A method and system for medical text quality assessment based on natural language processing

Abstract

This application provides a method and system for medical text quality assessment based on natural language processing, relating to the field of intelligent quality control technology for medical texts. The method first acquires the medical record text of the target patient and performs medical semantic parsing to extract actual clinical events and their timestamps. Then, it constructs an actual treatment diagram according to the time progression of the timestamps. Next, it loads a matching standard treatment diagram from a clinical knowledge base and performs event coverage verification to separate uncovered missing clinical events. Then, it performs temporal dependency review along the node connections of the actual treatment diagram to identify contradictory temporal connections that violate the evolution logic of the standard treatment diagram. Finally, it summarizes the missing clinical events and contradictory temporal connections to generate a document defect log. The technical solution of this application not only improves the automation level and accuracy of medical text quality assessment but also reduces the burden of manual review and the risk of missed reviews.

Description

Technical Field

[0001] This application relates to the field of intelligent quality control technology for medical texts, and in particular to a method and system for medical text quality assessment based on natural language processing. Background Technology

[0002] With the continuous advancement of medical informatization, electronic medical records have become the core carrier for medical institutions at all levels to record the diagnosis and treatment process, transmit medical information, and support medical quality management. The clinical events recorded in the medical record and their chronological order are directly related to the rationality of diagnostic reasoning, the correctness of treatment decisions, and the accuracy of subsequent medical quality evaluation. In the medical record content quality management system, whether the medical record fully records key diagnostic and treatment events and whether adjacent diagnostic and treatment events conform to the chronological order required by clinical practice have become two key dimensions for measuring the content quality of medical records.

[0003] Existing solutions primarily assess the quality of medical record texts through two methods: expert manual review and simple verification based on keyword matching. First, medical record quality control personnel or clinical experts read each medical record according to the department's quality control guidelines, relying on their personal experience to determine if there are any omissions of key events or inversions in the order of treatment. Second, some hospital information systems use preset keyword lists to match medical records, such as searching for fields like "preoperative discussion" and "informed consent," and then provide a coarse-grained compliance score. Other solutions incorporate rule bases, using the presence of certain specific sections in the medical record as a criterion for determining structural compliance.

[0004] However, existing solutions still suffer from the problem of balancing review efficiency and review depth. First, expert manual review heavily relies on the reviewers' clinical experience and diligence, making it difficult to guarantee the consistency and repeatability of review results. Furthermore, it faces a severe manpower bottleneck when dealing with medical institutions with a large volume of medical records. Second, keyword matching methods can only determine whether certain literal fields appear in the medical record, but cannot identify whether the actual clinical events corresponding to those fields have been completed, let alone determine whether the chronological order of different clinical events is reasonable. Third, rule bases based on chapter compliance only focus on the formal structure of the medical record, without performing semantic analysis on the diagnosis and treatment process recorded within the medical record. This makes it difficult to discover logical contradictions hidden in the text details, such as "prescribing medication before physical examination" or "preoperative discussion only after surgery." Fourth, traditional methods do not structurally compare the actual diagnosis and treatment process recorded in the medical record with the recognized standard diagnosis and treatment path for the disease. This makes it difficult to systematically expose the missing or chronologically disordered key diagnosis and treatment events in the medical record, resulting in important quality risks being concealed under large sections of text. Summary of the Invention

[0005] This application provides a medical text quality assessment method and system based on natural language processing to solve problems in the prior art, such as the omission or under-detection of key clinical events, difficulty in detecting reversed treatment sequences, poor consistency of review results, and the concealment of quality risks caused by the reliance on experience for manual review, keyword matching only recognizing literal fields, chapter compliance rules ignoring semantic connotations, and the failure to conduct structured comparison of actual diagnosis and treatment processes with standard diagnosis and treatment pathways.

[0006] Firstly, this application provides a method for assessing the quality of medical text based on natural language processing, including: Obtain the medical record text of the target patient to be reviewed, perform medical semantic parsing on the medical record text to be reviewed, and extract multiple actual clinical events recorded in the medical record text to be reviewed and the actual timestamp corresponding to each actual clinical event; By sequentially associating the multiple actual clinical events according to the time progression direction of the actual timestamps, an actual diagnosis and treatment graph is constructed. Load a standard diagnostic and treatment diagram matching the disease specialty type of the medical record text to be reviewed from a pre-set clinical knowledge base, perform an event coverage check for the multiple actual clinical events within the standard diagnostic and treatment diagram, and separate out the missing clinical events that are included in the standard diagnostic and treatment diagram but not covered by the actual diagnostic and treatment diagram. A temporal dependency review is performed along the node connection path of the actual diagnosis and treatment diagram to identify contradictory temporal connections within the actual diagnosis and treatment diagram whose sequential order violates the evolution logic of the standard diagnosis and treatment diagram. By summarizing the missing clinical events and connecting the contradictory timelines, a document defect log is generated for the medical record text to be reviewed.

[0007] Optionally, the medical record text of the target patient is obtained, and medical semantic parsing is performed on the medical record text to extract multiple actual clinical events recorded in the medical record text and the actual timestamp corresponding to each actual clinical event. The medical record text to be reviewed was divided into multiple basic segments according to the boundaries of punctuation marks; The system traverses and scans the pre-set medical terminology within the multiple basic segments to extract the medical behavior entities and medical object entities that constitute the actual clinical event. Based on the grammatical structure within the basic segment, establish the action-receiving relationship between the medical behavior entity and the medical object entity; Extract the temporal modifiers that are immediately adjacent to the action-patient relationship within the basic sentence segment; The time-modifying affixes are mapped to absolute time values ​​as the actual timestamps, and the action-actor association is established as the actual clinical event.

[0008] Optionally, the multiple actual clinical events are sequentially associated according to the time progression direction of the actual timestamps to construct an actual diagnosis and treatment map, including: Extract the actual timestamps corresponding to the actual clinical events and sort them in ascending order of values ​​to determine the sequential arrangement sequence of the actual clinical events; The actual clinical events are transformed into discrete event nodes in the actual diagnostic and treatment graph; Based on the adjacency relationship within the sequential arrangement sequence, establish connecting line segments with unidirectional pointing attributes between the discrete event nodes; Connecting the end-to-end connecting line segments with unidirectional pointing properties forms a continuous evolution path connecting the discrete event nodes; The continuous evolution path and the multiple discrete event nodes are combined to form an actual diagnostic and treatment map.

[0009] Optionally, a standard diagnostic chart matching the disease specialty type of the medical record text to be reviewed is loaded from a pre-set clinical knowledge base. Within the standard diagnostic chart, an event coverage check is performed on the multiple actual clinical events to separate missing clinical events included in the standard diagnostic chart but not covered by the actual diagnostic chart, including: Extract the medical behavior entities and medical object entities contained in the actual clinical events as the actual core elements; Analyze the standard event nodes contained within the standard diagnostic and treatment diagram, and extract the standard core elements corresponding to the standard event nodes; Perform a string comparison of each element between the actual core element and the standard core element; In the standard diagnostic diagram, standard event nodes that do not correspond to the actual core element with the same string are screened out as isolated standard nodes; The isolated standard node was identified as a missing clinical event.

