Multi-user group precise auxiliary diagnosis and treatment method and device for obesity diseases

Abstract

The invention provides a multi-user group precise auxiliary diagnosis and treatment device and method for obesity. The device comprises a data interaction panel, a function module based on interpretable machine learning and a consultation question and answer function module. The function module based on interpretable machine learning comprises an index variable set, a training data set and an XGBoost model; a user selects a response variable and a plurality of explanatory variables from the index variable set through the data interaction panel, and selects a certain scale of training data from the training data set; an XGBoost model is trained on the training data, a statistical result is output, and interpretability analysis is carried out on the statistical result; the data interaction panel displays the statistical result and the interpretability analysis thereof; and the consultation question and answer function module interprets a statistical result output by the XGBoost model and interpretability analysis thereof based on a large language model, and answers the problem about obesity of the user. The statistical method is integrated to the diagnosis and treatment process, so that the diagnosis and treatment pertinence of medical personnel can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and statistical technology, and in particular to a multi-user group precise auxiliary diagnosis and treatment method and device for obesity. Background Art

[0002] In recent years, the prevalence of overweight and obese people has continued to rise. As an independent chronic disease and an important pathogenic factor of many chronic diseases, obesity has become a major public health issue. Related diseases caused by obesity include abnormal blood sugar, abnormal blood lipids, hypertension, cardiovascular disease, mental and psychological abnormalities and other diseases, which can seriously endanger our physical health.

[0003] The World Health Organization (WHO) defines obesity as abnormal or excessive fat accumulation that has adverse effects on health. 2 ) is a universal standard for assessing systemic obesity. The index uses height to correct weight to reduce the impact of height on obesity assessment. It is calculated as: weight (kg) divided by height (m) squared.

[0004] However, using BMI as an indicator for assessing obesity has certain limitations. As people age, lean body mass (fat-free body mass) gradually decreases and body fat content gradually increases. Therefore, young people and the elderly with the same BMI will have different body fat ratios. At the same BMI level, people who regularly engage in high-intensity physical activities and professional athletes usually have lower body fat ratios than the general population.

[0005] The current diagnosis and treatment of obesity is mainly based on doctors' experience, with low efficiency in the use of hospital internal case databases and low utilization of statistical methods. As a disease affected by multiple factors, the diagnosis of obesity usually requires patients to provide multi-dimensional physical examination data. Therefore, if medical personnel only rely on experience to diagnose specific individuals, it is easy to cause misdiagnosis or ignore indicators that have important effects on obesity in a specific individual.

[0006] In diagnosis and treatment practice, even if relatively mature statistical tools are provided, medical personnel are generally unable to make good use of them due to insufficient statistical knowledge.

[0007] As of the issuance of this application, the main integrated technologies of the relevant diagnosis and treatment assistance methods in similar patents include knowledge graphs (for example, the Chinese patent document with publication number CN 118471486 A discloses an auxiliary diagnosis and treatment system and method based on a knowledge graph) and large language models (for example, the Chinese patent document with publication number CN 118471488A discloses a diagnosis and treatment assistance device and a diagnosis and treatment assistance method based on a large language model). There is no solution to combine explainable machine learning with large language models, the efficiency of using large amounts of medical data is not high, and the reference significance for medical diagnosis is relatively limited. Summary of the invention

[0008] The present invention provides a multi-user group precise auxiliary diagnosis and treatment method and device for obesity. By integrating statistical methods into the diagnosis and treatment process, it can effectively improve the targeted diagnosis and treatment of medical personnel and reduce medical diagnosis errors.

[0009] The technical solution of the present invention is as follows:

[0010] A multi-user group precision auxiliary diagnosis and treatment device for obesity, including a data interaction panel, a function module based on explainable machine learning, and a consultation question and answer function module;

[0011] The functional module based on interpretable machine learning includes an indicator variable set, a training data set, and an XGBoost model;

[0012] The user selects a response variable and several explanatory variables from the indicator variable set through the data interaction panel, and selects a certain scale of training data from the training data set; trains the XGBoost model on the training data, outputs the statistical results and performs interpretability analysis on the statistical results; the data interaction panel displays the statistical results and their interpretability analysis;

[0013] The consultation and question-answering function module interprets the statistical results and interpretability analysis output by the XGBoost model based on a large language model, and answers users' questions about obesity.

