Real-time triage guidance method and system based on dynamic inference engine and patient profile

Abstract

This application provides a real-time patient reception guidance method and system based on a dynamic reasoning engine and patient profiles. The method includes: real-time acquisition of audio streams from doctor-patient communication, performing speech recognition and semantic conversion, extracting semantic text such as symptom entities, and acquiring historical medical data; constructing patient profiles and identifying hierarchical tags for patient intentions; inputting the patient profiles, hierarchical tags for patient intentions, and standard operating procedures (SOPs) into a dynamic reasoning engine, performing multi-round reasoning based on a medical knowledge graph, and outputting judgment results for treatment nodes, reasoning chains, and communication risk warning information; generating standardized treatment action guidance and professional communication scripts and pushing them to the patient reception terminal; and conducting intelligent review and iterative optimization of guidance strategies after the patient reception is completed. Through the deep coupling of multimodal semantic perception, dynamic reasoning, and patient reception SOPs, accurate judgment of treatment nodes, real-time warning of communication risks, and full-link guidance and autonomous iterative optimization of the patient reception process are achieved, effectively standardizing the patient reception process.

Description

Technical Field

[0001] This application relates to the field of medical consultation, specifically to a method and system for real-time consultation guidance based on a dynamic reasoning engine and patient profiles. Background Technology

[0002] With the improvement of medical informatization, more and more medical institutions are introducing intelligent auxiliary systems to improve the efficiency and quality of diagnosis and treatment. In outpatient settings, doctors need to complete multiple steps within a limited time, including patient information collection, disease assessment, treatment plan formulation, and communication and explanation. This results in a heavy workload and is prone to fluctuations in the quality of diagnosis and treatment due to information omissions, communication difficulties, or lack of experience. Traditional electronic medical record systems only provide data recording functions and cannot provide doctors with real-time process guidance and decision support during the consultation process.

[0003] In existing technologies, some intelligent consultation systems collect patients' self-reported symptoms and use knowledge graphs or rule engines to output preliminary diagnostic suggestions. However, these systems have the following shortcomings: First, the input method is mainly text-based, failing to fully utilize the rich information in the natural voice communication between doctors and patients, resulting in low information collection efficiency and easy omission of key semantics; Second, the reasoning process is based solely on static matching of current symptoms with the knowledge base, lacking dynamic integration of patients' historical medical data and personalized characteristics, resulting in inference results that lack individualized targeting; Third, the system lacks deep integration with standard operating procedures (SOPs) for consultations, unable to determine the current stage of the consultation process, and unable to automatically adjust the scope of reasoning and guidance content based on the stage; Fourth, the output of existing systems usually only includes diagnostic suggestions, lacking standardized guidance on diagnostic and treatment actions and communication scripts for doctors, and unable to detect and warn of communication risks during the consultation process in real time; Fifth, existing systems lack intelligent review and autonomous iterative optimization capabilities after consultation, and the system performance cannot continuously improve with the accumulation of actual usage data.

[0004] Therefore, there is an urgent need for an intelligent patient reception guidance method that can comprehensively utilize multimodal speech perception, dynamic knowledge reasoning, patient reception SOP linkage, and real-time communication risk detection to fill the gaps in existing technologies in terms of full-process patient reception guidance, communication quality assurance, and system self-iterative optimization. Summary of the Invention

[0005] The purpose of this application is to provide a real-time patient reception guidance method and system based on a dynamic reasoning engine and patient profiling. This system can acquire real-time audio streams of doctor-patient communication, perform speech recognition and semantic conversion, extract semantic text such as symptom entities, and obtain historical medical data; construct patient profiles and identify hierarchical tags of treatment intentions; input the patient profiles, hierarchical tags of treatment intentions, and standard operating procedures (SOPs) into the dynamic reasoning engine, perform multi-round reasoning based on a medical knowledge graph, and output the judgment results of treatment nodes, reasoning chains, and communication risk warning information; generate standardized treatment action guidance and professional communication scripts and push them to the patient reception terminal; and perform intelligent review and iterative optimization of guidance strategies after the patient reception is completed. Through the deep coupling of multimodal semantic perception, dynamic reasoning, and patient reception SOPs, accurate judgment of treatment nodes, real-time warning of communication risks, and full-link guidance and autonomous iterative optimization of the patient reception process are achieved, effectively standardizing the patient reception process.

[0006] This application also provides a real-time patient reception guidance method based on a dynamic inference engine and patient profiles, including the following steps: Acquire patient identification data and audio stream data of doctor-patient communication, perform speech recognition and semantic conversion processing on the audio stream data, extract semantic text and standardized medical terms, and extract historical medical data from a pre-set electronic health record database in combination with patient identification data. Patient profiles are constructed based on semantic text and historical medical data, and patients' willingness to seek medical treatment is hierarchically identified based on a pre-trained willingness classification model, outputting hierarchical labels of willingness to seek medical treatment. The patient profile, the hierarchical labels of the willingness to seek medical treatment, and the preset standard operating procedures for receiving patients are input into the dynamic reasoning engine. The dynamic reasoning engine performs multiple rounds of reasoning based on the pre-trained medical knowledge graph and outputs the judgment result of the current diagnosis and treatment node, the reasoning chain, and communication risk warning information. Based on the judgment results and reasoning chain of the diagnosis and treatment nodes, standardized diagnosis and treatment action guidance and professional communication scripts are generated, and communication risk warning information and diagnosis and treatment action guidance are pushed to the patient reception terminal in real time. After the patient is admitted, the entire admission process is intelligently reviewed, a patient admission review report is generated, and the admission SOP and the guidance strategy of the inference engine are iteratively optimized based on the deviation data in the review report.

[0007] Optionally, in the real-time patient guidance method based on a dynamic inference engine and patient profile described in this application, the steps of acquiring patient identification data and audio stream data of doctor-patient communication, performing speech recognition and semantic conversion processing on the audio stream data, extracting semantic text and standardized medical terms, and extracting historical medical data from a preset electronic health record database in conjunction with patient identification data, specifically include: Audio signals from the patient reception environment are collected in real time using a microphone array, and the doctor's voice channel and the patient's voice channel are distinguished by a sound source separation algorithm. Noise filtering and speech endpoint detection are performed on the two speech streams respectively; The detected speech segments are automatically transcribed using speech recognition, and then semantic text is extracted using a preset medical named entity recognition model, including symptom entities, anatomical location entities, and time entities, and mapped to a standardized medical terminology system.

[0008] Optionally, in the real-time patient admission guidance method based on a dynamic inference engine and patient profiles described in this application, the step of constructing a patient profile based on semantic text and historical medical data, and performing hierarchical identification of patient admission intentions based on a pre-trained intention classification model, and outputting hierarchical labels for admission intentions, specifically includes: Patient profiles are constructed based on semantic text and historical medical data. These profiles include demographic characteristics, past medical history, medication records, allergy history, family history of genetic diseases, and lifestyle data. The extracted historical medical data is sorted chronologically and missing values ​​are filled to generate structured patient profile feature vectors; Semantic text is processed to obtain semantic text feature vectors, including symptom entity vectors, anatomical site entity vectors, and time entity vectors; The intention classification model takes as input a concatenation of symptom entity vectors, anatomical site entity vectors, time entity vectors, and patient profile feature vectors, and outputs a hierarchical label of medical visit intention, including one or more of the following: initial diagnosis, follow-up visit, symptom consultation, and prescription renewal.

[0009] Optionally, in the real-time patient reception guidance method based on dynamic reasoning engine and patient profile described in this application, the step of inputting patient profile, hierarchical tags of treatment intention, and preset standard operating procedures for patient reception into the dynamic reasoning engine, and the dynamic reasoning engine performing multi-round reasoning based on a pre-trained medical knowledge graph, outputting the judgment result of the current diagnosis and treatment node, the reasoning chain, and communication risk warning information, specifically includes: The dynamic reasoning engine employs a knowledge graph reasoning model based on an attention mechanism; The query vector is obtained by concatenating the patient profile feature vector and the symptom entity vector, and the association strength with the candidate disease node is calculated through a multi-head attention mechanism. The multi-head attention mechanism includes symptom combination matching analysis based on symptom and symptom co-occurrence matrix, complication probability analysis based on disease and complication causal graph, and medication conflict risk analysis based on drug and contraindication conflict detection matrix. The three analytical scores and the correlation strength are integrated into a comprehensive inference confidence score. The inference chain and the corresponding intermediate feature data of the inference chain are output according to the confidence score.

[0010] Optionally, in the real-time patient admission guidance method based on a dynamic reasoning engine and patient profile described in this application, obtaining the judgment result of the diagnosis and treatment node specifically includes: The intermediate features of the inference chain are matched with the preset SOP node sequence for receiving patients, and the judgment result of the current treatment node is output. The current diagnosis and treatment node judgment result reflects the node type of the current reception process in the SOP. The preset node types include one of the following: symptom inquiry node, physical sign collection node, preliminary diagnosis node, examination ordering node, and treatment plan formulation node. Based on the judgment result of the current diagnosis node, the corresponding SOP branch process is dynamically invoked, and the SOP branch process is used as a constraint condition input into the inference engine to limit the output range of the inference chain in the current round.

