AI nursing whiteboard system and method based on large model agent

Abstract

The invention discloses an AI nursing whiteboard system and method based on a large model agent, the system comprises a nursing whiteboard terminal, an AI intelligent chain platform and a database, the nursing whiteboard terminal comprises an AI nursing assistant, and a whiteboard control system is arranged in the nursing whiteboard terminal. Through the functions of integration of multi-source data, intelligent early warning, accurate question and answer and the like, various diagnosis and treatment data, such as vital signs, medical advices, nursing records and the like, of a patient can be rapidly integrated and presented in a clear and visual mode, a large amount of time and energy are saved, the patient can rapidly know the condition of the patient, and the diagnosis and treatment efficiency is improved. Vital signs and illness state changes of a patient can be monitored in real time, early warning can be immediately given out once abnormal conditions such as heart rate and blood pressure exceeding normal ranges occur, a nurse can know the abnormal conditions in the first time and take corresponding measures, the response speed to emergency situations of the patient is greatly increased, and when the nurse encounters a question in the aspect of medical knowledge, the nurse can timely take corresponding measures. Inquiry may be made by an AI care assistant.

Description

Technical Field

[0001] The present invention relates to the technical field of nursing whiteboards, and in particular to an AI nursing whiteboard system and method based on a large model intelligent agent. Background Art

[0002] Before the emergence of intelligent nursing whiteboard systems, traditional ward nursing whiteboards relied primarily on manual operation, which had many limitations. Data queries required switching between paper documents and multiple systems, which was inefficient. Furthermore, the whiteboards lacked intelligent functions, making it difficult to monitor and analyze patient data in real time and unable to proactively warn of abnormal situations. Furthermore, information transmission mostly relied on verbal communication, lacking a unified and efficient linkage mechanism. This led to poor communication between doctors, nurses, and patients, impacting the quality of nursing services and the patient experience.

[0003] In summary, the existing nursing whiteboards have the following shortcomings: (1) Single retrieval method: With static writing as the only display method, if nurses need to find detailed information of a specific patient or compare data of different patients, they can only search manually line by line and word by word. There is a lack of convenient interactive functions such as retrieval and screening. This method is not only inconvenient to operate, but also wastes a lot of time and cannot meet the needs of quickly obtaining information in emergency situations; (2) Lack of early warning capabilities: There is no function to automatically monitor the patient's vital signs data (such as body temperature, heart rate, blood pressure, etc.), and it is impossible to actively identify data anomalies. Only when nurses conduct regular inspections, manually record and compare and analyze data can they discover the risks of patients. This increases the risk of worsening of the patient's condition and makes it difficult to ensure patient safety; (3) Insufficient data analysis: It is just a simple display of patient information. It is impossible to integrate, count and deeply analyze the data. It is impossible to mine potential laws from a large amount of data, such as the development trend of the patient's condition, the evaluation of the effectiveness of nursing measures, etc., resulting in a lack of scientific data support for nursing decisions and difficulty in achieving personalized and precise nursing services. In summary, this application proposes an AI nursing whiteboard system and method based on a large model intelligent agent. Summary of the Invention

[0004] Based on the technical problems existing in the background technology, the present invention proposes an AI nursing whiteboard system and method based on a large model intelligent agent.

[0005] The present invention proposes an AI nursing whiteboard system based on a large model intelligent agent, including a nursing whiteboard terminal, an AI smart chain platform and a database. The nursing whiteboard terminal includes an AI nursing assistant and a built-in whiteboard control system.

[0006] The AI ​​Zhilian platform includes a voice interaction module, an intelligent decision-making engine module, a business data query module, a medical knowledge question-and-answer module, and a system monitoring and optimization module;

[0007] The database includes a basic information configuration library, a data structure knowledge library and a smart ward database.

[0008] Preferably, the voice interaction module is used to initiate interaction with the nurse and to load the guide language on the front end. When initiating interaction with the nurse, the nurse establishes a connection with the nursing whiteboard terminal through the handheld device using Bluetooth and Wi-Fi wireless communication technologies, and wakes up the AI ​​nursing assistant of the nursing whiteboard through the intercom function of the handheld device, accurately transmits the wake-up command to the whiteboard control system, and triggers the startup of the AI ​​nursing assistant;

[0009] When the guide words are loaded on the front end, the AI ​​nursing assistant front-end page loads and displays the question-and-answer guide words based on the background configuration. At the same time, you can choose whether to perform voice broadcasting, read the guide word configuration information from the background, complete the integration of the voice broadcast function through the speech synthesis SDK, and decide whether to speak based on the configuration switch.

