An AI-assisted medical guide system and method

Abstract

The application provides an AI-assisted medical guide system and method, first complaint information of a patient is collected, a pre-trained AI feature extraction model is used for processing to obtain a first feature vector set; an abnormality monitoring is used for elimination to obtain a second feature vector set, and the second feature vector set is matched with a preset department feature vector database to determine a target department. When the matching fails, a directional interrogation question is generated based on the department with the highest current matching degree, the second feature vector set is combined with the feedback answer of the patient to generate second complaint information, and repeated analysis is performed until the target department is determined, finally, the target department is recommended to the patient, and a recommendation reason is generated based on the department information of the day to display, which can reduce the interference of abnormal complaint features on department matching, and when the information is insufficient, the directional screening supplementary interrogation content is used to reduce invalid questioning, and an abnormality processing, department matching, secondary interrogation and reanalysis closed loop are formed, and the stability and effectiveness of automatic department division are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence natural language processing technology, and in particular to an AI-assisted medical triage system and method. Background Technology

[0002] With the continuous growth in demand for medical services, patients typically need to choose the appropriate department based on their symptoms before seeking medical care. Due to the highly specialized nature of medical knowledge, patients' judgment regarding the correct department for their symptoms is easily affected by factors such as incomplete expression, disordered time sequence, slips of the tongue, or memory lapses, leading to problems such as registering with the wrong department, repeated queuing, or repeated consultations. Furthermore, manual triage is easily limited by the volume of inquiries, staff experience, and service hours, making it difficult to consistently process large amounts of medical information.

[0003] Existing AI-powered triage solutions typically acquire patients' chief complaints, extract features using natural language processing models, and then determine the target department through classification models or standard information retrieval methods. While these solutions can achieve basic automatic departmentalization, they still suffer from the following technical problems: First, patient complaints often consist of symptom descriptions from multiple time points. Existing solutions usually classify or retrieve all features directly, lacking mechanisms for identifying and removing fragmented, ambiguous, and abnormal features, which may interfere with departmental matching. Second, when current information is insufficient, existing solutions often generate questions based on accompanying symptoms, fixed scenarios, or general interaction models, failing to further narrow down the question scope using representative features of the most likely department. Third, there is a lack of unified data flow and control relationships between abnormal data processing, departmental matching evaluation, secondary consultation triggering, and supplementary information re-analysis, making it difficult to balance departmental stability and consultation efficiency.

[0004] Therefore, the present invention provides an AI-assisted medical triage system and method. Summary of the Invention

[0005] This invention discloses an AI-assisted medical triage system and method, which evaluates and removes anomalies in the patient's chief complaint, and then matches the cleaned feature set with the department feature vector dataset. When the matching evaluation is insufficient, targeted questions are generated and sorted using the representative features of the most likely department, and supplementary answers are re-incorporated into the analysis to form a closed-loop consultation process.

[0006] This invention provides an AI-assisted medical triage system, comprising: The information acquisition module is used to acquire the patient's basic information, medical history data, symptom description, and medical record information, and at the same time record the time series information corresponding to each symptom description and each medical record information to generate the first chief complaint information; The AI ​​extraction module is used to extract features from the first chief complaint information using a pre-trained AI feature extraction model to obtain a first set of feature vectors. The information analysis module is used to detect and remove anomalies in the first feature vector set to obtain a second feature vector set, and to match the second feature vector set with a preset department feature vector database to determine the target department. The matching analysis module is used to generate targeted consultation questions based on the department with the highest matching degree when the matching fails. It combines the patient's feedback answer with the second feature vector set to generate second chief complaint information, and repeats the analysis until the target department is determined. The department recommendation module is used to recommend the target department to the patient, and generate and display the recommendation reasons based on the department information corresponding to the target department for the day.

[0007] In one implementable manner, the information acquisition module includes: The multi-source data acquisition unit is used to collect the patient's basic information, medical history data, symptom description, and medical record information respectively. The time sequence information annotation unit is used to sort the occurrence time corresponding to each symptom description and each medical record information for the first time, and then sort the order of the descriptions corresponding to each symptom description and each medical record information for the second time, so as to generate corresponding time sequence information. The information association and integration unit is used to associate and bind basic information, medical history data, symptom descriptions, medical record information and corresponding time series information to obtain a structured information combination; The chief complaint information generation unit is used to determine the first chief complaint information by combining the associated structured information and output it.

[0008] In one implementable manner, the AI ​​extraction module includes: The model calling unit is used to call the pre-trained BERT natural language processing model or Sentence-BERT natural language processing model based on Chinese medical corpus, and output feature vectors with dimensions of 384 to 1024. The vector generation unit is used to segment the first chief complaint information into paragraphs according to symptom features and information features, and input the obtained paragraphs into the natural language processing model to obtain the initial feature vector corresponding to each paragraph. The time binding unit is used to bind the time sequence information of the symptom description or medical record information to the corresponding paragraph in the first chief complaint information to each of the initial feature vectors; The normalization unit is used to perform L2 normalization on the initial feature vector of each bound time series to obtain the first feature vector set and transmit it to the information analysis module.

[0009] In one implementable manner, the information analysis module includes: The time-series sorting unit is used to sort the feature vectors in the first feature vector set according to the time-series information bound to each first feature vector to obtain a first feature vector sequence; An abnormal feature removal unit is used to evaluate each data point in the first feature vector sequence, generate an evaluation value sequence, mark anomalies based on the adjacent slope difference of the evaluation value sequence, define data points whose cumulative number of anomaly markings reaches a preset marking threshold as abnormal data points, remove the first feature vector corresponding to the abnormal data points, and obtain the second feature vector set. The department matching calculation unit is used to perform similarity calculations between the second feature vector set and the feature vector dataset corresponding to each department in the preset department feature vector database, and obtain the department matching evaluation value corresponding to each department. The target department determination unit is used to extract the maximum value among all department matching evaluation values ​​as the maximum department matching evaluation value. The maximum department matching evaluation value is compared with the preset department matching threshold. When the maximum department matching evaluation value is greater than or equal to the preset department matching threshold, the department is regarded as the target department. If the value is less than the maximum evaluation value for department matching and the corresponding department information, the matching analysis module will be sent to the matching analysis module.

[0010] In one implementable manner, the matching analysis module includes: The candidate department locking unit is used to extract the department with the highest evaluation value of department matching as a candidate department when the matching fails, and to retrieve the feature vector dataset corresponding to the candidate department. The target feature filtering unit is used to filter feature vectors whose local consultation evaluation values ​​meet the preset local consultation evaluation threshold from the feature vector dataset of the candidate departments as target feature vectors, and extract the target keywords corresponding to each target feature vector; The consultation question sorting unit is used to retrieve a pre-defined departmental consultation question dataset based on target keywords, calculate the evaluation value of each initial consultation question, filter out targeted consultation questions by using the average evaluation value of all initial consultation questions as a threshold, and determine the consultation order according to the evaluation value from high to low. The iterative analysis trigger unit is used to obtain the patient's feedback answers to each of the aforementioned targeted consultation questions, integrate the feedback answers with the second feature vector set to generate the second chief complaint information, and return to the feature extraction and analysis process to repeat until the target department is determined.

[0011] In one implementable manner, the anomaly removal unit includes: The neighborhood evaluation subunit is used to perform L2 normalization on each of the first feature vectors in the first feature vector sequence, construct a neighborhood with a preset distance as the radius, calculate the average similarity between each of the first feature vectors in the neighborhood and the neighborhood data points, take the maximum value of all the average similarity values ​​as the global maximum value, and take the ratio of the average similarity value corresponding to each of the first feature vectors to the global maximum value as the evaluation value of the corresponding first feature vector. The slope calculation subunit is used to construct a coordinate system by normalizing the time series of the first feature vector as the horizontal axis and the evaluation value as the vertical axis, calculate the slope of the line connecting adjacent data points and the difference between adjacent slopes, compare the difference with a preset difference threshold, and mark the data points in the corresponding window as abnormal and accumulate the number of markings when the difference is greater than or equal to the preset difference threshold. The labeling and removal subunit is used to define data points whose cumulative number of abnormal labelings is greater than or equal to a preset labeling threshold as abnormal data points and remove the first feature vector corresponding to the abnormal data points. The cleaning output subunit is used to form a second feature vector set from the remaining first feature vectors after removal, and output it to the department matching subunit.

