A medical treatment resource reservation dynamic optimization method, device, equipment and medium
By constructing a bimodal patient model and a multi-agent resource scheduling system, the problem of unreasonable resource utilization in traditional medical appointment methods has been solved. This has enabled accurate prediction of patient conditions and examination timeliness, as well as dynamic optimization of resources, thereby improving the efficiency of medical services.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 四川互慧软件有限公司
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-10
AI Technical Summary
Traditional medical appointment methods lack comprehensive consideration of patients' clinical urgency, differences in examination duration, and the real-time status of medical resources, resulting in excessively long waiting times for some patients, underutilization of resources, and impact on the efficiency of medical services.
By constructing a bimodal patient model with static clinical profile and dynamic state evolution, combining dynamic weighted clinical indicator assessment and dynamic quantification of clinical priority, the Transformer attention mechanism is used to predict examination duration, and an examination operation constraint matrix is constructed. The MADDPG multi-agent resource scheduling model under the CTDE framework is introduced to generate and dynamically update the appointment plan.
It enables precise characterization of the urgency of patients' conditions and the timeliness of examinations, improves the clinical rationality of appointment decisions, enhances resource utilization, reduces patient waiting time, and strengthens the system's flexibility and service efficiency.
Smart Images
Figure CN122000007B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical resource optimization strategies, and more particularly to a method, apparatus, equipment, and medium for dynamic optimization of medical resource appointment scheduling. Background Technology
[0002] With the continuous growth in demand for medical services, hospital resources for imaging examinations, laboratory tests, and functional examinations are becoming increasingly strained. Patients often face problems such as long waiting times, unreasonable examination arrangements, and low resource utilization during their visits.
[0003] Traditional medical appointment methods rely heavily on human experience or simple rule-based systems, typically allocating appointments on a first-come, first-served basis or within fixed time slots. This lack of consideration for the urgency of patients' clinical needs, the varying durations of examinations, and the real-time availability of medical resources can lead to excessively long waiting times for some patients and underutilization of certain medical equipment or examination resources, thus impacting overall medical service efficiency. Therefore, improving medical service efficiency and providing more efficient appointment methods is a pressing issue that needs to be addressed. Summary of the Invention
[0004] This invention provides a method, apparatus, equipment, and medium for dynamic optimization of medical resource reservations, which solves the technical problem of unreasonable reservation methods in the prior art and achieves the technical effect of improving the rationality of reservation methods.
[0005] In a first aspect, the present invention provides a method for dynamic optimization of medical resource appointment scheduling, comprising:
[0006] Acquire and preprocess multi-dimensional panoramic data, including data cleaning, standardization, and spatiotemporal feature joint encoding.
[0007] Based on preprocessed multi-dimensional panoramic data, a dual-modal patient model consisting of static clinical profile and dynamic state evolution is constructed. Combined with dynamic weighted clinical indicator evaluation and dynamic quantification of clinical priority, the patient's clinical characteristics are obtained.
[0008] A multi-dimensional coupled feature set is constructed based on the patient's clinical characteristics. The multi-dimensional coupled feature set is then input into an examination duration prediction model based on the Transformer attention mechanism, which has an online self-correction mechanism, to obtain the examination duration prediction result. An examination operation constraint matrix is also constructed to obtain the examination operation constraint rules.
[0009] Based on the inspection duration prediction results and inspection operation constraint rules, a MADDPG multi-agent resource scheduling model under the CTDE framework is constructed, and a joint optimization strategy for resource scheduling is obtained through joint training of multiple agents.
[0010] A reservation scheme is generated based on a joint optimization strategy for resource scheduling, and the reservation scheme is updated on a rolling basis based on a forward-looking rolling rescheduling mechanism.
[0011] Furthermore, based on the preprocessed multi-dimensional panoramic data, a bimodal patient model consisting of a static clinical profile and dynamic state evolution is constructed. Combined with dynamic weighted clinical indicator assessment and dynamic quantification of clinical priority, the patient's clinical characteristics are obtained, including:
[0012] Constructing a static clinical profiling model includes:
[0013]
[0014] in, For static feature vectors, For the patient's basic information feature vector, This provides a feature vector of the patient's past medical history, allergies, comorbidities, and medication history. The feature vector constrains the patient's clinical pathway and examination items. This is a feature vector representing the patient's historical appointment fulfillment rate and examination cooperation rate.
[0015] Constructing a dynamic state evolution model includes:
[0016] Constructing a time-series dynamic feature vector based on real-time physiological indicator data. ;
[0017] Based on temporal dynamic feature vectors Output a dynamic state evolution model, including:
[0018]
[0019]
[0020] in, for The hidden layer state of the LSTM network at time 1. This represents the hidden layer state from the previous time step. Here is the weight matrix of the LSTM network. for The dynamic state vector of the patient at any given time. It is the Sigmoid activation function. This is the weight matrix of the fully connected layer. For bias terms of fully connected layers;
[0021] According to the dynamic weighted clinical indicator evaluation system, which is related to the treatment stage and disease type;
[0022] Under the dynamic weighted clinical indicator evaluation system, clinical priority is dynamically quantified, including:
[0023]
[0024] in, for Clinical priority score at any given time. , and The weights provided for the dynamic weighted clinical indicator evaluation system The severity of the illness is scored. To check the timeliness score, Clinical pathway constraint score;
[0025] in,
[0026]
[0027] in, For the longest inspection window period, This represents the remaining inspectable time until the end of the current inspection window. This represents the disease timeliness coefficient.
