AI health medical service recommendation system based on multi-dimensional biological detection and use method thereof

Abstract

The invention relates to an AI health medical service recommendation system based on multi-dimensional biological detection and a use method of the AI health medical service recommendation system. User telomere length, intestinal flora and immune cell data are collected through a biological detection module, and a BioGraphFusion unit is utilized to construct a related causal path based on a biomedical knowledge graph and quantify association strength; coding each modal data through a two-layer graph convolutional network to construct a heterogeneous sub-graph, and fusing cross-modal features in combination with an attention mechanism of bio-association intensity scheduling; high-risk operation is intercepted through an ethical examination engine, multi-modal conclusion conflicts are arbitrated, the priority is calculated according to a clinical priority rule or the flow direction is controlled according to data sensitivity, and finally personalized health reports and medical recommendation containing a biological association path diagram are output. The system integrates knowledge graph-guided feature fusion and federal learning privacy protection technologies, constructs closed-loop management from detection to intervention, improves health assessment accuracy and medical recommendation security, and gives consideration to both data compliance and user privacy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical information technology, in particular to an AI health medical service recommendation system based on multi-dimensional biological detection and a use method thereof. BACKGROUND

[0002] Disease is the result of long-term interaction of multiple factors such as genes, environment and behavior. Genetic differences alone can cause a 40-70% difference in the efficacy of the same drug in different individuals. At the same time, with the intensification of global aging and the increasing burden of chronic diseases, the traditional medical system urgently needs to be upgraded intelligently. Therefore, integrating individual multi-dimensional biological data and combining AI technology to build a health service system is essentially an inevitable choice to address modern medical pain points and promote precision and intelligent changes. However, the traditional health service system still has many problems. For example, patent number IN202541009600A emphasizes the use of attention mechanism, ensemble learning and dynamic weighting for multi-modal fusion, but its fusion process completely relies on data-driven statistical association, and cannot embed "telomere aging-immune regulation" causal association into data fusion, resulting in non-causal association misjudgment and poor feature explanation. When multi-modal conclusions conflict, simply relying on artificial intervention by clinical doctors lacks data abuse interception capability, and when multi-modal conclusions conflict, relying on artificial arbitration cannot meet the clinical priority needs. SUMMARY

[0003] (I) Technical problems solved

[0004] In view of the problems existing in the existing medical service system, the present application adopts BioGraphFusion algorithm, constructs the causal path of telomere aging, flora metabolism and immune regulation based on biomedical knowledge graph and quantifies the correlation strength, forces the graph convolution network coding and the cross-modal attention mechanism to follow the biomedical rules, so that the fused features not only reflect the data law, but also fit the known biological mechanism, avoid non-causal association interference, improve medical explainability, solve the problem that multi-modal fusion relies on statistical association and lacks medical logic, is prone to misjudgment of non-causal association and poor explainability, and at the same time, through the ethics review engine, real-time interception of privacy leakage, unauthorized sharing and other illegal operations is realized; when conflicts occur, an arbitration unit is triggered, and based on the risk degree of clinical priority, automatic calculation is realized to replace artificial arbitration, improve efficiency and meet clinical practice, ensure that urgent risks are given priority, solve the problem that ethics review lags behind and multi-modal conclusion conflicts rely on artificial arbitration, which is low in efficiency and does not meet the clinical priority.

[0005] (II) Technical solutions

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions: a use method of an AI health medical service recommendation system based on multi-dimensional biological detection, comprising the following steps:

[0007] S1. Collect user telomere length, gut flora, and immune cell data through a biological detection module;

[0008] S2. BioGraphFusion fusion unit constructs causal paths and correlation strengths of telomere aging, flora metabolism, and immune regulation based on a biomedical knowledge graph;

[0009] S3. Use a two-layer graph convolutional network (GCN) to encode each biological modality data and construct a heterogeneous subgraph corresponding to each modality;

[0010] S4. Fuse each modality encoding feature through a biological correlation strength scheduled attention mechanism and perform attention weighted calculation;

[0011] S5. The fused features are intercepted by an ethics review engine for high-risk operations and trigger conflict arbitration when a conflict is detected;

[0012] S6. After the conflict arbitration unit of the ethics review engine is triggered, priority calculation is performed according to clinical priority rules;

[0013] S7. In the absence of conflicts, the data enters a hierarchical authorization unit that controls data flow based on sensitivity;

[0014] S8. Finally, the report display unit outputs personalized health reports and medical recommendations.

[0015] Preferably, in step S2, the BioGraphFusion fusion unit constructs causal paths and correlation strengths of telomere aging, flora metabolism, and immune regulation based on a biomedical knowledge graph, specifically including:

[0016] (1) Causal path construction: dynamically activate matching causal chains from the biomedical knowledge graph according to real-time biological detection data of the user;

[0017] (2) Correlation strength quantification: convert statistical correlations in medical literature into computable path strengths γ;

[0018] (3) Path conflict arbitration: when multiple paths compete, automatically select the main path according to evidence-based medicine levels and strength thresholds.

[0019] Step S2 can also be understood as a dynamic path generation system driven by medical evidence, which matches user detection data with causal chains in the biomedical knowledge graph in real time and quantifies correlation strength.

[0020] The core use of the BioGraphFusion fusion unit is to provide a medical evidence-driven causal logic basis for the entire system, that is, by dynamically matching user detection data with causal chains in the biomedical knowledge graph, the correlation paths between telomeres, flora and immunity are constructed and the strength is quantified, and the path conflicts are arbitrated to optimize the main path. This process provides biomedical rule constraints for the subsequent encoding of multi-modal data and cross-modal feature fusion, ensuring that feature fusion conforms to both data rules and known physiological mechanisms, avoiding misjudgment of non-causal associations, improving the medical interpretability and conclusion reliability of health assessment, and providing a scientific basis for personalized medical recommendations.

[0021] Further, the biomedical knowledge graph stores knowledge in the form of triples, which specifically include a first entity, a relationship and a second entity, wherein:

[0022] (a) The first entity and the second entity are selected from at least one of telomere-related biomarkers, gut flora-related biomarkers, immune cell-related biomarkers, metabolites, and disease states;

[0023] (b) The relationship is a causal association between the first entity and the second entity, and is selected from at least one of promotion, inhibition, mediation, and regulation.

[0024] The triples are used to provide structured data support for the real-time dynamic matching of user detection data and causal chains by the BioGraphFusion fusion unit in step S2, enabling efficient execution of causal path construction, correlation strength quantification, and path conflict arbitration based on a unified knowledge storage format.

[0025] Preferably, step S3 specifically includes:

[0026] (1) Constructing an initial graph structure for each biological modality data of telomere length, gut flora, and immune cells, wherein the constructed initial graph structure contains corresponding biomarker nodes and causal relationship edges; used to construct the initial graph into a data correlation graph with the causal relationship of the biomedical knowledge graph as the skeleton, ensuring that the "data-driven" of subsequent GCN encoding is always constrained by "medical logic-driven".

[0027] (2) Using the first layer graph convolution network GCN to aggregate the node and its direct neighborhood features to extract local correlations;

[0028] (3) Using the second layer graph convolution network GCN to aggregate the node and its second-order neighborhood features to fuse global causal logic;

[0029] (4) Finally generating a heterogeneous subgraph independent of each modality, wherein the node type, edge relationship type and feature embedding of the heterogeneous subgraph are adapted to the biomedical characteristics of the corresponding modality.

[0030] The causal constraint of the biomedical knowledge graph is deeply embedded in the graph coding process to make the features not only reflect the data rules but also conform to the known molecular mechanisms of the "telomere-gut flora-immune" system, providing a structured and specific intra-modal feature basis for subsequent cross-modal fusion.

[0031] Preferably, the attention weight distribution of the "biological correlation strength scheduled attention mechanism" in step S4 is scheduled based on the correlation strength γ quantified in step S2, and the encoding features of the three modalities of telomeres, gut flora and immune cells are weighted and integrated. The specific calculation steps include:

[0032] (1) Input features and initialize the correlation matrix: define the modal feature matrix according to the heterogeneous subgraph encoding features output by the three modalities output in step S3, and construct a cross-modal correlation matrix R based on the causal relationship of cross-modal entities in the biomedical knowledge graph;

[0033] (2) Attention weight calculation: the attention weight is distributed based on the "biological correlation strength of cross-modal entities", and after normalization of the correlation strength, the final attention weight matrix A is obtained by calculating the comprehensive weight through the cross-modal attention weight;

[0034] (3) Attention weighted calculation and feature fusion: based on the attention weight matrix A, the encoding features of each modality are weighted and fused, and the specific formula is:

[0035]

[0036] Where, is the stacking matrix of the encoding features of each modality; F t is the encoding feature matrix of the telomere modality; F m is the encoding feature matrix of the gut flora modality; F i is the encoding feature matrix of the immune cell modality.

[0037] The biological correlation strength scheduled attention mechanism here is based on the biological correlation strength quantified in step S2, and through the construction of a cross-modal correlation matrix, the calculation of attention weight and weighted fusion, the precise integration of the encoding features of the three modalities of telomeres, gut flora and immune cells is realized. By strengthening the synergistic effect of high-correlation cross-modal features and suppressing irrelevant feature interference, while preserving the core information of each modality, the cross-modal causal correlation is integrated to provide fusion features that conform to the biomedical logic for subsequent health assessment, ethical review and medical recommendation, improving the accuracy of evaluation and the scientific nature of recommendation.

[0038] Further, the attention weighted calculation and feature fusion process ensures the fusion effect by the following ways:

[0039] (a) The cross-modal features with high biological correlation strength are given high weights, and their synergies are strengthened in the fused features;

[0040] (b) The features with low or no biological correlation strength are given low weights to avoid interference from irrelevant features;

[0041] (c) The fused features retain the core information of each modality while integrating the cross-modal causal correlation information. By precisely allocating weights, the synergies of cross-modal features with high biological correlation strength are strengthened, and the interference of low or irrelevant features is suppressed. At the same time, on the basis of retaining the core information of telomeres, gut microbiota, and immune cells, the cross-modal causal correlation is integrated, making the fused features not only conform to the biomedical logic but also accurately reflect the multidimensional health status, providing a reliable feature basis for subsequent health assessment, ethical review, and medical recommendation, and improving the accuracy and scientificity of system analysis.