[0012] In one implementable manner, the department recommendation module includes: The department information retrieval unit is used to obtain basic information about the target department and the department's outpatient schedule for the day. The recommendation reason generation unit is used to generate corresponding recommendation reasons based on the matching criteria of the target department and the department information of the day; The registration guide generation unit is used to generate corresponding registration guide information based on the registration status of the target department on that day. The recommendation results display unit is used to output the target department, the reason for recommendation, and registration guidance information to the patient's end for display.

[0013] This invention provides an AI-assisted medical triage method, comprising: Step 1: Obtain the patient's basic information, medical history data, symptom description, and medical record information, and record the time series information corresponding to each symptom description and each medical record information to generate the first chief complaint information; Step 2: Extract features from the first chief complaint information using a pre-trained AI feature extraction model to obtain a first set of feature vectors; Step 3: Perform anomaly detection and removal on the first feature vector set to obtain the second feature vector set, and match the second feature vector set with the preset department feature vector database to determine the target department; Step 4: When a match fails, generate targeted consultation questions based on the department with the highest current matching degree, combine the patient's feedback answers with the second feature vector set to generate second chief complaint information, and repeat the analysis until the target department is determined; Step 5: Recommend the target department to the patient, and generate and display the recommendation reason based on the department information corresponding to the target department for the day.

[0014] In one feasible embodiment, step 4 includes: Step 41: When a match fails, extract the department with the highest departmental matching evaluation value as a candidate department, and retrieve the feature vector dataset corresponding to the candidate department. Step 42: Select feature vectors whose local consultation evaluation values ​​meet the preset local consultation evaluation threshold from the feature vector dataset of the candidate departments as target feature vectors, and extract the target keywords corresponding to each target feature vector; Step 43: Retrieve the preset departmental consultation question dataset based on the target keywords, calculate the evaluation value of each initial consultation question, use the average evaluation value of all initial consultation questions as the threshold to filter out targeted consultation questions, and determine the consultation order according to the evaluation value from high to low; Step 44: Obtain the patient's feedback answers to each of the aforementioned targeted consultation questions, integrate the feedback answers with the second feature vector set to generate the second chief complaint information, and return to the feature extraction and analysis process to repeat until the target department is determined.

[0015] The beneficial effects that this invention can achieve are: 1. Reducing the interference of vague, abrupt, or clinically unreasonable chief complaint features on department matching results.

[0016] 2. When there is insufficient information, select supplementary consultation content from the characteristics and questions most likely related to the current department, and reduce irrelevant questions.

[0017] 3. To create a closed data loop that connects abnormal data processing, department matching, secondary consultation, and reanalysis, thereby improving the stability and effectiveness of automatic departmental classification.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0019] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram illustrating the composition of an AI-assisted medical triage system according to an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the workflow of an AI-assisted medical triage method according to an embodiment of the present invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0022] Example 1: This example provides an AI-assisted medical triage system, such as... Figure 2 As shown, it includes: The information acquisition module is used to acquire the patient's basic information, medical history data, symptom description, and medical record information, and at the same time record the time series information corresponding to each symptom description and each medical record information to generate the first chief complaint information; The AI ​​extraction module is used to extract features from the first chief complaint information using a pre-trained AI feature extraction model to obtain a first set of feature vectors. The information analysis module is used to detect and remove anomalies in the first feature vector set to obtain a second feature vector set, and to match the second feature vector set with a preset department feature vector database to determine the target department. The matching analysis module is used to generate targeted consultation questions based on the department with the highest matching degree when the matching fails. It combines the patient's feedback answer with the second feature vector set to generate second chief complaint information, and repeats the analysis until the target department is determined. The department recommendation module is used to recommend the target department to the patient, and generate and display the recommendation reasons based on the department information corresponding to the target department for the day.

[0023] In this example, basic information refers to the patient's basic identity and health-related information, such as age, gender, and allergy history.

[0024] In this example, medical history data represents information related to the patient's past illnesses and treatments, such as previously diagnosed diseases, surgical history, and long-term medication history.

[0025] In this example, the symptom description refers to the patient's active description of the various physical discomfort symptoms they currently experience.

[0026] In this example, medical record information refers to medical records such as electronic medical record summaries generated during the patient's previous medical visits.

[0027] In this example, the time series information represents the time when the corresponding symptom description and medical record information occurred, or the chronological order of the patient's description of relevant content, which is used to reflect the logical relationship between the symptoms and the information.

[0028] In this example, the first chief complaint information represents the complete initial chief complaint data of the patient, which is formed by integrating the collected basic information, medical history data, symptom description, medical record information and corresponding time series information.

[0029] In this example, the pre-trained AI feature extraction model represents a BERT-like or Sentence-BERT-like natural language processing model adjusted based on Chinese medical corpus, used to convert unstructured complaint text into structured feature vectors.

[0030] In this example, the first feature vector set represents the aggregated set of all feature vectors carrying corresponding time series information obtained after the pre-trained AI feature extraction model extracts features from the first chief complaint information.

[0031] In this example, anomaly detection and removal refers to a feature cleaning operation that identifies and removes anomalous feature vectors from the first feature vector sequence based on neighborhood similarity evaluation and slope difference cumulative labeling.

[0032] In this example, the second feature vector set represents the set of all remaining valid feature vectors after the anomaly detection and elimination of the first feature vector set, which is used for subsequent department matching calculations.

[0033] In this example, the preset department feature vector database refers to a pre-built database containing feature vector datasets from multiple different departments. The feature vector datasets for each department are extracted based on data such as clinical medical records and medical guidelines of the corresponding department.

[0034] In this example, the targeted consultation questions refer to the consultation questions selected based on the representative features of the department with the highest current matching degree when the department fails to match. These questions are used to supplement the patient's medical information. The number of questions is 5-10, with a default of 8 questions.

[0035] In this example, the patient feedback answers represent the patient's responses to the targeted consultation questions, which are used to supplement and improve the chief complaint information.

[0036] In this example, the second chief complaint information refers to the new chief complaint data containing supplementary information generated after integrating the patient's feedback answers with the information corresponding to the second feature vector set that has undergone abnormal cleaning, which is used for iterative analysis.

[0037] In this example, the department information for the day refers to the real-time departmental diagnosis and treatment information for the target department, such as the attending physician, the time of day, and the availability of appointments.

[0038] In this example, the recommendation reason refers to the explanation content generated based on the symptom matching criteria of the target department and the department information on that day, which is used to explain the reason for recommending the department.

[0039] In this example, the construction of the department feature vector database is shown below: 1. Data Sources: Collect anonymized clinical medical record texts with clear departmental labels by department, and incorporate typical symptom descriptions from medical guidelines and papers; record the department, text fragment, keywords, source, and version date.

[0040] 2. Preprocessing: Remove identity information, standardize character, time and numerical representations, and perform medical synonym normalization, duplicate text removal and error sample cleanup; similar records of the same patient should not be included in the training set and validation set at the same time.

[0041] 3. Model and Dimensions: The example uses a BERT or Sentence-BERT-like encoding model finely tuned to Chinese medical corpus, and uses sentence vector pooling to obtain a 768-dimensional vector; the optional dimension is 384-1024, but the same model, dimension and pooling method must be used on the online side and the database side.

