Multi-modal health prescription generation and simulation evaluation method based on individualized digital twinborn body

Abstract

The invention discloses a multi-modal health prescription generation and simulation evaluation method based on an individualized digital twinborn body. The method comprises the following steps: firstly, constructing a multi-modal human body digital twinborn body model, representing individual physiological, pathological and behavioral states by hidden state variables, and training state transfer function and observation function parameters by using historical group data to form group-level twinborn bodies; and then performing parameter fine tuning in combination with multi-source health data of a new patient, and establishing an individualized digital twinborn body. The system generates multi-dimensional candidate prescriptions through a large model in combination with a medical knowledge base, carries out forward simulation prediction by using digital twin bodies, evaluates the curative effect, risk, cost and compliance of the prescriptions, and selects a final prescription by taking weighted utility or robust optimization as a criterion. According to the method, accurate and explainable prescription recommendation can be realized, closed-loop evaluation and safety monitoring are supported, and the method is suitable for the fields of individualized rehabilitation training, chronic disease management and clinical auxiliary decision making.

Description

A Multimodal Health Prescription Generation and Simulation Evaluation Method Based on Personalized Digital Twins Technical Field

[0001] This invention belongs to the field of artificial intelligence and digital health technology, and in particular relates to an intelligent prescription generation method based on multimodal health data fusion and digital twin modeling. It is applicable to the intelligent generation and simulation optimization of various types of health intervention programs, such as drug prescriptions, rehabilitation prescriptions, exercise prescriptions, nursing prescriptions, diet and sleep prescriptions. Background Technology

[0002] With the development of wearable devices, medical imaging, genomics, and the Internet of Things (IoT) technologies, the monitoring and management of human health has entered the era of multimodal data. In modern medical and rehabilitation processes, doctors typically prescribe medications, rehabilitation programs, or lifestyle modifications based on clinical test results, medical history, and experience. However, current technologies still face the following major challenges:

[0003] Prescription formulation relies on experience and lacks individualization and dynamism: Current medical and rehabilitation prescriptions are mostly based on doctors' experience, making it difficult to comprehensively consider individual physiological characteristics, behavioral habits, and real-time health status. Different patients respond significantly differently to the same intervention, and traditional prescriptions cannot achieve precise matching.

[0004] Low data utilization and fragmented multimodal information: Clinical data (medical records, laboratory tests, imaging) and wearable device data (exercise, heart rate, electromyography, blood oxygen, sleep, etc.) are usually scattered in different systems, lacking unified integration and modeling, resulting in incomplete health status assessment and lack of objective data support for prescription formulation.

[0005] Lack of prescription simulation and prediction mechanisms: Existing prescription generation methods cannot predict the intervention effect and potential risks before implementation, and it is also difficult to verify the rationality and safety of prescriptions through model simulation.

[0006] Lack of dynamic feedback and closed-loop optimization mechanism: Once the prescription is generated, it is difficult to automatically correct it according to the patient's real-time response and recovery progress, resulting in delayed or excessive intervention effects, affecting efficacy and safety.

[0007] In recent years, digital twin technology has gradually emerged in the industrial and medical fields. It achieves state synchronization, prediction, and control by constructing digital mapping models of physical objects. If digital twin technology can be combined with artificial intelligence algorithms to build personalized health digital twins, intelligent and dynamic management of prescription generation, simulation, and adjustment can be realized. Therefore, there is an urgent need for an intelligent method that integrates multimodal health data, utilizing digital twins to achieve prescription generation, effect prediction, and dynamic optimization, thereby improving the accuracy, safety, and sustainability of medical and rehabilitation interventions. Summary of the Invention

[0008] The purpose of this invention is to address the problems mentioned in the background art by proposing a multimodal health prescription generation and simulation evaluation method based on individualized digital twins. By constructing a dynamically updatable individual digital twin and integrating clinical data, wearable device data, and behavioral data, it achieves accurate modeling and dynamic mapping of multidimensional health status, thereby enabling intelligent generation, simulation evaluation, and closed-loop optimization of prescriptions for various types such as medication, rehabilitation, nursing, exercise, diet, and sleep.

[0009] To achieve the objectives of this invention, a method for generating and simulating health prescriptions based on individualized digital twins is disclosed, comprising the following steps:

[0010] Step 1: Representation of the digital twin; Construct a digital twin model to describe the evolution of an individual's health status, mapping the individual's physiological, pathological, functional, and behavioral states in virtual space;

[0011] Step 2: Establishing a baseline digital twin; Based on a large amount of multimodal time series data of historical patients, a deep learning model is used to train the parameters of the state transition function and the observation function at the population level to obtain a basic model that can reflect the pattern of population health changes;

[0012] Step 3: Establishment of a personalized digital twin; For each new individual, collect multi-source health data, including demographic characteristics, wearable signals, clinical test data, imaging information, and lifestyle records; Based on the baseline model, fine-tune the model parameters using individual data so that the digital twin can accurately reflect the individual's health status evolution characteristics and intervention response characteristics, thereby forming a personalized digital twin model;

[0013] Step 4: Large-scale model-driven candidate prescription generation; Based on individualized digital twins, extract multimodal observation data and historical intervention information of patients and embed them into vectors; Combine hybrid retrieval of knowledge base with RAG mechanism to obtain relevant evidence for constructing large-scale model input; Output a multi-dimensional, individualized candidate prescription set that meets safety constraints; Through the medical knowledge base retrieval module, obtain relevant medical evidence from clinical guidelines, pharmacopoeias, trial data, case databases and institutional standard operating procedures; Utilize a large language model fine-tuned in the medical field, combined with individual characteristics, retrieval evidence and safety constraints, to generate multiple candidate prescription schemes covering multiple dimensions of drug, exercise, diet and lifestyle interventions;

[0014] Step 5: Prescription Simulation and Multidimensional Evaluation; In a personalized digital twin environment, forward simulation is performed on each candidate prescription to predict future health trends after intervention; Based on the simulation results, each candidate prescription is evaluated in multiple dimensions, including efficacy indicators, risk indicators, prescription implementation costs, and patient compliance; A comprehensive weighted utility function is established to rank and screen candidate prescriptions and calculate the comprehensive utility score;

[0015] Step 6: Final prescription determination; Under the premise of safety constraints, all candidate solutions are hard-filtered to eliminate prescriptions with contraindications, exceeding limits or regulatory conflicts; Based on this, the optimal prescription is selected according to the comprehensive utility score, and an executable prescription containing evidence chain, simulation results, confidence assessment and compliance prediction is output for clinical experts to review and implement.

