Method and system for time sequence perception dynamic updating of chronic disease knowledge graph

Abstract

The present application relates to the field of medical care information science, and discloses a kind of time sequence perception dynamic updating method and system of chronic disease knowledge graph, comprising: obtaining heterogeneous medical health data in chronic disease clinical event input stream and extracting incremental data stream;When time inversion occurs, call time sequence constraint rule, calibrate timestamp using median smoothing algorithm to construct time sequence causal association chain;According to calibration data, calculate the medical semantic dynamic attenuation coefficient of main drug relationship edge, and when data monitoring time interval exceeds 90 days, use discrete difference dynamic updating coefficient;When the coefficient is less than 0.50 and the complication association subgraph reaches the joint drug risk critical threshold, cut off the existing long-term medical order relationship edge and write in situ replacement drug relationship edge, the present application responds to the pathology phase change in multiple disease coexistence scene in real time, eliminates the time sequence noise of asynchronous diagnosis and treatment data stream, avoids graph space closed loop deadlock.

Description

Technical Field

[0001] This invention relates to a time-series-aware dynamic update method and system for chronic disease knowledge graphs, belonging to the field of healthcare informatics technology. Background Technology

[0002] Currently, chronic disease decision support systems typically improve the credibility of decision-making results by establishing medical knowledge bases and treatment pathway guide maps. These systems often employ fixed-period batch refresh methods, updating the map content in a comprehensive manner according to a preset cycle, or manually setting fixed expiration periods for entity relationships to maintain the timeliness of existing entity relationships in the knowledge graph. However, with multiple chronic diseases coexisting and continuous generation of treatment data, the fixed map topology needs to constantly receive asynchronous data from different systems. Patients may simultaneously suffer from multiple interrelated chronic diseases, and the data recording and entry times of each system will fluctuate frequently. Therefore, outpatient registration, examination reports... In the timestamps of medical record review and hospitalization orders, the order of events is reversed in about 5% to 12% of cases. During follow-up, there may also be cases where the data monitoring interval exceeds 90 days. The existing system lacks a mechanism to determine causal relationships by combining the sequence of different complications and clinical events. When processing multi-source asynchronous data, it is difficult to dynamically adjust the credibility of existing relationships in the graph based on changes in multiple pathological states of patients. As a result, the originally valid historical paths may be covered by a large amount of noisy data, and the asynchronous updates of subgraphs associated with different complications may conflict with each other, thus forming a logical closed loop of relationship edges in the same graph topology, causing update deadlock and repeated oscillations.

[0003] Existing technologies typically address these issues by shortening the refresh cycle or adding manual filtering rules. However, simply increasing the refresh frequency or the number of rules not only increases computational resource overhead but may also frequently sever the correct long-term medical order triplet relationships, causing the decision support system to output incorrect reverse medication recommendations. Such approaches merely adjust the update frequency or the number of rules and still cannot solve the mismatch between the static semantic structure of the knowledge graph and the continuous changes in multiple physiological states of patients and the asynchronous arrival of data from different sources. Moreover, not only is the static topological architecture of the knowledge graph limited in dealing with scenarios of multiple coexisting diseases, but existing knowledge extraction and graph control methods also have shortcomings. For example, Chinese invention patent application CN120388712A discloses a method and system for constructing a diagnosis and treatment knowledge graph based on clinical pathways, which uses a domain pre-trained language model to extract triplets from clinical pathway texts and statically... While these technologies can be improved, they implicitly rely on static text corpora that are logically continuous and have complete rules. In the complex working conditions of real-world medical environments, data such as outpatient registration, examination report review, and inpatient medical orders are entered asynchronously by different systems, resulting in approximately 5% to 12% of timestamps being out of order. During follow-up, data monitoring intervals may exceed 90 days. The aforementioned existing technologies lack the ability to self-calibrate the temporal sequence of multi-source asynchronous medical data streams and cannot determine causal relationships by combining the sequence of different complications and clinical events. When faced with such inverted timestamps or sparse cross-sectional data, it is difficult to dynamically adjust the reliability of existing relationships in the graph based on changes in multiple pathological states of the patient. Originally valid historical paths are easily covered by a large amount of noisy data, which in turn frequently cuts off the correct long-term medical order triplet relationship edges, and may even form logical closed loops of relationship edges in the same graph topology, causing update deadlocks and repeated oscillations.

[0004] Therefore, the technical problem to be solved by this invention is how to collect the longitudinal clinical event time-series flow of patients, extract patient phenotypic event nodes, calculate the dynamic decay coefficient of medical semantics, establish a graph update method triggered by clinical events, extract the temporal causal association chain, determine the update boundary accordingly, and write the newly added entity relationship edge that reflects the temporal causal relationship in situ. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of this invention is as follows: A time-aware dynamic update method for a chronic disease knowledge graph, comprising the following steps:

[0006] Step S1: Obtain heterogeneous medical and health data from the chronic disease clinical event input stream, and extract incremental data streams containing patient phenotype event nodes and main medication relationship edges from the heterogeneous medical and health data;

[0007] Step S2: Retrieve the generation time sequence of the drug treatment node and the biochemical test node in the incremental data stream. When there is a time sequence inversion, call the engineering and diagnosis time sequence difference in the standard clinical pathway time sequence constraint rules, and use the median smoothing algorithm to calibrate the abnormal timestamps contained in the generation time sequence into calibrated time sequence data that conforms to the monotonically increasing causal property, so as to construct a time sequence causal association chain that converges unidirectionally from pre-intervention to post-outcome.

[0008] Step S3: Calculate the dynamic decay coefficient of medical semantics of the main drug relationship edge based on the calibration time series data. When the data monitoring time interval between adjacent patient phenotypic event nodes exceeds 90 days, suspend the dynamic decay calculation based on continuous time series, extract the historical full-process data to calculate the global baseline drift rate parameter, and compare it with the numerical change rate of the current sparse data points. Use discrete difference to correct the historical static semantic weights to dynamically update the dynamic decay coefficient of medical semantics.

[0009] Step S4: When the medical semantic dynamic decay coefficient is lower than the preset elimination threshold of 0.50, and the complication association subgraph indicated by the temporal causal association chain reaches the critical threshold of combined drug use risk, a relation edge blocking instruction is sent to the graph database to cut off the existing long-term medical order triple relation edge, and the alternative drug relation edge derived from the temporal causal association chain is written in place.

[0010] Preferably, step S2 includes the following sub-steps: Step S21, extracting the issuance timestamp of the drug treatment node and the review timestamp of the biochemical test node from the incremental data stream; Step S22, when the review timestamp is earlier than the issuance timestamp, retrieving the time difference between medical and clinical practice in the standard clinical pathway time constraint rules; Step S23, using the median smoothing algorithm to calculate the median of the historical timestamp sequence, and correcting the review timestamp to the algebraic sum of the issuance timestamp and the median to generate calibrated time series data.

[0011] Preferably, step S3 calculates the dynamic attenuation coefficient of the medical semantics of the primary drug relationship edge based on the calibration time series data, including the following sub-steps: step S31, extracting the time difference between adjacent chronic disease clinical events on the calibration time series data; step S32, determining the semantic attenuation increment based on the product of the time difference and the baseline attenuation rate; step S33, subtracting the semantic attenuation increment from the preceding medical semantic weight at the current moment to obtain the dynamic attenuation coefficient of the medical semantics of the primary drug relationship edge.

