AI dynamic decision engine-based outpatient queuing and calling scheduling system

Abstract

The invention discloses an outpatient queuing and calling scheduling system based on an AI dynamic decision engine. The outpatient queuing and calling scheduling system comprises a multi-source data acquisition and preprocessing module, a DeepSeek intelligent prediction and scheduling engine, a dynamic priority ranking and calling module and a cross-department resource collaboration module. According to the method, the whole outpatient service process is reconstructed through AI, a closed loop of data driving, intelligent prediction, dynamic scheduling and cross-section collaboration is achieved, a traditional static queue is converted into a dynamic system capable of sensing disease severity, special crowd requirements and resource loads, the medical resource allocation efficiency is improved, the average waiting time of patients is shortened, and the medical resource allocation efficiency is improved. The rescue response speed of critical patients is improved by multiple times, meanwhile, the diagnosis period of difficult cases is shortened through automatic triggering of multidisciplinary consultation (MDT), pregnant women / disabled persons / elderly patients are automatically recognized in cooperation with the multi-modal recognition technology, dynamic priority values are given, and the priority service missed recognition rate is obviously improved.

Description

Technical Field

[0001] This invention relates to the field of outpatient queueing and calling scheduling technology, and in particular to an outpatient queueing and calling scheduling system based on an AI dynamic decision engine. Background Technology

[0002] Traditional outpatient queuing systems operate on a "first-come, first-served" basis, relying on manual triage desks and basic electronic display call numbers. They lack dynamic priority adjustments and cross-departmental collaboration capabilities. Patients wait in a fixed sequence after registration, medical staff manually handle emergencies, and resource allocation depends on experience-based judgment, resulting in both peak-hour congestion and off-peak idleness. The shortcomings of existing outpatient queuing systems include: 1. Static and rigid queues: Sorting only by registration time fails to identify critically ill patients, leading to delayed SOS case response; 2. Resource silos: Doctors / equipment are fixedly bound between departments, preventing support from pediatricians when they are not in use. 1. Fluctuations in CT equipment utilization in general practice clinics; 2. Delayed identification of special populations: Patients need to actively declare or manual identification of vulnerable groups, resulting in a high rate of missed identification during priority calling and easily leading to doctor-patient conflicts; 3. Lack of predictive ability: Inability to predict peak visit times leads to delays in manpower / equipment allocation, with patients waiting for an average of more than 1 hour; 4. Inefficient cross-departmental collaboration: MDT consultations for difficult cases require manual coordination of more than 3 departments, with an average startup time of more than 1 hour, and equipment sharing relies on manual scheduling, resulting in high relocation time; In view of the above, this application proposes an outpatient queuing and calling scheduling system based on an AI dynamic decision engine. Summary of the Invention

[0003] Based on the technical problems existing in the background technology, this invention proposes an outpatient queuing and calling scheduling system based on an AI dynamic decision engine.

[0004] This invention proposes an outpatient queuing and calling scheduling system based on an AI dynamic decision engine, comprising a multi-source data acquisition and preprocessing module, a DeepSeek intelligent prediction and scheduling engine, a dynamic priority ranking and calling module, and a cross-departmental resource collaboration module. The multi-source data acquisition and preprocessing module, the DeepSeek intelligent prediction and scheduling engine, and the cross-departmental resource collaboration module are connected to each other.

[0005] The multi-source data acquisition and preprocessing module includes a patient behavior acquisition unit, a department status monitoring unit, an environmental and external data unit, and a data cleaning and feature engineering unit.

[0006] The DeepSeek intelligent prediction and scheduling engine includes a traffic prediction model unit, a resource allocation optimization unit, and an elastic number pool management unit.

[0007] The dynamic priority sorting and calling module includes a disease triage rule engine, a special population identification unit, and a multi-queue fusion algorithm unit.

[0008] The cross-departmental resource collaboration module includes a doctor flexible support unit, an examination equipment sharing unit, and a joint diagnosis and treatment triggering unit.

[0009] Preferably, the patient behavior acquisition unit is used for data source docking and time series annotation. When docking with the data source, it docks with the registration system, Bluetooth positioning, and APP interaction logs. The registration system obtains the patient's registration time, department, and chief complaint text in real time through the hospital HIS system API. When Bluetooth positioning is used, iBeacon devices are deployed in the waiting area to collect the patient's mobile phone MAC address and coordinates at 10-second intervals. The APP interaction logs are used to capture patient operations. When annotating the time series, the location data is aggregated in 15-minute intervals to generate a trajectory sequence, calculate the waiting time: current time - registration time, and perform a prediction of the risk of missing the appointment. When making the prediction, the current waiting time and the department's historical missed appointment rate are input into the LSTM to train the risk prediction model. The LSTM trains the risk prediction model and outputs the risk value.