[0010] Preferably, the intelligent decision engine module is used to process and classify historical data, creating two engines for medical knowledge question and answer and business knowledge question and answer, so as to correctly distribute the question and answer questions received by the voice interaction module;

[0011] The specific logical steps are as follows:

[0012] (1) Data cleaning and structuring: remove duplicate data and non-medical related text, use regular expressions to filter special symbols, and retain the core business fields of "department + drug + operation". For example, "

Emergency

[0013] (2) Pre-training model selection and fine-tuning: Based on the BERT model pre-trained in general fields, the hospital's internal electronic medical record text is loaded for domain transfer learning, focusing on optimizing the semantic understanding of the professional terms "creatinine value" and "nebulized inhalation". In the fine-tuning stage, high-frequency departments (such as "respiratory department" and "cardiology department") and drug names (such as "ambroxol" and "dopamine") in the business query data are input into the model as entity masks, forcing the model to learn the associated features of "department-drug-query type"; for example, the "cardiology department + metoprolol + inventory" pattern is used to identify the business query intent;

[0014] (3) Model architecture design, design of dual-channel feature fusion:

[0015] Two-channel embedding layer: Channel 1 is used to pre-train semantic features, freeze the parameters of the first six layers of BERT, extract common semantic vectors of medical terms, and identify knowledge keywords such as "taboo" and "combination". Channel 2 is for business dynamic features. It captures the combination pattern of "department + time + operation" through dynamic weights. The embedding layer is randomly initialized to learn dynamic parameters in business queries, including department name ("Emergency Department"), time words ("Today" and "Recent Week"), and operation words ("Query" and "Export").

[0016] The bidirectional LSTM layer receives the concatenated vector of the two channels (dimension = 2 × pre-trained vector dimension) and captures contextual dependencies. For example, when determining whether suctioning is necessary when the patient's blood oxygen saturation is below 90%, the BiLSTM layer extracts the temporal association between blood oxygen saturation, threshold, and nursing action.

[0017] Introducing an attention mechanism, forcing focus on the keywords "inventory," "trend," and "current" in business queries; focusing on the terms "taboo," "regulation," and "precautions" in medical knowledge queries. Using the attention weight matrix \(\alpha\), we quantify the contribution of each word to the classification (e.g., the weight of "inventory" in the business category is > 0.8).

[0018] (4) Business scenario-oriented optimization: perform sample weighting and data enhancement, give higher weights to samples of low-frequency departments in the business query category, avoid the model being biased towards high-frequency departments, and use EDA (Easy Data Augmentation) technology to perform synonym replacement and word order adjustment on medical knowledge texts, expand small sample category data, and use weighted cross entropy loss function. Set weights for business query category / medical knowledge category respectively according to the proportion of training set (such as 0.6 / 0.4), balance the difference in the number of samples of the two categories, and calculate the business category F1 and knowledge category F1 after each round of training. The F1 of both categories is required to be ≥92% to avoid overfitting of the model to one category;

[0019] (5) Confusion matrix driven optimization: Real-time confusion matrix analysis, deployment of online monitoring system, hourly statistics of confusion matrix, focusing on two types of errors, including business query → medical knowledge and medical knowledge → business query, and implementing incremental learning mechanism, summarizing classification error cases reported by clinical nurses every week, adding them to the training set after manual annotation, and fine-tuning the model through hot update mechanism to ensure the recognition accuracy of emerging business terms ≥95%.

[0020] Preferably, the business data query module performs business processing based on the business data query process divided by the voice interaction module and the intelligent decision engine module;

[0021] The specific logical steps are as follows:

[0022] (1) Hybrid database architecture design: Adapting to the characteristics of nursing business data: adopting a data layered storage strategy, the MySQL structured data layer is used to store standardized nursing records, including patient temperature sheets (fields: patient ID, department, measurement time, temperature value), doctor order execution tables (fields: doctor order ID, drug name, executing nurse, execution time), and defining the "patient-department-nursing operation" association relationship through the ER model, supporting multi-table JOIN complex queries (such as "querying the list of patients in the respiratory department with a temperature ≥ 38.5℃ and who have used ibuprofen in December 2023"). The Redis real-time cache layer is used to cache high-frequency query results (such as "the current number of inpatients in each department" and "common drug inventory thresholds"), setting a 5-minute expiration time to reduce the pressure on MySQL / MongoDB, and storing nurse login status and temporary query conditions (such as "locking the real-time vital sign data of the patient in bed 2");

[0023] (2) Text2SQL engine construction: A "department-data table-field" mapping dictionary is established through the medical Schema binding mechanism. When the nurse enters the query "FEV1 value trend of patients in bed 3 of the respiratory department in the past two days", the engine automatically matches the respiratory department schema, limits the query field to FEV1 value, and names the table "Respiratory Function Assessment Table for Patients in the Respiratory Department" to avoid cross-departmental unauthorized queries. It also designs an easy-to-verify layer and intercepts high-risk operations. It matches prohibited SQL patterns through regular expressions. If the intention of "full table query" is detected, it directly returns the prompt of "inadequate query permission" and forces the specification of specific fields. It also performs logical compliance checks and queries based on the query logic of nursing business rules. Its query logic includes "the time span of nursing record query shall not exceed 31 days". If the input is "query for nursing error records throughout 2023", it will be automatically split into 12 single-month queries to avoid excessive database pressure.