[0014] Preferably, the method for constructing the indicator variable set includes: selecting indicator variables related to obesity based on the experience and judgment of medical personnel to construct the indicator variable set.

[0015] The indicator variables related to obesity include physical examination indicator variables, serum insulin measurement indicator variables, glucose measurement indicator variables, serum C-peptide measurement indicator variables, glycosylated hemoglobin measurement indicator variables, five liver and kidney function / biochemistry indicator variables, seven thyroid function test indicator variables, vitamin D measurement, seven sex hormone measurement indicator variables, high-sensitivity CRP measurement (blood routine test) indicator variables, urine routine test indicator variables, hepatitis B qualitative + hepatitis C test indicator variables, diabetes autoantibody detection indicator variables, liver B-ultrasound, carotid artery B-ultrasound, uterine B-ultrasound, ovarian B-ultrasound, testicular B-ultrasound, adrenal B-ultrasound indicator variables, questionnaire survey indicator variables, bone density examination indicator variables.

[0016] Furthermore, the indicator variables related to obesity diseases include at least fasting blood glucose, glycosylated hemoglobin, triglycerides, cholesterol, systolic blood pressure, diastolic blood pressure, acanthosis nigricans, and uric acid.

[0017] Preferably, the method for constructing a training data set includes: selecting a number of pieces of relevant data of obesity patients from a built-in database of a medical institution to establish an initial data set; and preprocessing the data in the initial data set to obtain a training data set.

[0018] The data in the initial data set were preprocessed, including: renaming the indicator names in the initial data set to standard indicator names summarized by medical personnel, unifying the units of the indicators, and deleting abnormal values ​​in the indicator variables that did not conform to medical knowledge.

[0019] The response variable is the variable that needs to be predicted, and the explanatory variable is the variable used to predict the response variable.

[0020] Select a certain amount of training data from the training data set to train the XGBoost model. The larger the training data size, the better the model fits the data, but the training time will also increase accordingly.

[0021] The training data is randomly divided into a training set and a test set in a ratio of 7:3. The XGBoost model is trained on the training set, and the prediction performance of the XGBoost model is tested on the test set.

[0022] Preferably, 5-fold cross validation is used to select the model parameters of XGBoost.

[0023] Preferably, the output of the interpretability analysis includes:

[0024] (a) The distribution of the values ​​of the response and explanatory variables, differentiated by gender;

[0025] (b) global interpretability, including feature importance and cumulative local effect plots;

[0026] (c) Local interpretability, including local SHAP plots.

[0027] The consultation question and answer function module includes:

[0028] The filter module removes characters in user questions that cannot be recognized by the large language model and filters questions that are not related to obesity, and then inputs them into the retrieval enhancement module;

[0029] The retrieval enhancement module searches for fields related to the user's question from relevant authoritative data sets through a retrieval enhancement search method, and inputs the user's question and related fields into the large language model together;

[0030] The memory module stores the user's previous chat records and inputs the chat records into the large language model when the user asks subsequent questions;

[0031] The large language model outputs relevant answers to user questions based on user questions, related fields, and previous chat records.

[0032] The filter module uses Python's regular expression to remove characters in user questions that cannot be recognized by the large language model, and uses the text classification model to filter questions that are not related to obesity.

[0033] The method for constructing the text classification model includes:

[0034] A number of labeled question samples are constructed by using the LLM model and manually reviewing professional data on obesity, wherein the question samples include question samples related to obesity and question samples unrelated to obesity;

[0035] The DistilBERT model is trained using the labeled question samples to obtain a text classification model.

[0036] The training methods for large language models include:

[0037] (i) Collect professional journals and professional website materials and convert them into text format to construct relevant authoritative datasets; use ChatGPT with custom prompt words to expand the declarative text paragraphs in the relevant authoritative datasets into several question-answer pairs to construct a fine-tuning dataset;

[0038] (ii) The LLaMA model is trained using the fine-tuning dataset to obtain a large language model.