[0012] Preferably, in step S4, the temporal causal association chain indicates that the coexisting complication association subgraph has reached the critical threshold for combined drug use risk, including the following sub-steps: Step S41, identify complication nodes and chronic disease medication nodes belonging to the category of coexisting chronic diseases from the temporal causal association chain to construct a complication association subgraph; Step S42, count the dimension of frequent itemsets of drug combination effects in the complication association subgraph; Step S43, when the dimension of frequent itemsets reaches the preset combined drug use threshold, determine that the complication association subgraph has reached the critical threshold for combined drug use risk.

[0013] Preferably, step S4, which involves sending a relation edge blocking instruction to the graph database to sever existing long-term medical order triplet relation edges and writing alternative medication relation edges derived from the temporal causal association chain in situ, includes the following sub-steps: Step S44, sending a request to the graph database to disable relation edges between the head and tail entities in the long-term medical order triplet to change the time-state attribute of the long-term medical order triplet relation edges; Step S45, extracting new causal alternative node pairs from the temporal causal association chain; Step S46, writing alternative medication relation edges containing new causal alternative node pairs in situ after the long-term medical order triplet is disconnected to complete the reconstruction of the graph topology.

[0014] Preferably, when there are multiple semantic conflict update requests for complications in step S4, step S4 includes the following sub-steps: Step S47, call the standard chronic disease clinical risk weight matrix to extract the core vital signs evaluation indicators corresponding to the mutually exclusive triples in the multiple semantic conflict update requests for complications; Step S48, calculate the dynamic priority weight of each update request in real time based on the product of the core vital signs evaluation indicators and the standard chronic disease clinical risk weight matrix; Step S49, compare the dynamic priority weights numerically, assign the high-priority chain to the topology structure for in-situ update and redirect the low-priority chain to the temporary observation branch topology node for isolation and masking.

[0015] Preferably, the standard chronic disease clinical risk weight matrix called in step S47 includes risk sensitivity coefficients for heart rate, blood pressure, blood glucose, blood lipids and renal function indicators under different combinations of complications. The risk sensitivity coefficients are positively correlated with the degree of acute risk impact of each complication on vital signs.

[0016] Preferably, the alternative medication relationship edge derived from the temporal causal association chain in step S4 includes the following sub-steps: Step S410, mapping the intervention node and outcome node in the temporal causal association chain to the drug entity and efficacy entity in the standard clinical pathway; Step S411, retrieving the temporal constraint rules of the standard clinical pathway and verifying the causal rationality of the one-way dependency relationship between the drug entity and the efficacy entity; Step S412, after verifying the causal rationality of the one-way dependency relationship, encapsulating the drug entity and the efficacy entity into an alternative medication relationship edge.

[0017] Preferably, the in-situ writing of alternative medication relationship edges derived from the temporal causal association chain in step S4 further includes the following sub-steps: Step S413, performing common sense reasoning verification on the full topological path of the written chronic disease knowledge graph to search for the existence of mutually exclusive medication relationship edges or temporal inversion intervention links; Step S414, if mutually exclusive medication relationship edges are found, suspending the current in-situ writing operation and redirecting the alternative medication relationship edges to the manual review queue to maintain the stability of the graph topology.

[0018] A time-aware dynamic update system for a chronic disease knowledge graph, which implements a time-aware dynamic update method for a chronic disease knowledge graph, includes:

[0019] The data processing unit is used to acquire heterogeneous medical and health data from the input stream of chronic disease clinical events and extract incremental data streams containing patient phenotype event nodes and main medication relationship edges from the heterogeneous medical and health data.

[0020] The association unit is used to retrieve the generation time sequence of the drug treatment node and the biochemical test node in the incremental data stream. When there is a time sequence inversion, it calls the engineering and diagnosis time sequence difference in the standard clinical pathway time sequence constraint rules and uses the median smoothing algorithm to calibrate the abnormal timestamps contained in the generation time sequence into calibrated time sequence data that conforms to the causal monotonous growth property, so as to construct a time sequence causal association chain that converges unidirectionally from pre-intervention to post-outcome.

[0021] The calculation unit is used to calculate the dynamic decay coefficient of the medical semantics of the main drug relationship edge based on the calibration time series data. When the data monitoring time interval between adjacent patient phenotypic event nodes exceeds 90 days, the dynamic decay calculation based on continuous time series is suspended, the global baseline drift rate parameter is calculated by extracting the historical full-volume disease course data, and it is compared with the numerical change rate of the current sparse data points. The historical static semantic weight is corrected by discrete difference to dynamically update the dynamic decay coefficient of the medical semantics.

[0022] The graph update unit is used to send a relation edge blocking instruction to the graph database when the medical semantic dynamic decay coefficient is lower than the preset elimination threshold of 0.50 and the temporal causal association chain indicates that the coexisting complication association subgraph reaches the critical threshold of combined drug use risk. This instruction cuts off the existing long-term medical order triple relation edge and writes the alternative drug relation edge derived from the temporal causal association chain in situ.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. In the time-series-aware dynamic update of the chronic disease knowledge graph, the medical semantic dynamic decay coefficient of the main medication relationship edge is dynamically adjusted according to the time interval of adjacent patient phenotypic event nodes and the changes in physiological monitoring indicators of complications. This allows the credibility of existing diagnosis and treatment relationships in the graph to be updated in a timely manner as the patient's condition changes. When key physiological indicators change significantly, the weights of related relationship edges can be adjusted accordingly, thereby enabling the chronic disease knowledge graph to reflect the actual disease course changes of patients with multiple coexisting diseases more promptly and improving the adaptability of graph updates.

[0025] 2. By examining the sequential relationship between drug treatment nodes, biochemical test nodes, and clinical outcome nodes, and by using standard clinical pathway temporal constraint rules and median smoothing algorithm to calibrate abnormal timestamps, the reasonable temporal order of clinical events can be restored, reducing temporal interference caused by supplementary data and asynchronous recording, preventing the system from mistaking supplementary data for the latest changes in the patient's condition, and improving the reliability of the temporal causal relationship chain.

[0026] 3. When there is a long period of missing clinical events in patients and the monitoring data is relatively sparse, the dynamic decay coefficient of medical semantics can be updated discretely based on the changes in historical full-course data and current data. This can avoid the divergence of calculation results caused by continuing to calculate decay based on continuous time. As a result, the timeliness of the judgment of the entity relationship in the graph can be maintained during a long monitoring gap period, and the long-term lack of updating of knowledge relationships caused by the interruption of follow-up data can be reduced. Attached Figure Description

[0027] Figure 1 This is a flowchart of the entity relationship edge update process for the chronic disease knowledge graph of this invention;

[0028] Figure 2 This is a structural diagram of the chronic disease knowledge graph updating device of the present invention.

