Artificial intelligence-based triage method and device

Abstract

The present disclosure provides an artificial intelligence-based triage method and device, relates to the technical field of artificial intelligence, in particular to the technical field of medical artificial intelligence, medical informatization, man-machine interaction control and computer system resource optimization, and can be applied to an Internet hospital intelligent triage system, an online triage and registration platform, an intelligent pre-screening module based on a large model, a primary medical auxiliary triage terminal and the like. A specific embodiment of the method comprises: performing intent hierarchical identification on a chief complaint text input by a user to determine a current inquiry intent level; selecting a current inquiry form corresponding to the current inquiry intent level to determine a current round question; performing inquiry by using the current round question, and determining whether to terminate the inquiry after each round of interaction; and outputting a structured triage result and a recommendation reason based on the inquiry intent level when the inquiry is terminated.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, particularly to the fields of medical artificial intelligence, medical informatization, human-computer interaction control and computer system resource optimization technology, and can be applied to products such as intelligent triage systems for internet hospitals, online triage and registration platforms, intelligent pre-screening modules for consultation based on large models, and auxiliary triage terminals for primary healthcare. Background Technology

[0002] Currently, internet hospitals and online triage platforms generally adopt intelligent triage solutions to provide patients with pre-symptom screening and department recommendation services, effectively diverting patients and alleviating the pressure of offline triage. It is a core supporting module of smart medical informatization, relying on natural language processing and medical knowledge base to realize human-computer consultation interaction, and is widely used in various scenarios such as mobile mini-programs and in-hospital self-service terminals.

[0003] Currently, the mainstream triage and patient guidance implementation paths fall into two categories. The first is the fixed-rule questionnaire approach: after the user enters their symptoms and complaints, the system matches a preset rule tree, asks questions sequentially according to a standardized questionnaire, summarizes the answers, matches the relevant department, and outputs registration guidance. This approach is stable and has low implementation costs. The second is the large-scale model-based free-consultation approach, which relies on a large medical language model to parse the user's natural complaints, dynamically generates follow-up questions based on the dialogue context, and outputs triage suggestions after multiple rounds of interaction to complete the collection of medical information. This approach's interactive expression is more in line with natural human communication habits. Summary of the Invention

[0004] This disclosure presents an artificial intelligence-based triage method and apparatus.

[0005] In a first aspect, embodiments of this disclosure propose an artificial intelligence-based triage method, comprising: performing intent-layer recognition on the user's input complaint text to determine the current consultation intent level; selecting the current consultation format corresponding to the current consultation intent level to determine the current round of questions; conducting consultation using the current round of questions, and determining whether to terminate the consultation after each round of interaction; and outputting structured triage results and recommendation reasons based on the consultation intent level at the time of termination of consultation.

[0006] Secondly, this disclosure proposes an artificial intelligence-based triage device, comprising: an identification module configured to perform intent-layer recognition on the user-inputted chief complaint text to determine the current consultation intent level; a determination module configured to select the current consultation format corresponding to the current consultation intent level to determine the current round of questions; a consultation module configured to conduct consultation using the current round of questions and determine whether to terminate the consultation after each round of interaction; and an output module configured to output structured triage results and recommendation reasons based on the consultation intent level at the time of termination of consultation.

[0007] Thirdly, embodiments of this disclosure provide an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method as described in the first aspect.

[0008] Fourthly, embodiments of this disclosure provide a non-transitory computer-readable storage medium storing computer instructions for causing a computer to perform the methods described in the first aspect.

[0009] Fifthly, embodiments of this disclosure provide a computer program product including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0010] The key or essential features of the embodiments disclosed herein are not intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0011] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of this disclosure. Wherein: Figure 1 This is a flowchart of one embodiment of the AI-based triage method according to the present disclosure; Figure 2 This is a flowchart of yet another embodiment of the AI-based triage method according to the present disclosure; Figure 3 This is a diagram showing the overall architecture, module division, and business execution process of the intelligent triage system; Figure 4 This is a schematic diagram of a structure of an embodiment of an AI-based triage device according to the present disclosure; Figure 5 This is a block diagram of an electronic device used to implement the AI-based triage method in the embodiments of this disclosure. Detailed Implementation

[0012] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0013] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0014] This disclosure aims to address four core shortcomings of existing online medical triage systems: lack of dynamic optimization of information value in the inquiry process; lack of user and system-level load constraints; lack of engineered mandatory circuit breakers for high-risk conditions; and a simplistic stop-the-pathway logic that leads to a dilemma between insufficient information and excessive follow-up questions. Existing triage solutions include fixed-rule questionnaires and large-scale, unconstrained follow-up questions, neither of which can simultaneously meet the four requirements of consultation efficiency, server resource utilization, emergency safety, and multi-terminal compatibility. This disclosure designs four coupled functional modules to construct a complete closed loop. The implementation process, interaction scenarios, comparative alternatives, and product deployment benefits are fully explained below with reference to the accompanying diagrams.

[0015] Figure 1 A flowchart 100 of an embodiment of an AI-based triage method according to the present disclosure is shown. The AI-based triage method includes the following steps: Step 101: Perform intent layer recognition on the user's input complaint text to determine the current consultation intent level.

[0016] In this embodiment, the entity executing the AI-based triage method can perform intent-layer recognition on the user's input complaint text to determine the current consultation intent level.

[0017] The aforementioned execution entity can be a medical AI (Artificial Intelligence) processing service on the server backend, which includes NLP (Natural Language Processing) parsing sub-service, departmental probability calculation sub-service, and user profile reading sub-service working together to complete the three tasks of text parsing, probability calculation, and scenario layering in one go, unlike the existing triage system which calls the interface multiple times in a serial manner, causing delays.

[0018] The chief complaint text refers to the natural language description of the patient's condition entered by the user through terminals such as mini-programs, in-hospital self-service machines, and apps. For example, phrases like "chest tightness for two hours" or "coughing and fever for three days" include subjective descriptions of symptoms, duration, and pain. Intent-based hierarchical recognition differs from traditional single-condition judgment; it's a recognition logic that integrates multiple dimensions—condition risk, departmental uncertainty, user device, user age, and interaction tolerance—to comprehensively determine the consultation target. The consultation intent hierarchy can include four types of consultation tasks: high-risk emergency, rapid departmental identification, general detailed follow-up questioning, and low-tolerance conservative approach, each matched with a specific consultation logic.

[0019] This step addresses the shortcomings of existing technologies that can only provide a unified consultation and cannot differentiate between patients and equipment scenarios, by implementing differentiated consultation management through tiered approaches.

[0020] Step 102: Select the current consultation format corresponding to the current consultation intent level, and determine the current round of questions.

[0021] In this embodiment, the aforementioned execution entity can select the current consultation format corresponding to the current consultation intent level and determine the current round of questions.

[0022] The consultation format can be bound to a complete set of question-and-answer control rules corresponding to the intent level, including whether follow-up questions are allowed, question length constraints, and adaptive reward weighting rules. The current round of questions pushes the optimal consultation questions to the user in a single round of interaction, which are selected based on comprehensive technical benefits and are not randomly or in a fixed order.

