AI pre-inquiry analysis-based outpatient department accurate shunting and reservation guiding method

Abstract

The invention discloses an outpatient department accurate shunting and appointment guiding method based on AI pre-inquiry analysis. The method comprises the following steps that S1, a patient initiates pre-inquiry: the patient initiates a pre-inquiry request through a hospital APP, a self-service machine or an applet terminal; and S2: pushing the standardized inquiry questionnaire: the hospital APP, the self-service machine or the applet end pushes the standardized inquiry questionnaire through the shunting reservation system. By integrating multi-dimensional information such as texts, images, audios, patient ages and medical histories and by means of BERT semantic understanding, a medical knowledge graph and an XGBoost algorithm, the department matching accuracy is larger than or equal to 95%, the wrong number hanging rate is reduced to 3% or below, the limitation of single symptom matching is broken through, the cross-specialist differential diagnosis ability is improved, AI model driving replaces a rigid rule base, and the accuracy of diagnosis is improved. And iterative optimization is carried out through 1-20 thousand newly added data in each month, and rapid updating of the adaptive medical knowledge is realized.

Description

Technical Field

[0001] This invention relates to the field of smart outpatient technology, and in particular to a method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis. Background Technology

[0002] Traditional hospital outpatient triage and appointment systems employ a static, passive question-and-answer model, which cannot meet the high demands of modern smart hospitals for triage accuracy and patient experience. They also cannot handle high-concurrency consultation scenarios with tens of thousands of patients daily. Traditional systems suffer from low triage efficiency and a high misdiagnosis rate. In summary, traditional technologies have the following shortcomings: 1. Existing methods based on single symptom matching or fixed department trees only target keywords in the patient's chief complaint and do not integrate multi-dimensional information such as age, medical history, and symptom details, resulting in insufficient accuracy in triage recommendations; 2. Traditional software lacks the ability to deeply understand the semantics of patients' spoken descriptions, has weak contextual coherence in multi-turn dialogues, and poor information collection completeness; 3. The symptoms of diseases in different specialties are complex and intertwined, and the system lacks the ability to make differential diagnoses and inferences across specialties, resulting in low confidence in triage. 4. Traditional systems use rigid rule bases, and knowledge updates rely on manual maintenance, making it difficult to adapt to the rapid iteration of medical knowledge. Moreover, when faced with complex or rare diseases, the triage capability is particularly weak, and the system's intelligence and adaptability are poor. 5. Traditional procedures lack intelligent assessment of the urgency of the illness, and cannot identify potential signs of acute and critical illness to achieve priority treatment or early warning, leaving blind spots for patient safety; 6. The system data is isolated and fails to be linked with historical health records to provide personalized guidance. There is also a risk of storing and transmitting patient privacy information in plain text. Furthermore, the system lacks a medical risk warning mechanism and has insufficient privacy protection and personalized service capabilities. In summary, this application proposes a method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis. Summary of the Invention

[0003] Based on the technical problems existing in the background technology, this invention proposes a method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis.