[0010] Optionally, in the standard diagnostic graph, standard event nodes that do not correspond to the actual core element with the same string are screened out as isolated standard nodes, including: A set of candidate isolated nodes is formed by aggregating standard event nodes that do not correspond to the actual core elements with the same string. Retrieve the medical exemption trigger words that each candidate isolated node in the candidate isolated node set is bound to in the clinical knowledge base; Scan the medical exemption trigger words within the medical record text to be reviewed; When a scan hits, extract patient refusal entities or physiological contraindication entities that are immediately adjacent to the medical exemption trigger words; Remove nodes associated with the patient's refusal entity or the physiological contraindication entity from the candidate isolated node set, and establish the remaining standard event nodes after the removal operation as isolated standard nodes.

[0011] Optionally, a temporal dependency review is performed along the node connections of the actual diagnostic and treatment diagram to identify contradictory temporal connections within the actual diagnostic and treatment diagram whose sequential order violates the evolution logic of the standard diagnostic and treatment diagram, including: Extract the actual transition line between any two adjacent actual clinical events within the actual diagnostic and treatment diagram; Determine the subsequent clinical events to which the actual transition line points and the preceding clinical events that deviate from it; Locate the standard subsequent node matching the subsequent clinical event and the standard preceding node matching the preceding clinical event within the standard diagnostic and treatment map; Detect the direction of the standard connection line between the standard preceding node and the standard subsequent node within the standard diagnostic diagram; When the direction of the standard connection is opposite to that of the actual transition connection, the actual transition connection is locked as a contradictory timing connection.

[0012] Optionally, the missing clinical events and the contradictory timelines are summarized to generate a document defect log for the medical record text under review, including: Retrieve the disease specialty type corresponding to the missing clinical event in the standard diagnostic and treatment diagram; Retrieve the actual timestamps of the two actual clinical events connected at both ends of the actual diagnostic and treatment diagram by the contradictory timeline connection; The disease specialty type and the missing clinical event are combined and packaged to generate a missing item prompt; The actual timestamp and the contradictory time sequence are connected, merged, and encapsulated to generate a logical conflict entry; The omission prompts and logical conflict entries are vertically concatenated to form a document defect log.

[0013] Secondly, this application provides a medical text quality assessment system based on natural language processing, including: The acquisition module is used to acquire the medical record text of the target patient to be reviewed, perform medical semantic parsing on the medical record text to be reviewed, and extract multiple actual clinical events recorded in the medical record text to be reviewed and the actual timestamp corresponding to each actual clinical event. The construction module is used to sequentially associate the multiple actual clinical events according to the time progression direction of the actual timestamps to construct an actual diagnosis and treatment graph; The analysis module is used to load a standard diagnostic and treatment diagram that matches the disease specialty type of the medical record text to be reviewed from a preset clinical knowledge base, perform an event coverage check for the multiple actual clinical events within the standard diagnostic and treatment diagram, and separate out the missing clinical events that are included in the standard diagnostic and treatment diagram but not covered by the actual diagnostic and treatment diagram. The locking module is used to perform temporal dependency review along the node connection path of the actual diagnosis and treatment diagram, and lock contradictory temporal connections within the actual diagnosis and treatment diagram whose sequential order violates the evolution logic of the standard diagnosis and treatment diagram; The generation module is used to summarize the missing clinical events and the contradictory timelines to generate a document defect log for the medical record text to be reviewed.

[0014] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a medical text quality assessment method based on natural language processing as described in the first aspect above.

[0015] Fourthly, this application provides a computer storage medium storing a computer program, which, when executed by a computer, implements a medical text quality assessment method based on natural language processing as described in the first aspect.

[0016] This application extracts actual clinical events and timestamps by performing medical semantic parsing on the medical record text under review, and constructs an actual diagnosis and treatment graph based on this. This transforms the original unstructured natural language form of the diagnosis and treatment records into a computable graph structure, overcoming the limitation of traditional keyword matching, which can only identify literal fields and cannot understand the semantics of events. By loading a standard diagnosis and treatment graph that matches the disease specialty type of the medical record text under review, and performing event coverage verification within the standard diagnosis and treatment graph, it can systematically identify missing clinical events required by the standard diagnosis and treatment graph but not covered by the actual diagnosis and treatment graph, fundamentally solving the problem of the difficulty in exposing the omission of key diagnosis and treatment events.

[0017] Furthermore, by conducting temporal dependency reviews along the node connections in the actual treatment diagram and comparing the direction of the actual transitional connections with the direction of the standard connections between corresponding nodes in the standard treatment diagram, contradictory temporal connections that violate the logic of clinical evolution can be accurately identified, avoiding oversights in manual review or missing keyword matching that could lead to the neglect of reversed treatment order. By summarizing and encapsulating missing clinical events and contradictory temporal connections, a document defect log with structured entries is generated, enabling medical record quality control personnel to directly locate the location and type of each defect, thus improving the objectivity, consistency, and efficiency of medical text quality assessment as a whole.

[0018] These or other aspects of this application will become more apparent from the description of the following embodiments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart of a medical text quality assessment method based on natural language processing provided in this application is shown; Figure 2 A schematic diagram of the medical semantic parsing process provided in this application is shown; Figure 3 This application provides a schematic diagram of the structure of both the actual and standard diagnostic and treatment diagrams. Figure 4 This paper illustrates the logical diagram of the event coverage verification and timing dependency review provided in this application; Figure 5 This paper presents a schematic diagram of the structure of a medical text quality assessment system based on natural language processing provided in this application. Figure 6 A schematic diagram of the structure of a computing device provided in this application is shown. Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0022] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a chronological order, nor do they limit "first" and "second" to different types.

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] Figure 1 A flowchart of a medical text quality assessment method based on natural language processing is provided for embodiments of this application, as shown below. Figure 1 As shown, the method includes: Step 101: Obtain the medical record text of the target patient to be reviewed, perform medical semantic parsing on the medical record text to be reviewed, and extract multiple actual clinical events recorded in the medical record text to be reviewed and the actual timestamp corresponding to each actual clinical event.

[0025] In this step, the medical record text to be reviewed refers to the document written by a clinician for a target patient during a complete outpatient visit or hospitalization, recorded in natural language, including but not limited to components such as admission records, progress notes, preoperative discussion records, surgical records, and doctor's orders; medical semantic parsing refers to a natural language processing process designed specifically for the characteristics of medical texts, aiming to extract clinical events and their temporal attributes; actual clinical events refer to the medical acts with medical meaning actually recorded in the medical record text for the target patient, such as performing an examination on the patient, prescribing a certain medication, or performing a certain surgery on the patient; actual timestamp refers to the specific moment when the actual clinical event actually occurred, expressed in absolute time values ​​such as year, month, day, hour, and minute.

[0026] In this embodiment, firstly, the interface module of the hospital information system initiates a search request to the medical record database, retrieving all medical record texts corresponding to the patient's latest treatment based on the target patient's hospitalization number or visitation number, and storing them in a local buffer. Secondly, medical semantic parsing is performed on the medical record text to be reviewed, sequentially completing sub-steps such as basic segmentation, medical entity extraction, action agent-patient association establishment, time modifier affix extraction, and time value mapping. Finally, an event sequence consisting of several actual clinical events and their corresponding timestamps is output. The entire process is uniformly completed by a medical natural language processing engine deployed on the server side, and the output results are written to the actual event sequence buffer in the form of structured fields.

