A multimodal AI large model driven community chronic disease health management method and system

Abstract

The application relates to the technical field of data processing, and discloses a community chronic disease health management method and system driven by a multi-modal AI large model, which comprises the following steps: accessing multiple health monitoring data sources, uniformly mapping original measurement values to a standardized time grid and performing numerical normalization processing to complete multi-source information alignment. A large model is used to give semantic labels to the structured data, convert different modalities into unified semantic encoding, and detect abnormal jumps between adjacent records. User-specific semantic slots and fingerprints are constructed, repeated information is removed, and a dynamic health record is formed. On this basis, conflict analysis is implemented, data errors or real fluctuations are distinguished according to the change rate and the medical reachable range, and a confidence level is given to each record. A comprehensive health vector and a risk value are calculated by coupling multiple indexes, disease symptoms are automatically identified, and a risk level is divided. Finally, early warnings are generated for individuals and communities respectively.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a multimodal AI large model-driven community chronic disease health management method and system. Background Technology

[0002] With the accelerating aging of the population and changes in lifestyle, the demand for community-based chronic disease management continues to rise. Current community health management methods largely rely on single data sources for collection and manual analysis. Data comes from different devices and at different times, and includes numerical, textual descriptions, and medical images. This heterogeneous information lacks a unified time alignment and standardized processing mechanism. In practice, the recording formats of different devices vary greatly, measurement units are inconsistent, and timestamps are biased, making it difficult to compare and integrate multi-source data on the same dimension. Some solutions attempt to simply fill in missing data, but lack interval stability and accessibility checks based on medical guidelines, easily introducing outliers or distorted data, affecting the accuracy of subsequent health assessments.

[0003] At the semantic understanding level, existing technologies mostly employ rule engines or traditional machine learning models, which have limited capacity for structured processing of multimodal inputs. Numerical indicators, textual descriptions, and image information are often processed independently, failing to establish a unified semantic expression framework across modalities. Even when semantic encoding attempts exist, they fail to adequately consider the normalization of synonymous expressions and lack automatic detection and reliable propagation mechanisms for semantic jumps and intermodal conflicts. Consequently, the resulting health records contain a large amount of redundant and conflicting records, making it difficult to support accurate tracking of individual health status and analysis of changing trends.

[0004] In terms of risk warning and group management, existing methods mostly rely on static thresholds or single indicators for judgment, failing to incorporate the temporal changes in individual health status, slot stability, semantic confidence, and the coupling effect of multiple indicators into a unified risk quantification model. The ability to analyze risk diffusion and clustering trends at the community level is weak, making it impossible to adjust intervention priorities in a timely manner based on individual risk evolution and group distribution characteristics, resulting in delayed responses and inaccurate resource allocation in chronic disease management.

[0005] Therefore, there is an urgent need for a full-process community chronic disease health management method that can integrate multimodal data, unify semantic expression, automatically check for conflicts, assess confidence levels, and dynamically model health status and risks, so as to improve the intelligence, precision and timeliness of community chronic disease management. Summary of the Invention

[0006] This invention provides a community-based chronic disease health management method and system driven by a multimodal AI large model, which helps to solve the problems mentioned in the background art.

[0007] This invention provides the following technical solution: a community-based chronic disease health management method driven by a multimodal AI large model, comprising:

[0008] Multi-source data is mapped to a unified time grid and missing values ​​are filled in. Abnormal fluctuations are identified through interval stability checks.

[0009] Combining modality types with medical knowledge bases, semantic units are output from a large model. After consistency judgment and encoding, semantic sequences are constructed and transition anomalies are detected.

[0010] Construct semantic fingerprints to remove duplicate data, assign records to semantic slots to form a time chain, detect conflicts within the slots and evaluate stability to establish a health profile;

[0011] Analyze the rate of change and density of numerical values ​​within the tank, determine the type of conflict and correct data error-type conflicts, and output the conflict analysis results;

[0012] An initial confidence level is assigned to the semantic record. After conflict correction and neighborhood propagation, the enhanced confidence level is determined by combining the slot mean. The levels are then classified and outliers are marked.

[0013] Extract the status of each slot to construct a health vector, calculate the changes in health values ​​and the coupling strength between slots, classify the status according to the hierarchical threshold and mark the abnormal, and generate a unified health profile.

[0014] Track the direction and magnitude of health changes, determine the risk level by comprehensively considering trends, transitions and coupled risks, trigger immediate or continuous warnings and output disease sets;

[0015] By summarizing risk indicators of community users, identifying the status of risk spread, delineating high-risk groups and determining intervention priorities, and triggering individual or community early warnings.

[0016] Optionally, the step of mapping multi-source data to a unified time grid and filling in missing values, and identifying abnormal fluctuations through interval stability verification, includes:

[0017] Define the set of data sources to be accessed, and associate each data record with a user identifier, collection time, and original measurement value;

[0018] Establish a unified time grid, with its nodes divided at fixed time intervals;

[0019] Map the acquisition time of each record to the nearest time grid node;

[0020] For each data source, the dimensional range is determined based on the maximum and minimum values ​​of its historical data, and the original measurement values ​​are converted into standardized values ​​accordingly.

[0021] For data missing at a certain point in time, the mean of the standardized values ​​existing before and after that point in time is used to fill the missing data.

[0022] Finally, a unified data matrix is ​​constructed with users, time nodes, and data sources as dimensions. The stability of the data is verified by calculating the standardized numerical difference between adjacent time nodes. If the difference exceeds the range allowed by the preset medical guidelines, the data point is marked as abnormal.

[0023] Optionally, the step of combining modality types with the semantic units output by the large model based on medical knowledge, constructing a semantic sequence after consistency determination and encoding, and detecting abrupt transitions includes:

[0024] The modality type of each data source is identified by a preset mapping rule, including at least numerical, text, and image types;

[0025] Construct structured input units and introduce a knowledge base that stores medical rules related to chronic diseases;

[0026] Utilize large models to output predefined semantic labels based on input units;

[0027] Semantic tags are mapped to specific medical semantic entities and their corresponding attribute values, and the semantic relationship between them is determined based on the threshold range in medical guidelines, generating structured semantic units.

[0028] The consistency between modalities is verified by comparing the attribute values ​​of the same entity under different modalities.

[0029] Assign integer codes to entities, attribute values, and relations respectively, and perform combined calculations to generate unique semantic code values;

[0030] The semantic encoded values ​​of users are sorted by time to form a semantic sequence, and abnormal semantic jumps are identified by detecting abrupt changes in adjacent encoded values.

[0031] Optionally, the step of constructing semantic fingerprints to remove duplicate data, assigning records to semantic slots to form a time chain, detecting intra-slot conflicts, and evaluating stability to establish a health profile includes:

[0032] The semantic unit is redefined as a semantic primitive composed of entities, attributes, and relations;

[0033] A medical thesaurus is used to normalize different entities expressing the same concept and unify their numbering.

[0034] A unique semantic fingerprint value is generated by combining normalized entity, attribute, relation and time encoding;

[0035] Identify and reject completely duplicate data by comparing differences in semantic fingerprint values;

[0036] Create an independent semantic slot for each type of normalized entity, and store the verified attribute values ​​into the corresponding slot in chronological order;

[0037] Conflicts are detected by calculating the differences between attribute values ​​at different time points within the same slot. If the difference exceeds the threshold specified in medical guidelines, it is marked.

[0038] Ultimately, a user's structured health profile is formed, consisting of each semantic slot and its temporal data chain.

[0039] Optionally, the analysis includes the rate of change and density of values ​​within the analysis tank, determining the type of conflict and correcting data error-type conflicts, and outputting conflict analysis results, including:

[0040] Calculate the attribute value difference and time interval between adjacent records in the same semantic slot, and derive the rate of change accordingly;

[0041] The rate of change is compared with the maximum rate allowed by medical guidelines to initially determine a conflict;

[0042] A secondary verification is performed by combining the sampling time density and the range of physically reachable changes;

[0043] Based on the judgment results, the conflicts are divided into three types: conflicts caused by data errors, conflicts reflecting real physiological changes, and abnormal conflicts caused by excessively high sampling frequency.

[0044] For conflicts that are determined to be data errors, the average of the normal data before and after the conflict is used for correction;

[0045] The final output includes analysis results containing conflict flags, conflict types, physical reachability flags, and conflict zones.

[0046] Optionally, the step of assigning an initial confidence level to semantic records, determining a strengthened confidence level after conflict correction and neighborhood propagation, and combining this with the slot mean, classifying levels, and marking outliers includes:

[0047] An initial confidence level is set for each semantic record based on whether it is conflicting and the type of conflict, and then corrected using a physical reachability flag;

[0048] The neighborhood consistency difference is calculated using the attribute values ​​of adjacent records. If the difference is less than a preset threshold, the average confidence of the record itself and its neighbors is taken as the confidence after propagation; otherwise, the corrected confidence is maintained.