[0010] The department status monitoring unit is used for real-time data acquisition and structured time-series data generation. During real-time data acquisition, the HIS system records the timestamp of the doctor clicking "see a patient" and calculates the number of patients seen per hour to determine the doctor's patient-seeking speed. It also connects to the IoT sensors of the examination equipment to collect equipment status statistics, calculate occupancy rate, and obtain the current queue length through the queuing system API. During structured time-series data generation, Apache Kafka is used to transmit data in real time and store it in the InfluxDB time-series database to generate a time series table.

[0011] Preferably, the environment and external data unit is used for third-party API access and feature matrix construction. The third-party API access is used for weather data and holiday data. The weather data is obtained by calling the OpenWeatherMap API to obtain real-time PM2.5, temperature and humidity. The holiday data is synchronized with the government interface to obtain holiday markers. The feature matrix construction is used to associate external data with internal data to generate multi-dimensional features.

[0012] The data cleaning and feature engineering unit is used for unstructured data processing, structured data cleaning, and feature derivation. During unstructured data processing, the DeepSeek model is used to vectorize patient complaints and extract keywords as features. During structured data cleaning, historical averages are used to fill in missing equipment occupancy rates, and outliers in patient volume are removed based on the 3σ principle. During feature derivation, the "hourly pediatric patient growth rate" is calculated using the following formula:

[0013] Preferably, the operation logic steps of the traffic prediction model unit are as follows:

[0014] S101: Model Architecture Design: Input historical medical data and textual descriptions of external variables into the DeepSeek time series prediction engine. The DeepSeek time series prediction engine automatically extracts key features and generates predicted values ​​for the next 1-4 hours through DeepSeek's time series understanding capabilities. The DeepSeek time series prediction engine outputs structured prediction results.

[0015] S102: Notification Project: Design a dedicated notification template, including historical data: {Department} hourly patient volume in the past 24 hours = {list}, equipment occupancy rate = {value}%, doctor's consultation rate = {value} people / hour; external variables: current PM2.5 = {value}, whether it is a holiday = {0 / 1}, epidemic risk level = {low / medium / high}, predict the patient volume of the department in the next 1-4 hours, and give the error rate;

[0016] S103: Model Training and Optimization: Fine-tune DeepSeek's time series prediction module using historical data, focusing on optimizing the error rate and dynamically adjusting the input weights;

[0017] The operational logic steps of the resource allocation optimization unit are as follows:

[0018] S201: DeepSeek reinforcement learning: It performs state space description, action space description and reward function design. When describing the state space, it converts the real-time state into natural language input. When describing the action space, it defines the executable actions as natural language options. When designing the reward function, it uses DeepSeek's multi-objective optimization capabilities to balance doctor utilization, equipment utilization and patient satisfaction.

[0019] S202: Make dynamic decisions, the process of which is as follows:

[0020] S2021: Trigger a decision every 15 minutes and input the current state into DeepSeek;

[0021] S2022: The model outputs the optimal action;

[0022] S2023: Perform the action and update the queuing system;

[0023] S203: Perform emergency scheduling: When the number of people in the queue exceeds the threshold, DeepSeek is immediately invoked for real-time decision-making.

[0024] Preferably, the flexible appointment pool management unit is used for cross-department appointment pool design and appointment allocation. The shared rule for cross-department appointment pool design is to dynamically determine the appointment allocation conditions through DeepSeek's rule engine, and to convert the appointment allocation request into natural language input through a real-time decision engine, with DeepSeek outputting a decision within 30 seconds.

[0025] The appointment allocation process is as follows: monitor the status of each department → trigger allocation conditions → call DeepSeek decision → update the queuing system → push notification.

[0026] Preferably, the logical steps of the disease diagnosis triage rule engine are as follows:

[0027] S301: Knowledge Graph Construction: Injecting the mapping relationship between common symptoms and disease severity into the DeepSeek medical knowledge graph;

[0028] S302: Patient Chief Complaint Analysis: Connects to the pre-examination and triage system, extracts the patient's chief complaint text, identifies key symptoms and duration through DeepSeek's text semantic analysis, and calls the knowledge graph reasoning engine to output priority labels;

[0029] S303: Dynamic rule updates: Supports updating rules via natural language commands, which are automatically parsed and updated by DeepSeek to update the knowledge graph;

[0030] The special population identification unit has multimodal recognition technology and priority weight allocation. Its multimodal recognition technology includes voice interaction recognition and ID card number parsing. During voice interaction recognition, the patient reports by voice, and DeepSeek's speech-to-text + intent recognition module extracts key information. During ID card number parsing, the age is automatically calculated or special identifiers are identified through the ID card number. The weight rules for priority weight allocation are: pregnant women: +2, elderly people (≥65 years old): +1, and disabled people: +3. If the patient meets multiple conditions at the same time, the priority weights are superimposed.

[0031] Preferably, the operation logic steps of the multi-queue fusion algorithm unit are as follows:

[0032] S401: Queue type definition: Time sequence queue: sorted by registration time; Disease priority queue: sorted by "SOS > Emergency > Normal"; Special population queue: sorted by priority value;

[0033] S402: Weighting Strategy: DeepSeek Dynamic Weight Calculation: The weight allocation is optimized through a reinforcement learning model to minimize patient waiting time and medical resource utilization. The calculation and fusion formula is: Final Score = Disease Priority Weight + Special Population Weight + Time Order Weight.

