A functional medical active health management method and system of large model fine-tuning engineering

Abstract

This invention provides a functional medicine proactive health management method and system for large-scale model fine-tuning engineering. The method includes the following steps: S101, pre-training a general large-scale model using functional medicine literature and clinical guidelines; S102, fine-tuning the general large-scale model using medical tasks to form a medical-specific large-scale model; S103, integrating LSTM and Transformer temporal algorithms, and introducing autoencoder model reconstruction error as an anomaly scoring indicator; S104, integrating reinforcement learning and generative adversarial networks to output personalized health intervention suggestions based on user health risks; S105, collecting user health behavior data in real time through IoT wearable devices and portable detection instruments, preprocessing the health behavior data to generate a structured health dataset, and inputting it into the medical-specific large-scale model for user health assessment and management. This invention achieves a closed-loop process from data collection, accurate assessment, dynamic early warning to personalized intervention, meeting the needs of functional medicine proactive health management.

Description

Technical Field

[0001] This invention relates to the field of health management technology, and in particular to a functional medicine proactive health management method and system for large-scale model fine-tuning engineering. Background Technology

[0002] Current mainstream methods or systems in the field of health management mostly focus on single-dimensional data collection and basic health monitoring, relying primarily on traditional medical data (outpatient medical records, physical examination reports) or single-type sensor data (such as only blood pressure and blood glucose monitoring), lacking the ability to deeply integrate multi-source heterogeneous data. At the technology application level, most adopt early warning mechanisms based on fixed thresholds, equipped with general algorithm models, without combining medical vertical knowledge for targeted optimization; health management solutions are mainly based on standardized templates, unable to adapt to individual differences such as genetic background, metabolic characteristics, and lifestyle habits.

[0003] Meanwhile, current mainstream methods and systems in the field of health management generally suffer from data silos: data is not shared between medical institutions, and multi-omics data such as genomics, metabolomics, and proteomics are not effectively integrated with clinical data and IoT wearable data, resulting in a lack of comprehensive insights from the molecular to the behavioral levels in health assessments; moreover, they mostly adopt centralized computing architectures, which are highly dependent on network bandwidth and lack real-time response capabilities, making it impossible to achieve early and accurate warnings and dynamic interventions for health risks. Such systems have been applied in general health monitoring scenarios, but in the field of functional medicine and proactive health management, they are unable to meet the core requirements of precision, personalization, and forward-looking approaches. Summary of the Invention

[0004] Therefore, the purpose of this invention is to provide a functional medicine proactive health management method and system for large-scale model fine-tuning engineering, so as to solve or at least partially solve the above-mentioned problems existing in the prior art.

[0005] To achieve the above objectives, the first aspect of the present invention provides a functional medicine proactive health management method for large-scale model fine-tuning engineering, the method comprising the following steps: S101. Obtain functional medicine literature and clinical guidelines to pre-train the general model and strengthen its medical knowledge reserves. S102. Based on the general large model, set up a variety of medical tasks, and use the medical tasks to fine-tune the general large model to form a medical-specific large model. S103. Based on a large medical-specific model, it integrates LSTM and Transformer time series algorithms and introduces autoencoder model reconstruction error as an anomaly scoring index to adaptively identify users' health risks. S104. Based on a large medical-specific model, it integrates reinforcement learning and generative adversarial networks to output personalized health intervention suggestions according to the user's health risks. S105. Through IoT wearable devices and portable testing instruments, real-time collection of users' health behavior data is performed. The health behavior data is preprocessed to generate a structured health dataset, which is then input into a large medical model for health assessment and management of users.

[0006] Furthermore, step S101 specifically includes the following steps: S11. Obtain functional medicine literature, clinical guideline documents and health management record samples, perform deduplication and annotation, construct a functional medicine-specific corpus, and form a pre-training dataset; S12. Pre-train the general large model based on the pre-trained dataset, optimize the parameters of the general large model through the batch gradient descent algorithm, minimize the cross-entropy loss function, and strengthen the medical knowledge reserve of the general large model.