[0023] This step reads the intent level identifier output from the previous step, retrieves the preset consultation control rules, and then searches the medical knowledge graph to generate candidate questions. The optimal question for a single round is selected and output through a unified scoring formula, taking into account the needs of triage efficiency, medical safety, and user experience. This solves the pain points of redundant questions in traditional fixed questionnaires and unrestrained follow-up questions in large models.

[0024] Step 103: Conduct a consultation using the current round of questions, and determine whether to terminate the consultation after each round of interaction.

[0025] In this embodiment, the aforementioned execution entity can conduct a consultation using the current round of questions and determine whether to terminate the consultation after each round of interaction.

[0026] Each round of interaction involves the user receiving a question from the system and completing a full dialogue loop. Terminating the consultation is triggered by a stop signal from the state machine, at which point no new consultation questions are generated, and the system directly enters the department recommendation output path.

[0027] Each time a user submits an answer to the server, the system synchronously collects the character, round, and interface time data generated in this round of interaction, updates the global fatigue convergence factor in real time, and verifies four types of termination conditions in parallel. If any one of them is met, the follow-up questioning process is immediately cut off, realizing multi-dimensional joint control of interaction fatigue, system load, and disease risk to stop the consultation, overcoming the single defect of existing technologies that only rely on a fixed number of rounds to stop.

[0028] Step 104: Based on the consultation intent level at the time of termination of consultation, output structured triage results and recommendation reasons.

[0029] In this embodiment, the aforementioned executing entity can output structured triage results and recommendation reasons based on the consultation intent level at the time of termination of consultation.

[0030] Structured triage results are standardized outputs, including risk labels, a list of recommended departments, confidence scores for each department, and emergency / general care guidance. The reasons for the recommendations are presented in a readable text summary, explaining the basis for the triage decision (e.g., cough and high fever indicating respiratory medicine, chest pain triggering emergency care).

[0031] This step reads the entire consultation process data, departmental probability distribution, high-risk markers, and user's historical health records. It generates differentiated results according to the corresponding level output template. In high-risk scenarios, the emergency entrance is forcibly placed at the top. In scenarios where patients are fatigued and terminate early, the general practice recommendation weight is automatically increased to reduce the risk of misdiagnosis. This achieves tiered and differentiated safe output. High-risk patients are given priority to push to the emergency department, while low-tolerance patients are given a backup general practice, which makes up for the shortcomings of the existing system's homogeneous output and lack of a safety backup.

[0032] The AI-based triage method provided in this disclosure connects four modules in a complete and hierarchical manner to manage the entire consultation process, simultaneously addressing the four core shortcomings of existing fixed questionnaire and free large-scale model triage schemes: redundant and inefficient inquiries, uncontrollable interaction and server load, lack of mandatory circuit breaking for high-risk conditions, and the dilemma of a single stop strategy. It balances triage efficiency, medical safety, multi-terminal and multi-population adaptation, and system resource saving, and can be deployed in various smart healthcare products such as internet hospitals and primary care triage terminals.

[0033] Figure 2 A flow 200 of yet another embodiment of the AI-based triage method according to this disclosure is shown. This AI-based triage method includes the following steps: Step 201: Convert the unstructured medical expressions in the chief complaint text into structured medical feature vectors.

[0034] In this embodiment, the aforementioned execution entity can convert unstructured medical expressions in the chief complaint text into structured medical feature vectors.

[0035] Unstructured medical expressions are user-generated, colloquial descriptions of symptoms without standardized language, such as "I have chest pain, it started yesterday, and I can't breathe." Structured medical feature vectors are computer-computable digitized arrays. These arrays contain standardized fields such as symptoms, duration of illness, pain level, and high-risk indicators, serving as the input basis for subsequent departmental probability calculations.

[0036] This step transforms free speech into standardized data that the model can compute, which is a prerequisite for hierarchical recognition.

[0037] In some embodiments, structure transformation may include the following steps: First, the main complaint text is preprocessed to generate a preprocessed main complaint text.

[0038] The preprocessing may include, but is not limited to, at least one of the following: word segmentation, spelling normalization, synonymous symptom merging, noise filtering, etc. Word segmentation can split consecutive Chinese sentences into independent lexical units of symptoms, time and degree. Spelling normalization can correct wrong words, for example, "xiōng téng (wrongly written)" is unified as "xiōng tòng (chest pain)". Synonymous symptom merging can map terms such as "hard to breathe" and "dyspnea" to the same standard term. Noise filtering can delete meaningless modal particles and irrelevant small talk. For example, after preprocessing, the user input "Oh, my chest hurts so bad I can't catch my breath, ugh" is simplified to "chest pain, dyspnea".

[0039] Then, the medical language model is used to perform named entity recognition on the preprocessed chief complaint text, extract medical semantic entities, map the entities to a preset medical terminology set, and obtain corresponding standard symptom terms.

[0040] The medical language model is a special NLP model fine-tuned with massive medical records and national standard medical dictionaries. Named Entity Recognition (NER) specially extracts entities of symptoms, parts and duration; the medical terminology set stores industrially unified and standardized symptom vocabularies, eliminates oral ambiguities, and provides unified input for probability matrix calculation.

[0041] Finally, based on the standard symptom terms, structured medical feature vectors are constructed.

[0042] All extracted information including standardized symptoms, duration, severity, whether the condition is high-risk and the like is converted into a binary / floating-point array and stored in an in-memory vector. Vector fields are fixed and uniform, which ensures the standardization of the input format for subsequent probability matrices.

[0043] Step 202: identifying the current inquiry intention level based on the structured medical feature vector, risk, confidence and interaction constraints.

[0044] In this embodiment, the aforementioned execution body may identify the current inquiry intention level based on the structured medical feature vector, risk, confidence and interaction constraints.

[0045] The risk may be a trigger marker of high-risk core words, representing life-threatening symptoms such as myocardial infarction, stroke and acute abdomen. The confidence is the maximum probability value of each candidate department, with a numerical range from 0 to 1. A closer value to 1 indicates that the department indicated by the condition is clearer. The interaction constraints are user-side limitation conditions such as the user's age, terminal device bandwidth and historical question-and-answer tolerance.

[0046] In this step, department probability calculation and four-layer intention hierarchical determination are completed, and generating a high-risk circuit-breaking signal is the key safety innovation point of the present solution.

[0047] In some embodiments, the inquiry intention level identification may include the following steps: First, the structured medical feature vector is input into the preset symptom-department conditional probability matrix to generate the initial probability distribution vector of each candidate department, and the maximum probability value in the initial probability distribution vector is taken as the maximum confidence level.

[0048] The symptom-department conditional probability matrix is ​​a pre-trained two-dimensional statistical matrix that stores the incidence probability of each symptom in each department. The sum of the probability distribution vectors is fixed at 1. For example, after inputting the cold feature vector, the probability distribution vector P = [General Department 0.55, Respiratory Medicine 0.3, Otolaryngology 0.1, Others 0.05], with a maximum value of 0.55, which is the maximum confidence level.

[0049] Then, based on the triggering status of high-risk core words, maximum confidence, chief complaint complexity, account attributes, terminal bandwidth capacity, and historical interaction tolerance characteristics, intent is layered to determine the current consultation intent level.

