Field injury assessment scheduling system based on deep learning model

Abstract

The invention discloses a field injury assessment scheduling system based on a deep learning model, and relates to the technical field of intelligent medical scheduling, and the system comprises a collection and calculation module which collects multi-source data, combines fuzzy logic and slope analysis, calculates a comprehensive score, extracts a time sequence and image features, carries out the fusion through a multi-head attention mechanism, and carries out the calculation of the comprehensive score. The injury condition probability is adjusted in combination with the comprehensive score to obtain light, medium and heavy injury evaluation results; the optimization module is used for optimizing an ambulance path by using an ant colony algorithm in combination with an evaluation result, constructing an Actor-Critic network to optimize ant colony algorithm parameters, performing global optimization by using a genetic algorithm, and feeding back to the ant colony algorithm for iteration to obtain an optimal solution; according to the invention, accurate assessment of the injury condition of the wounded person is realized, the scheduling efficiency of the ambulance is improved, the wounded person can be treated and cured in time in the shortest time, high combination of the injury condition assessment result and path optimization scheduling is realized, and the response speed and accuracy of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent medical dispatch, and in particular to a on-site injury assessment and dispatch system based on a deep learning model. Background Art

[0002] With the development of modern technology, the application of deep learning technology has made remarkable progress in various fields. Especially in the field of medical and health, traditional on-site injury assessment methods often rely on manual judgment and experience, which are not only inefficient but also easily affected by subjective factors, resulting in insufficient accuracy and timeliness of casualty assessment. In recent years, with the rapid development of deep learning and artificial intelligence technologies, injury assessment systems based on deep learning have gradually replaced traditional methods. They can accurately analyze multi-source information such as vital sign data and injury image data, automatically conduct casualty assessment, and quickly obtain accurate injury assessment results. Such technologies not only improve the efficiency of casualty assessment but also significantly enhance the accuracy of assessment, providing strong support for emergency medical rescue.

[0003] However, most of the existing on-site injury assessment and dispatch systems are limited to simple rule-based assessments and lack an effective comprehensive analysis mechanism, making it difficult to provide accurate injury assessment results in complex environments. In addition, when existing on-site injury assessment and dispatch systems perform path planning, they usually only focus on a single indicator and fail to fully consider road condition changes and real-time dynamic data, resulting in low efficiency and accuracy in ambulance path selection, thereby affecting the rescue efficiency. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a on-site injury assessment and dispatch system based on a deep learning model, which solves the problems that most of the existing on-site injury assessment and dispatch systems are limited to simple rule-based assessments and lack an effective comprehensive analysis mechanism, making it difficult to provide accurate injury assessment results in complex environments. In addition, when existing on-site injury assessment and dispatch systems perform path planning, they usually only focus on a single indicator and fail to fully consider road condition changes and real-time dynamic data, resulting in low efficiency and accuracy in ambulance path selection, thereby affecting the rescue efficiency.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a on-site injury assessment and dispatch system based on a deep learning model, which includes an acquisition and calculation module that collects multi-source data, combines fuzzy logic and slope analysis, calculates a comprehensive score, extracts temporal and image features, and after fusing through a multi-head attention mechanism, adjusts the injury probability in combination with the comprehensive score to obtain light, moderate, and severe injury assessment results;

[0008] The multi-source data includes vital sign data, injury image data, and location information data;

[0009] An optimization module, in combination with the evaluation results, uses the ant colony algorithm to optimize the ambulance path, constructs an Actor-Critic network to optimize the parameters of the ant colony algorithm, and performs global optimization using the genetic algorithm, and then feeds back to the ant colony algorithm for iteration to obtain the optimal solution; a conversion and transmission module converts the optimal solution and transmits and issues it in real time through the MQTT protocol;

[0010] A storage and display module stores the data through a database and displays it through a visualization interface.

[0011] As a preferred solution of the on-site injury assessment and scheduling system based on the deep learning model of the present invention, wherein: collecting multi-source data and combining fuzzy logic and slope analysis, calculating the comprehensive score includes the following steps:

[0012] Collect the vital sign data, injury image data, and location information data of the wounded at consecutive time points through a physiological monitoring device, a portable X-ray machine, and a GPS device respectively;

[0013] The vital sign data includes the actual measured values of heart rate, respiratory rate, blood oxygen saturation, body temperature, and blood pressure;

[0014] Use empirical rules to set corresponding thresholds for each data in the vital sign data;

[0015] The corresponding thresholds include the lowest physiological reasonable value, the highest physiological reasonable value, the lower limit of the safe range, and the upper limit of the safe range corresponding to each data;

[0016] Use fuzzy logic rules to calculate the data reliability score for each data in the vital sign data;

[0017] According to each data in the vital sign data, calculate the change rate of each data and the corresponding data at the adjacent time point respectively, define it as the data slope of the corresponding data, then select the maximum data slope from the slopes of the corresponding data using the maximization operation, and calculate the data slope reliability score;

[0018] Use a logical operation formula to combine the reliability score and the slope reliability score to calculate the final data reliability score, and use a weighted average formula to combine the final data reliability scores of each data at each time point to calculate the comprehensive score of all data at each time point, and obtain the comprehensive score of the current wounded.

