A health management auxiliary decision-making method and device based on multi-mode body digital twinning and a medium

Abstract

The application discloses a health management auxiliary decision-making method and device based on multi-mode body digital twinning and a medium. The method comprises the following steps: performing receiving multi-source data and quality evaluation to calculate a quality score; performing constructing a unified state object to divide an access boundary and an audit track; performing extracting a high-correlation influence node to generate a change vector; performing generating a comprehensive execution sequencing result; performing semantic comparison and shunting update baseline, behavior blocking when matching medical characteristics, and outputting a health management data combination when not matching. The application can effectively block medical practice risks such as prescription overreach from the system root, and simultaneously output a high-landing value improvement method which is quantized and combined with the soft constraint preference of a target object.

Description

Technical Field

[0001] This application relates to the fields of medical and health data processing and artificial intelligence-assisted decision-making technology. Specifically, it relates to a method, device, and medium for assisting health management decision-making based on multimodal digital twins. The method and device provided in this application are not intended to automatically make any disease diagnoses, generate prescriptions, or clinical treatment plans; they are only used to assist in health data processing and security risk interception workflows. All high-risk plans and medical decisions generated are forcibly diverted and require manual verification and intervention. Furthermore, this application involves underlying architecture scheduling technologies such as digital twin state objects, state machine scheduling, a two-layer micro-state buffer pool, temporal characteristic baseline verification, concurrent write lock control, and event-driven differential synchronization. Background Technology

[0002] With the widespread adoption of smart medical examination instruments, wearable devices, and personal health record terminals, massive amounts of multi-source, heterogeneous health data are generated daily. Currently, it is difficult to effectively integrate data collected from multiple industries or modalities, resulting in data silos. Existing health management systems typically only provide simple displays of basic indicators and single-point suggestions based on static rules, lacking the ability to construct unified digital twin state objects. This makes it difficult to trace the output data of different models or evaluation logics back to their source, and the quality, collection origin, and time confidence of the data cannot be consistently applied to subsequent long-term extrapolation processes.

[0003] Based on this, there is a disconnect between existing health risk assessments and intervention projections. Systems often provide static scores or single-path suggestions, lacking high-pressure testing and multi-objective execution sequence projections based on individual soft and hard constraints. This results in recommendations that do not align with the target group's dietary preferences, time budgets, and financial limitations, leading to extremely low adherence. Furthermore, directly using conventional general-purpose large language models to generate recommendations can easily result in the output of clinically risky diagnostic conclusions, prescription suggestions, or even high-risk intervention methods that induce severe physiological burdens due to the generalization and divergence characteristics of large-scale model algorithms. This lacks underlying safety gating and logical blocking mechanisms.

[0004] On the other hand, as the sampling rate of high-frequency wearable physiological time-series data accepted by the system increases, concurrent ingestion of multiple data streams often leads to write contention bottlenecks in the underlying database. When high-frequency physiological time-series signals and low-frequency discrete events simultaneously initiate write requests to the core shared state object, existing optimistic locking or retry mechanisms are prone to write failures and lock storms, which in turn cause the entire auxiliary decision-making system's computation channel to be blocked by invalid physiological fluctuations and collapse. Summary of the Invention

[0005] The main purpose of this application is to provide a health management auxiliary decision-making method, device and medium based on multimodal digital twins, so as to overcome the technical defects of existing technologies, such as the difficulty in tracing multi-source data, the lack of individual constraints in health advice, the medical safety practice risks in the large model generation process, and the tendency of high-frequency data concurrent writing to trigger lock storms.

[0006] To achieve the above objectives, the first aspect of this application provides a health management auxiliary decision-making method based on multimodal digital twins, executed by a computer device, comprising: receiving multi-source phenotypic data and time-series physiological data of a target object, performing consistency assessment and calculating a quality score; constructing a unified state object, dividing data access boundaries and recording operation audit trajectories within the unified state object; extracting highly correlated influencing nodes in a health risk-driven network, generating physiological indicator change vectors under different intervention paths based on the health risk-driven network, and outputting a comprehensive execution ranking result for candidate follow-up data according to a preset multi-constraint objective function; semantically comparing the comprehensive execution ranking result with a preset medical interception rule dictionary, performing behavioral blocking and diverting to a manual review pipeline when matching medical diagnosis and prescription features, and outputting a health management data combination when not matching the medical diagnosis and prescription features, and using the health management data combination to update the iterative baseline of the unified state object.