[0007] Furthermore, step S102 specifically includes the following steps: S21. Based on a general large model, set up a variety of medical tasks, including health risk prediction, interpretation of functional medicine indicators and generation of personalized health intervention suggestions; S22. Based on various medical tasks, a multi-task learning framework is constructed. The self-attention mechanism of the general large model is optimized by adopting the Transformer variant architecture, and a multi-omics data-specific attention head is introduced. The multi-omics data-specific attention head is based on the standard attention head and divided according to the data type of the multi-omics data. This allows the general large model to learn the heterogeneous data processing requirements and generate a medical-specific large model.

[0008] Furthermore, the fusion of LSTM and Transformer timing algorithms specifically includes the following steps: S31. Collect users' historical health behavior data, sort them by timestamp to form a sequence, use the LSTM time series algorithm to capture the long-term gradual trend of historical health behavior data, and output the time series feature vector. S32. Based on time-series feature vectors, the Transformer time-series algorithm is used to mine the potential correlations between various health indicators in historical health behavior data through the attention mechanism, and output collaborative feature vectors.

[0009] Furthermore, the introduction of autoencoder model reconstruction error as an anomaly scoring metric specifically includes the following steps: S41. Construct an autoencoder model, standardize the historical health behavior data, and input the standardized historical health behavior data into the autoencoder model. Calculate the reconstruction error as an anomaly scoring index, as shown below:

[0010] in, For anomaly scoring indicators, As a health indicator dimension, For the original number Health index values, To reconstruct the first Health index values, This is the output vector reconstructed from the input data by the autoencoder model. The standardized historical health behavior data vector is input into the autoencoder model; S42. A weighted summation algorithm is used to fuse time-series feature vectors, collaborative feature vectors, and anomaly scoring indicators to generate a comprehensive health risk score, which is used to identify the user's health risks.

[0011] Furthermore, the reinforcement learning uses health behavior data as the state space, health intervention suggestions as the action space, and health improvement effects as the reward function, specifically including the following steps: S51. The state space integrates health behavior data, multi-omics data and comprehensive health risk scores to construct a high-dimensional state vector; S52. The action space is divided into multiple categories of health intervention actions, and each category of health intervention actions includes specific parameters, forming a set of health intervention actions; S53. Based on a high-dimensional state vector and a set of health intervention actions, a dual reward mechanism is designed to iteratively optimize health intervention suggestions. The reward function is expressed as follows:

[0012] in, For the reward function, These are the weighting coefficients. It is a high-dimensional state vector. A collection of health intervention actions, For the first A comprehensive health risk score for each step. For target health risk scoring, For long-term evaluation of the time step, For the first The high-dimensional state vector of the step, As a short-term reward mechanism, As a long-term reward mechanism, For the first A comprehensive health risk score for each step. For the first A comprehensive health risk score for each step.

[0013] Furthermore, the generative adversarial network includes a generator and a discriminator. The generator is used to produce diverse health intervention suggestions, and the discriminator is used to evaluate the suitability of the health intervention suggestions to the user. Specifically, it includes the following steps: S61. The generator takes a high-dimensional state vector as input and generates diverse health intervention suggestions through a multilayer perceptron and a convolutional layer. S62. The discriminator calculates a fit score between health intervention suggestions and users based on diverse health intervention suggestions and high-dimensional state vectors, as shown below:

[0014] in, For state distribution, To ensure the distribution of truly effective health intervention programs, Let be the generator loss function. Let the discriminator loss function be... For mathematical expectation operators, and The discriminator scores the suitability of the health intervention recommendations output by the generator to the user. It is Gaussian noise. For diverse health intervention recommendations, This represents the range of Gaussian noise distribution. S63. Integrate the health intervention suggestions that have been iteratively optimized in reinforcement learning with the health intervention suggestions that have been assessed by fit scores, and then screen them through medical-grade evaluation indicators to finally output personalized health intervention suggestions.

