Multi-disciplinary health consultation auxiliary management method based on large language model and related equipment

Abstract

The invention provides a multidisciplinary health consultation auxiliary management method based on a large language model and related equipment, and is applied to the technical field of data processing. The method comprises the following steps: processing physiological feature information of a target user and identity information of the target user to generate an initial physiological tag of the target user; processing a preset consultation result prediction model based on the training data set with the target feature data to generate a target consultation result prediction model; processing the initial physiological tag of the target user based on the target consultation result prediction model to generate access preference information and validity preference information; processing the access preference information and the validity preference information to generate preference alignment information; processing the question and answer information of the target user and the consultation purpose information of the target user to generate an initial prediction feature vector of the target user; and processing the initial prediction feature vector of the target user based on the preference alignment information to generate target prediction report information of the target user.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a multidisciplinary health consultation auxiliary management method based on a large language model and related equipment. Background Art

[0002] In the field of healthcare, accurate diagnosis is crucial and is the basis of all health guidance. In addition, with the growth of the global population and the acceleration of the aging process, the demand for healthcare has increased dramatically, resulting in an increasingly severe shortage of medical resources. Doctors have to face a large number of users every day, and the work pressure is huge and it is easy to cause fatigue, which further increases the possibility of misdiagnosis. For interdisciplinary patients with complex or serious diseases, each user has different characteristics, such as age, gender, and previous medical history, and the judgment methods they need should also be different. In a real environment, it is difficult for doctors to have enough energy and time to fully analyze the user's information.

[0003] Currently, most large language model research focuses on medical case reports, lacking large language models specifically designed for actual clinical diagnosis. In addition, generative large language models may produce false or fabricated facts, which may cause harm if clinicians rely on the recommended diagnosis of these models.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0005] The purpose of this application is to provide a multidisciplinary health consultation auxiliary management method and related equipment based on a large language model, which at least overcomes the problems existing in the prior art to a certain extent. The server comprehensively collects target users' multivariate information and key model data to build an accurate health assessment system. It starts from physiological characteristics, question and answer records, identity information, and consultation purpose information, supplemented by training data sets and preset models. When generating the initial physiological label, the user's identity information is deeply analyzed to obtain classification and medication details, and the physiological characteristics are integrated to accurately calculate the associated value, and accurately outline the user's health base profile. The training data set is carefully preprocessed, and the features are extracted according to age and gender groups to form an original library. After the set is divided, the target data representing physiological abnormality risk factors is locked through classifier prediction and algorithm training, so as to build a barrier for the accurate identification of health hazards. The initial label is processed by the prediction model, and after genetic image mining, portrait outlining, matching value calculation and other links, access preference and effectiveness preference information are generated, which is oriented for personalized health navigation. Further analysis of preference information to achieve alignment, deep processing of question and answer consultation materials to shape the initial prediction feature vector, provide exclusive accurate prediction report information for target users, and empower health management decision-making in all directions.

[0006] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present invention.

[0007] According to one aspect of the present application, a multidisciplinary health consultation auxiliary management method based on a large language model is provided, including: obtaining physiological characteristic information of a target user, question and answer information of the target user, identity information of the target user, consultation purpose information of the target user, a training data set and a preset consultation result prediction model; processing the physiological characteristic information of the target user and the identity information of the target user to generate an initial physiological label of the target user; preprocessing the training data set to generate a training data set with target characteristic data, wherein the target characteristic data is used to characterize the risk factors of abnormal physiological characteristics of the target user; processing the preset consultation result prediction model based on the training data set with the target characteristic data to generate a target consultation result prediction model; processing the initial physiological label of the target user based on the target consultation result prediction model to generate access preference information and validity preference information; processing the access preference information and the validity preference information to generate preference alignment information; processing the question and answer information of the target user and the consultation purpose information of the target user to generate an initial prediction feature vector of the target user; processing the initial prediction feature vector of the target user based on the preference alignment information to generate target prediction report information of the target user.

[0008] Another aspect of the present application is a multidisciplinary health consultation auxiliary management device based on a large language model, comprising: an acquisition module, used to acquire physiological characteristic information of a target user, question and answer information of the target user, identity information of the target user, consultation purpose information of the target user, a training data set and a preset consultation result prediction model; a processing module, used to process the physiological characteristic information of the target user and the identity information of the target user to generate an initial physiological label of the target user; pre-process the training data set to generate a training data set with target characteristic data, wherein the target characteristic data is used to characterize the risk factors of abnormal physiological characteristics of the target user; based on the training data set with the target characteristic data, the preset consultation result prediction model is processed to generate a target consultation result prediction model; based on the target consultation result prediction model, the initial physiological label of the target user is processed to generate access preference information and validity preference information; the access preference information and the validity preference information are processed to generate preference alignment information; the question and answer information of the target user and the consultation purpose information of the target user are processed to generate an initial prediction feature vector of the target user; the initial prediction feature vector of the target user is processed based on the preference alignment information to generate target prediction report information of the target user.

[0009] According to another aspect of the present application, an electronic device includes: a first processor; and a memory for storing executable instructions of the first processor; wherein the first processor is configured to execute the executable instructions to implement the above-mentioned multidisciplinary health consultation auxiliary management method based on a large language model.

[0010] According to another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a second processor, the multidisciplinary health consultation auxiliary management method based on a large language model is implemented.

[0011] The application provides a multidisciplinary health consultation auxiliary management method and related equipment based on a large language model. The server is used to build an accurate health assessment system, and comprehensively collects target users' multivariate information and key model data. It starts from physiological characteristics, question and answer records, identity information, and consultation purpose information, supplemented by training data sets and preset models. When generating the initial physiological label, the user's identity information is deeply analyzed to obtain classification and medication details, and the physiological characteristics are integrated to accurately calculate the associated value, and accurately outline the user's health base profile. The training data set is carefully preprocessed, and the features are extracted according to age and gender groups to form an original library. After the set is divided, the classifier is predicted and the algorithm is trained to lock the target data representing the risk factors of physiological abnormalities, and a barrier is built for the accurate identification of health hazards. The initial label is processed by the prediction model, and after genetic image mining, portrait outlining, matching value calculation and other links, access preferences and effectiveness preference information are generated, which is oriented for personalized health navigation. Further analysis of preference information reaches the state of alignment, and deep processing of question and answer consultation materials shapes the initial prediction feature vector, providing exclusive and accurate prediction report information for target users, and enabling health management decisions in all directions.

[0012] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A flowchart showing a multidisciplinary health consultation auxiliary management method based on a large language model provided by an embodiment of the present application is shown;

[0014] Figure 2 A schematic diagram of the structure of a multidisciplinary health consultation auxiliary management device based on a large language model provided in one embodiment of the present application is shown. DETAILED DESCRIPTION

[0015] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0016] Combine the following Figure 1 To describe the multidisciplinary health consultation auxiliary management method based on a large language model according to an exemplary embodiment of the present application. It should be noted that the following application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the embodiments of the present application are not limited in this regard. On the contrary, the embodiments of the present application can be applied to any applicable scenario.

[0017] In one embodiment, the present application also proposes a multidisciplinary health consultation auxiliary management method and related equipment based on a large language model. Figure 1 The flowchart of a multidisciplinary health consultation auxiliary management method based on a large language model according to an embodiment of the present application is schematically shown. Figure 1 As shown, the method is applied to a server, comprising:

[0018] S101, obtaining physiological characteristic information of a target user, question and answer information of the target user, identity information of the target user, consultation purpose information of the target user, a training data set and a preset consultation result prediction model.

