An intelligent management method, system, device and medium of hospital medical resources

Abstract

This application relates to an intelligent management method, system, device, and medium for hospital medical resources. The method includes: processing real-time heterogeneous event streams to generate a unified time-series event graph and extracting multi-dimensional hospital state feature vectors; updating the urgency values ​​in the attribute set of tasks to be scheduled based on dynamic priority adjustment instructions to obtain priority change summary features, and combining these with the multi-dimensional hospital state feature vectors to obtain an enhanced state vector; inputting the enhanced state vector into a resource scheduling strategy neural network model to obtain candidate resource scheduling schemes; performing deduction and verification of the candidate resource scheduling schemes based on preset medical safety red line rules to obtain effective candidate resource scheduling schemes; and using a context-aware dynamic reward function, selecting the effective candidate resource scheduling scheme with the highest expected utility evaluation value as the recommended scheduling trajectory. This method can optimize resource allocation efficiency and balance various scheduling objectives.

Description

Technical Field

[0001] This invention belongs to the field of resource management technology, and in particular relates to an intelligent management method, system, equipment and medium for hospital medical resources. Background Technology

[0002] In traditional hospital medical resource management, the scheduling of various medical resources is mainly addressed through manual experience rules and simple queuing models. For the allocation of expensive equipment such as operating rooms, MRI / CT, ​​special drugs, and medical staff such as surgeons and anesthesiologists, the allocation relies on the personal experience of a few managers to judge resource priorities and uses fixed, static rules to arrange medical tasks. Each department carries out scheduling work independently based on its own business needs, lacking cross-departmental collaboration. When encountering emergencies such as an influx of patients or equipment failures, the scheduling plan is adjusted manually on a temporary basis.

[0003] However, the above methods rely too heavily on personal experience and are difficult to cope with the complex spatiotemporal constraints and collaborative needs among large-scale, multi-resource operations. They are unable to effectively coordinate the simultaneous availability of multiple resources required for surgery and the connection between various stages of examination. Static priority rules lack flexibility and cannot dynamically adjust strategies according to the hospital's real-time operational status, making it difficult to adapt to dynamic priority adjustment needs such as changes in patient conditions and public health events. Each department operates independently, lacking a hospital-wide overall optimization perspective, which can easily lead to an imbalance in resource allocation. When faced with emergencies, manual rescheduling is slow and can easily cause chain delays. Furthermore, it cannot properly handle complex issues such as multi-objective optimization conflicts and safety risk control. Summary of the Invention

[0004] Therefore, it is necessary to provide an intelligent management method, system, device, and medium for hospital medical resources that can organically combine inverse reinforcement learning, hierarchical reward structure, context-aware mechanism, and safe exploration strategy to address the above-mentioned technical problems.

[0005] Firstly, this application provides an intelligent management method for hospital medical resources, including: The real-time heterogeneous event stream of the hospital is standardized and cross-system correlated to generate a unified time-series event graph. The multi-dimensional hospital state feature vector at the current moment is extracted from the unified time-series event graph. The multi-dimensional hospital state feature vector includes the overall hospital load index, the real-time status of each resource, the attribute set of all tasks to be scheduled, and the resource conflict matrix. The real-time heterogeneous event stream is collected in real time through various information systems and physical sensing nodes of the hospital. In response to receiving a dynamic priority adjustment instruction for a specific scheduling task, the urgency value in the attribute set of the task to be scheduled is updated based on the dynamic priority adjustment instruction to obtain a priority change summary feature. The priority change summary feature is then concatenated with a multi-dimensional hospital state feature vector to obtain an enhanced state vector. The enhanced state vector is input into a pre-trained resource scheduling policy neural network model to obtain the scheduling action probability distribution of all tasks to be scheduled, and at least one candidate resource scheduling scheme is generated based on the scheduling action probability distribution. Based on the preset medical safety red line rules, the candidate resource scheduling schemes are deduced and verified, and all candidate resource scheduling schemes that pass the verification are marked as valid candidate resource scheduling schemes. Based on the context-aware dynamic reward function, the expected utility evaluation value of each effective candidate resource scheduling scheme is calculated, and the effective candidate resource scheduling scheme with the highest expected utility evaluation value is taken as the recommended scheduling trajectory.

[0006] In one embodiment, the resource scheduling strategy neural network model is obtained through the following method: Based on a multidimensional performance comprehensive scoring function, all historical scheduling trajectories in the historical scheduling trajectory dataset are scored, and historical scheduling trajectories with scores in the top preset percentage are selected as expert demonstration trajectories, while historical scheduling trajectories with scores in the bottom preset percentage are marked as poor demonstration trajectories. The multidimensional performance comprehensive scoring function includes patient satisfaction, resource utilization, critical patient waiting time, cost control, and number of manual interventions. Based on a parameterized fundamental reward function neural network, a stochastic policy is trained in a high-fidelity simulation environment using a reinforcement learning algorithm, and multiple trajectories are obtained by sampling based on the stochastic policy; the expression for the fundamental reward function neural network is: ,in, This is a real-time state vector of the resource. In order to schedule trajectory actions, The weight parameters to be learned. For state-action pair features, The scalar reward value; Calculate the expected value of state-action pairs in the trajectory and the empirical expected value of state-action pairs in the expert demonstration trajectory. The gradient of the maximum entropy inverse reinforcement learning objective function based on contrastive loss is calculated using the expected value and the empirical expected value. Based on this gradient, the weight parameters of the parameterized base reward neural network are updated using gradient ascent until the difference between the expected value and the empirical expected value is less than a preset threshold, thus obtaining the corresponding expert-scheduled base reward function. The expression for the maximum entropy inverse reinforcement learning objective function based on contrastive loss is: ,in, For gradient, To demonstrate the trajectory to experts, This serves as a negative example. Marginal value, These are the weighting coefficients; Based on the basic reward function, a hierarchical reward function is constructed; the hierarchical reward function includes a basic reward function layer, an operational efficiency reward function layer, a cross-resource process coordination reward function layer, and a medical safety red line reward function layer. Based on real-time hospital context indicators extracted from multidimensional state feature vectors, a context-aware dynamic reward function is obtained by dynamically calculating the weight coefficients of each reward function and internal indicator in the hierarchical reward function through a weighted neural network. The expression of the context-aware dynamic reward function is as follows: ,in, and To adjust the hyperparameters that balance efficiency and the importance of coordinating objectives, Basic reward function layer, For the operational efficiency reward function layer, For cross-resource process coordination reward function layer, A reward function layer for medical safety red lines; Based on a historical scheduling trajectory dataset containing expert demonstration trajectories, an offline reinforcement learning algorithm is used to pre-train a neural network model for resource scheduling strategies to obtain a preliminary security strategy. Based on a high-fidelity simulation environment simulating the dynamic operation of a hospital, a neural network model for resource scheduling strategy is obtained by using a context-aware dynamic reward function as the optimization objective and inequality constraints derived from medical safety red line rules as the constraint conditions, and fine-tuning the initial safety strategy with a constrained strategy optimization algorithm.

[0007] In one embodiment, a hierarchical reward function is constructed based on the basic reward function, including: An operational efficiency reward function layer is constructed based on the dynamic weighted sum of multiple operational indicators; the expression of the operational efficiency reward function layer is as follows: ,in, This represents the average increase in resource utilization after this scheduling. and Tasks to be scheduled The urgency level and estimated time, Penalties for resource idleness caused by scheduling , and These are dynamic weighting coefficients; Based on a lightweight discrete event simulation model, the hospital process in the target cycle after scheduling is deduced, and a cross-resource process coordination reward function layer is constructed based on the rewards corresponding to the deduction results: the rewards corresponding to the deduction results include a positive reward if the scheduling makes the operation end time match the bed idle time; if the scheduling resolves or mitigates the future resource conflicts identified based on the resource conflict matrix, a positive reward is given according to the urgency value of the conflict task and the resource scarcity value. A medical safety red line reward function layer is constructed based on a hard penalty mechanism and a soft constraint proximity condition: the hard penalty mechanism corresponds to giving a negative reward and triggering an alarm if scheduling causes any task to violate the preset inviolable red line constraint; the soft constraint proximity condition corresponds to determining a negative reward based on the difference between the target indicator and the danger threshold if scheduling causes the target indicator to deteriorate to below the danger threshold. A hierarchical reward function is constructed based on a basic reward function, an operational efficiency reward function layer, a cross-resource process coordination reward function layer, and a medical safety red line reward function layer.

[0008] In one embodiment, based on real-time hospital context indicators extracted from a multi-dimensional state feature vector, a context-aware dynamic reward function is obtained by dynamically calculating the weight coefficients of each reward function and internal indicator in the hierarchical reward function through a weighted adjustment neural network, including: Based on the overall hospital load index of multidimensional hospital state feature vector, the backlog rate of emergency patients, ICU occupancy rate, proportion of major surgeries, time period pattern and emergency event markers are extracted to obtain the real-time context state vector. The real-time context state vector is input into a pre-trained weight adjustment neural network to obtain the set of weight coefficients for the context-aware dynamic reward function at the current time step; the expression for the set of weight coefficients is: ,in, Adjusting the weights of the neural network For real-time context state vector, Adjust the parameters of the neural network for the weights; The weighted neural network is trained through the following steps: Based on the multidimensional performance comprehensive scoring function, historical scheduling trajectories and multidimensional performance comprehensive scores of historical scheduling trajectories under different real-time context state vectors are extracted from the historical scheduling trajectory dataset. With the goal of maximizing the correlation between the multidimensional performance comprehensive score and the predicted cumulative reward induced by the context-aware dynamic reward function adjusted by weights, the parameters of the weight-adjusted neural network are trained to obtain the weight-adjusted neural network.

[0009] In one embodiment, based on a high-fidelity simulation environment simulating the dynamic operation of a hospital, using a context-aware dynamic reward function as the optimization objective and inequality constraints derived from medical safety red line rules as constraints, a constrained policy optimization algorithm is employed to fine-tune the initial safety policy, resulting in a resource scheduling policy neural network model, including: The hard penalty mechanism and soft constraint proximity condition in the medical safety red line reward function layer are transformed into a set of inequality constraints; the expression of the inequality constraints is as follows: , ,in To constrain the allowable threshold, This is a real-time state vector of the resource. To schedule trajectory actions; Based on the Lagrange relaxation method, the constrained policy optimization problem is transformed into unconstrained optimization, and a Lagrange function is constructed; the expression of the Lagrange function is: ,in, For strategy The expectation of context-aware dynamic rewards. For Lagrange multipliers; The update gradient is calculated based on the Lagrange function, and the parameters and Lagrange multipliers of the resource scheduling strategy neural network model are updated along the direction of the update gradient. The update formula for the parameters of the resource scheduling strategy neural network model is as follows: The updated form of the Lagrange multipliers is: .