[0028] Furthermore, a multi-dimensional coupled feature set is constructed based on the patient's clinical characteristics, and this feature set is input into an examination duration prediction model based on the Transformer attention mechanism, which has an online self-correction mechanism, to obtain the examination duration prediction result. Additionally, an examination operation constraint matrix is constructed to obtain the examination operation constraint rules, including:
[0029] Based on a multi-head self-attention mechanism, the association weights between different features in a multi-dimensional coupled feature set are constructed by determining the patient's clinical characteristics, including:
[0030]
[0031] in, For querying the matrix, The key matrix, For value matrices, For transpose, For feature dimensions;
[0032] The mean and standard deviation of the inspection duration prediction results are output through the fully connected layer, including:
[0033]
[0034]
[0035] in, For multi-dimensional coupled feature vectors, For Transformer encoders, Here is the encoder weight matrix. To examine the mean of the duration prediction results, To examine the standard deviation of the duration prediction results, The weights of the mean output layer, The bias of the mean output layer The weights of the variance output layer, This is the bias of the variance output layer;
[0036] Based on the mean and standard deviation, the confidence interval is determined as follows:
[0037]
[0038] Constructing the loss function includes:
[0039]
[0040] in, For loss function, For the sample size, For the first The actual inspection time for each sample For the first Predicted examination time for each sample The L2 regularization coefficient is... This is the model weight matrix;
[0041] Constructing the inspection operation constraint matrix includes: setting the minimum operation interval for different inspection combinations, resulting in... dimensional constraint matrix .
[0042] Furthermore, based on the inspection duration prediction results and inspection operation constraint rules, a MADDPG multi-agent resource scheduling model under the CTDE framework is constructed, and a joint optimization strategy for resource scheduling is obtained through joint training of multiple agents, including:
[0043] Building a patient intelligent agent cluster ,include:
[0044]
[0045] in, For the first A patient's intelligent agent;
[0046] Building a resource intelligent agent cluster ,include:
[0047]
[0048] in, For the first Individual resource intelligent agents;
[0049] The elements of Markov decision-making include: global state space, local observation space, action space, reward function, and state transition probabilities.
[0050] Design the reward function for building a cluster of patient intelligent agents, including:
[0051]
[0052] in, The reward value for the patient's intelligent agent. , as well as All are weights. For the first The expected waiting time for each sample For the first Maximum acceptable waiting time for a sample For the first The scheduled time slots for the allocation of each sample. For the first The ideal appointment time for each sample For the first The acceptable appointment time window for each sample. For the first Clinical priority score for each sample For the first Historical appointment compliance score for each sample;
[0053] Design the reward function for building a resource agent cluster, including:
[0054]
[0055] in, The reward value for the resource agent. For resource intelligent agents exist Utilization rate of time slots To achieve optimal resource utilization, For resource intelligent agents exist Cumulative load over a period of time For resource intelligent agents Average load, For the maximum rated load of the resource, To utilize resource-intelligent agents medical staff Real-time fatigue level during the period , as well as All are weights;
[0056] Global optimization based on reward function includes:
[0057]
[0058] in, For global weights, In order to find the best result, For the sample size, For the number of resource agents;
[0059] Perform policy gradient updates, including:
[0060]
[0061] in, No. The policy network parameters for each sample, To accumulate rewards, For experience replay pool, For the first Local observations of a sample This is the global state. A centralized action value function based on global information. For the first The policy network for each sample, For the first The action space of each sample For the global state space, As expected.
[0062] Furthermore, a reservation scheme is generated based on a joint resource scheduling optimization strategy, and the reservation scheme is updated on a rolling basis based on a forward-looking rolling rescheduling mechanism, including:
[0063] The resource scheduling joint optimization strategy and inspection operation constraint rules generate a sequence of diagnosis and treatment nodes, and generate an appointment plan by prioritizing them;
[0064] With the optimization objectives of minimizing the increase in expected waiting time, minimizing the number of appointment adjustments, and maximizing resource utilization, rolling optimization is performed, including:
[0065]
[0066] in, For the scrolling optimization function, , as well as All are weights. To account for the expected increase in waiting time, This represents the global fluctuation value of resource utilization. The number of reservations has been adjusted.
[0067] Furthermore, it also includes:
[0068] Construct a three-dimensional performance evaluation system, including:
[0069]
[0070] in, For the overall evaluation score, , as well as All are weights. Standardized scores for the clinical dimension, Standardize the scoring for the operational dimension. Standardize the scores for the patient dimension;
[0071] To maximize To achieve the objective, a Bayesian optimization algorithm is employed to adaptively optimize the global parameters, including:
[0072]
[0073] in, These are global policy network parameters.
[0074] Furthermore, preprocessing of multi-dimensional panoramic data includes:
[0075] The multi-dimensional panoramic data is classified, including continuous feature data, categorical feature data, and ordered feature data.
[0076] Missing value imputation and outlier removal are performed on multi-dimensional panoramic data; one-hot encoding is performed on categorical feature data; and label encoding is performed on ordered feature data.
[0077] Dimensionless processing of multi-dimensional panoramic data;
[0078] Perform spatiotemporal feature joint encoding, including:
[0079]
[0080] in, For time-coded vectors, For spatial encoding vectors, The estimated travel time for patients from their residence to the hospital. The straight-line distance from the patient's address to the hospital. It is a spatiotemporal joint encoding feature vector.
[0081] Secondly, the present invention provides a device for dynamically optimizing medical appointment scheduling, comprising:
[0082] The acquisition module is used to acquire and preprocess multi-dimensional panoramic data. The preprocessing includes data cleaning, standardization, and spatiotemporal feature joint encoding.
[0083] The patient clinical feature construction module is used to construct a bimodal patient model with static clinical profile and dynamic state evolution based on preprocessed multi-dimensional panoramic data, and to obtain patient clinical features by combining dynamic weight clinical indicator evaluation and dynamic quantification method of clinical priority.
[0084] The constraint module is used to construct a multi-dimensional coupled feature set based on the patient's clinical characteristics, and input the multi-dimensional coupled feature set into the examination duration prediction model based on the Transformer attention mechanism with an online self-correction mechanism to obtain the examination duration prediction result, and construct the examination operation constraint matrix to obtain the examination operation constraint rules;
[0085] The update module is used to construct a MADDPG multi-agent resource scheduling model under the CTDE framework based on the inspection duration prediction results and inspection operation constraint rules, and obtain a joint optimization strategy for resource scheduling through joint training of multiple agents.
[0086] The scheme generation module is used to generate reservation schemes based on the joint optimization strategy of resource scheduling, and to update the reservation schemes on a rolling basis based on the forward-looking rolling rescheduling mechanism.
[0087] Thirdly, the present invention provides an electronic device, comprising:
[0088] processor;
[0089] Memory used to store processor-executable instructions;
[0090] The processor is configured to execute a method for dynamically optimizing medical appointment scheduling, as provided in the first aspect.
