An artificial intelligence-based patient profiling method

Abstract

The application discloses a patient portrait construction method based on artificial intelligence, and particularly relates to the field of medical data processing, and is used for solving the problem of conflict medical evidence affecting the reliability of the portrait; through acquiring multi-source heterogeneous medical data of a patient, the patient's pointing information, clinical semantic information and occurrence time sequence information are uniformly processed, data segments are converged into state evidence units and evidence quality identifiers are generated, and a patient evidence unit set is obtained; a patient state evolution graph is constructed based on the patient evidence unit set, and a state gating process is used to control the state propagation process, so that a patient state representation containing state transmission information and conflict marking information is formed; then, the state evidence units are merged into a portrait theme node, and a standardized patient portrait is obtained; finally, a portrait interpretation rule is matched, and a portrait interpretation result is generated and associated.

Description

Technical Field

[0001] This invention relates to the field of medical data processing, and more specifically, to a method for constructing patient profiles based on artificial intelligence. Background Technology

[0002] With the development of medical informatization, regional medical collaboration, and health management platforms, patients continuously generate a large amount of multi-source, heterogeneous medical data during diagnosis, testing, medication, follow-up, home monitoring, and health management. This type of data typically comes from electronic medical record systems, testing systems, medication management systems, chronic disease follow-up systems, home monitoring terminals, and health management platforms. The data content includes structured fields, semi-structured records, and unstructured text. Because different business systems have different data collection purposes, field settings, terminology, and recording rules, the health status of the same patient may appear with different names, units, time granularities, or text descriptions. This leads to fundamental processing issues such as patient identity merging, medical terminology standardization, status expression conversion, and chronological alignment before patient profile construction.

[0003] Existing methods for constructing patient profiles typically begin by extracting information such as disease, symptoms, laboratory test results, medication use, and follow-up conclusions from patient-related data. Then, through label extraction, rule matching, feature concatenation, or model classification, several profile labels are generated to assist in patient stratification, risk identification, follow-up arrangements, or health management. While these methods can integrate patient information to some extent, their focus is usually on label generation or result classification, with insufficient attention paid to the semantic relationships, temporal evolution, and differences in evidence quality among various types of medical evidence prior to profile generation. When data from different sources are directly used as equally reliable information in profile aggregation, missing records, expired records, negative textual expressions, outliers, and inconsistencies between test results and follow-up descriptions are easily mixed into the same profile topic, causing profile topic state shifts or unstable confidence results.

[0004] Especially in chronic disease management, risk screening, and continuous health management scenarios, a patient's true condition is often not determined by a single record, but rather by multiple pieces of evidence with clinical semantic relevance and temporal sequence. For example, the same patient's test results, medication adjustment records, follow-up texts, and home monitoring data may describe different aspects of the same health problem; some of this evidence may exhibit local contradictions due to differences in collection time, data sources, patient self-reporting biases, or missing record fields. If there is a lack of a unified expression of state evidence units, a lack of assessment of the stability of state propagation, and a lack of control over the propagation path of conflicting evidence during the profile construction process, conflicting information may continue to spread along the profile construction chain, making it difficult for subsequent profile thematic nodes to distinguish between credible evidence, locally anomalous evidence, and conflicting evidence requiring verification.

[0005] Therefore, before generating patient profiles, a processing method is needed that can establish a unified state representation for multi-source heterogeneous medical data, organize medical relationships and temporal evolution relationships, record evidence quality, and control the propagation of conflicting evidence. This would ensure that patient profiles no longer rely solely on isolated labels or simple feature splicing, but rather form a clear data flow relationship between state evidence units, patient state evolution diagrams, state gate control processing, and profile topic aggregation. This processing method also needs to retain the topic evidence chain, topic confidence state, and conflict marker information when outputting the profile results, ensuring that the profile conclusions have a clear source of evidence and a verification path.

[0006] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an artificial intelligence-based patient profiling construction method. This method acquires multi-source heterogeneous medical data of patients, unifies patient-related information, clinical semantic information, and occurrence sequence information, converges data fragments into state evidence units, and generates evidence quality identifiers to obtain a patient evidence unit set. A patient state evolution graph is constructed based on a node selection rule that one patient corresponds to one patient state evolution graph and no connection edges are established between different patients. State propagation is controlled through state gate control to form a patient state representation containing state transmission information and conflict marker information. Then, based on clinical semantic relationship constraints and temporal continuity relationship constraints, the state evidence units are merged into profile topic nodes, generating topic states, topic confidence states, and topic evidence chains to obtain a standardized patient profile. Finally, profile interpretation rules are matched to generate and correlate the output profile interpretation results, thereby solving the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for constructing patient profiles based on artificial intelligence includes: Step S1: Obtain multi-source heterogeneous medical data of patients, unify the patient orientation information, clinical semantic information and occurrence time information, converge the data segments corresponding to the same patient and the same clinical semantic anchor point into state evidence units, and generate evidence quality identifiers for state evidence units to obtain a set of patient evidence units. Step S2: Based on the patient evidence unit set, construct the patient state evolution graph by using the state evidence units as nodes in a way that the same patient corresponds to one patient state evolution graph and no connection edges are established between state evidence units of different patients. Then, use state gate control to control the state propagation process in the patient state evolution graph to obtain a patient state representation containing state transmission information and conflict marking information. Step S3: Based on the patient state representation, state evidence units that have at least one clinical semantic relationship among the corresponding clinical semantic anchors, such as subordinate relationship, adjacent relationship, or common management goal, and whose occurrence sequence is connected, intersecting, or can form a continuous state description according to the temporal evolution relationship, are merged into the portrait topic node; within the portrait topic node, the state evidence units that are allowed to participate in the propagation are sorted according to the occurrence sequence to form a topic state sequence, the topic state is summarized according to the topic state sequence, and the topic confidence state and topic evidence chain are generated according to the evidence quality identifier, state transmission information, and conflict marker information to obtain a standardized patient portrait; Step S4: Based on the standardized patient profile, match each profile topic node with the profile interpretation rules to generate the profile interpretation result, and output the topic evidence chain, topic confidence status and conflict marker information.

[0009] Furthermore, in step S1, patient-related information is uniformly processed through the patient master index. The data processing server determines the affiliation of the same patient based on the patient's unique identifier, consultation number, management file number, and patient-related feature sequence. The medical named entity recognition model is used to extract disease names, symptom descriptions, test and examination names, medication names, follow-up management descriptions, status modifiers, and negation expressions from unstructured text, and outputs a candidate sequence of medical terms. Clinical semantic information is uniformly processed through the medical named entity recognition model and the medical semantic catalog, mapping structured entries, semi-structured fields, and medical text fragments to clinical semantic anchors.

[0010] Furthermore, in step S1, the data processing server forms an effective observation interval based on the effective time attribute of the clinical semantic anchor, and converges data segments belonging to the same patient, corresponding to the same clinical semantic anchor, and whose occurrence sequence falls within the same effective observation interval into state evidence units. It also generates evidence quality identifiers based on semantic integrity, temporal validity, and credible consistency, so that the state evidence units carry a unified state representation value and the original data index.

[0011] Furthermore, in step S2, the data processing server splits the patient evidence unit set according to the patient identifier, and uses the state evidence unit as the node of the patient state evolution graph. The same patient corresponds to one patient state evolution graph, and no connection edge is established between state evidence units of different patients. The data processing server determines the medical association relationship based on the subordinate relationship, adjacency relationship and common management goal between clinical semantic anchors, and determines the temporal evolution relationship based on the occurrence sequence and effective observation interval of the state evidence units. When two state evidence units simultaneously satisfy the medical association relationship and the temporal evolution relationship, a connection edge is formed with the state evidence unit that occurs earlier as the propagation source node and the state evidence unit that occurs later as the propagation target node. When two state evidence units occur at the same time, the superior clinical semantic anchor of the subordinate relationship in the medical semantic directory is used as the propagation target direction.

[0012] Furthermore, in step S2, the state gate control process includes: the data processing server composing a quality combination based on the judgment results of semantic integrity, temporal validity, and credibility consistency in the evidence quality identifier, and converting the quality combination into an evidence credibility value based on historically reviewed state evidence units or initial available judgments; for candidate associated node pairs, the medical relevance strength is first determined based on the relationship type of two clinical semantic anchors in the medical semantic catalog, and then the temporal proximity is determined based on the occurrence sequence and effective observation interval relationship of the two state evidence units, and the graph association strength is formed by the multiplicative constraint of the medical relevance strength and the temporal proximity; the evidence stability is determined based on the overall level of the evidence credibility value in the patient state evolution graph and the difference in evidence credibility values ​​between adjacent state evidence units, and the evidence stability is used as the gate control judgment input for the state propagation process.

[0013] Furthermore, in step S2, the data processing server forms a local state group based on consistency verification and determines the dissimilarity of conflicting state evidence units according to the unified state representation values ​​within the local state group. Subsequently, the evidence stability and dissimilarity are input into a monotonic Bayesian gate control model. This model is a constrained probabilistic graphical model for determining the risk of evidence propagation in the patient's state evolution graph, including evidence stability observation nodes, dissimilarity observation nodes, propagation risk state nodes, and gate control action nodes. The data processing server writes the evidence stability into the evidence stability observation node and the dissimilarity into the dissimilarity observation node. Through a conditional probability relationship corrected by monotonic constraints, the propagation risk confidence value remains constant as evidence stability decreases or dissimilarity increases. The propagation risk confidence value is output by the propagation risk state node, and the gate control action node outputs retention, local restriction, or blocking actions. The data processing server performs retention, local restriction, and blocking processing on the state propagation path based on the propagation risk confidence value.