[0016] Furthermore, in step 1, the digital twin model includes three types of variables: latent state, multidimensional observation data, and external intervention parameters. The latent state represents the internal health state that cannot be directly observed. The multidimensional observation data comes from multimodal health data, including wearable device data, clinical test data, imaging features, and behavioral data. The external intervention parameters describe the external input of prescriptions or rehabilitation training. The digital twin describes the dynamic evolution of an individual's health state over time through a state transition function and realizes the correspondence between the latent state and multimodal observation signals through an observation function, forming a virtual-real mapping mechanism.

[0017] Furthermore, a digital twin is represented as a set O:

[0018]

[0019] Set O contains the following elements:

[0020] x t The hidden state is represented by a vector, which represents the internal physiological, pathological, and functional state of an individual at time t. This state cannot be directly and fully observed.

[0021] y t : represents the multimodal observation data at time t, representing the observable data obtained from wearable devices, imaging systems, and laboratory testing methods, totaling M modalities; y t (m) : Represents the observation data of the m-th mode at time t, represented as a vector;

[0022] u t : Represents the external intervention at time t, expressed as a vector, including prescription or rehabilitation intervention parameters;

[0023] f θThe state transition function is a parameterized function model with the function form f. θ (*); State transition function f θ This represents the evolution process from time t to state t+1, and is influenced by intervention u. t and the effects of random disturbances;

[0024] g θ : Represents the multimodal observation function, which is a parameterized function model, with a total of M functions; g θ (m) The observation function for the m-th mode, in the form g θ (m) (*);g θ (m) Mapping unobservable latent states to measurable data enables the alignment of "digital twins ↔ real-world data";

[0025] θ: θ = {θ f ,θ g The trainable parameters, θ, originate from a group-level learning model and are optimized and adjusted based on individual characteristics. f The parameters representing the state transition function are denoted as a vector; θ g The parameters representing the observation function are represented as a vector;

[0026] The dynamic evolution of a human digital twin follows the following stochastic state-space model:

[0027] (1)

[0028] (2)

[0029] x t+1 : The hidden state space at time t+1; w t Process noise describes physiological fluctuations, measurement errors, and environmental disturbances; v t (m) : represents the observation noise of the m-th mode; Q, R m These are the covariance matrices of the process noise and the observation noise, respectively.

[0030] Equation (1) is the state transition equation, where f θ (*) is the state transition function, used to describe the evolution of the hidden state from time t to t+1; the state transition function is a parameterized function model, and the parameter is a vector θ. f Its input is the hidden state x at the current time. t and external intervention t The output is the state prediction value; process noise w is then superimposed on this value. t To obtain the hidden state x at the next time step. t+1 ;

[0031] Equation (2) is the observation equation, where g θ (m) (*) is the observation function for the m-th mode, used to map the hidden states to the observation data of the corresponding mode; the observation function is a parameterized function model with parameters θ. g Its input is the hidden state x at the current time. t The output is the observed predicted value; observation noise v is then superimposed on this value. t (m) The actual observed value y is obtained. t (m) .

[0032] Furthermore, in step 2, an inference network is introduced during the training process to estimate the hidden state. Combined with observation reconstruction, state consistency and parameter regularization constraints, the group-level model parameters are optimized for the initialization of the digital twin.

[0033] The baseline digital twin is established to complete the function f θ and g θ The establishment of the median parameter θ; the use of a large amount of historical multimodal time series data D from patients during the model training phase. pop :

[0034]

[0035] Where i represents the sample number; the sample size is N. patients ;T i y represents the time series quantity of the observations for sample i; m represents the number of observations; the maximum value is M; t (i,m) u represents the m-th observation data of the i-th sample at time t; t (i) This represents the external intervention for the i-th sample at time t; Representation: p represents the time from a to b at time t. t A set;

[0036] The hidden state x of the i-th sample at time t t (i) Since it cannot be directly observed, an inference network (with parameters of...) is introduced. This is used to obtain the distribution of the hidden states. The probability distribution of the hidden states is represented as:

[0037]

[0038] The inference network is as follows:

[0039] Inference network input: current and historical multimodal observation data y 1:t (i,1:m) and historical external intervention u 1:t-1 (i) ;

[0040] Inferring the internal structure of the network (parameters are) ): (1) Multimodal encoder (parameters are (1) The historical observation data of each modality of the sample and the external intervention at each time step are used to encode features using the MLP / 1D-CNN method; (2) Time series modeler (parameters are ): Use RNN / GRU / LSTM / Transformer to perform sequence modeling on the encoded features to obtain the final features; (3) The final features are passed through a fully connected layer (parameter is Output hidden state distribution parameters: mean μ t and variance σ t 2 ;

[0041] The final output of the inference network: the estimated hidden state. , is represented as:

[0042]

[0043] Where ε is random noise; I is the identity matrix; and ⊗ represents element-wise multiplication.

[0044] Furthermore, a group training loss function is established, and the function f is obtained through training. θ and g θ Intense parameters θ and inference network parameters Specifically, it includes the following steps:

[0045] Observation reconstruction loss L obs :

[0046]

[0047] State transition consistency loss L trans :

[0048]

[0049] in, Let represent the hidden state estimated at time t+1 for the i-th sample;

[0050] Group parameter regularization L reg :

[0051]

[0052] Where λ is the regularization coefficient;

[0053] The total loss is expressed as L pop :

[0054]

[0055] Where α is the weighting parameter;

[0056] The parameters are obtained using gradient descent population optimization, and the output state transition function f is output. θ and observation function g θ The group parameter θ pop And inferring the population parameters of the network. :

[0057]

[0058] in, Indicates that a set has been found Make the function Obtain the minimum value.

[0059] Furthermore, in step 3, for new patient i*, multimodal data is collected, including but not limited to: demographic information; wearable device data (IMU, sEMG, ECG, blood oxygen, sleep, etc.); clinical data (laboratory indicators, imaging features, medical record information); and behavioral and lifestyle data (exercise, diet, rest, and psychological state).

[0060] Multimodal data is represented as follows:

[0061]

[0062] in, T* represents the m-th observation data of sample i* at time t; T* represents the observation time series of sample i*.

[0063] The historical prescription data for new patients is represented as follows:

[0064]

[0065] in, This represents the prescription for sample i* at time t;

[0066] The process of fine-tuning the population parameters of the baseline digital twin using data from new patients can be represented as follows:

[0067]

[0068]

[0069] Where θ i* , The parameters representing the individualization of sample i*; A function representing the individualization of sample i*;

[0070] Fine-tuning the parameters of the population twins on individual data makes the digital twins more closely resemble the dynamic characteristics of new patients. During fine-tuning, a loss function is established:

[0071] Observation reconstruction loss :

[0072]

[0073] in, The hidden state of the new sample i* at time t is estimated and obtained using an inference network;

[0074] State transition consistency loss :

[0075]

[0076] in, The hidden state of the new sample i* at time t+1 is represented by the inference network.