[0024] (3) After generating the SQL statement using the DeepSeek-V3 model, the confidence is calculated using the built-in scoring function. The formula used is: confidence = field matching × 0.6 + table association rationality × 0.3 + business rule compliance × 0.1; when the confidence is ≥ 80%, the SQL statement is executed directly; when the confidence is < 80%, the manual review process is triggered, and the nurse needs to supplement the patient's ID or name key information;

[0025] (4) Knowledge graph-driven precise query: Construct a medical knowledge graph, define entities and relationships, and perform vector retrieval and recall optimization. Embed knowledge graph entities into 300-dimensional vectors and store them in the Milvus vector database. When processing queries, first extract the keywords "ambroxol" and "nursing matters" through BERT, generate query vectors, and recall related entities in the graph. The recall rate is increased from 45% of traditional keyword matching to 80%;

[0026] (5) Perform RAG fusion. According to the query intent, relevant data fragments are recalled from MySQL / MongoDB or knowledge graphs, and the recalled data is combined with the SQL statements generated by DeepSeek-V3. First, the graph is used to find out whether the two are contraindicated in combination, and then SQL is used to retrieve specific patients. The final result is accompanied by knowledge prompts.

[0027] Preferably, the medical knowledge question and answer module and the business data query module are parallel modules, and also perform business processing based on the business data query process divided by the voice interaction module and the intelligent decision engine module. The specific logical steps are as follows:

[0028] (1) Question parsing and scenario recognition: Input preprocessing: receiving natural language questions input by nurses through nursing whiteboards or handheld terminals, removing punctuation and converting them to lowercase;

[0029] (2) Large model calling strategy: automatically load corresponding domain parameters according to the problem type;

[0030] (3) Convert the original question into a structured prompt and output the result.

[0031] Preferably, the system monitoring and optimization module is used for performance indicator monitoring and system optimization and adjustment. During performance indicator monitoring, the system's performance indicators are monitored in real time through server-integrated monitoring tools, including data collection frequency, early warning analysis delay, and message push success rate. Monitoring code is implanted in key system nodes to collect performance data and store it in the Prometheus database, and the data is displayed through the Grafana visualization interface to promptly identify system performance bottlenecks.

[0032] When optimizing and adjusting the system, the system is optimized based on performance monitoring data. If the data collection delay is found to be too high, the data collection algorithm is optimized or collection nodes are added. When the early warning analysis efficiency is low, the large model parameters are adjusted or the configuration of the stream computing framework is optimized. The system is regularly maintained and upgraded to ensure that the system is always in the best operating state and to ensure the efficient development of smart ward services.

[0033] The present invention also proposes an AI nursing whiteboard method based on a large model intelligent agent, comprising the following steps:

[0034] S1: The nurse wakes up the AI ​​nursing assistant through the intercom function of the handheld device;

[0035] S2: After receiving the signal, the nursing whiteboard terminal loads and displays the question-and-answer guidance;

[0036] S3: Users input questions in natural language. The intelligent decision engine module collects the input questions and performs data cleaning and structuring.

[0037] S4: The intelligent decision engine module uses the BERT model fine-tuned through transfer learning and the dual-channel BiLSTM-Attention model to determine whether the question belongs to "business inquiry" or "medical knowledge question answering," and performs entity recognition and intent classification.

[0038] S5: If it is a business data query problem: follow the business query path through the business data query module and perform the following steps:

[0039] S501: Convert natural language questions into SQL queries, automatically match the medical schema, bind fields to data tables, ensure department isolation and field legitimacy, and verify SQL compliance. Then, calculate the SQL confidence level. If the confidence level is ≥80%, the SQL statement is executed directly. If the confidence level is <80%, a manual review process is triggered, requiring the nurse to provide key patient information such as the patient's ID or name.

[0040] S502: The cached results are retrieved from Redis first. If a query fails, the query is returned to the source MySQL. The query results are displayed graphically on the nursing whiteboard terminal.

[0041] S503: RAG fusion enhanced output: Based on the query intent, relevant data fragments are recalled from MySQL / MongoDB or knowledge graph (medical knowledge), and the recalled data is combined with the SQL statement generated by DeepSeek-V3. First, the graph is used to determine whether the two are contraindicated. Then, SQL is used to search for specific patients, and the final result is accompanied by knowledge prompts.

[0042] S6: If the question is a medical knowledge question and answer question: the medical knowledge question and answer module follows the large model question and answer path, and performs the following steps:

[0043] S601: Use the domain dictionary and question template to determine whether it is a medical knowledge question (threshold ≥ 0.9);

[0044] S602: If yes, then enter the large model reasoning process and fine-tune the weights based on the problem domain by calling different domain knowledge;

[0045] S603: Generate a structured prompt containing query objectives, constraints, and output requirements;

[0046] S7: The system monitoring and optimization module uses (Prometheus + Grafana) to monitor the system's performance indicators in real time, including data collection frequency, warning analysis delay, and message push success rate, and regularly optimizes system parameters based on performance monitoring data.

[0047] Compared with the existing technology, the beneficial effects of the present invention are:

[0048] By integrating multi-source data, intelligent early warning, precise question and answer and other functions, the present invention can quickly integrate various types of patient diagnosis and treatment data, such as vital signs, doctor's orders, nursing records, etc., and present them in a clear and intuitive manner, saving a lot of time and energy, enabling it to understand the patient's condition more quickly, and can monitor the patient's vital signs and condition changes in real time. Once an abnormal situation occurs, such as heart rate and blood pressure exceeding the normal range, an early warning will be issued immediately, and the nurse can be informed and take corresponding measures at the first time, greatly improving the response speed to the patient's emergency. When the nurse encounters questions about medical knowledge, he or she can ask through the AI ​​nursing assistant, which improves the convenience of question and answer and helps to improve the accuracy of nursing decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a block diagram of an AI nursing whiteboard system based on a large model agent proposed by the present invention;

[0050] Figure 2 This is a flow chart of the intelligent decision engine module in the AI ​​nursing whiteboard system based on a large model agent proposed by the present invention;

[0051] Figure 3 This is a flowchart of a business data query module in an AI nursing whiteboard system based on a large model agent proposed by the present invention;

[0052] Figure 4 This is a flowchart of the medical knowledge question-answering module in the AI ​​nursing whiteboard system based on a large model agent proposed by the present invention;

[0053] Figure 5 This is a flowchart of the AI ​​nursing whiteboard method based on a large model intelligent agent proposed by the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further explained below with reference to specific embodiments.

