Hospital resource scheduling system and method based on space-time diagram neural network

Abstract

The invention relates to the field of artificial intelligence, and provides a hospital resource scheduling system and method based on a space-time diagram neural network. The method comprises the following steps: acquiring multi-dimensional operation data of medical equipment and a medical service terminal through a multi-modal sensor network in the medical equipment and the medical service terminal; determining real-time medical scenes through a first-level scheduling model based on a time-space diagram neural network based on the geographic positions of the multi-dimensional operation data, the medical equipment and the medical service terminal, and predicting a resource scheduling chain matched with each real-time medical scene through a second-level scheduling model matched with different medical scenes; virtual resource mapping of the hospital model is constructed according to the resource scheduling chain, scheduling scheme execution effects of different candidate diagnosis and treatment paths in the resource scheduling chain are simulated in real time through a digital-intelligent twinborn technology, and a finally adopted target diagnosis and treatment path and a corresponding scheduling scheme are selected by a user with corresponding authority; therefore, intelligent scheduling of the medical resources is realized, and the medical resource scheduling efficiency is improved.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application relate to the field of data processing, and more particularly, embodiments of the present application relate to a hospital resource scheduling system and method based on a spatio-temporal graph neural network. BACKGROUND

[0002] The existing hospital operation system mainly includes a hardware layer and a software layer. The existing medical resource scheduling mode of the software layer has certain defects and is difficult to match the dynamics and complexity of the medical scene.

[0003] In the medical resource scheduling, the resource scheduling rules are mainly implemented through fixed code logic or simple configuration files, and the scheduling strategy is completely based on preset static rules, such as assigning a nuclear magnetic device according to the registration order, and manual experience judgment, which cannot dynamically adjust according to the sudden demand. For example, when the nuclear magnetic device needs to frequently replace the coil due to the switching of the examination project, the existing rules cannot identify the idle gap of the device in real time and assign the matching examination task, resulting in a high idle rate in the use of the device. The emergency patient cannot automatically adjust the examination priority and device allocation scheme due to the examination conflict of the previous barium meal examination, resulting in the extension of the patient waiting time. Therefore, it is urgent to design a new technical solution to solve the technical problems of scheduling rule solidification and dependence on manual experience in the above medical resource scheduling process. SUMMARY

[0004] In the present context, embodiments of the present application aim to provide a hospital resource scheduling system and method based on a spatio-temporal graph neural network, which can solve the technical problems of scheduling rule solidification and dependence on manual experience in the above medical resource scheduling process.

[0005] In a first aspect of the embodiments of the present application, a hospital resource scheduling system based on a spatio-temporal graph neural network is provided, comprising: a data acquisition layer, configured to deploy a multi-modal sensor network in a medical device and a medical service terminal; acquire multi-dimensional operation data of the medical device and the medical service terminal through the multi-modal sensor network; wherein the multi-dimensional operation data at least includes: device running state, device service type, and failure precursor signal; the multi-modal sensor network at least includes: sensors matched with different medical devices or medical service terminals, and the sensor type is associated with the deployed position and the device type; The intelligent decision layer is configured to determine real-time medical scenes where the medical devices and the medical service terminals are located based on multi-dimensional operation data, the medical devices, and geographical positions where the medical service terminals are located, determine resource scheduling chains matched with each real-time medical scene through a first scheduling model based on a spatio-temporal graph neural network, and determine next medical service nodes matched with current patients served by the medical devices and the medical service terminals through a second scheduling model matched with different medical scenes. The collaborative scheduling layer is configured to construct virtual resource mapping of a hospital model according to the resource scheduling chains, simulate execution effects of scheduling schemes of different candidate diagnosis and treatment paths in the resource scheduling chains in real time through digital twin technology, and push the execution effects to the medical service terminals or a cloud service platform, so that users with corresponding permissions select a target diagnosis and treatment path and a corresponding scheduling scheme finally adopted.

[0006] In a second aspect of the embodiments of the present application, a hospital resource scheduling method based on a spatio-temporal graph neural network is provided, which includes the following steps. Multi-dimensional operation data of the medical devices and the medical service terminals are collected through a multi-modal sensor network. Real-time medical scenes where the medical devices and the medical service terminals are located are determined based on multi-dimensional operation data, the medical devices, and geographical positions where the medical service terminals are located through a first scheduling model based on a spatio-temporal graph neural network, resource scheduling chains matched with each real-time medical scene are predicted through a second scheduling model matched with different medical scenes, and next medical service nodes matched with current patients served by the medical devices and the medical service terminals are determined. Virtual resource mapping of a hospital model is constructed according to the resource scheduling chains, execution effects of scheduling schemes of different candidate diagnosis and treatment paths in the resource scheduling chains are simulated in real time through digital twin technology, and the execution effects are pushed to the medical service terminals or a cloud service platform, so that users with corresponding permissions select a target diagnosis and treatment path and a corresponding scheduling scheme finally adopted.

[0007] In a third aspect of the embodiments of the present application, a terminal device is provided, comprising at least one processor, a memory and an input-output unit; wherein the memory is configured to store a computer program, and the processor is configured to invoke the computer program stored in the memory to execute the hospital resource scheduling method based on the spatio-temporal graph neural network according to any one of the first aspect.

[0008] In a fourth aspect of the embodiments of the present application, a computer-readable storage medium is provided, comprising instructions which, when executed on a computer, cause the computer to execute the hospital resource scheduling method based on the spatio-temporal graph neural network according to any one of the first aspect.

[0009] In a fifth aspect of the embodiments of the present application, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the hospital resource scheduling method based on the spatio-temporal graph neural network according to any one of the first aspect.

[0010] According to the hospital resource scheduling system and method based on the spatio-temporal graph neural network of the embodiments of the present application. Specifically, the data acquisition layer is configured to deploy a multi-modal sensor network in medical devices and medical service terminals; multi-dimensional operation data of the medical devices and the medical service terminals are collected through the multi-modal sensor network; wherein the multi-dimensional operation data at least includes: device operation state, device service type, failure precursor signal; the multi-modal sensor network at least includes: sensors matched with different medical devices or medical service terminals, the sensor type is associated with the deployed location and the device type. The intelligent decision-making layer is configured to determine the real-time medical scene where the medical devices and the medical service terminals are located based on the multi-dimensional operation data, the geographical location where the medical devices and the medical service terminals are located, through a first-level scheduling model based on the spatio-temporal graph neural network, and predict the resource scheduling chain matched with each real-time medical scene through a second-level scheduling model matched with different medical scenes, to determine the next medical service node matched with the current patient served by the medical devices and the medical service terminals; the resource scheduling chain at least includes: at least one candidate diagnosis and treatment path, each candidate diagnosis and treatment path includes: a plurality of medical service nodes corresponding to the current patient, diagnosis and treatment equipment used by each medical service node, doctor type, drug type. The collaborative scheduling layer is configured to construct a virtual resource mapping of a hospital model according to the resource scheduling chain, to simulate the scheduling scheme execution effect of different candidate diagnosis and treatment paths in the resource scheduling chain in real time through digital twin technology, and to push to the medical service terminal or the cloud service platform, so that a user with corresponding authority selects the target diagnosis and treatment path and the corresponding scheduling scheme finally adopted. Through the embodiments of the present application, intelligent scheduling of medical resources can be realized, resource demand prediction and dynamic adjustment can be completed in time, medical resource scheduling efficiency can be improved, and medical service configuration time can be shortened. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 A structure schematic diagram of a hospital resource scheduling system based on a spatio-temporal graph neural network is shown in the present application. Figure 2 A flow schematic diagram of a hospital resource scheduling method based on a spatio-temporal graph neural network is shown in the present application. Figure 3 A structure schematic diagram of a medium is shown in an embodiment of the present application. DETAILED DESCRIPTION

[0012] Reference will be made to the following Figure 1 , Figure 1 A flow schematic diagram of a hospital resource scheduling system and method based on a spatio-temporal graph neural network is provided in an embodiment of the present application. It should be noted that the embodiments of the present application can be applied to a variety of medical operation scenarios.

[0013] In view of at least one technical problem introduced above, the embodiments of the present application provide a hospital resource scheduling system and method based on a spatio-temporal graph neural network.

[0014] Figure 1 A structure schematic diagram of a hospital resource scheduling system based on a spatio-temporal graph neural network is shown in an embodiment of the present application. The system at least includes the following modules: The data acquisition layer is used to deploy a multi-modal sensor network in medical devices and medical service terminals; and multi-dimensional operation data of the medical devices and medical service terminals are collected through the multi-modal sensor network.

[0015] In the embodiments of the present application, the multi-dimensional operation data at least includes: device running state, device service type, and failure precursor signal. Further optionally, the multi-modal sensor network at least includes: sensors matched with different medical devices or medical service terminals, and the sensor type is associated with the deployed position and the device type.

[0016] The core function of the data acquisition layer is to build a perception network covering various medical devices and service terminals in the hospital, and to realize real-time capture and transmission of various dynamic information in the medical operation process. Conceptually, it breaks through the scattered and isolated data acquisition mode in the traditional medical system, integrates the originally fragmented device running, service development and other information into continuous and complete data streams through systematic deployment of perception devices, and provides a basic support for subsequent intelligent decision-making. Its working principle is based on the wide application of Internet of Things technology, and with the help of various sensors and transmission protocols, the state of physical entities in the medical scene is converted into processable digital signals, and then transmitted to the upper modules of the system through the network, forming a mapping from the physical world to the digital space.

[0017] In the implementation process, the deployment of the data acquisition layer needs to be combined with the actual layout and equipment distribution characteristics of the hospital, and differentiated sensing strategies are formulated for different departments and medical areas with different functions. For example, around the MRI equipment in the imaging department, the deployment of sensors will focus on the physical state monitoring needs of the equipment during operation. In the diagnosis and treatment area of the emergency department, the sensing equipment focuses more on the real-time capture of patient flow and emergency situations. Through such targeted deployment, the comprehensiveness and accuracy of sensing are ensured, avoiding information omission or redundant collection, while taking into account the particularity of the medical environment, such as the stability of equipment operation, the protection of patient privacy, etc., so that the sensing process neither interferes with normal medical services nor efficiently obtains the required information. In addition, the data acquisition layer can also cover supporting medical facilities or equipment, such as first aid equipment in subway stations, bus stations, and train stations, ambulances, school medical offices, community medical offices, and health clinics.

[0018] The multi-modal sensor network is the key carrier for the data acquisition layer to realize its functions. Its concept emphasizes the diversity and adaptability of sensor types, that is, according to the characteristics of different medical equipment or service terminals, the type of sensor that matches them is selected, so that each sensor can fully exert its sensing advantages in a specific scenario. This network is not simply a stack of sensors, but through a unified communication protocol and data standard, different types of sensors are connected as an organic whole to realize the coordinated collection and transmission of data. Its working principle is to use the sensitive characteristics of various sensors to specific physical quantities, such as some sensors are good at capturing mechanical vibrations of equipment, and others are suitable for monitoring subtle changes in biological signs. Through reasonable collocation, different dimensions of information needs in medical operations are covered.