[0010] In one embodiment, the urgency value in the attribute set of the task to be scheduled is updated based on the dynamic priority adjustment instruction to obtain a priority change summary feature. This priority change summary feature is then concatenated with a multi-dimensional hospital state feature vector to obtain an enhanced state vector, including: Based on the urgency value of the dynamic priority adjustment instruction, the urgency value of the target task to be scheduled is updated to obtain the priority change event; the priority change event includes the initiating role; Update the priority change event to the unified time series event graph, and update the multidimensional hospital status feature vector based on the updated unified time series event graph; The total number of priority change events and the average priority change magnitude of all priority change events that occurred within the preset time window are counted, and it is checked whether there are priority change events initiated by preset initiating roles to obtain priority change summary features. The priority change summary features are concatenated with the updated multidimensional hospital state feature vector to obtain the enhanced state vector.

[0011] In one embodiment, the real-time heterogeneous event stream of the hospital is standardized and cross-system correlated to generate a unified time-series event graph. A multi-dimensional hospital state feature vector at the current moment is then extracted from the unified time-series event graph, including: The real-time heterogeneous event stream is standardized according to the preset five-tuple format to obtain the standardized event stream; the five-tuple format includes subject, predicate, object, timestamp and attribute; Based on event identifiers, events linked together in the same medical process within a standardized event flow are defined as workflow instances. The real-time status of all resource entities is maintained based on timestamps to obtain the real-time status of resources. The real-time status of resources includes idle, reserved, occupied, and unavailable. Take snapshots of the dynamic network consisting of all workflow instances and real-time resource statuses at preset time intervals to generate a unified time-series event graph; From the unified time-series event graph, the overall hospital load index is statistically calculated and the micro-attribute vectors of each task to be scheduled are extracted. The overall hospital load index includes the number of patients waiting for emergency treatment, ICU bed occupancy rate, and average queue length for large equipment. The micro-attribute vectors include task type, average estimated time, current urgency level, list of required resources, ready time, and latest allowed time. By concatenating and encoding the overall hospital load index, real-time resource status, and micro-attribute vectors of each task to be scheduled, a multi-dimensional hospital status feature vector is obtained.

[0012] Secondly, this application also provides an intelligent management system for hospital medical resources, including: The event resource module is used to standardize and perform cross-system correlation processing on the real-time heterogeneous event stream of the hospital, generate a unified time-series event graph, and extract the multi-dimensional hospital status feature vector at the current moment from the unified time-series event graph. The multi-dimensional hospital status feature vector includes the overall hospital load index, the real-time status of each resource, the attribute set of all tasks to be scheduled, and the resource conflict matrix. The real-time heterogeneous event stream is collected in real time through various information systems and physical sensing nodes of the hospital. The emergency demand module is used to respond to the dynamic priority adjustment instruction for a specific scheduling task, update the urgency value in the attribute set of the task to be scheduled based on the dynamic priority adjustment instruction, obtain the priority change summary feature, and concatenate the priority change summary feature with the multi-dimensional hospital status feature vector to obtain the enhanced status vector. The candidate scheduling module is used to input the enhanced state vector into the pre-trained resource scheduling policy neural network model to obtain the scheduling action probability distribution of all tasks to be scheduled, and generate at least one candidate resource scheduling scheme based on the scheduling action probability distribution. The red line verification module is used to perform simulation verification of candidate resource scheduling schemes based on preset medical safety red line rules, and to mark all candidate resource scheduling schemes that pass the verification as valid candidate resource scheduling schemes. The recommended scheduling module is used to calculate the expected utility evaluation value of each effective candidate resource scheduling scheme based on the context-aware dynamic reward function, and to take the effective candidate resource scheduling scheme with the highest expected utility evaluation value as the recommended scheduling trajectory.

[0013] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above-mentioned intelligent management methods for hospital medical resources.

[0014] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of any of the above-mentioned intelligent management methods for hospital medical resources.

[0015] The aforementioned intelligent management methods, systems, equipment, and media for hospital medical resources standardize and cross-system correlate heterogeneous event streams collected in real time from various hospital systems to generate a unified time-series event graph, from which multi-dimensional state feature vectors are extracted. Responding to external dynamic priority adjustment commands, the system instantly updates the urgency of relevant tasks and generates change summaries, concatenating these summaries with the original state features to form an enhanced state vector, thus encoding dynamic intervention signals into model-understandable inputs. This enhanced state vector is input into a pre-trained resource scheduling strategy neural network model, which outputs the action probability distribution of all tasks to be scheduled and generates candidate scheduling schemes accordingly, achieving intelligent mapping from complex states to specific scheduling schemes. Pre-set medical safety red line rules are used to deduce and verify all candidate schemes, filtering out non-compliant schemes to ensure the safety baseline of the output schemes. A reward function with weights dynamically adjusted according to the hospital's real-time context calculates the expected utility evaluation value of each verified scheme, selecting the scheme with the highest evaluation value as the final recommendation. This technical effect improves the overall intelligence level and decision-making quality of hospital resource scheduling by responding in real time to dynamic priority changes, adaptively balancing multiple objectives, and strictly ensuring medical safety red lines. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the intelligent management method for hospital medical resources according to the present invention. Figure 2 This is a structural diagram of the intelligent management system for hospital medical resources according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] In one embodiment, such as Figure 1As shown, an intelligent management method for hospital medical resources is provided. This embodiment illustrates the method applied to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps: S101. Standardize and perform cross-system correlation processing on the real-time heterogeneous event stream of the hospital to generate a unified time-series event graph, and extract the multi-dimensional hospital state feature vector at the current moment from the unified time-series event graph; the multi-dimensional hospital state feature vector includes the overall hospital load index, the real-time status of each resource, the attribute set of all tasks to be scheduled, and the resource conflict matrix; the real-time heterogeneous event stream is collected in real time through the hospital's various information systems and physical sensing nodes.

[0020] Indicatively, the acquisition of real-time heterogeneous event streams is achieved by deploying a distributed data acquisition agent network across key information systems and physical sensing nodes within the hospital. These information systems include the Hospital Information System (HIS), Laboratory Information System (LIS), Picture Archiving and Communication System (PACS), Surgical Anesthesia System, Drug Management System, Human Resources System, and Equipment Management System. The physical sensing nodes include an IoT sensor network, video recognition equipment, and a logistics channel status monitoring module.

[0021] Furthermore, when standardizing the collected real-time heterogeneous event streams, each original event is converted into a unified "resource event triplet" format, which includes five core elements: subject, predicate, object, timestamp, and context attributes. The subject is the core object associated with the event, the predicate is the action behavior corresponding to the event, the object is the target object of the event, the timestamp records the time of the event with millisecond precision, and the context attributes are supplementary descriptive information of the event.

[0022] Specifically, cross-system association processing is implemented based on an event association algorithm that integrates rules and machine learning. A globally unified resource entity registry is established, assigning a unique identifier and status attributes to each physical and logical resource. The status attributes include five types: idle, reserved, occupied, occupied but about to be released, and unavailable. During the association process, patient ID, application number, and resource unique identifier are used as the core association keys. Combining time proximity and business process logic, related events scattered across different systems are linked to form complete workflow instances. For example, events such as "surgery application," "preoperative examination completed," "anesthesia assessment passed," "operating room allocation," and "surgery started" are associated as the same surgical workflow. At the same time, a globally unified time-series database is constructed to store all standardized events in timestamp order, forming a unified time-series event graph to dynamically reflect the relationship and state evolution trajectory between hospital resources and medical tasks.

[0023] The extraction of multidimensional hospital state feature vectors is achieved through statistical analysis and logical computation based on a unified time-series event graph. The overall hospital load index is obtained through the aggregation and statistics of global events in the event graph, reflecting the hospital's current overall operational pressure, including the number of emergency patients, ICU bed occupancy rate, operating room occupancy rate, average queue length for large medical equipment, critical drug inventory warning level, and the on-duty rate of medical staff on the current shift. The real-time status of each resource is extracted through a dynamic resource status tracker, updating the current status of each resource entity in real time based on status change events in the event graph. An attribute set for each scheduled task is constructed, with each task's attributes including task type, estimated time, urgency level, patient's condition level, required resource set list, ready time, latest allowed time, current waiting time, and related task information. The resource conflict matrix is ​​constructed by analyzing the resource requirements of all scheduled and pending tasks within a preset future time window. The rows of the matrix correspond to various hospital resources, and the columns correspond to each pending task. Matrix elements take values ​​of 1 or 0, where 1 indicates that the resource required by the task is already occupied by other scheduled tasks or has a high probability of being occupied within the expected time period, and 0 indicates that the resource is available within the expected time period. The calculation of the resource conflict matrix combines the resource occupancy periods of scheduled tasks in the event graph with the expected execution periods of pending tasks, and is achieved through time interval overlap determination.

[0024] S102. In response to receiving a dynamic priority adjustment instruction for a specific scheduling task, update the urgency value in the attribute set of the task to be scheduled based on the dynamic priority adjustment instruction to obtain a priority change summary feature, and concatenate the priority change summary feature with the multi-dimensional hospital state feature vector to obtain an enhanced state vector.

[0025] Furthermore, dynamic priority adjustment instructions can be received through a preset human-computer interaction interface, supporting authorized users such as hospital dispatchers and clinicians to adjust the priority of specific tasks to be scheduled in real time. The adjustment operation includes discrete level adjustment and continuous numerical adjustment. The adjustment instruction contains core information such as target task identifier, urgency value before adjustment, urgency value after adjustment, operator identifier, and adjustment timestamp.

[0026] Specifically, based on the adjusted urgency value in the adjustment instruction, the original urgency value in the task attribute set to be scheduled is replaced to ensure that the task's attribute information is consistent with the latest priority requirements. Simultaneously, the complete process information of this priority adjustment is recorded, including numerical changes before and after the adjustment, and the reason for the adjustment. The priority change summary feature is generated by statistical analysis of all priority adjustment events within a recent preset time window. This feature is used to quantitatively reflect the overall trend of recent priority adjustments. Its components include the total number of tasks with priority changes within the time window, the average change in urgency of each changed task, the percentage of changes initiated by high-authority operators, the ratio of tasks with increased priority to tasks with decreased priority, and the maximum single change magnitude.

[0027] S103. Input the enhanced state vector into the pre-trained resource scheduling strategy neural network model to obtain the scheduling action probability distribution of all tasks to be scheduled, and generate at least one candidate resource scheduling scheme based on the scheduling action probability distribution.

[0028] Optionally, the resource scheduling strategy neural network model adopts a structure combining a Transformer encoder and a multilayer perceptron (MLP). The Transformer encoder is used to process the sequential features in the enhanced state vector, such as the attribute sequence of the task to be scheduled and the temporal information of the resource state. It captures the complex relationships between tasks and resources, and between tasks themselves, through a self-attention mechanism. The MLP is used to perform nonlinear transformation and dimensionality compression on the features output by the Transformer encoder, and finally map them to the scheduling action space.