[0019] As a preferred solution of the on-site injury assessment and dispatch system based on the deep learning model of the present invention, the following steps are included: extracting temporal and image features, fusing them through the multi-head attention mechanism, and then adjusting the injury probability in combination with the comprehensive score to obtain the assessment results of minor, moderate, and severe injuries:

[0020] After constructing a feature extraction model using the deep convolutional neural network architecture, input the vital sign data and injury image data into the feature extraction model respectively. After obtaining the vital features, use the ReLU activation function to perform a non-linear transformation on the injury image data to obtain the image features;

[0021] Use the multi-head attention mechanism to perform weighted fusion on the vital features and image features to obtain the fused features;

[0022] Use the softmax function to transform the fused features to obtain the final injury assessment probability;

[0023] The final injury assessment probability includes the minor injury assessment probability, the moderate injury assessment probability, and the severe injury assessment probability;

[0024] Use element-wise multiplication to multiply the comprehensive score element-wise with the probabilities in the final injury assessment probability to obtain the adjusted probability. After sorting them in ascending order, take the largest probability as the assessment result of the current casualty, including minor injury, moderate injury, and severe injury.

[0025] As a preferred solution of the on-site injury assessment and dispatch system based on the deep learning model of the present invention, the following steps are included: combining the assessment results and using the ant colony algorithm to optimize the ambulance path:

[0026] Collect the positions of ambulances and hospitals, use the ambulance position as the starting point, and use the casualty position and hospital position as nodes;

[0027] Use the map API to obtain the actual driving distance of the complete path, and obtain the average speed by averaging the historical speeds of the ambulances in the map API through the HTTP request technology;

[0028] The complete path is from the ambulance position to the casualty position and then from the casualty position to the hospital position;

[0029] Define the ratio of the actual driving distance to the average speed as the response time;

[0030] The response time is divided into the response time from the ambulance to the casualty position and the response time from the casualty position to the hospital position;

[0031] Using historical regression analysis to set the road congestion coefficient, and after dynamically adjusting the response time from the ambulance location to the casualty location, further using the dynamic adjustment formula to calculate the response time from the casualty location to the hospital location, obtaining the adjusted response time from the casualty location to the hospital location, and calculating the total response time;

[0032] Using the rule of thumb to set constants, and using the proportional formula to combine the total response time and the constants to calculate the initial pheromone, and further using the proportional formula to combine the total response time to calculate the initial heuristic information;

[0033] Using the grid search method to set the weight coefficient, and using the path selection probability formula to calculate the path selection probability and generate a set of path selection probabilities;

[0034] When the path selection probability of the seriously injured exists in the set of path selection probabilities, directly use the path selection probability of the seriously injured as the path and perform dynamic adjustment. When there are l seriously injured in the set of path selection probabilities, sort the adjusted path selection probabilities in ascending order and select the largest path selection probability as the path. When there are no seriously injured in the set of path selection probabilities, use roulette wheel selection to select the path;

[0035] Normalize the path selection probability and the total response time of the selected path, and use the expert scoring method to set the attenuation coefficient and the weight coefficient, define the objective function, and minimize the response time and maximize the rescue efficiency;

[0036] Use theoretical analysis to set the evaporation coefficient, calculate the objective function value, perform iterative updates, use the evaporation coefficient to control the evaporation rate of the pheromone, and output the optimal solution;

[0037] The optimal solution refers to the optimal path with the minimized response time and the maximized rescue efficiency.

[0038] As a preferred solution of the on-site injury assessment and dispatching system based on the deep learning model of the present invention, wherein: constructing an Actor-Critic network to optimize the ant colony algorithm parameters includes the following steps:

[0039] Use the normalized state space to scale the response time, rescue efficiency, weight coefficients α and β, and evaporation coefficient in the optimal solution to obtain a scaled data set;

[0040] Use the multi-layer perceptron architecture to construct the Actor network and the Critic network in the A2C algorithm;

[0041] Input the scaled dataset into the Critic network to output the value of the dataset. Combine the value of the dataset using the n-step return method. After calculating the cumulative return, define the difference between the cumulative return and the dataset value as the advantage function. Set the feedback threshold using the Delphi method, compare the advantage function with the feedback threshold to obtain the updated Actor network, and input the scaled dataset into the updated Actor network to output a three-dimensional action vector;

[0042] The three-dimensional action vector includes weight coefficients α and β and a volatility coefficient;

[0043] Perform a smoothing operation on the three-dimensional action vector using double exponential smoothing filtering, write the smoothed three-dimensional action vector into the Redis key-value pair, and read it in real time through the ACO scheduling module to trigger the recalculation of the path selection probability;

[0044] Use the monitoring module to detect the performance metrics of the ACO scheduling module. When the comprehensive objective function value of the recalculated path selection probability is greater than the comprehensive objective function value of the optimal solution, roll back the historical optimal parameters and use the parameters in the optimal solution after rollback as the new parameters. Otherwise, use the recalculated parameters as the final parameters.

[0045] As a preferred solution of the on-site injury assessment scheduling system based on the deep learning model of the present invention, the following steps are included: After performing global optimization using the genetic algorithm, feedback it to the ant colony algorithm for iteration to obtain the optimal solution:

[0046] Use the hybrid coding technique to perform discrete coding on the optimal path in the optimal solution, and perform continuous coding on the weight coefficients α and β, the volatility coefficient, and the three-dimensional action vector in the optimal solution and then combine them as individuals to generate a population for initialization; Calculate the objective function value using the objective function of minimizing the response time and maximizing the rescue efficiency, and perform iterative optimization to output the optimal path;

[0047] Use the graph theory algorithm in the open source library to repair the optimal path, update the pheromone using the pheromone update rule, decode the updated pheromone using the decoding library, and re-enter the decoded pheromone into the ACO scheduling module for iteration to obtain a new optimal solution.