[0042] 4. Standardization and Data Storage: The data is centered based on the training set statistics, followed by L2 normalization. Each record stores the vector_id, department_id, vector, source text, standard keywords, data source, and model_version, and is grouped by department_id.

[0043] 5. Quality control: Cluster duplicate vectors with a cosine similarity higher than 0.95 within the same department to remove duplicates; samples with low similarity are subject to manual review instead of being directly deleted; training, validation and test data are divided according to the patient or source subject.

[0044] 6. Versioning and Calibration: After the model or data is updated, model_version and database_version are generated, and r, θΔ, m0, ηdept, ηlocal, K and Qmax are recalibrated; the production environment ensures that the online model and database versions are consistent.

[0045] The values ​​and settings of each parameter are shown in Table 1: Table 1 The working principle and beneficial effects of the above technical solution are as follows: First, the information acquisition module collects the patient's basic information, medical history data, symptom descriptions, and medical record information, simultaneously recording the time series information corresponding to each type of information and generating the first chief complaint information. Then, the AI ​​extraction module calls a pre-trained AI feature extraction model to extract features from the first chief complaint information, obtaining a first feature vector set carrying time-series correlation attributes. Subsequently, the information analysis module performs anomaly monitoring and removal on the first feature vector set, filtering out abnormal features lacking clinical rationality such as patient slips of the tongue, memory bias, and vague statements, to obtain a second feature vector set. The second feature vector set is then matched with a preset department feature vector database to determine the target department, starting from the input... This approach reduces noise interference, effectively improving the stability and accuracy of department matching. When a match fails, the matching analysis module generates targeted consultation questions around the department with the highest matching degree. It integrates the patient's feedback answers with the cleaned second feature vector set to generate the second chief complaint information, returning to the feature extraction and analysis stage for repeated iterations until the target department is determined. This avoids indiscriminate general questions, reduces irrelevant consultation content, and significantly improves the relevance of supplementary information and consultation efficiency. Finally, the department recommendation module recommends the target department to the patient and generates recommendation reasons based on the target department's information for the day. This reduces patients' blind registration and repeated department transfers, alleviates the service pressure on manual triage, and promotes the rational allocation of medical resources.

[0046] Example 2: Based on Example 1, the information acquisition module of the AI-assisted medical triage system includes: The multi-source data acquisition unit is used to collect the patient's basic information, medical history data, symptom description, and medical record information respectively. The time sequence information annotation unit is used to sort the occurrence time corresponding to each symptom description and each medical record information for the first time, and then sort the order of the descriptions corresponding to each symptom description and each medical record information for the second time, so as to generate corresponding time sequence information. The information association and integration unit is used to associate and bind basic information, medical history data, symptom descriptions, medical record information and corresponding time series information to obtain a structured information combination; The chief complaint information generation unit is used to determine the first chief complaint information by combining the associated structured information and output it.

[0047] In this example, the first sorting refers to the chronological order of each piece of information based on the actual occurrence time corresponding to the symptom description and medical record information, in order to establish the objective temporal logic of the information.

[0048] In this example, the second sorting refers to the sorting of symptom descriptions and medical record information without a clear time of occurrence, based on the order in which the patient states these information, in order to supplement and improve the temporal logic of content without clear time markers.

[0049] In this example, association binding means establishing a one-to-one mapping relationship between basic information, medical history data, symptom description, and medical record information and their corresponding time series information, so that each piece of information carries a clear time series attribute.

[0050] In this example, the structured information combination refers to a standardized data combination that is organized and well-organized, with various types of information corresponding one-to-one with time-series information after being associated and bound. It is the intermediate data form for generating the first chief complaint information.

[0051] The working principle and beneficial effects of the above technical solution are as follows: First, the multi-source data acquisition unit collects the patient's basic information, medical history data, symptom descriptions, and medical record information. Then, the time-series information annotation unit first sorts each symptom description against the actual occurrence time of each medical record, and then sorts the content without clear time markers against the corresponding order of descriptions, generating time-series information covering all information. Subsequently, the information association and integration unit associates and binds the four types of information with the corresponding time-series information to obtain a structured information combination. Finally, the chief complaint information generation unit determines the structured information combination as the first chief complaint information and outputs it. By using a two-stage hierarchical sorting method, both objective onset time and subjective description order are taken into account, solving the problems of fragmented and disordered patient chief complaint time, ensuring the integrity and logical coherence of the time-series information. The structured information combination after association and binding eliminates the problem of scattered and disordered multi-source data, providing standardized and clear time-series input for subsequent feature extraction and time-series anomaly detection, effectively avoiding interference from disordered symptom time sequence with the accuracy of subsequent department matching, and ensuring the stable operation of subsequent anomaly identification and department matching processes.

[0052] Example 3: Based on Example 1, the AI-assisted medical triage system includes an AI extraction module comprising: The model calling unit is used to call a pre-trained BERT-like natural language processing model or a Sentence-BERT-like natural language processing model adjusted based on Chinese medical corpus, and outputs feature vectors with dimensions of 384 to 1024. The vector generation unit is used to segment the first chief complaint information into paragraphs according to symptom features and information features, and input the obtained paragraphs into the natural language processing model to obtain the initial feature vector corresponding to each paragraph. The time binding unit is used to bind the time sequence information of the symptom description or medical record information to the corresponding paragraph in the first chief complaint information to each of the initial feature vectors; The normalization unit is used to perform L2 normalization on the initial feature vector of each bound time series to obtain the first feature vector set and transmit it to the information analysis module.

[0053] In this example, the anomaly detection process is as follows: Let the first feature vector sequence arranged in chronological order be V = {v1, v2, ..., vN}. First, L2 normalization is performed: v'i = vi / max(||vi||2, ε), where ε is a small positive number to prevent division by zero. The cosine similarity is sim(vi, vj) = v'i·v'j, and the cosine distance is d(vi, vj) = 1 - sim(vi, vj).

[0054] For the i-th data point, construct a neighborhood Ni = {j | j ≠ i and d(vi, vj) ≤ r} with a preset distance r; if Ni is empty, select the nearest data point. Calculate μi = (1 / |Ni|)Σj∈Nisim(vi, vj), let μmax = max(μ1, ..., μN), and evaluate the value Ei = μi / max(μmax, ε). The lower Ei is, the weaker the consistency between the feature and the local data.

[0055] The timestamp is normalized to xi = (ti - t1) / max(tN - t1, ε). The slope of adjacent points is ki = (Ei+1 - Ei) / max(xi+1 - xi, ε), and the difference between adjacent slopes is Δki = |ki - ki-1|. The preset difference threshold is θΔ = median(Δk) + λ·MAD(Δk), where λ is between 2.5 and 3.5, and 3.0 is recommended.

[0056] When Δki ≥ θΔ, data points within adjacent windows that form this slope difference are marked as anomalies; the number of markings is accumulated when the same data point is covered by adjacent windows. The preset marking threshold m0 is set to 1-2, with 2 recommended; a data point is identified as an anomaly when the accumulated number of markings reaches m0.

[0057] In this example, the Chinese medical corpus represents a collection of medical texts, including Chinese clinical medical records, medical guidelines, medical papers, and descriptions of typical symptoms, used for model fine-tuning to improve the model's semantic understanding of medical scenario texts.

[0058] In this example, the BERT-like natural language processing model represents a pre-trained language model based on the Transformer architecture, which has deep semantic understanding capabilities and can convert text content into corresponding semantic feature vectors.

[0059] In this example, the Sentence-BERT-like natural language processing model represents a sentence-level vector generation model optimized based on the BERT model. It can efficiently generate semantic feature vectors at the sentence level and is suitable for text similarity calculation scenarios.

[0060] In this example, the output dimension represents the number of feature dimensions contained in the feature vector generated by the natural language processing model, with a value ranging from 384 to 1024. The dimension value determines the richness of semantic information carried by the feature vector.