[0015] Furthermore, the user's health behavior data includes physiological indicators, lifestyle data, and psychological state data, and the health behavior data is synchronously integrated with multi-omics data such as genomics, metabolomics, proteomics, and clinical data.

[0016] A second aspect of the present invention provides a functional medicine active health management system for large-scale model fine-tuning engineering, the system comprising: Edge layer: includes IoT wearable devices, portable testing instruments and edge computing nodes. The IoT wearable devices and portable testing instruments are connected to the edge computing nodes via Bluetooth and WiFi. Cloud layer: includes cloud server cluster and distributed database. The cloud server cluster is used to deploy large-scale medical models. The edge computing nodes are connected to the cloud server cluster via the network. The distributed database is used to store users' health behavior data and health assessment and management results. Application layer: includes smart terminals, which are used to connect to the cloud layer via a network to receive the user's health assessment and management results.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a functional medicine-based proactive health management method using a large-scale model fine-tuning project. It pre-trains a general-purpose large-scale model using functional medicine literature, clinical guidelines, and health management records to enhance its medical knowledge reserves. Fine-tuning of the model using medical tasks results in a medical-specific large-scale model with medical-grade interpretation capabilities. By integrating LSTM and Transformer temporal algorithms and introducing an autoencoder model, it adaptively identifies users' health risks. Furthermore, by fusing reinforcement learning and generative adversarial networks, it outputs personalized health intervention suggestions that combine medical effectiveness with individual adaptability. Preprocessing of health behavior data creates a standardized format for the data. This invention achieves a closed-loop process from data collection, accurate assessment, dynamic early warning, to personalized intervention, meeting the needs of functional medicine-based proactive health management. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of a functional medicine proactive health management method for large-scale model fine-tuning engineering, provided as an embodiment of the present invention. Detailed Implementation

[0020] The principles and features of the present invention are described below with reference to the accompanying drawings. The listed embodiments are only used to explain the present invention and are not intended to limit the scope of the present invention.

[0021] Reference Figure 1 This embodiment provides a functional medicine proactive health management method for large-scale model fine-tuning engineering, the method comprising the following steps: S101. Obtain functional medicine literature, clinical guidelines and health management records to pre-train the general model and strengthen the medical knowledge reserves of the general model. S102. Based on the general large model, set up a variety of medical tasks, and use the medical tasks to fine-tune the general large model to form a medical-specific large model. S103. Based on a large medical-specific model, it integrates LSTM and Transformer time series algorithms and introduces autoencoder model reconstruction error as an anomaly scoring index to adaptively identify users' health risks. S104. Based on a large medical-specific model, it integrates reinforcement learning and generative adversarial networks to output personalized health intervention suggestions according to the user's health risks. S105. Through IoT wearable devices and portable testing instruments (such as metabolomics testing devices and protein testing devices), real-time health behavior data of users is collected. The health behavior data is preprocessed to generate a structured health dataset. The structured health dataset is then input into a large medical model to conduct health assessment and management of users, and output the health assessment and management results, including the user's health risks and personalized health intervention suggestions.

[0022] This embodiment achieves a closed-loop process from health data collection, accurate health assessment, dynamic health warning to personalized health intervention by integrating multi-source data, fine-tuning a large model, generating health dynamic warnings, and generating personalized health intervention suggestions. This overcomes the technical deficiencies in the current health management field and meets the needs of functional medicine for proactive health management. The technical deficiencies are as follows: (1) The health dimension is singular, making it impossible to locate the root cause of health problems; (2) It relies on fixed thresholds to trigger health warnings, making it impossible to capture subtle trend changes in health indicators and the synergistic effect of multiple health indicators, and making it difficult to issue effective warnings in the early stages of health risk formation; (3) Health intervention suggestions are generated based on general templates, which cannot adapt to the individual differences of users, resulting in low compliance with health intervention suggestions; (4) The general large model is fine-tuned with medical vertical knowledge, lacks professional medical literacy, and cannot accurately interpret complex health data or generate medical-grade professional suggestions.

