Adaptive optimization methods and systems for medical auxiliary decision-making integrating knowledge graphs

Abstract

This invention discloses a method and system for adaptive optimization of medical auxiliary decision-making based on knowledge graphs, relating to the field of smart healthcare. The method includes: acquiring medical-related data, performing time alignment and feature decomposition to construct a decision context state vector; performing feature perturbation analysis on observed feature components to measure their impact on the decision path, and constructing decision sensitivity components; constructing a knowledge graph containing medical behavior nodes and observation nodes, and generating an initial set of decision paths using node dependencies; reconstructing the path set using the state vector and sensitivity components to establish a target decision space; calculating path information gain and path redundancy indices for candidate paths in the space, and performing priority ranking to output auxiliary decisions. This application addresses the problem of existing medical auxiliary decision-making systems' inability to adaptively adjust decision schemes according to dynamically changing clinical states, achieving the effect of dynamically adapting decision schemes based on real-time changes in clinical states.

Description

Technical Field

[0001] This application relates to the field of smart healthcare, and in particular to an adaptive optimization method and system for medical auxiliary decision-making that integrates knowledge graphs. Background Technology

[0002] The accuracy of medical decision support directly impacts patient outcomes and is a crucial aspect of clinical practice. Current technologies commonly employ rule-matching systems based on clinical guidelines or statistical analysis tools based on historical medical records to assist physicians in developing treatment plans. This includes comparing patient test results with preset threshold conditions to trigger standard procedures and recommending reference treatment pathways by retrieving past treatment records of similar cases. However, these methods are limited by the generality and static nature of clinical guidelines, making it difficult to cover complex and ever-changing individualized clinical situations. Furthermore, statistical matching methods based on historical data often suffer from data sparsity when dealing with rare disease combinations lacking sufficient historical records, leading to deviations between recommended results and the patient's actual physiological state. They also fail to respond promptly to real-time changes in patient monitoring indicators.

[0003] Currently, medical decision support technologies face the challenge of adaptively adjusting decision-making plans based on the patient's dynamically changing clinical condition. Summary of the Invention

[0004] This application provides a medical auxiliary decision-making adaptive optimization method and system that integrates knowledge graphs. It acquires multi-source data and constructs a decision context state vector, constructs decision sensitivity components through feature perturbation analysis, builds a knowledge graph and generates an initial decision path set, reconstructs the initial decision path set using the decision context state vector and decision sensitivity components to establish a target decision space, calculates information gain and redundancy indices for candidate paths in the target decision space, prioritizes them, and outputs auxiliary decisions. These technical means solve the technical problem of existing medical auxiliary decision-making systems' inability to adaptively adjust decision schemes according to the dynamic changes in the patient's clinical state, achieving the technical effect of dynamically adapting decision schemes based on the patient's real-time changing clinical state.

[0005] This application provides an adaptive optimization method for medical auxiliary decision-making that integrates knowledge graphs, comprising: acquiring multi-source medical association data; performing time alignment and feature decomposition processing on the multi-source medical association data to construct a decision context state vector; performing feature perturbation analysis on the observed feature components in the decision context state vector, measuring the impact of features on decision path selection, and constructing a decision sensitivity component; constructing a knowledge graph containing medical behavior nodes and observation nodes, and generating an initial decision path set using the node dependencies in the knowledge graph; reconstructing the initial decision path set using the decision context state vector and the decision sensitivity component to establish a target decision space adaptively matching the current state; calculating path information gain values ​​and path redundancy indices for candidate paths in the target decision space, performing path priority ranking based on the calculation results, and outputting auxiliary decisions.

[0006] In a possible implementation, the initial decision path set is reconstructed using the decision context state vector and decision sensitivity components, and the following processes are performed: The difference between the maximum and minimum values ​​of the patient's current physiological indicators within a preset sliding time window is extracted from the decision context state vector to establish short-term fluctuations; simultaneously, the influence intensity identifiers of each physiological indicator on the decision path selection are obtained from the decision sensitivity components; for each path in the initial decision path set, the set of observation indicators associated with all medical behavior nodes is parsed; the short-term fluctuations are compared one by one using the set of observation indicators; if any comparison result exceeds a preset stability boundary and the corresponding influence intensity identifier is high influence intensity, the corresponding path is marked as an unstable path with perturbation of vital signs, and path elimination processing is performed; a pruned path set is established based on the path elimination processing results, and the initial decision path set is reconstructed using the pruned path set.

[0007] In a possible implementation, the initial decision path set is reconstructed using the decision context state vector and decision sensitivity components, and the following processing is also performed: parsing the patient's currently identified primary diagnosis data, the recorded list of comorbidities, and the sequence of executed medical actions from the decision context state vector; traversing the pruned path set, marking the node paths with the same primary diagnosis data and all containing at least one of the same comorbidities as path clusters of the same disease and comorbidity; extracting the shared pre-examination and examination behavior chain in the knowledge graph where each path cluster is located; when the length of the shared pre-examination and examination behavior chain exceeds the preset minimum common chain length, the path cluster of the same disease and comorbidity is placed in the shared pre-examination and examination behavior chain. Merge the terminal nodes of the behavior chain and retain the differentiated branches of each path after the terminal node to form a merged path that integrates multiple comorbidity treatment branches. Take the last behavior node of the executed medical behavior sequence as the current breakpoint, perform knowledge graph localization using the current breakpoint, and construct a set of direct successor behavior nodes. Use the set of direct successor behavior nodes to extract path segments starting from any direct successor behavior node from the pruned path set, and splice the path segments with the executed medical behavior sequence to form a continuous connection path. Establish a reconstructed path set based on the merged path and the continuous connection path, and use the pruned path set and the reconstructed path set to perform initial decision path set reconstruction.

[0008] In a possible implementation, the initial decision path set is reconstructed using the decision context state vector and decision sensitivity components, and the following processing is also performed: Based on the decision sensitivity components, sensitivity increase analysis is performed for the period approaching the clinical evaluation cycle, and at least one key observation feature whose sensitivity increase exceeds a preset surge threshold is marked as a decision priority driving feature; for each path in the reconstructed path set, the observation-decision trigger relationship between medical behavior nodes and decision priority driving features in the knowledge graph is queried; based on the query results, medical behavior nodes that directly depend on the decision priority driving features are marked as emergency driving nodes; emergency driving node relocation processing is performed on the paths marked with emergency driving nodes, and parallel candidate behavior node sorting and reconstruction is performed on the relocated paths to establish a topology rearranged path; the initial decision path set is reconstructed using the pruned path set, the reconstructed path set, and the topology rearranged path.

[0009] In a possible implementation, an emergency-driven node relocation process is performed on the path with marked emergency-driven nodes. The process involves: querying all evaluation nodes that have an observation-decision triggering relationship with the decision priority driving feature for the path with marked emergency-driven nodes; determining the earliest evaluation node in the original path sequence as the driving predecessor node, moving the emergency-driven node out of its original position and inserting it after the driving predecessor node; performing an anomaly analysis of the insertion behavior, and if the necessary forward dependency relationship between the emergency-driven node and other behavior nodes in the original path sequence is disrupted, then moving the emergency-driven node sequentially along the path until the necessary forward dependency relationship is fully satisfied, thus completing the emergency-driven node relocation.

[0010] In a possible implementation, the relocated path is reordered and reconstructed using parallel candidate behavior nodes, and the following processes are performed: the relocated path is subjected to node substitution relationship detection, and parallel candidate behavior nodes with multiple substitution relationships are taken as decision branch nodes; real-time data of each patient state dimension in the decision context state vector is obtained; the pre-admission condition set recorded in the knowledge graph is extracted for each parallel candidate behavior node; the real-time data and the pre-admission condition set are subjected to item-by-item deviation analysis to establish an admission gap sequence; after sorting the admission gap sequence from smallest to largest, the parallel candidate behavior nodes are reordered and reconstructed.