[0050] The high-risk core keyword database pre-stores keywords for acute conditions such as chest pain, severe abdominal pain, and confusion. Chief complaint complexity refers to the number of symptoms. Account attribute distinguishes between elderly / adults / children. Bandwidth distinguishes between low-spec mobile phones / computer terminals. Historical tolerance is calculated by reading the average number of previous consultations, and stratification is completed based on these six dimensions.

[0051] In some embodiments, in response to identifying preset high-risk core words or high-risk chief complaint combinations, it is determined to be a high-risk emergency type; in response to determining that the maximum confidence level is greater than a preset confidence threshold, it is determined to be a rapid departmental identification type; in response to determining that the chief complaint information is insufficient, the candidate departments are scattered and there are conditions for continuing multiple rounds of interaction, it is determined to be a normal detailed follow-up question type; in response to determining that the account attribute is elderly, low-configuration low-bandwidth terminal, or historical interaction shows tolerance is lower than a preset tolerance threshold, it is determined to be a low-tolerance conservative type.

[0052] The threshold can be configured according to business needs; for example, the threshold for rapid diagnosis can be set to 0.7. The four-category stratification fully covers all user scenarios. Existing technology only simply distinguishes between normal and emergency scenarios, lacking dedicated stratification logic for scenarios involving the elderly or those with low-end devices and low tolerance.

[0053] Finally, in response to determining that the current consultation intent level is high-risk emergency, a high-risk circuit breaker flag is generated, and an absolute circuit breaker signal is sent.

[0054] The absolute circuit breaker signal is an independent bypass control command that bypasses all conventional questioning processes and directly triggers emergency output. It does not require waiting for multiple rounds of questioning and answers, and achieves immediate interception of emergencies. Existing solutions do not have an independent bypass cutoff mechanism, and will continue to ask questions, delaying emergency guidance.

[0055] Step 203: Use the current consultation format to conduct follow-up questions to generate a candidate question matrix.

[0056] In this embodiment, the aforementioned execution entity can use the current consultation format to conduct follow-up questions on candidates and generate a candidate question matrix.

[0057] The candidate question matrix is ​​a two-dimensional data table. Each candidate question is bound to five sets of parameters: text, identification features, high-risk weight, complexity, and character length. It is generated in batches from the medical knowledge graph.

[0058] This step forms the basis for the data source of dynamic question selection.

[0059] In some embodiments, the high-risk emergency type corresponds to an immediate and forceful diagnostic output; the rapid departmental identification type corresponds to a short-round, rapid convergence consultation; the ordinary, detailed follow-up questioning type corresponds to a multi-round, information-gain consultation; and the low-tolerance, conservative type corresponds to a conservative consultation under fatigue constraints.

[0060] For example, there are four different rules for different types of consultations: follow-up questions are completely prohibited for high-risk and acute cases; for rapid departmental identification, a maximum of 2 rounds are allowed; for ordinary detailed follow-up questions, there is no hard limit on the number of rounds but it is subject to fatigue factor control; for low-tolerance conservative cases, there are short single-sentence questions, a maximum of 2 rounds, and strict character limits.

[0061] In some embodiments, in response to determining that the current consultation format allows for continued inquiry, the maximum confidence level is lower than a preset confidence threshold and the absolute circuit breaker has not been triggered, candidate follow-up questions are conducted; based on standardized symptom terms, relevant co-occurring symptoms, accompanying symptoms and differential diagnosis points are retrieved from the medical knowledge graph, symptom co-occurrence library or differential diagnosis rule base, and assembled into a candidate question matrix.

[0062] The medical knowledge graph stores the relationships between symptoms, complications, and differential diagnoses; the symptom co-occurrence database records common clinical complications; and the differentiation rule database stores the key points for distinguishing similar diseases during consultation. All three are combined to generate all candidate questions in batches and store them in a matrix.

[0063] Step 204: Calculate the comprehensive technical benefit value based on the candidate problem matrix.

[0064] In this embodiment, the aforementioned execution entity can calculate the comprehensive technical benefit value based on the candidate problem matrix.

[0065] The comprehensive technical benefits can be quantified by a weighted formula that measures the value of a single question, while simultaneously evaluating triage efficiency, medical safety, and user burden. This is the core innovative algorithm of this solution.

[0066] In some embodiments, calculating the overall technical benefit value may include the following steps: The first step is to remove confirmed information, duplicate semantic questions, and questions with low differentiation of candidate departments from the candidate question matrix.

[0067] Filtering optimization reduces computational load, avoids redundant or irrelevant questions from consuming computational resources, and lowers the overhead of model inference.

[0068] The second step involves simulating the update results of the candidate department probability distribution vector under different answer branches for the candidate questions in the candidate question matrix, and calculating the Shannon entropy difference based on the probability distribution before and after the question to obtain the information entropy reduction.

[0069] Information entropy reduction, also known as departmental differentiation enhancement, is a positive benefit. Entropy represents the degree of disorder in departmental probabilities. The greater the decrease in entropy after asking a question, the stronger the question's ability to differentiate between departments. For example, asking "Do you have a cough?" can significantly reduce the probability of a respiratory illness, resulting in a high entropy reduction.

[0070] The third step is to assign high-risk weights to candidate questions according to preset rules.

[0071] Questions involving the identification of acute conditions such as myocardial infarction and shock are given higher weight, increasing the overall score of such questions and enabling priority inquiry for acute conditions.

[0072] The fourth step is to calculate the fatigue cost of a single interaction for candidate questions based on the character length of the question text, lexical complexity, sentence complexity, and terminal understanding overhead estimation coefficient.

[0073] Single-interaction fatigue is a negative value quantified as the burden of reading and input for the user. The longer the sentence and the more difficult the words, the higher the fatigue cost, which is deducted from the total score as a penalty.

[0074] The fifth step involves weighting the information entropy reduction and high-risk identification gain as positive benefit items and the fatigue cost of a single interaction as a negative penalty item, and summing them according to preset weights to calculate the comprehensive technical benefit value of the candidate problem. The preset weights include information gain weight, high-risk weight, and fatigue cost penalty weight.

[0075] The formula for the comprehensive technical benefit value can be as follows: ; in, It is the comprehensive technical benefit value. It is a decrease in information entropy. It is a high-risk identification gain value. This is the fatigue cost value per interaction. It is the information gain weight. It is a high-risk weight. It is the fatigue cost penalty weight.

[0076] The sixth step is to increase the fatigue penalty weight in response to the identification of login account attributes as elderly people, children under guardianship, or low-configuration, low-bandwidth terminals; and to increase the information gain weight in response to the determination that the chief complaint involves difficult and complex symptoms or that the probability distribution of the current department is too dispersed.

[0077] Adaptive adjustment: for the elderly / low-end devices Automatic zoom in, prioritizing short and simple questions; complex and multi-disease chief complaints. Enlarge the scope and prioritize questions with high differentiation; all existing technical parameters are fixed and cannot be adapted to the scenario.

[0078] Step 205: Determine the current round of problems based on comprehensive technical benefits.