[0007] Furthermore, the step of outputting a comprehensive execution ranking result for the candidate follow-up data based on a preset multi-constraint objective function includes: obtaining multi-dimensional preference constraint data corresponding to the target object, wherein the multi-dimensional preference constraint data includes time dimension data, financial restriction data, and exercise ability assessment data; calculating the expected health benefit value of the candidate follow-up data based on the physiological indicator change vector; calculating the potential negative risk value and data missing penalty item of the candidate follow-up data during the execution process; performing a weighted fusion calculation on the multi-dimensional preference constraint data, the expected health benefit value, the potential negative risk value, and the data missing penalty item through the multi-constraint objective function to obtain a comprehensive score; and sorting all the candidate follow-up data in descending order according to the comprehensive score to generate the comprehensive execution ranking result.

[0008] Furthermore, the multi-constraint objective function is configured to discount the expected health benefit value according to a preset time distribution model. The specific calculation process includes: mapping the target object's work-rest time distribution characteristics into a time window availability coefficient matrix; calculating the time consumption required for each operation in the candidate follow-up data; performing a dot product operation on the time consumption and the time window availability coefficient matrix to generate a compliance resistance coefficient; subtracting a risk term associated with the product of the potential negative risk value and the data missing penalty term from the expected health benefit value; and dividing the intermediate value obtained after subtracting the risk term by the compliance resistance coefficient to obtain the corrected comprehensive score.

[0009] Furthermore, the step of semantically comparing the comprehensive execution ranking result with a preset medical interception rule dictionary, and blocking behavior and diverting the process to the manual review pipeline when matching medical diagnosis and prescription features includes: extracting the text description vector from the comprehensive execution ranking result; matching and verifying the text description vector with the scope of evidence in the clinical knowledge base based on the retrieval enhancement generation framework; retrieving a list of blocking keywords associated with disease diagnosis terms, drug replacement instructions, and high-risk physiological induction behaviors from the medical interception rule dictionary; if the text description vector contains any word in the list of blocking keywords, triggering a high-risk suspension instruction, intercepting the output process, and packaging the corresponding comprehensive execution ranking result along with the matched list of blocking keywords to the manual review queue.

[0010] Furthermore, the construction of a unified state object, and the division of data access boundaries and recording operation audit tracks within the unified state object, includes: allocating a population file recording area, a multi-frequency signal buffer area, a quality assessment recording area, and a prediction result retention area within the unified state object; pre-setting exclusive read / write permission masks matching the execution scope of multiple independent data processing agents running in the system; receiving data access requests initiated by any of the independent data processing agents against the unified state object, and performing unauthorized interception judgment by reading the corresponding exclusive read / write permission mask; structurally concatenating the target memory coordinates, operation timestamp, and corresponding agent identity identifier of the legitimate data change operation, and writing them into a read-only audit traceability log at the end of the unified state object.

[0011] Furthermore, the process of receiving multi-source phenotypic data and temporal physiological data of the target object includes: allocating parallel micro-state buffer rings and macro-state submission queues in a central scheduling and management program; guiding the high-frequency sampled temporal physiological data to be continuously written into consecutive storage nodes of the micro-state buffer ring, and intercepting high-frequency concurrent write requests to the main database; extracting features from the temporal physiological segments accumulated in the micro-state buffer ring, and when a preset state change condition is detected, transferring the corresponding temporal physiological segment to the macro-state submission queue to await asynchronous persistence processing.

[0012] Furthermore, the step of extracting features from the time-series physiological segments accumulated in the micro-state cache ring, and transferring the corresponding time-series physiological segments to the macro-state submission queue to await asynchronous persistence processing when a preset state mutation condition is detected, includes: instantiating a shadow twin object at the data access layer, controlling the shadow twin object to cache a copy of the previous stable version of the unified state object in memory; establishing a sliding observation window to continuously scan newly arrived data frames in the micro-state cache ring; calculating the statistical mean and variance parameters of the numerical values ​​contained in the data frame in real time; comparing the statistical mean and variance parameters with the historical baseline parameters recorded in the shadow twin object, and if the deviation exceeds a preset confidence interval, determining that a baseline drift event has occurred and confirming that the state mutation condition is met.

[0013] Furthermore, the data in the micro-state cache ring is marked as steady-state data when the state mutation condition is not met. The following control steps are performed on the steady-state data: differential timestamp records of the steady-state data are locally accumulated only at the memory level, and duplicate redundant physiological values ​​are discarded; fluctuations in input values ​​are monitored, and when the micro-state cache ring detects that the input value touches a preset hard physiological safety extreme value, finds a continuous heterogeneous distribution jump in the time dimension, or the time span from the last submission event reaches a preset heartbeat cycle threshold, a forced packaging operation is triggered; the locally accumulated differential timestamp records are structurally recombined with the last collected valid values ​​to generate a differential merged data packet with a priority weight lock; by sending the differential merged data packet, a mutation overwrite write operation for overlapping storage blocks is initiated to the macro-state submission queue.

[0014] A second aspect of this application provides a health management auxiliary decision-making device based on multimodal digital twins, comprising: at least one processor; and a memory communicatively connected to the at least one processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the aforementioned health management auxiliary decision-making method based on multimodal digital twins.

