CN-121983277-A - Outpatient department accurate diversion and reservation guiding method based on AI pre-consultation analysis

Abstract

The invention discloses an outpatient department accurate diversion and appointment guiding method based on AI pre-consultation analysis, which comprises the following steps of S1, initiating pre-consultation by a patient, initiating a pre-consultation request by a hospital APP, a self-help machine or an applet terminal, and S2, pushing a standardized questionnaire, and pushing the standardized questionnaire by the hospital APP, the self-help machine or the applet terminal through a diversion appointment system. According to the invention, through integrating multi-dimensional information such as texts, images, audios, ages and medical histories of patients and the like, by means of BERT semantic understanding, medical knowledge graph and XGBoost algorithm, the accuracy of department matching is more than or equal to 95%, the error number rate is reduced to below 3%, the limitation of single symptom matching is broken through, the cross-specialty differential diagnosis capability is improved, an AI model is used for driving a substitution rigidifying rule base, iterative optimization is carried out through 1-2 ten thousand pieces of newly added data per month, and the adaptive medical knowledge is updated rapidly.

Inventors

• CHEN WENJIE
• XIE JIAHAN

Assignees

• 厦门狄耐克物联智慧科技有限公司

Dates

Publication Date

20260505

Application Date

20251212

Claims (6)

