A medical text deep semantic parsing and intelligent query platform

Abstract

The application provides a medical text deep semantic analysis and intelligent query platform, and relates to the technical field of medical query platforms, and comprises obtaining multi-source medical text data, including outpatient medical records, admission records, test reports, medical records, doctor active input, patient active input and the like; while obtaining the medical text data, the patient identifier, the document type, the document generation time, the department information, the doctor information, the data source and the like are recorded synchronously; and the medical text data is subjected to standardization pretreatment. In the application, the medical text intelligent processing system developed for multiple scenarios of hospital clinical scientific research, medical record management, medical quality control and discipline construction focuses on the full-quantity structured and unstructured medical record text data of the whole hospital since the hospital was built, and overcomes the industry pain points in traditional medical text processing, such as shallow semantic understanding, low-efficiency information retrieval, not deep data mining and difficulty in efficient reuse of diagnosis and treatment information.

Description

Technical Field

[0001] This invention relates to the field of medical query platform technology, specifically to a deep semantic analysis and intelligent query platform for medical texts. Background Technology

[0002] A medical intelligent query platform refers to a digital system that utilizes artificial intelligence, big data, and cloud computing technologies to integrate medical resources and health data, providing the public with services such as disease consultation, medical guidance, and report interpretation. Most existing medical intelligent query platforms simply record or query patients' information. However, doctors often have strong personal habits and tend to simplify treatment plans, making it difficult to retrieve and view medical texts entered into the query platform, thus hindering accurate querying and interpretation. Furthermore, existing intelligent query platforms...

[0003] Therefore, this invention provides a deep semantic analysis and intelligent query platform for medical texts. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention aims to provide a deep semantic analysis and intelligent query platform for medical texts, solving the problems mentioned in the background. This invention is a medical text intelligent processing system developed for multiple scenarios including hospital clinical research, medical record management, medical quality control, and discipline development. It focuses on all structured and unstructured medical record text data collected since the hospital's establishment, overcoming industry pain points in traditional medical text processing such as shallow semantic understanding, inefficient information retrieval, insufficient data mining, and difficulty in efficiently reusing diagnostic and treatment information. Through core technologies such as deep semantic analysis, intelligent information extraction, multi-dimensional intelligent retrieval, and diagnostic and treatment data correlation analysis, it achieves standardized governance, intelligent interpretation, and precise querying of massive amounts of medical record text data. It deeply integrates medical professional knowledge with large-scale model semantic analysis capabilities, adapting to various applications. This system can efficiently and accurately identify medical entities, perform semantic structured conversion of text, mine diagnostic and treatment information, quickly retrieve research data, and automatically verify medical quality control information across all types of medical documents, including outpatient medical records, inpatient progress notes, surgical records, discharge summaries, doctor's orders, and laboratory reports. It comprehensively streamlines the transformation of medical record data from text to value, providing one-stop intelligent data support for clinical research project initiation, disease pattern research, treatment pathway optimization, medical quality control, and medical insurance compliance management. Combining professionalism, practicality, and innovation, it effectively improves the efficiency of hospital medical record data management and research transformation capabilities, contributing to the digital construction of smart hospitals and the upgrading of medical research levels. Through error correction, it corrects doctors' medical plans and performs deep learning based on doctors' feedback, thereby ensuring the accuracy of treatment.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a deep semantic parsing and intelligent query platform for medical texts, comprising an intelligent query system, wherein the intelligent query system performs parsing and querying through the following method:

[0006] S1. Acquire multi-source medical text data, including outpatient medical records, admission records, laboratory reports, medical records, doctor-initiated input, patient-initiated input, etc. While acquiring medical text data, simultaneously record patient identifiers, document types, document generation time, department information, doctor information, data sources, etc.

[0007] S2. Standardize and preprocess the medical text data. Standardize the acquired multi-source medical text data to obtain medical text in a unified format.

[0008] S3. Extract medical entities from medical texts. Perform medical entity recognition on standardized medical texts, extract core medical entities from the text, and record the source document, paragraph, original text fragment, character position, extraction confidence, etc. for each medical entity.

[0009] S4. Identify the semantic state of medical entities. Perform semantic state identification on the extracted medical entities to determine the true clinical meaning of the medical entity in the text. Through semantic state identification, avoid misjudging negative diseases, suspected diseases, past diseases, etc. as currently diagnosed diseases.

[0010] S5. Extract diagnosis and treatment events and determine the event time. Based on the medical entity and its context, extract the patient's diagnosis and treatment events and determine the occurrence time of each diagnosis and treatment event.

[0011] S6. Construct a patient time-series diagnosis and treatment map. Taking patients as the unit, merge the diagnosis and treatment times extracted from different medical texts in chronological order to construct a patient time-series diagnosis and treatment map. The time-series diagnosis and treatment map can be used to reconstruct the complete diagnosis and treatment process of patients from admission to discharge.

[0012] S7. Establish an original evidence chain index. For each medical entity, diagnosis and treatment event and graph relationship, an original evidence chain index is established. Doctors can view the original medical records corresponding to the system's error correction suggestions through the original evidence chain index to avoid black box prompts.

[0013] S8. Correcting doctors' treatment plans: Doctors input treatment plans for patients, and the system retrieves text data based on the patient's condition to compare the treatment plans and provides correction reminders for the doctors' treatment plans.

[0014] S9. Show the doctor the error correction structure and receive feedback. Show the doctor the error correction suggestions for the treatment plan, and the doctor will process the error correction suggestions.

[0015] S10. Perform safety error correction and risk level calculation on the medical plan. The doctor returns the treatment plan, and the original system information is overwritten according to the new treatment plan. The safety risk level of the treatment plan is calculated.

[0016] S11. Optimize based on doctor feedback, and include the doctor feedback results into the low confidence sample pool and the rule optimization sample pool for subsequent system optimization.

[0017] S12. The diagnostic results and process are input into the system to form a closed loop, and the doctor inputs the diagnostic process and results into the system.

[0018] Furthermore, the standardization preprocessing includes: removing invalid characters and garbled text, unifying text encoding format, identifying document type, dividing medical record paragraphs, extracting titles, body text, tables and examination result fields, expanding medical abbreviations, standardizing synonymous medical terms, normalizing time expressions, and standardizing encoding mapping for content such as diseases, surgeries, tests, and drugs.

[0019] Standardization preprocessing of medical text data includes the following steps:

[0020] S201. Clean the text, standardize the format, and correct garbled characters;

[0021] S202. Based on a large medical model, standardize the mapping of professional terminology and automatically connect to international medical standard coding systems such as ICD-10, SNOMEDCT, and LOINC. Eliminate differences in writing formats and expression habits among different physicians, and achieve standardized and unified processing of medical record text data throughout the hospital and its entire lifecycle.

[0022] Furthermore, the medical entities include: disease diagnosis, symptoms, signs, laboratory indicators, imaging findings, pathological results, drug names and dosages, surgical procedures, complications, adverse reactions, allergy history, family history, and past medical history, including the following steps:

[0023] S301. Based on large-scale model deep semantic understanding technology, accurately extract core medical entities from medical record texts, extract entities such as diseases, symptoms, signs, tests, examinations, drugs, surgeries, complications, allergy history, and past medical history, and combine them into diagnosis and treatment events.

[0024] S302. Automatically identify hidden medical information, symptom evolution logic, and medication sequence relationships in text.

[0025] Furthermore, the semantic states include affirmative state, negative state, suspected state, excluded state, past state, family history state, planned state, preoperative state, postoperative state, and outcome state.

[0026] Furthermore, the patient time-series diagnostic and treatment atlas includes patient nodes, disease nodes, symptom nodes, examination nodes, laboratory test nodes, drug nodes, surgical nodes, and prognostic nodes. The relationships within the atlas include chronological relationships, diagnostic basis relationships, treatment relationships, examination relationships, drug effect relationships, adverse drug reaction relationships, and complication relationships, and include the following steps:

[0027] S601. Calculate the joint confidence level for each entity and event, excluding data interference from denial, family history, past history, time expiration, or conflicting evidence.