[0015] A third aspect of this application provides a computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned health management auxiliary decision-making method based on multimodal digital twins.

[0016] The auxiliary decision-making system disclosed in this application ensures that every step of the health data operation logic throughout the entire process is securely, clearly, and locatably recorded by implementing feature data isolation and audit traceability based on a unified state object. By establishing a comprehensive scoring and descending-order ranking model that includes time constraints and potential risk calculations through a data inference module, it effectively avoids invalid recommendations that deviate from individual preferences and improves the objective execution rate of output suggestions. The adoption of a strict medical diagnosis blocking semantic matching and manual triage and review mechanism hard-blocks the inherent black-box characteristics and risks of unauthorized practice of large language models, ensuring the compliance of data suggestions from the underlying technical architecture. By adding a micro-state buffer and differential packet architecture, it not only reduces the conflict and preemption initiated by redundant data without feature jumps to the bus, but also successfully solves the underlying database deadlock failure caused by continuous access from multiple high-frequency concurrent sources. Attached Figure Description

[0017] Figure 1 This is a flowchart of a health management auxiliary decision-making method based on multimodal digital twins provided in an embodiment of this application.

[0018] Figure 2 This is a logical structure block diagram of the multi-agent and micro-architecture provided in the embodiments of this application.

[0019] Figure 3 This is a logical structure diagram of the multi-constraint objective function scoring and deduction evaluation process provided in the embodiments of this application.

[0020] Figure 4 This is a block diagram of the logical structure of the microstate cache ring and macrostate submission queue provided in the embodiments of this application.

[0021] Figure 5 This is a logical structure block diagram of a health management auxiliary decision-making device based on multimodal digital twins provided in the embodiments of this application.

[0022] Explanation of reference numerals in the attached figures

[0023] In the diagram: 501 - processor, 502 - memory. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are for illustrative purposes only and should not be considered as limiting the scope of protection of this application. It is also stated that all steps performed in this application involve legally collected data, and all actions are performed by computer only for data structure processing, feature extraction, and semantic blocking of risky terms.

[0025] It should be specifically noted that the execution logic of this application is an auxiliary health data processing system. The system's algorithm mechanism has completely isolated instructions related to automated disease diagnosis, prescription issuance, or medical intervention. For all instructions and parameter configurations with clinical medical risks, this application adopts a truncation and blocking logic, forcibly transferring them all to a manual verification process. Therefore, the technical solution disclosed in this application does not essentially involve disease diagnosis and treatment.

[0026] Combination Figure 1 and Figure 2 This application provides a health management auxiliary decision-making method based on multimodal digital twins. The method is mainly executed by a set of microservice application logic programs running on a backend server. Specifically, the method includes the following sequentially executed steps: receiving multi-source data and quality assessment S101 to semantic comparison and baseline update S105.

[0027] The system receives multi-source data and performs quality assessment (S101). This involves receiving multi-source phenotypic data and time-series physiological data of the target object, performing a consistency assessment, and calculating a quality score. Specifically, the front-end architecture requires the deployment of a front-end sensor gateway and a protocol stack. This protocol stack connects the data channels of medical examination instruments, wearable monitoring devices, and related data entry front-ends. When receiving discrete multi-source phenotypic data and time-series physiological data, the system first performs normalization mapping and time-series alignment on the data. During this process, a consistency assessment is performed to determine the reliability of each representative data point. The system relies on a defined quality evaluation equation to quantify and assign values ​​to each dimension of the indicators, calculating a quality score. This quality score follows a specific formula; specifically, the system performs a weighted calculation based on various evaluation coefficients: ;

[0028] In this formula, the Q variable represents the overall data quality score of the final output; the credibility variable represents the device credibility index of the data source; the completeness variable reflects the integrity assessment value of the data structure field; the timeliness variable measures the time decay value of the time-series data from the current collection timestamp of the system; the consistency variable characterizes the consistency value of whether there are contradictions between heterogeneous data from the same source; the range determination variable is used to verify whether the physiological value is within the allowable range that conforms to physical norms; and the operation permission variable represents the compliant operation permission authorized to call this part of the data. a1 to a6 are the corresponding weighted adjustment coefficients. In this embodiment, the values ​​of a1 to a6 are all preset to empirical constants between 0.1 and 0.3, and their weighted sum is configured to equal 1.0. Based on this, the influence caused by the measurement accuracy deviation of individual sensors or omissions in the data entry can be eliminated, thereby outputting an accurate calibration value Q. The calculated Q score is a probabilistic reference value distributed between 0 and 1. Furthermore, the system sets a processing diversion threshold logic for the calculated quality score Q. When the judgment... At this point, this portion of data will be deemed to have passed consistency quality control and will be directly written to the database in preparation for application. When determined... When the system determines that the data has limited usability, it will mark the data segment with a low confidence level. When Q < 0.50, the data is marked as unusable directly, and a re-collection and marking instruction is automatically triggered to the front end. By employing this quality control calculation and diversion detection mechanism, threshold screening can be used to exclude incomplete samples caused by measurement interference, thereby preventing subsequent long-link recommendation models from being contaminated by unwanted noise and resulting in a shift in decision direction.