[0029] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0031] A time-aware dynamic update method for a chronic disease knowledge graph includes the following steps:

[0032] Step S1: Obtain heterogeneous medical and health data from the chronic disease clinical event input stream, and extract incremental data streams containing patient phenotype event nodes and main medication relationship edges from the heterogeneous medical and health data;

[0033] Step S2: Retrieve the generation time sequence of the drug treatment node and the biochemical test node in the incremental data stream. When there is a time sequence inversion, call the engineering and diagnosis time sequence difference in the standard clinical pathway time sequence constraint rules, and use the median smoothing algorithm to calibrate the abnormal timestamps contained in the generation time sequence into calibrated time sequence data that conforms to the monotonically increasing causal property, so as to construct a time sequence causal association chain that converges unidirectionally from pre-intervention to post-outcome.

[0034] Step S3: Calculate the dynamic decay coefficient of medical semantics of the main drug relationship edge based on the calibration time series data. When the data monitoring time interval between adjacent patient phenotypic event nodes exceeds 90 days, suspend the dynamic decay calculation based on continuous time series, extract the historical full-process data to calculate the global baseline drift rate parameter, and compare it with the numerical change rate of the current sparse data points. Use discrete difference to correct the historical static semantic weights to dynamically update the dynamic decay coefficient of medical semantics.

[0035] Step S4: When the medical semantic dynamic decay coefficient is lower than the preset elimination threshold of 0.50, and the complication association subgraph indicated by the temporal causal association chain reaches the critical threshold of combined drug use risk, a relation edge blocking instruction is sent to the graph database to cut off the existing long-term medical order triple relation edge, and the alternative drug relation edge derived from the temporal causal association chain is written in place.

[0036] Preferably, step S2 includes the following sub-steps: Step S21, extracting the issuance timestamp of the drug treatment node and the review timestamp of the biochemical test node from the incremental data stream; Step S22, when the review timestamp is earlier than the issuance timestamp, retrieving the time difference between medical and clinical practice in the standard clinical pathway time constraint rules; Step S23, using the median smoothing algorithm to calculate the median of the historical timestamp sequence, and correcting the review timestamp to the algebraic sum of the issuance timestamp and the median to generate calibrated time series data.

[0037] Preferably, step S3 calculates the dynamic attenuation coefficient of the medical semantics of the primary drug relationship edge based on the calibration time series data, including the following sub-steps: step S31, extracting the time difference between adjacent chronic disease clinical events on the calibration time series data; step S32, determining the semantic attenuation increment based on the product of the time difference and the baseline attenuation rate; step S33, subtracting the semantic attenuation increment from the preceding medical semantic weight at the current moment to obtain the dynamic attenuation coefficient of the medical semantics of the primary drug relationship edge.

[0038] Preferably, in step S4, the temporal causal association chain indicates that the coexisting complication association subgraph has reached the critical threshold for combined drug use risk, including the following sub-steps: Step S41, identify complication nodes and chronic disease medication nodes belonging to the category of coexisting chronic diseases from the temporal causal association chain to construct a complication association subgraph; Step S42, count the dimension of frequent itemsets of drug combination effects in the complication association subgraph; Step S43, when the dimension of frequent itemsets reaches the preset combined drug use threshold, determine that the complication association subgraph has reached the critical threshold for combined drug use risk.

[0039] Preferably, step S4, which involves sending a relation edge blocking instruction to the graph database to sever existing long-term medical order triplet relation edges and writing alternative medication relation edges derived from the temporal causal association chain in situ, includes the following sub-steps: Step S44, sending a request to the graph database to disable relation edges between the head and tail entities in the long-term medical order triplet to change the time-state attribute of the long-term medical order triplet relation edges; Step S45, extracting new causal alternative node pairs from the temporal causal association chain; Step S46, writing alternative medication relation edges containing new causal alternative node pairs in situ after the long-term medical order triplet is disconnected to complete the reconstruction of the graph topology.

[0040] Preferably, when there are multiple semantic conflict update requests for complications in step S4, step S4 includes the following sub-steps: Step S47, call the standard chronic disease clinical risk weight matrix to extract the core vital signs evaluation indicators corresponding to the mutually exclusive triples in the multiple semantic conflict update requests for complications; Step S48, calculate the dynamic priority weight of each update request in real time based on the product of the core vital signs evaluation indicators and the standard chronic disease clinical risk weight matrix; Step S49, compare the dynamic priority weights numerically, assign the high-priority chain to the topology structure for in-situ update and redirect the low-priority chain to the temporary observation branch topology node for isolation and masking.

[0041] Preferably, the standard chronic disease clinical risk weight matrix called in step S47 includes risk sensitivity coefficients for heart rate, blood pressure, blood glucose, blood lipids and renal function indicators under different combinations of complications. The risk sensitivity coefficients are positively correlated with the degree of acute risk impact of each complication on vital signs.

[0042] Preferably, the alternative medication relationship edge derived from the temporal causal association chain in step S4 includes the following sub-steps: Step S410, mapping the intervention node and outcome node in the temporal causal association chain to the drug entity and efficacy entity in the standard clinical pathway; Step S411, retrieving the temporal constraint rules of the standard clinical pathway and verifying the causal rationality of the one-way dependency relationship between the drug entity and the efficacy entity; Step S412, after verifying the causal rationality of the one-way dependency relationship, encapsulating the drug entity and the efficacy entity into an alternative medication relationship edge.

[0043] Preferably, the in-situ writing of alternative medication relationship edges derived from the temporal causal association chain in step S4 further includes the following sub-steps: Step S413, performing common sense reasoning verification on the full topological path of the written chronic disease knowledge graph to search for the existence of mutually exclusive medication relationship edges or temporal inversion intervention links; Step S414, if mutually exclusive medication relationship edges are found, suspending the current in-situ writing operation and redirecting the alternative medication relationship edges to the manual review queue to maintain the stability of the graph topology.

[0044] A time-aware dynamic update system for a chronic disease knowledge graph, which implements a time-aware dynamic update method for a chronic disease knowledge graph, includes:

[0045] The data processing unit is used to acquire heterogeneous medical and health data from the input stream of chronic disease clinical events and extract incremental data streams containing patient phenotype event nodes and main medication relationship edges from the heterogeneous medical and health data.

[0046] The association unit is used to retrieve the generation time sequence of the drug treatment node and the biochemical test node in the incremental data stream. When there is a time sequence inversion, it calls the engineering and diagnosis time sequence difference in the standard clinical pathway time sequence constraint rules and uses the median smoothing algorithm to calibrate the abnormal timestamps contained in the generation time sequence into calibrated time sequence data that conforms to the causal monotonous growth property, so as to construct a time sequence causal association chain that converges unidirectionally from pre-intervention to post-outcome.

[0047] The calculation unit is used to calculate the dynamic decay coefficient of the medical semantics of the main drug relationship edge based on the calibration time series data. When the data monitoring time interval between adjacent patient phenotypic event nodes exceeds 90 days, the dynamic decay calculation based on continuous time series is suspended, the global baseline drift rate parameter is calculated by extracting the historical full-volume disease course data, and it is compared with the numerical change rate of the current sparse data points. The historical static semantic weight is corrected by discrete difference to dynamically update the dynamic decay coefficient of the medical semantics.

[0048] The graph update unit is used to send a relation edge blocking instruction to the graph database when the medical semantic dynamic decay coefficient is lower than the preset elimination threshold of 0.50 and the temporal causal association chain indicates that the coexisting complication association subgraph reaches the critical threshold of combined drug use risk. This instruction cuts off the existing long-term medical order triple relation edge and writes the alternative drug relation edge derived from the temporal causal association chain in situ.