[0034] S403: Real-time call generation: The queue is refreshed every 30 seconds, and a call sequence is generated by sorting the final scores.

[0035] Preferably, the doctor elastic support unit is used for the design of the global doctor resource pool and the distributed decision engine. It performs real-time status monitoring during the design of the global doctor resource pool, obtains the status of doctors in each department and patient backlog data in real time through the hospital HIS system, and defines support priorities, including: doctors in adjacent departments, doctors in the same specialty group, and the hospital's backup doctor pool.

[0036] The operational logic steps of the distributed decision engine are as follows:

[0037] S501: Input: Current department status, distribution of doctor resources;

[0038] S502: Processing: Call DeepSeek for global optimization to minimize total waiting time and doctor movement costs. Key decision factors include: doctor's professional matching degree, support path time, and impact on the original department after support.

[0039] S503: Output: Optimal support plan. After the decision is made, the HIS system will be automatically updated to update the doctor's schedule, and a notification will be pushed through the hospital's communication tools.

[0040] Preferably, the operating logic steps of the inspection equipment sharing unit are as follows:

[0041] S601: Device Status IoT Integration: Real-time data acquisition, obtaining device status through RFID tags or sensors, combined with traffic prediction results, to predict high-demand departments 30 minutes in advance;

[0042] S602: Path Planning: The path planning algorithm is calculated using a dynamic scheduling engine. The specific logical steps are as follows:

[0043] S6021: Input: Device location, target department, hospital map data;

[0044] S6022: Processing: By calculating the optimal path, the goal is to minimize the sum of travel time and equipment preparation time;

[0045] S6023: Key Constraints; Avoid congested areas during peak hours and prioritize the use of medical elevators;

[0046] S6024: Output: Device movement command;

[0047] S603: Automatic Navigation Execution: Integrated with AGV systems to enable autonomous movement of equipment.

[0048] The operational logic steps of the combined diagnosis and treatment triggering unit are as follows:

[0049] S701: Difficult Case Identification Engine: Performs multimodal data fusion, accesses patient electronic medical records (EMR), imaging data (PACS), and laboratory results (LIS), performs comprehensive analysis through the DeepSeek large model, and automatically identifies difficult cases through multidisciplinary team (MDT) trigger rules;

[0050] S702: Collaboration Platform: Consultation Suggestion Generation: Input: Patient data, medical history, examination results; Processing: Generates multidisciplinary consultation suggestions through a large model;

[0051] Dedicated lane creation: Add a "dedicated lane" to the queuing system, with a higher priority than SOS patients;

[0052] S703: Dynamic calling logic adjustment: Insert MDT cases at the head of the queue and push notifications to doctors in relevant departments.

[0053] Compared with existing technologies, the beneficial effects of this invention are:

[0054] 1. By using NLP chief complaint analysis technology, SOS cases such as chest pain / coma can be directly connected to the emergency care channel, bypassing the routine queue process, thus achieving a qualitative leap in the timeliness of arrival in the emergency room;

[0055] 2. By using multimodal recognition technology to automatically identify pregnant women / people with disabilities / elderly patients and assign them dynamic priority values, the rate of missed identification in priority services has been significantly improved.

[0056] 3. By using an intelligent prediction engine to achieve dynamic matching of doctor / equipment resources, the workload of pediatricians is effectively improved and the phenomenon of CT equipment being idle is significantly curbed.

[0057] 4. The flexible support mechanism for doctors has greatly reduced the time required for cross-departmental response. The AGV equipment combined with path optimization technology has significantly reduced the time required for sharing, dispatching and moving medical equipment. In addition, the automatic triggering of multidisciplinary consultation mechanism for difficult cases has led to a breakthrough improvement in the time required for expert teams to assemble.

[0058] This invention reconstructs the entire outpatient process using AI, achieving a closed loop of "data-driven, intelligent prediction, dynamic scheduling, and interdisciplinary collaboration." It transforms the traditional static queue into a dynamic system that can perceive the severity of illness, the needs of special populations, and resource load, thereby improving the efficiency of medical resource allocation, shortening the average patient waiting time, and increasing the response speed for critically ill patients by several times. At the same time, it shortens the diagnosis cycle of difficult cases through automated triggering of multidisciplinary team (MDT) consultations, and automatically identifies pregnant women, disabled persons, and elderly patients with dynamic priority weights using multimodal recognition technology, significantly improving the rate of missed priority services. In addition, the intelligent prediction engine enables dynamic matching of doctor and equipment resources, effectively improving the workload of pediatricians and significantly curbing the idleness of CT equipment. Attached Figure Description

[0059] Figure 1 This is a block diagram of an outpatient queuing and calling scheduling system based on an AI dynamic decision engine proposed in this invention.

[0060] Figure 2 The flowchart shows the DeepSeek intelligent prediction and scheduling engine in the outpatient queuing and calling scheduling system based on AI dynamic decision engine proposed in this invention.