[0028] S602. Generate a current patient profile based on the patient’s time-series diagnosis and treatment map, and determine the patient’s current diagnosis, symptoms, abnormal test results, medications, allergy history, complications, surgical history and risk factors.

[0029] Furthermore, the process of correcting the doctor's treatment plan includes the following steps:

[0030] S801. Obtain the doctor's current treatment plan. Obtain the doctor's current proposed diagnosis, examination, medication, treatment, surgery and follow-up plan from the doctor's workstation, prescription system, medical order system or surgical application system.

[0031] S802. Generate a patient status profile through a dynamic patient status profile generation algorithm. Based on the patient-level time-series diagnosis and treatment map, extract the patient's current diagnosis, symptoms, test results, imaging results, medication status, allergy history, past medical history, complications, and liver and kidney function status.

[0032] S803. Verify the diagnostic basis by matching the diagnosis filled in by the doctor with the symptoms, examinations, tests, imaging and pathological evidence in the patient status profile to determine whether the diagnostic basis is sufficient.

[0033] S804. Perform treatment pathway verification by comparing the treatment pathway proposed by the doctor with clinical pathways, guidelines, hospital treatment standards and the patient's current disease stage to determine whether the treatment sequence, intensity and timing are reasonable.

[0034] S805. Conduct medication safety verification based on the drug knowledge base, patient allergy history, liver and kidney function, age, weight, pregnancy status, current medication and past adverse reactions, and verify drug selection, dosage, frequency, route and combination relationships.

[0035] S806. Conduct a reasonableness check of the inspection and examination, determine whether there are any omissions, duplications, mismatches or incorrect timing of the proposed inspection items, and determine whether a certain medical entity or diagnosis and treatment event is real and valid through the evidence-state-time joint belief propagation algorithm.

[0036] S807. Perform a surgical plan suitability check to determine whether the proposed surgery matches the results of diagnosis, imaging, pathology, disease staging, and preoperative risk assessment.

[0037] S808. Generate risk level: Calculate the risk level based on the degree of deviation from the treatment plan, the patient's basic risk, the degree of rule conflict, and the degree of lack of evidence.

[0038] S809. Generate error correction suggestions. The error correction problem, risk level, triggering rules, patient evidence, original text fragments and suggested handling methods are output through the counterfactual error correction algorithm of the treatment plan. The judgment is made by the two-way evidence chain verification algorithm. The missing diagnosis and treatment error active discovery algorithm actively infers the diagnosis and treatment measures that should have occurred but did not occur. The standard path is dynamically adjusted according to the individual patient status by the dynamic calculation algorithm of the deviation degree of the diagnosis and treatment path. Then, it is judged whether the doctor's plan has deviated. The hierarchical risk gating algorithm determines whether to remind, how to remind, and whether to block based on the severity of risk, strength of evidence and doctor's historical feedback.

[0039] S810 receives doctor feedback and performs closed-loop optimization. It collects doctors' actions regarding error correction suggestions, such as acceptance, modification, ignoring, marking false alarms, or supplementing evidence, and uses this data for subsequent model and rule optimization. It dynamically adjusts rule weights, evidence weights, and reminder strategies through an adaptive weight update algorithm driven by doctor feedback.

[0040] Furthermore, the evidence-state-time joint belief propagation algorithm includes the following steps:

[0041] S8.11 Calculate the basic confidence score for entity extraction. For a medical entity x, obtain the basic confidence score given by the entity recognition model: Cner(x).

[0042] S8.12 Calculate the semantic state weight: Calculate the state weight based on the semantic state of entity x: StateWeight(x);

[0043] S8.13. Calculate the time validity weight. For a treatment event x, calculate the time difference Δt between it and the time when the current treatment plan was generated, and calculate the time validity weight:

[0044] TimeWeight(x) = exp(-λΔt),

[0045] Where λ is the time decay coefficient;

[0046] S8.14 Calculate document source reliability. Different medical documents have different reliability for different medical facts. Set document source reliability: SourceWeight(x). If the entity is "pathological diagnosis" and comes from a pathology report, then SourceWeight is higher; if the entity is "pathological diagnosis" but only comes from a paraphrase in the medical record, then SourceWeight is lower.

[0047] S8.15 Calculate the rule consistency weight. Perform a consistency check between entity x and the medical knowledge rules. If the entity or event is consistent with the medical rules, the RuleWeight is higher; if there is a conflict, the RuleWeight is lower.

[0048] S8.16 Calculate the conflict penalty term: If there is evidence in the same patient atlas that contradicts entity x, then calculate the conflict penalty: ConflictPenalty(x).

[0049] S8.17 Calculate the joint confidence level, and finally calculate the effective confidence level of the entity or time:

[0050] Credibility(x)=σ[aCner(x)+bStateWeight(x)+cTimeWeight(x)+dSourceWeight(x)+eRuleWeight(x)-fConflictPenalty(x)]

[0051] Where σ represents the normalization function, and a, b, c, d, e, and f are weighting coefficients.

[0052] Furthermore, the dynamic patient profile generation algorithm includes:

[0053] S8.21, Candidate State Extraction: Extract candidate state items from the patient-level time-series diagnosis and treatment graph: Candidate={x1, x2, ..., xn}, where each candidate state item corresponds to a medical entity or diagnosis and treatment event;

[0054] S8.22. For each candidate state item x, calculate the current valid score:

[0055] CurrentScore(x)=αCredibility(x)+βTimeWeight(x)+γContinuity(x)+δSeverity(x)+εSourceWeight(x);

[0056] Where Credibility(x) represents the joint confidence level, TimeWeight(x) represents the time validity, Continuity(x) represents the state continuity, Severity(x) represents the severity, and SourceWeight(x) represents the source reliability.

[0057] S8.23, Calculation of state continuity: For continuous states such as chronic diseases, long-term medication, allergy history, and previous surgical history, calculate the state continuity: Continuity(x) = 1 - DiscontinueEvidence(x), where DiscontinueEvidence(x) represents the evidence of state termination.

[0058] S8.24. State Conflict Resolution: If multiple conflicting pieces of evidence exist for the same state item, they are resolved based on the strength of the evidence. The system judges the conflict resolution based on document source, chronological order, and strength of evidence. The conflict resolution score is as follows:

[0059] ResolveScore(x)=Credibility(x)+SourceWeight(x)+TimeWeight(x)-ConflictPenalty(x)

[0060] Select the state with the highest ResolveScore as the current state.

[0061] S8.25. Generate a current patient status profile. When CurrentScore(x) is greater than the set threshold θ, x is included in the current patient status profile.

[0062] Furthermore, the counterfactual error correction algorithm for the treatment plan includes the following steps:

[0063] S8.31. Generate doctor's treatment plan path, converting the doctor's current treatment plan into a treatment plan path:

[0064] PathDoctor={DiagnosisDoctor, ExamDoctor, DrugDoctor, SurgeryDoctor,TreatmentDoctor, FollowupDoctor}

[0065] S8.32. Generate counterfactual safety solution paths. Based on the patient's current status profile, medical knowledge base, clinical pathway database, and drug knowledge base, generate one or more counterfactual safety solution paths:

[0066] PathSafe={DiagnosisSafe, ExamSafe, DrugSafe, SurgerySafe, TreatmentSafe, FollowupSafe};

[0067] S8.33. Perform path difference calculation to determine the difference between the doctor's proposed path and the counterfactual safety proposed path:

[0068] Diff(PathDoctor, PathSafe)={DiffDiagnosis, DiffExam, DiffDrug,DiffSurgery, DiffTreatment, DiffFollowup}

[0069] Among them, DiffDiagnosis represents diagnostic differences, DiffExam represents examination differences, DiffDrug represents medication differences, DiffSurgery represents surgical differences, DiffTreatment represents treatment pathway differences, and DiffFollowup represents follow-up differences.