[0049] Example 1: Current standardized medical information service nodes process the input flow of chronic disease clinical events accompanying type 2 diabetes and hypertension. In the heterogeneous medical and health data input into the hospital information system, 5% to 12% of outpatient registration, examination report review, and inpatient medical order records have reversed timestamp misalignment. In home follow-up or community management data, there are also sparse cross-sectional data with data monitoring intervals of more than 90 days between adjacent chronic disease clinical events. Traditional medical and health information systems adopt a fixed-cycle batch processing refresh architecture, treating timeliness degradation as a single-dimensional absolute time decay. When the entity attributes or paths in a single subgraph change, this architecture is difficult to establish a dynamic correspondence between multi-dimensional pathological states and semantic credibility. It is also difficult to distinguish between entity attribute changes caused by disease progression and data heterogeneity conflicts caused by recording noise when incremental diagnosis and treatment data are continuously input. It is easy to generate logical closed-loop deadlocks, update oscillations, or erroneous severing of long-term medical order triple relationship edges in the same topological space of the knowledge graph, thereby outputting high-risk reverse medication suggestions or incorrect complication prediction results.

[0050] The data processing unit collects heterogeneous medical and health data and identifies newly added or modified records using patients, event types, and generation times as indexes. For records reflecting patient conditions, such as diagnoses, test results, and vital signs, the data processing unit generates patient phenotypic event nodes. For medication orders, it extracts drug entities and their treatment relationships with corresponding patient phenotypic event nodes, generating master-drug relationship edges, thus forming a de-identified incremental data stream. The association unit retrieves the generation time sequence of drug treatment nodes and biochemical test nodes in the incremental data stream, extracting the prescription timestamp of the drug treatment node and the review timestamp of the biochemical test node, respectively. When the timestamp is earlier than the issuance timestamp, the associated unit retrieves the work-consultation time difference from the time constraint rules of the standard clinical pathway. These rules include the work-consultation time difference within the same treatment cycle, as well as the unidirectional dependency between drug intervention and biochemical test results and clinical outcomes. The work-consultation time difference is the standardized benchmark time interval required from the attending physician's issuance of a drug prescription to the laboratory department's review and output of the corresponding biochemical test report. It is stored in a structured time-domain configuration file as an integer value in minutes. For the routine outpatient clinical pathway for type 2 diabetes combined with hypertension, this work-consultation time difference is preset to 45 minutes based on the historical average outpatient process time.

[0051] The association unit extracts the issuance and review timestamps from the confirmed historical records under the same standard clinical pathway, forms a historical timestamp sequence based on the time interval between the two, and calculates the median of the sequence using a median smoothing algorithm. If a time sequence inversion is detected, the abnormal review timestamp is corrected to the algebraic sum of the issuance timestamp and the median, generating calibrated time-series data. After calibration, the association unit sorts the patient phenotypic event nodes according to the calibrated timestamps and connects the drug treatment node to the subsequently generated biochemical test node or clinical outcome node according to the time-series constraints of the standard clinical pathway. Only connections where the pre-intervention time is earlier than the post-outcome time and conforms to a unidirectional dependency relationship are retained, thereby constructing a time-series causal association chain that converges unidirectionally from the pre-intervention to the post-outcome.

[0052] The computation unit receives calibration time-series data, extracts the absolute time step between adjacent patient phenotypic event nodes, and converts the absolute time step into a time difference. The intrinsic baseline evolution rate constant for the specified chronic disease type in memory is used as the baseline decay rate. The computation unit determines the semantic decay increment based on the product of the time difference and the baseline decay rate, and then subtracts this semantic decay increment from the preceding medical semantic weights at the current moment to obtain the dynamic medical semantic decay coefficient of the primary medication relationship edge. To reflect the impact of fluctuations in complication physiological parameters on semantic credibility, the computation unit introduces a cross-domain coupling variance factor through a physiological and pathological time-effect function model to dynamically scale the semantic credibility of historical diagnostic and treatment protocols. The formula for calculating the dynamic medical semantic decay coefficient is as follows: ,in, The output medical semantic dynamic decay coefficient, This represents the absolute time step between adjacent patient phenotypic event nodes. The intrinsic baseline evolution rate constant for specifying the chronic disease type in memory is used as the baseline decay rate. The preset state adjustment mapping coefficient; The cross-domain coupling variance factor is determined by the statistical variance of the array of physiological monitoring indicators of complications read continuously within the sliding-time observation window. The right side of the formula is expressed as... and The product of the two numbers is used as the dividend, with the value 1 and... and The algebraic sum of the products is used as the divisor. The data processing unit reads the array of fasting blood glucose data collected continuously within the sliding time observation window and calculates its statistical variance, then multiplies the variance by a factor corresponding to the dimensions of the relevant physiological monitoring indicators. The variance term with physical dimensions is converted into a dimensionless value and then added to the value 1.

[0053] When the fasting blood glucose level in the chronic disease clinical event input stream crosses the clinical risk boundary of 7 mmol / L from the stable range and undergoes a non-steady-state change, the cross-domain coupling variance factor... Step increase, dynamic decay coefficient of medical semantics This results in a non-smooth, step-like downsizing. If the data monitoring time interval between adjacent patient phenotypic event nodes exceeds 90 days, the computation unit suspends the dynamic decay calculation based on continuous time series, activates the baseline drift adaptive calibration operator, extracts adjacent effective sampling points from the historical full-course data, calculates the ratio of the index change of each adjacent effective sampling point to the sampling interval, and takes the average value as the global baseline drift rate parameter. The difference between the current sparse data point's index value and the index value of the previous valid sample point is then divided by the current sparse time step. The slope of the single-point index change was obtained. The computing unit will and The comparison is performed, and the historical static semantic weights are corrected through discrete difference. The discretization fine-tuning formula is as follows: ,in, The preceding medical semantic weights are corrected using discrete difference. Historical static semantic weights Given the current sparse time step, This represents the global baseline drift rate. For the slope of change of a single-point index, the calculation unit will... As the preceding medical semantic weights at the current moment, the semantic decay increment is determined based on the current sparse time step and the baseline decay rate, and then from... The semantic attenuation increment is subtracted to dynamically update the medical semantic dynamic attenuation coefficient. .

[0054] The graph update unit identifies complication nodes and chronic disease medication nodes belonging to the category of coexisting chronic diseases from the temporal causal relationship chain, constructs a complication association subgraph, and merges the co-occurring chronic disease medication nodes into drug combination records within the same monitoring window. It also counts the drug combinations that appear repeatedly in different monitoring windows and uses the maximum number of drug types contained in each repeated drug combination as the frequent itemset dimension. The preset combination drug use threshold is calibrated according to the number of drug types allowed for the corresponding complication combination in the standard clinical pathway. When the frequent itemset dimension reaches the preset combination drug use threshold, the complication association subgraph is determined to have reached the critical threshold for combination drug use risk.