[0049] Calculate the average confidence of all records within the same semantic slot, and combine it with the post-propagation confidence to obtain the final confidence.

[0050] Discrete levels are assigned based on the final confidence score to indicate the level of confidence.

[0051] The final output includes the final confidence value, confidence level, and anomaly confidence markers.

[0052] Optionally, the step of extracting the status of each slot to construct a health vector, calculating the change in health value and the coupling strength between slots, classifying the status according to a hierarchical threshold and marking abnormalities, and generating a unified health profile includes:

[0053] Define state information for each semantic slot, including its corresponding entity, the latest valid attribute value, the average confidence level of the slot, and the number of records.

[0054] The user's comprehensive health status vector is calculated based on the current attribute values ​​of each slot and their confidence weights.

[0055] Calculate the change and rate of change of the health status vector, and classify the health status into different levels according to the medical grading standard;

[0056] By constructing coupling pairs between different semantic slots and calculating the change in their attribute differences, the linkage between indicators can be monitored.

[0057] If the change in health status or level exceeds a preset threshold, an anomaly is marked.

[0058] Finally, the above information is integrated to generate a unified health record model.

[0059] Optionally, the step of tracking the direction and magnitude of health changes, determining the risk level by considering trends, transitions, and coupled risks, triggering immediate or continuous early warnings, and outputting a set of diseases includes:

[0060] Analyze the direction of change in health status, the length of continuous unidirectional changes, and the cumulative magnitude to calculate the trend risk value;

[0061] Detect whether a health level has transitioned, and generate a transition risk value;

[0062] The number of times the change in coupling relationship exceeds a threshold is counted to calculate the coupling risk value;

[0063] The trend risk value, transition risk value, and coupling risk value are combined to obtain the comprehensive risk value;

[0064] Risk levels are determined based on the comparison between the comprehensive risk value and a preset threshold, and different levels of warnings are triggered accordingly.

[0065] This is achieved by matching the user's set of symptoms with a pre-established set of disease symptoms based on medical guidelines.

[0066] Optionally, the process of summarizing community user risk calculation group indicators, identifying the risk spread status, delineating high-risk groups and determining intervention priorities, and triggering individual or community early warnings includes:

[0067] Build a community user set and obtain the individual risk value for each user;

[0068] Based on this, we can calculate the overall average risk level of the community, the proportion of high-risk users, and the dispersion of risk distribution.

[0069] By monitoring the growth rate of average risk in the community, it can be determined whether there is a risk diffusion trend;

[0070] Identify and generate sets of high-risk users and sets of users with ongoing risks;

[0071] The priority of individual interventions is determined based on the user's overall risk score and the number of historical warnings.

[0072] When an individual's risk level reaches a preset standard, an individual warning is triggered and information about their possible diseases is pushed to them.

[0073] When the proportion of high-risk users in a community exceeds a preset threshold, a community-level group alert is triggered.

[0074] A system for implementing the community chronic disease health management method driven by the multimodal AI large model includes:

[0075] The data is fed into the preprocessing module, which performs unified time alignment, normalization and missing data completion on multi-source heterogeneous health data, and outputs a structured standard data matrix to provide basic input for subsequent semantic analysis.

[0076] The multimodal semantic parsing module is used to input standardized data into a large multimodal model, generate structured semantic information under the constraints of medical knowledge, and perform consistency and abnormal semantic detection.

[0077] The semantic deduplication and slot management module is used to eliminate semantic redundancy and duplicate data, and organize semantic information into a traceable time-series slot structure by entity;

[0078] The conflict identification and data correction module is used to identify abnormal changes, sampling anomalies, and physical inaccessibility in the data, and to automatically correct erroneous data.

[0079] The confidence assessment module is used to evaluate the credibility of each piece of semantic data and improve the overall stability through a neighborhood propagation mechanism.

[0080] The health status modeling module is used to integrate information from multiple indicators to build a dynamic health status model for users.

[0081] The risk assessment and early warning module is used to comprehensively assess user health risks and trigger early warnings from the perspectives of trends, mutations, and multi-indicator coupling.

[0082] The present invention has the following beneficial effects:

[0083] 1. By constructing a standardized mechanism for unified processing and fusion of multi-source heterogeneous data, this invention completely solves the problems of data silos and incomparability caused by the numerous brands of testing equipment, different data formats, and inconsistent collection frequencies in community chronic disease management. Through defining a unified time grid, performing standardized value mapping, and intelligently filling in missing data, the method successfully transforms raw measurements from different data sources into a standardized data matrix with unified physical meaning and mathematical dimensions. This significantly improves data usability. In real-world community applications, users may alternate between home self-testing devices and shared equipment at community health service centers. System errors and dimensional differences between devices often make it difficult to analyze data trends consistently. This invention sets an independent dimensional range for each data source and performs normalization calculations, compressing the raw measurements into a unified closed interval, thereby eliminating data bias caused by differences in device hardware. This allows even historical data recorded by different models of equipment to be compared and analyzed equally within the same mathematical framework, providing a clean and consistent input foundation for subsequent large-scale artificial intelligence model analysis. Secondly, the time grid mapping and data filling strategies in this method effectively address the pain points of sparsity and misalignment in time-series data. In real-world community management, users don't strictly adhere to precise second-level intervals when uploading data, and data uploads are often interrupted or missed. Mapping the collection time to the nearest available time node and using linear interpolation to fill in missing data nodes ensures the mathematical integrity of the time series. This prevents errors in extracting time features from AI models due to timestamp misalignment, ensuring the accuracy of trend calculations. More importantly, the defined interval stability check algorithm can screen for abnormal fluctuations immediately after data fusion. Its maximum allowable variation range is set according to medical guidelines. This allows the system to not only complete data fusion but also simultaneously perform a preliminary review of data rationality based on physiological principles, effectively filtering out false anomalies caused by momentary equipment interference or operational errors, ensuring data quality from the source for large-scale model analysis. This proactive data governance capability is unparalleled by traditional manual entry or simple aggregation systems.

[0084] 2. By establishing a modality adaptive mapping and semantic structured coding system, a deterministic leap from low-level physiological values ​​to high-level medical logic is achieved, greatly enhancing the interpretability and accuracy of multimodal large-scale models in chronic disease management. It details how to uniformly transform multiple modalities such as numerical, textual, and image data into standardized semantic units, and then construct rigorous user health profiles through semantic fingerprinting and slot mapping. This solves the common "illusion" problem and logical confusion in general large-scale models when processing professional medical data. By introducing a predefined medical knowledge base and semantic label mapping function, abstract measurement values ​​are forcibly bound to specific medical semantic entities and their corresponding hierarchical relationships. This forced mapping ensures that the output of the AI ​​large-scale model is not a textual description based on probabilistic guessing, but rather a determination based on deterministic medical guidelines. When multiple modalities simultaneously point to the same semantic entity, the modality consistency verification mechanism can automatically detect potential contradictions between numerical indicators and subjective descriptions and trigger conflict markers. This mechanism is crucial for community-based chronic disease management because it alerts healthcare professionals to hidden dangers where "data appears normal but patients complain of severe symptoms," effectively preventing missed diagnoses from relying solely on numerical monitoring. Furthermore, by constructing semantic primitives, semantic fingerprints, and semantic slot sets, complex multimodal time-series data is highly structured into slot-based time-series chains. This innovative data organization method brings significant improvements in storage and retrieval efficiency. By calculating semantic fingerprint dissimilarity, the system can identify and reject completely duplicated redundant data with extremely low computational cost, greatly reducing storage pressure. Simultaneously, stability detection and conflict functions within slots automatically identify and mark severe contradictions occurring in the same semantic slot within a short period before data is entered into the database. This structured preprocessing allows subsequent large models to avoid time-consuming pattern recognition in massive amounts of unstructured raw data, instead performing inference and prediction directly on clean, standard semantic record units. This not only accelerates response time but, more importantly, elevates the role of large models from "data cleaners" to true "clinical reasoning experts," enabling them to focus on uncovering deep health evolution patterns rather than dealing with data noise.