[0014] Further, in step S3, the data processing server reads the state transmission information and conflict marking information from the patient's state representation, and establishes topic merging candidate edges based on the subordinate relationships, adjacency relationships, common management goals, and state propagation paths in the state transmission information between clinical semantic anchors. Subsequently, a constrained connected component aggregation algorithm is used to determine the portrait topic nodes. This algorithm uses state evidence units as vertices, topic merging candidate edges as connections, and clinical semantic relationship constraints and temporal continuity constraints to obtain the set of mergeable state evidence units. This algorithm only considers local states marked as retained or valid within the current portrait topic node range in the state transmission information. Restricted paths are treated as connected edges, while blocked paths are not. Only the corresponding conflict state evidence units are attached as conflict marker information. Each candidate connected component that satisfies the clinical semantic relationship constraint and the temporal continuity relationship constraint corresponds to a portrait topic node, and a topic state, topic confidence state, and topic evidence chain are generated within the portrait topic node. Among them, the clinical semantic relationship constraint is that there is at least one clinical semantic relationship among the subordinate relationship, adjacent relationship, or common management goal between the clinical semantic anchors corresponding to the two state evidence units. The temporal continuity relationship constraint is that the two state evidence units are temporally connected, intersecting, or can form a continuous state description according to the temporal evolution relationship.

[0015] Furthermore, in step S4, the data processing server reads the portrait topic node, topic status, topic confidence status, and topic evidence chain from the standardized patient portrait, and matches the portrait interpretation rules in the order of priority for portrait topic nodes, secondly for clinical semantic anchors, and supplemented by common management goals, to generate the portrait meaning interpretation and evidence reliability interpretation corresponding to the portrait topic node.

[0016] Furthermore, in step S4, the data processing server organizes the main sources of evidence according to the occurrence sequence of the state evidence units in the subject evidence chain, and generates conflict evidence prompts based on the conflict marker information; then, a traceable path is formed according to the hierarchy of the portrait subject node, subject status, subject evidence chain, state evidence unit and original data index, and the traceable path, subject evidence chain, subject confidence status and conflict marker information are associated with the portrait interpretation result.

[0017] The technical effects and advantages of the patient profiling construction method based on artificial intelligence of the present invention are as follows: This invention unifies multi-source heterogeneous medical data into state evidence units and standardizes the expression of patient-related evidence by combining clinical semantic anchors, unified state representation values, and evidence quality identifiers. This reduces profiling biases caused by inconsistencies in terminology, time granularity, and recording quality between different business systems. By constructing a patient state evolution graph and performing state gate control processing, state evidence units with inexplicable contradictions or local anomalies are retained, partially restricted, or blocked, ensuring that conflicting evidence does not affect the profile subject state under the same conditions. By associating subject state, subject confidence state, subject evidence chain, and conflict marker information in the standardized patient profile, the profile result simultaneously presents the state conclusion, the reliability of the evidence, and the evidence source path, facilitating medical administrators to identify low-confidence evidence and conflicting evidence, and to review the patient profile conclusions. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a patient profiling construction method based on artificial intelligence according to the present invention. Figure 2 This is a schematic diagram illustrating the generation of the state evidence unit in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 - Figure 2 This invention provides a method for constructing patient profiles based on artificial intelligence, comprising: The following steps are all performed by the data processing server. The data processing server extracts patient-related evidence from multi-source heterogeneous medical data generated during diagnosis, treatment, testing, medication, follow-up, home monitoring, and health management, and sequentially completes the generation of state evidence units, construction of patient state evolution diagrams, state gate processing, profile theme aggregation, and profile interpretation output, thereby obtaining interpretable and traceable patient profile construction results.

[0021] Step S1: Obtain multi-source heterogeneous medical data of patients, unify the patient-directed information, clinical semantic information and occurrence time information, converge the data segments corresponding to the same patient and the same clinical semantic anchor point into state evidence units, and generate evidence quality identifiers for state evidence units to obtain a set of patient evidence units.

[0022] Step S1 begins with processing the patient's multi-source heterogeneous medical data. This multi-source heterogeneous medical data includes data fragments related to the patient's status from medical records, laboratory test records, medication records, follow-up records, home monitoring records, and health management records. The data processing server reads the patient-related information, clinical semantic information, and occurrence sequence information from each data fragment, and maintains the correspondence between each data fragment and the source record index, enabling the subsequently generated status evidence units to be traced back to the original data source.

[0023] The data processing server first standardizes the patient-related information. For data segments that can directly match the existing patient unique identifier, visit number, or management file number of the medical institution, the data segment is written into the corresponding patient master index; the patient master index is the correspondence between the unified patient identifier formed after the same patient is identified and the source record index. For data segments that cannot be directly matched, the data processing server extracts the patient-related feature sequence from the patient's name, gender, date of birth, contact information, anonymized ID number, and medical institution code, and generates a matching judgment result according to the strong consistency priority rule: if the anonymized ID number and date of birth are the same, they are determined to be the same patient; if the contact information and name are the same, and the gender and date of birth do not conflict, they are determined to be the same patient; if only the name is the same but there is a conflict in the date of birth, contact information, or anonymized ID number, they are determined to be different patients; if there are missing fields and the same patient judgment condition is not met, it is marked as a data segment to be verified. Data segments to be verified are not included in the patient master index; only the original data index and the verification mark are retained.

[0024] Once the patient master index is determined, the data processing server performs standardized processing on the occurrence time sequence information. The server extracts the record generation time, examination sampling time, test report time, medication start time, follow-up time, or home monitoring collection time from the data fragments, determining the occurrence time sequence according to the order in which the actual occurrence of the clinical event takes precedence over the record writing. When multiple time fields exist for the same data fragment, the time field that best represents the point in time when the patient's status was established is prioritized; if this cannot be determined, the record generation time is used, and the time source is written into the time sequence criteria in the evidence quality identifier. Thus, each data fragment establishes a unified correspondence between the patient identifier, occurrence time sequence, source record index, and original content.

[0025] After unifying the patient-related information and the time sequence of events, the data processing server standardizes the clinical semantic information. For structured entries, it reads the field names, field values, and business scenarios; for semi-structured fields, it splits the field titles, units, values, and remarks; for unstructured text, it uses a medical named entity recognition model to extract disease names, symptom descriptions, test names, medication names, follow-up management descriptions, status modifiers, and negation expressions.

[0026] The data processing server maps structured entries, semi-structured fields, and candidate sequences of medical terms output by the medical named entity recognition model to the medical semantic directory. The data processing server completes the clinical semantic anchor mapping in the order of terminology standardization, synonym matching, business scenario verification, and valid time attribute verification. When multiple candidate clinical semantic anchors exist, the priority anchor is determined based on the business scenario and valid time attribute, and unused candidate relationships are retained in the traceability relationships of the original data index.

[0027] The Medical Semantic Directory is a standardized semantic index set used to carry clinical semantic anchors and their interrelationships. It is used to consolidate patient-related data from different business sources, with different writing styles, and different field structures under the same medical meaning. The Medical Semantic Directory is compiled from existing diagnostic and treatment terms, laboratory and examination terms, medication terms, follow-up terms, and disease management terms used by medical institutions. Specifically, diagnostic and treatment terms cover disease names and symptoms; laboratory and examination terms cover test items, examination items, result units, and reference ranges; medication terms cover drug names, drug categories, administration methods, and treatment cycles; follow-up terms cover follow-up questions, management conclusions, and patient self-reported status; and disease management terms cover chronic disease management goals, risk stratification goals, and health management goals. The data processing server uniformly writes the standard names, synonyms, business scenarios, and effective time attributes of the above terms into the Medical Semantic Directory and generates a clinical semantic anchor for each standard medical meaning. Each clinical semantic anchor is a standardized semantic benchmark in the medical semantic catalog, used to indicate a medical meaning that can be used in patient profiling. Each clinical semantic anchor includes at least an anchor name, synonym expression, applicable business scenario, effective time attribute, state interval, and anchor relationship record. The anchor name records the standard expression of the medical meaning; the synonym expression records different ways of writing the same medical meaning in different business systems, text records, or field names; the applicable business scenario records the source scenarios in which the clinical semantic anchor can appear; the effective time attribute records the effective duration and observation direction of the corresponding state in the patient state expression; the state interval converts numerical or hierarchical results into a unified state representation value; and the anchor relationship record records the subordinate, adjacent, and common management goal relationships between the clinical semantic anchor and other clinical semantic anchors. All of the above serve as the basis for subsequent clinical semantic anchor mapping, state evidence unit convergence, patient state evolution graph construction, and profile topic aggregation. In one embodiment, for diabetes management scenarios, the medical semantic directory sets clinical semantic anchors such as glycated hemoglobin, fasting blood glucose, postprandial blood glucose, hypoglycemic medication, hypoglycemic reaction, and blood glucose control status. The anchor name for the glycated hemoglobin clinical semantic anchor is "Glycated Hemoglobin," and synonyms include HbA1c, glycated Hb, and glycated hemoglobin detection. Applicable business scenarios include laboratory test records and chronic disease follow-up records. The effective time attribute is used to indicate the observation period of the test result in glucose metabolism control evaluation, and the state interval is used to convert the test result into a unified state representation value of being below the target state, in the target state, or above the target state. Synonyms for the fasting blood glucose clinical semantic anchor include FPG, fasting glucose, and fasting blood glucose value, and applicable business scenarios include laboratory test records, home monitoring records, and follow-up records. Synonyms for the hypoglycemic medication clinical semantic anchor include oral hypoglycemic drugs, insulin therapy, and hypoglycemic regimen adjustment, and applicable business scenarios include medication records and follow-up records. The medical semantic catalog records glycated hemoglobin, fasting blood glucose, and postprandial blood glucose as common management targets, hypoglycemic medication and blood glucose control status as common management targets, and hypoglycemic reactions and hypoglycemic medication as adjacent relationships. Based on this, the data processing server can map expressions such as HbA1c, fasting glucose, elevated blood glucose, and hypoglycemic medication adjustments from different sources to corresponding clinical semantic anchors, and establish medical associations and temporal evolution relationships in subsequent steps.

