On-site injury assessment and dispatching system based on deep learning model

Abstract

This invention discloses a field injury assessment and dispatch system based on a deep learning model, belonging to the field of intelligent medical dispatch technology. It includes a data acquisition and calculation module, which collects multi-source data and combines it with fuzzy logic and slope analysis to calculate a comprehensive score, extract temporal and image features, fuses them through a multi-head attention mechanism, and adjusts the injury probability based on the comprehensive score to obtain assessment results for minor, moderate, and severe injuries. An optimization module, based on the assessment results, uses an ant colony algorithm to optimize ambulance routes. An Actor-Critic network is constructed to optimize the ant colony algorithm parameters, and a genetic algorithm is used for global optimization, which is then fed back to the ant colony algorithm for iteration to obtain the optimal solution. This invention achieves accurate assessment of the injuries of the wounded and improves the efficiency of ambulance dispatch, ensuring that the wounded receive timely treatment in the shortest possible time. Furthermore, it achieves a high degree of integration between injury assessment results and route optimization dispatch, thereby improving the response speed and accuracy of the invention.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical scheduling technology, and in particular to an on-site injury assessment and scheduling system based on a deep learning model. Background Technology

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

[0003] Existing on-site injury assessment and dispatch systems are mostly limited to simple rule-based assessments and lack effective comprehensive analysis mechanisms, making it difficult to provide accurate injury assessment results in complex environments. In addition, when planning routes, existing on-site injury assessment and dispatch systems usually only focus on a single indicator and fail to fully consider changes in road conditions and real-time dynamic data, resulting in low efficiency and accuracy in ambulance route selection, which in turn affects rescue efficiency. Summary of the Invention

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

[0005] Therefore, this invention provides a field injury assessment and dispatch system based on a deep learning model, which solves the problem that most existing field 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, existing field injury assessment and dispatch systems usually only focus on a single indicator when planning routes, failing to fully consider changes in road conditions and real-time dynamic data, resulting in low efficiency and accuracy of ambulance route selection, which in turn affects rescue efficiency.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a field injury assessment and scheduling system based on a deep learning model, which includes a data acquisition and calculation module, which acquires multi-source data and combines it with fuzzy logic and slope analysis, calculates a comprehensive score, extracts temporal and image features, fuses them through a multi-head attention mechanism, and adjusts the injury probability in combination with the comprehensive score to obtain the assessment results of minor, moderate and severe injuries.

[0008] Multi-source data includes vital signs data, injury image data, and location information data;

[0009] The optimization module, based on the evaluation results, uses the ant colony algorithm to optimize the ambulance route, constructs an Actor-Critic network to optimize the ant colony algorithm parameters, and performs global optimization using a genetic algorithm before feeding it 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 via the MQTT protocol.

[0010] The storage and display module stores data through a database and displays it through a visual interface.

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

[0012] The vital signs, injury images, and location information of the injured were collected at continuous time points using physiological monitoring equipment, portable X-ray machines, and GPS devices, respectively.

[0013] Vital signs data include actual measurements of heart rate, respiratory rate, blood oxygen saturation, body temperature, and blood pressure;

[0014] Use rules of thumb to set corresponding thresholds for each data point in the vital signs data;

[0015] The corresponding thresholds include the minimum physiologically reasonable value, the maximum physiologically reasonable value, the lower limit of the safety range, and the upper limit of the safety range for each data point;

[0016] Using fuzzy logic rules, calculate the data reliability score for each data point in the vital signs data;

[0017] For each data point in the vital signs data, calculate the rate of change between each data point and the corresponding data point at adjacent time points. Define the rate of change as the slope of the corresponding data point. Then, use the maximization operation to select the maximum data slope from the slopes of the corresponding data points and calculate the data slope reliability score.

[0018] The final reliability score of the data is calculated by combining the reliability score and the slope reliability score using logical operation formulas. Then, the final reliability score of all data at each time point is calculated by combining the final reliability score of each data point with the weighted average formula, thus obtaining the comprehensive score of the current injured person.

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

[0020] After constructing a feature extraction model using a deep convolutional neural network architecture, vital sign data and injury image data are input into the feature extraction model respectively to obtain vital signs. Then, the ReLU activation function is used to perform a nonlinear transformation on the injury image data to obtain image features.

[0021] A multi-head attention mechanism is used to weightedly fuse vital features and image features to obtain fused features;

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

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

[0024] Using element-wise multiplication, the overall score is multiplied element-wise with the probability in the final injury assessment probability to obtain the adjusted probability. After sorting in ascending order, the highest probability is taken as the current assessment result for the injured person, including minor injury, moderate injury, and serious injury.