[0011] In a possible implementation, the initial decision path set reconstruction is based on path information evolution consistency constraint control, and performs the following processes: constructing an information state representation sequence corresponding to the decision context state vector, mapping any candidate decision path to an information state transition sequence that evolves sequentially along the path nodes; in the information state transition sequence, using the information state representation sequence to calculate the information increment vector and uncertainty change amount for the information state changes between adjacent nodes, forming a path information evolution trajectory; during the execution of path pruning, path reorganization, and topology rearrangement processes, performing consistency verification on the path information evolution trajectory before and after reconstruction; when the reconstruction operation causes the path information evolution trajectory to fail to meet the consistency verification conditions, performing a rollback process on the current path reconstruction result.

[0012] In a possible implementation, the following processing is performed: The consistency verification includes: determining whether there is a node sequence in the reconstructed path whose information increment vector direction deviates from the corresponding interval of the pre-reconstructed path by more than a preset threshold; determining whether there are abnormal transition nodes in the reconstructed path whose information uncertainty change exceeds a preset discontinuity threshold, and completing the consistency verification based on the determination results.

[0013] In a possible implementation, the following process is performed: decision records are made for auxiliary decisions, a decision record set is established, and the decision record set is used for self-optimizing decision management under manual identification.

[0014] This application also provides a medical auxiliary decision-making adaptive optimization system integrating knowledge graphs, including: a decision context state vector construction module, used to acquire multi-source medical related data, perform time alignment and feature decomposition processing on the multi-source medical related data, and construct a decision context state vector; a decision sensitivity component construction module, used to perform feature perturbation analysis on the observed feature components in the decision context state vector, perform influence measurement on decision path selection, and construct decision sensitivity components; an initial decision path set generation module, used to construct a knowledge graph containing medical behavior nodes and observation nodes, and generate an initial decision path set using the node dependencies in the knowledge graph; a target decision space establishment module, used to reconstruct the initial decision path set using the decision context state vector and decision sensitivity components, and establish a target decision space adaptively matching the current state; and an auxiliary decision module, used to calculate path information gain values ​​and path redundancy indices for candidate paths in the target decision space, perform path priority ranking based on the calculation results, and output auxiliary decisions.

[0015] The proposed adaptive optimization method and system for medical auxiliary decision-making, which integrates knowledge graphs, first acquires multi-source medical data, performs time alignment and feature decomposition on this data, and constructs a decision context state vector. Next, it performs feature perturbation analysis on the observed feature components of the decision context state vector, measures the impact of features on decision path selection, and constructs a decision sensitivity component. Then, it constructs a knowledge graph containing medical behavior nodes and observation nodes, generates an initial decision path set using node dependencies in the knowledge graph, and reconstructs the initial decision path set using the decision context state vector and the decision sensitivity component to establish a target decision space adaptively matching the current state. Finally, it calculates path information gain and path redundancy indices for candidate paths in the target decision space, prioritizes paths based on the calculation results, and outputs the auxiliary decision. Through the above process, the proposed method and system achieve the technical effect of dynamically adapting decision schemes according to the patient's real-time changing clinical state. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1 This is a flowchart illustrating the adaptive optimization method for medical auxiliary decision-making based on knowledge graph fusion, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of the process for reconstructing the initial decision path set for execution, as provided in the embodiments of this application.

[0019] Figure 3 This is a schematic diagram of the structure of the medical auxiliary decision-making adaptive optimization system that integrates knowledge graphs, as provided in an embodiment of this application.

[0020] Figure labeling: Decision context state vector construction module 10, decision sensitivity component construction module 20, initial decision path set generation module 30, target decision space establishment module 40, and auxiliary decision module 50. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] This application provides an adaptive optimization method for medical auxiliary decision-making that integrates knowledge graphs, such as... Figure 1 As shown, the method includes: Step S100: Obtain multi-source medical association data, perform time alignment and feature decomposition processing on the multi-source medical association data, and construct a decision context state vector.

[0023] Specifically, electronic medical record data is acquired through the hospital information system interface, real-time vital sign monitoring data is acquired through the monitoring equipment data acquisition interface, laboratory test result data is acquired through the laboratory information system interface, and executed medical behavior data is acquired through the medical order system interface. All data is collected according to the timestamp of the patient's unique identifier as the primary key. Time alignment processing is performed on the acquired multi-source medical data. Using the current decision-making time as the baseline, a preset time window is traced back, and all data source records falling within this time window are sorted by timestamp. In cases where multiple records exist for the same timestamp, the record with the highest priority is selected based on the data source priority, from highest to lowest: real-time monitoring equipment data, laboratory test result data, electronic medical record data, and medical order execution record data. After time alignment, feature decomposition processing is performed on the characteristic components of each data source. Values ​​representing the same physiological indicator from different sources in the multi-source data are weighted and fused. The fusion weight is determined by normalizing the inverse of the data source priority. Feature decomposition processing also includes mapping diagnostic text data to discrete label vectors using the International Classification of Diseases (ICD) coding, and mapping medical behavior data to behavior coding vectors using current diagnostic and treatment operation coding. All the processed numerical feature components, discrete label vector components, and behavior coding vector components are then concatenated in a fixed-dimensional order to construct a decision context state vector, representing the patient's overall clinical status at the current decision moment. This vector includes physiological indicator components, diagnostic label components, executed medical behavior coding components, and timestamp components.

[0024] Step S200: Perform feature perturbation analysis on the observed feature components in the decision context state vector, measure the impact of features on decision path selection, and construct decision sensitivity components.

[0025] Specifically, for the j-th observed feature component, keeping the values ​​of all other feature components unchanged, the value of the j-th feature component is shifted multiple times within a preset perturbation range and with a preset step size. After each shift, a pre-built baseline decision path selection model is invoked, and the decision path selection result corresponding to that shift is recorded. The baseline decision path selection model employs a multi-class logistic regression model, with the decision context state vector as input and the selection probability distribution of each candidate decision path as output. After completing all shifts, the total number of times the j-th feature component causes a switch in the optimal path output by the baseline decision path selection model at each shift is counted. This total number is divided by the preset perturbation step number to obtain the original influence value of the feature component. After performing the above operation on each observed feature component, the original influence vector is obtained. Normalization is performed on the original influence vector by dividing the original influence value of each feature component by the sum of the original influence values ​​of all feature components to obtain the normalized influence value of each feature component. All normalized influence values ​​constitute the decision sensitivity component, characterizing the degree to which a unit change in each observed feature component affects the decision path selection result. The values ​​of each element in the decision sensitivity component range from 0 to 1. The higher the value, the more likely the decision path will switch when the feature changes, and the sum of all elements equals 1.

[0026] Step S300: Construct a knowledge graph containing medical behavior nodes and observation nodes, and use the node dependencies in the knowledge graph to generate an initial set of decision paths.

[0027] Specifically, the knowledge graph's nodes are categorized into two types: medical behavior nodes and observation nodes. Medical behavior nodes represent specific diagnostic and treatment procedures, medication administration, or surgical procedures, while observation nodes represent clinical laboratory tests, vital sign monitoring indicators, or symptom assessment scale entries. Directed edges in the knowledge graph represent dependencies between nodes, including: directed edges from observation nodes to medical behavior nodes, indicating that the medical behavior node is triggered when the value of the observation node meets a preset trigger condition; directed edges from medical behavior nodes to observation nodes, indicating the indicators to be observed after performing the medical behavior; and directed edges from one medical behavior node to another, indicating a sequential constraint between the two behaviors. The knowledge graph is constructed based on clinical treatment guidelines, drug instructions, expert consensus, and statistical patterns of historical treatment pathways, and is stored in triplet form. An initial set of decision paths is generated using the node dependencies in the knowledge graph. Specifically, a breadth-first traversal strategy is employed, starting from all medical behavior nodes marked as the starting node in the knowledge graph and traversing along directed edges. All reachable paths from the starting node to the ending node are recorded. Each path consists of a series of medical behavior nodes arranged sequentially, where any two adjacent medical behavior nodes in the path are directly connected by a directed edge in the knowledge graph or indirectly connected through an observation node. When an observation node exists in the path, it is recorded as a conditional constraint between its predecessor and successor medical behavior nodes in the path attributes.

