Clinical examination result analysis method and system based on medical knowledge graph

Abstract

The invention relates to the technical field of medical data analysis, and discloses a clinical examination result analysis method and system based on a medical knowledge graph. The method comprises the following steps: discretizing continuous test data of a patient into data segments with change modes, mapping a time sequence mode of the data segments to corresponding entity nodes of a knowledge graph, and constructing knowledge graph segments carrying dynamic information; and performing heuristic extension by identifying key mode paths in the analysis paths to generate a candidate analysis path set. And performing synchronous iteration comparison on the core analysis path and real-time inspection data, dynamically correcting the confidence coefficient weight of each path, and finally outputting a structured report according to the optimized main analysis link. According to the method, fusion of inspection data time sequence features and the knowledge graph is realized, and the depth of result analysis and the adaptive capacity to illness state evolution are improved through dynamic interaction optimization of paths and data.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, specifically to a method and system for analyzing clinical test results based on medical knowledge graphs. Background Technology

[0002] In the field of clinical diagnostic decision support, interpreting test results using medical knowledge graphs has become an important technical tool. Existing methods typically involve simply associating a patient's individual test values ​​with corresponding entity nodes in the knowledge graph. This association is inherently static and isolated, reflecting only the state at a specific point in time and failing to effectively characterize the dynamic changes in test indicators over time. As a static knowledge base, the knowledge graph itself lacks the capacity to carry and understand such temporal dynamic information, causing the analysis to remain at the level of judging anomalies in isolated values.

[0003] Existing analytical pathway generation methods largely rely on predefined fixed rules or one-off reasoning. In graph-based reasoning, once several candidate pathological or physiological pathways are generated based on initial evidence, their evaluation often depends on pre-defined, static weights or probability models. These pathways lack a continuous, closed-loop interactive verification mechanism with the patient's subsequent, continuously generated laboratory data. The confidence level of the pathway is largely determined in the early stages of generation, making it difficult to dynamically adjust based on newly arriving individualized patient data. This reduces the analytical system's adaptability to complex and evolving conditions and its accuracy in interpretation.

[0004] There is a need for a technology that can organically integrate the temporal variation patterns of test data into a knowledge graph structure, and enable graph-based analysis paths to interact and be optimized in real time with continuously updated patient data, thereby improving the depth and reliability of clinical test analysis results. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for analyzing clinical test results based on medical knowledge graphs, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a method for analyzing clinical test results based on medical knowledge graphs, the method comprising:

[0007] The system receives patient clinical test datasets and merges them with associated subgraph structures extracted from a medical knowledge graph. These subgraph structures contain entity nodes and relational edges related to each test.

[0008] The patient's clinical test data set is discretized and segmented to generate multiple data segment sequences corresponding to entity nodes in the sub-graph structure. Each data segment sequence has a corresponding data change pattern.

[0009] The data change patterns of each data segment sequence are mapped to the corresponding entity nodes of the medical knowledge graph, forming knowledge graph fragments with temporal pattern tags;

[0010] Identify key pattern paths in knowledge graph fragments with temporal pattern tags, and determine the initial focus of analysis based on the key pattern paths;

[0011] Based on the initial analysis focus, a heuristic expansion traversal is performed in the medical knowledge graph to generate an analysis path set containing multiple candidate analysis paths;

[0012] Each candidate analysis path in the analysis path set is quantitatively evaluated using predefined multi-dimensional measurement rules to select the core analysis path set;

[0013] The set of core analytical pathways is synchronously and iteratively compared with the set of patient clinical laboratory data, and the confidence weight of each pathway in the set of core analytical pathways is dynamically adjusted.

[0014] Based on the revised confidence weights, the main analysis path is selected from the core analysis path set. The main analysis path drives the result parsing engine to output a structured clinical test result analysis report.

[0015] Preferably, the process of receiving patient clinical test datasets and extracting associated sub-graph structures from the medical knowledge graph includes:

[0016] The system analyzes the patient's clinical laboratory data set to identify all test item names. Using the identified test item names as query keys, it performs matching and retrieval in the medical knowledge graph to locate all directly corresponding entity nodes. Starting from the directly corresponding entity nodes, it performs a finite-depth traversal along the relational edges in the medical knowledge graph. The traversal rule is to only traverse relation types that have a direct logical connection with the test result inference. All entity nodes and relational edges visited during the traversal are collected to form an initial relational subgraph. In the initial relational subgraph, the semantic relevance of each entity node to the original test item set is calculated, and entity nodes and their relational edges with semantic relevance below a preset threshold are filtered out to obtain a simplified subgraph structure.

[0017] Preferably, the process of discretizing the patient's clinical laboratory data set into segments to generate multiple data segment sequences corresponding to entity nodes in the sub-graph structure includes:

[0018] For each test item with time-series observations in the patient's clinical laboratory data set, the original time-series data is formed by sorting the observation time points. Based on the statistical distribution characteristics and clinically significant inflection points of the observations in the original time-series data, segmentation boundary points are adaptively determined. According to the determined segmentation boundary points, the original time-series data is divided into continuous data segments, each containing at least one observation. The mean, variance, and slope of the change trend of the observations within each data segment are calculated. The combination of statistics is defined as the data change pattern of the data segment. All continuous data segments under the same test item are arranged in chronological order to form the data segment sequence corresponding to the test item.