[0028] During the clinical semantic anchor mapping process, the data processing server synchronously generates unified state representation values. Unified state representation values ​​are numerical or hierarchical results that express the clinical meaning and strength of state evidence units at the same scale. For numerical test records, the data processing server first converts the original test values ​​to the standard units of the corresponding clinical semantic anchor according to the unit conversion rules in the medical semantic catalog. Then, it reads the state interval corresponding to the clinical semantic anchor and maps the converted test values ​​to unified state representation values ​​indicating whether the state is below the target state, in the target state, above the target state, or cannot be determined. The state interval is determined by the reference range bound to the clinical semantic anchor in the medical semantic catalog, the disease management target range, or the state division rules maintained by the medical institution. Clinical semantic anchors without a maintained reference range do not generate numerical unified state representation values, and insufficient semantic integrity is recorded in the evidence quality identifier. For medication and follow-up texts, the data processing server generates hierarchical unified state representation values ​​based on the entity category, state modifiers, and negation expressions output by the medical named entity recognition model. When a negation expression covers an entity, it is recorded as a non-existent or unknown state. When a state modifier points to an increase, worsening, or poor control, it is recorded as abnormal enhancement. When a state modifier points to a decrease, relief, or stability, it is recorded as relief or stability. State values ​​below the target state, in the target state, above the target state, as well as states not existing, relief or stability, and abnormal enhancement in the text state, are all converted into ordered numerical codes. State values ​​that cannot be determined or are unknown are not included in the state representation matrix calculation but are only retained in the evidence quality identifier. All unified state representation values ​​are bound to corresponding clinical semantic anchors for subsequent consistency verification and disambiguation calculation.

[0029] After completing the above standardization process, the data processing server converges data segments belonging to the same patient, corresponding to the same clinical semantic anchor, and with adjacent occurrence times into state evidence units. Specifically, the data processing server first groups data segments according to a unified patient identifier, then performs secondary grouping within each patient group according to clinical semantic anchors, and subsequently divides adjacent occurrence times based on the occurrence time sequence and the effective time attribute of the clinical semantic anchor. The effective observation interval is generated based on the occurrence time sequence of the data segments or state evidence units; for laboratory test results, the effective observation interval extends backward from the sampling time; for medication records, the effective observation interval is formed from the medication start time to the medication end time or the prescription expiration date; for follow-up records and home monitoring records, the effective observation interval is formed based on the observation length maintained by the medical semantic catalog, centered on the record occurrence time. When multiple data segments belong to the same patient, correspond to the same clinical semantic anchor, and their occurrence times fall within the same effective observation interval or adjacent observation intervals can jointly express a continuous state, these data segments are merged into a single state evidence unit.

[0030] When generating a state evidence unit, the data processing server retains the correspondence between patient identifiers, clinical semantic anchors, occurrence time sequences, original data indexes, and unified state representation values, and establishes an internal time sequence chain according to the order in which data fragments enter the state evidence unit. If multiple unified state representation values ​​exist within the same valid observation interval, the primary state representation is first determined according to source priority and occurrence time sequence, and then the remaining unified state representation values ​​are used as the basis for consistency verification. Source priority is fixed within the same patient profile construction task. The data processing server prioritizes test and examination records with clear sampling time, clear medical items, and structured results as the primary state representation; if such records do not exist, treatment records or medication records with clear prescription start times are selected; if still not available, follow-up records or home monitoring records with occurrence times are selected. Medical institution management rules are only used to determine the source order within the same priority level and do not change cross-source priority levels.

[0031] The data processing server generates evidence quality labels based on the performance of state evidence units in terms of semantic completeness, temporal validity, and credible consistency. Semantic completeness records whether a state evidence unit possesses the necessary semantic components to express the target clinical semantics; temporal validity records whether the occurrence sequence of a state evidence unit falls within the valid observation interval of the corresponding clinical semantic anchor; credible consistency records whether a state evidence unit exhibits any inexplicable contradictions with other state evidence units within the same local state group. If the main state representation within the same state evidence unit is consistent with the direction of consistency verification, credible consistency is recorded as consistent; if the two directions are different but can be explained by the sequence of occurrence, changes in medication, or follow-up instructions, credible consistency is recorded as an explainable difference; if the two directions are different and cannot be explained by time changes or medical correlations within the local state group, credible consistency is recorded as an inexplicable contradiction, and the corresponding original data index is written into the evidence quality label.

[0032] For example, in a diabetes management scenario, a patient's outpatient records, glycated hemoglobin (HbA1c) test results, descriptions of poor glycemic control in follow-up texts, and home monitoring blood glucose readings are all grouped into the same patient's master index by the data processing server. The HbA1c test results are mapped to clinical semantic anchors related to glucose metabolism control, and the descriptions of poor glycemic control in the follow-up texts are extracted by a medical named entity recognition model and grouped into adjacent clinical semantic anchors. If the HbA1c test results, follow-up descriptions, and home monitoring blood glucose readings are within adjacent valid observation intervals, the data processing server converges them into corresponding state evidence units and generates unified state representation values ​​and evidence quality identifiers.

[0033] After step S1 is completed, the data processing server obtains a set of patient evidence units. This set of patient evidence units uses state evidence units as the smallest profile evidence object. Each state evidence unit has completed the unification of patient orientation, merging of clinical semantic anchors, organization of occurrence time sequence, generation of unified state representation value, and generation of evidence quality label.

[0034] Step S2: Construct a patient state evolution graph based on the patient evidence unit set, and use state gate control to control the state propagation process in the patient state evolution graph to obtain a patient state representation containing state transmission information and conflict marker information.

[0035] Step S2 reads the patient evidence unit set output in step S1 and uses each state evidence unit as a node in the patient state evolution graph. Each patient corresponds to one patient state evolution graph, and no connection edges are established between state evidence units of different patients. The data processing server first splits the patient evidence unit set according to the patient identifier, and then reads the clinical semantic anchor, occurrence sequence, unified state representation value, and evidence quality identifier of each state evidence unit within each patient group, enabling the nodes to have the basis for entering medical association judgment and temporal evolution judgment.

[0036] The data processing server constructs connection edges based on the medical associations between clinical semantic anchors and the temporal evolution relationships between state evidence units. Medical associations are determined by anchor relationship records pre-maintained in the medical semantic catalog, including subordinate relationships, adjacency relationships, and shared management objectives. When the clinical semantic anchors of two state evidence units exhibit any of these anchor relationships in the medical semantic catalog, they are listed as candidate associated node pairs. Temporal evolution relationships are determined based on the order of occurrence and the effective observation interval of the two state evidence units. The state evidence unit that occurs earlier is considered a candidate propagation source node, and the state evidence unit that occurs later is considered a candidate propagation target node. When their effective observation intervals intersect, connect, or are separated but still within the effective duration, a candidate connection edge with calculable temporal proximity is formed. For candidate associated node pairs, the data processing server first determines the strength of medical relevance, then determines the temporal proximity, and subsequently converts both into graph association strength.

[0037] For state evidence units and state evidence unit The graph correlation strength is formed according to the multiplicative constraint of medical relevance strength and temporal proximity: ; in, State Evidence Unit With state evidence unit The graph association strength between two state evidence units is used to represent the degree to which state propagation is possible between them. The medical relevance strength is derived from the relationship type between two clinical semantic anchors in the medical semantic directory. Direct subordination, direct adjacency, and shared management target relationships correspond to the ordered strength levels maintained in the medical semantic directory, respectively. The ordered strength levels in the medical semantic directory are recorded as non-negative integers, with the lowest level indicating no association and the highest level indicating direct association. The data processing server divides the relationship level of the candidate associated node pair by the highest allowed relationship level in the medical semantic directory to obtain the medical relevance strength within the range of 0 to 1. When the relationship level is the lowest level It is 0. The temporal proximity is derived from the relationship between the occurrence time of two state evidence units and their effective observation intervals. When two effective observation intervals intersect or connect, Set to 1; when two valid observation intervals are separated, first calculate the time interval between the two valid observation intervals. Then, read the effective duration length from the effective time attribute of the two clinical semantic anchors, and take the smaller value as the normalized time length. ,according to Determine the temporal proximity; if If the value is 0, no connecting edges will be generated. Without entering the patient's state evolution diagram.