[0028] Step S400: Reconstruct the initial decision path set using the decision context state vector and decision sensitivity components to establish a target decision space that adaptively matches the current state.

[0029] Specifically, such as Figure 2 As shown, using the decision context state vector constructed in step S100 and the decision sensitivity component constructed in step S200, a reconstruction operation is performed on the initial decision path set generated in step S300. The reconstruction operation includes three parallel sub-processes: path pruning, path recombination, and topology rearrangement. The execution order of the three sub-processes is adaptively determined based on the stability of the patient's current state in the decision context state vector. When the fluctuation of the patient's vital signs exceeds a preset stability threshold, path pruning is performed first; when the number of comorbidities exceeds a preset number threshold, path recombination is performed first; and when there is an observation feature with a sudden increase in sensitivity, topology rearrangement is performed first. After all three sub-processes are completed, the pruned path set, the recombined path set, and the topology rearranged path are merged and deduplicated to obtain a target decision space that adaptively matches the current state, i.e., a set of candidate decision paths that adaptively matches the current patient state.

[0030] In one possible implementation, the initial decision path set is reconstructed using the decision context state vector and decision sensitivity components. Step S400 further includes step S410, which extracts the difference between the maximum and minimum values ​​of the patient's current physiological indicators within a preset sliding time window from the decision context state vector to establish short-term fluctuations. Simultaneously, the influence intensity identifier of each physiological indicator on the decision path selection is obtained from the decision sensitivity components. Specifically, all physiological indicator components are extracted from the decision context state vector. For each physiological indicator component, its monitoring value sequence within a preset sliding time window is determined. The length of the preset sliding time window is, for example, set to the past 6 hours, and the window is traced back to the left from the right endpoint of the current decision time. The maximum and minimum values ​​are found from the monitoring value sequence, and the difference between the maximum and minimum values ​​is calculated. The absolute value of this difference is taken as the short-term fluctuation of the physiological indicator. The above operation is performed on all physiological indicator components to obtain the short-term fluctuation corresponding to each physiological indicator. Simultaneously, sensitivity values ​​corresponding to each physiological indicator component are obtained from the decision sensitivity component. Each sensitivity value is compared with a preset influence intensity threshold. When the sensitivity value is greater than or equal to the preset influence intensity threshold, the influence intensity of the physiological indicator is marked as high influence intensity; otherwise, it is marked as low influence intensity. The preset influence intensity threshold is, for example, set to 1.5 times the average value of all elements in the decision sensitivity component.

[0031] Step S420: For each path in the initial decision path set, parse the set of observation indicators associated with all medical behavior nodes. Specifically, traverse each path in the initial decision path set. For each medical behavior node in the current path, query all observation nodes in the knowledge graph directly connected to that medical behavior node through directed edges. Summarize the physiological indicator names or test item names corresponding to all the queried observation nodes to form the set of observation indicators associated with that medical behavior node. Perform a union operation on the set of observation indicators associated with all medical behavior nodes in the current path to obtain the total set of observation indicators for that path. Perform the above parsing operation on each path in the initial decision path set.

[0032] Step S430: Short-term fluctuations are compared one by one using the set of observation indicators. If any comparison result exceeds a preset stability boundary and the corresponding influence intensity is identified as high influence intensity, the corresponding path is marked as an unstable path due to bodily disturbances, and path removal processing is performed. Specifically, for each path in the initial decision path set, the total set of observation indicators for that path is obtained, and the short-term fluctuations of each observation indicator in the total set of observation indicators are compared with the corresponding physiological indicator. A preset stability boundary database is obtained, which pre-stores the upper limit of the tolerance fluctuation corresponding to each physiological indicator. When the short-term fluctuation corresponding to a certain observation indicator is greater than the preset stability boundary value of the corresponding physiological indicator, it is determined that the comparison result exceeds the preset stability boundary. At this time, the influence intensity identifier corresponding to the physiological indicator is further queried. If the influence intensity identifier is high influence intensity, the path is marked as an unstable path due to bodily disturbances. After completing the above comparison and determination for all paths in the initial decision path set, all paths marked as unstable paths due to bodily disturbances are removed from the initial decision path set, and the unmarked paths are retained to form a pruned path set.

[0033] Step S440: Establish a pruned path set based on the path elimination processing results, and use the pruned path set to reconstruct the initial decision path set. Specifically, gather all paths remaining after the path elimination processing in step S430 to establish a pruned path set. Use this pruned path set to replace the original initial decision path set as part of the reconstructed candidate path set.

[0034] For example, the following scenario is given for the path pruning process described in steps S410 to S440. Assume that a patient's current decision context state vector contains two physiological indicators: heart rate and systolic blood pressure. The preset sliding time window is the past 4 hours. The heart rate monitoring sequence is [92, 95, 88, 96, 93, 94, 91, 97] beats / min, with a maximum of 97 beats / min and a minimum of 88 beats / min, and a short-term heart rate fluctuation of 9 beats / min. The systolic blood pressure monitoring sequence is [135, 132, 138, 130, 136, 134, 139, 133] mmHg, with a maximum of 139 mmHg and a minimum of 130 mmHg, and a short-term systolic blood pressure fluctuation of 9 mmHg. The sensitivity values ​​for heart rate and systolic blood pressure are 0.28, respectively, based on the central tendency of the decision-making sensitivity component. The preset influence threshold is 1.5 times the average of all sensitivity values. Assuming the threshold is 0.25 in this scenario, both heart rate and systolic blood pressure are marked as having high influence. The preset stability boundary database has a tolerance limit of 8 beats / min for heart rate and 10 mmHg for systolic blood pressure. The current initial decision path set includes path A with the total set of associated observation indicators {heart rate, body temperature}, path B with the total set of associated observation indicators {systolic blood pressure, diastolic blood pressure}, and path C with the total set of associated observation indicators {heart rate, systolic blood pressure}. The comparison results are shown in Table 1.

[0035] Table 1

[0036] As shown in Table 1, path A is marked as an unstable path due to short-term heart rate fluctuations exceeding the stability boundary and its heart rate influence intensity being identified as high. Path B is not marked because its short-term systolic blood pressure fluctuations do not exceed the stability boundary. Path C is also marked because its systolic blood pressure does not exceed the boundary, but its heart rate exceeds the boundary and is of high influence intensity. After performing path removal processing, the pruned path set only includes path B.

[0037] In one possible implementation, the initial decision path set is reconstructed using the decision context state vector and decision sensitivity components. Step S400 further includes step S450, parsing the patient's currently identified primary diagnosis data, the recorded list of comorbidities, and the sequence of executed medical actions from the decision context state vector. Specifically, a diagnostic label vector component is extracted from the decision context state vector to parse the patient's currently identified primary diagnosis data, which is represented using the International Classification of Diseases (ICD). A comorbidity label vector component is extracted from the decision context state vector, and all comorbidity labels are arranged in chronological order of appearance to establish a list of recorded comorbidities. A behavior encoding vector component is extracted from the decision context state vector and arranged in ascending order of execution timestamps to establish a sequence of executed medical actions.

[0038] Step S460: Traverse the trimmed path set and mark the node paths with the same master diagnosis data and all containing at least one intervention behavior for the same comorbidity as a path cluster of the same disease and comorbidity. Specifically, traverse each path in the trimmed path set. For the current path, first extract the codes of all medical behavior nodes in the path and match these codes with a pre-established comorbidity-intervention behavior mapping table to determine which comorbidities were treated with intervention behaviors in the path. Compare the current path with other paths in the trimmed path set. If two paths have the same master diagnosis data code and the comorbidity intervention behaviors contained in each path correspond to at least one of the same comorbidity types, then these two paths are assigned to the same path cluster of the same disease and comorbidity. Repeat the above comparison process until all paths in the trimmed path set have been clustered.

