Inpatient electronic medical record automatic generation method based on ai and data fusion

Abstract

The application discloses an inpatient electronic medical record automatic generation method based on AI and data fusion, and relates to the technical field of clinical medicine. The inpatient electronic medical record automatic generation method regards different medical orders, signs, tests and examinations as single documents, and distributes and updates the collected documents according to time sequence time windows. The initial inpatient electronic medical record is automatically generated through different time sequences and different medical record types. The initial inpatient electronic medical record is taken as input data, the summary generated by the doctor according to the patient's complaint and physical examination content is combined, the medical vertical field large model is used to summarize the patient's disease progression and give the corresponding treatment scheme, and the corresponding type of inpatient electronic medical record is generated according to the subjective medical examination manually input by the doctor.

Description

Technical Field

[0001] This invention relates to the field of clinical medical technology, specifically to a method for automatically generating inpatient electronic medical records based on AI and data fusion. Background Technology

[0002] Currently, doctors need to make daily rounds to obtain the latest changes in the condition of hospitalized patients and analyze the test results to form hospitalization records. A complete hospitalization record has corresponding rules. Doctors need to first comprehensively analyze the condition based on the patient's daily chief complaint, physical examination report, vital signs (blood pressure, heart rate, respiration, pulse, etc.) recorded by nurses, reports issued by the laboratory department, reports issued by the examination department, and hospitalization records, and then conduct further tests and analyses. Each step needs to be recorded in detail, and the purpose of laboratory reports and medications needs to be analyzed and explained. Finally, the doctor will prescribe, stop, or adjust the medication method and dosage. This is done by manually typing or copying and pasting templates to write a single traditional inpatient electronic medical record.

[0003] The scoring rules for writing hospital records vary across different provinces, municipalities, and autonomous regions. Doctors can only complete the writing of medical records within the prescribed time limit according to local rules. Through a series of operations such as using templates, the efficiency is low and the content is highly homogenized.

[0004] Because the development and treatment of a patient's disease is a unidirectional process that follows a time sequence, each record written during hospitalization must refer to and analyze all previous records. However, in busy or urgent situations, doctors may not have time to review past data, or the copying and pasting of templates may result in the loss of details about the patient's condition in the medical records. As a result, doctors can only use medications appropriate for the current condition, which may conflict with the correct medications that should be used according to the progression of the disease, thus affecting the optimal time for treatment. Summary of the Invention

[0005] The purpose of this invention is to generate inpatient electronic medical records in a dynamic chronological order. These records provide objective and interpretable summaries of the content processed by doctors and summarize the current progress of the patient's condition based on a large model of the medical vertical field, providing corresponding diagnosis and treatment suggestions.

[0006] To achieve the above-mentioned objectives, the present invention provides a method for automatically generating inpatient electronic medical records based on AI and data fusion, comprising the following steps: The system retrieves the patient's admission time based on preset time intervals and / or user requests, and then cycles through the corresponding patient information based on the admission time. The patient information includes name, age, gender, and nursing level. Based on patient information, medical orders, vital signs, test results, examinations, and other documents are collected and classified according to the document type conditions in the preset condition set to obtain event-type documents and non-event-type documents; The system is divided into multiple modules based on the type of medical record, and corresponding time limits are preset for different types of medical records. For event-type documents, the time variables in the preset condition set and the operation variables in the preset operation set are dynamically adjusted according to the generation time sequence of the event-type document. The corresponding operation set action instructions are executed according to the conditions in the preset condition set, and the final operation is obtained by combining the operation variables. The first initial inpatient electronic medical record of the corresponding type is generated according to the final operation. For non-event-type documents, the time variables in the preset condition set are dynamically adjusted according to the generation time sequence of the corresponding non-event-type documents. The non-event-type documents are dynamically sorted according to the time variables and filled into the module to generate the second initial inpatient electronic medical record. The contents of the first and second initial inpatient electronic medical records are used as input to the medical vertical category big model. Before generating the final medical record, the system performs clinical pathway self-verification: 1. Path tree construction: Admission diagnosis, key surgery, postoperative anti-infection, rehabilitation assessment, and discharge criteria.

[0007] 2. Matching degree calculation: Match score = 1 - (edit distance (actual path, standard path) / standard path length).

[0008] 3. Deviation Warning: If the matching degree is less than 0.85 or a critical node is missing: generate a three-level alarm: interface pop-up (level 1) → SMS reminder (level 2) → system lock (level 3).

[0009] 4. Differences Knowledge Package:

[0010]

[0011] Output medication analysis, diagnostic basis, and differential diagnosis, and fill them into the module to obtain the final inpatient electronic medical record.

