Federal learning-based intelligent transfer and self-rescue guidance aid decision-making method and system

Abstract

The invention provides an intelligent transfer and self-rescue guidance aid decision-making method and system based on federal learning, and relates to the technical field of intelligent medical treatment, and the method comprises the steps: obtaining emergency help information, constructing an emergency rescue knowledge graph, extracting physiological feature parameters to recognize diseases, and generating a self-rescue scheme; based on the disease and position of the user, medical resources are retrieved, and an ant colony algorithm is adopted to solve the multi-objective optimization function to obtain an optimal transfer path; and inputting the transfer path and the self-rescue scheme into a federal learning model to obtain a comprehensive decision. Accurate and efficient emergency rescue and self-rescue guidance are realized, and the medical resource allocation efficiency and the first-aid success rate are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical technology, and in particular to an intelligent transfer and self-rescue guidance and decision-making assistance method and system based on federated learning. Background Technology

[0002] With social development and increased awareness of healthcare, the public's demand for emergency medical rescue services is growing. When sudden illness or accidental injury occurs, patients or those around them often need to quickly obtain professional self-rescue guidance and be rapidly transported to the most suitable medical institution. Traditional emergency medical systems mainly rely on telephone call centers to provide rescue guidance and dispatch ambulances. However, with the development of artificial intelligence, big data, and mobile internet technologies, intelligent emergency medical decision-making systems have become a research hotspot.

[0003] In recent years, artificial intelligence-based emergency medical technology has made some progress. Some systems collect patients' physiological parameters and symptom descriptions, and combine them with medical knowledge bases to provide preliminary diagnoses and self-rescue suggestions; intelligent navigation technology is applied to ambulance dispatching and optimal route planning; and knowledge graph technology is used to build medical knowledge systems to support medical decision-making. However, organically integrating these technologies and solving the problem of medical data security protection has become a major challenge facing current emergency medical decision-making systems. Summary of the Invention

[0004] This invention provides a method and system for intelligent transportation and self-rescue guidance and decision-making based on federated learning, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a method for intelligent transportation and self-rescue guidance-assisted decision-making based on federated learning, comprising: Obtain emergency help requests sent by user terminals; Knowledge entities and relationships are extracted from the medical emergency knowledge base and historical rescue case database to construct an emergency rescue knowledge graph; based on the emergency help information, physiological characteristic parameters are extracted and the user's current symptoms are identified, path search is performed in the emergency rescue knowledge graph, and a self-rescue guidance plan is generated in combination with the on-site environment; Based on the user's current symptoms and location, matching medical resource needs are retrieved, multiple candidate medical institutions are identified, and a multi-objective optimization function is constructed by combining the current traffic status of each road segment. The ant colony algorithm is used to solve the problem and obtain the optimal transfer path. The optimal transfer route and the self-rescue guidance plan are input into the federated learning model to obtain a comprehensive decision, which is then sent to the user terminal.

[0006] Knowledge entities and relationships are extracted from the medical emergency knowledge base and historical rescue case database to construct an emergency rescue knowledge graph, including: The medical emergency knowledge base retrieves treatment guidelines and emergency care instructions, generates structured medical knowledge text, constructs a medical domain word vector space, maps medical terms in the medical knowledge text to semantic vectors, extracts textual features from the semantic vectors, and identifies medical knowledge entities based on the textual features. The medical knowledge entities are divided into multiple entity categories according to a medical professional tree classification system. The correlation degree between the entity categories is calculated. Entity pairs with a correlation degree greater than the correlation threshold are marked as candidate relationship pairs. The candidate relationship pairs are screened by confidence evaluation to obtain the initial knowledge graph. New types of knowledge entities and relationships are identified from the historical rescue case database, and the semantic compatibility of the new entities and relationships with the existing knowledge in the initial knowledge graph is verified. The verified new entities and relationships are then integrated into the initial knowledge graph to obtain the emergency rescue knowledge graph.

[0007] Based on the emergency assistance information, physiological characteristic parameters are extracted and the user's current symptoms are identified. A path search is performed within the emergency rescue knowledge graph, and a self-rescue guidance plan is generated based on the on-site environment, including: From the emergency help information, extract the user's symptom feature vector and physiological feature data, perform time alignment and synchronization calibration on the physiological feature data, and generate physiological feature vector; Based on the symptom feature vector and the physiological feature vector, and combined with the symptom-disease association network in the emergency rescue knowledge graph, the probability of symptom identification of the user at the current moment is calculated, and the user's current symptom is determined. The current symptom is taken as the starting state node, the target recovery state is determined as the target state node, the initial candidate rescue path from the starting state node to the target state node is calculated, and the initial candidate rescue path is pruned based on the upper limit of resource consumption, the upper limit of time consumption and environmental adaptability constraints to obtain the optimized candidate rescue path. Calculate the feasibility score for each optimized candidate rescue path, select the candidate rescue path with the highest feasibility score, decompose it according to the execution steps, identify the target rescue action, and obtain a self-rescue guidance plan.