[0019] In terms of implementation details, the construction of the multi-modal sensor network needs to fully consider the characteristics of the type and deployment location of medical equipment. The sensors connected to imaging equipment will focus on monitoring the coil state, mechanical structure stability, and other parameters directly related to the performance of the equipment during operation; while the sensors deployed at service terminals will focus more on patient interaction behavior, service request response status, and other information. At the same time, the installation method of the sensor will be optimized according to the structural characteristics of the equipment to avoid affecting the normal operation of the equipment, such as using non-invasive installation methods or adapting to the original interface of the equipment. This targeted deployment enables the sensor network to accurately capture the unique state information of each device and terminal, providing high-quality raw materials for subsequent data processing.

[0020] Further optionally, the sensors connected to the medical devices at least include: piezoelectric sensors connected to the imaging devices to monitor coil replacement pressure peaks in real time to predict mechanical failures; fiber optic strain sensors connected to the gantries to detect device vibration amplitudes to prevent image blurring; six-axis force sensors connected to surgical instruments to feedback surgical operation forces to prevent tissue damage; ultrasonic flow meters connected to life support devices to monitor oxygen flow deviations to ensure ventilation safety; millimeter wave radars connected to environmental perception sensors to track patient movement trajectories and trigger diversion alerts in case of congestion; flexible skin electrodes connected to wearable devices to continuously collect ECG signals to realize ST segment elevation real-time alerts; and ultraviolet intensity sensors connected to disinfection devices to monitor ultraviolet lamp tube attenuation to ensure disinfection effectiveness meets standards.

[0021] Multi-dimensional operation data is the sum of various information collected by the multi-modal sensor network in the data collection layer. Its concept covers all dynamic data related to device operation and service development in the medical operation process. These data collectively constitute a digital portrait reflecting the real-time operation status of the hospital. The principle is to convert device operating parameters and service execution conditions in the physical world into continuously updated digital information through sensors. The value of these information lies not only in recording the current status, but also in reflecting potential problems or needs through continuous trends, providing a basis for system prediction.

[0022] From the implementation details, the collection content of multi-dimensional operation data will be refined according to different sensing objects. The data of device operation status will include the start and stop time of the device, the energy consumption change during operation, and the working parameters of key components, which can intuitively reflect whether the device is in normal working state. The data of device service type is related to the current diagnosis and treatment projects and service objects of the device, helping the system understand the direction of resource use. The data of failure precursor signals is more subtle, which may include abnormal vibration and slight temperature fluctuation during device operation. These signals are often early signs of device failure. By capturing and analyzing them, potential problems can be detected in advance.

[0023] Through the synergy of the data acquisition layer, multi-modal sensor network, and multi-dimensional operation data, the system realizes comprehensive and real-time perception of the medical operation state. The technical effect is reflected in breaking the limitations of information collection in traditional medical systems, making dynamic changes that were originally difficult to capture perceptible, recordable, and analyzable. This improvement in perception capability puts various resource states and service processes in hospital operations under real-time monitoring, providing precise and timely data support for subsequent intelligent scheduling, as well as data basis for equipment maintenance and service optimization, promoting the transition of hospital operations from experience-based judgment to data-driven.

[0024] The intelligent decision layer is used to determine the real-time medical scene of each medical device and medical service terminal based on multi-dimensional operation data, medical devices, and the geographical location of each medical service terminal, through a first-level scheduling model based on a spatio-temporal graph neural network, and to predict the resource scheduling chain matched for each real-time medical scene through a second-level scheduling model matched for different medical scenes, to determine the next medical service node matched for the current patient served by the medical device and the medical service terminal.

[0025] The intelligent decision layer is the core processing hub of the hospital resource scheduling system based on a spatio-temporal graph neural network. It receives multi-dimensional operation data transmitted by the data acquisition layer, combines geographical location information of medical devices and medical service terminals, and realizes dynamic scheduling decisions for medical resources through intelligent model operations. Its core principle is to use the characteristics of spatio-temporal graph neural networks to convert various elements in the medical scene into nodes and associated relationships in the network, dynamically analyze these relationships, accurately determine the current medical scene, and call the corresponding scheduling model to generate an adaptive medical service path. In the implementation process, the intelligent decision layer does not simply execute preset instructions, but continuously learns the rules in real-time data to optimize the judgment and prediction capabilities of the model, enabling the decision-making process to adapt to complex and changing medical environments, such as automatically adjusting resource allocation logic during peak emergency periods or quickly switching to backup plans in the event of device anomalies.

[0026] Medical devices cover all kinds of instruments and devices used for diagnosis, treatment and monitoring in hospitals, from large imaging equipment (such as MRI, CT machine) to small monitoring instruments (such as ECG monitor), to various instruments used in surgery. They are the material basis for the development of medical services. These devices not only have specific functional properties, but also are associated with specific geographic location information, such as MRI equipment located in the imaging department, anesthesia machine in the operating room, etc. Their running status and use are transmitted in real time to the intelligent decision-making layer through the data collection layer, becoming an important basis for decision analysis. Medical service terminals are interactive nodes that connect patients and medical systems, including self-service registration machines in outpatient departments, operation terminals in nurse stations, and call equipment in patient rooms. They can not only receive patient service requests, but also deliver system-generated dispatch instructions, which are key connection points in the medical service process. Their geographic locations are usually distributed in various service areas of the hospital, ensuring that patients and medical staff can conveniently obtain or send information.

[0027] Medical scenarios are a comprehensive description of the environment and state of various medical activities in hospitals. It is worth noting that real-time medical scenarios are dynamic scenario states based on multi-dimensional operation data. Different medical scenarios have differentiated characteristics according to the nature of medical activities, participants, resource requirements, etc. For example, in the emergency scenario, patients are mostly sudden emergencies, requiring rapid mobilization of rescue equipment and specialist doctors. The outpatient scenario mainly focuses on routine diagnosis and treatment, with relatively stable processes and resource allocation focusing on efficiency and order. The surgery scenario involves the coordination of operating rooms, anesthesia equipment, surgical teams, and other factors, with extremely high requirements for time integration. The imaging examination scenario revolves around the use of various imaging equipment, requiring coordination of patient queuing and equipment operation rhythm. These medical scenarios are not fixed, but dynamically change with patient flow, equipment status, and time, such as the outpatient scenario temporarily presenting a busy state similar to an emergency due to a sudden batch of patients at a specific time.

[0028] In the embodiments of the present application, the resource scheduling chain at least includes: at least one candidate diagnosis and treatment path, each candidate diagnosis and treatment path including: a plurality of medical service nodes corresponding to the current patient, diagnosis and treatment equipment used by each medical service node, doctor type, and drug type.

[0029] The resource scheduling chain is a complete diagnosis and treatment resource configuration scheme generated by the intelligent decision layer for a specific patient or medical scene, which is used to connect various resources required by the patient from entering the medical system to completing the diagnosis and treatment. The candidate diagnosis and treatment path contained in each resource scheduling chain is a variety of possible diagnosis and treatment schemes generated by the system according to the patient's condition, current resource state and other factors. These schemes are not randomly combined, but are reasonable paths based on medical standards, resource conditions and historical experience. For example, for a patient suspected of having a heart attack, the system may generate two candidate diagnosis and treatment paths. One path is to first perform an emergency electrocardiogram examination, and then enter the catheter room for interventional treatment. The other path is to prepare the surgical team while performing the electrocardiogram examination. The two paths correspond to different resource allocation rhythms to adapt to different emergency levels.

[0030] Each medical service node in the candidate diagnosis and treatment path represents a specific link or step in the diagnosis and treatment process. These nodes are connected in sequence according to the diagnosis and treatment logic to form a complete diagnosis and treatment process. Each node is clearly associated with the required diagnosis and treatment equipment, doctor type and drug type. For example, the emergency electrocardiogram examination node requires an electrocardiogram machine, an emergency doctor and related recording consumables, while the interventional treatment node requires catheter room equipment, a cardiologist and anticoagulant drugs. The arrangement of these nodes not only conforms to the specifications of the medical process, but also can be flexibly adjusted in sequence or content according to real-time conditions. For example, when a device temporarily fails, the system will automatically replace the node that depends on the device with a node that uses other backup devices to ensure the continuity of the entire diagnosis and treatment path.

[0031] Through the operation of the intelligent decision layer, medical equipment, medical service terminals and various medical scenes form a dynamic response linkage. The generation of resource scheduling chains and candidate diagnosis and treatment paths no longer relies on manual planning, but is automatically optimized based on real-time data and intelligent models, so that each patient can obtain a diagnosis and treatment scheme that adapts to their condition and current resource state. This approach not only makes the use of medical resources more reasonable, reducing unnecessary waiting and resource waste, but also ensures that patients can be smoothly guided to the next appropriate medical service node in complex medical scenes, regardless of whether they are in the emergency, outpatient or surgery stages, promoting the entire medical service process to operate more efficiently and orderly. This fundamentally changes the traditional reliance on manual coordination, which is lagging and cumbersome, making medical services more targeted and timely.

[0032] As an embodiment, the intelligent decision layer is further configured to: adopt a multi-objective evolutionary algorithm, take the shortest patient waiting time and the highest device utilization as the multi-objective optimization core, iteratively update the scheduling rules of the first-level scheduling model based on historical hospital operation data, and iteratively update the scheduling rules of the second-level scheduling model based on local historical department operation data and a specialized medical experience database; and share model parameters of the first-level scheduling model and the localized second-level scheduling model through federated learning. In the embodiment of the application, the second-level scheduling model at least includes: an emergency resource allocation model, an operating room scheduling model, an imaging room scheduling model, an admission and discharge resource allocation model, and an outpatient specialized resource allocation model.

[0033] It can be understood that in the medical resource dynamic scheduling scenario, the emergency resource allocation model, the operating room scheduling model, the imaging room scheduling model, the admission and discharge resource allocation model, and the outpatient specialized resource allocation model are core scheduling tools for department localization needs, each of which focuses on a specific medical resource scenario, takes the shortest patient waiting time and the highest resource utilization as the optimization target, and realizes precise scheduling in combination with department historical operation data and specialized experience.

[0034] The emergency resource allocation model focuses on the high emergency, high priority, and strong resource constraint scenarios of the emergency department. The core is to realize the dynamic matching of rescue resources, manpower, and space. The dispatch objects include emergency medical personnel, rescue equipment, emergency beds, and ambulance reception resources. The key logic is to set priorities based on patient triage levels (such as ESI 1-5 levels) to reduce the waiting time of critically ill patients, combine historical data to predict peak traffic to reserve resources in advance, and dynamically coordinate resource conflicts (such as evaluating the condition of patients in observation beds to move space when rescue beds are full). The local iteration basis includes emergency historical triage data, rescue success rate, etc. Further, obtaining the emergency resource allocation model needs to be based on the full operation data of the emergency scene, including patient triage records (such as ESI classification, disease type), real-time state data of resources (medical staff scheduling, rescue equipment utilization rate, bed turnover), historical emergency event handling records (such as multiple critically ill patients simultaneously receiving treatment cases), etc. On the basis of data integration, it is necessary to combine the core principles of emergency and specialist diagnosis and treatment specifications to determine the minimum delay risk of critically ill patients and the minimum resource idle rate as the optimization goal, and select a multi-objective evolutionary algorithm that adapts to high dynamic scenarios as the core framework. Through deep learning algorithms to simulate the performance of different resource scheduling schemes (such as rescue bed allocation, consultation expert allocation) in historical scenarios, rules that take into account priority and efficiency are selected. After the model is built, it needs to access real-time data interfaces such as emergency triage systems and equipment Internet of Things monitoring systems to ensure that it can dynamically perceive changes in resource supply and demand, and continuously iterate through indicators such as rescue response time and resource conflict resolution efficiency in actual operation, while integrating the practical experience of emergency medical personnel (such as special disease resource demand preferences) to form a scheduling logic that adapts to local emergency characteristics.