[0039] Step S470: Extract the shared pre-condition inspection / check behavior chain from the knowledge graph of each path cluster containing the same disease and symptoms. Specifically, for each path cluster, obtain the sequence of medical behavior nodes contained in each path within the cluster. Starting from the first medical behavior node of each path, compare whether the medical behavior nodes at the same position are the same across paths and whether the dependencies of these nodes in the knowledge graph are consistent. Record the longest consecutive sequence of medical behavior nodes that are common to all paths and are in the same order. This sequence may contain inspection / check observation nodes as conditional constraints. Determine the longest consecutive sequence of medical behavior nodes as the shared pre-condition inspection / check behavior chain for the path cluster containing the same disease and symptoms.

[0040] Step S480: When the length of the shared pre-examination and examination behavior chain exceeds a preset minimum common chain length, the path clusters of the same disease and symptoms are merged at the end node of the shared pre-examination and examination behavior chain. After merging, the differentiated branches of each path after the end node are retained to form a merged path that integrates multiple comorbidity treatment branches. Specifically, a preset minimum common chain length threshold is obtained, which is, for example, set to include at least 2 medical behavior nodes. For each path cluster of the same disease and symptoms, it is determined whether the number of medical behavior nodes contained in its shared pre-examination and examination behavior chain is greater than the preset minimum common chain length. If it is greater, all paths in the path cluster of the same disease and symptoms are merged at the end node of the shared pre-examination and examination behavior chain. The merging method is as follows: the shared pre-examination and examination behavior chain is retained as the pre-common part of the merged path, and the differentiated node sequence of each path after the end node is attached as a branch structure after the end node. Each branch is arranged according to the node order of its original path to form a merged path that integrates multiple comorbidity treatment branches.

[0041] Step S490: The last action node of the executed medical action sequence is used as the current breakpoint. Knowledge graph localization is performed using this breakpoint to construct a set of direct successor action nodes. Specifically, the executed medical action sequence parsed in step S450 is obtained, and the last action node in the sequence is extracted and used as the current breakpoint. Using the node identifier of the current breakpoint as the query key, node localization is performed in the knowledge graph to find the corresponding position of the node. Starting from this position, all direct successor nodes are queried along the directed edges in the knowledge graph. A direct successor node is a medical action node directly connected to the current breakpoint by a directed edge and without any intermediate nodes. All queried direct successor medical action nodes are collected to construct a set of direct successor action nodes.

[0042] Step S4100: Using the set of direct successor behavior nodes, extract path segments starting from any direct successor behavior node from the trimmed path set. Concatenate these path segments with the executed medical behavior sequence to form a continuous path. Specifically, traverse each node in the set of direct successor behavior nodes. For the current node, search the trimmed path set for all paths starting from that node. For each searched path, extract a complete path segment from its starting node to its ending node. Concatenate this path segment to the executed medical behavior sequence, with the executed medical behavior sequence as the preceding part of the concatenated path and the extracted path segment as the following part. Connect the two parts end-to-end to form a continuous path. Perform the above search, extraction, and concatenation operations on each node in the set of direct successor behavior nodes to obtain all continuous paths.

[0043] Step S4110: Establish a reconstructed path set based on the merged path and the continuous connection path, and use the trimmed path set and the reconstructed path set to perform the initial decision path set reconstruction. Specifically, all merged paths obtained in step S480 and all continuous connection paths obtained in step S4100 are combined to establish a reconstructed path set. The paths in the reconstructed path set are merged with the paths in the trimmed path set. If there are duplicate paths in the reconstructed path set, the paths in the reconstructed path set are retained as unique versions. The merged path set replaces the original initial decision path set, and the initial decision path set is reconstructed.

[0044] In one possible implementation, the initial decision path set is reconstructed using the decision context state vector and decision sensitivity components. Step S400 further includes step S4120, which performs sensitivity increase analysis on the decision sensitivity components in the vicinity of the clinical evaluation cycle, and marks at least one key observation feature whose sensitivity increase exceeds a preset surge threshold as a decision priority driving feature. Specifically, the historical records of decision sensitivity components for a preset number of historical clinical evaluation cycles prior to the current decision time are obtained, with each clinical evaluation cycle set to, for example, 24 hours. For each observation feature component in the decision context state vector, the decision sensitivity component value at the current decision time and the decision sensitivity component value of the most recent historical clinical evaluation cycle are obtained, and the difference between the two is calculated as the sensitivity increase of the observation feature. The sensitivity increase of each observation feature is compared with a preset surge threshold, which is set to, for example, 0.15. When the sensitivity increase of an observation feature is greater than the preset surge threshold, the observation feature is marked as a decision priority driving feature.

[0045] Step S4130 involves querying the knowledge graph for each path in the reconstructed path set to determine the observation-decision triggering relationship between medical behavior nodes and decision priority driving features. Specifically, each path in the reconstructed path set is traversed. For each medical behavior node in the current path, using that medical behavior node as a successor node, all directed edges pointing to that medical behavior node are queried in the knowledge graph. For each queried directed edge, the observation feature name corresponding to the starting observation node of that directed edge is obtained. It is determined whether the observation feature name matches the decision priority driving feature marked in step S4120. If they match, an observation-decision triggering relationship is recorded between the medical behavior node and the decision priority driving feature. The above query operation is performed on all paths in the reconstructed path set.

[0046] Step S4140: Based on the query results, medical behavior nodes that directly depend on the decision priority driving feature are marked as emergency driving nodes. Specifically, for each path in the recombined path set, based on the query results of step S4130, all medical behavior nodes recorded as having an observation-decision triggering relationship with the decision priority driving feature are marked as emergency driving nodes. A path may contain multiple emergency driving nodes, or it may not contain any emergency driving nodes.

[0047] Step S4150: Perform emergency-driven node relocation processing on the paths marked with emergency-driven nodes, and perform parallel candidate behavior node sorting and reconstruction on the relocated paths to establish topologically rearranged paths; use the pruned path set, reassembled path set, and topologically rearranged paths to perform initial decision path set reconstruction. Specifically, for each path in the reassembled path set marked with an emergency-driven node, perform emergency-driven node position relocation processing in the path sequence. The goal of the relocation processing is to adjust the emergency-driven node to an earlier position in the path sequence so that it can be executed promptly after the triggering conditions are met. After completing the relocation processing, perform sorting and reconstruction on multiple parallel candidate behavior nodes in the same decision branch position in the path, and rearrange the execution order between parallel nodes according to the degree of matching between the pre-admission conditions of each node and the current patient status data. Establish the paths after relocation processing and sorting and reconstruction as topologically rearranged paths. Perform a union operation on all paths in the pruned path set, all paths in the reassembled path set, and all paths in the topologically rearranged paths. During merging, the uniqueness of the path's node sequence is used as the criterion; that is, when two paths have identical node sequences, only one is retained. The merged path set is the reconstructed initial decision path set, which is then sent to steps S4160 to S4190 for information evolution trajectory consistency verification. After passing the verification, it becomes the target decision space in step S500.

[0048] In one possible implementation, an emergency-driven node relocation process is performed on the path with marked emergency-driven nodes. Step S4150 further includes step S4151, querying all evaluation nodes on the path with marked emergency-driven nodes that have an observation-decision triggering relationship with the decision priority driving feature. Specifically, for a path in the reconstructed path set marked with an emergency-driven node, all medical behavior nodes marked as emergency-driven nodes in that path are obtained. For each emergency-driven node, all directed edges pointing to that emergency-driven node are queried in the knowledge graph. From these directed edges, all edges whose observation features corresponding to the starting observation node match the decision priority driving feature are selected, and the starting observation nodes of these edges are taken as evaluation nodes related to the emergency-driven node and the decision priority driving feature.

[0049] Step S4152: Determine the earliest evaluation node in the original path sequence as the driving predecessor node, and remove the emergency driving node from its original position and insert it after the driving predecessor node. Specifically, for a given emergency driving node and all its corresponding evaluation nodes, obtain the position indices of these evaluation nodes in the original node sequence of the current path, and select the evaluation node with the smallest position index as the driving predecessor node. Remove the emergency driving node from its original position, and then insert it immediately after the driving predecessor node in the sequence. If there are multiple emergency driving nodes in the path, perform the above removal and insertion operations sequentially according to the original positions of each emergency driving node from back to front, to avoid position changes affecting the determination of the original positions of unprocessed emergency driving nodes.