[0004] The method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis proposed in this invention includes the following steps: S1: Patient initiates pre-consultation: Patients initiate pre-consultation requests through the hospital's APP, self-service machine, or mini-program. S2: Pushing standardized consultation questionnaires: The hospital's APP, self-service machine or mini-program pushes standardized consultation questionnaires through the triage and appointment system. The standardized pre-consultation questionnaire information includes the uploaded items of symptoms, medical history and physical signs. S3: Complete the standardized questionnaire and upload data: Patients complete the standardized pre-consultation questionnaire based on the HL7 FHIR standard in the triage appointment system. The content covers information on symptoms and medical history. At the same time, they upload pictures of physical signs and audio data to provide input for subsequent multimodal data processing. S4: System calls HIS SDK to obtain historical medical records: The triage and appointment system calls the hospital's HIS system SDK through Java Native Access technology to obtain patients' historical medical record data; S5: HIS system returns medical record data: The hospital's HIS system returns the patient's historical medical record data to the triage and appointment system; S6: Multimodal data preprocessing: Multimodal data preprocessing is performed on the collected questionnaire data, uploaded materials and acquired historical medical record data, and the multimodal features are weighted and fused through an attention mechanism to output a 512-dimensional standardized feature vector; S7: AI Three-Level Decision Making: Based on the XGBoost multi-classification algorithm, the patient's condition is divided into four levels: emergency, priority, routine, and non-outpatient. Through the fusion architecture of knowledge graph and BERT classification model, the department is accurately matched to achieve a three-level mapping of "symptom-disease-department". The greedy algorithm is used in combination with the department resource indicators collected by Prometheus at the minute level to recommend the optimal appointment time. S8: Push personalized appointment plans: The triage appointment system pushes personalized appointment plans to patients, including the top 3 matching departments, available appointment time slots, and estimated waiting time. S9: Patient confirms appointment: The patient confirms the appointment plan. If necessary, the plan can be adjusted. The system supports re-triggering the resource scheduling model after "plan adjustment". S10: Synchronize appointment information to departmental resource pool: The triage appointment system synchronizes patients' appointment information to the resource pool of outpatient departments; S11: Real-time feedback on departmental waiting progress: Outpatient departments provide real-time feedback on waiting progress to the triage and appointment system, including patient number and number of people waiting; S12: WebSocket push real-time appointment status: The triage and appointment system pushes real-time appointment status to patients through WebSocket technology, and patients can check their waiting progress at any time; S13: Patients should visit the outpatient department according to their appointment time. S14: Record the actual department visited and the diagnosis result: The outpatient department records the actual department visited and the diagnosis result of the patient and feeds the information back to the hospital HIS system; S15: Pushing Patient Feedback Data: The hospital's HIS system pushes patient feedback data to the triage and appointment system; S16: Update the AI ​​model training set and perform iterative optimization: The triage and appointment system uses patient feedback data to update the AI ​​model training set, adding 10,000 to 20,000 valid patient data entries per month, and iteratively optimizing the disease classification and department matching models to ensure the accuracy of department matching.

[0005] Preferably, in step S3, the patient-side data is processed by a protocol conversion module to parse the request body of data from different sources using the OkHttp3 framework, extract the fields of "symptom description, body temperature, and medical history", and convert them into the "Patient" and "Observation" resource formats of HL7 FHIR. In S4, the acquired patient historical medical record data is used by the SDK calling module to call the SDKs of Winning Health and Chuangye Huikang through JNA to read the patient's historical diagnostic records. The interface response latency is ≤200ms, achieving 100% compatibility with mainstream manufacturers. Data acquired through S3 and S4 is standardized from multiple sources through HL7 FHIR protocol adaptation and Drools rule validation, providing "clean raw materials" for the preprocessing layer. During Drools rule validation, the data validation module intercepts invalid data based on Drools rules, and the classification storage module allocates storage media according to data type. Structured data is stored in a MySQL cluster using a dual primary key index of "patient ID + acquisition time"; unstructured data is stored in MinIO, stored according to the path "patient ID / year / month"; and active data from the past hour is stored in Redis with a 3600-second expiration time. After storage, the response time for a single data query is ≤50ms, meeting the real-time requirements of subsequent preprocessing. The Drools rules include: (1) if (chief complaint.contains("fever")) then the "body temperature" field must exist and the value ∈ [35,42]℃; (2) if (medical history.contains("diabetes")) then the "fasting blood glucose" field must exist and the value ∈ [3.9,6.1]mmol / L; if the validation fails, a prompt to fill in the missing information will be returned in real time.