[0029] For example, the offline pre-training stage of the resource scheduling strategy neural network model uses an offline dataset constructed from historical hospital scheduling trajectory data as training samples. This dataset contains a large number of state vectors at historical moments, corresponding scheduling actions, and feedback results after execution. It is trained using an offline reinforcement learning algorithm. The training objective is to enable the model to learn effective strategies from historical scheduling experience, while avoiding the model from learning high-risk actions through conservative constraints.

[0030] Specifically, after the enhanced state vector is input into the model, it is processed by a Transformer encoder to obtain a feature representation that integrates the associated information. This feature representation is then mapped through an MLP layer to a scheduling action score for each task to be scheduled. The scheduling action score reflects the suitability of the task for a specific scheduling action. Subsequently, a Softmax function is used to convert the scheduling action scores of all tasks to be scheduled into a scheduling action probability distribution. This distribution satisfies the constraint that the sum of all action probabilities is 1. ,in, This indicates that in the augmented state vector Next, the A scheduling action was taken for each pending task. The probability of; The first output of the model represents the... The scheduling action score for each pending task; This indicates the total number of tasks currently awaiting scheduling; This represents an exponential function; the denominator is the sum of the exponents of the scheduling action scores of all tasks to be scheduled, used to normalize the scores and ensure that the output results meet the basic requirements of probability distribution.

[0031] When generating candidate resource scheduling schemes based on the probability distribution of scheduling actions, a Top-K sampling strategy is adopted. Each candidate resource scheduling scheme explicitly includes core information such as the resource allocation result, execution time arrangement, and task execution order for each task to be scheduled, while ensuring that there are no logical conflicts between the resource allocation and time arrangement of all tasks in the scheme.

[0032] S104. Based on the preset medical safety red line rules, the candidate resource scheduling schemes are deduced and verified, and all candidate resource scheduling schemes that pass the verification are marked as valid candidate resource scheduling schemes.

[0033] Indicatively, medical safety red line rules include two categories: hard constraints and soft constraints / early warning rules. Hard constraints are mandatory requirements that cannot be violated, while soft constraints / early warning rules are restrictive requirements that must be controlled within safety thresholds. Hard constraints may include that waiting times for critically ill patients must not exceed preset medical safety limits, and that the allocation of medical resources must meet qualification matching requirements. Soft constraints / early warning rules may include that the average response time for emergency patients must not exceed preset thresholds, and that the difference in average waiting times for patients of different disease types must be controlled within preset fairness indicators.

[0034] Optionally, the simulation and verification of candidate resource scheduling schemes can be implemented using a lightweight discrete event simulation model. This involves inputting the candidate resource scheduling scheme into the simulation model, setting the simulation start time to the current moment, and the simulation duration to the execution cycle of all medical tasks covered by the scheme. During the simulation, the model simulates the execution process of each medical task according to the resource allocation rules and time period arrangements in the scheme, including resource consumption, process advancement, and state transitions. Simultaneously, it monitors in real time whether the scheme violates any medical safety red line rules during execution. If a candidate scheme violates any hard constraint rule during simulation execution, it is directly deemed to have failed verification and is eliminated. If the scheme does not violate any hard constraint rule but violates a soft constraint warning rule, the degree of soft constraint violation is calculated. If the degree of violation exceeds a preset tolerance threshold, the verification is deemed to have failed. If the degree of violation is within the tolerance threshold, the scheme is temporarily retained and will be considered in the subsequent utility evaluation stage.

[0035] S105. Based on the context-aware dynamic reward function, calculate the expected utility evaluation value of each effective candidate resource scheduling scheme, and take the effective candidate resource scheduling scheme with the highest expected utility evaluation value as the recommended scheduling trajectory.

[0036] For example, the context-aware dynamic reward function is: ,in, The total reward value corresponding to an effective candidate resource scheduling scheme is the expected utility evaluation value. To enhance the state vector; The set of scheduling actions corresponding to valid candidate resource scheduling schemes; The basic reward items are learned from expert demonstration trajectories through inverse reinforcement learning, reflecting the core preferences of expert scheduling decisions; This is an operational efficiency reward item used to evaluate the effectiveness of the solution in optimizing efficiency indicators such as resource utilization and task completion rate. This is a collaborative and coordinating reward item, used to evaluate the effectiveness of the plan in improving the smoothness of cross-departmental and cross-resource collaboration; This is a safety red line reward item, used to evaluate the degree to which the plan meets medical safety constraints; The dynamic weighting coefficient for operational efficiency rewards; This refers to the dynamic weighting coefficients for collaborative and coordinated reward items.

[0037] The expected utility evaluation value of each effective candidate resource scheduling scheme is calculated by determining its corresponding total reward value. The expected utility evaluation values ​​of all valid candidate resource scheduling schemes are obtained and ranked. The scheme with the highest evaluation value is selected as the recommended scheduling trajectory. The recommended scheduling trajectory contains complete information such as resource allocation scheme, task execution sequence arrangement, and explanation of key constraint satisfaction, which can be directly used to guide the hospital's actual resource scheduling execution.

[0038] The aforementioned intelligent management method for hospital medical resources involves real-time collection of heterogeneous event streams from various hospital information systems and physical sensing nodes. These streams are then standardized and cross-system correlated to construct a unified time-series event graph. From this graph, a multi-dimensional hospital state feature vector is extracted, containing overall hospital load indicators, real-time resource status, a set of task attributes to be scheduled, and a resource conflict matrix. This vector responds to dynamic priority adjustment commands, updates task urgency values, and generates priority change summary features. These features are then concatenated with the multi-dimensional hospital state feature vector to form an enhanced state vector. This enhanced state vector is input into a pre-trained resource scheduling strategy neural network model to generate candidate resource scheduling schemes. These schemes are then evaluated and selected based on pre-defined medical safety red line rules to identify effective schemes. Finally, a context-aware dynamic reward function is used to calculate the expected utility evaluation value of each effective candidate scheme, and the optimal scheme is selected as the recommended scheduling trajectory. This method achieves intelligent scheduling of hospital medical resources, dynamically responding to changes in task priority. While ensuring medical safety, it optimizes resource allocation efficiency, balances various scheduling objectives, improves the rationality and adaptability of scheduling decisions, and enhances robustness in complex hospital operation scenarios.

[0039] In one embodiment, the resource scheduling strategy neural network model is obtained through the following method: S11. Based on the multidimensional performance comprehensive scoring function, all historical scheduling trajectories in the historical scheduling trajectory dataset are scored, and historical scheduling trajectories with scores in the top preset percentage are selected as expert demonstration trajectories, while historical scheduling trajectories with scores in the bottom preset percentage are marked as poor demonstration trajectories. The multidimensional performance comprehensive scoring function includes patient satisfaction, resource utilization, critical patient waiting time, cost control, and number of manual interventions.

[0040] As an illustration, the historical scheduling trajectory dataset consists of complete scheduling records accumulated during the hospital's past operations. Each historical scheduling trajectory contains the hospital's status sequence within a specific time period, the corresponding scheduling action sequence, and the actual feedback data after execution, covering core information such as resource allocation results, task completion status, patient waiting time, and resource utilization rate.

[0041] Optionally, the multidimensional performance comprehensive scoring function is: ,in, Historical scheduling trajectory Overall score; For patient satisfaction; The resource utilization rate is calculated by weighting the ratio of the actual occupied time of various core resources in the scheduling trajectory to the total available time. The waiting time for critically ill patients is the average waiting time for all critically ill patients in the trajectory from when the task is ready to when it begins execution. This is a cost control item, encompassing the sum of scheduling-related costs such as resource idleness costs, drug loss costs, and additional equipment maintenance costs. This refers to the number of times human intervention is required, i.e., the number of times human schedulers make manual adjustments during the execution of the scheduling trajectory to correct scheduling deviations and handle emergencies. , , , , These are the weighting coefficients for each evaluation item.

[0042] S12. A parameterized basic reward function neural network is used to train a random policy in a high-fidelity simulation environment through reinforcement learning algorithms, and multiple trajectories are obtained by sampling based on the random policy; the expression of the basic reward function neural network is: ,in, This is a real-time state vector of the resource. In order to schedule trajectory actions, The weight parameters to be learned. For state-action pair features, This is a scalar reward value.

[0043] Specifically, This is a real-time status vector for resources, containing the current status and associated attribute information of various hospital resources; The scheduling trajectory action refers to the resource allocation decisions made under specific conditions; The weight parameters to be learned include the connection weights and bias terms between neurons in each layer of the neural network, and their initial values ​​are generated through random initialization. The state-action pair features are derived from the original state vector through a neural network. and dispatching actions The high-level abstract feature vectors extracted can capture the complex relationship between state and action; The output scalar reward value is used to quantify the quality of the state-action pair.

[0044] The high-fidelity simulation environment provides a realistic simulation scenario for reinforcement learning training, closely resembling actual hospital operations. Built upon agent modeling and discrete event simulation techniques, this environment accurately replicates the behavioral characteristics and treatment process logic of hospital resource entities, while incorporating uncertainties from real operations, such as fluctuations in patient arrival rates, randomness in surgical time, equipment malfunctions, and temporary leave requests from medical staff. Within this simulation environment, a randomized policy is trained using a reinforcement learning algorithm; the selected algorithm is the SoftActor-Critic (SAC) algorithm. During training, the randomized policy acts as an agent interacting with the high-fidelity simulation environment, and the agent adjusts its current state based on feedback from the environment. Output scheduling action After the environment performs the action, it updates its state and returns an immediate reward calculated based on a neural network using a fundamental reward function. Through continuous interactive iteration, the stochastic policy gradually learns effective action patterns under different states. When the policy training reaches a preset number of iterations or the reward value converges and stabilizes, a large number of trajectories are sampled in a simulation environment based on the stochastic policy. The sampling process covers a variety of hospital operation scenarios, and finally, multiple complete trajectories containing state sequences, action sequences, and reward sequences are obtained.

[0045] S13. Calculate the expected value of the state-action pair features in the trajectory, and calculate the empirical expected value of the state-action pair features in the expert demonstration trajectory.

[0046] Furthermore, each sampled trajectory is analyzed to extract all state-action pairs contained therein. , The trajectory is indexed, and the feature vector corresponding to each state-action pair is calculated using a pre-defined feature extraction function. Calculate the weighted sum of the feature vectors for each state-action pair across all sampled trajectories, with the weights being the cumulative reward value of the corresponding trajectory; divide the weighted sum by the sum of the cumulative reward values ​​of all trajectories to obtain the expected value of the state-action pair feature vectors, i.e., ,in, For random strategies The expected value of the state-action pair features of the sampled trajectory; This represents the total number of sampled trajectories. For the first The length of the trajectory and the number of state-action pairs it contains; For the first Trajectory number Feature vectors of state-action pairs; For the first The cumulative reward value of each trajectory.