[0079] In this embodiment, the aforementioned executing entity can determine the current round of problems based on comprehensive technical benefits.

[0080] Iterate through all candidate problems in the matrix and select... The maximum value is used as the question pushed in this round. The weighting varies under different consultation formats to achieve tiered and optimal question selection.

[0081] In some embodiments, the candidate questions in the candidate question matrix are traversed, and the candidate question with the highest comprehensive technical benefit value is selected as the current round question based on the current consultation format. Among them, the rapid departmental identification type corresponds to the question that can narrow down the departmental scope the fastest, the ordinary detailed follow-up question type corresponds to the high discrimination question with the largest reduction in information entropy, and the low tolerance conservative type corresponds to the question with shorter text and lower fatigue cost.

[0082] For example, for the same two questions, the one with moderate discrimination but which is for screening for myocardial infarction, scores higher after the high-risk weight is increased, and the system pushes it first, while taking into account the principle of prioritizing safety.

[0083] Step 206: In the current consultation session corresponding to the current round of questions, continuously collect underlying interaction and system load data, calculate the global cumulative fatigue convergence factor, and determine whether to terminate the consultation through a multi-condition parallel state machine after each round of interaction.

[0084] In this embodiment, the aforementioned execution entity can continuously collect underlying interaction and system load data in the current consultation session corresponding to the current round of questions, calculate the global cumulative fatigue convergence factor, and determine whether to terminate the consultation through a multi-condition parallel state machine after each round of interaction.

[0085] The consultation session is a single user's complete triage process, with independent data communication from inputting the chief complaint to outputting the department. The global cumulative fatigue convergence factor integrates three-dimensional load indicators: rounds, text, and interface latency, quantifying the dual pressure on both the user and the server. The multi-condition parallel Boolean state machine synchronously verifies four types of stopping conditions; any trigger immediately stops the consultation, corresponding to the complete logic of module three.

[0086] In some embodiments, calculating the global cumulative fatigue convergence factor may include the following steps: First, record the interaction round counter value, the total number of characters of input and output text, and the interface I / O latency data in the current consultation session.

[0087] The server logs collect three types of underlying hardware / interaction metrics in real time, without the need for additional database queries, resulting in higher real-time performance.

[0088] Then, the interaction round counter value, the total number of characters in the input and output text, and the interface I / O delay data are weighted and summed according to the fatigue factor weight to calculate the global cumulative fatigue convergence factor, and the global cumulative fatigue convergence factor is written into the global state bus.

[0089] The formula for the global cumulative fatigue convergence factor can be as follows: ; in, It is the global cumulative fatigue convergence factor. It is the value of the interaction round counter. It is the total number of characters in the input and output text. This is interface I / O latency data. , and It is the fatigue factor weight.

[0090] The global status bus enables data communication between all modules, and the fatigue values ​​can be read by the identification, inquiry, and output modules. Existing technologies do not have a unified status data channel.

[0091] In some embodiments, determining whether to terminate the consultation may include the following steps: First, configure the stop strategy corresponding to the current consultation format for the state machine.

[0092] High-risk, acute cases require an immediate cessation strategy; rapid subject identification requires a lower maximum number of rounds; general, detailed follow-up questions require multiple rounds of follow-up questions driven by higher information gain; and low-tolerance, conservative cases require a lower fatigue threshold and stricter character length constraints.

[0093] Four categories of stratified matching independent thresholds: high risk is stopped immediately; rapid subject determination has a low upper limit on the number of rounds; normal allows multiple rounds with high entropy gain; low tolerance for fatigue has an extremely low threshold.

[0094] Then, a multi-condition parallel Boolean decision state machine is initialized and maintained for the current consultation session. After each round of interaction, the termination conditions are polled or checked in parallel. The termination conditions include: whether the maximum probability value in the department probability distribution vector reaches the preset probability threshold, whether the global cumulative fatigue convergence factor exceeds the fatigue threshold, whether the interaction round counter reaches the maximum limit, and whether a high-risk circuit breaker signal is received.

[0095] Parallel verification differs from serial judgment in that it involves simultaneous calculation of four conditions, with high-risk signals having the highest priority and not waiting for other conditions to be calculated.

[0096] Finally, in response to the determination that any termination condition is met, a termination signal is output to prevent the generation of candidate problems; in response to the determination that the termination reason is the risk of insufficient information due to fatigue exceeding the limit or the number of rounds reaching the upper limit, the system switches to a safe and conservative triage mode; in response to the determination that the termination reason is a high-risk circuit breaker, the system enters the forced interception branch.

[0097] The safe and conservative triage model can increase the weight of general practitioners and internal medicine practitioners in advance to reduce the risk of missed diagnoses; the forced interception branch can directly display the emergency pop-up and emergency entry point, and the two independent output branches are a safety fallback logic that existing technologies do not have.

[0098] Step 207: Based on the final probability distribution, risk level, user intent level, and termination reason at the time of termination of consultation, output structured triage results and recommendation reasons.

[0099] In this embodiment, the aforementioned execution entity can output structured triage results and recommendation reasons based on the final probability distribution, risk level, user intent level, and termination reason at the time of termination of the consultation.

[0100] This step provides differentiated outputs based on the risks, stratification, and reasons for service suspension throughout the entire consultation process, thus achieving a logical closed loop between consultation and recommendation results.

[0101] In some embodiments, outputting structured triage results and recommendation rationale may include the following steps: The first step is to determine the risk level based on the high-risk circuit breaker markers, the final probability distribution, and the termination reasons.

[0102] The risk levels are divided into high, medium, and low. High-risk patients are required to seek emergency care, while those in medium and low risk are referred to general outpatient clinics.

[0103] The second step is to respond to the determination that the current consultation intent level is high-risk emergency, or to trigger a high-risk circuit breaker and enter the forced interception branch, and output emergency prompt labels and emergency medical advice.

[0104] The front-end page features a prominent red emergency notification, along with navigation to the nearest emergency department and a quick registration button, prioritizing access to medical care for emergency patients.

[0105] The third step involves determining whether the current consultation intent level is rapid departmental identification or general detailed follow-up questioning. If the high-risk circuit breaker is not triggered, the recommended department list, corresponding recommendation confidence level, and a summary of structured recommendation reasons are output.

[0106] The summary of the recommendation clearly explains the triage criteria, such as "cough and fever symptoms are highly matched with respiratory medicine, with a match confidence level of 0.86", which improves readability for both doctors and patients.

[0107] The fourth step is to respond to the determination that the current consultation intention level is low-tolerance conservative, or to passively terminate the consultation due to fatigue exceeding the limit or the maximum number of rounds. In this case, the recommendation weight of general practitioners, internal medicine, or the preset safety fallback departments is increased, and low-risk conservative suggestions are output.

[0108] To address the issue of misdiagnosis when insufficient information is collected during the inquiry process, a comprehensive approach is used to reduce the risk of misdiagnosis complaints.

[0109] The fifth step is to read the user's historical health records, past medical consultation records, or chronic disease management tags, and supplement the summary of the recommendation reasons.