[0012] Preferably, an alarm is triggered when the timer takes longer than a set time interval to acquire patient information, prompting the user to reset the time interval.

[0013] Preferably, the event-type documents include documents corresponding to key medical events such as surgery, consultation, transfer to another department, and blood transfusion.

[0014] Preferably, the medication analysis is obtained by analyzing the chief complaint during ward rounds, medication orders, general information, present medical history, vital signs, physical examination, and laboratory reports in the first and second initial inpatient electronic medical records based on a medical knowledge graph and a drug database.

[0015] Preferably, the characteristic information of the first initial inpatient electronic medical record and the second initial inpatient electronic medical record is extracted, and the diagnostic basis is obtained by searching the medical knowledge base using a medical knowledge graph in conjunction with the preliminary diagnosis.

[0016] Preferably, based on characteristic information and knowledge graph retrieval of medical knowledge base, a first strategy candidate disease group is obtained; based on characteristic information, a second strategy candidate disease group is obtained by screening for similar diseases; based on characteristic information, a third strategy candidate disease group is obtained by screening for text semantic similarity recognition; the first strategy candidate disease group, the second strategy candidate disease group, and the third strategy candidate disease group are combined through a simulation of the reasoning process of medical experts, taking into account the individual characteristics of patients and disease characteristics, and finally outputting the differential diagnosis of diseases.

[0017] Preferably, in the event-type document classification step, a dynamic weight value is assigned to each type of document, wherein the weight of surgical documents is W0=1.0, consultation documents is W0=0.9, laboratory documents is W0=0.6, and vital sign documents is W0=0.3; a weight decay function is set: W(t)=W0×e^(-λ·Δt), where Δt is the difference between the document generation time and the current recording time, the decay coefficient of critical value documents is λ=0.1, and that of routine documents is λ=0.3; high-weight documents W>0.8 are preferentially processed as event-type documents. Preferably, in the non-event-based document filling step: when the numerical difference between two documents in the same module exceeds the dynamic threshold σ, and the sum of the weights of the two documents is greater than 1.0, clinical conflict analysis is triggered. The dynamic threshold is calculated as σ = μ ± 3δ, where μ is the historical mean and δ is the standard deviation. The multimodal knowledge enhancement engine is automatically invoked to generate a conflict analysis report; the current filling operation is paused until the doctor confirms the treatment plan.

[0018] Preferably, a drug action chain reasoning mechanism is established: drug, target, pathological mechanism, contraindications; when a nephrotoxic drug is detected, the patient's creatinine clearance rate data is automatically associated; medication recommendations containing evidence-based medicine are output, and the evidence level and source guidelines are marked.

[0019] Preferably, a guideline compliance scoring mechanism is added, and the calculation formula is: matching degree = number of matching features / total number of features × 100%; the name of the clinical guideline on which the diagnosis is based and the matching degree percentage are marked in the diagnostic criteria.

[0020] Preferably, before generating the final inpatient electronic medical record, a clinical pathway tree is constructed, with the root node being the admission diagnosis and the leaf nodes being the treatment goals. The matching degree between the current medical record and the standard path is calculated: matching degree = 1 - (edit distance / standard path length). When a key node is missing and the matching degree is <85%, a level 3 alarm (sound / pop-up / SMS) is triggered.

[0021] Preferably, modifications made by doctors to system-generated content automatically generate a discrepancy knowledge package, which includes the modified content, the reason for the modification, and a timestamp. This package is then reviewed by a medical expert panel and updated to the medical knowledge base. Preferably, key conclusions in the final inpatient electronic medical record are marked with a source tracing tag, in the following format: [Conclusion Type] Content; [Level of Evidence] Level; [Tracing the Source] Knowledge Source; [Path Matching] Compliance Rate Percentage.

[0022] This invention proposes an automatic generation method for inpatient electronic medical records based on AI and data fusion. Using the patient's hospitalization to discharge as a static timeline, different medical orders, physical signs, tests, and examinations are treated as documents. The collected documents are allocated and updated according to time windows. Automatic initial inpatient electronic medical records are generated through different time sequences and different medical record types. Using the initial inpatient electronic medical records as input data, a summary is generated by combining the patient's chief complaint and physical examination content updated by the doctor. A large-scale model in the medical vertical field summarizes the patient's condition progression and provides the current corresponding treatment plan. Finally, the method combines the subjective medical examinations manually entered by the doctor to generate corresponding types of inpatient electronic medical records.