[0050] Step S4153: Perform an anomaly analysis of the insertion behavior. If the necessary preceding dependency between the emergency driving node and other behavior nodes in the original path sequence is disrupted, the emergency driving node is moved sequentially along the path until the necessary preceding dependency is fully satisfied, thus completing the relocation of the emergency driving node. Specifically, after the insertion operation in step S4152 is completed, check whether the emergency driving node at its new position conflicts with other medical behavior nodes in the path due to necessary preceding dependencies. A necessary preceding dependency is a constraint represented by a directed edge in the knowledge graph pointing from one medical behavior node to another, indicating that the latter can only be executed after the former has been executed. For the emergency driving node, query all directed edges in the knowledge graph that point from the emergency driving node to other behavior nodes, and all directed edges that point from other behavior nodes to the emergency driving node, to determine all necessary preceding dependencies related to the emergency driving node. Each dependency in the new sequence is checked one by one to see if its predecessor node precedes its successor node. If any dependency has a predecessor node that precedes its successor node, a conflict is identified. When a conflict exists, the urgent driving node is moved one position backward along the path sequence. After moving, the above check is performed again, and this process is repeated until all necessary preceding dependencies are satisfied. After the movement is complete, the urgent driving node is relocated.

[0051] In one possible implementation, after performing parallel candidate behavior node sorting and reconstruction on the relocated path, step S4150 further includes step S4154, which involves detecting the substitution relationship between nodes on the relocated path and using parallel candidate behavior nodes with multiple substitution relationships as decision branch nodes. Specifically, for a path that has completed the relocation of the emergency-driven node, each pair of medical behavior nodes in the path is traversed, and the existence of metadata identifiers in the knowledge graph indicating that the two nodes are substitution relationships is queried. A substitution relationship is a relationship between multiple medical behavior nodes that can be selected to achieve similar diagnostic and treatment effects in the same clinical context. Medical behavior nodes with substitution relationships are used as parallel candidate behavior nodes. When there are multiple parallel candidate behavior nodes with substitution relationships after a certain position in the path, these nodes are used as a decision branch node group.

[0052] Step S4155: Obtain real-time data for each patient state dimension in the decision context state vector. Specifically, extract real-time values ​​for all patient state dimensions from the decision context state vector constructed in step S100, including real-time values ​​of physiological indicators, real-time values ​​of laboratory indicators, real-time values ​​of symptom scores, and other dimension components to form a real-time data vector.

[0053] Step S4156: Extract the set of prerequisite conditions recorded in the knowledge graph for each parallel candidate behavior node. Specifically, for each parallel candidate behavior node identified in step S4154, query the knowledge graph for all prerequisite conditions associated with that node. Prerequisite conditions include, but are not limited to: physiological indicator range conditions, test indicator threshold conditions, time interval conditions, and conditions of previously executed prerequisite behaviors that must be met before executing the medical behavior node. Summarize all the retrieved prerequisite conditions to form the set of prerequisite conditions for that parallel candidate behavior node.

[0054] Step S4157: Perform a step-by-step deviation analysis between the real-time data and the pre-condition set to establish an admission gap sequence. Specifically, for each parallel candidate behavior node's pre-condition set, compare each condition in the condition set with the corresponding dimension value in the real-time data vector obtained in step S4155. For numerical range-type conditions, calculate the absolute difference between the real-time value and the optimal value specified by the condition as the deviation of that condition. For threshold-type conditions, calculate the absolute value of the difference between the real-time value and the threshold as the deviation. For category-type conditions, if the real-time category matches the category required by the condition, the deviation is 0; otherwise, the deviation is the preset maximum deviation value. The deviations of each condition are arranged in the order of their positions in the condition set to form the admission gap sequence for that node.

[0055] Step S4158 involves sorting the admission gap sequence from smallest to largest, and then reconstructing the order of parallel candidate behavior nodes. Specifically, for multiple parallel candidate behavior nodes in the same decision branch node group, the deviations in the admission gap sequence of each node are summed or weighted to obtain the overall admission gap score for each node. The weight coefficients are pre-set according to the importance level of each condition in clinical guidelines. The parallel candidate behavior nodes are sorted in ascending order of overall admission gap score, with the node with the smallest overall admission gap score ranked first, indicating that it has the highest degree of matching with the current patient status and should be executed first. The sorted parallel candidate behavior nodes replace the original node arrangement in the path, completing the reconstructed order.

[0056] For example, the following scenario example is given for the emergency-driven node relocation and parallel candidate behavior node sorting and reconstruction process described in steps S4151 to S4158. Assume the original node sequence of a path is [node D, node E, node F, node G, node H], the decision priority driving feature is blood oxygen saturation, node H is marked as an emergency-driven node, and the evaluation nodes corresponding to blood oxygen saturation are node E and node G. Node E has a position index of 2 in the original path sequence, and node G has a position index of 4, therefore the preceding driving node is node E. Moving node H from position 5 and inserting it after node E results in the sequence [node D, node E, node H, node F, node G]. Querying the necessary preceding dependencies reveals a dependency that node G must be executed after its predecessor node F. In the new sequence, node G is located after node F, satisfying this dependency, with no other conflicts, thus completing the relocation. After relocation, in the path, after node E, there exists a substitution relationship between node H and another parallel candidate behavior node K, forming a decision branch node group {node H, node K}. The knowledge graph records the preconditions for node H as {oxygen saturation greater than or equal to 90%, heart rate less than 120 beats / min}, and the preconditions for node K as {oxygen saturation greater than or equal to 92%, systolic blood pressure greater than or equal to 90 mmHg}. Current real-time data shows oxygen saturation of 91%, heart rate of 115 beats / min, and systolic blood pressure of 85 mmHg. The preconditions for node H are: oxygen saturation deviation of 91% - 90% = 1%, and heart rate deviation of 120 - 115 = 5 beats / min; the preconditions for node K are: oxygen saturation deviation of 92% - 91% = 1%, and systolic blood pressure deviation of 90 - 85 = 5 mmHg. The original deviations of different physiological dimensions are uniformly normalized to eliminate the difference in dimensions before the total score is calculated. It is assumed that the weight of both conditions after normalization is 1. The total score of the comprehensive admission gap of node H is 6 and the total score of the comprehensive admission gap of node K is 6. Since the total scores of the two are the same, they are sorted according to the preset parallel processing rules. The sorted parallel candidate behavior node order is [node H, node K], and the sorting reconstruction is completed.

[0057] In one possible implementation, the initial decision path set reconstruction is based on path information evolution consistency constraint control. Step S400 further includes step S4160, constructing an information state representation sequence corresponding to the decision context state vector, mapping any candidate decision path to an information state transition sequence that evolves sequentially along the path nodes. Specifically, a pre-trained encoder is used to encode the decision context state vector. The encoder adopts an autoencoder structure based on a multilayer perceptron, including an input layer, a hidden layer, and an output layer, and is trained by minimizing the reconstruction error. The decision context state vector is input into the encoder, and the activation values ​​of the encoder's hidden layer are extracted as the information state representation of the state vector, resulting in an information state representation vector corresponding to the decision context state vector. For any candidate decision path, the initial information state of the path is set to the aforementioned information state representation vector. For the first medical behavior node in the path, the observation feature encoding and behavior encoding corresponding to the node are obtained. This encoding is concatenated with the current information state representation vector and input into a trained gated recurrent unit network. The output of the gated recurrent unit network is the updated information state representation vector, representing the information state after advancing one node along the path. Repeat the above process to perform state update operations on each medical behavior node in the path in sequence, and obtain the information state representation vector sequence arranged in order along the path nodes, that is, the information state transition sequence.