[0048] As a preferred solution of the on-site injury assessment scheduling system based on the deep learning model of the present invention, the following steps are included: Collect vital sign data, injury image data, and location information data, and first perform preprocessing operations including KNN interpolation, outlier detection, standardization, and normalization.

[0049] As a preferred solution of the on-site injury assessment and dispatching system based on the deep learning model of the present invention, specifically: converting the optimal solution and transmitting and real-time issuing it through the MQTT protocol means converting the optimal solution into the JSON format, using the MQTT protocol to transmit the JSON-format optimal solution to the edge computing device, and then issuing the JSON-format optimal solution to the dispatching personnel through the edge computing device.

[0050] As a preferred solution of the on-site injury assessment and dispatching system based on the deep learning model of the present invention, specifically: storing data through the database means storing the JSON-format optimal solution issued by the edge computing device through the database, marking a unique ID and timestamp for each optimal solution, and sorting the JSON-format optimal solution marked with the timestamp in ascending order through the database.

[0051] As a preferred solution of the on-site injury assessment and dispatching system based on the deep learning model of the present invention, specifically: displaying through the visualization interface means extracting the JSON-format optimal solution in the database and using the Plotly.js tool to display the dispatching priority and response time.

[0052] The beneficial effects of the present invention are as follows: The present invention evaluates the reliability of vital sign data by using fuzzy logic rules, captures the data change trend by using the slope analysis method, thereby realizing the accurate assessment of the injuries of the wounded, and uses the multi-head attention mechanism for feature fusion, which not only improves the accuracy of the assessment, but also can dynamically adjust the assessment results. Moreover, by introducing the ant colony algorithm to optimize the ambulance path and combining the historical regression analysis method to dynamically adjust the road congestion coefficient, the ambulance dispatching efficiency is further improved, ensuring that the wounded can be treated in time within the shortest time. Secondly, by combining the advantages of the ant colony algorithm, the A2C algorithm and the genetic algorithm, a high-degree combination of the injury assessment result and the path optimization dispatching is realized, so that the present invention not only considers the injury assessment result in the path selection process, but also can feedback and adjust the optimization parameters in real time, thereby improving the response speed and accuracy of the present invention. Description of the Drawings

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0054] Figure 1 It is the structure diagram of the on-site injury assessment and dispatching system based on the deep learning model in Embodiment 1.

[0055] Figure 2 It is the flowchart of the injury assessment in Embodiment 1.

[0056] Figure 3 It is a flow chart for path optimization of feature extraction in Embodiment 1.

[0057] Figure 4 It is a structural diagram of the optimization module in Embodiment 1. Specific implementation manners

[0058] To make the above objects, features and advantages of the present invention more obvious and understandable, the specific implementation manners of the present invention will be described in detail below with reference to the accompanying drawings of the specification.

[0059] In the following description, many specific details are set forth to facilitate a thorough understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.

[0060] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that may be included in at least one implementation manner of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it an individual or alternative embodiment that is mutually exclusive with other embodiments.

[0061] Embodiment 1, referring to Figures 1 to 4 , is the first embodiment of the present invention. This embodiment provides a on-site injury assessment and dispatch system based on a deep learning model, including the following steps:

[0062] S1. Acquisition and calculation module: Collect multi-source data, combine fuzzy logic and slope analysis, calculate a comprehensive score, extract temporal and image features, and after fusion through a multi-head attention mechanism, adjust the injury probability in combination with the comprehensive score to obtain light, medium, and severe injury assessment results;

[0063] The multi-source data includes vital sign data, injury image data, and location information data;

[0064] Specifically, collecting multi-source data and combining fuzzy logic and slope analysis to calculate the comprehensive score includes the following steps:

[0065] Collect vital sign data, injury image data, and location information data of the wounded at consecutive time points through a physiological monitoring device, a portable X-ray machine, and a GPS device respectively;

[0066] The vital sign data includes the actual measured values of heart rate, respiratory rate, blood oxygen saturation, body temperature, and blood pressure;

[0067] Use a rule of thumb to set corresponding thresholds for each piece of data in the vital sign data, obtaining z, c, v, b, m, where z represents the heart rate threshold, c represents the respiratory rate threshold, v represents the blood oxygen saturation threshold, v represents the body temperature threshold, and m represents the blood pressure threshold;

[0068] The corresponding thresholds include the lowest physiological reasonable value, the highest physiological reasonable value, the lower limit of the safe range, and the upper limit of the safe range for each piece of data;

[0069] Use fuzzy logic rules to calculate the data reliability score for each piece of data in the vital sign data. The formula is:

[0070]

[0071] In the formula, r a,i represents the data reliability score of the a-th vital sign at time point i, z a,1 represents the lowest physiological reasonable value of the a-th vital sign, z a,2 represents the lower limit of the safe range of the a-th vital sign, z a,3 represents the upper limit of the safe range of the a-th vital sign, z a,4 represents the highest physiological reasonable value of the a-th vital sign, s a,i represents the actual measured value of the a-th vital sign at time point i;

[0072] According to each piece of data in the vital sign data, calculate the change rate of each piece of data with the corresponding data at the adjacent time point, which is defined as the data slope of the corresponding data. The formula is:

[0073]

[0074] In the formula, g a,i represents the data slope of the a-th vital sign at time point i, s a,i-1 represents the actual measured value of the a-th casualty at time point i - 1, t represents the time interval, s a,i represents the actual measured value of the a-th vital sign at time point i; Select the maximum data slope from the slopes of the corresponding data using the maximization operation, and calculate the data slope reliability score. The formula is:

[0075]

[0076] In the formula, ψ a,i represents the reliability score of the data slope of the a-th vital sign at time point i, g a,max represents the maximum data slope of the a-th vital sign, |·| represents the absolute value;

[0077] Use a logical operation formula to combine the reliability score and the slope reliability score to calculate the final reliability score of the data. The formula is:

[0078] φ a,i = min(ψ a,i , r a,i )

[0079] In the formula, φ a,i represents the final reliability score of the data of the a-th vital sign at time point i, and min(·) represents taking the minimum value;

[0080] Use a weighted average formula to combine the final reliability scores of each data at each time point to calculate the comprehensive score of all data at each time point, and obtain the comprehensive score of the current casualty.