[0061] In this example, paragraph segmentation refers to the processing operation of splitting the complete first chief complaint information into multiple independent text fragments according to the granular boundaries of symptom features and information features.

[0062] In this example, a paragraph represents an independent text segment obtained after paragraph segmentation, corresponding to a single symptom description or a single piece of information, and is the basic input unit for feature extraction.

[0063] In this example, the initial feature vector represents the feature vector directly output by the natural language processing model after encoding the input paragraph, before it has undergone temporal binding and normalization.

[0064] In this example, L2 normalization means performing a normalization operation on the feature vectors, unifying the L2 norm of the vectors to 1, which is used to eliminate the influence of vector magnitude differences on subsequent similarity calculation results.

[0065] The working principle and beneficial effects of the above technical solution are as follows: First, the model calling unit calls a BERT-type or Sentence-BERT-type natural language processing model finely tuned to Chinese medical corpus, outputting feature vectors with dimensions ranging from 384 to 1024. Then, the vector generation unit segments the first chief complaint information into paragraphs according to the granularity of symptom features and information features. The segmented independent paragraphs are then sequentially input into the natural language processing model to obtain the initial feature vector corresponding to each paragraph. Subsequently, the time binding unit binds the time series information of the symptom description or medical record information to which the corresponding paragraph belongs to each initial feature vector. Finally, the normalization unit normalizes all bound time... The initial feature vectors of the sequence undergo L2 normalization to generate the first set of feature vectors, which are then transmitted to the information analysis module. A language model finely tuned for the medical field is used to accurately capture the semantic features of medical texts, improving the accuracy of feature extraction of symptoms and medical information. Segmentation based on feature granularity ensures that each feature vector corresponds to an independent symptom or information unit. Temporal binding provides the necessary data foundation for subsequent temporal anomaly detection. L2 normalization eliminates the interference of vector magnitude differences on subsequent similarity calculations, providing standardized and high-quality feature inputs for subsequent anomaly removal and department matching, ensuring the stability and reliability of subsequent calculations.

[0066] Example 4: Based on Example 1, the information analysis module of the AI-assisted medical triage system includes: A time sequence sorting unit, configured to sort the feature vectors in the first feature vector set according to the time sequence information bound to each of the first feature vectors to obtain a first feature vector sequence; An abnormal feature culling unit, configured to evaluate each data point in the first feature vector sequence respectively, generate an evaluation value sequence, perform abnormality marking based on the adjacent slope difference of the evaluation value sequence, determine a data point with a cumulative number of abnormality markings reaching a preset marking threshold as an abnormal data point, and cull the first feature vector corresponding to the abnormal data point to obtain a second feature vector set; A department matching calculation unit, configured to perform similarity calculation on the second feature vector set and the feature vector dataset corresponding to each department in a preset department feature vector database respectively, so as to obtain a department matching evaluation value corresponding to each of the departments; A target department determination unit, configured to extract the maximum value among all the department matching evaluation values as the maximum department matching evaluation value, compare the maximum department matching evaluation value with a preset department matching threshold, and regard the department as a target department when the maximum department matching evaluation value is greater than or equal to the preset department matching threshold; If it is smaller, the maximum department matching evaluation value and the corresponding department information are sent to the matching analysis module.

[0067] In this example, the department matching calculation logic is: let the second feature vector set V2={z1, ..., zN2}, and the dataset of department c be Dc={uc1, ..., ucMc}. All vectors are processed by the same model and subjected to L2 normalization, and the cosine similarity is mapped to [0, 1]: s01(a, b)=(sim(a, b)+1) / 2.

[0068] Calculate s01 for each zi and all vectors in Dc, take the average of the top K similarities to obtain sic=(1 / K)ΣTop-Ks01(zi, ucj). K is 3-10, defaulting to 5; when Mc<K, K is set to Mc. This method is used to reduce the impact of the scale difference of datasets of different departments.

[0069] The department matching evaluation value Pc=Σi=1…N2sic; Pmax=maxcPc, and the candidate department c*=argmaxcPc. The preset department matching threshold θdept=N2·ηdept, where ηdept is 0.65-0.80, defaulting to 0.70. When Pmax≥θdept, c* is determined as the target department, otherwise it is transmitted to the matching analysis module.

[0070] ηdept is calibrated using a validation set with real department labels: select a value with a high recall rate on the premise that the upper limit of the target misclassification rate is met. Recalibration shall be performed after the version of the model or the department database changes.

[0071] In this example, the data point represents the coordinate point corresponding to each independent first feature vector in the first feature vector sequence in the coordinate system formed by the time sequence and the evaluation value, and is the basic calculation unit for anomaly detection.

[0072] In this example, the evaluation value represents the ratio of the mean similarity in the neighborhood centered on each data point to the mean of the global maximum similarity. It is used to quantify the degree of consistency between the corresponding feature vector and the local surrounding features. The lower the value, the weaker the fit between the feature and the local data.

[0073] In this example, the evaluation value sequence represents a numerical sequence formed by arranging the evaluation values ​​corresponding to all data points in the first feature vector sequence in chronological order, which is used to carry out slope difference anomaly detection.

[0074] In this example, the adjacent slope difference represents the absolute value of the difference between the slopes of the lines connecting two adjacent data points in a coordinate system with time as the horizontal axis and evaluation value as the vertical axis. It is used to reflect the fluctuation range of the evaluation value over time.

[0075] In this example, the anomaly marker indicates that when the difference between adjacent slopes exceeds a preset difference threshold, an anomaly identification operation is performed on the data points within the corresponding window range. The same data point can be covered by multiple adjacent windows and the number of markings is accumulated.

[0076] In this example, the preset marking threshold represents the critical value of the cumulative number of times a data point is marked to determine whether it is an abnormal data point. The value range is 1 to 2, and the default value is 2. When the cumulative number of abnormal markings of a data point reaches this threshold, it is determined to be an abnormal data point.

[0077] In this example, abnormal data points refer to data points whose cumulative number of abnormal markings reaches a preset marking threshold and whose corresponding feature vectors lack local consistency. Their corresponding feature vectors belong to abnormal features that need to be removed.

[0078] In this example, the first feature vector represents a semantic feature vector output by the AI ​​extraction module, which is bound to the corresponding time series information and has undergone normalization processing. It is the basic data unit for department matching and anomaly detection.

[0079] In this example, the feature vector dataset represents the set of all feature vectors corresponding to a single department within the preset department feature vector database, which is constructed based on medical data such as clinical medical records and medical guidelines of the corresponding department.

[0080] In this example, similarity calculation refers to the process of calculating the semantic similarity between two feature vectors based on the cosine similarity algorithm, which is used to measure the degree of matching between patient symptom features and department features.

[0081] In this example, the department matching evaluation value represents the total matching score obtained after calculating the similarity between the second feature vector set and the feature vector dataset of a single department. It is used to quantify the degree of fit between the department and the patient's complaint.

[0082] In this example, the preset department matching threshold is the critical matching score used to determine whether the target department can be directly identified. It is obtained by multiplying the number of vectors in the second feature vector set by a proportional coefficient. The proportional coefficient ranges from 0.65 to 0.80, with a default value of 0.70. When the maximum evaluation value of department matching reaches this threshold, the corresponding department can be identified as the target department.