[0061] Figure 3 This is a flowchart of the dynamic priority sorting and calling module in an outpatient queuing and calling scheduling system based on an AI dynamic decision engine, as proposed in this invention.

[0062] Figure 4 This is a flowchart of a cross-departmental resource collaboration module in an outpatient queuing and calling system based on an AI dynamic decision engine, as proposed in this invention. Detailed Implementation

[0063] The present invention will be further explained below with reference to specific embodiments.

[0064] Example

[0065] Reference Figure 1-4 This embodiment proposes an outpatient queuing and calling scheduling system based on an AI dynamic decision engine, including a multi-source data acquisition and preprocessing module, a DeepSeek intelligent prediction and scheduling engine, a dynamic priority ranking and calling module, and a cross-departmental resource collaboration module. The multi-source data acquisition and preprocessing module, the DeepSeek intelligent prediction and scheduling engine, and the cross-departmental resource collaboration module are connected to each other.

[0066] The multi-source data acquisition and preprocessing module includes a patient behavior acquisition unit, a departmental status monitoring unit, an environmental and external data unit, and a data cleaning and feature engineering unit.

[0067] The patient behavior data collection unit is used for data source integration and time series annotation. During data source integration, it connects with the registration system, Bluetooth positioning, and APP interaction logs. The registration system obtains the patient's registration time, department, and chief complaint (e.g., "fever for 3 days, cough") in real time through the hospital's HIS system API. For Bluetooth positioning, iBeacon devices are deployed in the waiting area to collect the patient's mobile phone MAC address and coordinates at 10-second intervals (e.g., "Patient A was located at coordinates (23,45) at 10:00"). The APP interaction logs are used to capture patient actions (e.g., clicking "re-queue after missed appointment"). The system uses the timestamp of the button to aggregate location data in 15-minute intervals during time series labeling, generating a trajectory sequence (e.g., "10:00-10:15 in the waiting area, 10:15 entering the consultation room"). It calculates the waiting time as: current time - registration time (e.g., patient A waits 45 minutes) and predicts the risk of missing an appointment. During prediction, the system inputs the current waiting time and the department's historical missed appointment rate (e.g., pediatrics' historical missed appointment rate is 15%) into the LSTM training risk prediction model. The LSTM training risk prediction model outputs a risk value (e.g., 0.72, representing a 72% probability of missing an appointment).

[0068] The department status monitoring unit is used for real-time data acquisition and structured time-series data generation. During real-time data acquisition, the HIS system records the timestamp of the doctor clicking "see a patient" and calculates the number of patients seen per hour (e.g., "Dr. Zhang in Pediatrics saw 12 patients between 10:00 and 11:00") to determine the doctor's patient-seeking speed. It also connects to the IoT sensors of the examination equipment to collect statistics on the equipment status ("idle" / "in use"), calculates the occupancy rate (e.g., "Ultrasound room occupancy rate 85%), and obtains the current queue length (e.g., "50 people are waiting in line in Pediatrics") through the queuing system API. During structured time-series data generation, Apache Kafka is used to transmit data in real time and store it in the InfluxDB time-series database to generate a time series table.

[0069] For example, at 10:30 AM, the system detected 50 people in the pediatric queue, an 85% occupancy rate for ultrasound machines, and a doctor's consultation rate of 12 patients per hour. Based on historical data, it predicted that the number of people in the queue would increase to 65 within the next hour, triggering module 2 (DeepSeek engine) to dynamically increase the number of pediatric appointments by 10%.

[0070] The environmental and external data unit is used for third-party API access and feature matrix construction. The third-party API access is used for weather data and holiday data. Weather data is obtained by calling the OpenWeatherMap API to obtain real-time PM2.5, temperature and humidity. Holiday data is synchronized with the government interface to obtain holiday markers. Feature matrix construction is used to associate external data with internal data to generate multi-dimensional features.

[0071] For example: Weather: On a certain day, PM2.5 = 120 (severe pollution). The system detected that the number of pediatric respiratory patients increased by 30% compared to the previous day. Based on historical data, it was found that when PM2.5 > 100, the number of patients increased by an average of 25%. Trigger module 2 to increase the number of respiratory appointment slots by 15%.

[0072] The data cleaning and feature engineering unit is used for unstructured data processing, structured data cleaning, and feature derivation. For unstructured data processing, the DeepSeek model is used to vectorize patient complaints and extract keywords as features. For structured data cleaning, historical averages are used to fill in missing values ​​for equipment occupancy rates, and outliers in patient volume are removed based on the 3σ principle. For feature derivation, the "hourly pediatric patient growth rate" is calculated using the following formula:

[0073] For example, if a patient complains of "recurrent cough with low-grade fever for 1 week", the system extracts the keywords "cough" and "low-grade fever" using Med-BERT and combines them with historical data to find that the average waiting time for patients with this type of complaint is 60 minutes, triggering module 2 to prioritize the allocation of resources to the respiratory department.