[0042] Preferably, in step S5, the ethical review engine scans the fused features in real time through the pre-set ethical rule library to intercept violations such as user privacy leakage, unauthorized data sharing, and high-risk intervention recommendations; when conflicts are detected in the multi-modal conclusions, conflict arbitration is automatically triggered.

[0043] Preferably, in step S6, the clinical priority rule is set according to the principle of emergency risk priority processing, which is divided into four levels according to the emergency degree, and then hierarchical calculation is performed, which specifically includes:

[0044] (1) Assign weights to each level: the weight coefficient is adjusted based on the basis that the priority 1 to 4 level coefficients decrease successively;

[0045] (2) Dynamically adjust the weight: dynamically adjust the weight in combination with the quantified evidence-based medicine level in step S2;

[0046] (3) Determine the conclusion according to the comprehensive score: determine the priority conclusion according to the comprehensive score, which satisfies the formula:

[0047] Comprehensive score = R x W p x C

[0048] Wherein, R is the risk level, W p is the level weight coefficient, and C is the evidence level correction coefficient.

[0049] When there is a conflict between multi-modal conclusions, scientific arbitration is performed through clinical priority rules, that is, based on the urgency level and weight assignment, dynamically adjusting the weight based on the evidence level, and finally determining the priority conclusion through comprehensive scoring. This process ensures that the system prioritizes urgent and reliable health risks, which not only conforms to the clinical principle of "urgent risk priority", but also avoids misjudgment due to single priority or evidence bias, providing reasonable conclusion basis for subsequent health report generation and medical recommendation, and improving the scientificity and clinical applicability of decision-making.

[0050] Preferably, the sensitivity grading of the data by the hierarchical authorization unit in step S7 specifically includes:

[0051] (a) High sensitivity level: core biomarker data directly related to individual health privacy, only allowed to be stored in the user's local terminal or in the encrypted channel of the cooperative medical institution authorized dynamically, prohibited from cross-institution transmission;

[0052] (b) Medium sensitivity level: data that can be transmitted to designated health management institutions within the scope of user authorization for personalized intervention scheme generation, and the transmission process uses federated learning technology for feature desensitization;

[0053] (c) Low sensitivity level: data that can be used for population health statistical analysis after anonymization, does not require real-time authorization but needs to be synchronized with user usage records periodically.

[0054] The core is to control data according to sensitivity, balance user privacy protection and reasonable application of data, where high sensitivity data is strictly limited in transmission range to protect core health privacy; medium sensitivity data is desensitized and transmitted within the authorized range to support personalized intervention scheme generation; low sensitivity data is anonymized for population analysis, taking into account the needs of public health research.

[0055] Preferably, a use method of an AI health medical service recommendation system based on multi-dimensional biological detection involves the module composition of the system, including:

[0056] (a) Biological detection module: for detecting the telomere length, gut flora composition and immune cell subpopulation ratio of the user;

[0057] (b) Data integration and processing module: including data acquisition unit, BioGraphFusion fusion unit and feature extraction unit, wherein the BioGraphFusion fusion unit is based on a biomedical knowledge graph, encodes features through two layers of GCN, and combines cross-modal attention mechanism to fuse data;

[0058] (c) Artificial intelligence analysis module: including health status evaluation unit, ethics review engine and medical service recommendation unit, wherein the ethics review engine intercepts operations that violate preset ethical rules in real time, and arbitrates conflicts in multi-modal data conclusions;

[0059] (d) User interaction module: including report display unit and hierarchical authorization unit, the hierarchical authorization unit controls data access rights according to data sensitivity grading, and supports users to dynamically adjust the authorization range.

[0060] Advantages

[0061] (1) BioGraphFusion algorithm is adopted, causal paths of telomere aging, flora metabolism and immune regulation are constructed based on biomedical knowledge graph, and correlation strength is quantified, and the graph convolution network coding and the cross-modal attention mechanism are forced to follow the biomedical rules, so that the fused features not only reflect the data law, but also conform to the known biological mechanism, avoid non-causal correlation interference, and improve the medical interpretability.

[0062] (2) The ethics review engine intercepts illegal operations such as privacy leakage and unauthorized sharing in real time; when conflicts occur, the arbitration unit is triggered, and the "acute risk priority" clinical priority rule is automatically calculated based on the "acute risk priority" clinical priority rule, replacing manual arbitration, improving efficiency and conforming to clinical practice, and ensuring that the emergency risk is handled in priority. BRIEF DESCRIPTION OF DRAWINGS

[0063] Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are for purposes of illustration only and are not considered a limitation of the present application. Moreover, like reference numerals are used to designate identical components throughout the specification. In the drawings:

[0064] Figure 1 is a system overall architecture diagram of an AI health medical service recommendation system and a use method thereof based on multi-dimensional biological detection according to the present application;

[0065] Figure 2 is a whole flow chart of an AI health medical service recommendation system and a use method thereof based on multi-dimensional biological detection according to the present application;

[0066] Figure 3 is a BioGraphFusion algorithm flow chart of an AI health medical service recommendation system and a use method thereof based on multi-dimensional biological detection according to the present application;

[0067] Figure 4 is an ethics review engine workflow diagram of an AI health medical service recommendation system and a use method thereof based on multi-dimensional biological detection according to the present application;

[0068] Figure 5 is a federal learning data desensitization flowchart of an AI health medical service recommendation system based on multi-dimensional biological detection and a use method thereof according to the present application; DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings in the examples of the present application Figure 1 -Appendix Figure 5 The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings in the examples of the present application

[0070] A use method of an AI health medical service recommendation system based on multi-dimensional biological detection, comprising the following steps:

[0071] S1. Collecting user telomere length, gut flora and immune cell data through a biological detection module;

[0072] S2. BioGraphFusion fusion unit, based on a biomedical knowledge graph, constructing causal paths and correlation strengths of telomere aging, flora metabolism and immune regulation;

[0073] S3. Using a two-layer graph convolution network GCN to encode each biological modality data respectively, and constructing a heterogeneous subgraph corresponding to each modality;

[0074] S4. Fusing each modality encoding feature through a biological correlation strength scheduling attention mechanism, and performing attention weighted calculation;

[0075] S5. The fused features are intercepted by an ethics review engine for high-risk operations and trigger conflict arbitration when a conflict is judged to exist;

[0076] S6. After the conflict arbitration unit of the ethics review engine is triggered, priority calculation is performed according to clinical priority rules;

[0077] S7. In the case of no conflict, the data will enter a hierarchical authorization unit which will control the data flow according to the sensitivity of the data;

[0078] S8. Finally, a personalized health report and medical recommendation are output by a report display unit.

[0079] The BioGraphFusion fusion unit in step S2 constructs causal paths and correlation strengths of telomere aging, microbial metabolism, and immune regulation based on a biomedical knowledge graph. This can also be understood as a dynamic path generation system driven by medical evidence, which matches user detection data with causal chains in the biomedical knowledge graph in real time and quantifies the correlation strength γ. Specifically, it includes:

[0080] (1) Causal path construction: According to real-time biological detection data of the user, the matching causal chain in the biomedical knowledge graph is dynamically activated. Specifically, the system first extracts core biomarkers (such as user telomere length, gut flora, immune cells, and other key data detected by the biological detection module) from the detection data, then accurately matches the causal chain from the biomedical knowledge graph through SPARQL query, and converts the result into a computable path, that is, locates the entity nodes corresponding to these markers in the knowledge graph, and generates the correlation chain between multi-modal data by traversing the direct / indirect causal edges between nodes. For example, when it is detected that the user's telomere length is short, the system searches the causal path related to telomere length in the knowledge graph. If it is found that the user has a low abundance of beneficial bacteria (such as Bifidobacterium) and a high abundance of harmful bacteria (such as Escherichia coli), combined with the causal relationships in the knowledge graph such as "an increase in harmful bacteria leads to the production of a large amount of harmful metabolites, which triggers an inflammatory response and accelerates telomere loss" and "a decrease in beneficial bacteria leads to insufficient production of short-chain fatty acids, which cannot effectively inhibit the telomere damage-related pathway", a causal path between telomere length and gut flora is constructed. At the same time, if it is detected that the NK cell activity in the immune cells is low at this time, according to the relationship "immune cell dysfunction (low NK cell activity) leads to the inability to remove damaged cells in time, thus accumulating telomere damage", the immune cell data is also included in this causal path, forming a complex causal correlation chain across multi-modal data, which fully reflects the internal relationship of the user's health status.

[0081] (2) Correlation strength quantification: Convert the statistical correlation in medical literature into a computable path strength γ. This can be understood as quantitatively representing the correlation between the links in the constructed causal path. The core basis includes:

[0082] (a) Evidence strength: The number of high-quality studies supporting the causal relationship in the statistical knowledge graph, the combined effect size of Meta analysis, and the P value.

[0083] (b) Entity correlation: Analyze the co-occurrence frequency and semantic similarity of two entities in the literature through text mining techniques.

[0084] (c) User data adaptability: Adjust the weight in combination with the individual characteristics of the user (such as age, underlying diseases).

[0085] Specific quantification method adopts three-layer calculation model, including basic weight calculation, correction coefficient calculation and path total intensity calculation:

[0086] 1) Basic weight calculation: based on evidence intensity assignment calculation, that is, according to the standard of evidence-based medicine, medical evidence supporting causal relationship is divided into four levels, which are:

[0087] Level 1: high-quality Meta analysis, multi-center RCT study, its weight coefficient α = 1.0;

[0088] Level 2: cohort study, case-control study, its weight coefficient α = 0.7-0.9;

[0089] Level 3: expert consensus, small sample clinical observation, its weight coefficient α = 0.4-0.6;

[0090] Level 4: animal experiment or cell experiment, its weight coefficient α = 0.1-0.3.

