Burn patient treatment scheme management system and method based on artificial intelligence

Abstract

The invention relates to the technical field of medical management, in particular to a burn patient treatment scheme management system and method based on artificial intelligence. And learning historical medical resource demand data through the # imgabs0 # model, predicting a future resource demand state, and inputting a prediction result into the # imgabs1 # model so as to dynamically adjust a resource allocation strategy. The reward function of the # imgabs2 model comprehensively considers the resource utilization efficiency, the management cost, the patient waiting time, the patient satisfaction degree and the resource overload penalty, and the state transition probability is dynamically adjusted through the prediction result of the # imgabs3 model, so that more reasonable medical resource allocation is realized.

Description

Technical Field

[0001] The present invention relates to the field of medical management technology, and in particular to a treatment plan management system and method for burn patients based on artificial intelligence. Background Art

[0002] Burns are a complex trauma that poses a severe challenge to medical resources and management systems. Its complexity is not only reflected in the severity and diversity of patients' injuries, but also in the huge demand for medical resources and the dynamic adjustment of treatment plans. Traditional burn treatment plan management methods usually rely on static resource allocation and scheduling strategies, which often seem to be inadequate in the face of individual differences in patients and uncertainties in the treatment process. Each burn patient's injury is unique, which leads to significant differences in the demand for different medical resources (such as beds, medical staff, operating rooms, etc.). At the same time, the treatment process of burn patients is often accompanied by complications, changes in the condition and other emergencies, which require the treatment management system to respond in time and make corresponding adjustments. For example, when a large-scale burn incident occurs, how to reasonably allocate medical resources in a short period of time to maximize the treatment of patients is a very complex and challenging problem. Traditional resource management methods, such as allocating beds and medical staff in a fixed ratio, are difficult to adapt to such dynamic needs. Therefore, there is an urgent need for an intelligent management system that can dynamically adjust resource allocation strategies based on the patient's real-time status and predicted resource needs, thereby improving the management efficiency and quality of burn treatment.

[0003] At present, there are many mature treatment methods for burns in clinical practice, such as wound debridement, anti-infection treatment, nutritional support, and rehabilitation treatment. However, how to effectively manage these treatment methods, including reasonable resource allocation, optimization of treatment scheduling, and dynamic adjustment of treatment plans according to the real-time status of patients, remains a huge challenge. Traditional treatment plan management methods often adopt static resource allocation strategies, which are difficult to flexibly adjust according to individual differences of patients and emergencies during treatment once formulated. An obvious shortcoming of this method is that it cannot fully take into account the dynamic changes in medical resource demand. For example, in the early stage of burns, more medical staff and operating room resources may be required for emergency and debridement surgery, while in the rehabilitation stage, more rehabilitation therapists and beds may be required. In addition, during the treatment process, various unexpected situations may occur, such as sudden deterioration of the patient's condition, shortage of medical resources, or the emergence of complications, which require the management system to be able to identify and adjust them in a timely manner. Therefore, static treatment plan management methods are difficult to cope with the complexity and dynamic changes of burn treatment, and it is urgent to introduce more intelligent and adaptive strategies to better support the decision-making of medical staff.

[0004] Therefore, how to combine the individual differences of patients, the dynamic adjustment of treatment plans, and the reasonable allocation of medical resources has become the key to improving the efficiency and quality of burn treatment management. Traditional treatment plan management methods usually rely on experience-based strategies, that is, managers allocate resources based on past experience and judgment of the current situation. However, this method has certain limitations. First, the experience of managers is limited, and it is difficult to fully consider all possible influencing factors, such as changes in patients' conditions, the dynamic nature of resource requirements, and other unpredictable emergencies. Secondly, this method is also relatively inefficient and difficult to adapt to the needs of large-scale burn incidents, and it is impossible to respond quickly and adjust dynamically. In addition, traditional resource allocation methods are usually difficult to predict future resource needs, so advance planning and prevention are impossible, which often leads to resource waste and inefficiency. This urgently requires a more intelligent and flexible management tool that can dynamically adjust resource allocation and treatment plans based on the real-time status and historical data of patients, so as to provide patients with more personalized care. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a treatment plan management system and method for burn patients based on artificial intelligence. Figure 1 As shown in the figure, the present invention provides a treatment plan management method for burn patients based on artificial intelligence. Through intelligent resource allocation and scheduling, the execution of existing treatment plans is optimized, the utilization efficiency of medical resources is improved, and the management cost is reduced. This method constructs a deeply integrated resource management model. The model first uses , that is To learn historical medical resource demand data, it is used to predict the resource demand status at different time points in the future. Specifically, The model contains a and a . The specific structure is , to process time series data, whose input consists of a random noise vector and historical resource demand state sequence ,in represent The resource demand status at the moment, including information such as bed occupancy, medical staff scheduling, and operating room availability, and the output is the predicted future resource demand status . The specific structure is Its input is the real future resource demand state or the generated future resource demand state , the output is The probability that the input is the true state. In order to improve the accuracy of the prediction, the model introduces a prediction loss function. The training goal of the model is to minimize its overall loss function , the loss function is defined by the following formula:

[0006] ;

[0007] in, represent loss, represent loss, represents the prediction loss, and is the weight coefficient. Loss function, The loss function and prediction loss function are expressed as:

[0008] ;

[0009] ;

[0010] ;

[0011] in, Represents the real resource demand state distribution, represents the distribution of the noise vector, represent Consider the real state is the probability of true data, represent Consider the generated state is the probability of true data, For the The predicted future resource demand state of samples, For the The real future resource demand state of samples, represents the total number of samples, represents the two-norm.