[0025] As a preferred embodiment of the on-site injury assessment and dispatch system based on a deep learning model of the present invention, the following steps are included: Optimizing the ambulance route using an ant colony algorithm based on the assessment results:

[0026] Collect the locations of ambulances and hospitals, using the ambulance location as the starting point and the locations of the injured and hospitals as nodes;

[0027] The actual driving distance of the complete route is obtained using the map API, and the historical speed of the ambulance in the map API is obtained through HTTP request technology and averaged to obtain the average speed.

[0028] The complete route is from the ambulance location to the injured person's location, and then from the injured person's location to the hospital location;

[0029] The ratio of actual driving distance to average speed is defined as response time;

[0030] Response time is divided into the response time from the ambulance to the location of the injured person and the response time from the location of the injured person to the location of the hospital;

[0031] The traffic congestion coefficient was set using historical regression analysis. After dynamically adjusting the response time from the ambulance location to the location of the injured person, the dynamic adjustment formula was used to calculate the response time from the location of the injured person to the location of the hospital, and the adjusted response time from the location of the injured person to the location of the hospital was obtained. The total response time was then calculated.

[0032] After setting constants using rules of thumb and calculating the initial pheromone by combining the total response time and constants using a proportional formula, the initial heuristic information is calculated by further combining the proportional formula with the total response time.

[0033] The weight coefficients are set using the grid search method, and the path selection probability set is generated by calculating the path selection probability using the path selection probability formula.

[0034] When the path selection probability of a seriously injured person exists in the path selection probability set, the path selection probability of the seriously injured person is directly used as the path and dynamically adjusted. When there are l seriously injured persons in the path selection probability set, the adjusted path selection probabilities are sorted in ascending order and the path selection probability with the largest probability is selected as the path. When there are no seriously injured persons in the path selection probability set, a roulette wheel is used to select the path.

[0035] The path selection probability and total response time of the selected path are normalized, and the attenuation coefficient and weight coefficient are set using the expert scoring method. The objective function is defined to minimize the response time and maximize the rescue efficiency.

[0036] The volatility coefficient is set using theoretical analysis, the objective function value is calculated, iterative updates are performed, the volatility coefficient is used to control the evaporation rate of pheromones, and the optimal solution is output.

[0037] The optimal solution refers to the path that minimizes response time and maximizes rescue efficiency.

[0038] As a preferred embodiment of the on-site injury assessment and scheduling system based on a deep learning model of the present invention, the following steps are included in constructing the Actor-Critic network to optimize the ant colony algorithm parameters:

[0039] The response time, rescue efficiency, weighting coefficients α and β, and volatility coefficient in the optimal solution are scaled using a normalized state space to obtain a scaled dataset.

[0040] The Actor network and Critic network in the A2C algorithm are constructed using a multilayer perceptron architecture;

[0041] The scaled dataset is input into the Critic network, which outputs the value of the dataset. The cumulative reward is calculated by combining the dataset value with the n-step reward method. The difference between the cumulative reward and the dataset value is defined as the dominance function. The Delphi method is used to set the feedback threshold. The dominance function is compared with the feedback threshold to obtain the updated Actor network. The scaled dataset is input into the updated Actor network to output a three-dimensional action vector.

[0042] The three-dimensional motion vector includes weighting coefficients α and β, as well as a volatility coefficient;

[0043] The three-dimensional action vector is smoothed by double exponential smoothing filter, and the smoothed three-dimensional action vector is written to a Redis key value and read in real time by the ACO scheduling module to trigger the recalculation of the path selection probability.

[0044] The monitoring module is used to detect the performance indicators 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 rolled-back optimal solution are used as the new parameters. Otherwise, the recalculated parameters are used as the final parameters.

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

[0046] The optimal path in the optimal solution is discretely encoded using a hybrid encoding technique. The weight coefficients α and β, volatility coefficient, and three-dimensional action vector in the optimal solution are continuously encoded and combined, and then used as individuals to generate a population for initialization. The objective function is calculated using the objective function of minimizing response time and maximizing rescue efficiency. After iterative optimization, the optimal path is output.

[0047] The optimal path is repaired using a graph theory algorithm from an open-source library. After updating the pheromone using the pheromone update rule, the updated pheromone is decoded using a decoding library. The decoded pheromone is then re-input into the ACO scheduling module for iteration to obtain a new optimal solution.

[0048] As a preferred embodiment of the on-site injury assessment and scheduling system based on a deep learning model of the present invention, the system collects vital sign data, injury image data, and location information data, and performs preprocessing operations including KNN interpolation, outlier detection, standardization, and normalization.

[0049] As a preferred embodiment of the on-site injury assessment and scheduling system based on a deep learning model of the present invention, the optimal solution is converted and transmitted and distributed in real time via the MQTT protocol. The optimal solution is converted into JSON format, and after the optimal solution in JSON format is transmitted to the edge computing device via the MQTT protocol, the optimal solution in JSON format is distributed to the scheduling personnel via the edge computing device.