[0055] When the medical semantic dynamic decay coefficient of the primary medication relationship edge is lower than the preset elimination threshold of 0.50, and the complication association subgraph reaches the critical threshold for combined medication risk, the graph update unit sends a request to the graph database to disable the relationship edge between the head entity and the tail entity in the long-term medical order triplet. This changes the time-state attribute of the long-term medical order triplet relationship edge to disabled, thereby severing the existing long-term medical order triplet relationship edge. Subsequently, intervention nodes and outcome nodes are extracted from the temporal causal association chain and mapped to drug entities and efficacy entities in the standard clinical pathway, respectively. Then, the temporal constraint rules of the standard clinical pathway are retrieved to verify whether the unidirectional dependency relationship between the drug entity and the efficacy entity conforms to the causal order. The verified drug entities and efficacy entities constitute a new causal replacement node pair and are encapsulated as a replacement medication relationship edge, which is written to the graph topology position where the long-term medical order triplet relationship edge was before it was disconnected.

[0056] After the alternative medication relationship edge is written, the graph update unit uses the standard clinical pathway temporal constraint rules and the existing mutually exclusive medication relationships in the graph database as constraints to perform common sense reasoning verification on the full topological path of the written chronic disease knowledge graph. It searches for the existence of mutually exclusive medication relationship edges or temporally reversed intervention links. If a mutually exclusive medication relationship edge is found, the current in-situ writing operation is suspended and the alternative medication relationship edge is redirected to the manual review queue. If no such conflict is found, the in-situ writing result is retained.

[0057] When multiple complication semantic conflict update requests simultaneously trigger the main graph update, and the parallel temporal causal association chains are logically mutually exclusive, the graph update unit calls the standard chronic disease clinical risk weight matrix. This matrix corresponds to each row with different complication combinations, and each column with heart rate, blood pressure, blood glucose, blood lipids, renal function indicators, and corresponding biochemical test indicators. The matrix elements are the risk sensitivity coefficients of each indicator. The risk sensitivity coefficients increase as the acute risk impact of the corresponding complication on vital signs increases. In this embodiment, the core evaluation indicators of vital signs corresponding to the mutually exclusive triples are extracted, including blood potassium concentration and creatinine value as a renal function indicator. Each indicator value is multiplied by the corresponding risk sensitivity coefficient in the matrix and summed to obtain the dynamic priority weight of each update request. The graph update unit performs numerical comparison of the dynamic priority weights, assigns high-priority chains to the original topology position for updating, and redirects low-priority chains to temporary observation branch topology nodes for isolation and masking storage. After the subsequent chronic disease clinical event input stream outputs new biochemical test indicator balanced data, the isolation and masking state of the low-priority chains is released, and they are gradually merged into the main graph.

[0058] Through the synergistic effect of cross-complication semantic temporal cross-gated topology adjustment mechanism and distributed temporal calibration closed loop, the system automatically restores the monotonic growth of temporal sequence when faced with timestamp distortion caused by supplementary data or non-ideal engineering environment and data loss caused by cross-hospital visits. This reduces the graph update response latency of the knowledge graph in complex multi-concurrent chronic disease scenarios from hours to less than 15ms, and reduces the occasional redundancy error rate of clinical pathway recommendations by more than 42%. It also completes the interpretable fusion of heterogeneous data in the medical information system and the dynamic credibility maintenance of the knowledge base of the chronic disease decision support system.

[0059] Example 2: The current standardized medical information service node establishes a sample database of long-term chronic disease clinical events containing 10,000 patients with diabetes and hypertension. This sample database is reconstructed from publicly available clinical datasets to form a simulated multi-center medical information warehouse, which includes 5% to 12% timestamp reverse misalignment data and multiple time-domain sparse sections with sampling intervals extended from 10 days to 120 days. This is used to test the operation results of the data processing unit, association unit, calculation unit, and map update unit under different input conditions.

[0060] The calculation unit determines the length of the sliding time observation window based on the patient's postprandial blood glucose fluctuation cycle and the average outpatient visit cycle. When the coefficient of variation of fasting blood glucose levels between adjacent sampling points increases, the window length converges to 30 days to balance the identification of pathological changes and the update frequency of the map topology; in the testing condition of diabetes mellitus combined with hypertension, The calibration period is 30 days.

[0061] The data processing unit retrieves the input stream sample library of long-term chronic disease clinical events and divides the samples into three groups: the present invention sample group using a complete graph to update the topology, the control group using traditional techniques, and a partially missing control group without a physiological and pathological time-effect function model. When the present invention sample group detects that the review timestamp of the biochemical test node is earlier than the prescription timestamp of the corresponding drug treatment node, it retrieves the time difference between medical and clinical procedures from the standard clinical pathway time-series constraint rules, corrects the time-series inversion using a median smoothing algorithm, and generates calibrated time-series data that conforms to causal monotonous growth. The calculation unit receives the calibrated time-series data, extracts the absolute time step between adjacent patient phenotypic event nodes, and runs the physiological and pathological time-effect function model. This model reads the complication physiological monitoring index data array within the sliding time observation window and calculates the cross-domain coupling variance factor. , and then and The product is used as the dividend, with 1 and 1 as the divisor. and The algebraic sum of the products, used as the divisor, yields the dynamic decay coefficient of medical semantics. .

[0062] During system initialization and configuration, the main control processor reads the pre-set chronic disease baseline risk mapping table from memory. The baseline value was set to 0.02. Before the system was put into operation, a simulation test was conducted using 100 desensitized sample streams with known disease progression history. During the stable period... The benchmark fluctuation did not cause When abnormal disturbances exceeding 5% occur, maintain the value at 0.02; when the coefficient of variation exceeds the preset safety boundary, adjust in increments of 0.005. Until the fasting blood glucose level crosses the clinical risk threshold, It exhibits a recognizable step-down pattern.

[0063] To test the update process under extremely long surveillance gaps, the sample simulated interruptions in patient home follow-up and underreporting of community data. When the data monitoring time interval between adjacent patient phenotypic event nodes exceeded 90 days, the control group, lacking long-term adaptive adjustment, exhibited... In cases where invalidation is identified and the long-term medical order triplet relationship edges are severed, the sample group of this invention suspends the dynamic decay calculation based on continuous time series, initiates the baseline drift adaptive calibration operator, and calculates the global baseline drift rate parameter from the historical full-course data. And compare it with the slope of the single-point index change of the current sparse data points. Comparison is performed, and the computing unit uses and product divided by The negative value of the resulting ratio is used as the exponent to correct the historical static semantic weights. ,get ; then with The preceding medical semantic weights, used as the basis for the current moment, are dynamically updated according to the process of determining the semantic decay increment. .

[0064] After completing the operation of each group, the measured values ​​of cross-domain coupling variance factor, topology reconstruction response delay, and recommendation redundancy error rate were recorded respectively. Under the input conditions of a timestamp misalignment ratio of 5% and a data monitoring interval of 30 days, sample number 1 adopted the complete processing procedure of the sample group of this invention. The measured value was 1.14, the topology reconstruction response latency was 12.6ms, and the recommended redundancy error rate was 1.3%; sample number 2 was treated as a control group. The measured value was 0.00, the topology reconstruction response latency was 1850.4ms, and the recommended redundancy error rate was 15.6%; sample number 3 was treated as a partially missing control group. The measured value is 1.08, the topology reconfiguration response latency is 412.3ms, and the recommended redundancy error rate is 8.4%.