[0019] In one embodiment, it is assumed that there is a 40-year-old male, 175cm tall, 75kg in weight, married and fertile. In recent months, headache symptoms have frequently occurred, and the pain is mostly concentrated in the bilateral temples and forehead, showing throbbing pain, which is sometimes light and sometimes heavy. In severe cases, it is as if the head is tightly bound and daily activities are restricted. There is no obvious regularity in the attack, or it aggravates after fatigue, or it breaks out during mood swings, and it lasts for several hours each time. Accompanied by mild blurred vision, occasional double images when looking at nearby objects, dryness and soreness in the eyes, especially after long-term use of the eyes. Memory has also declined, and recent trivial matters are often forgotten, work efficiency is reduced, and the rhythm of life is disturbed. On the medical consultation platform, he urgently asked: "Headache accompanied by blurred vision and memory loss, what is the sign of disease? What kind of examination is needed to confirm the diagnosis?" "What are the emergency measures to relieve headaches? How to prevent them in daily life?" "Are vision problems related to headaches? What eye protection measures should be taken to improve them?" "Are there any special methods or auxiliary means to improve memory?" His identity information shows that in the middle-aged stage, the body functions change, and under the dual pressure of family and career, health risks are prone to appear. Marital and child-rearing conditions add complexity to life, affecting work and rest and physical and mental state.

[0020] The purpose of consultation is to accurately identify the cause of the disease, quickly relieve headaches, restore visual clarity, improve memory efficiency, prevent potential risks such as brain lesions, eye diseases and neurasthenia, reshape efficient life and work order, and consolidate the foundation of health. The training data set comes from multiple channels. The physiological characteristic data includes brain images, eye examination reports, neurological function test values ​​and life habit records of a large number of people with similar symptoms in the same age group, and deeply analyzes the abnormal correlation of physiological indicators and the co-occurrence of symptoms. The question and answer data are collected from professional medical platforms, patient mutual assistance communities, and ophthalmology and neurology forum consultation exchanges, sorting out symptom descriptions, diagnostic ideas, treatment feedback and experience sharing, and exploring the core of patients' health demands and the evolution path of cognitive blind spots. Identity background and consultation purpose data are integrated from detailed medical records, lifestyle questionnaires, psychological assessment scales and expert health assessments. The interactive effects of life situations, psychological stress, and health pursuits are systematically sorted out. After strict screening, standardized processing, fine labeling and strict review, a solid data set framework is constructed.

[0021] In another embodiment, suppose there is a 40-year-old male, 175cm tall, 75kg in weight, married and with children as the target user. Its physiological characteristic information contains rich dimensions. From the basic physical signs, the resting heart rate is stable at about 70 beats / minute for a long time, the blood pressure is maintained in the ideal range of 120 / 80mmHg, and the vital capacity is 4000ml after professional testing, showing a good cardiopulmonary function foundation. In terms of blood test indicators, the cholesterol content is maintained at 5mmol / L, the triglyceride is at 1.2mmol / L, and various liver and kidney function indicators such as alanine aminotransferase and creatinine are within the normal range, reflecting the relative stability of the body's metabolic function. The body composition analysis data shows that the body fat rate is 20%, the muscle mass is moderate, the bone density is above the average level of peers, the bone mineral content measured by dual-energy X-ray absorptiometry is a healthy standard value, and the body water content accounts for about 60% of the body weight. These data together outline the internal composition of the body.

[0022] In the area of ​​question-and-answer information, users actively seek knowledge on health consultation platforms and conduct many inquiries around the intensity of daily fitness exercises. For example, "If you want to improve your cardiopulmonary endurance and avoid sports injuries, how many times a week should you do high-intensity interval training? How should you accurately control the duration and intensity of each session?" "To enhance muscle strength and endurance, how should you scientifically plan the movement selection and weight increase pattern for different muscle groups in a strength training plan?" Regarding nutritional diet combinations, they ask, "To reduce body fat percentage while maintaining muscle mass, what is the most reasonable daily intake ratio of protein, carbohydrates and fat? How do you choose from various food sources? "When the seasons change, how can we adjust our diet to supplement specific vitamins and minerals and enhance our body's immunity?" Regarding the physical discomfort caused by long-term desk work, the questions include "After sitting for 8 hours, how can we arrange the precise stretching movements and relaxation time for the shoulders, neck, and lumbar spine to effectively relieve muscle tension?" "What are the specific mechanisms of simple eye exercises and viewing time settings during work breaks to relieve visual fatigue and prevent myopia from getting worse?" His identity information is clear and unambiguous. His age of 40 symbolizes that his body functions have entered a specific transformation stage. His gender of male is associated with specific physiological characteristics and potential health risks. His height of 175cm and weight of 75kg provide key parameters for analyzing body mass index (BMI) and body load tolerance. His marital status and childbearing status affect health decisions and lifestyle shaping from the dimensions of lifestyle, family responsibilities, and psychological pressure.

[0023] The information on the purpose of consultation focuses on improving physical fitness in an all-round way, hoping to break through physical bottlenecks and strengthen physiological function reserves with the help of scientific planning. Preventing potential health risks is the core demand. In view of the gradual changes in physiological functions in the middle-aged stage, it aims to lay out health strategies in advance to prevent hidden dangers such as cardiovascular disease, bone degenerative lesions, and metabolic syndrome. Improving the quality of life is the key driving force, hoping to improve work efficiency and enrich the vitality of spare time by optimizing health status. Maintaining a good physical state is a long-term vision, pursuing a sustainable health model with balanced body and mind and stable functions. The training data set comes from a large number of diverse channels, and widely collects all-round information from many healthy people and people experiencing minor health fluctuations. From the perspective of physiological characteristics, it covers physical examination reports, physiological monitoring equipment data, and professional medical test values ​​of people of different ages, genders, and regions, accurately capturing the dynamic changes and differential characteristics of vital signs, metabolic indicators, and body composition. The question and answer records are mined from major health forums, professional consulting platforms, and community surveys, precipitating a large number of health consultation questions, professional answers, and user feedback in real scenarios, and deeply analyzing users' health concerns, knowledge blind spots, and demand evolution trends. Identity background and consultation purpose information are integrated from user registration information, lifestyle questionnaires, health tracking files and expert interview records, comprehensively sorting out the complex interactive relationship between living habits, social roles, environmental factors and health pursuits. After rigorous screening, normalization, multiple rounds of labeling and quality review, a large-scale, representative and complete data set foundation is constructed.

[0024] The architecture of the preset consultation result prediction model is as follows: Transformer module processes genetic data: used to process genetic sequencing information, using the advantages of Transformer in processing long sequence data to capture features in genetic data, such as identifying disease-related gene mutations. Transformer module processes medical imaging data: processes features extracted from medical imaging information (such as cardiovascular structure and function indicators, organ morphology and density, etc.), and accurately analyzes key information in the image with its powerful feature extraction capabilities. Multi-layer perceptron (MLP) module processes structured data: responsible for processing structured data such as lifestyle (daily steps, sleep duration and quality, dietary structure, etc.), family medical history, etc., and constructs a mapping relationship from these data to health risk assessment and consultation recommendations. Transformer module integrates information: used to integrate the output of genetic data, medical imaging data, and structured data processing modules, considering the timing and dependencies between different modules, and comprehensively and comprehensively assessing user health risks.

[0025] The model training process is as follows:

[0026] Pre-training stage: Based on the advanced general large language model base, the auxiliary consultation result prediction model is pre-trained on a large-scale medical corpus. i} represents the collected medical corpus, a sequence x i By i A token indicates that The pre-training goal is to minimize the negative log-likelihood, so that the model has a vast amount of general medical knowledge, can understand and process various medical-related information, and lay the foundation for subsequent fine-tuning for health consultation tasks.

[0027] Fine-tuning stage: Based on the self-bootstrapping strategy, the model is fine-tuned using a diagnostic thinking logic annotated dataset containing doctors’ diagnostic reasoning thinking logic. In this stage, the language model is trained to respond to a given input prompt I = ω 1:m , generating a reply R = v 1:n , the goal is to optimize p θ (R|I)=p θ (v 1:n |ω 1:m ), where n and m represent the length of the response and input prompt respectively. The final loss function is In this way, the model can simulate the diagnostic thinking logic of doctors and improve the accuracy of diagnosis.

[0028] Preference alignment optimization: A unified preference alignment framework is proposed, which includes diagnostic hierarchy preferences and usefulness preferences to meet diagnostic standards and the needs of medical professionals in clinical scenarios. The diagnostic hierarchy preference is guided by the disease classification hierarchy defined by the International Classification of Diseases (ICD) codes, so that model generation is consistent with disease classification standards; the usefulness preference is refined by the usefulness scoring model trained on expert annotations, making the model generation more informative, useful, and credible for diagnosis, while reducing the risk of harm or misleading information. The preference alignment process consists of two steps: preference data construction and preference optimization.