[0008] Calculate the initial candidate rescue path from the starting state node to the target state node. Prune the initial candidate rescue path based on resource consumption limits, time consumption limits, and environmental adaptability constraints to obtain an optimized candidate rescue path, including: Starting from the initial state node, the adjacent state nodes are traversed sequentially through depth search. The node feature vectors of the adjacent state nodes are extracted, and their reachability scores to the target state node are calculated. The adjacent state nodes with reachability scores greater than the reachability threshold are selected as the next search nodes, generating multiple initial candidate rescue paths. For each initial candidate rescue path, calculate the total resource consumption of all its state nodes, and calculate the execution time and transition delay time of each state node to obtain the total execution time. Based on the current environment, calculate the environment adaptability score. When any one of the indicators of the total resource consumption, the total execution time, and the environment adaptability score exceeds the default limit threshold, mark the candidate rescue path as an infeasible path to obtain multiple feasible candidate rescue paths. Calculate the similarity of rescue actions among the feasible candidate rescue paths, and merge the rescue paths with a similarity of rescue actions higher than the path merging threshold to obtain the optimized candidate rescue paths.

[0009] Based on the user's current symptoms and location, matching medical resource needs are retrieved, multiple candidate medical institutions are identified, and a multi-objective optimization function is constructed by combining the current traffic status of each road segment. This function is then solved using an ant colony optimization algorithm to obtain the optimal transfer path, including: Based on the user's current symptoms, the importance weights of various medical resources are calculated, and combined with the standard configuration requirements of various medical resources, candidate medical institutions that meet the basic requirements are selected. The system obtains information on bed occupancy, expert availability, and equipment usage of the candidate medical institutions, calculates resource availability scores, and obtains a comprehensive score for each candidate medical institution based on the resource availability scores and transport distance. The congestion intensity of road segments is calculated based on road traffic flow data and average vehicle speed data, and the real-time travel time of each road segment is predicted by combining traffic light timing data; based on the comprehensive score and the real-time travel time, a multi-objective optimization function is constructed, and the maximum allowable transfer time is set. Multiple search units are set in the road grid. The transfer path is determined based on the path pheromone concentration and dynamic heuristic information. The pheromone update amount is adjusted according to the multi-objective optimization function, and the optimal transfer path is obtained through iterative optimization.

[0010] Multiple search units are set up in the road grid. The transport path is determined based on the path pheromone concentration and heuristic strength. The pheromone update amount is adjusted according to the multi-objective optimization function, and the optimal transport path is obtained through iterative optimization, including: Multiple search units are randomly deployed in the road grid. Each search unit contains the current location, direction of movement, and visited paths. The heuristic intensity of each search unit is calculated by combining the historical heuristic intensity and the real-time travel time of each road segment. An initial pheromone concentration is randomly distributed in the road grid. The product of the heuristic intensity and the initial pheromone concentration is used as the path selection probability of the search unit. The next transfer direction is determined based on the path selection probability, and the transfer path formed by each search unit in the current iteration cycle is obtained. Calculate the multi-objective optimization function value for each transport path, map it to the pheromone increment, and update the pheromone concentration in the road grid according to the pheromone evaporation rate; Calculate the distribution entropy of the multi-objective optimization function value. When the change in distribution entropy over multiple consecutive iterations is less than the entropy change threshold, select the transfer path with the largest multi-objective function value from the transfer paths in the current iteration cycle as the optimal transfer path.

[0011] A second aspect of this invention provides an intelligent transfer and self-rescue guidance and decision-making support system based on federated learning, comprising: The first unit is used to obtain emergency help information sent by user terminals; The second unit is used to extract knowledge entities and relationships from the medical emergency knowledge base and the historical rescue case database to construct an emergency rescue knowledge graph; based on the emergency help information, extract physiological characteristic parameters and identify the user's current symptoms, perform path search in the emergency rescue knowledge graph, and generate a self-rescue guidance plan in combination with the on-site environment; The third unit is used to retrieve matching medical resource needs based on the user's current symptoms and location, determine multiple candidate medical institutions, and construct a multi-objective optimization function by combining the current traffic status of each road segment, and solve it using the ant colony algorithm to obtain the optimal transfer path. The fourth unit is used to input the optimal transfer path and the self-rescue guidance plan into the federated learning model to obtain a comprehensive decision, and then send the comprehensive decision to the user terminal.