[0055] Example

[0056] Reference Figure 1-5 ,This embodiment proposes an AI nursing whiteboard system based on a large model intelligent agent, including a nursing whiteboard terminal, an AI smart chain platform and a database. The nursing whiteboard terminal includes an AI nursing assistant, and the nursing whiteboard terminal has a built-in whiteboard control system;

[0057] The AI ​​Zhilian platform includes a voice interaction module, an intelligent decision-making engine module, a business data query module, a medical knowledge question-and-answer module, and a system monitoring and optimization module;

[0058] The voice interaction module is used to initiate interaction with the nurse and load the guidance words on the front end. When initiating interaction with the nurse, the nurse uses the handheld device to establish a connection with the nursing whiteboard terminal using Bluetooth and Wi-Fi wireless communication technologies, and wakes up the nursing whiteboard's AI nursing assistant through the intercom function of the handheld device. The wake-up command is accurately transmitted to the whiteboard control system, triggering the AI ​​nursing assistant to start;

[0059] When the front-end loads the guidance words, the AI ​​nursing assistant front-end page loads and displays the question-and-answer guidance words according to the background configuration. At the same time, you can choose whether to perform voice broadcast. The guidance words configuration information is read from the background, and the voice broadcast function is integrated through the speech synthesis SDK. The decision of whether to speak is based on the configuration switch.

[0060] The intelligent decision engine module processes and classifies historical data, creating two engines: medical knowledge Q&A and business knowledge Q&A. This allows for the correct distribution of Q&A questions from the voice interaction module.

[0061] The specific logical steps are as follows:

[0062] (1) Data cleaning and structuring: remove duplicate data and non-medical related text, use regular expressions to filter special symbols, and retain the core business fields of "department + drug + operation". For example, "

Emergency

[0063] (2) Pre-training model selection and fine-tuning: Based on the BERT model pre-trained in general fields, the hospital's internal electronic medical record text (including more than 300,000 medical records) is loaded for domain transfer learning, focusing on optimizing the semantic understanding of the professional terms "creatinine value" and "nebulized inhalation". In the fine-tuning stage, high-frequency departments (such as "respiratory department" and "cardiology department") and drug names (such as "ambroxol" and "dopamine") in the business query data are input into the model as entity masks, forcing the model to learn the association features of "department-drug-query type"; for example, the "cardiology department + metoprolol + inventory" pattern is used to identify the business query intent;

[0064] (3) Model architecture design, design of dual-channel feature fusion:

[0065] Two-channel embedding layer: Channel 1 is used to pre-train semantic features, freeze the parameters of the first six layers of BERT, extract common semantic vectors for medical terms (such as the molecular mechanism association features of "drug interaction"), and identify knowledge keywords such as "contraindications" and "combination use." Channel 2 is for business dynamic features. It uses dynamic weights to capture the combination pattern of "department + time + operation" (for example, "pediatrics + today + temperature trend" must belong to the business query category). The embedding layer is randomly initialized to learn dynamic parameters in business queries, including department name ("Emergency Department"), time terms ("today" and "past week"), and operation terms ("query" and "export").

[0066] The bidirectional LSTM layer receives the concatenated vector of the two channels (dimension = 2 × pre-trained vector dimension) and captures contextual dependencies. For example, when determining whether suctioning is necessary when the patient's blood oxygen saturation is below 90%, the BiLSTM layer extracts the temporal association between blood oxygen saturation, threshold, and nursing action.

[0067] Introducing an attention mechanism, forcing focus on the keywords "inventory," "trend," and "current" in business queries; focusing on the terms "taboo," "regulation," and "precautions" in medical knowledge queries. Using the attention weight matrix \(\alpha\), we quantify the contribution of each word to the classification (e.g., the weight of "inventory" in the business category is > 0.8).

[0068] (4) Business scenario-oriented optimization: perform sample weighting and data enhancement, give higher weights to samples of low-frequency departments (such as "burn department") in business query categories, avoid the model being biased towards high-frequency departments (such as "internal medicine"), and use EDA (Easy Data Augmentation) technology to perform synonym replacement (such as "joint use" → "combined use") and word order adjustment on medical knowledge texts, expand small sample category data, and use weighted cross entropy loss function to set weights (such as 0.6 / 0.4) for business query category / medical knowledge category according to the proportion of training set, balance the difference in the number of samples of the two categories, and calculate the business category F1 and knowledge category F1 after each round of training, requiring both categories F1 to be ≥92% to avoid overfitting of the model to one category;

[0069] (5) Confusion matrix driven optimization: Real-time confusion matrix analysis, deployment of online monitoring system, hourly statistics of confusion matrix, focusing on two types of errors, including business query → medical knowledge and medical knowledge → business query, and implementing incremental learning mechanism, summarizing classification error cases reported by clinical nurses every week, adding them to the training set after manual annotation, and fine-tuning the model through hot update mechanism to ensure the recognition accuracy of emerging business terms ≥ 95%;

[0070] For example, business query → medical knowledge, such as "Query the inventory of vancomycin in the ICU," was misclassified as knowledge. This may be because the model overweights the pharmacological properties of "vancomycin." Therefore, the attention weights of business keywords such as "inventory" and "ICU" need to be increased.