[0023] Through technological innovations in intelligent weighted time-series processing, multimodal knowledge fusion, and real-time clinical pathway verification, a leapfrog development from "automated medical record generation" to "intelligent support for clinical decision-making" has been achieved while remaining fully compatible with existing medical workflows. All technological improvements are implemented in a modular manner, retaining the core advantages of the original system while significantly enhancing medical quality and efficiency through deep AI empowerment, providing a new generation of electronic medical record system paradigms for the construction of smart hospitals. Attached Figure Description

[0024] Figure 1 The medication analysis module provided in this embodiment of the invention analyzes the effects of the medication to find the chief complaint, physical examination, test, and examination logic diagram required to form a diagnosis or to constitute such a diagnosis. Figure 2 A logic diagram for tracing drug action to obtain recommended medication based on chief complaint, physical examination, test results, and examinations, provided for embodiments of the present invention; Figure 3 A flowchart of the medication analysis module provided in an embodiment of the present invention; Figure 4 A flowchart of the diagnostic basis module provided in the embodiments of the present invention; Figure 5 The flowchart of the differential diagnosis module provided in the embodiment of the present invention is shown. Detailed Implementation

[0025] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0026] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0027] This invention provides a method for automatically generating inpatient electronic medical records based on AI and data fusion, comprising two main steps: medical record generation and content analysis. The specific steps for generating medical records are as follows: After a patient is admitted, a timer retrieves the patient's admission time based on a preset time interval and / or a user-initiated request. The timer then iterates through the corresponding patient information, including name, age, gender, and nursing care level. The preset time interval is determined by the amount of patient data; it should be greater than or equal to the time consumed by the timer in retrieving patient information. The preset time interval can be set to 24 hours, 1 hour, 30 minutes, etc. An alarm is triggered when the timer takes longer than the set time interval to retrieve patient information in a single instance, prompting the user to reset the time interval.

[0028] The task executor is bound to the timer and starts automatically after the time polling operation is completed. The task executor includes a type editor, a rule engine, a document collector, a document sorter, and a sequencer.

[0029] The document collector will poll patient data using a timer function, collecting documents such as patient orders, vital signs, laboratory tests, and examinations. To optimize the clinical priority of document processing, this system can be equipped with a dynamic weight allocation mechanism. This mechanism, while retaining the original document classification, assigns differentiated weight values ​​to various types of medical data: W0=1.0 (highest priority), consultation report W0=0.9, critical value test report (e.g., white blood cell count >12×10⁻⁶). 9 / L) W0=0.8, routine inspection documents W0=0.6, physical examination documents W0=0.3, weights dynamically decay over time: W(t)=W0×e^(-λ·Δt), where: Δt = Current record time - Document generation time (hours), critical document attenuation coefficient λ = 0.1 (slow attenuation), regular document λ = 0.3 (rapid attenuation).

[0030] Integration mechanism with the original system: High-weight documents (W(t)>0.8) are automatically included in the event-based document processing flow, while other documents maintain their original classification logic. This enhancement module does not change the basic classification rules, but only optimizes the priority of processing through weight optimization.

[0031] Based on the document type conditions in the preset condition set, documents are classified into event-type documents and non-event-type documents. Event-type documents mean that once such documents are collected, a separate medical record needs to be generated for recording, while non-event-type documents do not.

[0032] Event-based data includes critical medical events such as surgery, consultations, department transfers, and blood transfusions, while non-event-based data includes objective medical data such as body temperature, heart rate, and medication information. Each collected document will have a unique ID and will be labeled "unused" by default.

[0033] To improve data consistency, the document sorter integrates a clinical conflict resolution engine. When non-event-based documents are populated into the medical record module, conflict detection is automatically performed. If there are two documents (value A, value B) in the same module: the difference value D = |value A - value B|, the dynamic threshold σ = historical mean μ ± 3 × standard deviation δ, if D > σ and (W(A) + W(B)) > 1.0: activate the multimodal knowledge engine to retrieve evidence-based medicine evidence and generate a "Clinical Data Conflict Report" example: "abnormal combination of blood pressure 160 / 100 mmHg (weight 0.8) and heart rate 55 bpm (weight 0.7)" pause the current sorting process and wait for doctor confirmation.

[0034] Features: Triggered only when there is a significant difference in high-weight document combinations (ΣW>1.0), threshold dynamic calculation is based on departmental historical data, and related module filling is automatically paused before doctor confirmation.

[0035] For event-type documents, the corresponding action instructions of the operation set are executed according to the conditions in the preset condition set. The type editor executes the action instructions to achieve independent recording. For example, if the condition in the preset condition set is that the quality control scoring item requires that the blood transfusion record be completed within 24 hours after each blood transfusion, then the "Blood Transfusion" document is set as an event-type document. The preset condition set is [Document = "Blood Transfusion"] and [Blood Transfusion End Time < Blood Transfusion Record Time < (Blood Transfusion End Time + 24h)] and [Number of Types] ≥ 0]. The corresponding operation set is [Generate Blood Transfusion Record]. That is, the blood transfusion record will be automatically generated within 24 hours after the blood transfusion ends, and the status of the document will change from "Unused" to "Used".