[0083] The working principle and beneficial effects of the above technical solution are as follows: First, the temporal sorting unit sorts the feature vectors in the first feature vector set according to the time series information bound to each first feature vector, obtaining a first feature vector sequence with sequential logic. This converts the scattered and disordered features into structured temporal data, providing standardized input for subsequent anomaly detection. Then, the anomaly feature removal unit calculates an evaluation value for each data point in the first feature vector sequence. Based on the slope difference between adjacent evaluation value sequences, anomaly marking is performed. Data points whose cumulative anomaly marking count reaches a preset marking threshold are identified as anomaly data points and their corresponding first feature vectors are removed, resulting in a cleaned second feature vector set. This effectively filters out abnormal features lacking clinical rationality, such as patient slips of the tongue, memory biases, and vague expressions. To reduce the interference of noise data on department matching at the source and improve the stability of matching results, the department matching calculation unit then performs similarity calculations on the second feature vector set and the feature vector dataset corresponding to each department in the preset department feature vector database to obtain the department matching evaluation value for each department, quantifying the degree of matching between the feature set and each department. Finally, the target department determination unit extracts the maximum value among all department matching evaluation values ​​as the maximum department matching evaluation value, compares it with the preset department matching threshold, and directly determines the target department when the threshold condition is met; otherwise, the corresponding data is sent to the matching analysis module to trigger a targeted secondary consultation. By using a single threshold, result determination and process branch control are realized simultaneously, simplifying the control logic while ensuring the reliability of department classification judgment.

[0084] Example 5: Based on Example 1, the AI-assisted medical triage system, wherein the matching analysis module includes: The candidate department locking unit is used to extract the department with the highest evaluation value of department matching as a candidate department when the matching fails, and to retrieve the feature vector dataset corresponding to the candidate department. The target feature filtering unit is used to filter feature vectors whose local consultation evaluation values ​​meet the preset local consultation evaluation threshold from the feature vector dataset of the candidate departments as target feature vectors, and extract the target keywords corresponding to each target feature vector; The consultation question sorting unit is used to retrieve a pre-defined departmental consultation question dataset based on target keywords, calculate the evaluation value of each initial consultation question, filter out targeted consultation questions by using the average evaluation value of all initial consultation questions as a threshold, and determine the consultation order according to the evaluation value from high to low. The iterative analysis trigger unit is used to obtain the patient's feedback answers to each of the aforementioned targeted consultation questions, integrate the feedback answers with the second feature vector set to generate the second chief complaint information, and return to the feature extraction and analysis process to repeat until the target department is determined.

[0085] In this example, the candidate department indicates the department with the highest matching score when department matching fails. This department is the one with the highest matching degree with the patient's chief complaint and serves as the anchor for targeted secondary consultation.

[0086] In this example, the local consultation evaluation value represents the sum of similarities between a single feature vector and all other feature vectors in the candidate department feature vector dataset. It is used to measure the representativeness of the feature vector in the department feature distribution, and the higher the value, the stronger the representativeness.

[0087] In this example, the preset local consultation evaluation threshold represents the critical evaluation value used to screen target feature vectors. It is obtained by subtracting one from the total number of vectors in the candidate department feature vector dataset and then multiplying it by a scaling factor. The scaling factor ranges from 0.60 to 0.80, with a default value of 0.70. When the local consultation evaluation value of a feature vector reaches this threshold, it can be selected as a target feature vector.

[0088] In this example, the target feature vector represents the feature vector selected from the feature vector dataset of the candidate department, whose local consultation evaluation value meets the threshold requirement. It is a highly representative feature belonging to the core area of ​​the feature distribution of the department and is used to extract consultation keywords.

[0089] In this example, the target keywords refer to standard medical keywords extracted from the text corresponding to the target feature vector that have departmental identification value, and are used to retrieve and match the corresponding consultation questions.

[0090] In this example, the preset departmental consultation question dataset represents a pre-constructed set of consultation questions categorized and stored by department, which stores the corresponding consultation questions for each department and their associated keywords.

[0091] In this example, the initial consultation questions represent all consultation questions in the preset department consultation question dataset that are related to the target keywords and are to be screened. It is a candidate set of targeted consultation questions.

[0092] In this example, the targeted consultation questions refer to the consultation questions selected from the initial consultation questions that meet the mean threshold requirement. These are targeted questions generated around the core identification points of the candidate departments and are used to supplement the patient's chief complaint information.

[0093] In this example, the process of selecting target feature vectors is as follows: For the candidate department dataset D*={u1,…,uM}, calculate Lj=Σl≠js01(uj,ul). The larger Lj is, the closer uj is to the dense region of the feature distribution of that department, and the stronger its representativeness.

[0094] The preset local diagnostic evaluation threshold is θlocal=(M-1)·ηlocal, where ηlocal is between 0.60 and 0.80, with a default value of 0.70. Vectors with Lj≥θlocal are determined as target feature vectors; if more than 20 vectors are selected, the top 20 can be retained in descending order of Lj.

[0095] The database contains information such as: vector-source text-keywords-department association. Keywords corresponding to the target feature vector are read, and medical synonyms are normalized and deduplicated to obtain the target keyword set K*.

[0096] In this example, the process of weighted filtering and sorting of consultation questions is as follows: The initial question set Q is retrieved using K*. For each keyword k, the number of questions containing that keyword, fk, is counted and used as the keyword label value; the normalized weight wk = fk / Σh∈K*fh is then applied.

[0097] For an initial problem q, let K(q) be the set of target keywords it contains, and let the problem evaluation value Score(q) = Σk∈K(q)wk.

[0098] The mean threshold θq = (1 / |Q|)Σq∈QScore(q). Questions with a score (q) ≥ θq are selected and sorted in descending order of score; in case of a tie, questions covering more different keywords are prioritized, and if scores are still tied, the database order is maintained.

[0099] After mean filtering, the top Qmax questions can be retained, with Qmax ranging from 5 to 10 and a default of 8. The scores are also jointly calibrated based on average completion time, user interruption rate, and department identification accuracy.

[0100] The working principle and beneficial effects of the above technical solution are as follows: When department matching fails, the system initiates the process. First, the candidate department locking unit extracts the department with the highest matching evaluation value as a candidate department and retrieves the corresponding feature vector dataset. This precisely anchors the scope of the secondary consultation to the department with the highest probability, avoiding indiscriminately covering all departments with invalid questions. Then, the target feature filtering unit calculates the local consultation evaluation value of each feature vector in the candidate department feature vector dataset, selecting feature vectors whose local consultation evaluation values ​​meet a preset local consultation evaluation threshold as target feature vectors. Simultaneously, it extracts the target keywords corresponding to each target feature vector. The selected features are highly representative of the core area of ​​department feature distribution, accurately pointing to key information for department identification. Subsequently, the consultation question... The sorting unit retrieves a pre-defined dataset of departmental consultation questions based on target keywords, calculates the evaluation value of each initial consultation question, and uses the average evaluation value of all initial consultation questions as a screening threshold to obtain targeted consultation questions. The consultation order is determined by ranking the evaluation values ​​from highest to lowest, ensuring that questions with the highest identification value are asked first, further reducing consultation rounds and improving information supplementation efficiency. Finally, the iterative analysis triggering unit obtains the patient's feedback answers to the targeted consultation questions, integrates the feedback answers with the second feature vector set to generate the second chief complaint information, and returns to the feature extraction and analysis process for repeated execution until the target department is determined. This forms a complete closed loop of anomaly cleaning, department matching, targeted consultation, and reanalysis, continuously optimizing consultation efficiency while ensuring the accuracy of departmental classification and reducing the patient's answering burden and waiting time.

[0101] Example 6: Based on Example 4, the AI-assisted medical triage system, wherein the abnormal feature removal unit includes: The neighborhood evaluation subunit is used to perform L2 normalization on each of the first feature vectors in the first feature vector sequence, construct a neighborhood with a preset distance as the radius, calculate the average similarity between each of the first feature vectors in the neighborhood and the neighborhood data points, take the maximum value of all the average similarity values ​​as the global maximum value, and take the ratio of the average similarity value corresponding to each of the first feature vectors to the global maximum value as the evaluation value of the corresponding first feature vector. The slope calculation subunit is used to construct a coordinate system by normalizing the time series of the first feature vector as the horizontal axis and the evaluation value as the vertical axis, calculate the slope of the line connecting adjacent data points and the difference between adjacent slopes, compare the difference with a preset difference threshold, and mark the data points in the corresponding window as abnormal and accumulate the number of markings when the difference is greater than or equal to the preset difference threshold. The labeling and removal subunit is used to define data points whose cumulative number of abnormal labelings is greater than or equal to a preset labeling threshold as abnormal data points and remove the first feature vector corresponding to the abnormal data points. The cleaning output subunit is used to form a second feature vector set from the remaining first feature vectors after removal, and output it to the department matching subunit.