[0077] Group parameter regularization :

[0078]

[0079] Where λ is the regularization coefficient;

[0080] The total loss is expressed as :

[0081]

[0082] Where α is the weighting parameter;

[0083] The parameters are obtained using gradient descent population optimization, and the population parameters of the individualized state transition function and observation function, as well as the population parameters of the individualized inference network, are output:

[0084]

[0085] in, Represents a group Make the function Obtain the minimum value;

[0086] Ultimately, the individualized twins of the current time T are represented as a set. :

[0087]

[0088] in, Let represent the set of all estimated hidden states of sample i* between time 1 and T, and the set of multimodal observation data, respectively. Let i* represent the set of all external interventions for sample i* from time 1 to T-. The function and parameters representing the individualization of sample i*.

[0089] This individualized twin can be used to simulate prescriptions at time T in the future at time H.

[0090] Furthermore, step 4 specifically includes the following:

[0091] Step 4-1: Obtain existing observations and past prescriptions from the individualized digital twin:

[0092] ;

[0093] in, This represents the observation data of sample i* from time 1 to T in the m-th mode; for example: y (i*,1) Basic information (demographics, medical history, physiological indicators); y (i*,2) Functional / symptom assessment results; y (i*,3) Historical prescription execution and response data; data for each modality. After being embedded into a vector by a pre-defined encoder (embedding network), it is represented as z. (m) ;

[0094]

[0095] This represents the set of encoding results for all observation data of the m-th mode of sample i* from time 1 to T; This represents the encoding function corresponding to the m-th modal observation data; these embedded vectors are used for subsequent knowledge retrieval and large model input;

[0096] Step 4-2, Prescription Knowledge Base and Embedding; Construct a prescription knowledge base, including but not limited to clinical guidelines, pharmacopoeias / drug instructions, clinical trial results, validated medical records / case libraries, institutional SOPs, regulations and contraindication lists; Segment each document at the paragraph level and create a vector index, while retaining the original metadata (source, publication date, confidence label, applicable population); Use a medical domain-fine-tuned sentence vector model (e.g., Bio / Clinical embeddings) to ensure medical semantic fit;

[0097] A hybrid retrieval module for the knowledge base is constructed, which utilizes Sparse retrieval (based on inverted index algorithms such as BM25) and Dense retrieval (based on embedding vector similarity). This hybrid retrieval method retrieves and returns relevant fields, with each segment containing information such as text content, source, and evidence confidence.

[0098] Step 4-3: Enhanced RAG retrieval; Construct the question, using the patient's current information (y) to obtain a patient summary and main query prescription task, such as "For a 50-year-old stroke patient in the rehabilitation phase, with an FMA score of ×× and previous prescriptions of ××, what is the recommended prescription intervention plan?"; Simultaneously, the system generates several sub-questions to cover multi-dimensional intervention areas (e.g., medication, exercise, diet, sleep, etc.), ensuring that the generated results are targeted and comprehensive across different dimensions; The main query and sub-questions are input into the hybrid retrieval module to obtain the candidate evidence set D={d1,d2,…,d k}; Subsequently, the patient summary, retrieval evidence segments, and safety constraints (such as drug contraindications, intensity limits, and other hard constraints) are concatenated to form the input prompt of the large language model LLM;

[0099] Step 4-4: Large Model Invocation; Invoke the instruction-tuned / few-shot / fine-tuned medical large model; Input the prompt constructed in Step 4-3, including patient summary, multimodal observation embeddings, knowledge base evidence segments retrieved via RAG, as well as safety constraints and contraindication rules; The large model generates a list of multiple candidate prescriptions U={u1,...u,...,u n Each prescription contains prescriptions with multiple dimensions.

[0100] Furthermore, in step 5, after the candidate prescriptions are generated, the system will perform forward simulation and multi-dimensional simulation evaluation on each prescription scheme in the digital twin environment; through the state transition function f of the digital twin... θ The rolling prediction obtains the state distribution of the next H steps under a given prescription intervention u. The do function represents execution, and do(u) represents executing prescription u.

[0101]

[0102] For prescription u, from the current hidden state x t Perform L Monte Carlo forward simulations to obtain the estimated set of sample hidden state trajectories:

[0103]

[0104] That is, the set of estimated hidden states from future time t+1 to t+H, totaling L (j from 1 to L);

[0105] Using the observation function, each trajectory Generate the corresponding observation trajectory:

[0106]

[0107] Evaluation indices are calculated using the estimated hidden states of the samples.

[0108] Furthermore, the evaluation indicators include efficacy, risk, cost, and compliance;

[0109] Therapeutic efficacy is a quantitative indicator that measures the "positive clinical effect" brought about by a prescription, often defined as the expected gain of one or more clinical objective functions within H steps in the future; R eff (*) indicates immediate single-step utility (which can be determined by clinical parameters such as FMA score increment, gait symmetry, and ADL improvement), used to evaluate the immediate effect of a prescription under current conditions and intervention. Its specific form is constructed according to different clinical application scenarios to reflect the corresponding treatment goals, thus indicating the expected efficacy of the prescription. Defined as:

[0110]

[0111] Where, x t+h Let be the hidden state at time t+h; E is the expectation function, in the form E(*);

[0112] Following Monte Carlo sampling, the expected therapeutic effect of this prescription can be estimated as follows: :

[0113]

[0114] in, Let be the estimated value of the hidden state at time t+h on the j-th trajectory;

[0115] Risk represents the probability or expected loss that a prescription will cause an adverse outcome or exceed a safety threshold; risk is characterized using probabilistic risk, and risk is estimated as follows: :

[0116]

[0117] Among them, z safe It is a safety threshold;

[0118] Prescription cost is the resource consumption required to administer a prescription, including time cost. time Human resource costs staff And the monetary costs of materials / medicines and equipment c money To ensure that all cost items have a unified dimension, it is preferable to convert different resource consumptions into a unified monetary cost; formally, the total cost function of prescription u is defined as follows: , is represented as:

[0119]

[0120] c time The required training time for the patient; c staff c. Number of therapist hours required; money Equipment or medication costs; k1 is the conversion factor for the monetary cost per unit of training time; k2 is the conversion factor for the monetary cost per unit of therapist's working hours.

[0121] Prescription adherence: a predictor of the probability or expected completion rate of a patient actually following a prescription. u Estimated by the patient's historical behavior model, the estimated value is expressed as: ;

[0122] To provide a unified evaluation of different candidate prescriptions, a comprehensive utility function is constructed by integrating multiple dimensions such as efficacy, risk, cost, and adherence. This comprehensive utility function unifies the evaluation indicators from different dimensions into a single, comparable utility value, thereby enabling the ranking and screening of candidate prescriptions. The comprehensive utility function is expressed as follows:

[0123]

[0124] Among them, the overall utility score E wu w E ,w U ,w A ,w C These are the weighting coefficients.