[0019] Preferably, the process of mapping the data change patterns of each data segment sequence to the corresponding entity nodes of the medical knowledge graph to form knowledge graph fragments with temporal pattern tags includes:

[0020] A mapping table is established between data change patterns and predefined pattern descriptors in the medical knowledge graph. For each entity node in the subgraph structure corresponding to a test item, it is checked whether there is a corresponding data segment sequence in the patient's clinical test data set. If so, the data change pattern of each data segment in the data segment sequence is read, and the data change pattern is converted into the corresponding predefined pattern descriptor according to the mapping table. According to the order of the data segments, the series of predefined pattern descriptors obtained by conversion are used as time sequence markers and associated with the entity node. After completing the operation on all entity nodes in the subgraph structure that have corresponding data segment sequences, the resulting subgraph structure with complete time sequence markers is the knowledge graph fragment with time sequence pattern markers.

[0021] Preferably, the process of identifying key pattern paths in knowledge graph fragments with temporal pattern tags and determining the initial focus of analysis based on the key pattern paths includes:

[0022] In a knowledge graph segment with temporal pattern tags, a pattern saliency evaluation criterion is established. This criterion comprehensively considers the rarity of the pattern description, the magnitude of variation, and the deviation from normal references. All entity nodes with temporal patterns are traversed, and the saliency score of the temporal pattern tag on each entity node is calculated according to the pattern saliency evaluation criterion. Entity nodes with saliency scores exceeding the activation threshold are selected and marked as high-saliency nodes. In the knowledge graph segment with temporal pattern tags, the shortest path connecting any two high-saliency nodes is found. The shortest path constitutes a set of candidate critical paths. For each set of candidate critical paths, the saliency scores of all entity nodes on the path are aggregated, and the total saliency intensity of the path is calculated. The candidate critical path with the highest total saliency intensity is selected and determined as the critical pattern path for this round of analysis. The starting and ending entities of the critical pattern path are jointly set as the initial focus of analysis.

[0023] Preferably, the process of generating an analysis path set containing multiple candidate analysis paths by performing a heuristic expansion traversal in the medical knowledge graph based on the initial analysis focus includes:

[0024] Starting with the entities included in the initial analysis focus, a heuristic expansion rule is defined. This rule is based on the type weights of relation edges in the medical knowledge graph, the prior importance of entity nodes, and the co-occurrence probability of patterns learned from knowledge graph fragments with temporal pattern tags. Starting from the starting entity, the heuristic expansion rule is iteratively applied to select the next entity node to be visited until a predetermined path depth is reached or a path termination condition is met, forming an analysis path. By adjusting the parameters in the heuristic expansion rule or introducing random factors, multiple independent traversals are performed starting from the same set of starting entities, thereby generating multiple different analysis paths. All generated analysis paths are collected, and completely duplicated paths are removed. The remaining paths constitute a candidate analysis path set.

[0025] Preferably, the process of quantitatively evaluating each candidate analysis path in the analysis path set using predefined multi-dimensional measurement rules to select the core analysis path set includes:

[0026] The predefined multi-dimensional measurement rules include path logical coherence measurement, clinical evidence support measurement, pattern interpretation coverage measurement, and path simplicity measurement. For each candidate analysis path in the candidate analysis path set, its score on each measurement dimension is calculated. An integration weight is assigned to each measurement dimension. The scores of the candidate analysis path on each dimension are multiplied by the corresponding integration weights and then summed to obtain the comprehensive evaluation score of the candidate analysis path. A comprehensive evaluation score threshold is set, and candidate analysis paths with comprehensive evaluation scores higher than the comprehensive evaluation score threshold are selected. The selected candidate analysis paths are arranged in descending order of their comprehensive evaluation scores, and a predetermined number of paths with the highest ranking are selected to form the core analysis path set.

[0027] Preferably, the process of synchronously iteratively comparing the core analysis path set with the patient clinical laboratory data set, and dynamically adjusting the confidence weight of each path in the core analysis path set, includes:

[0028] Supplementary test data, which was not used in constructing the knowledge graph fragments with temporal pattern labels, is extracted from the patient clinical test data set. The entities and relationships involved in each analysis path in the core analysis path set are matched and verified with the supplementary test data. A data fit index for each path is calculated. An iterative feedback process is designed whereby, in each iteration, the confidence weight of each path is proportionally adjusted based on its current data fit index. Paths with high data fit indices have their confidence weights increased, and vice versa. During the iteration process, if the confidence weight of an analysis path remains below the elimination threshold for more than a preset number of iterations, the analysis path is removed from the core analysis path set. Iterative adjustments continue until the confidence weight changes of all paths remaining in the core analysis path set stabilize, or the maximum number of iterations is reached.

[0029] Preferably, the process of selecting the main analysis path from the core analysis path set based on the corrected confidence weight, and outputting a structured clinical laboratory result analysis report by driving the result parsing engine according to the main analysis path, includes:

[0030] In the core analysis path set after the confidence weight adjustment is stable, the single analysis path with the highest confidence weight is selected as the main analysis link. If multiple paths have the same and highest confidence weight, the main analysis link is selected by secondary sorting based on the logical coherence metric score of the path. The result parsing engine is configured, which has a built-in template and rule library for converting analysis paths into natural language descriptions and structured diagnostic suggestions. The selected main analysis link is input into the result parsing engine, which traverses each entity node and relation edge in the main analysis link. Combining the temporal pattern markers attached to the entity nodes, it calls the corresponding templates and rules to generate descriptive text and structured suggestion items. The generated descriptive text and structured suggestion items are organized according to the preset report format, filled into the report template, and finally a structured clinical test result analysis report is generated.