[0102] In this example, the preset distance represents the cosine distance threshold used to define the neighborhood range of each data point. It is a pre-calibrated algorithm parameter used to determine the range of surrounding data points participating in the local similarity calculation. It can be calibrated and adjusted according to the actual distribution of the department feature dataset.

[0103] In this example, the neighborhood data point represents the remaining data points within the neighborhood defined by the preset distance of the current data point, and is the reference object used to calculate the mean local similarity of the current data point.

[0104] In this example, the mean similarity represents the arithmetic mean of the cosine similarity between a single data point and all other data points in its neighborhood, used to quantify the degree of consistency between the feature vector corresponding to the data point and the surrounding local features.

[0105] In this example, the preset difference threshold is adaptively set based on the robust distribution of the evaluation value sequence, and its value range is: edian + λ·MAD; λ = 2.5-3.5, with a default value of 3.0.

[0106] In this example, the global maximum value represents the maximum value among the average similarity values ​​of all data points in the first feature vector sequence. It serves as a unified reference benchmark for normalizing the average similarity values ​​of each data point.

[0107] In this example, the ratio of the mean similarity to the global maximum value represents the value obtained by dividing the mean similarity of a single data point by the global maximum value, which is the evaluation value of that data point. The value ranges from 0 to 1 and is used to normalize and measure the local consistency level of each feature vector. The lower the value, the weaker the fit between the feature and the surrounding local features.

[0108] The working principle and beneficial effects of the above technical solution are as follows: After receiving the first feature vector sequence, the abnormal feature removal unit first performs L2 normalization on each first feature vector in the sequence. A neighborhood is constructed for each data point with a preset distance as the radius. The mean similarity between each first feature vector and all neighborhood data points is calculated. Then, the maximum value among all mean similarity values ​​is extracted as the global maximum value. The ratio of the mean similarity value corresponding to each first feature vector to the global maximum value is used as the evaluation value of that data point, quantifying the credibility of each symptom feature from the perspective of local feature consistency. Subsequently, the slope calculation unit normalizes the time series corresponding to the first feature vector and uses it as the horizontal axis, with the evaluation values ​​of each data point as the vertical axis to construct a rectangular coordinate system. The slope of the line connecting adjacent data points and the difference between adjacent slopes are calculated sequentially. The difference is compared with a preset difference threshold. When the difference value is greater than or equal to a preset threshold, the data points within the corresponding window are marked as abnormal, and the number of markings for each data point is accumulated to identify abrupt abnormal features from the perspective of temporal fluctuation. Then, the marking and removal subunit determines the data points whose accumulated abnormal marking count is greater than or equal to the preset marking threshold as abnormal data points and removes the corresponding first feature vector. Through the mechanism of combining local consistency detection and temporal fluctuation detection and cumulative marking judgment, it can accurately filter abnormal features that lack clinical rationality, such as patients' slips of the tongue, memory bias, and vague expressions, and reduce the probability of misjudgment caused by a single detection rule, avoiding the accidental deletion of effective symptom features. Finally, the cleaning and output subunit forms a second feature vector set from the first feature vectors remaining after removing abnormalities and outputs it to the department matching calculation unit, providing low-noise and high-reliability feature input for the subsequent department matching process, effectively improving the stability of department matching results and the accuracy of department classification.

[0109] Example 7: Based on Example 1, the AI-assisted medical triage system includes a department recommendation module comprising: The department information retrieval unit is used to obtain basic information about the target department and the department's outpatient schedule for the day. The recommendation reason generation unit is used to generate corresponding recommendation reasons based on the matching criteria of the target department and the department information of the day; The registration guide generation unit is used to generate corresponding registration guide information based on the registration status of the target department on that day. The recommendation results display unit is used to output the target department, the reason for recommendation, and registration guidance information to the patient's end for display.

[0110] In this example, the basic introductory information represents the target department's scope of diagnosis and treatment, specialty characteristics, and discipline positioning, which are routine departmental information used to help patients understand the department's basic diagnosis and treatment direction.

[0111] In this example, the daily outpatient information refers to the real-time outpatient information of the target department, including the list of outpatient doctors, outpatient time, and outpatient room.

[0112] In this example, the registration guidance information represents the registration-related guidance content generated based on the registration status of the target department on that day, including information such as the remaining number of appointments, registration channels, registration deadline, and precautions for visiting the hospital.

[0113] In this example, the patient end refers to the terminal carrier used to display the triage recommendation results to the patient, including various forms such as online applications, web pages, and offline self-service terminals.

[0114] The working principle and beneficial effects of the above technical solution are as follows: After the target department is determined, the system is activated. First, the department information retrieval unit obtains basic information about the target department and the department's outpatient schedule for the day, providing complete data support for the generation of subsequent recommendation content. Then, the recommendation reason generation unit combines the matching criteria of the target department with the department information for the day to generate corresponding recommendation reasons, allowing patients to clearly understand the core basis for the department recommendation and improving the credibility and patient acceptance of the recommendation results. Subsequently, the registration guidance generation unit generates corresponding registration guidance information based on the registration status of the target department for the day, clearly informing patients of the availability of appointments, registration channels, and related precautions, reducing patient confusion during the registration process. Finally, the recommendation result display unit integrates the target department, recommendation reasons, and registration guidance information and outputs them to the patient's end for display, realizing a closed-loop information process from department recommendation to registration guidance. This effectively reduces patients' blind registration and repeated inquiries to manual guidance, reduces the service pressure on manual guidance, and improves the efficiency of patients' information acquisition before medical treatment and their overall medical experience.

[0115] Example 8: This example provides an AI-assisted medical triage method, such as... Figure 1 As shown, it includes: Step 1: Obtain the patient's basic information, medical history data, symptom description, and medical record information, and record the time series information corresponding to each symptom description and each medical record information to generate the first chief complaint information; Step 2: Extract features from the first chief complaint information using a pre-trained AI feature extraction model to obtain a first set of feature vectors; Step 3: Perform anomaly detection and removal on the first feature vector set to obtain the second feature vector set, and match the second feature vector set with the preset department feature vector database to determine the target department; Step 4: When a match fails, generate targeted consultation questions based on the department with the highest current matching degree, combine the patient's feedback answers with the second feature vector set to generate second chief complaint information, and repeat the analysis until the target department is determined; Step 5: Recommend the target department to the patient, and generate and display the recommendation reason based on the department information corresponding to the target department for the day.

[0116] In this example, basic information refers to the patient's basic identity and health-related information, such as age, gender, and allergy history.

[0117] In this example, medical history data represents information related to the patient's past illnesses and treatments, such as previously diagnosed diseases, surgical history, and long-term medication history.

[0118] In this example, the symptom description refers to the patient's active description of the various physical discomfort symptoms they currently experience.

[0119] In this example, medical record information refers to medical records such as electronic medical record summaries generated during the patient's previous medical visits.

[0120] In this example, the time series information represents the time when the corresponding symptom description and medical record information occurred, or the chronological order of the patient's description of relevant content, which is used to reflect the logical relationship between the symptoms and the information.

[0121] In this example, the first chief complaint information represents the complete initial chief complaint data of the patient, which is formed by integrating the collected basic information, medical history data, symptom description, medical record information and corresponding time series information.