[0070] S8.34. Locate the error type and locate the error in the treatment plan based on the difference type;

[0071] S8.35. Calculate the counterfactual risk contribution. For each difference item di, calculate its risk contribution:

[0072] RiskContribution(di)=Severity(di)×EvidenceStrength(di)×PatientRisk(di)×RuleConflict(di)

[0073] Among them, Severity (di) indicates the severity of the difference itself, Evidence Strength (di) indicates the strength of the original evidence, Patient Risk (di) indicates the patient's basic risk, and Rule Conflict (di) indicates the degree of conflict with medical rules;

[0074] S8.36. Output error correction suggestions, select the difference items whose RiskContribution exceeds the threshold as error correction items, and generate error correction suggestions;

[0075] The bidirectional evidence chain verification algorithm includes the following steps:

[0076] S8.41, Extracting protocol elements: Extracting protocol elements from the doctor's protocol:

[0077] PlanElement={p1, p2,...,pn}

[0078] For example, diagnosis, medication, examination, surgery, and treatment plans;

[0079] S8.42. Search for supporting evidence. For each protocol element pi, search for supporting evidence in the patient timeline diagnosis and treatment atlas and the chain of evidence index: SupportEvidence(pi).

[0080] S8.43. Search for opposing evidence, and simultaneously search for opposing evidence: OpposeEvidence(pi)

[0081] S8.44 Calculate the evidence balance score.

[0082] EvidenceBalance(pi)=SupportScore(pi)-OpposeScore(pi),

[0083] Among them: SupportScore(pi)=ΣCredibility(supporti),

[0084] OpposeScore(pi)=ΣCredibility(opposei);

[0085] S8.45. Generate evidence-based error correction conclusions. If EvidenceBalance(pi) is lower than the threshold, output: "The current protocol element lacks sufficient supporting evidence or there is significant opposing evidence." If OpposeScore is significantly higher than SupportScore, output: "The current protocol element conflicts with medical record evidence."

[0086] Furthermore, the doctor feedback-driven adaptive weight update algorithm includes the following steps:

[0087] S9.1, Collect doctor feedback;

[0088] S9.2. The effectiveness of feedback should be judged in conjunction with the results of subsequent disease progression. If the doctor ignores the high-risk warning but a related adverse event occurs later, the rule should not be simply downgraded. If the doctor marks a false alarm and there is no subsequent risk event, and the senior doctor approves, the weight of the corresponding rule can be reduced.

[0089] S9.3 Calculate the rule performance score. For rule r, calculate:

[0090] RulePerformance(r)=TP / (TP+FP+FN+ε)

[0091] Where TP represents a correct alert, FP represents a false alarm, FN represents a missed alert, and ε is a constant to prevent the denominator from being zero.

[0092] S9.4 Update the rule weights. The rule weights are updated as follows: Wr(new) = μWr(old) + (1-μ)RulePerformance(r), where μ is the historical weight retention coefficient.

[0093] S9.5 Update algorithm weights: If a certain type of error correction has a high false alarm rate over a long period, reduce its low-risk alert weight; if a certain type of error correction leads to adverse consequences due to missed reports, increase the weight of that type of rule.

[0094] The beneficial effects of this invention are:

[0095] 1. Empowered by a professional medical big model: The big model is finely tuned for the medical field and deeply adapted to medical terminology and diagnosis and treatment logic. The semantic parsing accuracy is much higher than that of general big models, which is in line with the actual needs of hospital medical record data processing.

[0096] 2. Full-cycle medical record data adaptation: Covers all historical medical record data since the hospital's establishment, is compatible with the writing standards of both new and old medical records, and enables unified parsing and querying of long-term medical texts, breaking down barriers to the use of historical data.

[0097] 3. Highly efficient and intelligent processing: Enables batch and rapid parsing and retrieval of massive amounts of medical record texts, significantly shortening the time for medical data processing and scientific research data preparation, replacing repetitive manual work, and improving work efficiency.

[0098] 4. Data Compliance and Security: Adhering to medical data security management standards, the system is deployed locally and privately, ensuring that data does not leave the hospital and is encrypted and controlled throughout the entire process, thus protecting the security of medical records and patient privacy.

[0099] 5. Strong application in various scenarios: It accurately connects with core scenarios such as clinical research, medical record management, medical quality control, and medical insurance management. Its functions are practical and easy to operate, and users can quickly get started without a professional technical background.

[0100] 6. By correcting and demonstrating the doctor's treatment plan, and conducting deep learning based on the doctor's feedback, a better learning effect can be achieved, ensuring the accuracy of treatment results. Attached Figure Description

[0101] Figure 1 This is a flowchart of the overall structure of a medical text deep semantic parsing and intelligent query platform according to the present invention;

[0102] Figure 2 This is a flowchart of the intelligent error correction process in a medical text deep semantic parsing and intelligent query platform of the present invention;

[0103] Figure 3This is a flowchart of the standardized preprocessing of medical text data in a medical text deep semantic parsing and intelligent query platform of the present invention.

[0104] Figure 4 This is a flowchart of the evidence-state-temporal joint belief propagation algorithm in a medical text deep semantic parsing and intelligent query platform of the present invention;

[0105] Figure 5 This is a flowchart of a dynamic patient status profile generation algorithm in a medical text deep semantic parsing and intelligent query platform of the present invention.

[0106] Figure 6 This is a flowchart of a counterfactual error correction algorithm for treatment plans in a medical text deep semantic parsing and intelligent query platform of the present invention;

[0107] Figure 7 This is a flowchart of a two-way evidence chain verification algorithm in a medical text deep semantic parsing and intelligent query platform of the present invention;

[0108] Figure 8 This is a flowchart of an algorithm for proactively detecting definitive diagnostic errors in a medical text deep semantic parsing and intelligent query platform according to the present invention.

[0109] Figure 9 This is a flowchart of the dynamic calculation algorithm for the deviation of the diagnosis and treatment path in a medical text deep semantic parsing and intelligent query platform of the present invention;

[0110] Figure 10 This is a flowchart of a hierarchical risk gating algorithm in a medical text deep semantic parsing and intelligent query platform of the present invention;

[0111] Figure 11 This is a flowchart of the adaptive weight update algorithm driven by doctor feedback in a medical text deep semantic parsing and intelligent query platform of the present invention. Detailed Implementation

[0112] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0113] Please see Figures 1 to 11 This invention provides a technical solution: a deep semantic parsing and intelligent query platform for medical texts, including an intelligent query system, which performs parsing and querying through the following methods:

[0114] S1. Acquire multi-source medical text data, including outpatient medical records, admission records, laboratory reports, medical records, doctor-initiated input, patient-initiated input, etc. While acquiring medical text data, simultaneously record patient identifiers, document types, document generation time, department information, doctor information, data sources, etc.

[0115] S2. Standardize and preprocess the medical text data. Standardize the acquired multi-source medical text data to obtain medical text in a unified format.

[0116] S3. Extract medical entities from medical texts. Perform medical entity recognition on standardized medical texts, extract core medical entities from the text, and record the source document, paragraph, original text fragment, character position, extraction confidence, etc. for each medical entity.

[0117] S4. Identify the semantic state of medical entities. Perform semantic state identification on the extracted medical entities to determine the true clinical meaning of the medical entity in the text. Through semantic state identification, avoid misjudging negative diseases, suspected diseases, past diseases, etc. as currently diagnosed diseases.

[0118] S5. Extract diagnosis and treatment events and determine the event time. Based on the medical entity and its context, extract the patient's diagnosis and treatment events and determine the occurrence time of each diagnosis and treatment event.