[0058] Step S4170: In the information state transition sequence, the information increment vector and uncertainty change are calculated based on the information state changes between adjacent nodes using the information state representation sequence, forming a path information evolution trajectory. Specifically, for the information state transition sequence obtained in step S4160, two adjacent information state representation vectors in the sequence are taken sequentially, and the difference between the latter vector and the former vector is calculated to obtain the information increment vector between these two adjacent nodes. This vector represents the distribution of newly added information content in each dimension after executing the node in the path. For each information state representation vector in the information state transition sequence, the information entropy estimate of the vector is calculated, specifically by multiplying the logarithm of the sum of the squares of the values ​​of each dimension of the information state representation vector by a preset coefficient. The difference between the information entropy estimates of two adjacent information state representation vectors is calculated to obtain the uncertainty change; a positive value indicates an increase in uncertainty, and a negative value indicates a decrease in uncertainty. All information increment vectors are arranged in the order of the path nodes to obtain the information increment vector sequence; all uncertainty changes are arranged in the order of the path nodes to obtain the uncertainty change sequence. The path information evolution trajectory of this path is composed of the information state transition sequence, the information increment vector sequence, and the uncertainty change sequence.

[0059] Step S4180: During the path pruning, path recombination, and topology rearrangement processes, a consistency check is performed on the path information evolution trajectories before and after reconstruction. The consistency check includes: determining whether there are node sequences in the reconstructed path whose information increment vector direction deviates in the opposite direction from the corresponding interval of the pre-reconstruction path by more than a preset threshold; and determining whether there are abnormal transition nodes in the reconstructed path whose information uncertainty change suddenly increases by more than a preset discontinuity threshold. The consistency check is completed based on the determination results. Specifically, for any path in the initial decision path set, after path pruning, recombination, or topology rearrangement, a reconstructed path is obtained. The path information evolution trajectories of the pre-reconstruction path and the reconstructed path are calculated according to the methods in steps S4160 and S4170, respectively. The first determination of the consistency check is as follows: For each pair of adjacent nodes in the reconstructed path, the cosine value of the direction angle between the information increment vector corresponding to the pair of adjacent nodes and the information increment vector of the same clinical semantic interval in the pre-reconstruction path is calculated. If the cosine value is less than a preset direction consistency threshold, it is determined that the direction of the information increment vector of the node sequence has deviated in the opposite direction by more than the preset threshold. The preset direction consistency threshold is, for example, set to 0. The second criterion for consistency verification is as follows: For each node in the reconstructed path, calculate the amount of uncertainty change at that node's location. If the absolute value of this uncertainty change exceeds a preset multiple of the absolute value of the uncertainty change at the corresponding location in the pre-reconstruction path, and the sign of the uncertainty change at that node is inconsistent with the sign of the corresponding location before reconstruction, then that node is determined to be an abnormal transition node. The preset multiple is, for example, set to 3 times. Combining the results of both criteria, if either the first or second criterion is triggered, the reconstructed path is determined to not meet the consistency verification conditions.

[0060] In step S4190, when the reconstruction operation causes the path information evolution trajectory to fail to meet the consistency check conditions, a rollback process is performed on the current path reconstruction result. Specifically, when the consistency check in step S4180 determines that a reconstructed path does not meet the consistency check conditions, the reconstructed path is abandoned, the path is restored to its original path form before reconstruction, and the original path is added back to the candidate path set. Simultaneously, for all subsequent operations associated with this reconstruction operation, such as those in the same step where the reconstruction of other paths depends on this reconstruction result, a rollback process is also performed to ensure the integrity and consistency of the candidate path set.

[0061] Step S500: Calculate the path information gain value and path redundancy index for candidate paths in the target decision space, perform path priority sorting based on the calculation results, and output auxiliary decision.

[0062] Specifically, for each candidate path in the target decision space, based on the information state transition sequence and information increment vector sequence calculated in step S4170, the absolute values ​​of each dimension of the information increment vector corresponding to all nodes in the path are summed to obtain the path information gain value. This value represents the total amount of effective decision-making information increment that can be obtained after executing all medical behavior nodes along the path. For each candidate path in the target decision space, the set of observation indicators associated with all medical behavior nodes in the path is extracted, the number of times the same observation indicator appears repeatedly at different nodes is counted, and the sum of the repetition counts of each observation indicator minus 1 is obtained to obtain the path redundancy index. The higher the value, the more redundant behavior nodes there are in the path. Path priority ranking is performed based on the path information gain value and the path redundancy index. Specifically, the comprehensive priority score of each path is calculated according to a preset comprehensive evaluation formula. For example, the comprehensive evaluation formula can be: the comprehensive priority score equals the path information gain value multiplied by the preset information gain weight coefficient minus the path redundancy index multiplied by the preset redundancy penalty coefficient. The comprehensive priority scores are sorted from high to low, and the candidate path with the highest score is output as the optimal auxiliary decision path. The output method includes displaying the node sequence of the path and the corresponding execution instructions in text form on the terminal display device.

[0063] In one possible implementation, step S500 further includes step S510, which involves recording the auxiliary decision, establishing a decision record set, and using the decision record set for self-optimizing decision management under manual labeling. Specifically, after each execution of step S500 to complete the auxiliary decision output, a decision record is made for that auxiliary decision process. The content of the decision record includes: the decision context state vector at the current decision time, all candidate paths in the target decision space and their corresponding path information gain values ​​and path redundancy indices, path priority ranking results, and the optimal auxiliary decision path with the highest comprehensive priority score. The above content is stored in a structured manner according to the decision time timestamp as an index to establish a decision record set. After each auxiliary decision output, the optimal auxiliary decision path is presented to the clinician in a structured form. The clinician manually labels the displayed auxiliary decision on the terminal interactive interface. The manual labeling is divided into two types: adoption label and rejection label. When the clinician labels it as adoption, the decision record is marked as a valid decision record. When the clinician labels it as rejection, the clinician enters or selects the actual adopted decision path on the interactive interface, and stores the actual adopted path as a manually corrected path in the decision record. The self-optimizing decision management specifically performs the following operations: It periodically triggers a self-optimizing update cycle. Within each update cycle, it reads all entries marked as valid decision records from the decision record set, counts the adoption frequency of each candidate path under different decision context state vectors, and calculates the adoption rate of each path as its empirical success probability. It reads all entries in the decision record set containing manually corrected paths, calculates the difference metric between the manually corrected path and the optimal auxiliary decision path automatically output in step S500. The difference metric is calculated by converting both paths into node sequences and then calculating the edit distance. When the average difference metric exceeds a preset difference threshold, it triggers parameter adjustments to the comprehensive evaluation formula used for priority ranking in step S500. The adjustment method involves fine-tuning the information gain weight coefficient and redundancy penalty coefficient according to a preset step size until the average difference metric drops below the preset difference threshold. The updated parameters are then used in subsequent auxiliary decision priority ranking to achieve adaptive optimization of the decision strategy.

[0064] This application's embodiments address the technical problem of existing medical auxiliary decision-making systems' inability to adaptively adjust decision plans based on the patient's dynamically changing clinical state by acquiring multi-source data and constructing a decision context state vector, constructing decision sensitivity components through feature perturbation analysis, building a knowledge graph and generating an initial decision path set, reconstructing the initial decision path set using the decision context state vector and decision sensitivity components to establish a target decision space, calculating information gain and redundancy indices for candidate paths in the target decision space, prioritizing them, and outputting auxiliary decision-making results.

[0065] In the above text, refer to Figures 1-2 This paper describes in detail an adaptive optimization method for medical auxiliary decision-making based on the fusion of knowledge graphs according to embodiments of the present invention. Next, reference will be made to... Figure 3 This invention describes an adaptive optimization system for medical auxiliary decision-making based on a fused knowledge graph, according to an embodiment of the present invention.