[0035] The operating room scheduling model optimizes the scheduling efficiency of operating rooms, medical teams, instruments, etc. according to the characteristics of the operation. The key logic includes planned operation scheduling (formulating plans by combining operation types, estimated duration, etc., interspersing long and short operations to balance the load), emergency operation insertion (guaranteeing emergency needs by "flexible buffer time" or adjusting low-priority operations), and team coordination optimization (matching doctor expertise with operation type). Iteration is based on operation duration deviation rate, operating room idle time, etc. Further optionally, the construction of the operating room scheduling model begins with deep analysis of the entire operation process data, including historical operation records (type, estimated duration and actual duration deviation, participating team configuration), operating room resource data (type, quantity, maintenance period), emergency operation insertion records (reason, delay duration), etc. Based on these data, the core constraint conditions of operation scheduling need to be determined first. For example, the priority of elective surgery and emergency surgery, the professional matching degree of the operation team, and the adaptability of equipment function (such as negative pressure operating room for infectious surgery), and then a double module of planned scheduling and emergency adjustment is constructed through a multi-objective evolutionary algorithm: the planning module generates a basic plan by learning efficient patterns such as "long and short operation combination" and "similar operation centralized arrangement" in historical scheduling; the emergency module simulates the impact of different emergency operation insertion on the original plan, optimizes the setting of flexible buffer time and the adjustment strategy of low-priority operations. The model needs to be connected with the hospital operation reservation system and the anesthesia management system to realize the real-time synchronization of preoperative examination completion, team on-site state, etc. information, and continuously optimize according to the operating room utilization rate, planned execution deviation rate, etc. indicators, combined with the feedback of surgeons on the rationality of scheduling, to finally form a scheduling logic that takes into account both planning and flexibility.

[0036] The image room scheduling model is aimed at the scene of multi-source demand for image examination and strong equipment specificity, with the core of optimizing the examination sequence of CT, MRI and other equipment. The scheduling objects include image equipment, technician manpower, examination appointment queue, etc. The key logic is the precise matching of equipment and examination type (such as MRI adapting to soft tissue imaging), priority layered scheduling (emergency priority, inpatient cycle ordering) and efficient use of equipment (concentrating the same type of examination to reduce debugging time). The iteration basis includes the average daily examination volume of equipment, patient waiting time distribution, etc. Further optionally, the acquisition of the image room scheduling model needs to be supported by the whole chain data of image examination, including equipment basic data (type, function parameters, daily maximum load), historical examination records (patient source, examination site, time, failure reason), technician operation data (proficiency, operation time consumption), etc. On the basis of data, the matching rules of equipment and examination type need to be clarified (such as CT suitable for rapid positioning of hemorrhage, MRI suitable for soft tissue imaging), and then the layered scheduling logic is constructed through multi-objective evolutionary algorithm: the highest priority channel is set for emergency examination (such as stroke CTA), the time window is divided according to the diagnosis and treatment cycle (such as preoperative 1 day) for inpatient examination, and the outpatient examination is dynamically sorted combined with the appointment time and waiting time. The model needs to integrate equipment state monitoring (such as on / off state, maintenance state), patient examination reminding system, to reduce the empty window period caused by equipment failure or patient unpreparedness; at the same time, through the analysis of the data such as the average daily effective on-time of equipment, the waiting time distribution of patients from different sources, combined with the optimization suggestions of image technicians on examination process (such as concentrating the same part examination to reduce debugging time), the scheduling rules are continuously iterated to ensure that the model adapts to the characteristics of imaging equipment and multi-source examination demand.

[0037] The admission and discharge resource allocation model focuses on improving the turnover efficiency of hospital beds. The scheduling objects cover hospital beds (classified by department, disease, etc.) and discharge process resources (settlement window, cleaning team, etc.). The key logic includes bed demand prediction and pre-allocation (coordinated across departments in advance in combination with outpatient surgery volume), discharge plan-driven bed release (early start of discharge evaluation to shorten the idle interval), and special bed priority protection (such as isolation beds for infectious disease patients). Iterative basis includes bed turnover days, admission waiting time, etc. Further optionally, the construction of the admission and discharge resource allocation model needs to integrate bed full life cycle data, including department bed types (such as ICU, general bed, isolation bed), historical usage records (admission / discharge time, turnover days), admission waiting data (waiting time for different diseases, cross-department bed transfer records), discharge process data (settlement time, cleaning time), etc. Based on these data, the model needs to first establish a bed demand prediction module. By learning the outpatient surgery appointment volume, emergency hospital conversion rate, and historical same-period bed usage patterns, the bed shortage in each department can be predicted 24-48 hours in advance; then build a bed allocation and release mechanism: set rigid priority rules for special resources such as isolation beds and ICU beds (such as infectious disease patients using isolation beds first), and for general beds, simulate the feasibility and efficiency of cross-department allocation through algorithms; at the same time, embed the discharge plan in the model, estimate the length of stay through preoperative evaluation, and start the discharge preparation process simultaneously after surgery to shorten the bed idle interval. The model needs to interface with the hospital HIS system, settlement system, and logistics cleaning system to realize real-time information exchange, and optimize parameters based on bed turnover rate, admission waiting time, cross-department bed transfer success rate, and other indicators, combined with the feedback of hospital staff on bed allocation rationality, to form a scheduling logic that adapts to the hospital's bed supply and demand characteristics.

[0038] The outpatient specialist resource allocation model is designed to optimize the allocation of physician visits, examination rooms, and other resources in light of the large fluctuations in patient flow and the differentiated needs of different specialists. The key logic includes dynamic matching of visit resources (adjusting the number of physicians and the type of appointments based on the flow pattern), patient diversion and path optimization (reducing stress through pre-consultation and hierarchical diagnosis and treatment), and special needs adaptation (such as increasing the number of accompanying doctors for geriatric patients). Iterative criteria include the average daily number of patients seen by each specialist, waiting time, and other factors. Further, the acquisition of the outpatient specialist resource allocation model requires data on the operation of specialist clinics, including historical flow data (by time period and by disease type), physician resource data (visit times, areas of expertise, and efficiency), examination room and auxiliary equipment usage records (such as the use of nebulizers in pediatric departments and slit lamps in ophthalmology departments), and patient waiting data (the interval between registration and treatment, and the rate of missed appointments). Based on this data, the model first identifies the flow patterns of each specialist (such as the high volume of pediatric patients in the morning on Mondays and the high volume of patients with chronic conditions in the afternoon), and then uses a multi-objective evolutionary algorithm to develop dynamic resource matching rules: automatically increasing the number of visiting physicians and the number of open examination rooms during peak periods, adjusting the ratio of specialist appointments to regular appointments, and guiding patients with mild symptoms to general clinics through a pre-consultation system to reduce the pressure on specialist clinics. The model needs to interface with the outpatient registration system and the triage and appointment system to real-time sense changes in patient flow, and combine specialist preferences for visits (such as a specialist who is good at handling complex cases) and patient feedback on waiting experiences (such as the demand for accompanying doctors among elderly patients), and continuously iterate based on indicators such as the average daily number of patients seen by each specialist and waiting time, to ultimately develop resource scheduling logic that adapts to the characteristics of different specialist clinics and patient needs.

[0039] These models all use a multi-objective evolutionary algorithm as the core of optimization, iteratively adjust the scheduling rules based on historical data, and use federated learning to share parameters across departments (such as the use of emergency urgency evaluation logic in the imaging room priority algorithm), ultimately forming a globally coordinated and locally precise medical resource scheduling system.

[0040] Specifically, the intelligent decision-making layer uses a multi-objective evolutionary algorithm as the core optimization tool in the process of dynamic scheduling. The core idea is to find the optimal solution that balances the two core objectives of minimizing patient waiting time and maximizing equipment utilization in the medical context. This algorithm does not simply optimize a single objective, but rather simulates the selection, crossover, and mutation mechanisms in biological evolution to select a solution that can both reduce patient waiting time and improve equipment utilization efficiency from a large number of possible scheduling solutions, allowing different objectives that may conflict with each other to form a synergy. For example, when an imaging department device is temporarily idle, the algorithm will automatically evaluate whether to prioritize scheduling patients with longer waiting times or to reserve the device for possible emergency patients. Through learning from similar scenarios in historical data, the algorithm generates a scheduling strategy that takes both factors into account.

[0041] For the autonomous iteration of scheduling rules in the primary scheduling model, the intelligent decision-making layer makes full use of the rules embedded in historical hospital operation data, which includes patient flow at different times, usage peaks of various equipment, resource allocation records of different departments, etc. Through in-depth analysis of these data, the algorithm can identify parts of the scheduling rules that need to be optimized, such as automatically adjusting patient diversion logic during outpatient peak hours, or planning the usage sequence of backup equipment in advance during equipment maintenance periods, so that the rules can dynamically adapt to changes in hospital operation patterns. The localized autonomous iteration of the secondary scheduling model focuses more on combining the characteristics of each department, relying on local historical operation data and specialized medical experience databases, to make the scheduling rules more tailored to the actual needs of the department. Taking the operating room as an example, its scheduling model optimizes the sequence and interval of surgery arrangements based on historical surgery duration, resource consumption of different surgery types, and the coordination rhythm of the medical team, while incorporating the clinical experience of surgeons, such as reserving more preparation time for complex surgeries to ensure that the rules comply with departmental operation specifications and improve overall efficiency.

[0042] The various specific models included in the secondary scheduling model each play a role in different medical scenarios. The emergency resource allocation model focuses on the rapid deployment of resources in emergency scenarios, and can generate a complete resource chain from reception to treatment based on the patient's condition, the location of available rescue equipment, and the real-time status of specialized doctors. The operating room scheduling model coordinates preoperative preparation, surgery execution, and postoperative handover around the use of operating rooms to avoid gaps or conflicts between surgeries. The imaging room scheduling model balances the time needs of different types of examinations and equipment load based on the characteristics of imaging equipment, such as reasonably arranging the sequence of MRI and CT examinations to reduce time waste caused by equipment switching. The admission and discharge resource allocation model focuses on bed turnover efficiency, and plans ahead for bed cleaning, preparation, and allocation based on patient discharge plans and admission needs. The outpatient specialist resource allocation model optimizes the matching of outpatient doctors, examination rooms, and patients to reduce patient waiting time between different specialists.

[0043] To enable these secondary scheduling models to maintain local adaptability while sharing overall optimization experience, the intelligent decision layer introduces a federated learning mechanism. When training locally, the secondary scheduling models of each department will only upload encrypted parameter update information generated during model optimization to the shared platform, without revealing specific patient data or department operation details. For example, when the emergency department of a hospital optimizes the response speed of the emergency resource allocation model through local data, the emergency department models of other hospitals can learn from this optimization experience and improve their model performance without accessing specific case data. This approach not only protects data privacy but also breaks down information barriers between different departments and hospitals, allowing the primary scheduling model and each secondary scheduling model to evolve in collaboration, enabling the scheduling capability of the entire system to be rooted in local reality while absorbing global optimization experience, ultimately achieving dual improvement in efficiency and adaptability of medical resource scheduling.

