A medical question and answer decision support method, system and related devices

Abstract

The application discloses a medical question and answer decision support method, system and related device, and relates to the technical field of medical information processing. The method comprises the following steps: in response to receiving a medical question input by a user, performing natural language processing on the medical question to generate a structured query; acquiring in real time a plurality of system running state parameters including a knowledge base query response time, a current number of concurrent requests and the like; if the knowledge base feedback is successfully acquired, generating first response content based on the current system state and the feedback result through a first prediction model; if the feedback is not successfully acquired, generating second response content based on the current system state through a second prediction model; and finally outputting a decision support result. Through the double-model redundancy decision mechanism, the application dynamically selects a response generation path, effectively deals with high-concurrency access and service uncertainty, and guarantees the continuity of system service, the efficiency of response and the credibility of result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical question and answer decision support, in particular to a medical question and answer decision support method, system and related device. BACKGROUND

[0002] In the process of intelligent medical informatization, medical question and answer systems, as an important tool for providing instant professional information reference, are increasingly widely used. Such systems provide information support for users by processing unstructured problems input by users, relying on background knowledge bases.

[0003] In actual operation, the system faces complex challenges such as high concurrent user access, multi-source heterogeneous medical data integration, and real-time response requirements. It is difficult to ensure the continuity of services, the efficiency of responses, and the reliability of results. SUMMARY

[0004] The technical problem solved by the present application is to provide a medical question and answer decision support method, system and related device to ensure the continuity of medical question and answer system services, the efficiency of responses, and the reliability of output results.

[0005] To solve the above technical problems, one technical solution adopted by the present application is to provide a medical question and answer decision support method, which comprises: in response to receiving a medical question input by a user, performing natural language processing on the medical question to generate a structured query; acquiring system running state parameters in real time, the system running state parameters including multiple types of knowledge base query response time, current concurrent request number, user historical interaction behavior data, medical data source availability state, and model reasoning resource occupation rate; in response to successfully acquiring feedback results of the knowledge base for the structured query within a current processing period, generating first response content based on the current system running state parameters and the feedback results through a first prediction model; in response to not successfully acquiring feedback results of the knowledge base within the current processing period, generating second response content based on the current system running state parameters through a second prediction model; wherein the first response content or the second response content is output as a decision support result for the medical question.

[0006] In response to not successfully acquiring feedback results of the knowledge base within the current processing period, generating second response content based on the current system running state parameters through a second prediction model, comprising: if the feedback results of the knowledge base are not successfully acquired within the current processing period, and the change of the system running state parameters exceeds a predetermined state change threshold before the end of the current processing period, then generating the second response content based on the current system running state parameters through the second prediction model; if the feedback results of the knowledge base are still not successfully acquired until the end of the current processing period, then generating the second response content based on the current system running state parameters through the second prediction model.

[0007] The step of performing natural language processing on the medical question to generate a structured query further comprises: obtaining an initial estimated processing delay based on the current system running state parameter by using a second prediction model; and setting a waiting timeout time for connecting with the knowledge base based on the initial estimated processing delay.

[0008] The step of performing natural language processing on the medical question to generate a structured query comprises: identifying medical entities and attributes in the medical question by using a named entity recognition technology in response to receiving the medical question input by the user; determining mutual relationships between the identified medical entities and attributes based on a pre-constructed medical knowledge graph; and generating a structured query statement for querying the knowledge graph according to the medical entities, attributes and mutual relationships.

[0009] The step of generating the first response content based on the current system running state parameter and the feedback result by using the first prediction model comprises: determining an actual response time and a result data volume of the current query based on the feedback result; and generating the first response content by using the first prediction model based on the current system running state parameter, the actual response time and the result data volume.

[0010] The step of generating the first response content based on the current system running state parameter, the actual response time and the result data volume by using the first prediction model comprises: performing normalization processing on the current system running state parameter to extract system state features; performing normalization processing on the actual response time and the result data volume to extract network transmission and calculation features of the current query; fusing the system state features and the network transmission and calculation features to obtain target fusion features; and generating the first response content by using the first prediction model based on the target fusion features.

[0011] In a case where the feedback result is successfully obtained, the first prediction model is trained by using a plurality of first training samples, the first training samples including system running state parameters and knowledge base feedback results of a historical period as training data and response content generation delay or optimal response strategy of the historical period as training labels; and in a case where the feedback result is not successfully obtained, the second prediction model is trained by using a plurality of second training samples, the second training samples including system running state parameters of a historical period as training data and response content generation delay or optimal response strategy of the historical period as training labels.