[0085] 3. By introducing a dynamic conflict classification and grading confidence correction mechanism based on time density and physiological accessibility, the system successfully achieved intelligent identification of "errors" and "sudden illnesses" in monitoring data, avoiding the false cleansing of valid symptoms. In traditional outlier handling algorithms, data fluctuations exceeding a threshold are considered noise and removed or smoothed. However, in chronic disease management, drastic numerical changes are often the only early signal of acute events. Conflicts are determined by calculating the "rate of change" and the "maximum permissible rate of change," and further, the innovative concept of "physical accessibility markers" is introduced. By calculating the maximum physiologically permissible range of change within a time interval, the system can accurately distinguish between "data error-type conflicts" and "real change-type conflicts." This distinction has extremely high clinical value: for data error-type conflicts, the system performs automatic correction to ensure record continuity; while for real change-type conflicts, the system not only does not filter them but also retains and reinforces their marking as high-risk events. This advantage is further enhanced through a confidence evaluation system. The system not only assigns an initial confidence level to each record based on conflict type, but also performs secondary propagation and reinforcement through "neighborhood consistency judgment" and "average confidence level of semantic slots." Even if a data point is marked as low confidence due to large fluctuations, its confidence can still be restored to some extent through neighborhood propagation if its surrounding neighboring data is stable and shows a supporting trend; conversely, if an isolated high value appears amidst a sea of ​​low-value noise, its confidence will be further suppressed. This confidence calculation logic, similar to expert consultation, makes the health records output by the system no longer just a bunch of cold, impersonal numerical records, but a chain of medical evidence with "credibility" weights. For community doctors, they no longer need to spend a lot of energy distinguishing which data was entered by mistake and which data represents a real condition; they only need to focus on the semantic slots marked as "low confidence" but with "high physiological accessibility." This greatly reduces the workload of follow-up work in communities with limited medical resources and achieves effective allocation of precision medicine resources.

[0086] 4. By establishing a dynamic evolution model of individual health status based on semantic slot coupling strength and a community-based risk diffusion analysis mechanism, a leap from individual early warning to community-based prevention and control has been achieved, providing a data-driven decision support tool for grassroots public health management. This method does not simply provide a binary classification of "normal" or "abnormal," but rather constructs a multi-dimensional unified health record model by comprehensively calculating health status values, change rates, slot coupling strength, and semantic slot stability. In particular, the calculation of "coupling strength" and "coupling change" reveals whether the linkage between different physiological indicators is unbalanced. For example, an increase in the coupling between heart rate and blood pressure during increased exercise is normal, but if blood pressure rises while heart rate abnormally decreases, this decoupling phenomenon is often a sign of more insidious cardiovascular diseases. By quantifying this relationship, the system can detect potential systemic functional disorders earlier than single-indicator over-limit alarms. This system dynamics-based assessment perspective allows health management to evolve from "staring at individual numbers" to "assessing overall regulatory capacity." Secondly, by capturing the cumulative trend change magnitude and transition events, early detection of acute exacerbations of chronic diseases is achieved. Unlike static threshold alarms that only focus on absolute values, this invention calculates trend risk values, enabling it to capture hidden risks that "drift continuously in an unfavorable direction within the normal range." This sensitivity to trends is crucial for the long-term management of chronic diseases such as diabetes and hypertension, often securing valuable intervention windows. Finally, by extending these individual analytical capabilities to the community level, community health management agencies can monitor the macro-health status of chronic disease populations in their jurisdictions in real time by calculating the community's average risk, high-risk proportion, and risk dispersion. In particular, the introduction of "risk growth rate" and "risk diffusion indicators" gives the system the sensitivity of epidemic monitoring. When an abnormal increase in the proportion of high-risk users or a general decline in the health status of a population occurs in a certain area, the system can automatically trigger a community alert. This function has strong practical guiding significance for responding to peak seasons for seasonal diseases, the impact of extreme weather, or community environmental pollution events. It frees community doctors and administrators from tedious case statistics reports, allowing them to focus directly on high-risk clusters and groups with worsening health trends. This enables more efficient allocation of limited medical follow-up and health education resources, achieving an upgrade in public health service models from "passive treatment of individual cases" to "proactive management of community health." For elderly people in the community with various illnesses, the past reliance on manual record-keeping and classification—meaning someone with hypertension might be recorded separately, and someone with diabetes might be recorded separately—eliminates this redundant work. Furthermore, many diseases are interconnected; this system can automatically integrate multiple indicators to provide warnings of potential health risks, such as automatically identifying high-risk ASCVD patients. Attached Figure Description

[0087] Figure 1 This is a schematic diagram of the process of the present invention.

[0088] Figure 2 This is a schematic diagram of the system structure of the present invention.

[0089] Figure 3 This is the swimlane diagram for the timing decision of this invention.

[0090] Figure 4 Diagram of confidence propagation and evaluation mechanism. Detailed Implementation

[0091] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0092] Example 1, see Figures 1 to 4 A multimodal AI-driven community-based chronic disease health management method, comprising:

[0093] Multi-source data is mapped to a unified time grid and missing values ​​are filled in. Abnormal fluctuations are identified through interval stability checks.

[0094] Combining modality types with medical knowledge bases, semantic units are output from a large model. After consistency judgment and encoding, semantic sequences are constructed and transition anomalies are detected.

[0095] Construct semantic fingerprints to remove duplicate data, assign records to semantic slots to form a time chain, detect conflicts within the slots and evaluate stability to establish a health profile;

[0096] Analyze the rate of change and density of numerical values ​​within the tank, determine the type of conflict and correct data error-type conflicts, and output the conflict analysis results;

[0097] An initial confidence level is assigned to the semantic record. After conflict correction and neighborhood propagation, the enhanced confidence level is determined by combining the slot mean. The levels are then classified and outliers are marked.

[0098] Extract the status of each slot to construct a health vector, calculate the changes in health values ​​and the coupling strength between slots, classify the status according to the hierarchical threshold and mark the abnormal, and generate a unified health profile.

[0099] Track the direction and magnitude of health changes, determine the risk level by comprehensively considering trends, transitions and coupled risks, trigger immediate or continuous warnings and output disease sets;

[0100] By summarizing risk indicators of community users, identifying the status of risk spread, delineating high-risk groups and determining intervention priorities, and triggering individual or community early warnings.

[0101] The process of mapping multi-source data to a unified time grid and filling in missing values, and identifying abnormal fluctuations through interval stability checks, includes:

[0102] Define the set of data sources to be accessed, and associate each data record with a user identifier, collection time, and original measurement value;

[0103] Establish a unified time grid, with its nodes divided at fixed time intervals;

[0104] Map the acquisition time of each record to the nearest time grid node;

[0105] For each data source, the dimensional range is determined based on the maximum and minimum values ​​of its historical data, and the original measurement values ​​are converted into standardized values ​​accordingly.

[0106] For data missing at a certain point in time, the mean of the standardized values ​​existing before and after that point in time is used to fill the missing data.

[0107] Finally, a unified data matrix is ​​constructed with users, time nodes, and data sources as dimensions. The stability of the data is verified by calculating the standardized numerical difference between adjacent time nodes. If the difference exceeds the range allowed by the preset medical guidelines, the data point is marked as abnormal.

[0108] The process of combining modality types with medical knowledge base to output semantic units from a large model, constructing semantic sequences after consistency determination and encoding, and detecting abrupt transitions includes:

[0109] The modality type of each data source is identified by a preset mapping rule, including at least numerical, text, and image types;

[0110] Construct structured input units and introduce a knowledge base that stores medical rules related to chronic diseases;

[0111] Utilize large models to output predefined semantic labels based on input units;

[0112] Semantic tags are mapped to specific medical semantic entities and their corresponding attribute values, and the semantic relationship between them is determined based on the threshold range in medical guidelines, generating structured semantic units.

[0113] The consistency between modalities is verified by comparing the attribute values ​​of the same entity under different modalities.

[0114] Assign integer codes to entities, attribute values, and relations respectively, and perform combined calculations to generate unique semantic code values;

[0115] The semantic encoded values ​​of users are sorted by time to form a semantic sequence, and abnormal semantic jumps are identified by detecting abrupt changes in adjacent encoded values.

[0116] The process of constructing semantic fingerprints to remove duplicate data, assigning records to semantic slots to form a time chain, detecting intra-slot conflicts, and evaluating stability to establish a health profile includes:

[0117] The semantic unit is redefined as a semantic primitive composed of entities, attributes, and relations;

[0118] A medical thesaurus is used to normalize different entities expressing the same concept and unify their numbering.

[0119] A unique semantic fingerprint value is generated by combining normalized entity, attribute, relation and time encoding;

[0120] Identify and reject completely duplicate data by comparing differences in semantic fingerprint values;

[0121] Create an independent semantic slot for each type of normalized entity, and store the verified attribute values ​​into the corresponding slot in chronological order;

[0122] Conflicts are detected by calculating the differences between attribute values ​​at different time points within the same slot. If the difference exceeds the threshold specified in medical guidelines, it is marked.

[0123] Ultimately, a user's structured health profile is formed, consisting of each semantic slot and its temporal data chain.