[0091] It should be noted that the level weight needs to be dynamically adjusted, that is, if a study is refuted by subsequent studies, α is downgraded by one level.

[0092] According to the weight coefficient obtained by classification, the statistical correlation in the literature is converted into the basic weight factor w b , which satisfies the formula:

[0093]

[0094] Among them, ES study is the statistical effect size of a single article, which reflects the statistical quantity of causal relationship in the current study, which needs to be standardized according to the type of research; ES max,k is the upper limit of effect size corresponding to the evidence level, which is set based on the historical data of the knowledge graph of this level of research, and is used for normalization of effect size; P is the P value of statistical significance, which is taken from the hypothesis test result in the literature, which reflects the statistical reliability of the correlation, which is used to convert P value into positive weight.

[0095] 2) Correction coefficient calculation: combined with entity correlation and user data adaptability, the basic weight factor is dynamically adjusted, which can be understood as the realization of weighted fusion of entity correlation score and user adaptability score, satisfying the formula:

[0096] C = 0.5S rel + 0.5S ada

[0097] Among them, S rel is the entity correlation score, which is based on text mining technology to analyze the correlation between two entities in the knowledge graph; S adaUser adaptability score is to adjust the correlation strength according to the individual characteristics of the user (age, underlying disease, living habits, etc.); the two are fused by the weight of 0.5:0.5, to ensure that the two dimensions have balanced influence on the correction coefficient.

[0098] Entity correlation score S rel The calculation includes co-occurrence frequency normalization and semantic similarity calculation. Co-occurrence frequency normalization refers to counting the co-occurrence times of two entities (such as "telomere length" and "bifidobacterium") in medical literature, and comparing it with the maximum value of the total co-occurrence times of entities in the literature of the field to obtain the frequency factor F; semantic similarity calculation is to encode the text description of the entity through the BERT model, and then calculate the cosine similarity Sim after obtaining the semantic vector; the two are combined to obtain S rel = 0.8 + 0.4 * (0.6F + 0.4Sim).

[0099] Here 0.8 is the baseline value, 0.4 is the adjustment range, 0.6 is the co-occurrence frequency weight, and the semantic similarity weight is 0.4. Through the weighted result dynamic floating, it is ensured that S rel falls in the interval of 0.8-1.2. Through such setting, S rel can dynamically respond to the changes of entity correlation characteristics within a reasonable interval, and can balance the influence of different dimensions, providing reliable entity correlation basis for the correction coefficient calculation.

[0100] User adaptability score S ada The calculation includes feature matching degree calculation and human rights difference coefficient calculation. Feature matching degree calculation refers to matching the user characteristics (such as "65 years old + diabetes") with the applicable population characteristics of the causal relationship in the knowledge graph (such as "the association is more significant in the elderly diabetic population"), to obtain the matching degree M; the human rights difference coefficient calculation is to set the coefficient D according to the difference of effect size in different subgroups in clinical research; the two are combined to obtain S ada = 0.4 + 1.2MD.

[0101] Here 0.4 is the baseline value, 1.2 is the adjustment range, and the product of matching degree M and difference setting coefficient D realizes individual floating.

[0102] The core role of the correction coefficient is to make up for the limitations of the basic weight relying only on literature evidence, by introducing entity semantic correlation and user individual characteristics, to make the quantitative result closer to the real physiological correlation strength.

[0103] 3) Path total strength calculation: the chain aggregation calculation of the correlation strength of each link in the path after correction, the core is to convert the correlation strength of multiple links into a single path level quantitative index, the specific content is:

[0104] If there are n consecutive associated links in the causal path, and the modified associated strength of each link is W i b，i ×C i , where W b，i is the basic weight factor of the link, and C i is the corresponding correction coefficient, then at this time, the total strength of the path satisfies the formula:

[0105]

[0106] where, represents the multiplication operation of the modified associated strength W i from the first link to the nth link in the causal path.

[0107] When there are parallel branches in the causal path (such as a node affecting the next node through two sub-paths at the same time), the sub-path strength of each branch needs to be calculated first, and then aggregated into the total strength of the node through weighted summation. At this time, the total strength of the path satisfies the formula:

[0108]

[0109] where m is the number of parallel branches; γ j is the total strength of the jth branch, and w j is the weight of the branch, which is determined by the proportion of the evidence strength of the branch.

[0110] The total strength of the path integrates the associated strength of multiple links into a single indicator through chain multiplication (or parallel branch weighting), which can not only reflect the overall associated reliability of the entire path, but also locate the "weak link" in the path (such as a link W i is extremely low, which will significantly lower γ, indicating that there may be problems such as insufficient evidence or poor individual adaptability in this link), providing quantitative basis for subsequent health report generation and medical recommendation (such as preferentially recommending intervention schemes for high total strength paths).

[0111] (3) Path conflict arbitration: When multiple paths compete, the main path is automatically selected according to the evidence-based medicine level and strength threshold. That is, when multiple causal paths compete for the same health outcome (such as "telomere shortening" and "immune function decline"), the system automatically selects the main path through the double mechanism of "evidence-based medicine level weighting" and "strength threshold screening". The specific process is as follows:

[0112] 1) Identify conflicting paths and classify: First, identify all candidate paths for the same health outcome through path label matching, denoted as set P = {P1, P2, …, P k ​k is the number of paths, and is divided into conclusion conflict and strength competition according to the conflict type. The conclusion conflict refers to the opposite influence direction of the path on the health outcome. The strength competition refers to the consistent path conclusion but significant difference in correlation strength, and the dominant influence path needs to be determined.

[0113] 2) Evidence-based grade weighting calculation: Assign "evidence-based grade weight" E to each path w which is based on the highest evidence-based grade of each link in the path, that is, taking the highest medical evidence grade link in the path as the representative. At the same time, the path comprehensive score S is calculated combined with the total strength γ of the path, which satisfies the formula:

[0114]

[0115] Among them, is the historical average strength of the path in the biomedical knowledge graph, mainly reflecting the general applicability of the group. Through the weighted balance of evidence-based grade and individual strength, the deviation of a single index is avoided.

[0116] 3) Screening strength threshold and determining the main path: Screening is carried out through setting double thresholds, including absolute threshold and relative threshold. The absolute threshold means that if the comprehensive score S of a path is greater than or equal to 0.6 and the scores of other paths are less than 0.6, then the path is directly determined as the main path. The relative threshold means that if the scores of multiple paths all exceed 0.6, then the path with the highest score is selected, and it needs to meet the condition that "the difference between the highest score and the second highest score is greater than or equal to 0.1". This is to ensure that the main path is significantly superior. (Here, 0.6 is the preset high credibility threshold. For the convenience of understanding, the numerical value is directly used, but the preset data is not unique.)

[0117] If the above conditions are not met, secondary arbitration is started, which is specifically:

[0118] a) Preferentially retaining the path containing higher evidence-based grade (such as the path containing grade 1 evidence is preferred to the path containing grade 2 evidence);

[0119] b) If the evidence-based grade is the same, compare the path length (such as short path is more likely to be a direct causal relationship, then preferentially retain);

[0120] c) If still cannot be determined, the conflict path and score are synchronized to the artificial audit module, and the clinical experts intervene in arbitration.

[0121] Through the combination of evidence-based grade and strength threshold, it is ensured that the main path not only meets the authority of medical evidence, but also fits the individual data characteristics of the user, providing a reliable causal logic basis for subsequent personalized health recommendation, while avoiding the interference of low-quality or conflict paths on the conclusion.

[0122] The biomedical knowledge graph here stores knowledge in the form of triples, which specifically include a first entity, a relationship, and a second entity, wherein:

[0123] The first entity and the second entity are selected from at least one of a telomere-related biomarker, a gut microbiota-related biomarker, an immune cell-related biomarker, a metabolite, and a disease state;

[0124] The relationship is a causal association between the first entity and the second entity, and is selected from at least one of promotion, inhibition, mediation, and regulation;

[0125] And the pre-construction of the biomedical knowledge graph needs to follow the following rules:

[0126] (1) Data source screening: preferentially collect literature data in authoritative medical databases (such as PubMed, EMBASE, etc.), and include clinical guidelines and expert consensus published by well-known medical research institutions; for basic research data, obtain from top academic journals such as Cell, Nature, and Science, and integrate high-quality large population cohort study data;

[0127] (2) Entity and relationship extraction: use named entity recognition algorithms in natural language processing (NLP) technology to accurately extract entities such as genes, proteins, metabolites, diseases, and cell types (such as immune cell subgroups) from literature texts, i.e., use a deep learning-based BERT model to identify entities such as telomeres and specific bacteria in the gut microbiota. Relationship extraction uses techniques such as dependency syntax analysis and semantic role labeling to identify causal relationships such as "promotion", "inhibition", "cause", etc. For structured data (such as clinical test reports), directly extract entity and relationship data according to the established format.

[0128] The deep learning-based BERT model here is a deep learning-based pre-training language model proposed by Google in 2018, which has been widely used in the field of natural language processing (NLP) and belongs to a relatively common existing technology, so it will not be described in detail here.

[0129] (3) Ontology construction rules: adopt a combination of top-down and bottom-up methods to construct the ontology. Top-down is to define general concept frameworks such as "biological processes", "molecular functions", and "cell components" first, and then gradually refine them into specific concepts such as "telomere maintenance processes" and "microbiota metabolic functions"; bottom-up is to summarize new concepts and relationships from a large number of extracted entities and relationships. Use Web Ontology Language (OWL) to formally define concepts and relationships, and clearly define the hierarchical relationships and attribute characteristics between entities, such as clearly defining that "Bifidobacterium" belongs to the "gut beneficial bacteria" category and has the attribute of "producing short-chain fatty acids".

[0130] (4) Knowledge fusion and disambiguation: For knowledge from different sources, eliminate duplication and inconsistency through entity alignment technology, that is, similarity calculation based on entity name, attribute, description information, etc., for example, calculate the similarity of the description of "CD4+ T cells" in different literatures, and determine whether it is the same entity. At the same time, use semantic annotation, link open data (LOD), etc. to associate the extracted knowledge with existing authoritative knowledge graph (such as UMLS, Bio2RDF, etc.) to supplement missing information and correct incorrect information. For ambiguous terms, disambiguation processing is performed according to the context and domain knowledge.