[0038] After the patient state evolution graph is constructed, the data processing server converts the evidence quality identifier into evidence credibility values. The evidence credibility values ​​are obtained using an empirical probability smoothing method for quality combinations. The data processing server assembles the judgment results of semantic integrity, temporal validity, and credibility consistency into quality combinations, and encodes them into quality combination indexes in a fixed order. Semantic integrity, temporal validity, and credibility consistency are respectively designated as the 1st, 2nd, and 3rd quality dimensions. The judgment result of each dimension is recorded according to a preset enumeration value, and the three enumeration values ​​are concatenated sequentially to form the quality combination. For any quality combination, the data processing server retrieves the index of the same quality combination from the historically reviewed evidence units, counts the number of pieces of evidence corresponding to that quality combination that have been manually confirmed as usable for profile topic formation and the total number of pieces of evidence corresponding to that quality combination, and uses Bayesian smoothing to obtain the evidence confidence value: ; in, State Evidence Unit The credibility of the evidence; State Evidence Unit The corresponding quality combination is determined by semantic integrity, temporal validity, and trustworthy consistency. Quality combination in historically reviewed evidence units The number confirmed as usable for portrait subject formation; Quality combination in historically reviewed evidence units The corresponding total number of evidence units. When the historically reviewed evidence units are insufficient, the medical institution's patient management rules will first provide an initial usability determination for each quality combination, and this initial usability determination will be used as... and The initial statistical source; after subsequent manual review results enter the statistical sequence, the evidence confidence value is updated according to the same formula. When a quality combination does not have historically reviewed evidence units, if the quality combination is determined by the rules to be usable for portrait theme formation, then set... and If it is determined that manual review is required before use, then set... and If it is determined that it cannot be used for portrait theme creation, then set and .

[0039] After the evidence confidence value transformation is completed, the data processing server calculates the overall level of evidence confidence values ​​within the patient state evolution graph and calculates the smoothing loss caused by the difference in evidence confidence values ​​between adjacent state evidence units under the influence of graph correlation strength. Evidence stability is... express, The dimensionless statistic used to characterize the stability of evidence propagation in a patient state evolution graph is calculated as follows: ; in, It is a set of nodes in the patient state evolution graph, derived from state evidence unit nodes; The set of connecting edges in the patient state evolution graph is derived from medical associations and temporal evolution relationships; For nodes The degree of correlation contribution, according to calculate; and They are nodes and nodes The credibility of the evidence; For nodes With nodes The strength of the graph association between them; It is a stable constant in the denominator, used to prevent division by zero when the contribution of node association or the association strength of the connected edge graph is 0. Take a fixed positive number that is less than the minimum non-zero graph association strength in the historical samples; if there is no non-zero graph association strength in the historical samples, take the minimum positive floating-point number supported by the data processing server. The calculation of the state evolution diagram for the same batch of patients remains consistent. The first term on the left side of the formula represents the overall level of evidence confidence, and the second term represents the smoothing loss. The difference between the two yields the evidence stability.

[0040] While calculating the stability of the evidence, the data processing server performs consistency checks and conflict localization on the patient state evolution graph. The data processing server forms a local state group based on the first-order adjacency range of the target state evidence unit. This local state group includes the target state evidence unit and its adjacent state evidence units connected by edges. If an adjacent state evidence unit does not belong to the same patient as the target state evidence unit, or if its connecting edge does not simultaneously contain both medical association and temporal evolution relationship, it will not be included in this local state group. For conflicting state evidence units that are locally restricted, a local state group is formed only within their restricted range.

[0041] The data processing server reads the unified state representation values ​​of each state evidence unit within the local state group and arranges them according to their occurrence sequence to form a state representation matrix. ,in Represents the local state group. State representation matrix. Each row corresponds to a state evidence unit within the local state group that enters the matrix calculation, and each column corresponds to a state representation dimension in the unified state representation value vector. Unified state representation values ​​with unknown or undeterminable states are not included in this matrix calculation but are retained only with the evidence quality identifier. Subsequently, consistency verification is performed based on the medical correlation between clinical semantic anchors and the temporal evolution relationship between state evidence units. If the direction of change, temporal order, and medical correlation of the unified state representation values ​​of adjacent state evidence units can be jointly explained, they are recorded as interpretable differences. If there are opposite directions, abrupt changes in strength, or inconsistencies with the medical meaning of adjacent anchors among the unified state representation values, and the credible consistency in the evidence quality identifier has already indicated the existence of unexplainable contradictions, then the corresponding state evidence unit is identified as a conflicting state evidence unit.

[0042] For local state groups The data processing server uses the state representation matrix as a basis. Determine the robust center and the stable distribution range. Robust Center Local state group The central state vector in the unified state representation value space, according to the state representation matrix The median of each state representation dimension is used to reduce the influence of a single extreme state evidence unit on the local state group center. Stable distribution range. The contraction covariance estimation yielded the following: ; in, Local state group The stable distribution range is used to calculate the degree of standardized deviation of the evidence units in conflict states; Local state group The empirical covariance matrix is ​​derived from the state representation matrix. ; For the reason The diagonal stable matrix formed by the diagonal elements is used to reduce the irreversible risk caused by the strong correlation between different state representation dimensions; The contraction coefficient is given by the local state group. Number of state evidence units involved in the calculation of the state representation matrix and the dimension of the unified state representation value vector Confirmed, according to calculate; For matrix stability constants, when When the average of the diagonal elements is greater than 0, Take the product of this average value and the preset numerical precision coefficient, when When the average of the diagonal elements is 0, Take the smallest positive floating-point number supported by the data processing server, and Maintain consistency in the calculation of the state evolution diagram for the same batch of patients; This is the identity matrix. All of the above matrices are formed from the unified state representation values ​​within the local state group.

[0043] For conflict state evidence units Its standardization deviation is express: ; in, For conflict state evidence unit The unified state representation value vector; This represents the transpose operation. If multiple conflicting state evidence units exist in the same patient state evolution graph, the data processing server's standardization deviation for all conflicting state evidence units is... When performing quantile aggregation, first, the standardized deviation degrees corresponding to all conflicting state evidence units in the current patient state evolution graph are combined into a deviation degree sequence. ,in, This represents the number of conflicting state evidence units in the current patient state evolution graph. For the first The standardized deviation of each conflict state evidence unit from the stable distribution range of its local state group. Subsequently, the data processing server sequences the deviations. Sort the values ​​from smallest to largest to obtain an ordered deviation sequence. ,in, This represents the minimum standardized deviation after sorting. This indicates the maximum degree of standardized deviation after sorting; The target quantile used in quantile aggregation is... express, For quantile positions used to represent higher levels of conflict but not entirely determined by a single extreme deviation. Target quantile. Priority is given to determining the target quantile based on historically confirmed conflict localization results. Specifically, the data processing server reads the deviation degree of conflict state evidence units in historical patient state evolution map samples that have been manually verified, and selects the quantile that covers most high-conflict samples and does not fall into the position of the maximum extreme value as the target quantile. In the absence of historical samples, the target quantile... Use the upper quartile, i.e. Determine the target quantile. Then, the data processing server determines the quantile positions as follows: ; in, For the target quantile in the ordered deviation sequence The position in the middle, The number of evidence units in a conflict state. For the target quantile. If If it is an integer, then... As a degree of ambiguity ;like If it is not an integer, then let Not greater than integer position, Not less than The integer positions are determined, and the degree of discrepancy is determined by linear interpolation of adjacent positions. : ; in, , which is the degree of ambiguity, used to characterize the overall deviation of conflicting state evidence units in the current patient state evolution diagram from the stable distribution range of the local state group; and These represent the standardized deviations located on either side of the target quantile in the ordered deviation sequence. If there is only one conflicting state evidence unit in the current patient state evolution graph, then the standardized deviation of that conflicting state evidence unit is directly used as the dissimilarity. If there are no conflicting evidence units, then the degree of discrepancy is... Recorded as 0. Through the above quantile aggregation, the dissimilarity is... It can reflect the high level of deviation in multiple conflict state evidence units, while reducing the dominance of a single extreme outlier on the overall conflict judgment.

[0044] After calculating the stability and ambiguity of the evidence, the data processing server performs state propagation and state gate control. In this embodiment, state propagation is a path-based state transfer process based on the patient state evolution graph. State propagation does not change the unified state representation value formed by the state evidence units in step S1; instead, it writes the propagable unified state representation value, propagation direction, and propagation path into the state transfer information. For connection edges in the patient state evolution graph, the data processing server uses the state evidence unit that occurs earlier as the propagation source node and the state evidence unit that occurs later as the propagation target node. If two state evidence units occur at the same time, the higher-level clinical semantic anchor point of the subordinate relationship in the medical semantic directory is used as the propagation target direction. Candidate paths for state propagation consist of connection edges whose graph association strength reaches the associative boundary.

[0045] The data processing server will ensure stability Write the stability observation node into the monotonic Bayesian gated model, and include the dissimilarity. The propagation risk confidence value is written into the dissimilarity observation node and outputs by the propagation risk state node according to the conditional probability relationship. The monotonic Bayesian gate control model is a constrained probabilistic graphical model constructed for determining the risk of evidence propagation in the patient state evolution graph. It sets evidence stability and ambiguity as observation nodes, and propagation risk state and gate control action as constrained output nodes. It corrects the conditional probability relationship through monotonic constraints so that the confidence value of propagation risk does not decrease as evidence stability decreases or ambiguity increases.