[0027] Specifically, the medical semantic parsing process is as follows: Figure 2 As shown. Figure 2As shown, the medical record text 2-1 to be reviewed is first split into multiple basic segments by the segment segmentation module 2-2, then the medical entity extraction module 2-3 scans the medical behavior entities and medical object entities in it, then the action agent-patient association module 2-4 establishes the action agent-patient association between the two types of entities based on the grammatical structure, and finally the time mapping module 2-5 extracts and maps time modifiers and affixes, outputs several actual clinical events and their actual timestamps, and writes them into the actual event sequence 2-6.

[0028] For example, in a medical record quality control scenario at a tertiary general hospital, the quality control system extracts a pending inpatient medical record from the respiratory medicine department's information system. The text contains descriptions such as "A chest CT scan was performed on the patient at 8:00 AM on January 10, 2026," "Levofloxacin was prescribed for intravenous infusion on the patient at 2:00 PM on January 10, 2026," and "A complete blood count was repeated at 9:00 AM on January 12, 2026." After parsing the medical record text, the medical semantic parsing engine sequentially outputs the actual clinical events "chest CT scan," "levofloxacin prescription," and "complete blood count repeat," along with their corresponding timestamps "2026-01-10 08:00," "2026-01-10 14:00," and "2026-01-12 09:00," and writes them into the actual event sequence buffer, preparing for the subsequent construction of the actual diagnosis and treatment graph.

[0029] Step 102: Construct an actual diagnosis and treatment map by sequentially associating the multiple actual clinical events according to the time progression direction of the actual timestamps.

[0030] In this step, the actual treatment graph refers to a directed graph data structure with actual clinical events as nodes and the temporal sequence between events as directed edges, used to depict the actual treatment process experienced by the patient in a computable form; the time progression direction refers to the direction from the earlier actual timestamp to the later actual timestamp.

[0031] The reason why it is necessary to construct the actual diagnosis and treatment graph after obtaining the actual event sequence in step 101 is that a simple event list can only reflect the events themselves and cannot directly express the temporal dependencies between events. Subsequent event coverage verification and temporal dependency review both require graph structure as input. Therefore, the actual event sequence must be upgraded to an actual diagnosis and treatment graph to support subsequent graph comparison operations.

[0032] In this embodiment, firstly, all actual clinical events and their corresponding timestamps are read from the actual event sequence buffer, and each actual clinical event is transformed into a discrete event node in the actual diagnosis and treatment map. Secondly, all discrete event nodes are sorted according to the increasing direction of the actual timestamp values ​​to determine the sequential arrangement sequence of actual clinical events. Thirdly, according to the adjacency relationship within the sequential arrangement sequence, a connecting line segment with a unidirectional pointing attribute is established between two adjacent discrete event nodes, wherein the unidirectional pointing attribute direction is from the earlier event to the later event. Finally, all connecting line segments with unidirectional pointing attributes are connected end to end to form a continuous evolution path connecting the discrete event nodes, and this continuous evolution path is combined with all discrete event nodes to form the actual diagnosis and treatment map.

[0033] Following the implementation case of step 101, the quality control system transforms the actual event sequence "chest CT scan", "prescription of levofloxacin", and "re-examination of blood routine" into discrete event nodes N1, N2, and N3 respectively. These nodes are then sorted from earliest to latest according to their actual timestamps to obtain the sequential arrangement sequence N1→N2→N3. A one-way connection segment e12 is then established between N1 and N2, and a one-way connection segment e23 is established between N2 and N3. Finally, the node set {N1, N2, N3} and the continuous evolution path N1→N2→N3 are combined to form the actual diagnosis and treatment map for this patient.

[0034] Step 103: Load a standard diagnostic and treatment diagram matching the disease specialty type of the medical record text to be reviewed from a preset clinical knowledge base, perform an event coverage check for the multiple actual clinical events within the standard diagnostic and treatment diagram, and separate out the missing clinical events that are included in the standard diagnostic and treatment diagram but not covered by the actual diagnostic and treatment diagram.

[0035] In this step, the clinical knowledge base refers to an electronic knowledge resource database pre-established by the hospital or industry organization, organized by disease specialty type. Internally, it stores standard treatment diagrams corresponding to each disease specialty type in a graph structure. A standard treatment diagram is a directed graph, based on clinical guidelines, expert consensus, and summarizing past high-quality cases, representing the key treatment events that should be included in the disease specialty and the order in which these events should be followed. Specifically, the construction process of the standard treatment diagram includes: retrieving the clinical pathway text for a specific disease, extracting "must-execute" treatment nodes using a pre-defined predicate logic template, and establishing directed edges between nodes according to the sequential logic in the specifications. Disease specialty type refers to disease classification labels such as respiratory medicine (pulmonary infection), cardiovascular medicine (coronary heart disease), and general surgery (appendicitis). Missing clinical events refer to key clinical events included in the standard treatment diagram but not covered by the actual treatment diagram, which should appear in the medical record.

[0036] The reason for comparing the actual medical records with the standard medical records for event coverage is that the actual medical records can only reproduce what has already happened and cannot find the steps that "should have been done but were not". Only by introducing a standard medical record that matches the disease specialty of the medical record as a reference baseline can the key medical events that have been missed in the medical record be systematically exposed.

[0037] In this embodiment, firstly, the quality control system determines the disease specialty type of the medical record based on the diagnosis field on the first page of the medical record text to be reviewed, and loads a standard diagnosis and treatment diagram matching the disease specialty type from the clinical knowledge base; secondly, it extracts the medical behavior entity and medical object entity corresponding to each discrete event node in the actual diagnosis and treatment diagram as the actual core element; thirdly, it parses all the standard event nodes covered in the standard diagnosis and treatment diagram, and extracts the medical behavior entity and medical object entity corresponding to each standard event node as the standard core element; then, it performs a string comparison between the actual core element and the standard core element; finally, it screens out the standard event nodes in the standard diagnosis and treatment diagram that do not have the same string corresponding to the actual core element as isolated standard nodes, and establishes all isolated standard nodes as missing clinical events.

[0038] Furthermore, to avoid incorrectly classifying missing events due to reasonable omissions of certain events for individual patient reasons, this embodiment adds a medical exemption determination step. Specifically, firstly, standard event nodes that do not correspond to the same actual core elements are collected to form a candidate isolated node set; secondly, the medical exemption trigger words bound to each candidate isolated node in the clinical knowledge base are retrieved. These medical exemption trigger words can be set as fixed words such as "refusal to sign," "drug allergy," and "abnormal liver function," indicating that the patient cannot or should not receive the treatment; thirdly, the medical exemption trigger words are scanned within the medical record text to be reviewed; then, when a scan hits, the patient refusal entity or physiological contraindication entity adjacent to the medical exemption trigger word is extracted; finally, nodes associated with the patient refusal entity or physiological contraindication entity are removed from the candidate isolated node set, and the remaining standard event nodes after the removal operation are established as isolated standard nodes, thereby improving the accuracy of missing event determination.

[0039] Specifically, the diagram of the event coverage check is shown below. Figure 3 As shown. Figure 3 As shown, standard diagnosis and treatment Figure 3-1 It contains standard event nodes S1, S2, S3, and S4 connected in a predetermined evolutionary order, and the actual diagnosis and treatment Figure 3-2 It contains discrete event nodes N1, N2, and N3, which will be used in actual diagnosis and treatment. Figure 3-2 The actual core elements and standard diagnosis and treatment Figure 3-1After comparing the standard core elements of each standard event node one by one, it was found that standard event node S3 did not match any actual event node with the same actual core elements. Therefore, S3 was sent to the medical exemption determination stage as a candidate isolated node. If the medical exemption trigger words corresponding to the candidate isolated node did not match the corresponding exemption entity in the medical record text to be reviewed, then S3 was finally output as a missing clinical event to the defect summary buffer.