[0044] As an embodiment, the intelligent decision layer adopts a multi-objective evolutionary algorithm, taking the shortest patient waiting time and the highest device utilization as the core of multi-objective optimization. When iteratively optimizing the scheduling rules of the primary scheduling model based on historical hospital operation data, the following steps are used: Based on the improved non-dominated sorting genetic algorithm (VA-NSGA-III), on the basis of the reference point layering of NSGA-III, a vector angle screening strategy is introduced. By calculating the weight vector direction of the device utilization and patient waiting time solution set in the multi-objective optimization function space of the primary scheduling model, similar angle solutions are removed to ensure the uniform distribution of the Pareto front solution set. In combination with the priority of the emergency scene, the weight vector distribution in the multi-objective optimization function space is adjusted in real time; Through time series clustering analysis of the hospital operation conflict case library in historical data, high-frequency fault modes are identified and a robust rule library is generated, which is embedded in the algorithm constraints of the primary scheduling model; When the sensor detects a failure precursor signal indicating device anomalies, in combination with the multi-objective optimization function space and algorithm constraints, the retraining of the primary scheduling model is automatically triggered, the scheduling rules of the primary scheduling model are dynamically optimized, and the sub-objective of minimizing the fault recovery time is added to the standby device scheduling scheme.

[0045] Specifically, when the intelligent decision layer iteratively optimizes the primary scheduling model using a multi-objective evolutionary algorithm, the core is to find the optimal solution that dynamically balances patient waiting time and device utilization through VA-NSGA-III. This algorithm does not simply optimize a single objective, but simulates the selection and screening mechanism in biological evolution to generate adaptive scheduling rules in complex medical scenarios.

[0046] In specific implementation, first, the initial optimization direction of the two core targets of device utilization and patient waiting time is set based on the reference point stratification of NSGA-III. Unlike the general processing of solution set in traditional algorithms, VA-NSGA-III introduces a vector angle screening strategy to identify and eliminate solutions with similar angles by calculating the direction of the weight vector solution set in the multi-objective optimization function space. This means that during the generation of scheduling rules, the algorithm will actively avoid multiple scheduling schemes with similar effects, ensuring that the final retained Pareto front solution set can uniformly cover different optimization directions, including schemes focusing on reducing patient waiting time and schemes focusing on improving device utilization, providing more comprehensive reference dimensions for subsequent decision-making. At the same time, VA-NSGA-III algorithm adjusts the weight vector distribution dynamically combined with the priority of emergency scenes, for example, when the system identifies a heart attack patient or other emergency situations, it will automatically increase the weight of patient waiting time in the optimization function, making the generated scheduling rules more inclined to respond quickly to emergency needs.

[0047] To make the scheduling rules robust in complex scenarios, VA-NSGA-III algorithm performs time series clustering analysis on the historical data of hospital operation conflict case library. These cases cover various problems such as "barium meal conflict" and "scheduling interruption caused by temporary equipment failure", etc. By analyzing the occurrence time, scene characteristics and solution of these cases, the algorithm can identify high-frequency fault patterns that repeatedly occur and convert these patterns into a robust rule library. For example, when it is found that a certain type of examination equipment is prone to efficiency decline due to continuous use at a specific time period, the rule library will automatically record this rule and embed it into the algorithm's constraint conditions, so that the primary scheduling model naturally avoids the time window or resource allocation method that may cause conflicts when generating scheduling rules.

[0048] When the sensor captures the abnormal failure signal of the equipment, the algorithm will quickly start the retraining mechanism of the primary scheduling model. This process is not isolated adjustment, but reorganizes the logical association in the scheduling rules in combination with the current multi-objective optimization function space and existing algorithm constraints. For example, when the coil sensor of the nuclear magnetic equipment detects abnormal current fluctuation, VA-NSGA-III algorithm will minimize the fault recovery time as a new sub-target in the standby equipment scheduling scheme, ensuring the balance between patient waiting time and device utilization, while prioritizing how to quickly switch to standby equipment to reduce the impact of the fault on the overall scheduling. This dynamic strategy optimization makes the rules of the primary scheduling model no longer a fixed preset logic, but can be adjusted automatically according to the real-time changes of the equipment state, making the scheduling scheme more flexible and adaptable when dealing with unexpected situations.

[0049] Through such an iterative process, the primary scheduling model can continuously learn the rules from historical data and real-time status, and the generated scheduling rules not only meet the actual needs of the medical scene, but also find the best balance point between multiple optimization objectives. Whether it is regular outpatient scheduling or emergency treatment, the model can quickly generate adaptive solutions, avoiding long patient waiting times and reducing the idle waste of equipment resources, allowing the scheduling of medical resources to remain efficient and accurate in complex and changing scenarios.

[0050] The collaborative scheduling layer is used to construct a virtual resource mapping of the hospital model according to the resource scheduling chain, simulate the execution effect of the scheduling scheme of different candidate diagnosis and treatment paths in the resource scheduling chain in real time through digital twin technology, and push it to the medical service terminal or cloud service platform. The user with corresponding authority selects the target diagnosis and treatment path and the corresponding scheduling scheme finally adopted.

[0051] The collaborative scheduling layer is a key link connecting intelligent decision-making and actual medical service execution. Its core role is to convert the abstract candidate diagnosis and treatment paths in the resource scheduling chain into an intuitive simulation scenario through digital twin technology, providing a concrete reference for the final decision. This process is not a simple process replication, but a digital mirror based on physical resources, building a virtual space synchronized with the real hospital operation status, so that the execution effect of each candidate diagnosis and treatment path can be real-time deduced in the virtual environment.

[0052] Further optionally, the collaborative scheduling layer, when constructing a virtual resource mapping of the hospital model according to the resource scheduling chain and simulating the execution effect of the scheduling scheme of different candidate diagnosis and treatment paths in the resource scheduling chain in real time through digital twin technology, is specifically used for: constructing a physical resource digital mirror based on the hospital BIM model, taking bed mapping as a dynamic occupancy state node, and taking equipment mapping as a real-time running model with fault probability; inputting the candidate diagnosis and treatment paths, simulating the conflict probability, resource utilization rate, and time cost of each candidate diagnosis and treatment path through the digital twin engine; visualizing the simulation data of different candidate diagnosis and treatment paths based on the conflict probability, resource utilization rate, and time cost of each candidate diagnosis and treatment path, and marking the resource consumption difference and patient waiting time fluctuation range of each candidate diagnosis and treatment path.

[0053] In the embodiments of the present application, the hospital BIM model is a digital three-dimensional model of the hospital physical space and facilities based on building information modeling (BIM) technology. It is not simply a digitalization of architectural drawings, but integrates various information such as hospital building structure, medical equipment, pipeline system, and space layout into a three-dimensional digital carrier rich in data, realizing the whole life cycle information integration from architectural design to operation management. In terms of composition, the hospital BIM model covers the main building structure of the hospital, including the precise size and spatial relationship of walls, floors, doors, windows, etc. The layout and partition of various medical function areas, such as outpatient clinic, ward, operating room, imaging department, and laboratory department, and the medical equipment, pipeline system (such as air conditioning, water supply and drainage, electricity, medical gas pipeline), fire fighting facilities distributed in these areas, each element is endowed with detailed attribute information, such as equipment model, installation location, operating parameters, maintenance record, pipeline material, pipe diameter, direction, etc. This comprehensive information integration makes the model not only able to intuitively show the physical form of the hospital, but also able to reflect the correlation between elements and the operation logic.

[0054] In the hospital resource scheduling system based on space-time graph neural network, the hospital BIM model is the basic framework for the collaborative scheduling layer to construct virtual resource mapping. It provides spatial coordinates and structural support for the digital mirror of physical resources, allowing the digital nodes of beds, equipment, and other resources to accurately correspond to their locations in the real hospital, ensuring the spatial accuracy of virtual resource mapping. For example, when simulating the movement of a patient from the emergency room to the operating room in a certain diagnosis and treatment path, the corridors, elevators, and passages in the BIM model will become the basis for virtual reasoning, combined with the digital mirror of equipment, to realistically simulate the spatial obstacles or route optimization points that the patient may encounter during the movement.

[0055] In addition, the information richness of the hospital BIM model also supports more detailed operation simulation. For example, when simulating operating room scheduling, the size of the operating room, the layout of internal equipment, and the distance from the sterilization supply room in the model will affect the simulation accuracy of the preparation time, equipment transfer path of the operation; when planning the placement of imaging department equipment, the model can combine the room's load-bearing capacity, power configuration, radiation protection requirements, and other information to evaluate the feasibility of equipment installation and use. This simulation based on accurate space and attribute information enables the digital twin technology of the collaborative scheduling layer to be closer to the actual operation scenario, providing a more reliable spatial and resource basis for the evaluation of candidate diagnosis and treatment paths.

[0056] The dynamic nature of the hospital BIM model also allows it to adapt to hospital expansion and resource adjustment. When the hospital adds new equipment, adjusts the department layout, or undergoes partial renovation, the model can update the information of relevant elements in real time, ensuring that the virtual resource mapping is always consistent with the physical space, providing a continuous and reliable digital foundation for long-term operation scheduling. It can be seen that the hospital BIM model is not only a static digital copy, but also a dynamic evolving digital twin foundation that supports the entire operation management system to accurately simulate and efficiently schedule complex scenarios in the hospital.

[0057] In the above embodiment, when constructing the virtual resource mapping of the hospital model, the collaborative scheduling layer takes the hospital BIM model as the basic framework and converts the physical resources in reality into corresponding nodes in the digital space one by one. For example, a bed is no longer a simple numbered record, but is mapped to a node with a dynamic occupancy state, which can reflect the state changes of the bed, such as idle, use, and cleaning. Various medical equipment is converted into a real-time operation model with a fault probability. Its digital mirror not only contains the basic functional parameters of the equipment, but also dynamically updates the probability of possible faults according to historical fault data and real-time operation status, so that the virtual model can accurately simulate various situations that may occur in actual operation. This mapping is not a static copy, but a real-time data transmission through the data acquisition layer, which keeps the state of virtual resources and physical resources synchronized, ensuring that the digital mirror can accurately reflect the real-time operation status of the hospital.

[0058] When the intelligent decision-making layer generates candidate diagnosis and treatment paths, the collaborative scheduling layer inputs these paths into the digital twin engine to start real-time simulation of different scheduling schemes. For example, for a patient who needs to be transferred from the emergency room to the operating room, there may be two candidate paths: one is to perform preoperative examination before entering the operating room, and the other is to perform preoperative preparation and operating room allocation simultaneously. The digital twin engine will simulate the execution process of these two paths in the virtual space, considering factors such as resource connection, equipment status, personnel coordination, etc. in the simulation process, and then generating the conflict probability that each path may face, such as the overlap risk of examination equipment and operating room usage time. Resource utilization rate, such as the idle time proportion of preoperative examination equipment. Time cost, such as the total time cost from the patient arriving at the emergency room to entering the operating room, and other key information.