[0006] Preferably, in step S6, during multimodal data preprocessing, One-Hot encoding is used for discrete data; and Z-Score normalization is used for continuous data, with the following formula: The system is designed to eliminate dimensional differences and complete missing data based on a medical knowledge graph. For text data, the BERT-BiLSTM-CRF model is used, with the patient's free description text as input. The BERT layer obtains the contextual semantic vector, the BiLSTM layer captures temporal dependencies, and the CRF layer outputs entity labels. For image data, the ResNet-50 network is used, with MedicalNet pre-trained weights loaded. Convolutional operations are performed on skin rash images, and the output of the last fully connected layer is extracted as a 2048-dimensional feature vector. For audio data, cough audio is pre-emphasized, framed, and windowed. 13-dimensional MFCC coefficients are extracted and input into the GMM model to classify "dry cough / wet cough".

[0007] Preferably, in step S6, during the weighted fusion of multimodal features, an attention weight calculation function is constructed: ,in The medical relevance score for the i-th feature is obtained by weighting and summing the structured features, text features, image features, and audio features, and then reducing the dimensionality to 512 dimensions using PCA to provide standardized input for the AI ​​decision-making layer.

[0008] Preferably, in step S7, the logical steps for classifying the urgency of the illness are as follows: S7011: Training data and feature selection: 1.2 million historical outpatient data were used, and the top 30 key features were selected by SHAP value. Features with SHAP value ≥ 0.1 were included in the model. S7012: Model Training and Inference: Construct an XGBoost multi-classification model and set four levels of labels: (1) Emergency: Needs emergency treatment within 1 hour, sample proportion 5%; (2) Priority: Needs medical treatment within 48 hours, sample proportion 15%; (3) Routine: Needs medical treatment within 7 days, sample proportion 70%; (4) Non-outpatient: Needs home care, sample proportion 10%. S7013: The model uses 5-fold cross-validation, with a test set accuracy ≥ 0.93, an emergency case recall rate ≥ 0.98, and a single data inference time ≤ 100ms.

[0009] Preferably, the logical steps for precise department matching in S7 are as follows: S7021: Disease Prediction: Input 512-dimensional fused features, use BERT classification model to output TOP3 possible diseases and confidence scores; S7022: Disease-Department Mapping: Construct a knowledge graph of the association between "disease-department" and store standardized mapping relationships; S7023: Cross-department symptom conflict correction: For the cross-department symptom "headache", the patient's historical data is combined for correction, and the final department matching accuracy rate is ≥0.95.

[0010] Compared with existing technologies, the beneficial effects of this invention are: 1. By integrating multimodal data such as text, images, and audio, and combining multi-dimensional information such as age, medical history, and symptom details, it breaks through the limitations of single symptom matching. At the same time, with the help of the BERT semantic understanding model and medical knowledge graph, it realizes cross-specialty differential diagnosis and reasoning, with a department matching accuracy rate of ≥95% and the wrong appointment rate reduced to below 3%. It also links with the patient's historical health records, and through data completion and conflict correction, it further improves the confidence of triage and avoids misdiagnosis of complex conditions and rare diseases. 2. It adopts AI model-driven approach to replace rigid rule base, and iterates and optimizes with 10,000 to 20,000 new medical data every month to adapt to rapid updates in medical knowledge. It has strong contextual relevance in multi-turn dialogues, can deeply understand patients' spoken descriptions, ensures the integrity of information collection through entity recognition and feature extraction, and dynamically adjusts decision logic for different scenarios. It can still run stably in high-concurrency scenarios with tens of thousands of patients per day. 3. Based on the XGBoost algorithm, a four-level disease classification is implemented, with an emergency case recall rate of ≥0.98. It accurately identifies signs of acute and critical illness and prioritizes treatment, eliminating blind spots in patient safety. Furthermore, data transmission adopts standardized protocols and encrypted storage, and structured and unstructured data are classified and stored in isolation to avoid the risk of plaintext transmission of privacy information. A medical risk early warning mechanism is constructed, which combines historical medical records and real-time symptom data to provide early warning of potential health risks. 4. By dynamically scheduling departmental resources through a greedy algorithm and recommending optimal time slots based on real-time indicators such as the number of doctors on duty and the number of people waiting, the utilization rate of departments during peak hours reaches 85%-90%. At the same time, combined with online pre-consultation and personalized appointment schemes, the appointment success rate is ≥92%, avoiding invalid visits. In addition, by using WebSocket to push real-time appointment status, the traditional waiting time of more than 60 minutes is reduced to within 30 minutes, and the total medical treatment time is reduced by 40%. This invention integrates text, images, audio, and multi-dimensional information such as patient age and medical history. Leveraging BERT semantic understanding, medical knowledge graphs, and the XGBoost algorithm, it achieves a department matching accuracy rate of ≥95%, reduces the mis-registration rate to below 3%, overcomes the limitations of single-symptom matching, and improves cross-specialty differential diagnosis capabilities. Furthermore, it replaces a rigid rule base with an AI model-driven approach, iterating and optimizing with 10,000-20,000 new data entries monthly to adapt to rapid updates in medical knowledge. It also supports high-concurrency scenarios with tens of thousands of patients daily, with end-to-end decision-making time ≤500ms. It identifies acute and critical cases through a four-level disease classification system and prioritizes their treatment, while standardized encrypted transmission and categorized storage ensure privacy and security. Moreover, it dynamically schedules departmental resources using a greedy algorithm, combined with online pre-consultation, personalized appointments, and real-time WebSocket appointment push notifications, reducing waiting time to within 30 minutes and total medical time by 40%. This achieves a comprehensive improvement in triage accuracy, system intelligence, patient safety, and overall experience. Attached Figure Description