[0040] Following the implementation case of step 102, the disease specialty of the medical record text to be reviewed is respiratory pulmonary infection. The quality control system loads the standard diagnosis and treatment diagram corresponding to this specialty from the clinical knowledge base, which includes standard event nodes such as "chest CT examination", "pathogen submission", "antibiotic use", and "efficacy review". After comparing the actual core elements of the actual diagnosis and treatment diagram with the standard core elements of each standard event node one by one, it was found that the standard event node "pathogen submission" has no corresponding node with the same name in the actual diagnosis and treatment diagram and is listed as a candidate isolated node. The system continues to scan the medical exemption trigger words "family refusal" and "critical condition making it difficult to collect" bound to "pathogen submission" in the medical record text, but no corresponding exemption entity is found. Therefore, "pathogen submission" is finally identified as a missing clinical event and output.

[0041] Step 104: Perform a temporal dependency review along the node connection path of the actual diagnosis and treatment diagram to identify contradictory temporal connections within the actual diagnosis and treatment diagram that violate the evolution logic of the standard diagnosis and treatment diagram.

[0042] In this step, the temporal dependency review refers to the process of checking the directional compliance of each transition line between adjacent events in the actual medical record based on the objective sequential constraints between events in the standard medical record; contradictory temporal lines refer to those transition lines in the actual medical record whose directions are opposite to those between the same-name nodes specified in the standard medical record. The existence of contradictory temporal lines means that the order of diagnosis and treatment recorded in the medical record violates the logic of clinical evolution.

[0043] The reason why a time-dependent review is still needed after the missing clinical events have been obtained in step 103 is that even if all key clinical events have been recorded, there may still be quality defects due to the reversed order of diagnosis and treatment. For example, the situation of issuing postoperative medical orders before writing preoperative discussions cannot be identified by event coverage verification alone. It must be exposed by comparing the direction of the connection.

[0044] In this embodiment, firstly, the actual transition line between any two adjacent actual clinical events within the actual diagnostic diagram is extracted; secondly, the subsequent clinical event and the preceding clinical event that the actual transition line points to are determined; thirdly, the standard subsequent node matching the subsequent clinical event and the standard preceding node matching the preceding clinical event are located within the standard diagnostic diagram; then, the direction of the standard connection between the standard preceding node and the standard subsequent node within the standard diagnostic diagram is detected; finally, when the direction of the standard connection and the direction of the actual transition line are in opposite states, the actual transition line is locked as a contradictory temporal connection. By performing the above comparison on all transition lines in the actual diagnostic diagram, all contradictory temporal connections can be obtained.

[0045] Specifically, the overall logic of event coverage verification and time-series dependency review is as follows: Figure 4 As shown. Figure 4 As shown, the medical record text to be reviewed is processed by the semantic parsing module 4-1 to form an actual event sequence, which is then processed by the actual diagnosis and treatment diagram construction module 4-2 to form an actual diagnosis and treatment diagram. This diagram, along with the standard diagnosis and treatment diagram loaded from the clinical knowledge base 4-3, is then sent to the event coverage verification module 4-4 and the time sequence dependency review module 4-5. The event coverage verification module 4-4 outputs missing clinical events, and the time sequence dependency review module 4-5 outputs contradictory time sequence connections. The two are finally merged into the defect summary and encapsulation module 4-6 to generate a document defect log.

[0046] Following the implementation case of step 103, suppose there are two actual clinical events in the actual diagnosis and treatment diagram: "prescription of levofloxacin" and "pathogen testing". The actual transition line points from "prescription of levofloxacin" to "pathogen testing". After matching the corresponding nodes in the standard diagnosis and treatment diagram, it is found that the standard line should point from "pathogen testing" to "antibiotic use", that is, the direction of the standard line is opposite to the direction of the actual transition line. Based on this, the quality control system locks the actual transition line as a contradictory timing line and outputs it to the defect summary buffer, thereby exposing the timing error of "medication first, then testing" in the medical record.

[0047] Step 105: Summarize the missing clinical events and the contradictory timelines to generate a document defect log for the medical record text to be reviewed.

[0048] In this step, the document defect log refers to a report document that records all quality defects in the medical record text to be reviewed in the form of structured entries. Each entry contains at least key information such as defect type, defect content, and the location or time of defect occurrence, which facilitates subsequent reading, classification and processing by medical record quality control personnel.

[0049] The reason for summarizing and encapsulating missing clinical events and contradictory timelines into a document defect log is that outputting a single type of problem makes it difficult for quality control personnel to form a panoramic understanding of the overall quality of the medical record. Presenting omissions and timelines in a log format allows quality control personnel to see the distribution of all defects in the medical record in the same output.

[0050] In this embodiment, firstly, the disease specialty type corresponding to the missing clinical event in the standard treatment diagram is retrieved; secondly, the actual timestamps of the two actual clinical events connected by the contradictory timeline at both ends of the actual treatment diagram are retrieved; thirdly, the disease specialty type and the missing clinical event are merged and encapsulated to generate omission prompt entries, each omission prompt entry containing at least the fields "disease specialty type" and "missing event name"; then, the actual timestamps and the contradictory timeline are merged and encapsulated to generate logical conflict entries, each logical conflict entry containing at least the fields "preceding event name", "subsequent event name", "preceding event actual timestamp", and "subsequent event actual timestamp"; finally, all omission prompt entries and all logical conflict entries are vertically spliced ​​according to a preset layout order to assemble a document defect log, and output to the medical record quality control terminal in the form of a structured file.

[0051] Following the implementation case of step 104, the quality control system merges the missing clinical event "Pathogen Submission" with its disease specialty type "Respiratory Medicine Lung Infection" into a missing item "[Omission] Respiratory Medicine Lung Infection - Pathogen Submission Not Recorded"; it merges the contradictory timeline "Prescription of Levofloxacin → Pathogen Submission" and its two ends' actual timestamps into a logical conflict item "[Timeline] 2026-01-10 14:00 Levofloxacin Prescription earlier than 2026-01-11 09:00 Pathogen Submission"; finally, the above two items are vertically spliced ​​together to assemble the document defect log of the medical record, which is output to the medical record quality control terminal for subsequent review and processing by quality control personnel.

[0052] In summary, this application achieves a core breakthrough by sequentially performing medical semantic parsing, constructing actual diagnostic diagrams, verifying event coverage against standard diagnostic diagrams, reviewing the temporal dependencies of node connection directions, and summarizing and encapsulating defects in the medical record text under review. This constitutes a complete intelligent quality control chain from text to graph, and then from graph to structured quality defect logs. In the context of medical quality control where the total number of medical records continues to grow, each medical record can be objectively and consistently assessed without relying on manual experience to read it sentence by sentence. This completely eliminates the limitations of traditional keyword matching, which only looks at the literal meaning, and chapter compliance rules, which ignore semantics. Simultaneously, it can accurately identify two types of problems: omission of key diagnosis and treatment events and reversal of the order of diagnosis and treatment. This enables medical institutions to obtain a deeper and more stable quality management capability for the content of medical records at a lower human cost.