[0119] S6. Construct a patient time-series diagnosis and treatment map. Taking patients as the unit, merge the diagnosis and treatment times extracted from different medical texts in chronological order to construct a patient time-series diagnosis and treatment map. The time-series diagnosis and treatment map can be used to reconstruct the complete diagnosis and treatment process of patients from admission to discharge.

[0120] S7. Establish an original evidence chain index. For each medical entity, diagnosis and treatment event and graph relationship, an original evidence chain index is established. Doctors can view the original medical records corresponding to the system's error correction suggestions through the original evidence chain index to avoid black box prompts.

[0121] S8. Correcting doctors' treatment plans: Doctors input treatment plans for patients, and the system retrieves text data based on the patient's condition to compare the treatment plans and provides correction reminders for the doctors' treatment plans.

[0122] S9. Show the doctor the error correction structure and receive feedback. Show the doctor the error correction suggestions for the treatment plan, and the doctor will process the error correction suggestions.

[0123] S10. Perform safety error correction and risk level calculation on the medical plan. The doctor returns the treatment plan, and the original system information is overwritten according to the new treatment plan. The safety risk level of the treatment plan is calculated.

[0124] S11. Optimize based on doctor feedback, and include the doctor feedback results into the low confidence sample pool and the rule optimization sample pool for subsequent system optimization.

[0125] S12. The diagnostic results and process are input into the system to form a closed loop, and the doctor inputs the diagnostic process and results into the system.

[0126] In this embodiment, the standardized preprocessing includes: removing invalid characters and garbled text, unifying the text encoding format, identifying document types, dividing medical record paragraphs, extracting titles, body text, tables and examination result fields, expanding medical abbreviations, standardizing synonymous medical terms, normalizing time expressions, and performing standard encoding mapping on content such as diseases, surgeries, tests, and drugs.

[0127] Standardization preprocessing of medical text data includes the following steps:

[0128] S201. Clean the text, standardize the format, and correct garbled characters;

[0129] S202. Based on a large medical model, standardize the mapping of professional terminology and automatically connect to international medical standard coding systems such as ICD-10, SNOMEDCT, and LOINC. Eliminate differences in writing formats and expression habits among different physicians, and achieve standardized and unified processing of medical record text data throughout the hospital and its entire lifecycle.

[0130] The multi-source medical text data includes outpatient medical records, admission records, initial medical progress records, daily medical progress records, surgical records, anesthesia records, medical orders, laboratory reports, imaging reports, pathology reports, discharge summaries, medical record cover sheets, and follow-up records.

[0131] When medical text data is standardized, for example, “coronary heart disease”, “CHD”, and “coronary atherosclerotic heart disease” are grouped into the same standard disease concept; and “postoperative day 3”, “three days after admission”, and “March 4, 2025” are uniformly converted into calculable time information.

[0132] Specifically, it is compatible with all types of unstructured and semi-structured medical record texts in the hospital, automatically completes text cleaning, format unification, and garbled character correction, completes the standardized mapping of professional terms based on a medical-specific large model, and automatically connects to international medical standard coding systems such as ICD-10, SNOMEDCT, and LOINC, eliminating differences in writing formats and expression habits among different physicians, and realizing the standardized and unified processing of medical record text data throughout the entire hospital and its life cycle.

[0133] Medical entities include: disease diagnosis, symptoms, signs, laboratory indicators, imaging findings, pathological results, drug names and dosages, surgical procedures, complications, adverse reactions, allergy history, family history, and past medical history, and include the following steps:

[0134] S301. Based on large-scale model deep semantic understanding technology, accurately extract core medical entities from medical record texts, extract entities such as diseases, symptoms, signs, tests, examinations, drugs, surgeries, complications, allergy history, and past medical history, and combine them into diagnosis and treatment events.

[0135] S302. Automatically identify hidden medical information, symptom evolution logic, and medication sequence relationships in text.

[0136] Semantic states include affirmative, negative, suspected, excluded, past history, family history, planned, preoperative, postoperative, and outcome states. For example, if the text is "Patient denies chest pain," then "chest pain" is marked as a negative state; if the text is "Patient has a history of hypertension," then "hypertension" is marked as a past history; if the text is "Consider lung infection," then "lung infection" is marked as a suspected state; if the text is "Father has diabetes," then "diabetes" is marked as a family history. By identifying these semantic states, we can avoid mistakenly identifying negative, suspected, past, or family history diseases as the patient's current diagnosed disease.

[0137] Based on medical entities and their contextual relationships, extract the patient's diagnostic and treatment events and determine the occurrence time of each event. Diagnostic and treatment events include: onset events, admission events, diagnosis events, examination events, laboratory events, medication events, surgical events, complication events, etc. Each diagnostic and treatment event includes at least the patient identifier, event type, medical entity, event time, and event status. For example, "The patient developed fever and elevated white blood cell count on the 5th day after surgery, suspected postoperative infection, and was given cephalosporin antibiotics" can be parsed as: fever event, elevated white blood cell count event, suspected postoperative infection event, and cephalosporin antibiotic administration event.

[0138] The patient timeline diagnosis and treatment atlas includes patient nodes, disease nodes, symptom nodes, examination nodes, laboratory test nodes, medication nodes, surgical nodes, and prognostic nodes. The relationships within the atlas include chronological relationships, diagnostic basis relationships, treatment relationships, examination relationships, drug effect relationships, adverse drug reaction relationships, and complication relationships. It includes the following steps:

[0139] S601. Calculate the joint confidence level for each entity and event, excluding data interference from denial, family history, past history, time expiration, or conflicting evidence.

[0140] S602. Generate a current patient profile based on the patient’s time-series diagnosis and treatment map, and determine the patient’s current diagnosis, symptoms, abnormal test results, medications, allergy history, complications, surgical history and risk factors.

[0141] Specifically, taking patients as the unit, the diagnosis and treatment events extracted from different medical texts are merged in chronological order to construct a patient chronological diagnosis and treatment map. The patient chronological diagnosis and treatment map includes patient nodes, disease nodes, etc. The relationships in the map include: chronological relationships, diagnostic basis relationships, treatment relationships, etc. Through the patient-level chronological diagnosis and treatment map, the complete diagnosis and treatment process of the patient from admission, diagnosis, examination, treatment, surgery, complication management to discharge follow-up can be reconstructed.

[0142] Correcting a doctor's treatment plan involves the following steps:

[0143] S801. Obtain the doctor's current treatment plan. Obtain the doctor's current proposed diagnosis, examination, medication, treatment, surgery and follow-up plan from the doctor's workstation, prescription system, medical order system or surgical application system.

[0144] S802. Generate a patient status profile through a dynamic patient status profile generation algorithm. Based on the patient-level time-series diagnosis and treatment map, extract the patient's current diagnosis, symptoms, test results, imaging results, medication status, allergy history, past medical history, complications, and liver and kidney function status.

[0145] S803. Verify the diagnostic basis by matching the diagnosis filled in by the doctor with the symptoms, examinations, tests, imaging and pathological evidence in the patient status profile to determine whether the diagnostic basis is sufficient.

[0146] S804. Perform treatment pathway verification by comparing the treatment pathway proposed by the doctor with clinical pathways, guidelines, hospital treatment standards and the patient's current disease stage to determine whether the treatment sequence, intensity and timing are reasonable.

[0147] S805. Conduct medication safety verification based on the drug knowledge base, patient allergy history, liver and kidney function, age, weight, pregnancy status, current medication and past adverse reactions, and verify drug selection, dosage, frequency, route and combination relationships.

[0148] S806. Conduct a reasonableness check of the inspection and examination, determine whether there are any omissions, duplications, mismatches or incorrect timing of the proposed inspection items, and determine whether a certain medical entity or diagnosis and treatment event is real and valid through the evidence-state-time joint belief propagation algorithm.

[0149] S807. Perform a surgical plan suitability check to determine whether the proposed surgery matches the results of diagnosis, imaging, pathology, disease staging, and preoperative risk assessment.