[0012] To solve the above technical problems, another technical solution adopted by the present application is to provide a medical question and answer decision support system, which comprises a natural language processing module, a state monitoring module, a first answer generation module, a second answer generation module and a result output module. The natural language processing module is configured to perform natural language processing on a medical question input by a user and generate a structured query in response to receiving the medical question. The state monitoring module is configured to acquire system running state parameters in real time, wherein the system running state parameters comprise multiple parameters such as knowledge base query response time, current number of concurrent requests, user historical interaction behavior data, medical data source availability state and model inference resource occupation rate. The first answer generation module is configured to generate first answer content based on the current system running state parameters and feedback results of the knowledge base for the structured query by using a first prediction model in response to successfully acquiring the feedback results of the knowledge base for the structured query in a current processing period. The second answer generation module is configured to generate second answer content based on the current system running state parameters by using a second prediction model in response to failing to successfully acquire the feedback results of the knowledge base in the current processing period. The result output module is configured to output the first answer content or the second answer content as a decision support result for the medical question.

[0013] To solve the above technical problems, another technical solution adopted by the present application is to provide a medical question and answer decision support device, which comprises a memory and at least one processor. The memory stores instructions. The at least one processor invokes the instructions in the memory to enable the medical question and answer decision support device to perform the steps of the medical question and answer decision support method according to any one of the above embodiments.

[0014] To solve the above technical problems, another technical solution adopted by the present application is to provide a computer readable storage medium, which stores instructions. The instructions are executed by a processor to implement the steps of the medical question and answer decision support method according to any one of the above embodiments.

[0015] Compared with the prior art, the present application has the following beneficial effects: by monitoring multiple system running state parameters such as knowledge base query response time, number of concurrent requests and data source availability state in real time, and combining a dual-path redundant decision mechanism formed by the first prediction model and the second prediction model, reliable answer content can be dynamically generated in the case of normal or abnormal knowledge base service. This technical solution effectively solves the problems of response delay or service interruption caused by high system load, network delay or data source failure, thereby significantly improving the service continuity, response speed and result reliability of the medical question and answer system in a high-concurrency and unstable environment. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is a flowchart of an embodiment of the medical question and answer decision support method of the present application.

[0017] Figure 2 is a structural framework schematic diagram of an embodiment of the medical question and answer decision support system of the present application.

[0018] Figure 3 is a structural framework schematic diagram of an embodiment of the medical question and answer decision support device of the present application. DETAILED DESCRIPTION

[0019] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be described clearly and completely below with reference to the drawings, obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0020] The terms "comprising" and "having" and any variations thereof as mentioned in the embodiments of the present application are intended to cover the inclusion not the exclusion of any steps or elements listed. For example, a process, method, system, product or apparatus that comprises a list of steps or elements is not limited to the listed steps or elements but can optionally further include other steps or elements not listed or can optionally further include steps or elements inherent to such process, method, product or apparatus.

[0021] The term "plurality" herein means two or more than two. In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more features. In the description of the present application, the meaning of "plurality" is two or more than two, unless otherwise specifically limited.

[0022] In order to facilitate the understanding of the embodiments, first, a medical question and answer decision support method disclosed by the embodiments of the present application will be described in detail. As shown in Figure 1 Figure 1 is a flowchart of an embodiment of the medical question and answer decision support method of the present application, and the method comprises the following steps.

[0023] Step S11: In response to receiving a medical question input by a user, performing natural language processing on the medical question to generate a structured query.

[0024] ​Specifically, the medical question can include a user's inquiry about symptoms, diseases, drugs, treatment plans, etc. through text or voice input. Natural language processing can include processing steps such as word segmentation, part-of-speech tagging, named entity recognition, and syntax analysis. Structured queries can include SPARQL (SPARQL Protocol and RDF Query Language) query statements, SQL (Structured Query Language) query statements, or other knowledge graph or database-based query language forms.

[0025] In some embodiments, the natural language processing module can be configured with BERT (Bidirectional Encoder Representations from Transformers), BiLSTM-CRF (Bidirectional Long Short-Term Memory with Conditional Random Field), and other models for entity recognition and relationship extraction.

[0026] In some application scenarios, a user inputs "How should a diabetic control blood sugar?", and after natural language processing, the medical entities "diabetes" and "blood sugar" are identified, the attribute "control" is identified, and a corresponding structured query statement is generated, such as a SPARQL statement for querying a medical knowledge graph.

[0027] Step S12: Real-time acquisition of system running state parameters, including multiple types of knowledge base query response time, current number of concurrent requests, user historical interaction behavior data, medical data source availability state, and model inference resource occupation rate.

[0028] Specifically, system running state parameters can be periodically or event-drivenly collected by system built-in monitoring agents or external monitoring tools. The knowledge base query response time can refer to the time interval experienced from sending a query to receiving a complete response. The current number of concurrent requests can refer to the total number of user requests currently being processed by the system. User historical interaction behavior data can include user click stream, session length, query history, and other information. The medical data source availability state can indicate whether various databases or API services can be normally accessed. The model inference resource occupation rate can include the current usage proportion of CPU, GPU, memory, and other hardware resources.