[0031] Preferably, the present invention also includes a clinical test result analysis system based on medical knowledge graph, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the clinical test result analysis method based on medical knowledge graph described above.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] Continuous clinical laboratory data is discretized into sequences of data segments with specific change patterns. These temporal patterns are then mapped to corresponding entity nodes in a knowledge graph, forming knowledge graph fragments with temporal labels. This transforms the knowledge graph from a static relational network into a dynamic structure capable of carrying individualized temporal evolution information for patients. Based on this, the relationships between entities not only include medical logical connections but also contain coupling patterns of change over time. The subsequent process of identifying key pattern paths from the knowledge graph fragments essentially involves mining the combinational patterns of laboratory indicators under specific temporal logic. This reveals the synergistic or sequential changes of different laboratory tests during disease progression, enhancing the ability to analyze disease dynamics.

[0034] The generated core analytical pathway set is synchronously and iteratively compared with continuously generated patient clinical laboratory data, and the confidence weights of each pathway are dynamically adjusted. This establishes a real-time closed loop between analytical reasoning and data validation. The evaluation of candidate pathways no longer relies on preset static rules, but rather achieves adaptive weight adjustment by continuously receiving new data and matching it with expected pathway patterns. The focus of analysis can dynamically shift based on data feedback, avoiding pathway locking due to the one-sidedness of initial data. The final determined main analytical pathway is the optimal interpretation converged after multiple rounds of data interaction validation. Its conclusions have higher consistency with the actual temporal progression of the patient's condition, reducing the risk of misjudgment from single-time-point static analysis and enhancing the clinical adaptability and reliability of the analysis results. Attached Figure Description

[0035] Figure 1 This is a schematic diagram illustrating the working principle of the clinical test result analysis method based on medical knowledge graphs as described in this invention.

[0036] Figure 2 A flowchart for extracting the associated sub-map structure;

[0037] Figure 3 A flowchart for identifying key pattern paths and determining the initial focus of analysis;

[0038] Figure 4 A bar chart showing the multi-dimensional measurement scores of candidate analysis paths in the analysis of clinical test results based on medical knowledge graphs;

[0039] Figure 5 A bar chart showing the coverage and completeness of structured report dimensions in the analysis of clinical test results for medical knowledge graphs. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see Figure 1 This invention provides a method for analyzing clinical test results based on a medical knowledge graph. The method includes: receiving a set of clinical test data from a patient, which contains multiple test indicators and their time-series observations; simultaneously extracting sub-graph structures associated with these test indicators from a pre-constructed medical knowledge graph, whereby the sub-graph structures contain entity nodes related to diseases, symptoms, and physiological indicators, as well as relational edges such as causal and indicative relationships between them; discretizing and segmenting the patient's clinical test data set, dividing the continuous time-series data of each test item into multiple clinically significant data segments, each representing a data change pattern, thereby generating a corresponding data segment sequence for each test item; mapping the data change patterns represented by these data segment sequences to corresponding entity nodes in the medical knowledge graph, attaching time-series pattern labels to these nodes, forming knowledge graph fragments with time-series pattern labels; and identifying key pattern paths within these fragments by evaluating pattern saliency, thereby determining the initial focus of analysis. Based on this initial analytical focus, a heuristic expansion traversal is performed within the medical knowledge graph to explore potentially related medical concepts and pathways, generating a set containing multiple candidate analytical pathways. Each pathway in this set is quantitatively evaluated using predefined multi-dimensional measurement rules, selecting a core set of analytical pathways with more rigorous logic and stronger evidence support. The selected core analytical pathway set is then iteratively compared with the patient's original clinical laboratory data, dynamically adjusting the confidence weights of each pathway based on the degree of data fit. Finally, based on the adjusted confidence weights, a main analytical link is selected from the core analytical pathway set to drive the results parsing engine, generating a structured clinical laboratory results analysis report.

[0042] In one embodiment of the present invention, see [reference] Figure 2The process of receiving a patient's clinical laboratory data set involves receiving a structured electronic data file. This data set includes test item names such as "white blood cell count," "serum creatinine," and "glycated hemoglobin," along with their corresponding multiple measurements and timestamps. The data parsing module identifies all test item names within the data set. Using the identified test item names as query keys, the system performs a matching search within a medical knowledge graph. This knowledge graph is a pre-built graph database containing entities such as diseases, symptoms, test indicators, and drugs, along with their relationships. The matching search locates all directly corresponding entity nodes; for example, the test item name "serum creatinine" matches the entity node named "creatinine" in the medical knowledge graph.

[0043] Starting from these directly corresponding entity nodes, the system traverses the relational edges in the medical knowledge graph at a finite depth. The traversal depth can be configured to 3 layers, and the traversal rule is to only traverse relation types that have a direct logical connection with the inference of the test results. All entity nodes and relational edges visited during the traversal are collected to form an initial relational subgraph. In the initial relational subgraph, the semantic relevance between each entity node and the original set of test items needs to be calculated; the semantic relevance can be calculated using cosine similarity based on graph embedding or statistical methods based on co-occurrence frequency. Entity nodes and their relational edges with semantic relevance below a preset threshold are filtered out. The preset threshold can be set empirically to 0.5, resulting in a simplified subgraph structure.

[0044] In some embodiments, the process of discretizing and segmenting the patient clinical laboratory data set is performed independently for each test item with time-series observations in the set. For each test item with time-series observations in the patient clinical laboratory data set, the original time-series data is formed by sorting the observation time points. Based on the statistical distribution characteristics and clinically significant inflection points of the observations in the original time-series data, the segmentation boundary points are adaptively determined; one method for adaptively determining the segmentation boundary points is to use a sliding window to calculate the local mean and variance, and when the change in statistical characteristics between adjacent windows exceeds a dynamic threshold... If so, a segmentation boundary point is set. Dynamic threshold. The calculation formula is:

[0045] ;

[0046] in: This represents the mean of the observations within a sliding window. This represents the standard deviation of the observations within the sliding window. It is an adjustable sensitivity coefficient. The original time series data is divided into continuous data segments based on the determined segmentation boundary points, and each data segment contains at least one observation.