[0012] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] The beneficial effects of this application are as follows: This invention constructs an emergency rescue knowledge graph. By extracting knowledge entities and relationships from a medical emergency knowledge base and a historical rescue case database, it can quickly identify a user's current symptoms and generate targeted self-rescue guidance plans, thereby improving the efficiency and scientific nature of self-rescue in emergency situations and effectively reducing the risk of the condition worsening.

[0015] This invention uses an ant colony algorithm to construct a multi-objective optimization function, which comprehensively considers the user's symptoms, location information, and real-time traffic status of each road segment. It can scientifically determine the optimal transfer route, greatly shorten the emergency response time, and improve the efficiency of medical resource allocation.

[0016] This invention utilizes a federated learning model to make comprehensive decisions on self-rescue guidance plans and transport routes, ensuring patient information security while achieving optimal matching of medical resources and patient needs. It forms a complete closed-loop decision support system from self-rescue guidance to medical transport, significantly improving the overall quality of emergency services. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the intelligent transportation and self-rescue guidance and decision-making assistance method based on federated learning, as described in an embodiment of the present invention. Figure 2 A schematic diagram illustrating the process of generating an emergency rescue and self-rescue guidance plan. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a flowchart illustrating the intelligent transportation and self-rescue guidance and decision-making assistance method based on federated learning, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Obtain emergency help requests sent by user terminals; Knowledge entities and relationships are extracted from the medical emergency knowledge base and historical rescue case database to construct an emergency rescue knowledge graph; based on the emergency help information, physiological characteristic parameters are extracted and the user's current symptoms are identified, path search is performed in the emergency rescue knowledge graph, and a self-rescue guidance plan is generated in combination with the on-site environment; Based on the user's current symptoms and location, matching medical resource needs are retrieved, multiple candidate medical institutions are identified, and a multi-objective optimization function is constructed by combining the current traffic status of each road segment. The ant colony algorithm is used to solve the problem and obtain the optimal transfer path. The optimal transfer route and the self-rescue guidance plan are input into the federated learning model to obtain a comprehensive decision, which is then sent to the user terminal.

[0021] In one optional implementation, knowledge entities and relationships are extracted from a medical emergency knowledge base and a historical rescue case database to construct an emergency rescue knowledge graph, including: The medical emergency knowledge base retrieves treatment guidelines and emergency care instructions, generates structured medical knowledge text, constructs a medical domain word vector space, maps medical terms in the medical knowledge text to semantic vectors, extracts textual features from the semantic vectors, and identifies medical knowledge entities based on the textual features. The medical knowledge entities are divided into multiple entity categories according to a medical professional tree classification system. The correlation degree between the entity categories is calculated. Entity pairs with a correlation degree greater than the correlation threshold are marked as candidate relationship pairs. The candidate relationship pairs are screened by confidence evaluation to obtain the initial knowledge graph. New types of knowledge entities and relationships are identified from the historical rescue case database, and the semantic compatibility of the new entities and relationships with the existing knowledge in the initial knowledge graph is verified. The verified new entities and relationships are then integrated into the initial knowledge graph to obtain the emergency rescue knowledge graph.

[0022] In constructing an emergency rescue knowledge graph, it is necessary to obtain treatment guidelines and emergency rescue instructions from the medical emergency knowledge base, and access multiple authoritative medical data sources, such as the standardized treatment document library of emergency medical centers and emergency rescue guidelines issued by various specialized societies. These resources are collected through web crawling technology, and long texts are segmented into semantically complete small paragraphs using document segmentation technology in natural language processing. For unstructured PDF documents, optical character recognition technology is applied for digital conversion. At the same time, regular expressions and pattern matching methods are used to extract document titles, chapters and key paragraphs to form an initial knowledge text set.

[0023] To construct a word vector space for the medical field, a distributed representation learning method is employed. In practice, medical texts are segmented using a medical-specific word segmentation dictionary to ensure the integrity of professional terminology. For example, "ventricular fibrillation after acute myocardial infarction" is treated as a single term rather than split into smaller parts. Word vector training utilizes a 300-dimensional vector space with a window size of 5, a negative sampling count of 5, and 15 iterations. The generated word vectors effectively capture the semantic relationships between medical terms. In practical applications, the cosine similarity between "myocardial infarction" and "chest pain" is 0.82, while the similarity with "fracture" is only 0.14, demonstrating that the vector space accurately reflects the semantic distance between medical concepts.