[0011] Optionally, in the real-time patient reception guidance method based on a dynamic reasoning engine and patient profiles described in this application, the output of communication risk warning information specifically includes: Communicating risk warning information includes risk type labels and corresponding response suggestions; Risk type labels include risk of missing key information, risk of abnormal patient emotions, and risk of information asymmetry between doctors and patients; When the necessary medical history information required for the reasoning chain does not appear in the patient profile or the current semantic text, the risk of missing key information is triggered. Based on a pre-set voice emotion recognition model, the patient's voice signal is classified into emotions. When the recognition result is negative emotion and the confidence level exceeds a pre-set threshold, the risk of abnormal patient emotions is triggered. When doctors use professional terminology more frequently than a preset density threshold and patients show signs of confusion in their feedback, the risk of information asymmetry between doctors and patients is triggered.

[0012] Optionally, in the real-time patient reception guidance method based on a dynamic reasoning engine and patient profile described in this application, before outputting the judgment result of the current diagnosis and treatment node, the reasoning chain, and the communication risk warning information, it further includes: Obtain the patient's historical medical visit time series data, and output the patient's disease progression risk score within a preset future time period through a pre-trained disease progression prediction model; Assess the severity of each symptom and output a symptom severity score; Symptoms with severity scores above a preset threshold are marked as high-risk symptoms and given higher weight in the inference chain; By inputting disease progression risk scores and high-risk symptom markers as contextual features into the dynamic inference engine, "urgent treatment suggestions" are automatically displayed in advance for high-risk patients.

[0013] Secondly, this application provides a real-time patient reception guidance system based on a dynamic inference engine and patient profiles. The system includes a memory and a processor. The memory stores a program for a real-time patient reception guidance method based on a dynamic inference engine and patient profiles. When the program for the real-time patient reception guidance method based on a dynamic inference engine and patient profiles is executed by the processor, it performs the following steps: Acquire patient identification data and audio stream data of doctor-patient communication, perform speech recognition and semantic conversion processing on the audio stream data, extract semantic text and standardized medical terms, and extract historical medical data from a pre-set electronic health record database in combination with patient identification data. Patient profiles are constructed based on semantic text and historical medical data, and patients' willingness to seek medical treatment is hierarchically identified based on a pre-trained willingness classification model, outputting hierarchical labels of willingness to seek medical treatment. The patient profile, the hierarchical labels of the willingness to seek medical treatment, and the preset standard operating procedures for receiving patients are input into the dynamic reasoning engine. The dynamic reasoning engine performs multiple rounds of reasoning based on the pre-trained medical knowledge graph and outputs the judgment result of the current diagnosis and treatment node, the reasoning chain, and communication risk warning information. Based on the judgment results and reasoning chain of the diagnosis and treatment nodes, standardized diagnosis and treatment action guidance and professional communication scripts are generated, and communication risk warning information and diagnosis and treatment action guidance are pushed to the patient reception terminal in real time. After the patient is admitted, the entire admission process is intelligently reviewed, a patient admission review report is generated, and the admission SOP and the guidance strategy of the inference engine are iteratively optimized based on the deviation data in the review report.

[0014] Optionally, in the real-time patient guidance system based on a dynamic inference engine and patient profiles described in this application, the steps of acquiring patient identification data and audio stream data of doctor-patient communication, performing speech recognition and semantic conversion processing on the audio stream data, extracting semantic text and standardized medical terms, and extracting historical medical data from a preset electronic health record database in conjunction with patient identification data, specifically include: Audio signals from the patient reception environment are collected in real time using a microphone array, and the doctor's voice channel and the patient's voice channel are distinguished by a sound source separation algorithm. Noise filtering and speech endpoint detection are performed on the two speech streams respectively; The detected speech segments are automatically transcribed using speech recognition, and then semantic text is extracted using a preset medical named entity recognition model, including symptom entities, anatomical location entities, and time entities, and mapped to a standardized medical terminology system.

[0015] Optionally, in the real-time patient guidance system based on a dynamic inference engine and patient profiles described in this application, the step of constructing a patient profile based on semantic text and historical medical data, and performing hierarchical identification of patient's willingness to seek medical treatment based on a pre-trained willingness classification model, and outputting hierarchical labels for willingness to seek medical treatment, specifically includes: Patient profiles are constructed based on semantic text and historical medical data. These profiles include demographic characteristics, past medical history, medication records, allergy history, family history of genetic diseases, and lifestyle data. The extracted historical medical data is sorted chronologically and missing values ​​are filled to generate structured patient profile feature vectors; Semantic text is processed to obtain semantic text feature vectors, including symptom entity vectors, anatomical site entity vectors, and time entity vectors; The intention classification model takes as input a concatenation of symptom entity vectors, anatomical site entity vectors, time entity vectors, and patient profile feature vectors, and outputs a hierarchical label of medical visit intention, including one or more of the following: initial diagnosis, follow-up visit, symptom consultation, and prescription renewal.

[0016] As described above, the real-time patient reception guidance method and system based on a dynamic reasoning engine and patient profiles provided in this application collects audio streams of doctor-patient communication in real time, performs speech recognition and semantic conversion, extracts semantic text such as symptom entities, and obtains historical medical data; constructs patient profiles and identifies hierarchical tags of treatment intentions; inputs the patient profiles, hierarchical tags of treatment intentions, and standard operating procedures for patient reception into the dynamic reasoning engine, performs multi-round reasoning based on a medical knowledge graph, and outputs judgment results of treatment nodes, reasoning chains, and communication risk warning information; generates standardized treatment action guidance and professional communication scripts and pushes them to the patient reception terminal; after the patient reception is completed, intelligent review is performed and the guidance strategy is iteratively optimized. Through the deep coupling of multimodal semantic perception, dynamic reasoning, and patient reception SOPs, accurate judgment of treatment nodes, real-time warning of communication risks, and full-link guidance and autonomous iterative optimization of the patient reception process are achieved, effectively standardizing the patient reception process.

[0017] Other features and advantages of this application will be set forth in the following description and will be apparent in part from the description or may be learned by practicing embodiments of this application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a real-time patient reception guidance method based on a dynamic inference engine and patient profile provided in an embodiment of this application; Figure 2A flowchart illustrating the acquisition of semantic text, standardized medical terminology, and historical medical data in the real-time patient reception guidance method based on a dynamic inference engine and patient profile provided in this application embodiment; Figure 3 This is a schematic diagram illustrating the implementation steps of the real-time patient admission guidance method based on a dynamic inference engine and patient profile provided in this application embodiment. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0022] Please refer to Figure 1 , Figure 1 This is a flowchart of a real-time patient admission guidance method based on a dynamic inference engine and patient profiles, as described in some embodiments of this application. This real-time patient admission guidance method based on a dynamic inference engine and patient profiles is used in terminal devices, such as computers and mobile phones. The real-time patient admission guidance method based on a dynamic inference engine and patient profiles includes the following steps: S11. Obtain patient identification data and audio stream data of doctor-patient communication, perform speech recognition and semantic conversion processing on the audio stream data, extract semantic text and standardized medical terms, and extract historical medical data from the preset electronic health record database in combination with patient identification data. S12. Construct patient profiles based on semantic text and historical medical data, and perform stratified identification of patients' willingness to seek medical treatment based on a pre-trained willingness classification model, and output stratified labels of willingness to seek medical treatment. S13. Input the patient profile, the hierarchical label of the willingness to seek medical treatment, and the preset standard operating procedures for receiving patients into the dynamic reasoning engine. The dynamic reasoning engine performs multiple rounds of reasoning based on the pre-trained medical knowledge graph and outputs the judgment result of the current diagnosis and treatment node, the reasoning chain, and the communication risk warning information. S14. Based on the judgment results and reasoning chain of the diagnosis and treatment nodes, generate standardized diagnosis and treatment action guidance and professional communication scripts, and push communication risk warning information and diagnosis and treatment action guidance to the patient reception terminal in real time. S15. After the consultation is completed, the entire consultation process is intelligently reviewed, a consultation review report is generated, and the consultation SOP and the guidance strategy of the inference engine are iteratively optimized based on the deviation data in the review report.

[0023] It should be noted that the core concept of this application lies in deeply coupling multimodal speech perception, dynamic reasoning engine, and pre-set standard operating procedures (SOPs) for patient reception, forming a closed-loop system from information collection, profile construction, dynamic reasoning to guidance push and review iteration. The data collection step uses a multimodal speech perception module deployed in the reception environment to acquire doctor-patient dialogue audio in real time, rather than relying on manual recording by doctors, thus significantly shortening information collection time and reducing the probability of information omission. After speech recognition and semantic conversion, the collected audio extracts structured semantic text and standardized medical terminology. Simultaneously, historical medical data is retrieved from the electronic health record database based on patient identification, providing a data foundation for subsequent patient profile construction. The patient profile construction and intention recognition step integrates the semantic text acquired during the current reception with historical medical data to generate a multi-dimensional patient profile covering demographic characteristics, past medical history, medication records, allergy history, family history, and lifestyle habits. A pre-trained intention classification model is used to determine the patient's main purpose for this visit (e.g., initial diagnosis, follow-up visit, symptom consultation, or prescription renewal), outputting hierarchical labels for the patient's intention to seek medical attention. The dynamic reasoning process is the core technology of this application. Patient profiles, tiered tags indicating patient intent, and pre-defined standard operating procedures (SOPs) are input into the dynamic reasoning engine. The engine, based on a pre-trained medical knowledge graph, iterates through multiple rounds, generating three types of outputs: 1) the current treatment node's judgment result (i.e., which stage of the SOP the current treatment process is in); 2) the reasoning chain (i.e., candidate diagnoses and reasoning paths sorted by confidence level); and 3) communication risk warning information (i.e., real-time detection of communication risks such as information omissions, emotional fluctuations, and information asymmetry). To guide the generation and delivery of steps, based on the treatment node and reasoning chain, matching items are retrieved from a pre-defined script template library and SOP action library to generate personalized, standardized treatment action guidance and professional communication scripts, which, along with risk warning information, are pushed to the treatment terminal for doctors' real-time reference. The review and iterative optimization process involves extracting full-process data after the treatment is completed. By comparing the reasoning results with the final diagnosis, and the use of scripts with the actual communication effect, a structured review report is generated. Based on deviation data, the inference engine parameters and SOP benchmarks are automatically updated through mechanisms such as reinforcement learning.