1. 1. The outpatient department accurate diversion and reservation guiding method based on AI pre-consultation analysis is characterized by comprising the following steps: S1, a patient initiates a pre-consultation, namely, the patient initiates a pre-consultation request through a hospital APP, a self-service machine or a small program end; S2, pushing a standardized questionnaire, namely pushing the standardized questionnaire by a hospital APP, a self-service machine or a small program end through a shunt reservation system, wherein standardized pre-questionnaire information comprises symptoms, medical history and sign uploading items; S3, filling in a standardized questionnaire and uploading data, namely filling in a standardized pre-questionnaire based on HL7 FHIR standard in a shunt reservation system by a patient, wherein the content covers information of symptoms and medical history, and uploading sign pictures and audio data to provide input for subsequent multi-mode data processing; S4, the system calls an HIS SDK to obtain a history medical record, wherein the diversion reservation system calls the SDK of the hospital HIS system through Java NATIVE ACCESS technology to obtain patient history medical record data; S5, the HIS system returns medical record data, namely, the hospital HIS system returns the history medical record data of the patient to the shunting appointment system; S6, preprocessing the multi-modal data, namely preprocessing the acquired questionnaire data, the uploaded data and the acquired history medical record data, weighting and fusing the multi-modal features through a attention mechanism, and outputting 512-dimensional standardized feature vectors; s7, AI three-level decision-making, namely dividing the disease condition of a patient into four levels of emergency, priority, routine and non-outpatient service based on XGBoost multi-classification algorithm, carrying out department accurate matching through a knowledge graph and BERT classification model fusion framework, realizing 'symptom-disease-department' three-level mapping, adopting greedy algorithm, and recommending optimal reservation period by combining department resource indexes acquired by Prometheus in minute; S8, pushing an individualized reservation scheme, namely pushing the individualized reservation scheme to a patient by a shunting reservation system, wherein the individualized reservation scheme comprises a TOP3 matched department, a reserved time period and estimated waiting time; S9, patient confirmation reservation, namely, patient confirmation reservation scheme, if an adjustable scheme is needed, the system supports scheme adjustment and then re-triggers the resource scheduling model; S10, synchronizing reservation information to a department resource pool, wherein a diversion reservation system synchronizes reservation information of a patient to the resource pool of an outpatient department; s11, feeding back the waiting progress of the department in real time, wherein the outpatient department feeds back the waiting progress to the diversion reservation system in real time, and the waiting progress comprises the serial numbers of patients and the waiting number of people; s12, a WebSocket pushing real-time waiting state, wherein the shunt reservation system pushes the real-time waiting state to a patient through a WebSocket technology, and the patient can check the waiting progress at any time; s13, the patient visits according to the reserved time period, wherein the patient visits an outpatient department according to the reserved time period; s14, recording actual consultation departments and diagnosis results, namely recording the actual consultation departments and diagnosis results of patients by the outpatient departments, and feeding information back to a hospital HIS system; S15, pushing the diagnosis feedback data, namely pushing the diagnosis feedback data of the patient to a shunting appointment system by a hospital HIS system; And S16, updating an AI model training set for iterative optimization, wherein the diversion reservation system updates the AI model training set by utilizing the diagnosis feedback data, newly adds 1-2 ten thousand pieces of effective diagnosis data per month, and iteratively optimizes the illness state grading and department matching models to ensure the accuracy of department matching.
2. 2. The AI-pre-consultation analysis-based outpatient department accurate diversion and appointment guiding method according to claim 1, wherein in the step S3, patient-side data analyze request bodies of different sources through OkHttp frames by a protocol conversion module, extract fields of 'symptom description, body temperature and medical history' and convert the fields into 'Patient' resource format of HL7 FHIR; In the step S4, the SDK calling module calls Wei Ning healthy and entrepreneur intelligent Kang Anshang SDK through JNA, reads the historical diagnosis record of the patient, and realizes 100% of main stream manufacturer compatibility, wherein the response delay of the interface is less than or equal to 200 ms; The data obtained through S3 and S4 are all subjected to HL7 FHIR protocol adaptation and Drools rule verification to realize multi-source data standardization, clean raw materials are provided for a preprocessing layer, invalid data are intercepted based on Drools rules through a data verification module during Drools rule verification, storage media are distributed according to data types through a classification storage module, wherein structured data are stored in a MySQL cluster, a 'patient ID+acquisition time' double-primary key index is adopted, unstructured data are stored in MinIO according to a 'patient ID/year/month' path, active data are stored in Redis for 1 hour, 3600S expiration time is set, and when the data are read later, the single data inquiry response time is less than or equal to 50ms, so that the real-time requirement of subsequent preprocessing is met; The Drools rule includes (1) if (complaint, contins ("fever")) the field must exist for "body temperature" and the value e 35,42℃, (2) if (medical history, contins ("diabetes")) the field must exist for "fasting blood glucose" and the value e 3.9,6.1 mmol/L, and the refill prompt is returned in real time when the verification fails.
3. 3. The method for accurately shunting and guiding appointment in outpatient departments based on AI pre-consultation analysis according to claim 1, wherein in the step S6, when multi-mode data preprocessing is performed, one-Hot encoding is adopted for discrete data, and Z-Score standardization is adopted for continuous data, wherein the formula is as follows: the method comprises the steps of eliminating dimension differences, supplementing missing data based on medical knowledge patterns, adopting a BERT-BiLSTM-CRF model for text data, inputting a patient free description text, acquiring upper Wen Yuyi vector and lower Wen Yuyi vector by a BERT layer, capturing time sequence dependence by a BiLSTM layer, outputting entity labels by a CRF layer, adopting ResNet-50 network for image data, loading MEDICALNET pre-training weight, carrying out convolution operation on skin rash pictures, extracting the final full-connection layer output as 2048-dimensional feature vectors, carrying out pre-emphasis, framing and windowing treatment on cough audio data, extracting 13-dimensional MFCC coefficients, and inputting GMM model classification 'dry cough/wet cough'.
4. 4. The AI-pre-consultation-analysis-based outpatient department accurate diversion and appointment guiding method according to claim 1, wherein in the step S6, a attention weight calculation function is constructed during multi-modal feature weighted fusion: wherein And (3) carrying out weighted summation on the structural features, the text features, the image features and the audio features for the medical relevance score of the ith type of features, and then reducing the dimension to 512 dimensions through PCA (principal component analysis), thereby providing standardized input for an AI decision layer.
5. 5. The outpatient department accurate diversion and appointment guiding method based on AI pre-consultation analysis of claim 1, wherein in S7, the logic step of grading the emergency degree of the illness state is as follows: s7011, training data and feature selection, namely screening TOP30 key features through SHAP values by adopting 120 ten thousand pieces of historical clinic data, wherein the SHAP values are more than or equal to 0.1 of feature inclusion model; S7012, model training and reasoning, namely constructing XGBoost multi-classification models, and setting four-level labels (1), emergency treatment within 1 hour, sample accounting for 5 percent, (2), priority, treatment within 48 hours, sample accounting for 15 percent, (3), conventional, treatment within 7 days, sample accounting for 70 percent, and (4), non-outpatient treatment, wherein the sample accounts for 10 percent; S7013, 5-fold cross verification is adopted for the model, the accuracy of a test set is more than or equal to 0.93, the recall rate of an emergency case is more than or equal to 0.98, and the single data reasoning time is less than or equal to 100ms.
6. 6. The outpatient department accurate diversion and reservation guiding method based on AI pre-consultation analysis of claim 1, wherein in S7, the logic steps during department accurate matching are as follows: S7021, disease prediction, namely inputting 512-dimensional fusion characteristics, and outputting possible diseases and confidence of TOP3 by adopting a BERT classification model; S7022, disease-department mapping, namely constructing a disease-department association knowledge graph and storing a standardized mapping relation; s7023, correcting cross-department symptom conflict, namely correcting the 'headache' cross-department symptom by combining with the historical data of a patient, wherein the final department matching accuracy rate is more than or equal to 0.95.