[0024] In the medical knowledge entity recognition stage, a dual strategy combining rule-based and deep learning is employed. The rule-based part includes medical terminology dictionary matching and regular expression pattern recognition, which can quickly capture standardized medical entities. For example, matching "body temperature > 39°C" is recognized as the entity of "high fever". The deep learning part adopts a bidirectional long short-term memory network structure, with word vectors as the input layer, 128 hidden layer nodes, and conditional random field for sequence labeling in the output layer. The model achieves an F1 score of 0.92 on the validation set, and the entity recognition results include various entities such as "cardiopulmonary injury", "traumatic shock", "dyspnea", "loss of consciousness", "chest compressions", and "endotracheal intubation".

[0025] Based on the international medical professional classification system, the identified entities are divided into multiple categories, specifically adopting a four-level classification structure. The first level includes six major categories: diseases, symptoms, treatments, drugs, equipment, and first aid skills. The second level is further subdivided into 43 subcategories, such as cardiovascular diseases and respiratory diseases under diseases. The third level is further refined into 217 specific categories. The fourth level is the specific entity. Through this hierarchical classification, "cardiopulmonary resuscitation" can be correctly classified into the category of "first aid skills - basic life support techniques - cardiac resuscitation".

[0026] The inter-entity correlation calculation adopts a multi-dimensional feature fusion method. Preliminary correlation is constructed by the co-occurrence frequency of entities. For example, "cardiac arrest" and "cardiopulmonary resuscitation" co-occur 723 times in the corpus. The cosine similarity of the vectors between entities is calculated to capture semantic correlation. Combined with entity type constraint rules, for example, the correlation weight between "symptom" type entities and "disease" type entities is increased by 0.2. The above features are fused to calculate the comprehensive correlation score. The correlation threshold is set at 0.65. Entity pairs with a value greater than this threshold are marked as candidate relation pairs.

[0027] In the candidate relation pair screening phase, a relation classifier assesses the confidence level of each candidate entity relation pair. The relation classifier receives the contextual syntactic information of the entity pair and outputs the relation type and confidence level. For example, for the entity pair "chest compressions" and "cardiac arrest," the system determines the relation type as "treatment method" with a confidence level of 0.94, and therefore retains this relation. Candidate relations with a confidence level below 0.7 are filtered out to ensure the accuracy of the initial knowledge graph. Through the above steps, the initial knowledge graph contains 5,823 entities and 12,746 relations.

[0028] To continuously expand and improve the knowledge graph, a historical rescue case database is accessed for new knowledge discovery. This database contains over 10,000 real emergency rescue records, each including information such as injury description, on-site treatment measures, transfer process, and treatment results. Entity recognition algorithms are applied to the case texts to discover potential new entities. For newly identified entities, such as "explosive decompression injury," a special injury caused by a new type of disaster, the semantic similarity between the entity and existing entities in the knowledge graph is calculated. When the similarity with the closest entity, "shockwave injury," is 0.76, it is considered a potential new entity.

[0029] In the new knowledge verification phase, semantic compatibility testing ensures that the new knowledge is consistent with the existing knowledge system. Semantic compatibility assessment includes three stages: entity type consistency check, relational logic conflict detection, and expert rule verification. New entities and relations that pass verification are integrated into the existing knowledge graph, and the association weights between related entities are updated. For example, the newly discovered relation "mask ventilation" is an "emergency treatment method" for "airway obstruction," which does not conflict with existing knowledge and has passed verification; therefore, it has been successfully integrated into the knowledge graph.

[0030] The final emergency rescue knowledge graph contains 6,205 medical knowledge entities and 14,362 relationships, covering the main areas of emergency medicine and providing comprehensive knowledge support for emergency rescue decision-making.

[0031] Figure 2 A flowchart illustrating a method for generating an emergency rescue self-rescue guidance plan. In one optional implementation, based on the emergency request information, physiological characteristic parameters are extracted and the user's current symptoms are identified. A path search is performed within the emergency rescue knowledge graph, and a self-rescue guidance plan is generated by combining the on-site environment, including: From the emergency help information, extract the user's symptom feature vector and physiological feature data, perform time alignment and synchronization calibration on the physiological feature data, and generate physiological feature vector; Based on the symptom feature vector and the physiological feature vector, and combined with the symptom-disease association network in the emergency rescue knowledge graph, the probability of symptom identification of the user at the current moment is calculated, and the user's current symptom is determined. The current symptom is taken as the starting state node, the target recovery state is determined as the target state node, the initial candidate rescue path from the starting state node to the target state node is calculated, and the initial candidate rescue path is pruned based on the upper limit of resource consumption, the upper limit of time consumption and environmental adaptability constraints to obtain the optimized candidate rescue path. Calculate the feasibility score for each optimized candidate rescue path, select the candidate rescue path with the highest feasibility score, decompose it according to the execution steps, identify the target rescue action, and obtain a self-rescue guidance plan.