[0131] (5) Update and maintenance: Regularly (such as monthly or quarterly) update the knowledge graph to track the latest medical research progress. Through the setting of automatic monitoring mechanism, when new high-impact research is published, it is timely incorporated into the knowledge graph. In addition, for existing causal relationships, if new research results contradict them, re-evaluate the evidence strength, and modify or update the causal relationship according to the latest evidence to ensure the timeliness and accuracy of the knowledge graph.

[0132] Step S3 specifically includes:

[0133] (1) Construct an initial graph structure for each biological modality data of telomere length, intestinal flora, and immune cells, wherein the constructed initial graph structure includes corresponding biomarker nodes and causal relationship edges;

[0134] The construction of the initial graph strictly depends on the correlation strength-based weight quantified in step S2, ensuring that the graph structure embeds biomedical causal logic from the source, specifically including:

[0135] (a) Telomere modality initial graph: nodes include "telomere length measured value", "telomerase activity detection value", "telomere damage marker", etc. These data are all from the detection data of step S1; the edges are the causal relationships between nodes (such as "telomerase activity rises leading to telomere length rising" and "telomere damage rising leading to telomere length falling", etc.), and the edge weights directly use the basic weights calculated in step S2. Intestinal flora modality initial graph: nodes include "dominant flora relative abundance", "characteristic metabolite concentration", etc.; the edges are the interaction relationships between flora (such as "Bifidobacterium rising leading to Escherichia coli falling") and flora-metabolite relationships (such as "Bifidobacterium rising leading to short-chain fatty acid rising"), and the edge weights are also based on the basic weights of step S2.

[0136] (b) Immune cell modality initial graph: nodes include "immune cell subpopulation proportion", "cytokine level", etc.; edges are cell function regulation relationships (such as "NK cell activity rising leading to abnormal cell clearance rate rising") and cytokine secretion relationships (such as "CD4 +The edge weight is also based on the base weight determination of step S2.

[0137] (2) The first layer graph convolutional network GCN is used to aggregate the features of the node and its direct neighborhood to extract local correlation. The core is to capture the local regulatory relationship between the nodes in each modality and its direct neighborhood. The specific implementation is as follows:

[0138] 1) The original detection data of each node is converted into a 256-dimensional feature vector as the initial embedding of the node.

[0139] 2) For each node, the features of its direct neighborhood nodes are aggregated through a 3*3 convolution kernel, and the calculation satisfies the formula:

[0140]

[0141] where w ij is the edge weight between nodes i and j, that is, the base weight in step S2), W1 is the trainable weight matrix of the first layer graph convolutional network GCN, and b1 is the bias term.

[0142] 3) The first layer embedding features containing local causal correlation are generated, and the most direct biological regulatory relationship within the modality is retained.

[0143] (3) The second layer graph convolutional network GCN is used to aggregate the features of the node and its second-order neighborhood to fuse global causal logic.

[0144] The second layer GCN further integrates the association between the node and the second-order neighborhood based on the local features, that is, the nodes indirectly connected through two causal edges, to realize the fusion of global causal logic. The specific steps are as follows:

[0145] 1) The embedding features output by the first layer GCN are taken as the input h i (1) , and the edge weight is updated to the value w ij ′=w ij ×C, C is the correction coefficient) adjusted by the correction coefficient in step S2.

[0146] 2) The features of the node and its second-order neighborhood are aggregated through a 5*5 convolution kernel, and the calculation satisfies the formula:

[0147]

[0148] where W2 is the trainable weight matrix of the second layer GCN, b2 is the bias term, and the LeakyReLU activation function is used to avoid gradient disappearance of sparse features.

[0149] 3) Generate the second layer of embedding features that integrate the global causal chain, so that the features not only reflect direct associations, but also embody multi-link biological causal logic.

[0150] (4) Finally generate each modality independent heterogeneous subgraph, the node type, edge relationship type and feature embedding of the heterogeneous subgraph are adapted to the corresponding modality biomedical characteristics.

[0151] After two layers of GCN coding, each modality data is generated as a heterogeneous subgraph. The so-called heterogeneous subgraph is a graph structure that is constructed independently for telomere length, gut flora and immune cell biological modal data, and is adapted to the corresponding modality biomedical characteristics. Its core characteristics include:

[0152] a) Heterogeneity: The subgraphs of different modalities are completely independent in node type and edge relationship type, and do not cross each other. That is, the nodes of the telomere-related heterogeneous subgraph are telomere-related biomarkers, and the edges are the regulatory relationship between nodes; the nodes of the gut flora-related heterogeneous subgraph are gut flora markers and metabolites, and the edges are the interaction or metabolic generation relationship of the flora; and the nodes of the immune cell-related heterogeneous subgraph are immune cell markers and cytokines, and the edges are the cell function regulation or cytokine secretion relationship.

[0153] b) Modality adaptability: The structure and features of each subgraph are adapted to the biomedical characteristics of the corresponding modality. After two layers of GCN coding, the subgraph contains not only the local association features of the nodes within the modality (such as direct regulatory relationship), but also the global causal logic (such as multi-link chain association), and finally presents in the form of adjacency matrix (storing node association weight) and node embedding matrix (storing the features output by the second layer of GCN).

[0154] c) Structured storage: Each subgraph is stored in the form of "adjacency matrix + node embedding matrix", the adjacency matrix records the corrected association weight between nodes, and the node embedding matrix records the h i (2) Features output by the second layer of GCN.

[0155] The attention weight distribution of the "biological correlation strength scheduled attention mechanism" in step S4 is scheduled based on the correlation strength γ (which can also be understood as the path strength) quantified in step S2 as the core, and the encoded features of the three modalities of telomeres, gut flora and immune cells are weighted and integrated. That is, after the completion of the heterogeneous subgraph calculation in step S3, the node embedding features of the generated heterogeneous subgraph, that is, the encoded features of the three modalities of telomeres, gut flora and immune cells, are weighted and integrated after three-step calculation of biological correlation strength calculation, attention weight calculation and feature fusion, so that the fused features not only retain the specificity information of each modality, but also highlight the key causal correlation between cross-modalities, such as showing the key causal correlation of "gut flora metabolites affecting immune cell activity and causing telomere damage". The specific calculation steps include:

[0156] (1) Input features and initialize correlation matrix: define the modality feature matrix based on the heterogeneous subgraph encoding features output by the three modalities in step S3, and construct the cross-modal correlation matrix R based on the causal relationship between cross-modal entities in the biomedical knowledge graph;

[0157] Define the modality feature matrix based on the heterogeneous subgraph encoding features output by the three modalities, respectively denoted as:

[0158] a) Telomere modality feature matrix Where n t is the number of telomere-related nodes, and d is the feature dimension, which is consistent with the embedding dimension output by the GCN in step S3;

[0159] b) Gut flora modality feature matrix Where n m is the number of flora-related nodes;

[0160] c) Immune cell modality feature matrix Where n i is the number of immune-related nodes.

[0161] After inputting the telomere, gut flora and immune cell three modality feature matrices as input features into the attention mechanism, an initialized correlation matrix is generated, that is, a cross-modal correlation matrix R is constructed based on the causal relationship between cross-modal entities in the biomedical knowledge graph. Here, the initial value of R a，b is the correlation strength γ between two entities quantified in step S2, and γ is 0 in the case of no correlation.

[0162] For ease of understanding, further examples are given to describe the causal relationship here, for example, short-chain fatty acids belonging to the gut flora modality can inhibit interleukin-6 (IL-6) belonging to the immune cell modality, and interleukin-6 (IL-6) can promote telomere damage belonging to the telomere modality.

[0163] (2) Attention weight calculation: based on the "biological correlation strength of cross-modal entities", the attention weight distribution is performed, and after the correlation strength is normalized, the final attention weight matrix A is obtained by calculating the comprehensive weight through the cross-modal attention weight;

[0164] That is, the distribution of attention weight is based on the "biological correlation strength γ of cross-modal entities", and in the specific calculation, the correlation strength is first normalized, that is, the correlation strength γ in the previously constructed cross-modal correlation matrix R is normalized to obtain the normalized correlation matrix At this time,

[0165] After the normalization is completed, the intra-modal weight is calculated based on the internal correlation strength of the modal heterogeneous subgraph for the features within a single modal, so as to ensure that important features within the modal can obtain higher weight. The internal correlation strength of the heterogeneous subgraph is the weight of the adjacency matrix of the heterogeneous subgraph in step S3.

[0166] Then, the cross-modal attention weight between different modalities is calculated based on the normalized correlation matrix The cross-modal attention weight between each modality of telomere, flora and immune modality is calculated respectively, and finally the intra-modal weight and the cross-modal weight are integrated by weighted summation to obtain the final attention weight matrix

[0167] And for the cross-modal attention weight, the normalized correlation matrix is determined; and for the weight of the immune modality attention feature of the telomere modality, the normalized matrix is determined.

[0168] (3) Attention weight calculation and feature fusion: based on the attention weight matrix A, the modal encoding features are weighted and fused, and the specific formula is:

[0169]

[0170] Wherein, is the stacking matrix of the encoding features of each modality; F t is the encoding feature matrix of the telomere modality; F m is the encoding feature matrix of the intestinal flora modality; F i is the encoding feature matrix of the immune cell modality.

[0171] ​Further, the attention-weighted calculation and feature fusion process ensure the fusion effect by the following ways:

[0172] (a) Cross-modal features with high biological correlation strength are given high weights, and their synergies are strengthened in the fused features. Here, "high biological correlation strength" specifically refers to the quantification of the cross-modal entity causal path strength (i.e., correlation strength) γ, that is, the quantification in step S2 reaches a preset threshold. If the system preset threshold γ is 0.6 and the quantified γ ≥ 0.6, this type of feature is called a cross-modal feature with high biological correlation strength, and this type of feature usually corresponds to a core causal chain supported by high-quality evidence in the biomedical knowledge graph.