[0110] By integrating users' chronic diseases and past medical records, recommendations can be made that are more in line with their long-term health conditions. Most existing triage systems cannot link with historical medical data.

[0111] The AI-based triage method provided in this disclosure relies on four coupled functional modules to form a complete closed-loop control link. It achieves multi-dimensional optimization throughout the entire process, from chief complaint analysis and hierarchical question selection, dynamic benefit selection, load state machine management and hierarchical security output. At the same time, it overcomes the four inherent defects of traditional fixed questionnaires and unconstrained large model consultations, and takes into account the four technical effects of system computing power saving, patient user experience, emergency medical safety, and multi-terminal adaptation. It can be implemented on various online and offline smart triage hardware and software platforms.

[0112] Compared with existing technology products, the embodiments disclosed herein have five major advantages: 1. Improve triage convergence efficiency: Rely on information entropy reduction to select the best questions, significantly reduce the number of invalid question rounds, and shorten the triage path; 2. Reduce system computing power and network transmission overhead: Control interaction characters, rounds, and interface latency to reduce large language model inference calls and save token and bandwidth resources; 3. Enhance emergency medical safety: Pre-emptive high-risk word identification and follow-up high-risk gain dual screening, emergency cases are directly isolated, reducing the probability of missed diagnosis of critical illnesses; 4. Adaptable to diverse users and hardware: Adaptively adjusts revenue weights, automatically reducing the intensity of follow-up questions for the elderly and those with low-end devices; 5. Triage results are interpretable: they include departmental confidence levels and judgment criteria, facilitating the connection between online registration and offline referral services.

[0113] The AI-based triage method provided in this disclosure is adaptable to three types of front-end carriers: mobile mini-programs, in-hospital self-service terminals, and tablet consultation devices. The complete interaction process is as follows: 1. User-side operation: Input the chief complaint of illness using text / voice and submit; 2. Backend module 1 analyzes symptoms and determines the level of consultation intent; 3. In non-high-risk scenarios, the optimal single-round question is pushed out, and users answer in a loop; 4. Differentiated result display after triggering any termination condition; high-risk cases directly redirect to the emergency page; automatically shorten question length and reduce rounds for elderly / low-end devices. The entire interaction system caters to multiple scenarios including online internet hospitals and offline primary care triage, significantly reducing the number of large model calls and saving server computing power and bandwidth costs.

[0114] Using the main complaint "I have a cold" as an example, the entire process of the four modules is traversed to intuitively demonstrate the operational logic of the whole solution: 1. Upon entering Module 1, the system performs hierarchical identification: First, it performs NLP named entity recognition on the chief complaint "I have a cold," standardizing the colloquial "cold" into the medical term "upper respiratory tract infection." It then traverses the built-in high-risk core vocabulary, and if no high-risk keywords such as chest pain, severe bleeding, altered consciousness, or severe dyspnea are matched, no high-risk absolute circuit breaker signal is generated. The system retrieves the preset symptom-department conditional probability matrix and calculates the current department probability distribution: General Practice / Internal Medicine 0.55, Respiratory Medicine 0.30, Otolaryngology 0.10, and other departments combined 0.05. The current maximum confidence level is 0.55, lower than the preset rapid department identification threshold of 0.7-0.8. The current chief complaint is only a single mild symptom and does not belong to complex or difficult cases. Combined with the fact that the logged-in account is an ordinary adult and the mobile phone has sufficient bandwidth, it is determined to be a typical detailed follow-up question type. If the logged-in account is marked as a 75-year-old elderly user and the device is an old, low-spec terminal, the system will directly switch to a low-tolerance conservative type, tightening the consultation constraints in advance.

[0115] 2. Enter Module Two: Dynamic Follow-up Decision Making: Based on the medical knowledge graph and symptom co-occurrence database, a candidate question matrix is ​​generated in batches. The candidate content includes "Do you have a fever? What temperature?", "Do you have a cough or phlegm?", "Do you have a sore throat?", "Do you have chest tightness or difficulty breathing?", and "How many days have the symptoms lasted?". The information entropy reduction value is calculated for each candidate question. High-risk gains Fatigue overhead per interaction Substitute into the scoring formula The overall benefit ranking is completed; among them, "Do you have a fever or cough?" can significantly narrow down the distinction between respiratory medicine and general practice, has a high entropy reduction value, and has the highest overall score due to its short text and low fatigue cost. The system will push this question to the user as the question in this round.

[0116] 3. Branch Scenario A: The user replies "No fever, just stuffy nose and itchy throat, symptoms lasted 1 day". The system updates the department probability distribution, and the confidence level of the general practice increases to 0.86, reaching the preset confidence threshold. Module 3's multi-condition parallel state machine verifies that the confidence level meets the standard, outputs a termination signal, and directly jumps to Module 4, outputting the structured triage result: recommending general practice / general internal medicine, marking the matching confidence level, and attaching a summary of the recommendation reason "common upper respiratory tract infection, stay at home, drink more water and rest, no need for emergency".

[0117] 4. Branch Scenario B: A user replies "High fever of 39℃, severe cough, chest tightness," this question has a high-risk weight. The system automatically escalates and adds questions to differentiate between chest pain and difficulty breathing. If high-risk patient complaints such as chest pain and persistent difficulty breathing are identified in subsequent responses, Module 1 immediately generates a high-risk circuit breaker flag and issues an absolute circuit breaker signal. The state machine prioritizes triggering the forced interception branch, and Module 4 places a red emergency room reminder label at the top, pushes suggestions for fever clinics and emergency medical treatment, and simultaneously links to the nearest emergency room registration portal.

[0118] 5. Branch Scenario C: If the user is an elderly person, the probability of visiting the department remains dispersed after three rounds of interaction, resulting in a global cumulative fatigue convergence factor. If the fatigue threshold is exceeded, the state machine in module three determines that the fatigue over-limit termination condition is triggered and automatically switches to the safe and conservative triage mode; module four increases the weight of general practice and internal medicine as a fallback recommendation, outputs conservative treatment suggestions, prompts to prioritize general practice and follow up with specialists when necessary, and reduces the risk of mis-triage due to insufficient information.

[0119] 6. Branch Scenario D: The user directly replies "It's just a common cold, no need to ask any more questions." The system recognizes that the user has actively terminated the interaction and has low interaction tolerance. It automatically switches to the low-tolerance conservative output branch and directly outputs basic general care triage suggestions, ending the entire consultation process in advance and reducing the user's interaction burden.

[0120] Figure 3 The diagram illustrates the overall architecture, module division, and business execution flowchart of the intelligent triage system.

[0121] This block diagram is divided into a five-layer logical structure: The first layer is the user input layer, used to receive the user's chief complaint text; the second layer contains four core functional modules: Module 1, User Chief Complaint and Intent Layered Recognition Module; Module 2, Layered Driving Consultation Format Selection and Dynamic Follow-up Questioning Decision Module; Module 3, Fatigue Constraint and State Machine Stop Control Module; and Module 4, Security Layering and Department Recommendation Module; the third layer is the business control path, including the highest priority high-risk circuit breaker bypass, multi-round interactive follow-up questioning feedback loop, and the underlying data supply path; the fourth layer is the global state bus, used to uniformly store real-time operational data such as interaction rounds, text character count, I / O latency, global cumulative fatigue convergence factor, and department probability distribution vector; the fifth layer is the data and knowledge foundation, configured with a medical terminology set / NER dictionary, symptom-department conditional probability matrix, medical knowledge graph / identification rule base, high-risk core terminology library, user health records / historical consultation records, and fatigue and benefit weight configuration items, providing unified data support for the four upper-layer modules.