[0059] To make the simulation results easier to understand and compare, the collaborative scheduling layer will visualize the simulation data of different candidate diagnosis and treatment paths. In the visualization presentation, not only will the conflict probability, resource utilization, and time cost of each path be clearly displayed, but also the resource consumption difference between different paths will be marked, such as which path will save more equipment usage time, which path may need additional manpower cost, and the fluctuation range of patient waiting time will also be presented, allowing medical staff to intuitively feel the stability of different paths in time connection. This visualization is not simply a pile of data charts, but a combination of hospital space layout and process logic, presenting the connection status of each link in the form of dynamic flowchart, so that users with corresponding authority can quickly grasp the pros and cons of different paths.

[0060] Taking the examination scheduling of the imaging department as an example, when there are two candidate diagnosis and treatment paths, one is to arrange magnetic resonance examination according to the order of patient arrival, and the other is to cross-arrange according to the complexity of examination items, the digital twin engine will simulate the running rhythm of equipment, the change of patient waiting queue, the workload of technicians and other scenarios in virtual space under these two arrangement methods. Through simulation, it can be found that arranging in order may lead to long waiting time for subsequent patients after complex examination, while cross-arrangement may have conflict risk of equipment parameter switching. The collaborative scheduling layer will present these potential problems in a visual way, marking the differences in resource consumption and waiting time between the two paths, helping doctors or scheduling personnel to choose a better solution according to actual situation.

[0061] Through this simulation and visualization mechanism, the collaborative scheduling layer effectively makes up for the limitations of relying on experience judgment in traditional scheduling, making the originally abstract decision-making process perceptible and comparable. Medical staff no longer need to rely on subjective speculation to evaluate the feasibility of the path, but can make choices based on the results of real simulation in virtual space, which is more in line with actual needs, ensuring the efficiency of medical services, reducing the risk of resource conflicts and process delays, and making the finally selected target diagnosis and treatment path achieve the expected effect in actual execution, promoting the transformation of medical resource scheduling from experience-driven to data-driven.

[0062] In the embodiments of the present application, by constructing an integrated architecture of global perception, intelligent decision-making and collaborative scheduling, the fundamental change of the medical resource scheduling mode is realized. The allocation of medical resources is no longer dependent on the lag response of artificial experience, but a real-time closed loop from data collection to decision execution is formed, so that each link of medical services can be dynamically adjusted according to actual needs, the utilization of resources is more efficient, the waiting and connection time of patients in the diagnosis and treatment process is greatly shortened, and the medical resources in different departments, different equipment and different scenarios can be collaboratively linked, breaking the information barriers and resource islands existing in traditional operation, and making the configuration of medical services more accurate and forward-looking. The embodiments of the present application can realize intelligent scheduling of medical resources, timely complete resource demand prediction and dynamic adjustment, improve the efficiency of medical resource scheduling, and shorten the medical service configuration time.

[0063] As an embodiment, when the intelligent decision-making layer iterates the scheduling rules of the secondary scheduling model based on the local historical department operation data and the specialty medical experience database of each department, it is specifically used for: dynamically updating the matching relationship between each department and different medical scenarios based on the corresponding diagnosis and treatment scenarios of each department; extracting the department business characteristics of each department based on the local historical department operation data of each department, and selecting a model parameter updating mode matched with the change trend of the department business characteristics; using the selected model parameter updating mode, combining the local historical department operation data of each department and the specialty medical experience database, updating the model parameters of the secondary scheduling model matched for each department, and using a differential privacy protection algorithm to share encrypted model gradients with other departments or other medical institutions matched with the same secondary scheduling model.

[0064] Specifically, the local autonomous iteration of the intelligent decision-making layer on the scheduling rules of the secondary scheduling model is to enable the scheduling model of each department to adapt to its own business characteristics and optimize the decision logic in continuous learning, while absorbing external experience through safe parameter sharing. This process is not isolated rule adjustment, but dynamic adaptation based on department diagnosis and treatment scenarios and business characteristics, so that the model can be rooted in local reality and improve its generalization ability in collaboration.

[0065] Based on the dynamic update of the matching relationship corresponding to the diagnosis and treatment scene of each department, the essence is to form a flexible association link between the department and the medical scene. The core business of each department determines the type of scene it often faces, and these scenes will adjust with the changes in the operating state. The model needs to capture this change and update the corresponding relationship. For example, the core business of the imaging department revolves around the examination process of the imaging equipment. When the department discovers through historical data that the coil replacement frequency has reached a certain level, it means that equipment maintenance needs have become a key factor affecting operation. At this time, the association between the department and the "equipment maintenance scene" will be automatically strengthened, and the model will focus more on the rules of equipment state monitoring and backup resource allocation. The diagnosis and treatment scene of the emergency department is closely related to patient flow. When the peak regularity of patient concentration in the hospital is monitored at a certain time of the day, the department will automatically associate with the "tide flow scheduling model", so that the scheduling rules will tilt to respond to flow fluctuations, such as deploying mobile medical personnel in advance or optimizing the diagnosis room switching rhythm. This dynamic matching ensures that the focus of the model is always consistent with the current core scene of the department, avoiding the lack of adaptation caused by rigid rules.

[0066] Extracting department business features and selecting matching parameter update methods is the key to making the model iteration pace consistent with the characteristics of department data. The operation data of different departments presents different change rules, some are relatively stable, and some are volatile. The model needs to adopt differentiated learning strategies. For outpatient departments, whose business processes are stable and data changes are gentle, their business features are more reflected in slow trend adjustment. At this time, incremental learning through online gradient descent is more appropriate for fine-tuning parameters. This method does not require large-scale retraining, but only by continuously absorbing subtle changes in new data, the model can gradually adapt to the seasonal fluctuations of outpatient flow or the fine-tuning of department layout. For emergency departments, whose data fluctuates dramatically and sudden situations are frequent, the business features often show sudden patient flow or sudden changes in resource demand. Deep Q network (DQN) can better respond to this dynamic nature. It converts the scheduling process into a process similar to trial-and-error learning, using resource conflict rate as feedback signal. When there is a conflict in the operating room scheduling, it will learn effective coping strategies from actions such as delaying non-emergency surgery, and through the reward mechanism, it will strengthen such decisions, so that the rules can quickly adapt to the needs of emergency scenes.

[0067] In the model parameter updating process, the department will deeply integrate local historical operation data with specialized medical experience database, so that explicit operation data and implicit clinical experience will participate in rule optimization together. For example, the scheduling model of the operating room will not only analyze the data rules such as historical operation time and equipment use interval, but also integrate the experience of surgeons about the rhythm of operation connection, so that the updated parameters not only conform to the objective laws presented by the data, but also meet the implicit requirements of clinical operation. After the parameter updating is completed, in order to avoid data privacy leakage, the model will use differential privacy protection algorithm to encrypt the gradient information generated by the update, and only share the gradient data after the confusion with other departments or medical institutions using the same type of secondary scheduling model. This sharing is not the direct transfer of decision logic, but through the cloud platform to aggregate the optimization direction in similar scenarios, such as the scheduling models of the imaging rooms of multiple hospitals sharing convolution layer features, while retaining their own fully connected layer parameters related to local equipment characteristics. They not only absorb the common experience of different institutions in the optimization of imaging examination process, but also do not expose the specific operation details of each department.

[0068] This localized and autonomous iterative mechanism makes the rules of the secondary scheduling model always resonate with the department business. For example, the outpatient model can adapt to the changes in patient's visiting habits without interfering with the daily process. The emergency model can quickly generate a response strategy that has been verified by practice when dealing with sudden traffic. And the encrypted parameter sharing across institutions allows each department to avoid reinventing the wheel and optimize decisions based on a wider range of experience while protecting privacy. Ultimately, the scheduling rules of each department become a dynamic system that has both local characteristics and common wisdom, enabling the secondary scheduling model to respond accurately to local needs and be flexible in dealing with unknown changes in complex medical scenarios.

[0069] As an optional embodiment, the multi-modal sensor network further includes a wearable device equipped on an ambulance and a corresponding wireless sensor. In the process of collecting multi-dimensional operation data of medical devices and medical service terminals through the multi-modal sensor network, the data collection layer is specifically configured to collect patient vital sign data from the wearable device through the wireless sensor, and encrypt and transmit the patient vital sign data and real-time diagnosis and treatment information input by the ambulance doctor to the hospital edge computing node closest to the location of the ambulance and / or the cloud service platform of the destination hospital of the ambulance in real time.

[0070] Further, the intelligent decision layer, when determining the next medical service node matched by the current patient served by the medical device and the medical service terminal based on the real-time medical scene in which the medical device and the medical service terminal are respectively located, and the resource scheduling chain matched by each real-time medical scene through the secondary scheduling model matched by different medical scenes, specifically: through the hospital edge computing node and / or the cloud service platform, according to the patient vital sign data and the real-time diagnosis and treatment information, and the real-time geographic location of the ambulance, the current real-time medical scene is predicted as an emergency scene through the first-level scheduling model; through the emergency resource allocation model matched by the emergency scene, the resource demand change within the golden treatment time window is predicted to obtain the resource scheduling chain matched by the patient in the ambulance; the emergency resource allocation model is constructed based on a spatio-temporal graph neural algorithm; the predicted resource scheduling chain is compared with the current real-time medical resource operation situation of the hospital, and the corresponding candidate hospital is matched for the current patient according to the comparison result.

[0071] It can be understood that the extension of the multi-modal sensor network in the emergency scene makes the ambulance not a isolated mobile unit, but a dynamic perception node connecting pre-hospital emergency and in-hospital resources. The wearable devices equipped on the ambulance continuously capture the patient's vital signs, and these subtle physiological signals are transmitted in real time through wireless sensors, and are fused with the real-time diagnosis and treatment information (such as preliminary diagnosis and emergency measures taken) input by the on-board doctor. In order to protect data security, these information will be encrypted in real time, and then transmitted to the nearest hospital edge computing node of the ambulance, and also synchronized to the cloud service platform of the destination hospital. This dual-path transmission not only ensures the timely processing of data, but also provides an information basis for cross-hospital collaboration.

[0072] When these data arrive at the hospital edge computing node or the cloud service platform, the intelligent decision layer starts the identification and resource scheduling process of the emergency scene. The first-level scheduling model relies on the spatio-temporal graph neural network to analyze the patient's vital sign data, real-time diagnosis and treatment information, and geographic location information of the ambulance. The spatio-temporal graph neural network is good at processing data with spatial correlation and time series characteristics. Here, it regards the ambulance, patient, and hospitals along the way as nodes in the network, and the correlation between nodes reflects the distance of geographic location, the accessibility of resources, and other relationships. Through the analysis of the change of node state over time, the model can accurately determine that the current real-time medical scene is an emergency scene, and set the tone for subsequent resource scheduling.

[0073] The emergency resource allocation model based on emergency scene matching is also constructed based on the spatio-temporal graph neural algorithm. The core task of the model is to predict the resource demand change within the golden treatment time window, so as to generate an adaptive resource scheduling chain. The model will convert the patient's vital sign trend and real-time diagnosis and treatment information into specific resource needs, such as predicting whether surgical intervention is needed according to the patient's heart rate and blood oxygen changes, and then deducing the required operating room type, specialist qualifications, and essential medicines. At the same time, the model will combine the real-time location of the ambulance, calculate the time cost to reach different hospitals, and integrate these factors to generate a complete resource scheduling chain from the ambulance to the in-hospital diagnosis and treatment, clearly defining the examination and treatment nodes that the patient should link to after admission and the required resource support.