[0074] DeepSeek's intelligent prediction and scheduling engine includes a traffic prediction model unit, a resource allocation optimization unit, and an elastic number pool management unit.

[0075] The operational logic steps of the traffic prediction model unit are as follows:

[0076] S101: Model Architecture Design: Input historical medical data and textual descriptions of external variables into the DeepSeek time series prediction engine. The DeepSeek time series prediction engine automatically extracts key features and generates predicted values ​​for the next 1-4 hours through DeepSeek's time series understanding capabilities. The DeepSeek time series prediction engine outputs structured prediction results.

[0077] S102: Notification Project: Design a dedicated notification template, including historical data: {Department} hourly patient volume in the past 24 hours = {list}, equipment occupancy rate = {value}%, doctor's consultation rate = {value} people / hour; external variables: current PM2.5 = {value}, whether it is a holiday = {0 / 1}, epidemic risk level = {low / medium / high}, predict the patient volume of the department in the next 1-4 hours, and give the error rate;

[0078] S103: Model Training and Optimization: Fine-tune DeepSeek's time series prediction module using historical data, focusing on optimizing the error rate and dynamically adjusting the input weights;

[0079] The operational logic steps of the resource allocation optimization unit are as follows:

[0080] S201: DeepSeek reinforcement learning: It performs state space description, action space description and reward function design. When describing the state space, it converts the real-time state into natural language input. When describing the action space, it defines the executable actions as natural language options. When designing the reward function, it uses DeepSeek's multi-objective optimization capabilities to balance doctor utilization, equipment utilization and patient satisfaction.

[0081] S202: Make dynamic decisions, the process of which is as follows:

[0082] S2021: Trigger a decision every 15 minutes and input the current state into DeepSeek;

[0083] S2022: The model outputs the optimal action;

[0084] S2023: Perform the action and update the queuing system;

[0085] S203: Perform emergency dispatch: When the number of people in the queue exceeds the threshold, DeepSeek is immediately invoked for real-time decision-making;

[0086] The flexible appointment pool management unit is used for cross-department appointment pool design and appointment allocation. The shared rule for cross-department appointment pool design is to dynamically determine the appointment allocation conditions through DeepSeek's rule engine, and to convert the appointment allocation request into natural language input through the real-time decision engine, with DeepSeek outputting the decision within 30 seconds.

[0087] The appointment allocation process is as follows: monitor the status of each department → trigger allocation conditions → call DeepSeek decision → update the queuing system → push notification;

[0088] The dynamic priority sorting and calling module includes a disease-based triage rule engine, a special population identification unit, and a multi-queue fusion algorithm unit;

[0089] The logical steps of the disease diagnosis triage rule engine are as follows:

[0090] S301: Knowledge Graph Construction: Injecting the mapping relationship between common symptoms and disease severity into the DeepSeek medical knowledge graph;

[0091] S302: Patient Chief Complaint Analysis: Connects to the pre-examination and triage system, extracts the patient's chief complaint text, identifies key symptoms and duration through DeepSeek's text semantic analysis, and calls the knowledge graph reasoning engine to output priority labels;

[0092] S303: Dynamic rule updates: Supports updating rules via natural language commands, which are automatically parsed and updated by DeepSeek to update the knowledge graph;

[0093] For example: Input: Patient's chief complaint: "Chest pain accompanied by profuse sweating, lasting for 30 minutes";

[0094] deal with:

[0095] DeepSeek analyzes the symptoms: "chest pain" (weight +3), "profuse sweating" (weight +2), "lasting for 30 minutes" (weight +1);

[0096] Knowledge graph matching: Associated disease "acute myocardial infarction", marked with priority "SOS";

[0097] Output: The patient is marked as "SOS priority" and skips the regular queue to enter the emergency channel directly;

[0098] The special population identification unit has multimodal recognition technology and priority weight allocation. Its multimodal recognition technology includes voice interaction recognition and ID card number parsing. During voice interaction recognition, patients report by voice, and DeepSeek's speech-to-text + intent recognition module extracts key information. During ID card number parsing, the age is automatically calculated or special identifiers are identified through the ID card number. The weight rules for priority weight allocation are: pregnant women: +2, elderly people (≥65 years old): +1, and disabled people: +3. If a patient meets multiple conditions at the same time, the priority weights are added together.

[0099] The execution logic steps of the multi-queue fusion algorithm unit are as follows:

[0100] S401: Queue type definition: Time sequence queue: sorted by registration time; Disease priority queue: sorted by "SOS > Emergency > Normal"; Special population queue: sorted by priority value;

[0101] S402: Weighting Strategy: DeepSeek Dynamic Weight Calculation: The weight allocation is optimized through a reinforcement learning model to minimize patient waiting time and medical resource utilization. The calculation and fusion formula is: Final Score = Disease Priority Weight + Special Population Weight + Time Order Weight.

[0102] S403: Real-time call generation: The queue is refreshed every 30 seconds, and a call sequence is generated by sorting the final scores;

[0103] For example: Queue status:

[0104] Time sequence queue: Patient A (registered at 8:00, regular).