[0029] A unified state object S102 is constructed, within which data access boundaries and record operation audit trails are defined. Specifically, when high-confidence data is deposited into the system's core logic layer, the central scheduling intelligence creates a structured storage object called the unified state object. This unified state object is not merely a data context cache for single operations, but a structured digital twin model with mandatory traceability features. Within the unified state object, the system pre-allocates a population file recording area, a multi-frequency signal buffer area, a quality assessment recording area, and a prediction result retention area, each strictly separated by contiguous memory address blocks. The system sets a set of core data structure fields to manage memory information. These core fields include at least an identity identifier field for identity mapping, a data version field for version traceability, a body shape feature field reflecting an individual's body shape, a signal feature field for loaded waveforms, and security review and audit trail fields for record security screening and auditing. Based on this, for multiple professional-grade independent data processing agents running in the system, dedicated read / write permission masks matching the scope of tasks executed by the agents are pre-set in the configuration table. When a data access request is received from any independent data processing agent targeting a specific area of ​​the unified state object, the control chip intercepts and extracts the request header information. It then reads the corresponding dedicated read / write permission mask and performs a bitwise AND operation to determine if the access is unauthorized. If the access is deemed legitimate, the operation is allowed. Furthermore, after the data is written, the system will structurally concatenate the target memory coordinates of the legitimate data change operation, the operation timestamp accurate to microseconds, and the corresponding agent identifier, and write this data to a read-only audit traceability log at the end of the unified state object. By setting up a digital twin object with permission masks and read-only traceability logs, the principle of isolated writes can prevent cross-contamination of model contexts in a multi-threaded environment, thereby enabling logical tracing and responsibility identification for each diagnostic method.

[0030] The system extracts highly correlated impact nodes (S103) from the health risk-driven network. Specifically, the system employs a three-level processing engine, where a data network module parses the previously constructed unified state object. This module extracts the logical boundaries between scattered data points and, based on established clinical literature statistical relationships or high-incidence correlation matrices, constructs a health risk-driven network graph containing multiple data points. The system traverses the data status of each edge and node within this network and assigns hazard scores to various potential driving factors, thereby extracting highly correlated impact nodes. The scoring formula used is as follows: Specifically, the system performs normalized weighted calculations for risk factors across various dimensions:

[0031] ;

[0032] In this scoring formula, the driving score variable represents the priority score for the currently selected network node as a high-risk intervention target; the anomaly variable represents the absolute deviation of the node's original measurement value from the standard range; the modifiability variable represents the degree to which the feature can be changed under external environmental intervention; the evidence support variable represents the support coefficient of the evidence-based papers that associate the node with risk; the connectivity centrality variable reflects the connectivity centrality of the node in the graph topology; the focus fit variable indicates its focus fit for the current set phased task; and the uncertainty variable, as a negative penalty term, refers to the data blind box ambiguity caused by the current missing or interfered measurement of some relevant fields. Coefficients b1 to b6 are all configured with empirical thresholds of 0.1 to 0.4, specifically 0.2. By adopting this driving network calculation equation, the disordered physiological indicators can be quantified and sorted using an asymmetric weighted penalty rule, thereby extracting entity targets with execution value and risk relevance for subsequent path deduction.

[0033] A comprehensive execution ranking result S104 is generated. Based on the health risk-driven network, physiological indicator change vectors under different intervention paths are generated, and a comprehensive execution ranking result is output for the candidate follow-up data according to a preset multi-constraint objective function. (Refer to reference...) Figure 3 The logical structure diagram shows that in this step, the processing flow delves into the trend prediction module composed of a professional time series model. The system retrieves historical baseline trend data and generates a vector of physiological indicator changes for the next few weeks or months based on the potential increases or decreases caused by various intervention methods. To derive practically usable auxiliary recommendations, the system invokes the intervention simulation module, comparing not only the physiological benefits from a purely medical perspective but also simultaneously acquiring multi-dimensional preference constraint data corresponding to the target object. This preference constraint data specifically includes time dimension data, financial constraint data, and exercise ability assessment data obtained from physical fitness questionnaires. Based on the generated physiological indicator change vector, this processing module extracts the expected health benefit values ​​of candidate follow-up data obtained after simulating various actions. Simultaneously, for certain action flows containing extreme recommendations, the algorithm extracts potential negative risk values ​​that could burden other organs and data missing penalty items due to incomplete data. A fusion scoring and descending order processing is performed using a preset multi-constraint objective function. The comprehensive scoring logic here is based on the following specific formula; specifically, the system performs a comprehensive scoring calculation covering benefits, costs, and resistance:

[0034] ;

[0035] In this formula, the comprehensive score variable is the final comprehensive effectiveness score of each candidate implementation suggestion package calculated and output; the expected benefit variable represents the quantified expected health benefit value; the compliance ability variable represents the action compliance ability combined with individual feedback assessment; the preference alignment variable represents the degree of alignment between the plan and the habit preferences declared by the target subject in the questionnaire; the guideline support variable represents the degree of support of the clinical knowledge base guidelines behind the extracted recommendation plan; and as a negative penalty logic, the potential negative risk variable corresponding to w5 multiplication is the potential negative risk value; the data missing penalty variable is the data missing penalty item mentioned above; the financial consumption variable represents the estimated degree of financial restriction consumption; and the physical load variable represents the life occupation and physical exhaustion load brought about by the action itself. The values ​​of coefficients w1 to w8 are all preset to controllable normalization factors between 0.05 and 0.25.

[0036] Furthermore, the multi-constraint objective function is configured to perform deep discounting calculations on the expected health benefit value according to a preset time distribution model. The specific technical actions of this mechanism are as follows: The trajectory model in the log is read, and the target object's daily routine distribution characteristics are mapped to a time window availability coefficient matrix composed of a two-dimensional time-segment grid. The time consumption required for each operation in a current candidate follow-up data is calculated. This time consumption is input into the processor's underlying layer and multiplied by the time window availability coefficient matrix in tensor dimension to obtain the compliance resistance coefficient that hinders the implementation of the behavior. The final equation for the discounting calculation is: based on the initial expected health benefit value, a risk compensation term associated with the product of the potential negative risk value and the data missing penalty term is subtracted; the intermediate value obtained after subtracting the risk term is forcibly divided by the compliance resistance coefficient obtained through the above multiplication, thereby obtaining the corrected comprehensive score. By adopting a discounting mechanism that incorporates compliance resistance coefficients and a multi-dimensional preference superposition descending order mechanism, a rigorous negative decay algorithm can be used to filter out extreme paper solutions that seem effective but cannot be implemented under individual real-world constraints. This achieves the technical effect of converging the solution ranking results towards personalization and feasibility.

[0037] Semantic comparison and diversion update baseline S105 involves semantically comparing the comprehensive execution ranking result with a preset medical interception rule dictionary. When a match is found with medical diagnosis and prescription features, behavioral blocking is implemented and the request is diverted to the manual review pipeline. If no match is found, a health management data combination is output and used to update the iterative baseline of the unified state object. Specifically, after all candidate suggestions are sorted in descending order and organized into report fragments, the system does not directly deliver them for display. This is because if some generated text is generated from a general model base, its generalization characteristics could easily violate the bottom line of medical industry regulations. Therefore, the system internally configures a filtering gateway composed of an evidence retrieval agent and a security review module. The system extracts the text description vector sequence from the comprehensive execution ranking result constructed in the preceding process. The system invokes the hybrid model routing gateway configuration logic to send the text request to an internally deployed retrieval enhancement generation framework based on vector distance calculation. Using the text segmentation and encoding logic of this retrieval enhancement generation framework, the text description vector is placed into the clinical knowledge base evidence application scope containing numerous evidence-based rules for strong consistency matching verification. Within the same action cycle, the control program searches a strict contraindication list in a preset medical interception rule dictionary to see if it contains blocking keywords associated with absolute disease diagnoses, instructions for adjusting or replacing any medication dosage, or high-risk physiological induction behaviors. When the comparison reveals that the text description vector contains any word or semantic phrase with extremely high similarity in the blocking keyword list, the system reverses its underlying judgment, triggering a high-risk suspension instruction and forcibly blocking the outgoing output of that content. After the content is blocked, the system does not discard it but encapsulates the corresponding comprehensive execution ranking result along with the blocked keyword list that triggered the matching alarm, transferring it to a manual review queue for approval by a licensed pharmacist or clinical staff. Only when all results are unmatched and deemed compliant does the system normally deliver the health management data combination. Furthermore, in the timeline after the recommendation is issued, the system, based on the configured follow-up task plan, tracks the actual indicators after port backflow execution and feeds them back to the original unified state object to update the iterative baseline parameters in the prediction evolution module. This security isolation framework, based on hard rule dictionary matching and interception, can utilize pre-set non-compliant expression vector blocking barriers to block false medical promises and unsafe risk outputs caused by the randomness of large models, thereby ensuring the engineering compliance of this application as an auxiliary prompt for security information positioning.