[0011] Figure 1 This is a diagram illustrating the overall framework of the AI-based pre-diagnosis analysis-based outpatient department precise triage and appointment guidance method proposed in this invention. Figure 2 This is a sequence diagram of the outpatient department precise triage and appointment guidance method based on AI pre-diagnosis analysis proposed in this invention; Figure 3 This is a data acquisition protocol adaptation block diagram for the outpatient department precise triage and appointment guidance method based on AI pre-diagnosis analysis proposed in this invention. Figure 4 This is a multimodal feature fusion diagram of the outpatient department precise triage and appointment guidance method based on AI pre-diagnosis analysis proposed in this invention; Figure 5 This is an AI three-level decision-making flowchart for the outpatient department precise triage and appointment guidance method based on AI pre-diagnosis analysis proposed in this invention. Detailed Implementation

[0012] The present invention will be further explained below with reference to specific embodiments.

[0013] Example Reference Figure 1-5 This embodiment proposes a method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis, including the following steps: S1: Patient initiates pre-consultation: Patients initiate pre-consultation requests through the hospital's APP, self-service machine, or mini-program. S2: Pushing standardized consultation questionnaires: The hospital's APP, self-service machine or mini-program pushes standardized consultation questionnaires through the triage and appointment system. The standardized pre-consultation questionnaire information includes the uploaded items of symptoms, medical history and physical signs. S3: Complete the standardized questionnaire and upload data: Patients complete the standardized pre-consultation questionnaire based on the HL7 FHIR standard in the triage appointment system. The content covers information on symptoms and medical history. At the same time, they upload pictures of physical signs and audio data to provide input for subsequent multimodal data processing. The patient-side data is processed by a protocol conversion module, which parses the request body of data from different sources using the OkHttp3 framework, extracts the fields of "symptom description, body temperature, and medical history", and converts them into the "Patient" and "Observation" resource formats of HL7 FHIR. S4: System calls HIS SDK to obtain historical medical records: The triage and appointment system calls the hospital's HIS system SDK through Java Native Access technology to obtain patients' historical medical record data. The SDK calling module calls the SDKs of Weining Health and Chuangye Huikang through JNA to read the patient's historical diagnosis records. The interface response latency is ≤200ms, achieving 100% compatibility with mainstream manufacturers. Data acquired through S3 and S4 is standardized from multiple sources through HL7 FHIR protocol adaptation and Drools rule validation, providing "clean raw materials" for the preprocessing layer. During Drools rule validation, the data validation module intercepts invalid data based on Drools rules, and the classification storage module allocates storage media according to data type. Structured data (ICD-11 symptom codes, medical history duration) is stored in a MySQL cluster using a dual primary key index of "patient ID + collection time"; unstructured data (skin rash images, cough audio) is stored in MinIO, stored according to the path "patient ID / year / month"; and active data from the past hour is stored in Redis with a 3600-second expiration time. After storage, the response time for a single data query is ≤50ms, meeting the real-time requirements of subsequent preprocessing. The Drools rules include: (1) if (chief complaint.contains("fever")) then the "body temperature" field must exist, and the value ∈ [35,42]℃; (2) if (medical history.contains("diabetes")) then the "fasting blood glucose" field must exist, and the value ∈ [3.9,6.1]mmol / L; when the validation fails, a prompt to fill in the missing information (such as "please fill in the body temperature value") is returned in real time to ensure that the data validation pass rate is ≥98%; S5: HIS system returns medical record data: The hospital's HIS system returns the patient's historical medical record data to the triage and appointment system; S6: Multimodal data preprocessing: Multimodal data preprocessing is performed on the collected questionnaire data, uploaded materials and acquired historical medical record data, and the multimodal features are weighted and fused through an attention mechanism to output a 512-dimensional standardized feature vector; When preprocessing multimodal data, discrete data (such as "allergy history: yes / no" and "symptom type: headache / abdominal pain") are encoded using One-Hot encoding, for example, "allergy history = yes" is encoded as [1,0], and "symptom type = headache" is encoded as [1,0,0,...]. Continuous data are standardized using Z-Score, with the following formula: This system is used to eliminate dimensional differences (such as the influence of units in body temperature "°C" and medical history "years") and, based on a medical knowledge graph, to complete missing data. For example, when a patient complains of "headache + nausea," it automatically associates the fields "may be accompanied by vomiting, photophobia," achieving a data completion accuracy of ≥92%. For text data, the BERT-BiLSTM-CRF model is used. The input is a patient's freely descriptive text (such as "recurrent headaches for the past 3 days, each lasting 2 hours, accompanied by nausea"). The BERT layer obtains the contextual semantic vector, the BiLSTM layer captures temporal dependencies, and the CRF layer outputs entity labels (such as "symptom name: headache" and "duration of attack: 2 hours / time"). The entity recognition F1 score is ≥0.94, and the processing time for a single text is ≤100ms. For image data, a ResNet-50 network is used, loaded with MedicalNet pre-trained weights (based on 100,000+). (Medical image training) Convolutional operations are performed on skin rash images, and the output of the last fully connected layer is extracted as a 2048-dimensional feature vector. The lesion region recognition accuracy is ≥0.91 (e.g., distinguishing image features between "eczema" and "urticaria"). For audio data, cough audio is pre-emphasized (high-pass filtering), framed (frame length 20ms, frame shift 10ms), and windowed (Hamming window) to extract 13-dimensional MFCC coefficients. These coefficients are then input into the GMM model to classify "dry cough / wet cough", with a classification accuracy of ≥0.88. When performing weighted fusion of multimodal features, construct the attention weight