[0036] The rule engine categorizes medical record quality control scoring criteria into medical record type, number of types, time range, and generation time. It pre-sets a set of conditions and operations. The generation time for each medical record type cannot exceed its preset time range. The ratio of condition variables to operation variables in the set must be at least 1, and these variables can be connected using AND / OR relationships. For example, the quality control scoring criteria require the initial progress note to be completed within 8 hours of admission, and this record type must occur once during a single hospitalization. Therefore, its generation condition must satisfy [Admission Time < Record Time < (Admission Time + 8h)] and [Number of Types = 1]. The admission time to admission time + 8h is the preset time range, and the generation time is any time variable within this range. The corresponding operation set is [Generate Initial Progress Note]. If the patient meets the generation conditions in the condition set upon admission, the operation set will be executed according to the structured layout preset by the type editor, generating a record type titled "Initial Progress Note". Since the original preset module content contains empty values, the content displayed in the first medical record can be empty. It needs to wait for the document collector to collect relevant documents and then fill it in.

[0037] For non-event-based documents, all documents will be assigned to different modules by the document sorter according to the pre-defined structured layout in the type editor when the timer is triggered. For example, in a non-event-based scenario, data such as body temperature, pulse, and blood pressure collected within 8 hours of admission will be assigned to the "Physical Examination" module in the initial medical record, while medical orders such as medications collected during this period will be assigned to the treatment plan module. At the same time, the label of the assigned documents will change from "Unused" to "Used".

[0038] First, distinguish the types of medical records involved in inpatient medical records. Currently, according to national policy documents, complete inpatient electronic medical records are divided into several types, including: admission record, initial progress note, daily progress note, ward round record of senior physician, interim summary, transfer record, consultation record, invasive procedure record, preoperative summary, preoperative discussion, surgical record, first postoperative progress note, admission and discharge record within 24 hours, and discharge record. These records use the unidirectional time from admission to discharge as a static timeline and record the complete treatment process according to the different medical treatments used during hospitalization.

[0039] The type editor is used to edit the number of different record types that can be generated and the content layout framework. For example, setting the number of generated records (type count) for the initial progress note to 1 means that there will be exactly one copy of this record type during a single complete hospitalization. Alternatively, setting the number of generated records (type count) for routine progress notes to ≥0 means that this record type can be either not generated or can be generated multiple times during a single complete hospitalization. The type editor allows you to edit the structured layout of a record type according to actual needs, and freely adjust each content module. Modules include, but are not limited to, general patient information, admission complaint, ward round complaint, present illness, past medical history, vital signs, physical examination, laboratory reports, examination reports, preliminary diagnosis, diagnostic basis, differential diagnosis, treatment plan, new treatment advances, medication orders, medication analysis, treatment process, and precautions. For example, the preset structured layout of the initial progress note can include several modules such as general patient information, admission complaint, present illness, past medical history, physical examination, preliminary diagnosis, diagnostic basis, differential diagnosis, and treatment plan.

[0040] The sequencer operates synchronously when documents are sorted. Its function is to, within the linear time sequence from a patient's admission to discharge and within the limited time ranges set for different medical record types, send instructions to the rule engine for the times of non-event-type documents collected by the document collector, dynamically adjust the time variables in its condition set, sort the documents by the document sorter according to the dynamic time sequence and update the generation time, so as to ensure that all documents can be correctly assigned to different types of medical records, and avoid time paradox of documents under the linear time sequence, that is, no situation where a medical record at 10 a.m. contains a document with a time of 3 p.m. occurs. For example, the generation time of the first course record is within the time range from the admission time to admission time + 8h. Since there are null values in the original preset module content, it is necessary to wait for the document collector to collect the data required by each module. When document a1 with a time of n1 (admission time < n1 < admission time + 8h) is collected, the sequencer dynamically adjusts the time variable in the condition set to n1, and then the document sorter assigns the data to the corresponding module of the first course record, so the time of the first course record triggered by a1 is n1; when document a2 with a time of n2 (n1 < n2 < admission time + 8h) is collected, the time variable is dynamically adjusted to n2 and the document sorter continues to assign the data, so the time of the first course record is updated to n2 by a2; when document a3 with a time of n3 (admission time < n3 < n1) is collected, the sequencer does not affect the time variable in the condition set, so the medical record content updated by data a3 does not change the generation time of the first course record.