[0038] It should be noted that the above-described architecture for data modeling and predictive review is only an example of the main business flow. Those skilled in the art can also realize the design concept of this application at the data cleaning and interception level by using other alternative underlying feature assembly sequences and configuring related filter terms.

[0039] In another embodiment of this application, in conjunction with reference to Figure 4 This paper presents a micro-state buffer queuing principle structure diagram to address the technical challenge of lock storms and crashes in the shared database of the central unified state object caused by multiple wearable vital sign data sources concurrently pushing data at extremely high sampling rates. This application deploys a dual-buffering and differential merging architecture at the intersection of the central scheduling agent and the data communication bus to handle the injection of high-frequency heterogeneous data. Specifically, the process of receiving time-series physiological data is modified as follows: the system allocates two sets of parallel queue logic in the central scheduling management program—a memory-level micro-state buffer ring and a macro-state submission queue for interfacing with long-term database write locks. When high-frequency sampled time-series physiological data is pushed in from the underlying link, the system forcibly cuts off its request to directly acquire persistent locks, instead guiding it uniformly and continuously writing it into the continuous storage nodes of the micro-state buffer ring, intercepting high-frequency concurrent write impacts on the main database. To screen out truly valuable medical mutation points, the system specifically instantiates and deploys a shadow twin object at the data access layer and controls this shadow twin object to cache a copy of the previous stable version of the central unified state object in memory. Within the memory runtime area, the program establishes a sliding observation window consisting of a timeline and pointers. This window continuously scans newly arriving data frame segments in the micro-state cache ring and calculates the statistical mean and variance parameters of various physiological values ​​contained in the data frame in real time. The system activates the state change drift judgment logic calculation engine to perform a high-frequency difference comparison between the calculated statistical mean and variance parameters and the historical baseline parameters recorded in the shadow twin object. If the deviation is within the stable fluctuation range, the system ignores it; however, if the calculated deviation exceeds the preset confidence interval, a baseline drift event is determined to have occurred, and the state change condition initiated to the bus is confirmed to be met. The core calculation formula for this drift judgment condition is specifically configured as follows: Specifically, the system performs a judgment deduction of variance and confidence multiplier:

[0040]

[0041] In this baseline verification formula, the trigger state variable serves as a Boolean trigger flag; the current time series mean variable represents the mean of the current segment captured within the sliding window; the historical baseline mean and historical baseline variance variables represent the macroscopic mean and variance of the previous historical period stored in the shadow twin, respectively; the coefficient K serves as a scaling multiplier for fluctuation tolerance, typically ranging from 2.0 to 3.0, representing the upper limit of statistical confidence. By employing sliding window time series feature comparison in conjunction with isolated memory pool technology, it is possible to filter and shield physiologically useless data that only exhibits minor, normal fluctuations using local computation, thereby reducing the traffic load entering the main business line at the source.

[0042] Furthermore, in this high-concurrency debouncing implementation branch, for data in the micro-state cache ring that do not meet the aforementioned state change conditions in the above checks, and whose amplitude conforms to the formula's safety threshold, they are temporarily marked as steady-state data. The system implements local folding and dimensionality reduction control logic for the steady-state data. The specific steps are as follows: without triggering physical disk writes, the system only locally accumulates the differential timestamp records of the steady-state data at the memory cache level, and immediately discards those redundant physiological numerical entities that are repetitive and unchanging. At the same time, the program process initiates a second monitoring defense line to detect the extreme limits of the input values. When the micro-state cache ring, in addition to not detecting variance drift, discovers through threshold judgment that the input value has touched the preset hard physiological safety extreme value. The logic for triggering the absolute threshold for high-frequency data state transitions is as follows: The absolute boundary condition (newly accessed data value greater than the upper limit of the physiological safety limit) or (newly accessed data value less than the lower limit of the physiological safety limit) is met; or a continuous heterogeneous distribution transition is detected in the time sequence dimension; or the time span between the local accumulator and the last submission event reaches a preset heartbeat cycle threshold. As soon as any of the above critical conditions is triggered, the control module immediately triggers a forced packaging operation. At this time, the scheduler structurally reassembles the differential timestamp record group accumulated in the previous local accumulator with the last valid value collected to trigger the alarm, generating a differential merged data packet with a request priority weight lock. During the final step of system execution, the differential merged data packet is sent through the network port, formally initiating an exclusive mutation overwrite write operation targeting the underlying overlapping storage block within the macroscopic state submission queue. This dual-trigger event-driven differential synchronization mechanism, consisting of time heartbeats and extreme value jumps, can leverage edge memory isolation and packaged acquisition logic to alleviate system-level resource exhaustion or read-write deadlock issues caused by high-frequency probe direct injection into shared core objects, thereby improving the robustness of the data ingestion and transfer framework.