[0071] Medical knowledge → business query: For example, "blood sugar needs to be monitored after insulin injection" is misclassified as a query type, and feature extraction of modal words such as "need" and "standard" needs to be strengthened;

[0072] The business data query module performs business processing based on the business data query process divided by the voice interaction module and the intelligent decision engine module;

[0073] The specific logical steps are as follows:

[0074] (1) Hybrid database architecture design: Adapting to the characteristics of nursing business data: adopting a data layered storage strategy, the MySQL structured data layer is used to store standardized nursing records, including patient temperature sheets (fields: patient ID, department, measurement time, temperature value), doctor order execution tables (fields: doctor order ID, drug name, executing nurse, execution time), and defining the "patient-department-nursing operation" association relationship through the ER model, supporting multi-table JOIN complex queries (such as "querying the list of patients in the respiratory department with a temperature ≥ 38.5℃ and who have used ibuprofen in December 2023"). Its business scenario is: the head nurse regularly counts the compliance rate of nursing operations in each department, and calculates the compliance rate by linking the "nursing record" with the "operation specification table" through SQL statements;

[0075] The Redis real-time cache layer is used to cache high-frequency query results (such as "current number of patients in each department" and "common drug inventory thresholds"). It has a 5-minute expiration time to reduce the pressure on MySQL / MongoDB. It also stores nurse login status and temporary query conditions (such as "lock the real-time vital sign data of the patient in bed 2"). The business scenario is that when a nurse's handheld terminal requests patient vital sign data every second, the cached result is first obtained from Redis. If the result is expired, the cached result is returned to MySQL to query the latest data. This reduces the response latency from 200ms to 30ms.

[0076] (2) Text2SQL engine construction: Establish a "department-data table-field" mapping dictionary through the medical Schema binding mechanism, such as: Respiratory Department → Table: Patient Respiratory Function Assessment Table → Fields: FEV1 value, blood oxygen saturation, number of nebulized inhalations; Emergency Department → Table: Rescue Record → Fields: Rescue start time, use of emergency drugs, changes in vital signs;

[0077] When a nurse enters the query "FEV1 value trend of patients in bed 3 of the respiratory department in the past two days", the engine automatically matches the respiratory department schema, limits the query field to FEV1 value, and names the table "Respiratory Function Assessment Table for Patients in the Respiratory Department" to avoid unauthorized queries across departments. It also designs an easy-to-verify layer and intercepts high-risk operations. It uses regular expressions to match prohibited SQL patterns. If the intention of "full table query" is detected (such as "export all patient data"), it will directly return a prompt of "inadequate query permission", forcing the specification of specific fields, logical compliance checks, and query logic based on nursing business rules. Its query logic includes "the time span of nursing record queries must not exceed 31 days". If the input is "query for nursing error records throughout 2023", it will automatically split into 12 single-month queries to avoid excessive database pressure;

[0078] (3) After the DeepSeek-V3 model generates the SQL statement, the confidence is calculated using the built-in scoring function. The formula used is: confidence = field matching × 0.6 + table association rationality × 0.3 + business rule compliance × 0.1; when the confidence is ≥ 80% (such as "query cardiology department metoprolol inventory", the field / table matching is clear), the SQL is executed directly; when the confidence is < 80% (such as "special nursing measures for a certain patient in surgery", "a certain patient" is unclear), the manual review process is triggered, and the nurse needs to supplement the patient's ID or name key information;

[0079] (4) Knowledge graph-driven precise query: construct a medical knowledge graph, define entities and relationships, and perform vector retrieval and recall optimization. Embed knowledge graph entities into 300-dimensional vectors and store them in the Milvus vector database. When processing queries, first extract the keywords "ambroxol" and "nursing matters" through BERT to generate query vectors, and recall related entities in the graph (such as "ambroxol-nursing contraindications-avoid use with central antitussive drugs"). The recall rate is increased from 45% of traditional keyword matching to 80%;

[0080] Entities include diseases (e.g., pneumonia), medications (e.g., amoxicillin), and nursing items (e.g., sputum suctioning, nebulized inhalation). Relationships include "disease-common medication" (pneumonia → amoxicillin), "drug-nursing contraindications" (amoxicillin → avoid use with probenecid), and "nursing item-applicable disease" (sputum suctioning → chronic obstructive pulmonary disease). The data source is "disease-medication" association data extracted from structured electronic medical records and "procedure-contraindication" relationships extracted from nursing specification documents, constructing a graph containing over 500,000 entities and over 1 million relationships.