[0150] S808. Generate risk level: Calculate the risk level based on the degree of deviation from the treatment plan, the patient's basic risk, the degree of rule conflict, and the degree of lack of evidence.

[0151] S809. Generate error correction suggestions. The error correction problem, risk level, triggering rules, patient evidence, original text fragments and suggested handling methods are output through the counterfactual error correction algorithm of the treatment plan. The judgment is made by the two-way evidence chain verification algorithm. The missing diagnosis and treatment error active discovery algorithm actively infers the diagnosis and treatment measures that should have occurred but did not occur. The standard path is dynamically adjusted according to the individual patient status by the dynamic calculation algorithm of the deviation degree of the diagnosis and treatment path. Then, it is judged whether the doctor's plan has deviated. The hierarchical risk gating algorithm determines whether to remind, how to remind, and whether to block based on the severity of risk, strength of evidence and doctor's historical feedback.

[0152] S810 receives doctor feedback and performs closed-loop optimization. It collects doctors' actions regarding error correction suggestions, such as acceptance, modification, ignoring, marking false alarms, or supplementing evidence, and uses this data for subsequent model and rule optimization. It dynamically adjusts rule weights, evidence weights, and reminder strategies through an adaptive weight update algorithm driven by doctor feedback.

[0153] The evidence-state-temporal joint belief propagation algorithm includes the following steps:

[0154] S8.11 Calculate the basic confidence score for entity extraction. For a medical entity x, obtain the basic confidence score given by the entity recognition model: Cner(x).

[0155] S8.12 Calculate the semantic state weight: Calculate the state weight based on the semantic state of entity x: StateWeight(x);

[0156] S8.13. Calculate the time validity weight. For a treatment event x, calculate the time difference Δt between it and the time when the current treatment plan was generated, and calculate the time validity weight:

[0157] TimeWeight(x) = exp(-λΔt),

[0158] Where λ is the time decay coefficient;

[0159] S8.14 Calculate document source reliability. Different medical documents have different reliability for different medical facts. Set document source reliability: SourceWeight(x). If the entity is "pathological diagnosis" and comes from a pathology report, then SourceWeight is higher; if the entity is "pathological diagnosis" but only comes from a paraphrase in the medical record, then SourceWeight is lower.

[0160] S8.15 Calculate the rule consistency weight. Perform a consistency check between entity x and the medical knowledge rules. If the entity or event is consistent with the medical rules, the RuleWeight is higher; if there is a conflict, the RuleWeight is lower.

[0161] S8.16 Calculate the conflict penalty term: If there is evidence in the same patient atlas that contradicts entity x, then calculate the conflict penalty: ConflictPenalty(x).

[0162] S8.17 Calculate the joint confidence level, and finally calculate the effective confidence level of the entity or time:

[0163] Credibility(x)=σ[aCner(x)+bStateWeight(x)+cTimeWeight(x)+dSourceWeight(x)+eRuleWeight(x)-fConflictPenalty(x)]

[0164] Where σ represents the normalization function, and a, b, c, d, e, and f are weighting coefficients.

[0165] The algorithm for generating dynamic patient profiles includes:

[0166] S8.21, Candidate State Extraction: Extract candidate state items from the patient-level time-series diagnosis and treatment graph: Candidate={x1, x2, ..., xn}, where each candidate state item corresponds to a medical entity or diagnosis and treatment event;

[0167] S8.22. For each candidate state item x, calculate the current valid score:

[0168] CurrentScore(x)=αCredibility(x)+βTimeWeight(x)+γContinuity(x)+δSeverity(x)+εSourceWeight(x);

[0169] Where Credibility(x) represents the joint confidence level, TimeWeight(x) represents the time validity, Continuity(x) represents the state continuity, Severity(x) represents the severity, and SourceWeight(x) represents the source reliability.

[0170] S8.23, Calculation of state continuity: For continuous states such as chronic diseases, long-term medication, allergy history, and previous surgical history, calculate the state continuity: Continuity(x) = 1 - DiscontinueEvidence(x), where DiscontinueEvidence(x) represents the evidence of state termination.

[0171] S8.24. State Conflict Resolution: If multiple conflicting pieces of evidence exist for the same state item, they are resolved based on the strength of the evidence. The system judges the conflict resolution based on document source, chronological order, and strength of evidence. The conflict resolution score is as follows:

[0172] ResolveScore(x)=Credibility(x)+SourceWeight(x)+TimeWeight(x)-ConflictPenalty(x)

[0173] Select the state with the highest ResolveScore as the current state.

[0174] S8.25. Generate a current patient status profile. When CurrentScore(x) is greater than the set threshold θ, x is included in the current patient status profile.

[0175] The counterfactual error correction algorithm for treatment plans includes the following steps:

[0176] S8.31. Generate doctor's treatment plan path, converting the doctor's current treatment plan into a treatment plan path:

[0177] PathDoctor={DiagnosisDoctor, ExamDoctor, DrugDoctor, SurgeryDoctor,TreatmentDoctor, FollowupDoctor}

[0178] S8.32. Generate counterfactual safety solution paths. Based on the patient's current status profile, medical knowledge base, clinical pathway database, and drug knowledge base, generate one or more counterfactual safety solution paths:

[0179] PathSafe={DiagnosisSafe, ExamSafe, DrugSafe, SurgerySafe, TreatmentSafe, FollowupSafe};

[0180] S8.33. Perform path difference calculation to determine the difference between the doctor's proposed path and the counterfactual safety proposed path:

[0181] Diff(PathDoctor, PathSafe)={DiffDiagnosis, DiffExam, DiffDrug,DiffSurgery, DiffTreatment, DiffFollowup}

[0182] Among them, DiffDiagnosis represents diagnostic differences, DiffExam represents examination differences, DiffDrug represents medication differences, DiffSurgery represents surgical differences, DiffTreatment represents treatment pathway differences, and DiffFollowup represents follow-up differences.

[0183] S8.34. Locate the error type and locate the error in the treatment plan based on the difference type;

[0184] S8.35. Calculate the counterfactual risk contribution. For each difference item di, calculate its risk contribution:

[0185] RiskContribution(di)=Severity(di)×EvidenceStrength(di)×PatientRisk(di)×RuleConflict(di)

[0186] Among them, Severity (di) indicates the severity of the difference itself, Evidence Strength (di) indicates the strength of the original evidence, Patient Risk (di) indicates the patient's basic risk, and Rule Conflict (di) indicates the degree of conflict with medical rules;

[0187] S8.36. Output error correction suggestions, select the difference items whose RiskContribution exceeds the threshold as error correction items, and generate error correction suggestions;

[0188] The bidirectional evidence chain verification algorithm includes the following steps:

[0189] S8.41, Extracting protocol elements: Extracting protocol elements from the doctor's protocol:

[0190] PlanElement={p1, p2,...,pn}

[0191] For example, diagnosis, medication, examination, surgery, and treatment plans;

[0192] S8.42. Search for supporting evidence. For each protocol element pi, search for supporting evidence in the patient timeline diagnosis and treatment atlas and the chain of evidence index: SupportEvidence(pi).

[0193] S8.43. Search for opposing evidence, and simultaneously search for opposing evidence: OpposeEvidence(pi)

[0194] S8.44 Calculate the evidence balance score.

[0195] EvidenceBalance(pi)=SupportScore(pi)-OpposeScore(pi),

[0196] Among them: SupportScore(pi)=ΣCredibility(supporti),

[0197] OpposeScore(pi)=ΣCredibility(opposei);

[0198] S8.45. Generate evidence-based error correction conclusions. If EvidenceBalance(pi) is lower than the threshold, output: "The current protocol element lacks sufficient supporting evidence or there is significant opposing evidence." If OpposeScore is significantly higher than SupportScore, output: "The current protocol element conflicts with medical record evidence."