[0043] like Figure 5As shown, based on the same inventive concept, another embodiment of this application provides a health management auxiliary decision-making device based on multimodal digital twins. This embodiment focuses on describing the physical circuit hardware architecture that carries the aforementioned software logic system. The device, as an independent computing cluster node or intelligent terminal facility, includes at least one processor 501 and a memory 502 interconnected via a bus. The processor 501 can be a central processing unit, or a general-purpose processor, digital signal processor, application-specific integrated circuit, field-programmable gate array, or other electronic gating device with programmable logic operation capabilities, used to handle the parallel floating-point logic calculation instruction scheduling for the data quality equation, complex matrix multiplication discount derivation, and high-voltage concurrent anti-jitter ring in this application. The memory 502 serves as a data loading area, including but not limited to volatile cache units such as random access memory, and non-volatile persistent storage devices such as flash memory chips and magnetic media blocks. Within the memory 502, fixed sections are allocated to record the aforementioned cache sequence containing secure retrieval large language model parameters and shadow twins. Specifically, the memory 502 stores a large set of pre-programmed machine instruction codes executable by the at least one processor. When the device is powered on and starts up from the microsystem to enter the normal execution state, the above instructions are read and executed by the at least one processor 501, thereby controlling the entire physical device and its configured front-end and back-end sensor communication interfaces and database link interfaces, and coordinating the execution of the complete set of health management auxiliary decision-making methods based on multi-modal digital twins, including receiving multi-source vital sign data for digital twin mapping and archiving, using predictive inference modal fusion discounting indicators to give scores and rankings, and executing medical risk statement suspension and solving multi-thread blocking based on double buffer ring differential dimensionality reduction.

[0044] Furthermore, another embodiment of this application provides a computer-readable storage medium. This computer-readable storage medium can be any tangible, non-transient digital media product capable of containing or embedding instruction code, including but not limited to USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks, and other media capable of persistently or temporarily storing combinations of program code. The computer-readable storage medium internally contains and stores dedicated computer instruction files. When these computer instructions are read, compiled, and executed by a microprocessor mounted within a data processing device or computing cluster, they drive the computer environment to complete various data collection, deduction, comparison, and joint judgment operation workflows based on time-sensitive differential buffering and security semantic screening and interception, as described in all the preceding method embodiments.

[0045] The technical solutions of this application involve specific industry application logic mechanisms such as digital twin state objects, state machine scheduling, two-layer microstate buffer pools, timing feature baseline verification, concurrent write lock control, and event-driven differential synchronization. All of the above technical features are fully disclosed herein as a tightly linked and inseparable combination of technologies to achieve the overall purpose. Those skilled in the art can also realize the technical concept of this application in related fields by using equivalent mathematical operations that have similar utility, heterogeneous adjustments to local storage media, or fine-tuning and reorganizing the timing of software processes, as long as the core purpose is to handle auxiliary health information anti-jitter deduction and blocking logic.

Claims

1. A health management auxiliary decision-making method based on multimodal digital twins, characterized in that, Performed by computer equipment, the process includes: receiving multi-source phenotypic data and time-series physiological data of the target object, performing consistency assessment and calculating quality scores; constructing a unified state object, defining data access boundaries and recording operation audit trajectories within the unified state object; extracting highly correlated influencing nodes from the health risk-driven network, generating physiological indicator change vectors under different intervention paths based on the health risk-driven network, and outputting a comprehensive execution ranking result for candidate follow-up data according to a preset multi-constraint objective function; semantically comparing the comprehensive execution ranking result with a preset medical interception rule dictionary, blocking behavior and diverting to the manual review pipeline when matching medical diagnosis and prescription features, and outputting a health management data combination when not matching the medical diagnosis and prescription features, and using the health management data combination to update the iterative baseline of the unified state object.

2. The health management auxiliary decision-making method based on multimodal digital twin as described in claim 1, characterized in that, The step of outputting a comprehensive execution ranking result for candidate follow-up data based on a preset multi-constraint objective function includes: obtaining multi-dimensional preference constraint data corresponding to the target object, wherein the multi-dimensional preference constraint data includes time dimension data, financial restriction data, and exercise ability assessment data; calculating the expected health benefit value of the candidate follow-up data based on the physiological indicator change vector; calculating the potential negative risk value and data missing penalty item of the candidate follow-up data during the execution process; performing a weighted fusion calculation on the multi-dimensional preference constraint data, the expected health benefit value, the potential negative risk value, and the data missing penalty item through the multi-constraint objective function to obtain a comprehensive score; and sorting all the candidate follow-up data in descending order according to the comprehensive score to generate the comprehensive execution ranking result.