[0029] S102: Process the physiological characteristic information of the target user and the identity information of the target user to generate an initial physiological tag of the target user.

[0030] In one embodiment, the target user's identity information is processed to generate a target user's classification label and target user's medication type information. Assume that the target user is a 70-year-old female, 160 cm tall, 70 kg in weight, with a history of coronary heart disease and osteoporosis. Based on her age, height, weight, and diseases, she can be classified as a "high-risk group for elderly chronic diseases." This classification takes into account the user's advanced age and the two chronic diseases of coronary heart disease and osteoporosis, which means that the user has a higher risk of recurrence of cardiovascular events and fractures. For coronary heart disease, she may be taking aspirin for antiplatelet aggregation and prevention of thrombosis; she may also take statins such as atorvastatin to regulate blood lipids and stabilize atherosclerotic plaques. For osteoporosis, calcium (such as calcium carbonate) and vitamin D supplements may be taken to increase bone density, and bisphosphonates (such as sodium alendronate) may be used to inhibit osteoclast activity and reduce bone resorption.

[0031] The target user's classification label, the target user's medication type information and the target user's physiological characteristic information are processed to generate the target user's characteristic data association value, and the target user's characteristic data association value is processed to generate the target user's initial physiological label. The target user's identity information is processed to generate the target user's classification label and related health maintenance information. Suppose there is a 70-year-old female target user with a height of 160cm and a weight of 70kg. Taking into account her age, height, weight and other factors, she can be classified as a "focused elderly health population". This classification fully considers the user's advanced age, which means that the user has a certain risk of changes in physical function. In order to maintain her physical function, she may take some targeted health care measures, such as taking some nutritional supplements that help maintain cardiovascular function; she may also use some nutritional supplements that help bone health to keep the bones in good condition.

[0032] The method includes a calculation formula for obtaining a characteristic data association value of a target user, and the calculation formula is:

[0033]

[0034] Among them, a i is the weight coefficient corresponding to the th physiological characteristic information, n is the number of physiological characteristic information; X i is the target user’s i-th physiological characteristic information value, X 0i is the reference value of the th physiological characteristic information; σ i is the standard deviation of the ith physiological characteristic information value in the population, which is used for standardization processing to make different physiological characteristic information comparable; b j is the weight coefficient corresponding to the j-th medication type information, m is the number of medication type information; Y j is the target user’s j-th medication type information value, Y 0j is the reference value of the j-th medication type information; c k is the weight coefficient corresponding to the kth hierarchical label information, p is the number of hierarchical label information; Z k is the kth graded label information value of the target user, Z 0k is the reference value of the kth classification label information; Z maxk and Z mink are the maximum and minimum values ​​of the kth classification label information in the population, which are used for standardization. The parameters in the formula can be understood and illustrated as follows:

[0035] For the first part

[0036] Assume X i is the user's blood lipid index, for example, low-density lipoprotein cholesterol (LDL-C) is 3.5mmol / L, which is the ideal LDL-C value (such as 2.6mmol / L), σ i is the standard deviation of LDL-C values ​​in healthy people (assumed to be 0.5mmol / L), a i is the weight coefficient of the impact of blood lipids on overall health (assuming it is 0.3). Then this part is calculated as

[0037] For the second part

[0038] Assumption Y j is the dose of aspirin taken by the user (e.g. 100 mg per day), Y 0j is the standard recommended dose (also 100 mg per day), b jis the weight coefficient of the effect of the dosage on health (assuming it is 0.2). Then this part is calculated as 0.2×(100-100)=0.

[0039] For the third part

[0040] Assume Z k is the user's specific level in the "elderly chronic disease high risk population" classification (for example, level 4, assuming the classification is 1-5), Z 0k is a normal level (such as level 1), Z maxk is the highest level (level 5), Z mink is the lowest level (level 1), c k is the weight coefficient of the graded health impact (assumed to be 0.5). This part is calculated as

[0041] The total feature data association value Z is the sum of the above three parts: Z=0.972+0+0.375=1.347.

[0042] Based on the calculated feature data association value Z=1.347, an initial physiological label is further generated. For example, combined with the above calculation results, an initial physiological label such as "70-year-old female, suffering from coronary heart disease and osteoporosis, blood lipids slightly higher than the ideal value, medication basically meets the standard, but overall at a high risk of chronic diseases, need to pay close attention to cardiovascular and bone health" may be generated. This label can provide a more specific basis for subsequent health management and treatment. For example, the consultant can decide whether to adjust the dosage of medication, increase the frequency of examinations, etc. based on this label.

[0043] S103, preprocessing the training data set to generate a training data set with target feature data.

[0044] In one implementation, the training data set is grouped to generate a grouped training data set, wherein the grouped training data set includes physiological characteristic information of users of different age groups and genders. Assume that we have a training data set containing physiological characteristic information of 1,000 users. The age range of these users ranges from 20 to 80 years old, and the gender includes males and females. We group them according to age group and gender. For example, males aged 20-30 are grouped into one group, females aged 20-30 are grouped into one group, males aged 31-40 are grouped into one group, and so on. In this way, we get multiple groups of different age groups and genders. Each group contains physiological characteristic information of several users, such as height, weight, blood pressure, heart rate, blood sugar, blood lipids and other data. For example, in the 31-40 year old male group, there are data of 50 users, including their average height of 175cm, average weight of 80kg, average blood pressure of 130 / 85mmHg, etc.

[0045] Perform feature extraction on the grouped training data set to generate an original feature library. For each group, we perform feature extraction. Taking the 31-40-year-old male group as an example, we may extract the following features: The average blood pressure of the group is 130 / 85mmHg as a feature, which reflects the blood pressure level of men in this age group. Calculate the average body mass index (BMI) of the group, assuming it is 25.5, which is also a feature, reflecting the body fatness of the group. Count the proportion of users in the group whose blood sugar is higher than the normal range (assuming fasting blood sugar>6.1mmol / L), for example, 10%, which is also a meaningful feature. By performing such feature extraction operations on all groups, we have constructed an original feature library. This library contains various physiological feature statistics of different age groups and genders.

[0046] The original feature library is processed to generate a training set and a validation set. For example, we randomly select 70% of the data as the training set and 30% of the data as the validation set. Assuming that there are 100 feature vectors in the original feature library, 70 feature vectors enter the training set and 30 feature vectors enter the validation set. The data in the training set will be used to train the classifier, while the validation set is used to evaluate the performance of the classifier.

[0047] Based on the classifier, the validation set is predicted and processed to generate prediction results. We select a classifier, such as a support vector machine (SVM) classifier. The validation set is input into the SVM classifier for prediction. Assume that there is a feature vector in the validation set that represents the physiological characteristic data of the 45-50-year-old female group, including average blood pressure, BMI, and abnormal blood sugar ratio. The SVM classifier predicts whether the group of users has a risk of abnormal physiological characteristics based on its trained model. For example, the prediction result may be that the group has a higher risk of hypertension (with a probability of 70%). The training set is trained based on the preset algorithm to generate a validation set class prediction result, and the training set is trained using a preset algorithm (such as a stochastic gradient descent algorithm). Assume that during the training process, the model learns the relationship between different physiological characteristics and risk factors. Then, we use the trained model to predict the validation set and obtain the validation set class prediction result. For example, for the 45-50-year-old female group in the validation set, the trained model predicts that their risk of hypertension is 65%.

[0048] The prediction results and the validation set prediction results are processed to generate target feature information, which is used to characterize the risk factors of abnormal physiological characteristics of the target user. Compare the prediction results (70% risk of hypertension) and the validation set prediction results (65% risk of hypertension). By analyzing and processing these results, such as calculating the difference between the two and considering other relevant factors, we can generate target feature information. For example, we found that in the 45-50 year old female group, when the BMI is higher (greater than 25) and the proportion of abnormal blood sugar is greater than 10%, the risk of hypertension increases significantly. This finding is the target feature information, which characterizes the risk factors of abnormal physiological characteristics (hypertension) of the target user (45-50 year old female).