[0199] The algorithm for proactively detecting missing diagnostic errors includes the following steps:

[0200] S8.51. Determine the patient's current clinical scenario. Based on the patient's profile, determine the current clinical scenario: Scene = f(Profile).

[0201] S8.52. Generate a set of items to be executed, based on the current clinical scenario: RequiredSet(Scene) = {r1, r2, ..., rn}

[0202] S8.53. Extract the doctor's executed or planned items. Extract the set of executed items from the doctor's plan and patient time series graph: PerformedSet={p1, p2, ..., pm}

[0203] S8.54 Calculate the set of missing items: MissingSet = RequiredSet(Scene) - PerformedSet

[0204] S8, 55, Missing Item Risk Filtering: Not all missing items require alerts; therefore, the missing item risk is calculated based on patient risk and item necessity.

[0205] MissingRisk(r)=Necessity(r)×PatientRisk(Profile)×TimeUrgency(r)×EvidenceStrength(Scene);

[0206] If MissingRisk(r) exceeds the threshold, a missing error correction prompt is generated.

[0207] The algorithm for dynamically calculating the deviation of the treatment pathway includes the following steps:

[0208] S8.61. Obtain the standard clinical pathway, which is represented as: PathStd={Step1, Step2, ..., Stepn}

[0209] S8.62. Perform individualized path modification based on patient profile. Based on factors such as patient age, complications, liver and kidney function, disease stage, and allergy history, the standard path is modified to generate an individualized reference path: PathPersonal=Adjust(PathStd, Profile)

[0210] S8.63 Calculate node matching degree, NodeMatch = |PathDoctor∩PathPersonal| / |PathPersonal|

[0211] S8.64. Calculate the order match degree to determine whether the order of treatment steps in the doctor's plan conforms to the individualized reference path: OrderMatch = CorrectOrder / TotalOrder

[0212] S8.65. Calculate the time match to determine whether the time arrangement for treatment, examination, follow-up, and re-examination is reasonable: TimeMatch = CorrectTime / TotalTime

[0213] S8.66 Calculate the intensity matching degree to determine whether the treatment intensity matches the severity of the patient's condition:

[0214] IntensityMatch=f(TreatmentIntensity,DiseaseSeverity)

[0215] S8.67 Calculate the overall deviation score: DeviationScore = 1 - [aNodeMatch + bOrderMatch + cTimeMatch + dIntensityMatch];

[0216] Where a, b, c, and d are weights; when the DeviationScore exceeds the threshold, a path deviation warning is output.

[0217] The hierarchical risk gating algorithm includes the following steps:

[0218] S8.71 Calculate the overall alert strength: AlertScore = aRiskScore + bEvidenceStrength + cPatientRisk + dRuleScore - eFalseAlarmRate;

[0219] Where: FalseAlarmRate represents the historical false alarm rate of this type of rule, and a, b, c, d, and e are weights;

[0220] S8.72. Determine the alert strategy and output the alert strategy based on the AlertScore:

[0221] AlertScore ≥ 0.90: Block submission or superior review;

[0222] 0.75≤AlertScore<0.90: Strong alert;

[0223] 0.55≤AlertScore<0.75: Normal alert;

[0224] 0.35≤AlertScore<0.55: Weak alert;

[0225] AlertScore < 0.35: No pop-up window will be displayed; only the log will be recorded.

[0226] S8.73, Special High-Risk Rules Mandatory Upgrade: For high-risk issues such as drug allergies, severe drug interactions, inconsistencies between the left and right sides of surgery, and untreated critical values, even if the historical false alarm rate is high, the alert level will not be lower than the strong alert level.

[0227] When abnormal situations occur, diagnostic error correction prompts are generated, such as the current diagnosis lacking supporting symptoms, signs, laboratory or imaging evidence, or the current diagnosis being based only on suspected symptoms and not forming a definitive diagnosis. For example, if a doctor records "pulmonary infection" but the patient has no fever, elevated white blood cell count, elevated C-reactive protein, imaging evidence of inflammation, or anti-infection treatment records, the system will prompt "insufficient evidence for pulmonary infection diagnosis." When the system identifies a patient with postoperative infection risk but the doctor has not arranged for follow-up inflammatory marker testing or anti-infection treatment, it will prompt "suspected omission of key treatment steps." The patient's allergy history indicates a "cephalosporin allergy." When a doctor prescribes cephalosporin antibiotics, the system generates a high-risk error correction prompt and displays the original allergy history as evidence. If the patient is scheduled for surgery but lacks coagulation function, electrocardiogram, or anesthesia assessment, the system prompts "Necessary preoperative examinations missing." If an imaging report indicates a lesion on the left side, but the surgical application specifies a right-side surgical site, the system prompts "Suspected inconsistency in surgical site," and generates an error correction suggestion: "The patient's admission record indicates an allergy to cephalosporins. The current prescription includes ceftriaxone, indicating a high-risk drug conflict. It is recommended to change the antibiotic or supplement the allergy assessment and the reasons for continuing use."

[0228] The doctor feedback-driven adaptive weight update algorithm includes the following steps:

[0229] S9.1, Collect doctor feedback;

[0230] S9.2. The effectiveness of feedback should be judged in conjunction with the results of subsequent disease progression. If the doctor ignores the high-risk warning but a related adverse event occurs later, the rule should not be simply downgraded. If the doctor marks a false alarm and there is no subsequent risk event, and the senior doctor approves, the weight of the corresponding rule can be reduced.

[0231] S9.3 Calculate the rule performance score. For rule r, calculate:

[0232] RulePerformance(r)=TP / (TP+FP+FN+ε)

[0233] Where TP represents a correct alert, FP represents a false alarm, FN represents a missed alert, and ε is a constant to prevent the denominator from being zero.

[0234] S9.4 Update the rule weights. The rule weights are updated as follows: Wr(new) = μWr(old) + (1-μ)RulePerformance(r), where μ is the historical weight retention coefficient.

[0235] S9.5 Update algorithm weights: If a certain type of error correction has a high false alarm rate over a long period, reduce its low-risk alert weight; if a certain type of error correction leads to adverse consequences due to missed reports, increase the weight of that type of rule.

[0236] Workflow: Acquire multi-source medical text data, including outpatient medical records, admission records, laboratory reports, medical records, doctor-initiated input, and patient-initiated input. Simultaneously, record patient identifiers, document types, document generation times, department information, doctor information, and data sources. Perform standardized preprocessing on the acquired multi-source medical text data to obtain medical text in a unified format. Extract medical entities from the medical text. Perform medical entity recognition on the standardized medical text, extracting core medical entities from the text. For each medical entity, record its source document, paragraph, original text segment, character position, extraction confidence, etc. Identify the semantic state of the medical entities. Perform semantic state recognition on the extracted medical entities to determine the true clinical meaning represented by the medical entity in the text. Through semantic state recognition, avoid misclassifying negative diseases, suspected diseases, and past diseases as current confirmed diseases. Extract treatment events and determine event times. Based on medical entities and their contextual relationships, extract patient treatment events and determine the occurrence time of each treatment event. Construct a patient time-series treatment map. Using a chronological approach, the system integrates treatment times extracted from different medical texts to construct a patient time-series treatment graph. This graph reconstructs the complete treatment process from admission to discharge, establishing an original evidence chain index for each medical entity, treatment event, and graph relationship. Doctors can use this index to view the original medical records corresponding to system error correction suggestions, avoiding black-box prompts and allowing for correction of treatment plans. Doctors input their treatment plans for patients, and the system retrieves text data based on the patient's condition to compare the treatment plans and provide feedback to the doctor. The system provides error correction reminders for treatment plans, displays the error correction structure to doctors and receives feedback, presents error correction suggestions to doctors, and doctors process these suggestions, performing safety error correction and risk level calculations on the medical plans. Doctors then return treatment plans, overwriting the original system information with the new plans and calculating the safety risk level of the plans. The system is optimized based on doctor feedback, incorporating the doctor's feedback results into a low-confidence sample pool and a rule optimization sample pool for subsequent system optimization. This process forms a closed loop by inputting diagnostic results and procedures into the system, with doctors inputting the diagnostic process and results into the system.