calculation function: ,in The medical relevance score for the i-th type of feature (e.g., chest pain text feature). =0.8, patient occupational characteristics =0.1), after weighted summation of structured features, text features, image features, and audio features, the dimensionality is reduced to 512 dimensions using PCA to provide standardized input for the AI ​​decision-making layer; S7: AI Three-Level Decision Making: Based on the XGBoost multi-classification algorithm, patient conditions are classified into four levels: emergency, priority, routine, and non-outpatient. The recall rate for emergency cases is ≥0.98. Through a knowledge graph and BERT classification model fusion architecture, accurate department matching is performed to achieve a three-level mapping of "symptom-disease-department". The department matching accuracy is ≥0.95. A greedy algorithm is used, combined with department resource indicators (number of attending physicians, number of patients waiting, and average consultation time) collected by Prometheus at the minute level, to recommend the optimal appointment time. The resource scheduling response time is ≤200ms. The logical steps for classifying the urgency of a medical condition are as follows: S7011: Training Data and Feature Selection: 1.2 million historical outpatient data (including symptoms, signs, final diagnosis, and consultation priority) were used. The top 30 key features (such as "duration of chest pain", "systolic blood pressure > 180 mmHg", and "confusion") were selected by SHAP value. Features with SHAP value ≥ 0.1 were included in the model (e.g., the SHAP value of chest pain duration = 0.8, which is the highest weight feature). S7012: Model Training and Inference: Construct an XGBoost multi-classification model and set four levels of labels: (1) Emergency: Needs emergency treatment within 1 hour, sample proportion 5%; (2) Priority: Needs medical treatment within 48 hours, sample proportion 15%; (3) Routine: Needs medical treatment within 7 days, sample proportion 70%; (4) Non-outpatient: Needs home care, sample proportion 10%. S7013: The model uses 5-fold cross-validation, with a test set accuracy ≥ 0.93, an emergency case recall rate ≥ 0.98, and a single data inference time ≤ 100ms; The logical steps for precise department matching are as follows: S7021: Disease Prediction: Input 512-dimensional fused features, use BERT classification model (pre-trained weights based on PubMed Central medical text dataset) to output TOP3 possible diseases and confidence levels. For example, "abdominal pain + diarrhea + fever 38.5℃" is predicted as "acute gastroenteritis (confidence level 92%), bacterial dysentery (confidence level 85%), viral enteritis (confidence level 78%)". The accuracy of TOP3 disease prediction is ≥0.91. S7022: Disease-Department Mapping: Construct a knowledge graph of disease-department associations, store standardized mapping relationships, such as "acute gastroenteritis → gastroenterology", "grade 3 hypertension → cardiology (hypertension clinic)", "retinopathy → ophthalmology (fundus clinic)", and support subdivided matching of departments; S7023: Cross-departmental symptom conflict correction: For cross-departmental symptoms of "headache", correction is made based on the patient's historical data. For example, "patient with a history of sinusitis → headache is the first choice for matching otolaryngology (confidence 90%)" and "patient with blurred vision → headache is the first choice for matching ophthalmology (confidence 88%)". The final department matching accuracy rate is ≥0.95. S8: Push personalized appointment plans: The triage appointment system pushes personalized appointment plans to patients, including the top 3 matching departments, available appointment time slots, and estimated waiting time. S9: Patient confirms appointment: The patient confirms the appointment plan. If necessary, the plan can be adjusted. The system supports re-triggering the resource scheduling model after "plan adjustment". S10: Synchronize appointment information to departmental resource pool: The triage appointment system synchronizes patients' appointment information to the resource pool of outpatient departments; S11: Real-time feedback on departmental waiting progress: Outpatient departments provide real-time feedback on waiting progress to the triage and appointment system, including patient number and number of people waiting; S12: WebSocket push real-time appointment status: The triage and appointment system pushes real-time appointment status to patients through WebSocket technology, and patients can check their waiting progress at any time; S13: Patients should visit the outpatient department according to their appointment time. S14: Record the actual department visited and the diagnosis result: The outpatient department records the actual department visited and the diagnosis result of the patient and feeds the information back to the hospital HIS system; S15: Pushing Patient Feedback Data: The hospital's HIS system pushes patient feedback data to the triage and appointment system; S16: Update the AI ​​model training set and perform iterative optimization: The triage and appointment system uses patient feedback data to update the AI ​​model training set, adding 10,000-20,000 valid patient data entries per month, and iteratively optimizing the disease grading and department matching models to ensure that the department matching accuracy improves by ≥0.5% per month; This embodiment integrates text, images, audio, and multi-dimensional information such as patient age and medical history. Leveraging BERT semantic understanding, medical knowledge graphs, and the XGBoost algorithm, it achieves a department matching accuracy rate of ≥95%, reduces the mis-registration rate to below 3%, overcomes the limitations of single-symptom matching, and improves cross-specialty differential diagnosis capabilities. Furthermore, it replaces a rigid rule base with an AI model-driven approach, iterating and optimizing with 10,000-20,000 new data entries monthly to adapt to rapid updates in medical knowledge. It also supports high-concurrency scenarios with tens of thousands of patients daily, with end-to-end decision-making time ≤500ms. It identifies acute and critical cases through a four-level disease classification system and prioritizes their treatment, while standardized encrypted transmission and categorized storage ensure privacy and security. Moreover, it dynamically schedules departmental resources using a greedy algorithm, combined with online pre-consultation, personalized appointments, and WebSocket real-time appointment push notifications, reducing waiting time to within 30 minutes and total medical time by 40%. This achieves a comprehensive improvement in triage accuracy, system intelligence, patient safety, and overall experience.