[0049] S104: Process the preset consultation result prediction model based on the training data set with the target feature data to generate a target consultation result prediction model.

[0050] In one embodiment, in this comprehensive health management system, our goal is to comprehensively assess the health risks of users and provide personalized health advice. The training data set contains multi-dimensional target feature data. For example, in addition to basic physiological data (such as age, height, and weight), it also includes features extracted from the user's gene sequencing information (such as certain disease-related gene mutations), features extracted from medical imaging information (such as cardiovascular structure and function indicators, organ morphology and density, etc.), user lifestyle characteristics (such as daily steps, sleep duration and quality, dietary structure, etc.), family history characteristics (such as whether there is a genetic tendency for a specific disease in the family), etc. Suppose we have data for 5,000 users. Take one of the users as an example. The user is 40 years old, 170 cm tall, and weighs 75 kg. Gene sequencing found that there are gene mutations related to cardiovascular disease. Medical imaging shows that the coronary artery has slight calcification, the average number of steps per day is 5,000, and the sleep duration per night is 7 hours but the sleep quality is average (monitored by wearable devices). There are elders in his family who suffer from cardiovascular disease.

[0051] The preset consultation result prediction model is a complex model built on the Transformer architecture, which aims to integrate multi-source data and accurately predict the user's health risks and provide personalized consultation advice. Model architecture composition: Encode and decode gene sequencing information using the Transformer's powerful ability to process sequence data. Through the multi-head self-attention mechanism, it can simultaneously focus on different positions and features in the gene sequence, effectively capturing the complex relationship between gene variants and their potential impact on health. For example, when analyzing cardiovascular disease-related genes, it can comprehensively consider the mutations of multiple gene loci and judge the comprehensive impact of their interactions on disease risk. For medical imaging data, the Transformer module can convert the spatial information in the image into a sequence form for processing. It can identify different regional features in the image and learn the association patterns between these features. For example, when analyzing brain images, it can capture the characteristic differences between the white matter lesion area and the surrounding normal tissue, as well as the relationship between these differences and potential health problems such as cognitive dysfunction.

[0052] The multi-layer perceptron (MLP) module processes structured data: it is responsible for processing structured data such as lifestyle and family medical history. MLP performs nonlinear transformations on input data through multiple hidden layers to learn complex patterns and relationships in the data. For example, it can determine how these factors jointly affect the user's risk of developing a certain disease based on information such as the user's diet structure, exercise frequency, and family medical history. Transformer module integrates information: this module is used to integrate outputs from genetic data, medical imaging data, and structured data processing modules. It can capture the timing and dependencies between different types of data, and comprehensively analyze the information provided by each module to generate comprehensive and accurate user health risk assessments and personalized consulting recommendations.

[0053] Model initialization: When the model is initialized, the parameters of each module (such as the attention mechanism weights in the Transformer module, the neuron weights in the MLP module, etc.) are randomly assigned initial values. At this point, the model does not have the ability to accurately predict the user's health risks and provide effective consultation, and these parameters need to be optimized through subsequent training processes. Input the training data set with the target feature data into the preset consultation result prediction model to start the model training process.

[0054] Training strategy based on gradient descent: For different types of data modules, a gradient descent-based optimization algorithm is used for training. For the Transformer modules of genetic data and medical imaging data, the parameters are updated by minimizing the loss function between the predicted results and the actual health risk labels (determined based on long-term follow-up studies and clinical diagnosis, such as whether a specific disease will occur in the next 5 years). The loss function can use common classification loss functions such as cross entropy loss to measure the difference between the probability distribution predicted by the model and the true label. During the training process, the gradient of the loss function to the model parameters is calculated, and the parameter values ​​are adjusted according to the gradient direction so that the model prediction results gradually approach the actual situation. For MLP modules of structured data such as lifestyle, the gradient descent algorithm is also used for training. By comparing the health risk assessment results predicted by the model with the actual results, the loss is calculated and the gradient is back-propagated, and the neuron weights in the MLP module are updated to improve the model's processing ability and prediction accuracy for structured data.

[0055] Adjustment and optimization during training: During the training process, the model's hyperparameters (such as learning rate, regularization parameters, etc.) and network structure (such as increasing or decreasing the number of layers of the Transformer module, adjusting the number of heads, etc.) are continuously adjusted to improve the performance of the model. Learning rate adjustment: Adopt adaptive learning rate strategies, such as Adagrad, Adadelta or Adam algorithms, to dynamically adjust the learning rate according to the parameter update during the training process. In the early stage of training, a larger learning rate is used to enable the model to quickly converge to a better parameter range; as the training progresses, the learning rate is gradually reduced to avoid the model oscillating around the minimum value and ensure that the model can stably converge to the optimal solution. Regularization: In order to prevent the model from overfitting, regularization techniques such as L1 and L2 regularization are introduced. L1 regularization will make the model parameters sparse toward 0 and reduce the complexity of the model; L2 regularization will penalize the sum of squares of the model parameters to limit the parameters from being too large, making the model smoother and improving the generalization ability. Network structure adjustment: Through methods such as cross-validation, the impact of different network structures on model performance is evaluated. For example, try increasing the number of layers in the Transformer module and observe the changes in the model's accuracy, recall, and other indicators on the training set and validation set. If the model performance is improved after increasing the number of layers, keep the new structure; otherwise, restore the original structure or make other adjustments.

[0056] During the training process, the validation data set is used to evaluate the model regularly, focusing on indicators such as accuracy, recall, and F1-score. Accuracy refers to the proportion of the number of samples correctly predicted by the model to the total number of samples, reflecting the accuracy of the model; recall is the proportion of the number of correctly predicted positive samples to the actual number of positive samples, measuring the coverage of the model for positive samples; F1-score is the harmonic mean of the accuracy and recall rates, which can more comprehensively evaluate the performance of the model. After a large number of iterative training (for example, 5,000 iterations), when the comprehensive evaluation indicators of the model on the training data set and the validation data set reach a satisfactory level (such as F1-score reaches 0.8), or the loss function of the model converges to a smaller value, we believe that the model training is complete and the target consultation result prediction model is obtained. At this point, the model can accurately predict the risk of users in different health fields (such as cardiovascular disease, diabetes, cancer, etc.) and provide corresponding personalized health consultation suggestions. For example, for a user with a high risk of cardiovascular disease, the model can accurately recommend that he or she increase aerobic exercise (such as at least 150 minutes of moderate-intensity aerobic exercise per week, such as brisk walking, jogging, etc.), improve diet (reduce saturated fat and cholesterol intake, increase dietary fiber and unsaturated fat intake, etc.), and regularly undergo cardiovascular-related examinations (such as blood lipid tests every six months, heart ultrasound examinations once a year, etc.).

[0057] S105, processing the initial physiological tag of the target user based on the target consultation result prediction model to generate access preference information and validity preference information.

[0058] In one embodiment, the initial physiological label of the target user is processed to generate the target user's gene sequencing information and the target user's medical imaging information. For a 55-year-old male user, gene sequencing is a detailed analysis of his DNA. This may reveal some gene variations related to body functions. For example, some gene variations may be related to body metabolic function and vascular function. This gene information can provide a basis for assessing the user's long-term health risks. Gene sequencing technology usually involves extracting a user's DNA sample and then sequencing it through a high-throughput sequencer. The sequencing results will be analyzed by bioinformatics and compared with known gene databases to determine whether there are gene variations with potential health effects. In addition to certain vascular conditions found in heart-related examinations (not specifically indicating coronary atherosclerotic plaques), other imaging examination results may also be included. For example, cardiac ultrasound can assess the structure and function of the heart, including ventricular size, myocardial thickness, heart valve function, etc. Vascular ultrasound in other parts can detect whether there are certain changes in blood vessels. CT scans can provide more detailed body structure information, and these imaging examination results together constitute the user's medical imaging information, providing an intuitive basis for assessing physical condition.