[0050] As a preferred embodiment of the on-site injury assessment and scheduling system based on a deep learning model of the present invention, the following is provided: storing data through a database refers to storing the optimal solutions in JSON format issued by the edge computing device through a database, marking each optimal solution with a unique ID and a timestamp, and sorting the optimal solutions in JSON format after marking the timestamp in the database in ascending order.

[0051] As a preferred embodiment of the on-site injury assessment and scheduling system based on a deep learning model of the present invention, the optimal solution is extracted from the database in JSON format through a visual interface, and the scheduling priority and response time are displayed using the Plotly.js tool.

[0052] The beneficial effects of this invention are as follows: This invention uses fuzzy logic rules to assess the reliability of vital sign data and employs slope analysis to capture data change trends, thereby achieving accurate assessment of the injured person's condition. Furthermore, it uses a multi-head attention mechanism for feature fusion, which not only improves the accuracy of the assessment but also allows for dynamic adjustment of the assessment results. Moreover, by introducing an ant colony algorithm to optimize ambulance routes and combining it with historical regression analysis to dynamically adjust traffic congestion coefficients, it further improves ambulance dispatch efficiency, ensuring that the injured receive timely treatment in the shortest possible time. Secondly, by combining the advantages of ant colony algorithms, A2C algorithms, and genetic algorithms, it achieves a high degree of integration between injury assessment results and route optimization scheduling. This allows the invention to not only consider injury assessment results during route selection but also provide real-time feedback and adjust optimization parameters, thereby improving the response speed and accuracy of the invention. Attached Figure Description

[0053] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is a structural diagram of the on-site injury assessment and scheduling system based on a deep learning model in Example 1.

[0055] Figure 2 This is a flowchart of the injury assessment in Example 1.

[0056] Figure 3 This is a flowchart of feature extraction and path optimization in Example 1.

[0057] Figure 4 This is a structural diagram of the optimization module in Example 1. Detailed Implementation

[0058] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0059] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0060] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation 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 a single or selective embodiment that is mutually exclusive with other embodiments.

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

[0062] S1, the data acquisition and calculation module, collects multi-source data and combines it with fuzzy logic and slope analysis to calculate a comprehensive score, extracts temporal and image features, and then integrates them through a multi-head attention mechanism. Finally, it adjusts the injury probability based on the comprehensive score to obtain the assessment results for minor, moderate and severe injuries.

[0063] Multi-source data includes vital signs data, injury image data, and location information data;

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

[0065] The vital signs, injury images, and location information of the injured were collected at continuous time points using physiological monitoring equipment, portable X-ray machines, and GPS devices, respectively.

[0066] Vital signs data include actual measurements of heart rate, respiratory rate, blood oxygen saturation, body temperature, and blood pressure;

[0067] Using rules of thumb, a corresponding threshold is set for each data point in the vital signs data, resulting in 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 minimum physiologically reasonable value, the maximum physiologically reasonable value, the lower limit of the safety range, and the upper limit of the safety range for each data point;

[0069] Using fuzzy logic rules, a data reliability score is calculated for each data point in the vital signs data. The formula is as follows:

[0070]

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

[0072] For each data point in the vital signs data, calculate the rate of change between each data point and its corresponding data point at adjacent time points. This rate is defined as the slope of the corresponding data point, and the formula is as follows:

[0073]

[0074] In the formula, g a,i s represents the slope of the data for the a-th vital sign at time point i. a,i-1 This represents the actual measured value of the data for the a-th injured person at time point i-1, where t represents the time interval, s. a,i Let represent the actual measured value of the a-th vital sign at time point i; the maximum data slope is selected from the slopes of the corresponding data using a maximization operation, and the data slope reliability score is calculated using the following formula:

[0075]

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

[0077] The final reliability score of the data is calculated by combining logical operation formulas with reliability scores and slope reliability scores. The formula is as follows:

[0078] φ a,i =min(ψ) a,i ,r a,i )

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

[0080] By combining the weighted average formula with the final reliability score of each data point at each time point, a comprehensive score for all data at each time point is calculated to obtain the current comprehensive score for the injured person.