3. The health management auxiliary decision-making method based on multimodal digital twin as described in claim 2, characterized in that, The multi-constraint objective function is configured to discount the expected health benefits according to a preset time distribution model. The specific calculation process includes: mapping the target object's work-rest time distribution characteristics into a time window availability coefficient matrix; calculating the time consumption required for each operation in the candidate follow-up data; performing a dot product operation on the time consumption and the time window availability coefficient matrix to generate a compliance resistance coefficient; subtracting a risk term associated with the product of the potential negative risk value and the data missing penalty term from the expected health benefits; and dividing the intermediate value obtained after subtracting the risk term by the compliance resistance coefficient to obtain the corrected comprehensive score.

4. The health management auxiliary decision-making method based on multimodal digital twin as described in claim 1, characterized in that, The step of semantically comparing the comprehensive execution ranking result with a preset medical interception rule dictionary, and blocking behavior and diverting the process to the manual review pipeline when matching medical diagnosis and prescription features, includes: extracting text description vectors from the comprehensive execution ranking result; matching and verifying the text description vectors with the scope of evidence in the clinical knowledge base based on a retrieval enhancement generation framework; retrieving a list of blocking keywords associated with disease diagnosis terms, drug replacement instructions, and high-risk physiological induction behaviors from the medical interception rule dictionary; if the text description vector contains any word in the blocking keyword list, triggering a high-risk suspension instruction, intercepting the output process, and packaging the corresponding comprehensive execution ranking result along with the matched blocking keyword list and transferring it to the manual review queue.

5. The health management auxiliary decision-making method based on multimodal digital twin as described in claim 1, characterized in that, The construction of a unified state object, and the division of data access boundaries and recording operation audit tracks within the unified state object, includes: allocating a population file recording area, a multi-frequency signal buffer area, a quality assessment recording area, and a prediction result retention area within the unified state object; pre-setting exclusive read / write permission masks matching the execution scope of multiple independent data processing agents running in the system; receiving data access requests initiated by any of the independent data processing agents against the unified state object, and performing unauthorized interception judgment by reading the corresponding exclusive read / write permission mask; structurally concatenating the target memory coordinates, operation timestamp, and corresponding agent identity identifier of the legitimate data change operation, and writing them into a read-only audit traceability log at the end of the unified state object.

6. The health management auxiliary decision-making method based on multimodal digital twin as described in claim 1, characterized in that, The process of receiving multi-source phenotypic data and temporal physiological data of the target object includes: allocating parallel micro-state buffer rings and macro-state submission queues in a central scheduling and management program; guiding the high-frequency sampled temporal physiological data to be continuously written into the continuous storage nodes of the micro-state buffer ring, and intercepting high-frequency concurrent write requests to the main database; extracting features from the temporal physiological segments accumulated in the micro-state buffer ring, and when a preset state change condition is detected, transferring the corresponding temporal physiological segment to the macro-state submission queue to wait for asynchronous persistence processing.

7. The health management auxiliary decision-making method based on multimodal digital twin as described in claim 6, characterized in that, The step of extracting features from the time-series physiological segments accumulated in the micro-state cache ring and transferring the corresponding time-series physiological segments to the macro-state submission queue for asynchronous persistence processing when a preset state mutation condition is detected includes: instantiating a shadow twin object at the data access layer and controlling the shadow twin object to cache a copy of the previous stable version of the unified state object in memory; establishing a sliding observation window to continuously scan newly arrived data frames in the micro-state cache ring; calculating the statistical mean and variance parameters of the numerical values ​​contained in the data frame in real time; comparing the statistical mean and variance parameters with the historical baseline parameters recorded in the shadow twin object, and if the deviation exceeds a preset confidence interval, determining that a baseline drift event has occurred and confirming that the state mutation condition is met.

8. The health management auxiliary decision-making method based on multimodal digital twin as described in claim 6, characterized in that, Data in the micro-state cache ring is marked as steady-state data when the state mutation condition is not met. The following control steps are performed on the steady-state data: differential timestamp records of the steady-state data are locally accumulated only at the memory level, and duplicate redundant physiological values ​​are discarded; fluctuations in input values ​​are monitored, and when the micro-state cache ring detects that the input value touches a preset hard physiological safety extreme value, finds a continuous heterogeneous distribution jump in the time dimension, or the time span from the last submission event reaches a preset heartbeat cycle threshold, a forced packaging operation is triggered; the locally accumulated differential timestamp records are structurally recombined with the last collected valid values ​​to generate a differential merged data packet with a priority weight lock; by sending the differential merged data packet, a mutation overwrite write operation for overlapping storage blocks is initiated to the macro-state submission queue.

9. A health management auxiliary decision-making device based on multimodal digital twins, characterized in that, include: At least one processor; A memory communicatively connected to the at least one processor, the memory storing instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the health management auxiliary decision-making method based on multimodal digital twins as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the health management auxiliary decision-making method based on multimodal digital twins as described in any one of claims 1 to 8.

Images