A remote medical quality control abnormal data automatic attribution analysis method and system

Abstract

The application relates to the technical field of data processing, and discloses a remote medical quality control abnormal data automatic attribution analysis method and system. The method comprises the following steps: collecting patient physiological parameters, medical equipment operation states and diagnosis and treatment operation records into a multi-modal data set of remote medical quality control; setting a dynamic threshold, identifying abnormal points deviating from the dynamic threshold in the multi-modal data set, and obtaining an abnormal propagation data set; vectorizing medical entities and associated relationships in a preset integrated medical knowledge base to obtain an initial graph; updating node and edge weights in the initial graph in real time to obtain a dynamic knowledge graph; performing logical deduction on the abnormal propagation data set, tracing potential root causes leading to the abnormal propagation data set, and obtaining an attribution hypothesis set; bidirectionally calibrating causal relationships between the dynamic knowledge graph and the attribution hypothesis set to obtain an attribution analysis report; and the application can improve the efficiency of remote medical quality control abnormal data automatic attribution analysis.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an automatic attribution analysis method and system for abnormal data in remote medical quality control. Background Technology

[0002] Remote medical quality control involves multiple types of data, including patient physiological parameters, medical equipment operating status, and treatment records. Existing technologies struggle to efficiently collect and standardize this multimodal data. Significant differences in data formats and complex correlation logic result in a lack of comprehensiveness in capturing anomalies. Furthermore, threshold settings are often fixed and cannot be adaptively adjusted based on dynamic changes in historical data, thus affecting the accuracy of anomaly identification and making it difficult to form complete and realistic anomaly propagation correlations.

[0003] Current attribution analysis of abnormal data relies on static medical knowledge bases, which cannot incorporate new diagnostic and treatment evidence in real time to update the relationships between medical entities, resulting in a lack of timeliness and accuracy in causal reasoning. Furthermore, the lack of a two-way calibration mechanism between the knowledge graph and the reasoning assumptions during the attribution process makes logical deduction prone to bias and difficult to accurately trace the root cause of abnormal data. This leads to low efficiency in attribution analysis and fails to meet the needs of remote medical quality control for rapid location and timely handling of abnormal issues. Therefore, improving the efficiency of automatic attribution analysis of abnormal data in remote medical quality control has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides an automatic attribution analysis method and system for abnormal data in remote medical quality control, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an automatic attribution analysis method for abnormal data in remote medical quality control, comprising:

[0006] S1. Collect patient physiological parameters, medical equipment operating status, and diagnosis and treatment operation records into a multimodal dataset for remote medical quality control.

[0007] S2. Based on the mean and standard deviation in the historical data of the remote medical quality control, set the dynamic threshold of the remote medical quality control, and identify the outliers in the multimodal dataset that deviate from the dynamic threshold to obtain the abnormal propagation dataset of the remote medical quality control.

[0008] S3. Vectorize the medical entities and relationships in the preset integrated medical knowledge base to obtain the initial map of the remote medical quality control;

[0009] S4. Use the new evidence in the multimodal dataset to update the node and edge weights in the initial graph in real time to obtain the dynamic knowledge graph of the remote medical quality control.

[0010] S5. Based on the dynamic knowledge graph, perform logical deduction on the abnormal propagation dataset to trace the potential root causes of the abnormal propagation dataset and obtain the attribution hypothesis set for the remote medical quality control.

[0011] S6. Perform bidirectional calibration of the causal relationship between the dynamic knowledge graph and the attribution hypothesis set to obtain the attribution analysis report of the remote medical quality control.

[0012] In a preferred embodiment, the step of compiling the collected patient physiological parameters, medical equipment operating status, and treatment operation records into a multimodal dataset for remote medical quality control includes:

[0013] The format is parsed to obtain the standard data record of the remote medical quality control by parsing the patient's physiological parameters, the operating status of medical equipment and the diagnosis and treatment operation records, and the parsed data is mapped to fields.

[0014] The standard data records are synchronously organized in chronological order to obtain the multimodal data sequence of the remote medical quality control.

[0015] The multimodal data sequences are aggregated into a multimodal dataset for remote medical quality control.

[0016] In a preferred embodiment, the step of setting a dynamic threshold for remote medical quality control based on the mean and standard deviation of the historical data of the remote medical quality control, and identifying outliers in the multimodal dataset that deviate from the dynamic threshold to obtain the anomaly propagation dataset of the remote medical quality control includes:

[0017] By sampling through a sliding window, the mean and standard deviation of the historical data of the remote medical quality control are statistically analyzed to generate the benchmark parameters of the remote medical quality control.

[0018] Based on the benchmark parameters, the dynamic thresholds of different types of data streams in the multimodal dataset are defined to obtain the threshold boundary range of remote medical quality control;

[0019] Based on the threshold boundary range, the real-time data stream in the multimodal dataset is traversed to identify data points that exceed the threshold boundary range, thereby obtaining the anomaly set of the multimodal dataset;

[0020] Based on the temporal order and device correlation of the data in the anomaly set, an anomaly propagation chain of the anomaly set is constructed;

[0021] Based on the topology of the anomaly propagation chain, causal anomalies in the anomaly set are aggregated to obtain the anomaly propagation dataset for remote medical quality control.

[0022] In a preferred embodiment, the step of traversing the real-time data stream in the multimodal dataset based on the threshold boundary range to identify data points exceeding the threshold boundary range, thereby obtaining an outlier set of the multimodal dataset, includes:

[0023] Based on the data source type, extract the classification features of the real-time data in the multimodal dataset;

[0024] The classification features are mapped to the threshold boundary range to filter out the preliminary set of outliers in the multimodal dataset;

[0025] Redundant data points are removed from the initial set of outliers to obtain the set of outliers in the multimodal dataset.

[0026] In a preferred embodiment, the vectorization of medical entities and relationships in the pre-defined integrated medical knowledge base to obtain the initial map of the remote medical quality control includes:

[0027] The medical entity description text in the pre-defined integrated medical knowledge base is parsed, and key medical features are extracted using natural language processing technology to generate the entity feature set of the integrated medical knowledge base.

[0028] Multi-dimensional feature encoding is performed on the medical entities in the entity feature set to convert semantic information into numerical vectors, thereby obtaining the entity vector set of the entity feature set;

[0029] Identify the relationships between medical entities in the integrated medical knowledge base, quantify the relationship type and strength into relationship weight values, and obtain the relationship weight set of the medical entities;

[0030] Based on the entity vector and the relation weight set, an initial graph for remote medical quality control is constructed, with the medical entity as the node and the association relationship as the edge.

[0031] In a preferred embodiment, identifying the relationships between medical entities in the integrated medical knowledge base, quantifying the relationship type and strength into relationship weight values, and obtaining the relationship weight set of the medical entities includes:

[0032] The relationships between medical entities in the integrated medical knowledge base are analyzed, the relationship types are identified, and the relationship type set of the integrated medical knowledge base is obtained.

[0033] Based on the level of medical evidence, the strength of the relationships between medical entities in the integrated medical knowledge base is assessed to obtain the relationship strength level of the medical entities.

[0034] Based on the set of association types and the relationship strength level, the association relationships of medical entities are quantified into relationship weight values ​​for the medical entities, wherein the formula for calculating the relationship weight value is as follows:

[0035] ;

[0036] In the formula, The relation weight value, This represents the importance coefficient of the type of relationship between entities in the medical entity. This serves as the baseline value for the strength of the association between the medical entities. As a preset structural importance factor, This is a moderating index for the importance of inter-entity association types within the medical entity. This refers to the time decay adjustment coefficient in the integrated medical knowledge base. This refers to the time decay factor in the integrated medical knowledge base;

[0037] The relationship weight values ​​are bound to the association relationships to obtain the relationship weight set of the medical entity.

[0038] In a preferred embodiment, the step of updating the node and edge weights in the initial graph in real time using new evidence from the multimodal dataset to obtain the dynamic knowledge graph of the remote medical quality control includes:

[0039] Extract new evidence data from the multimodal dataset that is associated with medical entities in the initial atlas, and generate an updated evidence set for the multimodal dataset;

[0040] The credibility and timeliness of the new evidence in the updated evidence set are evaluated to obtain the evidence weight coefficient of the updated evidence set;

[0041] Based on the evidence weight coefficients, the feature vector nodes in the initial graph are incrementally updated to obtain the updated feature vector nodes of the initial graph;

[0042] Based on the changes in the relationships between entities in the updated evidence set, the relation weight values ​​of the corresponding edges in the initial graph are dynamically adjusted to obtain the updated edge weights of the updated evidence set.

[0043] By integrating all the updated feature vector nodes and the updated weights, the initial knowledge graph is reconstructed to obtain the dynamic knowledge graph of the remote medical quality control.

[0044] In a preferred embodiment, the step of performing logical deduction on the abnormal propagation dataset based on the dynamic knowledge graph to trace the potential root causes leading to the abnormal propagation dataset and obtain the attribution hypothesis set for remote medical quality control includes:

[0045] Starting from the abnormal propagation point in the abnormal propagation dataset, the upstream related nodes in the dynamic knowledge graph are traced backward to obtain the potential node candidate set for the remote medical quality control.