[0122] The overall business data flow process is as follows: After the user submits the chief complaint text, it is sent to Module 1 to complete text preprocessing and medical entity recognition, generate a structured feature vector E and a department probability distribution vector P, and complete the user intent level determination. Once an absolute circuit breaker signal is generated, the user can directly jump to Module 4 along the high-risk circuit breaker bypass. If the consultation is allowed to continue, the user enters Module 2 to generate a candidate question matrix Q, complete the comprehensive technical benefit calculation, select the optimal question for this round, and receive the user's response through the follow-up question feedback loop, and cyclically update the department probability. The running data generated by each round of interaction is written to the global state bus in real time. Module 3 reads the bus data, calculates the global cumulative fatigue convergence factor, and verifies multiple termination conditions in parallel through a multi-condition parallel Boolean state machine. After outputting the termination signal and the final department probability distribution, the user enters Module 4 to complete the risk level determination, generate a structured recommendation reason summary based on historical health records, and finally output the stratified triage results.

[0123] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of an artificial intelligence-based triage device, which is similar to... Figure 1 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0124] like Figure 4As shown, the AI-based triage device 400 of this embodiment may include: an identification module 401, a determination module 402, a consultation module 403, and an output module 404. The identification module 401 is configured to perform intent-layer recognition on the user's input complaint text to determine the current consultation intent level; the determination module 402 is configured to select the current consultation format corresponding to the current consultation intent level and determine the current round of questions; the consultation module 403 is configured to conduct consultation using the current round of questions and determine whether to terminate the consultation after each round of interaction; the output module 404 is configured to output a structured triage result and recommendation reason based on the consultation intent level at the time of termination.

[0125] In this embodiment, the specific processing of the identification module 401, the determination module 402, the consultation module 403, and the output module 404 in the AI-based triage device 400, and the resulting technical effects, can be found in the following references: Figure 1 The relevant descriptions of steps 101-104 in the corresponding embodiments will not be repeated here.

[0126] In some optional implementations of this embodiment, the identification module 401 is further configured to: convert unstructured medical expressions in the chief complaint text into structured medical feature vectors; and identify the current consultation intent level based on the structured medical feature vectors, risk, confidence, and interaction constraints.

[0127] In some optional implementations of this embodiment, the recognition module 401 is further configured to: preprocess the chief complaint text to generate a preprocessed chief complaint text, wherein the preprocessing includes at least one of the following: word segmentation, spelling normalization, merging of synonymous symptoms, and noise filtering; use a medical language model to perform named entity recognition on the preprocessed chief complaint text, extract medical semantic entities, and map them to a preset medical terminology set to obtain corresponding standard symptom terms; and construct a structured medical feature vector based on the standard symptom terms.

[0128] In some optional implementations of this embodiment, the identification module 401 is further configured to: input the structured medical feature vector into a preset symptom-department conditional probability matrix, generate an initial probability distribution vector for each candidate department, and use the maximum probability value in the initial probability distribution vector as the maximum confidence level; perform intent layering based on the high-risk core word triggering state, maximum confidence level, chief complaint complexity, account attributes, terminal bandwidth capability, and historical interaction tolerance features to determine the current consultation intent level; in response to determining that the current consultation intent level is a high-risk emergency type, generate a high-risk circuit breaker flag and send an absolute circuit breaker signal.

[0129] In some optional implementations of this embodiment, the identification module 401 is further configured to: determine as a high-risk emergency type in response to identifying preset high-risk core words or high-risk chief complaint combinations; determine as a rapid departmental identification type in response to determining that the maximum confidence level is greater than a preset confidence threshold; determine as a normal detailed follow-up question type in response to determining that the chief complaint information is insufficient, the candidate departments are scattered and there are conditions for continuing multiple rounds of interaction; and determine as a low-tolerance conservative type in response to determining that the account attribute is elderly, low-configuration low-bandwidth terminal or historical interaction display tolerance is lower than a preset tolerance threshold.

[0130] In some optional implementations of this embodiment, the determining module 402 is further configured to: use the current consultation format to conduct follow-up questions on candidates and generate a candidate question matrix; calculate the comprehensive technical benefit value based on the candidate question matrix; and determine the current round of questions based on the comprehensive technical benefit.

[0131] In some optional implementations of this embodiment, the high-risk emergency type corresponds to the immediate strong output form; the rapid departmental identification type corresponds to the short-round rapid convergence consultation form; the ordinary detailed follow-up questioning type corresponds to the multi-round information gain consultation form; and the low tolerance conservative type corresponds to the conservative consultation form under fatigue constraints.

[0132] In some optional implementations of this embodiment, the determining module 402 is further configured to: in response to determining that the current consultation format allows for continued inquiry, the maximum confidence level is lower than a preset confidence threshold and the absolute circuit breaker has not been triggered, conduct candidate follow-up questions; based on standardized symptom terms, retrieve relevant co-occurring symptoms, accompanying symptoms and differential diagnosis points from a medical knowledge graph, a symptom co-occurrence library or a differential diagnosis rule base, and assemble them into a candidate question matrix.

[0133] In some optional implementations of this embodiment, the determining module 402 is further configured to: remove confirmed information, duplicate semantic questions, and questions with low differentiation of candidate departments from the candidate question matrix; for candidate questions in the candidate question matrix, simulate the update results of the candidate question on the probability distribution vector of candidate departments under different answer branches, and calculate the Shannon entropy difference based on the probability distribution before and after the question to obtain the information entropy reduction; configure high-risk weights for candidate questions according to preset rules; calculate the single-interaction fatigue cost value of candidate questions based on the character length of the question text, vocabulary complexity, sentence complexity, and terminal understanding overhead estimation coefficient; and calculate the comprehensive technical benefit value of candidate questions by weighting and summing the information entropy reduction and high-risk identification gain value as positive benefit items and the single-interaction fatigue cost value as negative penalty items according to preset weights, wherein the preset weights include information gain weight, high-risk weight, and fatigue cost penalty weight.

[0134] In some optional implementations of this embodiment, the determining module 402 is further configured to: traverse the candidate questions in the candidate question matrix, and select the candidate question with the largest comprehensive technical benefit value as the current round question in combination with the current consultation format. Among them, the rapid department type corresponds to the question that can narrow down the department scope the fastest, the ordinary detailed follow-up question type corresponds to the high discrimination question with the largest reduction in information entropy, and the low tolerance conservative type corresponds to the question with shorter text and lower fatigue cost.

[0135] In some optional implementations of this embodiment, the determining module 402 is further configured to: increase the fatigue overhead penalty weight in response to identifying that the login account attribute is elderly, child guardian or low-configuration low-bandwidth terminal; and increase the information gain weight in response to determining that the chief complaint involves difficult and complex symptoms or that the probability distribution of the current department is too scattered.