[0066] The adaptive optimization system for medical auxiliary decision-making based on knowledge graphs, according to embodiments of the present invention, addresses the technical problem of existing medical auxiliary decision-making systems' inability to adaptively adjust decision-making schemes based on the dynamic changes in the patient's clinical state. It achieves the technical effect of dynamically adapting decision-making schemes to the patient's real-time changing clinical state. The adaptive optimization system for medical auxiliary decision-making based on knowledge graphs includes: a decision context state vector construction module 10, a decision sensitivity component construction module 20, an initial decision path set generation module 30, a target decision space establishment module 40, and an auxiliary decision-making module 50.

[0067] The decision context state vector construction module 10 is used to acquire multi-source medical correlation data, perform time alignment and feature decomposition processing on the multi-source medical correlation data, and construct a decision context state vector. The decision sensitivity component construction module 20 is used to perform feature perturbation analysis on the observed feature components in the decision context state vector, perform feature impact measurement on decision path selection, and construct decision sensitivity components. The initial decision path set generation module 30 is used to construct a knowledge graph containing medical behavior nodes and observation nodes, and generate an initial decision path set using the node dependency relationships in the knowledge graph. The target decision space establishment module 40 is used to reconstruct the initial decision path set using the decision context state vector and decision sensitivity components, and establish a target decision space that adaptively matches the current state. The auxiliary decision module 50 is used to calculate the path information gain value and path redundancy index of candidate paths in the target decision space, perform path priority sorting according to the calculation results, and output auxiliary decisions.

[0068] The specific configuration of the target decision space establishment module 40 is described in detail below: As mentioned above, the target decision space establishment module 40 further includes the following components: a short-term fluctuation quantity establishment unit, which extracts the difference between the maximum and minimum values ​​of the patient's current physiological indicators within a preset sliding time window from the decision context state vector to establish short-term fluctuation quantities, and simultaneously obtains the influence intensity identifier of each physiological indicator on the decision path selection from the decision sensitivity component; an observation indicator set parsing unit, which parses the observation indicator set associated with all medical behavior nodes for each path in the initial decision path set; a path elimination processing unit, which compares the short-term fluctuation quantities one by one with the observation indicator set, and if any comparison result exceeds the preset stability boundary and the corresponding influence intensity identifier is high influence intensity, then the corresponding path is marked as an unstable path of vital sign disturbance, and path elimination processing is performed; and a pruning path set establishment unit, which establishes a pruning path set based on the path elimination processing result, and uses the pruning path set to perform the initial decision path set reconstruction.

[0069] The target decision space establishment module 40, which utilizes the decision context state vector and decision sensitivity components to reconstruct the initial decision path set, may further include: a patient diagnosis data parsing unit for parsing the patient's currently identified primary diagnosis data, recorded comorbidity list, and executed medical behavior sequence from the decision context state vector; a path cluster marking unit for traversing the trimmed path set and marking the node paths with the same primary diagnosis data and all containing at least one of the same comorbidities as path clusters of the same disease and symptom; a shared pre-examination and examination behavior chain extraction unit for extracting the shared pre-examination and examination behavior chain in the knowledge graph where each path cluster of the same disease and symptom is located; and a path merging unit for merging the path clusters of the same disease and symptom when the length of the shared pre-examination and examination behavior chain exceeds the preset minimum common chain length. Clusters are merged at the end nodes of the shared pre-examination and inspection behavior chain, and the differentiated branches of each path after the end node are retained after merging, forming a merged path that integrates multiple comorbidity treatment branches; the direct successor behavior node set construction unit is used to take the last behavior node of the executed medical behavior sequence as the current breakpoint, and use the current breakpoint to perform knowledge graph localization to construct the direct successor behavior node set; the continuous connection path construction unit is used to use the direct successor behavior node set to extract path segments starting from any direct successor behavior node from the pruned path set, and splice the path segments with the executed medical behavior sequence to form a continuous connection path; the reconstructed path set establishment unit is used to establish a reconstructed path set based on the merged path and the continuous connection path, and use the pruned path set and the reconstructed path set to perform initial decision path set reconstruction.

[0070] The target decision space establishment module 40, which utilizes the decision context state vector and decision sensitivity component to reconstruct the initial decision path set, may further include: a decision priority-driven feature marking unit for performing sensitivity increase analysis based on the decision sensitivity component in the near clinical evaluation cycle, marking at least one key observation feature whose sensitivity increase exceeds a preset surge threshold as a decision priority-driven feature; an observation-decision triggering relationship query unit for querying the observation-decision triggering relationship between medical behavior nodes and decision priority-driven features in the knowledge graph for each path in the reconstructed path set; an emergency-driven node marking unit for marking medical behavior nodes that directly depend on the decision priority-driven features as emergency-driven nodes based on the query results; a topology rearrangement path establishment unit for performing emergency-driven node relocation processing on the paths marked with emergency-driven nodes, and performing parallel candidate behavior node sorting and reconstruction on the relocated paths to establish a topology rearrangement path; and using the pruned path set, reconstructed path set, and topology rearrangement path to reconstruct the initial decision path set.

[0071] Specifically, the path relocation process for the marked emergency driving nodes involves performing emergency driving node relocation. The topology reordering path establishment unit may further include: a path query subunit for querying all evaluation nodes that have an observation-decision triggering relationship with the decision priority driving feature for the marked emergency driving nodes; a driving precursor node determination subunit for determining the earliest evaluation node in the original path sequence as the driving precursor node, moving the emergency driving node from its original position and inserting it after the driving precursor node; and an emergency driving node relocation subunit for performing anomaly analysis of the insertion behavior. If the necessary forward dependency between the emergency driving node and other behavioral nodes in the original path sequence is disrupted, the emergency driving node is moved sequentially along the path until the necessary forward dependency is fully satisfied, thus completing the emergency driving node relocation.

[0072] The topology reordering path establishment unit further includes: a decision branch node determination subunit for detecting the substitution relationship between nodes in the relocated path and selecting parallel candidate behavior nodes with multiple substitution relationships as decision branch nodes; a real-time data acquisition subunit for acquiring real-time data of each patient state dimension in the decision context state vector; a pre-admission condition set extraction subunit for extracting the pre-admission condition set recorded in the knowledge graph for each parallel candidate behavior node; an admission gap quantity sequence establishment subunit for performing item-by-item deviation analysis between the real-time data and the pre-admission condition set to establish an admission gap quantity sequence; and a sorting and reconstruction subunit for sorting the parallel candidate behavior nodes in ascending order of the admission gap quantity sequence.

[0073] The initial decision path set reconstruction is based on path information evolution consistency constraint control. The target decision space establishment module 40 may further include: an information state representation sequence construction unit for constructing an information state representation sequence corresponding to the decision context state vector, mapping any candidate decision path to an information state transition sequence that evolves sequentially along the path nodes; a path information evolution trajectory construction unit for calculating the information increment vector and uncertainty change amount of the information state change between adjacent nodes using the information state representation sequence in the information state transition sequence, forming a path information evolution trajectory; a consistency verification unit for performing consistency verification on the path information evolution trajectory before and after reconstruction during path pruning, path reorganization, and topology rearrangement; and a rollback processing unit for performing rollback processing on the current path reconstruction result when the reconstruction operation causes the path information evolution trajectory to fail to meet the consistency verification conditions.

[0074] The consistency verification unit may further include: the consistency verification includes: determining whether there is a node sequence in the reconstructed path where the direction of the information increment vector deviates from the corresponding interval of the pre-reconstructed path by more than a preset threshold; determining whether there is an abnormal transition node in the information uncertainty change of the reconstructed path where the change suddenly increases by more than a preset discontinuity threshold, and completing the consistency verification based on the determination result.

[0075] The detailed description of the specific configuration of the auxiliary decision-making module 50 is explained as follows: As mentioned above, the auxiliary decision-making module 50 may further include: recording auxiliary decisions, establishing a decision record set, and using the decision record set to perform self-optimizing decision management under manual identification.

[0076] The medical auxiliary decision-making adaptive optimization system based on knowledge graph fusion provided in this embodiment of the invention can execute the medical auxiliary decision-making adaptive optimization method based on knowledge graph fusion provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the method execution.