[0012] Then, the model will The predicted future resource demand state is integrated into It is used to dynamically adjust resource allocation strategies. The state space of the model consists of medical resource states Definitions, including bed occupancy rates , medical staff scheduling , operating room availability , Drug Inventory and patient waiting time , expressed as The action space is managed by Definitions, including adjustments to medical staff schedules , Bed allocation adjustment , Operating room schedule adjustment , drug procurement adjustments and postponement or advancement of non-urgent surgeries , expressed as For guidance The decision model defines a comprehensive reward function , which is rewarded by resource utilization efficiency , Management cost penalty , Patient waiting time penalty , Patient Satisfaction Rewards and resource overload penalties Composition, expressed as Among them, the resource utilization efficiency reward The calculation method is: , management cost penalty The calculation method is: , patient waiting time penalty The calculation method is: , Patient Satisfaction Award The calculation method is: , resource overload penalty The calculation method is: ,in, , , , , , and They are all weight coefficients, which are used to adjust the impact of each indicator. Representative Action The cost, Waiting time for patients, Status Patient satisfaction, Status In order to better The prediction results are integrated into In the model, the present invention also utilizes Future resource demand status predicted by the model , dynamically adjust State transition probability , the specific calculation formula is:

[0013] ;

[0014] in, is the weight coefficient used to balance Predicted and true state transition probabilities, Represents the predicted future resource demand state The similarity with all possible states in the state space, Represents the true state transition probability.

[0015] In order to train the deep fusion model, the present invention adopts an iterative training method. First, the historical medical resource management data is used for training. Model, Optimization The parameters of the model make The model can accurately predict future resource demand status. During the model training process, a batch of historical resource demand data is sampled from the real data set in the form of , and then from the random noise distribution Sample a noise vector in , and then use Generate the future state based on the current historical state sequence and noise vector, and then use To distinguish the real state from the generated state, and calculate of loss, The loss and prediction loss are then updated based on these losses. Then, use the trained Model and actual data training Model, using The model selects the appropriate strategy based on resource status and actions. During model training, the current state is first sampled from the real data set And actions , and then use the trained The model is used to predict the future resource demand state, and then the reward function is used to calculate the reward value of the current state and action, and then the reward value based on the predicted future state is calculated. The transition probability is updated The transition probability in is finally used The policy iteration or value iteration algorithm updates strategy, so that The expected cumulative reward of is maximized. Through iterative training Model and Model, so that the two models work together, so that the model can better manage resources.

[0016] On the other hand, Figure 2 As shown, the present invention provides a treatment plan management system for burn patients based on artificial intelligence.

[0017] The system includes the following core modules:

[0018] 1. Based on Resource demand prediction module: This module uses The model learns historical medical resource demand data and predicts the resource demand status at different time points in the future. The model includes a and a . use ,For example or , to process time series data. The input is a random noise vector and historical resource demand state sequence ,in express The resource demand status at all times, including bed occupancy, medical staff scheduling, operating room availability, etc. The output is the predicted future resource demand state . use , whose input is the real or generated future resource demand state, and the output is The probability that the input is the true state. To improve the prediction accuracy, this module also calculates the prediction loss. This module optimizes by minimizing the following loss function Model:

[0019] ;

[0020] in, represent loss, represent loss, represents the prediction loss, and is the weight coefficient. Loss function, The loss function and prediction loss function are expressed as:

[0021] ;

[0022] ;

[0023] ;

[0024] in, Represents the real resource demand state distribution, represents the distribution of the noise vector, represent Consider the real state is the probability of true data, represent Consider the generated state is the probability of true data, For the The predicted future resource demand state of samples, For the The real future resource demand state of samples, represents the total number of samples, represents the two-norm.

[0025] 2. Based on Resource allocation decision module: This module uses , and The module predicts the future resource demand state as input to dynamically adjust the resource allocation strategy. The state space of this module is defined as the medical resource state , the specific form is , including bed occupancy rate , medical staff scheduling , operating room availability , Drug Inventory and patient waiting time The action space is defined as the management operations , the specific form is , including adjustments to medical staff shifts , Bed allocation adjustment , Operating room schedule adjustment , drug procurement adjustments and postponement or advancement of non-urgent surgeries In addition, this module defines a comprehensive reward function , in the form of , used to evaluate resource allocation strategies, this reward function comprehensively considers resource utilization efficiency rewards , Management cost penalty , Patient waiting time penalty , Patient Satisfaction Rewards and resource overload penalties Among them, the resource utilization efficiency reward The calculation method is , management cost penalty The calculation method is , patient waiting time penalty The calculation method is , Patient Satisfaction Award The calculation method is , resource overload penalty The calculation method is ,in, , , , , , and They are all weight coefficients, which are used to adjust the impact of each indicator. Representative Action The cost, Waiting time for patients, Status Patient satisfaction, Status The degree of overload of medical resources.

[0026] 3. State transfer dynamic adjustment module: This module uses The module predicts the future resource demand status , dynamically adjust The state transition probability of the model , the transition probability is calculated by the following formula:

[0027] ;

[0028] in, is the weight coefficient used to balance Predicted and true state transition probabilities; Represents the predicted future resource demand state Similarity to all possible states in the state space; Represents the true state transition probability.

[0029] 4. Model training module: This module is responsible for model training, including Model training and The module first trains the model using historical medical resource management data. Model, Optimization The parameters of the model make The model can accurately predict future resource demand status. During the model training process, the module samples a batch of historical resource demand data from the real data set in the form of , and then from the random noise distribution Sample a noise vector in , and then use Generate the future state based on the current historical state sequence and noise vector, and then use To distinguish the real state from the generated state, and calculate of loss, The loss and prediction loss are then updated based on these losses. The module then uses the trained Model and actual data training Model, using The model selects the appropriate strategy based on resource status and actions. During model training, the module first samples the current state from the real data set And actions , and then use the trained The model is used to predict the future resource demand state, and then the reward function is used to calculate the reward value of the current state and action, and then the reward value based on the predicted future state is calculated. The transition probability is updated The transition probability in is finally used The policy iteration or value iteration algorithm updates strategy, so that The expected cumulative reward of is maximized. This module is trained iteratively Model and model, so that the two models work together.

[0030] The above modules work together to achieve intelligent management of treatment plans for burn patients. The system dynamically adjusts resource allocation and scheduling based on predicted future resource needs, thereby improving resource utilization efficiency, shortening patient waiting time, and reducing management costs.