[0014] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis, characterized in that: Includes the following steps: S1: Patient initiates pre-consultation: Patients initiate pre-consultation requests through the hospital's APP, self-service machine, or mini-program. S2: Pushing standardized consultation questionnaires: The hospital's APP, self-service machine or mini-program pushes standardized consultation questionnaires through the triage and appointment system. The standardized pre-consultation questionnaire information includes the uploaded items of symptoms, medical history and physical signs. S3: Complete the standardized questionnaire and upload data: Patients complete the standardized pre-consultation questionnaire based on the HL7 FHIR standard in the triage appointment system. The content covers information on symptoms and medical history. At the same time, they upload pictures of physical signs and audio data to provide input for subsequent multimodal data processing. S4: System calls HIS SDK to obtain historical medical records: The triage and appointment system calls the hospital's HIS system SDK through Java Native Access technology to obtain patients' historical medical record data; S5: HIS system returns medical record data: The hospital's HIS system returns the patient's historical medical record data to the triage and appointment system; S6: Multimodal data preprocessing: Multimodal data preprocessing is performed on the collected questionnaire data, uploaded materials and acquired historical medical record data, and the multimodal features are weighted and fused through an attention mechanism to output a 512-dimensional standardized feature vector; S7: AI Three-Level Decision Making: Based on the XGBoost multi-classification algorithm, the patient's condition is divided into four levels: emergency, priority, routine, and non-outpatient. Through the fusion architecture of knowledge graph and BERT classification model, the department is accurately matched to achieve a three-level mapping of "symptom-disease-department". The greedy algorithm is used in combination with the department resource indicators collected by Prometheus at the minute level to recommend the optimal appointment time. S8: Push personalized appointment plans: The triage appointment system pushes personalized appointment plans to patients, including the top 3 matching departments, available appointment time slots, and estimated waiting time. S9: Patient confirms appointment: The patient confirms the appointment plan. If necessary, the plan can be adjusted. The system supports re-triggering the resource scheduling model after "plan adjustment". S10: Synchronize appointment information to departmental resource pool: The triage appointment system synchronizes patients' appointment information to the resource pool of outpatient departments; S11: Real-time feedback on departmental waiting progress: Outpatient departments provide real-time feedback on waiting progress to the triage and appointment system, including patient number and number of people waiting; S12: WebSocket push real-time appointment status: The triage and appointment system pushes real-time appointment status to patients through WebSocket technology, and patients can check their waiting progress at any time; S13: Patients should visit the outpatient department according to their appointment time. S14: Record the actual department visited and the diagnosis result: The outpatient department records the actual department visited and the diagnosis result of the patient and feeds the information back to the hospital HIS system; S15: Pushing Patient Feedback Data: The hospital's HIS system pushes patient feedback data to the triage and appointment system; S16: Update the AI ​​model training set and perform iterative optimization: The triage and appointment system uses patient feedback data to update the AI ​​model training set, adding 10,000 to 20,000 valid patient data entries per month, and iteratively optimizing the disease classification and department matching models to ensure the accuracy of department matching.