[0046] In one embodiment, the monotonic Bayesian gate control model belongs to the discrete probabilistic graphical model for determining the risk of state propagation in patient state evolution graphs. The data processing server uses historical patient state evolution graph samples as the training dataset. Each sample includes at least evidence stability, dissimilarity, confirmed conflict location results, state propagation path processing records, and subsequent portrait topic state review results. The subsequent portrait topic state review results are used to indicate whether the propagation of conflict evidence leads to a shift in the portrait topic state. Before training, the data processing server divides evidence stability into three intervals: insufficient credibility, generally credible, and credibly stable; dissimilarity into three intervals: low deviation, moderate deviation, and high deviation; and gate control actions into three categories: retention, local restriction, and blocking. The model structure includes evidence stability observation nodes, dissimilarity observation nodes, propagation risk state nodes, and gate control action nodes. The evidence stability observation nodes and dissimilarity observation nodes both point to the propagation risk state nodes, and the propagation risk state nodes point to the gate control action nodes. During model training, the data processing server statistically analyzes the frequency of portrait topic state shifts under each combination of stability and ambiguity intervals, forming a conditional probability table for transmission risk state nodes. It then statistically analyzes the adoption rates of three types of gate control actions—retention, local restriction, and blocking—under different transmission risk states, forming a conditional probability table for gate control action nodes. When samples are insufficient, the conditional probability tables are initialized by the medical institution's patient management rules and updated after the accumulation of manual review samples. The training objective is to ensure that the transmission risk confidence value is consistent with the manually confirmed portrait topic state shift results, and that the gate control actions are consistent with the manually reviewed adoption actions. During model validation, the data processing server uses validation samples to check the consistency between the transmission risk confidence value and the portrait topic state shift label, and performs monotonic constraint correction on the conditional probability tables to ensure that the transmission risk probability does not decrease when stability decreases or ambiguity increases. When the model is invoked, the data processing server inputs the stability and heterosity of the current patient state evolution graph, queries the conditional probability table after monotonic correction, outputs the propagation risk confidence value, and outputs the retention, local restriction or blocking action by the gate action node.

[0047] Propagation risk confidence value This section describes the conditional probability of a shift in the profile's subject state after the propagation of conflicting evidence, given the current combination of evidence stability and dissimilarity intervals. The data processing server first divides evidence stability into intervals from low to high according to the quantile boundaries in historical samples, and divides dissimilarity into intervals from low to high according to the quantile boundaries in historical samples. Then, it calculates the empirical probability of propagation risk under each combination of evidence stability and dissimilarity intervals, forming a conditional probability table for propagation risk. Both evidence stability and dissimilarity intervals use three levels of discrete intervals: low, medium, and high. When the number of historical samples is sufficient for quantile statistics, the three quantile boundaries of the historical sequences of evidence stability and dissimilarity are used to form three levels of intervals. When the number of historical samples is insufficient, based on the patient management rules of medical institutions, evidence stability is first mapped to three levels: insufficient credibility, generally credible, and credibly stable; dissimilarity is mapped to three levels: low deviation, moderate deviation, and high deviation. When the sample is insufficient, the patient management rules of medical institutions are only used to generate the initial conditional probability relationship. This initial conditional probability relationship is formed by mapping the rule conditions to the stability interval and the ambiguity interval, mapping the rule processing conclusion to the retention, local restriction or blocking action, and is updated after the accumulation of manually reviewed samples.

[0048] The data processing server performs monotonic constraint verification on the propagation risk conditional probability table: Without decreasing the dissimilarity, a decrease in the proof stability must not lead to a decrease in the propagation risk probability; without increasing the proof stability, an increase in the dissimilarity must not lead to a decrease in the propagation risk probability. If the conditional probability relationship does not satisfy this constraint, a two-dimensional order-preserving regression is used to process the conditional probability table. During correction, when the dissimilarity interval is fixed as a column, the pooled adjacent violation algorithm is executed along the direction of proof stability from high to low to ensure that a decrease in proof stability does not decrease the propagation risk probability; when the proof stability interval is fixed as a row, the pooled adjacent violation algorithm is executed along the direction of dissimilarity from low to high to ensure that an increase in dissimilarity does not decrease the propagation risk probability. If row correction disrupts the monotonicity of the column direction, or column correction disrupts the monotonicity of the row direction, row and column corrections are performed alternately until the conditional probability table satisfies the monotonic constraint in both directions. The corrected conditional probability table serves as the conditional probability relationship for the propagation risk state nodes, outputting the propagation risk confidence value. .

[0049] In one embodiment, the data processing server first reads the stability, ambiguity, conflict localization results, and subsequent profile topic status review results for each sample from historical patient state evolution map samples. It then divides the stability into three intervals: insufficient credibility, generally credible, and credibly stable; and the ambiguity into three intervals: low deviation, moderate deviation, and high deviation. Subsequently, the data processing server calculates the proportion of profile topic state shifts under each combination of stability and ambiguity intervals, forming a conditional probability table for transmission risk. If a certain interval combination lacks historical samples, it is filled with the initial risk level given by the medical institution's patient management rules. After completing the initial conditional probability table, the data processing server checks whether the table satisfies monotonic constraints. When a decrease in stability occurs but a decrease in transmission risk occurs, or an increase in ambiguity occurs but a decrease in transmission risk occurs, adjacent violation intervals are merged and reassigned the same transmission risk level until each interval combination satisfies the relationship that a decrease in stability does not reduce risk, and an increase in ambiguity does not reduce risk. The corrected propagation risk conditional probability table serves as the parameter table for the monotonic Bayesian gated model, and is used to query the propagation risk confidence value when inputting stability and dissimilarity.

[0050] The gate control action node determines the state propagation path processing method based on local constraint limits and blocking limits. Local constraint limits and blocking limits are determined by the distribution of propagation risk confidence values ​​in the validation samples. The data processing server categorizes validation samples into three groups based on the manual review conclusions: no portrait topic state shift, partial portrait topic state shift, and significant portrait topic state shift. If the topic state of the portrait topic node remains unchanged before and after the propagation of conflicting evidence, it is classified as no portrait topic state shift. If the topic state change only occurs within the range of directly adjacent clinical semantic anchors within the same portrait topic node and does not change the final topic state of that portrait topic node, it is classified as partial portrait topic state shift. If the propagation of conflicting evidence causes a change in the final topic state of the portrait topic node, or causes the topic confidence state to decrease from high or medium confidence to low confidence, it is classified as significant portrait topic state shift. The propagation risk confidence values ​​of samples causing partial portrait topic state shift are sorted in ascending order, and the quartiles are used as local constraint limits; the propagation risk confidence values ​​of samples causing significant portrait topic state shift are sorted in ascending order, and the quartiles are used as blocking limits. If the blocking limit is lower than the local constraint limit, the blocking limit will be adjusted to the adjacent higher quantile value that is not lower than the local constraint limit.

[0051] When the propagation risk confidence value reaches the blocking threshold, the data processing server blocks the outward propagation path of the conflict state evidence unit and writes the conflict state evidence unit, its corresponding original data index, deviation degree, propagation risk confidence value, and gate action into the conflict marker information. When the propagation risk confidence value reaches the local limitation threshold but not the blocking threshold, the data processing server restricts the conflict state evidence unit to state propagation only within the same clinical semantic anchor point or directly adjacent anchor points, and writes the limitation range into the state transmission information. When the propagation risk confidence value does not reach the local limitation threshold, the data processing server retains the original state propagation path and writes the propagable path into the state transmission information. The state transmission information records at least the propagation source node, the propagation target node, the index of the transmitted unified state representation value, the graph association strength, the gate action, and the propagation range. The transmitted unified state representation value index preferentially points to the main state representation in the state evidence unit; other unified state representation values ​​that serve as the basis for consistency verification within the same state evidence unit are not directly transmitted as the subject state formation objects, but enter the subject evidence chain along with the original data index; if a local restriction path allows a certain consistency verification basis to participate in the subject state formation within the restriction range, then the consistency verification basis is marked as locally valid evidence, and its index and participation range are recorded in the state transmission information.

[0052] For example, in a diabetes management scenario, if both glycated hemoglobin and follow-up blood glucose control descriptions indicate poor blood glucose control, while home monitoring blood glucose records consistently show normal levels over similar periods, and no medication adjustments or data collection conditions can explain the difference, then the state evidence unit corresponding to the home monitoring blood glucose record is identified as a conflicting state evidence unit. The data processing server calculates the standardized deviation of this state evidence unit from the stable distribution range of the local state group and further generates a dissimilarity score. When the stability score is low and the dissimilarity score is high, the propagation risk confidence value output by the monotonic Bayesian gated model reaches the blocking threshold. Therefore, the outward propagation path of this home monitoring node towards the glucose metabolism control profile is blocked, while the conflict marker information is retained.

[0053] After step S2 is completed, the data processing server obtains a patient state representation containing state propagation information and conflict marker information. The patient state representation records the state propagation path in the patient state evolution graph after preservation, local restriction, or blocking processing, and retains traceable markers of conflict state evidence units.