[0105] Disease priority cohort: Patient B (SOS).

[0106] Special population cohort: Patient C (pregnant woman, weight +2).

[0107] Weight calculation:

[0108] Patient A: Ordinary (10) + Time sequence (earlier registration time, weight +10) = 20.

[0109] Patient B: SOS(100) = 100.

[0110] Patient C: General (10) + Special population (2×3=6) + Time sequence (late registration time, weight +5)=21;

[0111] Calling sequence: Patient B (SOS) → Patient C (Special Population) → Patient A (Regular);

[0112] The cross-departmental resource collaboration module includes a flexible doctor support unit, an examination equipment sharing unit, and a joint diagnosis and treatment triggering unit;

[0113] The physician elastic support unit is used for the design of the global physician resource pool and the distributed decision engine. It performs real-time status monitoring during the design of the global physician resource pool, obtains the status of physicians and patient backlog data in each department in real time through the hospital HIS system, and defines support priorities, including: physicians in adjacent departments, physicians in the same specialty group, and the hospital's backup physician pool.

[0114] The operational logic steps of the distributed decision engine are as follows:

[0115] S501: Input: Current department status, distribution of doctor resources;

[0116] S502: Processing: Call DeepSeek for global optimization to minimize total waiting time and doctor movement costs. Key decision factors include: doctor's professional matching degree, support path time, and impact on the original department after support.

[0117] S503: Output: Optimal support plan. After the decision is made, the HIS system will be automatically triggered to update the doctor's schedule, and a notification will be pushed through the hospital's communication tools.

[0118] The steps for checking the operating logic of the shared device unit are as follows:

[0119] S601: Device Status IoT Integration: Real-time data acquisition, obtaining device status through RFID tags or sensors, combined with traffic prediction results, to predict high-demand departments 30 minutes in advance;

[0120] S602: Path Planning: The path planning algorithm is calculated using a dynamic scheduling engine. The specific logical steps are as follows:

[0121] S6021: Input: Device location, target department, hospital map data;

[0122] S6022: Processing: By calculating the optimal path, the goal is to minimize the sum of travel time and equipment preparation time;

[0123] S6023: Key Constraints; Avoid congested areas during peak hours and prioritize the use of medical elevators;

[0124] S6024: Output: Device movement command;

[0125] S603: Automatic Navigation Execution: Integrated with AGV systems to enable autonomous movement of equipment;

[0126] The operational logic steps of the combined diagnosis and treatment trigger unit are as follows:

[0127] S701: Difficult Case Identification Engine: Performs multimodal data fusion, accesses patient electronic medical records (EMR), imaging data (PACS), and laboratory results (LIS), performs comprehensive analysis through the DeepSeek large model, and automatically identifies difficult cases through multidisciplinary team (MDT) trigger rules;

[0128] S702: Collaboration Platform: Consultation Suggestion Generation: Input: Patient data, medical history, examination results; Processing: Generates multidisciplinary consultation suggestions through a large model;

[0129] Dedicated lane creation: Add a "dedicated lane" to the queuing system, with a higher priority than SOS patients;

[0130] S703: Dynamic call number logic adjustment: Insert MDT cases at the head of the queue and push notifications to doctors in relevant departments;

[0131] This embodiment reconstructs the entire outpatient process using AI, achieving a closed loop of "data-driven, intelligent prediction, dynamic scheduling, and interdisciplinary collaboration." It transforms the traditional static queue into a dynamic system that can perceive the severity of illness, the needs of special populations, and resource load, thereby improving the efficiency of medical resource allocation, shortening the average waiting time for patients, and increasing the response speed for critically ill patients by several times. At the same time, the automated triggering of multidisciplinary team (MDT) consultations shortens the diagnosis cycle for difficult cases. Furthermore, the use of multimodal recognition technology to automatically identify pregnant women, people with disabilities, and elderly patients and assign them dynamic priority values ​​significantly improves the rate of missed priority services. In addition, the intelligent prediction engine enables dynamic matching of doctor and equipment resources, effectively improving the workload of pediatricians and significantly curbing the idleness of CT equipment.