[0173] More specifically, for example, the previous example "short-chain fatty acids belonging to the gut microbiota modality can inhibit interleukin-6 (IL-6) belonging to the immune cell modality" cross-modal feature pair, if its correlation strength γ = 0.75, which is greater than the preset threshold 0.6, then this cross-modal feature pair belongs to the high correlation strength feature, and in the attention weight matrix A, the weight value of this feature pair will be significantly higher than that of other feature pairs. In the fusion, the encoded features of short-chain fatty acids and interleukin-6 (IL-6) are weighted and summed by high weights, so that the synergistic effect of "short-chain fatty acids inhibiting interleukin-6 (IL-6)" in the fused features is amplified, which is specifically manifested as a significant increase in the corresponding dimension value in the fused feature vector.

[0174] This strengthening mechanism ensures that the system preferentially captures key causal correlations such as "metabolites of gut microbiota to immune regulation", which is consistent with the known mechanism in biomedical science that "short-chain fatty acids protect telomeres by inhibiting pro-inflammatory factors", avoiding excessive statistical noise that obscures the core correlation and affects the accuracy of the results.

[0175] (b) Features with low biological correlation strength or no correlation are given low weights to avoid interference from irrelevant features. Low biological correlation strength means that the correlation strength γ of the cross-modal entity pair is < 0.3, or there is no clear causal relationship recorded in the biomedical knowledge graph, i.e., γ = 0. This type of feature pair is usually a weak or pseudo correlation in statistics.

[0176] For example, the cross-modal feature pair "a rare gut microbiota with extremely low abundance affects telomere length", due to its rarity, has no causal research support for the relationship between the microbiota and telomere in the biomedical knowledge graph, which means γ = 0. Therefore, the weight of this feature pair in the attention weight matrix A is set close to 0. In the fusion, the encoded feature of the rare microbiota has a negligible effect on the fusion result, so as to avoid its random fluctuations interfering with the core evaluation of the "telomere aging" result.

[0177] This mechanism aims to reduce unnecessary disturbance to the core evaluation of the image, that is, to solve the problem of misjudgment of seasonal changes and statistical co-occurrence of telomere length as related key in traditional data fusion, improve the accuracy of fusion, and thus improve the diagnostic accuracy of the system.

[0178] (c) The fused features retain the core information of each modality while integrating the cross-modality causal correlation information. Retaining the core information of each modality means that the key biomarker features of telomeres, gut flora, and immune cells can still be traced in the fused features; and integrating cross-modality causal correlation means that the feature contains chainwise causal logic between multi-modal entities (such as "flora affecting metabolites affecting immune factors affecting telomeres").

[0179] For example, the core features of the telomere modality, such as "telomere length measured value" and "telomerase activity", still exist as independent dimensions in the fused feature matrix F 融合 , and their values are highly correlated with the embedded features of the corresponding nodes in the telomere feature matrix F t encoded in step S3, which ensures that some relevant basic detection results are not obscured by cross-modality fusion. This is called retaining modality core information.

[0180] Integrating cross-modality causal correlation means that in the fused features, for example, "short-chain fatty acids belonging to the gut flora modality affect interleukin-6 (IL-6) belonging to the immune cell modality, causing telomere damage belonging to the telomere modality", such chainwise correlation is integrated through the transitivity of the attention weight matrix A. The high weight of the gut flora modality on the immune cell modality and the high weight of the immune cell modality on the telomere modality are superimposed, so that the overall contribution of this chain in the fused features is the synergistic weight after the addition of the two, which is much higher than the independent contribution of a single modality, and can clearly reflect the multi-link causal logic.

[0181] In step S5, the ethics review engine scans the fused features in real time through the pre-set ethics rule library, intercepts violations such as user privacy leakage, unauthorized data sharing, and high-risk intervention recommendations; when conflicts are detected in the multi-modal conclusions, automatic conflict arbitration is triggered.

[0182] The core of the ethics review engine is the pre-set ethics rule library, and the rule entries of the pre-set ethics rule library are based on medical ethics guidelines, data security regulations, and clinical operation specifications. Each type of rule contains quantifiable judgment criteria, which are manifested as:

[0183] (1) Privacy protection rules: used to clearly define the processing boundaries of high-sensitivity data, and triggered by data tag matching to intercept;

[0184] The high-sensitive data targeted by the privacy protection rules specifically include: a. Telomere length raw detection value, telomerase activity data, which are directly related to aging and disease risk, with a privacy level of one; b. Gut microbiota whole genome sequencing data, dominant flora absolute abundance, which can reflect individual diet, disease history, with a privacy level of one; c. Immune cell subpopulation ratio, such as CD4 + / CD8 + T cell ratio, associated with sensitive health information such as immune deficiency, with a privacy level of one;

[0185] In addition, the privacy level of medium-sensitive data (such as gut microbiota relative abundance, cytokine concentration) is recorded as level two, and the privacy level of ordinary health indicators (such as detection date) is recorded as level three.

[0186] The interception triggered by data tag matching is realized by embedding the tag system through metadata, that is, each type of data is automatically associated with a "privacy level" tag after collection in step S1, and the corresponding level of each data is marked and stored in the data header file.

[0187] At the same time, the system will preset an "authorized scenario whitelist", which can include health assessment, clinical consultation, and user self-query, etc. When it is detected that data with a privacy level of one is used in scenarios outside the whitelist, interception is triggered immediately.

[0188] At the same time, the privacy protection rules clearly state that data with a privacy level of one is only allowed to be transmitted within the user's local terminal and the encrypted channel of the cooperating medical institution. That is, once it is detected that data with a privacy level of one is flowing to an unrecorded IP address, regardless of its purpose, it will trigger interception. It should be noted that to achieve this interception trigger, the IP whitelist of cooperating institutions needs to be built into the system, and the target IP of data transmission needs to be checked in real time. When the matching success rate of non-whitelist IP is > 0, it is determined to be illegal, and interception is triggered immediately.

[0189] Further, to achieve interception triggering, the API interface tag matching of the data transmission destination is required, that is, the interface tag "auth = 1" is preset for whitelist scenarios, and the tag "auth = 0" is preset for non-whitelist scenarios. When data with a privacy level of one is matched with the non-whitelist scenario preset tag "auth = 0" interface, the interception mechanism is started.

[0190] (2) Data misuse interception rules: used to define the compliance range of data use, automatically identify misuse risks through keyword matching and semantic analysis;

[0191] The quantifiable scenarios of explicit data misuse include: a. Discriminatory decision-making, that is, using biological data for screening in non-health fields (such as employment, education), and their judgment keywords include "recruitment", "admission", "assessment" and other scene-related words. The system identifies such keywords in the text through an NLP model, and if the matching degree is > 80%, it is determined to be misused. B. Identity identifiable information leakage, that is, through the BERT derivative model to detect the combination of "age, gender, detection agency and detection time" in the health report or data output, through which the individual can be uniquely identified; or directly mention the name, ID number and other identification-related information, if the recognition accuracy is > 95%, it is determined to be leaked.

[0192] Further, to implement this rule interception, the system needs to build a "misuse keyword library" containing more than 120 sensitive keywords (including but not limited to "dismissal", "refusal", "premium adjustment" and other content), and set a weight for each keyword. In addition, the system performs semantic vector conversion on the data output text and calculates the cosine similarity with the keyword library vector. If the similarity is ≥ 0.7, it triggers a misuse risk warning; for ambiguous scenarios (such as "health assessment for employee welfare optimization"), the system will determine the core intent through context semantic analysis. If welfare optimization involves differential treatment, that is, according to the level of rating (such as dividing the welfare level according to the flora data), it is determined to be misused.

[0193] The core functions of the NLP (Natural Language Processing) model and the BERT derivative (entity recognition) model here are to complete keyword recognition, semantic vector conversion, context semantic analysis and other basic natural language processing tasks. These technologies have become mature and widely used in the field. As early as 2024, the public literature "RoBERTa Model in Natural Language Processing: Evolution of BERT" showed that NLP models have been applied in the field of medical data compliance, such as keyword interception in "insurance underwriting" and "employment screening" scenarios. The BERT derivative model is essentially a downstream task adaptation of BERT. According to the above-mentioned literature, it can be understood that RoBERTa, ALBERT and other derivative models have become the mainstream solution for entity recognition in 2023-2024. Therefore, no detailed description is given.

[0194] (3) Clinical safety guidelines: used to regulate the safety of medical recommendations, by setting risk levels for intervention recommendations, when the recommended content does not meet the level conditions, triggering interception.

[0195] Setting risk levels for intervention recommendations can also be understood as pre-setting risk levels for all possible medical recommendations. The specific quantitative standards can refer to the following examples: (the division is not unique)

[0196]

[0197]

[0198] It should be noted that according to the table content, when the risk level of the recommended content is greater than or equal to three, the condition of "contraindication screening pass" must be met, and the screening fails to be intercepted; for example, when recommending an anti-inflammatory drug, the anti-inflammatory drug belongs to a prescription-level intervention, and the risk level is divided into three levels, so the system will perform contraindication screening on the user, that is, automatically check whether the user has a "gastric ulcer" history. If the screening fails, that is, the user has a history of gastric ulcer, it is directly intercepted. Intervention suggestions with a risk level equal to five are directly intercepted in any scenario, and the source of the recommendation is recorded.

[0199] The risk level and the contraindication screening result are associated through a logic gate circuit. In addition, it should be noted that the rule base is linked with the latest regulations (including but not limited to the revised provisions of GDPR, the domestic "Personal Information Protection Law" and other relevant regulations). The keyword library, risk level standard and authorized scene white list are automatically synchronized and updated every month to ensure compliance. In addition, all interception events generate tamper-proof blockchain logs (including time, triggering rules, and processing results), and at the same time, an alarm is pushed to the system administrator, and high-risk events are automatically synchronized to the user authorization panel for confirmation by the user whether to be exempted.