[0054] Step S3: Based on the patient state representation, state evidence units that have at least one clinical semantic relationship among the corresponding clinical semantic anchors, such as subordinate relationship, adjacent relationship, or common management goal, and whose occurrence sequence is connected, intersecting, or can form a continuous state description according to the temporal evolution relationship, are merged into the portrait topic node; within the portrait topic node, the state evidence units that are allowed to participate in the propagation are sorted according to the occurrence sequence to form a topic state sequence, the topic state is summarized according to the topic state sequence, and the topic confidence state and topic evidence chain are generated according to the evidence quality identifier, state transmission information, and conflict marker information to obtain a standardized patient portrait.

[0055] Step S3 uses the patient state representation and patient state evolution graph output in step S2 as the starting point for processing. The data processing server reads the state transit information and conflict marker information from the patient state representation, and simultaneously reads the state evidence units, clinical semantic anchors, medical associations, temporal evolution relationships, graph association strength, and evidence quality identifiers from the patient state evolution graph. The state transit information is used to determine whether the state evidence unit is allowed to participate in topic aggregation, and the conflict marker information is used to identify the scope of state evidence units that have been locally restricted or blocked during aggregation.

[0056] The data processing server first locates the patient state evolution graph corresponding to a single patient based on the patient identifier, and then establishes topic merging candidate edges within this patient state evolution graph. Topic merging candidate edges are temporary connections used to determine whether two state evidence units can enter the same profile topic node. The establishment of a topic merging candidate edge requires simultaneously satisfying three conditions: First, there is at least one clinical semantic relationship between the clinical semantic anchors of the two state evidence units, including a subordinate relationship, an adjacent relationship, or a common management goal; second, the two state evidence units are temporally connected, intersecting, or there is a state propagation path between them preserved in step S2; third, the state transmission information does not mark the propagation path between them as blocked, and the local restricted path does not exceed the range of the same clinical semantic anchor or directly adjacent anchors. Common management goals are only derived from management goals in the medical semantic catalog that have already been mapped to clinical semantic anchors.

[0057] When the clinical semantic anchors of two state evidence units belong to the same common management goal, or when they have a subordinate or adjacent relationship in the medical semantic catalog, and when they are temporally connected, intersecting, or can form a continuous state description according to a temporal evolution relationship, the data processing server further reads the state transmission information formed in step S2. If the corresponding propagation path is preserved, a complete topic merging candidate edge is established; if the corresponding propagation path is locally restricted, a topic merging candidate edge is established only within the same clinical semantic anchor or directly adjacent anchors; if the corresponding propagation path is blocked, the propagation path is not used as a topic merging candidate edge, but the traceability relationship between the conflicting state evidence units and the original data index is preserved.

[0058] After the candidate edges for topic merging are formed, the data processing server uses a constrained connected component aggregation algorithm to determine the portrait topic nodes. This algorithm uses state evidence units as vertices, topic merging candidate edges as connections, and calculates the set of mergeable state evidence units based on clinical semantic relationship constraints and temporal continuity constraints. The clinical semantic relationship constraint requires that there is at least one clinical semantic relationship between the clinical semantic anchors corresponding to two state evidence units, including a subordinate relationship, an adjacency relationship, or a shared management goal. The temporal continuity constraint requires that two state evidence units are temporally connected, intersecting, or can form a continuous state description according to a temporal evolution relationship. The algorithm only uses locally restricted paths marked as retained or valid within the current portrait topic node range in the state transmission information as connected edges; blocked paths are not used as connected edges, but only the corresponding conflicting state evidence units are attached as conflict marker information. After the above processing, each candidate connected component that satisfies the clinical semantic relationship constraints and temporal continuity constraints is generated as a portrait topic node. When the same state evidence unit simultaneously meets the merging conditions of multiple candidate connected components, it is preferentially merged into the common management target directly bound to its clinical semantic anchor point; when there are multiple directly bound common management targets, the candidate connected component with the highest cumulative graph association strength value is merged; when the cumulative values ​​are the same, the candidate connected component containing the most main state sources within the most recent valid observation interval is merged, and the unselected candidate connected components are only retained for retrospective reference. In one embodiment, the glycated hemoglobin state evidence unit, fasting blood glucose state evidence unit, and follow-up blood glucose control description state evidence unit all point to the common management target of glucose metabolism control, and the state propagation paths between the three are all preserved. Then, the data processing server uses the three as vertices and the topic merging candidate edges as the connection relationship to form the same candidate connected component and generate a glucose metabolism control profile topic node. If the home monitoring blood glucose state evidence unit is marked as locally restricted in step S2, the candidate connected component is only added when its restriction range is consistent with the glucose metabolism control profile topic node; if its propagation path is blocked, it does not participate as a connected edge in the topic state formation of the profile topic node, but is only attached to the profile topic node with conflict marking information.

[0059] After the portrait theme node is formed, the data processing server determines the theme state. The data processing server reads the unified state representation value, occurrence sequence, evidence quality identifier, and state transmission information of the state evidence units within the portrait theme node, and sorts the state evidence units within the portrait theme node that are still allowed to participate in propagation according to their occurrence sequence, forming a theme state sequence. The theme state sequence is a sequence of state expressions obtained by arranging the state evidence units within the portrait theme node that can participate in the formation of the theme state in chronological order. Blocked conflict state evidence units do not enter the theme state sequence; locally restricted conflict state evidence units only enter the theme state sequence if the restriction scope is consistent with the current portrait theme node.

[0060] After the topic state sequence is formed, the data processing server determines the direction of the state effect of each state evidence unit within the portrait topic node based on the state transmission information. For state evidence units whose propagation path is preserved, their unified state representation value is entered into the topic state judgment according to the order of occurrence. For state evidence units whose propagation path is locally restricted, the data processing server first determines whether their clinical semantic anchor point belongs to the same clinical semantic anchor point or a directly adjacent anchor point of the current portrait topic node. If this range is met, they only participate in the judgment as a local state supplement. For state evidence units carrying conflict marker information and whose propagation path is blocked, their unified state representation value only enters the topic evidence chain and does not change the topic state.

[0061] The topic status is obtained through temporal constraint state induction. The data processing server first reads the unified state representation value of each state evidence unit in the topic status sequence, and then identifies continuous consistent segments and state inflection points according to the order of occurrence. A continuous consistent segment is a temporal segment in which adjacent state evidence units maintain consistency in the direction and clinical meaning of the unified state representation value; a state inflection point is the position where an interpretable change occurs in the direction or strength of the unified state representation value between adjacent state evidence units. Consistency in the direction of the unified state representation value is determined by the state direction attribute corresponding to the clinical semantic anchor and the ordered numerical code; if the unified state representation value codes of adjacent state evidence units are the same, or both are in the code corresponding to the target state, they are determined to be consistent in direction; unified state representation values ​​whose states are unknown or cannot be determined do not participate in the formation of continuous consistent segments, but only enter the topic evidence chain and topic confidence state judgment. The topic status is primarily determined by consecutive consistent segments within the most recent valid observation interval. When multiple consecutive consistent segments exist within the most recent valid observation interval, the consecutive consistent segment covering the most recent occurrence time sequence is selected first. If multiple candidates still exist, the consecutive consistent segment with the highest average level of confidence value for the corresponding state evidence unit is selected. If they are still the same, the consecutive consistent segment with the most main state sources in the topic evidence chain is selected. When there are insufficient state evidence units within the most recent valid observation interval to form consecutive consistent segments, the data processing server reads adjacent state evidence units with preserved propagation paths to the current profile topic node and determines the topic status by combining this with the evidence quality identifier.

[0062] After the topic status is determined, the data processing server determines the topic confidence status. The data processing server first determines the topic status formation link, which consists of continuous, consistent segments used to form the topic status and their adjacent propagation paths. The topic confidence status is determined in the following order: low confidence first, medium confidence second, and high confidence last. The data processing server first checks for the existence of low confidence triggering conditions. If no low confidence triggering conditions exist, it then checks for the existence of medium confidence triggering conditions. Only when neither low nor medium confidence triggering conditions exist is the topic confidence status recorded as high confidence. If the evidence quality identifiers of all state evidence units within the topic status formation link do not record semantic integrity deficiencies, incomplete temporal validity, or unexplainable contradictions in confidence consistency, and all state propagation paths within the topic status formation link are reserved propagation paths, and the current profile topic node does not have conflict marker information attached to the topic status formation link corresponding to that topic, then the topic confidence status is recorded as high confidence. If at least one locally restricted path exists within the topic state formation link, or if at least one evidence quality identifier is incomplete but the state evidence unit is not the primary state source within the most recent valid observation interval, or if the conflict marker information attached to the current profile topic node does not block the topic state formation link, then the topic confidence status is recorded as moderately confident. If the primary state source within the most recent valid observation interval has an incomplete evidence quality identifier, or if any propagation path in the topic state formation link connecting the most recent valid observation interval is blocked, or if the conflict state evidence unit corresponding to the conflict marker information belongs to the same clinical semantic anchor point as the topic state formation link and the gate control action achieves blocking, then the topic confidence status is recorded as low confident.

[0063] Simultaneously with determining the subject confidence status, the data processing server generates subject confidence evidence. Subject confidence evidence is a traceable processing record used to explain the source of the subject confidence status. It is formed by the availability of evidence quality indicators, the propagation path processing method in the status transmission information, the conflict status evidence units in the conflict marking information, and the gate control actions. Subject confidence evidence is written into the portrait subject node along with the subject confidence status, enabling subsequent portrait interpretation results to simultaneously display the status conclusion and the source of reliability.