[0046] The topological attributes and causal association strength of the nodes in the potential node candidate set are evaluated to obtain the node influence metric of the potential node candidate set;

[0047] Based on the node influence metric, key root cause nodes are selected from the potential node candidate set to obtain the high-influence node set of the dynamic knowledge graph.

[0048] Based on the connection path between the high-influence node set and the anomaly propagation point, construct a multi-level causal inference chain for the anomaly propagation dataset;

[0049] The multi-level causal reasoning chain is subjected to clinical logic consistency verification to eliminate reasoning paths that do not conform to medical common sense, thereby obtaining the attribution hypothesis set for the remote medical quality control.

[0050] In a preferred embodiment, the bidirectional calibration of the causal relationships between the dynamic knowledge graph and the attribution hypothesis set to obtain the attribution analysis report for the remote medical quality control includes:

[0051] Extract the causal paths related to the attribution hypothesis set from the dynamic knowledge graph to obtain the causal path set of the remote medical quality control;

[0052] Extract the key nodes and their relationships in the multi-level causal reasoning chain to obtain the hypothetical causal chain set of the remote medical quality control;

[0053] By comparing the set of causal paths with the set of hypothetical causal chains in both directions, logical conflicts and missing links are identified to obtain the set of differences in the remote medical quality control.

[0054] Based on a pre-set clinical logic rule base, the node features and edge weights in the dynamic knowledge graph are corrected, and the reasoning chain in the attribution hypothesis set is logically optimized to obtain the calibrated causal relationship of the remote medical quality control.

[0055] Based on clinical importance, the calibrated causal relationships are compiled to obtain the attribution analysis report of the telemedicine quality control.

[0056] To address the aforementioned problems, the present invention also provides an automatic attribution analysis system for abnormal data in remote medical quality control, the system comprising:

[0057] The multimodal data aggregation module is used to aggregate the collected patient physiological parameters, medical equipment operating status, and diagnosis and treatment operation records into a multimodal dataset for remote medical quality control.

[0058] An anomaly data identification module is used to set a dynamic threshold for remote medical quality control based on the mean and standard deviation in the historical data of the remote medical quality control, and to identify anomalies in the multimodal dataset that deviate from the dynamic threshold, thereby obtaining the anomaly propagation dataset of the remote medical quality control.

[0059] The knowledge graph construction module is used to vectorize the medical entities and relationships in the preset integrated medical knowledge base to obtain the initial graph of the remote medical quality control.

[0060] The dynamic knowledge graph update module is used to update the node and edge weights in the initial graph in real time using new evidence in the multimodal dataset to obtain the dynamic knowledge graph of the remote medical quality control.

[0061] The attribution hypothesis generation module is used to perform logical deduction on the abnormal propagation dataset based on the dynamic knowledge graph, trace the potential root causes of the abnormal propagation dataset, and obtain the attribution hypothesis set for the remote medical quality control.

[0062] The attribution analysis report output module is used to bidirectionally calibrate the causal relationship between the dynamic knowledge graph and the attribution hypothesis set to obtain the attribution analysis report of the remote medical quality control.

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] 1. This invention can perform format parsing, field mapping, and time synchronization of collected patient physiological parameters, medical equipment operating status, and medical operation records to form a standardized multimodal dataset, ensuring the integrity and consistency of data integration. Simultaneously, it sets dynamic thresholds based on the mean and standard deviation of historical remote medical quality control data, and constructs anomaly propagation chains by combining the temporal sequence of outliers with equipment correlation, effectively improving the accuracy and comprehensiveness of outlier data identification. This provides a reliable data foundation for subsequent attribution analysis and facilitates the efficient capture of outlier data in remote medical quality control.

[0065] 2. This invention constructs an initial knowledge graph by vectorizing and integrating medical entities and relationships in a medical knowledge base. It then uses new evidence from multimodal data to update the graph node features and edge weights in real time, forming a dynamic knowledge graph that ensures the timeliness and accuracy of medical information. Based on this dynamic knowledge graph, it performs logical deduction on anomaly propagation datasets, filters key root cause nodes by combining node influence metrics, and optimizes the reasoning logic by bidirectionally calibrating the causal relationship between the knowledge graph and the attribution hypothesis set. This significantly improves the accuracy of automatic attribution of abnormal data in remote medical quality control, while also enhancing the overall efficiency of attribution analysis, providing efficient and reliable decision support for remote medical quality control. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating an automatic attribution analysis method for abnormal data in remote medical quality control, provided in an embodiment of the present invention.

[0067] Figure 2 This is a functional block diagram of an automatic attribution analysis system for abnormal remote medical quality control data provided in an embodiment of the present invention;

[0068] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0069] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0070] This application provides an automatic attribution analysis method for abnormal data in remote medical quality control. The executing entity of this method includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the automatic attribution analysis method for abnormal data in remote medical quality control can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0071] Reference Figure 1 The diagram shown is a flowchart illustrating an automatic attribution analysis method for abnormal remote medical quality control data according to an embodiment of the present invention. In this embodiment, the automatic attribution analysis method for abnormal remote medical quality control data includes:

[0072] S1. Collect patient physiological parameters, medical equipment operating status, and diagnosis and treatment operation records into a multimodal dataset for remote medical quality control.

[0073] In this embodiment of the invention, the step of compiling the collected patient physiological parameters, medical equipment operating status, and treatment operation records into a multimodal dataset for remote medical quality control includes:

[0074] The format is parsed to obtain the standard data record of the remote medical quality control by parsing the patient's physiological parameters, the operating status of medical equipment and the diagnosis and treatment operation records, and the parsed data is mapped to fields.

[0075] The standard data records are synchronously organized in chronological order to obtain the multimodal data sequence of the remote medical quality control.

[0076] The multimodal data sequences are aggregated into a multimodal dataset for remote medical quality control.

[0077] First, the raw data formats of patient physiological parameters, medical equipment operating status, and treatment operation records are obtained separately. Patient physiological parameters may include text formats such as XML and JSON, as well as device-specific binary formats. Medical equipment operating status is mostly in log format generated by the device system, while treatment operation records exist in the form of electronic medical record text or structured forms. Corresponding parsing methods are used for different data formats. Key fields are directly extracted from text format data, binary format data is converted into readable information using decoding rules provided by the device manufacturer, and form data is extracted item by item. A standard field set for remote medical quality control is preset, covering fixed fields such as patient identification, data type, core parameter values, collection / recording time, and data source. The parsed data fields are precisely matched with the standard field set. For example, "heart rate value" in physiological parameters is mapped to the "physiological parameter - core value" field, and "operating voltage" in device status is mapped to the "device parameter - voltage" field. At the same time, the data format is unified to ensure that each data record has a consistent structure and complete information, resulting in standard data records for remote medical quality control.

[0078] Extract the collection or recording timestamp marked in each standard data record. This timestamp is the precise moment when the data was generated and is the core basis for data synchronization and organization. Based on the timestamp, sort all standard data records in chronological order. If multiple data records of different types correspond to the same timestamp during the sorting process, these data are grouped into a "synchronization data group" and marked with "multi-type synchronization". After sorting, check the temporal continuity of the data sequence. If there are gaps with time intervals exceeding a preset threshold, mark the corresponding position with "data missing" and the missing time range to ensure that the data sequence is clearly traceable in the time dimension and obtain the multimodal data sequence for remote medical quality control.

[0079] The aggregation rules for the multimodal dataset were determined, using "unique patient identifier" as the primary aggregation dimension. All multimodal data sequences of the same patient were grouped into one category, ensuring the independence and integrity of patient data. Under the patient dimension, subsets were divided according to "remote medical service cycle." Each service cycle corresponds to a complete remote medical service process, and the multimodal data sequences within that cycle were fully included in the corresponding subset. Metadata information was added to each aggregated dataset, including dataset number, basic patient information, data collection start and end times, a list of medical equipment involved, and information on personnel involved in the diagnosis and treatment. At the same time, a dataset index was established, containing query keywords such as patient identifier, service cycle, and data type, to facilitate rapid retrieval and access later. Through this hierarchical aggregation and information supplementation, the multimodal data sequences were integrated into a well-structured and comprehensive multimodal dataset for remote medical quality control.

[0080] The beneficial effects include: parsing the collected patient physiological parameters, medical equipment operating status, and treatment operation records; mapping the parsed data to fields; unifying data specifications and structure; and obtaining standard data records for remote medical quality control. The generated standard data records are then synchronously organized according to chronological order, clarifying the temporal relationships between data points to ensure logical consistency, resulting in a multimodal data sequence for remote medical quality control. This multimodal data sequence is then aggregated, integrating data information of different types and dimensions to ultimately form a multimodal dataset for remote medical quality control.

[0081] S2. Based on the mean and standard deviation in the historical data of the remote medical quality control, set the dynamic threshold of the remote medical quality control, and identify the outliers in the multimodal dataset that deviate from the dynamic threshold to obtain the abnormal propagation dataset of the remote medical quality control.