[0081] The vital sign data provides the basic physiological condition of the casualty. The injury image can assist in evaluating the location and severity of the injury. The location information helps to determine the location of the casualty and the timeliness of receiving medical services. By comprehensively using multi-source data, the health condition of the casualty can be evaluated more accurately. In an emergency environment, the casualty may be in a state of incomplete information. Multi-source data fusion can make up for the deficiencies of a single data source and improve the accuracy of the evaluation. Moreover, fuzzy logic rules are used to judge whether the data is within a reasonable range and generate the reliability score of each data. This enables fuzzy logic to effectively provide a meaningful evaluation even when there are certain errors or uncertainties in the data. Fuzzy logic can effectively handle complex and uncertain data and avoid misjudgment caused by data noise. Through fuzzy rules, the rationality of each data point can be evaluated, thereby improving the robustness and accuracy of the present invention. Secondly, by calculating the change rate of the vital sign data through slope analysis, the dynamic trend of the casualty's injury change can be captured. Especially in the change speed of physiological data such as heart rate and respiratory rate, slope analysis can help identify sudden changes and timely reflect the physiological changes of the casualty. Slope analysis can provide important information about the trend of the casualty's injury change, helping first aid personnel identify the rapid changes in the casualty's condition, thus providing an important basis for decision-making. By performing a weighted average on the reliability score and the slope reliability score, the comprehensive score of the casualty is finally obtained. This score combines multiple factors and can quantitatively evaluate the casualty's injury in multiple dimensions. Moreover, the comprehensive score integrates multiple physiological data and injury factors, making the injury assessment not limited to a single data point, but comprehensively analyzing the casualty's condition from multiple angles. This can provide a more comprehensive and accurate decision-making support for medical personnel, helping to improve the reaction speed and accuracy in the first aid process.

[0082] Further, the time series and image features are extracted, fused through the multi-head attention mechanism, and combined with the comprehensive score to adjust the injury probability, and the assessment results of minor, moderate, and severe injuries include the following steps:

[0083] After constructing a feature extraction model using a deep convolutional neural network architecture, the vital sign data is input into the feature extraction model to obtain vital features;

[0084] After inputting the injury image data into the feature extraction model, the ReLU activation function is used to perform a non-linear transformation on the injury image data to obtain image features;

[0085] The multi-head attention mechanism is used to perform weighted fusion on the vital features and image features to obtain fused features;

[0086] The softmax function is used to transform the fused features to obtain the final injury assessment probability;

[0087] The final injury assessment probability includes the minor injury assessment probability, the moderate injury assessment probability, and the severe injury assessment probability;

[0088] Using element-wise multiplication, the comprehensive score is multiplied element-wise with the probabilities in the final injury assessment probability to obtain the adjusted probability, and the adjusted probability is sorted in ascending order;

[0089] The largest probability in the sorted result is used as the assessment result of the current wounded, including minor injury, moderate injury, and severe injury.

[0090] By inputting the vital signs and image features into the CNN model respectively, high-level semantic information can be extracted, and the image features are non-linearly activated to suppress the adverse effects below zero and enhance the propagation of positive information, enabling the neural network to show stronger convergence ability and generalization ability when dealing with complex image features. Moreover, through the multi-head attention mechanism, weighted fusion of the vital signs and image features can be achieved, information cross-enhancement can be realized, and the understanding and discrimination accuracy of complex injury patterns can be improved. Using the softmax function can normalize the model output into a probability distribution, which is convenient for intuitively evaluating the possibilities of minor injury, moderate injury, and severe injury, providing a standardized output result for the present invention and facilitating subsequent dispatching. Secondly, introducing the comprehensive score as an external weight to adjust the assessment probability is a soft guidance mechanism. This method can not only alleviate the over-reliance of the model on a single data source but also enhance the individual adaptability of the assessment and improve the processing ability of the present invention for on-site complex situations. By sorting to determine the injury level with the highest probability, the final assessment decision output is formed to ensure that the wounded can obtain the first-aid resource response most in line with the actual situation in multi-level triage.

[0091] S2. Optimization module: Based on the evaluation results, the ant colony algorithm is used to optimize the ambulance path. An Actor-Critic network is constructed to optimize the parameters of the ant colony algorithm, and the genetic algorithm is used for global optimization and then fed back to the ant colony algorithm for iteration to obtain the optimal solution;

[0092] Specifically, using the ant colony algorithm to optimize the ambulance path based on the evaluation results includes the following steps:

[0093] Collect the positions of ambulances and hospitals. Take the ambulance position as the starting point, and take the casualty position and hospital position as nodes;

[0094] Use the map API to obtain the actual driving distance of the complete path, and obtain the average speed by averaging the historical speeds of the ambulances in the map API through HTTP request technology;

[0095] The complete path is from the ambulance position to the casualty position and then from the casualty position to the hospital position;

[0096] Define the ratio of the actual driving distance to the average speed as the response time;