[0023] Step S101 specifically includes the following steps: S11. Construct a functional medicine-specific corpus, covering 100,000+ functional medicine articles, 50,000+ clinical metabolic diseases, chronic disease management and other sub-fields of guidelines and documents, and 30,000+ structured health management records, etc., and deduplicatize and annotate them, and form a pre-training dataset after adopting the BioBERT annotation standard.

[0024] S12. Pre-train the general-purpose large model based on the pre-training dataset, optimize the parameters of the general-purpose large model using the batch gradient descent algorithm, minimize the cross-entropy loss function, and enhance the medical knowledge reserve of the general-purpose large model, as shown below:

[0025] in, To minimize the cross-entropy loss function, The number of samples in the corpus. The length of a single sample sequence. For the first The first sample The real label for each location Given the preceding input sequence, For general large model learnable parameters, To ensure that the general large model has learnable parameters Given a preceding input sequence At that time, the general large model predicted the first The first sample Real labels for each location The conditional probability. A unified hyperparameter setting: learning rate ( The iteration count was 100 rounds, the batch size was 32, and the dropout coefficient was 0.15 to avoid overfitting.

[0026] Step S102 specifically includes the following steps: S21. Based on a general large model, set up a variety of medical tasks, including health risk prediction, interpretation of functional medicine indicators and generation of personalized health intervention suggestions.

[0027] S22. Based on various medical tasks, a multi-task learning framework is constructed. A Transformer variant architecture is used to optimize the self-attention mechanism of the general-purpose large model, and a multi-omics data-specific attention head is introduced. This multi-omics data-specific attention head is based on the standard attention head, but is divided according to the data type of the multi-omics data (e.g., independent and dedicated attention heads are created for genomics, metabolomics, and proteomics, forcing each attention head to be responsible only for the internal dependencies or associations of its corresponding omics data). This accurately captures the dependencies between multi-omics data (e.g., co-expression patterns of gene expression, co-occurrence patterns of mutations), learns the associations between multi-omics data, reduces interference and information confusion between different omics data, and enables the general-purpose large model to learn the heterogeneous data processing requirements. A joint loss function is defined to balance the weights of various medical tasks, generating a medical-specific large model with medical-grade interpretation capabilities. The joint loss function is expressed as follows:

[0028] in, For the joint loss function, For the cross-entropy loss of the health risk prediction task, Semantic matching loss for interpreting functional medicine indicators. The perplexity loss for generating personalized health recommendations. , , These are the weighting coefficients, uniformly set as follows: , , The Transformer variant architecture parameters are unified as follows: 8 attention heads, 6 encoder layers, 6 decoder layers, 1024 hidden layer dimensions, and 25% attention heads dedicated to multi-omics data, enhancing the feature capture capability of genomic and metabolomics data.

[0029] Traditional machine learning models, which use algorithms like logistic regression and ordinary random forests to replace large-scale model fine-tuning and time-series fusion models, lack the ability to adapt to specific medical knowledge, cannot handle heterogeneous health data, and provide insufficient professionalism in health advice, failing to meet medical-grade standards. This embodiment, however, addresses these issues by employing a hierarchical pre-training and multi-task learning optimization strategy based on large-scale functional medicine literature, clinical guidelines, and health management records, along with a Transformer variant architecture adapted to heterogeneous health data.

[0030] In step S103, the fusion of LSTM and Transformer timing algorithms specifically includes the following steps: S31. The LSTM time series algorithm uses a gating mechanism to handle long-term dependencies and identify gradual trends in health data such as metabolic indicators. It collects historical health behavior data from users, such as data from the past 6 months (sampling frequency of once per hour), and sorts it by timestamp to form a sequence, as shown below:

[0031] in, For sequence, For the first Step-by-step multi-dimensional health indicator vector, For time step.