[0047] Similarly, the empirical expected value of the state-action pair features in the expert demonstration trajectory is... ,in, Provide expert demonstration trajectories of state-action pairs with expected values ​​of features. The total number of expert demonstration trajectories; For the first The length of the expert demonstration trajectory; For the first The first expert demonstration trajectory The feature vector of a state-action pair.

[0048] S14. Calculate the gradient of the maximum entropy inverse reinforcement learning objective function based on the expected value and the empirical expected value. Based on the gradient, update the learning weight parameters of the parameterized basic reward neural network using the gradient ascent method until the difference between the expected value and the empirical expected value is less than a preset threshold, thus obtaining the corresponding expert-scheduled basic reward function. The expression for the maximum entropy inverse reinforcement learning objective function based on contrastive loss is: ,in, For gradient, To demonstrate the trajectory to experts, This serves as a negative example. Marginal value, These are the weighting coefficients.

[0049] Specifically, The objective function is the neural network weight parameters relative to the basic reward function. The gradient is used to guide the direction of parameter updates; Provide the expected value of the reward function gradient for all state-action pairs in the expert demonstration trajectory; For random strategies The expected gradient of the reward function for all state-action pairs in the sampled trajectory; Weighting coefficients for comparison losses; The marginal value is used to set the minimum reward gap between the expert demonstration trajectory and the poor demonstration trajectory, ensuring that the learned reward function can clearly distinguish between good and bad trajectories; For a single expert demonstration trajectory, This is a single bad example trajectory; The cumulative reward value for the expert demonstration trajectory is obtained by summing the reward values ​​of all state-action pairs in the trajectory; This is the cumulative reward value for poor demonstration trajectories, calculated in the same way as expert trajectories; To compare the loss term, when the reward difference between the expert trajectory and the bad trajectory is less than the marginal value... When a positive loss is generated, the reward function is driven to increase the gap between the two; when the gap is greater than or equal to... The time loss is 0, avoiding over-optimization.

[0050] After the gradient calculation is completed, the weight parameters of the basic reward function neural network are updated using the gradient ascent method. The step size for parameter updates is controlled by the learning rate, which employs an adaptive adjustment strategy. A larger learning rate is set in the early stages of training to accelerate convergence, and then gradually decreased as training progresses to ensure parameter stability. After each parameter update, the random policy needs to be retrained based on the updated reward function, and trajectories are sampled. The expected value of the features and the expected value of the experience are repeatedly calculated until the difference between them is less than a preset threshold. The preset threshold is determined based on the dimension and value range of the feature vectors, and the difference is quantified using Euclidean distance. When the Euclidean distance is less than this threshold, it indicates that the behavioral characteristics of the random policy have highly matched the expert decision-making model. At this point, parameter updates are stopped, and the final basic reward function is obtained. This function can accurately capture the core preferences and value orientations of expert scheduling decisions.

[0051] S15. Based on the basic reward function, construct a hierarchical reward function; the hierarchical reward function includes a basic reward function layer, an operational efficiency reward function layer, a cross-resource process coordination reward function layer, and a medical safety red line reward function layer.

[0052] For example, the basic reward function layer is the basic reward function. This is used to capture the implicit decision-making logic formed by experts in long-term scheduling practice, and can reflect the overall performance of scheduling actions. Operational efficiency reward function layer. The focus is on optimizing resource utilization efficiency and task execution speed to encourage increased utilization of core resources, accelerate the execution of high-priority tasks, and reduce idle waste of critical resources. This is achieved through a cross-resource process coordination reward function layer. Designed to address the complexities of multi-department and multi-resource collaboration within hospitals, this system aims to alleviate cross-process bottlenecks and improve the smoothness of the treatment process. It includes three sub-items: surgery-bed matching rewards, examination-treatment process integration rewards, and resource conflict mitigation rewards. (Medical safety red line reward function layer) It is a key constraint layer to ensure the safety of dispatching, and adopts a strong penalty mechanism to ensure that dispatching actions do not violate the bottom line of medical safety.

[0053] S16. Based on real-time hospital context indicators extracted from multi-dimensional state feature vectors, the weight coefficients of each reward function and internal indicator in the hierarchical reward function are dynamically calculated through a weighted neural network to obtain the context-aware dynamic reward function; the expression of the context-aware dynamic reward function is: ,in, and To adjust the hyperparameters that balance efficiency and the importance of coordinating objectives, Basic reward function layer, For the operational efficiency reward function layer, For cross-resource process coordination reward function layer, A reward function layer for medical safety red lines.

[0054] Optionally, real-time hospital context indicators are extracted from a multi-dimensional hospital status feature vector. These indicators aim to quantitatively reflect the hospital's current overall operational status and resource supply and demand, providing a basis for dynamic weight adjustments. They include emergency patient backlog rate, ICU bed occupancy rate, percentage of major surgeries, whether the hospital is in an emergency response state, and critical drug inventory warning level. Specifically, the emergency patient backlog rate is calculated as the ratio of the number of patients currently waiting for emergency treatment to the emergency service capacity; the ICU bed occupancy rate is the ratio of the number of currently occupied ICU beds to the total number of ICU beds; the percentage of major surgeries is the ratio of the number of major surgeries scheduled for the day to the total number of surgeries; whether the hospital is in an emergency response state is a binary identifier used to indicate whether the hospital is facing emergency scenarios such as mass casualties or public health emergencies; and the critical drug inventory warning level is a quantitative indicator reflecting the adequacy of inventory of critical drugs such as those in short supply and those in cold chain logistics.

[0055] For example, the weighted neural network is constructed using a lightweight multilayer perceptron. The network's input is a vector of real-time contextual indicators for the hospital, and its output is the weight coefficients of each reward function and its internal indicators within the hierarchical reward function. The weighted neural network is trained based on historical operational data of the hospital and corresponding optimal weight annotation data. The training objective is to ensure that the hierarchical reward function corresponding to the network's output weight coefficients maximizes the hospital's long-term overall performance. During training, a mean squared error loss function is used, and the network parameters are iteratively optimized using a stochastic gradient descent algorithm to ensure that the network learns the mapping relationship between contextual indicators and optimal weights.

[0056] in, This is a hyperparameter of the operational efficiency reward function layer, used to adjust the relative importance of efficiency targets in the total reward; Hyperparameters for the reward function layer in cross-resource process coordination are used to adjust the relative importance of collaborative objectives in the total reward. and It is not a fixed value, but rather a dynamic output by a weighted neural network based on real-time contextual indicators of the hospital.

[0057] S17. Based on the historical scheduling trajectory dataset containing expert demonstration trajectories, an offline reinforcement learning algorithm is used to pre-train the neural network model of the resource scheduling strategy to obtain a preliminary security strategy.

[0058] As an illustration, each trajectory in the dataset has been converted into the standard transfer tuple format for offline reinforcement learning. ,in for The state vector at time t, for Time-based scheduling actions, The immediate reward value is calculated based on a context-aware dynamic reward function. To carry out the operation The state vector obtained at the next moment.

[0059] Optionally, the pre-training process employs the Implicit Q-Learning (IQL) algorithm from offline reinforcement learning algorithms. Specifically, the state-value function is first trained. This function is used to approximate the expected value of the optimal Q-value under the current dataset distribution, and parameter optimization is achieved by minimizing the quantile regression loss; secondly, the Q-network is trained. By minimizing the Bellman error, the Q-network can accurately estimate the future cumulative reward of state-action pairs. Finally, the policy network, namely the resource scheduling policy neural network model, is trained. Policy optimization is achieved by maximizing the expected value of the policy under the evaluation of the Q-network. At the same time, a constraint mechanism is introduced to ensure that the action distribution of the policy output does not deviate too far from the historical data distribution. This enables the resource scheduling policy neural network model to learn effective scheduling patterns in historical data and has the ability to make safe and efficient scheduling decisions in normal operation scenarios.

[0060] S18. Based on a high-fidelity simulation environment simulating the dynamic operation of a hospital, with the context-aware dynamic reward function as the optimization objective and the inequality constraints derived from the medical safety red line rules as the constraint conditions, a constrained strategy optimization algorithm is used to fine-tune the initial safety strategy, resulting in a neural network model for the resource scheduling strategy.

[0061] Specifically, the optimization objective of the fine-tuning process is a context-aware dynamic reward function. This means maximizing the total reward value of scheduled actions through strategy adjustments to achieve comprehensive optimization of multiple objectives, including operational efficiency, process coordination, and medical safety. The constraints are derived from the medical safety red line rules and are expressed in the form of inequality constraints, specifically including... , , ,in, To limit waiting time for critically ill patients For dispatching operations The resulting waiting time for critically ill patients This is a preset safe threshold for waiting time for critically ill patients; To constrain resource qualification matching, For dispatching operations The degree of matching between the qualifications of resources and tasks; Due to emergency response time constraints, For dispatching operations The resulting average response time for emergency patients This is a preset emergency response time threshold. Furthermore, the Lagrange method from constrained policy optimization algorithms can be used for policy fine-tuning. By introducing Lagrange multipliers, the constrained optimization problem is transformed into an unconstrained optimization problem, achieving synergistic optimization of the optimization objective and constraints.

[0062] In one embodiment, a hierarchical reward function is constructed based on the basic reward function, including: S21. Based on the dynamic weighted sum of multiple operational indicators, construct an operational efficiency reward function layer; the expression for the operational efficiency reward function layer is: ,in, This represents the average increase in resource utilization after this scheduling. and Tasks to be scheduled The urgency level and estimated time, Penalties for resource idleness caused by scheduling , and These are dynamic weighting coefficients.

[0063] For example, This represents the increase in the average utilization rate of core resources after this scheduling. Core resources include large medical equipment such as operating rooms and CT / MRI machines, expert medical staff, and key treatment sites. The calculation method is the difference between the average utilization rate of all core resources after the scheduling action and the average utilization rate before the scheduling. The average utilization rate is obtained by weighted summation of the utilization rates of individual resources, with the weight being the strategic importance coefficient of the resource. This refers to the set of tasks to be scheduled for execution in this scheduling process. Tasks to be scheduled The urgency level value; Tasks to be scheduled The estimated time is determined based on historical statistics of the execution time of similar tasks, and after adjustments for factors such as task complexity and individual patient differences. Tasks to be scheduled The "value density" metric is used to measure the priority benefit of completing the task per unit of time, and encourages the scheduling system to prioritize tasks with high urgency and short duration. The penalty value for the idle critical resources caused by this scheduling is the penalty value for the idle resources. Critical resources refer to resources that are highly scarce and irreplaceable. The penalty value for the idle resources is calculated by multiplying the idle time by the opportunity cost per unit time of the resource. The opportunity cost per unit time of the resource is comprehensively evaluated based on factors such as the resource purchase cost, maintenance cost, and frequency of use.