2. The method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis according to claim 1, characterized in that, In S3, the patient-side data is processed by the protocol conversion module, which parses the request body of data from different sources through the OkHttp3 framework, extracts the fields of "symptom description, body temperature, and medical history", and converts them into the "Patient" and "Observation" resource formats of HL7 FHIR. In S4, the acquired patient historical medical record data is used by the SDK calling module to call the SDKs of Winning Health and Chuangye Huikang through JNA to read the patient's historical diagnostic records. The interface response latency is ≤200ms, achieving 100% compatibility with mainstream manufacturers. Data acquired through S3 and S4 is standardized from multiple sources through HL7 FHIR protocol adaptation and Drools rule validation, providing "clean raw materials" for the preprocessing layer. During Drools rule validation, the data validation module intercepts invalid data based on Drools rules, and the classification storage module allocates storage media according to data type. Structured data is stored in a MySQL cluster using a dual primary key index of "patient ID + acquisition time"; unstructured data is stored in MinIO, stored according to the path "patient ID / year / month"; and active data from the past hour is stored in Redis with a 3600-second expiration time. After storage, the response time for a single data query is ≤50ms, meeting the real-time requirements of subsequent preprocessing. The Drools rules include: (1) if (chief complaint.contains("fever")) then the "body temperature" field must exist and the value ∈ [35,42]℃; (2) if (medical history.contains("diabetes")) then the "fasting blood glucose" field must exist and the value ∈ [3.9,6.1]mmol / L; if the validation fails, a prompt to fill in the missing information will be returned in real time.