[0122] In this example, the pre-trained AI feature extraction model represents a BERT-like or Sentence-BERT-like natural language processing model adjusted based on Chinese medical corpus, used to convert unstructured complaint text into structured feature vectors.

[0123] In this example, the first feature vector set represents the aggregated set of all feature vectors carrying corresponding time series information obtained after the pre-trained AI feature extraction model extracts features from the first chief complaint information.

[0124] In this example, anomaly detection and removal refers to a feature cleaning operation that identifies and removes anomalous feature vectors from the first feature vector sequence based on neighborhood similarity evaluation and slope difference cumulative labeling.

[0125] In this example, the second feature vector set represents the set of all remaining valid feature vectors after the anomaly detection and elimination of the first feature vector set, which is used for subsequent department matching calculations.

[0126] In this example, the preset department feature vector database refers to a pre-built database containing feature vector datasets from multiple different departments. The feature vector datasets for each department are extracted based on data such as clinical medical records and medical guidelines of the corresponding department.

[0127] In this example, the targeted consultation question refers to the consultation question selected based on the representative features of the department with the highest current matching degree when the department fails to match. This question is used to supplement the patient's medical information.

[0128] In this example, the patient feedback answers represent the patient's responses to the targeted consultation questions, which are used to supplement and improve the chief complaint information.

[0129] In this example, the second chief complaint information refers to the new chief complaint data containing supplementary information generated after integrating the patient's feedback answers with the information corresponding to the second feature vector set that has undergone abnormal cleaning, which is used for iterative analysis.

[0130] In this example, the department information for the day refers to the real-time departmental diagnosis and treatment information for the target department, such as the attending physician, the time of day, and the availability of appointments.

[0131] In this example, the recommendation reason refers to the explanation content generated based on the symptom matching criteria of the target department and the department information on that day, which is used to explain the reason for recommending the department.

[0132] The working principle and beneficial effects of the above technical solution are as follows: First, the information acquisition module collects the patient's basic information, medical history data, symptom descriptions, and medical record information, simultaneously recording the time series information corresponding to each type of information and generating the first chief complaint information. Then, the AI ​​extraction module calls a pre-trained AI feature extraction model to extract features from the first chief complaint information, obtaining a first feature vector set carrying time-series correlation attributes. Subsequently, the information analysis module performs anomaly monitoring and removal on the first feature vector set, filtering out abnormal features lacking clinical rationality such as patient slips of the tongue, memory bias, and vague statements, to obtain a second feature vector set. The second feature vector set is then matched with a preset department feature vector database to determine the target department, starting from the input... This approach reduces noise interference, effectively improving the stability and accuracy of department matching. When a match fails, the matching analysis module generates targeted consultation questions around the department with the highest matching degree. It integrates the patient's feedback answers with the cleaned second feature vector set to generate the second chief complaint information, returning to the feature extraction and analysis stage for repeated iterations until the target department is determined. This avoids indiscriminate general questions, reduces irrelevant consultation content, and significantly improves the relevance of supplementary information and consultation efficiency. Finally, the department recommendation module recommends the target department to the patient and generates recommendation reasons based on the target department's information for the day. This reduces patients' blind registration and repeated department transfers, alleviates the service pressure on manual triage, and promotes the rational allocation of medical resources.

[0133] Example 9: Based on Example 8, the AI-assisted medical triage method, step 4 includes: Step 41: When a match fails, extract the department with the highest departmental matching evaluation value as a candidate department, and retrieve the feature vector dataset corresponding to the candidate department. Step 42: Select feature vectors whose local consultation evaluation values ​​meet the preset local consultation evaluation threshold from the feature vector dataset of the candidate departments as target feature vectors, and extract the target keywords corresponding to each target feature vector; Step 43: Retrieve the preset departmental consultation question dataset based on the target keywords, calculate the evaluation value of each initial consultation question, use the average evaluation value of all initial consultation questions as the threshold to filter out targeted consultation questions, and determine the consultation order according to the evaluation value from high to low; Step 44: Obtain the patient's feedback answers to each of the aforementioned targeted consultation questions, integrate the feedback answers with the second feature vector set to generate the second chief complaint information, and return to the feature extraction and analysis process to repeat until the target department is determined.

[0134] In this example, the candidate department indicates the department with the highest matching score when department matching fails. This department is the one with the highest matching degree with the patient's chief complaint and serves as the anchor for targeted secondary consultation.

[0135] In this example, the local consultation evaluation value represents the sum of similarities between a single feature vector and all other feature vectors in the candidate department feature vector dataset. It is used to measure the representativeness of the feature vector in the department feature distribution, and the higher the value, the stronger the representativeness.

[0136] In this example, the preset local consultation evaluation threshold represents the critical evaluation value used to screen target feature vectors. It is obtained by subtracting one from the total number of vectors in the candidate department feature vector dataset and then multiplying it by a proportional coefficient. The proportional coefficient ranges from 0.60 to 0.80, with a recommended value of 0.70. When the local consultation evaluation value of a feature vector reaches this threshold, it can be selected as a target feature vector.

[0137] In this example, the target feature vector represents the feature vector selected from the feature vector dataset of the candidate department, whose local consultation evaluation value meets the threshold requirement. It is a highly representative feature belonging to the core area of ​​the feature distribution of the department and is used to extract consultation keywords.

[0138] In this example, the target keywords refer to standard medical keywords extracted from the text corresponding to the target feature vector that have departmental identification value, and are used to retrieve and match the corresponding consultation questions.

[0139] In this example, the preset departmental consultation question dataset represents a pre-constructed set of consultation questions categorized and stored by department, which stores the corresponding consultation questions for each department and their associated keywords.

[0140] In this example, the initial consultation questions represent all consultation questions in the preset department consultation question dataset that are related to the target keywords and are to be screened. It is a candidate set of targeted consultation questions.

[0141] In this example, the targeted consultation questions refer to the consultation questions selected from the initial consultation questions that meet the mean threshold requirement. These are targeted questions generated around the core identification points of the candidate departments and are used to supplement the patient's chief complaint information.

[0142] The working principle and beneficial effects of the above technical solution are as follows: When department matching fails, the system initiates the process. First, the candidate department locking unit extracts the department with the highest matching evaluation value as a candidate department and retrieves the corresponding feature vector dataset. This precisely anchors the scope of the secondary consultation to the department with the highest probability, avoiding indiscriminately covering all departments with invalid questions. Then, the target feature filtering unit calculates the local consultation evaluation value of each feature vector in the candidate department feature vector dataset, selecting feature vectors whose local consultation evaluation values ​​meet a preset local consultation evaluation threshold as target feature vectors. Simultaneously, it extracts the target keywords corresponding to each target feature vector. The selected features are highly representative of the core area of ​​department feature distribution, accurately pointing to key information for department identification. Subsequently, the consultation question... The sorting unit retrieves a pre-defined dataset of departmental consultation questions based on target keywords, calculates the evaluation value of each initial consultation question, and uses the average evaluation value of all initial consultation questions as a screening threshold to obtain targeted consultation questions. The consultation order is determined by ranking the evaluation values ​​from highest to lowest, ensuring that questions with the highest identification value are asked first, further reducing consultation rounds and improving information supplementation efficiency. Finally, the iterative analysis triggering unit obtains the patient's feedback answers to the targeted consultation questions, integrates the feedback answers with the second feature vector set to generate the second chief complaint information, and returns to the feature extraction and analysis process for repeated execution until the target department is determined. This forms a complete closed loop of anomaly cleaning, department matching, targeted consultation, and reanalysis, continuously optimizing consultation efficiency while ensuring the accuracy of departmental classification and reducing the patient's answering burden and waiting time.