Description

Outpatient department accurate diversion and reservation guiding method based on AI pre-consultation analysis Technical Field The invention relates to the technical field of intelligent outpatient service, in particular to an outpatient department accurate diversion and reservation guiding method based on AI pre-consultation analysis. Background The traditional hospital outpatient diagnosis and appointment system adopts a static and passive question-answer mode, can not meet the high requirements of modern intelligent hospitals on diagnosis accuracy and patient experience, can not cope with the high concurrence consultation scene of ten-thousand patients in daily consultation, has low diagnosis efficiency and high misdiagnosis rate, and has the following defects in combination: 1. the existing method based on single symptom matching or fixed department tree is only aimed at keywords of patient complaints, does not integrate multidimensional information such as age, medical history, symptom details and the like, and results in insufficient accuracy of triage recommendation; 2. Traditional software lacks deep semantic understanding capability for patient spoken language description, and has weak relevance of multi-turn dialogue context and poor information collection integrity; 3. The symptoms of different specialty diseases represent complex interweaving, the system lacks the differential diagnosis and reasoning capability of the cross specialty, and the differential diagnosis confidence is low; 4. The traditional system adopts a stiff rule base, knowledge updating depends on manual maintenance, is difficult to adapt to rapid iteration of medical knowledge, has weak diagnosis separating capability when facing complex diseases and rare diseases, and has poor system intellectualization and self-adaptation capability; 5. The traditional flow lacks an intelligent evaluation function for the emergency degree of the illness state, and can not identify the symptom of the potential critical serious illness so as to realize the prior diagnosis or early warning, and the safety of the patient has a dead zone; 6. The system data is isolated, cannot be linked with the historical health record to provide personalized guidance, meanwhile, risks of plaintext storage and transmission of patient privacy information exist, and a system medical risk early warning mechanism is absent, so that privacy protection and personalized service capability are insufficient; In view of the above, the application provides an outpatient department accurate diversion and reservation guiding method based on AI pre-consultation analysis. Disclosure of Invention Based on the technical problems in the background technology, the invention provides an outpatient department accurate diversion and reservation guiding method based on AI pre-consultation analysis. The invention provides an outpatient department accurate diversion and reservation guiding method based on AI pre-consultation analysis, which comprises the following steps: S1, a patient initiates a pre-consultation, namely, the patient initiates a pre-consultation request through a hospital APP, a self-service machine or a small program end; S2, pushing a standardized questionnaire, namely pushing the standardized questionnaire by a hospital APP, a self-service machine or a small program end through a shunt reservation system, wherein standardized pre-questionnaire information comprises symptoms, medical history and sign uploading items; S3, filling in a standardized questionnaire and uploading data, namely filling in a standardized pre-questionnaire based on HL7 FHIR standard in a shunt reservation system by a patient, wherein the content covers information of symptoms and medical history, and uploading sign pictures and audio data to provide input for subsequent multi-mode data processing; S4, the system calls an HIS SDK to obtain a history medical record, wherein the diversion reservation system calls the SDK of the hospital HIS system through Java NATIVE ACCESS technology to obtain patient history medical record data; S5, the HIS system returns medical record data, namely, the hospital HIS system returns the history medical record data of the patient to the shunting appointment system; S6, preprocessing the multi-modal data, namely preprocessing the acquired questionnaire data, the uploaded data and the acquired history medical record data, weighting and fusing the multi-modal features through a attention mechanism, and outputting 512-dimensional standardized feature vectors; s7, AI three-level decision-making, namely dividing the disease condition of a patient into four levels of emergency, priority, routine and non-outpatient service based on XGBoost multi-classification algorithm, carrying out department accurate matching through a knowledge graph and BERT classification model fusion framework, realizing 'symptom-disease-department' three-level mapping, adopting greedy algorithm, and