[0031] Through the above technical solution, the present invention realizes intelligent management of treatment plans for burn patients, optimizes the allocation and scheduling of medical resources, improves resource utilization efficiency, reduces management costs, shortens patients' waiting time, and improves patients' satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flowchart of an artificial intelligence-based method for managing treatment plans for burn patients.

[0033] Figure 2 This is a system structure diagram of a burn patient treatment plan management system based on artificial intelligence. DETAILED DESCRIPTION

[0034] The deep fusion proposed by the present invention will be described in detail below. and This paper proposes a treatment plan management method for burn patients. Different from the traditional treatment plan management method, this method aims to optimize the treatment process through intelligent resource allocation and scheduling, rather than directly formulating specific treatment plans.

[0035] In the present invention, the present invention will It is used to predict the resource requirements during the treatment of burn patients, rather than predicting the patient's condition. To learn historical medical resource demand data and generate predictions of future resource demand status. The resource demand state is recorded as , which can include the hospital bed occupancy rate, the medical staff's shift schedule, and the operating room availability. The goal of this invention is to The model can be based on the historical state sequence Generate future resource demand state .

[0036] To achieve this goal, of for ,For example or , to process time series data. Specifically, The input is a random noise vector and the historical state sequence , the output is the predicted future resource demand state . is , whose input is the real state or Generated , the output is The probability that the input state is the true state.

[0037] In order to The generated future resource demand state is not only realistic, but also can reflect the individual differences and treatment responses of patients. The present invention uses the following loss function:

[0038] ;

[0039] in, represent of loss, represent of losses, and is a weight coefficient used to balance Loss and loss. The loss function is defined as:

[0040] ;

[0041] in, represents the expectation of real resource demand status data, Represented by Expectations for the generated data. represents the distribution of real data, is the noise vector The distribution of represent Consider the real state is the probability of true data, represent Consider the generated state is the probability of true data.

[0042] The loss function is defined as:

[0043] ;

[0044] This loss function encourages Generate more deceptive of the sample.

[0045] In order to The prediction results are better than Framework combination, the present invention introduces a loss function , which is used to measure the difference between the predicted future resource demand state and the actual future resource demand state, and is defined as follows:

[0046] ;

[0047] in, Indicates The predicted future resource demand state of samples, Indicates The real future resource demand state of samples, represents the total number of samples, represents the two-norm. This loss function can promote Generate a more accurate picture of future resource requirements, thereby improving quality of decision making.

[0048] final, The overall loss function becomes:

[0049] ;

[0050] in, is a weight coefficient used to balance the generation adversarial loss and prediction loss. This loss function encourages Generate realistic and accurate future resource demand states to Provide more reliable basis.

[0051] exist In , we first define the state, action, and reward function.

[0052] state( ):state Defined as the hospital in time The status of medical resources, which can include the following aspects:

[0053] Bed occupancy rate ( ): The occupancy rate of hospital beds, which can be expressed as a percentage, for example 80% of the beds are occupied.

[0054] Medical staff scheduling ( ): The hospital's medical staff scheduling, including the number of doctors and nurses on duty, for example, 5 doctors and 10 nurses on duty.

[0055] Operating room availability ( ): Availability of operating rooms, for example, whether an operating room is available or the operating room scheduling situation.

[0056] Drug Inventory ( ): The hospital's drug inventory, including the inventory of various commonly used drugs, for example, the inventory of a certain antibiotic is 1,000 units.

[0057] Patient waiting time ( ): Patient waiting time, for example, how long a patient has to wait before being assigned a bed.

[0058] So, the state can be expressed as:

[0059] ;

[0060] action( ):action Defined as the time period in which the treatment plan manager The management operations that can be selected include the following aspects:

[0061] Adjustment of medical staff shifts ( ): Adjust the schedule of medical staff, for example, increase or decrease the number of doctors and nurses.

[0062] Bed allocation adjustment ( ): Adjust the allocation of beds, for example, increase or decrease the number of beds used for burn patients.

[0063] Adjustment of operating room schedule ( ): Adjust the operating room schedule, for example, increase or decrease the operating room usage time.

[0064] Drug procurement adjustments ( ): Adjust drug purchases, for example, increase or decrease the purchase volume of a certain drug.

[0065] Postponement or advancement of non-urgent surgery ( ): Adjust the scheduling of non-emergency surgeries, for example, whether to postpone or advance non-emergency surgeries.

[0066] So, the action can be expressed as:

[0067] ;

[0068] Reward function ( ): Reward function To evaluate the status Take action The reward function of the present invention is to maximize the resource utilization efficiency and treatment management quality, while taking into account the cost and risk factors. The reward function consists of the following parts:

[0069] Resource efficiency reward ( ): Used to measure whether medical resources are used effectively, for example, whether the bed utilization rate is reasonable and whether medical staff work efficiently.

[0070] Management cost penalty ( ): Used to measure the cost of treatment management, such as drug procurement costs and medical staff salaries.

[0071] Patient waiting time penalty ( ): Used to punish patients for waiting too long. The longer the waiting time, the greater the penalty.

[0072] Patient Satisfaction Award ( ): Patient satisfaction, the more satisfied, the greater the reward.

[0073] Resource overload penalty ( ): Used to punish situations where resources are overloaded, for example, there are not enough beds or medical staff.

[0074] The resource utilization efficiency bonus is calculated using the following formula:

[0075] ;

[0076] in, , and is a constant coefficient used to adjust the weights of different resource utilization efficiencies. Indicates the current bed occupancy rate, represents the target bed occupancy rate (usually set within a certain range), Indicates the available medical staff, Indicates the total number of medical staff, is the utilization rate of the operating room, e.g. This means the operating room is in use. This formula can prompt the model to select actions that help improve the efficiency of medical resource utilization.

[0077] To calculate the management cost penalty, the present invention uses the following formula:

[0078] ;

[0079] in, is a constant coefficient used to adjust the weight of the cost penalty, Representative Action costs, such as medical staff salaries, drug procurement costs, etc.