[0143] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An AI-assisted medical triage system, characterized in that, include: The information acquisition module is used to acquire the patient's basic information, medical history data, symptom description, and medical record information, and at the same time record the time series information corresponding to each symptom description and each medical record information to generate the first chief complaint information; The AI ​​extraction module is used to extract features from the first chief complaint information using a pre-trained AI feature extraction model to obtain a first set of feature vectors. The information analysis module is used to detect and remove anomalies in the first feature vector set to obtain a second feature vector set, and to match the second feature vector set with a preset department feature vector database to determine the target department. The matching analysis module is used to generate targeted consultation questions based on the department with the highest matching degree when the matching fails. It combines the patient's feedback answer with the second feature vector set to generate second chief complaint information, and repeats the analysis until the target department is determined. The department recommendation module is used to recommend the target department to the patient, and generate and display the recommendation reasons based on the department information corresponding to the target department for the day.

2. The AI-assisted medical triage system as described in claim 1, characterized in that, The information acquisition module includes: The multi-source data acquisition unit is used to collect the patient's basic information, medical history data, symptom description, and medical record information respectively. The time sequence information annotation unit is used to sort the occurrence time corresponding to each symptom description and each medical record information for the first time, and then sort the order of the descriptions corresponding to each symptom description and each medical record information for the second time, so as to generate corresponding time sequence information. The information association and integration unit is used to associate and bind basic information, medical history data, symptom descriptions, medical record information and corresponding time series information to obtain a structured information combination; The chief complaint information generation unit is used to determine the first chief complaint information by combining the associated structured information and output it.

3. The AI-assisted medical triage system as described in claim 1, characterized in that, The AI ​​extraction module includes: The model calling unit is used to call the pre-trained BERT natural language processing model or Sentence-BERT natural language processing model based on Chinese medical corpus, and output feature vectors with dimensions of 384 to 1024. The vector generation unit is used to segment the first chief complaint information into paragraphs according to symptom features and information features, and input the obtained paragraphs into the natural language processing model to obtain the initial feature vector corresponding to each paragraph. The time binding unit is used to bind the time sequence information of the symptom description or medical record information to the corresponding paragraph in the first chief complaint information to each of the initial feature vectors; The normalization unit is used to perform L2 normalization on the initial feature vector of each bound time series to obtain the first feature vector set and transmit it to the information analysis module.

4. The AI-assisted medical triage system as described in claim 1, characterized in that, The information analysis module includes: The time-series sorting unit is used to sort the feature vectors in the first feature vector set according to the time-series information bound to each first feature vector to obtain a first feature vector sequence; An abnormal feature removal unit is used to evaluate each data point in the first feature vector sequence, generate an evaluation value sequence, mark anomalies based on the adjacent slope difference of the evaluation value sequence, define data points whose cumulative number of anomaly markings reaches a preset marking threshold as abnormal data points, remove the first feature vector corresponding to the abnormal data points, and obtain the second feature vector set. The department matching calculation unit is used to perform similarity calculations between the second feature vector set and the feature vector dataset corresponding to each department in the preset department feature vector database, and obtain the department matching evaluation value corresponding to each department. The target department determination unit is used to extract the maximum value among all department matching evaluation values ​​as the maximum department matching evaluation value. The maximum department matching evaluation value is compared with the preset department matching threshold. When the maximum department matching evaluation value is greater than or equal to the preset department matching threshold, the department is regarded as the target department. If the value is less than the maximum evaluation value for department matching and the corresponding department information, the matching analysis module will be sent to the matching analysis module.

5. The AI-assisted medical triage system as described in claim 1, characterized in that, The matching analysis module includes: The candidate department locking unit is used to extract the department with the highest evaluation value of department matching as a candidate department when the matching fails, and to retrieve the feature vector dataset corresponding to the candidate department. The target feature filtering unit is used to filter feature vectors whose local consultation evaluation values ​​meet the preset local consultation evaluation threshold from the feature vector dataset of the candidate departments as target feature vectors, and extract the target keywords corresponding to each target feature vector; The consultation question sorting unit is used to retrieve a pre-defined departmental consultation question dataset based on target keywords, calculate the evaluation value of each initial consultation question, filter out targeted consultation questions by using the average evaluation value of all initial consultation questions as a threshold, and determine the consultation order according to the evaluation value from high to low. The iterative analysis trigger unit is used to obtain the patient's feedback answers to each of the aforementioned targeted consultation questions, integrate the feedback answers with the second feature vector set to generate the second chief complaint information, and return to the feature extraction and analysis process to repeat until the target department is determined.

6. The AI-assisted medical triage system as described in claim 4, characterized in that, The abnormal feature removal unit includes: The neighborhood evaluation subunit is used to perform L2 normalization on each of the first feature vectors in the first feature vector sequence, construct a neighborhood with a preset distance as the radius, calculate the average similarity between each of the first feature vectors in the neighborhood and the neighborhood data points, take the maximum value of all the average similarity values ​​as the global maximum value, and take the ratio of the average similarity value corresponding to each of the first feature vectors to the global maximum value as the evaluation value of the corresponding first feature vector. The slope calculation subunit is used to construct a coordinate system by normalizing the time series of the first feature vector as the horizontal axis and the evaluation value as the vertical axis, calculate the slope of the line connecting adjacent data points and the difference between adjacent slopes, compare the difference with a preset difference threshold, and mark the data points in the corresponding window as abnormal and accumulate the number of markings when the difference is greater than or equal to the preset difference threshold. The labeling and removal subunit is used to define data points whose cumulative number of abnormal labelings is greater than or equal to a preset labeling threshold as abnormal data points and remove the first feature vector corresponding to the abnormal data points. The cleaning output subunit is used to form a second feature vector set from the remaining first feature vectors after removal, and output it to the department matching subunit.

7. The AI-assisted medical triage system as described in claim 1, characterized in that, The department recommendation module includes: The department information retrieval unit is used to obtain basic information about the target department and the department's outpatient schedule for the day. The recommendation reason generation unit is used to generate corresponding recommendation reasons based on the matching criteria of the target department and the department information of the day; The registration guide generation unit is used to generate corresponding registration guide information based on the registration status of the target department on that day. The recommendation results display unit is used to output the target department, the reason for recommendation, and registration guidance information to the patient's end for display.

8. A medical triage method based on AI assistance, characterized in that, include: Step 1: Obtain the patient's basic information, medical history data, symptom description, and medical record information, and record the time series information corresponding to each symptom description and each medical record information to generate the first chief complaint information; Step 2: Extract features from the first chief complaint information using a pre-trained AI feature extraction model to obtain a first set of feature vectors; Step 3: Perform anomaly detection and removal on the first feature vector set to obtain the second feature vector set, and match the second feature vector set with the preset department feature vector database to determine the target department; Step 4: When a match fails, generate targeted consultation questions based on the department with the highest current matching degree, combine the patient's feedback answers with the second feature vector set to generate second chief complaint information, and repeat the analysis until the target department is determined; Step 5: Recommend the target department to the patient, and generate and display the recommendation reason based on the department information corresponding to the target department for the day.

9. The AI-assisted medical triage method as described in claim 8, characterized in that, Step 4 includes: Step 41: When a match fails, extract the department with the highest departmental matching evaluation value as a candidate department, and retrieve the feature vector dataset corresponding to the candidate department. Step 42: Select feature vectors whose local consultation evaluation values ​​meet the preset local consultation evaluation threshold from the feature vector dataset of the candidate departments as target feature vectors, and extract the target keywords corresponding to each target feature vector; Step 43: Retrieve the preset departmental consultation question dataset based on the target keywords, calculate the evaluation value of each initial consultation question, use the average evaluation value of all initial consultation questions as the threshold to filter out targeted consultation questions, and determine the consultation order according to the evaluation value from high to low; Step 44: Obtain the patient's feedback answers to each of the aforementioned targeted consultation questions, integrate the feedback answers with the second feature vector set to generate the second chief complaint information, and return to the feature extraction and analysis process to repeat until the target department is determined.

Images