[0024] By combining multimodal speech perception and semantic transformation, the system achieves automatic and efficient collection of doctor-patient communication information, overcoming the shortcomings of traditional text input methods, such as low efficiency and missing information. By incorporating patient profiles and treatment intentions into the inference input in a hierarchical manner, the inference results are personalized and adaptable, rather than a general diagnostic output. Through the constraint linkage of the consultation SOP, the system ensures that the scope of inference always remains within the framework of standardized diagnosis and treatment procedures, avoiding the risk of deviation from clinical norms that may be caused by open reasoning. Through the closed-loop design of post-consultation review and iterative optimization, the system has the ability to continuously self-evolve, constantly approaching the optimal guidance strategy as the data used accumulates.

[0025] Please refer to Figure 2 , Figure 2 This is a flowchart illustrating the acquisition of semantic text, standardized medical terminology, and historical medical data in a real-time patient guidance method based on a dynamic inference engine and patient profiles, as described in some embodiments of this application. According to embodiments of the present invention, the acquisition of patient identification data and audio stream data of doctor-patient communication, performing speech recognition and semantic conversion processing on the audio stream data, extracting semantic text and standardized medical terminology, and extracting historical medical data from a pre-set electronic health record database in conjunction with patient identification data specifically includes: S21. Audio signals in the reception environment are collected in real time through a microphone array, and the doctor's voice channel and the patient's voice channel are distinguished by a sound source separation algorithm. S22. Perform noise filtering and voice endpoint detection on the two voice channels respectively; S23. The detected speech segments are automatically transcribed using speech recognition, and then semantic text is extracted using a preset medical named entity recognition model, including symptom entities, anatomical location entities, and time entities, and mapped to a standardized medical terminology system.

[0026] It should be noted that "microphone array" refers to a sound acquisition device deployed in a patient reception environment, containing at least two microphone units, which achieves sound source localization and separation by analyzing the phase difference of the signals acquired by each microphone. "Sound source separation algorithm" specifically includes speech separation techniques based on independent component analysis or deep learning. Its core purpose is to separate the speech signals of doctors and patients in mixed audio into independent channels, enabling subsequent processing to accurately distinguish the speakers. "Noise filtering" includes the suppression of environmental noise (such as air conditioning and equipment operation sounds) and transient noise (such as door closing sounds and object collision sounds). "Speech endpoint detection" refers to a technique that automatically marks the start and end points of speech segments in a continuous audio stream, used to segment the continuous audio stream into independent speech segments. The "Medical Named Entity Recognition Model" refers to a neural network model trained on annotated medical corpora that can identify and classify entity words in transcribed text into three categories: symptom entities, referring to subjective feelings or observable abnormalities described by patients, such as "chest tightness," "headache," and "coughing up blood"; anatomical location entities, referring to body parts related to the symptoms, such as "right lower abdomen," "precordial region," and "left knee joint"; and temporal entities, referring to time information such as onset time, duration, and frequency, such as "three days ago," "lasting for two weeks," and "every morning." Mapping to a standardized medical terminology system involves converting the identified non-standard colloquial expressions into standard terms in a standard medical coding system (such as ICD-11 or SNOMED-CT), for example, mapping "palpitation" to "heart palpitation" and "stomach ache" to "abdominal pain." This embodiment, through sound source separation and independent dual-channel processing, enables the system to accurately distinguish between the doctor's and patient's statements, avoiding misidentification of the doctor's questions as the patient's symptom descriptions, thereby improving the accuracy of semantic extraction. By jointly extracting and standardizing the three types of entities, the symptom, location, and time information are structured into inference inputs, providing precise query conditions for subsequent knowledge graph retrieval. For example, when only "abdominal pain" is extracted without the location information of "right lower abdomen," the inference engine can only provide a broad range of abdominal diseases; however, when both the symptom entity "abdominal pain" and the anatomical location entity "right lower abdomen" are extracted, the inference engine can accurately pinpoint specific diseases such as appendicitis, significantly improving the accuracy and efficiency of inference.

[0027] For example, if a patient says, "I've had pain in my lower right abdomen for two days," the system extracts the symptom entity "abdominal pain," the anatomical location entity "lower right abdomen," and the time entity "two days." After standardized terminology mapping, "abdominal pain" corresponds to the standard term "abdominal pain [R10.4]," and "lower right abdomen" corresponds to the anatomical location "right iliac fossa." If the patient's description is vague, such as "It's just that this area is uncomfortable, but I can't explain it clearly," the system fails to extract specific symptom and anatomical location entities. In this case, the system will mark the information completeness score of this semantic acquisition as low, triggering a subsequent progressive consultation guidance mechanism.

[0028] According to an embodiment of the present invention, the step of constructing a patient profile based on semantic text and historical medical data, and performing hierarchical identification of the patient's willingness to seek medical treatment based on a pre-trained willingness classification model, and outputting hierarchical labels for the willingness to seek medical treatment, specifically includes: Patient profiles are constructed based on semantic text and historical medical data. These profiles include demographic characteristics, past medical history, medication records, allergy history, family history of genetic diseases, and lifestyle data. The extracted historical medical data is sorted chronologically and missing values ​​are filled to generate structured patient profile feature vectors; Semantic text is processed to obtain semantic text feature vectors, including symptom entity vectors, anatomical site entity vectors, and time entity vectors; The intention classification model takes as input a concatenation of symptom entity vectors, anatomical site entity vectors, time entity vectors, and patient profile feature vectors, and outputs a hierarchical label of medical visit intention, including one or more of the following: initial diagnosis, follow-up visit, symptom consultation, and prescription renewal.

[0029] It should be noted that "demographic characteristics" include basic information such as the patient's age, gender, occupation, and marital / reproductive status; "past medical history" refers to the names and dates of diseases diagnosed during each of the patient's visits; "medication records" include historical medication information such as drug names, dosages, treatment cycles, and discontinuation dates; "allergy history" covers records related to drug allergies and food allergies; "family history" refers to records of confirmed hereditary diseases among the patient's immediate family members; and "lifestyle data" includes smoking history, alcohol consumption history, exercise habits, dietary habits, and sleep patterns. When "time-series sorting" historical medical data, the system arranges the records according to the order of visits; "missing value imputation" refers to filling in missing fields in the time-series data (such as missing lifestyle data from a particular visit) using methods such as forward imputation or statistical imputation based on similar patients. The generated "structured patient profile feature vector" is a multi-dimensional numerical vector, with each dimension corresponding to a profile feature, used as input for subsequent model calculations. "Semantic text feature vector" refers to the extracted semantic text encoded into a fixed-length dense vector representation using a pre-trained language model (such as BERT or its medical-domain fine-tuned version; BERT is a Transformer-based bidirectional encoder representation model). The symptom entity vector, anatomical location entity vector, and time entity vector correspond to the encoding results of different entity types. The "Intention Classification Model" is a multi-classifier trained on labeled data. Its input concatenated vector dimension is the sum of the dimensions of the symptom entity vector, anatomical location entity vector, time entity vector, and patient profile feature vector. The output is the probability distribution of four categories (initial diagnosis, follow-up visit, symptom consultation, and prescription renewal), with the label corresponding to the highest probability being the final output.

[0030] This embodiment constructs a patient profile containing multi-dimensional information, enabling the system to comprehensively understand the patient's health background, rather than relying solely on information from a single visit. Allergy history information within the profile is particularly crucial; when generating treatment guidance, it automatically filters out treatment suggestions involving allergic medications, effectively preventing medication-related accidents. By identifying the purpose of the visit through a willingness classification model, the system can automatically adjust the retrieval strategy of the dialogue template and the focus of the inference engine based on different visit intentions. For example, for newly diagnosed patients, the focus is on symptom differential diagnosis; for follow-up patients, the focus is on treatment effect evaluation and medication adjustment suggestions, thereby improving the relevance and practicality of the guidance information.

[0031] In this embodiment, a 60-year-old female patient presented with a cough. Her patient profile showed a history of chronic bronchitis, a penicillin allergy, and a 30-year smoking history. The symptom entities extracted from the semantic text were "cough" and "sputum," the time entity was "more than two months," and the anatomical location entity was "chest." The intention classification model took the concatenated vector of these data as input and output a probability of 0.78 for "follow-up visit" and a probability of 0.15 for "initial diagnosis." Based on this, the model determined the patient's intention to seek medical attention was for a follow-up visit. In boundary scenarios, when historical medical data is severely lacking (e.g., no history for initial visits), many dimensions in the generated profile feature vector are empty. In this case, the intention classification model mainly relies on the semantic text feature vector for judgment, resulting in an overall low confidence level in the output probability distribution. However, it can still provide a relatively reasonable inclination judgment based on the current symptom description.