[0080] To calculate the patient waiting time penalty, the present invention uses the following formula:

[0081] ;

[0082] in, is a constant coefficient used to adjust the weight of the waiting time penalty, represents the waiting time of patients. This formula can encourage the model to select actions that help reduce the waiting time of patients.

[0083] To calculate the patient satisfaction reward, the present invention uses the following formula:

[0084] ;

[0085] in, is a constant coefficient used to adjust the weight of patient satisfaction, Represents in state This formula can prompt the model to select actions that can improve patient satisfaction.

[0086] To calculate the resource overload penalty, the present invention uses the following formula:

[0087] ;

[0088] in, is a constant coefficient used to adjust the weight of resource overload penalty. Indicates in status The overload penalty for medical resources is high when there are insufficient beds or medical staff. This formula can prompt the model to choose actions that help alleviate resource overload.

[0089] The final reward function is:

[0090] ;

[0091] Each part of these reward functions needs to be adjusted according to the actual situation so that the model can better learn the optimal treatment management strategy. It should be noted that here, the present invention has innovated the reward function. Compared with the traditional reward function, the present invention combines several factors such as resource utilization efficiency, management cost, patient waiting time, patient satisfaction and resource overload to achieve a more comprehensive optimization effect, thereby achieving more effective management of the treatment process of burn patients and ensuring efficient use of medical resources.

[0092] In the traditional In the state transition probability is a predefined probability distribution. However, in actual burn treatment management, due to the influence of factors such as changes in resource requirements and changes in patient conditions, state transitions are highly uncertain. In order to solve this problem, the present invention uses Predicted future resource demand status To help estimate Specifically, the present invention is based on the current state and actions ,use Predicting future states Then, the present invention is based on renew The state transition probability. The prediction is not completely accurate, so the present invention also needs to consider the uncertainty of the prediction and use a weight coefficient to adjust the use Predict the probability of the state.

[0093] definition It is the real state. is the true transition probability, and then the present invention defines The predicted transition probability is:

[0094] ;

[0095] in, is a weight coefficient used to balance The predicted and actual transition probabilities. Represents the predicted future resource demand state The similarity with all possible states in the state space. If With a state in the state space The more similar the state is, the higher the probability of transferring to that state. If the prediction is close to the actual situation, This method can make Able to make better use of To further enhance the stability of the model, the real probability When it is not known exactly, the present invention can be initialized using a uniform distribution of all possible states.

[0096] In order to train the deep fusion and The present invention uses an iterative training algorithm. The algorithm first uses historical medical resource management data to train Model, and then use the trained Model and actual data to train Model. The entire algorithm flow is as follows:

[0097] 1. Initialization:

[0098] initialization of and network parameters.

[0099] initialization The state transition probability , can be initialized to a uniform distribution.

[0100] Set the number of training iterations and other training parameters.

[0101] 2. Model training:

[0102] For each training iteration , perform the following steps:

[0103] Sample a batch of historical resource demand data from a real dataset .

[0104] From random noise distribution Sample a noise vector in .

[0105] According to the current historical state sequence and the noise vector ,use of Generate Future State .

[0106] use To distinguish the real state and the resulting state ,calculate Loss ,and Loss .

[0107] Calculating prediction loss .

[0108] renew of and The network parameters make The overall loss minimize.

[0109] 3. Model training:

[0110] For each training iteration , perform the following steps:

[0111] Sample the current state from a real dataset And actions

[0112] Use trained Model to predict future resource demand status .

[0113] Using the reward function Calculate the reward value for the current state and action.

[0114] Based on the predicted future state , calculated based on The transition probability , and update The transition probability in .

[0115] use The policy iteration or value iteration algorithm updates Strategy , so that The expected cumulative reward of maximizing.

[0116] 4. Circuit training: Repeat steps 2 and 3 until Model and The models converged.

[0117] Through this iterative training algorithm, the present invention can simultaneously train Model and model, enabling them to work better together, thereby achieving optimized management of the treatment process for burn patients.

[0118] Below, the present invention will introduce the specific implementation details of the model in detail, including and The model's network structure, training environment, and hyperparameter settings.

[0119] Model by and composition. The role is to use the random noise vector and the historical state sequence Generate future resource demand state . The role of is to distinguish the real future resource demand state and Generated future resource demand state .

[0120] :The present invention adopts Network as The basic structure of The network performs well in processing time series data. Specifically, The network receives a noise vector and the historical state sequence As input, output a predicted future state . Among them, the noise vector is a 100-dimensional random vector, the historical state sequence In each is a 10-dimensional vector, representing the dimensions of resource requirements. Able to receive inputs of different dimensions, the present invention first uses a fully connected layer to transform each Mapped into a 100-dimensional vector.

[0121] Input layer: receives a 100-dimensional noise vector and the historical state sequence For each historical state , first use a fully connected layer to map it to a 100-dimensional vector, and then connect all the historical state sequences together and input them into the LSTM network.

[0122] Layer: Use 2 layers Layers, each containing 256 hidden units, are used to learn the relationship between historical states and extract features.

[0123] Fully connected layer: The output of is mapped to a 10-dimensional vector as a prediction of the future resource status .

[0124] :The present invention adopts Network as The basic structure of The network performs well on image and state data. Specifically, The network receives a real future resource demand state or Generated future resource demand state As input, output a probability value, indicating the probability that the input is the true state. Able to process vector input, the present invention first uses a fully connected layer to map the 10-dimensional input state to 64 dimensions.

[0125] Input layer: Receives a 10-dimensional future resource demand state as input and maps it to 64 dimensions through a fully connected layer.

[0126] Convolutional layer: Use 3 convolutional layers, each layer uses 64 convolution kernels, and the convolution kernel size is 3x3. Use one after each convolutional layer. Activation function.

[0127] Fully connected layer: A fully connected layer is used to map the output of the convolutional layer to a scalar value representing the probability that the input is the true state.