[0124] The analysis analyzes the rate of change and density of numerical values ​​within the analysis tank, determines the type of conflict, corrects data error-type conflicts, and outputs the conflict analysis results, including:

[0125] Calculate the attribute value difference and time interval between adjacent records in the same semantic slot, and derive the rate of change accordingly;

[0126] The rate of change is compared with the maximum rate allowed by medical guidelines to initially determine a conflict;

[0127] A secondary verification is performed by combining the sampling time density and the range of physically reachable changes;

[0128] Based on the judgment results, the conflicts are divided into three types: conflicts caused by data errors, conflicts reflecting real physiological changes, and abnormal conflicts caused by excessively high sampling frequency.

[0129] For conflicts that are determined to be data errors, the average of the normal data before and after the conflict is used for correction;

[0130] The final output includes analysis results containing conflict flags, conflict types, physical reachability flags, and conflict zones.

[0131] The process of assigning an initial confidence level to semantic records, followed by conflict correction and neighborhood propagation, and then combining this with the slot mean to determine a reinforced confidence level, classifying levels and marking outliers, includes:

[0132] An initial confidence level is set for each semantic record based on whether it is conflicting and the type of conflict, and then corrected using a physical reachability flag;

[0133] The neighborhood consistency difference is calculated using the attribute values ​​of adjacent records. If the difference is less than a preset threshold, the average confidence of the record itself and its neighbors is taken as the confidence after propagation; otherwise, the corrected confidence is maintained.

[0134] Calculate the average confidence of all records within the same semantic slot, and combine it with the post-propagation confidence to obtain the final confidence.

[0135] Discrete levels are assigned based on the final confidence score to indicate the level of confidence.

[0136] The final output includes the final confidence value, confidence level, and anomaly confidence markers.

[0137] The process involves extracting the status of each slot to construct a health vector, calculating changes in health values ​​and inter-slot coupling strength, classifying states according to hierarchical thresholds and marking anomalies, and generating a unified health profile, including:

[0138] Define state information for each semantic slot, including its corresponding entity, the latest valid attribute value, the average confidence level of the slot, and the number of records.

[0139] The user's comprehensive health status vector is calculated based on the current attribute values ​​of each slot and their confidence weights.

[0140] Calculate the change and rate of change of the health status vector, and classify the health status into different levels according to the medical grading standard;

[0141] By constructing coupling pairs between different semantic slots and calculating the change in their attribute differences, the linkage between indicators can be monitored.

[0142] If the change in health status or level exceeds a preset threshold, an anomaly is marked.

[0143] Finally, the above information is integrated to generate a unified health record model.

[0144] The system tracks the direction and magnitude of health changes, determines risk levels by considering trends, transitions, and coupled risks, triggers immediate or continuous warnings, and outputs a set of diseases, including:

[0145] Analyze the direction of change in health status, the length of continuous unidirectional changes, and the cumulative magnitude to calculate the trend risk value;

[0146] Detect whether a health level has transitioned, and generate a transition risk value;

[0147] The number of times the change in coupling relationship exceeds a threshold is counted to calculate the coupling risk value;

[0148] The trend risk value, transition risk value, and coupling risk value are combined to obtain the comprehensive risk value;

[0149] Risk levels are determined based on the comparison between the comprehensive risk value and a preset threshold, and different levels of warnings are triggered accordingly.

[0150] This is achieved by matching the user's set of symptoms with a pre-established set of disease symptoms based on medical guidelines.

[0151] The process of summarizing community user risk calculation group indicators, identifying the risk spread status, delineating high-risk groups and determining intervention priorities, and triggering individual or community early warnings includes:

[0152] Build a community user set and obtain the individual risk value for each user;

[0153] Based on this, we can calculate the overall average risk level of the community, the proportion of high-risk users, and the dispersion of risk distribution.

[0154] By monitoring the growth rate of average risk in the community, it can be determined whether there is a risk diffusion trend;

[0155] Identify and generate sets of high-risk users and sets of users with ongoing risks;

[0156] The priority of individual interventions is determined based on the user's overall risk score and the number of historical warnings.

[0157] When an individual's risk level reaches a preset standard, an individual warning is triggered and information about their possible diseases is pushed to them.

[0158] When the proportion of high-risk users in a community exceeds a preset threshold, a community-level group alert is triggered.

[0159] Example 2: A system for implementing the community chronic disease health management method driven by the multimodal AI large model, comprising:

[0160] The data is fed into the preprocessing module, which performs unified time alignment, normalization and missing data completion on multi-source heterogeneous health data, and outputs a structured standard data matrix to provide basic input for subsequent semantic analysis.

[0161] The multimodal semantic parsing module is used to input standardized data into a large multimodal model, generate structured semantic information under the constraints of medical knowledge, and perform consistency and abnormal semantic detection.

[0162] The semantic deduplication and slot management module is used to eliminate semantic redundancy and duplicate data, and organize semantic information into a traceable time-series slot structure by entity;

[0163] The conflict identification and data correction module is used to identify abnormal changes, sampling anomalies, and physical inaccessibility in the data, and to automatically correct erroneous data.

[0164] The confidence assessment module is used to evaluate the credibility of each piece of semantic data and improve the overall stability through a neighborhood propagation mechanism.

[0165] The health status modeling module is used to integrate information from multiple indicators to build a dynamic health status model for users.

[0166] The risk assessment and early warning module is used to comprehensively assess user health risks and trigger early warnings from the perspectives of trends, mutations, and multi-indicator coupling.

[0167] Example 3: A community-based chronic disease health management method driven by a multimodal AI large model, comprising:

[0168] Multi-source data is mapped to a unified time grid and missing values ​​are filled in. Abnormal fluctuations are identified through interval stability checks.

[0169] Combining modality types with medical knowledge bases, semantic units are output from a large model. After consistency judgment and encoding, semantic sequences are constructed and transition anomalies are detected.

[0170] Construct semantic fingerprints to remove duplicate data, assign records to semantic slots to form a time chain, detect conflicts within the slots and evaluate stability to establish a health profile;

[0171] Analyze the rate of change and density of numerical values ​​within the tank, determine the type of conflict and correct data error-type conflicts, and output the conflict analysis results;

[0172] An initial confidence level is assigned to the semantic record. After conflict correction and neighborhood propagation, the enhanced confidence level is determined by combining the slot mean. The levels are then classified and outliers are marked.

[0173] Extract the status of each slot to construct a health vector, calculate the changes in health values ​​and the coupling strength between slots, classify the status according to the hierarchical threshold and mark the abnormal, and generate a unified health profile.

[0174] Track the direction and magnitude of health changes, determine the risk level by comprehensively considering trends, transitions and coupled risks, trigger immediate or continuous warnings and output disease sets;

[0175] By summarizing risk indicators of community users, identifying the status of risk spread, delineating high-risk groups and determining intervention priorities, and triggering individual or community early warnings.

[0176] The process of mapping multi-source data to a unified time grid and filling in missing values, and identifying abnormal fluctuations through interval stability checks, includes:

[0177] The set of data sources accessed by the system is as follows: ;

[0178] in, Represents a collection of data sources. Indicates the first One data source, Indicates the data source index number. Indicates the total number of data sources;

[0179] For any data source Define its output data as: ;

[0180] in, Indicates the first The first data source 1 data record This represents a unique user identifier, generated during system registration. Indicates the collection time. Represents the original measurement value. This represents a data sequence index, generated by sorting by time.

[0181] Define a unified time grid: ;

[0182] in, Represents a unified time grid set. Indicates a time index. This represents the total number of time points. Indicates the first Each time point, specifically expressed as: ;

[0183] in, Indicates the start time. Indicates the time interval for data collection;

[0184] Collection time Mapped to the nearest time node:

[0185] ;

[0186] in, This represents the index of the mapped time node. This indicates the index corresponding to the minimum difference.

[0187] For each data source Set its dimensional range as ;

[0188] in, Indicates the first Minimum measurable value of each data source Indicates the first Maximum measurable value for each data source Historical data statistics can be used to analyze: ;

[0189] Define standardized values:

[0190] ;

[0191] in, Represents standardized data, with a range of ;

[0192] If time nodes exist In the data source If there is no corresponding data, then define ;

[0193] Otherwise, define ;

[0194] in, This indicates that the data exists.

[0195] for The data source calculates its corresponding fill value. :

[0196] ;

[0197] in, Indicates data source At the time point The time index of the previous existing data. Indicates data source At the time point The time index of the next existing data;

[0198] Constructing a unified data matrix: ;

[0199] in, Indicates user At the time point From data source The standardized value;

[0200] Construct an interval stability check algorithm:

[0201] ;

[0202] in, Indicates the difference between adjacent time points;

[0203] like Then mark the user At the time point From data source The standardized values ​​are abnormal;

[0204] in, This indicates the maximum permissible range of variation, provided by medical guidelines.