[0064] S22. Based on a lightweight discrete event simulation model, the hospital process of the target cycle after scheduling is deduced, and a cross-resource process coordination reward function layer is constructed based on the rewards corresponding to the deduction results: the rewards corresponding to the deduction results include a positive reward if the scheduling makes the operation end time match the bed idle time; if the scheduling resolves or mitigates the future resource conflict identified based on the resource conflict matrix, a positive reward is given according to the urgency value of the conflict task and the resource scarcity value.

[0065] Specifically, the resource allocation plan and task execution sequence of the current scheduling action are input into a lightweight discrete event simulation model. The model uses the current hospital resource status, the attributes of the tasks to be scheduled, and process constraint rules as initial conditions, simulating the progression of the treatment process event by event in a time sequence, including key aspects such as task execution, resource occupation and release, and process node connections. During the simulation, the model records in real time the matching status of surgery completion time and bed idle time, the resolution of resource conflicts, and other core coordination indicators, providing data support for reward calculation.

[0066] One type of reward corresponding to the simulation results is the surgery-bed matching reward. A positive reward is given when the simulation shows that the scheduling action ensures the expected end time of a surgery and the expected idle time of a ward or ICU bed fall within a preset reasonable matching range. The matching degree is quantified by calculating the time difference between the surgery end time and the bed idle time; the smaller the time difference, the higher the matching degree, and the larger the corresponding reward value. If the time difference is 0, the highest matching reward is given. Another type is the resource conflict mitigation reward. During the simulation, the resource conflict matrix is ​​called to identify potential resource conflict changes within the future target period after the scheduling action is executed. If the scheduling action transforms elements marked as conflict in the original resource conflict matrix into non-conflict elements, or significantly reduces the degree of conflict, a positive reward is calculated based on the urgency value of the conflict-related task and the resource scarcity value. The urgency value of the conflicting task is... Resource scarcity is a quantitative indicator, which is comprehensively evaluated based on factors such as the difficulty of acquiring resources in the market, purchase costs, maintenance complexity, and the number of alternative resources. The higher the value, the scarcer the resource. The reward value is a weighted sum of the urgency value and the resource scarcity value. The weight is determined by hospital process collaboration experts to ensure that higher rewards are given for resolving conflicts between high-urgency tasks and highly scarce resources.

[0067] S23. Construct a medical safety red line reward function layer based on hard penalty mechanism and soft constraint proximity condition: The hard penalty mechanism corresponds to giving a negative reward and triggering an alarm if scheduling causes any task to violate the preset inviolable red line constraint; the soft constraint proximity condition corresponds to determining a negative reward based on the difference between the target indicator and the danger threshold if scheduling causes the target indicator to deteriorate to less than the danger threshold.

[0068] Optionally, the final reward value of the medical safety red line reward function layer is the sum of the reward values ​​corresponding to the hard penalty mechanism and the soft constraint proximity condition.

[0069] S24. Construct a layered reward function based on the basic reward function, the operational efficiency reward function layer, the cross-resource process coordination reward function layer, and the medical safety red line reward function layer.

[0070] In one embodiment, based on real-time hospital context indicators extracted from a multi-dimensional state feature vector, a context-aware dynamic reward function is obtained by dynamically calculating the weight coefficients of each reward function and internal indicator in the hierarchical reward function through a weighted adjustment neural network, including: S31. Based on the multidimensional hospital state feature vector, extract the hospital's overall load index, including emergency patient backlog rate, ICU occupancy rate, proportion of major surgeries, time period pattern, and emergency event markers to obtain a real-time context state vector.

[0071] For illustration, the emergency patient backlog rate is extracted based on emergency patient-related data from the hospital's overall load indicators, calculated as the ratio of the number of emergency patients currently waiting for treatment to the emergency service capacity. The ICU occupancy rate is extracted directly from the ICU bed occupancy rate indicator in the multidimensional hospital status feature vector, calculated as the ratio of the number of currently occupied ICU beds to the total number of ICU beds. The major surgery percentage is extracted based on the scheduled surgical tasks for the day, calculated as the ratio of the number of major surgeries scheduled for the day to the total number of surgeries scheduled for the day. The time period pattern is extracted to distinguish the resource allocation characteristics of different hospital operating periods, with specific values ​​determined according to the time interval to which the current moment belongs, covering four patterns: daytime regular periods, nighttime periods, weekend periods, and holiday periods. The emergency event flag is a binary identifier used to indicate whether the hospital is in an emergency response state, with a value of 1 or 0, where 1 indicates the existence of an emergency and 0 indicates normal operation.

[0072] By arranging the above five indicators in a preset order, a real-time context state vector is formed. Its vector form is ,in For emergency room patient backlog rate, For ICU occupancy rate, The proportion of major surgeries, This is a numerical identifier for the time period pattern. This serves as a marker for sudden events.

[0073] S32. Input the real-time context state vector into the pre-trained weight adjustment neural network to obtain the set of weight coefficients for the context-aware dynamic reward function at the current time step; the expression for the set of weight coefficients is: ,in, Adjusting the weights of the neural network For real-time context state vector, Adjust the parameters of the neural network for the weights; The weighted neural network is trained through the following steps: Based on the multidimensional performance comprehensive scoring function, historical scheduling trajectories and multidimensional performance comprehensive scores of historical scheduling trajectories under different real-time context state vectors are extracted from the historical scheduling trajectory dataset. With the goal of maximizing the correlation between the multidimensional performance comprehensive score and the predicted cumulative reward induced by the context-aware dynamic reward function adjusted by weights, the parameters of the weight-adjusted neural network are trained to obtain the weight-adjusted neural network.

[0074] Indicative, For the weight-adjusting neural network, a lightweight multilayer perceptron (MLP) structure is used, with the input layer dimension and the real-time context state vector being... The dimensions are consistent, and the hidden layer uses the ReLU activation function to introduce non-linear mapping capability, ensuring that it can learn the complex relationship between the context state and the weight coefficients; The set of parameters for weighting a neural network, including connection weights and bias terms between neurons in each layer of the network; The hyperparameter weights of the operational efficiency reward function layer at the current moment are used to adjust the relative importance of operational efficiency rewards in the total reward of the context-aware dynamic reward function; The hyperparameter weights of the cross-resource process coordination reward function layer at the current moment are used to adjust the relative importance of process coordination rewards in the total reward; , , These are the dynamic weighting coefficients for the three indicators in the current operational efficiency reward function layer: average utilization increment, total task value density, and resource idle penalty. They are used to adjust the contribution intensity of each indicator in the operational efficiency reward calculation.

[0075] Optionally, the historical scheduling trajectory dataset contains complete scheduling records accumulated during the hospital's past operations. It is segmented into trajectory segments based on "shifts" along the time dimension. Each trajectory segment corresponds to a hospital state sequence, scheduling action sequence, and execution result data within a specific time period. For each trajectory segment, the real-time context state vector at the start time of that trajectory segment is calculated. Simultaneously, a multidimensional performance comprehensive scoring function is used to calculate the comprehensive score of this trajectory segment. This refers to the multi-dimensional performance comprehensive score of historical scheduling trajectories.

[0076] The goal is to maximize the correlation between the multidimensional performance composite score and the predicted cumulative reward induced by the weighted context-aware dynamic reward function. The parameters of the weighted adjustment neural network are trained accordingly. ,in, Adjust the training loss value of the neural network for the weights; The function for calculating the Pearson correlation coefficient; Historical scheduling trajectory A multi-dimensional performance comprehensive score; The scheduling strategy induced by adjusting the weight coefficients of the neural network output for weight adjustment; For this inducement strategy in the historical context state The expected cumulative reward is calculated by running the inducement strategy in a high-fidelity simulation environment, simulating the operational scenario corresponding to the historical scheduling trajectory, and statistically analyzing the cumulative output value of the context-aware dynamic reward function during the strategy execution process. This is the context-aware dynamic reward function with adjusted weights, where the weights are determined by the weights of the neural network in the historical context. Output below.

[0077] The training process uses the stochastic gradient descent algorithm to minimize the loss function. Iteratively update the weights to adjust the parameters of the neural network. That is, initialization parameters For each historical context state, the value is random; The process involves adjusting the set of weight coefficients output by the weighted neural network; adjusting the context-aware dynamic reward function based on these weight coefficients; loading the initial state and operational scenario parameters corresponding to historical scheduling trajectories into a high-fidelity simulation environment; running the resource scheduling strategy neural network model to calculate the expected cumulative reward; calculating the correlation loss value of this batch of data; and calculating the loss function with respect to the parameters using the backpropagation algorithm. The gradient; updating parameters along the gradient descent direction. Repeat the above steps until the loss function value converges to the preset threshold or reaches the maximum number of iterations. At this point, the trained weighted neural network is obtained.

[0078] Real-time context state vector The trained weight adjustment neural network is input, and the network outputs the set of weight coefficients at the current time step by learning the mapping relationship between the context and the weights. By substituting this set of weight coefficients into the hierarchical reward function, a context-aware dynamic reward function can be obtained, enabling the reward function to dynamically adapt to the real-time operation scenario of the hospital.

[0079] In one embodiment, based on a high-fidelity simulation environment simulating the dynamic operation of a hospital, using a context-aware dynamic reward function as the optimization objective and inequality constraints derived from medical safety red line rules as constraints, a constrained policy optimization algorithm is employed to fine-tune the initial safety policy, resulting in a resource scheduling policy neural network model, including: S41. Transform the hard penalty mechanism and soft constraint proximity condition in the medical safety red line reward function layer into a set of inequality constraints; the expression for the inequality constraints is: , ,in To constrain the allowable threshold, This is a real-time state vector of the resource. To schedule trajectory actions.

[0080] Indicative, For the first A constraint function is used to quantify scheduling actions. Current real-time resource status The degree to which a certain medical safety constraint is met; For the first The allowable threshold of a constraint, that is, the maximum value allowed for the constraint function; This represents the total number of constraint functions.

[0081] Optionally, for the inviolable red line constraint corresponding to the hard penalty mechanism, the constraint function... Defined as the difference between the actual value and the safety threshold, such as the waiting time constraint function for critically ill patients. ,in For dispatching operations The resulting waiting time for critically ill patients The safe threshold for waiting time for critically ill patients, and the corresponding allowable threshold. That is, requirements Ensure that the waiting time for critically ill patients does not exceed the safety threshold; for key operational indicator constraints corresponding to soft constraint proximity conditions, the constraint function... Defined as the difference between the actual value of the indicator and the danger threshold, for example, the emergency response time constraint function is... ,in For dispatching operations The resulting average response time for emergency patients The dangerous threshold for emergency response time, and the corresponding allowable threshold. The preset tolerance deviation value allows the actual value of the indicator to slightly exceed the danger threshold, so as to retain the flexibility of strategy optimization.