[0237] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A deep semantic analysis and intelligent query platform for medical texts, comprising an intelligent query system, characterized in that, The intelligent query system performs parsing and querying using the following methods: S1. Acquire multi-source medical text data, including outpatient medical records, admission records, laboratory reports, medical records, doctor-initiated input, patient-initiated input, etc. While acquiring medical text data, simultaneously record patient identifiers, document types, document generation time, department information, doctor information, data sources, etc. S2. Standardize and preprocess the medical text data. Standardize the acquired multi-source medical text data to obtain medical text in a unified format. S3. Extract medical entities from medical texts. Perform medical entity recognition on standardized medical texts, extract core medical entities from the text, and record the source document, paragraph, original text fragment, character position, extraction confidence, etc. for each medical entity. S4. Identify the semantic state of medical entities. Perform semantic state identification on the extracted medical entities to determine the true clinical meaning of the medical entity in the text. Through semantic state identification, avoid misjudging negative diseases, suspected diseases, past diseases, etc. as currently diagnosed diseases. S5. Extract diagnosis and treatment events and determine the event time. Based on the medical entity and its context, extract the patient's diagnosis and treatment events and determine the occurrence time of each diagnosis and treatment event. S6. Construct a patient time-series diagnosis and treatment map. Taking patients as the unit, merge the diagnosis and treatment times extracted from different medical texts in chronological order to construct a patient time-series diagnosis and treatment map. The time-series diagnosis and treatment map can be used to reconstruct the complete diagnosis and treatment process of patients from admission to discharge. S7. Establish an original evidence chain index. For each medical entity, diagnosis and treatment event and graph relationship, an original evidence chain index is established. Doctors can view the original medical records corresponding to the system's error correction suggestions through the original evidence chain index to avoid black box prompts. S8. Correcting doctors' treatment plans: Doctors input treatment plans for patients, and the system retrieves text data based on the patient's condition to compare the treatment plans and provides correction reminders for the doctors' treatment plans. S9. Show the doctor the error correction structure and receive feedback. Show the doctor the error correction suggestions for the treatment plan, and the doctor will process the error correction suggestions. S10. Perform safety error correction and risk level calculation on the medical plan. The doctor returns the treatment plan, and the original system information is overwritten according to the new treatment plan. The safety risk level of the treatment plan is calculated. S11. Optimize based on doctor feedback, and include the doctor feedback results into the low confidence sample pool and the rule optimization sample pool for subsequent system optimization. S12. The diagnostic results and process are input into the system to form a closed loop, and the doctor inputs the diagnostic process and results into the system.

2. The medical text deep semantic parsing and intelligent query platform according to claim 1, characterized in that: The standardized preprocessing includes: removing invalid characters and garbled text, unifying text encoding format, identifying document type, dividing medical record paragraphs, extracting titles, body text, tables and examination result fields, expanding medical abbreviations, standardizing synonymous medical terms, normalizing time expressions, and standardizing encoding mapping for content such as diseases, surgeries, tests, and drugs. Standardization preprocessing of medical text data includes the following steps: S201. Clean the text, standardize the format, and correct garbled characters; S202. Relying on a large medical model, standardize the mapping of professional terminology and automatically connect to international medical standard coding systems such as ICD-10, SNOMED CT, and LOINC. Eliminate differences in writing formats and expression habits among different physicians, and achieve standardized and unified processing of medical record text data throughout the hospital and its entire lifecycle.

3. The medical text deep semantic parsing and intelligent query platform according to claim 1, characterized in that: The medical entities include: disease diagnosis, symptoms, signs, laboratory indicators, imaging findings, pathological results, drug names and dosages, surgical procedures, complications, adverse reactions, allergy history, family history, and past medical history, including the following steps: S301. Based on large-scale model deep semantic understanding technology, accurately extract core medical entities from medical record texts, extract entities such as diseases, symptoms, signs, tests, examinations, drugs, surgeries, complications, allergy history, and past medical history, and combine them into diagnosis and treatment events. S302. Automatically identify hidden medical information, symptom evolution logic, and medication sequence relationships in text.

4. The medical text deep semantic parsing and intelligent query platform according to claim 1, characterized in that: The semantic states include affirmative state, negative state, suspected state, excluded state, past state, family history state, planned state, preoperative state, postoperative state, and outcome state.

5. The medical text deep semantic parsing and intelligent query platform according to claim 1, characterized in that: The patient time-series diagnostic and treatment atlas includes patient nodes, disease nodes, symptom nodes, examination nodes, laboratory test nodes, drug nodes, surgical nodes, and prognostic nodes. The relationships within the atlas include chronological relationships, diagnostic basis relationships, treatment relationships, examination relationships, drug effect relationships, adverse drug reaction relationships, and complication relationships. The atlas includes the following steps: S601. Calculate the joint confidence level for each entity and event, excluding data interference from denial, family history, past history, time expiration, or conflicting evidence. S602. Generate a current patient profile based on the patient’s time-series diagnosis and treatment map, and determine the patient’s current diagnosis, symptoms, abnormal test results, medications, allergy history, complications, surgical history and risk factors.

6. The medical text deep semantic parsing and intelligent query platform according to claim 1, characterized in that: The process of correcting doctors' treatment plans includes the following steps: S801. Obtain the doctor's current treatment plan. Obtain the doctor's current proposed diagnosis, examination, medication, treatment, surgery and follow-up plan from the doctor's workstation, prescription system, medical order system or surgical application system. S802. Generate a patient status profile through a dynamic patient status profile generation algorithm. Based on the patient-level time-series diagnosis and treatment map, extract the patient's current diagnosis, symptoms, test results, imaging results, medication status, allergy history, past medical history, complications, and liver and kidney function status. S803. Verify the diagnostic basis by matching the diagnosis filled in by the doctor with the symptoms, examinations, tests, imaging and pathological evidence in the patient status profile to determine whether the diagnostic basis is sufficient. S804. Perform treatment pathway verification by comparing the treatment pathway proposed by the doctor with clinical pathways, guidelines, hospital treatment standards and the patient's current disease stage to determine whether the treatment sequence, intensity and timing are reasonable. S805. Conduct medication safety verification based on the drug knowledge base, patient allergy history, liver and kidney function, age, weight, pregnancy status, current medication and past adverse reactions, and verify drug selection, dosage, frequency, route and combination relationships. S806. Conduct a reasonableness check of the inspection and examination, determine whether there are any omissions, duplications, mismatches or incorrect timing of the proposed inspection items, and determine whether a certain medical entity or diagnosis and treatment event is real and valid through the evidence-state-time joint belief propagation algorithm. S807. Perform a surgical plan suitability check to determine whether the proposed surgery matches the results of diagnosis, imaging, pathology, disease staging, and preoperative risk assessment. S808. Generate risk level: Calculate the risk level based on the degree of deviation from the treatment plan, the patient's basic risk, the degree of rule conflict, and the degree of lack of evidence. S809. Generate error correction suggestions. The error correction problem, risk level, triggering rules, patient evidence, original text fragments and suggested handling methods are output through the counterfactual error correction algorithm of the treatment plan. The judgment is made by the two-way evidence chain verification algorithm. The missing diagnosis and treatment error active discovery algorithm actively infers the diagnosis and treatment measures that should have occurred but did not occur. The standard path is dynamically adjusted according to the individual patient status by the dynamic calculation algorithm of the deviation degree of the diagnosis and treatment path. Then, it is judged whether the doctor's plan has deviated. The hierarchical risk gating algorithm determines whether to remind, how to remind, and whether to block based on the severity of risk, strength of evidence and doctor's historical feedback. S810 receives doctor feedback and performs closed-loop optimization. It collects doctors' actions regarding error correction suggestions, such as acceptance, modification, ignoring, marking false alarms, or supplementing evidence, and uses this data for subsequent model and rule optimization. It dynamically adjusts rule weights, evidence weights, and reminder strategies through an adaptive weight update algorithm driven by doctor feedback.