[0041] For event-type documents, while dynamically adjusting the time variables of the condition set in the rule engine, the sequencer also sends instructions to the operation set to affect the operation variables therein, and the instructions include creation, deletion and update of different medical record types.

[0042] Affecting the time variables of the condition set, for example: there is a surgical operation within 8 hours after the patient is admitted, that is, the document collector collects a surgical document c, which is an event-type document with a document time of m1 (admission time < m1 < admission time + 8h). According to the quality control requirements, a pre-operation chief surgeon ward round record is required within 24 hours before the operation. Therefore, under the action of the sequencer, the time variable m2 (m2 < m1) in the condition set is added first, then a pre-operation chief surgeon ward round record with the time of m2 is created under the action of the operation set. The non-event-type documents subsequently collected between m2 and m1 are assigned to the pre-operation chief surgeon ward round record by the document sorter, and the time variable of the record is dynamically adjusted at the same time, so as to ensure that m2 is not less than the latest document time before m1 and avoid time paradox. Therefore, the time of the pre-operation chief surgeon ward round record triggered by c is m2, and it is continuously updated backward in the subsequent time period from m2 to m1.

[0043] The time variables of the condition set affect the operation variables of the operation set at the same time. For example: a patient has a surgical operation on the 3rd day after admission, the operation time is m3, and there is a daily ward round record before time m3 on the 3rd day, with the record time being m4. Since a chief surgeon's ward round record needs to be generated before the operation and there is a daily ward round record in the same time period at this moment, the sequencer will first add a time variable m5 (admission time + 48h < m5 < m3) to the condition set, then modify the operation variables in the operation set, delete the daily ward round record of the day, convert the "used" document in the daily ward round record back to "unused", and then create a preoperative chief surgeon's ward round record with time m5. At this time, if m5 > m4, all "unused" documents before time m5 will be sorted into this preoperative chief surgeon's ward round record, and the label will be converted to "used" at the same time; if m5 < m4, all "unused" documents before time m5 will be sorted into this preoperative chief surgeon's ward round record, and the label will be converted to "used" at the same time. If there are "unused" documents in the time period from m5 to m4, dynamic adjustment is realized under the previous effect of the sequencer on the condition variables of the condition set, until all documents are redistributed, so as to ensure that there is no time paradox situation.

[0044] Regarding content analysis, the document contents contained in different medical record types are taken as inputs and put into a medical domain-specific large model, and the final required analysis content is output through the calculation of the medical domain-specific large model and the recall path formed by the medical knowledge graph. The medical domain-specific large model includes a medication analysis module, a diagnosis basis module and a differential diagnosis module.

[0045] To enhance the decision support capability, each analysis module is connected with a multimodal knowledge engine: Medication analysis enhancement: Establish a drug action chain reasoning mechanism: drug → target → pathological mechanism → contraindication. When analyzing vancomycin medication: automatically associate the patient's renal function indicators, match the drug metabolism path and the contraindication database, and output an enhanced suggestion: "According to the creatinine clearance rate of 28ml / min, it is recommended to adjust the dose to 1g q48h".

[0046] Diagnosis basis enhancement: Add a new guideline compliance score: Matching degree = (number of guideline-conforming features / total number of features) × 100%. Output example: "The diagnosis of acute pancreatitis conforms to the 2023 version of the guideline level Ⅲ evidence (matching degree 92%)".

[0047] Differential diagnosis enhancement: Combined with the clinical path tree for reverse reasoning, the weight of candidate diseases that deviate from the standard path is automatically reduced.

[0048] The modules that serve as inputs for calculation include, but are not limited to: general patient information, chief complaint upon admission, chief complaint during ward rounds, present medical history, past medical history, vital signs, physical examination, auxiliary examinations, medication orders, and preliminary diagnosis. The modules that need to be calculated as outputs include: medication analysis, diagnostic basis, and differential diagnosis.

[0049] Firstly, the medication analysis module's main logic starts with medication, analyzing its potential effects, and then seeking the corresponding diagnosis or the chief complaint, physical examination, laboratory tests, and other necessary evidence required to establish that diagnosis (such as...). Figure 1 As shown), it also combines the capabilities of large-scale medical vertical models to provide suggestions on disease diagnosis and treatment plans (such as...). Figure 2 As shown in the figure, the process requires a knowledge graph and a drug database as dependencies.

[0050] A knowledge graph database is a database composed of triples as the smallest storage unit. That is, "entity-relationship-entity" is a triple component, such as "proteinuria-symptom-kidney disease", which means that "proteinuria" is a symptom of "kidney disease".