[0125] Furthermore, in step 6, after completing the simulation evaluation and multi-dimensional quantitative scoring of candidate prescriptions, the system enters the "final prescription selection" stage. This stage takes safety as a hard constraint, weighted utility / robustness as the primary criterion, and retains the right of manual review by clinical experts. In specific implementation, a hard filter is first applied to all candidate prescriptions to eliminate any prescriptions that violate medical or regulatory constraints (such as prescriptions that exceed the dosage limit, have clear contraindications, or conflict with the patient's previous medical history).

[0126] After any proposals that violate medical or regulatory constraints are put forward, the final choice will be made:

[0127]

[0128] in, This means finding a u in set U such that E wu The maximum value is obtained; the final selected prescription u* includes a complete chain of evidence (simulation trajectory, supporting literature paragraphs, confidence intervals, compliance predictions, and cost estimates) and an executable prescription is generated.

[0129] Compared with the prior art, the significant progress of this invention is as follows: 1) Digital twin modeling method based on multimodal dynamic state space: A multimodal state space modeling method combining wearable sensing, imaging and clinical indicators is proposed. The method realizes the dynamic tracking and prediction of hidden states through state transition function and observation function, and introduces inference network for latent variable estimation. The model enables fine-tuning of group model parameters at the individual level, thereby constructing individualized digital twins and providing a reliable dynamic basis for prescription simulation; 2) A large-scale prescription generation mechanism integrating medical knowledge and retrieval enhancement: A structured medical knowledge base is established, combining sparse retrieval and semantic retrieval to form a hybrid knowledge retrieval module; through the RAG framework, patient digital twin information and knowledge base evidence are jointly input into a medically fine-tuned large language model, thereby generating candidate prescriptions covering multi-dimensional interventions such as medication, exercise, diet, and sleep, achieving deep integration of medical knowledge and individual status; 3) A multi-dimensional prescription evaluation system based on forward simulation of digital twins: A forward rolling prediction mechanism based on digital twins is proposed, which obtains the future H-step state distribution through Monte Carlo sampling, performs multi-dimensional simulation evaluation of candidate prescriptions in terms of efficacy, risk, cost, and compliance, and constructs a weighted utility function for final plan decision-making. This mechanism realizes quantitative prediction and safety verification of prescription-level intervention effects.

[0130] To more clearly illustrate the functional characteristics and structural parameters of the present invention, further explanation is provided below in conjunction with the accompanying drawings and specific embodiments. Attached Figure Description

[0131] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0132] Figure 1 is a schematic diagram of the process for generating and simulating the evaluation of multimodal health prescriptions based on individualized digital twins. Detailed Implementation

[0133] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0134] As shown in Figure 1, a method for generating and simulating health prescriptions based on individualized digital twins includes the following steps:

[0135] Step 1: Representation of the Digital Twin: Construct a digital twin model to describe the evolution of an individual's health status, mapping the individual's physiological, pathological, functional, and behavioral states in virtual space. The model includes three types of variables: latent states, observed signals, and external intervention parameters. Latent states represent internal health states that cannot be directly observed. Observed signals are derived from multimodal health data, including wearable device data, clinical test data, imaging features, and behavioral data. External intervention parameters describe the external inputs of prescriptions or rehabilitation training. The digital twin describes the dynamic evolution of an individual's health status over time through a state transition function and establishes a correspondence between latent states and multimodal observed signals through an observation function, forming a virtual-real mapping mechanism.

[0136] Step 2: Baseline Digital Twin Establishment: Based on a large amount of multimodal time series data of historical patients, a deep learning model is used to train the parameters of the state transition function and observation function at the population level to obtain a basic model that can reflect the health change pattern of the population. During the training process, an inference network is introduced to estimate the hidden state. Combined with constraints such as observation reconstruction, state consistency and parameter regularization, the population-level model parameters are optimized for the initialization of the digital twin.

[0137] Step 3: Establishment of a Personalized Digital Twin: For each new individual, collect multi-source health data, including demographic characteristics, wearable signals, clinical test data, imaging information, and lifestyle records. Based on the baseline model, fine-tune the model parameters using individual data so that the digital twin can accurately reflect the individual's health status evolution characteristics and intervention response characteristics, thereby forming a personalized digital twin model.

[0138] Step 4: Large-scale model-driven candidate prescription generation: Based on individualized digital twins, extract multimodal observation data and historical intervention information of patients and embed them into vectors; through the medical knowledge base retrieval module, obtain relevant medical evidence from clinical guidelines, pharmacopoeias, trial data, case databases and institutional standard operating procedures; using a large language model finely tuned in the medical field, combined with individual characteristics, retrieval evidence and safety constraints, generate multiple candidate prescription schemes covering multiple dimensions such as drug, exercise, diet and lifestyle interventions.

[0139] Step 5, Prescription Simulation and Multidimensional Evaluation: In a personalized digital twin environment, forward simulation is performed on each candidate prescription to predict future health trends after intervention; based on the simulation results, each candidate prescription is evaluated in multiple dimensions, including efficacy indicators, risk indicators, prescription implementation costs, and patient compliance; a comprehensive weighted utility function is established to rank and screen candidate prescriptions and calculate the overall utility score.

[0140] Step 6: Final prescription determination: Under the premise of safety constraints, all candidate solutions are hard-filtered to eliminate prescriptions with contraindications, exceeding limits or regulatory conflicts; on this basis, the optimal prescription is selected according to the comprehensive utility score, and an executable prescription containing evidence chain, simulation results, confidence assessment and compliance prediction is output for clinical experts to review and implement.

[0141] Furthermore, in step 1, the digital twin model includes three types of variables: latent state, multidimensional observation data, and external intervention parameters. The latent state represents the internal health state that cannot be directly observed. The multidimensional observation data comes from multimodal health data, including wearable device data, clinical test data, imaging features, and behavioral data. The external intervention parameters describe the external input of prescriptions or rehabilitation training. The digital twin describes the dynamic evolution of an individual's health state over time through a state transition function and realizes the correspondence between the latent state and multimodal observation signals through an observation function, forming a virtual-real mapping mechanism.