[0136] In some optional implementations of this embodiment, the consultation module 403 is further configured to: continuously collect underlying interaction and system load data in the current consultation session corresponding to the current round of questions, calculate the global cumulative fatigue convergence factor, and determine whether to terminate the consultation through a multi-condition parallel state machine after each round of interaction.

[0137] In some optional implementations of this embodiment, the consultation module 403 is further configured to: record the interaction round counter value, the total number of characters of input and output text, and the interface I / O delay data in the current consultation session; calculate the global cumulative fatigue convergence factor by weighted summation of the interaction round counter value, the total number of characters of input and output text, and the interface I / O delay data according to the fatigue factor weight, and write the global cumulative fatigue convergence factor into the global state bus.

[0138] In some optional implementations of this embodiment, the consultation module 403 is further configured to: configure a stop strategy corresponding to the current consultation format for the state machine; initialize and maintain a multi-condition parallel Boolean decision state machine for the current consultation session; after each round of interaction, poll or perform parallel verification of the termination conditions, wherein the termination conditions include: whether the maximum probability value in the department probability distribution vector reaches a preset probability threshold, whether the global cumulative fatigue convergence factor exceeds the fatigue threshold, whether the interaction round counter reaches the maximum limit, and whether a high-risk circuit breaker signal is received; in response to determining that any termination condition is met, output a termination signal to prevent the continued generation of candidate questions; in response to determining that the termination reason is the risk of insufficient information due to fatigue exceeding the limit or the round reaching the upper limit, switch to a safe and conservative triage mode; in response to determining that the termination reason is a high-risk circuit breaker, enter the forced interception branch.

[0139] In some optional implementations of this embodiment, the high-risk emergency type corresponds to an immediate stop strategy; the rapid subject determination type corresponds to a lower round limit; the ordinary detailed probing type corresponds to multi-round probing driven by higher information gain; and the low-tolerance conservative type corresponds to a lower fatigue threshold and stricter character length constraints.

[0140] In some optional implementations of this embodiment, the output module 404 is further configured to output structured triage results and recommendation reasons based on the final probability distribution, risk level, user intent level, and termination reason at the time of termination of the consultation.

[0141] In some optional implementations of this embodiment, the output module 404 is further configured to: determine the risk level based on the high-risk circuit breaker flag, the final probability distribution, and the termination reason; in response to determining that the current consultation intent level is high-risk emergency type, or triggering a high-risk circuit breaker and entering a forced interception branch, output an emergency prompt label and emergency medical treatment suggestion; in response to determining that the current consultation intent level is rapid departmental identification type or ordinary detailed follow-up questioning type, when no high-risk circuit breaker is triggered, output a list of recommended departments, corresponding recommendation confidence levels, and a structured summary of recommendation reasons; in response to determining that the current consultation intent level is low-tolerance conservative type, or passively terminated due to fatigue exceeding the limit or the maximum number of rounds, increase the recommendation weight of general practice, general internal medicine, or preset safety backup departments, and output low-risk conservative suggestions; read the user's historical health records, previous consultation records, or chronic disease management tags to supplement the summary of recommendation reasons.

[0142] The collection, storage, use, processing, transmission, provision, and disclosure of any type of information, such as user personal information, in this technical solution comply with relevant laws and regulations and do not violate public order and good morals.

[0143] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0144] Figure 5 A schematic block diagram of an example electronic device 500 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0145] like Figure 5As shown, device 500 includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 502 or a computer program loaded from storage unit 508 into random access memory (RAM) 503. RAM 503 may also store various programs and data required for the operation of device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via bus 504. Input / output (I / O) interface 505 is also connected to bus 504.

[0146] Multiple components in device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0147] The computing unit 501 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as an AI-based triage method. For example, in some embodiments, the AI-based triage method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program can be loaded and / or installed on device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of the AI-based triage method described above can be performed. Alternatively, in other embodiments, computing unit 501 may be configured to perform an AI-based triage method by any other suitable means (e.g., by means of firmware).

[0148] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0149] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0150] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0151] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0152] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0153] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, distributed system servers, or servers incorporating blockchain technology.

[0154] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution provided in this disclosure can be achieved, and this is not limited herein.

[0155] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A triage and referral method based on artificial intelligence, comprising: Perform intent-layer recognition on the user's input chief complaint text to determine the current consultation intent level; Select the current consultation format corresponding to the current consultation intent level, and determine the current round of questions; The current round of questions is used to conduct a consultation, and it is determined whether to terminate the consultation after each round of interaction; Based on the consultation intent level at the time of termination, the system outputs structured triage results and recommendation reasons.

2. The method according to claim 1, wherein, The process of performing intent-layer recognition on the user's input complaint text to determine the current consultation intent level includes: Convert the unstructured medical expressions in the chief complaint text into structured medical feature vectors; Based on the structured medical feature vector, risk, confidence level, and interaction constraints, the current consultation intent level is identified.

3. The method according to claim 2, wherein, The step of converting unstructured medical expressions in the chief complaint text into structured medical feature vectors includes: The main complaint text is preprocessed to generate a preprocessed main complaint text. The preprocessing includes at least one of the following: word segmentation, spell normalization, synonym symptom merging, and noise filtering. The preprocessed chief complaint text is subjected to named entity recognition using a medical language model to extract medical semantic entities and map them to a preset medical terminology set to obtain the corresponding standard symptom terms. Based on the standard symptom terminology, the structured medical feature vector is constructed.

4. The method according to claim 2, wherein, The process of identifying the current consultation intent level based on the structured medical feature vector, risk, confidence level, and interaction constraints includes: The structured medical feature vector is input into a preset symptom-department conditional probability matrix to generate an initial probability distribution vector for each candidate department, and the maximum probability value in the initial probability distribution vector is taken as the maximum confidence level. Intent layering is performed based on the triggering state of high-risk core words, the maximum confidence level, the complexity of the chief complaint, account attributes, terminal bandwidth capacity, and historical interaction tolerance characteristics to determine the current consultation intent level; In response to determining that the current consultation intent level is a high-risk emergency, a high-risk circuit breaker flag is generated and an absolute circuit breaker signal is sent.

5. The method according to claim 4, wherein, The intent layering based on high-risk core keyword triggering status, maximum confidence level, chief complaint complexity, account attributes, terminal bandwidth capacity, and historical interaction tolerance features determines the current consultation intent level, including: In response to the identification of preset high-risk core words or high-risk chief complaint combinations, it is determined to be a high-risk acute illness type; In response to determining that the maximum confidence level is greater than a preset confidence threshold, the classification is determined to be rapid; In response to the insufficient information on the chief complaint, the scattered distribution of candidate departments, and the availability of conditions for further multiple rounds of interaction, the case was determined to be a typical detailed follow-up questioning type. In response to determining that the account attribute is elderly, low-configuration low-bandwidth terminal, or historical interaction display tolerance is lower than a preset tolerance threshold, it is determined to be of the low tolerance conservative type.