[0128] The model mainly includes state space, action space, transition probability, reward function and strategy. In the present invention, the state space is composed of the utilization of medical resources, including bed occupancy rate, busyness of medical staff, operating room scheduling, etc. The action space is composed of resource management operations, such as adjusting the scheduling of medical staff, adjusting the allocation of beds, or adjusting the scheduling of operating rooms. The state transition probability is calculated by The reward function is composed of resource utilization efficiency, cost, and patient satisfaction, and the strategy is the optimal resource allocation plan calculated based on the reward function.

[0129] State Space: State Defined as the hospital in time The medical resource status is a 10-dimensional vector, which can be expressed as ,in represents the bed occupancy rate, It means that medical staff are scheduled to work. Indicates operating room availability, represents the drug inventory, Indicates the patient waiting time. Each dimension is represented by a numerical value. For example, the bed occupancy rate can be represented by a value between 0 and 1, representing the current occupancy level of the bed.

[0130] Action Space: Action Defined as the time period in which the treatment plan manager The management operation that can be selected is a 5-dimensional vector, which can be expressed as .in, Representing the medical staff to adjust their shifts, Represents the adjustment of bed allocation, Representing the operating room scheduling adjustment, Represents drug procurement adjustments, Indicates whether non-urgent surgery is postponed. Each dimension is represented by a numerical value. For example, the adjustment of medical staff shifts can be represented by a positive or negative value to represent the increase or decrease of personnel. For example, if Representatives added two medical staff.

[0131] Transition probability: state transition probability Depend on The prediction results and actual data are jointly determined. The specific calculation method has been described in detail in the previous content.

[0132] Reward function: Reward function It is composed of resource utilization efficiency, management cost, patient waiting time and patient satisfaction. The specific calculation method has been described in detail in the previous content.

[0133] Strategy: Strategy is a function that maps states to actions, that is, The present invention uses Algorithms are used to learn the optimal strategy. In the present invention, a Table, recording each action in each state value.

[0134] Q-Value calculation: Initialize the Q value table , for all and ,

[0135] Q-Value Update: Update using the following formula value: .in is the learning rate, is the discount factor, is in state All possible actions.

[0136] The model of the present invention is used Deep learning frameworks.

[0137] Hyperparameter settings:

[0138] :

[0139] Learning rate: 0.0002;

[0140] Learning rate: 0.0002;

[0141] Optimizer parameters: ;

[0142] Training batch size: 64;

[0143] Number of training iterations: =50000;

[0144] ( Loss weight): 1.0;

[0145] (weight of prediction loss): 1.0;

[0146] Number of hidden units: 256;

[0147] Convolution kernel size: 3x3;

[0148] Number of convolution kernels: 64;

[0149] :

[0150] Learning rate( ): ;

[0151] Discount factor ( ): ;

[0152] Exploration Rate : It starts at 1 and then gradually decays to 0.1 during training;

[0153] Number of training iterations: =50000;

[0154] Parameters in the reward function , , , , , , ;

[0155] (For combination The weights of the predicted results and the true probability): ;

[0156] Other parameters:

[0157] Time window length n = 5 (length of historical state)

[0158] These hyperparameters can be adjusted according to specific experimental results to achieve the best performance.

[0159] The following will introduce how to build a simulation environment to verify the deep fusion proposed in this invention. and The effectiveness of the treatment plan management method for burn patients is determined. Due to the ethical and practical difficulties of testing directly in a real hospital environment, the present invention chooses to build a simulated burn treatment center environment. By conducting experiments in this simulated environment, the present invention can evaluate the performance of the method proposed by the present invention under different resource constraints and can safely compare different management strategies. In addition, in order to more comprehensively evaluate the performance of the method proposed by the present invention, the present invention will compare the method of the present invention with multiple baseline methods.

[0160] The simulation environment of the present invention is designed to simulate the daily operation of a burn treatment center as realistically as possible. The environment includes the following key components:

[0161] Bed management: A certain number of beds are set up in the simulation environment to treat burn patients. Each bed has a status attribute, indicating whether the bed is occupied. The total number of beds is a configurable parameter that can be adjusted according to different experimental scenarios.

[0162] Medical staff management: Two types of medical staff are set up in the simulation environment: doctors and nurses. The number of each type of medical staff is a configurable parameter. In addition, each medical staff has a status attribute, indicating whether it is currently working. The scheduling of medical staff can be dynamically adjusted to meet the treatment needs of different patients.

[0163] Operating room management: There are multiple operating rooms in the simulation environment. Each operating room has a status attribute indicating whether it is currently available. The operating room schedule can also be dynamically adjusted to meet the needs of the patient's surgery. The operating room resources can be dynamically adjusted for surgeries of different urgency.

[0164] Patient management: In the simulation environment, burn patients will enter the treatment center at certain time intervals. Each patient has an admission time and a corresponding series of treatment processes, which will consume a certain amount of medical resources. The severity of the patient's condition and the required treatment process will affect the demand for different resources. At the same time, sudden complications will also affect the demand for resources.

[0165] Time management: The simulation environment uses discrete time steps, each time step represents a period of time, for example, 1 hour. In each time step, the system updates the status based on the current resource status and patient status, and The decision of the model is to take corresponding management actions.

[0166] Resource requirements: Each patient's treatment process will generate demand for medical resources, such as beds, medical staff, and operating rooms. These resource requirements are dynamic and depend on the patient's condition and treatment stage. The patient's treatment process can be simulated as a series of continuous resource demand events.

[0167] Complications: The simulation environment introduces the concept of complications, such as infection, dyspnea, etc. The occurrence of these complications will increase the demand for medical resources and will directly affect the patient's hospital stay and may affect the demand for other resources. The probability of complications can be adjusted according to different experimental scenarios.

[0168] For training Model, the present invention needs to generate a large amount of medical resource demand data. Since the real medical resource demand data is difficult to obtain, the present invention uses the following method to generate simulation data:

[0169] 1. Basic resource demand model: The present invention first establishes a basic resource demand model. The model simulates the patient's demand for different medical resources according to the severity of the patient's condition and the stage of treatment. Specifically, it is assumed that the treatment process of a patient includes the following stages:

[0170] Admission stage: During the admission stage, patients need preliminary assessment and wound treatment, which requires a certain number of beds and medical staff.