[0082] In this embodiment of the invention, the step of setting a dynamic threshold for remote medical quality control based on the mean and standard deviation of the historical data of the remote medical quality control, and identifying outliers in the multimodal dataset that deviate from the dynamic threshold to obtain the abnormal propagation dataset of the remote medical quality control includes:

[0083] By sampling through a sliding window, the mean and standard deviation of the historical data of the remote medical quality control are statistically analyzed to generate the benchmark parameters of the remote medical quality control.

[0084] Based on the benchmark parameters, the dynamic thresholds of different types of data streams in the multimodal dataset are defined to obtain the threshold boundary range of remote medical quality control;

[0085] Based on the threshold boundary range, the real-time data stream in the multimodal dataset is traversed to identify data points that exceed the threshold boundary range, thereby obtaining the anomaly set of the multimodal dataset;

[0086] Based on the temporal order and device correlation of the data in the anomaly set, an anomaly propagation chain of the anomaly set is constructed;

[0087] Based on the topology of the anomaly propagation chain, causal anomalies in the anomaly set are aggregated to obtain the anomaly propagation dataset for remote medical quality control.

[0088] The sliding window duration is set to a fixed length, determined based on the regular cycle of telemedicine services, ensuring that each window covers a complete short-term data fluctuation cycle. The window moves at fixed time intervals, with each move including data for a new time period while removing older data outside the window's range. For the historical telemedicine quality control data within each sliding window, the central tendency and dispersion of all data points are statistically analyzed according to data type. Central tendency represents the average level of the data, while dispersion represents the prevalence of data deviation from the average level. The central tendency and dispersion values ​​obtained from each window are arranged in chronological order, and the statistical results from the most recent windows are integrated to form a benchmark parameter for telemedicine quality control that reflects the normal fluctuation characteristics of recent data.

[0089] For different types of data streams in the multimodal dataset, corresponding baseline parameters are invoked respectively. For data streams of physiological parameters and equipment operating parameters, the upper limit is determined by expanding upwards to the magnitude corresponding to the dispersion value, and the lower limit is determined by expanding downwards by the same magnitude, thus forming the normal fluctuation range of this type of data stream. For data streams of medical operation records, the judgment range of early, delayed, or incorrectly ordered operations is defined based on the regular operation time intervals and sequences recorded in the baseline parameters. The upper and lower limits or judgment ranges of various data streams are clearly recorded, and the corresponding data stream types and applicable scenarios are labeled to obtain the threshold boundary range of remote medical quality control.

[0090] The system iterates through the real-time data streams in the multimodal dataset by data type. For each real-time data point, it extracts its core values ​​or key information. The extracted information is then compared with the corresponding threshold boundary range. If the values ​​of physiological parameters or equipment parameters are higher than the upper threshold or lower than the lower threshold, they are identified as numerical anomalies. If the time or sequence of the diagnostic and treatment operation records exceeds the judgment range, they are identified as operation anomalies. For each data point identified as anomaly, its data content, data stream type, timestamp of occurrence, and associated patient or equipment identifier are recorded. All anomalies are then organized in chronological order to obtain the anomaly set of the multimodal dataset.

[0091] Arrange all anomalies in the anomaly set in chronological order from earliest to latest timestamp to determine the order in which each anomaly appeared. Analyze the device correlation of the anomalies, identify the medical device or operating object to which each anomaly belongs through data source identification, and check whether different anomalies come from the same device, devices with a connection relationship, or devices associated with the same patient. Connect the anomalies that are consecutive in time and have device correlation in sequence to form a chain structure. The first anomaly to appear is the starting point of the chain, and the anomalies that are subsequently affected by it are the subsequent nodes. For example, if the ventilator pressure anomaly is the starting point, the subsequent patient blood oxygen anomaly is the second node, forming an abnormal propagation path and obtaining the anomaly propagation chain of the anomaly set.

[0092] The topology of the anomaly propagation chain is analyzed to identify pairs of anomalies with a direct causal relationship, where the occurrence of one anomaly directly leads to the occurrence of the next. These anomalies are then grouped into multiple subgroups based on causal relationships. Each subgroup contains a set of anomalies with a continuous causal association, arranged in causal order. Causal relationship descriptions and propagation time interval information are added to each subgroup. All subgroups are then integrated to form a structured dataset that clearly demonstrates how anomalies propagate from their source and trigger chain reactions. This dataset is the anomaly propagation dataset for remote medical quality control.

[0093] The system iterates through the real-time data streams in the multimodal dataset, identifying the source type of each data point and determining whether it belongs to a specific category: patient physiological parameters, medical equipment operating status, or treatment operation records. For patient physiological parameter data, classification features such as parameter name, measurement unit, and numerical type are extracted, for example, "heart rate once per minute - continuous numerical value". For medical equipment operating status data, classification features such as equipment number, operating parameter name, and status identifier are extracted, for example, "ventilator - airway pressure - kPa - continuous numerical value". For treatment operation record data, classification features such as operation type, execution time attribute, and operator identifier are extracted, for example, "intravenous administration - time-based data - medical staff ID". All classification features are recorded in the structure "data ID - source type - classification feature details" to ensure a one-to-one correspondence between features and data, thus obtaining the classification features of the real-time data in the multimodal dataset.

[0094] The preset threshold boundary ranges are divided according to the data source type. Each source type has a unique threshold standard for different classification features. For example, the threshold boundary for "heart rate" in patient physiological parameters is the normal fluctuation range, the threshold boundary for "airway pressure" in medical equipment operation status is the allowable working range of the equipment, and the threshold boundary for "intravenous administration interval" in diagnosis and treatment operation records is the clinical standard time range. The classification features of each extracted real-time data are precisely mapped to the thresholds of the same type and feature in the threshold boundary range. The core information of the data is compared one by one to see if it exceeds the threshold boundary. If the data information is higher than the upper limit of the threshold, lower than the lower limit of the threshold, or does not conform to the logic of the threshold, it is determined to be an abnormal data point. The classification features of the data point, the specific circumstances of exceeding the threshold, the timestamp of occurrence, and the associated objects are recorded. All abnormal data points are summarized to form the preliminary abnormal point set of the multimodal dataset.

[0095] All data points in the initial anomaly set were analyzed to identify redundant data points, mainly categorized into two types: duplicate anomalies and misjudged anomalies. Duplicate anomalies refer to multiple identical records of the same data appearing in the set due to duplicate transmission or misjudgment. Misjudged anomalies refer to isolated data points that exceeded the threshold once due to momentary fluctuations in data transmission or temporary equipment interference, but subsequent continuous data returned to normal and had no other anomaly correlation. For duplicate anomalies, only one record containing complete information was retained, and the remaining identical entries were deleted. For misjudged anomalies, the continuous real-time data before and after the data point was traced to verify whether the anomaly was only a single momentary exceedance without any causal correlation. Once confirmed, the misjudged anomaly was removed from the set. After removing redundant data points, the remaining anomalies were organized in chronological order of data appearance, and an anomaly level label was added to each data point to form a concise, non-redundant, and misjudged multimodal dataset anomaly set.

[0096] The beneficial effects are as follows: historical data for remote medical quality control is extracted using a sliding window sampling method. Statistical calculations are performed on each item of the extracted historical data to obtain the overall mean and standard deviation. These two statistical results are then integrated to form the baseline parameters for remote medical quality control. Based on these baseline parameters, dynamic thresholds are defined for different types of data streams in the multimodal dataset, such as clinical physiological parameters and equipment operating indicators. The numerical fluctuation range of each type of data stream under normal conditions is clarified, ultimately forming the threshold boundary range for remote medical quality control. Using the determined threshold boundary range as a unified judgment standard, each real-time data stream in the multimodal dataset is examined one by one. The specific value of each real-time data stream is compared with the upper and lower limits of the threshold boundary range. All data points whose values ​​exceed the range are selected and aggregated to form the outlier set of the multimodal dataset.

[0097] This study analyzes the generation time of all anomalous data in the anomaly set, arranges the data chronologically, and analyzes the associated medical devices for each anomalous data point. It clarifies the device correlations between different anomalous data points due to device sharing, data transmission, etc. Based on the time sequence and device correlations, the scattered anomalies are linked into a complete chain, constructing an anomaly propagation chain for the anomaly set. The topology of the anomaly propagation chain is analyzed in depth to clarify the connection methods, positional relationships, and interaction logic of each anomaly point in the chain. From this, causally related anomalies are accurately identified—that is, the occurrence of one anomaly point directly or indirectly leads to the generation of another anomaly point. These anomalies with clear causal relationships are then grouped and integrated to obtain the anomaly propagation dataset for remote medical quality control.