[0091] Fourthly, the present invention provides a non-transitory computer-readable storage medium, which, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to perform a method for dynamic optimization of medical resource appointment as provided in the first aspect.
[0092] One or more technical solutions provided in this invention have at least the following technical effects or advantages:
[0093] This invention integrates multi-dimensional panoramic data to create a unified model of patient information, clinical pathways, and resource status, enabling intelligent scheduling and dynamic optimization of medical examination resources. Through data cleaning, standardization, and spatiotemporal feature co-coding, the usability and consistency of raw data are improved, providing a reliable data foundation for subsequent analysis. By constructing a bimodal patient model combining static clinical profiles and dynamic state evolution, and integrating dynamic weighted clinical indicator evaluation and clinical priority quantification methods, the invention achieves a precise characterization of the urgency of the patient's condition and the timeliness of examination needs, thereby improving the clinical rationality of scheduling decisions.
[0094] This invention utilizes a Transformer attention-based examination duration prediction model to accurately predict the actual time consumption for different patients and examination items, and ensures the standardization and executability of the examination process by constructing an examination operation constraint matrix. Furthermore, it introduces the MADDPG multi-agent resource scheduling model under the CTDE framework to achieve collaborative optimization between patient needs and medical resources, improving resource utilization and reducing patient waiting time. Finally, a prospective rolling rescheduling mechanism dynamically updates the appointment plan, enabling the system to adjust the appointment plan promptly in the event of sudden changes or resource fluctuations, thereby improving the overall scheduling flexibility, stability, and service efficiency. Attached Figure Description
[0095] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0096] Figure 1 This is a flowchart illustrating a method for dynamically optimizing medical resource reservations provided by the present invention. Detailed Implementation
[0097] This invention provides a method for dynamically optimizing medical resource reservations, which solves the technical problem of unreasonable reservation methods in the prior art.
[0098] The technical solution of this invention is to solve the above-mentioned technical problems, and the overall idea is as follows:
[0099] A method for dynamically optimizing medical resource appointments includes: acquiring and preprocessing multi-dimensional panoramic data, wherein preprocessing includes data cleaning, standardization, and spatiotemporal feature joint encoding; based on the preprocessed multi-dimensional panoramic data, constructing a bimodal patient model with static clinical profile and dynamic state evolution, and combining dynamic weighted clinical indicator evaluation and dynamic quantification of clinical priority to obtain patient clinical characteristics; constructing a multi-dimensional coupled feature set based on patient clinical characteristics, and inputting the multi-dimensional coupled feature set into an examination duration prediction model based on Transformer attention mechanism with online self-correction mechanism to obtain examination duration prediction results, and constructing an examination operation constraint matrix to obtain examination operation constraint rules; based on the examination duration prediction results and examination operation constraint rules, constructing a MADDPG multi-agent resource scheduling model under the CTDE framework, and obtaining a joint optimization strategy for resource scheduling through multi-agent joint training; generating an appointment plan based on the joint optimization strategy for resource scheduling, and updating the appointment plan on a rolling basis based on a look-ahead rolling rescheduling mechanism.
[0100] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.
[0101] First, it should be clarified that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.
[0102] In this invention, the sum of the weights of the same formula is 1, and each weight belongs to (0, 1).
[0103] This invention provides, for example Figure 1 The method for dynamically optimizing medical resource appointments, as shown, includes steps S11-S15:
[0104] Step S11: Acquire and preprocess multi-dimensional panoramic data, including data cleaning, standardization, and spatiotemporal feature joint encoding.
[0105] Preprocessing of multi-dimensional panoramic data includes: classifying the multi-dimensional panoramic data, including continuous feature data, categorical feature data, and ordered feature data; imputing missing values and removing outliers in the multi-dimensional panoramic data; performing one-hot encoding on categorical feature data and label encoding on ordered feature data; and performing dimensionless processing on the multi-dimensional panoramic data.
[0106] Perform spatiotemporal feature joint encoding, including:
[0107]
[0108] in, For time-coded vectors, For spatial encoding vectors, The estimated travel time for patients from their residence to the hospital. The straight-line distance from the patient's address to the hospital. It is a spatiotemporal joint encoding feature vector.
[0109] Specifically, multi-dimensional panoramic data can include data collection from clinical, resource, spatiotemporal, and environmental aspects.
[0110] This includes basic patient information, appointment time, appointment type (emergency / outpatient / inpatient / physical examination), symptom description, appointment items, and historical appointment fulfillment records;
[0111] This includes key time points such as admission assessment, examination, medication, surgery, and follow-up; compliance requirements for pathways; clinical connections between examination items; and constraints on examination window periods.
[0112] This includes real-time usage of the hospital's human resources, equipment resources, and consumable resources; full historical reservation data; and operational data for scenarios such as holidays, academic conferences, and infectious disease outbreaks.
[0113] This includes monitoring real-time equipment load, real-time location of medical staff, and status data of the examination room environment;
[0114] This includes data on the spatial location of patients' residences and the hospital area, real-time traffic conditions, public transportation timetables, congestion index around the hospital area, and data on weekdays, holidays, and morning and evening rush hours.
[0115] This includes the patient's past medical history, allergy history, comorbidities, medication history, contraindications for examinations, previous examination compliance, and historical time-series data of physiological indicators.
[0116] The original data can be cleaned by filling missing values (filling continuous features with the mean and categorical features with the mode), removing outliers (removing abnormal data that deviates from the normal range based on the 3σ principle), and deduplicating duplicate data.
[0117] One-hot coding is used for categorical features (such as appointment type, examination items, and equipment models), while label coding is used for ordinal features (such as symptom scores and physical assessments).
[0118] The Z-Score standardization method can be used to make all continuous features dimensionless. The core formula is as follows:
[0119]
[0120] In the formula, For the first Item continuous type original features, This is the mean of the feature across the entire historical dataset. This is the standard deviation of the feature across the entire historical dataset. These are the standardized feature values.
[0121] To address the shortcomings of the fragmented spatiotemporal features, a joint encoding vector of patient travel spatial features and appointment time features can be constructed, as detailed in the formula above.