[0171] Surgical Phase: During the surgical phase, the patient undergoes surgical treatment, which requires an operating room and a surgical team.

[0172] Rehabilitation stage: During the rehabilitation stage, the patient needs rehabilitation treatment, which requires a bed and a rehabilitation therapist.

[0173] Each stage will have different demands for different medical resources.

[0174] For example: Representing patients In time The demand for beds, Representing patients In time The demand for medical personnel, Representing patients In time The resource demand for each patient at each time step can be expressed as The specific value of each resource is determined by the current time stage and the severity of the disease.

[0175] The present invention can generate a resource requirement sequence for each patient during the entire treatment process by setting the resource requirement intensity and duration of each stage.

[0176] 2. Random factors: In order to increase the diversity and authenticity of data, the present invention introduces random factors into the basic resource demand model:

[0177] Randomness of patient admission time: The patient's admission time is randomly generated within a certain time range, simulating the uncertainty of the patient's visit.

[0178] Randomness of illness severity: The severity of the patient's illness is randomly generated within a certain range to simulate the differences between different patients' conditions. The severity of the illness will affect the intensity of resource requirements at each stage.

[0179] Random nature of complications: Complications occur randomly and may increase the demand for medical resources and prolong hospital stay.

[0180] Randomness of treatment response: Even for the same severity of illness, different patients may respond differently to treatment, which will also affect resource requirements.

[0181] Specifically, for each patient , the present invention can generate a random hospital stay , a random severity of illness , and a randomized complication rate Then, based on these random variables, we can determine the patient's demand for resources at each stage. The specific value of each resource is determined by variables such as the current time stage, severity of the disease, complications, and treatment response.

[0182] 3. Resource demand data generation: By combining the basic resource demand model and random factors, the present invention simulates the treatment process of a large number of patients, thereby generating a large amount of medical resource demand data. This data includes the demand for different medical resources for each patient at each time step. The present invention organizes this data into a time series form as The present invention generates a series of resource demand states, for example, , each state will be represented by different dimensions, including bed occupancy rate ( ), medical staff scheduling ( ), operating room availability ( ), drug inventory ( ), and patient waiting time ( );

[0183] 4. Data preprocessing: Before inputting the data into the model, the present invention needs to preprocess the data. This includes normalizing the data to eliminate the differences between different resource units. At the same time, in order to facilitate model training, the present invention organizes the resource demand data into a fixed-length sequence.

[0184] Through the above method, the present invention can generate a large amount of simulated medical resource demand data, which will be used to Model training and evaluation.

[0185] The specific settings of the experiments are described in detail below, including model parameters, evaluation metrics, and baseline methods.

[0186] 1. Model parameters: The present invention uses the model structure and hyperparameter settings described above, including and The learning rate, number of training iterations, batch size, and The learning rate, discount factor, etc. of the model. These parameters can be adjusted during the experiment to obtain the best performance. In order to ensure that the model can converge, the present invention adopts an early stopping strategy during the actual training process. That is, if the evaluation index of the model does not improve within a specific number of iterations, the present invention will stop training.

[0187] 2. Evaluation indicators: In order to evaluate the performance of the method proposed in the present invention, the present invention uses the following evaluation indicators:

[0188] Average Bed Occupancy Rate: Calculates the average bed occupancy rate during the simulation time. This indicator can evaluate whether the hospital's bed resources are fully utilized. The higher the bed utilization rate, the more efficient the resource utilization.

[0189] ;

[0190] in, represents the total time step, Represents at time The bed occupancy rate, Represents the maximum number of beds.

[0191] Average Staff Utilization Rate: Calculates the average staff utilization rate during the simulation time. This indicator can evaluate whether the medical staff is used efficiently. The higher the medical staff utilization rate, the higher the resource utilization efficiency. The present invention calculates the average utilization rate of doctors and nurses respectively.

[0192] ;

[0193] in, Represents at time Available medical staff, Represents the total number of medical staff.

[0194] Average Operating Room Utilization Rate: Calculates the average operating room utilization rate during the simulation time. This indicator can be used to evaluate whether the operating room is used efficiently. The higher the operating room utilization rate, the more efficient the resource utilization.

[0195] ;

[0196] in, Indicates time The use of operating rooms.

[0197] Average Patient Waiting Time: Calculates the average time patients wait for beds during the simulation time. This indicator can evaluate whether the allocation of medical resources can meet the needs of patients in a timely manner. The shorter the patient waiting time, the more reasonable the resource allocation.

[0198] ;

[0199] in, represents the total number of patients, Representative The waiting time of each patient.

[0200] Total Management Cost: Calculates the total resource management cost during the simulation time. This indicator can evaluate whether the model considers economic benefits while managing resources. The lower the total management cost, the higher the resource management efficiency.

[0201] ;

[0202] in, Representative time Time Action cost.

[0203] Patient Satisfaction: Calculate the average patient satisfaction during the simulation time. This indicator can evaluate whether the model can take into account the patient's experience. The higher the patient satisfaction, the better the resource management effect. The present invention collects patient satisfaction through a questionnaire survey. For example, patients will choose a score from 1 to 5 based on their satisfaction.

[0204] ;

[0205] in, Representative Patient satisfaction rating.

[0206] 3. Baseline method: In order to compare the performance of the method proposed in the present invention, the present invention selects the following commonly used resource allocation strategies as baseline methods. These methods are used in actual hospital management and can therefore be used as objects of comparison for the present invention.

[0207] Static Allocation: This method uses a fixed resource allocation strategy, that is, it allocates beds and medical staff in a fixed ratio without dynamic adjustment. This method is simple and easy to implement, but it cannot adapt to the dynamic changes in patient flow and resource requirements. This method can be used as a baseline for all methods. Specifically, this method pre-sets the number of beds, medical staff, and operating rooms, and then receives patients based on these resources, but does not make dynamic adjustments.