[0051] The drug database is a database of drug information, mainly containing drug names, drug classifications, and drug effects. Drug classifications are categorized into three levels: site of action, type of action, and drug name. For example, "Urapidil Hydrochloride Injection" (drug name) belongs to "Antihypertensive Drugs" (type of action) under "Cardiovascular System" (site of action classification), and "Nifedipine Capsules" (drug name) belongs to "Calcium Channel Blockers" (type of action) under "Cardiovascular System" (site of action classification). When performing calculations in the medication analysis module, the doctor's prescriptions are first classified by type of action, based on the drug database. Next, the drug's effect is located in the database, and then... Figure 1 As shown, the system searches for the corresponding diagnosis using a knowledge graph database, and then retrieves the corresponding patient vital signs information from four modules: ward round complaints, physical examination, laboratory reports, and examination reports. This information is then used as input to a large-scale medical vertical model, which ultimately provides medication analysis results. Figure 3 As shown. The specific workflow of the medication analysis module is as follows: After prescribing medication based on the patient's condition, the doctor receives the patient's medical history, the doctor's diagnosis, and the specific medication information.

[0052] Retrieve detailed information about medications prescribed by doctors from a pre-defined knowledge base, including the indications, dosage, and adverse reactions of the medications.

[0053] This system integrates patient medical history, physician diagnoses, and medication information, and combines these with specific prompts to generate professional medication analysis content through a large-scale medical model. This analysis includes drug suitability, potential risks, and optimization suggestions tailored to the individual patient's situation.

[0054] The diagnostic criteria module extracts patient case characteristic information and combines it with preliminary diagnosis and disease tracing diagnostic criteria in the medical knowledge graph.

[0055] For example: A patient's case features: "Female child, 1 year and 1 month old; acute onset, course of illness 1 day; her mother reported that 1 day ago the child began to have yellow, watery stools about 10 times a day without any obvious cause, small in amount, no purulent or bloody stools, vomiting gastric contents after eating, no coffee-ground material, non-projectile, no fever, chills, convulsions, cough, sputum, nasal congestion, runny nose, shortness of breath, cyanosis. She came to our department for further diagnosis and treatment, and was admitted for hospitalization in our department with a diagnosis of 'acute gastroenteritis'."

[0056] Case characteristics: The patient's mental state and sleep are fair after the illness, with a slightly poor appetite, normal urination, and unknown weight gain or loss. The patient has a history of good health. Physical examination: T 37.7℃, HR 120 bpm, R 32 bpm, weight 10.5 kg. The patient is alert but slightly lethargic, without signs of dehydration, skin elasticity is fair, development is normal, nutrition is moderate, no deformities of the head or facial features, no sunken eye sockets, rosy lips, pharyngeal congestion, supple neck, clear breath sounds in both lungs without rales, normal cardiac auscultation, soft and flat abdomen, active bowel sounds, no abnormalities found in the spine or limbs, and no abnormalities found in the neurological system. Ancillary examinations: currently unavailable. The preliminary diagnosis was "acute gastroenteritis".

[0057] The medical knowledge graph retrieves information from the medical knowledge base containing: "The following is information related to acute gastroenteritis: Symptoms of acute gastroenteritis: vomiting, diarrhea, nausea. Clinical manifestations of acute gastroenteritis: restlessness, high fever, mucus in stool, loss of appetite, vomiting of coffee-ground-like material. Signs of acute gastroenteritis: hypokalemia." The model analyzes the received three parts of information and ultimately provides the basis for the doctor's diagnosis of "acute gastroenteritis," thus explaining the logical reasoning behind the doctor's diagnosis. The specific process of the diagnostic basis is detailed below. Figure 4 As shown, the process is as follows: First, collect the patient's medical history characteristics and preliminary diagnosis. For example, enter "pneumonia (right side)" for the preliminary diagnosis.

[0058] The preliminary diagnosis is standardized by converting it into a standardized diagnostic name. For example, "pneumonia (right side)" is converted to "pneumonia". This process is achieved through pre-defined standardization rules to ensure the standardization and consistency of diagnostic information. Based on the standardized diagnostic name, medical information such as symptoms, signs, and clinical manifestations related to the disease is retrieved from a medical knowledge base using a medical knowledge graph. The knowledge base stores verified medical knowledge, providing basic data support for subsequent analysis.

[0059] Postoperative infection early warning case (enhanced function demonstration): Basic process: Surgical documents (W=1.0) trigger event-based processing, generating surgical records, and postoperative vital signs documents are sorted into daily medical records in chronological order.