[0081] (5) Perform RAG fusion. According to the query intent (such as business data query / knowledge consultation), relevant data fragments are recalled from MySQL / MongoDB or knowledge graphs, and the recalled data is combined with the SQL statements generated by DeepSeek-V3. For example, when querying "Are there any patients in the respiratory department who use ambroxol and codeine at the same time?", first find out through the graph that the two are contraindicated in combination, and then retrieve specific patients through SQL. The final result is accompanied by a knowledge prompt, "Note: The combination of ambroxol and codeine may cause respiratory depression";

[0082] The medical knowledge question-and-answer module and the business data query module are parallel modules. They also perform business processing based on the business data query process derived from the voice interaction module and the intelligent decision engine module. The specific logical steps are as follows:

[0083] (1) Question parsing and scenario recognition: Input preprocessing: receiving natural language questions input by nurses through nursing whiteboards or handheld terminals, removing punctuation and converting them to lowercase;

[0084] (2) Large model calling strategy: automatically load corresponding domain parameters according to the problem type (e.g., loading pharmacology fine-tuning weights for “drug interaction” and loading nursing fine-tuning weights for “nursing standards”);

[0085] (3) Convert the original question into a structured prompt and output the result;

[0086] The system monitoring and optimization module is used for performance indicator monitoring, system optimization, and adjustment. During performance indicator monitoring, the server-integrated monitoring tool provides real-time monitoring of system performance indicators, including data collection frequency, early warning analysis delay, and message push success rate. Monitoring code is embedded in key system nodes to collect performance data, store it in the Prometheus database, and display it through the Grafana visualization interface to promptly identify system performance bottlenecks.

[0087] When optimizing and adjusting the system, the system is optimized based on performance monitoring data. If data collection delay is too high, the data collection algorithm is optimized or collection nodes are added. If early warning analysis efficiency is low, the large model parameters are adjusted or the configuration of the stream computing framework is optimized. The system is regularly maintained and upgraded to ensure that the system is always in the best operating state and to ensure the efficient development of smart ward services.

[0088] The database includes a basic information configuration database, a data structure knowledge base, and a smart ward database;

[0089] The present invention also proposes an AI nursing whiteboard method based on a large model intelligent agent, comprising the following steps:

[0090] S1: The nurse wakes up the AI ​​nursing assistant through the intercom function of the handheld device;

[0091] S2: After receiving the signal, the nursing whiteboard terminal loads and displays the question-and-answer guidance;

[0092] S3: Users input questions in natural language. The intelligent decision engine module collects the input questions and performs data cleaning and structuring.

[0093] S4: The intelligent decision engine module uses the BERT model fine-tuned through transfer learning and the dual-channel BiLSTM-Attention model to determine whether the question belongs to "business inquiry" or "medical knowledge question answering," and performs entity recognition and intent classification.

[0094] S5: If it is a business data query problem: follow the business query path through the business data query module and perform the following steps:

[0095] S501: Convert natural language questions into SQL queries, automatically match the medical schema, bind fields to data tables, ensure department isolation and field legitimacy, and verify SQL compliance. Then, calculate the SQL confidence level. If the confidence level is ≥80%, the SQL statement is executed directly. If the confidence level is <80%, a manual review process is triggered, requiring the nurse to provide key patient information such as the patient's ID or name.

[0096] S502: The cached results are retrieved from Redis first. If a query fails, the query is returned to the source MySQL. The query results are displayed graphically on the nursing whiteboard terminal.

[0097] S503: RAG fusion enhanced output: Based on the query intent, relevant data fragments are recalled from MySQL / MongoDB or knowledge graph (medical knowledge), and the recalled data is combined with the SQL statement generated by DeepSeek-V3. First, the graph is used to determine whether the two are contraindicated. Then, SQL is used to search for specific patients, and the final result is accompanied by knowledge prompts.

[0098] S6: If the question is a medical knowledge question and answer question: the medical knowledge question and answer module follows the large model question and answer path, and performs the following steps:

[0099] S601: Use the domain dictionary and question template to determine whether it is a medical knowledge question (threshold ≥ 0.9);

[0100] S602: If yes, then enter the large model reasoning process and fine-tune the weights based on the problem domain by calling different domain knowledge;

[0101] S603: Generate a structured prompt containing query objectives, constraints, and output requirements;

[0102] S7: The system monitoring and optimization module uses Prometheus + Grafana to monitor the system's performance indicators in real time, including data collection frequency, early warning analysis delay, and message push success rate. System parameters are regularly optimized based on performance monitoring data.

[0103] This embodiment can quickly integrate various types of patient diagnosis and treatment data, such as vital signs, doctor's orders, nursing records, etc., by integrating multi-source data, intelligent early warning, precise question and answer, etc., and present them in a clear and intuitive manner, saving a lot of time and energy, enabling it to understand the patient's condition more quickly, and can monitor the patient's vital signs and condition changes in real time. Once an abnormal situation occurs, such as heart rate and blood pressure exceeding the normal range, an early warning will be issued immediately, and the nurse can be informed and take corresponding measures at the first time, greatly improving the response speed to the patient's emergency. When the nurse encounters questions about medical knowledge, he or she can ask through the AI ​​nursing assistant, which improves the convenience of question and answer and helps to improve the accuracy of nursing decisions.