[0074] After generating the resource scheduling chain, the intelligent decision-making layer compares it with the current real-time medical resource operation situation of each hospital. For example, whether the required operating room is idle, whether the corresponding specialist is on duty, and whether the essential medicines are in stock. According to the comparison results, the system will match the most suitable candidate hospital for the current patient, and if the nearest hospital has sufficient resources and meets the treatment conditions, it will be recommended first; if the hospital is resource-constrained, the system will filter out other feasible candidate solutions based on geographical location and resource conditions.

[0075] This closed-loop process from pre-hospital perception to in-hospital resource scheduling makes emergency care no longer limited by spatial and temporal distance. For example, when a heart attack patient is transported by an ambulance, the wearable device captures abnormal changes in his electrocardiogram, and combined with the doctor's initial diagnosis, the system quickly identifies through the spatio-temporal graph neural network that it is an emergency scene that requires emergency intervention treatment. The emergency resource allocation model then predicts the demand for catheter rooms and cardiologists within the golden treatment time window, and compares the resource conditions of the hospitals along the way to match the hospital that can provide the fastest interventional treatment for the patient. Throughout the process, the real-time flow of data and dynamic analysis of the model allow the hospital resources to be prepared in advance before the patient arrives, the golden treatment time is fully utilized, and the efficiency and success rate of emergency treatment are greatly improved. This also forms a seamless and coordinated system of pre-hospital and in-hospital medical resources.

[0076] Further optionally, the intelligent decision layer, when predicting the resource demand change in the golden treatment time window through the emergency scene matching emergency resource allocation model to obtain the resource scheduling chain matched by the patient in the ambulance, is specifically configured to: take the hospital location, the geographic coordinates of the ambulance, and the patient's vital signs as nodes, take the road travel time, the hospital receiving capacity, the historical response receiving rate, and the specialty matching degree as edge weights, and construct a real-time emergency medical resource graph structure; predict the resource demand change in the golden treatment time window by processing the patient's vital sign flow (such as the ST segment elevation trend) through a ConvLSTM layer, to obtain a candidate path containing diagnosis and treatment equipment, doctor types, and drug types matched with the resource demand change; calculate the candidate hospital matching degree score corresponding to the candidate path; take the weighted sum of the department idle rate in the candidate path, the specialist doctor on-duty time, the drug inventory matching degree, and the diagnosis and treatment equipment scheduling position as the candidate hospital matching degree score; and combine the candidate path and the corresponding candidate hospital matching degree score into a resource scheduling chain.

[0077] When the intelligent decision layer predicts the resource demand change in the golden treatment time window by means of the emergency resource allocation model, the core is to construct a graph structure that can reflect the correlation of various elements in the emergency scene in real time, and generate a resource scheduling chain that adapts to the patient's demand through time series analysis of vital sign data. This process is not simply information superimposition, but deep mining of spatial correlation and time trend through algorithm, so that the resource scheduling not only fits the patient's condition change, but also meets the actual carrying capacity of the hospital.

[0078] When constructing the real-time emergency medical resource graph structure, the system takes the hospital location, the geographic coordinates of the ambulance, and the patient's vital signs as the core nodes in the graph. These nodes are not isolated, but are connected through edge weights with practical significance. The road travel time reflects the spatial distance and accessibility between the ambulance and each hospital, the hospital receiving capacity reflects the resource surplus that the hospital can currently accept new patients, the historical response receiving rate reflects the response efficiency of the hospital to emergency requests based on past data, and the specialty matching degree measures the degree of fit between the hospital department and the patient's condition. These edge weights are dynamically updated, so that the graph structure can reflect the changes in the emergency scene in real time. For example, when a hospital suddenly receives a batch of patients and the receiving capacity decreases, the corresponding edge weight will be adjusted accordingly, ensuring that the graph structure is always synchronized with the actual situation.

[0079] In processing patient vital signs flow, the ConvLSTM layer plays a key role in capturing dynamic trends in time series data, which can extract the evolution of the patient's condition from the patient's vital signs (such as the amplitude and speed of ST segment elevation). Through analysis of these rules, the system can predict the patient's needs for diagnostic and treatment equipment (such as a specific model of monitor, interventional therapy equipment), physician types (such as a cardiologist, emergency physician), and drug types (such as anticoagulants, vasodilators) within the golden treatment time window, and then generate multiple candidate paths. These candidate paths are not generated out of thin air, but are based on the likelihood of disease progression and the regular configuration logic of hospital resources, ensuring that each path has practical execution value.

[0080] When calculating the candidate hospital matching score, the system will consider various resource factors involved in the candidate path. Department idle rate reflects the current busy degree of the corresponding department, specialist on-duty time relates to the timeliness of treatment, drug inventory matching degree ensures the availability of the required drugs, and diagnostic and treatment equipment scheduling position determines when the equipment can be put into use. These factors are given corresponding weights and summed up to obtain the candidate hospital matching score. This score is not an absolute value, but a relative measure of the adaptability of each hospital to meet the patient's current treatment needs, providing a quantifiable basis for comparing each candidate path.

[0081] After combining the candidate path with the corresponding candidate hospital matching score into a resource scheduling chain, a complete resource scheduling scheme is formed. For example, when a patient with suspected heart attack is on an ambulance, the system constructs a graph structure that takes the hospitals along the way, the ambulance location, and the patient's ECG changes as nodes, and reflects the accessibility and matching degree of each hospital through edge weights; the ConvLSTM layer analyzes the trend of ST segment elevation and predicts the need for a catheter room, a cardiologist, and anticoagulants, generating multiple candidate paths to different hospitals; then by calculating factors such as department idle rate and doctor on-duty time, the matching score is obtained, and finally a resource scheduling chain containing different hospital options and their adaptability is formed.

[0082] This generation method makes the resource scheduling chain both targeted and flexible, accurately matching the patient's condition needs and providing the best choice based on the real-time resource status of each hospital. Doctors can make the final decision based on the candidate path and matching score in the resource scheduling chain, combined with the actual situation, to ensure that the patient receives the most timely and effective treatment within the golden treatment time window, making the emergency resource scheduling more scientific and efficient, and effectively shortening the transition time from pre-hospital care to in-hospital treatment.

[0083] Further optionally, the intelligent decision layer, when matching the corresponding candidate hospital for the current patient according to the comparison result, is specifically configured to: determine whether the destination hospital or the nearest hospital to the location meets the medical treatment condition; if the candidate hospital matching degree score of the destination hospital or the nearest hospital to the location is higher than a set threshold, the destination hospital or the nearest hospital to the location meets the medical treatment condition; if the candidate hospital matching degree score of the destination hospital or the nearest hospital to the location is not higher than the set threshold, the destination hospital or the nearest hospital to the location does not meet the medical treatment condition; if it is determined that the medical treatment condition is met, whether the destination hospital or the nearest hospital to the location can receive the current patient is determined, and the determination result is uploaded to the cloud service platform; if it is determined that the medical treatment condition is not met, whether the destination hospital or the nearest hospital to the location can receive the current patient is determined, and the determination result is uploaded to the cloud service platform, so as to request the patient to start a further candidate hospital search task.

[0084] Specifically, when the intelligent decision layer matches the candidate hospital for the current patient according to the comparison result, the core is to determine the medical institution most suitable for receiving the patient through the preset judgment logic combined with the candidate hospital matching degree score, so as to ensure that the patient can be effectively treated within the golden treatment time window. This process is not a simple list screening, but a closed-loop decision based on objective scoring and dynamic feedback, which not only gives priority to the distance factor, but also takes into account the actual treatment capacity of the hospital.

[0085] Determining whether the destination hospital or the nearest hospital to the location meets the medical treatment condition is essentially to measure the adaptation degree of hospital resources and patient demand through comparison of the candidate hospital matching degree score and the set threshold. The setting of this threshold is not fixed, but is determined based on the effective experience of resource matching in a large number of past emergency cases, combined with the minimum requirement of different conditions for resources. It is like a baseline, which ensures that the hospital meeting the condition can at least meet the core resource demand of the current diagnosis and treatment path of the patient. When the matching degree score of the candidate hospital exceeds this threshold, it means that the key factors such as department idle rate, doctor on-duty time and drug inventory of the hospital can well adapt to the resource scheduling chain of the patient, and have the basic conditions to receive and treat the patient; otherwise, it means that the hospital has obvious gaps in some resource links, and may not be able to provide the required treatment for the patient in time.

[0086] If the destination hospital or the nearest hospital to the location meets the medical treatment condition, the intelligent decision layer will actively send a request to the hospital to receive the patient, inquire whether the patient can be admitted, and synchronize the judgment result to the cloud service platform. This process is not a one-way instruction, but an interactive process with a confirmation nature, aiming to avoid misjudgment caused by real-time resource changes, such as the sudden reception of other emergency patients in a short time, which changes the originally qualified resource condition. The cloud service platform records this result, not only to retain the dispatch trace, but also to provide information basis for subsequent possible cross-hospital cooperation.

[0087] When the destination hospital or the nearest hospital to the location does not meet the medical treatment condition, the system will also send a receiving inquiry to the hospital and upload the judgment result, but at the same time, it will request the cloud service platform to start a further candidate hospital search task for the patient. This dual-track processing method not only respects the importance of distance factor in emergency treatment (even if the hospital does not meet the standard, it does not rule out the possibility of meeting the demand through emergency resource allocation), but also expands the selection range by starting a new search task to avoid delay in treatment due to insufficient resources in a single hospital. The new search task is not a indiscriminate search, but a search based on the patient's resource scheduling chain in a larger geographical range to filter hospitals with higher matching scores, ensuring that the recommended candidate hospitals still meet the patient's diagnosis and treatment needs.

[0088] For example, a patient who needs emergency surgery has a matching score that does not meet the threshold due to the full occupancy of the operating room in the destination hospital. At this time, the system will first inquire whether the hospital can receive the patient by adjusting the surgery arrangement, and at the same time, start a search in the cloud service platform to find a hospital with an idle operating room and a specialist on site in the surrounding area. If the destination hospital replies that it can receive the patient by adjusting the resources, the system confirms that the hospital is a candidate; if it cannot receive the patient, it will again filter from the newly searched hospitals to ensure that the patient always has a suitable receiving institution to choose from.

[0089] This matching method not only reflects the priority consideration of the distance factor, but also ensures the reliability of resource matching through dynamic judgment and multiple searches, avoiding delay in treatment due to insufficient resources in a single hospital. At the same time, the information synchronization of the cloud service platform makes the whole process traceable, and provides data support for resource coordination between different hospitals, making the emergency dispatch efficient and flexible, and maximizing the guarantee of timely and effective medical treatment for patients.

[0090] It can be further understood that in some embodiments, the process of determining the patient resource scheduling chain by the intelligent decision layer is a coherent logical chain from data fusion to scene recognition, and to resource matching, the core of which is to deeply combine the real-time state of the patient, the distribution of medical resources, and the diagnosis and treatment specification to generate a complete diagnosis and treatment path that adapts to individual needs. This process begins with the multi-dimensional data transmitted by the data acquisition layer, including the patient's vital signs (such as heart rate, blood pressure, ST segment changes, etc.), the real-time state of medical equipment (such as the idle condition of the operating room, the running parameters of the imaging equipment), the patient's geographic location (such as the location of the ambulance, the number of the outpatient clinic), and the real-time dynamics of medical staff (such as the current working state of the specialist doctor). These data are not isolated, but are integrated into a dynamically updated information pool to provide a basis for subsequent scene judgment.