[0065] After increasing the timestamp misalignment rate to 12% and adjusting the data monitoring interval to 45 days, sample number 4 was obtained through the sample group processing of this invention. The measured value is 2.56, the topology reconstruction response latency is 14.3ms, and the recommended redundancy error rate is 1.8%; sample number 5 was obtained after processing the control group. The measured value was 0.00, the topology reconstruction response latency was 2431.1 ms, and the recommended redundancy error rate was 22.4%; sample number 6 was obtained after processing with a partially missing control group. The measured value is 2.31, the topology reconfiguration response latency is 518.7ms, and the recommended redundancy error rate is 11.2%.

[0066] Under the condition that the timestamp misalignment rate remains at 12% and the data monitoring interval is extended to 120 days, the sample group of this invention processes the data of the ultra-long monitoring gap period according to the discrete difference correction process. Sample number 7 The measured value was 5.83, the topology reconstruction response latency was 14.8ms, and the recommended redundancy error rate was 2.1%; sample number 8 was treated as a control group. The measured value was 0.00, the topology reconstruction response latency was 3612.5ms, and the recommended redundancy error rate was 43.7%; sample number 9 was treated as a partially missing control group. The measured value is 5.12, the topology reconfiguration response latency is 894.2ms, and the recommended redundancy error rate is 19.5%.

[0067] As the proportion of timestamp misalignment and the data monitoring time interval increase, the data measured by the sample group of this invention... The physiological parameters of complications increase with fluctuations. When the corresponding inflection point threshold is reached, A non-smooth, step-like downgrade occurs; after it approaches and falls below the elimination threshold of 0.50, the graph update unit enters the in-situ writing process of the blocking and alternative medication relationship edges of the long-term medical order triplet relationship edges. When multiple complication semantic conflict update requests occur simultaneously, the graph update unit calls the standard chronic disease clinical risk weight matrix to calculate the dynamic priority weight, assigns the high-priority chain to the original topology position for updating, and redirects the low-priority chain to the temporary observation branch topology node for isolation and masking.

[0068] With 5% to 12% timestamp reverse misalignment data and a data monitoring interval of 30 to 120 days, the knowledge graph update response latency of the sample group of this invention remained within 15ms. The topology reconstruction response latency and recommendation redundancy error rate of the control group and the partially missing control group increased with the increase of timestamp misalignment ratio and data monitoring interval. The above operating data show that after the data processing unit, association unit, calculation unit and graph update unit process the incremental data stream in sequence according to the calibration time series data, they can maintain the unidirectional order of the time series causal association chain and complete the dynamic update of the chronic disease knowledge graph topology.

[0069] Example 3: The current standardized medical information service node receives a sparse chronic disease clinical event input stream from patients with type 2 diabetes. The data monitoring time interval between adjacent patient phenotypic event nodes exceeds 90 days. The computing unit is run by a parallel processor chip containing a dual-path independent register array, with a floating-point operation frequency of no less than 2GHz. It reads the patient's historical full-process disease data collected by the data processing unit and de-identified and anonymized through the background data bus. External data acquisition nodes miss reports, forming a continuous monitoring gap of up to 120 days. The computing unit suspends the dynamic decay calculation based on continuous time series, starts the baseline drift adaptive calibration operator, and adjusts the global baseline drift rate parameter. slope of change of single-point index Perform step-by-step calculations.

[0070] The computing unit first retrieves the sliding time observation window of the previous 30 days for the current sparse sampling point. The fasting blood glucose levels from the last five historical sampling points were read. The five historical values ​​were 5.8 mmol / L, 6.0 mmol / L, 6.1 mmol / L, 6.3 mmol / L, and 6.4 mmol / L, with adjacent differences of 0.2 mmol / L, 0.1 mmol / L, 0.2 mmol / L, and 0.1 mmol / L, respectively. The algebraic average of the four adjacent differences was calculated, yielding 0.15 mmol / L per main sampling period, which was used as the current value. And store it in the first temporary register.

[0071] Subsequently, the computing unit reads the current 120-day sparse time step. The measured fasting blood glucose level at the endpoint, along with the values ​​at adjacent preceding sampling points, showed a difference of 1.2 mmol / L between the endpoint measurement of 7.6 mmol / L and a preceding value of 6.4 mmol / L. Dividing this difference by 120 days yields 0.01 mmol / L per day, which is taken as the current level. And store it in the second temporary register.

[0072] The computing unit synchronously reads the data from the first temporary register. In the second temporary register , with the current sparse time step and The product of these is the total drift, which is then divided by... The resulting ratio is used as a compensatory correction term in the decay index and is compared with the historical static semantic weights. Performing a cascaded product yields the preceding medical semantic weights after discrete difference correction. The computing unit is As the preceding medical semantic weight at the current moment, the semantic decay increment is determined by multiplying the time difference by the baseline decay rate, and then... After deducting the semantic decay increment, the medical semantic dynamic decay coefficient of the primary drug relation edge is dynamically updated. .

[0073] After the above processing, the dynamic attenuation coefficient of medical semantics At the current sparse time step The value is updated to 0.42, which is lower than the preset elimination threshold of 0.50. The association unit identifies the coexisting hypertension association subgraph along the temporal causal association chain and determines that the complication association subgraph has reached the critical threshold for combined medication risk. Based on this, the graph update unit sends a relation edge blocking instruction to the graph database, changes the time status attribute of the long-term medical order triple relation edge of type 2 diabetes, cuts off the long-term medical order triple relation edge, and writes the alternative medication relation edge derived from the temporal causal association chain at its original topological position.

[0074] After the above updates were completed, the response latency of the medical information service node to map updates under complex and multi-complication conditions remained within 15ms, and the occasional redundancy error rate of clinical pathway recommendations was reduced by more than 42%.

[0075] Example 4: This example combines Figures 1 to 2 A description of the time-series-aware dynamic update method and system for chronic disease knowledge graphs, such as... Figure 1As shown, step S1 obtains heterogeneous medical and health data from the chronic disease clinical event input stream, extracts incremental data streams containing patient phenotypic event nodes and primary medication relationship edges from the heterogeneous medical and health data, step S2 retrieves the generation time sequence of nodes in the incremental data stream, calls the medical and clinical time sequence difference when there is time sequence inversion, and uses the median smoothing algorithm to calibrate abnormal timestamps into calibrated time sequence data to construct a time sequence causal association chain that converges unidirectionally from pre-intervention to post-outcome, step S3 calculates the medical semantic dynamic decay coefficient of the primary medication relationship edge based on the calibrated time sequence data, and extracts the change rate of the comparison value of the disease course data when the time interval exceeds 90 days, and uses discrete difference to correct the historical static semantic weight to dynamically update the medical semantic dynamic decay coefficient, step S4 when the decay coefficient is lower than the elimination threshold of 0.50 and the complication association subgraph reaches the critical threshold of combined medication risk, sends a relationship edge blocking instruction to the graph database to cut off the existing long-term medical order triplet relationship edge, and writes the alternative medication relationship edge in place.