7. The medical text deep semantic parsing and intelligent query platform according to claim 6, characterized in that: The evidence-state-time joint confidence propagation algorithm includes the following steps: S8.11 Calculate the basic confidence score for entity extraction. For a medical entity x, obtain the basic confidence score given by the entity recognition model: Cner(x). S8.12 Calculate the semantic state weight: Calculate the state weight based on the semantic state of entity x: StateWeight(x); S8.

13. Calculate the time validity weight. For a treatment event x, calculate the time difference Δt between it and the time when the current treatment plan was generated, and calculate the time validity weight: TimeWeight(x) = exp(-λΔt), Where λ is the time decay coefficient; S8.14 Calculate document source reliability. Different medical documents have different reliability for different medical facts. Set document source reliability: SourceWeight(x). If the entity is "pathological diagnosis" and comes from a pathology report, then SourceWeight is higher; if the entity is "pathological diagnosis" but only comes from a paraphrase in the medical record, then SourceWeight is lower. S8.15 Calculate the rule consistency weight. Perform a consistency check between entity x and the medical knowledge rules. If the entity or event is consistent with the medical rules, the RuleWeight is higher; if there is a conflict, the RuleWeight is lower. S8.16 Calculate the conflict penalty term: If there is evidence in the same patient atlas that contradicts entity x, then calculate the conflict penalty: ConflictPenalty(x). S8.17 Calculate the joint confidence level, and finally calculate the effective confidence level of the entity or time: Credibility(x)=σ[aCner(x)+bStateWeight(x)+cTimeWeight(x)+dSourceWeight(x)+eRuleWeight(x)-fConflictPenalty(x)] Where σ represents the normalization function, and a, b, c, d, e, and f are weighting coefficients.

8. The medical text deep semantic parsing and intelligent query platform according to claim 7, characterized in that: The algorithm for generating dynamic patient profiles includes: S8.21, Candidate State Extraction: Extract candidate state items from the patient-level time-series diagnosis and treatment graph: Candidate={x1, x2, ..., xn}, where each candidate state item corresponds to a medical entity or diagnosis and treatment event; S8.

22. For each candidate state item x, calculate the current valid score: CurrentScore(x)=αCredibility(x)+βTimeWeight(x)+γContinuity(x)+δSeverity(x)+εSourceWeight(x); Where Credibility(x) represents the joint confidence level, TimeWeight(x) represents the time validity, Continuity(x) represents the state continuity, Severity(x) represents the severity, and SourceWeight(x) represents the source reliability. S8.23, Calculation of state continuity: For continuous states such as chronic diseases, long-term medication, allergy history, and previous surgical history, calculate the state continuity: Continuity(x) = 1 - DiscontinueEvidence(x), where DiscontinueEvidence(x) represents the evidence of state termination. S8.

24. State Conflict Resolution: If multiple conflicting pieces of evidence exist for the same state item, they are resolved based on the strength of the evidence. The system judges the conflict resolution based on document source, chronological order, and strength of evidence. The conflict resolution score is as follows: ResolveScore(x)=Credibility(x)+SourceWeight(x)+TimeWeight(x)-ConflictPenalty(x) Select the state with the highest ResolveScore as the current state. S8.

25. Generate a current patient status profile. When CurrentScore(x) is greater than the set threshold θ, x is included in the current patient status profile.

9. A medical text deep semantic parsing and intelligent query platform according to claim 6, characterized in that: The counterfactual error correction algorithm for the treatment plan includes the following steps: S8.

31. Generate doctor's treatment plan path, converting the doctor's current treatment plan into a treatment plan path: PathDoctor={DiagnosisDoctor, ExamDoctor, DrugDoctor, SurgeryDoctor,TreatmentDoctor, FollowupDoctor} S8.

32. Generate counterfactual safety solution paths. Based on the patient's current status profile, medical knowledge base, clinical pathway database, and drug knowledge base, generate one or more counterfactual safety solution paths: PathSafe={DiagnosisSafe, ExamSafe, DrugSafe, SurgerySafe, TreatmentSafe, FollowupSafe}; S8.

33. Perform path difference calculation to determine the difference between the doctor's proposed path and the counterfactual safety proposed path: Diff(PathDoctor, PathSafe)={DiffDiagnosis, DiffExam, DiffDrug,DiffSurgery, DiffTreatment, DiffFollowup} Among them, DiffDiagnosis represents diagnostic differences, DiffExam represents examination differences, DiffDrug represents medication differences, DiffSurgery represents surgical differences, DiffTreatment represents treatment pathway differences, and DiffFollowup represents follow-up differences. S8.

34. Locate the error type and locate the error in the treatment plan based on the difference type; S8.

35. Calculate the counterfactual risk contribution. For each difference item di, calculate its risk contribution: RiskContribution(di)=Severity(di)×EvidenceStrength(di)×PatientRisk(di)×RuleConflict(di) Among them, Severity (di) indicates the severity of the difference itself, Evidence Strength (di) indicates the strength of the original evidence, Patient Risk (di) indicates the patient's basic risk, and Rule Conflict (di) indicates the degree of conflict with medical rules; S8.

36. Output error correction suggestions, select the difference items whose RiskContribution exceeds the threshold as error correction items, and generate error correction suggestions; The bidirectional evidence chain verification algorithm includes the following steps: S8.41, Extracting protocol elements: Extracting protocol elements from the doctor's protocol: PlanElement={p1, p2,...,pn} For example, diagnosis, medication, examination, surgery, and treatment plans; S8.

42. Search for supporting evidence. For each protocol element pi, search for supporting evidence in the patient timeline diagnosis and treatment atlas and evidence chain index: SupportEvidence(pi). S8.

43. Search for opposing evidence, and simultaneously search for opposing evidence: OpposeEvidence(pi) S8.44 Calculate the evidence balance score. EvidenceBalance(pi)=SupportScore(pi)-OpposeScore(pi), Among them: SupportScore(pi)=ΣCredibility(supporti), OpposeScore(pi)=ΣCredibility(opposei); S8.

45. Generate evidence-based error correction conclusions. If EvidenceBalance(pi) is lower than the threshold, output: "The current protocol element lacks sufficient supporting evidence or there is obvious opposing evidence." If OpposeScore is significantly higher than SupportScore, output: "The current protocol element conflicts with medical record evidence." 10. A medical text deep semantic parsing and intelligent query platform according to claim 6, characterized in that: The doctor feedback-driven adaptive weight update algorithm includes the following steps: S9.1, Collect doctor feedback; S9.

2. The effectiveness of feedback should be judged in conjunction with the results of subsequent disease progression. If the doctor ignores the high-risk warning but a related adverse event occurs later, the rule should not be simply downgraded. If the doctor marks a false alarm and there is no subsequent risk event, and the senior doctor approves, the weight of the corresponding rule can be reduced. S9.3 Calculate the rule performance score. For rule r, calculate: RulePerformance(r)=TP / (TP+FP+FN+ε) Where TP represents a correct alert, FP represents a false alarm, FN represents a missed alert, and ε is a constant to prevent the denominator from being zero. S9.4 Update the rule weights. The rule weights are updated as follows: Wr(new) = μWr(old) + (1-μ)RulePerformance(r), where μ is the historical weight retention coefficient. S9.5 Update algorithm weights: If a certain type of error correction has a high false alarm rate over a long period, reduce its low-risk alert weight; if a certain type of error correction leads to adverse consequences due to missed reports, increase the weight of that type of rule.

Images