[0132] In this embodiment, the multi-source data acquisition module extracts structured and unstructured data in real time from systems such as HIS / LIS / PACS, cleans it, and then injects it into the data lake. The DeepSeek intelligent prediction engine generates traffic prediction and resource scheduling strategies based on historical data. The results are simultaneously input into the dynamic priority ranking and calling module and the cross-departmental resource collaboration module. The dynamic priority ranking and calling module generates a priority queue based on the patient's chief complaint and special population tags, which drives the triggering of the MDT dedicated channel. The cross-departmental resource collaboration module optimizes resource allocation through the doctor flexible support unit and the examination equipment sharing unit. It is executed through the joint diagnosis and treatment triggering unit, and its execution results are fed back to the DeepSeek intelligent prediction engine to form a closed loop.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An outpatient queuing and calling scheduling system based on an AI dynamic decision engine, characterized in that, It includes a multi-source data acquisition and preprocessing module, a DeepSeek intelligent prediction and scheduling engine, a dynamic priority ranking and calling module, and a cross-departmental resource collaboration module. The multi-source data acquisition and preprocessing module, the DeepSeek intelligent prediction and scheduling engine, and the cross-departmental resource collaboration module are connected. The multi-source data acquisition and preprocessing module includes a patient behavior acquisition unit, a department status monitoring unit, an environmental and external data unit, and a data cleaning and feature engineering unit. The DeepSeek intelligent prediction and scheduling engine includes a traffic prediction model unit, a resource allocation optimization unit, and an elastic number pool management unit. The dynamic priority sorting and calling module includes a disease triage rule engine, a special population identification unit, and a multi-queue fusion algorithm unit. The cross-departmental resource collaboration module includes a doctor flexible support unit, an examination equipment sharing unit, and a joint diagnosis and treatment triggering unit.

2. The outpatient queuing and calling scheduling system based on an AI dynamic decision engine according to claim 1, characterized in that, The patient behavior acquisition unit is used for data source integration and time series annotation. During data source integration, it connects with the registration system, Bluetooth positioning, and APP interaction logs. The registration system obtains the patient's registration time, department, and chief complaint text in real time through the hospital HIS system API. During Bluetooth positioning, iBeacon devices are deployed in the waiting area to collect the patient's mobile phone MAC address and coordinates at 10-second intervals. The APP interaction logs are used to capture patient operations. During time series annotation, the location data is aggregated in 15-minute intervals to generate a trajectory sequence, calculate the waiting time: current time - registration time, and perform a prediction of the risk of missing the appointment. During the prediction, the current waiting time and the department's historical missed appointment rate are input into the LSTM training risk prediction model. The LSTM training risk prediction model outputs a risk value. The department status monitoring unit is used for real-time data acquisition and structured time-series data generation. During real-time data acquisition, the HIS system records the timestamp of the doctor clicking "see a patient" and calculates the number of patients seen per hour to determine the doctor's patient-seeking speed. It also connects to the IoT sensors of the examination equipment to collect equipment status statistics, calculate occupancy rate, and obtain the current queue length through the queuing system API. During structured time-series data generation, Apache Kafka is used to transmit data in real time and store it in the InfluxDB time-series database to generate a time series table.

3. The outpatient queuing and calling scheduling system based on an AI dynamic decision engine according to claim 1, characterized in that, The environment and external data unit is used for third-party API access and feature matrix construction. The third-party API access is used for weather data and holiday data. The weather data is obtained by calling the OpenWeatherMap API to obtain real-time PM2.5, temperature and humidity. The holiday data is synchronized with the government interface to obtain holiday markers. The feature matrix construction is used to associate external data with internal data to generate multi-dimensional features. The data cleaning and feature engineering unit is used for unstructured data processing, structured data cleaning, and feature derivation. During unstructured data processing, the DeepSeek model is used to vectorize patient complaints and extract keywords as features. During structured data cleaning, historical averages are used to fill in missing equipment occupancy rates, and outliers in patient volume are removed based on the 3σ principle. During feature derivation, the "hourly pediatric patient growth rate" is calculated using the following formula:

4. The outpatient queuing and calling system based on an AI dynamic decision engine according to claim 1, characterized in that, The operational logic steps of the traffic prediction model unit are as follows: S101: Model Architecture Design: Input historical medical data and textual descriptions of external variables into the DeepSeek time series prediction engine. The DeepSeek time series prediction engine automatically extracts key features and generates predicted values ​​for the next 1-4 hours through DeepSeek's time series understanding capabilities. The DeepSeek time series prediction engine outputs structured prediction results. S102: Notification Project: Design a dedicated notification template, including historical data: {Department} hourly patient volume in the past 24 hours = {list}, equipment occupancy rate = {value}%, doctor's consultation rate = {value} people / hour; external variables: current PM2.5 = {value}, whether it is a holiday = {0 / 1}, epidemic risk level = {low / medium / high}, predict the patient volume of the department in the next 1-4 hours, and give the error rate; S103: Model Training and Optimization: Fine-tune DeepSeek's time series prediction module using historical data, focusing on optimizing the error rate and dynamically adjusting the input weights; The operational logic steps of the resource allocation optimization unit are as follows: S201: DeepSeek reinforcement learning: It performs state space description, action space description and reward function design. When describing the state space, it converts the real-time state into natural language input. When describing the action space, it defines the executable actions as natural language options. When designing the reward function, it uses DeepSeek's multi-objective optimization capabilities to balance doctor utilization, equipment utilization and patient satisfaction. S202: Make dynamic decisions, the process of which is as follows: S2021: Trigger a decision every 15 minutes and input the current state into DeepSeek; S2022: The model outputs the optimal action; S2023: Perform the action and update the queuing system; S203: Perform emergency scheduling: When the number of people in the queue exceeds the threshold, DeepSeek is immediately invoked for real-time decision-making.