[0091] Subsequently, the first-level scheduling model based on the spatio-temporal graph neural network appears, which converts these data into nodes and associated relationships in the network. The patient's vital signs constitute nodes reflecting the patient's condition, medical equipment and medical staff constitute resource nodes, and geographic location constitutes spatially associated edge weights. By analyzing the state changes and association strengths of these nodes, the model can accurately identify the current medical scene of the patient, such as an emergency rescue scene, a routine outpatient scene, or a postoperative recovery scene. Different scenes correspond to different diagnosis and treatment logic, and this step defines the scope for subsequent resource matching.

[0092] After the medical scene is determined, the intelligent decision layer will call the second-level scheduling model (such as the emergency resource allocation model, the outpatient scheduling model, etc.) that matches the scene. These models rely on the specialized medical experience database and historical operation data, and have learned a large number of diagnosis and treatment path rules in similar scenarios. Taking the emergency scene as an example, when the model identifies that the patient needs interventional therapy, it will automatically associate the standard path framework containing "emergency examination → catheter room preparation → cardiologist on site → anticoagulant drug allocation" and other links, and then adjust the framework according to the patient's specific condition (such as the severity of heart attack) and the real-time state of the current resources (such as whether the nearest catheter room is idle and whether the corresponding doctor is on duty) to generate multiple possible candidate diagnosis and treatment paths.

[0093] Each candidate diagnosis and treatment path will list in detail the resource elements required by each medical service node: which type of imaging equipment is needed at the examination node, which technician will operate; which type of specialist doctor is needed at the treatment node, with what qualifications; which drugs are needed at the medication node, and how much dosage, etc. These elements are not randomly combined, but strictly follow the clinical diagnosis and treatment specification, and are screened in combination with the availability of real-time resources to ensure that each link in the path has actual operability.

[0094] Finally, the intelligent decision layer evaluates these candidate paths by calculating resource matching degree, time connection efficiency, conflict probability, and other indicators to filter out the optimal path combination as the resource scheduling chain. This chain not only includes the complete diagnosis and treatment steps from the starting point to the end point, but also marks the resource requirements and time estimates for each step, providing a clear execution blueprint for subsequent collaborative scheduling. For example, for a patient with a sudden heart attack, the final generated resource scheduling chain may include specific links such as "emergency ambulance continuous monitoring → immediate electrocardiogram examination after admission → entering the catheter room within 30 minutes → performed by the chief of cardiology department → preoperative use of specific anticoagulant drugs", etc. Each link clearly specifies the corresponding resources and time requirements, ensuring that the patient can receive continuous medical services within the golden treatment time window.

[0095] Throughout the process, the model continuously receives new real-time data and dynamically optimizes the path. For example, when a device suddenly fails, the service chain will automatically replace it with a backup device path. When a doctor is temporarily called away on an emergency, the new arrival time is recalculated and the subsequent link connections are adjusted. This dynamic adjustment capability ensures that the resource scheduling chain always synchronizes with patient needs and resource status, truly realizing the precise transformation from "standardized process" to "individualized path".

[0096] In addition to the emergency scenarios described above, real-time medical scenarios also include outpatient diagnosis and treatment scenarios, surgical diagnosis and treatment scenarios, imaging examination scenarios, and inpatient care scenarios, among others. Each scenario has its unique business process and resource demand characteristics.

[0097] The outpatient diagnosis and treatment scenario focuses on the complete process from registration to treatment, examination, and medication. Patient flow fluctuates regularly over time and involves the coordinated cooperation of multiple departments (such as internal medicine, pediatrics, and laboratory). The surgical diagnosis and treatment scenario revolves around the entire surgical process, including preoperative preparation (such as anesthesia assessment and instrument sterilization), intraoperative operation (such as coordination of different surgical teams), and postoperative recovery. The time connection and resource precision requirements are extremely high. The imaging examination scenario focuses on the use of various imaging equipment (such as CT and ultrasound), and needs to coordinate patient queuing order, equipment operation status, and technician work arrangement to avoid equipment idling or long patient waiting times. The inpatient care scenario focuses on daily nursing, examination arrangement, and rehabilitation treatment during the patient's hospitalization period, involving bed management, nurse scheduling, and drug delivery.

[0098] In these scenarios, the process of predicting resource scheduling chains and determining the next medical service node by the secondary scheduling model also follows the logic of "scenario recognition - path generation - resource matching", but adjusts the specific implementation method according to the scenario characteristics.

[0099] Taking the outpatient clinic scenario as an example, the matched outpatient scheduling model first integrates the patient's registration information, medical history, preliminary symptoms, and other data, combined with the real-time reception status of each department in the outpatient clinic (such as the progress of the doctor's reception, the idle situation of the examination room), to identify the current stage of the patient's diagnosis and treatment (such as just completing registration, waiting for test results). Subsequently, the model calls the outpatient specialist resource allocation rule library to generate a basic service chain framework including "waiting for diagnosis → doctor's inquiry → issuing examination order → laboratory examination → returning to the examination room for re-examination → pharmacy for medicine", and dynamically adjusts the order of each link according to the urgency of the patient's symptoms and the priority of the examination items (such as blood routine test requiring fast results). When determining the next medical service node, the model will query the idle rate of the laboratory equipment (such as whether the blood analyzer is queued) and the workload of the examination technician in real time. If the current examination department is in a state of resource shortage, it will automatically recommend other idle similar equipment or adjust the examination order to ensure that the patient can complete the diagnosis and treatment in the shortest path.

[0100] In the surgical diagnosis and treatment scenario, the operating room scheduling model will construct a service chain for the whole process of surgery based on the surgical notification sheet (including the type of surgery, the estimated duration, and the patient's basic information) and the real-time state of the operating room (such as whether the operating room is disinfected, and whether the anesthesia equipment is available). For example, for a laparoscopic surgery, the service chain may include "preoperative preparation room (anesthesia assessment) → operating room 3 (surgery execution) → recovery room (postoperative observation) → ward (postoperative care)". When predicting, the model will focus on calculating the time connection points of each link, such as the matching degree of the end time of the surgery and the idle time of the recovery room bed, and the arrival time of the nurses needed for postoperative care, to determine the next node. If the preoperative preparation has been completed, the next node will be automatically set as the operating room, and the patient's information will be pushed to the surgical team terminal simultaneously; if the surgery is approaching the end, the recovery room will be notified in advance to prepare for reception.

[0101] The imaging room scheduling model in the imaging examination scenario will generate a service chain including "registration → waiting area → equipment examination → report generation" based on the patient's examination application (such as chest CT), the real-time running state of the equipment (such as whether the equipment is under maintenance, and the current length of the examination queue), and the technician's scheduling situation. The model predicts the start time of the current patient's examination by analyzing the average duration of similar examinations in historical data, and determines the next node in combination with the patient's overall diagnosis and treatment path (such as returning to the doctor's office after examination). If the patient has completed the examination and the report has been generated, the next node will be set as the original examination room, and the examination results will be pushed to the doctor's terminal; if the examination needs to be scheduled, the earliest available time will be recommended, and the patient's overall diagnosis and treatment path will be updated simultaneously.

[0102] The inpatient care scenario's admission and discharge resource allocation model integrates the patient's inpatient information (such as bed number, attending physician), daily care plan (such as infusion time, rehabilitation training), examination arrangement (such as daily blood glucose monitoring), and generates a personalized inpatient service chain. When determining the next node, the real-time working position of the nurse (such as through indoor positioning), the medicine distribution progress (such as whether the intravenous infusion medicine has arrived at the ward), for example, when the patient completes the morning care, the next node may be "rehabilitation department physiotherapy", which will notify the rehabilitation technician of the patient's bed information and rehabilitation needs in advance to ensure accurate resource docking.

[0103] The secondary scheduling model in these scenarios dynamically optimizes the service chain by continuously absorbing real-time data (such as patient flow changes, equipment failures), determines the next medical service node in accordance with the diagnosis and treatment specifications, and adapts to the real-time fluctuations of resources, ultimately achieving efficient circulation of medical resources and smooth connection of patient diagnosis and treatment paths in various scenarios.

[0104] After introducing the method of the example embodiment of the present application, next, with reference to Figure 2 A hospital resource scheduling method based on a spatio-temporal graph neural network is described for the example embodiment of the present application, which includes: S201, deploying a multi-modal sensor network in medical devices and medical service terminals; S202, collecting multi-dimensional operation data of medical devices and medical service terminals through the multi-modal sensor network; wherein the multi-dimensional operation data at least includes: device running state, device service type, fault precursor signal; the multi-modal sensor network at least includes: sensors matched with different medical devices or medical service terminals, the sensor type is associated with the deployed location and the device type; S203, based on the multi-dimensional operation data, the geographical location where the medical devices and the medical service terminals are respectively located, determining the real-time medical scene where the medical devices and the medical service terminals are respectively located through a first scheduling model based on a spatio-temporal graph neural network, and predicting a resource scheduling chain matched with each real-time medical scene through a secondary scheduling model matched with different medical scenes, to determine the next medical service node matched with the current patient served by the medical devices and the medical service terminals; wherein the resource scheduling chain at least includes: at least one candidate diagnosis and treatment path, each candidate diagnosis and treatment path includes: a plurality of medical service nodes corresponding to the current patient, diagnosis and treatment equipment used by each medical service node, doctor type, medicine type; S204, constructing a virtual resource mapping of a hospital model according to the resource scheduling chain, simulating the execution effect of the scheduling scheme of different candidate diagnosis and treatment paths in the resource scheduling chain in real time through digital twin technology, and pushing to the medical service terminal or the cloud service platform, so that a user with corresponding authority selects the target diagnosis and treatment path and the corresponding scheduling scheme finally adopted.

[0105] The above method can realize the functions described in the above system embodiments, and the specific implementation of each function will not be repeated here.

[0106] After introducing the method and system of the exemplary embodiments of the present application, next, a terminal device of the exemplary embodiments of the present application is described, which is used to implement the hospital resource scheduling method based on the space-time graph neural network introduced in the above embodiments. The terminal device can realize each step described in the above method embodiments, and the specific implementation of each step will not be repeated here.

[0107] After introducing the method, system and terminal device of the exemplary embodiments of the present application, next, reference is made to Figure 3 The computer-readable storage medium of the exemplary embodiments of the present application is described, please refer to Figure 3 The computer-readable storage medium shown is an optical disc 30, which stores a computer program (i.e. program product) thereon, and the computer program, when run by a processor, will realize each step described in the above method embodiments. The specific implementation of each step will not be repeated here.