[0060] Conflict detection triggered: Data collected 6 hours post-surgery, {"Type":"Temperature","Value":38.5℃,"Weight":0.42}{"Type":"White Blood Cells","Value":15×10 9 / L,"weight":0.58}, calculate the difference: |38.5-37.0|=1.5,|15.0-6.0|=9.0>threshold 8.2, the weight and 1.0 trigger contradictory analysis.

[0061] Multimodal engine response: Search evidence-based medicine databases Output: "Postoperative fever with elevated white blood cell count: Risk of infection (Level A evidence)" Clinical pathway validation: The detection lacked the "postoperative prophylactic antibiotics" node. A 72% match rate triggers a Level 2 alert. Final medical record marker: [Pathway compliance rate] 72% [Conflict Resolution] Antibiotic prescription has been supplemented (by a certain doctor).

[0062] The system integrates patient medical history characteristics, preliminary diagnosis, and disease-related information retrieved from a knowledge base. Based on this, and combined with specific prompts, a pre-trained model generates a diagnostic basis for the patient. This diagnostic basis integrates individual patient characteristics and disease-related medical information, ensuring the professionalism and relevance of the output. The final output provides professionally sound diagnostic information, offering doctors comprehensive and accurate diagnostic references to assist them in making more informed medical decisions.

[0063] The differential diagnosis module primarily outputs two aspects: 1. The disease to be identified; 2. An explanation of the identified disease in conjunction with the patient's specific circumstances. The differential diagnosis process is as follows: It receives the patient's primary diagnosis and medical history as input, including but not limited to the patient's symptoms, signs, examination results, and past medical history.

[0064] Multi-strategy disease recall generates candidate disease groups to be identified through the following three strategies: Strategy 1: Disease Recall Based on a Large-Scale Model Knowledge Base. Medical vertical models utilize the medical knowledge learned during their pre-training phase, combined with the patient's specific circumstances, to automatically output a set of diseases to be identified. This process relies on the model's understanding and reasoning abilities regarding medical knowledge.

[0065] Strategy Two: ICD-10-based disease recall. By matching ICD-10 codes in medical records, diseases of the same type as the primary diagnosis are automatically selected as candidate diseases and input into a large-scale medical model. The model then selects the most likely disease to be identified from these candidate diseases.

[0066] Strategy 3: Disease Recall Based on Text Semantic Similarity. The system utilizes text semantic similarity calculation technology to perform semantic analysis on disease names, symptom descriptions, or medical record texts to recall diseases highly related to the patient's condition. Finally, several diseases to be identified are generated as candidate groups and input into the large model. The model then selects the most likely disease based on its own logical reasoning.

[0067] Upon receiving candidate disease groups generated by the three strategies mentioned above, and combining the medical knowledge learned during the pre-training phase, the system performs comprehensive analysis and screening of the candidate diseases. By simulating the reasoning process of medical experts, and considering the individual characteristics of the patient and the characteristics of the disease, the system ultimately outputs the differential diagnosis disease that best matches the current patient's condition.

[0068] The generated differential diagnosis results are output in a structured form, providing doctors with accurate diagnostic references and assisting them in making more scientific medical decisions.

[0069] like Figure 5 As shown, taking the primary diagnosis and patient information as input, the medical vertical model outputs differential diagnoses using its three disease identification recall strategies. Strategy 1: The medical vertical model automatically outputs diseases to be identified by combining its own knowledge base with patient information. Strategy 2: Using ICD-10 encoding, the medical record system automatically inputs similar diseases as candidate diseases into the medical vertical model based on matching rules. Strategy 3: Related diseases are recalled using textual semantic similarity, resulting in five diseases to be identified. These five diseases are then input as a candidate group into the large model. The large model combines the knowledge learned during its pre-training phase to filter the diseases, thus providing the differential diagnosis that best matches the patient's current situation.

[0070] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for automatically generating inpatient electronic medical records based on AI and data fusion, characterized in that, Includes the following steps: First, the patient's admission time is obtained according to the preset time interval and the user's active request. Then, the corresponding patient information is circulated according to the admission time. The patient information includes name, age, gender, and nursing level, etc. Then, based on the patient information, medical orders, vital signs, test results, examinations, and other documents are collected and classified according to the document type conditions in the preset condition set to obtain event-type documents and non-event-type documents; Finally, it is divided into multiple modules according to the requirements of medical record type, and corresponding time limits are preset according to different medical record types; For event-type documents, the time variables in the preset condition set and the operation variables in the preset operation set are dynamically adjusted according to the generation time sequence of the event-type document. The corresponding operation set action instructions are executed according to the conditions in the preset condition set, and the final operation is obtained by combining the operation variables. The first initial inpatient electronic medical record of the corresponding type is generated according to the final operation. For non-event-type documents, the time variables in the preset condition set are dynamically adjusted according to the generation time sequence of the corresponding non-event-type documents. The non-event-type documents are dynamically sorted according to the time variables and filled into the module to generate the second initial inpatient electronic medical record. The contents of the first and second initial inpatient electronic medical records are used as input to the medical vertical category big model, and the output includes medication analysis, diagnostic basis and differential diagnosis, which are then filled into the module to obtain the final inpatient electronic medical record.