[0039] The present invention also provides a method for performing auxiliary diagnosis and treatment based on the auxiliary diagnosis and treatment device, comprising:

[0040] When diagnosing obesity, medical personnel select one indicator variable from the indicator variable set as the response variable and select several indicator variables as explanatory variables;

[0041] Select a certain scale of training data from the training data set;

[0042] Train the XGBoost model on the training data, output the statistical results and perform interpretable analysis on the statistical results; the data interaction panel displays the statistical results and their interpretable analysis;

[0043] The consultation and question-answering function module interprets the statistical results and interpretability analysis output by the XGBoost model based on a large language model, and answers users' questions about obesity.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] (1) The current diagnosis and treatment of obesity is mainly based on doctors' experience, with low efficiency in utilizing the hospital's internal case database and low utilization of statistical methods. Obesity is a disease affected by multiple factors, and its diagnosis usually requires patients to provide multi-dimensional physical examination data. Therefore, if medical personnel only rely on experience to diagnose specific individuals, it is easy to cause misdiagnosis or ignore indicators that have important effects on obesity in a specific individual. The diagnostic and treatment device provided by the present invention can effectively improve the targeted diagnosis and treatment of medical personnel and reduce medical diagnostic errors by integrating statistical methods into the diagnostic and treatment process.

[0046] (2) In diagnosis and treatment practice, even if relatively mature statistical tools are provided, medical personnel are generally unable to make good use of them due to insufficient statistical knowledge. The diagnosis and treatment device provided by the present invention converts highly professional statistical results into easy-to-understand sentences through the application of a large language model, so that medical personnel can understand them, thereby improving the diagnosis and treatment level. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the multi-user group precise auxiliary diagnosis and treatment method for obesity of the present invention;

[0048] Figure 2 A schematic diagram of the construction process of functional modules based on explainable machine learning;

[0049] Figure 3 Schematic diagram of the process of training the LLaMA model;

[0050] Figure 4 A flowchart for training a text classification model;

[0051] Figure 5 This is a diagram of the user question and answer process. DETAILED DESCRIPTION

[0052] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0053] like Figure 1 As shown, the multi-user group precise auxiliary diagnosis and treatment method for obesity of the present invention is based on the multi-user group precise auxiliary diagnosis and treatment device for obesity, and the multi-user group precise auxiliary diagnosis and treatment device for obesity includes the following functional modules:

[0054] 1. Based on the functional modules of interpretable machine learning, the construction process is as follows Figure 2 As shown, including:

[0055] (1) Based on the experience of medical personnel, 120 indicator variables related to obesity were selected (height, weight, BMI, waist circumference, hip circumference, waist-to-hip ratio, waist-to-height ratio, systolic blood pressure, diastolic blood pressure, instant heart rate, acanthosis nigricans, testicular volume (male), breast stage (female), pubic hair in physical examination; serum insulin (0'), serum insulin (30'), serum insulin (60'), serum insulin (90'), serum insulin (120'), fasting insulin in serum insulin measurement; blood glucose (0'), blood glucose (30'), blood glucose (60'), blood glucose (90') in glucose measurement). 0'), blood glucose (120'), fasting blood glucose; serum C-peptide (0'), serum C-peptide (30'), serum C-peptide (60'), serum C-peptide (90'), serum C-peptide (120'), fasting serum C-peptide in serum C-peptide determination; glycosylated hemoglobin in glycosylated hemoglobin determination; lipoprotein A, triglycerides, total cholesterol, apolipoprotein A, apolipoprotein B, alanine aminotransferase / alanine aminotransferase, aspartate aminotransferase / glutamyl aminotransferase, glutamyl transpeptidase, uric acid, serum ferritin, total bile acid, total bilirubin, serum albumin, serum calcium, serum magnesium, serum phosphorus in liver and kidney function / biochemical five items; vitamin 25-hydroxyvitamin D in vitamin D determination; thyrotropin-releasing hormone, thyroxine, triiodothyronine in 7 thyroid function tests; estradiol, follicle-stimulating hormone, luteinizing hormone, prolactin, testosterone in 7 sex hormone tests; hepatitis antigens in hepatitis B qualitative + hepatitis C; diabetes autoantibody test; white blood cells, neutrophil percentage, lymphocyte percentage, monocyte percentage, platelets, hemoglobin, high-sensitivity C-reactive protein in high-sensitivity CRP determination (blood routine); IL-2, IL-4, IL-6, IL-10, TNF, IFN-R in cytokine flow cytometry; urine routine Alpha 1-microglobulin, microalbumin in urine, immunoglobulin G in urine, protein in urine / creatinine in urine; fatty liver in liver B-ultrasound; thickness of left common carotid artery in carotid artery B-ultrasound; long diameter of uterus, wide diameter of uterus, thickness of uterus in uterus B-ultrasound; long diameter of left ovary, long diameter of right ovary, wide diameter of left ovary, wide diameter of right ovary, thickness of left ovary, thickness of right ovary in ovarian B-ultrasound; long diameter of left testicle, long diameter of right testicle, wide diameter of left testicle, wide diameter of right testicle, thickness of left testicle, thickness of right testicle in testicular B-ultrasound; adrenal gland in adrenal B-ultrasound; bone mineral content, body fat percentage, bone density in bone density examination;The questionnaire survey included children's dietary intake, children's snack intake, sugary drink intake, mother's height, mother's weight, mother's BMI, father's height, father's weight, father's BMI, skinfold thickness, heart rate, serum vitamin D, Turner stage of penis (male), Turner stage of breast (female), Turner stage of pubic hair, age of menarche, age of nocturnal emission, number of exercise days per week, number of exercise hours per day, number of sitting hours per day, number of sleep hours per day, number of face screen time hours per day), which constituted the indicator variable set. Based on the built-in database of the medical institution, 11,000 patient-related data were selected to establish the initial data set. ;

[0056] (2) Preprocessing the initial data set includes renaming the indicator names in the initial data set to standard indicator names summarized by medical personnel, unifying the units of the indicators, and deleting abnormal values ​​in the indicator variables that do not conform to medical knowledge to obtain the model training data set.

[0057] (3) To obtain effective diagnosis and treatment assistance, the specific training and operation methods of the model are as follows:

[0058] When diagnosing obesity, medical personnel need to first select a response variable (the variable to be predicted, such as body weight) and several explanatory variables (the variables used for prediction, such as uric acid) that the indicator variables focus on, and then select an appropriate training data set size. The larger the data set size, the better the model fits the data, but the training time will also increase accordingly. After completing the selection, the model training data set is randomly divided into a training set and a test set in a ratio of 7:3.

[0059] Use 5-fold cross validation to select XGBoost model parameters (such as nround, max.depth, eta). Finally, train the XGBoost model on the training set and test the model prediction performance on the test set.

[0060] The trained model generates outputs that are returned to the front end for interpretable analysis. This is to visualize why the model makes specific predictions, help medical professionals better understand how the model makes predictions, and enhance the transparency and credibility of the model. Models with good prediction performance need to be analyzed using interpretable machine learning methods. Model outputs include:

[0061] 1) The distribution of the response variable and explanatory variable values, differentiated by gender.

[0062] 2) For global interpretability, feature importance and cumulative local effect (ALE) plots are employed.

[0063] Feature Importance: Calculate the contribution of each feature to the final prediction. You can choose the number of times a feature appears in all tree nodes or the average gain it can bring when the feature is used to split a node as a measure of contribution.

[0064] Users can toggle to see how important the selected explanatory variables are to the predicted response variable.

[0065] The ALE plot can show the local impact of a single feature on model prediction, excluding the impact of possible interactions between features.

[0066] Users can switch to view the impact of the value of any selected explanatory variable on the response variable predicted by the model (including confidence interval), and support zooming in on the area for viewing.

[0067] 3) For local interpretability, local SHAP graphs are used.

[0068] The local SHAP graph explains the feature contribution of the model in the prediction process for a single patient, and analyzes the data in a more targeted manner.