[0098] The various data source types of real-time data in the multimodal dataset are clearly defined, including patient physiological parameters, medical equipment operating status, and medical operation records. For each source type of real-time data, classification features that can distinguish its data attributes are extracted one by one, forming the classification features of the real-time data in the multimodal dataset. The extracted classification features of the real-time data are mapped to the determined threshold boundary ranges, with each classification feature matched to a corresponding threshold interval. The real-time data values ​​corresponding to the classification features are compared with the upper and lower limits of the threshold boundary range, and all data points with values ​​exceeding the corresponding threshold intervals are filtered out. These data points are integrated to obtain the preliminary outlier set of the multimodal dataset. All data points in the preliminary outlier set are checked one by one, and redundant data points that occur repeatedly, have no actual correlation, or are caused by data transmission delays are identified and removed to ensure that the remaining data points are all real and valid outliers, thus obtaining the final outlier set of the multimodal dataset.

[0099] S3. Vectorize the medical entities and relationships in the preset integrated medical knowledge base to obtain the initial map of the remote medical quality control;

[0100] In this embodiment of the invention, the vectorization of medical entities and their relationships in the pre-defined integrated medical knowledge base to obtain the initial map of the remote medical quality control includes:

[0101] The medical entity description text in the pre-defined integrated medical knowledge base is parsed, and key medical features are extracted using natural language processing technology to generate the entity feature set of the integrated medical knowledge base.

[0102] Multi-dimensional feature encoding is performed on the medical entities in the entity feature set to convert semantic information into numerical vectors, thereby obtaining the entity vector set of the entity feature set;

[0103] Identify the relationships between medical entities in the integrated medical knowledge base, quantify the relationship type and strength into relationship weight values, and obtain the relationship weight set of the medical entities;

[0104] Based on the entity vector and the relation weight set, an initial graph for remote medical quality control is constructed, with the medical entity as the node and the association relationship as the edge.

[0105] The process involves identifying the relationships between medical entities in the integrated medical knowledge base, quantifying the relationship type and strength into relationship weight values, and obtaining a set of relationship weights for the medical entities, including:

[0106] The relationships between medical entities in the integrated medical knowledge base are analyzed, the relationship types are identified, and the relationship type set of the integrated medical knowledge base is obtained.

[0107] Based on the level of medical evidence, the strength of the relationships between medical entities in the integrated medical knowledge base is assessed to obtain the relationship strength level of the medical entities.

[0108] Based on the set of association types and the relationship strength level, the association relationships of medical entities are quantified into relationship weight values ​​for the medical entities, wherein the formula for calculating the relationship weight value is as follows:

[0109] ;

[0110] In the formula, The relation weight value, This represents the importance coefficient of the type of relationship between entities in the medical entity. This serves as the baseline value for the strength of the association between the medical entities. As a preset structural importance factor, This is a moderating index for the importance of inter-entity association types within the medical entity. This refers to the time decay adjustment coefficient in the integrated medical knowledge base. This refers to the time decay factor in the integrated medical knowledge base;

[0111] The relationship weight values ​​are bound to the association relationships to obtain the relationship weight set of the medical entity.

[0112] The pre-defined integrated medical knowledge base contains structured and unstructured texts such as medical terminology, disease definitions, symptom descriptions, treatment plans, and equipment operation specifications. The medical entity description texts record the entity's attributes, classifications, and related descriptions in detail. These description texts are parsed sentence by sentence, and key medical features that can uniquely identify the entity are extracted by identifying the core nouns, adjectives, and verb phrases in the text. For example, the key features of "myocardial infarction" include "coronary artery occlusion," "retrosternal pain," and "ST segment elevation on electrocardiogram." All key features of all entities are organized into a structure of "entity name - feature category - feature details" to ensure that each feature is directly associated with the entity, forming the entity feature set of the integrated medical knowledge base.

[0113] For each medical entity in the entity feature set, key features are encoded from multiple fixed dimensions such as clinical attributes, pathological features, treatment associations, and equipment associations. For each dimension, a specific numerical value is assigned based on the presence or degree of the feature. For example, in the "location of onset" dimension, if the entity is "cerebral infarction", then "brain" corresponds to a value of 1 and "heart" corresponds to a value of 0. In the "disease course" dimension, "acute" corresponds to a value of 2, "chronic" corresponds to a value of 1, and "subacute" corresponds to a value of 0. The values ​​of the same entity in all dimensions are arranged in a fixed order to form a sequence of values ​​that can reflect the semantic information of the entity. Each sequence is the numerical vector of the entity. The entity vector set of the entity feature set is obtained by summing the vectors of all entities.

[0114] The system iterates through the descriptive statements connecting medical entities in the integrated medical knowledge base, identifying the types of relationships between entities, including causal relationships, accompanying relationships, treatment relationships, and diagnostic relationships. For each type of relationship, the strength of the relationship is determined based on its frequency of mention in the knowledge base and its degree of acceptance in clinical practice; the higher the frequency of mention and the higher the degree of acceptance, the stronger the relationship. The relationship type is represented by a fixed identifier, and the relationship strength is converted into a corresponding weight value. All relationships are recorded in the structure of "Entity A - Relationship Type - Entity B - Weight Value" to obtain the relationship weight set of medical entities.

[0115] Each medical entity in the entity vector set is used as a node in the initial graph. The node attributes contain a numerical vector of the entity, and each value in the vector corresponds to the feature of the node in a specific dimension. According to the association relationships in the relation weight set, connecting edges are drawn between two related entity nodes. The attributes of the edges are labeled with the association type and the corresponding relation weight value. The larger the weight value, the more prominent the visual appearance of the edge. During the drawing process, it is ensured that each node is only connected to nodes with clear association relationships. The nodes are arranged according to entity categories, so that entities of the same type are concentrated in one area, forming a network graph with a clear structure that contains entity semantic vectors and association weight information. This graph is the initial graph of remote medical quality control.

[0116] The process involves traversing all content in the integrated medical knowledge base, including structured medical terminology association tables, unstructured literature descriptions, and guideline recommendations, to identify all relationships between medical entities. For each relationship, the relationship type is defined based on its inherent logic; common types include causal relationships, accompanying relationships, treatment relationships, diagnostic relationships, and subordinate relationships. Each relationship type is given a clear definition and characteristic description. For example, a causal relationship must satisfy the logic that "the existence of the preceding entity will cause the following entity," and an accompanying relationship must satisfy the characteristic that "two entities appear simultaneously in time or space but have no direct causal relationship." All identified relationship types are organized according to the structure of "type name - definition - typical instance" to ensure that there are no duplicate or ambiguous types, resulting in the set of relationship types in the integrated medical knowledge base.

[0117] Referring to the evidence grading system in evidence-based medicine, this system divides medical evidence into multiple levels, including systematic reviews based on a large number of randomized controlled trials, single randomized controlled trials, cohort studies, case-control studies, and case reports. The higher the grade, the stronger the reliability and persuasiveness of the evidence. For each relationship between medical entities in the integrated medical knowledge base, the source of supporting evidence is traced to determine the corresponding evidence grade. For example, if a relationship is confirmed by multiple systematic reviews, the evidence grade is the highest; if it is mentioned only by a single case report, the evidence grade is the lowest. Based on the level of evidence, the strength of the relationship is divided into multiple levels, with the highest evidence grade corresponding to the strongest relationship and the lowest evidence grade corresponding to the weakest relationship. Each strength grade is accompanied by a description of the source of evidence, thus obtaining the relationship strength level of the medical entities.

[0118] For each association type in the association type set, a pre-defined basic weight tendency is established. For example, causal and therapeutic relationships directly influence clinical decisions, so their basic weight tendency is higher than that of accompaniment and subordination relationships. The basic weight tendency is adjusted based on the relationship strength level. The highest association strength results in a higher weight based on the basic tendency, while the lowest association strength results in a lower weight. For example, if a causal relationship corresponds to the highest strength level, the final weight value is high; if an accompaniment relationship corresponds to the lowest strength level, the final weight value is low. By combining the inherent attributes of the association type with the strength of evidence support in this way, each medical entity association is converted into a corresponding relationship weight value. The weight value directly reflects the importance and reliability of the association.

[0119] Each medical entity association is bound to its corresponding association weight value. The binding content includes the names of the two entities involved in the association, the association type, the association strength level, the association weight value, and a summary of the evidence source supporting the association. All binding results are organized according to a fixed structure of "Entity A - Association Type - Entity B - Association Strength Level - Association Weight Value - Evidence Source" to ensure that each record is complete and unique, with no duplicate or missing associations. Finally, a set of association weights for medical entities containing quantitative association information between all medical entities is formed.

[0120] The importance coefficient of the type of relationship between entities is determined by medical experts based on their professional knowledge and clinical experience. In the process of determination, experts will systematically sort out the various relationships between medical entities, and combine their actual value in disease diagnosis and treatment and medical research to assign a clear quantitative value to each type of relationship. This value is the importance coefficient of the type of relationship between entities.

[0121] The baseline value for the strength of association is obtained through statistical analysis of data from an integrated medical knowledge base. Specifically, all associated records of the target medical entity are first extracted from the knowledge base. Core data such as the number of times the entity pairs co-occur and the frequency of associated mentions are statistically analyzed in clinical cases, medical literature, and treatment guidelines. Then, data standardization methods are used to convert these statistical data into values ​​between 0 and 1. The converted data is the baseline value for the strength of association, which reflects the basic strength of the association between entities.