[0032] Based on sequences, the LSTM time-series algorithm is used to capture the long-term gradual trend of historical health behavior data. The LSTM time-series algorithm gate unit is represented as follows:

[0033] in, , , These are the input gate, forget gate, and output gate, respectively. In cellular state, For the first The hidden state of the time step. It is the Sigmoid activation function. For element-wise multiplication, For the first The hidden state of the time step. For the first Multidimensional health indicators of time steps , , , These are the weight matrices for the current input gate, forget gate, output gate, and cell state, respectively. Input for the current time (i.e.) ); , , , These are the weight matrices from the hidden state in the previous time step to the input gate, forget gate, output gate, and cell state, respectively. The hidden state from the previous moment; This is a bias term for the candidate cell state values. The LSTM temporal algorithm parameters are adapted to the above hyperparameter settings, specifically: hidden layer dimension 128, dropout coefficient 0.2, learning rate 0.0005, number of iterations 50, and output temporal feature vector, represented as follows:

[0034] in, This is a time-series feature vector.

[0035] S32. The Transformer time-series algorithm analyzes the synergistic effect of multiple health indicators using a self-attention mechanism. Based on the time-series feature vector, the Transformer time-series algorithm uses an attention mechanism to mine the potential correlations between various health indicators in historical health behavior data and outputs a synergistic feature vector, as shown below:

[0036] in, For querying the matrix, The key matrix, For value matrices, The dimension of the key vector is set to 128. To output the projection matrix, For the first Each attention head output, This is a transpose operation.

[0037] In step S103, the introduction of autoencoder model reconstruction error as an anomaly scoring index specifically includes the following steps: S41. Construct an autoencoder model, standardize historical health behavior data, and input the standardized historical health behavior data into the autoencoder model. Compress features through the encoder, reconstruct the output through the decoder, and calculate the reconstruction error as an anomaly scoring index, as shown below:

[0038] in, For anomaly scoring indicators, As a health indicator dimension, For the original number Health index values, To reconstruct the first Health index values, This is the output vector reconstructed from the input data by the autoencoder model. This is a standardized historical health behavior data vector input to the autoencoder model. A unified reconstruction error threshold is set. 0.05, then It was determined to be an abnormal scoring indicator.

[0039] S42. A weighted summation algorithm is used to fuse time-series feature vectors, collaborative feature vectors, and anomaly scoring indicators to generate a comprehensive health risk score, which is used to identify users' health risks. This is represented as follows:

[0040] in, To achieve a comprehensive health risk score, Collaborative feature vectors. If <0.3 indicates low risk. For medium risk, For high risk, For extremely high risk, core health risk points and related indicators will be output simultaneously.

[0041] This embodiment addresses the problem that using only a single LSTM or Transformer algorithm to build an early warning model, without integrating autoencoder anomaly detection, cannot adaptively identify users' health risks, and lacks sufficient analysis of the synergistic effects of multiple health indicators.

[0042] In step S104, the reinforcement learning uses health behavior data as the state space, health intervention suggestions as the action space, and health improvement effects as the reward function, specifically including the following steps: S51. The state space integrates the user's real-time physiological indicators, historical lifestyle data, multi-omics data, and comprehensive health risk score to construct a high-dimensional state vector, represented as follows:

[0043] in, For the first The high-dimensional state vector of the time step. This is a feature vector of physiological indicators. For historical lifestyle data feature vectors, These are feature vectors for multi-omics data.

[0044] S52. The action space is divided into health intervention actions such as diet, exercise, detection, and rest. Each type of health intervention action includes specific parameters (such as intensity, duration, and frequency of exercise actions), forming a set of health intervention actions, as shown below:

[0045] in, A collection of health intervention actions, For specific dietary parameters, Specific parameters of motion To detect specific parameters, These are the specific parameters for the work and rest schedule.