5. The outpatient queuing and calling scheduling system based on an AI dynamic decision engine according to claim 1, characterized in that, The flexible appointment pool management unit is used for cross-department appointment pool design and appointment allocation. The shared rule for cross-department appointment pool design is to dynamically determine the appointment allocation conditions through DeepSeek's rule engine, and to convert the appointment allocation request into natural language input through the real-time decision engine, with DeepSeek outputting the decision within 30 seconds. The appointment allocation process is as follows: monitor the status of each department → trigger allocation conditions → call DeepSeek decision → update the queuing system → push notification.

6. The outpatient queuing and calling scheduling system based on an AI dynamic decision engine according to claim 1, characterized in that, The logical steps of the disease diagnosis triage rule engine are as follows: S301: Knowledge Graph Construction: Injecting the mapping relationship between common symptoms and disease severity into the DeepSeek medical knowledge graph; S302: Patient Chief Complaint Analysis: Connects to the pre-examination and triage system, extracts the patient's chief complaint text, identifies key symptoms and duration through DeepSeek's text semantic analysis, and calls the knowledge graph reasoning engine to output priority labels; S303: Dynamic rule updates: Supports updating rules via natural language commands, which are automatically parsed and updated by DeepSeek to update the knowledge graph; The special population identification unit has multimodal recognition technology and priority weight allocation. Its multimodal recognition technology includes voice interaction recognition and ID card number parsing. During voice interaction recognition, the patient reports by voice, and DeepSeek's speech-to-text + intent recognition module extracts key information. During ID card number parsing, the age is automatically calculated or special identifiers are identified through the ID card number. The weight rules for priority weight allocation are: pregnant women: +2, elderly people (≥65 years old): +1, and disabled people: +3. If the patient meets multiple conditions at the same time, the priority weights are superimposed.

7. The outpatient queuing and calling scheduling system based on an AI dynamic decision engine according to claim 1, characterized in that, The operation logic steps of the multi-queue fusion algorithm unit are as follows: S401: Queue type definition: Time order queue: sorted by registration time; Disease priority queue: sorted by "SOS > Emergency > Normal"; Special population queue: sorted by priority value; S402: Weighting Strategy: DeepSeek Dynamic Weight Calculation: The weight allocation is optimized through a reinforcement learning model to minimize patient waiting time and medical resource utilization. The calculation and fusion formula is: Final Score = Disease Priority Weight + Special Population Weight + Time Order Weight. S403: Real-time call generation: The queue is refreshed every 30 seconds, and a call sequence is generated by sorting the final scores.

8. The outpatient queuing and calling scheduling system based on an AI dynamic decision engine according to claim 1, characterized in that, The doctor elastic support unit is used for the design of the global doctor resource pool and the distributed decision engine. It performs real-time status monitoring during the design of the global doctor resource pool, obtains the status of doctors in each department and patient backlog data in real time through the hospital HIS system, and defines support priorities, including: doctors in adjacent departments, doctors in the same specialty group, and the hospital's backup doctor pool. The operational logic steps of the distributed decision engine are as follows: S501: Input: Current department status, distribution of doctor resources; S502: Processing: Call DeepSeek for global optimization to minimize total waiting time and doctor movement costs. Key decision factors include: doctor's professional matching degree, support path time, and impact on the original department after support. S503: Output: Optimal support plan. After the decision is made, the HIS system will be automatically updated to update the doctor's schedule, and a notification will be pushed through the hospital's communication tools.

9. The outpatient queuing and calling scheduling system based on an AI dynamic decision engine according to claim 1, characterized in that, The operating logic steps of the inspection equipment sharing unit are as follows: S601: IoT integration of equipment status: Real-time data acquisition, obtaining equipment status through RFID tags or sensors, combined with traffic prediction results, to predict high-demand departments 30 minutes in advance; S602: Path Planning: The path planning algorithm is calculated using a dynamic scheduling engine. The specific logical steps are as follows: S6021: Input: Device location, target department, hospital map data; S6022: Processing: By calculating the optimal path, the goal is to minimize the sum of travel time and equipment preparation time; S6023: Key Constraints; Avoid congested areas during peak hours and prioritize the use of medical elevators; S6024: Output: Device movement command; S603: Automatic Navigation Execution: Integrated with AGV systems to enable autonomous movement of equipment. The operational logic steps of the combined diagnosis and treatment triggering unit are as follows: S701: Difficult Case Identification Engine: Performs multimodal data fusion, accesses patient electronic medical records, image data, and test results, conducts comprehensive analysis through the DeepSeek large model, and automatically identifies difficult cases through multidisciplinary diagnosis and treatment trigger rules; S702: Collaboration Platform: Consultation Suggestion Generation: Input: Patient data, medical history, examination results; Processing: Generates multidisciplinary consultation suggestions through a large model; Dedicated lane creation: Add a "dedicated lane" to the queuing system, with higher priority than SOS patients; S703: Dynamic calling logic adjustment: Insert MDT cases at the head of the queue and push notifications to doctors in relevant departments.

Images