[0200] The meaning of multi-modal conclusion conflict is that the fusion features output in step S4 correspond to different modal health assessment conclusions, which exist contradictions, then trigger conflict arbitration, specific scenarios include:

[0201] (1) Risk direction conflict: different modal data point to opposite health trends; that is, the indicators of telomeres, gut microbiota and immune cells may present reverse signals due to compensatory mechanisms, detection window period differences, etc. Typical scenarios include conflicts between gut microbiota and immune cell modalities and conflicts between telomere modalities and cross-modal fusion features. Specifically:

[0202] The gut microbiota modality feature gets an increase in Bifidobacterium abundance, which affects an increase in short-chain fatty acids, thereby indicating an "immune protection trend", and its risk direction is low inflammation risk; the immune cell modality feature gets a decrease in CD4 + T cell proportion, which affects an increase in IL-17 concentration, thereby indicating an "immune imbalance trend", and its risk direction is high inflammation risk. The conflict between the gut microbiota modality and the immune cell modality may be caused by "short-term improvement of the microbiota has not reversed long-term immune damage", which belongs to the clinically common transition state contradiction, and needs to be judged by conflict arbitration to determine the dominant trend.

[0203] The direct detection of telomere mode indicates that the length of telomere is stable, and the risk direction is low aging risk; the cross-modal fusion feature, which is the fusion feature of intestinal flora mode and immune cell mode, indicates that the increase of E. coli abundance affects the increase of TNF-a, which in turn affects the increase of γ-H2AX, a marker of telomere damage, and the risk direction is high aging risk. Since telomere length is a static indicator, while damage markers are dynamic changes, the conflict between telomere mode and cross-modal fusion feature reflects that potential damage has not yet been reflected in length changes, and dynamic risk needs to be prioritized.

[0204] For the determination of risk direction conflict, the system identifies the direction conflict by labeling the modal features, that is, the risk direction label is assigned to the health assessment conclusion of each mode, where "+1" represents positive risk (high risk), "-1" represents negative risk (low risk), and when the labels of different modes are multiplied by "-1", it is determined that there is a risk direction conflict. For example, if the intestinal flora mode label is "-1" (low risk) and the immune cell mode label is "+1" (high risk), the product of the two is "-1", which triggers conflict arbitration at this time.

[0205] (2) Risk level contradiction: the conclusion direction is consistent, but the risk level difference exceeds the preset threshold. That is, the evaluation intensity of different modes on the same health problem differs beyond the clinically acceptable range. It is important to note that the setting of the preset threshold is not a subjective choice, but is based on the fault tolerance space of clinical decision-making. Specifically, the system divides health risks into 1-10 levels, with level 1 being the lowest risk and level 10 being the highest risk, corresponding to specific clinical intervention needs:

[0206] 1-3 levels: no medical intervention is needed, such as "diet adjustment" and other treatment recommendations;

[0207] 4-6 levels: clinical monitoring is needed, such as "review every 3 months" and other treatment recommendations;

[0208] 7-10 levels: immediate intervention is needed, such as "drug treatment or surgical treatment" and other treatment recommendations.

[0209] According to the above level division, the preset threshold is initially set to a level difference of ≥3 levels. In clinical practice, a risk level difference of 3 levels means that the intervention strategy jumps from "no need to handle" to "immediate treatment", and if not arbitrated, it is likely to lead to misdiagnosis.

[0210] According to the typical conflict scenario, further describe the impact, for example, the telomere length shortening rate in the telomere mode is slow, so it is determined that its health risk level is 3, which belongs to low risk; the abundance of pro-inflammatory bacteria in the intestinal flora mode increases significantly, so it is determined that its health risk level is 7, which belongs to high risk; if the conclusion of the telomere mode may delay intervention, and the conclusion of the intestinal flora mode may over-medicate, the time effectiveness of the biological mechanism needs to be balanced through arbitration, and the risk level difference between the telomere mode and the intestinal flora mode is 4 levels (≥ 3 levels), so the conflict arbitration is triggered.

[0211] For the conflict judgment of risk level contradiction, the system generates a risk level vector for the evaluation conclusion of each mode, and then judges the contradiction through the absolute value of the vector difference, that is, if |level A - level B| ≥ 3, the conflict arbitration is triggered.

[0212] The conflict judgment of risk level contradiction is linked with the biomedical knowledge graph, that is, the correlation strength γ needs to be introduced when calculating the difference, and in the case of level contradiction between high correlation strength γ mode and low high correlation strength γ mode, the system automatically gives higher weight to the high correlation strength γ mode to assist the preliminary judgment.

[0213] The setting of clinical priority rules in step S6 is set according to the principle of emergency risk priority processing, which is divided into four levels according to the degree of emergency, and then the hierarchical calculation is carried out, which specifically includes:

[0214] (1) Assign weight coefficients to each level: the weight coefficients are adjusted based on the basis that the priority coefficients 1 to 4 decrease in turn;

[0215] (2) Dynamically adjust the weight: dynamically adjust the weight in combination with the evidence-based medicine level quantified in step S2;

[0216] (3) Determine the conclusion according to the comprehensive score: determine the priority conclusion according to the comprehensive score, which satisfies the formula:

[0217] Comprehensive score = R × W p × C

[0218] Wherein, R is the risk level, corresponding to the health risk quantification evaluation result output by the fusion feature in step S4; W p is the level weight coefficient, corresponding to the fixed weight of the clinical priority level; C is the evidence-based level correction coefficient, which is set based on the evidence-based medicine level (1-4 levels) of the biomedical knowledge graph in step S2, and higher level evidence corresponds to higher correction coefficient.

[0219] The specific four levels and application standards set according to the principle of emergency risk priority processing can be referred to the following table:

[0220]

[0221] Note that the division is based on the "Disease Risk Stratification Standard" in the "Emergency Medicine Clinical Guidelines (2024 Edition)", combined with the interactive characteristics of the "Telomere-Microbiota-Immune" system (such as immune indicators often more immediate than telomere indicators reflecting acute risk), to ensure that the levels are consistent with biomedical mechanisms.

[0222] In addition, the core of the weight coefficient adjustment is to prioritize the processing of emergency situations based on the decreasing priority levels 1 to 4. The weight coefficients in the table are obtained by Delphi method research, scoring "decision weight for different risk types", taking the mean value and normalizing to get the initial division. The data is not unique, but the weight coefficients should be integrated to 1 to ensure the exclusivity of weight distribution.

[0223] Dynamic weight adjustment means adjusting the weight based on the evidence-based medicine level. This means that if the evidence-based medicine level of a high-priority conclusion (evidence-based medicine level based on step S2) is lower than that of a low-priority conclusion, the system automatically adjusts the weight of the low-priority conclusion. In other words, if the conclusion supported by high evidence-based level evidence is more reliable, even if the priority is low, the weight should be appropriately increased; and if the conclusion supported by low evidence-based level evidence is low, even if the priority is high, the weight should be reduced to avoid misjudgment.

[0224] The dynamic adjustment formula is: adjusted weight = basic weight coefficient x evidence-based level correction coefficient.

[0225] For example, if the system detects two conflicting conclusions, including an immune cell modality indicating "acute infection risk" with a clinical priority of level 1, which is the highest priority, but the medical evidence supporting this conclusion is "animal experiment" with a evidence-based level of 4, which is the lowest evidence-based level; while the gut microbiota modality indicates "chronic intestinal inflammation progression risk" with a clinical priority of level 3, which is lower than acute, but the evidence supporting this conclusion is "multi-center RCT study" with a evidence-based level of 1, which is the highest evidence-based level. At this time, the system has conflicts between the two, and the system will not completely adopt the conclusion with the highest priority just because "acute infection risk has higher priority", but will adjust the weight and derive a new conclusion.

[0226] Further explanation, the real-time scanning of the fused feature F 融合 is realized by "rule engine + feature matching" dual mechanism, the specific process is as follows:

[0227] Scan trigger time: synchronized with step S4 feature fusion, every time a batch of fused features is generated, the scanning is immediately started to ensure that high-risk operations are intercepted before data transfer.

[0228] Feature analysis: the fused feature matrix F 融合The analysis is divided into two parts: data labels and association conclusions. The data labels include the types of user biomarkers involved (such as "telomere length" and "bifidobacterium abundance") and the data sensitivity level. The association conclusions include health risk assessments generated based on the characteristics (such as "telomere aging risk") and preliminary intervention recommendations.

[0229] Rule matching and interception: If the data label matches the unauthorized use scenario in the privacy protection rule (such as high-sensitive data marked as "for third-party research" without user authorization record), the data transmission is immediately intercepted and frozen, and a violation log is generated. If the association conclusion contains high-risk intervention recommendations in the clinical safety criteria (such as "recommendation of certain experimental drugs" with a risk level of 5, the recommendation is automatically shielded and replaced with a prompt "suggestion to consult a clinician").

[0230] The sensitivity classification of the data by the hierarchical authorization unit in step S7 specifically includes:

[0231] (a) High sensitivity level: core biomarker data that can directly associate individual health privacy, only allowed to be stored in the user's local terminal or encrypted channel in the cooperative medical institution with dynamic authorization, prohibited from cross-institution transmission;

[0232] (b) Medium sensitivity level: data that can be transmitted to designated health management institutions within the scope of user authorization for personalized intervention scheme generation, and the transmission process uses federated learning technology for feature desensitization;

[0233] (c) Low sensitivity level: data that can be used for population health statistical analysis after anonymization, no real-time authorization is required but needs to be synchronized with user usage records periodically.

[0234] The data sensitivity classification is based on the dual dimensions of "biomedical privacy risk" and "clinical application value". The biomedical privacy risk refers to core biomarkers that can directly associate individual identity and health privacy (such as genetic level data), and the leakage of these data may lead to discrimination (such as insurance refusal, employment restrictions). The clinical application value refers to the fact that high-sensitive data is crucial for personalized medicine (such as telomere length guiding anti-aging solutions), while low-sensitive data can serve population health research (such as analysis of bacterial distribution trends).