[0122] Step S12: Based on the preprocessed multi-dimensional panoramic data, construct a dual-modal patient model with static clinical profile and dynamic state evolution, and combine dynamic weighted clinical indicator evaluation and dynamic quantification of clinical priority to obtain the patient's clinical characteristics.
[0123] Specifically, it includes:
[0124] Constructing a static clinical profiling model includes:
[0125]
[0126] in, For static feature vectors, For the patient's basic information feature vector, This provides a feature vector of the patient's past medical history, allergies, comorbidities, and medication history. The feature vector constrains the patient's clinical pathway and examination items. This is a feature vector representing the patient's historical appointment fulfillment rate and examination cooperation rate.
[0127] To analyze the temporal changes in patients' conditions and physiological indicators, a dynamic state evolution model of patients is constructed using a Long Short-Term Memory (LSTM) network to capture the real-time patterns of changes in patients' conditions.
[0128] Real-time physiological data (heart rate, blood pressure, body temperature, blood oxygen saturation, etc.) of patients are collected through medical devices, and real-time symptom scores, physical sign assessments, and disease severity grading data are collected through a clinical assessment system.
[0129] Constructing a dynamic state evolution model includes:
[0130] Constructing a time-series dynamic feature vector based on real-time physiological indicator data. ;
[0131] Based on temporal dynamic feature vectors Output a dynamic state evolution model, including:
[0132]
[0133]
[0134] in, for The hidden layer state of the LSTM network at time 1. This represents the hidden layer state from the previous time step. Here is the weight matrix of the LSTM network. for The dynamic state vector of the patient at any given time. It is the Sigmoid activation function. This is the weight matrix of the fully connected layer. For bias terms of fully connected layers;
[0135] According to the dynamic weighted clinical indicator evaluation system, which is related to the treatment stage and disease type;
[0136] For patients with different disease types and at different stages of treatment, a dynamic weighted clinical indicator evaluation system is constructed, which sets differentiated influence weights for different clinical indicators, and the weights are adaptively adjusted according to the patient's disease type and treatment stage.
[0137] For example, the weight of heart rate is set to 0.35 for patients with cardiovascular disease and 0.1 for patients with respiratory infection; the weight of body temperature is set to 0.4 for patients with infectious disease and 0.1 for patients with chronic disease follow-up. The above are just examples and can be adjusted according to the actual situation.
[0138] Based on three core dimensions—the severity of the patient's condition, the timeliness of the examination, and the constraints of the clinical pathway—a dynamic quantitative model for clinical priority is constructed, which outputs a priority score with a value range of (0-1). The higher the score, the higher the clinical priority of the patient.
[0139] Under the dynamic weighted clinical indicator evaluation system, clinical priority is dynamically quantified, including:
[0140]
[0141] in, for Clinical priority score at any given time. , and The weights provided for the dynamic weighted clinical indicator evaluation system The severity of the illness is scored. To check the timeliness score, Clinical pathway constraint score;
[0142] These are dynamic weighting coefficients for three dimensions, which adaptively adjust based on the patient's disease type, treatment stage, and appointment type (e.g., emergency room patients). Set to 0.7, for patients in the standard clinical pathway Set to 0.6, for elective examination of patients. Set to 0.5).
[0143] in,
[0144]
[0145] in, For the longest inspection window period, This represents the remaining inspectable time until the end of the current inspection window. This represents the disease timeliness coefficient.
[0146] in, (Severity of illness score) , (The score is calculated based on the time deviation between the current inspection node and the key nodes of the clinical pathway; the larger the deviation, the higher the score.) All scores were normalized before calculation.
[0147] Step S13: Construct a multi-dimensional coupled feature set based on the patient's clinical characteristics, and input the multi-dimensional coupled feature set into an examination duration prediction model based on the Transformer attention mechanism with an online self-correction mechanism to obtain the examination duration prediction result, and construct an examination operation constraint matrix to obtain the examination operation constraint rules.
[0148] By integrating four dimensions of features, we can comprehensively capture the core factors affecting examination time and construct a multi-dimensional coupled feature set, including: Patient-dimensional features: age, gender, cooperation level, previous examination time, comorbidities, contraindications, and body type; Operation-dimensional features: examination type, number of scan sequences, plain / contrast scan, whether anesthesia is required, operation complexity, and standard operation time specified by clinical guidelines; Equipment-dimensional features: equipment model, scanning speed, failure rate, current load, and average time of historical single examinations; Medical staff-dimensional features: technician's professional proficiency, historical time distribution of similar examinations, continuous working time, and fatigue index.
[0149] The self-attention mechanism of the Transformer encoder is adopted to automatically capture the influence weights of different features on the inspection time, solving the nonlinear fitting problem of multi-factor coupling. At the same time, the predicted time and 95% confidence interval are output, providing a redundant buffer for the allocation of appointment time slots, as follows:
[0150] The multi-dimensional coupled feature set is transformed into a fixed-dimensional feature vector through an embedding layer and then input into the Transformer encoder.
[0151] Based on a multi-head self-attention mechanism, the association weights between different features in a multi-dimensional coupled feature set are constructed by determining the patient's clinical characteristics, including:
[0152]
[0153] in, For querying the matrix, The key matrix, For value matrices, For transpose, For feature dimensions;
[0154] The mean and standard deviation of the inspection duration prediction results are output through the fully connected layer, including:
[0155]
[0156]
[0157] in, For multi-dimensional coupled feature vectors, For Transformer encoders, Here is the encoder weight matrix. To examine the mean of the duration prediction results, To examine the standard deviation of the duration prediction results, The weights of the mean output layer, The bias of the mean output layer The weights of the variance output layer, This is the bias of the variance output layer;
[0158] Based on the mean and standard deviation, the 95% confidence interval for the examination duration is determined as follows:
[0159]
[0160] The allocation of appointment slots is based on the upper limit to avoid subsequent appointment conflicts caused by check timeouts.
[0161] After each inspection is completed, the error between the actual inspection time and the predicted value is fed back to the model in real time. The model weights are updated online using the mini-batch stochastic gradient descent method to achieve continuous self-optimization of the model.