[0205] By constructing a standardized mechanism for unified processing and fusion of multi-source heterogeneous data, this invention completely solves the problems of data silos and incomparability caused by the numerous brands of testing equipment, different data formats, and inconsistent collection frequencies in community chronic disease management. Through defining a unified time grid, performing standardized value mapping, and intelligently filling in missing data, the method successfully transforms raw measurements from different data sources into a standardized data matrix with unified physical meaning and mathematical dimensions. This significantly improves data usability. In real-world community applications, users may alternate between home self-testing devices and shared equipment at community health service centers. System errors and dimensional differences between devices often make it difficult to analyze data trends consistently. This invention sets an independent dimensional range for each data source and performs normalization calculations, compressing the raw measurements into a unified closed interval, thereby eliminating data bias caused by differences in device hardware. This allows even historical data recorded by different models of equipment to be compared and analyzed equally within the same mathematical framework, providing a clean and consistent input foundation for subsequent large-scale artificial intelligence model analysis. Furthermore, the time grid mapping and data filling strategies in this method effectively address the pain points of sparsity and misalignment in time-series data. In real-world community management, users don't strictly adhere to precise second-level intervals when uploading data, and data uploads are often interrupted or missed. Mapping the collection time to the nearest available time node and using linear interpolation to fill in missing data nodes ensures the mathematical integrity of the time series. This prevents errors in extracting time features from AI models due to timestamp misalignment, ensuring the accuracy of trend calculations. More importantly, the defined interval stability check algorithm can screen for abnormal fluctuations immediately after data fusion. Its maximum allowable variation range is set according to medical guidelines. This allows the system to not only complete data fusion but also simultaneously perform a preliminary review of data rationality based on physiological principles, effectively filtering out false anomalies caused by momentary equipment interference or operational errors, ensuring data quality from the source for large-scale model analysis. This proactive data governance capability is unparalleled by traditional manual entry or simple aggregation systems.

[0206] The process of combining modality types with medical knowledge base to output semantic units from a large model, constructing semantic sequences after consistency determination and encoding, and detecting abrupt transitions includes:

[0207] Define the modal identification function: ;

[0208] in, Indicates data source Modal type, This represents the modal mapping function, specifically a predefined mapping function, determined by the data source type identifier;

[0209] The modal types include: Indicates numeric type. Represents text type. Indicates image type;

[0210] Define a unified input structure: ;

[0211] in, Represents a standardized input unit;

[0212] Constructing large model input vectors : ;

[0213] in, The medical knowledge base, or structured rule base, is used to store standardized medical rule information related to various chronic diseases. Its content includes: medical entity definitions corresponding to various health indicators, graded standard intervals corresponding to indicators, semantic mapping relationship between indicators and diseases, and a set of labels and judgment criteria used to constrain the output of large models. In essence, it is a queryable set of structured rules used to establish a deterministic correspondence between input data and medical standards.

[0214] Define the output of the large model: ;

[0215] in, Indicates the semantic trigger result. This represents a set of semantic tags, consisting of predefined medical tags, specifically: , Indicates the first A semantic tag;

[0216] Define semantic entities : ;

[0217] in, This represents a semantic tag-to-entity mapping function, which is a tag-entity mapping function based on a medical knowledge base and is implemented through a table lookup method. For example, "hypertension" → entity: blood pressure status;

[0218] Extract attribute values : ;

[0219] Define relational functions: ;

[0220] in, Indicates semantic relations. This represents a relation determination function used to determine whether an attribute value falls within a threshold range. , Indicates the first Threshold ranges for semantic labels, derived from medical guidelines, such as blood pressure grading standards;

[0221] Generate structured semantic units : ;

[0222] Calculate modal consistency: ;

[0223] in, Indicates the difference between modes. This represents the attribute values ​​of the same entity corresponding to different modalities;

[0224] Based on intermodal differences Perform a consistency check: ;

[0225] If the consistency criteria are met, a consistency flag is triggered. ;

[0226] Otherwise, an inconsistency flag is triggered. ;

[0227] in, This indicates the permissible error range, obtained from medical standards. Indicates a semantic conflict marker;

[0228] Set entity code : ;

[0229] Integer encoding refers to representing a semantic entity with a unique integer number, which is obtained by direct mapping from the medical coding table.

[0230] The medical coding table is a structured mapping table that predefines various medical entities and assigns them unique integer numbers one by one;

[0231] Based on attribute value Computed attribute encoding: ;

[0232] in, Represents discrete values ​​of an attribute;

[0233] Set relationship code : ;

[0234] Perform semantic encoding combination:

[0235] ;

[0236] in, This represents a semantically unique encoded value;

[0237] Constructing the semantic structure of time series: ;

[0238] in, Represents a user semantic sequence;

[0239] Constructing a semantic jump detection algorithm: ;

[0240] in, Indicates the amount of semantic encoding change;

[0241] like If so, the semantic encoding change of the marker will result in an abnormal semantic jump;

[0242] in, This indicates the threshold for determining semantic jumps, which is set based on historical data.

[0243] The process of constructing semantic fingerprints to remove duplicate data, assigning records to semantic slots to form a time chain, detecting intra-slot conflicts, and evaluating stability to establish a health profile includes:

[0244] Define semantic primitives Used for renaming to avoid duplicates: ;

[0245] in, Represents the entity primitive, that is , Represents the attribute primitive, that is , Represents relation primitives, i.e. ;

[0246] Construct an entity normalization function to eliminate semantic-level "synonymous but different encodings": ;

[0247] in, Indicates the normalized entity number. This represents a set of semantically equivalent entities, constructed using a medical thesaurus. For example, "blood pressure value" and "blood pressure measurement value" are grouped into the same category.

[0248] The medical thesaurus is specifically constructed by professionals who extract terms from a standard medical terminology set and use "corresponding to the same physiological indicator and having the same measurement method and consistent numerical judgment range" as the judgment condition, and then merge different expressions that meet the condition one by one.

[0249] Define time encoding: ;

[0250] in, Represents time-discrete coding;

[0251] Constructing semantic fingerprints:

[0252] ;

[0253] in, Represents semantic fingerprint values;

[0254] Calculate fingerprint differences: ;

[0255] in, This represents the difference between two semantic fingerprint values;

[0256] like If so, it is determined to be completely duplicate data;

[0257] Construct a semantic slot set: ;

[0258] in, Indicates the first Each semantic slot corresponds to one entity semantic. Indicates the number of semantic entity categories;

[0259] Perform slot mapping: ;

[0260] Constructing the insertion function:

[0261] ;

[0262] in, Indicates the insertion mark, when Indicates that write is allowed, when This indicates a refusal to write;

[0263] The semantic record unit is constructed as follows: ;

[0264] in, Indicates the first Semantic record, Indicates in semantic slot Sort by time Each attribute value , : This indicates that the time-sorted mapping is used to generate a sequence index, which is generated by sorting the same semantic slots in ascending order of time;

[0265] Construct a time series chain for the slots: ;

[0266] in, Indicates the first An ordered set formed by arranging all semantic records in each semantic slot in chronological order. Indicates the number of semantic records;

[0267] Constructing the conflict function: ;

[0268] in, This represents the difference between different attribute primitives within the same semantic slot.

[0269] like Then the tag will be a semantic slot. Inner Semantic record and the first A record pair consisting of semantic records Mark as conflicting record pairs;

[0270] in, This represents the threshold for the maximum attribute difference, determined according to medical guidelines.

[0271] Building a health record: ;

[0272] in, This represents the user's health record;

[0273] Constructing an in-slot stability detection algorithm:

[0274] ;

[0275] in, This indicates the stability of the slot.

[0276] By establishing a modality adaptive mapping and semantic structured coding system, a deterministic leap from low-level physiological values ​​to high-level medical logic is achieved, greatly enhancing the interpretability and accuracy of multimodal large-scale models in chronic disease management. The paper details how to unify the transformation of various modalities such as numerical, textual, and image data into standardized semantic units, and then construct a rigorously structured user health profile through semantic fingerprinting and slot mapping. It solves the "illusion" problem and logical confusion commonly encountered by general-purpose large-scale models when processing specialized medical data. By introducing a predefined medical knowledge base and semantic label mapping function, abstract measurement values ​​are forcibly bound to specific medical semantic entities and their corresponding hierarchical relationships. This forced mapping ensures that the output of the AI ​​large-scale model is not a textual description based on probabilistic guesswork, but rather a deterministic medical guideline judgment. When multiple modalities simultaneously point to the same semantic entity, the modality consistency verification mechanism can automatically detect potential contradictions between numerical indicators and subjective descriptions and trigger conflict markers. This mechanism is crucial for community-based chronic disease management because it alerts healthcare professionals to hidden dangers where "data appears normal but patients complain of severe symptoms," effectively preventing missed diagnoses from relying solely on numerical monitoring. Furthermore, by constructing semantic primitives, semantic fingerprints, and semantic slot sets, complex multimodal time-series data is highly structured into slot-based time-series chains. This innovative data organization method brings significant improvements in storage and retrieval efficiency. By calculating semantic fingerprint dissimilarity, the system can identify and reject completely duplicated redundant data with extremely low computational cost, greatly reducing storage pressure. Simultaneously, stability detection and conflict functions within slots automatically identify and mark severe contradictions occurring in the same semantic slot within a short period before data is entered into the database. This structured preprocessing allows subsequent large models to avoid time-consuming pattern recognition in massive amounts of unstructured raw data, instead performing inference and prediction directly on clean, standard semantic record units. This not only accelerates response time but, more importantly, elevates the role of large models from "data cleaners" to true "clinical reasoning experts," enabling them to focus on uncovering deep health evolution patterns rather than dealing with data noise.