[0047] In practice, the mean, variance, and trend slope of the internal observations are calculated for each data segment. The mean reflects the average level, the variance reflects the degree of fluctuation, and the trend slope is obtained through linear fitting. The combination of these three statistics—mean, variance, and trend slope—is defined as the data variation pattern of the data segment. Arranging all consecutive data segments under the same test item in chronological order constitutes the data segment sequence corresponding to the test item. It can be understood that the generation of the data segment sequence transforms continuous time-series data into a discrete pattern sequence with clear statistical characteristics and clinical significance. Optionally, when calculating the trend slope, for data segments containing only a single observation, the slope can be defined as 0 or a special identifier value.

[0048] In some embodiments, for the "serum creatinine" test item, the patient's clinical test dataset provides daily measurements over the past month, with the original time-series data containing 30 observation points. An adaptive segmentation algorithm may identify two segmentation boundary points, thereby dividing the 30 observation points into three consecutive data segments. The first data segment contains the observations from the previous 10 days, and its mean, variance, and slope are calculated to form the first data variation pattern. The second and third data segments follow the same pattern, and the final data segment sequence corresponding to the "serum creatinine" test item consists of these three time-ordered data segments and their data variation patterns. It is understandable that different test items, due to their different observation frequencies and numerical variation characteristics, will generate data segment sequence lengths and pattern features that vary.

[0049] In one embodiment of the present invention, establishing a mapping table between data change patterns and predefined pattern descriptors in a medical knowledge graph is the first step. This mapping table is either a static configuration table stored in the system database or a dynamic set of calculation rules. It explicitly defines the feature range of the data change pattern corresponding to each predefined pattern descriptor. The features of the data change pattern are characterized by a combination of statistical measures such as mean, variance, and slope of the change trend. For each predefined pattern descriptor, the mapping table sets its corresponding mean change range. arrive Variance threshold and slope determination conditions The matching process of mapping relationships can be achieved by calculating the similarity between the feature vectors of actual data change patterns and the feature templates of predefined pattern descriptors. To achieve similarity The calculation formula is:

[0050] ;

[0051] in: The function represents the mean of the observations in the data segment. Used to evaluate whether the mean conforms to the mean range of a predefined pattern descriptor; The function represents the variance of the observed values ​​in the data segment. Used to evaluate whether the variance exceeds a predefined variance threshold for a pattern descriptor. ; The slope of the trend representing the data segment, the function Used to evaluate whether the slope meets the criteria of a predefined pattern descriptor. ; , and These are the weighting coefficients assigned to different features.

[0052] In practical implementation, for each entity node in the sub-graph structure corresponding to a test item, the system searches for a corresponding data segment sequence in the patient's clinical test data set. This search is achieved by comparing the entity node's identifier with the test item name. If a corresponding data segment sequence exists, the system reads the data change pattern of each data segment, which includes pre-calculated mean, variance, and trend slope. Based on the mapping table, the data change pattern is converted into a corresponding predefined pattern descriptor. The conversion process involves inputting the data change pattern features of the data segment into the mapping table, and finding the predefined pattern descriptor with the highest similarity through similarity calculation or rule matching as the conversion result. Following the chronological order of the data segments, the converted series of predefined pattern descriptors are used as time-series markers and associated with the entity nodes. This association can be achieved by adding a list named "Time-Series Pattern Marker" to the entity node's attribute fields, storing the predefined pattern descriptors in sequence.

[0053] In some embodiments, for an entity node corresponding to "white blood cell count," its corresponding data segment sequence in the patient's clinical laboratory data set contains three consecutive data segments. The data change pattern of the first data segment is determined by a mapping table and matches the predefined pattern descriptor "rapid rise"; the data change pattern of the second data segment matches "high plateau"; and the data change pattern of the third data segment matches "slow decline." In chronological order, the system uses these three predefined pattern descriptors—"rapid rise," "high plateau," and "slow decline"—as time-series markers and stores them sequentially in the "time-series pattern marker" attribute of the "white blood cell count" entity node. After performing the above operation on all entity nodes with corresponding data segment sequences in the subgraph structure, the resulting subgraph structure with complete time-series markers is the knowledge graph fragment with time-series pattern markers. It can be understood that entity nodes without corresponding data segment sequences will not carry time-series pattern markers.

[0054] Optionally, the mapping table supports dynamic updates, allowing the addition of new predefined pattern descriptors or adjustment of feature matching rules for existing descriptors based on new medical knowledge or clinical practice feedback. In some embodiments, the mapping process between data change patterns and predefined pattern descriptors may not be one-to-one; a data change pattern may simultaneously have a similar degree of matching with multiple predefined pattern descriptors. In this case, the system can select the descriptor with the highest matching degree, or record multiple possible descriptors in the time-series tag along with their confidence levels. It can be understood that time-series pattern tagging injects patient-specific dynamic temporal information into the static medical concepts in the knowledge graph. Optionally, when attaching time-series tags, in addition to recording the predefined pattern descriptors, the system can also simultaneously record the statistical values ​​or time ranges of the original data segments that generated the tag.