[0104] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. An AI nursing whiteboard system based on a large model agent, characterized by: It includes a nursing whiteboard terminal, an AI smart chain platform and a database. The nursing whiteboard terminal includes an AI nursing assistant and a built-in whiteboard control system. The AI ​​Zhilian platform includes a voice interaction module, an intelligent decision-making engine module, a business data query module, a medical knowledge question-and-answer module, and a system monitoring and optimization module; The database includes a basic information configuration library, a data structure knowledge library and a smart ward database.

2. The AI ​​nursing whiteboard system based on a large model agent according to claim 1 is characterized in that: The voice interaction module is used to initiate interaction with the nurse and load the guidance language on the front end. When initiating interaction with the nurse, the nurse uses Bluetooth and Wi-Fi wireless communication technology to establish a connection with the nursing whiteboard terminal through the handheld device, and wakes up the AI ​​nursing assistant of the nursing whiteboard through the intercom function of the handheld device, accurately transmits the wake-up command to the whiteboard control system, and triggers the AI ​​nursing assistant to start; When the guide words are loaded on the front end, the AI ​​nursing assistant front-end page loads and displays the question-and-answer guide words based on the background configuration. At the same time, you can choose whether to perform voice broadcasting, read the guide word configuration information from the background, complete the integration of the voice broadcast function through the speech synthesis SDK, and decide whether to speak based on the configuration switch.

3. The AI ​​nursing whiteboard system based on a large model agent according to claim 1 is characterized in that: The intelligent decision engine module is used to process and classify historical data, creating two engines for medical knowledge question and answer and business knowledge question and answer, so as to correctly distribute the questions and answers received by the voice interaction module; The specific logical steps are as follows: (1) Data cleaning and structuring: remove duplicate data and non-medical related text, use regular expressions to filter special symbols, and retain the core business fields of "department + drug + operation"; (2) Pre-training model selection and fine-tuning: Based on the BERT model pre-trained in general domains, the hospital’s internal electronic medical record text is loaded for domain transfer learning, focusing on optimizing the semantic understanding of the professional terms "creatinine value" and "nebulized inhalation". During the fine-tuning stage, the high-frequency department and drug names in the business query data are input into the model as entity masks, forcing the model to learn the association features of "department-drug-query type"; (3) Model architecture design, design of dual-channel feature fusion: Two-channel embedding layer: Channel 1 is used to pre-train semantic features, freeze the parameters of the first six layers of BERT, extract common semantic vectors of medical terms, and identify knowledge keywords such as "taboo" and "combination." Channel 2 is for business dynamic features. It uses dynamic weights to capture the combination pattern of "department + time + operation." The embedding layer is randomly initialized to learn dynamic parameters in business queries, including department names, time words, and operation words. Bidirectional LSTM layer: Receives the concatenated vectors from the two channels, captures contextual dependencies, and uses BiLSTM to extract the temporal association between "blood oxygen saturation-threshold-nursing action"; Introducing an attention mechanism, for business queries, it focuses on the keywords "inventory," "trend," and "current." For medical queries, it focuses on the terms "taboo," "regulation," and "precautions." The attention weight matrix \(\alpha\) is used to quantify the contribution of each word to the classification. (4) Business scenario-oriented optimization: perform sample weighting and data enhancement, give higher weights to samples of low-frequency departments in the business query category, avoid the model being biased towards high-frequency departments, and use EDA technology to perform synonym replacement and word order adjustment on medical knowledge texts, expand small sample category data, and use weighted cross entropy loss function to set weights for business query category / medical knowledge category according to the proportion of training sets, balance the difference in the number of samples of the two categories, and calculate the business category F1 and knowledge category F1 after each round of training, requiring both categories F1 to be ≥92% to avoid overfitting of the model to one category; (5) Confusion matrix driven optimization: Real-time confusion matrix analysis, deployment of online monitoring system, hourly statistics of confusion matrix, focusing on two types of errors, including business query → medical knowledge and medical knowledge → business query, and implementing incremental learning mechanism, summarizing classification error cases reported by clinical nurses every week, adding them to the training set after manual annotation, and fine-tuning the model through hot update mechanism to ensure the recognition accuracy of emerging business terms ≥95%.