[0235] The data sensitivity level classification standard here refers to the "Personal Health Information Protection Framework" in the "Health Medical Data Security Guide (2024 Edition)", and combines the uniqueness of "telomere-bacteria-immunity" data (such as the fact that whole genome data of intestinal flora is more easily located than relative abundance data), ensuring that the classification logic meets the privacy protection needs of the medical scene.

[0236] The specific types of high-sensitivity data include telomere length raw detection values, telomerase activity quantitative results, intestinal flora whole genome sequencing data (including strain-specific gene sequences), absolute abundance, immune cell subpopulation ratio, and disease susceptibility assessment-related content generated based on the above data.

[0237] High-sensitivity data is only used to generate health reports for the user himself and is prohibited from being used for any form of population research or third-party commercial purposes. It is also only allowed to be stored in the private cloud of the user's local terminal or cooperative medical institutions using AES-256 encryption algorithm, with the key generated and exclusively held by the user terminal. In addition, high-sensitivity data is prohibited from being transmitted across institutions, and is only temporarily decrypted and transmitted to the designated emergency institution when the user initiates "emergency authorization" through the dynamic authorization panel, and the cache is automatically deleted after transmission.

[0238] The specific types of medium-sensitivity data include intestinal flora relative abundance, characteristic metabolite concentration, immune cell factor concentration range, and intervention plan recommendations generated based on the above data.

[0239] Medium-sensitivity data is only used to generate personalized intervention plans and is prohibited from being used for individual identity recognition or discriminatory decisions. It can be transmitted between institutions with user authorization, but must be desensitized through federated learning technology, that is, only model features are transmitted, not raw data. In addition, for medium-sensitivity data, the user can pre-set an "authorization whitelist" in the dynamic panel, which specifies the institutions authorized to access the data, and set the authorization time limit, which will automatically revoke the authorization when it expires.

[0240] The specific types of low-sensitivity data include detection date, detection institution name, anonymized description of health indicator trends (such as "telomere length of users in a certain age group shows a shortening trend" and "female intestinal flora diversity is higher than male"), and de-identified intervention effect statistics based on the above data.

[0241] Low-sensitivity data is only used for population health statistical analysis or optimization of algorithm models. The processing of low-sensitivity data requires the removal of individual identifiers through "k-anonymity" algorithm, that is, anonymization processing, and the removal of specific accurate content, such as age accurate to age intervals such as "40-50", rather than specific age, to avoid locating individuals through data combination. In addition, it should be noted that the system pushes a "low-sensitivity data usage report" to the user every quarter, listing data usage and statistical results, so that the user can turn off the population analysis authorization at any time.

[0242] Further explanation, the dynamic control process of hierarchical authorization, specifically includes:

[0243] (1) Automatic generation of hierarchical labels: After the completion of data collection in step S1, the system automatically adds a "sensitivity label" to each piece of data and matches it with the pre-set "biomarker sensitivity dictionary". It is important to note that the label is bound throughout the data flow and cannot be tampered with.

[0244] 2. Control the flow direction according to the sensitivity level: For high-sensitivity level data, store it in a hardware security area, trigger multi-factor authentication during transmission, and record the transmission log through blockchain to ensure data traceability; For medium-sensitivity level data, use the "local training + parameter aggregation" mode through the federated learning framework, only model gradients can be exchanged between institutions, and homomorphic encryption is used to ensure that the original information is not leaked during parameter transmission; For low-sensitivity level data, an independent de-identification engine is used for anonymization processing to ensure that the processed data cannot be reversed to identify.

[0245] The multi-factor authentication here refers to when the hospital system initiates a high-sensitivity data access request, the user terminal detects the high-sensitivity data access request and automatically enters verification, which includes biometric verification (comparing the pre-stored biological template), device feature verification (verifying whether the device initiating the request is in the "authorized whitelist"), and dynamic credential verification (verifying the verification code). After passing the three verifications, the system will temporarily decrypt the high-sensitivity data for the hospital system to access.

[0246] For the characteristics of medium-sensitivity level data, the federated learning framework here uses a lightweight federated sub-model, which inputs the feature vector of medium-sensitivity data into the model as an input item, and completes model training locally, i.e. on the server of the collection institution, only outputting the learned feature parameters. After training, the local model output does not contain any individual representation information and only reflects the statistical regularity of the data-related parameters, including "feature weight", "activation function threshold", etc. Then, the system will automatically filter non-privacy sensitive parameters through the pre-set parameter security filter, i.e. eliminate parameters that may reverse the original data, and only retain model structure parameters.

[0247] In addition, after completing the local training of the federated learning model, parameter aggregation is needed, i.e. through the federated learning coordinator, which can be a neutral health management alliance or other institution, to collect local parameters from each institution and fuse these parameters into a global model. The specific operation includes:

[0248] (1) Parameter upload and encrypted transmission: The parameters of each local institution need to be homomorphically encrypted before transmission, and then uploaded to the federated coordinator; homomorphic encryption means that the parameters can be calculated but not decrypted during transmission.

[0249] (2) Global parameter fusion: After receiving the local parameters delivered by each institution, the federal coordinator aggregates the parameters by weighted average to generate a global model; the weighted average aggregation parameter refers to assigning weights according to the data volume and data quality of each institution, and the specific calculation formula is:

[0250] W = a1 = E(W1) + a2 x E(W2) + … + a n x E(W n )

[0251] Wherein, a1 + a2 + … + a n = 1, a is the weight assigned according to the parameters of each institution.

[0252] To realize the use method of the AI health medical service recommendation system based on multi-dimensional biological detection as described above involves the module composition of the system, including:

[0253] (a) Biological detection module: for detecting the telomere length, intestinal flora composition and immune cell subpopulation ratio of the user;

[0254] (b) Data integration and processing module: including data acquisition unit, BioGraphFusion fusion unit and feature extraction unit, wherein the BioGraphFusion fusion unit is based on biomedical knowledge graph, encodes features through two layers of GCN, and combines cross-modal attention mechanism to fuse data;

[0255] (c) Artificial intelligence analysis module: including health status assessment unit, ethics review engine and medical service recommendation unit, wherein the ethics review engine intercepts operations that violate preset ethical rules in real time, and arbitrates conflicts in multi-modal data conclusions;

[0256] (d) User interaction module: including report display unit and hierarchical authorization unit, the hierarchical authorization unit controls data access authority according to data sensitivity grading, and supports users to dynamically adjust the authorization range.

[0257] Example 1: Health management service recommendation process for middle-aged male user (45 years old) based on multi-dimensional biological detection

[0258] Main background: User Zhang, 45 years old, IT industry practitioner, long-term sitting, irregular diet, fatigue increased in recent half year, occasional abdominal distension, actively to the community health management center for multi-dimensional biological detection, hope to obtain personalized health suggestion. The system generates intervention scheme through three core detections of telomere length, intestinal flora and immune cells combined with AI analysis. The specific process is as follows:

[0259] S1: biological data collection: the community health management center uses relevant equipment to collect data, and the relevant data of telomere detection is obtained. The length of telomere is 5.1 kb (reference value: 4.8-6.2 kb for a 45-year-old male normal range); the relevant data of intestinal flora detection is that the flora composition is composed of 18% relative abundance of Bifidobacterium (reference value 20%-30%), 22% relative abundance of Escherichia coli (reference value <15%), and 35 mmol / L concentration of short-chain fatty acid (reference value 40-60 mmol / L); and the relevant data of immune cell detection is that the proportion of CD4 + T cells is 32% (reference value 30%-40%), and the concentration of IL-6 is 12 pg / ml (reference value <7 pg / ml, slightly elevated).

[0260] S2: dynamically activate the causal chain in the knowledge graph in the BioGraphFusion fusion unit: the main content is:

[0261] (1) Core causal path: intestinal flora imbalance (decrease in Bifidobacterium and increase in Escherichia coli) affects the decrease in short-chain fatty acids, which in turn causes the increase in IL-6 (pro-inflammatory) leading to accelerated telomere damage;

[0262] At the same time, 3 multi-center RCT studies (evidence level 1) supporting "short-chain fatty acids inhibit IL-6 secretion" and 2 cohort studies (evidence level 2) supporting "IL-6 elevation is positively correlated with telomere shortening rate" are matched from the knowledge graph.

[0263] (2) Correlation strength quantification, i.e. calculating the correlation strength γ: based on the evidence level of 1 of the 3 RCT studies, the basic weight of the path "decrease in Bifidobacterium affecting decrease in short-chain fatty acids" is 0.7, and according to the characteristics of insufficient dietary fiber intake in the user's diet, the adaptability is high, and the correction coefficient is 1.1, and after comprehensive calculation, the final value is = 0.7 x 1.1 = 0.77;

[0264] And based on the 2 cohort studies, the basic weight of the path "IL-6 increase causing telomere damage" is 0.6, and according to the characteristics of the user without underlying diseases, the adaptability is moderate, and the correction coefficient is 1.0, and the final value of this path γ ＝0.6×1.0＝0.6 .

[0265] S3. Constructing a heterogeneous subgraph

[0266] First, construct the initial graph structure, where the nodes of the telomere modality are "telomere length 5.1 kb" and "telomere damage marker (not detected, default normal)", and the edge is "IL-6 increase causing telomere damage", and its weight is 0.6, taken from γ of S2;

[0267] The nodes of the intestinal flora mode are "Bifidobacterium 18%" and "Escherichia coli 22%", and the edge is "Bifidobacterium decline affecting short-chain fatty acid decline", and the weight thereof is 0.77; "Escherichia coli rise affecting short-chain fatty acid decline", and the weight thereof is 0.5;

[0268] The nodes of the immune mode are "CD4 + T cell 32%" and "IL-6 12 pg / ml", and the edge is "short-chain fatty acid decline affecting IL-6 rise", and the weight thereof is 0.8 based on the knowledge graph.