[0032] According to an embodiment of the present invention, the step of inputting patient profiles, hierarchical tags of medical treatment intentions, and preset standard operating procedures for patient reception into a dynamic reasoning engine, and the dynamic reasoning engine performing multi-round reasoning based on a pre-trained medical knowledge graph, outputting the judgment result of the current diagnosis and treatment node, the reasoning chain, and communication risk warning information, specifically includes: The dynamic reasoning engine employs a knowledge graph reasoning model based on an attention mechanism; The query vector is obtained by concatenating the patient profile feature vector and the symptom entity vector, and the association strength with the candidate disease node is calculated through a multi-head attention mechanism. The multi-head attention mechanism includes symptom combination matching analysis based on symptom and symptom co-occurrence matrix, complication probability analysis based on disease and complication causal graph, and medication conflict risk analysis based on drug and contraindication conflict detection matrix. The three analytical scores and the correlation strength are integrated into a comprehensive inference confidence score. The inference chain and the corresponding intermediate feature data of the inference chain are output according to the confidence score.

[0033] It should be noted that the "knowledge graph reasoning model based on attention mechanism" refers to representing nodes such as diseases, symptoms, examinations, and drugs in the medical knowledge graph as low-dimensional embedding vectors, and calculating the semantic association between the query vector (i.e., the patient's current state) and each candidate disease node through an attention mechanism. The "multi-head attention mechanism" is an extended form of the attention mechanism, using multiple independent attention heads in parallel to capture the association patterns between the query and candidate nodes from different subspaces, avoiding fine-grained features that single-head attention might overlook. The "symptom and symptom co-occurrence matrix" is a symmetric square matrix where each element M[i][j] represents the conditional probability of symptom i and symptom j occurring simultaneously in historical medical data statistics. When a patient has multiple symptoms simultaneously, the system queries the co-occurrence probabilities of each symptom pair in the co-occurrence matrix and comprehensively calculates the matching score between the current symptom combination and typical disease symptom patterns. The "disease and complication causal graph" is a subgraph of causal relationships between diseases constructed based on the medical knowledge graph, annotating various complications that may be caused by the primary disease and their probabilities. The "Drug and Contraindication Conflict Detection Matrix" is a drug-disease relationship matrix that stores a list of contraindications for each drug. When the candidate diagnostic suggestions output by the inference engine involve a specific drug, this matrix is ​​automatically retrieved to detect whether the patient's current profile of diseases or allergies constitutes a contraindication for that drug. The fusion method for the comprehensive inference confidence is to weight and sum the scores of the three association analyses with the multi-head attention association strength, and the weight coefficients are determined by hyperparameter search on the validation set. The intermediate feature data of the inference chain refers to the attention weight matrices and intermediate embedding vectors of each layer in the inference process. These data will serve as inputs for subsequent steps such as diagnosis and treatment node judgment. This embodiment, by introducing triple association analysis, breaks through the limitations of traditional single knowledge graph retrieval, enabling the inference results to not only consider the direct matching of symptoms and diseases, but also comprehensively consider the co-occurrence patterns between symptoms, the evolutionary path from the primary disease to complications, and medication safety constraints, thereby realizing multi-dimensional comprehensive inference and improving the clinical reliability and safety of diagnostic suggestions. In particular, the drug and contraindication conflict detection can eliminate unsafe medication regimens at the inference stage, effectively preventing medical accidents caused by improper medication.

[0034] For example, the patient is a 55-year-old male with symptoms of "chest pain" and "shortness of breath". His past medical history in the profile includes "hypertension" and "type 2 diabetes", and his medication records include "metformin" and "nifedipine". The dynamic reasoning engine calculation process is as follows: (1) Multi-head attention association strength: The patient profile feature vector and the symptom entity vector are concatenated as query Q, and the embedding of each candidate disease node is K and V. The association strength score of the candidate disease "unstable coronary heart disease" is calculated to be 0.82; the association degree of "gastroesophageal reflux" is 0.45. (2) Triple analysis: The co-occurrence conditional probability of "chest pain-shortness of breath" in the symptom co-occurrence matrix is ​​0.78, and the combination matching score is relatively high; the causal map of disease complications shows that the probability of "hypertension" patients developing "angina pectoris" is 0.65; the drug contraindication conflict detection shows that there is no contraindication for the candidate drugs. The overall inference confidence score is a weighted fusion of the association strength and the scores of the three analyses. Ultimately, "unstable angina pectoris with coronary heart disease" ranked first in the inference chain with an overall confidence score of 0.85. In the boundary scenario, if the patient is a newly diagnosed patient with no past medical history or medication records in the profile and only reports "dizziness," the symptom co-occurrence analysis scores low due to the presence of only a single symptom, and the complication analysis lacks analytical basis due to the absence of a primary disease. The overall inference confidence score decreases, with the highest confidence score in the inference chain being only 0.40, which will trigger a rule-matching downgrade strategy in the subsequent process.

[0035] According to an embodiment of the present invention, obtaining the judgment result of the diagnosis and treatment node specifically includes: The intermediate features of the inference chain are matched with the preset SOP node sequence for patient reception, and the judgment result of the current treatment node is output. The current diagnosis and treatment node judgment result reflects the node type of the current reception process in the SOP. The preset node types include one of the following: symptom inquiry node, physical sign collection node, preliminary diagnosis node, examination ordering node, and treatment plan formulation node. Based on the judgment result of the current diagnosis node, the corresponding SOP branch process is dynamically invoked, and the SOP branch process is used as a constraint condition input into the inference engine to limit the output range of the inference chain in the current round.

[0036] It should be noted that the core of this embodiment lies in the deep coupling of the dynamic inference engine with the preset Standard Operating Procedure (SOP) for patient reception. The "Patient Reception SOP" refers to a pre-defined patient reception process according to medical quality management standards, dividing a standard patient reception into multiple sequentially advancing or conditionally branching nodes. Each node defines the diagnostic and treatment operations to be completed and the information to be collected at that stage. "Intermediate features of the inference chain" refer to the output attention weight matrix and intermediate embedding vectors; these features reflect the inference engine's internal representation of the patient's current state. The "matching" process is implemented through a node classifier. This classifier takes the intermediate features of the inference chain and the embedding vectors of the SOP node sequence as input, and classifies them using a multilayer perceptron to map the current patient reception state to one of the five preset node types in the SOP. The specific implementation of "SOP branch process as a constraint input to the inference engine" is as follows: the SOP branch process corresponding to the current node type is transformed into the attention mask matrix of the inference engine, and only the attention channels of candidate disease node paths related to the current SOP node are retained, while candidate paths unrelated to the current node are masked, so that the output range of the inference engine is limited to the range of diseases and treatment actions that the current SOP node is concerned with.

[0037] It should be noted that this embodiment uses the Standard Operating Procedure (SOP) as a constraint on the inference engine, ensuring that the inference results always remain within the framework of the standardized diagnosis and treatment process. This avoids problems such as "overdiagnosis" (inferring a large number of disease suggestions unrelated to the current stage of diagnosis and treatment, interfering with the doctor's judgment) or "diagnostic jump" (jumping to the treatment plan before sufficient symptom data is collected, violating clinical norms) that may result from open reasoning. Simultaneously, the dynamic judgment of node types enables the guidance system to perceive the pace of the patient reception process and automatically recommend different types of standardized operations and scripts at different nodes, thereby achieving true real-time guidance throughout the entire process.

[0038] For example, a patient with a cough enters the internal medicine clinic, and the doctor is collecting symptom information. At this time, the intermediate features of the inference engine (mainly the attention distribution of symptom entity vectors) are matched with the SOP node sequence. The probability of the node classifier outputting "symptom inquiry node" is 0.91, and the probability of "preliminary diagnosis node" is 0.06. The system determines that it is currently in the symptom inquiry node. Based on this determination, the SOP branch process of the symptom inquiry node is invoked, constraining the output range of the inference engine to "recommended consultation questions to be collected further," rather than outputting complete diagnostic suggestions. Subsequently, when the doctor completes all symptom collection and enters the physical examination stage, the intermediate features of the inference engine (which now contain richer symptom combinations and preliminary physical examination information) will be mapped to the "physical examination node," and the SOP constraint condition will switch accordingly. If the doctor skips physical examination and directly attempts to order examinations, the node classifier will still map the current state to the "physical examination node." The system restricts the inference engine from outputting examination suggestions through the SOP constraint and pushes a prompt "It is recommended to complete basic physical examination first," thereby achieving standardized supervision of the diagnosis and treatment process.

[0039] According to an embodiment of the present invention, the output of communication risk warning information specifically includes: Communicating risk warning information includes risk type labels and corresponding response suggestions; Risk type labels include risk of missing key information, risk of abnormal patient emotions, and risk of information asymmetry between doctors and patients; When the necessary medical history information required for the reasoning chain does not appear in the patient profile or the current semantic text, the risk of missing key information is triggered. Based on a pre-set voice emotion recognition model, the patient's voice signal is classified into emotions. When the recognition result is negative emotion and the confidence level exceeds a pre-set threshold, the risk of abnormal patient emotions is triggered. When doctors use professional terminology more frequently than a preset density threshold and patients show signs of confusion in their feedback, the risk of information asymmetry between doctors and patients is triggered.