[0081] Vital signs data provide the basic physiological condition of the injured, injury images can help assess the location and severity of injuries, and location information helps determine the location of the injured and the timeliness of receiving medical services. The comprehensive use of multi-source data allows for a more accurate assessment of the injured's health status. In emergency environments, injured individuals may be in a state of incomplete information; multi-source data fusion can compensate for the shortcomings of a single data source, improving the accuracy of the assessment. Furthermore, using fuzzy logic rules to determine whether the data is within a reasonable range and generating a reliability score for each data point allows fuzzy logic to effectively provide a meaningful evaluation even when there is some error or uncertainty in the data. Fuzzy logic can effectively handle complex and uncertain data, avoiding misjudgments due to data noise. Through fuzzy rules, the reasonableness of each data point can be evaluated, thereby improving the robustness and accuracy of the invention. Secondly, through oblique... Rate analysis calculates the rate of change of vital signs data, capturing the dynamic trend of changes in the patient's condition, especially in physiological data such as heart rate and respiratory rate. Slope analysis can help identify sudden changes and reflect the patient's physiological changes in a timely manner. Slope analysis can provide important information about the trend of changes in the patient's condition, helping emergency personnel identify rapid changes in the patient's condition, thus providing an important basis for decision-making. By weighted averaging the reliability score and the slope reliability score, a comprehensive score for the patient is finally obtained. This score combines multiple factors and can quantitatively assess the patient's condition from multiple dimensions. Furthermore, the comprehensive score integrates multiple physiological data and injury factors, so that the injury assessment is not limited to a single data point, but comprehensively analyzes the patient's condition from multiple perspectives. This can provide medical personnel with a more comprehensive and accurate decision support, and help improve the response speed and accuracy in the emergency process.

[0082] Furthermore, temporal and image features are extracted, fused using a multi-head attention mechanism, and then combined with a comprehensive score to adjust the injury probability, resulting in mild, moderate, and severe injury assessments. This process includes the following steps:

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

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

[0085] A multi-head attention mechanism is used to weightedly fuse 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 probability of minor injury assessment, the probability of moderate injury assessment, and the probability of severe injury assessment;

[0088] Using element-wise multiplication, the overall score is multiplied element-wise with the probability in the final injury assessment probability to obtain the adjusted probability, which is then sorted in ascending order.

[0089] The highest probability in the ranking results is used as the current assessment result for the injured person, including minor injury, moderate injury, and serious injury.

[0090] By inputting vital signs and image features into a CNN model, high-level semantic information can be extracted. Nonlinear activation of image features suppresses the adverse effects of values ​​below zero, enhances the propagation of positive information, and enables the neural network to exhibit stronger convergence and generalization capabilities when processing complex image features. Furthermore, by using a multi-head attention mechanism to weightedly fuse vital signs and image features, cross-information enhancement can be achieved, improving the accuracy of understanding and judging complex injury patterns. Using the softmax function, the model output can be normalized into a probability distribution, facilitating an intuitive assessment of the likelihood of minor, moderate, and severe injuries. This provides standardized output results for this invention and facilitates subsequent scheduling. Secondly, the introduction of a comprehensive score as an external weight to adjust the assessment probability is a soft-guided mechanism. This approach not only alleviates the model's over-reliance on a single data source but also enhances the individual adaptability of the assessment, improving the invention's ability to handle complex on-site situations. By ranking and determining the injury level with the highest probability, the final assessment decision output is formed, ensuring that the injured receive the most realistic emergency resource response during multi-level triage.

[0091] S2, Optimization Module: Based on the evaluation results, the ant colony algorithm is used to optimize the ambulance route. An Actor-Critic network is constructed to optimize the parameters of the ant colony algorithm. After global optimization using a genetic algorithm, the results are fed back to the ant colony algorithm for iteration to obtain the optimal solution.

[0092] Specifically, based on the evaluation results, optimizing ambulance routes using the ant colony algorithm includes the following steps:

[0093] Collect the locations of ambulances and hospitals, using the ambulance location as the starting point and the locations of the injured and hospitals as nodes;

[0094] The actual driving distance of the complete route is obtained using the map API, and the historical speed of the ambulance in the map API is obtained through HTTP request technology and averaged to obtain the average speed.

[0095] The complete route is from the ambulance location to the injured person's location, and then from the injured person's location to the hospital location;

[0096] The ratio of actual driving distance to average speed is defined as response time;

[0097] Response time is divided into the response time from the ambulance to the location of the injured person and the response time from the location of the injured person to the location of the hospital;

[0098] Historical regression analysis was used to determine the traffic congestion coefficient, and the response time from the ambulance location to the injured person's location was dynamically adjusted. The formula is as follows:

[0099] t 2,qe =t qe ·(1+c qe )

[0100] In the formula, t 2,qe t represents the response time from the adjusted ambulance location q to the injured person's location e. qe c represents the response time from the initial ambulance location q to the injured person's location e. qe This represents the traffic congestion coefficient between the ambulance's location q and the injured person's location e.

[0101] The dynamic adjustment formula is then used to calculate the response time from the location of the wounded to the location of the hospital, thus obtaining the adjusted response time from the location of the wounded to the location of the hospital.