[0082] S42. Based on the Lagrange relaxation method, the constrained policy optimization problem is transformed into an unconstrained optimization problem, and the Lagrange function is constructed; the expression of the Lagrange function is: ,in, For strategy The expectation of context-aware dynamic rewards. It is a Lagrange multiplier.

[0083] For example, For example, the value of the Lagrange function reflects the optimization performance of the strategy and the degree of constraint satisfaction. This is the parameter set of the neural network model for resource scheduling strategies, including the connection weights and bias terms of each layer of the network; Let Lagrange multiplier vectors be the vectors of multipliers, each multiplier... Corresponding to an inequality constraint This is used to adjust the penalty intensity of the corresponding constraint during the optimization process; For strategy The expected cumulative reward value, i.e. ,in For parameters The resource scheduling strategy neural network model, The strategy executes the trajectory in a high-fidelity simulation environment. Context-aware dynamic rewards The mathematical expectation reflects the overall optimization performance of the strategy; For strategy execution trajectory constraint functions The mathematical expectation, constraint function In the trajectory The value of is the average value of the constraint function corresponding to each state-action pair in the trajectory; For the first The allowable threshold for each constraint; To constrain the penalty term, when the policy violates the constraint... When the penalty term is positive, the Lagrange function value decreases, thus suppressing this type of policy that violates the constraints; when the policy satisfies the constraints... When the penalty term is non-positive, the Lagrange function value and the expected cumulative reward value... The convergence of approaches encourages strategies to optimize within constraints.

[0084] S43. Calculate the update gradient based on the Lagrange function, and update the parameters and Lagrange multipliers of the resource scheduling strategy neural network model along the update gradient direction; the update formula for the parameters of the resource scheduling strategy neural network model is: The updated form of the Lagrange multipliers is: .

[0085] For example, This is the set of parameters for the neural network model of the resource scheduling strategy to be updated. The learning rate is a policy parameter used to control the step size of each parameter update. The value is adjusted adaptively, with a larger value set in the early stage of training to speed up convergence and a smaller value gradually reduced as training progresses to ensure parameter stability. Expected cumulative reward value For strategy parameters The gradient, in the direction of parameter adjustment, points to the direction that improves the overall performance of the strategy; For the first Lagrange multipliers corresponding to each constraint; Expectation of the constraint function For strategy parameters The gradient is directed in the direction of parameter adjustment that exacerbates the constraint violation. To constrain the gradient penalty term, Lagrange multipliers are used. Adjusting the weights of constraint gradients so that when a constraint is frequently violated, the corresponding... Increasing the value of the constraint gradient penalty term enhances its influence, guiding the parameter update direction away from the direction that leads to constraint violations, thereby reducing constraint violations.

[0086] The first one to be updated One Lagrange multiplier; The learning rate of the Lagrange multiplier is used to control the step size of the multiplier update. It is set to a preset fixed value to ensure the stability of the multiplier update. For the first in the strategy execution trajectory The expectation of a constraint function; For the first The allowable threshold for each constraint; To constrain the degree of violation index, when this index is positive, it indicates that the strategy violates the constraint, and the multiplier... Increasing this value strengthens the penalty for this constraint in subsequent parameter updates; when this index is non-positive, it indicates that the policy satisfies the constraint, and the multiplier... Keep it unchanged or reduce it to avoid excessive penalties that limit strategy optimization. The function is used to ensure the Lagrange multipliers .

[0087] In one embodiment, the urgency value in the attribute set of the task to be scheduled is updated based on the dynamic priority adjustment instruction to obtain a priority change summary feature. This priority change summary feature is then concatenated with a multi-dimensional hospital state feature vector to obtain an enhanced state vector, including: S51. Based on the urgency value of the dynamic priority adjustment instruction, update the urgency value of the target task to be scheduled, and obtain the priority change event; the priority change event includes the initiating role.

[0088] The dynamic priority adjustment instruction includes the target task identifier, the adjusted urgency value, the initiating role identifier, the adjustment timestamp, and the adjustment reason. Specifically, it replaces the original urgency value in the target task attribute set and generates a priority change event. The structured format of the priority change event includes a unique event identifier, the target task identifier, the urgency value before adjustment, the urgency value after adjustment, the initiating role, the adjustment timestamp, and the adjustment reason code.

[0089] S52. Update the priority change event to the unified time-series event graph, and update the multi-dimensional hospital status feature vector based on the updated unified time-series event graph.

[0090] Furthermore, based on the adjustment timestamp of the priority change event, the event is inserted into the corresponding time sequence position in the graph to ensure the temporal integrity of the graph; the event links in the graph associated with the target scheduled task are updated.

[0091] S53. Calculate the total number of priority change events and the average priority change magnitude of all priority change events that occurred within the preset time window in the past, and check whether there are priority change events initiated by the preset initiating role, and obtain the priority change summary characteristics.

[0092] For example, the total number of events is counted using a distributed counting method. By traversing the priority change events within the time window in the unified time-series event graph, the number of valid events is accumulated to form the "total number of priority change events" indicator. This indicator reflects the frequency of recent priority adjustments; frequent changes usually indicate significant fluctuations in the hospital's operational status. The average priority change magnitude is calculated by extracting the pre-adjustment urgency value and the post-adjustment urgency value for each valid priority change event, and calculating the change magnitude of a single event, i.e., the absolute value of the difference between the post-adjustment value and the pre-adjustment value. The arithmetic mean of the change magnitudes of all events is then used to obtain the average priority change magnitude. This indicator reflects the overall intensity of recent priority adjustments; a larger average magnitude indicates a concentrated adjustment of high-urgency tasks.

[0093] S54. Concatenate the priority change summary features with the updated multidimensional hospital status feature vector to obtain the enhanced status vector.

[0094] Optionally, feature splicing adopts a vector concatenation strategy, with the splicing order explicitly set as "updated multidimensional hospital status feature vector + priority change summary feature".

[0095] In one embodiment, the real-time heterogeneous event stream of the hospital is standardized and cross-system correlated to generate a unified time-series event graph. A multi-dimensional hospital state feature vector at the current moment is then extracted from the unified time-series event graph, including: S61. Standardize the real-time heterogeneous event stream according to the preset five-tuple format to obtain a standardized event stream; the five-tuple format includes subject, predicate, object, timestamp and attribute.

[0096] The preset five-tuple format is a unified event description framework designed for hospital diagnosis and treatment scenarios and resource management characteristics. Its core function is to eliminate the heterogeneity of event formats across different systems and achieve structured alignment of events across systems.

[0097] S62. Based on event identifiers, link events of the same medical process in the standardized event flow into workflow instances, and maintain the real-time status of all resource entities based on timestamps to obtain the real-time status of resources; the real-time status of resources includes idle, reserved, occupied and unavailable.

[0098] Event identifiers are the core indexes for cross-system event association. They consist of a globally unified set of association keys, including four core association keys: patient unique ID, treatment application number, resource unique identifier, and workflow instance ID. Different types of events select the corresponding association key combination according to the business scenario.

[0099] Specifically, initial matching is performed based on event identifiers, aggregating standardized events with the same association key combination into event groups to ensure that events within an event group belong to the same subject or the same diagnostic and treatment task. Further, logical validation is performed based on preset business process rules. During validation, if the order of events within an event group violates the business process rules, the event group is automatically split and re-matched to ensure that the linked workflow instances conform to clinical diagnostic and treatment logic. Each workflow instance includes an event sequence, a set of associated subjects, a set of associated resources, and the current process status.

[0100] The maintenance of real-time resource status is implemented based on an event-driven state machine model. The initial state of a resource entity is idle, and state transitions are triggered by specific predicates in a standardized event stream. For example, the state transition rules are explicit and exclusive: when a "reservation" type predicate event is received, the resource state changes from idle to reserved.

[0101] S63. Take snapshots of the dynamic network consisting of all workflow instances and real-time resource statuses at preset time intervals to generate a unified time-series event graph.

[0102] This is an illustrative dynamic network comprised of the real-time states of all workflow instances and resources. Network nodes include workflow instance nodes and resource entity nodes, while network edges include association edges between workflow instances and entities, occupancy / reservation edges between workflow instances and resources, and mapping edges between resource entities and their states. When taking a snapshot of this dynamic network, the attribute information of all nodes and the relationships between all edges at the current moment are captured simultaneously, forming a static network snapshot.

[0103] The generation of the unified time-series event graph is based on the aggregation and indexing of snapshot sequences. Each static network snapshot serves as a time slice of the graph and is stored in the time-series database in timestamp order. The graph structure adopts an attribute graph model, with nodes divided into three categories: workflow nodes, resource nodes, and main nodes, and edges divided into three categories: associated edges, occupied edges, and reserved edges. All types of nodes and edges carry timestamp attributes.

[0104] S64. From the unified time-series event graph, statistically calculate the overall hospital load index and extract the micro-attribute vectors of each task to be scheduled; the overall hospital load index includes the number of patients waiting for emergency treatment, ICU bed occupancy rate, and average queue length of large equipment; the micro-attribute vectors include task type, average estimated time, current urgency level, list of required resources, ready time, and latest allowed time.

[0105] For example, the number of patients waiting for emergency treatment is counted by filtering nodes in the graph whose subject type is "patient" and whose associated workflow instance has a process status of "emergency registration completed" and has not had an "emergency treatment started" event. The number of such nodes is accumulated, and patients who have canceled their registration, been referred, or completed their treatment are automatically excluded during the statistical process. ,in, This refers to the total number of ICU beds. This represents the number of beds currently occupied. in, This refers to the total number of large equipment. For the first The queuing length of large equipment is summed to cover all preset large equipment, ensuring that the indicator reflects the overall queuing pressure of large equipment.

[0106] Task types are represented using standardized codes, such as 01 for surgical tasks, 02 for examination tasks, and 03 for treatment tasks. The estimated average execution time is based on statistical analysis of historical workflow instances of similar tasks in the graph. Historical instances with the same task type and complexity as the current task are selected, their execution times are extracted, and the arithmetic mean is calculated as the estimated average execution time of the current task. The current urgency level is directly extracted from the attribute fields of the workflow instance nodes. The list of required resources is formed by traversing the associated resource edges of the workflow instance nodes and extracting the unique identifiers of the associated resource nodes. The ready time is the timestamp when the workflow instance meets all the prerequisites, and the timestamp of the last prerequisite event is determined as the ready time. The latest allowed time is determined based on medical safety rules and task type.

[0107] S65. The overall hospital load index, real-time resource status, and micro-attribute vectors of each task to be scheduled are concatenated and encoded to obtain a multi-dimensional hospital status feature vector.