[0055] In one embodiment of the present invention, see [reference] Figure 3 In a knowledge graph segment with temporal pattern labels, a pattern significance evaluation criterion is established. This criterion comprehensively considers the rarity of the pattern description, the magnitude of variation, and the deviation from the normal reference range. Rarity is quantified by the reciprocal of the frequency of the predefined pattern descriptor in historical healthy population data. The magnitude of variation is measured by the ratio of the range or standard deviation of the observed values ​​within the data segment to the individual baseline value. The deviation from the normal reference range is calculated based on the degree to which the mean of the data segment deviates from the upper and lower limits of the clinical normal value. All entity nodes with temporal patterns are traversed, and the significance score of the temporal pattern label on each entity node is calculated according to the pattern significance evaluation criterion. The significance score can be calculated by weighted summation after scoring the three dimensions mentioned above. Entity nodes with significance scores exceeding the activation threshold (a preset numerical limit) are selected and marked as high-significance nodes. In the knowledge graph segment with temporal pattern labels, the shortest path connecting any two high-significance nodes is found. The shortest path is found using breadth-first search or Dijkstra's algorithm in graph theory, and the shortest paths form a set of candidate critical paths. For each candidate critical path, the saliency scores of all entity nodes on the path are aggregated to calculate the total saliency strength of the path. One method for calculating the total saliency strength is... The formula is:

[0056] ;

[0057] in: The first candidate critical path The significance score of each entity node. Represents the total number of entity nodes on the candidate critical path. It is a decay coefficient between 0 and 1, used to adjust the contribution of nodes according to their order in the path. The candidate critical path with the highest total significance is selected and identified as the critical pattern path for this round of analysis. The starting and ending entities of the critical pattern path are jointly set as the initial focus of analysis.

[0058] In some embodiments, the knowledge graph fragment with time-series pattern label contains two highly significant nodes, “serum potassium” and “electrocardiogram QT interval”. The system finds the shortest path connecting these two nodes, which contains the entity “hyperkalemia”. This path is calculated to have the highest total significance intensity and is thus identified as the key pattern path. The initial analysis focus is set to “serum potassium” and “electrocardiogram QT interval”.

[0059] It is understandable that the process of heuristically expanding the medical knowledge graph based on the initial analysis focus starts with the entities contained in the initial analysis focus. Heuristic expansion rules are defined based on the type weights of relation edges in the medical knowledge graph, the prior importance of entity nodes, and the co-occurrence probabilities of patterns learned from knowledge graph fragments with temporal pattern labels. The type weights of relation edges are predefined; for example, the weight of the "cause" relation is higher than that of the "related to" relation. The prior importance of entity nodes can be set according to the frequency of their appearance in medical textbooks or guidelines. The pattern co-occurrence probability is calculated by statistically analyzing the frequency of different pattern descriptors appearing simultaneously in the same patient from the knowledge graph fragments with temporal pattern labels. Starting from the starting entity, the heuristic expansion rules are iteratively applied to select the next entity node to be visited. The selection process can be based on a scoring function that comprehensively evaluates the weights of each relation edge originating from the current node, as well as the importance and co-occurrence probability of the target node. The node with the highest score is selected as the next node to be visited. The iterative process continues until a predetermined path depth is reached or a path termination condition is met. The path depth is, for example, set to the fourth node after the starting point, and the path termination condition is, for example, encountering an entity node representing a clear diagnosis, thus forming an analysis path. Multiple different analysis paths are generated by adjusting the parameters in the heuristic extension rules or introducing random factors, such as randomly selecting one of the top two nodes in the scoring function for visit, and performing multiple independent traversals starting from the same set of starting entities. All generated analysis paths are then collected, and completely duplicated paths are removed; the remaining paths constitute a candidate analysis path set.

[0060] Optionally, the pattern saliency evaluation criteria can also incorporate consideration of the temporal pattern tag length, i.e., the length of the predefined pattern descriptor sequence attached to an entity node. Longer sequences may reflect more complex dynamic processes. In some embodiments, the pattern co-occurrence probability in the heuristic expansion rule can be dynamically updated. As new entity nodes with patterns are visited during the traversal, the co-occurrence probability matrix is ​​adjusted in real time to influence subsequent node selection. It is understood that the initial analysis focus may contain more than two entities. In this case, the starting point of the heuristic expansion traversal is a set of entities, and the traversal can begin from each entity individually or from a virtual central node. Optionally, loop structures are allowed in the generated analysis path, but a maximum loop length must be set to avoid infinite loops.

[0061] In one embodiment of the present invention, predefined multi-dimensional measurement rules include path logical coherence measurement, clinical evidence support measurement, pattern interpretation coverage measurement, and path simplicity measurement. The path logical coherence measurement assesses the medical logical rationality of the relationships between adjacent entity nodes in a candidate analysis path, generating scores by querying the prior strength of relationship types in a medical knowledge base or encoding the path subgraph using a pre-trained graph neural network model. The clinical evidence support measurement assesses the degree of matching between the medical entities involved in the candidate analysis path and abnormal patterns in the current patient clinical test data set, achieved by calculating the degree of deviation of the test data corresponding to each node in the path from the normal range and performing a weighted average. The pattern interpretation coverage measurement assesses how many abnormal patterns marked as highly significant in a knowledge graph segment with temporal pattern labeling can be explained by the candidate analysis path, calculated as the proportion of highly significant nodes traversed by the path to the total number of highly significant nodes. The path simplicity measurement tends to select shorter paths, and its score is inversely proportional to the total number of entity nodes in the path, calculated as: Path Simplicity Measurement Score = ,in This represents the total number of entity nodes in the candidate analysis path. See Table 1.

[0062] Table 1: Multidimensional Measurement Table of Candidate Analysis Paths

[0063] For each candidate analysis path in the candidate analysis path set, the system calculates its score across four dimensions: path logical coherence, clinical evidence support, pattern interpretation coverage, and path simplicity. An integration weight is assigned to each dimension; these weights are pre-defined by domain experts or optimized using machine learning methods. The overall evaluation score of the candidate analysis path is obtained by multiplying its score across each dimension by its corresponding integration weight and then summing the results. Set a comprehensive evaluation score threshold and filter out candidate analysis paths whose comprehensive evaluation scores are higher than the threshold. Sort the selected candidate analysis paths in descending order of their comprehensive evaluation scores, and select a predetermined number of paths with the highest rankings, such as the top five paths, to form the core analysis path set.