[0069] Personal SHAP graph, users can switch to view the SHAP contribution value of each explanatory variable for any patient to determine the most important variables for prediction, and can view a simple explanation of the graph. This section supports medical staff to provide targeted explanations for the causes of obesity in specific patients.

[0070] The construction of the data analysis panel can run the XGBoost model on the background server and integrate multiple interpretable machine learning methods (feature importance, ALE diagram and personal SHAP diagram) for prediction and visual interpretation.

[0071] The parameters that users can select include: the number of data set samples, response variables, explanatory variables, and histogram step size. Users can freely select the response variables they want to predict and any combination of explanatory variables of interest, and the background can fit the model for prediction and data analysis. The response variables and explanatory variables can be searched and selected by entering keywords.

[0072] Both the response variables and explanatory variables were mapped based on the indicators in the standard indicator set related to obesity established by professional doctors and the examination indicators in the database to ensure the consistency and correctness of the indicators.

[0073] The data analysis results are displayed in four small panels. Panel 1 is the data overview. Users can switch to view the distribution of any selected response variable and explanatory variable values, distinguished by gender. Panel 2 is the feature importance. Users can switch to view the importance of the selected explanatory variable for prediction. Panel 3 is the ALE chart. Users can switch to view the trend (including confidence interval) of the value of any selected explanatory variable on the model prediction, and support zooming in on the area. Panel 4 is the personal SHAP chart. Users can switch to view the SHAP contribution value of each explanatory variable of any patient to determine the most important variable for prediction, and can view a simple explanation of the chart.

[0074] The data set acquisition is changed from file import to real-time query of MySQL database. The corresponding inspection indicators are queried from the database in real time according to the explanatory variables and response variables selected on the page.

[0075] 2. Consultation and question-answering function module, the construction process is as follows:

[0076] In order to facilitate medical personnel to understand the statistical results, the present invention integrates the LLM (Large Language Model) large language model to interpret the statistical model.

[0077] Methods to improve LLM's professionalism include fine-tuning instructions, RAG (Retrival Augmented Generation), and input data set text adjustment. LLM will provide two different dialogue modes, professional medical consultation and public medical consultation, as well as two different versions in Chinese and English, based on the user's attributes. Among them, professional medical consultation provides result interpretation for doctors and other professional medical personnel to assist in medical diagnosis; the public medical consultation function serves patients and their families, providing them with easy-to-understand result interpretation.

[0078] In addition to explaining the interpretable machine learning results, LLM also provides a general obesity knowledge question-and-answer service, where patients and medical staff can learn about obesity-related information in the form of conversations. In terms of usage, LLM has a variety of ways to improve the user experience. The large language model is used in three places in the integrated software, which is used to interpret medical reports through command fine-tuning, to have convenient and fast chats with users, and to use multiple modes for comprehensive chat inquiries.

[0079] The data set used for fine-tuning of the LLM of the present invention includes several relevant authoritative books and dozens of academic journals. The main part of the LLM model of the present invention is the LLaMA series model (the latest version is LLaMA3.1), on which multiple modules such as retrieval enhancement search and question filtering are integrated. The training of the model adopts the lora method, and the parameters of the original model are fine-tuned by adding a low-rank matrix.

[0080] like Figure 3 As shown, the training of the LLaMA model includes:

[0081] (1) In order to obtain more professional, accurate and targeted answers to the specific field of obesity, the present invention adopts the mode of self-built data sets to collect professional journals and professional website materials, including obesity-related works and reports published by WHO and Oxford University, and journals published in recent years by the World Obesity Federation (Obesity Reviews, Pediatric Obesity, Clinical Obesity, Obesity Science & Practice) (see Table 1), and uses the Langchain framework to cut and convert the input data set (pdf format) into text format. At the same time, the LLM large language model is used to further cut the text to generate semantically complete sentences, and the processed data is stored in the FAISS database for database establishment.

[0082] Table 1 Sources of fine-tuning datasets

[0083]

[0084] Note: In the table, USDA: United States Department of Agriculture; HHS: United States Department of Health and Human Services.

[0085] (2) The present invention selects appropriate model parameters and downloads and configures the required large language model (llama-3-7B) from the huggingface website. The model has good performance and performs outstandingly in the AI ​​chat module.