[0277] The analysis analyzes the rate of change and density of numerical values ​​within the analysis tank, determines the type of conflict, corrects data error-type conflicts, and outputs the conflict analysis results, including:

[0278] Calculate adjacent differences :

[0279] ;

[0280] Calculate time interval : ;

[0281] Calculate the rate of change of the difference between adjacent attribute primitives in the same slot. : ;

[0282] Set the maximum allowed rate of change. Obtained from medical guidelines;

[0283] like If it is, then it is marked as a conflict point, that is ,

[0284] Otherwise, it is marked as a non-conflict point, i.e. ;

[0285] in, Indicates a conflict symbol;

[0286] Calculate local time density : ;

[0287] Set density threshold It is determined based on the system sampling interval;

[0288] like If so, it is determined to be a high-density sampling anomaly;

[0289] Calculate the achievable range of variation: ;

[0290] in, This indicates the maximum permissible change within the time interval;

[0291] like If so, it is determined to be a data error and ;

[0292] Otherwise, the data is considered normal and ;

[0293] in, This indicates a physical reachability flag, where 1 indicates reachability and 0 indicates unreachability.

[0294] If the following conditions are met: If so, then the data error is set as a data error type conflict, that is... , Indicates the conflict type identifier;

[0295] If the following conditions are met: If so, then the data error is set to a real change type conflict, that is... ;

[0296] If the following conditions are met: If so, the data error is set to a sampling anomaly conflict, i.e. ;

[0297] Construct conflict intervals : ;

[0298] The conflict zone This refers to the conflict flag corresponding to any index in the set. All are 1, and if there is a position adjacent to its index outside this set, the conflict flag corresponding to that position is 0;

[0299] when At that time, perform data correction:

[0300] .

[0301] Output conflict analysis results : .

[0302] The process of assigning an initial confidence level to semantic records, followed by conflict correction and neighborhood propagation, and then combining this with the slot mean to determine a reinforced confidence level, classifying levels and marking outliers, includes:

[0303] Define the initial confidence level for each semantic record. ;

[0304] The specific rules for determining the value are as follows:

[0305] like Then set ;

[0306] like Then set ;

[0307] like Then set ;

[0308] like Then set ;

[0309] Initial confidence level Perform reachability correction:

[0310] ;

[0311] in, Indicates the corrected confidence level;

[0312] Construct a neighborhood index set: ;

[0313] in, Indicates the first A set of neighborhood indexes for each semantic record;

[0314] Perform neighborhood consistency determination:

[0315] ;

[0316] in, This represents the neighborhood consistency difference;

[0317] Set consistency threshold This is obtained from statistics on the normal fluctuation range in medicine;

[0318] like Then set the confidence level after propagation:

[0319] ;

[0320] Otherwise, set the confidence level after propagation to 1. ;

[0321] in, Indicates the confidence level after propagation;

[0322] Calculate the average confidence of semantic slots :

[0323] ;

[0324] Calculate the final slot reinforcement confidence level :

[0325] ;

[0326] Perform confidence level discrete classification:

[0327] ;

[0328] in, Indicates the confidence level: High = 2, Medium = 1, Low = 0;

[0329] like If so, the semantic record is marked as an anomalous confidence point, that is... ;

[0330] Otherwise, mark the semantic record as a normal confidence point, i.e. ;

[0331] in, Indicates an anomaly confidence flag;

[0332] Constructing confidence time series: ;

[0333] Output confidence results : .

[0334] By introducing a dynamic conflict classification and grading confidence correction mechanism based on time density and physiological accessibility, the system successfully achieved intelligent identification of "errors" and "sudden illnesses" in monitoring data, avoiding the false cleansing of valid symptoms. In traditional outlier handling algorithms, data fluctuations exceeding a threshold are considered noise and removed or smoothed. However, in chronic disease management, drastic numerical changes are often the only early signal of acute events. The system determines conflicts by calculating the "rate of change" and the "maximum permissible rate of change," and further introduces the innovative concept of "physical accessibility markers." By calculating the maximum physiologically permissible range of change within a time interval, the system can accurately distinguish between "data error-type conflicts" and "real change-type conflicts." This distinction has high clinical value: for data error-type conflicts, the system performs automatic correction to ensure record continuity; while for real change-type conflicts, the system not only does not filter them but also retains and reinforces their markers as high-risk events. This advantage is further enhanced through a confidence evaluation system. The system not only assigns an initial confidence level to each record based on conflict type, but also performs secondary propagation and reinforcement through "neighborhood consistency judgment" and "average confidence level of semantic slots." Even if a data point is marked as low confidence due to large fluctuations, its confidence can still be restored to some extent through neighborhood propagation if its surrounding neighboring data is stable and shows a supporting trend; conversely, if an isolated high value appears amidst a sea of ​​low-value noise, its confidence will be further suppressed. This confidence calculation logic, similar to expert consultation, makes the health records output by the system no longer just a bunch of cold, impersonal numerical records, but a chain of medical evidence with "credibility" weights. For community doctors, they no longer need to spend a lot of energy distinguishing which data was entered by mistake and which data represents a real condition; they only need to focus on the semantic slots marked as "low confidence" but with "high physiological accessibility." This greatly reduces the workload of follow-up work in communities with limited medical resources and achieves effective allocation of precision medicine resources.

[0335] The process involves extracting the status of each slot to construct a health vector, calculating changes in health values ​​and inter-slot coupling strength, classifying states according to hierarchical thresholds and marking anomalies, and generating a unified health profile, including:

[0336] For each semantic slot Define state elements:

[0337] ;

[0338] in, Indicates the first Semantic slot state This indicates the entity identifier of the slot (from step 3). This indicates the number of semantic records within a semantic slot. This indicates the currently valid attribute value, specifically: , The attribute value representing the latest recorded time;

[0339] Constructing a health vector:

[0340] ;

[0341] in, Indicates at a point in time The following health status value, Indicates semantic slots At the point of time The attribute value, Indicates semantic slots At the point of time Confidence level, Indicates the total number of slots;

[0342] Calculate the change in health status : ;

[0343] Calculate the rate of change in health status :

[0344] ;

[0345] Constructing coupling pairs: ;

[0346] in, Indicates semantic slot combination;

[0347] Calculate coupling strength : ;

[0348] Calculate the coupling change : ;

[0349] To avoid continuous unstable values, state division is performed:

[0350] ;

[0351] in, Indicates health level. These represent the health level thresholds, which are set according to the medical health grading standards.

[0352] like Then mark It is abnormal, that is ;

[0353] in, Indicates an anomaly marker. The threshold for changes in health status is determined by statistical analysis of historical data.

[0354] Build a unified health record : ;

[0355] Output a unified health record model: .

[0356] The system tracks the direction and magnitude of health changes, determines risk levels by considering trends, transitions, and coupled risks, triggers immediate or continuous warnings, and outputs a set of diseases, including:

[0357] Marking changes in health status Direction of change:

[0358] ;

[0359] in, Indicates the direction of change. Indicates an increase, Indicates stability. Indicates a decrease;

[0360] Define the length that changes continuously in the same direction :

[0361] ;

[0362] Calculate the strength of the trend:

[0363] ;

[0364] in, Indicates the cumulative change in the trend;

[0365] Calculate the trend risk value : ;

[0366] Calculate changes in health status : ;

[0367] like If so, it is determined to be a jump event;

[0368] Set jump risk value ;

[0369] Count of abnormal couplings : ;

[0370] in, This indicates an indicator function, which is 1 if it meets the condition, and 0 otherwise. The threshold for coupling changes is obtained from historical health indicators.

[0371] Set coupling risk value ;

[0372] Perform risk synthesis:

[0373] ;

[0374] in, This represents the overall risk value;

[0375] Construct a risk range:

[0376] ;

[0377] in, Indicates the risk level. This represents the risk threshold, obtained from statistical analysis of historical clinical data.

[0378] like Then set an immediate warning, that is ;

[0379] If the following conditions are met consecutively: Then set a continuous alert, that is ;

[0380] Otherwise, set it not to issue a warning, i.e. ;

[0381] in, Indicates a warning sign;

[0382] Output risk analysis results ;

[0383] Define disease set : ;

[0384] For each disease Define its symptom set : ;

[0385] in, Indicates the first a disease, Indicates the number of diseases. express A collection of symptoms Indicates the first One symptom, The number of symptoms is indicated, obtained from medical guidelines.