[0064] In some embodiments, supplementary laboratory data not used in constructing knowledge graph fragments with temporal pattern labels is extracted from the patient's clinical laboratory data set. This supplementary data may include imaging report summaries, pathological biopsy results text, or individual laboratory indicators not included in the initial temporal analysis. The entities and relationships involved in each analysis path in the core analysis path set are matched and validated against the supplementary laboratory data. For example, if an analysis path points to the entity "acute glomerulonephritis," the system searches the supplementary laboratory data for keywords such as "glomeruli," "nephritis," and "biopsy" to determine if there are supporting or negative descriptions. A data fit index is calculated for each path. The calculation formula is:

[0065] ;

[0066] in: This represents the number of items in the supplementary test data that have supporting evidence for the pathway medicine judgment. The number of projects representing the amount of clearly contradictory evidence. This is a very small positive value used to prevent the denominator from being zero. Design an iterative feedback process where, in each iteration, the confidence weight of each path is adjusted proportionally based on its current data consistency index. Paths with higher data consistency indices have their confidence weight increased, and vice versa. The adjustment rule can be: new confidence weight = old confidence weight × (1 + ... ×( - )),in: It's the learning rate. This is the average of the data fit indexes in the current core analysis path set. During the iteration process, if the confidence weight of an analysis path remains below the elimination threshold for more than a preset number of iterations (e.g., the weight is below 0.2 for three consecutive iterations), the analysis path will be removed from the core analysis path set. Iterative adjustments continue until the confidence weight changes of all paths remaining in the core analysis path set tend to stabilize (e.g., the maximum weight change between two consecutive iterations is less than 0.01), or the maximum number of iterations is reached (e.g., 20 iterations).

[0067] It is understandable that supplementary test data may exist in the form of unstructured text. In this case, natural language processing techniques are needed for information extraction and medical concept matching to make a judgment on support or contradiction. In some embodiments, the calculation of the data consistency index not only considers the supplementary test data but may also re-examine the initial clinical test data set to check for any overlooked weak patterns that support the current path. Optionally, the adjustment of confidence weights in the iterative feedback process can introduce a simulated annealing strategy, allowing for larger adjustments in early iterations to explore the space and smaller adjustments in later iterations to achieve convergence and stability. It is understandable that the dynamic correction process makes the core analysis path set more closely match the patient's complete chain of evidence. Optionally, removed paths are recorded, and if their data consistency index rises in subsequent iterations due to newly discovered supporting evidence, the system may consider reintroducing them into the core analysis path set.

[0068] See Figure 4 This is a bar chart showing the multi-dimensional scores of candidate analysis paths in clinical laboratory results analysis based on a medical knowledge graph. The differences in scores across various dimensions for different paths can help select the core path best suited to clinical needs. The balance of multi-dimensional scores (e.g., no significant weakness in P002) is a crucial guarantee of the reliability of clinical laboratory results analysis. Scores for path logical coherence and pattern interpretation coverage directly reflect the associative matching ability of the medical knowledge graph in clinical analysis. Transforming abstract multi-dimensional evaluation indicators into intuitive score comparisons helps quickly identify the strengths and weaknesses of each candidate path, avoiding biases from subjective judgment.

[0069] In one embodiment of the present invention, in the core analysis path set after the confidence weight adjustment is stable, the single analysis path with the highest confidence weight is selected as the main analysis link. The system traverses the current confidence weight value of each analysis path in the core analysis path set and identifies the path corresponding to the maximum value through comparison operations. If multiple paths have the same and highest confidence weight, the main analysis link is selected by secondary sorting based on the logical coherence metric score of the paths. The secondary sorting process arranges these paths with the same weight in descending order of their logical coherence metric score and selects the path with the highest ranking. A result parsing engine is configured. The result parsing engine is an independent software module. The result parsing engine has built-in templates and rule bases for converting analysis paths into natural language descriptions and structured diagnostic suggestions. The templates and rule bases contain preset text generation patterns and diagnostic suggestion frameworks for different medical entities and relationship types.

[0070] In specific implementation, the selected main analysis link is input into the result parsing engine, which traverses each entity node and relation edge in the main analysis link. Combining the temporal pattern markers attached to the entity nodes, the result parsing engine calls the corresponding templates and rules to generate descriptive text and structured suggestions. For the first entity node "fever" in the main analysis link, which has the temporal pattern marker "recurrent high fever," the result parsing engine matches the descriptive pattern "the [entity] of the patient exhibiting [pattern]" in the template and rule base, thus generating the descriptive text "the patient exhibits recurrent high fever." For the relation edge "cause" and its connected next entity "bacterial infection," the result parsing engine matches rules to generate "consider it caused by [cause]," and continues to parse the possible test evidence patterns attached to the "bacterial infection" entity. In some embodiments, the process of the result parsing engine generating descriptive text and structured suggestions follows a recursive logic. The recursive logic starts from the starting entity of the main analysis link, sequentially processes each node and edge until the ending entity, and semantically connects the text fragments generated in each step to form a coherent paragraph. The generated structured recommendations may include categories such as "recommended examinations", "differential diagnoses", and "treatment principles", with each recommendation associated with an entity or relation fragment in the main analysis link.