4. The AI ​​nursing whiteboard system based on a large model agent according to claim 1 is characterized in that: The business data query module performs business processing based on the business data query process divided by the voice interaction module and the intelligent decision engine module; The specific logical steps are as follows: (1) Hybrid database architecture design: Adapting to the characteristics of nursing business data: adopting a data layered storage strategy, the MySQL structured data layer is used to store standardized nursing records, including patient temperature sheets and doctor's order execution tables, and the "patient-department-nursing operation" relationship is defined through the ER model, supporting multi-table JOIN complex queries, and the Redis real-time cache layer is used to cache high-frequency query results, setting a 5-minute expiration time to reduce the pressure on MySQL / MongoDB, and storing nurse login status and temporary query conditions; (2) Text2SQL engine construction: A "department-data table-field" mapping dictionary is established through the medical Schema binding mechanism. When the nurse enters the query "FEV1 value trend of patients in bed 3 of the respiratory department in the past two days", the engine automatically matches the respiratory department schema, limits the query field to FEV1 value, and names the table "Respiratory Function Assessment Table for Patients in the Respiratory Department" to avoid cross-departmental unauthorized queries. It also designs an easy-to-verify layer and intercepts high-risk operations. It matches prohibited SQL patterns through regular expressions. If the intention of "full table query" is detected, it directly returns the prompt of "inadequate query permission", forcing the specification of specific fields, logical compliance checks, and queries based on the query logic of nursing business rules. Its query logic includes "the time span of nursing record query shall not exceed 31 days". If you enter "query the nursing error record for the whole year of 2023", it will automatically split into 12 single-month queries to avoid excessive database pressure. (3) After generating the SQL statement using the DeepSeek-V3 model, the confidence is calculated using the built-in scoring function. The formula used is: confidence = field matching × 0.6 + table association rationality × 0.3 + business rule compliance × 0.1; when the confidence is ≥ 80%, the SQL statement is executed directly; when the confidence is < 80%, the manual review process is triggered, and the nurse needs to supplement the patient's ID or name key information; (4) Knowledge graph-driven precise query: Construct a medical knowledge graph, define entities and relationships, and perform vector retrieval and recall optimization. Embed knowledge graph entities into 300-dimensional vectors and store them in the Milvus vector database. When processing queries, first extract the keywords "ambroxol" and "nursing matters" through BERT, generate query vectors, and recall related entities in the graph. The recall rate is increased from 45% of traditional keyword matching to 80%; (5) Perform RAG fusion. According to the query intent, relevant data fragments are recalled from MySQL / MongoDB or knowledge graphs, and the recalled data is combined with the SQL statements generated by DeepSeek-V3. First, the graph is used to find out whether the two are contraindicated in combination, and then SQL is used to retrieve specific patients. The final result is accompanied by knowledge prompts.

5. The AI ​​nursing whiteboard system based on a large model agent according to claim 1 is characterized in that: The medical knowledge question and answer module and the business data query module are parallel modules. They also perform business processing based on the business data query process divided by the voice interaction module and the intelligent decision engine module. The specific logical steps are as follows: (1) Question parsing and scenario recognition: Input preprocessing: receiving natural language questions input by nurses through nursing whiteboards or handheld terminals, removing punctuation and converting them to lowercase; (2) Large model calling strategy: automatically load corresponding domain parameters according to the problem type; (3) Convert the original question into a structured prompt and output the result.

6. The AI ​​nursing whiteboard system based on a large model agent according to claim 1 is characterized in that: The system monitoring and optimization module is used for performance indicator monitoring and system optimization and adjustment. During performance indicator monitoring, the server-integrated monitoring tool is used to monitor the system's performance indicators in real time, including data collection frequency, early warning analysis delay, and message push success rate. Monitoring code is implanted at key system nodes to collect performance data and store it in the Prometheus database. This data is then displayed through the Grafana visualization interface to promptly identify system performance bottlenecks. When optimizing and adjusting the system, optimize the system based on performance monitoring data. If the data collection delay is too high, optimize the data collection algorithm or add collection nodes. When early warning analysis is inefficient, adjust large model parameters or optimize the configuration of the stream computing framework, and regularly maintain and upgrade the system to ensure that the system is always in the best operating state and to ensure the efficient development of smart ward services.

7. A method for AI nursing whiteboard based on large model intelligent agent, characterized in that: The following steps are involved: S1: The nurse wakes up the AI ​​nursing assistant through the intercom function of the handheld device; S2: After receiving the signal, the nursing whiteboard terminal loads and displays the question-and-answer guidance; S3: Users input questions in natural language. The intelligent decision engine module collects the input questions and performs data cleaning and structuring. S4: The intelligent decision engine module uses a BERT model fine-tuned through transfer learning and a dual-channel BiLSTM-Attention model to determine whether a question is a "business query" or a "medical knowledge question and answer" and performs entity recognition and intent classification. S5: If it is a business data query problem: follow the business query path through the business data query module and perform the following steps: S501: Convert natural language questions into SQL queries, automatically match the medical schema, bind fields to data tables, ensure department isolation and field legitimacy, and verify SQL compliance. Then, calculate the SQL confidence level. If the confidence level is ≥80%, the SQL statement is executed directly. If the confidence level is <80%, a manual review process is triggered, requiring the nurse to provide key patient information such as the patient's ID or name. S502: The cached results are retrieved from Redis first. If a query fails, the query is returned to the source MySQL. The query results are displayed graphically on the nursing whiteboard terminal. S503: RAG fusion enhanced output: Based on the query intent, relevant data fragments are recalled from MySQL / MongoDB or the knowledge graph. The recalled data is combined with the SQL statement generated by DeepSeek-V3. First, the graph is used to determine whether the two are contraindicated. Then, SQL is used to search for specific patients. The final result is accompanied by knowledge prompts. S6: If the question is a medical knowledge question and answer question: the medical knowledge question and answer module follows the large model question and answer path, and performs the following steps: S601: Use the domain dictionary and question template to determine whether it is a medical knowledge question (threshold ≥ 0.9); S602: If yes, then enter the large model reasoning process and fine-tune the weights based on the problem domain by calling different domain knowledge; S603: Generate a structured prompt containing query objectives, constraints, and output requirements; S7: The system monitoring and optimization module uses (Prometheus + Grafana) to monitor the system's performance indicators in real time, including data collection frequency, warning analysis delay, and message push success rate, and regularly optimizes system parameters based on performance monitoring data.

Images