[0108] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0109] Based on the same inventive concept, this application also provides an intelligent management system for hospital medical resources to implement the intelligent management method for hospital medical resources described above. The solution provided by this system is similar to the solution described in the above method; therefore, the specific limitations of one or more embodiments of the intelligent management system for hospital medical resources provided below can be found in the limitations of the intelligent management method for hospital medical resources described above, and will not be repeated here.

[0110] In one exemplary embodiment, such as Figure 2 As shown, an intelligent management system for hospital medical resources is provided, including: Event Resource Module 201 is used to standardize and perform cross-system correlation processing on the real-time heterogeneous event stream of the hospital, generate a unified time-series event graph, and extract the multi-dimensional hospital status feature vector at the current moment from the unified time-series event graph. The multi-dimensional hospital status feature vector includes the overall hospital load index, the real-time status of each resource, the attribute set of all tasks to be scheduled, and the resource conflict matrix. The real-time heterogeneous event stream is collected in real time through various information systems and physical sensing nodes of the hospital. The emergency demand module 202 is used to respond to the acquisition of a dynamic priority adjustment instruction for a specific scheduling task, update the urgency value in the attribute set of the task to be scheduled based on the dynamic priority adjustment instruction, obtain the priority change summary feature, and concatenate the priority change summary feature with the multi-dimensional hospital status feature vector to obtain the enhanced status vector. The candidate scheduling module 203 is used to input the enhanced state vector into the pre-trained resource scheduling strategy neural network model to obtain the scheduling action probability distribution of all tasks to be scheduled, and generate at least one candidate resource scheduling scheme based on the scheduling action probability distribution. The red line verification module 204 is used to perform simulation verification on candidate resource scheduling schemes based on preset medical safety red line rules, and to mark all candidate resource scheduling schemes that pass the verification as valid candidate resource scheduling schemes. The recommended scheduling module 205 is used to calculate the expected utility evaluation value of each effective candidate resource scheduling scheme based on the context-aware dynamic reward function, and to take the effective candidate resource scheduling scheme with the highest expected utility evaluation value as the recommended scheduling trajectory.

[0111] In one embodiment, a model building module is also included, for: Based on a multidimensional performance comprehensive scoring function, all historical scheduling trajectories in the historical scheduling trajectory dataset are scored, and historical scheduling trajectories with scores in the top preset percentage are selected as expert demonstration trajectories, while historical scheduling trajectories with scores in the bottom preset percentage are marked as poor demonstration trajectories. The multidimensional performance comprehensive scoring function includes patient satisfaction, resource utilization, critical patient waiting time, cost control, and number of manual interventions. Based on a parameterized fundamental reward function neural network, a stochastic policy is trained in a high-fidelity simulation environment using a reinforcement learning algorithm, and multiple trajectories are obtained by sampling based on the stochastic policy; the expression for the fundamental reward function neural network is: ,in, This is a real-time state vector of the resource. In order to schedule trajectory actions, The weight parameters to be learned. For state-action pair features, The scalar reward value; Calculate the expected value of state-action pairs in the trajectory and the empirical expected value of state-action pairs in the expert demonstration trajectory. The gradient of the maximum entropy inverse reinforcement learning objective function based on contrastive loss is calculated using the expected value and the empirical expected value. Based on this gradient, the weight parameters of the parameterized base reward neural network are updated using gradient ascent until the difference between the expected value and the empirical expected value is less than a preset threshold, thus obtaining the corresponding expert-scheduled base reward function. The expression for the maximum entropy inverse reinforcement learning objective function based on contrastive loss is: ,in, For gradient, To demonstrate the trajectory to experts, This serves as a negative example. Marginal value, These are the weighting coefficients; Based on the basic reward function, a hierarchical reward function is constructed; the hierarchical reward function includes a basic reward function layer, an operational efficiency reward function layer, a cross-resource process coordination reward function layer, and a medical safety red line reward function layer. Based on real-time hospital context indicators extracted from multidimensional state feature vectors, a context-aware dynamic reward function is obtained by dynamically calculating the weight coefficients of each reward function and internal indicator in the hierarchical reward function through a weighted neural network. The expression of the context-aware dynamic reward function is as follows: ,in, and To adjust the hyperparameters that balance efficiency and the importance of coordinating objectives, Basic reward function layer, For the operational efficiency reward function layer, For cross-resource process coordination reward function layer, A reward function layer for medical safety red lines; Based on a historical scheduling trajectory dataset containing expert demonstration trajectories, an offline reinforcement learning algorithm is used to pre-train a neural network model for resource scheduling strategies to obtain a preliminary security strategy. Based on a high-fidelity simulation environment simulating the dynamic operation of a hospital, a neural network model for resource scheduling strategy is obtained by using a context-aware dynamic reward function as the optimization objective and inequality constraints derived from medical safety red line rules as the constraint conditions, and fine-tuning the initial safety strategy with a constrained strategy optimization algorithm.

[0112] In one embodiment, a hierarchical reward function module is also included for: An operational efficiency reward function layer is constructed based on the dynamic weighted sum of multiple operational indicators; the expression of the operational efficiency reward function layer is as follows: ,in, This represents the average increase in resource utilization after this scheduling. and Tasks to be scheduled The urgency level and estimated time, Penalties for resource idleness caused by scheduling , and These are dynamic weighting coefficients; Based on a lightweight discrete event simulation model, the hospital process in the target cycle after scheduling is deduced, and a cross-resource process coordination reward function layer is constructed based on the rewards corresponding to the deduction results: the rewards corresponding to the deduction results include a positive reward if the scheduling makes the operation end time match the bed idle time; if the scheduling resolves or mitigates the future resource conflicts identified based on the resource conflict matrix, a positive reward is given according to the urgency value of the conflict task and the resource scarcity value. A medical safety red line reward function layer is constructed based on a hard penalty mechanism and a soft constraint proximity condition: the hard penalty mechanism corresponds to giving a negative reward and triggering an alarm if scheduling causes any task to violate the preset inviolable red line constraint; the soft constraint proximity condition corresponds to determining a negative reward based on the difference between the target indicator and the danger threshold if scheduling causes the target indicator to deteriorate to below the danger threshold. A hierarchical reward function is constructed based on a basic reward function, an operational efficiency reward function layer, a cross-resource process coordination reward function layer, and a medical safety red line reward function layer.

[0113] In one embodiment, a weighting module is also included, for: Based on the overall hospital load index of multidimensional hospital state feature vector, the backlog rate of emergency patients, ICU occupancy rate, proportion of major surgeries, time period pattern and emergency event markers are extracted to obtain the real-time context state vector. The real-time context state vector is input into a pre-trained weight adjustment neural network to obtain the set of weight coefficients for the context-aware dynamic reward function at the current time step; the expression for the set of weight coefficients is: ,in, Adjusting the weights of the neural network For real-time context state vector, Adjust the parameters of the neural network for the weights.

[0114] In one embodiment, the model building module is further configured to: The hard penalty mechanism and soft constraint proximity condition in the medical safety red line reward function layer are transformed into a set of inequality constraints; the expression of the inequality constraints is as follows: , ,in To constrain the allowable threshold, This is a real-time state vector of the resource. To schedule trajectory actions; Based on the Lagrange relaxation method, the constrained policy optimization problem is transformed into unconstrained optimization, and a Lagrange function is constructed; the expression of the Lagrange function is: ,in, For strategy The expectation of context-aware dynamic rewards. For Lagrange multipliers; The update gradient is calculated based on the Lagrange function, and the parameters and Lagrange multipliers of the resource scheduling strategy neural network model are updated along the direction of the update gradient. The update formula for the parameters of the resource scheduling strategy neural network model is as follows: The updated form of the Lagrange multipliers is: .

[0115] In one embodiment, the emergency demand module 202 is further configured to: Based on the urgency value of the dynamic priority adjustment instruction, the urgency value of the target task to be scheduled is updated to obtain the priority change event; the priority change event includes the initiating role; Update the priority change event to the unified time series event graph, and update the multidimensional hospital status feature vector based on the updated unified time series event graph; The total number of priority change events and the average priority change magnitude of all priority change events that occurred within the preset time window are counted, and it is checked whether there are priority change events initiated by preset initiating roles to obtain priority change summary features. The priority change summary features are concatenated with the updated multidimensional hospital state feature vector to obtain the enhanced state vector.

[0116] In one embodiment, the event resource module 201 is further configured to: The real-time heterogeneous event stream is standardized according to the preset five-tuple format to obtain the standardized event stream; the five-tuple format includes subject, predicate, object, timestamp and attribute; Based on event identifiers, events linked together in the same medical process within a standardized event flow are defined as workflow instances. The real-time status of all resource entities is maintained based on timestamps to obtain the real-time status of resources. The real-time status of resources includes idle, reserved, occupied, and unavailable. Take snapshots of the dynamic network consisting of all workflow instances and real-time resource statuses at preset time intervals to generate a unified time-series event graph; From the unified time-series event graph, the overall hospital load index is statistically calculated and the micro-attribute vectors of each task to be scheduled are extracted. The overall hospital load index includes the number of patients waiting for emergency treatment, ICU bed occupancy rate, and average queue length for large equipment. The micro-attribute vectors include task type, average estimated time, current urgency level, list of required resources, ready time, and latest allowed time. By concatenating and encoding the overall hospital load index, real-time resource status, and micro-attribute vectors of each task to be scheduled, a multi-dimensional hospital status feature vector is obtained.

[0117] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0118] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0119] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0120] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for intelligent management of hospital medical resources, characterized in that, The method includes: The real-time heterogeneous event stream of the hospital is standardized and cross-system correlated to generate a unified time-series event graph. A multi-dimensional hospital state feature vector at the current moment is extracted from the unified time-series event graph. The multi-dimensional hospital state feature vector includes the overall hospital load index, the real-time status of each resource, the attribute set of all tasks to be scheduled, and the resource conflict matrix. The real-time heterogeneous event stream is collected in real time through various information systems and physical sensing nodes of the hospital. In response to receiving a dynamic priority adjustment instruction for a specific scheduling task, the urgency value in the attribute set of the task to be scheduled is updated based on the dynamic priority adjustment instruction to obtain a priority change summary feature. The priority change summary feature is then concatenated with the multidimensional hospital status feature vector to obtain an enhanced status vector. The enhanced state vector is input into a pre-trained resource scheduling strategy neural network model to obtain the scheduling action probability distribution of all the tasks to be scheduled, and at least one candidate resource scheduling scheme is generated based on the scheduling action probability distribution. The candidate resource scheduling schemes are deduced and verified based on the preset medical safety red line rules, and all candidate resource scheduling schemes that pass the verification are marked as valid candidate resource scheduling schemes. Based on the context-aware dynamic reward function, the expected utility evaluation value of each of the effective candidate resource scheduling schemes is calculated, and the effective candidate resource scheduling scheme with the highest expected utility evaluation value is taken as the recommended scheduling trajectory.