[0040] It should be noted that this embodiment proposes for the first time a real-time detection mechanism for three types of communication risks during the patient reception process. Among them, the "risk of missing key information" refers to the situation where the inference engine discovers during the inference process that one or more pieces of medical history information crucial to accurate diagnosis (such as drug allergy history, previous related surgical history, or key family history) are missing from both the patient profile and the current semantic text. Based on this, the system generates a missing information warning, prompting the doctor to supplement the inquiry with information in the specified direction. The judgment criterion for triggering this risk is based on the essential information fields of each candidate disease node marked in the knowledge graph of the inference engine. The system is triggered when the essential fields of high-confidence candidate diseases in the inference chain are missing. The "voice emotion recognition model" refers to a deep learning model trained on emotional corpus that can identify negative emotion categories such as anger, anxiety, fear, and sadness from the prosodic features (including fundamental frequency changes, speech rate changes, energy distribution, etc.) and acoustic features of the patient's voice signal. When the confidence level of the identified negative emotion exceeds a preset threshold (generally 0.75 in actual deployment), a patient emotional abnormality risk warning is triggered. The detection of "information asymmetry risk between doctors and patients" is achieved through two independent sub-mechanisms: First, the ratio of the frequency of professional terms in the transcribed text from the doctor's voice channel to the total vocabulary is statistically analyzed. When this ratio exceeds a preset density threshold (generally 0.15 in actual deployment), it is considered that the doctor has used too much professional terminology. Second, semantic analysis is performed on the transcribed text from the patient's voice channel to detect whether there are language signals indicating confusion (such as "What does this mean?", "I don't quite understand", "Could you say it again?", etc.). When both conditions are met simultaneously, the risk of information asymmetry between doctors and patients is triggered. Through real-time detection of three types of communication risks, the system has the ability to monitor and warn of non-technical factors (communication quality) during the consultation process. The warning of the risk of missing key information can effectively prevent misdiagnosis or missed diagnosis due to insufficient consultation; the warning of the risk of abnormal patient emotions can help doctors promptly detect patients' anxiety or panic and adjust their communication methods to soothe patients; the warning of the risk of information asymmetry between doctors and patients can prompt doctors to translate professional terms into colloquial expressions that are easy for patients to understand, thereby improving patients' understanding and compliance with the treatment plan. These three types of risk warnings complement traditional diagnostic support, enabling the intelligent guidance system to upgrade from a simple technical decision-making aid to a comprehensive patient reception aid that includes communication quality assurance.

[0041] For example, an elderly patient presents with chest tightness. The inference engine detects that the crucial information needed for this high-confidence candidate disease (coronary artery disease), namely "whether the patient is taking aspirin or other anticoagulants," is missing from the patient profile and the current conversation. The system triggers a risk of missing crucial information and pushes a suggestion to "ask the patient if they are currently taking anticoagulants." Simultaneously, the voice emotion recognition model detects an elevated fundamental frequency and increased speech rate in the patient's voice, identifying it as "anxiety" with a confidence level of 0.88. The system triggers a risk of abnormal patient emotions and pushes a response suggestion to "reassure the patient and explain that most chest tightness is not necessarily a serious heart problem." Furthermore, if the doctor uses professional terms such as "myocardial ischemia," "ST segment changes," or "coronary angiography" in the current conversation, with a terminology density exceeding the threshold of 0.22, and the patient's feedback includes a confused signal such as "What does coronary angiography mean?", the system triggers a risk of information asymmetry between doctor and patient and pushes a suggestion to "explain the purpose and necessity of each examination to the patient in plain language." In boundary scenarios, if the patient remains silent or responds with monosyllabic words (such as "um", "yes", "no"), the speech emotion recognition model may not be able to obtain enough prosodic features due to the short speech samples, and may not be able to trigger an emotional abnormality warning. In this case, no intervention is performed by default to avoid false alarms.

[0042] According to an embodiment of the present invention, before outputting the current diagnosis and treatment node judgment result, inference chain, and communication risk warning information, the method further includes: Obtain the patient's historical medical visit time series data, and output the patient's disease progression risk score within a preset future time period through a pre-trained disease progression prediction model; Assess the severity of each symptom and output a symptom severity score; Symptoms with severity scores above a preset threshold are marked as high-risk symptoms and given higher weight in the inference chain; By inputting disease progression risk scores and high-risk symptom markers as contextual features into the dynamic inference engine, "urgent treatment suggestions" are automatically displayed in advance for high-risk patients.

[0043] It should be noted that this embodiment adds two pre-processing steps—disease progression prediction and symptom severity assessment—before dynamic inference, giving the inference engine a forward-looking predictive capability. "Historical medical visit time-series data" refers to the time-series records of each patient's medical visits. Each record includes the visit time, chief complaint, diagnosis, examination results, and medication information, arranged chronologically to form a time-series sequence. The "disease progression prediction model" employs a time-series prediction architecture based on a Long Short-Term Memory (LSTM) network or a Transformer encoder. Using historical medical visit time-series data as input, it predicts the probability of a target disease or complication occurring within a predetermined time period (typically 30 days in actual deployment). The model is obtained through supervised training using labeled historical case data (with actual outcomes already labeled). The disease progression risk score is a continuous value between 0 and 1, with higher values ​​indicating a greater risk of disease progression. The "Symptom Severity Assessment" is based on a pre-trained severity assessment model. This model takes the frequency (e.g., "several times a day"), duration (e.g., "lasts for 2 hours"), and impact (e.g., "unable to work") of the patient's self-reported symptoms as input features and outputs a severity score for each symptom. The score is also a continuous value between 0 and 1. A preset threshold is typically set to 0.70 in actual deployments. Symptoms with severity scores higher than this threshold are marked as high-risk symptoms, and the attention channel corresponding to the high-risk symptom is multiplied by a weight enhancement coefficient greater than 1 (typically 1.5 in actual deployments) in the inference chain. "Automatic upfront display of emergency treatment suggestions for high-risk patients" means that when generating inference results, the system will upfront display cases with risk scores higher than a preset emergency threshold (typically 0.80 in actual deployments) in the push list and add a prominent emergency label. By introducing disease progression prediction and symptom severity assessment before inference, the inference engine can not only provide diagnostic suggestions based on current symptoms but also predict the disease's development trend and risk of deterioration, thereby achieving priority identification and emergency alerts for high-risk patients. This forward-looking reasoning ability enables the system to upgrade from a passive "symptom-response diagnosis" to an active "risk-predictive guidance," which has important clinical value for the triage and timely intervention of critically ill patients.

[0044] For example, a 65-year-old patient with a history of chronic obstructive pulmonary disease (COPD) presented with the chief complaint of "worsening cough, increased sputum volume, and yellow sputum color over the past three days." Their historical consultation timeline data (including four visits within the past 12 months, recording a stable period of COPD and one mild acute exacerbation) was extracted and input into an LSTM disease progression prediction model. The model output a risk score of 0.82 for moderate to severe acute exacerbations within the next 30 days. The severity assessment model assigned severity scores of 0.85, 0.78, and 0.72 for "worsening cough," "increased sputum volume," and "yellow sputum color," respectively. The first two, exceeding the 0.70 threshold, were marked as high-risk symptoms. During inference, the inference engine multiplied the attention weights of high-risk symptoms by an enhancement coefficient of 1.5, resulting in higher overall confidence for candidate disease pathways associated with COPD acute exacerbations. Meanwhile, because the disease progression risk score of 0.82 exceeds the emergency threshold of 0.80, the patient's case is displayed prominently in the push notification with the note "Emergency treatment recommended: High risk of acute exacerbation of COPD". If the patient has no medical history (newly diagnosed patient), the disease progression prediction model cannot output a valid prediction result due to the lack of time-series input data. In this case, the system defaults to setting the risk score to 0 (no prediction data) and relies solely on symptom severity assessment for inference, without triggering emergency prompts related to disease progression.

[0045] It is worth mentioning that the generation of the patient review report specifically includes: Extract the semantic text, reasoning process records, dialogue usage records, and final diagnosis results from the entire consultation process, and calculate the reasoning accuracy rate, dialogue adoption rate, and communication risk warning effectiveness rate. The actual values ​​of each indicator are compared with the preset benchmark values ​​to identify deviation items and corresponding triggering reasons, and a structured deviation analysis report is output. Deviation data includes inference path deviations, script replacement records, and abnormal communication node time consumption information.

[0046] It should be noted that in this embodiment, the calculation formula for "inference accuracy" is: inference accuracy equals the number of consultations where the candidate diagnosis ranked first in the inference chain matches the doctor's final confirmed disease label, divided by the total number of consultations. In actual deployment, if the top three candidate diagnoses in the inference chain include the final confirmed disease, it is counted as "predicted hit," and the hit rate is recorded as an auxiliary indicator. "Script adoption rate" refers to the degree of consistency between the communication script actually used by the doctor and the script recommended by the system, which is achieved through semantic similarity calculation: the semantic vector of the doctor's actual speech is calculated with the semantic vector of the recommended script, and the similarity exceeding the preset threshold (generally 0.70 in actual deployment) is considered adoption. "Communication risk warning effectiveness rate" refers to the proportion of risk warnings issued in which the doctor actually takes corresponding measures (such as supplementary inquiries, comforting patients, changing to colloquial language, etc.), which is determined by analyzing the semantic changes in the doctor's subsequent dialogue after the warning is issued. "Preset benchmark value" is the average level of each indicator obtained from historical accumulated data, such as the average inference accuracy benchmark of all departments, the average script adoption rate benchmark of the same department, etc. The deviation data includes "Inference Path Deviation," which records the deviation path between the ranking in the inference chain and the actual diagnosis; "Script Replacement Record," which records the alternative scripts actually used when the doctor did not adopt the recommended scripts; and "Communication Node Time Consumption Anomaly Information," which records significant deviations between the actual time consumed at each treatment node and the SOP preset time benchmark. This embodiment provides an objective data foundation for medical quality management and system optimization by systematically extracting and quantifying key quality indicators throughout the entire patient reception process. The generation of the debriefing report makes the quality of each patient reception traceable, quantifiable, and comparable, facilitating the identification of systemic problems (such as a persistently low inference accuracy rate for a certain type of disease) and individual anomalies (such as an abnormal script adoption rate for a particular doctor), thereby providing precise improvement directions for subsequent iterative optimization.