[0059] The target user's gene sequencing information and the target user's medical imaging information are processed to generate the target user's portrait information. The target user's portrait information is a comprehensive distillation of the gene sequencing information and medical imaging information. It is not limited to the current physical condition related to the situation. It may also include predictions of future changes in body function based on these findings. For example, based on the type and location of gene mutations, combined with vascular-related findings, the risk level of certain changes in body function in the next 5-10 years is predicted. In addition, the portrait information can also take into account factors such as the user's family medical history and lifestyle. For example, if the user has a family history of certain health problems and there are related gene mutations, then the risk of related health problems will be further increased. The portrait information will integrate these factors together to form a comprehensive and personalized description of the user's health.

[0060] The target user profile information is processed based on the target consultation result prediction model to generate the target user's access match value. The model is built based on big data and machine learning algorithms. The training data may come from a large number of clinical data of people with different physical conditions and healthy people, including genetic data, imaging data, lifestyle, etc. By learning from these data, the model can identify the relationship between different feature combinations and changes in physical function. When processing the profile information of a 55-year-old male user, the model will take all relevant factors into consideration. For example, the model may calculate the user's degree of need for different health examinations and interventions based on the user's genetic variation, vascular-related findings, age, gender, and other physiological indicators. This degree of need is expressed in the form of an access match value. For this user, the higher access match value given by the model means that he needs to pay closer attention to his health. In addition to regular specific physical function examinations, other related examinations may be required, such as monitoring of certain metabolic indicators and review of health risk assessments.

[0061] The access matching values ​​of the target users are processed to generate access preference information and effectiveness preference information. Access preferences also include preferences for different examination methods, medical treatment time, consultant gender, etc. For example, some users may prefer non-invasive examination methods, or hope to see a doctor in the morning on weekdays so as not to affect work. For this user, a certain comprehensive intervention (not specifically indicating that it is a combination of drug intervention and lifestyle adjustment) is considered to be a more effective solution. This is based on a comprehensive judgment of clinical research and individual circumstances. For example, for patients with certain signs of changes in physical function, certain interventions can help improve physical function and reduce the risk of health problems. Effectiveness preference information may also consider factors such as the adaptability of different interventions and user compliance. For example, if a user has an adverse reaction to a certain intervention, other alternative measures will be considered when formulating the plan to ensure the effectiveness of the plan and user compliance.

[0062] In another implementation, the method further includes a calculation formula for obtaining an access matching value of a target user, the calculation formula being:

[0063] Among them, x represents the initial physiological label of the target user. This is the starting point of the entire calculation process, which contains the basic physiological characteristics of the target user, such as height, weight, blood pressure, heart rate and other basic indicators, as well as family medical history and other related information; y wRepresents the high-risk related features extracted from the target user's gene sequencing information and medical imaging information. For example, specific gene mutations associated with serious diseases found in gene sequencing, or features of suspected lesions (such as tumors, atherosclerotic plaques, etc.) found in medical imaging (such as CT, MRI, etc.); l Represents low-risk related features extracted from the target user's gene sequencing information and medical imaging information. For example, no obvious disease-related mutations were found in gene sequencing, and medical imaging showed normal physiological structure characteristics (such as normal organ morphology, no stenosis of blood vessels, etc.). π θ (y w |x) represents the prediction model based on the target consultation result (the model parameter is θ). Given the initial physiological label x of the target user, it is predicted that the target user has high-risk characteristics y w The probability distribution of . For example, for a user with a specific initial physiological label (such as age, gender, family history, etc.), the model predicts the probability that he or she has a certain high-risk genetic feature (such as breast cancer-related gene mutation); β represents the weight coefficient, which is used to adjust the relative importance of high-risk features and low-risk features in calculating the access matching value. It is determined according to factors such as disease type and clinical experience. If the high-risk feature is more important in judging the disease in the diagnosis of a certain disease, the weight of the part related to the high-risk feature in the formula can be increased; σ represents the base of the logarithm, which is used to scale the calculation results and map the calculation results to a suitable numerical range to facilitate the interpretation and application of the access matching value. For example, by selecting an appropriate , the final value can be mapped to 0 to 1 or other meaningful intervals.

[0064] Suppose we have a health management system and the target user is a 45-year-old male. The following is an example setting of the parameters in the formula: The parameters in the formula are explained as follows:

[0065] x: Relevant information of the target user, where x includes the target user's basic health information, such as age (45 years old), gender (male), lifestyle (e.g., exercising 3 times a week for 30 minutes each time, has a history of smoking but has quit smoking for 2 years), family medical history (father has cardiovascular disease), etc. w and l : The occurrence of different events, y w : Events associated with increased health risks. For example, w It can indicate the occurrence of cardiovascular related health problems (such as high blood pressure, abnormal blood lipids, etc.) in the next year. l , events related to maintaining a healthy state: for example, y l It can represent an event in which cardiovascular-related indicators remain normal (blood pressure, blood lipids, blood sugar, etc. are within the normal range) within the next year.

[0066] π θ (y w |x): Predict y based on the target consultation result prediction model w The probability of cardiovascular health problems in the next year is 0.3, assuming that based on the system model, for this 45-year-old male, considering his age, family history, and previous smoking history, the probability of cardiovascular health problems in the next year is 0.3, that is, π θ (y w |x)=0.3. θ (y l |x): The probability predicted by the target consultation result prediction model. Accordingly, the probability of maintaining normal cardiovascular indicators in the next year is π θ (y l |x)=0.7. ref (y w |x): y in the reference population w The probability of occurrence of cardiovascular-related health problems in the general population of the same age and gender is 0.2, that is, π ref (y w |x)=0.2.π ref (y l |x): y in the reference population l The probability of occurrence of cardiovascular disease is 0.8, and the average probability of maintaining normal cardiovascular indicators in the general population of the same age and gender is 0.8, that is, π ref (y l |x)=0.8. β: Weight coefficient, assuming β=0.5, it is used to adjust the weights of different parts in the formula, which here indicates the importance attached to the probability difference. σ: The base of the logarithm, assuming σ=10, it is used to scale the calculation results so that the results are within the appropriate numerical range.

[0067] Substituting the above parameters into the formula middle:

[0068] First calculate the inner logarithmic part: Similarly,

[0069] Then calculate the difference between the two: 0.088-(-0.029)=0.117.

[0070] Finally, compute the outer logarithm: The calculated L = -0.93 is the target user's access matching value, which reflects the difference in health risk between the target user and the reference population. A negative L value indicates that the user has a relatively high risk of developing cardiovascular-related health problems in the next year (because π θ (y w |x)>π ref (y w |x) and π θ (y l |x)<π ref (y l |x)). The health management system can develop more targeted health management plans for users based on this access matching value, such as recommending more frequent cardiovascular health checks and providing lifestyle intervention guidance.

[0071] S106: Process the access preference information and the validity preference information to generate preference alignment information.

[0072] In one implementation, the access preference information and the validity preference information are processed to generate risk gene position information, genetic risk information, image texture information, and image shape features. Assume that there is a health management platform, and the user is a 45-year-old male. The platform collects his access preference information and validity preference information. Access preference information: Users tend to choose health check-ups on weekends, prefer medical institutions that are closer to home, and have a higher acceptance of non-invasive examination methods. For example, when choosing a physical function assessment, users are more willing to accept non-invasive examinations such as physical fitness tests, and are resistant to invasive examination methods. Validity preference information: Users believe that in terms of maintaining a healthy state of body, a program that combines nutritional supplements with regular exercise is more effective. For example, to maintain a good physical condition, users prefer a comprehensive program of taking vitamin supplements and doing aerobic exercise more than 3 times a week.