[0097] The response time is divided into the response time from the ambulance to the casualty position and the response time from the casualty position to the hospital position;

[0098] Use the historical regression analysis method to set the road condition congestion coefficient and dynamically adjust the response time from the ambulance position to the casualty position. The formula is:

[0099] t 2,qe =t qe ·(1 + c qe )

[0100] In the formula, t 2,qe represents the adjusted response time from the ambulance position q to the casualty position e, t qe represents the initial response time from the ambulance position q to the casualty position e, and c qe represents the road condition congestion coefficient from the ambulance position q to the casualty position e;

[0101] Further use the dynamic adjustment formula to calculate the response time from the casualty position to the hospital position to obtain the adjusted response time from the casualty position to the hospital position;

[0102] According to the adjusted response time, calculate the total response time. The formula is:

[0103] t 4,q→e→h =t 2,qe +t 3,eh

[0104] In the formula, t 4,q→e→h represents the total response time of the complete path q→e→h, t3,eh Denote the response time from the adjusted casualty location e to the hospital location h;

[0105] Set a constant using the rule of thumb, and calculate the initial pheromone by combining the total response time and the constant using the proportional formula. The formula is:

[0106]

[0107] In the formula, denotes the initial pheromone of the complete path q→e→h, and Q denotes the constant;

[0108] Further calculate the initial heuristic information by combining the total response time using the proportional formula. The formula is:

[0109]

[0110] In the formula, η q→e→h denotes the initial heuristic information of the complete path q→e→h;

[0111] Set the weight coefficient using the grid search method, and calculate the path selection probability and generate a set of path selection probabilities using the path selection probability formula. The formula is:

[0112]

[0113] In the formula, P q→e→h denotes the path selection probability of the complete path q→e→h, N denotes the total number of casualty locations, and α and β denote the weight coefficients of the initial pheromone and the initial heuristic information respectively;

[0114] When the path selection probability of the severely injured casualty exists in the set of path selection probabilities, directly use the path selection probability of the severely injured casualty as the path and perform dynamic adjustment. The formula is:

[0115]

[0116] In the formula, denotes the path selection probability of the complete path q→e→h of the mth severely injured casualty after adjustment, Z m denotes the severe injury assessment probability of the mth severely injured casualty, denotes the path selection probability of the complete path q→e→h of the mth severely injured casualty;

[0117] When there are l severely injured casualties in the set of path selection probabilities, sort the adjusted path selection probabilities in ascending order and select the largest path selection probability as the path;

[0118] When there are no severely injured casualties in the set of path selection probabilities, use roulette wheel selection to select the path;

[0119] Normalize the path selection probability and total response time of the selected path, and use the expert scoring method to set the decay coefficient and weight coefficient;

[0120] Define the objective function to minimize response time and maximize rescue efficiency. The formula is:

[0121]

[0122] f3=σ1·f1+σ2·(1-f2)

[0123] Where f1 represents the objective function value of response time, U represents the total number of ambulances, f2 represents the objective function value of rescue efficiency, G represents the total number of seriously injured people, and λ represents the attenuation coefficient. represents the total response time of the complete path q→e→h for the mth seriously injured person, exp(·) represents the exponential decay function, f3 represents the value of the comprehensive objective function, σ1 represents the weight coefficient of the response time, and σ2 represents the weight coefficient of the rescue efficiency;

[0124] Theoretical analysis is used to set the volatility coefficient, calculate the objective function value, count all objective function values, and select the maximum objective function value to update the path according to the path selection probability. During the update process, the volatility coefficient is used to control the evaporation rate of the pheromone. During the iterative process, when the number of iterations reaches the maximum number, the optimal solution is output;

[0125] The optimal solution refers to the best path that minimizes response time and maximizes rescue efficiency.

[0126] By incorporating the ratio of the road congestion coefficient to the historical speed average into the response time calculation, this method effectively addresses the problem of static path planning ignoring actual traffic conditions, making path evaluation more consistent with the on-site environment and improving scheduling reliability. Furthermore, by combining initial pheromones with heuristic information weights to calculate path probabilities, this method reflects the dual considerations of actual traffic accessibility and medical urgency. This mechanism ensures the rationality of resource prioritization, especially when dealing with multi-objective scheduling (multiple injured patients). Introducing the patient's injury probability as a path adjustment factor, medical judgment is incorporated into the algorithm's decision-making logic, achieving a deep integration of "medical care and scheduling," prioritizing resources for the seriously injured and maximizing lifesaving efficiency. Furthermore, by introducing dual objective functions of minimizing response time and maximizing rescue efficiency and using an expert scoring method to set weights, this method enables comprehensive optimal solution search. This dual objective design overcomes the drawback of traditional scheduling systems that prioritize single-time cost optimization. By combining theoretical analysis with a volatility coefficient to naturally decay pheromone concentration over time, the algorithm is guided to achieve a dynamic balance between exploration and exploitation. By controlling the pheromone iteration process, path selection gradually approaches the global optimum.