[0386] Constructing disease sets through medical guidelines ;

[0387] Extracting disease sets from medical guidelines The symptoms corresponding to each disease in the Chinese text;

[0388] Obtain the user's symptom set after obtaining the user's consent: ;

[0389] in, This indicates that the user already has a set of symptoms. Indicates the first One symptom;

[0390] Define the symptom matching function:

[0391] ;

[0392] in, Symptom matching indicator: 1 indicates a match, 0 indicates a mismatch. Indicates disease The corresponding number One symptom;

[0393] Determine whether a disease has been diagnosed:

[0394] ;

[0395] in, The hallmark of disease establishment Indicates disease The total number of corresponding symptoms is obtained from medical guidelines. This means that disease c is only considered valid if all symptoms are met.

[0396] Output the user's set of diseases: .

[0397] The process of summarizing community user risk calculation group indicators, identifying the risk spread status, delineating high-risk groups and determining intervention priorities, and triggering individual or community early warnings includes:

[0398] Building a community user base: ;

[0399] in, Represents a collection of community users. Indicates the first One user, Indicates the total number of users;

[0400] Perform user risk mapping ;

[0401] in, Indicates user In time The risk value;

[0402] Calculate the average risk of the community :

[0403] ;

[0404] Calculate the high-risk ratio : ;

[0405] in, This indicates the number of high-risk users, i.e., those that meet the following criteria. The number of users;

[0406] Calculate the risk dispersion:

[0407] ;

[0408] in, Indicates the dispersion of risk distribution;

[0409] Calculate the risk growth rate : ;

[0410] like If so, it is determined to be a risk diffusion, that is... ;

[0411] Otherwise, it is determined to be non-risk diffusion, i.e. ;

[0412] in, Indicates a risk spread indicator. This represents the growth threshold, obtained from historical data statistics;

[0413] Constructing a high-risk set : ;

[0414] Constructing a continuous risk set : ;

[0415] Define intervention priorities : ;

[0416] in, Indicates the number of historical alerts for the user;

[0417] like Then, an individual warning will be issued, that is... and output the set of diseases corresponding to the user. ;

[0418] Community early warning determination:

[0419] ;

[0420] in, Community early warning threshold This indicates that a community warning has been issued. At that time, no community warning was issued. This indicates the high-risk ratio threshold, which is determined by historical community risk statistics.

[0421] By establishing a dynamic evolution model of individual health status based on semantic slot coupling strength and a community-based risk diffusion analysis mechanism, a leap from individual early warning to community-based prevention and control has been achieved, providing a data-driven decision support tool for grassroots public health management. This method does not simply provide a binary classification of "normal" or "abnormal," but rather constructs a multi-dimensional unified health record model by comprehensively calculating health status values, change rates, slot coupling strength, and semantic slot stability. In particular, the calculation of "coupling strength" and "coupling change" reveals whether the linkage between different physiological indicators is unbalanced. For example, an increase in the coupling between heart rate and blood pressure during increased exercise is normal, but if blood pressure rises while heart rate abnormally decreases, this decoupling phenomenon is often a sign of more insidious cardiovascular diseases. By quantifying this relationship, the system can detect potential systemic functional disorders earlier than single-indicator over-limit alarms. This system dynamics-based assessment perspective allows health management to evolve from "staring at individual numbers" to "assessing overall regulatory capacity." Secondly, by capturing the cumulative trend change magnitude and transition events, early detection of acute exacerbations of chronic diseases is achieved. Unlike static threshold alarms that only focus on absolute values, this invention calculates trend risk values, enabling it to capture hidden risks that "drift continuously in an unfavorable direction within the normal range." This sensitivity to trends is crucial for the long-term management of chronic diseases such as diabetes and hypertension, often securing valuable intervention windows. Finally, by extending these individual analytical capabilities to the community level, community health management agencies can monitor the macro-health status of chronic disease populations in their jurisdictions in real time by calculating the community's average risk, high-risk proportion, and risk dispersion. In particular, the introduction of "risk growth rate" and "risk diffusion indicators" gives the system the sensitivity of epidemic monitoring. When an abnormal increase in the proportion of high-risk users or a general decline in the health status of a population occurs in a certain area, the system can automatically trigger a community alert. This function has strong practical guiding significance for responding to peak seasons for seasonal diseases, the impact of extreme weather, or community environmental pollution events. It frees community doctors and administrators from tedious case statistics reports, allowing them to focus directly on high-risk clusters and groups with worsening health trends. This enables more efficient allocation of limited medical follow-up and health education resources, achieving an upgrade in public health service models from "passive treatment of individual cases" to "proactive management of community health." For elderly people in the community with various illnesses, the past reliance on manual record-keeping and classification—meaning someone with hypertension might be recorded separately, and someone with diabetes might be recorded separately—eliminates this redundant work. Furthermore, many diseases are interconnected; this system can automatically integrate multiple indicators to provide warnings of potential health risks, such as automatically identifying high-risk ASCVD patients.

[0422] In the technical solution disclosed in this invention, the various judgment markers defined and generated, including but not limited to data anomaly markers, semantic jump markers, conflict type markers, confidence anomaly markers, status level markers, and risk warning markers, are not isolated or merely intermediate variables in the process description, but constitute a hierarchical and progressive data quality control and decision support constraint system that runs through the entire process. The role mechanism and design intent of the above-mentioned marker variables in this invention are specifically reflected in the following aspects: (1) as implicit constraints for data screening and semantic construction. In the data processing and semantic modeling stage, various marker variables implicitly determine whether data enters the next calculation stage through rule mapping or logical judgment. For example, data existence markers Used to control the execution of the missing value imputation algorithm; anomaly markers in interval stability checks are used to eliminate the interference of abrupt changes in the normalized matrix; semantic conflict markers. These markers are used to indicate inconsistencies between multimodal data, thereby triggering corresponding deduplication or conflict resolution logic when constructing semantic fingerprints or health profiles. Although such markers do not appear repeatedly in every subsequent formula in the form of explicit mathematical symbols, their judgment results have substantially participated in the calculation of subsequent steps by controlling the program execution branches or data flow, and are implicit technical features that can be determined without doubt by those skilled in the art.

[0423] (2) As a weighting adjustment factor for confidence assessment and state calculation. In the process of confidence propagation and health state vector synthesis, the aforementioned labeling indirectly but necessarily changes the final health record model and risk value calculation results by affecting the confidence value or the range of state values. For example, records labeled as "data error type conflict" are assigned zero confidence, thereby affecting the subsequent health vector calculation results. The contribution is zero in the calculation, thus automatically masking outlier data. This implicit, rule-based weight adjustment mechanism is the core means by which this invention achieves automatic data correction and noise reduction, and it relies on the various labeled variables defined above.

[0424] (3) As a cumulative effect factor in risk synthesis and early warning decision-making. In the risk analysis stage, although the comprehensive risk value The direct components of risk are trend risk, transition risk, and coupling risk, but the underlying data constituting these risk items are all selected or calculated under the constraints of the aforementioned labeled variables. For example, records marked as abnormal confidence points will not be included in the calculation of the rate of change in health status; data points marked as high-density sampling anomalies will not participate in the accumulation of trend strength. Therefore, labeled variables indirectly improve the accuracy and sensitivity of risk assessment through cumulative effects, and their existence is a necessary prerequisite for ensuring the reliability of risk analysis results.

[0425] (4) As a traceable basis for the interpretability of system output. When generating individual early warnings or community intervention suggestions at the application layer, the system can explain the source of risk to community doctors or managers based on the stored labeled information. These labeled variables, as traceable metadata, significantly enhance the transparency and credibility of the AI ​​large model output, and are a key technical means to solve the AI ​​"black box" problem in the medical and health field. In summary, the various labeled variables defined in this invention together constitute a unified constraint and interpretation framework that runs through the entire process of data processing, semantic parsing, state modeling, and risk decision-making. Its operation includes both explicit mathematical substitution and implicit logical control and data filtering that can be clearly deduced by those skilled in the art based on the context. This design avoids redundant calculations and improves the stability and interpretability of the system when it is implemented in engineering, and is not a redundant expression without practical meaning.

[0426] 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.