[0086] (3) Figure 4 As shown, training the text classification model includes:

[0087] In order to provide more professional medical consultation services, the present invention uses a self-built training data set, which includes a large number of labeled samples constructed by LLM and manual reference to professional books (question samples related to obesity are marked as 1, and those not related are marked as 0). After training with this data set, the fasttext model can generalize its judgment ability to texts that it has never encountered, thereby filtering user questions that are not related to the specific field of obesity, making the answers more accurate and improving the user experience.

[0088] (4) In order to enhance the performance of the large language model on our target task, we fine-tuned the model using more than 3,000 training data that are highly relevant to the target task. We generated this data using a combination of manual and large model methods to construct a fine-tuning dataset. Specifically, we extracted hundreds of paragraphs from relevant papers and books, and used the ChatGPT3.5 API with a custom prompt to expand these declarative paragraphs into multiple question-answer pairs. Ultimately, the model, after fine-tuning and modifying parameters, can be more specialized in answering questions in the field of obesity.

[0089] like Figure 5 As shown in the figure, when it is put into use, the user's question is first passed into the text classification model. If the model determines that the user's question is not related to the field of obesity, the AI ​​Chatbot is programmed to output a specific answer; if the model determines that the user's question is related to the field of obesity, the processed self-built data set data in the FAISS database is read, mainly including the cut professional journals and website texts, and the similarity search function under the lanchain framework is used to calculate the similarity between the data in the data set and the question, and the data with high similarity to the question is output as a reference to guide the generation of the final answer.

[0090] The obesity health intelligent question-answering AIDoctor system consists of multiple modules to ensure the scalability, maintainability and security of the system. Specifically, the system architecture can be divided into the following main modules:

[0091] Filter module: This module pre-processes the user input, mainly responsible for removing characters that cannot be recognized by the large language model and filtering out questions that are irrelevant to the task set by this system. The character cleaning work is completed by Python regular expressions, and the question filtering system is completed by a DistilBERT model trained on the task-related question dataset set by this system as a classifier, which has the advantages of taking into account both accuracy and efficiency. The classifier can classify unlearned input samples as relevant or irrelevant. If relevant, it will pass and enter the next step (retrieval enhancement module). If irrelevant, it will refuse to answer and end the task.

[0092] Retrieval Enhanced Search: The retrieval enhanced search method, also known as RAG, quickly searches for the fields most relevant to the question input to the model by retrieving the relevant authoritative data sets built into the system. Passing this field along with the user's question as input to the large language model can improve the quality and credibility of the model's answer. Adding this module alleviates the problem of insufficient credibility of the output of the large language model to a certain extent.

[0093] The main part of the large language model: This part is implemented by a Llama model that has been fine-tuned on the downstream dataset. Llama is an open source large language model series, and this system currently uses the Llama3.1-8b-instruct model in the latest Llama3.1 version. This model has been supervised fine-tuned on tasks related to this system before being deployed in this product.

[0094] History record: Since the large language model itself has no memory, in order to allow it to associate the information in previous questions and answers when processing the user's subsequent additional questions to make the conversation smoother, it is necessary to store the user's previous chat records and input them into the large language model together when the user asks subsequent questions. In the next round of conversation, the stored input and output history records are used together with the user's new questions as the new round of input for the large language model, which can make the model's response more coherent and natural.

[0095] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-user group precise auxiliary diagnosis and treatment device for obesity, characterized in that: Including data interaction panel, functional modules based on explainable machine learning, and consultation and question-and-answer functional modules; The functional module based on interpretable machine learning includes an indicator variable set, a training data set, and an XGBoost model; The user selects a response variable and several explanatory variables from the indicator variable set through the data interaction panel, and selects a certain scale of training data from the training data set; Train the XGBoost model on the training data, output the statistical results and perform interpretable analysis on the statistical results; the data interaction panel displays the statistical results and their interpretable analysis; The consultation and question-answering function module interprets the statistical results and interpretability analysis output by the XGBoost model based on a large language model, and answers users' questions about obesity.