[0053] To address the problem in existing technologies where keyword matching alone is insufficient for reliably extracting clinical events and time attributes from natural language medical record texts, some embodiments, according to step 101, perform medical semantic parsing on the medical record text to be reviewed, extracting multiple actual clinical events recorded in the medical record text and the actual timestamp corresponding to each actual clinical event, including: Step 201: Divide the medical record text to be reviewed into multiple basic segments according to the boundaries of punctuation marks.

[0054] In this step, punctuation mark boundaries refer to characters with strong semantic segmentation functions, such as periods, semicolons, exclamation marks, question marks, and line breaks in common Chinese punctuation marks; basic segments refer to continuous character sequences that are sandwiched between two adjacent punctuation mark boundaries and contain relatively complete semantic groups.

[0055] In this embodiment, the medical record text to be reviewed is first read into a string buffer in character order; then, the string is scanned and segmented according to a preset list of punctuation boundary characters; next, the substrings between adjacent boundaries obtained from the scan are written as a basic segment into a segment queue; finally, several basic segments arranged in the original text order are output. This process is implemented using a rule-based finite state machine algorithm, supporting the segmentation of various boundary combinations such as "XXX,XXX." and "XXX;XXX.".

[0056] Step 202: Traverse and scan the preset medical professional terms in the multiple basic segments to extract the medical behavior entities and medical object entities that constitute the actual clinical event.

[0057] In this step, medical terminology refers to a collection of words with medical meanings compiled in advance from medical terminology dictionaries and the hospital's unique terminology; medical action entities refer to words that characterize medical actions, such as "examination," "administration," "surgery," and "discussion"; and medical object entities refer to words that characterize the object to which the medical action is performed, such as "chest CT," "levofloxacin," "appendectomy," and "preoperative assessment."

[0058] In this embodiment, the basic text segments are first read in one by one as the smallest processing unit. Then, based on a double-array Trie tree structure, efficient multi-pattern string matching is performed on the pre-set medical professional terms, scanning all matching positions in each segment. Next, the matched terms are categorized into medical behavior entities or medical object entities based on their category. Finally, the set of medical behavior entities and the set of medical object entities corresponding to each basic text segment are output. The double-array Trie tree structure used is a variant of a dictionary tree that uses two arrays to implement node navigation, featuring fast search speed and low memory usage.

[0059] Step 203: Based on the grammatical structure within the basic segment, establish the action-receiving relationship between the medical behavior entity and the medical object entity.

[0060] In this step, the action agent-receiver relationship refers to the "action-object" binary relationship formed by the medical behavior entity as the agent and the medical object entity as the recipient, which essentially corresponds to the verb-object structure in Chinese sentences.

[0061] In this embodiment, firstly, Chinese word segmentation and part-of-speech tagging based on the Conditional Random Field (CRF) algorithm are performed on each basic text segment to obtain a word sequence with part-of-speech tags. Secondly, the word sequence is parsed using a dependency parser to obtain a dependency tree representing the dependency relationships between words. Thirdly, the shortest path from the medical behavior entity to the medical object entity is searched along action-receptor type dependency edges such as "verb-object relationship" and "prepositional object relationship" in the dependency tree. Then, the medical behavior entity and the medical object entity connected by the shortest path are bound into an action-application-receive association. Finally, all action-application-receive associations corresponding to the basic text segment are output. The CRF algorithm uses a parametric model pre-trained on a medical NER (Named Entity Recognition) dataset. The feature templates include: medical affix features, contextual word segmentation features, and medical record chapter position features to improve the recognition accuracy in unstructured medical text. The dependency parser can be, for example, a neural dependency parser based on a transfer system, which can be specifically set according to actual engineering performance requirements.

[0062] Step 204: Extract the temporal modifiers that are immediately adjacent to the action-actor relationship within the basic segment.

[0063] In this step, time-modifying affixes refer to word segments used to describe the time attribute of the action before and after the action-patient relationship, such as "8:00 AM on January 10, 2026", "the day after surgery", "the day of admission", etc.

[0064] In this embodiment, firstly, a search window of fixed character length is defined forward and backward within the basic segment around the established action-actor relationship; secondly, string matching is performed within the search window based on a preset set of time expression regular expressions, which covers various types such as absolute time, relative time, and date phrases; thirdly, the matched time expressions are extracted as time modifier affixes; finally, the time modifier affixes are bound to the corresponding action-actor relationship, awaiting subsequent mapping processing.

[0065] Step 205: Map the time modifiers to absolute time values ​​as the actual timestamps, and establish the action-actor association as the actual clinical event.

[0066] In this step, the absolute time value refers to the precise time value in terms of year, month, day, hour, minute, and second, which facilitates subsequent sorting and comparison by time value.

[0067] In this embodiment, it is first determined whether the time modifier belongs to an absolute time expression or a relative time expression; secondly, for absolute time expressions, the field is directly parsed and mapped to an absolute time value according to the preset time format rules; thirdly, for relative time expressions, the known admission date or the previous known absolute time in the medical record text is used as an anchor point for addition and subtraction operations, and finally mapped to an absolute time value; then, the absolute time value is used as the actual timestamp of the action-response association; finally, the action-response association with the actual timestamp is output as an actual clinical event to the actual event sequence buffer.

[0068] In some embodiments, as described in step 102, the multiple actual clinical events are sequentially associated according to the time progression direction of the actual timestamps to construct an actual diagnosis and treatment map, including: Step 301: Extract the actual timestamps corresponding to the actual clinical events and sort them in ascending order to determine the sequential arrangement of the actual clinical events.

[0069] In this embodiment, all actual clinical events and their corresponding timestamps are first read from the actual event sequence buffer; then, each actual clinical event and its actual timestamp are combined into a key-value pair, and sorted in ascending order using the actual timestamp as the key; the sorting operation uses a stable sorting algorithm to ensure that events with the same timestamp maintain their original order; finally, a sequentially arranged sequence in ascending order of time values ​​is output.

[0070] Step 302: Transform the actual clinical events into discrete event nodes of the actual diagnosis and treatment graph.

[0071] In this embodiment, firstly, a globally unique node number is assigned to each actual clinical event; secondly, the actual clinical event name, medical behavior entity, medical object entity, and actual timestamp are written as node attributes into the corresponding discrete event node; thirdly, all discrete event nodes are registered in the node table; and finally, node object references are provided for the creation of subsequent connecting segments.

[0072] Step 303: According to the adjacent relationship within the sequential arrangement sequence, establish connecting line segments with unidirectional pointing attributes between the discrete event nodes.

[0073] In this step, the unidirectional pointing attribute means that the connecting line segment can only point from one end to the other, and cannot be reached in the opposite direction. It is used to express the irreversible progression of clinical events in time.

[0074] In this embodiment, firstly, all adjacent event pairs are extracted from the sequential arrangement sequence with a sliding window length of two; secondly, for each adjacent event pair, the earlier discrete event node is taken as the starting point and the later discrete event node is taken as the ending point; thirdly, a one-way connection segment from the starting point to the ending point is created based on the directed graph data structure; finally, all one-way connection segments are registered in the edge table.

[0075] Step 304: Connect the end-to-end connecting line segments with unidirectional pointing properties to form a continuous evolution path connecting the discrete event nodes.

[0076] In this embodiment, the edge table is first traversed in chronological order, and each unidirectional connecting line segment is added to the evolution path queue. Then, it is checked whether the starting node of the current line segment is consistent with the ending node of the previous line segment to ensure that the beginning and end are connected. Next, all the line segments that are connected are connected in sequence. Finally, a continuous evolution path that runs through all discrete event nodes is obtained.

