patients to manage their health more effectively and reduce unnecessary medical interventions. However, when predicting, there will be deviations with the user's actual situation. If the time period for collecting user data for prediction is long, it will also affect the prediction results. Therefore, the user classification unit 300 also includes a node division module, which is used to define a window size, such as data from the past 3 months. The user data collected by the data collection module 310 in the sliding window is input into the model selection module 330 for prediction. Whenever new blood sugar monitoring data is received, the window is moved forward by a time step and the latest data is added to the window. Whenever the window is updated, the prediction results are updated according to the user's latest data and time changes, such as regularly collecting the user's blood sugar monitoring data and based on the latest data.

[0088] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A diabetes risk prediction model based on a multimodal large model, characterized by: It comprises a multi-data fusion collection unit (100), a model building selection unit (200) and a user classification unit (300); The multi-data fusion acquisition unit (100) is used to mine historical data of diabetic patients, the history including historical data of diabetic patients with family history and historical data of diabetic patients without family history; The model construction selection unit (200) uses the historical data of the diabetic patients with a family history in the multi-data fusion acquisition unit (100) as an initial training set to train a carrying prediction model for analyzing time series data and predicting blood sugar fluctuations and complication risks; and uses the historical data of the diabetic patients without a family history as a training set to train a normal prediction model for predicting future disease risks; The user classification unit (300) is used to collect current user data, which includes gene data, family history data, clinical data and lifestyle data, and obtain a gene chip from the gene data to analyze the current user's diabetes status; If the user carries the diabetes gene and the family history data reflects no family history, the current user data will be input into the carrier prediction model to predict blood sugar fluctuations and complication risks; If the user carries a diabetes gene and the family history data reflects a family history, the family data corresponding to the current user data will be used as a training set to retrain the actual gene training model to predict blood sugar fluctuations and complication risks; If the user does not carry the diabetes gene, and the family history data reflects a family history, the current user data is input into the actual carrier prediction model to predict blood sugar fluctuations and complication risks; If the user does not carry the diabetes gene and the family history data reflects no family history, the current user data is input into the normal prediction model to predict future disease risk.

2. The diabetes risk prediction model based on a multimodal large model according to claim 1, characterized in that: The multi-data fusion collection unit (100) comprises a crawler mining module (110) and a family history classification module (120); The crawler mining module (110) is used to crawl the disease data of all diabetic patients from medical records, wherein the disease data includes family history data, clinical data and lifestyle data; The family history classification module (120) is used to retrieve disease data, analyze whether there is a disease record in the corresponding family history data for classification and storage; if there is a disease record, the diabetic patients are marked as having a family history and stored in the first database; if there is no disease record, the diabetic patients are marked as having no family history and stored in the second database.

3. The diabetes risk prediction model based on a multimodal large model according to claim 2, characterized in that: The model construction selection unit (200) trains the prediction model in the first database, using a recurrent neural network to train historical data of patients with family history of diabetes, including the following steps: Select the disease data corresponding to the diabetic patients marked as having a family history in the first database, clean the data, process missing values, normalize the data, and convert it into a format suitable for recurrent neural network input; Extract periodic time features as the input layer, use RNN units as hidden layers to process sequence data, capture time dependencies, and design the output layer based on predicting blood sugar fluctuations and complication risks. The expression is: h t =f(h t-1 ,x t ,W h ,b h ) y t =g(h t ,W y ,b y ) Among them, h t is the hidden state at time step t, and the previous hidden state h t-1 , x t is the input at time step t, W h and b h are the weights and biases of the hidden layer, f is the activation function, usually tanh or ReLU, y t is the output at time step t, W y and b y are the weight and bias of the output layer, and g is the activation function of the output layer.

4. The diabetes risk prediction model based on a multimodal large model according to claim 3, characterized in that: The normal prediction model is trained in the model construction selection unit (200) in the second database, and a deep learning algorithm is used to train the historical data of diabetic patients without family history, including the following steps: Select the disease data corresponding to the diabetic patients marked as having no family history in the second database, clean the data, process the missing values, and normalize them, so that all numerical features are standardized into feature vectors; A multi-layer perceptron model is constructed. The input layer accepts the patient's feature vector, and the hidden layer uses a nonlinear activation function to extract and transform the input data. Finally, the output layer uses a sigmoid activation function and outputs the disease risk probability, with a value between 0 and 1.

5. The diabetes risk prediction model based on multimodal large model according to claim 4, characterized in that: The user classification unit (300) includes a data collection module (310), a gene detection module (320), a model selection module (330) and an update adjustment module (340); The data collection module (310) is used to collect the user's family history data through family medical history records, output family history conditions, including family history and no family history, obtain clinical data from the electronic health record system, and collect lifestyle data through a questionnaire filled out by the user; The gene detection module (320) is used to obtain the user's gene information through gene detection, and use a known gene chip related to diabetes to perform targeted detection. If the diabetes gene exists, a gene signal is output; if the diabetes gene does not exist, a gene-free signal is output; The model selection module (330) is used to simultaneously receive the family history and targeted detection results of the data collection module (310) and the gene detection module (320), including the following gestures: Posture 1: receiving the gene signal and no family history, inputting the clinical data and lifestyle data in the data acquisition module (310) to carry the gene training model, and outputting the predicted blood sugar fluctuation and complication risk results; Posture 2: receiving a gene signal and having a family history, transmitting a signal to an update adjustment module (340), for collecting family data of the current user's family members and inputting it into a first database, wherein the family data includes family clinical data and family lifestyle data, and using a recurrent neural network to train the family clinical data and lifestyle data again to form an actual gene-carrying training model, inputting the clinical data and family lifestyle data in the data collection module (310), and outputting predicted blood sugar fluctuations and complication risks; Posture 3: receiving a no gene signal and a family history, inputting the clinical data and lifestyle data in the data acquisition module (310) into the actual gene-carrying training model, and outputting the predicted blood sugar fluctuation and complication risk; Posture 4: receiving no gene signal and no family history, inputting the clinical data and lifestyle data in the data acquisition module (310) into the normal prediction model to predict future disease risk.

6. The diabetes risk prediction model based on multimodal large model according to claim 5, characterized in that: The updating and adjusting module (340) collects family data of the current user's family members, uses a pre-designed questionnaire, and uses a text classification algorithm to classify the text data to identify clinical data and lifestyle data.

7. The diabetes risk prediction model based on a multimodal large model according to claim 6, characterized in that: The questionnaire was divided into two categories: clinical data and lifestyle data; Clinical data: including but not limited to disease diagnosis and medical records; Lifestyle data: including but not limited to eating habits, exercise habits, sleeping habits and emotional state.

8. The diabetes risk prediction model based on multimodal large model according to claim 5, characterized in that: The user classification unit (300) also includes a node division module, which is used to define a window size, use the user data collected by the data acquisition module (310) in the sliding window to input the model selection module (330) for prediction, and whenever new blood glucose monitoring data is received, the window is moved forward by a time step and the latest data is added to the window. Whenever the window is updated, the prediction result is updated according to the user's latest data and time changes.

Images