[0127] Furthermore, constructing an Actor-Critic network to optimize the parameters of the ant colony algorithm includes the following steps:

[0128] Use the normalized state space to scale the response time, rescue efficiency, weight coefficients α and β, and evaporation coefficient in the optimal solution to obtain a scaled data set;

[0129] Use a multi-layer perceptron architecture to construct the Actor network and the Critic network in the A2C algorithm;

[0130] Both the Actor network and the Critic network include an input layer, a hidden layer, and an output layer;

[0131] Both the hidden layers of the Actor network and the Critic network are set with two fully connected layers and contain ReLU activation functions;

[0132] The output layer of the Actor network contains a Tanh function;

[0133] Define a policy gradient loss function and a mean squared error loss function, and use the Adam optimizer to perform iterative optimization of the parameters of the Actor network and the Critic network. During the iteration process, when the decrease values of the policy gradient loss function and the mean squared error loss function no longer decrease significantly, output the optimized Actor network and Critic network;

[0134] Input the scaled data set into the Critic network and output the value of the data set;

[0135] Use the n-step return method combined with the value of the data set to calculate the cumulative return;

[0136] Define the difference between the cumulative return and the value of the data set as the advantage function;

[0137] Use the Delphi method to set a feedback threshold, compare the advantage function with the feedback threshold, and when the advantage function is greater than the feedback threshold, update and increment the policy gradient of the Actor network, otherwise update and decrement the policy gradient of the Actor network;

[0138] Input the scaled data set into the updated Actor network and output a three-dimensional action vector;

[0139] The three-dimensional action vector includes weight coefficients α and β and the evaporation coefficient;

[0140] Use double exponential smoothing filtering to smooth the three-dimensional action vector, write the smoothed three-dimensional action vector into the Redis key-value, and read it in real time through the ACO scheduling module to trigger the recalculation of the path selection probability;

[0141] The monitoring module is used to detect the performance metrics of the ACO scheduling module. When the comprehensive objective function value of the recalculated path selection probability is greater than the comprehensive objective function value of the optimal solution, the historical optimal parameters are rolled back, and the parameters in the optimal solution after rollback are used as the new parameters. Otherwise, the recalculated parameters are used as the final parameters.

[0142] The dynamic three-dimensional action vector is output through the Actor network in A2C to achieve the adaptive update of the key parameters of ACO. This method not only reduces the empirical dependence on manual parameter setting but also enables ACO to have the ability of environment perception and adjustment, enhancing the applicability of the algorithm in complex and changeable scenarios. The Critic network evaluates the state value, combines the n-step return method and the advantage function, and introduces a policy gradient adjustment mechanism to avoid the policy falling into local optimality. At the same time, the two-way update strategy of comparing the advantage function with the feedback threshold further improves the stability of algorithm convergence and the directionality of policy evolution. Redis and the ACO scheduling module form a closed-loop control of parameter update - path recalculation - performance detection, with extremely high response efficiency. Combining the continuous comparison of the objective function value by the monitoring module not only ensures the superiority of the new solution but also introduces a parameter rollback mechanism, adding redundant protection to the scheduling process. Secondly, the feedback optimization of the A2C network not only controls the parameter output but also obtains the performance metric feedback of the ACO module in real time to form a closed-loop iteration. Moreover, the value function and the advantage function output by the Critic form a guiding role for the Actor policy. Supplementary with the feedback threshold set by the Delphi method, it effectively avoids overfitting or fluctuations in policy updates.

[0143] Furthermore, after global optimization using the genetic algorithm, it is fed back to the ant colony algorithm for iteration. The steps to obtain the optimal solution are as follows:

[0144] Use the hybrid coding technique to perform discrete coding on the optimal path in the optimal solution, and perform continuous coding on the weight coefficients α and β,

[0145] the evaporation coefficient, and the three-dimensional action vector in the optimal solution;

[0146] Combine the discrete coding and the continuous coding and use them as individuals to generate a population for initialization;

[0147] Calculate the objective function value using the objective function of minimizing the response time and maximizing the rescue efficiency. Use each objective function value as the fitness value of the corresponding individual. Randomly select individuals from the population as the next generation using the tournament method, and generate new individuals using crossover and mutation operations. During the iteration process, when the iteration count reaches the maximum number, output the optimal path;

[0148] Use the graph theory algorithm in the open-source library to repair the optimal path and update the pheromone using the pheromone update rule;

[0149] The updated pheromone is decoded using the decoding library and the decoded pheromone is re-input into the ACO scheduling module for iteration to obtain a new optimal solution.

[0150] By utilizing hybrid coding technology to compatibly handle the optimization problems of structural paths (discrete) and policy parameters (continuous), the present invention bridges the gap between path search and algorithm performance tuning. This allows the present invention to generate a diverse population during the initialization phase, effectively enhancing the coverage of the search space and avoiding early convergence. Furthermore, it increases the diversity of the initial distribution of individual qualities, improving the algorithm's global search capability. Furthermore, the dual objective function of minimum response time and maximum rescue efficiency is used for individual evaluation, meeting the dual requirements of efficiency and effectiveness in actual emergency dispatch and achieving a dynamic trade-off between service efficiency and rescue quality. Furthermore, a graph theory algorithm is introduced to structurally repair the legitimacy and feasibility of paths, ensuring that the paths meet actual operational constraints. The GA optimization results are then fed back to the ACO, and pheromone updates are used to strengthen the valid paths, ensuring the physical feasibility of the path solutions. Furthermore, the pheromone-enhanced feedback mechanism accelerates the convergence of the ACO algorithm. The path parameters optimized by the GA are decoded into a scheduling strategy that the ACO algorithm can recognize, enabling the application and feedback of the optimization results in the ACO model, forming a multi-round, alternating, and linked optimization closed loop.