[0162] Constructing the loss function includes:
[0163]
[0164] in, For loss function, For the sample size, For the first The actual inspection time for each sample For the first Predicted examination duration for each sample The L2 regularization coefficient is... This is the model weight matrix;
[0165] Constructing the inspection operation constraint matrix includes: setting the minimum operation interval for different inspection combinations, resulting in... dimensional constraint matrix ;
[0166] Specifically:
[0167] Establish a hierarchy of operation types:
[0168] Based on clinical pathways and treatment guidelines, the connection between different types of examinations and the pre- and post-examination constraints should be clarified. For example, an MRI should be performed 24 hours after a CT contrast-enhanced examination, and pathological examination results should be used as a prerequisite for subsequent invasive examinations.
[0169] Construct a dynamic inspection time node and operation interval constraint matrix:
[0170] Based on clinical guidelines and historical data, standardized minimum operation intervals were set for different test combinations, forming... dimensional constraint matrix , where matrix elements Representative inspection After completion, proceed to inspection. The minimum time interval required to begin provides a hard constraint for optimizing the subsequent inspection order.
[0171] Step S14: Based on the inspection duration prediction results and inspection operation constraint rules, construct the MADDPG multi-agent resource scheduling model under the CTDE framework, and obtain the joint optimization strategy for resource scheduling through multi-agent joint training.
[0172] To address the shortcomings of static weights, single-objective optimization, and scheduling failure in extreme scenarios, a multi-agent deep deterministic policy gradient (MADDPG) scheduling model is constructed under the centralized training and distributed execution (CTDE) framework. It is divided into patient agent clusters and resource agent clusters to achieve bilateral Nash equilibrium optimization of patient clinical needs and hospital operational efficiency, which is completely different from the existing genetic algorithms and static reverse computation schemes.
[0173] Specifically, it includes:
[0174] Building a patient intelligent agent cluster ,include:
[0175]
[0176] in, For the first Each patient has an independent intelligent agent; the optimization goal is to minimize their own waiting time, maximize the timeliness of examinations, and maximize adherence to appointment times.
[0177] Building a resource intelligent agent cluster ,include:
[0178]
[0179] in, For the first Each resource agent is a separate resource agent; each group of "inspection equipment + operator technician" corresponds to an independent resource agent, with the optimization goal of maximizing its own resource utilization, minimizing load fluctuations, and reducing the fatigue of the corresponding medical staff.
[0180] The elements of Markov decision-making include: global state space, local observation space, action space, reward function, and state transition probabilities.
[0181] The global state space includes all patients' clinical priorities, predicted examination durations, spatiotemporal characteristics, and appointment constraints; the real-time status, time slot occupancy, and load data of all resources; and global information such as the hospital-wide resource supply-demand ratio and the risk level of emergencies.
[0182] The local observation space contains local information that each agent can directly observe. The patient agent observes its own clinical characteristics and appointment needs, while the resource agent observes its own equipment status and time slot occupancy.
[0183] The action space includes actions of the patient agent, such as selecting appointment slots and target resource units; and actions of the resource agent, such as accepting or rejecting appointment requests, adjusting slot allocation, and optimizing examination order.
[0184] The reward function includes individual and global two-level reward functions to achieve balanced optimization under bilateral constraints;
[0185] The state transition probability includes the temporal transition of the global state based on the joint actions of all agents;
[0186] Design the reward function for building a cluster of patient intelligent agents, including:
[0187]
[0188] in, The reward value for the patient's intelligent agent. , as well as All are weights. For the first The expected waiting time for each sample For the first Maximum acceptable waiting time for a sample For the first The scheduled time slots for the allocation of each sample. For the first The ideal appointment time for each sample For the first The acceptable appointment time window for each sample. For the first Clinical priority score for each sample For the first Historical appointment compliance score for each sample;
[0189] Design the reward function for building a resource agent cluster, including:
[0190]
[0191] in, The reward value for the resource agent. For resource intelligent agents exist Utilization rate of time slots To achieve optimal resource utilization, For resource intelligent agents exist Cumulative load over a period of time For resource intelligent agents Average load, For the maximum rated load of the resource, To utilize resource-intelligent agents medical staff Real-time fatigue level during the period , as well as All are weights;
[0192] Global optimization based on reward function includes:
[0193]
[0194] in, For global weights, In order to find the best result, For the sample size, For the number of resource agents;
[0195] The model aims to maximize the global cumulative reward. Through multi-agent joint training, it achieves Nash equilibrium between patient and resource agents. This means that no single agent can increase its reward value by changing its strategy alone, thus achieving optimal global resource allocation.
[0196] The MADDPG algorithm, employing the CTDE framework, optimizes the policy network of all agents using global information during the centralized training phase, while each agent makes decisions based solely on its local observations during the distributed execution phase. This approach balances global optimization accuracy with real-time scheduling efficiency, including:
[0197]
[0198] in, No. The policy network parameters for each sample, To accumulate rewards, For experience replay pool, For the first Local observations of a sample This is the global state. A centralized action value function based on global information. For the first The policy network for each sample, For the first The action space of each sample For the global state space, As expected.
[0199] Step S15: Generate a reservation plan based on the joint optimization strategy of resource scheduling, and update the reservation plan on a rolling basis based on the forward-looking rolling rescheduling mechanism.
[0200] This step generates personalized reservation plans based on the results of global resource scheduling optimization. It also pioneers a forward-looking rolling rescheduling mechanism to achieve real-time monitoring and dynamic adjustment of the entire reservation process, thus solving the shortcomings of existing technologies such as passive adjustment and delayed response.
[0201] Specifically, this includes: generating a sequence of diagnosis and treatment nodes by combining resource scheduling joint optimization strategies and inspection operation constraint rules, and generating an appointment plan by prioritizing them;
[0202] Generate initial diagnosis and treatment node sequence: Based on the scheduling optimization results and combined with the examination operation constraint matrix, generate the patient examination order and diagnosis and treatment node sequence that meet the requirements of the clinical pathway;
[0203] By combining patient clinical priority and resource operation priority, the diagnosis and treatment nodes are re-ranked to ensure that examination nodes for critically ill patients are allocated first.