[0077] Step 305: Combine the continuous evolution path and the multiple discrete event nodes to form an actual diagnostic and treatment map.

[0078] In this embodiment, firstly, the set of discrete event nodes is used as the vertex set of the graph; secondly, all unidirectional connecting line segments on the continuous evolution path are used as the edge set of the graph; thirdly, the vertex set and edge set are encapsulated into a unified directed graph data structure object; finally, the directed graph is output as the actual diagnosis and treatment graph for the patient and written into the actual diagnosis and treatment graph buffer.

[0079] In some embodiments, according to step 103, a standard diagnostic chart matching the disease specialty type of the medical record text to be reviewed is loaded from a preset clinical knowledge base. Within the standard diagnostic chart, an event coverage check is performed on the multiple actual clinical events to separate missing clinical events included in the standard diagnostic chart but not covered by the actual diagnostic chart, including: Step 401: Extract the medical behavior entities and medical object entities contained in the actual clinical event as the actual core elements.

[0080] In this step, the actual core element refers to the binary combination consisting of the medical behavior entity and the medical object entity, which is the smallest set of elements that characterizes the essential semantics of the actual clinical event.

[0081] In this embodiment, firstly, all discrete event nodes in the actual diagnosis and treatment diagram are traversed; secondly, the medical behavior entity and the medical object entity are read from the attribute fields of each discrete event node; thirdly, the two entities are merged into a tuple as the actual core element of the node; finally, the actual core elements of all nodes are gathered into a set of actual core elements.

[0082] Step 402: Analyze the standard event nodes contained within the standard diagnostic diagram and extract the standard core elements corresponding to the standard event nodes.

[0083] In this embodiment, firstly, a standard diagnosis and treatment diagram matching the disease specialty type of the medical record text to be reviewed is loaded from the clinical knowledge base; secondly, all standard event nodes in the standard diagnosis and treatment diagram are traversed; thirdly, the standard medical behavior entity and standard medical object entity of each standard event node are read from the attribute field of the node and formed into a tuple; finally, all tuples are gathered into a standard core element set.

[0084] Step 403: Perform a string comparison of each element between the actual core element and the standard core element.

[0085] In this embodiment, firstly, each standard core element in the standard core element set is used as the outer loop variable; secondly, each actual core element in the actual core element set is used as the inner loop variable; thirdly, in each iteration of the outer loop, the current standard core element is compared with all actual core elements one by one, and the comparison method can be selected as complete equality comparison or equivalent comparison with synonym normalization; finally, a binary determination result is generated for each standard core element to determine whether a match exists in the actual core element set. The synonym normalization can be, for example, based on the institute's unique terminology lookup table, and can be specifically set according to the actual engineering needs.

[0086] Step 404: Screen out standard event nodes that do not correspond to the actual core element with the same string in the standard diagnostic diagram as isolated standard nodes.

[0087] In this embodiment, firstly, based on the binary judgment result output in step 403, the standard event nodes corresponding to all standard core elements that "do not match" are selected; secondly, these standard event nodes are temporarily stored in the candidate isolated node set; thirdly, the nodes that are reasonably omitted are further screened out in the subsequent medical exemption judgment step; finally, the remaining standard event nodes after removing the exemption nodes are established as isolated standard nodes.

[0088] Step 405: The isolated standard node is identified as a missing clinical event.

[0089] In this embodiment, the set of isolated standard nodes is first traversed; then, a corresponding missing clinical event description object is constructed for each isolated standard node, including fields such as the missing event name, the disease specialty type to which the missing event belongs, and the expected occurrence stage of the missing event in the standard diagnosis and treatment diagram; next, the description object is written into the "omission class" partition of the defect summary buffer; finally, all missing clinical events are output for use by the subsequent log generation module.

[0090] In some embodiments, as described in step 404, screening out standard event nodes that do not correspond to the same actual core element as isolated standard nodes in the standard diagnostic graph includes: Step 501: Gather standard event nodes that do not correspond to the actual core elements with the same string to form a candidate isolated node set.

[0091] In this embodiment, firstly, based on the judgment result of the element string comparison, all unmatched standard event nodes are collected into a temporary container; secondly, a "candidate isolated" status flag is attached to each collected node; thirdly, all nodes in the temporary container are grouped according to disease specialty type and stage; finally, a set of candidate isolated nodes is output.

[0092] Step 502: Retrieve the medical exemption trigger words that each candidate isolated node in the candidate isolated node set is bound to in the clinical knowledge base.

[0093] In this step, medical exemption trigger words refer to a set of fixed words that are pre-bound to each standard event node in the clinical knowledge base, representing that the event may be omitted under reasonable clinical reasons, such as "patient refusal", "history of drug allergy", "child-Pugh liver function grade C", etc.

[0094] In this embodiment, the identifier of the candidate isolated node is first used as the query key to initiate a query to the exemption dictionary submodule of the clinical knowledge base; secondly, the exemption dictionary submodule returns a list of medical exemption trigger words bound to the node; thirdly, the list is associated with the candidate isolated node and cached; and finally, a matching template is provided for subsequent text scanning.

[0095] Step 503: Scan the medical exemption trigger words within the medical record text to be reviewed.

[0096] In this embodiment, an Aho-Corasick multi-pattern matching automaton is first constructed using the list of medical exemption trigger words corresponding to candidate isolated nodes. Then, the automaton is run once using the medical record text to be reviewed as the string to be scanned. Next, the specific words and positions of each hit are recorded. Finally, a list of hit events is output. The Aho-Corasick automaton is a string algorithm that supports efficient matching of multiple pattern strings simultaneously and is commonly used in keyword scanning engineering scenarios.

[0097] Step 504: When the scan hits, extract the patient refusal entity or physiological contraindication entity that is immediately adjacent to the medical exemption trigger word.

[0098] In this step, the patient refusal entity refers to the entity phrase that appears immediately after the medical exemption trigger word and indicates that the patient or their family members have clearly refused a certain treatment, such as "family members sign to refuse" or "the patient does not agree"; the physiological contraindication entity refers to the entity phrase that appears immediately after the medical exemption trigger word and indicates that the patient has a certain contraindication to a certain treatment, such as "severe allergy" or "abnormal coagulation function".

[0099] In this embodiment, firstly, with the hit position of each hit as the center, a fixed character length adjacency window is drawn on the left and right sides of the medical record text to be reviewed; secondly, secondary entity recognition is performed within the adjacency window based on a preset entity dictionary; thirdly, the recognized entities are classified into patient refusal entities or physiological contraindication entities according to their categories; finally, the recognition results are backfilled into the corresponding candidate isolated nodes.

[0100] Step 505: Remove nodes associated with the patient's refusal entity or the physiological contraindication entity from the candidate isolated node set, and establish the remaining standard event nodes after the removal operation as isolated standard nodes.

[0101] In this embodiment, the candidate isolated node set is first traversed; then, each node is checked to see if it contains a patient refusal entity or a physiological contraindication entity; next, nodes with backfilling are removed and marked with a "reasonable exemption" flag; finally, the remaining candidate isolated nodes are uniformly marked as isolated standard nodes for output.

[0102] In some embodiments, as described in step 104, a temporal dependency review is performed along the node connections of the actual diagnostic and treatment map to identify contradictory temporal connections within the actual diagnostic and treatment map whose sequential order violates the evolution logic of the standard diagnostic and treatment map, including: Step 601: Extract the actual transition line between any two adjacent actual clinical events within the actual diagnostic and treatment diagram.