[0427] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A community-based chronic disease health management method driven by a multimodal AI large model, characterized in that, include: Multi-source data is mapped to a unified time grid and missing values ​​are filled. Abnormal fluctuations are identified through interval stability verification. The modal types of the multi-source data include at least numerical, textual, and image types. Combining modality types with medical knowledge bases, semantic units are output from a large model. After consistency determination and encoding, a semantic sequence is constructed and abrupt transition anomalies are detected. Specifically: The modality type of each data source is identified by a preset mapping rule, including at least numerical, text, and image types; Construct structured input units and introduce a knowledge base that stores medical rules related to chronic diseases; Utilize large models to output predefined semantic labels based on input units; Semantic tags are mapped to specific medical semantic entities and their corresponding attribute values, and the semantic relationship between them is determined based on the threshold range in medical guidelines, generating structured semantic units. The consistency between modalities is verified by comparing the attribute values ​​of the same entity under different modalities. Assign integer codes to entities, attribute values, and relations respectively, and perform combined calculations to generate unique semantic code values; The user's semantic encoded values ​​are sorted by time to form a semantic sequence, and abnormal semantic jumps are identified by detecting abrupt changes in adjacent encoded values. A semantic fingerprint is constructed to remove duplicate data, records are assigned to semantic slots to form a time chain, intra-slot conflicts are detected and stability is evaluated to establish a health profile, specifically as follows: The semantic unit is redefined as a semantic primitive composed of entities, attributes, and relations; A medical thesaurus is used to normalize different entities expressing the same concept and unify their numbering. A unique semantic fingerprint value is generated by combining normalized entity, attribute, relation and time encoding; Identify and reject completely duplicate data by comparing differences in semantic fingerprint values; Create an independent semantic slot for each type of normalized entity, and store the validated attribute values ​​into the corresponding slot in chronological order; Conflicts are detected by calculating the differences between attribute values ​​at different time points within the same slot. If the difference exceeds the threshold specified in medical guidelines, it is marked. Ultimately, a user's structured health profile is formed, consisting of each semantic slot and its temporal data chain; Analyze the rate of change and density of numerical values ​​within the tank, determine the type of conflict and correct data error-type conflicts, and output the conflict analysis results; An initial confidence level is assigned to the semantic record. After conflict correction and neighborhood propagation, the enhanced confidence level is determined by combining the slot mean. The levels are then classified and outliers are marked. Extract the status of each slot to construct a health vector, calculate the changes in health values ​​and the coupling strength between slots, classify the status according to the hierarchical threshold and mark the abnormal, and generate a unified health profile. Track the direction and magnitude of health changes, determine the risk level by comprehensively considering trends, transitions and coupled risks, trigger immediate or continuous warnings and output disease sets; By summarizing risk indicators of community users, identifying the status of risk spread, delineating high-risk groups and determining intervention priorities, and triggering individual or community early warnings.

2. The community chronic disease health management method driven by a multimodal AI large model according to claim 1, characterized in that, The process of mapping multi-source data to a unified time grid and filling in missing values, and identifying abnormal fluctuations through interval stability checks, includes: Define the set of data sources to be accessed, and associate each data record with a user identifier, collection time, and original measurement value; Establish a unified time grid, with nodes divided at fixed time intervals; Map the acquisition time of each record to the nearest time grid node; For each data source, the dimensional range is determined based on the maximum and minimum values ​​of its historical data, and the original measurement values ​​are converted into standardized values ​​accordingly. For data missing at a certain point in time, the mean of the standardized values ​​existing before and after that point in time is used to fill the missing data. Finally, a unified data matrix is ​​constructed with users, time nodes, and data sources as dimensions. The stability of the data is verified by calculating the standardized numerical difference between adjacent time nodes. If the difference exceeds the range allowed by the preset medical guidelines, the standardized value at the time node corresponding to the standardized numerical difference is marked as abnormal.

3. The community chronic disease health management method driven by a multimodal AI large model according to claim 1, characterized in that, The analysis analyzes the rate of change and density of numerical values ​​within the analysis tank, determines the type of conflict, corrects data error-type conflicts, and outputs the conflict analysis results, including: Calculate the attribute value difference and time interval between adjacent records in the same semantic slot, and derive the rate of change accordingly; The rate of change is compared with the maximum rate allowed by medical guidelines to initially determine a conflict; A secondary verification is performed by combining the sampling time density and the range of physically reachable changes; Based on the judgment results, the conflicts are divided into three types: conflicts caused by data errors, conflicts reflecting real physiological changes, and abnormal conflicts caused by excessively high sampling frequency. For conflicts that are determined to be data errors, the average of the normal data before and after the conflict is used for correction; The final output includes analysis results containing conflict flags, conflict types, physical reachability flags, and conflict zones.

4. The community chronic disease health management method driven by a multimodal AI large model according to claim 1, characterized in that, The process of assigning an initial confidence level to semantic records, followed by conflict correction and neighborhood propagation, and then combining this with the slot mean to determine a reinforced confidence level, classifying levels and marking outliers, includes: An initial confidence level is set for each semantic record based on whether it is conflicting and the type of conflict, and then corrected using a physical reachability flag; The neighborhood consistency difference is calculated using the attribute values ​​of adjacent records. If the difference is less than a preset threshold, the average confidence of the record itself and its neighbors is taken as the confidence after propagation; otherwise, the corrected confidence is maintained. Calculate the average confidence of all records within the same semantic slot, and combine it with the post-propagation confidence to obtain the final confidence. Discrete levels are assigned based on the final confidence score to indicate the level of confidence. The final output includes the final confidence value, confidence level, and anomaly confidence markers.

5. The community chronic disease health management method driven by a multimodal AI large model according to claim 1, characterized in that, The process involves extracting the status of each slot to construct a health vector, calculating changes in health values ​​and inter-slot coupling strength, classifying states according to hierarchical thresholds and marking anomalies, and generating a unified health profile, including: Define state information for each semantic slot, including its corresponding entity, the latest valid attribute value, the average confidence level of the slot, and the number of records. The user's comprehensive health status vector is calculated based on the current attribute values ​​of each slot and their confidence weights. Calculate the change and rate of change of the health status vector, and classify the health status into different levels according to the medical grading standard; By constructing coupling pairs between different semantic slots and calculating the change in their attribute differences, the linkage between indicators can be monitored. If the change in health status or level exceeds a preset threshold, an anomaly is marked. Finally, the above information is integrated to generate a unified health record model.

6. The community chronic disease health management method driven by a multimodal AI large model according to claim 1, characterized in that, The system tracks the direction and magnitude of health changes, determines risk levels by considering trends, transitions, and coupled risks, triggers immediate or continuous warnings, and outputs a set of diseases, including: Analyze the direction of change in health status, the length of continuous unidirectional changes, and the cumulative magnitude to calculate the trend risk value; Detect whether a health level has transitioned, and generate a transition risk value; The number of times the change in coupling relationship exceeds a threshold is counted to calculate the coupling risk value; The trend risk value, transition risk value, and coupling risk value are combined to obtain the comprehensive risk value; Risk levels are determined based on the comparison between the comprehensive risk value and a preset threshold, and different levels of warnings are triggered accordingly. This is achieved by matching the user's set of symptoms with a pre-established set of disease symptoms based on medical guidelines.

7. The community chronic disease health management method driven by a multimodal AI large model according to claim 1, characterized in that, The process of summarizing community user risk calculation group indicators, identifying the risk spread status, delineating high-risk groups and determining intervention priorities, and triggering individual or community early warnings includes: Build a community user set and obtain the individual risk value for each user; Based on this, we can calculate the overall average risk level of the community, the proportion of high-risk users, and the dispersion of risk distribution. By monitoring the growth rate of average risk in the community, it can be determined whether there is a risk diffusion trend; Identify and generate sets of high-risk users and sets of users with ongoing risks; The priority of individual interventions is determined based on the user's overall risk score and the number of historical warnings. When an individual's risk level reaches a preset standard, an individual warning is triggered and information about their possible diseases is pushed to them. When the proportion of high-risk users in a community exceeds a preset threshold, a community-level group alert is triggered.

8. A system employing the multimodal AI large model-driven community chronic disease health management method described in claim 1, characterized in that, include: The data is fed into the preprocessing module, which performs unified time alignment, normalization, and missing data completion on multi-source heterogeneous health data, and outputs a structured standard data matrix to provide basic input for subsequent semantic analysis. The multimodal semantic parsing module is used to input standardized data into a large multimodal model, generate structured semantic information under the constraints of medical knowledge, and perform consistency and abnormal semantic detection. The semantic deduplication and slot management module is used to eliminate semantic redundancy and duplicate data, and organize semantic information into a traceable time-series slot structure by entity; The conflict identification and data correction module is used to identify abnormal changes, sampling anomalies, and physical inaccessibility in the data, and to automatically correct erroneous data. The confidence assessment module is used to evaluate the credibility of each piece of semantic data and improve the overall stability through a neighborhood propagation mechanism. The health status modeling module is used to integrate information from multiple indicators to build a dynamic health status model for users. The risk assessment and early warning module is used to comprehensively assess user health risks and trigger early warnings from the perspectives of trends, mutations, and multi-indicator coupling.

Images