[0073] By analyzing the user's genetic test data, combined with their access preferences and effectiveness preferences. Suppose that the user is found to have variations in certain gene loci related to physical function. For example, a specific variation in a gene related to metabolism may be related to the body's energy metabolism efficiency. The variation of this gene locus is the risk gene locus information. Based on the above gene variation, the user's genetic risk is further evaluated. Taking physical function decline as an example, if the variation of metabolism-related genes may affect energy metabolism, combined with similar physical function changes in the family (assuming that there are related situations in the family), the user's genetic risk of significant changes in physical function in the next 10 years is comprehensively assessed to be 30%, which is the genetic risk information. Suppose the user has undergone an imaging examination (such as MRI) of certain parts of the body (such as muscles and bones). By analyzing the image data, image texture information is extracted, such as the uniformity of muscle texture and the uniformity of bone density. At the same time, image shape features are also extracted, such as the size ratio of muscles and the morphology of bones. If it is found that there is local unevenness in muscle texture and slight changes in bone morphology, these may be signs of potential changes in physical function.

[0074] The risk gene position information, genetic risk information, image texture information and image shape features are processed to generate an access decision tree. The above risk gene position information, genetic risk information, image texture information and image shape features are integrated. For example, whether further physical function assessment is needed is the decision goal: if the risk gene position information shows that there is a gene variation related to physical function (yes / no), and the genetic risk information indicates that the genetic risk of physical function change is higher than 20% (yes / no), and the image texture information shows that the muscle texture is uneven (yes / no), and the image shape features show that the bone morphology is abnormal (yes / no), then the following decision tree is constructed: Root node: Is there a gene variation related to physical function? If yes, enter the next layer of nodes: Is the genetic risk higher than 20%? If yes, enter the next layer of nodes: Is the muscle texture uneven? If yes, finally enter the leaf node: Is there an abnormality in the bone morphology? If yes, the decision result is "a comprehensive physical function assessment is needed immediately."

[0075] The access decision tree is processed to generate a time decay factor and a cost-effectiveness coefficient. The time factor is considered based on the results of the access decision tree and the actual situation of the user. For example, if the decision result is that an assessment is required, but the user cannot proceed immediately due to busy work, etc. Assuming that the user decides to postpone the assessment for 3 months, the possibility of changes in physical function may increase over time. Through data analysis and model calculation, the time decay factor is 1.2 (indicates that the risk increases by 20% for every month of postponement).

[0076] In terms of cost, consider the cost of assessment, nutritional supplementation, possible work loss, etc. For example, the cost of a comprehensive physical function assessment is 2,000 yuan, the subsequent possible nutritional supplementation cost is 500 yuan per month, and the possible monthly work loss cost is 1,000 yuan. In terms of benefits, consider the risk of physical function decline and health loss that can be avoided through assessment and subsequent plans. Assuming that through timely assessment and implementation of the plan, the risk of significant changes in physical function in the next 10 years can be reduced from 30% to 10%, according to the relevant model calculation, the cost-effectiveness coefficient is 0.8 (indicates that for every 1 yuan of cost invested, 0.8 yuan of health benefits can be obtained).

[0077] The time decay factor and the cost-effectiveness coefficient are processed to generate preference alignment information. The time decay factor and the cost-effectiveness coefficient are processed comprehensively. For example, if the time decay factor is high (such as 1.2), it means that the risk of postponing the assessment increases rapidly, and the cost-effectiveness coefficient is relatively low (such as 0.8), then the preference alignment information may be "It is recommended that users adjust their schedules and conduct physical function assessments as soon as possible to reduce the risk of physical function changes. Although the current cost-effectiveness ratio is not very ideal, from a health perspective, timely assessment is more important." If the time decay factor is low and the cost-effectiveness coefficient is high, then the preference alignment information may be "Users can arrange assessments at a suitable time according to their own circumstances. The current risk of postponing the assessment is relatively small, and the cost-effectiveness ratio is reasonable."

[0078] S107, processing the question-and-answer information of the target user and the consulting purpose information of the target user to generate an initial prediction feature vector of the target user.

[0079] In one implementation, the question and answer information of the target user is processed to generate access keywords, and the access keywords are processed to generate question and answer semantic features. Assume that the target user asks on the health consultation platform: "I always feel tired recently, it is difficult to concentrate at work, and the sleep quality at night is not good, and I wake up easily due to many dreams. What should I do?" The system analyzes and extracts this question and answer information through natural language processing technology to generate access keywords: "fatigue", "inattention", "poor sleep quality", and "many dreams and easy awakening". These keywords concisely summarize the main problems and consultation focuses currently faced by the user, providing a basis for subsequent analysis. The above access keywords are further semantically analyzed to explore their deep meanings and associations. For example: "fatigue" may be related to the body's energy metabolism, psychological pressure or potential physiological dysfunction. "Inattention" may imply fluctuations in brain neural function, which may be affected by lack of sleep, malnutrition or mental state. "Poor sleep quality" and "many dreams and easy awakening" involve sleep cycle disorders, abnormal excitability of the nervous system, and possible psychological stress factors. These semantic understandings are converted into question and answer semantic feature vectors, in which each dimension represents a different semantic feature strength. For example, in a ten-dimensional vector space, the dimension value corresponding to "fatigue" is relatively high, while the values ​​of other dimensions with low correlation with fatigue are close to zero, thereby accurately representing the semantic connotation of user questions and laying the foundation for accurate matching of consulting resources.

[0080] The target user's consultation purpose information is processed to generate access type information, the access type information is processed to generate a prediction weighting coefficient, and the prediction weighting coefficient and question-answer semantic features are processed to generate the target user's initial prediction feature vector. The target user indicates during registration or interaction that the purpose of the consultation is to improve work efficiency and improve the quality of life. After the system processes the consultation purpose information, it determines that the access type information is "work and life health optimization". This type covers various consultation needs related to work stress management, daily life habit adjustment, and physical and mental health balance, helping the system to grasp the user's consultation intention from a macro level, so as to more accurately screen and recommend adapted knowledge and services. Based on a large number of case data and analysis results of similar access types in the past ("work and life health optimization"), combined with the specific situation of the current user (such as age, gender, occupation, etc.), the prediction weighting coefficient is calculated and generated through a machine learning algorithm. For example, for a 35-year-old office white-collar male, when dealing with the problem of "fatigue, sleep and concentration", due to the common stressors and health risk patterns of this age group and occupational group, he may give a higher weight (such as 0.7) to consulting advice related to stress relief and energy improvement, while giving a lower weight (such as 0.3) to the investigation of rare health problems. This allows the system to focus more on solutions with a high probability of being effective in subsequent predictions, thereby improving the accuracy of recommendations.

[0081] Integrate the prediction weighting coefficient and the question-answer semantic feature vector. Take the weighting coefficient [0.7, 0.3] (only two dimensions are used as examples) and the question-answer semantic feature vector [0.6, 0.4] (corresponding to the above two feature dimensions) as an example, and generate the elements of the initial prediction feature vector through dot multiplication, and get [0.42, 0.12] (the actual dimension depends on the specific model design). This vector is a quantitative expression of user needs. After entering the prediction model, it can accurately locate the health management strategy, lifestyle advice or further diagnosis direction that best meets the user's needs, and open up precise navigation for personalized health services.

[0082] S108: Process the initial prediction feature vector of the target user based on the preference alignment information to generate target prediction report information of the target user.

[0083] In one embodiment, the initial predicted feature vector of the target user is processed based on the preference alignment information to generate several feature matching values. Assuming that the preference alignment information indicates that the user prefers long-term stable maintenance and precise resource allocation for health management, the initial predicted feature vector of the target user includes [nutritional intake preference, exercise habit tendency, stress coping mode], with values ​​of [0.6, 0.4, 0.3] respectively. The system sets dynamic thresholds and weights for each dimension according to the preference alignment information (such as nutritional intake threshold 0.5, weight 0.4; exercise habit threshold 0.3, weight 0.3; stress coping threshold 0.2, weight 0.3). Calculate the feature matching value: the nutritional intake dimension is (0.6-0.5)×0.4=0.04; the exercise habit dimension is (0.4-0.3)×0.3=0.03; the stress coping dimension is (0.3-0.2)×0.3=0.03. These values ​​quantify the fit between user characteristics and the ideal health model, helping to accurately locate the entry point for health management.