[0029] In some embodiments, the state monitoring module can implement metric collection and visualization through monitoring tools such as Prometheus, Grafana, etc.

[0030] In some application scenarios, the system running state parameters can include: the average response time of the knowledge base is 350 ms, the current number of concurrent requests is 125, the user recently frequently queries cardiovascular diseases, the status of the medical image data source is normal, and the GPU occupancy rate is 78%.

[0031] Step S13: In response to successfully obtaining the feedback result of the knowledge base for the structured query in the current processing period, generating first response content based on the current system running state parameters and the feedback result through a first prediction model.

[0032] Specifically, the current processing period can be a fixed time window, for example, 500 ms. Successfully obtaining the feedback result can mean receiving complete and valid answer data returned by the knowledge base within the period. The first prediction model can be a regression or classification model based on a neural network, the input of which can include normalized system state features and query result features, and the output of which can include recommended answer content or generation strategy.

[0033] In some embodiments, feature fusion can be achieved through splicing, attention mechanism or feature weighting, etc.

[0034] In some application scenarios, if the knowledge base returns "diabetes blood sugar control methods include diet adjustment, exercise and drug intervention", and the current system load is low, the first prediction model can generate a detailed point-by-point answer; if the system load is high, a concise summary answer can be generated.

[0035] Step S14: In response to not successfully obtaining the feedback result of the knowledge base in the current processing period, generating second response content based on the current system running state parameters through a second prediction model. The first response content or the second response content is output as the decision support result for the medical problem.

[0036] The first response content or the second response content is output as the decision support result for the medical problem.

[0037] Specifically, not successfully obtaining the feedback result can include timeout, network error, knowledge base non-response, etc. The second prediction model can be a lightweight model or a rule-based alternative system, which only depends on system state parameters to generate response content with a certain credibility, such as returning cache answers, general suggestions or prompting to retry later.

[0038] In some embodiments, the second prediction model can be a decision tree, a random forest, or a lookup table based on historical state-response mappings.

[0039] In some application scenarios, if the knowledge base query times out and the system concurrency is high, the second prediction model can return: "The current system is busy, please try to simplify your question or try again later"; if it is detected that the medical data source is unavailable, it can return: "The related data source is temporarily under maintenance, please consult a clinician".

[0040] It should be noted that the "predetermined state change threshold" described above can be set according to actual system configuration and performance requirements. In one embodiment, the state change threshold can be set to: the number of concurrent requests increases by more than 50% within 100 milliseconds, and / or the average response time of the knowledge base query increases by more than 100 milliseconds. Those skilled in the art can adjust the threshold according to the system tolerance. The above values are only examples and should not be construed as limiting the scope of the present application.

[0041] In some embodiments, in response to not successfully obtaining the feedback result of the knowledge base within the current processing period, a second response content is generated based on the current system running state parameter through a second prediction model, including: if the feedback result of the knowledge base is not successfully obtained within the current processing period, and the change of the system running state parameter before the end of the current processing period exceeds a predetermined state change threshold, then a second response content is generated based on the current system running state parameter through a second prediction model; if the feedback result of the knowledge base is not successfully obtained until the end of the current processing period, then a second response content is generated based on the current system running state parameter through a second prediction model.

[0042] Specifically, the predetermined state change threshold can include a change amplitude limit of one or more system running state parameters. For example, if the number of concurrent requests increases by more than 20% within 50ms, or the response time of the knowledge base query increases by more than 100ms within 100ms, it can be considered as exceeding the state change threshold. This can be used to trigger the backup response mechanism in advance when the system state fluctuates significantly, without having to wait for the processing period to end completely.

[0043] In some embodiments, the state monitoring module can configure a sliding time window (such as 100ms) to calculate the change rate of each state parameter in real time. If the change rate of any parameter exceeds its corresponding threshold, the state is immediately marked as abnormal. The second prediction model can receive the current parameter values and output a backup response suitable for the instantaneous state. This model can be trained based on successful response records under abnormal system state in history.

[0044] In some application scenarios, assume that the current processing cycle is set to 800 ms. At 500 ms of the cycle, it is monitored that the number of concurrent requests increases from 100 to 180 (exceeding the preset "50% increase" threshold), and the knowledge base response time is extended from 200 ms to 500 ms. At this time, the system can judge that the state mutates, and immediately call the second prediction model. Based on the current state of high concurrency and high delay, the model generates a response: "The current query pressure is large, and a brief suggestion is provided for you in priority: pay attention to diet and exercise for blood sugar control of diabetes. It is recommended to obtain detailed guidance again later." If no response from the knowledge base is received until the end of the 800 ms cycle, another alternative response content is generated based on the current state (concurrent number, delay, etc.).