[0142] Specifically, in one embodiment, the digital twin model is represented as O:

[0143]

[0144] Where: x t : Latent state, which represents the internal physiological, pathological and functional state of an individual at time t, and this state cannot be directly and fully observed;

[0145] y t : represents the multimodal observation data at time t, representing the observable data obtained from wearable devices, imaging systems, and laboratory testing methods, totaling M modalities; y t(m) : Represents the observation data of the m-th mode at time t;

[0146] u t : Represents the external intervention at time t, including prescription or rehabilitation intervention parameters;

[0147] f θ : State transition function, function form f θ (*); State transition function f θ This represents the evolution process from time t to state t+1, and is influenced by intervention u. t and the effects of random disturbances;

[0148] g θ : Represents a multimodal observation function, with a total of M functions; g θ (m) The observation function for the m-th mode, in the form g θ (m) (*);g θ (m) Mapping unobservable latent states to measurable data enables the alignment of "digital twins ↔ real-world data";

[0149] θ: θ = {θ f ,θ g The parameters of the function are derived from the group-level learning model and optimized based on individual characteristics, θ. f θ represents the parameter of the state transition function. g These represent the parameters of the observed functions;

[0150] The dynamic evolution of a human digital twin follows the following stochastic state-space model:

[0151] (1)

[0152] (2)

[0153] x t+1 : The hidden state space at time t+1; w t Process noise describes physiological fluctuations, measurement errors, and environmental disturbances; v t (m) : represents the observation noise of the m-th mode; Q, R m Equation (1) is the covariance matrix of process noise and observation noise, respectively; Equation (2) is the state transition equation, which describes the dynamic changes of an individual's health status in the time dimension; Equation (3) is the observation equation, which represents the nonlinear mapping relationship between the latent state and multimodal observation data.

[0154] Specifically, in one embodiment, in step 2, an inference network is introduced during the training process to estimate the hidden state, and the group-level model parameters are optimized by combining observation reconstruction, state consistency and parameter regularization constraints for the initialization of the digital twin.

[0155] The baseline digital twin is established to complete the function f θ and g θ The establishment of the median parameter θ; the use of a large amount of historical multimodal time series data D from patients during the model training phase. pop :

[0156]

[0157] Where i represents the sample number; the sample size is N. patients ;T i y represents the time series quantity of the observations for sample i; m represents the number of observations; the maximum value is M; t (i,m) u represents the m-th observation data of the i-th sample at time t; t (i) This represents the external intervention for the i-th sample at time t; Representation: p represents the time from a to b at time t. t A set;

[0158] The hidden state x of the i-th sample at time t t (i) Since it cannot be directly observed, an inference network (with parameters of...) is introduced. This is used to obtain the distribution of the hidden states. The probability distribution of the hidden states is represented as:

[0159]

[0160] The inference network is as follows:

[0161] Inference network input: current and historical multimodal observation data y 1:t (i,1:m) and historical external intervention u 1:t-1 (i) ;

[0162] Inferring the internal structure of the network (parameters are) ): (1) Multimodal encoder (parameters are (1) The historical observation data of each modality of the sample and the external intervention at each time step are used to encode features using the MLP / 1D-CNN method; (2) Time series modeler (parameters are ): Use RNN / GRU / LSTM / Transformer to perform sequence modeling on the encoded features to obtain the final features; (3) The final features are passed through a fully connected layer (parameter is Output hidden state distribution parameters: mean μ t and variance σ t 2 ;

[0163] The final output of the inference network: the estimated hidden state. , is represented as:

[0164]

[0165] Where ε is random noise; I is the identity matrix; and ⊗ represents element-wise multiplication.

[0166] Specifically, in one embodiment, a group training loss function is established while simultaneously training to obtain the function f. θ and g θ Intense parameters θ and inference network parameters Specifically, it includes the following steps:

[0167] Observation reconstruction loss L obs :

[0168]

[0169] State transition consistency loss L trans :

[0170]

[0171] in, Let represent the hidden state estimated at time t+1 for the i-th sample;

[0172] Group parameter regularization L reg :

[0173]

[0174] Where λ is the regularization coefficient;

[0175] The total loss is expressed as L pop :

[0176]

[0177] Where α is the weighting parameter;

[0178] The parameters are obtained using gradient descent population optimization, and the output state transition function f is output. θ and observation function g θ The group parameter θ pop And inferring the population parameters of the network. pop :

[0179]

[0180] in, Indicates that a set has been found Make the function Obtain the minimum value.

[0181] Specifically, in one embodiment, in step 3, for the new patient i*, multimodal data is collected, including but not limited to: demographic information; wearable device data (IMU, sEMG, ECG, blood oxygen, sleep, etc.); clinical data (laboratory indicators, imaging features, medical record information); and behavioral and lifestyle data (exercise, diet, rest, and psychological state).

[0182] Multimodal data is represented as follows:

[0183]

[0184] in, T* represents the m-th observation data of sample i* at time t; T* represents the observation time series of sample i*.

[0185] The historical prescription data for new patients is represented as follows:

[0186]

[0187] in, This represents the prescription for sample i* at time t;

[0188] The process of fine-tuning the population parameters of the baseline digital twin using data from new patients can be represented as follows:

[0189]

[0190]

[0191] Where θ i* , The parameters representing the individualization of sample i*; A function representing the individualization of sample i*;

[0192] Fine-tuning the parameters of the population twins on individual data makes the digital twins more closely resemble the dynamic characteristics of new patients. During fine-tuning, a loss function is established:

[0193] Observation reconstruction loss :

[0194]

[0195] in, The hidden state of the new sample i* at time t is estimated and obtained using an inference network;

[0196] State transition consistency loss :

[0197]

[0198] in, The hidden state of the new sample i* at time t+1 is represented by the inference network.

[0199] Group parameter regularization :

[0200]

[0201] Where λ is the regularization coefficient;

[0202] The total loss is expressed as :

[0203]

[0204] Where α is the weighting parameter;

[0205] The parameters are obtained using gradient descent population optimization, and the population parameters of the individualized state transition function and observation function, as well as the population parameters of the individualized inference network, are output:

[0206]

[0207] in, Represents a group Make the function Obtain the minimum value;

[0208] Ultimately, the individualized twins of the current time T are represented as a set. :

[0209]

[0210] in, Let represent the set of all estimated hidden states of sample i* between time 1 and T, and the set of multimodal observation data, respectively. Let i* represent the set of all external interventions for sample i* from time 1 to T-. The function and parameters representing the individualization of sample i*.

[0211] This individualized twin can be used to simulate prescriptions at time T in the future at time H.

[0212] Specifically, in one embodiment, step 4 includes the following:

[0213] Step 4-1: Obtain existing observations and past prescriptions from the individualized digital twin:

[0214] ;

[0215] in, This represents the observation data of sample i* from time 1 to T in the m-th mode; for example: y (i*,1) Basic information (demographics, medical history, physiological indicators); y (i*,2) Functional / symptom assessment results; y (i*,3) Historical prescription execution and response data; data for each modality. After being embedded into a vector by a pre-defined encoder (embedding network), it is represented as z. (m) ;

[0216]

[0217] This represents the set of encoding results for all observation data of the m-th mode of sample i* from time 1 to T; This represents the encoding function corresponding to the m-th modal observation data; these embedded vectors are used for subsequent knowledge retrieval and large model input;

[0218] Step 4-2, Prescription Knowledge Base and Embedding; Construct a prescription knowledge base, including but not limited to clinical guidelines, pharmacopoeias / drug instructions, clinical trial results, validated medical records / case libraries, institutional SOPs, regulations and contraindication lists; Segment each document at the paragraph level and create a vector index, while retaining the original metadata (source, publication date, confidence label, applicable population); Use a medical domain-fine-tuned sentence vector model (e.g., Bio / Clinical embeddings) to ensure medical semantic fit;

[0219] A hybrid retrieval module for the knowledge base is constructed, which utilizes Sparse retrieval (based on inverted index algorithms such as BM25) and Dense retrieval (based on embedding vector similarity). This hybrid retrieval method retrieves and returns relevant fields, with each segment containing information such as text content, source, and evidence confidence.