[0151] The three algorithms, ACO, A2C, and GA, work together through precise interaction to form an efficient optimization framework. First, A2C optimizes ACO's path selection parameters through reinforcement learning to ensure a more accurate path selection strategy. ACO finds a preliminary path through local search and passes the path solution to GA for global optimization. GA avoids local optimality problems through global optimization and feeds the results back to ACO for further adjustment. The linkage between the three ensures the accuracy, efficiency, and global coordination of path selection, which enables the present invention to comprehensively improve the quality of path selection from local optimization to global optimization and then to strategy adjustment. This multi-level optimization framework ensures that path selection is both efficient and accurate, reduces response time, and improves emergency efficiency.

[0152] S3, the conversion and transmission module, converts the optimal solution, transmits it through the MQTT protocol, and sends it in real time;

[0153] Specifically, the optimal solution is converted and transmitted and distributed in real time through the MQTT protocol, which means that the optimal solution is converted into JSON format, the optimal solution in JSON format is transmitted to the edge computing device using the MQTT protocol, and then the optimal solution in JSON format is distributed to the dispatcher through the edge computing device.

[0154] By expressing the optimal scheduling results in JSON format, the scheduling information is made universal and extensible. Based on the MQTT protocol for transmission, the transmission efficiency and extensibility can be improved, and the "multi-point synchronization" of scheduling tasks can be achieved, avoiding the latency bottleneck and expansion obstacles in traditional point-to-point communication. Secondly, using edge computing devices to issue tasks can shorten the scheduling response time.

[0155] S4. Storage and display module, which stores data through a database and displays it through a visualization interface;

[0156] Specifically, storing data through a database means storing the JSON-formatted optimal solutions sent by the edge computing devices through the database, marking each optimal solution with a unique ID and timestamp, and sorting the JSON-formatted optimal solutions marked with timestamps in ascending order through the database.

[0157] Storing the JSON-formatted optimal solutions sent by the edge computing devices through the database and marking each solution with a unique ID and timestamp can effectively solve the redundancy and duplication problems in data management, and each optimal solution has an independent identity identifier (ID) and a clear generation time, ensuring conflict-free storage in the database.

[0158] Furthermore, displaying through a visualization interface means extracting the JSON-formatted optimal solutions from the database and using the Plotly.js tool to display the scheduling priority and response time.

[0159] By extracting the JSON-formatted optimal solutions from the database and using the Plotly.js tool to display the scheduling priority and response time, users can see the processing order of tasks in real time, and in a visual way, decision-makers can intuitively understand the priority ranking of each task in the current system.

[0160] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not restrictive. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered by the scope of the claims of the present invention.

Claims

1. On-site injury assessment and dispatch system based on deep learning model, characterized by: include, The acquisition and calculation module collects multi-source data and combines it with fuzzy logic and slope analysis to calculate a comprehensive score. It extracts time series and image features, fuses them through a multi-head attention mechanism, and adjusts the injury probability based on the comprehensive score to obtain light, moderate, and severe injury assessment results. The multi-source data includes vital sign data, injury image data and location information data; The optimization module combines the evaluation results and uses the ant colony algorithm to optimize the ambulance path. It builds an actor-critic network to optimize the ant colony algorithm parameters, and uses the genetic algorithm for global optimization. The results are fed back to the ant colony algorithm for iteration to obtain the optimal solution. The conversion and transmission module converts the optimal solution and transmits and distributes it in real time through the MQTT protocol; The storage and display module stores data through the database and displays it through a visual interface.

2. The on-site injury assessment and dispatch system based on a deep learning model according to claim 1, characterized in that: The method of collecting multi-source data, combining fuzzy logic and slope analysis, and calculating the comprehensive score includes the following steps: The patient's vital signs data, injury image data, and location information data are collected at consecutive time points using physiological monitoring equipment, portable X-ray machines, and GPS equipment. The vital signs data include actual measured values of heart rate, respiratory rate, blood oxygen saturation, body temperature and blood pressure; Use empirical rules to set corresponding thresholds for each data in the vital sign data; The corresponding thresholds include the lowest physiological reasonable value, the highest physiological reasonable value, the lower limit of the safety range and the upper limit of the safety range corresponding to each data; Using fuzzy logic rules, a data reliability score for each data in the vital sign data is calculated; According to each data in the vital sign data, the change rate of each data and the corresponding data at the adjacent time point is calculated respectively, and the data slope of the corresponding data is defined as the data slope. Then, the maximum data slope is selected from the slopes of the corresponding data using the maximization operation, and the data slope reliability score is calculated; The reliability score and slope reliability score are combined using a logical operation formula to calculate the final reliability score of the data. The weighted average formula is used to combine the final reliability score of each data at each time point to calculate the comprehensive score of all data at each time point to obtain the comprehensive score of the current injured person.

3. The on-site injury assessment and dispatch system based on a deep learning model according to claim 2, characterized in that: The extraction of time series and image features, fusing them through a multi-head attention mechanism, and adjusting the injury probability in combination with a comprehensive score to obtain light, moderate, and severe injury assessment results include the following steps: After building a feature extraction model using a deep convolutional neural network architecture, the vital signs data and injury image data are input into the feature extraction model respectively. After obtaining the vital features, the injury image data is nonlinearly transformed using the ReLU activation function to obtain image features. Use the multi-head attention mechanism to perform weighted fusion of life features and image features to obtain fused features; Use the softmax function to transform the fused features to obtain the final injury assessment probability; The final injury assessment probability includes a light injury assessment probability, a moderate injury assessment probability, and a severe injury assessment probability; Using element-wise multiplication, multiply the comprehensive score element-wise with the probabilities in the final injury assessment probability to obtain the adjusted probabilities. After sorting them in ascending order, take the largest probability as the assessment result for the current casualty, including minor injuries, moderate injuries, and severe injuries.