[0045] In some embodiments, before the step of performing natural language processing on the medical question to generate a structured query, further comprising: obtaining an initial estimated processing delay based on the current system running state parameters through a second prediction model; and setting a waiting timeout time for establishing a connection with the knowledge base by using the initial estimated processing delay.

[0046] Specifically, the initial estimated processing delay can refer to the prediction of the time that the system may consume for this interaction before actually interacting with the knowledge base. This prediction can be based on the system running state parameters collected at the current time, and calculated through a pre-trained prediction model (an application instance of the second prediction model). The obtained estimated delay value can be used to dynamically configure the timeout waiting time (Timeout Duration) when the system establishes a network connection with the knowledge base or initiates a query request.

[0047] In some specific embodiments, the second prediction model for predicting the initial processing delay can be a lightweight machine learning model, such as a Gradient Boosting Decision Tree (GBDT) or a Support Vector Regression (SVR). The input features of the model can include but are not limited to: current number of concurrent requests, historical average response time at the same period, medical data source availability state index, and current queue length of the model inference service. The output of the model can be a continuous numerical value (unit: ms) representing the predicted delay. After the system calls this model, the returned predicted delay value can be multiplied by a safety factor (for example, 1.5) as the waiting timeout time for this interaction with the knowledge base.

[0048] In some application scenarios, it is assumed that the system initiates a new query process. Before performing natural language processing in step S11, the state monitoring module reports the current system running state parameters as: concurrent request number = 85, knowledge base cluster state = good, and GPU memory occupancy rate = 65%. The system calls the time delay prediction function of the second prediction model, and the model predicts that the expected processing time delay of this query is about 420 milliseconds based on the current input. Subsequently, the system sets the timeout time for establishing a connection with the knowledge base and waiting for its response to be 420 * 1.5 = 630 milliseconds. This means that if no valid response from the knowledge base is received within 630 milliseconds, the system will determine that this query is timed out, and may trigger the backup response generation process of step S14 in advance without waiting for a fixed and possibly longer processing period to end. In this way, the timeout time can be personalized and dynamically set, reducing unnecessary waiting when the system load is light, and reserving more sufficient response time when the load is heavy or the state is poor, thereby optimizing the user experience and system resource utilization as a whole.

[0049] In one specific embodiment, the second prediction model used to predict the initial processing time delay is a lightweight Gradient Boosting Decision Tree (GBDT) model. The model is trained using historical data, and its input features include: the current number of concurrent requests, the average response time of knowledge base queries in the past one minute, a medical data source availability status flag (1 for available and 0 for unavailable), and the current task queue length of the model inference service. The output is a continuous value representing the predicted time delay (unit: milliseconds). The system regularly (e.g., every 5 minutes) updates the model online using the latest system log data to adapt to the dynamic changes in system load.

[0050] In some embodiments, in response to receiving the user input medical question, performing natural language processing on the medical question to generate a structured query includes: in response to receiving the user input medical question, identifying medical entities and attributes in the medical question using named entity recognition technology; determining the mutual relationship between the identified medical entities and attributes based on a pre-constructed medical knowledge graph; and generating a structured query statement for querying the knowledge graph according to the medical entities, attributes, and mutual relationships.

[0051] In particular, the Named Entity Recognition (NER) technique can be used to identify and extract medical-related entities with specific meanings from unstructured user input text. The medical entities can include disease names, clinical symptoms, drug names, examination items, body parts, medical operations, etc. The attributes can include modification information of the entities, such as degree (e.g., “severe” in “severe pain”), frequency (e.g., “daily” in “three times a day”), state (e.g., “postoperative”), time (e.g., “for a week”), etc. The pre-constructed medical knowledge graph can be a large knowledge base stored in a graph structure, containing a large number of medical concepts and their semantic relationships (e.g., “disease-symptom”, “drug-treat disease”, “examination-diagnose disease”). The process of generating a structured query statement can be to map the identified entities and attributes into standardized query conditions according to the schema or ontology defined in the knowledge graph, and combine them into a query language (e.g., SPARQL, Cypher) statement that the knowledge graph query engine can understand and execute.

[0052] In some embodiments, named entity recognition can be implemented with sequence labeling tools (e.g., spaCy, StanfordNLP) based on deep learning models (e.g., BERT, BiLSTM-CRF). The construction of the knowledge graph can be based on medical standard terminology sets (e.g., SNOMED CT (Systematized Nomenclature of Medicine -- Clinical Terms), UMLS (Unified Medical Language System)) and clinical guidelines, using graph databases (e.g., Neo4j, Nebula Graph, etc.) for storage and management. The relationship determination module can determine the most likely mutual relationship by calculating the co-occurrence frequency, path distance, or embedding-based similarity of entities in the knowledge graph. The structured query statement generator can be a template or rule-based program that fills the identified entity IDs, attribute values, and determined relationship types into predefined query templates.