[0047] For example, in a patient review calculation, the entire patient visit data was extracted: the first candidate diagnosis in the reasoning chain was "upper respiratory tract infection," which was ultimately diagnosed as "acute bronchitis," with a reasoning accuracy rate of 0 (mismatch). However, the third candidate diagnosis in the reasoning chain was indeed "acute bronchitis," so the prediction hit was counted as 1. The dialogue usage data showed that the doctor actually used 3 out of the 5 system-recommended dialogues, resulting in a dialogue adoption rate of 0.60. Two communication risk warnings were issued (one for information omission and one for emotional abnormality). The doctor responded to the information omission warning with supplementary inquiries but did not take significant intervention for the emotional abnormality warning, resulting in a warning effectiveness rate of 0.50. After comparing these indicators with the department's baseline values ​​(reasoning accuracy 0.75, dialogue adoption rate 0.65, warning effectiveness rate 0.70), the reasoning path deviation item (overlapping symptoms of upper respiratory tract infection and acute bronchitis leading to reasoning bias) and the warning effectiveness rate deviation item were identified, generating a structured deviation analysis report.

[0048] It is worth mentioning that the guiding strategies for iteratively optimizing the patient reception SOP and inference engine specifically include: Based on the reasoning path deviation in the debriefing report, the model parameters of the dynamic reasoning engine are updated through reinforcement learning algorithm, and the reward function is a weighted combination of reasoning accuracy and word adoption rate. Based on the script replacement records, update the priority weight of the corresponding script template in the script template library; Based on abnormal communication node time information, the preset time benchmark of the corresponding node in the consultation SOP is adaptively adjusted.

[0049] It should be noted that this embodiment defines three specific paths for iterative optimization. The "reinforcement learning algorithm" can employ reinforcement learning methods such as policy gradient algorithms or deep Q-networks. Here, "state" is defined as the patient's state representation at each time step during the consultation process (including patient profiles and vector representations of collected symptoms), "action" is defined as the candidate diagnosis ranking strategy output by the inference engine, and "reward" is defined as a weighted combination of the consistency between the final diagnosis and the inference result (inference accuracy) and the script adoption rate. In actual deployment, the weighting coefficients are generally set to 0.60 for inference accuracy and 0.40 for script adoption rate; these ratios can be adjusted according to the actual management needs of the medical institution. Model parameter updates use the gradient descent method, and the learning rate is generally set to a small value (e.g., 0.0001) in actual deployment to ensure stable convergence of the model. "Script template priority weight" refers to the probability weight of each script template in the script template library being retrieved and recommended. When doctors frequently replace a certain script template with an alternative script, the priority weight of the replaced template is automatically reduced, and the weight of the alternative script is increased (if the alternative script does not exist in the library, it is added as a new template), so that the script recommendations gradually approach the actual usage preferences of doctors. "SOP Node Duration Benchmark" refers to the suggested completion time for each node predefined in the SOP. When the review report shows that the actual time spent on a certain node consistently deviates significantly from the benchmark (e.g., more than 1.5 times the benchmark or less than 0.5 times the benchmark), the duration benchmark for that node is automatically fine-tuned, making the SOP more closely match the actual work rhythm.

[0050] This embodiment achieves autonomous evolution across three dimensions: reasoning ability, dialogue quality, and process efficiency, through the parallel operation of three optimization paths. Reinforcement learning-driven model updates enable the inference engine to learn from each diagnostic deviation, continuously improving diagnostic accuracy; dynamic adjustment of dialogue priorities allows the system to gradually adapt to local medical culture and doctors' individual communication habits; and adaptive adjustment of SOP duration benchmarks continuously aligns standardized processes with actual best practices. This multi-dimensional self-iterative mechanism is a key innovation that distinguishes this system from traditional static decision support systems.

[0051] For example, after accumulating data from 50 consecutive patient visits, statistics showed that the inference accuracy for coronary artery disease improved from an initial 0.70 to 0.81 (due to the effectiveness of reinforcement learning parameter updates), the adoption rate of the phrases related to "chest tightness" symptoms increased from 0.55 to 0.72 (the recommended phrases after adjusting the phrase template weights better suited the preferences of the physician population), and the standard operating procedure (SOP) duration for the "vital signs collection node" in the internal medicine outpatient department was adjusted from 120 seconds to 90 seconds (closer to the actual average time). If the initial data volume is insufficient (less than 100 patient visits) during system deployment, the update increment of iterative optimization is automatically reduced (by lowering the learning rate and weight adjustment step size) to avoid overfitting noisy data under small sample conditions and to ensure the stability of model updates.

[0052] It is worth mentioning that when outputting the current diagnosis and treatment node judgment result, reasoning chain, and communication risk warning information, a context-aware strategy adjustment step is also included, specifically including: Obtain the current department category, consultation time, and attending physician's qualification level, input the pre-trained context-aware decision model, and output the reasoning depth level and the level of detail of the dialogue; When the patient reception period is peak, the inference engine selects fast mode and performs single-layer knowledge graph retrieval; When the patient reception period is off-peak, the inference engine selects deep mode and performs multi-level iterative inference. When the doctor's qualification level is junior, a detailed explanation of the diagnosis and treatment actions and a demonstration of the language will be sent. When a doctor's qualification level is advanced, concise conclusive guidance and key risk warnings will be sent.

[0053] It should be noted that this embodiment introduces a dynamic adjustment mechanism for reasoning strategies based on multi-dimensional context. "Department Category" includes standard departmental classifications such as internal medicine, surgery, pediatrics, obstetrics and gynecology, and emergency medicine; "Consultation Time" is divided into peak hours (e.g., 8:00-11:00 AM on weekdays) and off-peak hours (other times) according to the actual operational rhythm of the medical institution, and the criteria for this division can be flexibly adjusted through system configuration; "Attending Physician Qualification Level" is divided into two levels: primary (resident physicians and below) and advanced (attending physicians and above). The "Context-Aware Decision Model" is a lightweight decision tree model or a shallow multilayer perceptron, taking the concatenation of departmental category code, time period label, and physician qualification level code as input, and outputting the reasoning depth level (fast or deep) and suggestion detail level (concise or detailed). The depth of inference directly affects the execution strategy of the inference engine: In fast mode, the inference engine performs a single-layer retrieval of the knowledge graph, that is, it only performs one round of calculation on the association strength between the query vector and each candidate disease node, and outputs the k candidate diseases with the highest association strength (k is usually 5 in actual deployment); in deep mode, the inference engine performs multi-layer iterative inference, that is, after each round of retrieval, the retrieved candidate disease nodes are added as new context nodes to the query process, and multi-hop inference is performed to explore deeper relationships such as complications and differential diagnoses. The iteration depth is usually 3 layers in actual deployment. The level of detail of the suggestion affects the output format of the guidance information: in concise mode, the output is a simplified string containing only the diagnostic conclusion and key risk warnings; in detailed mode, the output is a complete paragraph containing symptom interpretation, differential diagnosis basis, further examination suggestions, and exemplary language.

[0054] This embodiment addresses the adaptability issue of a "one-size-fits-all" inference strategy in real-world medical scenarios through context-aware dynamic strategy adjustment. A fast mode is used during peak hours to maintain a high throughput while ensuring basic diagnostic quality; a deep mode is used during off-peak hours to fully utilize computing resources for in-depth inference and improve diagnostic accuracy. Detailed guidance is provided to junior physicians to compensate for their lack of clinical experience; concise conclusions are offered to senior physicians to avoid redundant information interfering with their clinical judgment. This adaptive strategy mechanism enables the system to flexibly respond to the differentiated needs of different departments, time periods, and user groups.

[0055] For example, during the peak internal medicine outpatient period at 9:30 AM, a junior resident sees a patient. The context-aware decision model takes into account the three labels "internal medicine," "peak," and "junior," and outputs a reasoning depth level of "fast" and a suggestion detail level of "detailed." In fast mode, a single-layer knowledge graph retrieval is performed, taking approximately 0.3 seconds to output the top-5 candidate diagnoses. For each candidate diagnosis, detailed symptom interpretations and further examination suggestions are provided, along with exemplary communication scripts. However, during the off-peak period at 3:00 PM, when a chief physician sees the patient, the output reasoning depth is set to "deep" and the suggestion detail level to "concise." A three-layer iterative reasoning process is performed, taking approximately 1.2 seconds to output a complete reasoning chain including complication analysis and differential diagnosis. However, the guidance information only includes the diagnostic conclusion and key risk warnings (such as "consider ruling out aortic dissection"), omitting basic interpretations and exemplary communication scripts.

[0056] It is worth mentioning that it also includes: When the dynamic inference engine fails to generate a valid inference chain within a preset time, it switches to a preset rule-based symptom-disease mapping table for quick matching, marks it as rule matching mode, and prompts medical staff for manual review. When the highest confidence level in the inference chain is lower than the preset confidence threshold and the candidate disease is a rare disease, a special search of the rare disease knowledge base is triggered. Retrieve differential diagnosis information, gene testing indications, and specialist referral recommendations from a rare disease-specific knowledge graph, and add a "recommend genetic consultation" prompt to the patient guidance information; When a patient presents with signs of multiple complications, a multi-label classification model is used to output a combination of complication diagnostic suggestions.