[0122] The preset structural importance factor is a fixed value pre-set by medical informatics experts in combination with the hierarchical structure of the medical knowledge system. When setting it, the experts will focus on the positional attributes of entities in the medical knowledge network, such as the impact of structural differences between core diagnostic entities and auxiliary reference entities on the association relationship. By repeatedly verifying the rational adjustment effect of different values ​​on the weight calculation results, a fixed value is finally determined as the preset structural importance factor.

[0123] The inter-entity association type importance moderating index was determined through retrospective analysis of historical weight assessment data. In the analysis process, a large number of completed medical entity association weight assessment cases and corresponding practical application feedback were collected. For different association types, the impact of multiple candidate indices on the weight calculation results was tested. The values ​​that can make the calculation results highly match actual clinical needs and medical cognition were selected and determined as the inter-entity association type importance moderating index.

[0124] The time decay adjustment coefficient is derived from historical data mining of the integrated medical knowledge base. During the mining process, the medical entity relationships recorded at different time points in the knowledge base are first extracted. The changing patterns of the importance of these relationships in clinical applications over time are analyzed. The average magnitude of the decay of the importance of the relationships over time is statistically analyzed. Based on the correspondence between the decay magnitude and the weight adjustment requirements, the specific value of the time decay adjustment coefficient is calculated and determined.

[0125] The time decay factor is calculated based on the time attribute of the association. The calculation first obtains the time interval between the first recorded time of the medical entity association and the current calculation time. A linear decay rule is used to substitute the time interval into the preset decay formula and convert the interval time into a value between 0 and 1. When the time interval is 0, the value is 1. The longer the interval time, the smaller the value. The final value is the time decay factor, which is used to quantify the degree of time decay of the association.

[0126] The relation weight value is a comprehensive quantitative result of the current importance of the relationship between medical entities. Its value directly corresponds to the applicability of the relationship in the current medical scenario.

[0127] The inter-entity relationship type importance coefficient is calculated by raising the inter-entity relationship type importance adjustment index to a power. This is used to precisely control the influence of relationship type importance on the final weight. When the adjustment index is greater than 1, it will amplify the effect of relationship type importance, making the weight of important relationship types more significantly increased. When the adjustment index is less than 1, it will reduce its effect, avoiding excessive influence of a single relationship type on the weight result.

[0128] The baseline value of the association strength is added to a preset structural importance factor. This incorporates the structural characteristics of the medical knowledge system into the basic association strength between entities, ensuring that the weight calculation not only reflects the natural strength of the association but also the positional value of the entity within the knowledge structure. Structurally more important entity associations receive a reasonable weight increase through this calculation. Subtracting the time decay factor from 1, multiplying it by the time decay adjustment coefficient, and then adding 1 constitutes the time decay correction term for the weight calculation. This correction term accurately quantifies the change in the importance of the association over time. When the association record is old, the time decay factor is small, and the correction term is large, which reduces the final association weight value, aligning with the practical need in the medical field where the latest association information is more valuable.

[0129] The numerator integrates the importance of association type, the strength of basic association, and the influence of structural characteristics to form a comprehensive value that reflects the core value of the association. The denominator reflects the interference effect of time factors through a time decay correction term. The relationship weight value obtained by dividing the two achieves a comprehensive and accurate quantification of the current importance of the association between medical entities, providing a reliable basis for medical decision-making.

[0130] The beneficial effects are as follows: The descriptive text of all medical entities in the pre-defined integrated medical knowledge base is analyzed sentence by sentence. Natural language processing technology is used to semantically decompose the text content and extract core information, accurately extracting key medical features such as disease names, drug components, symptoms, and treatment procedures. These features are then categorized and organized according to entity type, generating an entity feature set for the integrated medical knowledge base. From multiple dimensions such as the functional attributes, clinical application scenarios, and types of associated objects, each medical entity in the entity feature set is feature-encoded, transforming the originally abstract semantic information into calculable and comparable numerical vectors. Each medical entity corresponds to a unique numerical vector, and all vectors are integrated to form an entity vector set for the entity feature set. The interactions and relationships between all medical entities in the integrated medical knowledge base are comprehensively analyzed, clarifying the specific types of relationships between entities, such as causal relationships between causes and symptoms, compatibility relationships between drugs and indications, and auxiliary relationships between examination items and disease diagnosis. Simultaneously, the strength of the association is determined based on the frequency of occurrence in clinical practice and the degree of medical consensus acceptance. The association type and strength are uniformly converted into specific relationship weight values, summarizing to form a relationship weight set for medical entities. Using the medical entities corresponding to the numerical vectors in the entity vector set as nodes of the graph, and the associations corresponding to the relation weight values ​​in the relation weight set as edges connecting the nodes, the vector information of each medical entity is bound to the corresponding association weight value. The connection structure of nodes and edges is built according to the real association logic between entities, and finally the initial graph of remote medical quality control is constructed.

[0131] This study comprehensively analyzes the relationships between all medical entities in an integrated medical knowledge base, clarifying the specific types of each relationship, such as causal relationships between diseases and symptoms, treatment relationships between drugs and diseases, and auxiliary relationships between examination items and diagnoses. These different types of relationships are categorized and organized to form a set of relationship types for the integrated medical knowledge base. Based on the medical evidence level standard, which covers different levels of evidence including results from randomized controlled trials, recommendations from authoritative medical guidelines, and expert consensus opinions, the strength of the relationships between medical entities in the integrated medical knowledge base is assessed. This determines the reliability and scope of influence of each relationship in clinical practice, classifying corresponding relationship strength levels. Combining the determined set of relationship types and relationship strength levels, specific quantitative values ​​are assigned to each medical entity relationship. The importance of the relationship type and the strength level jointly determine the numerical value; the more critical the type and the higher the strength level, the larger the corresponding numerical value, thus obtaining the relationship weight value of the medical entity. Each medical entity relationship is bound to its corresponding relationship weight value, ensuring that each relationship has a clear weight identifier. All bound relationships and weight values ​​are summarized and integrated to ultimately form a set of relationship weights for medical entities.

[0132] S4. Use the new evidence in the multimodal dataset to update the node and edge weights in the initial graph in real time to obtain the dynamic knowledge graph of the remote medical quality control.

[0133] In this embodiment of the invention, the step of using new evidence from the multimodal dataset to update the node and edge weights in the initial graph in real time to obtain the dynamic knowledge graph of the remote medical quality control includes:

[0134] Extract new evidence data from the multimodal dataset that is associated with medical entities in the initial atlas, and generate an updated evidence set for the multimodal dataset;

[0135] The credibility and timeliness of the new evidence in the updated evidence set are evaluated to obtain the evidence weight coefficient of the updated evidence set;

[0136] Based on the evidence weight coefficients, the feature vector nodes in the initial graph are incrementally updated to obtain the updated feature vector nodes of the initial graph;

[0137] Based on the changes in the relationships between entities in the updated evidence set, the relation weight values ​​of the corresponding edges in the initial graph are dynamically adjusted to obtain the updated edge weights of the updated evidence set.

[0138] By integrating all the updated feature vector nodes and the updated weights, the initial knowledge graph is reconstructed to obtain the dynamic knowledge graph of the remote medical quality control.

[0139] The multimodal dataset is traversed, and new evidence data directly related to these entities is selected by comparing them with the existing list of medical entities in the initial atlas. For example, the "tachycardia" entity in the atlas related to "heart rate consistently higher than 120 beats / minute" in patient physiological parameters, the "ventilator malfunction" entity in the atlas related to "ventilator pressure fluctuations" in medical equipment logs, and the "postoperative infection prevention" entity in the atlas related to "insufficient duration of postoperative antibiotic use" in treatment operation records. The selected new evidence data is then extracted, retaining core information, data generation time, and source identifier, and the names of the entities in the initial atlas that are associated with it are labeled. The data is then organized according to the structure of "associated entity - new evidence content - source - time" to form the updated evidence set of the multimodal dataset.

[0140] The system uses a pre-defined standard to assess the credibility of evidence, judging it from three dimensions: the authority of the data source, the completeness of the data, and the degree of clinical consensus. Each dimension is assigned a fixed score according to its level. The timeliness indicator is based on the interval between the data generation time and the current time. The shorter the interval, the higher the timeliness and the higher the corresponding score. The credibility score and the timeliness score are added together in a 1:1 ratio to obtain the comprehensive score of each new piece of evidence. The comprehensive score is then converted into an evidence weight coefficient between 0 and 1. The higher the score, the closer the coefficient is to 1, thus obtaining the evidence weight coefficient of the updated evidence set.

[0141] For each new piece of evidence associated with the initial graph entity in the updated evidence set, the corresponding feature vector node is located. It is then analyzed whether the new evidence has supplemented or corrected the entity's feature dimensions. For example, if the new evidence mentions "a new feature of younger onset of a certain disease," then the corresponding value is added to the "age distribution of onset" dimension of the disease entity vector. If the new evidence corrects the original features, then the value of the corresponding dimension is adjusted. The feature vector is adjusted in conjunction with the evidence weight coefficient. The closer the coefficient is to 1, the greater the adjustment of the feature dimension value. If the coefficient is close to 0, only a slight correction is made or the record is retained without adjustment. For entity nodes not associated with new evidence, their original feature vectors are kept unchanged. Finally, the updated feature vector nodes of the initial graph are obtained.