[0053] In some embodiments, based on the current system running state parameters and the feedback results, the first response content is generated by a first prediction model, including: based on the feedback results, determining the actual response time and result data volume of this query; based on the current system running state parameters, the actual response time and the result data volume, generating the first response content by the first prediction model.

[0054] In particular, the actual response time can refer to the time interval (usually in milliseconds) experienced from sending a structured query to the knowledge base, to completely receiving and parsing the returned feedback results. The result data volume can refer to the size of the result set returned by this query (e.g., measured in bytes or number of records). The system running state parameters, the actual response time, and the result data volume can collectively serve as input features of the first prediction model, which can output recommended response content or generation strategies (e.g., deciding the level of detail of the answer, whether to include additional explanations, whether to suggest further operations, etc.).

[0055] In some embodiments, the first prediction model can be a multi-task learning model based on a deep neural network (such as LSTM or Transformer), whose input layer can receive a normalized multi-dimensional feature vector, and whose output layer can produce a semantic representation of the response content or directly generate natural language text. The model training process can use historical interaction logs, where the input features include historical system states and query response indicators, and the training labels can be the best response manually annotated or the strategy optimized based on user feedback (such as satisfaction scores).

[0056] In some application scenarios, suppose a user query is about "treatment options for children with allergic rhinitis", and the knowledge base successfully returns feedback results containing 5 treatment suggestions, totaling 1024 bytes, with an actual response time of 280 milliseconds. The current system running state parameters are: knowledge base query response time (average of the last 5 times) = 250 milliseconds, current number of concurrent requests = 88, medical data source availability status = all normal, model inference resource occupancy rate = 65%. After receiving these inputs, the first prediction model may, after calculation, generate a first response content that integrates the 5 treatment suggestions and adds a disclaimer stating "the above suggestions should be used in conjunction with a clinical doctor's diagnosis". If the actual response time for this query is longer (e.g., 800 milliseconds) and the result data volume is large (e.g., 10 records), the model may generate a more concise summary answer to improve information delivery efficiency.

[0057] In some embodiments, based on the current system running state parameters, the actual response time, and the result data volume, the first response content is generated by the first prediction model, including: normalizing the current system running state parameters to extract system state features; normalizing the actual response time and the result data volume to extract network transmission and computing features of this query; fusing the system state features and the network transmission and computing features to obtain target fusion features; and generating the first response content based on the target fusion features through the first prediction model.

[0058] In particular, the normalization process can include a procedure of mapping original parameters of different dimensions and ranges to a uniform numerical interval (e.g., [0, 1] or [-1, 1]) through linear or nonlinear transformation. The system state feature can be a low-dimensional representation reflecting the overall load and health of the system, extracted from the normalized system running state parameters. The network transmission and computing feature can be a feature reflecting the efficiency and resource consumption of this query execution, extracted from the normalized actual response time and result data volume. Feature fusion can be a combination, concatenation or weighted summation of feature vectors from different sources to form a unified feature representation containing multi-dimensional information. The target fusion feature will serve as the input of the first prediction model to generate the final first response content.

[0059] In some embodiments, the normalization of system running state parameters can adopt the Min-Max normalization method, such as dividing the number of concurrent requests by the maximum supported concurrency of the system (e.g., 200) and dividing the resource occupancy rate by 100. The actual response time can be normalized by dividing by the historical maximum response time (e.g., 1000 ms), and the result data volume can be normalized by dividing by the maximum data volume allowed to be returned by a single query (e.g., 10 KB). Feature extraction can be achieved through a fully connected neural network layer or principal component analysis to reduce multiple normalized parameters to a fixed-length feature vector. Feature fusion can use a simple vector concatenation method, or input two feature vectors into an attention network to learn the weighted fused representation.

[0060] In some application scenarios, assuming that the current system running state parameters are: number of concurrent requests = 120, maximum supported concurrency = 200, normalized to 0.6; GPU occupancy rate = 75%, normalized to 0.75; medical data source availability status can be represented by a binary flag (available as 1, unavailable as 0). Actual response time = 320 ms, historical maximum response time = 1000 ms, normalized to 0.32; result data volume = 8 KB, maximum allowed data volume = 10 KB, normalized to 0.8. System state feature extraction can result in a 3-dimensional vector [0.72, 0.68, 0.95], and network transmission and computing feature extraction can result in a 2-dimensional vector [0.35, 0.82]. The fused target fusion feature can be a 5-dimensional vector [0.72, 0.68, 0.95, 0.35, 0.82]. After receiving this vector, the first prediction model (e.g., a neural network) outputs the final generated first response content, such as a reply text with moderate detail level, containing 5 treatment suggestions and accompanying precautions.