[0057] It should be noted that this embodiment designs specific response mechanisms for three types of boundary scenarios. The "preset time" is generally set to 2 seconds in actual deployment. If the inference engine cannot output a valid inference chain within 2 seconds due to insufficient input data or knowledge graph matching failure (i.e., the inference chain is empty or the confidence of all candidate diseases is below the minimum effective threshold of 0.10), it automatically switches to a degradation strategy. The "rule-based symptom-disease mapping table" is a pre-built static mapping table that stores the direct association rules between common symptom combinations and corresponding diseases (e.g., "fever + cough + sputum" maps to "acute bronchitis" and "pneumonia"). This mapping table is pre-compiled by clinical experts based on treatment guidelines, does not rely on machine learning model inference, and has deterministic and rapid response characteristics. The "preset confidence threshold" is generally set to 0.50 in actual deployment. When the highest confidence in the inference chain is below this threshold but the system identifies a candidate disease as a rare disease, it does not directly abandon the diagnosis but automatically triggers a "rare disease knowledge base retrieval." The "Rare Disease-Specific Knowledge Graph" is a subgraph extracted independently from the general medical knowledge graph, containing only knowledge related to rare diseases. It covers diseases listed in the National Rare Disease Catalog, their differential diagnosis information, gene testing indications, and specialist referral pathways. The "Multi-Label Classification Model" works as follows: Unlike traditional multi-classification models (where each input outputs only one category), the multi-label classification model allows the same input to belong to multiple categories simultaneously. For patients with multiple complications, this model can simultaneously output multiple possible complication diagnostic labels and their joint probability distribution, rather than just a single diagnosis.

[0058] This embodiment significantly improves usability and security in atypical and complex scenarios by designing specialized response mechanisms for three boundary scenarios: reasoning failure, rare diseases, and complications. The degradation strategy ensures that the system can provide some level of reference advice under any circumstances, preventing it from becoming "unresponsive" due to reasoning failure; the rare disease-specific retrieval compensates for the insufficient coverage of rare diseases in general reasoning models, reducing the risk of missed diagnoses; and the multi-label classification output of complication diagnosis combinations better meets the actual clinical needs of treating multiple coexisting diseases.

[0059] For example, a patient reports "butterfly-shaped facial rash, joint pain, and fever." The inference engine provides a high-confidence diagnosis of "systemic lupus erythematosus" without triggering any boundary scenario mechanisms. In scenario one, the patient only reports "fatigue" with no other symptoms or historical records. The inference engine cannot provide a valid inference chain within 2 seconds, so the system switches to rule matching mode, outputting a list of possible diseases for "fatigue," including "anemia, hypothyroidism, chronic fatigue syndrome, etc.", and noting "Rule matching mode, please manually review." In scenario two, the patient presents with "periodic fever, unexplained weight loss, and multiple swollen lymph nodes." The inference engine has a maximum confidence of 0.42, and the candidate disease is the rare disease "adult Still's disease." This triggers a rare disease-specific search, retrieving the differential diagnostic criteria (excluding infections, tumors, and other rheumatic and immunological diseases), gene testing indications (such as HLA typing), and specialist referral recommendations (rheumatology and immunology) from the rare disease knowledge graph, along with a "recommendation for genetic consultation" prompt.

[0060] It is worth mentioning that it also includes: When the information completeness score of the semantic text collected in real time is lower than the preset completeness threshold, the progressive consultation guidance sub-process is initiated: Based on the current symptom set, the next optimal consultation question is calculated using the information gain maximization algorithm, and the corresponding script is generated and pushed to the receiving terminal. When patient profile data is missing or its quality is below a preset quality threshold, the inference weight coefficient of the patient profile is automatically reduced and the weight coefficient of the real-time semantic text is increased accordingly. When the number of concurrent consultation requests exceeds the preset concurrency threshold, the inference requests are batch-processed and scheduled, and requests with the same symptom type are merged, inferred in batches, and then returned separately.

[0061] It should be noted that this embodiment enhances the system's robustness in three aspects: information collection, data quality, and system concurrency. The "Information Completeness Score" is a comprehensive indicator used to measure the degree to which the currently collected information supports an accurate diagnosis. It is calculated as follows: the set of information fields required for each candidate disease in the inference chain is compared with the currently collected information set, and the ratio of the number of collected fields to the total number of required fields is calculated. This ratio is the information completeness score, with a value between 0 and 1. The "Preset Completeness Threshold" is generally set to 0.60 in actual deployment. When the score is lower than this threshold, the system determines that the current information is insufficient to support a reliable diagnosis and initiates progressive consultation guidance. The "Information Gain Maximization Algorithm" works as follows: for the currently collected symptom set, the information entropy (conditional entropy) about the candidate disease category that can be reduced after answering each candidate consultation question (each candidate question corresponds to an uncollected information field) is calculated. The question with the largest reduction in conditional entropy is selected as the next optimal consultation question. For example, if the currently collected symptom is "chest pain," candidate consultation questions include "pain nature (squeezing / stabbing / dull)" and "pain duration." The system calculates the reduction in information entropy for candidate diseases such as "angina" and "intercostal neuralgia" after answering these two questions, and prioritizes the question with the largest reduction. A "preset quality threshold" is used to assess the data quality of the patient profile. It is calculated by the ratio of the number of non-empty fields in the profile to the total number of fields, typically set to 0.50 in actual deployments. When the profile quality falls below this threshold, the system automatically reduces the weight coefficient of the profile feature vector in the comprehensive inference confidence fusion (typically from 0.50 to 0.20 in actual deployments), while correspondingly increasing the weight coefficient of the real-time semantic text feature vector (from 0.50 to 0.80) to ensure that inference primarily relies on information collected during the current consultation, avoiding the impact of unreliable historical data on inference quality. In actual deployment, the "preset concurrency threshold" is set to 70% of the system's maximum concurrent processing capacity as the threshold for triggering batch processing; "same symptom type" refers to requests where the intersection of the sets of symptom entities exceeds a certain proportion (generally 50% in actual deployment); "batch inference" refers to averaging the symptom feature vectors of multiple requests in the same batch to construct a unified query vector, which is then input into the inference engine at once. The inference engine can complete the inference of the entire batch of requests in one forward propagation, and the inference results are returned to each receiving terminal.

[0062] The progressive consultation guidance mechanism guides doctors to obtain the most critical information with the fewest questions possible through an intelligent information collection strategy, which not only speeds up the efficiency of information collection but also ensures the sufficiency of diagnostic information. The dynamic weight adjustment mechanism when profiles are missing enables the system to maintain basic reasoning ability even when historical data is unavailable, avoiding system failures caused by first-time patients or missing data. The concurrent batch processing scheduling mechanism improves the system's large-scale concurrent processing capability while ensuring diagnostic quality, enabling the system to smoothly handle a large number of concurrent requests during peak outpatient periods.

[0063] For example, a new patient only reports a "headache" with no other medical history. The current information completeness score is calculated to be 0.25 (far below the threshold of 0.60), and progressive consultation guidance is initiated. The information gain maximization algorithm finds that "headache location (whole head / one side / occipital region)" has the highest information gain (0.45), followed by "headache nature (persistent / intermittent / throbbing)" with an information gain of 0.32. The system prioritizes pushing the consultation question "Where is the pain mainly concentrated? Is it all over the head or just one side?" and generates corresponding scripts. After the doctor asks follow-up questions based on the guidance and obtains the patient's answer, the information completeness score improves to 0.58, approaching the threshold, and the system continues to push the second optimal consultation question. In a boundary scenario, a large tertiary hospital's internal medicine outpatient department had 32 concurrent consultation requests during the peak hours on Monday morning, exceeding the system's preset concurrent threshold (25). The system groups 32 requests by symptom type: 12 of the requests have "cough" as the main symptom. The system merges these 12 requests into a batch and constructs a batch query vector based on the average of the symptom feature vectors of the 12 requests. After one inference is completed, the results are mapped back to their respective reception terminals. This reduces the computational load that originally required 12 independent inferences to 1, significantly reducing the system's computational load during peak periods.

[0064] Please refer to Figure 3 , Figure 3 This is a schematic diagram illustrating the implementation steps of the real-time patient admission guidance method based on a dynamic inference engine and patient profile provided in this application embodiment.

[0065] The present invention also discloses a real-time patient consultation guidance system based on a dynamic inference engine and patient profiles, including a memory and a processor. The memory stores a real-time patient consultation guidance method program based on a dynamic inference engine and patient profiles. When the processor executes the real-time patient consultation guidance method program based on a dynamic inference engine and patient profiles, it implements the steps of the real-time patient consultation guidance method based on a dynamic inference engine and patient profiles.

[0066] This invention discloses a real-time patient reception guidance method and system based on a dynamic reasoning engine and patient profiling. It collects real-time audio streams of doctor-patient communication, performs speech recognition and semantic conversion, extracts semantic text such as symptom entities, and acquires historical medical data. It constructs patient profiles and identifies hierarchical tags indicating patient intention. The patient profiles, hierarchical tags indicating patient intention, and standard operating procedures (SOPs) are input into the dynamic reasoning engine. Based on a medical knowledge graph, multiple rounds of reasoning are performed, outputting judgment results for treatment nodes, reasoning chains, and communication risk warning information. Standardized treatment action guidance and professional communication scripts are generated and pushed to the patient reception terminal. After the patient reception is completed, intelligent review and iterative optimization of guidance strategies are performed. Through the deep coupling of multimodal semantic perception, dynamic reasoning, and patient reception SOPs, accurate judgment of treatment nodes, real-time warning of communication risks, and full-link guidance and autonomous iterative optimization of the patient reception process are achieved, effectively standardizing the patient reception process.