[0204] Based on the upper limit of the 95% confidence interval of the inspection duration prediction, a precise appointment time slot is allocated to each inspection node, while a 10% buffer time slot is reserved to deal with unexpected timeout situations.
[0205] Based on a multi-level fuzzy evaluation method, combined with the patient's travel time and space characteristics, clinical constraints, and personal preferences, a personalized appointment strategy is generated. At the same time, appointment time, precautions, and travel suggestions are pushed to the patient's terminal.
[0206] To address the shortcomings of reactive adjustments after the fact, a forward-looking rolling rescheduling mechanism based on risk prediction is constructed, as follows:
[0207] The scheduling step is 15 minutes, with a 2-hour forward prediction window. A global rescheduling is performed every 15 minutes. At the same time, a hard stability constraint is set: appointments confirmed 24 hours in advance are not subject to mandatory adjustments, but only optional optimizations are made to ensure the patient's appointment experience.
[0208] By using an LSTM time-series prediction model, three types of scheduling risks are identified in advance. When the risk probability exceeds a preset threshold, rescheduling is automatically triggered.
[0209] Predict the probability of equipment failure, temporary reassignment of medical staff, and overtime occupancy of examination rooms within the next 2 hours;
[0210] Predict the increase in emergency room patients and the demand for additional testing within the next 2 hours;
[0211] Based on the patient's real-time travel trajectory, road conditions, and historical compliance, the probability of the patient missing an appointment / being late can be predicted.
[0212] With the optimization objectives of minimizing the increase in expected waiting time, minimizing the number of appointment adjustments, and maximizing resource utilization, rolling optimization is performed, including:
[0213]
[0214] in, For the scrolling optimization function, , as well as All are weights. To account for the expected increase in waiting time, This represents the global fluctuation value of resource utilization. The number of reservations has been adjusted.
[0215] Constraints: Clinical priority score Patients experiencing increased waiting times The waiting time for other patients increased. .
[0216] In addition, it can also implement full-process management of appointments;
[0217] We send appointment reminders to patients three times a day, 24 hours, 2 hours, and 30 minutes before the appointment, through multiple channels such as SMS, APP push, and official account messages. The reminders include the appointment time, clinic location, pre-examination preparation requirements, and travel suggestions.
[0218] The hospital information system monitors the appointment execution status in real time, including patient check-in, examination progress, whether the time limit has been exceeded, and equipment operation status.
[0219] In response to emergencies such as examinations exceeding the time limit, patients arriving late, or temporary equipment malfunctions, real-time local dynamic adjustments are triggered, including adjustments to the examination order, reassignment of examination equipment / doctors, and secondary optimization of appointment slots;
[0220] After the examination is completed, a follow-up reminder is automatically generated and issued. At the same time, the patient's appointment fulfillment status is recorded and updated to the patient compliance database, providing data support for subsequent appointment optimization.
[0221] It can construct a three-dimensional effect evaluation system of "clinical-operational-patient", and realize fully automated adaptive optimization of model parameters, weight coefficients and scheduling rules based on Bayesian optimization algorithm, forming a closed-loop optimization cycle of the whole link, which solves the defects of existing technology optimization that rely on manual and lack adaptive ability.
[0222] Continuously collect full real-time data on the execution of the appointment plan, including appointment completion status, actual examination duration, patient waiting time, resource utilization rate, equipment load balancing, patient satisfaction, reasons for delays / breach of contract, and handling results of emergencies.
[0223] Construct a three-dimensional performance evaluation system, including:
[0224]
[0225] in, For the overall evaluation score, , as well as All are weights. Standardized scores for the clinical dimension, Standardize the scoring for the operational dimension. Standardize the scores for the patient dimension;
[0226] To maximize To achieve the objective, a Bayesian optimization algorithm is employed to adaptively optimize the global parameters, including:
[0227]
[0228] in, These are global policy network parameters.
[0229] The optimal parameter combination obtained from parameter optimization is fed back in real time to the aforementioned resource scheduling optimization, duration prediction model, and clinical priority quantification model, enabling real-time updates of the model and scheduling strategy. At the same time, the full execution data is fed back to data preprocessing to update the feature engineering and model training datasets, forming a closed-loop self-optimization cycle of data acquisition, modeling, scheduling, execution, evaluation, and optimization, continuously improving the adaptability and optimization effect of the method.
[0230] In summary, this invention integrates multi-dimensional panoramic data to create a unified model of patient information, clinical pathways, and resource status, enabling intelligent scheduling and dynamic optimization of medical examination resources. Through data cleaning, standardization, and spatiotemporal feature co-coding, the usability and consistency of raw data are improved, providing a reliable data foundation for subsequent analysis. By constructing a bimodal patient model combining static clinical profiles and dynamic state evolution, and integrating dynamic weighted clinical indicator evaluation and clinical priority quantification methods, the urgency of patients' conditions and their time-sensitive examination needs are accurately characterized, thereby improving the clinical rationality of scheduling decisions.
[0231] This invention utilizes a Transformer attention-based examination duration prediction model to accurately predict the actual time consumption for different patients and examination items, and ensures the standardization and executability of the examination process by constructing an examination operation constraint matrix. Furthermore, it introduces the MADDPG multi-agent resource scheduling model under the CTDE framework to achieve collaborative optimization between patient needs and medical resources, improving resource utilization and reducing patient waiting time. Finally, a prospective rolling rescheduling mechanism dynamically updates the appointment plan, enabling the system to adjust the appointment plan promptly in the event of sudden changes or resource fluctuations, thereby improving the overall scheduling flexibility, stability, and service efficiency.
[0232] Based on the same inventive concept, this invention provides a device for dynamically optimizing medical appointment scheduling, comprising:
[0233] The acquisition module is used to acquire and preprocess multi-dimensional panoramic data. The preprocessing includes data cleaning, standardization, and spatiotemporal feature joint encoding.
[0234] The patient clinical feature construction module is used to construct a bimodal patient model with static clinical profile and dynamic state evolution based on preprocessed multi-dimensional panoramic data, and to obtain patient clinical features by combining dynamic weight clinical indicator evaluation and dynamic quantification method of clinical priority.