[0061] In some embodiments, the first prediction model is trained using a plurality of first training samples, the first training samples including, as training data, system running state parameters and knowledge base feedback results of a historical period, and, as training labels, response content generation time delays or optimal response strategies of the historical period; and the second prediction model is trained using a plurality of second training samples, the second training samples including, as training data, system running state parameters of a historical period, and, as training labels, response content generation time delays or optimal response strategies of the historical period.

[0062] Specifically, the construction process of the first training samples and the second training samples can include extracting relevant data records from system historical running logs. For the first training samples, each training data can include system running state parameters (such as the number of concurrent requests, resource occupancy, etc.) collected in a historical processing period and feedback results (such as query result data, response status code) returned by the knowledge base in the period, and the corresponding training label can be the time (i.e., response content generation time delay) actually measured from obtaining the feedback to generating the final response in the historical period, or the most appropriate response strategy category (i.e., optimal response strategy) annotated by an expert for the query. For the second training samples, each training data can include system running state parameters collected in a historical period (no knowledge base feedback result because the feedback is not successfully obtained), and the training label can also be the standby response generation time delay actually measured in the period or the standby response strategy annotated by an expert. The goal of model training is to learn the mapping relationship from the input features to the output labels.

[0063] In some specific embodiments, the training data of the first prediction model (such as an XGBoost regression model) can be a feature matrix, each row of which represents a historical sample, the column features include normalized historical concurrency, CPU occupancy, knowledge base response time, returned result data volume, etc., and the label can be a continuous value time delay (unit: millisecond) or a discrete value strategy ID. The training data feature matrix of the second prediction model (such as a random forest classifier) only includes system running state parameters, and the label is a discrete standby strategy ID. The training process can be completed by minimizing the mean square error of the predicted time delay and the real time delay (for regression tasks), or minimizing the cross-entropy loss of the predicted strategy and the real strategy (for classification tasks).

[0064] In some application scenarios, suppose a record of a successful query is selected from historical logs: at a certain moment, the system state is concurrency = 90, GPU occupancy = 60%, the knowledge base returns 5 suggestions about "diabetic diet" (data volume is about 2KB), and the actual generation of detailed answers takes 450ms. This can constitute a first training sample: the training data is [90, 60%, 5, 2KB] (after normalization), and the training label is 450 (regression task) or "strategy ID: 3" (classification task, ID = 3 represents generating detailed answers). Another historical record shows that at a certain moment, the knowledge base feedback was not successfully obtained due to network jitter, at which time the system state was concurrency = 150, GPU occupancy = 85%, and the system finally generated a cache answer and took 120ms. This can constitute a second training sample: the training data is [150, 85%] (after normalization), and the training label is 120 (regression task) or "strategy ID: 1" (classification task, ID = 1 represents returning a cache answer). Through the training of a large number of such samples, the first prediction model can learn to generate detailed answers when the system load is moderate and there is knowledge base result, and the second prediction model can learn to quickly return a cache answer when the load is high and there is no result.

[0065] In the above scheme, in response to receiving a user input medical question, the medical question is subjected to natural language processing to generate a structured query; real-time system running state parameters are obtained, including knowledge base query response time, number of concurrent requests, user historical behavior, data source availability, and inference resource occupancy rate, and the like; if the knowledge base feedback result is successfully obtained within the current processing period, the first prediction model is used to generate first answer content based on the current system state and the feedback result; if the feedback is not successfully obtained, the second prediction model is used to generate second answer content based on the system state; and finally, either result is output as decision support for the user question. In this way, by real-time monitoring of multi-dimensional system state and combining a double-model redundant decision mechanism, the generation path of the answer is dynamically selected, effectively dealing with high concurrency access, data source heterogeneity, and service uncertainty, and ensuring the continuity of system service, the efficiency of response, and the credibility of the result.

[0066] Please refer to Figure 2 , Figure 2 is a structural framework schematic diagram of an embodiment of the medical question and answer decision support system of the present application. As shown in Figure 2As shown, the medical question and answer decision support system 20 comprises a natural language processing module 21, a state monitoring module 22, a first answer generation module 23, a second answer generation module 24 and a result output module 25. The natural language processing module 21 is configured to, in response to receiving a medical question input by a user, perform natural language processing on the medical question to generate a structured query; the state monitoring module 22 is configured to acquire system running state parameters in real time, the system running state parameters comprising multiple types of parameters including knowledge base query response time, current number of concurrent requests, historical interaction behavior data of the user, availability state of medical data sources and model inference resource occupation rate; the first answer generation module 23 is configured to, in response to successfully acquiring feedback results of the knowledge base for the structured query within a current processing period, generate first answer content based on the current system running state parameters and the feedback results by using a first prediction model; the second answer generation module 24 is configured to, in response to failing to successfully acquire the feedback results of the knowledge base within the current processing period, generate second answer content based on the current system running state parameters by using a second prediction model; and the result output module 25 is configured to output the first answer content or the second answer content as a decision support result for the medical question.