6. The method according to claim 1, wherein, The step of selecting the current consultation format corresponding to the current consultation intent level and determining the current round of questions includes: Using the current consultation format, candidate follow-up questions are conducted to generate a candidate question matrix; Based on the candidate problem matrix, calculate the comprehensive technical benefit value; Based on the comprehensive technical benefits, the current round problem is determined.

7. The method according to claim 6, wherein, High-risk acute cases correspond to immediate and forceful diagnostic output; rapid departmental identification corresponds to a short-round, rapid convergence consultation; ordinary, detailed follow-up questioning corresponds to a multi-round, information-gain consultation; and low-tolerance, conservative cases correspond to a conservative consultation under fatigue constraints.

8. The method according to claim 7, wherein, The step of using the current consultation format to conduct follow-up questions and generate a candidate question matrix includes: In response to determining that the current consultation format allows for continued inquiry, the maximum confidence level is lower than a preset confidence threshold, and the absolute circuit breaker has not been triggered, candidate follow-up questions are conducted; Based on standardized symptom terminology, relevant co-occurring symptoms, accompanying symptoms, and key identification points are retrieved from medical knowledge graphs, symptom co-occurrence databases, or differential diagnosis rule bases, and assembled into the candidate question matrix.

9. The method according to claim 8, wherein, The calculation of the comprehensive technical benefit value based on the candidate problem matrix includes: Remove confirmed information, duplicate semantic questions, and questions with low differentiation of candidate departments from the candidate question matrix; For the candidate questions in the candidate question matrix, the update results of the candidate department probability distribution vector under different answer branches are simulated, and the Shannon entropy difference is calculated based on the probability distribution before and after the question to obtain the information entropy reduction. The candidate questions are assigned high-risk weights according to preset rules; The fatigue cost of a single interaction for the candidate question is calculated based on the character length of the question text, lexical complexity, sentence complexity, and terminal understanding overhead estimation coefficient. The information entropy reduction and high-risk identification gain are taken as positive benefit items, and the fatigue cost of a single interaction is taken as a negative penalty item. They are weighted and summed according to preset weights to calculate the comprehensive technical benefit value of the candidate problem. The preset weights include information gain weight, high-risk weight, and fatigue cost penalty weight.

10. The method according to claim 9, wherein, The process of determining the current round of problems based on the comprehensive technological benefits includes: The candidate questions in the candidate question matrix are traversed, and the candidate question with the largest comprehensive technical benefit value is selected as the current round question in combination with the current consultation format. Among them, the rapid department type corresponds to the question that can narrow down the department scope the fastest, the ordinary detailed follow-up question type corresponds to the high discrimination question with the largest reduction in information entropy, and the low tolerance conservative type corresponds to the question with shorter text and lower fatigue cost.

11. The method according to claim 10, wherein, The calculation of the comprehensive technical benefit value based on the candidate problem matrix further includes: In response to the identification that the login account attributes are elderly, child guardians, or low-configuration, low-bandwidth terminals, the fatigue overhead penalty weight is increased; In response to determining that the chief complaint involves difficult or complex symptoms or that the probability distribution of the current department is too dispersed, the information gain weight is increased.

12. The method according to claim 1, wherein, The step of using the current round of questions for consultation, and determining whether to terminate the consultation after each round of interaction, includes: In the current consultation session corresponding to the current round of questions, the underlying interaction and system load data are continuously collected, the global cumulative fatigue convergence factor is calculated, and after each round of interaction, a multi-condition parallel state machine is used to determine whether to terminate the consultation.

13. The method according to claim 12, wherein, In the current consultation session corresponding to the current round of questions, the underlying interaction and system load data are continuously collected, and the global cumulative fatigue convergence factor is calculated, including: Record the interaction round counter value, total number of characters of input and output text, and interface I / O latency data in the current consultation session; The interaction round counter value, the total number of characters in the input and output text, and the interface I / O delay data are weighted and summed according to the fatigue factor weight to calculate the global cumulative fatigue convergence factor, and the global cumulative fatigue convergence factor is written into the global state bus.

14. The method according to claim 13, wherein, The step of determining whether to terminate the consultation after each round of interaction using a multi-condition parallel state machine includes: Configure the state machine with the stop strategy corresponding to the current consultation format; The current consultation session is initialized and maintained as a multi-condition parallel Boolean decision state machine. After each round of interaction, the termination conditions are polled or checked in parallel. The termination conditions include: whether the maximum probability value in the department probability distribution vector reaches the preset probability threshold, whether the global cumulative fatigue convergence factor exceeds the fatigue threshold, whether the interaction round counter reaches the maximum limit, and whether a high-risk circuit breaker signal is received. In response to the determination that any termination condition is met, a termination signal is output to prevent the generation of candidate problems. In response to the risk of insufficient information due to determining that the termination is caused by excessive fatigue or reaching the maximum number of rounds, switch to a safe and conservative triage mode. In response to the determination that the termination reason is a high-risk circuit breaker, the system enters the mandatory interception branch.

15. The method according to claim 14, wherein, High-risk, acute cases require an immediate cessation strategy; rapid subject identification requires a lower maximum number of rounds; general, detailed follow-up questions require multiple rounds of follow-up questions driven by higher information gain; and low-tolerance, conservative cases require a lower fatigue threshold and stricter character length constraints.

16. The method according to claim 1, wherein, The structured triage results and recommendation reasons are output based on the consultation intent level at the time of termination, including: Based on the final probability distribution, risk level, user intent level, and termination reason at the time of termination of the consultation, the structured triage result and recommendation reason are output.

17. The method according to claim 16, wherein, The structured triage results and recommendation reasons are output based on the final probability distribution, risk level, user intent level, and termination reason at the time of termination of the consultation, including: The risk level is determined based on the high-risk circuit breaker marker, the final probability distribution, and the termination reason; In response to determining that the current consultation intent level is high-risk emergency, or triggering a high-risk circuit breaker and entering a forced interception branch, an emergency prompt label and emergency medical advice are output. In response to determining that the current consultation intent level is either rapid department identification or general detailed follow-up questioning, when the high-risk circuit breaker is not triggered, the system outputs a list of recommended departments, the corresponding recommendation confidence level, and a summary of the structured recommendation reasons. In response to determining that the current consultation intent level is low-tolerance conservative, or due to fatigue exceeding the limit or passive termination due to the maximum number of rounds, the recommendation weight of general practice, internal medicine or preset safety fallback departments is increased, and low-risk conservative suggestions are output. Read the user's historical health records, past medical records, or chronic disease management tags to supplement the summary of the recommendation reasons.

18. A triage and referral device based on artificial intelligence, comprising: The recognition module is configured to perform intent-layer recognition on the user's input chief complaint text to determine the current consultation intent level. The determination module is configured to select the current consultation format corresponding to the current consultation intent level and determine the current round of questions; The consultation module is configured to conduct a consultation using the current round of questions and determine whether to terminate the consultation after each round of interaction. The output module is configured to output structured triage results and recommendation reasons based on the consultation intent level at the time of termination of the consultation.

19. An electronic device comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-17.

20. A non-transitory computer-readable storage medium storing computer instructions for causing the computer to perform the method of any one of claims 1-17.

21. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-17.

Images