[0046] S53. Based on a high-dimensional state vector and a set of health intervention actions, a dual reward mechanism is designed to balance short-term immediate effects and long-term health benefits. The health intervention recommendations are iteratively optimized, and the reward function is expressed as follows:

[0047] in, For the reward function, These are the weighting coefficients. It is a high-dimensional state vector. A collection of health intervention actions, For the first A comprehensive health risk score based on time steps. For target health risk scoring, The time step for long-term evaluation is 3 months. For the first The high-dimensional state vector of the time step. As a short-term reward mechanism, As a long-term reward mechanism, For the first The comprehensive health risk score is calculated step by step. In this embodiment, the DQN algorithm is used to optimize the policy network with a learning rate of 0.001, an experience replay pool capacity of 100,000, a target network update frequency of 100 steps, and a discount factor of 0.9.

[0048] In step S104, the generative adversarial network includes a generator and a discriminator. The generator is used to produce diverse health intervention suggestions, and the discriminator is used to evaluate the suitability of the health intervention suggestions to the user. Adversarial training optimizes the quality of the health intervention suggestions, specifically including the following steps: S61. The generator takes a high-dimensional state vector as input and generates diverse health intervention suggestions through a multilayer perceptron (MLP) and convolutional layers, as shown below:

[0049] in, For diverse health intervention recommendations, It is Gaussian noise (mean 0, variance 0.1). A generator for health intervention recommendations; S62. The discriminator calculates a fit score between health intervention suggestions and users based on diverse health intervention suggestions and high-dimensional state vectors, as shown below:

[0050] in, Rate the fit.

[0051] The adversarial loss function of the generative adversarial network is expressed as follows:

[0052] in, The state distribution (i.e., the probability distribution of the user's high-dimensional state vector) defines the real user health state scenario and ensures that the patterns learned by the generator and discriminator are consistent with the actual changes in the user's health state. The distribution of real and effective health intervention programs (i.e., the probability distribution of real and effective health intervention recommendations) provides positive sample supervision signals for the discriminator, enabling the discriminator to learn to distinguish between real and effective health intervention programs and health intervention recommendations generated by the generator. The range of Gaussian noise distribution. Let be the generator loss function. Let the discriminator loss function be... For mathematical expectation operators, The discriminator scores the fit between the diverse health intervention suggestions output by the generator and the user's high-dimensional state vector. Both the generator and discriminator have a learning rate of 0.0002, a batch size of 16, and 200 iterations, using the Adam optimizer.

[0053] S63. Integrate the health intervention suggestions that have been iteratively optimized in reinforcement learning with the health intervention suggestions that have been assessed by fit, and screen them through medical-grade evaluation indicators to finally output personalized health intervention suggestions that combine medical effectiveness and individual fit.

[0054] In step S105, the user's health behavior data includes physiological indicators, lifestyle habits, and psychological state data. This health behavior data is synchronously integrated with multi-omics data from genomics, metabolomics, proteomics, and clinical data. After ETL cleaning, standardization encoding (ICD-10, ATC, etc.), deduplication, and desensitization, a structured health dataset is generated. The physiological indicators include heart rate variability, blood oxygen saturation, and skin resistance, reflecting the basic physiological function of the human body. The lifestyle habits data includes dietary structure, exercise intensity, and sleep patterns, reflecting behavioral factors affecting health status. The psychological state data includes stress index and mood fluctuations, reflecting the level of holistic physical and mental health.

[0055] Another embodiment of the present invention provides a functional medicine active health management system for large-scale model fine-tuning engineering, the system comprising: Edge layer: includes IoT wearable devices (integrating sensors such as heart rate, blood oxygen, and skin resistance), portable testing instruments, and edge computing nodes (such as embedded servers). The IoT wearable devices and portable testing instruments are connected to the edge computing nodes via Bluetooth 5.0 and WiFi 6 to transmit the user's structured health dataset to the cloud layer, enabling real-time data transmission.