[0102] The total response time is calculated based on the adjusted response time, using the following formula:

[0103] t 4,q→e→h =t 2,qe +t 3,eh

[0104] In the formula, t 4,q→e→h t represents the total response time for the complete path q→e→h.3,eh This indicates the response time from the adjusted location of the wounded soldier e to the hospital location h;

[0105] Using rules of thumb to set constants, and combining the proportion formula with the total response time and constants, the initial pheromone is calculated. The formula is:

[0106]

[0107] In the formula, Let Q represent the initial pheromone level for the complete path q→e→h, where Q is a constant.

[0108] Further, using the proportional formula in conjunction with the total response time, the initial heuristic information is calculated, as follows:

[0109]

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

[0111] We use a grid search method to set weight coefficients and calculate the path selection probability using the path selection probability formula to generate a path selection probability set. The formula is:

[0112]

[0113] In the formula, P q→e→h Let q→e→h represent the path selection probability of the complete path, N represent the total number of wounded locations, and α and β represent the weight coefficients of the initial pheromone and initial heuristic information, respectively.

[0114] When the path selection probability of a seriously injured person exists in the path selection probability set, the path selection probability of the seriously injured person is directly used as the path, and dynamically adjusted. The formula is as follows:

[0115]

[0116] In the formula, Z represents the path selection probability of the adjusted m-th seriously injured patient's complete path q→e→h. m This represents the probability of severe injury assessment for the m-th seriously injured patient. Let represent the path selection probability of the m-th seriously injured patient along the complete path q→e→h;

[0117] When there are l seriously injured people in the path selection probability set, the adjusted path selection probabilities are sorted in ascending order, and the path with the highest path selection probability is selected as the path.

[0118] When there are no seriously injured people in the path selection probability set, use roulette wheel to select the path;

[0119] The path selection probability and total response time of the selected path are normalized, and the attenuation coefficient and weight coefficient are set using the expert scoring method.

[0120] Define the objective function to minimize response time and maximize rescue efficiency, as shown in the formula:

[0121]

[0122] f3 = σ1·f1 + σ2·(1-f2)

[0123] In the formula, 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 patients, and λ represents the attenuation coefficient. Let f(·) represent the total response time of the m-th seriously injured person's complete path q→e→h, exp(·) represent the exponential decay function, f3 represent the comprehensive objective function value, σ1 represent the weighting coefficient of response time, and σ2 represent the weighting coefficient of rescue efficiency.

[0124] The volatilization coefficient is set using theoretical analysis, the objective function value is calculated, all objective function values ​​are statistically analyzed, and the path is updated according to the path selection probability based on the maximum objective function value. During the update process, the volatilization coefficient is used to control the evaporation rate of the pheromone. During the iteration process, the optimal solution is output when the number of iterations reaches the maximum number.

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

[0126] By introducing the ratio of traffic congestion coefficient to historical average speed to calculate response time, this method effectively solves the problem of static path planning ignoring real-time traffic conditions and makes path evaluation more consistent with the on-site environment, improving scheduling reliability. Furthermore, by combining initial pheromone and heuristic information weights to calculate path probability, it reflects the dual consideration of actual traffic accessibility and medical urgency. Especially when dealing with multi-objective scheduling (multiple injured persons), this mechanism ensures the rationality of resource priority allocation. Introducing the probability of injury severity as a path adjustment factor incorporates medical judgment into the algorithm's decision-making logic, achieving deep integration of "medical care and scheduling." This allows for priority allocation of resources to the seriously injured, maximizing life-saving efficiency. Secondly, the introduction of a dual objective function of minimizing response time and maximizing rescue efficiency, along with the use of expert scoring to set weights, enables this invention to perform comprehensive optimal solution seeking. The dual-objective design overcomes the shortcomings of traditional scheduling systems that focus on single-time cost optimization. Combining theoretical analysis to set an evaporation coefficient allows pheromone concentration to decay naturally over time, guiding the algorithm to dynamically balance exploration and development. By controlling the pheromone iteration process, the path selection gradually approaches the global optimum.

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

[0128] The response time, rescue efficiency, weighting coefficients α and β, and volatility coefficient in the optimal solution are scaled using a normalized state space to obtain a scaled dataset.

[0129] The Actor network and Critic network in the A2C algorithm are constructed using a multilayer perceptron architecture;

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

[0131] Both the Actor network and the Critic network have two fully connected layers in their hidden layers, and both contain the ReLU activation function.

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

[0133] Define the policy gradient loss function and the mean squared error loss function, and use the Adam optimizer to iteratively optimize the parameters of the Actor network and the Critic network. During the iteration process, when the decrease value of the policy gradient loss function and the mean squared error loss function no longer decreases significantly, the optimized Actor network and Critic network are output.

[0134] The scaled dataset is fed into the Critic network, which outputs the value of the dataset.

[0135] The cumulative return is calculated using the n-step reward method combined with the value of the dataset;

[0136] The difference between the cumulative return and the value of the dataset is defined as the dominance function;

[0137] The Delphi method is used to set a feedback threshold. The advantage function is compared with the feedback threshold. When the advantage function is greater than the feedback threshold, the policy gradient of the Actor network is updated and increased; otherwise, the policy gradient of the Actor network is updated and decreased.