[0235] The constraint module is used to construct a multi-dimensional coupled feature set based on the patient's clinical characteristics, and input the multi-dimensional coupled feature set into the examination duration prediction model based on the Transformer attention mechanism with an online self-correction mechanism to obtain the examination duration prediction result, and construct the examination operation constraint matrix to obtain the examination operation constraint rules;
[0236] The update module is used to construct a MADDPG multi-agent resource scheduling model under the CTDE framework based on the inspection duration prediction results and inspection operation constraint rules, and obtain a joint optimization strategy for resource scheduling through joint training of multiple agents.
[0237] The scheme generation module is used to generate reservation schemes based on the joint optimization strategy of resource scheduling, and to update the reservation schemes on a rolling basis based on the forward-looking rolling rescheduling mechanism.
[0238] Based on the same inventive concept, the present invention also provides an electronic device, comprising:
[0239] processor;
[0240] Memory used to store processor-executable instructions;
[0241] The processor is configured to execute a method for dynamically optimizing medical resource appointments as described above.
[0242] Based on the same inventive concept, the present invention also provides a non-transitory computer-readable storage medium, which, when the instructions in the storage medium are executed by the processor of an electronic device, enables the electronic device to execute a method for dynamic optimization of medical treatment resource appointment as described above.
[0243] Since the electronic device described in this embodiment is an electronic device used to implement the information processing method in the embodiments of the present invention, those skilled in the art can understand the specific implementation methods and various variations of the electronic device in this embodiment based on the information processing method described in the embodiments of the present invention. Therefore, how the electronic device implements the method in the embodiments of the present invention will not be described in detail here. Any electronic device used by those skilled in the art to implement the information processing method in the embodiments of the present invention falls within the scope of protection of the present invention.
[0244] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0245] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0246] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0247] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0248] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0249] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for dynamically optimizing medical visit resource reservations, characterized in that, include: Acquire and preprocess multi-dimensional panoramic data, including data cleaning, standardization, and joint encoding of spatiotemporal features; Based on preprocessed multi-dimensional panoramic data, a bimodal patient model consisting of a static clinical profile and dynamic state evolution is constructed. This model is then combined with dynamic weighted clinical indicator assessment and dynamic quantification of clinical priority to obtain the patient's clinical characteristics, including: Constructing a static clinical profiling model includes: in, For static feature vectors, For the patient's basic information feature vector, This provides a feature vector of the patient's past medical history, allergies, comorbidities, and medication history. The feature vector constrains the patient's clinical pathway and examination items. This is a feature vector representing the patient's historical appointment fulfillment rate and examination cooperation rate. Constructing a dynamic state evolution model includes: Constructing a time-series dynamic feature vector based on real-time physiological indicator data. ; Based on temporal dynamic feature vectors Output a dynamic state evolution model, including: in, for The hidden layer state of the LSTM network at time 1. This represents the hidden layer state from the previous time step. Here is the weight matrix of the LSTM network. for The dynamic state vector of the patient at any given time. It is the Sigmoid activation function. This is the weight matrix of the fully connected layer. For bias terms of fully connected layers; According to the dynamic weighted clinical indicator evaluation system, which is related to the treatment stage and disease type; Under the dynamic weighted clinical indicator evaluation system, clinical priority is dynamically quantified, including: in, for Clinical priority score at any given time. , and The weights provided for the dynamic weighted clinical indicator evaluation system The severity of the illness is scored. To check the timeliness score, Clinical pathway constraint score; in, in, For the longest inspection window period, This represents the remaining inspectable time until the end of the current inspection window. This refers to the disease's timeliness coefficient. A multi-dimensional coupled feature set is constructed based on the patient's clinical characteristics. The multi-dimensional coupled feature set is then input into an examination duration prediction model based on the Transformer attention mechanism, which has an online self-correction mechanism, to obtain the examination duration prediction result. An examination operation constraint matrix is also constructed to obtain the examination operation constraint rules. Based on the inspection duration prediction results and inspection operation constraint rules, a MADDPG multi-agent resource scheduling model under the CTDE framework is constructed, and a joint optimization strategy for resource scheduling is obtained through joint training of multiple agents. A reservation scheme is generated based on the aforementioned joint resource scheduling optimization strategy, and the reservation scheme is updated on a rolling basis using a forward-looking rolling rescheduling mechanism; including: The resource scheduling joint optimization strategy and inspection operation constraint rules generate a sequence of diagnosis and treatment nodes, and generate an appointment plan by prioritizing them; With the optimization objectives of minimizing the increase in expected waiting time, minimizing the number of appointment adjustments, and maximizing resource utilization, rolling optimization is performed, including: in, For the scrolling optimization function, , as well as All are weights. To account for the expected increase in waiting time, This represents the global fluctuation value of resource utilization. The number of reservations has been adjusted.
2. The method for dynamic optimization of medical resource appointment as described in claim 1, characterized in that, A multi-dimensional coupled feature set is constructed based on patient clinical characteristics. This feature set is then input into a Transformer attention-based examination duration prediction model with an online self-correction mechanism to obtain examination duration prediction results. Additionally, an examination operation constraint matrix is constructed to obtain examination operation constraint rules, including: Based on a multi-head self-attention mechanism, the association weights between different features in a multi-dimensional coupled feature set are constructed by determining the patient's clinical characteristics, including: in, For querying the matrix, The key matrix, For value matrices, For transpose, For feature dimensions; The mean and standard deviation of the inspection duration prediction results are output through the fully connected layer, including: in, For multi-dimensional coupled feature vectors, For Transformer encoders, Here is the encoder weight matrix. To examine the mean of the duration prediction results, To examine the standard deviation of the duration prediction results, The weights of the mean output layer, The bias of the mean output layer The weights of the variance output layer, This is the bias of the variance output layer; Based on the mean and standard deviation, the confidence interval is determined as follows: Constructing the loss function includes: in, For loss function, For the sample size, For the first The actual inspection time for each sample For the first Predicted examination time for each sample The L2 regularization coefficient is... This is the model weight matrix; Constructing the inspection operation constraint matrix includes: setting the minimum operation interval for different inspection combinations, resulting in... dimensional constraint matrix .