[0076] like Figure 2 As shown, within the independent embedded hardware carrier, a one-way configuration connection is established between the main control processor running the anti-interference pre-calibration process and the working memory used to temporarily store calibration data and timing differences, through writing calibration parameters. Simultaneously, the main control processor executes the anti-interference pre-calibration process and sends control logic to the physiological and pathological time-effect function model of the computing unit through task scheduling. The working memory temporarily stores calibration data and timing differences and establishes a one-way transmission channel to the physiological and pathological time-effect function model of the computing unit by reading calibration timing data. Furthermore, the patient phenotype events and long-term medical orders edges of the chronic disease knowledge graph database establish a one-way input connection to the physiological and pathological time-effect function model of the computing unit. The physiological and pathological time-effect function model of the computing unit establishes a state trigger association with the graph update unit relationship edge cutting and in-situ substitution through attenuation coefficient comparison, so that the graph update unit relationship edge cutting and in-situ substitution establish a closed-loop rewrite path to the patient phenotype events and long-term medical orders edges of the chronic disease knowledge graph database through in-situ writing of substitution relationship edges.

[0077] Example 5: Before deploying incremental data streams to a new heterogeneous distributed server cluster, the current standardized medical information service node executes a parameter initialization program. The main control processor fixes the program's entry point through the basic register address segment. Its chip configuration has no less than 32 processing cores, with a single-core processing frequency of no less than 2.4GHz. The target server cluster experiences network bandwidth jitter and cross-service node communication delays, resulting in a 2% to 5% concurrent control delay deviation. The main control processor reads the incremental data stream from the de-identified sample stream. If the cross-domain coupling variance factor driven by fluctuations in physiological parameters of complications is directly calculated without calibration, changes in hardware topology may cause memory addressing conflicts on the memory data bus, hindering the calculation of the medical semantic dynamic decay coefficient and the updating of the map topology. Therefore, the parameter initialization program sets the intrinsic baseline evolution rate constant before the calculation begins. Hardware delay compensation is performed so that it can be used as a reference decay rate in subsequent calculations.

[0078] The parameter initialization program executes three processing stages in sequence. In the first stage, the main control processor controls the counter to continuously perform high-frequency polling sampling of the data packet response time 100 times when the target server cluster is in an idle state. The sampled values ​​are summed by the accumulator and the arithmetic mean is calculated to obtain the scalar value of the node's inherent communication delay. After the scalar value is measured to be 1.2ms, it is stored in the basic temporary storage register.

[0079] In the second stage, the main control processor reads the preset standard evolution rate benchmark value corresponding to type 2 diabetes from the memory and uses this value as the intrinsic benchmark evolution rate constant. For the medical data benchmark, the processor multiplies the node's inherent communication latency scalar value by a dynamic conversion constant to obtain a correction amount used to offset the impact of hardware communication latency. This correction amount is then subtracted from a preset standard evolution rate benchmark value of 0.05, with a correction amount of 0.01. The resulting data is then... The value is set to 0.04 and stored in the working memory. This compensation process only corrects the computational latency deviation introduced by the new heterogeneous distributed server cluster and does not change the... The corresponding chronic disease types and their baseline attenuation relationships.

[0080] In the third stage, the main control processor reads data from the working memory. The physiological and pathological time-effect function model of the calculation unit is input with a baseline decay rate of 0.04. The calculation unit receives the calibration time series data, extracts the time difference between adjacent patient phenotypic event nodes, and then uses this time difference to... The semantic decay increment is determined by the product of the two factors. Then, the semantic decay increment is subtracted from the preceding medical semantic weights at the current time to obtain the dynamic medical semantic decay coefficient of the main drug relationship edge. The graph update unit compares the coefficient with the preset elimination threshold of 0.50 and performs subsequent graph update judgment accordingly.

[0081] With the initial parameter settings mentioned above, the main control processor converts the inherent communication latency of nodes caused by the replacement of underlying devices into a measurable compensation amount, and corrects the calculation input without changing the baseline attenuation relationship. In the new heterogeneous distributed server cluster, the update response latency of the chronic disease knowledge graph is kept within 15ms, and the occasional redundancy error rate of clinical pathway recommendations is reduced by more than 42%.

[0082] Example 6: The current master control computing node runs an anti-interference pre-calibration process on an independent embedded hardware carrier to handle timing update oscillations caused by multiple asynchronous data streams in the standardized medical information service node. This hardware carrier is configured with no less than 4GB of error-correcting memory to maintain the physical stability of storage access. Core vital signs evaluation indicators such as blood potassium concentration and renal function creatinine are collected by different sensors. Differences in sampling frequency will cause time offsets for each indicator. The processing unit executes a standardized time delay mapping procedure, reads the time offset distribution of each indicator within a preset 128 sampling periods, statistically analyzes the time offset histogram, and calibrates the time offset with the highest frequency as the total delay of the baseline system. , serving as a unified time-domain benchmark for graph update processing of multi-source heterogeneous inputs.

[0083] Processing unit based on The procedure for constructing an anti-oscillation smoothing operator for updating the boundary of the chronic disease knowledge graph is as follows: When the type 2 diabetes glycated hemoglobin index in the chronic disease clinical event input stream exhibits non-steady-state fluctuations within a preset fluctuation range of 0.3%, the processing unit retrieves the weight correction factor matrix stored in the working register and maps the fluctuation amplitude to a graph update dynamic damping coefficient used to correct historical static semantic weights. The mathematical representation of the anti-oscillation smoothing operator is as follows: ,in, Update the dynamic damping coefficients in the spectrum. Sensitivity factor for pathological index variation. The absolute value of the instantaneous change of the core vital signs assessment indicators within the sliding time observation window. This is the preset threshold for tolerance of index fluctuations.

[0084] When the fluctuation amplitude of the input noise is close to At that time, the processing unit utilizes Nonlinear smoothing correction is applied to historical static semantic weights to suppress frequent oscillations of relation edge weights near update critical points and reduce logical update lock-ups triggered by minor noise. Verified through engineering reproduction with a sample library of 10,000 cases, the anti-interference pre-calibration process and anti-oscillation smoothing operator reduce the frequency of semantic error conflict triggering of complications under high-frequency data disturbances from 8% to below 0.2%, ensuring that the dynamic structural evolution of the chronic disease knowledge graph in a heterogeneous network environment remains in a logically convergent state that is completely synchronized with clinical pathological patterns.

[0085] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A time-aware dynamic update method for a chronic disease knowledge graph, characterized in that, Includes the following steps: Step S1: Obtain heterogeneous medical and health data from the chronic disease clinical event input stream, and extract incremental data streams containing patient phenotype event nodes and main medication relationship edges from the heterogeneous medical and health data; Step S2: Retrieve the generation time sequence of the drug treatment node and the biochemical test node in the incremental data stream. When there is a time sequence inversion, call the engineering and diagnosis time sequence difference in the standard clinical pathway time sequence constraint rules, and use the median smoothing algorithm to calibrate the abnormal timestamps contained in the generation time sequence into calibrated time sequence data that conforms to the monotonically increasing causal property, so as to construct a time sequence causal association chain that converges unidirectionally from pre-intervention to post-outcome. Step S3: Calculate the dynamic decay coefficient of medical semantics of the main drug relationship edge based on the calibration time series data. When the data monitoring time interval between adjacent patient phenotypic event nodes exceeds 90 days, suspend the dynamic decay calculation based on continuous time series, extract the historical full-process data to calculate the global baseline drift rate parameter, and compare it with the numerical change rate of the current sparse data points. Use discrete difference to correct the historical static semantic weights to dynamically update the dynamic decay coefficient of medical semantics. Step S4: When the medical semantic dynamic decay coefficient is lower than the preset elimination threshold of 0.50, and the complication association subgraph indicated by the temporal causal association chain reaches the critical threshold of combined drug use risk, a relation edge blocking instruction is sent to the graph database to cut off the existing long-term medical order triple relation edge, and the alternative drug relation edge derived from the temporal causal association chain is written in place.