2. The method according to claim 1, characterized in that, The neural network model for the resource scheduling strategy is obtained through the following method: Based on the multidimensional performance comprehensive scoring function, all historical scheduling trajectories in the historical scheduling trajectory dataset are scored, and historical scheduling trajectories with scores in the top preset percentage are selected as expert demonstration trajectories, while historical scheduling trajectories with scores in the bottom preset percentage are marked as poor demonstration trajectories. The multidimensional performance comprehensive scoring function includes patient satisfaction, resource utilization, critical patient waiting time, cost control, and number of manual interventions. A parameterized basic reward function neural network is used to train a random policy in a high-fidelity simulation environment through a reinforcement learning algorithm, and multiple trajectories are obtained by sampling based on the random policy; the expression of the basic reward function neural network is as follows: ,in, This is a real-time state vector of the resource. To schedule trajectory actions, The weight parameters to be learned. For state-action pair features, The scalar reward value; Calculate the expected value of the state-action pair features in the trajectory, and calculate the empirical expected value of the state-action pair features in the expert demonstration trajectory; The gradient of the maximum entropy inverse reinforcement learning objective function based on contrastive loss is calculated based on the expected value and the empirical expected value. Then, based on the gradient, the weight parameters to be learned in the parameterized base reward neural network are updated using gradient ascent until the difference between the expected value and the empirical expected value is less than a preset threshold, thus obtaining the corresponding expert-scheduled base reward function. The expression for the maximum entropy inverse reinforcement learning objective function based on contrastive loss is: ,in, For gradient, To demonstrate the trajectory to experts, This serves as a negative example. Marginal value, These are the weighting coefficients; Based on the aforementioned basic reward function, a hierarchical reward function is constructed; the hierarchical reward function includes a basic reward function layer, an operational efficiency reward function layer, a cross-resource process coordination reward function layer, and a medical safety red line reward function layer. Based on the real-time hospital context indicators extracted from the multidimensional state feature vector, the weight coefficients of each reward function and internal indicator in the hierarchical reward function are dynamically calculated through a weighted adjustment neural network to obtain the context-aware dynamic reward function; the expression of the context-aware dynamic reward function is as follows: ,in, and To adjust the hyperparameters that balance efficiency and the importance of coordinating objectives, For the basic reward function layer, This refers to the operational efficiency reward function layer. For the cross-resource process coordination reward function layer, This is the reward function layer for the medical safety red line; Based on the historical scheduling trajectory dataset containing the expert demonstration trajectories, an offline reinforcement learning algorithm is used to pre-train the neural network model of the resource scheduling strategy to obtain a preliminary security strategy. Based on the high-fidelity simulation environment simulating the dynamic operation of a hospital, the context-aware dynamic reward function is used as the optimization objective, and the inequality constraints derived from the medical safety red line rules are used as the constraints. A constrained strategy optimization algorithm is employed to fine-tune the initial safety strategy, resulting in the neural network model of the resource scheduling strategy.

3. The method according to claim 2, characterized in that, The construction of a hierarchical reward function based on the basic reward function includes: Based on the dynamic weighted sum of multiple operational indicators, the operational efficiency reward function layer is constructed; the expression of the operational efficiency reward function layer is as follows: ,in, This represents the average increase in resource utilization after this scheduling. and Tasks to be scheduled The urgency level and estimated time, Penalties for resource idleness caused by scheduling , and These are dynamic weighting coefficients; Based on a lightweight discrete event simulation model, the hospital process for the target cycle after scheduling is deduced, and the cross-resource process coordination reward function layer is constructed based on the rewards corresponding to the deduction results: the rewards corresponding to the deduction results include a positive reward if scheduling makes the operation end time match the bed idle time; if scheduling resolves or mitigates future resource conflicts identified based on the resource conflict matrix, a positive reward is given according to the urgency value of the conflict task and the resource scarcity value. The medical safety red line reward function layer is constructed based on a hard penalty mechanism and a soft constraint proximity condition: the hard penalty mechanism corresponds to giving a negative reward and triggering an alarm if scheduling causes any task to violate the preset inviolable red line constraint; the soft constraint proximity condition corresponds to determining a negative reward based on the difference between the target indicator and the danger threshold if scheduling causes the target indicator to deteriorate to less than the danger threshold. The hierarchical reward function is constructed based on the basic reward function, the operational efficiency reward function layer, the cross-resource process coordination reward function layer, and the medical safety red line reward function layer.

4. The method according to claim 3, characterized in that, The context-aware dynamic reward function is obtained by dynamically calculating the weight coefficients of each reward function and internal indicator in the hierarchical reward function through a weighted neural network based on the real-time hospital context indicators extracted from the multidimensional state feature vector, including: Based on the overall hospital load index of the multidimensional hospital state feature vector, the emergency patient backlog rate, ICU occupancy rate, proportion of major surgeries, time period pattern and emergency event markers are extracted to obtain the real-time context state vector. The real-time context state vector is input into a pre-trained weight adjustment neural network to obtain the set of weight coefficients for the context-aware dynamic reward function at the current time step; the expression for the set of weight coefficients is: ,in, Adjust the neural network for the weights. The real-time context state vector, Adjust the parameters of the neural network for the weights; The weight adjustment neural network is trained through the following steps: Based on the multidimensional performance comprehensive scoring function, the historical scheduling trajectory and the multidimensional performance comprehensive score of the historical scheduling trajectory under different real-time context state vectors are extracted from the historical scheduling trajectory dataset. With the goal of maximizing the correlation between the multidimensional performance comprehensive score and the predicted cumulative reward induced by the context-aware dynamic reward function adjusted by the weights, the parameters of the weight-adjusted neural network are trained to obtain the weight-adjusted neural network.

5. The method according to claim 3, characterized in that, The high-fidelity simulation environment based on the dynamic operation of a simulated hospital, with the context-aware dynamic reward function as the optimization objective and the inequality constraints derived from the medical safety red line rules as constraints, employs a constrained policy optimization algorithm to fine-tune the initial safety strategy, resulting in the neural network model of the resource scheduling strategy, including: The hard penalty mechanism and soft constraint proximity condition in the medical safety red line reward function layer are transformed into a set of inequality constraints; the expression of the inequality constraints is as follows: , ,in To constrain the allowable threshold, This is a real-time state vector of the resource. To schedule trajectory actions; Based on the Lagrange relaxation method, the constrained policy optimization problem is transformed into unconstrained optimization, and a Lagrange function is constructed; the expression of the Lagrange function is as follows: ,in, For strategy The expectation of context-aware dynamic rewards. For Lagrange multipliers; The update gradient is calculated based on the Lagrange function, and the parameters of the resource scheduling strategy neural network model and the Lagrange multipliers are updated along the direction of the update gradient; the update formula for the parameters of the resource scheduling strategy neural network model is as follows: The updated form of the Lagrange multiplier is: .

6. The method according to claim 1, characterized in that, The urgency value in the attribute set of the task to be scheduled is updated based on the dynamic priority adjustment instruction to obtain a priority change summary feature. This priority change summary feature is then concatenated with the multidimensional hospital state feature vector to obtain an enhanced state vector, including: Based on the urgency value of the dynamic priority adjustment instruction, the urgency value of the target task to be scheduled is updated to obtain a priority change event; the priority change event includes the initiating role; The priority change event is updated in the unified time-series event graph, and the multidimensional hospital status feature vector is updated based on the updated unified time-series event graph. The total number of all priority change events that occurred within a preset time window and the average priority change magnitude are counted, and it is checked whether there are any priority change events initiated by a preset initiating role, so as to obtain the priority change summary feature. The priority change summary feature is concatenated with the updated multidimensional hospital state feature vector to obtain the enhanced state vector.

7. The method according to claim 1, characterized in that, The process involves standardizing and cross-system correlation of the hospital's real-time heterogeneous event streams to generate a unified time-series event graph. A multi-dimensional hospital state feature vector for the current moment is then extracted from this unified time-series event graph, including: The real-time heterogeneous event stream is standardized according to a preset five-tuple format to obtain a standardized event stream; the five-tuple format includes subject, predicate, object, timestamp, and attribute. Based on event identifiers, events of the same medical process in the standardized event flow are linked as workflow instances, and the real-time status of all resource entities is maintained based on the timestamps to obtain the real-time status of resources; the real-time status of resources includes idle, scheduled, occupied, and unavailable; Take snapshots of the dynamic network consisting of all the workflow instances and the real-time status of the resources at preset time intervals to generate the unified time-series event graph; From the unified time-series event graph, the overall hospital load index is statistically calculated and the micro-attribute vectors of each task to be scheduled are extracted; the overall hospital load index includes the number of patients waiting for emergency treatment, ICU bed occupancy rate, and average queue length for large equipment; the micro-attribute vectors include task type, average estimated time, current urgency level, list of required resources, ready time, and latest allowed time. The overall hospital load index, the real-time resource status, and the micro-attribute vectors of each task to be scheduled are concatenated and encoded to obtain the multi-dimensional hospital status feature vector.

8. An intelligent management system for hospital medical resources, characterized in that, The system includes: The event resource module is used to standardize and perform cross-system correlation processing on the real-time heterogeneous event stream of the hospital, generate a unified time-series event graph, and extract a multi-dimensional hospital status feature vector at the current moment from the unified time-series event graph. The multi-dimensional hospital status feature vector includes the overall hospital load index, the real-time status of each resource, the attribute set of all tasks to be scheduled, and the resource conflict matrix. The real-time heterogeneous event stream is collected in real time through various information systems and physical sensing nodes of the hospital. The emergency demand module is used to respond to the acquisition of a dynamic priority adjustment instruction for a specific scheduling task, update the urgency value in the attribute set of the task to be scheduled based on the dynamic priority adjustment instruction, obtain a priority change summary feature, and concatenate the priority change summary feature with the multi-dimensional hospital status feature vector to obtain an enhanced status vector. The candidate scheduling module is used to input the enhanced state vector into a pre-trained resource scheduling strategy neural network model to obtain the scheduling action probability distribution of all the tasks to be scheduled, and generate at least one candidate resource scheduling scheme based on the scheduling action probability distribution. The red line verification module is used to perform deduction and verification of the candidate resource scheduling schemes based on the preset medical safety red line rules, and to mark all candidate resource scheduling schemes that pass the verification as valid candidate resource scheduling schemes. The recommended scheduling module is used to calculate the expected utility evaluation value of each effective candidate resource scheduling scheme based on the context-aware dynamic reward function, and to take the effective candidate resource scheduling scheme with the highest expected utility evaluation value as the recommended scheduling trajectory.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images