[0056] The cloud layer comprises a cloud server cluster, a distributed database, and federated learning nodes. The cloud server cluster is used to deploy a large-scale medical model. The edge computing nodes are connected to the cloud server cluster via a network to input structured health datasets into the large-scale medical model. The distributed database is used to store users' health behavior data and health assessment and management results. The federated learning nodes are used to configure independent encryption modules to support data anonymization and encrypted transmission, thus mitigating the privacy risks caused by centralized hardware storage.

[0057] Application Layer: This layer includes smart terminals, which may include enterprise employee terminals (such as smartphones, tablets, etc.), community elderly terminals (such as smartphones, wearable watches, etc.), and medical care terminals (such as doctor workstations, nurse tablets, etc.). These smart terminals connect to the cloud layer via a network to receive users' health assessment and management results and to provide feedback on user performance.

[0058] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A functional medicine active health management method of large model fine-tuning engineering, characterized in that, The method includes the following steps: S101. Obtain functional medicine literature and clinical guidelines to pre-train the general model and strengthen its medical knowledge reserves. S102. Based on the general large model, set up a variety of medical tasks, and use the medical tasks to fine-tune the general large model to form a medical-specific large model. S103. Based on a large medical-specific model, it integrates LSTM and Transformer time series algorithms and introduces autoencoder model reconstruction error as an anomaly scoring index to adaptively identify users' health risks. S104. Based on a large medical-specific model, it integrates reinforcement learning and generative adversarial networks to output personalized health intervention suggestions according to the user's health risks. S105. Through IoT wearable devices and portable testing instruments, real-time collection of users' health behavior data is performed. The health behavior data is preprocessed to generate a structured health dataset, which is then input into a large medical model for health assessment and management of users.

2. The functional medical active health management method of large model fine-tuning engineering according to claim 1, characterized in that, Step S101 specifically includes the following steps: S11. Obtain functional medicine literature, clinical guidelines and health management records, perform deduplication and annotation, construct a functional medicine-specific corpus, and form a pre-training dataset; S12. Pre-train the general large model based on the pre-trained dataset, optimize the parameters of the general large model through the batch gradient descent algorithm, minimize the cross-entropy loss function, and strengthen the medical knowledge reserve of the general large model.

3. The functional medicine active health management method of fine-tuning engineering of a large model according to claim 2, characterized in that, Step S102 specifically includes the following steps: S21. Based on a general large model, set up a variety of medical tasks, including health risk prediction, interpretation of functional medicine indicators and generation of personalized health intervention suggestions; S22. Based on various medical tasks, a multi-task learning framework is constructed. The self-attention mechanism of the general large model is optimized by adopting the Transformer variant architecture, and a multi-omics data-specific attention head is introduced. The multi-omics data-specific attention head is based on the standard attention head and divided according to the data type of the multi-omics data. This allows the general large model to learn the heterogeneous data processing requirements and generate a medical-specific large model.

4. The functional medicine active health management method of fine-tuning engineering of a large model according to claim 3, characterized in that, The fusion of LSTM and Transformer timing algorithm specifically includes the following steps: S31. Collect users' historical health behavior data, sort them by timestamp to form a sequence, use the LSTM time series algorithm to capture the long-term gradual trend of historical health behavior data, and output the time series feature vector. S32. Based on time-series feature vectors, the Transformer time-series algorithm is used to mine the potential correlations between various health indicators in historical health behavior data through the attention mechanism, and output collaborative feature vectors.