[0103] In this step, the actual transition line refers to the connecting line segment with unidirectional pointing attribute that connects two adjacent discrete event nodes in the actual diagnosis and treatment diagram.

[0104] In this embodiment, the edge table is first traversed according to the storage order of the edges, using the actual diagnosis and treatment diagram as the object; then, the starting node reference and ending node reference of each edge are read; the starting and ending point information are sent to the review queue; and finally, all transitional connections are taken as the targets to be reviewed.

[0105] Step 602: Determine the subsequent clinical event to which the actual transition line points and the preceding clinical event to which it deviates.

[0106] In this embodiment, the actual clinical events corresponding to the starting node are first read from the transition connection as the preceding clinical events; the actual clinical events corresponding to the ending node are read as the following clinical events; the actual core elements of the preceding and following clinical events are extracted as keywords for subsequent matching and retrieval; and finally, the pair is sent to the next positioning stage.

[0107] Step 603: Locate the standard subsequent node matching the subsequent clinical event and the standard preceding node matching the preceding clinical event within the standard diagnostic and treatment diagram.

[0108] In this embodiment, firstly, the actual core elements of the preceding clinical event are used as the query key to perform a string matching search in the standard diagnostic and treatment graph node table; secondly, the matched standard event node is used as the standard preceding node; thirdly, the actual core elements of the subsequent clinical event are used as the query key to perform a string matching search in the standard diagnostic and treatment graph node table again; finally, the matched standard event node is used as the standard subsequent node.

[0109] Step 604: Detect the direction of the standard connection between the standard preceding node and the standard subsequent node within the standard diagnostic diagram.

[0110] In this embodiment, firstly, the standard preceding node and the standard following node are used as inputs to search for the edge connecting the two in the edge table of the standard diagnostic graph; secondly, the starting and ending node references of the edge are read to determine the direction of the standard connection; thirdly, if there is no direct connection in the edge table, the reachable direction between the two is determined by breadth-first search on the standard diagnostic graph; finally, the direction identifier is output, and the direction identifier has only two values: "preceding → following" or "following → preceding".

[0111] Step 605: When the direction of the standard connection is opposite to that of the actual transition connection, the actual transition connection is locked as a contradictory timing connection.

[0112] In this embodiment, firstly, the actual transition line direction is compared with the direction identifier output in step 604; secondly, when the two directions are opposite, the edge object corresponding to the actual transition line is marked as a "contradictory time sequence line"; thirdly, the edge object, along with the actual timestamps of its preceding and subsequent events, is written into the "time sequence" partition of the defect summary buffer; finally, the above comparison is performed on all transition lines in the actual diagnosis and treatment diagram, and the complete set of contradictory time sequence lines is output.

[0113] In some embodiments, as described in step 105, the missing clinical events and the contradictory timelines are summarized to generate a document defect log for the medical record text to be reviewed, including: Step 701: Retrieve the disease specialty type corresponding to the missing clinical event in the standard diagnostic and treatment chart.

[0114] In this embodiment, the search begins with the source standard event node of the missing clinical event as the starting point, and a retrieval is initiated to the meta-information submodule of the clinical knowledge base. The meta-information submodule then returns the disease specialty type to which the node belongs. Next, the disease specialty type is used as the membership attribute of the missing clinical event and filled back into the description object of the missing clinical event. Finally, this provides a classification basis for subsequent entry encapsulation.

[0115] Step 702: Retrieve the actual timestamps of the two actual clinical events connected by the contradictory timeline at both ends of the actual diagnosis and treatment diagram.

[0116] In this embodiment, firstly, based on the edge object references of the contradictory time sequence connection, the starting node and the ending node are extracted respectively; secondly, the actual timestamps are read from the attribute fields of the two nodes; thirdly, the actual timestamps of the preceding event and the subsequent event are recorded in the "previous time" and "subsequent time" fields respectively; finally, time evidence is provided for the encapsulation of logical conflict entries.

[0117] Step 703: Combine and encapsulate the disease specialty type and the missing clinical event to generate a missing item prompt.

[0118] In this embodiment, firstly, the omission prompt entry structure is allocated according to the preset entry template; secondly, the disease specialty type field, the missing event name field, and the expected occurrence stage of the missing event field are sequentially filled into the entry structure; thirdly, a unique number and an "omission type" label are added to the entry; finally, the omission prompt entry is output to the log assembly queue.

[0119] Step 704: Merge and encapsulate the actual timestamp and the contradictory time sequence to generate a logical conflict entry.

[0120] In this embodiment, firstly, a logical conflict entry structure is allocated according to a preset entry template; secondly, the four fields of the preceding event name, the following event name, the preceding time, and the following time are filled into the entry structure in sequence; then, a "time series" type label and a brief description of the conflict reason are added; finally, the logical conflict entry is output to the log assembly queue.

[0121] Step 705: Vertically concatenate the omission prompt entries and the logical conflict entries to assemble them into a document defect log.

[0122] In this embodiment, firstly, all missing prompts and all logically conflicting entries are retrieved from the log assembly queue according to a preset sorting rule. The sorting rule can be arranged from high to low according to the severity of the defect or in chronological order. Secondly, the patient's basic information and quality control time are written into the medical record text to be reviewed using a preset log header template. Thirdly, all entries are sequentially spliced ​​vertically to form the complete log text. Finally, the assembled document defect log is output in a structured document format to the medical record quality control terminal for subsequent review and processing.

[0123] Figure 5 This application provides a schematic diagram of the structure of a medical text quality assessment system based on natural language processing, as shown in the embodiments below. Figure 5 As shown, the system includes: The acquisition module 51 is used to acquire the medical record text of the target patient to be reviewed, perform medical semantic parsing on the medical record text to be reviewed, and extract multiple actual clinical events recorded in the medical record text to be reviewed and the actual timestamp corresponding to each actual clinical event. Construction module 52 is used to sequentially associate the multiple actual clinical events according to the time progression direction of the actual timestamps to construct an actual diagnosis and treatment graph; Analysis module 53 is used to load a standard diagnosis and treatment diagram matching the disease specialty type of the medical record text to be reviewed from a preset clinical knowledge base, perform an event coverage check for the multiple actual clinical events within the standard diagnosis and treatment diagram, and separate out the missing clinical events contained in the standard diagnosis and treatment diagram but not covered by the actual diagnosis and treatment diagram. The locking module 54 is used to perform time-series dependency review along the node connection path of the actual diagnosis and treatment diagram, and lock contradictory time-series connections within the actual diagnosis and treatment diagram that violate the evolution logic of the standard diagnosis and treatment diagram. The generation module 55 is used to summarize the missing clinical events and the contradictory time sequence connections to generate a document defect log for the medical record text to be reviewed.

[0124] Figure 5 The aforementioned medical text quality assessment system based on natural language processing can perform... Figure 1The implementation principle and technical effects of the medical text quality assessment method based on natural language processing described in the illustrated embodiment will not be repeated here. The specific methods by which each module and unit of the medical text quality assessment system based on natural language processing in the above embodiments are described in detail in the embodiments related to this method, and will not be elaborated upon here.

[0125] In one possible design, Figure 5 The medical text quality assessment system based on natural language processing, as shown in the embodiment, can be implemented as a computing device, such as... Figure 6 As shown, the computing device may include a storage component 61 and a processing component 62; The storage component 61 stores one or more computer instructions, wherein the one or more computer instructions are invoked and executed by the processing component 62.