2. The time-aware dynamic update method for a chronic disease knowledge graph according to claim 1, characterized in that, Step S2 includes the following sub-steps: Step S21, extract the invoicing timestamp of the drug treatment node and the review timestamp of the biochemical testing node from the incremental data stream; Step S22: When the review timestamp is earlier than the issuance timestamp, retrieve the clinical pathway timing difference from the standard clinical pathway timing constraint rules. Step S23: Calculate the median of the historical timestamp sequence using the median smoothing algorithm, and correct the audit timestamp to the algebraic sum of the issuance timestamp and the median to generate calibrated time series data.

3. The time-aware dynamic update method for a chronic disease knowledge graph according to claim 1, characterized in that, In step S3, the medical semantic dynamic decay coefficient of the primary drug relationship edge is calculated based on the calibration time series data, including the following sub-steps: Step S31, extract the time difference between adjacent chronic disease clinical events on the calibration time series data; Step S32, determine the semantic decay increment based on the product of the time difference and the baseline decay rate; Step S33, subtract the semantic decay increment from the preceding medical semantic weight at the current time to obtain the medical semantic dynamic decay coefficient of the primary drug relationship edge.

4. The time-aware dynamic update method for a chronic disease knowledge graph according to claim 1, characterized in that, Step S4, where the temporal causal association chain indicates that the coexisting complication association subgraph has reached the critical threshold for combined medication risk, includes the following sub-steps: Step S41, identifying complication nodes and chronic disease medication nodes belonging to the category of coexisting chronic diseases from the temporal causal association chain to construct a complication association subgraph; Step S42, counting the dimension of frequent itemsets of drug combination effects in the complication association subgraph; Step S43, when the dimension of frequent itemsets reaches the preset combined medication threshold, determining that the complication association subgraph has reached the critical threshold for combined medication risk.

5. The time-aware dynamic update method for a chronic disease knowledge graph according to claim 1, characterized in that, Step S4 involves sending a relation edge blocking command to the graph database to sever existing long-term medical order triplet relation edges and writing alternative medication relation edges derived from the temporal causal association chain in situ. This includes the following sub-steps: Step S44: Sending a request to the graph database to disable relation edges between the head and tail entities in the long-term medical order triplet to change the expiration status attribute of the long-term medical order triplet relation edges; Step S45: Extracting new causal alternative node pairs from the temporal causal association chain; Step S46: Writing alternative medication relation edges containing new causal alternative node pairs in situ after the long-term medical order triplet is disconnected to complete the reconstruction of the graph topology.

6. The time-aware dynamic update method for a chronic disease knowledge graph according to claim 1, characterized in that, When there is a semantic conflict update request for multiple complications in step S4, step S4 includes the following sub-steps: Step S47, call the standard chronic disease clinical risk weight matrix and extract the core vital signs evaluation indicators corresponding to the mutually exclusive triples in the semantic conflict update request for multiple complications. Step S48: Calculate the dynamic priority weight of each update request in real time based on the product of the core vital signs evaluation indicators and the standard chronic disease clinical risk weight matrix. Step S49: Compare the dynamic priority weights numerically, assign high-priority chains to the original topology structure for updating, and redirect low-priority chains to temporary observation branch topology nodes for isolation and masking.

7. The time-aware dynamic update method for a chronic disease knowledge graph according to claim 6, characterized in that, The standard chronic disease clinical risk weight matrix called in step S47 contains risk sensitivity coefficients for heart rate, blood pressure, blood glucose, blood lipids and renal function indicators under different combinations of complications. The risk sensitivity coefficients are positively correlated with the degree of acute risk impact of each complication on vital signs.

8. The time-aware dynamic update method for a chronic disease knowledge graph according to claim 1, characterized in that, The alternative medication relationship edge derived from the temporal causal association chain in step S4 includes the following sub-steps: Step S410, mapping the intervention node and outcome node in the temporal causal association chain to the drug entity and efficacy entity in the standard clinical pathway; Step S411, retrieving the temporal constraint rules of the standard clinical pathway and verifying the causal rationality of the one-way dependency relationship between the drug entity and the efficacy entity; Step S412, after verifying the causal rationality of the one-way dependency relationship, encapsulating the drug entity and the efficacy entity into an alternative medication relationship edge.

9. The time-aware dynamic update method for a chronic disease knowledge graph according to claim 1, characterized in that, Step S4, which involves writing alternative medication relationship edges derived from the temporal causal association chain in situ, also includes the following sub-steps: Step S413, performing common sense reasoning verification on the full topological path of the chronic disease knowledge graph after writing to search for the existence of mutually exclusive medication relationship edges or temporal inversion intervention links; Step S414, if mutually exclusive medication relationship edges are found, suspending the current in-situ writing operation and redirecting the alternative medication relationship edges to the manual review queue to maintain the stability of the graph topology.

10. A time-aware dynamic update system for a chronic disease knowledge graph, used to implement the time-aware dynamic update method for a chronic disease knowledge graph as described in claim 1, characterized in that, include: The data processing unit is used to acquire heterogeneous medical and health data from the input stream of chronic disease clinical events and extract incremental data streams containing patient phenotype event nodes and main medication relationship edges from the heterogeneous medical and health data. The association unit is used to retrieve the generation time sequence of the drug treatment node and the biochemical test node in the incremental data stream. When there is a time sequence inversion, it calls the engineering and diagnosis time sequence difference in the standard clinical pathway time sequence constraint rules and uses the median smoothing algorithm to calibrate the abnormal timestamps contained in the generation time sequence into calibrated time sequence data that conforms to the causal monotonous growth property, so as to construct a time sequence causal association chain that converges unidirectionally from pre-intervention to post-outcome. The calculation unit is used to calculate the dynamic decay coefficient of the medical semantics of the main drug relationship edge based on the calibration time series data. When the data monitoring time interval between adjacent patient phenotypic event nodes exceeds 90 days, the dynamic decay calculation based on continuous time series is suspended, the global baseline drift rate parameter is calculated by extracting the historical full-volume disease course data, and it is compared with the numerical change rate of the current sparse data points. The historical static semantic weight is corrected by discrete difference to dynamically update the dynamic decay coefficient of the medical semantics. The graph update unit is used to send a relation edge blocking instruction to the graph database when the medical semantic dynamic decay coefficient is lower than the preset elimination threshold of 0.50 and the temporal causal association chain indicates that the coexisting complication association subgraph reaches the critical threshold of combined drug use risk. This instruction cuts off the existing long-term medical order triple relation edge and writes the alternative drug relation edge derived from the temporal causal association chain in situ.

Images