[0084] Several feature matching values ​​are processed separately to generate target grouping information, where the target grouping information is used to characterize the target gene features corresponding to different groups and the weight information matching the target gene features. Target grouping information is generated by clustering based on feature matching values. For example, clusters with similar feature matching values ​​are grouped together. If the nutritional intake and exercise habits match values ​​are close, they are classified as the "active life group". The target gene features focus on genes related to energy metabolism and muscle function, and the weights are allocated according to the proportion of feature matching values, such as nutritional intake 0.57 (0.04 / (0.04+0.03)) and exercise habits 0.43; the stress coping matching value is a unique "mental adjustment group", and the target gene features involve neurotransmitter regulation genes with a weight of 1. This grouping draws a blueprint for precision health intervention and optimizes health strategies according to genetic characteristics and weights.

[0085] The target grouping information is processed to generate the target access list information, wherein the target access list information includes the access time interval sequence of different groups and the access list information of different groups. For the "active life group", the access time interval sequence is set according to its genetic characteristics and weights, such as physical examination of key metabolic indicators every 3 months, and receiving nutrition and exercise program push every 2 weeks; the access list includes fitness course recommendations and personalized nutrition recipes. The "mind adjustment group" sets the stress hormone detection interval every six months according to genetic characteristics, and the access list covers meditation training guidance and psychological counseling resources. This information is a customized dynamic health navigation for users, which accurately delivers health resources according to genetic characteristics and weights to improve the timeliness and quality of health management.

[0086] Process the target access list information to generate abnormal difference values. Monitor user feedback on the target access list, such as the deviation in the implementation of the nutrition plan of the "energizing life group" and the proportion of exercise duration that does not meet the standard, and compare the abnormal difference values ​​between the mean and the target value of the same group. For example, if the mean intake of key nutrients in the nutrition plan is 20% lower than the target, it will be counted as 0.2, and if the exercise duration does not meet the standard by 30%, it will be counted as 0.3. If the default rate of meditation frequency in the "mind adjustment group" is 40%, it will be counted as 0.4. This value quantifies the deviation of user behavior and provides a key basis for accurately adjusting health plans.

[0087] The abnormal difference values ​​are processed to generate the target prediction report information of the target users, and the abnormal difference values, grouping information and preference data are integrated into a report. For example, the energy metabolism genes of the "energetic life group" fluctuate due to nutritional imbalance and lack of exercise. According to the preference, it is recommended to optimize the diet structure, increase the intensity and frequency of exercise, and set a 2-month improvement period with progress indicators; the neurotransmitter genes of the "mental adjustment group" are disturbed by stress, and it is recommended to increase meditation for 15 minutes a day and enjoy psychological counseling once a month to help stabilize emotions and optimize gene expression. The report customizes the health path for users, based on genetic science and guided by behavioral data, to drive personalized health upgrades.

[0088] In another embodiment, the method further includes a calculation formula for obtaining an abnormal difference value, the calculation formula being: Where n is the number of packets in the access list information, ω i is the weight coefficient of the ith group, TAL i is the target access time interval of the ith group, PAL i is the actual access time interval of the ith group, P i is the power coefficient of the ith group.

[0089] Suppose we have a health management system that groups users and manages access time intervals.

[0090] The parameters are set as follows: Number of groups (n): Assume that we divide users into 3 groups, that is, n = 3. These 3 groups may be divided according to factors such as the user's health risk level, age, gender, etc. For example, the first group is young users with low health risks, the second group is middle-aged users with certain health risks, and the third group is elderly users or users with high health risks.

[0091] Weight coefficient (ω i ): For the first group (i=1), assume that its weight coefficient ω1=0.2, which means that the relative importance of this group in the overall evaluation is low. For the second group (i=2), ω2=0.3, its importance is moderate. For the third group (i=3), ω3=0.5, this group has the highest importance, probably because the health risk of the third group of users is the greatest and needs closer attention.

[0092] Target access time interval (TAL i ): The target visit interval TAL1 for the first group (young people with low health risks) is set to 12 months, that is, a health check-up is conducted once a year. The target visit interval TAL2 for the second group (middle-aged people with certain health risks) is set to 6 months, that is, a health check-up is conducted every six months. The target visit interval TAL3 for the third group (elderly or high health risks) is set to 3 months, that is, a health check-up is conducted once a quarter.

[0093] PAL i ): Assume that the actual access time interval PAL1 of the first group of users is 15 months, that is, they actually perform a health check once every 15 months. The actual access time interval PAL2 of the second group of users is 8 months, that is, they actually perform a health check once every 8 months. The actual access time interval PAL3 of the third group of users is 4 months, that is, they actually perform a health check once every 4 months.

[0094] Power coefficient (P i ): For the first group, let P1 = 2. For the second group, let P2 = 1. For the third group, let P3 = 3. These power coefficients may be determined based on the health risk sensitivity of different groups or other factors.

[0095] Calculate the abnormal difference value (ADV) according to the formula

[0096] For the first group:

[0097] For the second group:

[0098] For the third group:

[0099] The total abnormal difference value ADV = 1.8 + 0.6 + 0.5 = 2.9. The abnormal difference value ADV = 2.9 reflects the overall deviation between the user's actual access time interval and the target access time interval. The larger the value, the greater the deviation between the user's actual health check frequency and the target frequency set according to their grouping, and the health management strategy may need to be adjusted. For example, for the first group of users, their actual access time interval is 3 months longer than the target interval, and because the power coefficient is 2, this deviation is amplified in the calculation, resulting in a larger contribution to the overall abnormal difference value. Based on this result, the health management system can remind the first group of users to perform health checks more regularly, or re-evaluate the rationality of the grouping strategy and the target access time interval.

[0100] In this application, the server obtains multiple types of information of the target user, including physiological characteristics, question and answer records, identity information, and consultation purpose information, and prepares a training data set and a preset consultation result prediction model. When generating the initial physiological label, the physiological characteristics and identity information of the target user are processed. The classification label and medication type information are derived from the identity information, and then these information are combined with the physiological characteristic information to calculate the characteristic data association value, thereby generating the initial physiological label.

[0101] Preprocessing is required for the training data set. First, the data set is grouped by age group and gender to obtain the physiological characteristic data set of users in different groups. Then, features are extracted from the grouped data to form an original feature library. Next, the original feature library is divided into a training set and a validation set. The validation set is predicted by a classifier, and the training set is trained with a preset algorithm. The prediction results of both are processed to obtain target feature data that can characterize the presence of abnormal risk factors in the user's physiological characteristics. The initial physiological labels are processed based on the target consultation result prediction model to generate access preference and validity preference information. The process includes obtaining gene sequencing and medical imaging related information from the initial physiological labels, generating user portrait information, calculating the access matching value through the model, and then processing to obtain preference information.

[0102] Further process the access preference and effectiveness preference information, generate risk gene positions and other related information in turn, build an access decision tree, calculate the time decay factor and cost-effectiveness coefficient, and finally obtain the preference alignment information. By processing the target user's question and answer information and consultation purpose information, the initial prediction feature vector is obtained. This process includes generating access keywords, extracting question and answer semantic features, determining access type information, and calculating prediction weighting coefficients. Finally, the initial prediction feature vector is processed according to the preference alignment information to obtain multiple feature matching values, and then the target grouping information is generated. The information contains the target gene features corresponding to different groups and their matching weights. Based on the target grouping information, the target access list information is generated, including the access time interval sequence and access list of different groups. The target access list information is processed to generate abnormal difference values, and then the abnormal difference values ​​are processed to finally generate the target prediction report information of the target user. The server comprehensively collects multi-dimensional information and key model data of the target user to build the cornerstone of accurate health assessment. The initial physiological label is shaped from identity and physiological characteristics, and the user's health characteristics are deeply mined. The training data set is preprocessed to accurately identify abnormal risk factors and build a solid defense line for health warning. Leverage the power of the model to generate preference information, dig deep into needs from genetic images, and guide the direction of personalized health.