[0077] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0078] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An adaptive optimization method for medical auxiliary decision-making integrating knowledge graphs, characterized in that, The method includes: Acquire multi-source medical correlation data, perform time alignment and feature decomposition processing on the multi-source medical correlation data, and construct a decision context state vector; Perform feature perturbation analysis on the observed feature components in the decision context state vector, measure the impact of features on decision path selection, and construct decision sensitivity components; Construct a knowledge graph containing medical behavior nodes and observation nodes, and use the node dependencies in the knowledge graph to generate an initial set of decision paths; By utilizing the decision context state vector and decision sensitivity components, the initial decision path set is reconstructed to establish a target decision space that adaptively matches the current state. Calculate the path information gain value and path redundancy index for candidate paths in the target decision space, perform path priority ranking based on the calculation results, and output auxiliary decision.

2. The adaptive optimization method for medical auxiliary decision-making fusion based on knowledge graphs as described in claim 1, characterized in that, Reconstructing the initial decision path set using the decision context state vector and decision sensitivity components includes: Extract the difference between the maximum and minimum values ​​of the patient's current physiological indicators within a preset sliding time window from the decision context state vector to establish short-term fluctuations, and simultaneously obtain the influence intensity identifier of each physiological indicator on the decision path selection from the decision sensitivity component. For each path in the initial decision path set, analyze the set of observation indicators associated with all medical behavior nodes; By comparing short-term fluctuations one by one using the set of observation indicators, if any comparison result exceeds the preset stability boundary and the corresponding influence intensity is marked as high influence intensity, the corresponding path is marked as an unstable path of vital sign disturbance and path removal is performed. A set of pruned paths is established based on the path elimination process, and the initial decision path set is reconstructed using the pruned path set.

3. The adaptive optimization method for medical auxiliary decision-making fusion based on knowledge graphs as described in claim 2, characterized in that, Reconstructing the initial decision path set using the decision context state vector and decision sensitivity components also includes: The patient's currently identified primary diagnosis data, recorded list of comorbidities, and sequence of executed medical actions are parsed from the decision context state vector. Traverse the set of trimmed paths and mark the node paths with the same master diagnostic data and all containing at least one of the same comorbidities as path clusters of the same disease and comorbidity. For each path cluster with the same disease and symptoms, extract the shared pre-testing and inspection behavior chain in the knowledge graph where the path is located; When the length of the shared pre-examination and inspection behavior chain exceeds the preset minimum common chain length, the path clusters of the same disease and symptoms are merged at the end node of the shared pre-examination and inspection behavior chain, and the differentiated branches of each path after the end node are retained after merging, forming a merged path that integrates multiple comorbidity treatment branches. The last action node of the executed medical action sequence is taken as the current breakpoint. After performing knowledge graph localization using the current breakpoint, a set of direct successor action nodes is constructed. Using the set of direct successor behavior nodes, a path segment starting from any direct successor behavior node is extracted from the set of pruned paths, and the path segment is spliced ​​with the executed medical behavior sequence to form a continuous and connected path; A set of recombined paths is established based on the merged path and the continuous connection path. The initial decision path set is reconstructed using the pruned path set and the recombined path set.

4. The adaptive optimization method for medical auxiliary decision-making fusion based on knowledge graphs as described in claim 3, characterized in that, Reconstructing the initial decision path set using the decision context state vector and decision sensitivity components also includes: Based on the decision sensitivity component, perform sensitivity increase analysis in the near clinical evaluation cycle, and mark at least one key observation feature whose sensitivity increase exceeds a preset surge threshold as a decision priority driving feature; For each path in the recombined path set, query the observation-decision trigger relationship between medical behavior nodes and decision priority driving features in the knowledge graph; Based on the query results, medical behavior nodes that directly depend on the decision priority driving features are marked as emergency-driven nodes; For paths with marked emergency driving nodes, perform emergency driving node relocation processing, and perform parallel candidate behavior node sorting and reconstruction on the relocated paths to establish topology reordering paths; The initial decision path set is reconstructed using the trimmed path set, recombined path set, and topology rearranged path.

5. The adaptive optimization method for medical auxiliary decision-making fusion based on knowledge graphs as described in claim 4, characterized in that, Perform emergency drive node relocation processing on paths with marked emergency drive nodes, including: For the path query of the marked emergency-driven nodes, all evaluation nodes that have an observation-decision trigger relationship with the decision priority-driven feature; The earliest evaluation node in the original path sequence is determined as the driving preceding node, and the emergency driving node is moved out of its original position and inserted after the driving preceding node. An abnormal impact analysis of the insertion behavior is performed. If the necessary forward dependency between the emergency driving node and other behavior nodes in the original path sequence is disrupted, the emergency driving node is moved backward along the path one position at a time until the necessary forward dependency is fully satisfied, thus completing the relocation of the emergency driving node.

6. The adaptive optimization method for medical auxiliary decision-making fusion based on knowledge graphs as described in claim 5, characterized in that, The relocated path is then reconstructed using a parallel candidate behavior node ranking system, including: The relocation path is used to detect the substitution relationship between nodes, and parallel candidate behavior nodes with multiple substitution relationships are used as decision branch nodes; Obtain real-time data for each patient state dimension in the decision context state vector; For each parallel candidate behavior node, extract the set of pre-admission conditions recorded in the knowledge graph; The real-time data is compared with the set of pre-admission conditions to perform a deviation analysis on each item, and an admission gap sequence is established. After sorting the alignment gap sequence in ascending order, the parallel candidate behavior nodes are sorted and reconstructed.

7. The adaptive optimization method for medical auxiliary decision-making fusion based on knowledge graphs as described in claim 1, characterized in that, The initial decision path set reconstruction is based on path information evolution consistency constraint control, including: Construct an information state representation sequence corresponding to the decision context state vector, and map any candidate decision path into an information state transition sequence that evolves sequentially along the path nodes. In the information state transition sequence, the information increment vector and uncertainty change amount are calculated for the information state changes between adjacent nodes using the information state representation sequence, forming the path information evolution trajectory; During the execution of path pruning, path reorganization, and topology rearrangement, the consistency of the path information evolution trajectory before and after reconstruction is verified. When the reconstruction operation causes the path information evolution trajectory to fail to meet the consistency verification conditions, a rollback process is performed on the current path reconstruction result.

8. The adaptive optimization method for medical auxiliary decision-making fusion based on knowledge graphs as described in claim 7, characterized in that, The consistency check includes: Determine whether there is a node sequence in the reconstructed path whose information increment vector direction deviates from the corresponding interval of the path before reconstruction by more than a preset threshold. Determine whether there are any abnormal transition nodes where the information uncertainty of the reconstructed path suddenly increases beyond a preset discontinuity threshold, and complete the consistency verification based on the determination result.

9. The adaptive optimization method for medical auxiliary decision-making fusion based on knowledge graphs as described in claim 1, characterized in that, Decision records are made for auxiliary decision-making, a decision record set is established, and the decision record set is used for self-optimizing decision management under manual identification.

10. A medical auxiliary decision-making adaptive optimization system integrating knowledge graphs, characterized in that, The system is used to implement the adaptive optimization method for medical auxiliary decision-making based on the fusion of knowledge graphs as described in any one of claims 1-9, the system comprising: The decision context state vector construction module is used to acquire multi-source medical correlation data, perform time alignment and feature decomposition processing on the multi-source medical correlation data, and construct a decision context state vector. The decision sensitivity component construction module is used to perform feature perturbation analysis on the observed feature components in the decision context state vector, perform influence measurement on the decision path selection, and construct decision sensitivity components. The initial decision path set generation module is used to construct a knowledge graph containing medical behavior nodes and observation nodes, and to generate an initial decision path set using the node dependencies in the knowledge graph. The target decision space establishment module is used to reconstruct the initial decision path set using the decision context state vector and decision sensitivity components, and establish a target decision space that adaptively matches the current state. The decision support module is used to calculate the path information gain value and path redundancy index of candidate paths in the target decision space, perform path priority ranking based on the calculation results, and output decision support.

Images