[0108] It should be noted that examples of the computer-readable storage medium can also include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other optical, magnetic storage medium, which will not be repeated here. The above described embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit it, the protection scope of the present application is not limited to this, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art within the technical scope disclosed by the present application, can still modify or easily think of changes to the technical solutions described in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A hospital resource scheduling system based on spatiotemporal graph neural network, characterized in that: The system comprises: A data collection layer is configured to deploy a multimodal sensor network in medical devices and medical service terminals; collect multidimensional operational data of medical devices and medical service terminals through the multimodal sensor network; the multidimensional operational data includes at least: device operating status, device service type, and fault warning signals; the multimodal sensor network includes at least: sensors matched to different medical devices or medical service terminals, where the sensor type is associated with the deployment location and device type; The intelligent decision-making layer is used to determine the real-time medical scenarios in which the medical equipment and medical service terminals are located based on multi-dimensional operational data, the geographical locations of the medical equipment and medical service terminals, and through a first-level scheduling model based on a spatiotemporal graph neural network. The intelligent decision-making layer is used to predict the resource scheduling chain that matches each real-time medical scenario through a second-level scheduling model that matches different medical scenarios, so as to determine the next medical service node that matches the current patient served by the medical equipment and medical service terminal. The resource scheduling chain includes at least one candidate diagnosis and treatment path, each candidate diagnosis and treatment path includes: multiple medical service nodes corresponding to the current patient, the diagnosis and treatment equipment, doctor type, and drug type used by each medical service node; The collaborative scheduling layer is used to build a virtual resource mapping of the hospital model based on the resource scheduling chain. It uses digital twin technology to simulate the execution effects of scheduling plans for different candidate diagnosis and treatment paths in the resource scheduling chain in real time, and pushes them to the medical service terminal or cloud service platform. Users with corresponding permissions can select the final target diagnosis and treatment path and the corresponding scheduling plan.

2. The hospital resource scheduling system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The intelligent decision-making layer is also used to: A multi-objective evolutionary algorithm is used, with minimizing patient waiting time and maximizing equipment utilization as the core of multi-objective optimization. The scheduling rules of the first-level scheduling model are autonomously iterated through historical hospital operation data. The scheduling rules of the second-level scheduling model are localized and autonomously iterated through local historical department operation data and specialized medical experience databases of each department. The second-level scheduling model includes at least: an emergency resource allocation model, an operating room scheduling model, an imaging room scheduling model, an admission and discharge resource allocation model, and an outpatient specialist resource allocation model. Federated learning is used to share model parameters for the first-level scheduling model and the second-level scheduling model localized in each department.

3. The hospital resource scheduling system based on spatiotemporal graph neural network according to claim 2 is characterized in that: The intelligent decision-making layer uses a multi-objective evolutionary algorithm, with minimizing patient waiting time and maximizing equipment utilization as the core of multi-objective optimization. It uses historical hospital operation data to autonomously iterate the scheduling rules of the first-level scheduling model. Specifically, it is used to: Based on the improved non-dominated sorting genetic algorithm (VA-NSGA-III), a vector angle screening strategy is introduced based on the reference point stratification of NSGA-III. By calculating the vector directions of the weight vector solution sets of equipment utilization and patient waiting time in the multi-objective optimization function space of the first-level scheduling model, similar angle solutions are eliminated to ensure a uniform distribution of the Pareto front solution set. The weight vector distribution in the multi-objective optimization function space is adjusted in real time based on the priority of emergency scenarios. By performing time series cluster analysis on the hospital operation conflict case library in historical data, we can identify high-frequency failure modes and generate a robustness rule library, which is then embedded in the algorithmic constraints of the first-level scheduling model. When the sensor detects a fault warning signal indicating equipment abnormality, the multi-objective optimization function space and algorithm constraints are combined to automatically trigger retraining of the first-level scheduling model, dynamically optimize the scheduling rules of the first-level scheduling model, and add the sub-goal of minimizing the fault recovery time to the backup equipment scheduling plan.

4. The hospital resource scheduling system based on spatiotemporal graph neural network according to claim 2 is characterized in that: The intelligent decision-making layer, through each department's local historical department operation data and specialized medical experience database, performs localized and autonomous iteration of the scheduling rules of the secondary scheduling model. Specifically, it is used to: Based on the diagnosis and treatment scenarios corresponding to each department, dynamically update the matching relationship between each department and different medical scenarios; Based on the local historical department operation data of each department, the department business characteristics of each department are extracted, and the model parameter update method that matches the changing trend of the department business characteristics is selected; The selected model parameter update method is adopted, combined with the local historical department operation data and specialized medical experience database of each department, to update the model parameters of the secondary scheduling model matched by each department, and use the differential privacy protection algorithm to share the encrypted model gradient with other departments or other medical institutions matching the same secondary scheduling model.

5. The hospital resource scheduling system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The multimodal sensor network also includes: wearable devices and corresponding wireless sensors equipped in emergency vehicles; The data collection layer, when collecting multi-dimensional operational data of medical equipment and medical service terminals through a multimodal sensor network, is specifically used to: Collecting patient vital sign data from wearable devices through wireless sensors, encrypting the patient vital sign data and real-time diagnosis and treatment information input by the onboard doctor in real time, and transmitting the data to the edge computing node of the hospital closest to the ambulance's location and / or the cloud service platform of the destination hospital of the ambulance; The intelligent decision-making layer determines the real-time medical scenarios in which the medical equipment and medical service terminals are located based on multi-dimensional operational data, the geographical locations of the medical equipment and medical service terminals, and a first-level scheduling model based on a spatiotemporal graph neural network. It also predicts the resource scheduling chain that matches each real-time medical scenario through a second-level scheduling model that matches different medical scenarios, so as to determine the next medical service node that matches the current patient served by the medical equipment and medical service terminal. Specifically, it is used to: Through the hospital edge computing nodes and / or cloud service platform, based on the patient's vital signs data and the real-time diagnosis and treatment information, as well as the real-time geographic location of the ambulance, a first-level scheduling model is used to predict that the current real-time medical scenario is an emergency scenario; Through the emergency resource allocation model matched with emergency scenarios, the resource demand changes within the golden treatment time window are predicted to obtain the resource scheduling chain matched with the patients in the ambulance. The emergency resource allocation model is constructed based on the spatiotemporal graph neural algorithm. The resource scheduling chain obtained based on the prediction is compared with the hospital's current real-time medical resource operation status, and the corresponding candidate hospital is matched to the current patient based on the comparison results.

6. The hospital resource scheduling system based on spatiotemporal graph neural network according to claim 5 is characterized in that: The intelligent decision-making layer, using an emergency resource allocation model matched to emergency scenarios, predicts resource demand changes within the golden treatment time window to obtain a resource scheduling chain that matches patients in ambulances. Specifically, it is used to: A real-time emergency medical resource graph is constructed using hospital locations, ambulance geographic coordinates, and patient vital signs as nodes, and road travel time, hospital admission capacity, historical response acceptance rate, and specialty matching as edge weights. The ConvLSTM layer processes the patient's vital signs stream and predicts changes in resource demand within the golden treatment time window to obtain candidate paths that match the changes in resource demand, including diagnostic equipment, doctor types, and drug types. Calculate the candidate hospital matching score corresponding to the candidate pathway; take the weighted sum of the department vacancy rate, specialist arrival time, drug inventory matching, and diagnostic equipment scheduling ranking in the candidate pathway as the candidate hospital matching score; The candidate paths and the corresponding candidate hospital matching scores are combined into a resource scheduling chain.

7. The hospital resource scheduling system based on spatiotemporal graph neural network according to claim 6 is characterized in that: The intelligent decision-making layer is specifically used to match the current patient with a corresponding candidate hospital based on the comparison results: Determine whether the destination hospital or the nearest hospital meets the medical treatment conditions; if the candidate hospital matching score of the destination hospital or the nearest hospital is higher than the set threshold, then the destination hospital or the nearest hospital meets the medical treatment conditions; if the candidate hospital matching score of the destination hospital or the nearest hospital is not higher than the set threshold, then the destination hospital or the nearest hospital does not meet the medical treatment conditions; If it is determined that the medical treatment conditions are met, the destination hospital or the nearest hospital will be checked to see if it can accept the current patient, and the judgment result will be uploaded to the cloud service platform; If it is determined that the conditions for medical treatment are not met, the destination hospital or the nearest hospital will be asked whether it can accept the current patient, and the judgment result will be uploaded to the cloud service platform to request the patient to initiate further candidate hospital search tasks.

8. The hospital resource scheduling system based on spatiotemporal graph neural network according to claim 1 is characterized in that: The collaborative scheduling layer, when constructing a virtual resource mapping of the hospital model based on the resource scheduling chain and simulating the execution effects of scheduling plans for different candidate diagnosis and treatment pathways in the resource scheduling chain in real time through digital twin technology, is specifically used to: Build a digital mirror of physical resources based on the hospital BIM model, with bed mapping as a dynamic occupancy status node and equipment mapping as a real-time operation model with failure probability; Input candidate diagnosis and treatment pathways, and use the digital twin engine to simulate and generate the conflict probability, resource utilization, and time cost of each candidate diagnosis and treatment pathway; Based on the conflict probability, resource utilization, and time cost of each candidate treatment pathway, the simulation data of different candidate treatment pathways are visualized, and the resource consumption difference and patient waiting time fluctuation range of each candidate treatment pathway are marked.

9. The hospital resource scheduling system based on spatiotemporal graph neural network according to claim 1 is characterized in that: Sensors connected to medical devices include at least: The sensors connected to the imaging device include at least: a piezoelectric sensor that monitors coil replacement pressure peaks in real time to predict mechanical failure; The sensors connected to the rack include at least: an optical fiber strain sensor for detecting the vibration amplitude of the equipment to prevent image blur; The sensors connected to the surgical instruments include at least: a six-dimensional force sensor and a robotic arm joint, which are used to feedback the surgical operation force to prevent tissue damage; Sensors connected to life support equipment include at least: ultrasonic flowmeters and ventilator tubing to monitor oxygen flow deviations to ensure ventilation safety; The sensors connected to the environmental perception sensors include at least: millimeter-wave radar, which is used to track the movement trajectory of patients and trigger diversion alarms in case of congestion; The sensors connected to the wearable device include at least: flexible epidermal electrodes for continuously collecting ECG signals to realize real-time alarm of ST segment elevation; The sensors connected to the disinfection equipment include at least: an ultraviolet intensity sensor, which is used to monitor the attenuation of the ultraviolet lamp to ensure that the disinfection effect meets the standards.

10. A hospital resource scheduling method based on spatiotemporal graph neural network, characterized in that: The system executes the hospital resource scheduling system based on spatiotemporal graph neural network according to any one of claims 1 to 9, wherein the method comprises: Deploy multimodal sensor networks in medical devices and medical service terminals; Collect multi-dimensional operational data of medical equipment and medical service terminals through a multi-modal sensor network; the multi-dimensional operational data includes at least: equipment operating status, equipment service type, and fault warning signals; the multi-modal sensor network includes at least: sensors matched to different medical equipment or medical service terminals, and the sensor type is associated with the deployment location and equipment type; Based on multi-dimensional operational data, the geographical locations of medical equipment and medical service terminals, a first-level scheduling model based on a spatiotemporal graph neural network is used to determine the real-time medical scenarios in which the medical equipment and medical service terminals are located. A second-level scheduling model that matches different medical scenarios is used to predict the resource scheduling chain that matches each real-time medical scenario, so as to determine the next medical service node that matches the current patient served by the medical equipment and medical service terminal. The resource scheduling chain includes at least one candidate diagnosis and treatment path, and each candidate diagnosis and treatment path includes multiple medical service nodes corresponding to the current patient, the diagnosis and treatment equipment, doctor type, and drug type used by each medical service node. A virtual resource mapping of the hospital model is constructed based on the resource scheduling chain. The execution effects of the scheduling plans of different candidate treatment paths in the resource scheduling chain are simulated in real time through digital twin technology, and pushed to the medical service terminal or cloud service platform. Users with corresponding permissions can select the final target treatment path and the corresponding scheduling plan.

Images