[0103] In one embodiment, if Figure 2 As shown, the present application also provides a multidisciplinary health consultation auxiliary management device based on a large language model, comprising:

[0104] An acquisition module 201 is used to acquire physiological characteristic information of a target user, question and answer information of a target user, identity information of a target user, consultation purpose information of a target user, a training data set, and a preset consultation result prediction model;

[0105] The processing module 202 is used to process the physiological characteristic information of the target user and the identity information of the target user to generate an initial physiological label of the target user; pre-process the training data set to generate a training data set with target characteristic data, wherein the target characteristic data is used to characterize the risk factors of abnormal physiological characteristics of the target user; process the preset consultation result prediction model based on the training data set with target characteristic data to generate a target consultation result prediction model; process the initial physiological label of the target user based on the target consultation result prediction model to generate access preference information and validity preference information; process the access preference information and the validity preference information to generate preference alignment information; process the question and answer information of the target user and the consultation purpose information of the target user to generate an initial prediction feature vector of the target user; process the initial prediction feature vector of the target user based on the preference alignment information to generate target prediction report information of the target user.

[0106] Each embodiment in this application is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for evaluating the multidisciplinary health consultation auxiliary management method based on a large language model, electronic device, electronic device, and readable storage medium embodiment, since it is basically similar to the above-mentioned embodiment of the multidisciplinary health consultation auxiliary management method based on a large language model, the description is relatively simple, and the relevant parts can be referred to the partial description of the embodiment of the multidisciplinary health consultation auxiliary management method based on a large language model.

Claims

1. A multidisciplinary health consultation auxiliary management method based on a large language model, characterized in that: include: Obtaining physiological characteristic information of the target user, question and answer information of the target user, identity information of the target user, consultation purpose information of the target user, a training data set, and a preset consultation result prediction model; Processing the physiological characteristic information of the target user and the identity information of the target user to generate an initial physiological tag of the target user; Preprocessing the training data set to generate a training data set with target feature data, wherein the target feature data is used to characterize risk factors of abnormal physiological characteristics of the target user; Processing the preset consultation result prediction model based on the training data set with target feature data to generate a target consultation result prediction model; Processing the initial physiological tag of the target user based on the target consultation result prediction model to generate access preference information and validity preference information; Processing the access preference information and the validity preference information to generate preference alignment information; Processing the question-and-answer information of the target user and the consulting purpose information of the target user to generate an initial prediction feature vector of the target user; The initial prediction feature vector of the target user is processed based on the preference alignment information to generate target prediction report information of the target user.

2. The method according to claim 1, characterized in that Processing the physiological characteristic information of the target user and the identity information of the target user to generate an initial physiological tag of the target user includes: Processing the identity information of the target user to generate a classification label of the target user and medication type information of the target user; Processing the target user's classification label, the target user's medication type information, and the target user's physiological characteristic information to generate a characteristic data association value for the target user; Processing the characteristic data association value of the target user to generate an initial physiological tag of the target user; The method includes a calculation formula for obtaining a characteristic data association value of a target user, and the calculation formula is: Among them, a i is the weight coefficient corresponding to the th physiological characteristic information, n is the number of physiological characteristic information, X i is the target user’s i-th physiological characteristic information value, X 0i is the reference value of the physiological characteristic information, σ i is the standard deviation of the ith physiological characteristic information value in the population, b j is the weight coefficient corresponding to the j-th medication type information, m is the number of medication type information, Y j is the target user’s j-th medication type information value, Y 0j is the reference value of the j-th medication type information, c k is the weight coefficient corresponding to the kth hierarchical label information, p is the number of hierarchical label information, Z k is the kth graded label information value of the target user, Z 0k is the reference value of the k-th classification label information, Z maxk and Z mink are the maximum and minimum values ​​of the k-th classification label information in the population, respectively.

3. The method according to claim 1, characterized in that Preprocessing the training data set to generate a training data set with target feature data includes: Performing grouping processing on the training data set to generate a grouped training data set, wherein the grouped training data set includes physiological characteristic information of users of different age groups and genders; Perform feature extraction on the grouped training data set to generate an original feature library; Processing the original feature library to generate a training set and a verification set; Performing prediction processing on the verification set based on the classifier to generate a prediction result; Performing training processing on the training set based on a preset algorithm to generate a prediction result of the validation set class; The prediction result and the validation set prediction result are processed to generate target feature information, where the target feature information is used to characterize risk factors for abnormal physiological characteristics of the target user.

4. The method according to claim 1, characterized in that The initial physiological tag of the target user is processed based on the target consultation result prediction model to generate access preference information and validity preference information, including: Processing the initial physiological label of the target user to generate gene sequencing information of the target user and medical imaging information of the target user; Processing the gene sequencing information of the target user and the medical imaging information of the target user to generate target user portrait information; Processing the target user portrait information based on the target consultation result prediction model to generate an access matching value for the target user; Processing the access matching value of the target user to generate access preference information and validity preference information; The method also includes a calculation formula for obtaining the access matching value of the target user, and the calculation formula is: Among them, x represents the initial physiological label of the target user; y w Represents the high-risk-related features extracted from the target user's gene sequencing information and medical imaging information; y l represents the low-risk-related features extracted from the target user's gene sequencing information and medical imaging information; β represents the weight coefficient; σ represents the base of the logarithm.

5. The method according to claim 4, characterized in that Processing the access preference information and the validity preference information to generate preference alignment information includes: Processing the access preference information and the validity preference information to generate risk gene position information, genetic risk information, image texture information, and image shape features; Processing the risk gene locus information, the genetic risk information, the image texture information and the image shape feature to generate an access decision tree; Processing the access decision tree to generate a time decay factor and a cost-effectiveness coefficient; The time decay factor and the cost-effectiveness coefficient are processed to generate preference alignment information.

6. The method according to claim 5, characterized in that Processing the question-and-answer information of the target user and the consulting purpose information of the target user to generate an initial prediction feature vector of the target user includes: Processing the question and answer information of the target user to generate access keywords; Processing the access keywords to generate question-answer semantic features; Processing the consulting purpose information of the target user to generate access type information; Processing the access type information to generate a prediction weighting coefficient; The prediction weight coefficient and the question-answer semantic feature are processed to generate an initial prediction feature vector of the target user.

7. The method according to claim 6, characterized in that The initial prediction feature vector of the target user is processed based on the preference alignment information to generate target prediction report information of the target user, including: Processing the initial predicted feature vector of the target user based on the preference alignment information to generate a number of feature matching values; Processing a plurality of feature matching values ​​respectively to generate target grouping information, wherein the target grouping information is used to characterize target gene features corresponding to different groups and weight information matching the target gene features; Processing the target group information to generate target access list information, wherein the target access list information includes access time interval sequences of different groups and access list information of different groups; Processing the target access list information to generate an abnormal difference value; Processing the abnormal difference value to generate target prediction report information for the target user; The method also includes a calculation formula for obtaining an abnormal difference value, and the calculation formula is: Where n is the number of packets in the access list information, ω i is the weight coefficient of the ith group, TAL i is the target access time interval of the ith group, PAL i is the actual access time interval of the ith group, P i is the power coefficient of the ith group.

8. A multidisciplinary health consultation auxiliary management device based on a large language model, characterized in that: The device comprises: An acquisition module, used to acquire physiological characteristic information of a target user, question and answer information of a target user, identity information of a target user, consultation purpose information of a target user, a training data set, and a preset consultation result prediction model; A processing module is used to process the physiological characteristic information of the target user and the identity information of the target user to generate an initial physiological label of the target user; pre-process the training data set to generate a training data set with target characteristic data, wherein the target characteristic data is used to characterize the risk factors of abnormal physiological characteristics of the target user; based on the training data set with target characteristic data, the preset consultation result prediction model is processed to generate a target consultation result prediction model; based on the target consultation result prediction model, the initial physiological label of the target user is processed to generate access preference information and validity preference information; the access preference information and the validity preference information are processed to generate preference alignment information; the question and answer information of the target user and the consultation purpose information of the target user are processed to generate an initial prediction feature vector of the target user; the initial prediction feature vector of the target user is processed based on the preference alignment information to generate target prediction report information of the target user.

9. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; Wherein, the first processor is configured to execute the multidisciplinary health consultation auxiliary management method based on a large language model as described in any one of claims 1 to 7 by executing the executable instructions.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the multidisciplinary health consultation auxiliary management method based on a large language model as described in any one of claims 1 to 7 is implemented.

Images