2. The multi-user group precise auxiliary diagnosis and treatment device for obesity according to claim 1, characterized in that: The method for constructing the indicator variable set includes: selecting indicator variables related to obesity based on the experience and judgment of medical personnel to construct the indicator variable set.

3. The multi-user group precise auxiliary diagnosis and treatment device for obesity according to claim 2 is characterized in that: The indicator variables related to obesity include physical examination indicator variables, serum insulin measurement indicator variables, glucose measurement indicator variables, serum C-peptide measurement indicator variables, glycosylated hemoglobin measurement indicator variables, five liver and kidney function / biochemistry indicator variables, seven thyroid function test indicator variables, vitamin D measurement, seven sex hormone measurement indicator variables, high-sensitivity CRP measurement indicator variables, urine routine indicator variables, hepatitis B qualitative + hepatitis C detection indicator variables, diabetes autoantibody detection indicator variables, liver B-ultrasound, carotid artery B-ultrasound, uterine B-ultrasound, ovarian B-ultrasound, testicular B-ultrasound, adrenal B-ultrasound indicator variables, questionnaire survey indicator variables, and bone density examination indicator variables.

4. The multi-user group precise auxiliary diagnosis and treatment device for obesity according to claim 1, characterized in that: The method for constructing a training data set includes: selecting a number of relevant data of patients with obesity from a built-in database of a medical institution to establish an initial data set; and preprocessing the data in the initial data set to obtain a training data set.

5. The multi-user group precise auxiliary diagnosis and treatment device for obesity according to claim 4, characterized in that: The data in the initial data set were preprocessed, including: renaming the indicator names in the initial data set to standard indicator names summarized by medical personnel, unifying the units of the indicators, and deleting abnormal values ​​in the indicator variables that did not conform to medical knowledge.

6. The multi-user group precise auxiliary diagnosis and treatment device for obesity according to claim 1, characterized in that: The training data was randomly divided into a training set and a test set in a ratio of 7:3; the XGBoost model was trained on the training set, and the prediction performance of the XGBoost model was tested on the test set; the model parameters of XGBoost were selected using 5-fold cross validation.

7. The multi-user group precise auxiliary diagnosis and treatment device for obesity according to claim 1, characterized in that: Outputs from interpretability analysis include: (a) The distribution of the values ​​of the response and explanatory variables, differentiated by gender; (b) global interpretability, including feature importance and cumulative local effect plots; (c) Local interpretability, including local SHAP plots.

8. The multi-user group precise auxiliary diagnosis and treatment device for obesity according to claim 1, characterized in that: The consultation question and answer function module includes: The filter module removes characters in user questions that cannot be recognized by the large language model and filters questions that are not related to obesity, and then inputs them into the retrieval enhancement module; The retrieval enhancement module searches for fields related to the user's question from relevant authoritative data sets through a retrieval enhancement search method, and inputs the user's question and related fields into the large language model together; The memory module stores the user's previous chat records and inputs the chat records into the large language model when the user asks subsequent questions; The large language model outputs relevant answers to user questions based on user questions, related fields, and previous chat records.

9. The multi-user group precise auxiliary diagnosis and treatment device for obesity according to claim 1, characterized in that: The training methods for large language models include: (i) Collect professional journals and professional website materials and convert them into text format to construct relevant authoritative datasets; use ChatGPT with custom prompt words to expand the declarative text paragraphs in the relevant authoritative datasets into several question-answer pairs to construct a fine-tuning dataset; (ii) The LLaMA model is trained using the fine-tuning dataset to obtain a large language model.

10. A method for auxiliary diagnosis and treatment based on the auxiliary diagnosis and treatment device according to any one of claims 1 to 9, characterized in that: include: When diagnosing obesity, medical personnel select one indicator variable from the indicator variable set as the response variable and select several indicator variables as explanatory variables; Select a certain scale of training data from the training data set; Train the XGBoost model on the training data, output the statistical results and perform interpretable analysis on the statistical results; the data interaction panel displays the statistical results and their interpretable analysis; The consultation and question-answering function module interprets the statistical results and interpretability analysis output by the XGBoost model based on a large language model, and answers users' questions about obesity.

Images