[0208] First-Come-First-Served (FCFS): This method allocates beds and operating rooms according to the order in which patients are admitted to the hospital. Patients who are admitted first have priority in obtaining resources. This method is also relatively simple, but it may lead to a waste of resources. For example, beds are allocated first to patients with mild symptoms, while patients with severe symptoms are delayed in treatment due to lack of beds. This method only considers the factor of time, and does not consider other factors, such as the severity of the disease.

[0209] Priority-Based Allocation (PBA): This method allocates beds and operating rooms according to the severity of the patient's condition, and patients with serious conditions are given priority in resources. This method is more reasonable than the first-come, first-served method, but it also has certain limitations. For example, it only considers the severity of the disease, but not other factors, such as resource utilization. The present invention can use a priority list, which will put those seriously ill patients at the front. Specifically, in a simulated environment, the present invention will score the patient according to the severity of the disease. The more serious the disease, the higher the score, and then give priority to patients with high scores for treatment.

[0210] Rule-Based Allocation (RBA): This method uses a series of predefined rules to allocate resources. For example, resource allocation rules can be formulated based on the severity of the patient's condition and the remaining amount of resources. This method is more flexible than the static allocation method, but it is still difficult to adapt to complex and dynamically changing environments. The present invention can define a set of rules in advance, which may allocate resources based on factors such as the severity of the condition, resource utilization, or time. For example, the rule may be "When the bed occupancy rate is higher than 80%, beds are allocated first to critically ill patients", or the rule may be "When the operating room utilization rate is lower than 20%, non-emergency surgeries are prioritized", etc.

[0211] Random Allocation (RA): This method randomly allocates resources without considering the patient's condition or resource needs. This method is generally used as a baseline in research and is not usually used in real-world settings.

[0212] By comparing with these baseline methods, the present invention can more comprehensively evaluate the advantages of the dynamic management method proposed by the present invention.

[0213] First, generate a large amount of simulated resource demand data for Model training.

[0214] Then, use the trained Model and generated simulation data to train the deep fusion of the present invention and Model.

[0215] Next, we use the trained model to conduct simulation experiments to evaluate the resource management effect of the model and compare it with the above baseline methods.

[0216] Finally, analyze the experimental results and draw conclusions.

[0217] Through the above experimental settings, the present invention can evaluate the performance of the method proposed in the present invention under different resource constraints, and can be compared with a variety of baseline methods, so as to verify the effectiveness and superiority of the method proposed in the present invention.

[0218] In order to comprehensively evaluate the method proposed in the present invention, the present invention considers multiple experimental scenarios, including different resource constraints and different patient flows, to simulate treatment plan management in different situations. Specifically, the present invention simulates the following scenarios:

[0219] Sufficient resources scenario: simulates the situation of sufficient medical resources, such as sufficient beds, sufficient medical staff, and sufficient operating rooms. In this scenario, the main evaluation is whether the model can reasonably utilize resources.

[0220] Resource shortage scenario: simulates the situation of medical resource shortage, such as shortage of beds, shortage of medical staff, and insufficient operating rooms. In this scenario, the main evaluation model is whether it can maximize resource utilization and minimize patient waiting time under the condition of limited resources.

[0221] Peak patient flow scenario: simulates peak patient flow, such as when a large-scale burn incident occurs and patients are admitted to the hospital in large numbers. In this scenario, the main evaluation is whether the model can quickly adjust the resource allocation strategy to respond to emergencies.

[0222] For each scenario, the present invention will run 10 experiments and then calculate the average and standard deviation of the indicators to ensure the stability of the experimental results.

[0223] The present invention conducted experiments in three different scenarios: sufficient resources, scarce resources, and peak patient flow. In each scenario, the present invention used the fusion model proposed by the present invention and multiple baseline methods to perform simulations and calculate various evaluation indicators. In order to ensure the stability of the results, the present invention ran 10 independent experiments in each scenario and calculated the mean and standard deviation of each indicator. The following table shows the evaluation indicator results of the fusion model and baseline method in different scenarios.

[0224] Scenario with sufficient resources:

[0225] In the scenario of sufficient resources, the present invention simulates the situation of sufficient medical resources, such as sufficient beds, sufficient medical staff, and sufficient operating rooms. Table 1 shows the experimental results in this scenario.

[0226] Table 1:

[0227] ;

[0228] In the scenario of sufficient resources, the fusion model is significantly better than all baseline methods in all indicators. The fusion model can make more effective use of medical resources. For example, the utilization rate of beds, medical staff, and operating rooms is higher than all baseline methods. At the same time, the average waiting time of patients is also significantly reduced, indicating that the fusion model can provide treatment to patients more promptly, and the patient satisfaction is also higher than the baseline method. In addition, the total management cost is also significantly lower than the baseline method. It can be seen that the static allocation method does not make dynamic adjustments, so the resource utilization rate is not high, and the patient's waiting time is also long. The first-come-first-served method does not consider the patient's condition, which will lead to waste and unfairness of resources. For example, seriously ill patients may not get beds first, so the effect is worse. Priority allocation and rule allocation methods can alleviate the problem of first-come-first-served, but they are still not as effective as the fusion model. Random allocation has the worst effect, which shows that blindly allocating resources is unreasonable.

[0229] Resource-constrained scenarios:

[0230] In the resource shortage scenario, the present invention simulates the situation of medical resource shortage, such as shortage of beds, shortage of medical staff, and insufficient operating rooms. Table 2 shows the experimental results in this scenario.

[0231] Table 2:

[0232] ;

[0233] In the scenario of resource shortage, the fusion model still outperforms all baseline methods in various indicators. Although in this case, the overall resource utilization is higher than that in the case of sufficient resources, the fusion model can still make more effective use of limited medical resources. For example, the utilization of beds, medical staff, and operating rooms is still higher than all baseline methods. In addition, the fusion model can significantly reduce the average waiting time of patients and improve the average satisfaction of patients. At the same time, the total management cost of the fusion model is also significantly lower than that of the baseline method, which shows that the fusion model can allocate and manage resources more effectively under the condition of limited resources. The static allocation method is difficult to adjust according to the actual situation under the condition of resource shortage, so the overall performance is poor. The first-come-first-served method will cause the average waiting time of patients to be longer due to limited resources. Priority allocation and rule allocation can alleviate the situation of resource shortage, but the effect is still not as good as the fusion model. Random allocation has the worst effect, which shows that the randomness of resource allocation will lead to a large waste of medical resources.