[0064] Subsequently, the data processing server generates a thematic evidence chain. Within each profile thematic node, the data processing server reads the source of the state evidence unit, the original data index, the clinical semantic anchor, the state transmission information, the conflict marker information, and the gate control action according to the occurrence sequence and state propagation path of the state evidence unit, forming a chain record that can represent the process of the formation of the thematic state. The thematic evidence chain uses state evidence units as chain objects, the state propagation path after preservation, partial restriction, or blockage as the inter-chain relationship, and attaches conflicting state evidence units to the corresponding chain segments with conflict marker information.

[0065] The subject evidence chain records the evidence paths that participated in the formation of the subject status and the conflicting paths that did not participate in the formation of the subject status but affected the subject confidence status. The evidence paths that participated in the formation of the subject status connect valid subject evidence and locally valid evidence in chronological order, indicating which state evidence units formed the subject status. Conflict paths read the conflict marker information from step S2, indicating which state evidence units were partially restricted or blocked due to inexplicable contradictions, high deviations, or the propagation risk confidence value reaching the corresponding limit. If the same profile subject node contains multiple clinical semantic anchors, the data processing server organizes the subject evidence chain according to the subordinate and adjacent relationships between the clinical semantic anchors, ensuring that the chain order is consistent with the subject status formation order.

[0066] For example, in a diabetes management scenario, the state evidence units corresponding to glycated hemoglobin, fasting blood glucose, hypoglycemic medication, and follow-up blood glucose control descriptions form state propagation paths in step S2. After the data processing server reads the state transmission information, if the propagation path between glycated hemoglobin and fasting blood glucose is preserved, and both point to the glucose metabolism control management goal, then a topic merging candidate edge is established, and both are incorporated into the glucose metabolism control profile topic node. If home monitoring blood glucose recording is blocked, it does not enter the glucose metabolism control topic state formation chain segment, but is attached to the profile topic node as conflict marker information, and the original data index and gate control action are preserved.

[0067] After step S3 is completed, the data processing server obtains a standardized patient profile. The standardized patient profile consists of multiple profile topic nodes. Each profile topic node contains a topic state, a topic confidence state, and a topic evidence chain, and retains the correspondence between the profile topic node and the patient identifier, clinical semantic anchor, state transmission information, and conflict marker information.

[0068] Step S4: Based on the standardized patient profile, match each profile topic node with the profile interpretation rules to generate the profile interpretation result, and output the topic evidence chain, topic confidence status and conflict marker information.

[0069] Step S4 uses the standardized patient profile output in step S3 as the starting point for processing. The data processing server reads the topic status, topic confidence status, topic confidence basis, topic evidence chain, clinical semantic anchor, status transmission information, and conflict marker information of each topic node in the standardized patient profile one by one; at the same time, it reads the profile interpretation rules. The profile interpretation rules are derived from the medical institution's profile display specifications, patient management business rules, and clinical semantic interpretation templates, and establish a mapping relationship with the corresponding clinical semantic anchor or profile topic node.

[0070] The data processing server first completes candidate matching between portrait topic nodes and portrait interpretation rules, forming a candidate set of interpretation rules. This candidate set consists of portrait interpretation rules that are compatible with the current portrait topic node's clinical semantic anchor, portrait topic node name, and topic status. The data processing server performs matching in the following order: portrait topic node priority, clinical semantic anchor second, and supplementary common management goals. When a portrait interpretation rule is directly mapped to the current portrait topic node, that rule is used first. When no direct mapping exists, the server matches topic status interpretation rules, confidence status interpretation rules, conflict evidence suggestion rules, and evidence chain display rules based on the clinical semantic anchor corresponding to the status evidence unit within the portrait topic node. When the same portrait topic node contains multiple clinical semantic anchors, the rule matching order is determined according to the subordinate relationship, adjacency relationship, and common management goals used in step S3 when forming the topic evidence chain, ensuring that the interpretation rules and topic status form a consistent link. Each profile interpretation rule should include at least the applicable object, the applicable status range, the confidence status applicable range, the interpretation template, the evidence reliability expression item, and the conflict prompt expression item. The applicable object is used to record the profile topic node, clinical semantic anchor point, or common management goal corresponding to the rule. The applicable status range is used to match the topic status, and the confidence status applicable range is used to match the topic confidence status.

[0071] After the portrait interpretation rules are matched, the data processing server determines the applicability of the candidate interpretation rules. Applicability conditions include whether the subject status falls within the state expression range corresponding to the rule, whether the subject confidence state meets the reliability expression range corresponding to the rule, whether the conflict marker information triggers the conflict evidence prompting rule, and whether the subject evidence chain has a displayable original data index and propagation path. If multiple portrait interpretation rules simultaneously meet the applicability conditions, they are organized in the output order of subject status interpretation rules, confidence state interpretation rules, conflict evidence prompting rules, and evidence chain display rules. When multiple candidates exist for the same type of portrait interpretation rule, the data processing server determines a unique rule based on the priority of directly mapping to the portrait subject node, mapping to a clinical semantic anchor point, or mapping to a common management target; if multiple candidates still exist within the same priority, the rule consistent with the clinical semantic anchor point corresponding to the nearest valid observation interval in the subject status formation chain is selected. After rule adjudication, a portrait subject node outputs only one portrait interpretation result under the same interpretation type.

[0072] The data processing server then generates an interpretation of the patient profile's meaning and an interpretation of the reliability of the evidence. When generating the interpretation, the server reads the topic state of the profile's topic nodes and converts the structured state expression into medical management meaning according to the topic state interpretation rules. The medical management meaning is an explanatory text surrounding a specific aspect of the patient's profile, directly corresponding to the clinical orientation of the profile's topic nodes, without altering the topic state itself. The server reads the state evidence units involved in the formation of the topic state in the topic evidence chain according to the order in which the topic state is formed, confirming which clinical semantic anchors and their timing support the topic state; subsequently, it connects the topic state, clinical semantic anchors, the continuity of the timing, and the topic state interpretation rules to form the interpretation of the patient profile's meaning.

[0073] The reliability interpretation of evidence is generated based on the subject confidence status and the basis for subject confidence. The data processing server reads the subject confidence status and converts high, medium, or low confidence into a reliability description according to the confidence status interpretation rules; simultaneously, it reads the availability of evidence quality identifiers in the subject confidence basis, the propagation path processing method in the status transmission information, and the gate control action in the conflict marking information to form a reliability source description. If the subject status of the portrait subject node is mainly formed by state evidence units with complete evidence quality identifiers and preserved propagation paths, the reliability interpretation points to a stable evidence source; if there are local restricted paths, incomplete evidence quality identifiers, or conflict-indicating evidence, the reliability interpretation simultaneously writes the corresponding reasons; if a key propagation path is blocked, the reliability interpretation writes low confidence and manual review.

[0074] After the interpretation of the portrait's meaning and the interpretation of the reliability of the evidence are formed, the data processing server associates the thematic evidence chain with conflict evidence prompts. The data processing server organizes the main sources of evidence according to the chronological order of occurrence of the state evidence units in the thematic evidence chain, and then marks the propagation path processing method between each state evidence unit according to the state transmission information; for the propagation path that is retained, its participation in the formation of the thematic state is recorded; for the propagation path that is partially restricted, its scope of participation is recorded; for the propagation path that is blocked, it is recorded that it did not enter the formation of the thematic state but is retained as conflict prompt evidence.

[0075] Conflict evidence prompts are triggered by conflict evidence prompt rules. The triggering condition is that the current portrait topic node is attached with conflict marker information, or the topic confidence status is reduced due to conflict marker information. The data processing server reads the conflict status evidence unit, original data index, deviation degree, propagation risk confidence value, and gate control action from the conflict marker information, and generates conflict evidence prompts according to the conflict evidence prompt rules. The deviation degree originates from the dissimilarity calculation process in step S2, the propagation risk confidence value originates from the monotonic Bayesian gate control model, and the gate control action corresponds to the retention, partial restriction, or blocking processing results. When conflict evidence prompts are written into the portrait interpretation results, a correspondence is established between them and the portrait meaning interpretation and evidence reliability interpretation under the same portrait topic node.

[0076] The data processing server determines the output order of the traceable path based on the evidence chain presentation rules. The traceable path is organized hierarchically according to the portrait subject node, subject status, subject evidence chain, status evidence unit, and original data index, enabling medical administrators to trace back from the portrait interpretation result to the status evidence unit, and then from the status evidence unit back to the original data index retained in step S1. For conflicting status evidence units, the data processing server synchronously records the status propagation path that is partially restricted or blocked in the traceable path, ensuring that the evidence can be viewed without being confused with the evidence paths involved in the formation of the subject status.

[0077] Finally, the data processing server aggregates the portrait interpretation results for each portrait topic node according to the patient identifier, maintaining the correspondence between each portrait interpretation result and the corresponding portrait topic node, topic status, topic confidence status, topic evidence chain, and conflict marker information. During the output organization process, the data processing server first outputs the portrait meaning interpretation corresponding to the portrait topic node, then outputs the topic confidence status and evidence reliability interpretation, followed by the main evidence source, conflict evidence hints, and traceability path. If the same patient has multiple portrait topic nodes, they are arranged according to the management objective order specified in the medical institution's portrait display guidelines; if a portrait topic node contains conflict marker information, conflict evidence hints are synchronously associated at the output position of that portrait topic node, ensuring that conflict evidence is not removed from the portrait topic node display.