[0067] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0068] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0069] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

Claims

1. A real-time patient admission guidance method based on a dynamic reasoning engine and patient profiles, characterized in that: include: Acquire patient identification data and audio stream data of doctor-patient communication, perform speech recognition and semantic conversion processing on the audio stream data, extract semantic text and standardized medical terms, and extract historical medical data from a pre-set electronic health record database in combination with patient identification data. Patient profiles are constructed based on semantic text and historical medical data, and patients' willingness to seek medical treatment is hierarchically identified based on a pre-trained willingness classification model, outputting hierarchical labels of willingness to seek medical treatment. The patient profile, the hierarchical labels of the willingness to seek medical treatment, and the preset standard operating procedures for receiving patients are input into the dynamic reasoning engine. The dynamic reasoning engine performs multiple rounds of reasoning based on the pre-trained medical knowledge graph and outputs the judgment result of the current diagnosis and treatment node, the reasoning chain, and communication risk warning information. Based on the judgment results and reasoning chain of the diagnosis and treatment nodes, standardized diagnosis and treatment action guidance and professional communication scripts are generated, and communication risk warning information and diagnosis and treatment action guidance are pushed to the patient reception terminal in real time. After the patient is admitted, the entire admission process is intelligently reviewed, a patient admission review report is generated, and the admission SOP and the guidance strategy of the inference engine are iteratively optimized based on the deviation data in the review report.

2. The real-time patient admission guidance method based on a dynamic reasoning engine and patient profile as described in claim 1, characterized in that, The process of acquiring patient identification data and audio stream data of doctor-patient communication involves performing speech recognition and semantic conversion processing on the audio stream data to extract semantic text and standardized medical terminology, and combining the patient identification data with historical medical data extracted from a pre-set electronic health record database. Specifically, this includes: Audio signals from the patient reception environment are collected in real time using a microphone array, and the doctor's voice channel and the patient's voice channel are distinguished by a sound source separation algorithm. Noise filtering and speech endpoint detection are performed on the two speech streams respectively; The detected speech segments are automatically transcribed using speech recognition, and then semantic text is extracted using a preset medical named entity recognition model, including symptom entities, anatomical location entities, and time entities, and mapped to a standardized medical terminology system.

3. The real-time patient admission guidance method based on a dynamic reasoning engine and patient profile as described in claim 1, characterized in that, The process of constructing patient profiles based on semantic text and historical medical data, and performing stratified identification of patient visits based on a pre-trained intention classification model to output stratified labels for visits, specifically includes: Patient profiles are constructed based on semantic text and historical medical data. These profiles include demographic characteristics, past medical history, medication records, allergy history, family history of genetic diseases, and lifestyle data. The extracted historical medical data is sorted chronologically and missing values ​​are filled to generate structured patient profile feature vectors; Semantic text is processed to obtain semantic text feature vectors, including symptom entity vectors, anatomical site entity vectors, and time entity vectors; The intention classification model takes as input a concatenation of symptom entity vectors, anatomical site entity vectors, time entity vectors, and patient profile feature vectors, and outputs a hierarchical label of medical visit intention, including one or more of the following: initial diagnosis, follow-up visit, symptom consultation, and prescription renewal.

4. The real-time patient admission guidance method based on a dynamic reasoning engine and patient profile as described in claim 1, characterized in that, The process involves inputting patient profiles, stratified tags of treatment intentions, and preset standard operating procedures for patient reception into a dynamic reasoning engine. The dynamic reasoning engine performs multiple rounds of reasoning based on a pre-trained medical knowledge graph, outputting the judgment result of the current treatment node, the reasoning chain, and communication risk warning information, specifically including: The dynamic reasoning engine employs a knowledge graph reasoning model based on an attention mechanism; The query vector is obtained by concatenating the patient profile feature vector and the symptom entity vector, and the association strength with the candidate disease node is calculated through a multi-head attention mechanism. The multi-head attention mechanism includes symptom combination matching analysis based on symptom and symptom co-occurrence matrix, complication probability analysis based on disease and complication causal graph, and medication conflict risk analysis based on drug and contraindication conflict detection matrix. The three analytical scores and the correlation strength are integrated into a comprehensive inference confidence score. The inference chain and the corresponding intermediate feature data of the inference chain are output according to the confidence score.

5. The real-time patient admission guidance method based on a dynamic reasoning engine and patient profile as described in claim 4, characterized in that, The acquisition of the diagnostic and treatment node judgment results specifically includes: The intermediate features of the inference chain are matched with the preset SOP node sequence for patient reception, and the judgment result of the current treatment node is output. The current diagnosis and treatment node judgment result reflects the node type of the current reception process in the SOP. The preset node types include one of the following: symptom inquiry node, physical sign collection node, preliminary diagnosis node, examination ordering node, and treatment plan formulation node. Based on the judgment result of the current diagnosis node, the corresponding SOP branch process is dynamically invoked, and the SOP branch process is used as a constraint condition input into the inference engine to limit the output range of the inference chain in the current round.

6. The real-time patient admission guidance method based on a dynamic reasoning engine and patient profile according to claim 4, characterized in that, The output of risk warning information specifically includes: Communicating risk warning information includes risk type labels and corresponding response suggestions; Risk type labels include risk of missing key information, risk of abnormal patient emotions, and risk of information asymmetry between doctors and patients; When the necessary medical history information required for the reasoning chain does not appear in the patient profile or the current semantic text, the risk of missing key information is triggered. Based on a pre-set voice emotion recognition model, the patient's voice signal is classified into emotions. When the recognition result is negative emotion and the confidence level exceeds a pre-set threshold, the risk of abnormal patient emotions is triggered. When doctors use professional terminology more frequently than a preset density threshold and patients show signs of confusion in their feedback, the risk of information asymmetry between doctors and patients is triggered.

7. The real-time patient admission guidance method based on a dynamic reasoning engine and patient profile as described in claim 1, characterized in that, Before outputting the current diagnosis and treatment node's judgment result, reasoning chain, and communication risk warning information, it also includes: Obtain the patient's historical medical visit time series data, and output the patient's disease progression risk score within a preset future time period through a pre-trained disease progression prediction model; Assess the severity of each symptom and output a symptom severity score; Symptoms with severity scores above a preset threshold are marked as high-risk symptoms and given higher weight in the inference chain; By inputting disease progression risk scores and high-risk symptom markers as contextual features into the dynamic inference engine, "urgent treatment suggestions" are automatically displayed in advance for high-risk patients.

8. A real-time patient reception guidance system based on a dynamic reasoning engine and patient profiles, characterized in that, The system includes a memory and a processor. The memory contains a real-time patient admission guidance method program based on a dynamic inference engine and patient profiles. When executed by the processor, the real-time patient admission guidance method program based on the dynamic inference engine and patient profiles performs the following steps: Acquire patient identification data and audio stream data of doctor-patient communication, perform speech recognition and semantic conversion processing on the audio stream data, extract semantic text and standardized medical terms, and extract historical medical data from a pre-set electronic health record database in combination with patient identification data. Patient profiles are constructed based on semantic text and historical medical data, and patients' willingness to seek medical treatment is hierarchically identified based on a pre-trained willingness classification model, outputting hierarchical labels of willingness to seek medical treatment. The patient profile, the hierarchical labels of the willingness to seek medical treatment, and the preset standard operating procedures for receiving patients are input into the dynamic reasoning engine. The dynamic reasoning engine performs multiple rounds of reasoning based on the pre-trained medical knowledge graph and outputs the judgment result of the current diagnosis and treatment node, the reasoning chain, and communication risk warning information. Based on the judgment results and reasoning chain of the diagnosis and treatment nodes, standardized diagnosis and treatment action guidance and professional communication scripts are generated, and communication risk warning information and diagnosis and treatment action guidance are pushed to the patient reception terminal in real time. After the patient is admitted, the entire admission process is intelligently reviewed, a patient admission review report is generated, and the admission SOP and the guidance strategy of the inference engine are iteratively optimized based on the deviation data in the review report.

9. The real-time patient admission guidance system based on a dynamic reasoning engine and patient profile as described in claim 8, characterized in that, The process of acquiring patient identification data and audio stream data of doctor-patient communication involves performing speech recognition and semantic conversion processing on the audio stream data to extract semantic text and standardized medical terminology, and combining the patient identification data with historical medical data extracted from a pre-set electronic health record database. Specifically, this includes: Audio signals from the patient reception environment are collected in real time using a microphone array, and the doctor's voice channel and the patient's voice channel are distinguished by a sound source separation algorithm. Noise filtering and speech endpoint detection are performed on the two speech streams respectively; The detected speech segments are automatically transcribed using speech recognition, and then semantic text is extracted using a preset medical named entity recognition model, including symptom entities, anatomical location entities, and time entities, and mapped to a standardized medical terminology system.

10. The real-time patient admission guidance system based on a dynamic reasoning engine and patient profile as described in claim 9, characterized in that, The process of constructing patient profiles based on semantic text and historical medical data, and performing stratified identification of patient visits based on a pre-trained intention classification model to output stratified labels for visits, specifically includes: Patient profiles are constructed based on semantic text and historical medical data. These profiles include demographic characteristics, past medical history, medication records, allergy history, family history of genetic diseases, and lifestyle data. The extracted historical medical data is sorted chronologically and missing values ​​are filled to generate structured patient profile feature vectors; Semantic text is processed to obtain semantic text feature vectors, including symptom entity vectors, anatomical site entity vectors, and time entity vectors; The intention classification model takes as input a concatenation of symptom entity vectors, anatomical site entity vectors, time entity vectors, and patient profile feature vectors, and outputs a hierarchical label of medical visit intention, including one or more of the following: initial diagnosis, follow-up visit, symptom consultation, and prescription renewal.

Images