[0067] In some embodiments, the second answer generation module 24 is specifically configured to, in response to failing to successfully acquire the feedback results of the knowledge base within the current processing period, generate the second answer content based on the current system running state parameters by using the second prediction model, including: if the feedback results of the knowledge base are not successfully acquired within the current processing period and the system running state parameters change by more than a predetermined state change threshold before the end of the current processing period, generating the second answer content based on the current system running state parameters by using the second prediction model; and if the feedback results of the knowledge base are not successfully acquired until the end of the current processing period, generating the second answer content based on the current system running state parameters by using the second prediction model.

[0068] In some embodiments, the system further comprises, before the step of performing natural language processing on the medical question to generate a structured query, the steps of: obtaining an initial estimated processing delay by using the second prediction model based on the current system running state parameters; and setting a waiting timeout time for establishing a connection with the knowledge base by using the initial estimated processing delay.

[0069] In some embodiments, the natural language processing module 21 is specifically configured to, in response to receiving a medical question input by a user, perform natural language processing on the medical question to generate a structured query, including: in response to receiving the medical question input by the user, identifying medical entities and attributes in the medical question by using a named entity recognition technology; determining mutual relationships between the identified medical entities and attributes based on a pre-constructed medical knowledge graph; and generating a structured query statement for querying the knowledge graph according to the medical entities, the attributes and the mutual relationships.

[0070] In some embodiments, the first response generation module 23 is specifically configured to generate the first response content based on the current system running state parameter and the feedback result through the first prediction model, including: determining the actual response time and the result data volume of the current query based on the feedback result; and generating the first response content through the first prediction model based on the current system running state parameter, the actual response time and the result data volume.

[0071] In some embodiments, the first response generation module 23 is specifically configured to generate the first response content based on the current system running state parameter, the actual response time and the result data volume through the first prediction model, including: performing normalization processing on the current system running state parameter to extract system state features; performing normalization processing on the actual response time and the result data volume to extract network transmission and computing features of the current query; fusing the system state features and the network transmission and computing features to obtain target fusion features; and generating the first response content through the first prediction model based on the target fusion features.

[0072] In some embodiments, the system, in the case of successfully obtaining the feedback result, the first prediction model is trained by using a plurality of first training samples, the first training samples include the system running state parameter and the knowledge base feedback result of a historical period as training data, and the response content generation delay or the best response strategy of the historical period as training labels; in the case of unsuccessfully obtaining the feedback result, the second prediction model is trained by using a plurality of second training samples, the second training samples include the system running state parameter of a historical period as training data, and the response content generation delay or the best response strategy of the historical period as training labels.

[0073] In the above scheme, the structured query is generated by performing natural language processing on the medical problem input by the user in response to receiving the input; the system running state parameters are obtained in real time, including knowledge base query response time, concurrent request number, user historical behavior, data source availability and inference resource occupation rate and other indicators; if the knowledge base feedback result is successfully obtained within the current processing period, the first response content is generated based on the current system state and the feedback result through the first prediction model; if the feedback is not successfully obtained, the second response content is generated based on the system state through the second prediction model; finally, any result is output as decision support for the user's question. In this way, by real-time monitoring of multi-dimensional system state and combining the double model redundant decision mechanism, the response generation path is dynamically selected, effectively dealing with high concurrency access, data source heterogeneity and service uncertainty, and ensuring the continuity of system service, the efficiency of response and the credibility of result.

[0074] Figure 3is a schematic diagram of a structural framework of an embodiment of the medical question and answer decision support device. The medical question and answer decision support device 30 can vary greatly in configuration or performance, and can include one or more processors 31 and a memory 32, the processor 31 can be configured to communicate with the memory 32, and execute a series of instructions stored in the memory on the medical question and answer decision support device to implement the steps of the medical question and answer decision support method described above. Those skilled in the art can understand that, Figure 3 The medical question and answer decision support device structure shown does not constitute a limitation on the medical question and answer decision support device provided by the present application, and can include more or fewer components than shown, or combine certain components, or different component arrangements.

[0075] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium, or a volatile computer readable storage medium, and the computer readable storage medium stores instructions, which, when executed on a computer, cause the computer to perform the steps of the medical question and answer decision support method.