[0142] The updated evidence set is analyzed to identify changes in the relationships between medical entities. If new evidence strengthens existing relationships, the relationship weights of the corresponding edges in the initial graph are found and increased proportionally according to the evidence weight coefficient. If new evidence weakens existing relationships, the weights are decreased proportionally according to the coefficient. If new evidence discovers new entity relationships, new edges are added between the corresponding entities, and the initial relationship weights are set based on the evidence weight coefficient. For edges with no changes in relationships, the original weights are kept unchanged, resulting in the updated edge weights of the updated evidence set.

[0143] All updated feature vector nodes are integrated with their updated edge weights. Initial graph nodes and edge weights not associated with new evidence are retained as is to ensure no graph information is lost. The updated feature vector nodes are used as the new node base, and the graph is reorganized and arranged according to entity categories to ensure that similar entities are concentrated in one area. The edges corresponding to the updated edge weights are used to connect related entities, with the line thickness varying according to the weight value, the larger the weight value, the thicker the line. The graph is supplemented with annotation information for each updated content to ensure that the update trajectory is traceable. Through this reconstruction method of retaining the core and updating incrementally, a dynamic knowledge graph for remote medical quality control that reflects the latest medical evidence in real time is formed.

[0144] The beneficial effects are as follows: A comprehensive search of the multimodal dataset is conducted to identify all new data that is directly or indirectly related to medical entities in the initial atlas. This new data includes newly added clinical records, equipment operation monitoring results, and changes in patient physiological parameters—all new evidence that can supplement or correct the atlas. This data is categorized and organized according to entity association dimensions to generate an updated evidence set for the multimodal dataset. The source of each piece of new evidence in the updated evidence set is verified. Credibility is assessed by determining whether the source is an authoritative medical institution and whether the data collection process is standardized. The timeliness is also assessed by checking the interval between the generation time of the new evidence and the current quality control time. Based on the evaluation results of these two indicators, each piece of new evidence is assigned a corresponding quantitative value, resulting in an evidence weight coefficient for the updated evidence set. Using the calculated evidence weight coefficient as a basis, the feature vector nodes in the initial atlas are incrementally updated. For entities corresponding to new evidence with high weight coefficients, the numerical distribution of their feature vectors is adjusted, retaining the original effective feature information while incorporating the feature changes brought about by the new evidence. Nodes without new evidence support are not repeatedly updated, resulting in updated feature vector nodes for the initial atlas.

[0145] The changes in the relationships between medical entities in the updated evidence set are analyzed one by one, including new relationships, strengthening or weakening of existing relationships, and termination of relationships. Based on the evidence weight coefficients corresponding to these changes, the relationship weight values ​​corresponding to the edges connecting related entities in the initial graph are adjusted accordingly, strengthening or weakening the edge weight values ​​to obtain the updated edge weights of the updated evidence set. All updated feature vector nodes and adjusted updated edge weights are systematically integrated, and the connection structure of nodes and edges is reconstructed according to the real relationship logic between medical entities. The initial knowledge graph is reconstructed as a whole, filling information gaps in the original graph, correcting biased relationships, and finally obtaining a dynamic knowledge graph for remote medical quality control.

[0146] S5. Based on the dynamic knowledge graph, perform logical deduction on the abnormal propagation dataset to trace the potential root causes of the abnormal propagation dataset and obtain the attribution hypothesis set for the remote medical quality control.

[0147] In this embodiment of the invention, the step of performing logical deduction on the abnormal propagation dataset based on the dynamic knowledge graph to trace the potential root causes leading to the abnormal propagation dataset and obtain the attribution hypothesis set for remote medical quality control includes:

[0148] Starting from the abnormal propagation point in the abnormal propagation dataset, the upstream related nodes in the dynamic knowledge graph are traced backward to obtain the potential node candidate set for the remote medical quality control.

[0149] The topological attributes and causal association strength of the nodes in the potential node candidate set are evaluated to obtain the node influence metric of the potential node candidate set;

[0150] Based on the node influence metric, key root cause nodes are selected from the potential node candidate set to obtain the high-influence node set of the dynamic knowledge graph.

[0151] Based on the connection path between the high-influence node set and the anomaly propagation point, construct a multi-level causal inference chain for the anomaly propagation dataset;

[0152] The multi-level causal reasoning chain is subjected to clinical logic consistency verification to eliminate reasoning paths that do not conform to medical common sense, thereby obtaining the attribution hypothesis set for the remote medical quality control.

[0153] Extract all anomalous propagation points from the anomalous propagation dataset and identify the corresponding entity node in the dynamic knowledge graph for each anomalous propagation point. Starting from these nodes, trace upstream related nodes in the dynamic knowledge graph in the reverse direction of the association relationship. That is, if there is a causal relationship between node A and node B that "node B is caused by node A", then node A is the upstream node of node B. If there is a device association between node C and node D that "node D is controlled by node C", then node C is the upstream node of node D. Continue tracing to the edge node of the graph, collect all upstream nodes passed through, and organize them according to the structure of "anomalous propagation point - upstream node level - node name" to obtain a potential node candidate set for remote medical quality control.

[0154] The topological attributes of each node in the potential node candidate set are analyzed, including the number of connections the node has in the dynamic knowledge graph, whether it is at the intersection of multiple associated paths, and the path distance to the anomaly propagation point. At the same time, the causal relationship strength between the node and the anomaly propagation point is evaluated. Referring to the relationship weight description of the edges in the dynamic knowledge graph, the stronger and more direct the relationship, the higher the causal relationship strength. Combining the topological attributes and the causal relationship strength, the influence capacity of each node is qualitatively described. For example, "many connections, being a hub node and having a direct and strong association with the anomaly point" corresponds to high influence, while "few connections, not being a hub node and having an indirect and weak association with the anomaly point" corresponds to low influence. The node influence metric value of the potential node candidate set is obtained.

[0155] Based on the node influence metric, nodes with high influence metrics are retained. These nodes are closely associated with the anomaly propagation points in the dynamic knowledge graph and occupy key positions in the network structure, potentially playing a decisive role in the generation and propagation of anomalies. Nodes with medium to low influence metrics are removed, as these nodes have little or only an indirect impact on anomaly propagation. For the retained high-influence nodes, details of their association relationships in the dynamic knowledge graph are added and organized according to the structure of "node name - association details - influence metric" to obtain the set of high-influence nodes in the dynamic knowledge graph.

[0156] For each node in the high-influence node set, all connected paths between it and the anomaly propagation point are found in the dynamic knowledge graph. The paths must follow the direction of the association relationship. According to the hierarchical relationship of the nodes in the path, the path is decomposed into a multi-level structure, and the node name and the association type with the nodes above and below are labeled in each level. For example, the high-influence node "sensor failure" is associated with the anomaly propagation point "physiological parameter anomaly" through the intermediate node "equipment parameter distortion", forming a two-level causal inference chain of "sensor failure - equipment parameter distortion - physiological parameter anomaly". All multi-level paths corresponding to high-influence nodes are summarized to obtain the multi-level causal inference chain of the anomaly propagation dataset.

[0157] The multi-level causal reasoning chain is compared with authoritative medical guidelines, clinical diagnosis and treatment standards, and equipment operating principles to verify the logical consistency of each reasoning chain; all reasoning chains that conform to clinical logic are retained, and each reasoning chain is used as a hypothesis of a potential root cause. After summarizing, the attribution hypothesis set for remote medical quality control is obtained.

[0158] The beneficial effect is that, using each anomalous propagation point in the anomalous propagation dataset as the starting point for reverse tracing, the system searches upstream layer by layer along the relationships between entities in the dynamic knowledge graph, covering all upstream entities that may have a causal relationship with the anomalous propagation point. These nodes are then aggregated to obtain a potential candidate set of nodes for remote medical quality control. The topological attributes of each node in the potential candidate set are analyzed, including the node's connection density and network position in the dynamic knowledge graph. Simultaneously, the causal relationship strength between each node and the anomalous propagation point is calculated. These two evaluation results are combined and quantified to obtain a node influence metric for the potential candidate set. Using this node influence metric as a selection criterion, the metric values ​​of all nodes in the potential candidate set are compared, prioritizing the retention of nodes with higher metric values ​​and more significant impact on the anomalous propagation point. These nodes are the key root cause nodes leading to the anomalies, and the resulting set of high-influence nodes in the dynamic knowledge graph is obtained.

[0159] This study analyzes all connectivity paths between each node in the high-influence node cluster and the point of anomalous propagation. These paths are then organized and integrated according to their hierarchical relationship and causal transmission order, constructing a multi-level, progressive causal reasoning chain from the root cause node to the anomalous propagation point, clearly illustrating the transmission process of the anomalous event. Referring to clinical treatment guidelines, medical common sense, and relevant medical guidelines, each multi-level causal reasoning chain is validated, checking whether each reasoning path conforms to clinical logic. Reasoning paths that contradict medical common sense or lack clinical rationality are eliminated, while logically sound and clinically consistent reasoning results are retained, ultimately yielding the attribution hypothesis set for remote medical quality control.