4. The on-site injury assessment and dispatching system based on a deep learning model according to claim 3, characterized in that: Combined with the assessment results, using the ant colony algorithm to optimize the ambulance route includes the following steps: Collect the locations of ambulances and hospitals. Take the ambulance location as the starting point, and take the casualty location and hospital location as nodes. Use the map API to obtain the actual driving distance of the complete route, and average the historical speeds of the ambulances in the map API through HTTP request technology to obtain the average speed. The complete route is from the ambulance location to the casualty location, and then from the casualty location to the hospital location. Define the ratio of the actual driving distance to the average speed as the response time. The response time is divided into the response time from the ambulance to the casualty location and the response time from the casualty location to the hospital location. Use historical regression analysis to set the road congestion coefficient, dynamically adjust the response time from the ambulance location to the casualty location, and then use the dynamic adjustment formula to calculate the response time from the casualty location to the hospital location to obtain the adjusted response time from the casualty location to the hospital location, and calculate the total response time. Use the empirical rule to set a constant, and use the proportional formula to combine the total response time and the constant to calculate the initial pheromone. Then, further use the proportional formula to combine the total response time to calculate the initial heuristic information. Use the grid search method to set the weight coefficient, and use the path selection probability formula to calculate the path selection probability and generate a set of path selection probabilities. When the path selection probability of a severely injured casualty exists in the set of path selection probabilities, directly take the path selection probability of the severely injured casualty as the path and perform dynamic adjustment. When there are l severely injured casualties in the set of path selection probabilities, sort the adjusted path selection probabilities in ascending order and select the largest path selection probability as the path. When there are no severely injured casualties in the set of path selection probabilities, use roulette wheel selection to select the path. Normalize the path selection probability and the total response time of the selected path, and use the expert scoring method to set the attenuation coefficient and weight coefficient, define the objective function, and minimize the response time and maximize the rescue efficiency. Use theoretical analysis to set the evaporation coefficient, calculate the objective function value, perform iterative updates, and use the evaporation coefficient to control the evaporation rate of the pheromone, and output the optimal solution. The optimal solution refers to the optimal path with the minimized response time and the maximized rescue efficiency.

5. The on-site injury assessment and dispatching system based on a deep learning model according to claim 4, characterized in that: Constructing the Actor-Critic network to optimize the ant colony algorithm parameters includes the following steps: Use the normalized state space to scale the response time, rescue efficiency, weight coefficients α and β, and evaporation coefficient in the optimal solution to obtain a scaled dataset. Use the multi-layer perceptron architecture to construct the Actor network and Critic network in the A2C algorithm. Input the scaled dataset into the Critic network to output the value of the dataset. Then, use the n-step return method to combine the value of the dataset. After calculating the cumulative return, define the difference between the cumulative return and the dataset value as the advantage function. Use the Delphi method to set the feedback threshold, compare the advantage function with the feedback threshold to obtain the updated Actor network, and input the scaled dataset into the updated Actor network to output a three-dimensional action vector; The three-dimensional action vector includes weight coefficients α and β and a volatility coefficient; Perform a smoothing operation on the three-dimensional action vector using double exponential smoothing filtering, write the smoothed three-dimensional action vector into the Redis key-value pair, and read it in real time through the ACO scheduling module to trigger the recalculation of the path selection probability; Use the monitoring module to detect the performance metrics of the ACO scheduling module. When the comprehensive objective function value of the recalculated path selection probability is greater than the comprehensive objective function value of the optimal solution, roll back the historical optimal parameters and use the parameters in the optimal solution after rollback as the new parameters. Otherwise, use the recalculated parameters as the final parameters.

6. The on-site injury assessment and dispatching system based on a deep learning model according to claim 5, characterized in that: After performing global optimization using the genetic algorithm and feeding it back to the ant colony algorithm for iteration, obtaining the optimal solution includes the following steps: Use the hybrid coding technique to discretely encode the optimal path in the optimal solution, and continuously encode the weight coefficients α and β, the volatility coefficient, and the three-dimensional action vector in the optimal solution and then combine them as individuals to generate a population for initialization; calculate the objective function value using the objective function of minimizing the response time and maximizing the rescue efficiency, and perform iterative optimization to output the optimal path; Use the graph theory algorithm in the open source library to repair the optimal path, update the pheromone using the pheromone update rule, decode the updated pheromone using the decoding library, and re-enter the decoded pheromone into the ACO scheduling module for iteration to obtain a new optimal solution.

7. The on-site injury assessment and dispatching system based on a deep learning model according to claim 1, characterized in that: Collect the vital sign data, injury image data, and location information data, and first perform preprocessing operations including KNN imputation, outlier detection, standardization, and normalization.

8. The on-site injury assessment and dispatch system based on a deep learning model according to claim 6, characterized in that: The conversion of the optimal solution and its transmission and real-time distribution through the MQTT protocol means converting the optimal solution into the JSON format, using the MQTT protocol to transmit the JSON-format optimal solution to the edge computing device, and then the edge computing device distributes the JSON-format optimal solution to the dispatcher.

9. The on-site injury assessment and dispatch system based on a deep learning model according to claim 8, wherein: The storage of data through the database means storing the JSON-format optimal solution sent by the edge computing device through the database, marking each optimal solution with a unique ID and timestamp, and sorting the JSON-format optimal solutions marked with timestamps in ascending order through the database.

10. The on-site injury assessment and dispatch system based on a deep learning model according to claim 9, characterized in that: The display through the visualization interface means extracting the JSON-format optimal solution from the database and using the Plotly.js tool to display the scheduling priority and response time.

Images