[0220] Step 4-3: Enhanced RAG retrieval; Construct the question, using the patient's current information (y) to obtain a patient summary and main query prescription task, such as "For a 50-year-old stroke patient in the rehabilitation phase, with an FMA score of ×× and previous prescriptions of ××, what is the recommended prescription intervention plan?"; Simultaneously, the system generates several sub-questions to cover multi-dimensional intervention areas (e.g., medication, exercise, diet, sleep, etc.), ensuring that the generated results are targeted and comprehensive across different dimensions; The main query and sub-questions are input into the hybrid retrieval module to obtain the candidate evidence set D={d1,d2,…,d k}; Subsequently, the patient summary, retrieval evidence segments, and safety constraints (such as drug contraindications, intensity limits, and other hard constraints) are concatenated to form the input prompt of the large language model LLM;

[0221] Step 4-4: Large Model Invocation; Invoke the instruction-tuned / few-shot / fine-tuned medical large model; Input the prompt constructed in Step 4-3, including patient summary, multimodal observation embeddings, knowledge base evidence segments retrieved via RAG, as well as safety constraints and contraindication rules; The large model generates a list of multiple candidate prescriptions U={u1,...u,...,u n Each prescription contains prescriptions with multiple dimensions.

[0222] Specifically, in one embodiment, in step 5, after the candidate prescription is generated, the system will perform forward simulation and multi-dimensional simulation evaluation on each prescription scheme in the digital twin environment; through the digital twin state transition function f θ The rolling prediction obtains the state distribution of the next H steps under a given prescription intervention u. The do function represents execution, and do(u) represents executing prescription u.

[0223]

[0224] For prescription u, from the current hidden state x t Perform L Monte Carlo forward simulations to obtain the estimated set of sample hidden state trajectories:

[0225]

[0226] That is, the set of estimated hidden states from future time t+1 to t+H, totaling L (j from 1 to L);

[0227] Using the observation function, each trajectory Generate the corresponding observation trajectory:

[0228]

[0229] Evaluation indices are calculated using the estimated hidden states of the samples.

[0230] Specifically, in one embodiment, the evaluation indicators include efficacy, risk, cost, and compliance;

[0231] Therapeutic efficacy is a quantitative indicator that measures the "positive clinical effect" brought about by a prescription, often defined as the expected gain of one or more clinical objective functions within H steps in the future; R eff (*) indicates the immediate effect of a single step (which can be determined by clinical parameters such as FMA score increment, gait symmetry, and ADL improvement), and the expected therapeutic effect of the prescription. Defined as:

[0232]

[0233] Where, x t+h Let be the hidden state at time t+h; E is the expectation function, in the form E(*);

[0234] Following Monte Carlo sampling, the expected therapeutic effect of this prescription can be estimated as follows: :

[0235]

[0236] in, Let be the estimated value of the hidden state at time t+h on the j-th trajectory;

[0237] Risk represents the probability or expected loss that a prescription will cause an adverse outcome or exceed a safety threshold; risk is characterized using probabilistic risk, and risk is estimated as follows: :

[0238]

[0239] Among them, z safe It is a safety threshold;

[0240] Prescription cost is the resource consumption required to administer a prescription, including time cost. time Human resource costs staff And the monetary costs of materials / medicines and equipment cmoney To ensure that all cost items have a unified dimension, it is preferable to convert different resource consumptions into a unified monetary cost; formally, the total cost function of prescription u is defined as follows: , is represented as:

[0241]

[0242] c time The required training time for the patient; c staff c. Number of therapist hours required; money Equipment or medication costs; k1 is the conversion factor for the monetary cost per unit of training time; k2 is the conversion factor for the monetary cost per unit of therapist's working hours.

[0243] Prescription adherence: a predictor of the probability or expected completion rate of a patient actually following a prescription. u Estimated by the patient's historical behavior model, the estimated value is expressed as: ;

[0244] To provide a unified evaluation of different candidate prescriptions, a comprehensive utility function is constructed by integrating multiple dimensions such as efficacy, risk, cost, and adherence. This comprehensive utility function unifies the evaluation indicators from different dimensions into a single, comparable utility value, thereby enabling the ranking and screening of candidate prescriptions. The comprehensive utility function is expressed as follows:

[0245]

[0246] Among them, the overall utility score E wu w E ,w U ,w A ,w C These are the weighting coefficients.

[0247] Specifically, in one embodiment, in step 6, after completing the simulation evaluation and multi-dimensional quantitative scoring of candidate prescriptions, the system enters the "final prescription selection" stage. This stage takes safety as a hard constraint, weighted utility / robustness as the primary criterion, and retains the right of manual review by clinical experts. In practice, a hard filter is first applied to all candidate prescriptions to eliminate any prescriptions that violate medical or regulatory constraints (e.g., prescriptions that exceed the dosage limit, have clear contraindications, or conflict with the patient's previous medical history).

[0248] After any proposals that violate medical or regulatory constraints are put forward, the final choice will be made:

[0249]

[0250] in, This means finding a u in set U such that E wu The maximum value is obtained; the final selected prescription u* includes a complete chain of evidence (simulation trajectory, supporting literature paragraphs, confidence intervals, compliance predictions, and cost estimates) and an executable prescription is generated.