3. The method for dynamic optimization of medical resource appointment as described in claim 1, characterized in that, Based on the inspection duration prediction results and inspection operation constraints, a MADDPG multi-agent resource scheduling model under the CTDE framework is constructed. A joint optimization strategy for resource scheduling is obtained through joint training of multiple agents, including: Building a patient intelligent agent cluster ,include: in, For the first A patient's intelligent agent; Building a resource intelligent agent cluster ,include: in, For the first Individual resource intelligent agents; The elements of Markov decision-making include: global state space, local observation space, action space, reward function, and state transition probabilities. Design the reward function for building a cluster of patient intelligent agents, including: in, The reward value for the patient's intelligent agent. , as well as All are weights. For the first The expected waiting time for each sample For the first Maximum acceptable waiting time for a sample For the first The scheduled time slots for the allocation of each sample. For the first The ideal appointment time for each sample For the first The acceptable appointment time window for each sample. For the first Clinical priority score for each sample For the first Historical appointment compliance score for each sample; Design the reward function for building a resource agent cluster, including: in, The reward value for the resource agent. For resource intelligent agents exist Utilization rate of time slots To achieve optimal resource utilization, For resource intelligent agents exist Cumulative load over a period of time For resource intelligent agents Average load, For the maximum rated load of the resource, To utilize resource-intelligent agents medical staff Real-time fatigue level during the period , as well as All are weights; Global optimization based on reward function includes: in, For global weights, In order to find the best result, For the sample size, For the number of resource agents; Perform policy gradient updates, including: in, No. The policy network parameters for each sample, To accumulate rewards, For experience replay pool, For the first Local observations of a sample This is the global state. A centralized action value function based on global information. For the first The policy network for each sample, For the first The action space of each sample For the global state space, As expected.
4. The method for dynamic optimization of medical resource appointment as described in claim 1, characterized in that, Also includes: Construct a three-dimensional performance evaluation system, including: in, For the overall evaluation score, , as well as All are weights. Standardized scores for the clinical dimension, Standardize the scoring for the operational dimension. Standardize the scores for the patient dimension; To maximize To achieve the objective, a Bayesian optimization algorithm is employed to adaptively optimize the global parameters, including: in, These are global policy network parameters.
5. The method for dynamic optimization of medical resource reservation as described in claim 1, characterized in that, Preprocessing multi-dimensional panoramic data, including: The multi-dimensional panoramic data is classified, and the data types include continuous feature data, categorized feature data, and ordered feature data; Missing values are filled and outliers are removed from the multi-dimensional panoramic data; one-hot encoding is performed on the categorical feature data, and label encoding is performed on the ordered feature data; The multi-dimensional panoramic data is processed to be dimensionless; Perform spatiotemporal feature joint encoding, including: in, For time-coded vectors, For spatial encoding vectors, The estimated travel time for patients from their residence to the hospital. The straight-line distance from the patient's address to the hospital. It is a spatiotemporal joint encoding feature vector.
6. A device for dynamically optimizing medical visit resource reservations, characterized in that, include: The acquisition module is used to acquire and preprocess multi-dimensional panoramic data. The preprocessing includes data cleaning, standardization, and spatiotemporal feature joint encoding. The patient clinical characteristic construction module is used to build a bimodal patient model based on preprocessed multi-dimensional panoramic data, consisting of a static clinical profile and dynamic state evolution. It combines dynamic weighted clinical indicator evaluation and dynamic quantification of clinical priority to obtain patient clinical characteristics, including: Constructing a static clinical profiling model includes: in, For static feature vectors, For the patient's basic information feature vector, This provides a feature vector of the patient's past medical history, allergies, comorbidities, and medication history. The feature vector constrains the patient's clinical pathway and examination items. This is a feature vector representing the patient's historical appointment fulfillment rate and examination cooperation rate. Constructing a dynamic state evolution model includes: Constructing a time-series dynamic feature vector based on real-time physiological indicator data. ; Based on temporal dynamic feature vectors Output a dynamic state evolution model, including: in, for The hidden layer state of the LSTM network at time 1. This represents the hidden layer state from the previous time step. Here is the weight matrix of the LSTM network. for The dynamic state vector of the patient at any given time. It is the Sigmoid activation function. This is the weight matrix of the fully connected layer. For bias terms of fully connected layers; According to the dynamic weighted clinical indicator evaluation system, which is related to the treatment stage and disease type; Under the dynamic weighted clinical indicator evaluation system, clinical priority is dynamically quantified, including: in, for Clinical priority score at any given time. , and The weights provided for the dynamic weighted clinical indicator evaluation system The severity of the illness is scored. To check the timeliness score, Clinical pathway constraint score; in, in, For the longest inspection window period, This represents the remaining inspectable time until the end of the current inspection window. This refers to the disease's timeliness coefficient. The constraint module is used to construct a multi-dimensional coupled feature set based on the patient's clinical characteristics, and input the multi-dimensional coupled feature set into the examination duration prediction model based on the Transformer attention mechanism with an online self-correction mechanism to obtain the examination duration prediction result, and construct the examination operation constraint matrix to obtain the examination operation constraint rules; The update module is used to construct a MADDPG multi-agent resource scheduling model under the CTDE framework based on the inspection duration prediction results and inspection operation constraint rules, and obtain a joint optimization strategy for resource scheduling through joint training of multiple agents. The scheme generation module is used to generate reservation schemes based on the resource scheduling joint optimization strategy, and to continuously update the reservation schemes based on a forward-looking rolling rescheduling mechanism; including: The resource scheduling joint optimization strategy and inspection operation constraint rules generate a sequence of diagnosis and treatment nodes, and generate an appointment plan by prioritizing them; With the optimization objectives of minimizing the increase in expected waiting time, minimizing the number of appointment adjustments, and maximizing resource utilization, rolling optimization is performed, including: in, For the scrolling optimization function, , as well as All are weights. To account for the expected increase in waiting time, This represents the global fluctuation value of resource utilization. The number of reservations has been adjusted.
7. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute a method for dynamically optimizing medical appointment scheduling as described in any one of claims 1 to 5.
8. A non-transitory computer-readable storage medium, characterized in that, When the instructions in the storage medium are executed by the processor of the electronic device, the electronic device is able to perform a method for dynamic optimization of medical visit resource reservation as described in any one of claims 1 to 5.