[0138] Input the scaled dataset into the updated Actor network, which outputs 3D action vectors.

[0139] The three-dimensional motion vector includes weighting coefficients α and β, as well as a volatility coefficient;

[0140] The three-dimensional action vector is smoothed by double exponential smoothing filter, and the smoothed three-dimensional action vector is written to a Redis key value and read in real time by the ACO scheduling module to trigger the recalculation of the path selection probability.

[0141] The monitoring module is used to detect the performance indicators 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 rolled-back optimal solution are used as the new parameters. Otherwise, the recalculated parameters are used as the final parameters.

[0142] By outputting dynamic 3D action vectors through the Actor network in A2C, the ACO key parameters are adaptively updated. This approach not only reduces the reliance on experience for manual parameter setting but also enables ACO to have environmental awareness and adjustment capabilities, enhancing the algorithm's applicability in complex and ever-changing scenarios. The Critic network evaluates the value of the state, and combined with the n-step reward method and the advantage function, a policy gradient adjustment mechanism is introduced to avoid the policy getting trapped in local optima. At the same time, the bidirectional update strategy of comparing the advantage function and the feedback threshold further improves the stability of the algorithm's 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, which has extremely high response efficiency. Combined with the continuous comparison of the objective function value by the monitoring module, not only is the superiority of the new solution guaranteed, but a parameter rollback mechanism is also introduced to add redundancy protection to the scheduling process. Furthermore, the feedback optimization of the A2C network not only controls the parameter output but also obtains the performance index feedback of the ACO module in real time, forming a closed-loop iteration. Moreover, the value function and advantage function output by the Critic guide the Actor policy, and with the feedback threshold set by the Delphi method, overfitting or fluctuations in policy updates are effectively avoided.

[0143] Furthermore, after global optimization using a genetic algorithm, the results are fed back to the ant colony algorithm for iterative processing to obtain the optimal solution, which includes the following steps:

[0144] Hybrid coding techniques are used to discretely encode the optimal path in the optimal solution, and the weight coefficients α and β in the optimal solution are...

[0145] The volatility coefficient and three-dimensional motion vector are continuously encoded;

[0146] Discrete and continuous codes are combined and treated as individuals to generate a population for initialization;

[0147] The objective function is calculated by minimizing response time and maximizing rescue efficiency. Each objective function value is used as the fitness value of the corresponding individual. Individuals are randomly selected from the population as the next generation using a tournament. New individuals are generated using crossover and mutation operations. During the iteration process, the optimal path is output after the maximum number of iterations is reached.

[0148] The optimal path is repaired using graph theory algorithms from an open-source library, and pheromone is updated using pheromone update rules.

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

[0150] By utilizing hybrid coding technology, this invention can compatiblely handle optimization problems involving both discrete structural paths and continuous policy parameters, bridging the gap between path search and algorithm performance tuning. This allows the generation of diverse populations during the initialization phase, effectively enhancing search space coverage, avoiding early convergence, and increasing the diversity of initial individual quality distributions, thus improving the algorithm's global search capability. Furthermore, the use of a dual objective function—minimum response time and maximum rescue efficiency—for individual evaluation satisfies the dual requirements of efficiency and effectiveness in actual emergency dispatch, achieving a dynamic trade-off between service efficiency and rescue quality. Secondly, graph theory algorithms are introduced to structurally repair path legitimacy and feasibility, ensuring paths meet practical operational constraints. Then, the GA optimization results are fed back to ACO, and pheromone updates strengthen effective paths, guaranteeing the physical feasibility of path solutions. The pheromone-enhanced feedback mechanism accelerates ACO algorithm convergence. Finally, the path parameters obtained from GA optimization are decoded into scheduling strategies recognizable by the ACO algorithm, enabling the application and feedback of optimization results within the ACO model, forming a multi-round, alternating optimization loop.

[0151] The three algorithms, ACO, A2C, and GA, work together in a precise and collaborative manner to form a highly efficient optimization framework. First, A2C optimizes the path selection parameters of ACO through reinforcement learning, ensuring a more accurate path selection strategy. ACO finds an initial path through local search and hands the path solution over to GA for global optimization. GA avoids local optima problems through global optimization and feeds the results back to ACO for further adjustments. The interaction between the three ensures the accuracy, efficiency, and global coordination of path selection. This allows 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 response efficiency.

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

[0153] Specifically, the optimal solution is converted and transmitted and distributed in real time via the MQTT protocol. The optimal solution is converted into JSON format, and then transmitted to the edge computing device via the MQTT protocol. The edge computing device then distributes the optimal solution in JSON format to the scheduler.