[0078] For example, in a diabetes management scenario, a standardized patient profile includes a glucose metabolism control profile topic node. If the topic status of this profile topic node is jointly formed by glycated hemoglobin, fasting blood glucose, and follow-up blood glucose control descriptions, the data processing server generates the meaning of the glucose metabolism control status according to the topic status interpretation rules. If the topic confidence status is moderately reliable, and the reason is due to local restrictions or blockages in home monitoring blood glucose records, then the evidence reliability interpretation records that the topic status is mainly supported by test and follow-up evidence, while indicating that there are differences between home monitoring evidence and the local status group. The data processing server outputs glycated hemoglobin test records, fasting blood glucose records, follow-up records, and their occurrence sequence as traceable paths, and outputs restricted home monitoring blood glucose records as conflict evidence.

[0079] After step S4 is completed, the data processing server obtains interpretable and traceable patient profile construction results. These results include explanations of the profile's meaning, the reliability of evidence, the main sources of evidence, conflict evidence hints, and traceable paths for each profile topic node. Topic evidence chains, topic confidence states, and conflict marker information are all output in conjunction with the profile interpretation results, forming a complete processing loop from multi-source heterogeneous medical data, state evidence units, patient state evolution diagrams, state gate control processing, profile topic nodes, to the profile interpretation results.

[0080] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing patient profiles based on artificial intelligence, characterized in that, include: Step S1: Obtain multi-source heterogeneous medical data of patients, unify the patient orientation information, clinical semantic information and occurrence time information, converge the data segments corresponding to the same patient and the same clinical semantic anchor point into state evidence units, and generate evidence quality identifiers for state evidence units to obtain a set of patient evidence units. Step S2: Based on the patient evidence unit set, construct the patient state evolution graph by using the state evidence units as nodes in a way that the same patient corresponds to one patient state evolution graph and no connection edges are established between state evidence units of different patients. Then, use state gate control to control the state propagation process in the patient state evolution graph to obtain a patient state representation containing state transmission information and conflict marking information. Step S3: Based on the patient state representation, state evidence units that have at least one clinical semantic relationship among the corresponding clinical semantic anchors, such as subordinate relationship, adjacent relationship, or common management goal, and whose occurrence sequence is connected, intersecting, or can form a continuous state description according to the temporal evolution relationship, are merged into the portrait topic node; within the portrait topic node, the state evidence units that are allowed to participate in the propagation are sorted according to the occurrence sequence to form a topic state sequence, the topic state is summarized according to the topic state sequence, and the topic confidence state and topic evidence chain are generated according to the evidence quality identifier, state transmission information, and conflict marker information to obtain a standardized patient portrait; Step S4: Based on the standardized patient profile, match each profile topic node with the profile interpretation rules to generate the profile interpretation result, and output the topic evidence chain, topic confidence status and conflict marker information.

2. The method for constructing a patient profile based on artificial intelligence according to claim 1, characterized in that, In step S1, the patient identification information is uniformly processed through the patient master index, and the data processing server determines the affiliation of the same patient based on the patient's unique identifier, consultation number, management file number, and patient identification feature sequence. The medical named entity recognition model is used to extract disease names, symptom descriptions, test and examination names, medication names, follow-up management descriptions, status modifiers and negation expressions from unstructured text, and output a candidate sequence of medical terms. Clinical semantic information is processed in a unified manner through the medical named entity recognition model and the medical semantic catalog, mapping structured entries, semi-structured fields and medical text fragments to clinical semantic anchors.

3. The method for constructing a patient profile based on artificial intelligence according to claim 2, characterized in that, In step S1, the data processing server forms an effective observation interval based on the effective time attribute of the clinical semantic anchor. Data segments belonging to the same patient, corresponding to the same clinical semantic anchor, and whose occurrence time falls within the same effective observation interval are converged into state evidence units. Evidence quality identifiers are generated based on semantic integrity, temporal validity, and credible consistency, so that the state evidence units carry a unified state representation value and the original data index.

4. The method for constructing a patient profile based on artificial intelligence according to claim 1, characterized in that, In step S2, the data processing server splits the patient evidence unit set according to the patient identifier, and uses the state evidence unit as the node of the patient state evolution graph. The same patient corresponds to one patient state evolution graph, and no connection edge is established between state evidence units of different patients. The data processing server determines the medical association relationship based on the subordinate relationship, adjacency relationship and common management goal between clinical semantic anchors, and determines the temporal evolution relationship based on the occurrence sequence and effective observation interval of the state evidence units. When two state evidence units simultaneously satisfy the medical association relationship and the temporal evolution relationship, a connection edge is formed with the state evidence unit that occurs earlier as the propagation source node and the state evidence unit that occurs later as the propagation target node. When two state evidence units occur at the same time, the higher-level clinical semantic anchor point of the subordinate relationship in the medical semantic catalog is used as the direction of propagation.

5. The method for constructing a patient profile based on artificial intelligence according to claim 4, characterized in that, In step S2, the state gate control process includes: the data processing server composing a quality combination based on the judgment results of semantic integrity, temporal validity, and credibility consistency in the evidence quality identifier, and converting the quality combination into an evidence credibility value based on historically reviewed state evidence units or initial available judgments; for candidate associated node pairs, the medical relevance strength is first determined based on the relationship type of two clinical semantic anchors in the medical semantic catalog, and then the temporal proximity is determined based on the occurrence sequence and effective observation interval relationship of the two state evidence units, and the graph association strength is formed by the multiplicative constraint of the medical relevance strength and the temporal proximity; the evidence stability is determined based on the overall level of the evidence credibility value in the patient state evolution graph and the difference in evidence credibility value between adjacent state evidence units, and the evidence stability is used as the gate control judgment input for the state propagation process.

6. The method for constructing a patient profile based on artificial intelligence according to claim 5, characterized in that, In step S2, the data processing server forms a local state group based on consistency verification and determines the dissimilarity of conflicting state evidence units according to the unified state representation values ​​within the local state group. Subsequently, the evidence stability and dissimilarity are input into the monotone Bayesian gate control model. The monotone Bayesian gate control model is a constrained probabilistic graphical model for determining the risk of evidence propagation in the patient state evolution graph, including evidence stability observation nodes, dissimilarity observation nodes, propagation risk state nodes, and gate control action nodes. The data processing server writes the stability of the proof into the stability observation node and the dissimilarity into the dissimilarity observation node. Through the conditional probability relationship corrected by monotonic constraints, it ensures that the propagation risk confidence value does not decrease as the stability of the proof decreases or the dissimilarity increases. The propagation risk confidence value is output by the propagation risk state node, and the gate control action node outputs the retention, partial restriction, or blocking action. The data processing server performs retention, partial restriction, and blocking processing on the state propagation path based on the propagation risk confidence value.

7. The method for constructing a patient profile based on artificial intelligence according to claim 1, characterized in that, In step S3, the data processing server reads the state transmission information and conflict marker information from the patient's state representation, and establishes topic merging candidate edges based on the subordinate relationships, adjacency relationships, common management goals, and state propagation paths in the state transmission information between clinical semantic anchors. Subsequently, a constrained connected component aggregation algorithm is used to determine the portrait topic nodes. This algorithm uses state evidence units as vertices, topic merging candidate edges as connections, and clinical semantic relationship constraints and temporal continuity constraints to obtain the set of mergeable state evidence units. This algorithm only considers locally bounded elements marked as retained or valid within the current portrait topic node range in the state transmission information. Control paths are treated as connected edges, while blocking paths are not. Only the corresponding conflict state evidence units are attached as conflict marker information. Each candidate connected component that satisfies the clinical semantic relationship constraint and the temporal continuity relationship constraint corresponds to a portrait topic node, and a topic state, topic confidence state, and topic evidence chain are generated within the portrait topic node. Among them, the clinical semantic relationship constraint is that there is at least one clinical semantic relationship among the subordinate relationship, adjacent relationship, or common management goal between the clinical semantic anchors corresponding to the two state evidence units. The temporal continuity relationship constraint is that the two state evidence units are temporally connected, intersecting, or can form a continuous state description according to the temporal evolution relationship.

8. The method for constructing a patient profile based on artificial intelligence according to claim 7, characterized in that, In step S4, the data processing server reads the portrait topic nodes, topic status, topic confidence status, and topic evidence chain from the standardized patient profile, and matches the portrait interpretation rules in the order of priority for portrait topic nodes, secondly for clinical semantic anchors, and supplemented by common management goals, to generate the portrait meaning interpretation and evidence reliability interpretation corresponding to the portrait topic nodes.

9. The method for constructing a patient profile based on artificial intelligence according to claim 8, characterized in that, In step S4, the data processing server organizes the main sources of evidence according to the occurrence sequence of the state evidence units in the subject evidence chain, and generates conflict evidence prompts based on conflict marker information. Then, a traceable path is formed according to the hierarchy of portrait subject node, subject status, subject evidence chain, state evidence unit and original data index, and the traceable path, subject evidence chain, subject confidence status and conflict marker information are associated with the portrait interpretation result.

Images