[0160] S6. Perform bidirectional calibration of the causal relationship between the dynamic knowledge graph and the attribution hypothesis set to obtain the attribution analysis report of the remote medical quality control.

[0161] In this embodiment of the invention, the bidirectional calibration of the causal relationship between the dynamic knowledge graph and the attribution hypothesis set to obtain the attribution analysis report of the remote medical quality control includes:

[0162] Extract the causal paths related to the attribution hypothesis set from the dynamic knowledge graph to obtain the causal path set of the remote medical quality control;

[0163] Extract the key nodes and their relationships in the multi-level causal reasoning chain to obtain the hypothetical causal chain set of the remote medical quality control;

[0164] By comparing the set of causal paths with the set of hypothetical causal chains in both directions, logical conflicts and missing links are identified to obtain the set of differences in the remote medical quality control.

[0165] Based on a pre-set clinical logic rule base, the node features and edge weights in the dynamic knowledge graph are corrected, and the reasoning chain in the attribution hypothesis set is logically optimized to obtain the calibrated causal relationship of the remote medical quality control.

[0166] Based on clinical importance, the calibrated causal relationships are compiled to obtain the attribution analysis report of the telemedicine quality control.

[0167] Extract the entity nodes and causal relationship types involved in each hypothesis from the attribution hypothesis set. Use these as search criteria to traverse the dynamic knowledge graph and locate all causal paths that contain these entity nodes and match the relationship types. These paths must completely cover the association chain from the potential root cause node to the abnormal propagation point, including all intermediate nodes in the path and the association descriptions between nodes. Organize the extracted paths into a structure of "starting node - intermediate node - ending node - association type sequence" to ensure that each path can correspond to a certain hypothesis in the attribution hypothesis set, thus obtaining the causal path set for remote medical quality control.

[0168] By reviewing the multi-level causal reasoning chain, we identify the key nodes in each chain that play a core role in causal transmission. These nodes are usually turning points in abnormal propagation or entities directly related to clinical decisions. At the same time, we extract the relationships between key nodes and clarify the direction and nature of the relationships. We summarize this information according to the structure of "key node sequence - relationship sequence - reasoning chain level" to ensure that it is logically consistent with the original multi-level causal reasoning chain, thus obtaining the hypothetical causal chain set for remote medical quality control.

[0169] Each path in the causal path set is compared node-by-node and relationship-by-relationship with the corresponding chain in the hypothetical causal chain set to check whether the node names match completely and whether the types and directions of the associations are consistent. At the same time, the integrity of the path and chain is checked. If one side contains a certain node or relationship while the other side is missing, it is determined to be a missing link. All logical conflicts and missing links are recorded, and the specific location and content of the conflict or missing link are marked to obtain the set of differences in remote medical quality control.

[0170] The pre-defined clinical logic rule base contains standardized causal logic based on authoritative medical textbooks, treatment guidelines, and equipment operation manuals, such as "abnormal equipment parameters should precede abnormal physiological parameters" and "incorrect drug dosage will directly affect efficacy or cause adverse reactions." Based on this rule base, for logical conflicts in the set of discrepancies, the feature descriptions of corresponding nodes and the weight relationships of edges in the dynamic knowledge graph are modified. For missing links, intermediate nodes or relationships conforming to the rule base are added to the inference chain of the attribution hypothesis set. Through the above modifications and optimizations, the causal relationship between the dynamic knowledge graph and the attribution hypothesis set is logically completely consistent, resulting in a calibrated causal relationship for remote medical quality control.

[0171] The calibrated causal relationships are ranked according to clinical importance. The criteria for determining importance include: whether it directly threatens the patient's life, the frequency of occurrence in the spread of abnormalities, and whether it affects the core diagnosis and treatment process. Based on the ranking results, the starting cause, propagation path, involved entities, correlations, and clinical evidence of each causal relationship are organized into structured entries. The entries must contain a complete chain of "root cause - intermediate process - final abnormality" and indicate key information added during the correction process. After all entries are summarized, a document with standardized language, clear logic, and highlighted key points is formed. This document is the attribution analysis report for telemedicine quality control.

[0172] The beneficial effects are as follows: A comprehensive search of the dynamic knowledge graph is conducted to locate all causal paths associated with the reasoning paths in the attribution hypothesis set. These paths must encompass the logical connections between the root cause nodes, intermediate transmission nodes, and anomaly propagation points involved in the hypothesis set. All causal paths meeting the criteria are summarized and organized to obtain the causal path set for remote medical quality control. From the multi-level causal reasoning chain, nodes playing a key role in anomaly attribution are selected, including the root cause node that triggers the anomaly, the intermediate nodes that transmit the anomaly, and the final anomaly propagation point. Simultaneously, the relationships and corresponding weight information between these key nodes are extracted and integrated according to the hierarchical order of the reasoning chain to obtain the hypothesis causal chain set for remote medical quality control. A bidirectional cross-comparison is performed between the causal path set and the hypothesis causal chain set. On the one hand, it verifies whether the paths in the causal path set are logically consistent with the hypothesis causal chain set; on the other hand, it verifies whether the reasoning in the hypothesis causal chain set can be supported by the causal path set. Logical contradictions, conflicting relationships, and missing necessary links are identified, and these are summarized to form a set of discrepancies in remote medical quality control.

[0173] Based on a pre-defined clinical logic rule base, which includes clinical treatment guidelines and medical causal relationship criteria, this study addresses issues with concentrated discrepancies by modifying the feature information of nodes and the weight values ​​of edges in the dynamic knowledge graph to better align with clinical practice. Simultaneously, it logically adjusts the reasoning chains within the attribution hypothesis set, supplementing missing links and correcting conflicting logic to obtain calibrated causal relationships for remote medical quality control. Following the principle of prioritizing clinical importance, key causal relationships affecting patient safety and treatment outcomes are highlighted first, followed by secondary related content. The calibrated causal relationships are then organized and compiled according to a standardized report structure, clarifying the root cause of the abnormality, its propagation path, its scope of impact, and clinical recommendations, ultimately yielding an attribution analysis report for remote medical quality control.

[0174] like Figure 2 The diagram shown is a functional block diagram of an automatic attribution analysis system for abnormal remote medical quality control data provided in an embodiment of the present invention.

[0175] The remote medical quality control anomaly data automatic attribution analysis system 100 described in this invention can be installed in an electronic device. Depending on the functions implemented, the remote medical quality control anomaly data automatic attribution analysis system 100 may include a multimodal data collection module 101, an anomaly data identification module 102, a knowledge graph construction module 103, a dynamic knowledge graph update module 104, an attribution hypothesis generation module 105, and an attribution analysis report output module 106. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0176] In this embodiment, the functions of each module / unit are as follows:

[0177] The multimodal data collection module 101 is used to collect the collected patient physiological parameters, medical equipment operating status and diagnosis and treatment operation records into a multimodal dataset for remote medical quality control.

[0178] The abnormal data identification module 102 is used to set a dynamic threshold for remote medical quality control based on the mean and standard deviation in the historical data of remote medical quality control, and to identify abnormal points in the multimodal dataset that deviate from the dynamic threshold, so as to obtain the abnormal propagation dataset of remote medical quality control.

[0179] The knowledge graph construction module 103 is used to vectorize the medical entities and relationships in the preset integrated medical knowledge base to obtain the initial graph of the remote medical quality control.

[0180] The dynamic knowledge graph update module 104 is used to update the node and edge weights in the initial graph in real time using new evidence in the multimodal dataset to obtain the dynamic knowledge graph of the remote medical quality control.

[0181] The attribution hypothesis generation module 105 is used to perform logical deduction on the abnormal propagation dataset based on the dynamic knowledge graph, trace the potential root causes of the abnormal propagation dataset, and obtain the attribution hypothesis set of the remote medical quality control.

[0182] The attribution analysis report output module 106 is used to bidirectionally calibrate the causal relationship between the dynamic knowledge graph and the attribution hypothesis set to obtain the attribution analysis report of the remote medical quality control.

[0183] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0184] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0185] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0186] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0187] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for automatic attribution analysis of abnormal data in remote medical quality control, characterized in that, The method includes: S1. Collect patient physiological parameters, medical equipment operating status, and diagnosis and treatment operation records into a multimodal dataset for remote medical quality control. S2. Based on the mean and standard deviation in the historical data of the remote medical quality control, set the dynamic threshold of the remote medical quality control, and identify the outliers in the multimodal dataset that deviate from the dynamic threshold to obtain the abnormal propagation dataset of the remote medical quality control. S3. Vectorize the medical entities and relationships in the preset integrated medical knowledge base to obtain the initial map of the remote medical quality control; S4. Use the new evidence in the multimodal dataset to update the node and edge weights in the initial graph in real time to obtain the dynamic knowledge graph of the remote medical quality control. S5. Based on the dynamic knowledge graph, perform logical deduction on the abnormal propagation dataset to trace the potential root causes of the abnormal propagation dataset and obtain the attribution hypothesis set for the remote medical quality control. S6. Perform bidirectional calibration of the causal relationship between the dynamic knowledge graph and the attribution hypothesis set to obtain the attribution analysis report of the remote medical quality control.