2. The method for automatically generating inpatient electronic medical records based on AI and data fusion according to claim 1, characterized in that: An alarm will be triggered when the timer takes longer than the set time interval to acquire patient information in a single instance, prompting the user to reset the time interval. The event-based documents include documents corresponding to key medical events such as surgery, consultation, transfer to another department, and blood transfusion.

3. The method for automatically generating inpatient electronic medical records based on AI and data fusion according to claim 1, characterized in that: The medication analysis was obtained by analyzing the chief complaints during ward rounds, medication orders, general information, present medical history, vital signs, physical examinations, and laboratory reports in the first and second initial electronic medical records of hospitalization, based on a medical knowledge graph and a drug database. The characteristic information of the first and second initial inpatient electronic medical records is extracted, and combined with the preliminary diagnosis, the medical knowledge graph is used to search the medical knowledge base to conduct disease tracing diagnosis to obtain the diagnostic basis.

4. The method for automatically generating inpatient electronic medical records based on AI and data fusion according to claim 1, characterized in that: Based on characteristic information and knowledge graph retrieval of medical knowledge base, a first strategy candidate disease group is obtained. Based on characteristic information, the second strategy candidate disease group is obtained by screening for similar diseases. Based on characteristic information, the third strategy candidate disease group is obtained by screening for text semantic similarity recognition. The first strategy candidate disease group, the second strategy candidate disease group, and the third strategy candidate disease group are combined through the reasoning process of simulating medical experts, taking into account the individual characteristics of patients and disease characteristics, and finally outputting the differential diagnosis.

5. The method for automatically generating inpatient electronic medical records based on AI and data fusion according to claim 1, characterized in that: In the event-based document classification process; Assign dynamic weight values ​​to each type of document, with surgical documents having a weight of W0=1.0, consultation documents having a weight of W0=0.9, laboratory documents having a weight of W0=0.6, and physical examination documents having a weight of W0=0.

3. Set the weight decay function: W(t)=W0×e^(-λ·Δt), where Δt is the difference between the document generation time and the current recording time. The decay coefficient λ=0.1 for critical documents and λ=0.3 for regular documents. Documents with high weight (W>0.8) are processed as event-type documents first. In the non-event-based document population step: When the numerical difference between two documents within the same module exceeds the dynamic threshold σ, and the sum of the weights of the two documents is greater than 1.0, clinical contradiction analysis is triggered. The dynamic threshold calculation formula is σ=μ±3δ, where μ is the historical mean and δ is the standard deviation. The multimodal knowledge enhancement engine is automatically invoked to generate a contradiction analysis report. Pause the current filling procedure until the doctor confirms the treatment plan.

6. The method for automatically generating inpatient electronic medical records based on AI and data fusion as described in claim 1, characterized in that: a drug action chain reasoning mechanism is established: drug, target, pathological mechanism, contraindications; When nephrotoxic drugs are detected, the patient's creatinine clearance rate data is automatically linked. Output medication recommendations that include evidence-based medicine, indicating the level of evidence and the source guidelines.

7. The method for automatically generating inpatient electronic medical records based on AI and data fusion according to claim 1, characterized in that: Add a guideline compliance scoring mechanism, and indicate the name of the clinical guideline on which the diagnosis is based and the percentage of matching degree in the diagnostic criteria; Before generating the final inpatient electronic medical record, construct a clinical pathway tree with the root node representing the admission diagnosis and the leaf nodes representing the treatment goals. Calculate the matching degree between the current medical record and the standard pathway. When a key node is missing and the matching degree is <85%, trigger a level 3 alarm.

8. The method for automatically generating inpatient electronic medical records based on AI and data fusion according to claim 1, characterized in that: When doctors modify the content generated by the system, a discrepancy knowledge package is automatically generated. The discrepancy knowledge package includes the modified content, the reason for the modification, and the timestamp. After being reviewed by a medical expert group, it is updated to the medical knowledge base.

9. The method for automatically generating inpatient electronic medical records based on AI and data fusion according to claim 1, characterized in that: Key conclusions in the final inpatient electronic medical record are marked with source tracing markers.

Images