3. The method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis according to claim 1, characterized in that, In step S6, during multimodal data preprocessing, One-Hot encoding is used for discrete data, and Z-Score normalization is used for continuous data. The formula is as follows: To eliminate dimensional differences and complete missing data based on a medical knowledge graph, the BERT-BiLSTM-CRF model is used for text data. The patient's free description text is input, the BERT layer obtains the context semantic vector, the BiLSTM layer captures the temporal dependency, and the CRF layer outputs the entity label. For image data, the ResNet-50 network is used, with MedicalNet pre-trained weights loaded. Convolution operations are performed on skin rash images, and the output of the last fully connected layer is extracted as a 2048-dimensional feature vector. For audio data, cough audio is pre-emphasized, framed, and windowed. 13-dimensional MFCC coefficients are extracted and input into the GMM model to classify "dry cough / wet cough".

4. The method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis according to claim 1, characterized in that, In step S6, during the weighted fusion of multimodal features, an attention weight calculation function is constructed: ,in The medical relevance score for the i-th feature is obtained by weighting and summing the structured features, text features, image features, and audio features, and then reducing the dimensionality to 512 dimensions using PCA to provide standardized input for the AI ​​decision-making layer.

5. The method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis according to claim 1, characterized in that, In S7, the logical steps for classifying the urgency of the illness are as follows: S7011: Training data and feature selection: 1.2 million historical outpatient data were used, and the top 30 key features were selected by SHAP value. Features with SHAP value ≥ 0.1 were included in the model. S7012: Model Training and Inference: Construct an XGBoost multi-classification model and set four levels of labels: (1) Emergency: Needs emergency treatment within 1 hour, sample proportion 5%; (2) Priority: Needs medical treatment within 48 hours, sample proportion 15%; (3) Routine: Needs medical treatment within 7 days, sample proportion 70%; (4) Non-outpatient: Needs home care, sample proportion 10%. S7013: The model uses 5-fold cross-validation, with a test set accuracy ≥ 0.93, an emergency case recall rate ≥ 0.98, and a single data inference time ≤ 100ms.

6. The method for precise triage and appointment guidance in outpatient departments based on AI pre-diagnosis analysis according to claim 1, characterized in that, In S7, the logical steps for precise department matching are as follows: S7021: Disease Prediction: Input 512-dimensional fused features, use BERT classification model to output TOP3 possible diseases and confidence scores; S7022: Disease-Department Mapping: Construct a knowledge graph of the association between "disease-department" and store standardized mapping relationships; S7023: Cross-department symptom conflict correction: For the cross-department symptom "headache", the patient's historical data is combined for correction, and the final department matching accuracy rate is ≥0.95.

Images