[0126] The processing component 62 is used to perform the above. Figure 1 The embodiment describes a method for assessing the quality of medical text based on natural language processing.

[0127] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device, such as a personal computer, server, or network device, to execute the methods described in the various embodiments or some parts of the embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for assessing the quality of medical text based on natural language processing, characterized in that, include: Obtain the medical record text of the target patient to be reviewed, perform medical semantic parsing on the medical record text to be reviewed, and extract multiple actual clinical events recorded in the medical record text to be reviewed and the actual timestamp corresponding to each actual clinical event; By sequentially associating the multiple actual clinical events according to the time progression direction of the actual timestamps, an actual diagnosis and treatment graph is constructed. Load a standard diagnostic and treatment diagram matching the disease specialty type of the medical record text to be reviewed from a pre-set clinical knowledge base, perform an event coverage check for the multiple actual clinical events within the standard diagnostic and treatment diagram, and separate out the missing clinical events that are included in the standard diagnostic and treatment diagram but not covered by the actual diagnostic and treatment diagram. A temporal dependency review is performed along the node connection path of the actual diagnosis and treatment diagram to identify contradictory temporal connections within the actual diagnosis and treatment diagram whose sequential order violates the evolution logic of the standard diagnosis and treatment diagram. By summarizing the missing clinical events and connecting the contradictory timelines, a document defect log is generated for the medical record text to be reviewed.

2. The method according to claim 1, characterized in that, Obtain the medical record text of the target patient to be reviewed, perform medical semantic parsing on the medical record text to be reviewed, and extract multiple actual clinical events recorded in the medical record text to be reviewed and the actual timestamp corresponding to each actual clinical event; The medical record text to be reviewed was divided into multiple basic segments according to the boundaries of punctuation marks; The system traverses and scans the pre-set medical terminology within the multiple basic segments to extract the medical behavior entities and medical object entities that constitute the actual clinical event. Based on the grammatical structure within the basic segment, establish the action-receiving relationship between the medical behavior entity and the medical object entity; Extract the temporal modifiers that are immediately adjacent to the action-patient relationship within the basic sentence segment; The time-modifying affixes are mapped to absolute time values ​​as the actual timestamps, and the action-actor association is established as the actual clinical event.

3. The method according to claim 1, characterized in that, By sequentially associating the multiple actual clinical events according to the time progression direction of the actual timestamps, an actual diagnosis and treatment graph is constructed, including: Extract the actual timestamps corresponding to the actual clinical events and sort them in ascending order of values ​​to determine the sequential arrangement sequence of the actual clinical events; The actual clinical events are transformed into discrete event nodes in the actual diagnostic and treatment graph; Based on the adjacency relationship within the sequential arrangement sequence, establish connecting line segments with unidirectional pointing attributes between the discrete event nodes; Connecting the end-to-end connecting line segments with unidirectional pointing properties forms a continuous evolution path connecting the discrete event nodes; The continuous evolution path and the multiple discrete event nodes are combined to form an actual diagnostic and treatment map.

4. The method according to claim 1, characterized in that, A standard diagnostic chart matching the disease specialty type of the medical record text to be reviewed is loaded from a pre-built clinical knowledge base. Within the standard diagnostic chart, an event coverage check is performed on the multiple actual clinical events to separate missing clinical events included in the standard diagnostic chart but not covered by the actual diagnostic chart, including: Extract the medical behavior entities and medical object entities contained in the actual clinical events as the actual core elements; Analyze the standard event nodes contained within the standard diagnostic and treatment diagram, and extract the standard core elements corresponding to the standard event nodes; Perform a string comparison of each element between the actual core element and the standard core element; In the standard diagnostic diagram, standard event nodes that do not correspond to the actual core element with the same string are screened out as isolated standard nodes; The isolated standard node was identified as a missing clinical event.

5. The method according to claim 4, characterized in that, The standard diagnostic graph identifies standard event nodes that do not correspond to the same actual core element as isolated standard nodes, including: A set of candidate isolated nodes is formed by aggregating standard event nodes that do not correspond to the actual core elements with the same string. Retrieve the medical exemption trigger words that each candidate isolated node in the candidate isolated node set is bound to in the clinical knowledge base; Scan the medical exemption trigger words within the medical record text to be reviewed; When a scan hits, extract patient refusal entities or physiological contraindication entities that are immediately adjacent to the medical exemption trigger words; Remove nodes associated with the patient's refusal entity or the physiological contraindication entity from the candidate isolated node set, and establish the remaining standard event nodes after the removal operation as isolated standard nodes.

6. The method according to claim 1, characterized in that, A temporal dependency review is performed along the node connections of the actual treatment diagram to identify contradictory temporal connections within the actual treatment diagram whose sequential order violates the evolution logic of the standard treatment diagram, including: Extract the actual transition line between any two adjacent actual clinical events within the actual diagnostic and treatment diagram; Determine the subsequent clinical events to which the actual transition line points and the preceding clinical events that deviate from it; Locate the standard subsequent node matching the subsequent clinical event and the standard preceding node matching the preceding clinical event within the standard diagnostic and treatment map; Detect the direction of the standard connection line between the standard preceding node and the standard subsequent node within the standard diagnostic diagram; When the direction of the standard connection is opposite to that of the actual transition connection, the actual transition connection is locked as a contradictory timing connection.

7. The method according to claim 1, characterized in that, By summarizing the missing clinical events and connecting them with the contradictory timelines, a document defect log is generated for the medical record text under review, including: Retrieve the disease specialty type corresponding to the missing clinical event in the standard diagnostic and treatment diagram; Retrieve the actual timestamps of the two actual clinical events connected at both ends of the actual diagnostic and treatment diagram by the contradictory timeline connection; The disease specialty type and the missing clinical event are combined and packaged to generate a missing item prompt; The actual timestamp and the contradictory time sequence are connected, merged, and encapsulated to generate a logical conflict entry; The omission prompts and logical conflict entries are vertically concatenated to form a document defect log.

8. A medical text quality assessment system based on natural language processing, characterized in that, include: The acquisition module is used to acquire the medical record text of the target patient to be reviewed, perform medical semantic parsing on the medical record text to be reviewed, and extract multiple actual clinical events recorded in the medical record text to be reviewed and the actual timestamp corresponding to each actual clinical event. The construction module is used to sequentially associate the multiple actual clinical events according to the time progression direction of the actual timestamps to construct an actual diagnosis and treatment graph; The analysis module is used to load a standard diagnostic and treatment diagram that matches the disease specialty type of the medical record text to be reviewed from a preset clinical knowledge base, perform an event coverage check for the multiple actual clinical events within the standard diagnostic and treatment diagram, and separate out the missing clinical events that are included in the standard diagnostic and treatment diagram but not covered by the actual diagnostic and treatment diagram. The locking module is used to perform temporal dependency review along the node connection path of the actual diagnosis and treatment diagram, and lock contradictory temporal connections within the actual diagnosis and treatment diagram whose sequential order violates the evolution logic of the standard diagnosis and treatment diagram; The generation module is used to summarize the missing clinical events and the contradictory timelines to generate a document defect log for the medical record text to be reviewed.

9. A computing device, characterized in that, It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are invoked and executed by the processing component to implement a medical text quality assessment method based on natural language processing as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The device contains a computer program that, when executed by a computer, implements a medical text quality assessment method based on natural language processing as described in any one of claims 1 to 7.

Images