5. The functional medicine active health management method of fine-tuning engineering of a large model according to claim 4, characterized in that, The introduction of autoencoder model reconstruction error as an anomaly scoring metric specifically includes the following steps: S41. Construct an autoencoder model, standardize the historical health behavior data, and input the standardized historical health behavior data into the autoencoder model. Calculate the reconstruction error as an anomaly scoring index, as shown below: in, For anomaly scoring indicators, As a health indicator dimension, For the original number Health index values To reconstruct the first Health index values This is the output vector reconstructed from the input data by the autoencoder model. The standardized historical health behavior data vector is input into the autoencoder model; S42. A weighted summation algorithm is used to fuse time-series feature vectors, collaborative feature vectors, and anomaly scoring indicators to generate a comprehensive health risk score, which is used to identify the user's health risks.

6. The functional medicine proactive health management method for large-scale model fine-tuning engineering according to claim 5, characterized in that, The reinforcement learning method uses health behavior data as the state space, health intervention suggestions as the action space, and health improvement effects as the reward function, and specifically includes the following steps: S51. The state space integrates health behavior data, multi-omics data and comprehensive health risk scores to construct a high-dimensional state vector; S52. The action space is divided into multiple categories of health intervention actions, and each category of health intervention actions includes specific parameters, forming a set of health intervention actions; S53. Based on a high-dimensional state vector and a set of health intervention actions, a dual-reward mechanism is designed to iteratively optimize health intervention suggestions, as shown below: in, For the reward function, These are the weighting coefficients. It is a high-dimensional state vector. A collection of health intervention actions, For the first A comprehensive health risk score for each step. For target health risk scoring, For long-term evaluation time steps, For the first The high-dimensional state vector of the step, As a short-term reward mechanism, As a long-term reward mechanism, For the first A comprehensive health risk score for each step. For the first A comprehensive health risk score for each step.

7. The functional medicine proactive health management method for large-scale model fine-tuning engineering according to claim 6, characterized in that, The generative adversarial network includes a generator and a discriminator. The generator is used to produce diverse health intervention suggestions, and the discriminator is used to evaluate the suitability of the health intervention suggestions to the user. Specifically, it includes the following steps: S61. The generator takes a high-dimensional state vector as input and generates diverse health intervention suggestions through a multilayer perceptron and a convolutional layer. S62. The discriminator calculates a fit score between health intervention suggestions and users based on diverse health intervention suggestions and high-dimensional state vectors, as shown below: in, For state distribution, To ensure the distribution of truly effective health intervention programs, Let be the generator loss function. Let the discriminator loss function be... For mathematical expectation operators, and The discriminator scores the suitability of the health intervention recommendations output by the generator to the user. It is Gaussian noise. For diverse health intervention recommendations, This represents the range of Gaussian noise distribution. S63. Integrate the health intervention suggestions that have been iteratively optimized in reinforcement learning with the health intervention suggestions that have been assessed by fit scores, and then screen them through medical-grade evaluation indicators to finally output personalized health intervention suggestions.

8. The functional medicine proactive health management method for large-scale model fine-tuning engineering according to claim 1, characterized in that, The user's health behavior data includes physiological indicators, lifestyle data, and psychological state data, and the health behavior data is synchronously integrated with multi-omics data such as genomics, metabolomics, proteomics, and clinical data.

9. A functional medicine proactive health management system for large-scale model fine-tuning engineering, employing any one of the functional medicine proactive health management methods for large-scale model fine-tuning engineering as described in any one of claims 1-8, characterized in that, The system includes: Edge layer: includes IoT wearable devices, portable testing instruments and edge computing nodes. The IoT wearable devices and portable testing instruments are connected to the edge computing nodes via Bluetooth and WiFi. Cloud layer: includes cloud server cluster and distributed database. The cloud server cluster is used to deploy large-scale medical models. The edge computing nodes are connected to the cloud server cluster via the network. The distributed database is used to store users' health behavior data and health assessment and management results. Application layer: includes smart terminals, which are used to connect to the cloud layer via a network to receive the user's health assessment and management results.

Images