[0269] After the initial graph is constructed, encoding is performed based on a two-layer graph convolution network GCN, wherein the first layer GCN aggregates direct correlations and extracts local features; and the second layer GCN aggregates second-order correlations and generates global features, specifically, the intestinal flora mode subgraph outputs "flora imbalance-promoting inflammation" features, the immune mode subgraph outputs "mild inflammation" features, and the telomere mode subgraph outputs "low telomere length" features.

[0270] S4: The encoding features of each mode are fused through a biological correlation strength scheduling attention mechanism, and attention weighting calculation is performed. First, the attention mechanism scheduling is performed, that is, the cross-modal correlation matrix is initialized: based on the value of γ in S2, the correlation strength between the intestinal flora mode and the immune cell mode is 0.77, the correlation strength between the immune cell mode and the telomere mode is 0.6, and the rest of the irrelevant items are set to 0.

[0271] Then, the attention weight calculation is performed: after normalizing the cross-modal correlation matrix, the weight of the "Bifidobacterium affecting short-chain fatty acid affecting IL-6" chain is 0.7, which belongs to a high correlation weight, the weight of "Escherichia coli affecting IL-6" is 0.2, which belongs to a low correlation weight, and the weight of "CD4 + T cell affecting telomere" is 0.1, which belongs to no direct correlation weight;

[0272] Then, the feature fusion is performed: the synergistic effect of the high-weight chain is strengthened, and the fused features finally point to "flora imbalance driving mild inflammation and potentially accelerating telomere aging".

[0273] S5. Ethical review and conflict arbitration: the system scans the fused features and finds no high-risk operations such as "data used for insurance underwriting, intervention suggestion experimental drugs"; and after multi-modal conclusion consistency determination, it is determined that the flora, immune, and telomere data all point to "chronic inflammation related risk", there is no directional conflict, but the intestinal flora mode suggests "medium risk", the telomere mode suggests "low risk", and the difference between the two is <3 levels, so no arbitration is needed.

[0274] S6. Control the flow of data by sensitivity without conflict arbitration, that is, first conduct data sensitivity classification, where telomere length raw value (5.1 kb) and IL-6 accurate concentration (12 pg / ml) belong to high sensitive level data, which are only stored in the user's local health APP, and the community center can only view them after the user scans the code and authorizes; the relative abundance of flora (18% of Bifidobacterium) belongs to medium sensitive level data, which are transmitted to the cooperative health management company after being desensitized by federated learning for generating intervention programs; the detection date (June 2025) and the age interval (45-50 years old) belong to low sensitive level data, which are used for "middle-aged male flora health trend" group analysis after anonymization, and the user can view the statistical results in the APP.

[0275] S7. Generate personalized reports and intervention recommendations after data flow processing is completed. The specific health report content includes:

[0276] Visual display of biological correlation path diagram with support evidence (such as "based on 3 RCT studies") marked for each correlation;

[0277] Risk assessment: The patient has a risk of chronic inflammation and accelerated telomere aging;

[0278] Intervention suggestions: Intervention suggestions include diet adjustment, exercise guidance, and review plan, etc.

Claims

1. A method for using an AI-based health and medical service recommendation system based on multi-dimensional biological detection, characterized in that, Includes the following steps: S1. Collect user telomere length, gut microbiota, and immune cell data through the biodetection module; S2.BioGraphFusion fusion unit, based on biomedical knowledge graph, constructs causal pathways and correlation strength γ of telomere aging, microbiome metabolism and immune regulation; S3. Use a two-layer graph convolutional network (GCN) to encode the data of each biological modality and construct heterogeneous subgraphs corresponding to each modality; S4. The attention mechanism of biological association strength scheduling is used to fuse the coding features of each modality and perform attention weighting calculation; S5. The merged features intercept high-risk operations through the ethics review engine and trigger conflict arbitration when a conflict is detected; S6. Once the conflict arbitration unit of the ethics review engine is triggered, it will perform priority calculation according to the clinical priority rules. S7. If there is no conflict, the data will enter the hierarchical authorization unit, which will control the data flow according to the sensitivity level. S8. Finally, the report display unit outputs a personalized health report and medical recommendations.

2. The method of using the AI ​​health and medical service recommendation system based on multi-dimensional biological detection according to claim 1, characterized in that, Step S2 specifically includes: Causal path construction: Based on real-time biological detection data from users, dynamically activate matching causal chains from the biomedical knowledge graph; Association strength quantification: converting statistical correlations in medical literature into computable path strength γ; Path conflict arbitration: When multiple paths compete, the primary path is automatically selected based on evidence-based medicine level and intensity threshold.

3. The method of using the AI ​​health and medical service recommendation system based on multi-dimensional biological detection according to claim 2, characterized in that, The biomedical knowledge graph stores knowledge in the form of triples, where each triple specifically includes a first entity, a relation, and a second entity, wherein: The first entity and the second entity are selected from at least one of telomere-related biomarkers, gut microbiota-related biomarkers, immune cell-related biomarkers, metabolites, and disease states; The relationship is a causal link between the first entity and the second entity, selected from at least one of promotion, inhibition, mediation, and regulation.

4. The method of using the AI ​​health and medical service recommendation system based on multi-dimensional biological detection according to claim 1, characterized in that, Step S3 specifically includes: Initial graph structures were constructed for each biological modality data, including telomere length, gut microbiota, and immune cells. The constructed initial graph structures included corresponding biomarker nodes and causal relationship edges. The first-layer graph convolutional network (GCN) is used to aggregate the features of nodes and their immediate neighbors to extract local associations. The second-layer graph convolutional network (GCN) is used to aggregate node and second-order neighborhood features to fuse global causal logic. Finally, each modality-independent heterogeneous subgraph is generated, and the node type, edge relationship type, and feature embedding of the heterogeneous subgraph are all adapted to the biomedical characteristics of the corresponding modality.

5. The method of using the AI ​​health and medical service recommendation system based on multi-dimensional biological detection according to claim 1, characterized in that, The attention weight allocation in step S4, "attention mechanism for biological association strength scheduling," is based on the quantified association strength γ from step S2 as the core scheduling criterion. It integrates the weighted features of the three modalities: telomeres, gut microbiota, and immune cells. The specific calculation steps include: Input features and initialize the association matrix: Define modal feature matrices based on the heterogeneous subgraph encoding features of the three modal outputs in step S3, and construct a cross-modal association matrix R based on the causal relationships of cross-modal entities in the biomedical knowledge graph; Attention weight calculation: Attention weights are assigned based on the "biological association strength of cross-modal entities". After normalizing the association strength, the final attention weight matrix A is obtained by calculating the comprehensive weight through cross-modal attention weights. Attention weighted calculation and feature fusion: Based on the attention weight matrix A, the encoded features of each modality are weighted and fused. The specific formula is as follows: in, It is a stacked matrix of modality-coded features; F t The coding feature matrix of telomere modes; F m F is the encoding feature matrix of the gut microbiota modality; i This is the encoding feature matrix for immune cell modalities.

6. The method of using the AI ​​health and medical service recommendation system based on multi-dimensional biological detection according to claim 5, characterized in that, The attention-weighted computation and feature fusion process ensures the fusion effect through the following methods: Cross-modal features with strong biological associations are given high weights, and their synergistic effect is enhanced in the fusion features. Features with low or no biological association are assigned low weights to avoid interference from irrelevant features. The fused features retain the core information of each modality while integrating cross-modal causal relationship information.

7. The method of using the AI ​​health and medical service recommendation system based on multi-dimensional biological detection according to claim 5, characterized in that, In step S5, the ethics review engine scans the fused features in real time through a pre-set ethics rule base to intercept violations such as user privacy leaks, unauthorized data sharing, and high-risk intervention suggestions; when a conflict is detected in the multimodal conclusions, conflict arbitration is automatically triggered.

8. The method of using the AI ​​health and medical service recommendation system based on multi-dimensional biological detection according to claim 1, characterized in that, The clinical priority rules in step S6 are set according to the principle of prioritizing urgent risks, and are divided into four levels according to the degree of urgency, followed by a tiered calculation, specifically including: Assign weight coefficients to each level: the weight coefficients are adjusted based on the principle that the coefficients of priority levels 1 to 4 decrease sequentially; Dynamically adjust the weights: Combine the quantified evidence-based medicine level in step S2 to dynamically adjust the weights; Conclusion determined by overall score: The priority conclusion is determined based on the overall score, satisfying the formula: Overall score = R × W p ×C Where R represents the risk level, W p C is the hierarchical weighting coefficient, and C is the evidence-based level correction coefficient.

9. The method of using the AI ​​health and medical service recommendation system based on multi-dimensional biological detection according to claim 1, characterized in that, The sensitivity classification of data by the hierarchical authorization unit in step S7 specifically includes: High Sensitivity Level: Core biomarker data that can be directly linked to an individual's health privacy is only allowed to be stored on the user's local terminal or within an encrypted channel of a dynamically authorized partner medical institution, and cross-institutional transmission is prohibited; Medium sensitivity level: Data that can be transmitted to designated health management institutions within the scope authorized by the user for the generation of personalized intervention plans, and the transmission process uses federated learning technology for feature desensitization; Low Sensitivity Level: Data that has been anonymized and can be used for group health statistical analysis. No real-time authorization is required, but usage records need to be periodically synchronized with users.

10. A system for implementing an AI-based health and medical service recommendation method based on multi-dimensional biometric detection according to any one of claims 1-9, comprising: Biodetection module: used to detect the user's telomere length, gut microbiota composition, and immune cell subset ratio; The data integration and processing module includes a data acquisition unit, a BioGraphFusion fusion unit, and a feature extraction unit. The BioGraphFusion fusion unit is based on a biomedical knowledge graph and fuses data through two layers of GCN-encoded features combined with a cross-modal attention mechanism. Artificial intelligence analysis module: includes a health status assessment unit, an ethics review engine, and a medical service recommendation unit. The ethics review engine intercepts operations that violate pre-set ethical rules in real time and arbitrates conflicts in conclusions from multimodal data. User interaction module: includes report display unit and hierarchical authorization unit. The hierarchical authorization unit controls data access permissions according to data sensitivity and supports users to dynamically adjust the authorization scope.

Images