[0234] Peak patient flow scenarios:

[0235] In the patient flow peak scenario, the present invention simulates the situation where patients are admitted to the hospital in large numbers when a large-scale burn incident occurs. Table 3 shows the experimental results in this scenario.

[0236] Table 3:

[0237] ;

[0238] In the scenario of peak patient flow, the fusion model still significantly outperforms all baseline methods in all indicators. Although the utilization rate of all resources is relatively high, the fusion model can still use medical resources more effectively, reduce patient waiting time, and improve patient satisfaction. The total management cost of the fusion model is also lower than all baseline methods, which shows that even in the face of a sudden increase in patient flow, the fusion model can still make better resource allocation strategies. The static allocation method performs poorly because it cannot cope with the sudden increase in patient flow. The first-come-first-served method results in the longest waiting time for patients. The priority and rule allocation methods cannot fully adapt to the sudden increase in patient flow, so they still cannot achieve the best results. Random allocation still performs the worst.

[0239] From the above experimental results, we can clearly see that the deep fusion proposed by the present invention and The treatment plan management method of the proposed method outperforms all baseline methods in different scenarios. From a statistical point of view, the advantages of our method over multiple baseline methods are highly statistically significant in most indicators. In particular, in indicators related to resource utilization, such as bed utilization, medical staff utilization, and operating room utilization, the differences between the fusion model and all baseline methods reach the highest statistical significance. In indicators such as patient waiting time and management cost, although the difference significance with the priority allocation method is relatively low, it still shows the obvious advantage of the fusion model. These statistical results further verify the reliability and effectiveness of the fusion model in practical applications, proving that its performance improvement is not accidental, but a reflection of the advantages of the model itself.

[0240] The results show that deep fusion and The method can significantly improve the management efficiency and quality of treatment plans for burn patients. Compared with a variety of traditional resource allocation strategies, including static allocation, first-come-first-served, priority allocation, rule allocation, and random allocation, the method proposed in the present invention can dynamically adjust resource allocation according to changes in patient flow and resource demand, thereby making more efficient use of medical resources. This method shows good performance under different resource constraints, indicating that the model has good robustness and adaptability.

[0241] When resources are sufficient, this method can avoid resource waste, improve resource utilization efficiency, and shorten patient waiting time. When resources are tight, this method can prioritize the most urgent patient needs, reduce patient waiting time, and improve patient satisfaction. In the case of peak patient flow, this method can quickly respond to the increase in patient flow, pre-allocate resources in advance, and avoid paralysis of the medical system. In contrast, other baseline methods cannot achieve optimal resource management effects due to their lack of dynamic adjustment capabilities or the lack of ability to predict the future.

[0242] The significance of these results is that the method can help hospital managers better plan and allocate medical resources, thereby improving the efficiency and quality of overall medical services. This can help hospitals better respond to various emergencies, such as large-scale burn incidents, or situations where resources are scarce. For example, when resources are tight, hospitals can use the model of the present invention to prioritize the allocation of limited beds and medical staff to the patients who need them most, thereby maximizing the use of limited resources. In the case of peak patient flow, hospitals can also use the model of the present invention to pre-allocate resources in advance to cope with sudden increases in patient flow.

[0243] In addition, the method can reduce the operating costs of the medical system. Through more reasonable resource allocation, the method can avoid resource waste, improve resource utilization efficiency, and thus reduce medical expenses. This is of great significance for reducing the medical burden on patients and improving the economic benefits of the medical system.

[0244] The results show that this method is a very promising tool for managing treatment plans for burn patients. It can not only improve the utilization of medical resources, but also reduce patient waiting time, improve patient satisfaction, and reduce the operating costs of the medical system. Compared with other baseline methods, the model proposed in this paper can show better performance in multiple scenarios considered, indicating that the model of this invention has high practical application value.

Claims

1. A method for managing treatment plans for burn patients based on artificial intelligence, characterized in that: The method comprises the following steps: use The model predicts the future state of medical resource demand; Will The future medical resource demand state predicted by the model is The input of the model is The model dynamically adjusts resource allocation strategies; according to Model prediction results, dynamic adjustment The state transition probability of the model and iterative training Model and Model.

2. The method according to claim 1, characterized in that: Said The reward function of the model Rewarded by resource efficiency , Management cost penalty , Patient waiting time penalty , Patient Satisfaction Rewards and resource overload penalties Composition, expressed as , where resource efficiency rewards Calculated by the following formula: ； in, , and is the weight coefficient, is the bed occupancy rate, is the target bed occupancy rate, is the number of available medical staff, is the total number of medical staff, The utilization rate of operating rooms.

3. The method according to claim 1, characterized in that The state transition probability of the model according to Future resource demand status predicted by the model Dynamic adjustment is performed, and the specific calculation formula is: ； in, is the weight coefficient used to balance Predicted and true state transition probabilities; Represents the predicted future resource demand state Similarity to all possible states in the state space; Represents the true state transition probability.

4. The method according to claim 1, characterized in that: The method is implemented by the following iterative training steps Model and Model training: First, use historical medical resource management data to train Model; Then use the trained Model and actual data training Model; The above training steps are iterated repeatedly until the model converges.

5. An artificial intelligence-based burn patient treatment plan management system, characterized in that: The system includes the following modules: based on The resource demand prediction module uses Model, learn historical medical resource demand data and predict resource demand status at different time points in the future; based on The resource allocation decision module uses , and The module uses the predicted future resource demand state as input to dynamically adjust the resource allocation strategy; State transfer dynamic adjustment module, which uses The module predicts the future resource demand status and adjusts it dynamically The state transition probability of the model; Model training module, which implements Model Iterative training between models.

Images