2. The method for automatic attribution analysis of abnormal data in remote medical quality control as described in claim 1, characterized in that, The process of compiling collected patient physiological parameters, medical equipment operating status, and treatment operation records into a multimodal dataset for remote medical quality control includes: The format is parsed to obtain the standard data record of the remote medical quality control by parsing the patient's physiological parameters, the operating status of medical equipment and the diagnosis and treatment operation records, and the parsed data is mapped to fields. The standard data records are synchronously organized in chronological order to obtain the multimodal data sequence of the remote medical quality control. The multimodal data sequences are aggregated into a multimodal dataset for remote medical quality control.

3. The method for automatic attribution analysis of abnormal data in remote medical quality control as described in claim 1, characterized in that, The process involves setting a dynamic threshold for remote medical quality control based on the mean and standard deviation of historical data, and identifying outliers in the multimodal dataset that deviate from the dynamic threshold to obtain an anomaly propagation dataset for remote medical quality control, including: By sampling through a sliding window, the mean and standard deviation of the historical data of the remote medical quality control are statistically analyzed to generate the benchmark parameters of the remote medical quality control. Based on the benchmark parameters, the dynamic thresholds of different types of data streams in the multimodal dataset are defined to obtain the threshold boundary range of remote medical quality control; Based on the threshold boundary range, the real-time data stream in the multimodal dataset is traversed to identify data points that exceed the threshold boundary range, thereby obtaining the anomaly set of the multimodal dataset; Based on the temporal order and device correlation of the data in the anomaly set, an anomaly propagation chain of the anomaly set is constructed; Based on the topology of the anomaly propagation chain, causal anomalies in the anomaly set are aggregated to obtain the anomaly propagation dataset for remote medical quality control.

4. The method for automatic attribution analysis of abnormal data in remote medical quality control as described in claim 3, characterized in that, The step of traversing the real-time data stream in the multimodal dataset based on the threshold boundary range, identifying data points exceeding the threshold boundary range, and obtaining the outlier set of the multimodal dataset includes: Based on the data source type, extract the classification features of the real-time data in the multimodal dataset; The classification features are mapped to the threshold boundary range to filter out the preliminary set of outliers in the multimodal dataset; Redundant data points are removed from the initial set of outliers to obtain the set of outliers in the multimodal dataset.

5. The method for automatic attribution analysis of abnormal data in remote medical quality control as described in claim 1, characterized in that, The vectorized preset integrated medical knowledge base contains medical entities and their relationships, resulting in the initial map of the remote medical quality control, including: The medical entity description text in the pre-defined integrated medical knowledge base is parsed, and key medical features are extracted using natural language processing technology to generate the entity feature set of the integrated medical knowledge base. Multi-dimensional feature encoding is performed on the medical entities in the entity feature set to convert semantic information into numerical vectors, thereby obtaining the entity vector set of the entity feature set; Identify the relationships between medical entities in the integrated medical knowledge base, quantify the relationship type and strength into relationship weight values, and obtain the relationship weight set of the medical entities; Based on the entity vector and the relation weight set, an initial graph for remote medical quality control is constructed, with the medical entity as the node and the association relationship as the edge.

6. The method for automatic attribution analysis of abnormal data in remote medical quality control as described in claim 5, characterized in that, The process involves identifying the relationships between medical entities in the integrated medical knowledge base, quantifying the relationship type and strength into relationship weight values, and obtaining a set of relationship weights for the medical entities, including: The relationships between medical entities in the integrated medical knowledge base are analyzed, the relationship types are identified, and the relationship type set of the integrated medical knowledge base is obtained. Based on the level of medical evidence, the strength of the relationships between medical entities in the integrated medical knowledge base is assessed to obtain the relationship strength level of the medical entities. Based on the set of association types and the relationship strength level, the association relationships of medical entities are quantified into relationship weight values ​​for the medical entities, wherein the formula for calculating the relationship weight value is as follows: ； In the formula, The relation weight value, This represents the importance coefficient of the type of relationship between entities in the medical entity. This serves as the baseline value for the strength of the association between the medical entities. As a preset structural importance factor, This is a moderating index for the importance of inter-entity association types within the medical entity. This refers to the time decay adjustment coefficient in the integrated medical knowledge base. This refers to the time decay factor in the integrated medical knowledge base; The relationship weight values ​​are bound to the association relationships to obtain the relationship weight set of the medical entity.

7. The method for automatic attribution analysis of abnormal data in remote medical quality control as described in claim 1, characterized in that, The step of updating the node and edge weights in the initial graph in real time using new evidence from the multimodal dataset to obtain the dynamic knowledge graph of the remote medical quality control includes: Extract new evidence data from the multimodal dataset that is associated with medical entities in the initial atlas, and generate an updated evidence set for the multimodal dataset; The credibility and timeliness of the new evidence in the updated evidence set are evaluated to obtain the evidence weight coefficient of the updated evidence set; Based on the evidence weight coefficients, the feature vector nodes in the initial graph are incrementally updated to obtain the updated feature vector nodes of the initial graph; Based on the changes in the relationships between entities in the updated evidence set, the relation weight values ​​of the corresponding edges in the initial graph are dynamically adjusted to obtain the updated edge weights of the updated evidence set. By integrating all the updated feature vector nodes and the updated weights, the initial graph is reconstructed to obtain the dynamic knowledge graph of the remote medical quality control.

8. The method for automatic attribution analysis of abnormal data in remote medical quality control as described in claim 1, characterized in that, The step of performing logical deduction on the abnormal propagation dataset based on the dynamic knowledge graph to trace the potential root causes leading to the abnormal propagation dataset and obtain the attribution hypothesis set for remote medical quality control includes: Starting from the abnormal propagation point in the abnormal propagation dataset, the upstream related nodes in the dynamic knowledge graph are traced backward to obtain the potential node candidate set for the remote medical quality control. The topological attributes and causal association strength of the nodes in the potential node candidate set are evaluated to obtain the node influence metric of the potential node candidate set; Based on the node influence metric, key root cause nodes are selected from the potential node candidate set to obtain the high-influence node set of the dynamic knowledge graph. Based on the connection path between the high-influence node set and the anomaly propagation point, construct a multi-level causal inference chain for the anomaly propagation dataset; The multi-level causal reasoning chain is subjected to clinical logic consistency verification to eliminate reasoning paths that do not conform to medical common sense, thereby obtaining the attribution hypothesis set for the remote medical quality control.

9. The method for automatic attribution analysis of abnormal data in remote medical quality control as described in claim 8, characterized in that, The bidirectional calibration of the causal relationships between the dynamic knowledge graph and the attribution hypothesis set yields the attribution analysis report for the remote medical quality control, including: Extract the causal paths related to the attribution hypothesis set from the dynamic knowledge graph to obtain the causal path set of the remote medical quality control; Extract the key nodes and their relationships in the multi-level causal reasoning chain to obtain the hypothetical causal chain set of the remote medical quality control; By comparing the set of causal paths with the set of hypothetical causal chains in both directions, logical conflicts and missing links are identified to obtain the set of differences in the remote medical quality control. Based on a pre-set clinical logic rule base, the node features and edge weights in the dynamic knowledge graph are corrected, and the reasoning chain in the attribution hypothesis set is logically optimized to obtain the calibrated causal relationship of the remote medical quality control. Based on clinical importance, the calibrated causal relationships are compiled to obtain the attribution analysis report of the telemedicine quality control.

10. An automatic attribution analysis system for abnormal data in remote medical quality control, characterized in that, The system includes: The multimodal data aggregation module is used to aggregate the collected patient physiological parameters, medical equipment operating status, and diagnosis and treatment operation records into a multimodal dataset for remote medical quality control. An anomaly data identification module is used to set a dynamic threshold for remote medical quality control based on the mean and standard deviation in the historical data of the remote medical quality control, and to identify anomalies in the multimodal dataset that deviate from the dynamic threshold, thereby obtaining an anomaly propagation dataset for the remote medical quality control. The knowledge graph construction module is used to vectorize the medical entities and relationships in the preset integrated medical knowledge base to obtain the initial graph of the remote medical quality control. The dynamic knowledge graph update module is used to update the node and edge weights in the initial graph in real time using new evidence in the multimodal dataset to obtain the dynamic knowledge graph of the remote medical quality control. The attribution hypothesis generation module is used to perform logical deduction on the abnormal propagation dataset based on the dynamic knowledge graph, trace the potential root causes of the abnormal propagation dataset, and obtain the attribution hypothesis set for the remote medical quality control. The attribution analysis report output module is used to bidirectionally calibrate the causal relationship between the dynamic knowledge graph and the attribution hypothesis set to obtain the attribution analysis report of the remote medical quality control.

Images