[0154] By expressing the optimal scheduling results in JSON format, the scheduling information becomes universal and scalable. Transmission based on the MQTT protocol improves transmission efficiency and scalability, and enables "multi-point synchronization" of scheduling tasks, avoiding the latency bottlenecks and scalability obstacles in traditional point-to-point communication. Furthermore, using edge computing devices for task distribution can shorten the scheduling response time.

[0155] S4, Storage and Display Module: Stores data in a database and displays it through a visual interface;

[0156] Specifically, storing data through a database refers to storing the optimal solutions in JSON format issued by the edge computing device in a database, marking each optimal solution with a unique ID and timestamp, and then sorting the optimal solutions in JSON format after marking the timestamps in the database in ascending order.

[0157] By storing the optimal solutions in JSON format issued by edge computing devices in a database and marking each solution with a unique ID and timestamp, the redundancy and duplication problems in data management can be effectively solved. Furthermore, each optimal solution has an independent identifier (ID) and a clear generation time, ensuring conflict-free storage in the database.

[0158] Furthermore, the optimal solution is extracted from the database in JSON format and displayed through a visual interface, and the scheduling priority and response time are displayed using the Plotly.js tool.

[0159] By extracting the optimal solution in JSON format from the database and using the Plotly.js tool to display scheduling priority and response time, users can see the processing order of tasks in real time, and through visualization, 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 intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A field injury assessment dispatch system based on a deep learning model, characterized in that: Comprise, The acquisition and calculation module combines fuzzy logic and slope analysis to calculate the comprehensive score, extract time sequence and image features, fuse them through the multi-head attention mechanism, adjust the injury probability combined with the comprehensive score, and obtain the light, moderate and severe injury evaluation results; The multi-source data includes vital sign data, injury image data and location information data; The optimization module uses the ant colony algorithm to optimize the ambulance path combined with the evaluation results, constructs the Actor-Critic network to optimize the ant colony algorithm parameters, and uses the genetic algorithm for global optimization, then feeds back to the ant colony algorithm for iteration to obtain the optimal solution; The conversion and transmission module converts the optimal solution and transmits it through the MQTT protocol and real-time delivery; The storage and display module stores the data through the database and displays it through the visual interface; The acquisition and calculation module combines fuzzy logic and slope analysis to calculate the comprehensive score includes the following steps: Respectively through physiological monitoring equipment, portable X-ray machine and GPS equipment to collect the vital sign data, injury image data and location information data of the wounded at consecutive time points; The vital sign data includes the actual measurement values of heart rate, respiratory rate, blood oxygen saturation, body temperature and blood pressure; Use empirical rules to set corresponding threshold values for each data in the vital sign data; The corresponding threshold values 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; Use fuzzy logic rules to calculate the data reliability score of each data in the vital sign data; According to each data in the vital sign data, calculate the change rate of each data and the corresponding data at adjacent time points, define the data slope of the corresponding data, and use the maximum operation to select the maximum data slope from the slope of the corresponding data, and calculate the data slope reliability score; Use the logical operation formula to combine the reliability score and the slope reliability score to calculate the final data reliability score, and use the weighted average formula to combine the final data reliability score 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; The extraction of time sequence and image features, after fusion through the multi-head attention mechanism, combined with the comprehensive score to adjust the injury probability, and obtain the light, moderate and severe injury evaluation results include the following steps: After using the deep convolutional neural network architecture to build a feature extraction model, input the vital sign data and injury image data into the feature extraction model respectively, obtain the vital features, use the ReLU activation function to perform nonlinear transformation on the injury image data, and obtain the image features; Use the multi-head attention mechanism to weight and fuse the vital features and image features to obtain the fused features; Use the softmax function to convert the fused features to obtain the final injury evaluation probability; The final injury evaluation probability includes light injury evaluation probability, moderate injury evaluation probability and severe injury evaluation probability; The comprehensive score is multiplied with the probability in the final injury assessment probability by using element-by-element multiplication to obtain an adjusted probability, and after being sorted in ascending order, the largest probability is taken as the assessment result of the current wounded person, including light injury, moderate injury and severe injury; The combination of the assessment result and the use of the ant colony algorithm to optimize the ambulance path includes the following steps: Collect the positions of the ambulance and the hospital, take the position of the ambulance as the starting point, and take the positions of the wounded and the hospital as nodes; Use the map API to obtain the actual driving distance of the complete path, and use the HTTP request technology to obtain the historical speed of the ambulance in the map API to obtain the average speed; The complete path is from the position of the ambulance to the position of the wounded, and then from the position of the wounded to the position of the hospital; The ratio of the actual driving distance to the average speed is defined as the response time; The response time is divided into the response time from the position of the ambulance to the position of the wounded and the response time from the position of the