[0071] The generated descriptive text and structured suggestions are organized according to a preset report format, which defines the chapter titles, paragraph order, and presentation of structured suggestions. This is then populated into a report template, a document framework containing fixed titles, paragraph position identifiers, and variable placeholders, ultimately generating a structured clinical laboratory results analysis report. It's understandable that the results parsing engine needs to handle complex combinations of temporal pattern tags and entities / relationships when calling templates and rules. In some embodiments, when entity nodes in the main analysis chain do not have attached temporal pattern tags, the results parsing engine uses a default descriptive template, generating text solely based on the medical meaning of entities and relationships. Optionally, the results parsing engine can introduce a text fluency optimization step when generating descriptive text. This step uses a pre-trained language model to fine-tune the initially assembled text, improving the naturalness and professionalism of the sentences. If, during the main analysis link selection process, multiple paths are found to have identical confidence weights and logical coherence scores, the system can employ a random selection strategy or introduce a third metric (such as path timeliness) for adjudication. The adjudication formula is as follows:

[0072] ;

[0073] in: Represents the final selected path. This represents the average of the latest timestamps of the test data corresponding to all entities in the candidate path. This indicates the parameter corresponding to the maximum value. It can be understood that the structured clinical laboratory results analysis report is ultimately output as an electronic document or directly integrated into a designated interface of the hospital information system. Optionally, after the report is generated, the system can record the composition of the main analysis chain and the selection criteria for use in auditing and model optimization.

[0074] See Figure 5 This is a bar chart showing the coverage and completeness of structured report dimensions in clinical laboratory results analysis using a medical knowledge graph. It displays the coverage and completeness of the six core dimensions of a structured clinical laboratory report: "Basic Information, Symptom Description, Laboratory Evidence, Diagnostic Recommendations, Treatment Principles, and Differential Diagnosis." Quantitative indicators visually reflect the completion quality of each dimension of the structured report, helping to quickly identify weak dimensions such as "Treatment Principles" and "Differential Diagnosis." Dimensions with low completeness (such as differential diagnosis) can guide the results analysis engine to supplement corresponding templates and rules, improving the comprehensiveness of the report content. The high coverage / completeness of the "Basic Information" and "Symptom Description" dimensions ensures the reliability of the report's transmission of basic clinical information; while dimensions requiring optimization clearly indicate directions for report improvement.

[0075] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0076] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A clinical laboratory result analysis method based on medical knowledge graphs, characterized in that, Includes the following steps: The system receives patient clinical test datasets and merges them with associated subgraph structures extracted from a medical knowledge graph. These subgraph structures contain entity nodes and relational edges related to each test. The patient's clinical test data set is discretized and segmented to generate multiple data segment sequences corresponding to entity nodes in the sub-graph structure. Each data segment sequence has a corresponding data change pattern. The data change patterns of each data segment sequence are mapped to the corresponding entity nodes of the medical knowledge graph, forming knowledge graph fragments with temporal pattern tags; Identify key pattern paths in knowledge graph fragments with temporal pattern tags, and determine the initial focus of analysis based on the key pattern paths; Based on the initial analysis focus, a heuristic expansion traversal is performed in the medical knowledge graph to generate an analysis path set containing multiple candidate analysis paths; Each candidate analysis path in the analysis path set is quantitatively evaluated using predefined multi-dimensional measurement rules to select the core analysis path set; The set of core analytical pathways is synchronously and iteratively compared with the set of patient clinical laboratory data, and the confidence weight of each pathway in the set of core analytical pathways is dynamically adjusted. Based on the revised confidence weights, the main analysis path is selected from the core analysis path set. The main analysis path drives the result parsing engine to output a structured clinical test result analysis report.

2. The clinical test result analysis method based on medical knowledge graph according to claim 1, characterized in that, The process of receiving patient clinical test datasets and extracting associated sub-graph structures from the medical knowledge graph includes: The system analyzes the patient's clinical laboratory data set to identify all test item names. Using the identified test item names as query keys, it performs matching and retrieval in the medical knowledge graph to locate all directly corresponding entity nodes. Starting from the directly corresponding entity nodes, it performs a finite-depth traversal along the relational edges in the medical knowledge graph. The traversal rule is to only traverse relation types that have a direct logical connection with the test result inference. All entity nodes and relational edges visited during the traversal are collected to form an initial relational subgraph. In the initial relational subgraph, the semantic relevance of each entity node to the original test item set is calculated, and entity nodes and their relational edges with semantic relevance below a preset threshold are filtered out to obtain a simplified subgraph structure.

3. The clinical test result analysis method based on medical knowledge graph according to claim 2, characterized in that, The process of discretizing and segmenting patient clinical laboratory data sets to generate multiple data segment sequences corresponding to entity nodes in the sub-graph structure includes: For each test item with time-series observations in the patient's clinical laboratory data set, the original time-series data is formed by sorting the observation time points. Based on the statistical distribution characteristics and clinically significant inflection points of the observations in the original time-series data, segmentation boundary points are adaptively determined. According to the determined segmentation boundary points, the original time-series data is divided into continuous data segments, each containing at least one observation. The mean, variance, and slope of the change trend of the observations within each data segment are calculated. The combination of statistics is defined as the data change pattern of the data segment. All continuous data segments under the same test item are arranged in chronological order to form the data segment sequence corresponding to the test item.

4. The clinical test result analysis method based on medical knowledge graph according to claim 3, characterized in that, The process of mapping the data change patterns of each data segment sequence to the corresponding entity nodes of the medical knowledge graph, forming knowledge graph fragments with temporal pattern labels, includes: A mapping table is established between data change patterns and predefined pattern descriptors in the medical knowledge graph. For each entity node in the subgraph structure corresponding to a test item, it is checked whether there is a corresponding data segment sequence in the patient's clinical test data set. If so, the data change pattern of each data segment in the data segment sequence is read, and the data change pattern is converted into the corresponding predefined pattern descriptor according to the mapping table. According to the order of the data segments, the series of predefined pattern descriptors obtained by conversion are used as time sequence markers and associated with the entity node. After completing the operation on all entity nodes in the subgraph structure that have corresponding data segment sequences, the resulting subgraph structure with complete time sequence markers is the knowledge graph fragment with time sequence pattern markers.