[0251] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0252] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for generating and simulating health prescriptions based on personalized digital twins, characterized in that, Includes the following steps: Step 1: Digital Twin Representation; Construct a digital twin model to describe the evolution of an individual's health status, mapping the individual's physiological, pathological, functional, and behavioral states in virtual space. Step 2: Baseline Digital Twin Establishment; Based on a large amount of historical multimodal time-series data from patients, use a deep learning model to train the parameters of the state transition function and observation function at the population level, obtaining a basic model that reflects the pattern of population health changes. Step 3: Individualized Digital Twin Establishment; Collect multi-source health data for new individuals; Based on the baseline model, fine-tune the model parameters using individual data to ensure the digital twin accurately reflects the individual's health status evolution characteristics and intervention response characteristics, thus forming an individualized digital twin model. Step 4: Large Model-Driven Candidate Prescription Generation; Based on the individualized digital twin, extract patient multimodal observation data and historical intervention information and embed them into vectors; Combine hybrid retrieval of the knowledge base with the RAG mechanism to obtain relevant evidence for constructing the large model input; Output a multi-dimensional, individualized, and safety-constrained set of candidate prescriptions. Step 5: Prescription Simulation and Multi-Dimensional Evaluation. In a personalized digital twin environment, forward simulation is performed on each candidate prescription to predict future health trends after intervention. The simulation results are then used to evaluate each candidate prescription across multiple dimensions. A comprehensive weighted utility function is established to rank and screen candidate prescriptions, calculating the comprehensive utility score. Step 6: Final prescription determination. Under safety constraints, all candidate prescriptions are rigorously filtered to eliminate those with contraindications, exceeding limits, or regulatory conflicts. Based on this, the optimal prescription is selected according to the comprehensive utility score, and an executable prescription containing the evidence chain, simulation results, confidence assessment, and compliance prediction is output for clinical experts to review and implement.

2. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 1, characterized in that, In step 1, the digital twin model includes three types of variables: latent state, multidimensional observation data, and external intervention parameters. The latent state represents the internal health state that cannot be directly observed. The multidimensional observation data comes from multimodal health data, including wearable device data, clinical test data, imaging features, and behavioral data. The external intervention parameters describe the external input of prescriptions or rehabilitation training. The digital twin describes the dynamic evolution of an individual's health state over time through a state transition function and realizes the correspondence between the latent state and multimodal observation signals through an observation function, forming a virtual-real mapping mechanism.

3. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 2, characterized in that, A digital twin is represented as a set O: Set O contains the following elements: x t The hidden state, represented as a vector, indicates the internal physiological, pathological, and functional state of an individual at time t. This state cannot be directly and comprehensively observed. t : represents the multimodal observation data at time t, representing the observable data obtained from wearable devices, imaging systems, and laboratory testing methods, totaling M modalities; y t (m) : Represents the observation data of the m-th mode at time t, represented as a vector; u t : Represents the external intervention at time t, shown as a vector including prescription or rehabilitation intervention parameters; f θ The state transition function is a parameterized function model with the function form f. θ (*); State transition function f θ This represents the evolution process from time t to state t+1, and is influenced by intervention u. t and the effects of random disturbances; g θ : Represents the multimodal observation function, which is a parameterized function model, with a total of M functions; g θ (m) The observation function for the m-th mode, in the form g θ (m) (*);g θ (m) Mapping unobservable latent states to measurable data enables alignment between "digital twins" and "real-world data". θ: θ = {θ f ,θ g The trainable parameters, θ, originate from a group-level learning model and are optimized and adjusted based on individual characteristics. f The parameters representing the state transition function are denoted as a vector; θ g The parameters representing the observation function are denoted as a vector; the dynamic evolution of the human digital twin follows the following stochastic state-space model: （1） (2)x t+1 : The hidden state space at time t+1; w t Process noise describes physiological fluctuations, measurement errors, and environmental disturbances; v t (m) : represents the observation noise of the m-th mode; Q, R m Let f be the covariance matrix of the process noise and the observation noise, respectively; Equation (1) is the state transition equation, where f θ (*) is the state transition function, used to describe the evolution of the hidden state from time t to t+1; the state transition function is a parameterized function model, and the parameter is a vector θ. f Its input is the hidden state x at the current time. t and external intervention t The output is the state prediction value; process noise w is then superimposed on this value. t To obtain the hidden state x at the next time step. t+1 Equation (2) is the observation equation, where g θ (m) (*) is the observation function for the m-th mode, used to map the hidden states to the observation data of the corresponding mode; the observation function is a parameterized function model with parameters θ. g Its input is the hidden state x at the current time. t The output is the observed predicted value; observation noise v is then superimposed on this value. t (m) The actual observed value y is obtained. t (m) .

4. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 1, characterized in that, In step 2, an inference network is introduced during training to estimate the hidden state. Combined with observation reconstruction, state consistency and parameter regularization constraints, the parameters of the population-level model are optimized for the initialization of the digital twin. The baseline digital twin is established to complete the function f θ and g θ The establishment of the median parameter θ; the use of a large amount of historical multimodal time series data D from patients during the model training phase. pop : Where i represents the sample number; the sample size is N. patients ;T i y represents the time series quantity of the observations for sample i; m represents the number of observations; the maximum value is M; t (i,m) u represents the m-th observation data of the i-th sample at time t; t (i) This represents the external intervention for the i-th sample at time t; Representation: p represents the time from a to b at time t. t The set of; the hidden state x of the i-th sample at time t. t (i) Since it cannot be directly observed, a parameter is introduced as follows. An inference network is used to obtain the distribution of the hidden states; the probability distribution of the hidden states is represented as: The inference network input is current and historical multimodal observation data y 1:t (i,1:m) and historical external intervention u 1:t-1 (i) ; Inferring the internal structural parameters of the network ={ 1, 2, 3} includes: a multimodal encoder, with parameters as follows 1. For each modality of this sample, historical observation data and external interventions at each time step are used for feature encoding using the MLP / 1D-CNN method; a time series modeler with parameters is...

2. Use RNN / GRU / LSTM / Transformer to perform sequence modeling on the encoded features to obtain the final features; the final features are then passed through a fully connected layer with parameters of...

3. Output hidden state distribution parameters: mean μ t and variance σ t 2 The final output of the inference network is the estimated hidden state. , is represented as: Where ε is random noise; I is the identity matrix; and ⊗ represents element-wise multiplication.

5. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 4, characterized in that, Establish a group training loss function and simultaneously train to obtain function f θ and g θ Intense parameters θ and inference network parameters ,specific The steps include: observing the reconstruction loss L obs : State transition consistency loss L trans : in, Let L represent the estimated hidden state of the i-th sample at time t+1; population parameter regularization L reg : Where λ is the regularization coefficient; the total loss is expressed as L pop : Here, α is the weight parameter; the parameters are obtained using gradient descent population optimization, and the output state transition function f is given. θ and observation function g θ The group parameter θ pop And inferring the population parameters of the network. pop ： in, This indicates that a set (θ, ) makes the function Obtain the minimum value.

6. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 1, characterized in that, In step 3, for new patient i*, multimodal data is collected, including but not limited to: demographic information; wearable device data; clinical data, behavioral and lifestyle data; the multimodal data is represented as follows: in, T* represents the m-th observation of sample i* at time t; T* represents the time series of observations for sample i*; the historical prescription data of new patients are represented as: in, Let i* be the prescription for sample i at time t; the process of fine-tuning the population parameters of the baseline digital twin using data from new patients is represented as: in ， The parameters representing the individualization of sample i*; The function represents the individualization of sample i*; fine-tuning the parameters of the population twin on individual data makes the digital twin more closely resemble the dynamic characteristics of new patients. During fine-tuning, a loss function is established: observation reconstruction loss. ： in, The hidden state of the new sample i* at time t is estimated and obtained using an inference network; state transition consistency loss. ： in, The hidden state of the new sample i* at time t+1 is estimated and obtained using an inference network; population parameter regularization. ： Where λ is the regularization coefficient; the total loss is expressed as ： Where α is the weight parameter; the parameters are obtained using gradient descent population optimization, outputting the population parameters of the individualized state transition function and observation function, as well as the population parameters of the individualized inference network: in, Represents a group Make the function The minimum value is obtained; ultimately, the individualized twins of the current time T are represented as a set. ： in, Let represent the set of all estimated hidden states of sample i* between time 1 and T, and the set of multimodal observation data, respectively. Let i* represent the set of all external interventions for sample i* from time 1 to T-. The function and parameters representing the individualization of sample i*.

7. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 1, characterized in that, Step 4 specifically includes the following: Step 4-1: Obtain existing observations and past prescriptions from the individualized digital twin: ； in, y represents the observation data of sample i* from time 1 to T in the m-th mode; (i*,1) Basic information; y (i*,2) Functional / symptom assessment results; y (i*,3) Historical prescription execution and response data; data for each modality. After being embedded into a vector by a pre-defined encoder, it is represented as z. (m) ; This represents the set of encoding results for all observation data of the m-th mode of sample i* from time 1 to T; This represents the encoding function corresponding to the m-th modal observation data, with the function form as follows: These embedded vectors are used for subsequent knowledge retrieval and large model input; Step 4-2, Prescription Knowledge Base and Embedding; Construct a prescription knowledge base, the types and sources of which include, but are not limited to, clinical guidelines, pharmacopoeias / drug instructions, clinical trial results, validated medical records / case libraries, institutional SOPs, regulations and contraindication lists; Segment each document at the paragraph level and build a vector index, while retaining the original metadata; Use a sentence vector model finely tuned for the medical domain to ensure medical semantic fit; Construct a hybrid retrieval module for the knowledge base, utilizing Sparse Retrieval: Based on the BM25 inverted index algorithm; Denseretrieval: Based on embedding vector similarity; a hybrid retrieval method combining the two retrieves and returns relevant fields, each containing text content, source, and evidence confidence information; Step 4-3, Retrieval Enhancement (RAG); Constructing questions, using the patient's current information (y) to obtain the patient summary and main query prescription task; Simultaneously, the system generates several sub-questions to cover multi-dimensional intervention areas, ensuring that the generated results are targeted and comprehensive in different dimensions; The main query and sub-questions are input into the hybrid retrieval module to obtain the candidate evidence set D={d1,d2,…,d k }; Subsequently, the patient summary, retrieval evidence segments, and safety constraints are concatenated to form the input prompt of the large language model LLM; Step 4-4, Large Model Invocation; Invoke the instruction-tuned / few-shot / fine-tuned medical large model; Input the prompt constructed in step 4-3, including the patient summary, multimodal observation embedding, knowledge base evidence segments retrieved through RAG, as well as safety constraints and taboo rules; The large model generates multiple candidate prescription lists U={u1,...u,...,u n Each prescription contains prescriptions with multiple dimensions.

8. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 1, characterized in that, In step 5, after the candidate prescriptions are generated, the system will perform forward simulation and multi-dimensional simulation evaluation of each prescription scheme in the digital twin environment; through the state transition function f of the digital twin... θ The rolling prediction obtains the state distribution of the next H steps under a given prescription intervention u. The do function represents execution, and do(u) represents executing prescription u. For prescription u, from the current hidden state x t Perform L Monte Carlo forward simulations to obtain the estimated set of sample hidden state trajectories: That is, the set of estimated hidden states from future time t+1 to t+H, totaling L (j from 1 to L); Using the observation function, each trajectory Generate the corresponding observation trajectory: Evaluation indices are calculated using the estimated hidden states of the samples.

9. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 8, characterized in that, The evaluation indicators include efficacy, risk, cost, and compliance; the efficacy is a quantitative indicator that measures the "positive clinical effect" brought about by the prescription, defined as the expected gain of one or more clinical objective functions within H steps in the future; R eff (*) indicates single-step immediate utility, used to evaluate the immediate effect of the prescription under the current state and intervention conditions. Its specific form is constructed according to different clinical application scenarios to reflect the corresponding treatment goals; then the expected efficacy of the prescription is... Defined as: Where, x t+h Let be the hidden state at time t+h; E is the expectation function, in the form E(*); after Monte Carlo sampling, the expected therapeutic effect of this prescription is estimated as follows: ： in, Let be the estimated value of the hidden state at time t+h on the j-th trajectory; the risk represents the probability or expected loss that the prescription will lead to an adverse outcome or exceed a safety threshold; the risk is characterized using probabilistic risk, and the risk estimate is . ： Among them, z safe It is a safety threshold; prescription cost is the resource consumption required for prescription implementation, including time cost c. time Human resource costs staff And the monetary costs of materials / medicines and equipment c money ; Convert different resource consumptions into a uniform monetary cost; Define the total cost function of prescription u as follows: , is represented as: c time The required training time for the patient; c staff c. Number of therapist hours required; money Equipment or medication costs; k1 is the conversion factor for monetary cost per unit of training time; k2 is the conversion factor for monetary cost per unit of therapist's working hours; prescription adherence: the probability of a patient actually fulfilling a prescription or the expected completion rate A. u Estimated by the patient's historical behavior model, the estimated value is expressed as: To provide a unified evaluation of different candidate prescriptions, a comprehensive utility function is constructed by integrating multiple dimensions of indicators, including efficacy, risk, cost, and adherence. This comprehensive utility function unifies the evaluation indicators from different dimensions into a single, comparable utility value, thereby enabling the ranking and screening of candidate prescriptions. The comprehensive utility function is expressed as follows: Among them, the overall utility score E wu w E ,w U ,w A ,w C These are the weighting coefficients.

10. The method for generating and simulating health prescriptions based on individualized digital twins according to claim 1, characterized in that, In step 6, after completing the simulation evaluation and multi-dimensional quantitative scoring of candidate prescriptions, the system enters the "final prescription selection" stage. This stage uses safety as a hard constraint, weighted utility / robustness as the primary criterion, and retains the right of manual review by clinical experts. Specifically, a hard filter is first applied to all candidate prescriptions to eliminate any solutions that violate medical or regulatory constraints; after identifying any solutions that violate medical or regulatory constraints, the final selection is made. in, This means finding a u in set U such that E wu The maximum value is obtained; the final selected prescription u*, including the complete chain of evidence, is then used to generate an executable prescription.

Images