wounded to the position of the hospital; The historical regression analysis method is used to set the road congestion coefficient, and the response time from the position of the ambulance to the position of the wounded is dynamically adjusted, and then the dynamic adjustment formula is used to calculate the response time from the position of the wounded to the position of the hospital, to obtain the adjusted response time from the position of the wounded to the position of the hospital, and to calculate the total response time; An empirical rule is used to set a constant, and a proportional formula is used to combine the total response time and the constant to calculate the initial pheromone, and then a proportional formula is used to combine the total response time to calculate the initial heuristic information; A grid search method is used to set a weight coefficient, and a path selection probability formula is used to calculate the path selection probability to generate a path selection probability set; When the path selection probability of the severely wounded person exists in the path selection probability set, the path selection probability of the severely wounded person is directly taken as the path and dynamically adjusted, and the formula is: ； In the formula, Indicates the adjusted number Complete path of seriously injured patients The probability of path selection. Indicates the first The probability of severe injury assessment for each seriously injured patient. Indicates the first Complete path of seriously injured patients The probability of path selection; When there is one seriously injured person in the path selection probability set, the adjusted path selection probability is sorted in ascending order, and the largest path selection probability is selected as the path. When there is no seriously injured person in the path selection probability set, the path is selected by roulette. The path selection probability and the total response time of the selected path are normalized, and an expert scoring method is used to set a decay coefficient and a weight coefficient to define an objective function to minimize the response time and maximize the rescue efficiency; A theoretical analysis method is used to set a volatility coefficient to calculate the target function value, which is iteratively updated, and the volatility coefficient is used to control the evaporation rate of the pheromone to output the optimal solution; The optimal solution refers to the optimal path that minimizes the response time and maximizes the rescue efficiency; The steps for constructing the Actor-Critic network to optimize the ant colony algorithm parameters include: Using normalized state space to analyze response time, rescue efficiency, and weighting coefficients in the optimal solution. and The volatile coefficient is scaled to obtain a scaled dataset; The Actor network and Critic network in the A2C algorithm are constructed using a multi-layer perceptron architecture; The scaled data set is input into the Critic network to output the value of the data set, and the n-step return method is used to combine the value of the data set to calculate the cumulative return, and the difference between the cumulative return and the data set value is defined as the advantage function, the Delphi method is used to set the feedback threshold, the advantage function is compared with the feedback threshold to obtain the updated Actor network, and the scaled data set is input into the updated Actor network to output a three-dimensional action vector; The three-dimensional motion vector includes a weight coefficient and and a volatility coefficient; The three-dimensional action vector is smoothed using a double exponential smoothing filter, and the smoothed three-dimensional action vector is written into the Redis key value, which is read in real time through the ACO scheduling module to trigger the path selection probability recalculation. The performance indicators of the ACO scheduling module are detected using a monitoring module. When the comprehensive objective function value of the re-computed 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 rolled-back optimal solution are used as new parameters. Otherwise, the re-computed parameters are used as final parameters. After global optimization using the genetic algorithm, the feedback is fed back to the ant colony algorithm for iteration to obtain an optimal solution, including the following steps: The optimal path in the optimal solution is discretely coded using a hybrid coding technique, and the weight coefficients in the optimal solution are continuously coded using a real coding technique and The optimal path in the optimal solution is discretely coded using a hybrid coding technique, and the weight coefficients in the optimal solution are continuously coded using a real coding technique The objective function value is calculated using the objective function of minimizing the response time and maximizing the rescue efficiency, and after iterative optimization, the optimal path is output. The optimal path is repaired using a graph theory algorithm in an open source library, and after pheromone update using a pheromone update rule, the updated pheromone is decoded using a decoding library, and the decoded pheromone is re-input into the ACO scheduling module for iteration to obtain a new optimal solution. 2.The on-site injury assessment and dispatching system based on a deep learning model, as claimed in claim 1, wherein: The vital sign data, injury image data, and location information data are collected, and preprocessing operations including KNN interpolation, outlier detection, and standardization and normalization are performed. 3.The on-site injury assessment and dispatching system based on a deep learning model, as claimed in claim 2, wherein: The optimal solution is converted into JSON format, and the JSON format optimal solution is transmitted to the edge computing device using the MQTT protocol, and then the JSON format optimal solution is issued to the dispatcher through the edge computing device. 4.The on-site injury assessment and dispatching system based on a deep learning model, as claimed in claim 3, wherein: The JSON format optimal solution issued by the edge computing device is stored through the database, and each optimal solution is marked with a unique ID and a timestamp, and the JSON format optimal solution after marking the timestamp is sorted in ascending order through the database. 5.The on-site injury assessment and dispatching system based on a deep learning model, as claimed in claim 4, wherein: The JSON format optimal solution in the database is extracted, and the Plotly.js tool is used to display the dispatching priority and response time.

Images