[0076] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working process of the system or system, unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0077] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art, or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the embodiments of the method of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0078] The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A medical question-and-answer decision support method, characterized in that, The method includes: in response to receiving a medical question input by a user, performing natural language processing on the medical question to generate a structured query; and acquiring system operating status parameters in real time, including multiple parameters such as knowledge base query response time, current concurrent request count, user historical interaction behavior data, medical data source availability status, and model inference resource utilization rate. In response to successfully obtaining the feedback result of the knowledge base for the structured query within the current processing cycle, a first response content is generated through a first prediction model based on the current system operating status parameters and the feedback result; specifically, this includes: determining the actual response time and result data volume of this query based on the feedback result; and generating the first response content through the first prediction model based on the current system operating status parameters, the actual response time, and the result data volume. In response to the failure to obtain feedback from the knowledge base within the current processing cycle, a second response is generated based on the current system operating status parameters using a second prediction model. Specifically, this includes: if the feedback from the knowledge base is not successfully obtained within the current processing cycle, and the change in the system operating status parameters exceeds a predetermined state change threshold before the end of the current processing cycle, then a second response is generated based on the current system operating status parameters using the second prediction model; if the feedback from the knowledge base is still not successfully obtained by the end of the current processing cycle, then a second response is generated based on the current system operating status parameters using the second prediction model. The first response content or the second response content is output as a decision support result for the medical problem.

2. The medical question-and-answer decision support method according to claim 1, characterized in that, Before the step of performing natural language processing on the medical problem to generate a structured query, the method further includes: obtaining an initial estimated processing delay based on the current system operating status parameters through the second prediction model; and using the initial estimated processing delay to set a timeout for establishing a connection with the knowledge base.

3. The medical question-and-answer decision support method according to claim 1, characterized in that, The step of responding to a medical question input by a user by performing natural language processing on the medical question to generate a structured query includes: responding to a medical question input by a user by using named entity recognition technology to identify medical entities and attributes in the medical question; determining the relationships between the identified medical entities and attributes based on a pre-built medical knowledge graph; and generating a structured query statement for querying the knowledge graph based on the medical entities, attributes, and relationships.

4. The medical question-and-answer decision support method according to claim 3, characterized in that, The step of generating a first response content based on the current system operating status parameters, the actual response time, and the amount of result data through the first prediction model includes: normalizing the current system operating status parameters and extracting system status features; normalizing the actual response time and the amount of result data and extracting network transmission and computation features for this query; fusing the system status features and the network transmission and computation features to obtain target fusion features; and generating the first response content based on the target fusion features through the first prediction model.

5. The medical question-and-answer decision support method according to claim 1, characterized in that, If the feedback result is successfully obtained, the first prediction model is trained using multiple first training samples. The first training samples include system operation status parameters and knowledge base feedback results from historical periods as training data, and response content generation delay or optimal response strategy from historical periods as training labels. If the feedback result is not successfully obtained, the second prediction model is trained using multiple second training samples. The second training samples include system operating status parameters of historical periods as training data, and response content generation delay or optimal response strategy of the historical periods as training labels.

6. A medical question-and-answer decision support system, characterized in that, The system includes: a natural language processing module, used to respond to a medical question input by a user, perform natural language processing on the medical question, and generate a structured query; and a status monitoring module, used to acquire system operating status parameters in real time, including multiple parameters such as knowledge base query response time, current concurrent request count, user historical interaction behavior data, medical data source availability status, and model inference resource utilization rate. The first response generation module is used to respond to the successful acquisition of the feedback result of the knowledge base for the structured query within the current processing cycle, and to generate the first response content through the first prediction model based on the current system operating status parameters and the feedback result; specifically, it includes: determining the actual response time and result data volume of this query based on the feedback result; and generating the first response content through the first prediction model based on the current system operating status parameters, the actual response time, and the result data volume. The second response generation module is used to generate a second response based on the current system operating status parameters and a second prediction model in response to the failure to obtain feedback results from the knowledge base within the current processing cycle. The result output module is used to output the first response content or the second response content as a decision support result for the medical problem. Specifically, it includes: if the feedback results from the knowledge base are not successfully obtained within the current processing cycle, and the change in the system operating status parameters exceeds a predetermined state change threshold before the end of the current processing cycle, then the second response content is generated based on the current system operating status parameters and the second prediction model; if the feedback results from the knowledge base are still not successfully obtained until the end of the current processing cycle, then the second response content is generated based on the current system operating status parameters and the second prediction model.

7. A medical question-and-answer decision support device, characterized in that, The medical question-and-answer decision support device includes: a memory and at least one processor, the memory storing instructions; the at least one processor invokes the instructions in the memory to cause the medical question-and-answer decision support device to perform the steps of the medical question-and-answer decision support method as described in any one of claims 1-5.

8. A computer-readable storage medium storing instructions thereon, characterized in that, When the instruction is processed and executed, it implements the steps of the medical question-answering decision support method as described in any one of claims 1-5.

Images