5. The clinical test result analysis method based on medical knowledge graph according to claim 4, characterized in that, The process of identifying key pattern paths in knowledge graph fragments with temporal pattern tags and determining the initial focus of analysis based on these key pattern paths includes: In a knowledge graph segment with temporal pattern tags, a pattern saliency evaluation criterion is established. This criterion comprehensively considers the rarity of the pattern description, the magnitude of variation, and the deviation from normal references. All entity nodes with temporal patterns are traversed, and the saliency score of the temporal pattern tag on each entity node is calculated according to the pattern saliency evaluation criterion. Entity nodes with saliency scores exceeding the activation threshold are selected and marked as high-saliency nodes. In the knowledge graph segment with temporal pattern tags, the shortest path connecting any two high-saliency nodes is found. The shortest path constitutes a set of candidate critical paths. For each set of candidate critical paths, the saliency scores of all entity nodes on the path are aggregated, and the total saliency intensity of the path is calculated. The candidate critical path with the highest total saliency intensity is selected and determined as the critical pattern path for this round of analysis. The starting and ending entities of the critical pattern path are jointly set as the initial focus of analysis.

6. The clinical test result analysis method based on medical knowledge graph according to claim 5, characterized in that, Based on the initial analysis focus, the process of performing a heuristic expansion traversal in the medical knowledge graph to generate a set of analysis paths containing multiple candidate analysis paths includes: Starting with the entities included in the initial analysis focus, a heuristic expansion rule is defined. This rule is based on the type weights of relation edges in the medical knowledge graph, the prior importance of entity nodes, and the co-occurrence probability of patterns learned from knowledge graph fragments with temporal pattern tags. Starting from the starting entity, the heuristic expansion rule is iteratively applied to select the next entity node to be visited until a predetermined path depth is reached or a path termination condition is met, forming an analysis path. By adjusting the parameters in the heuristic expansion rule or introducing random factors, multiple independent traversals are performed starting from the same set of starting entities, thereby generating multiple different analysis paths. All generated analysis paths are collected, and completely duplicated paths are removed. The remaining paths constitute a candidate analysis path set.

7. The clinical test result analysis method based on medical knowledge graph according to claim 6, characterized in that, The process of quantitatively evaluating each candidate analysis path in the analysis path set using predefined multi-dimensional measurement rules to select the core analysis path set includes: The predefined multi-dimensional measurement rules include path logical coherence measurement, clinical evidence support measurement, pattern interpretation coverage measurement, and path simplicity measurement. For each candidate analysis path in the candidate analysis path set, its score on each measurement dimension is calculated. An integration weight is assigned to each measurement dimension. The scores of the candidate analysis path on each dimension are multiplied by the corresponding integration weights and then summed to obtain the comprehensive evaluation score of the candidate analysis path. A comprehensive evaluation score threshold is set, and candidate analysis paths with comprehensive evaluation scores higher than the comprehensive evaluation score threshold are selected. The selected candidate analysis paths are arranged in descending order of their comprehensive evaluation scores, and a predetermined number of paths with the highest ranking are selected to form the core analysis path set.

8. The clinical test result analysis method based on medical knowledge graph according to claim 7, characterized in that, The process of synchronously and iteratively comparing the core analytical pathway set with the patient clinical laboratory data set, and dynamically adjusting the confidence weights of each pathway in the core analytical pathway set, includes: Supplementary test data, which was not used in constructing the knowledge graph fragments with temporal pattern labels, is extracted from the patient clinical test data set. The entities and relationships involved in each analysis path in the core analysis path set are matched and verified with the supplementary test data. A data fit index for each path is calculated. An iterative feedback process is designed whereby, in each iteration, the confidence weight of each path is proportionally adjusted based on its current data fit index. Paths with high data fit indices have their confidence weights increased, and vice versa. During the iteration process, if the confidence weight of an analysis path remains below the elimination threshold for more than a preset number of iterations, the analysis path is removed from the core analysis path set. Iterative adjustments continue until the confidence weight changes of all paths remaining in the core analysis path set stabilize, or the maximum number of iterations is reached.

9. The clinical test result analysis method based on medical knowledge graph according to claim 8, characterized in that, Based on the corrected confidence weights, the process of selecting the main analysis path from the core analysis path set, and then using the main analysis path to drive the result parsing engine to output a structured clinical laboratory result analysis report includes: In the core analysis path set after the confidence weight adjustment is stable, the single analysis path with the highest confidence weight is selected as the main analysis link. If multiple paths have the same and highest confidence weight, the main analysis link is selected by secondary sorting based on the logical coherence metric score of the path. The result parsing engine is configured, which has a built-in template and rule library for converting analysis paths into natural language descriptions and structured diagnostic suggestions. The selected main analysis link is input into the result parsing engine, which traverses each entity node and relation edge in the main analysis link. Combining the temporal pattern markers attached to the entity nodes, it calls the corresponding templates and rules to generate descriptive text and structured suggestion items. The generated descriptive text and structured suggestion items are organized according to the preset report format, filled into the report template, and finally a structured clinical test result analysis report is generated.

10. A clinical test result analysis system based on medical knowledge graph, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the clinical test result analysis method based on medical knowledge graph as described in any one of claims 1 to 9.

Images