[0032] In this embodiment, emergency help requests containing user physiological information are received. This information comes from various channels, such as smart wearable devices, mobile terminal devices, or smart medical devices. Emergency help requests typically contain multimodal information, including text descriptions, voice recordings, and device monitoring data. The received emergency help requests are parsed and preprocessed to convert information of different formats into a standardized data format.

[0033] For symptom descriptions in text form, word segmentation technology is used to divide the text into word units. For example, "I feel chest tightness, shortness of breath, and rapid heartbeat" is segmented into word units such as "I," "feel," "chest tightness," "shortness of breath," "and," "heartbeat," and "rapid." Part-of-speech tagging technology is used to identify symptom-related words such as "chest tightness," "shortness of breath," and "rapid heartbeat." A medical symptom knowledge base is established, containing common symptom terms, synonyms, and related descriptions. This knowledge base includes standard names of symptoms, common expressions, and corresponding feature values. For example, for the symptom "chest tightness," the knowledge base stores multiple expressions such as "chest tightness," "chest pressure," and "chest pressure sensation," and maps them to a unified symptom identifier. A text matching algorithm is used to match the identified symptom descriptions with the standard symptoms in the knowledge base, generating a standardized set of symptom features.

[0034] For each identified symptom, its severity, duration, frequency, and other attribute information are extracted. For example, from "severe headache lasting two hours", the symptom is extracted as "headache", the severity is "severe", and the duration is "two hours". These attributes are associated with the corresponding symptoms to construct a structured symptom description. Based on the extracted symptoms and their attributes, a symptom feature vector is constructed. This vector is a multi-dimensional array, with each dimension corresponding to a specific symptom or attribute. The identified symptoms are mapped to the corresponding dimensions of the vector. Values ​​are assigned to each dimension based on the presence, severity, and other attributes of the symptom. For example, in a feature space containing 100 common symptoms, if a user is identified as having "chest tightness" with a severity of "moderate", then a value of 0.6 is assigned to the corresponding "chest tightness" dimension (assuming the severity range is 0-1).

[0035] For physiological characteristic data, various physiological indicators of users, such as heart rate, blood pressure, blood oxygen saturation, and respiratory rate, are extracted from emergency help information. These data come from different devices and are collected at different times and frequencies. In order to accurately assess the user's health status, it is necessary to perform time alignment and synchronization calibration on these heterogeneous data.

[0036] The system checks the timestamp formats of each data source and converts different formats to a standard time format. For example, it converts "2025-09-15 08:30:45" and "15 / 09 / 2025 08:30:45" to the ISO 8601 format "2025-09-15T08:30:45Z". It also identifies the sampling frequency of each data source; for example, heart rate data is sampled once per second, while blood pressure data is sampled once per minute. A unified time reference is selected, such as using the minimum sampling interval as the reference, to create a uniformly distributed time point sequence. For data with low sampling frequencies, interpolation algorithms are used to fill in missing time points. Specifically, for linearly changing physiological indicators, such as body temperature, linear interpolation is used; for indicators with large fluctuations, such as heart rate, the system uses cubic spline interpolation to maintain the smoothness of the curve.

[0037] Considering the time deviations between different devices, the time series patterns of the data, such as the correlation between heart rate and respiratory rate, are analyzed to identify the time offset between data sequences. The time delay between different physiological index sequences is calculated using the maximum cross-correlation method, and the timestamps are adjusted accordingly to ensure that data from different sources are aligned in time. Outliers and noise in the data are detected and processed. For example, for the heart rate data sequence [75, 76, 150, 74, 75], 150 is identified as an outlier and replaced with the median of the surrounding values ​​using median filtering.

[0038] Based on time-aligned and synchronized physiological data, a physiological feature vector is constructed. This vector contains multiple dimensions, each corresponding to a physiological indicator or its statistical feature. For time-series physiological data, statistical features such as mean, standard deviation, and peak value are extracted to capture the distribution characteristics and trends of the data. For example, for heart rate data, not only is the current heart rate value recorded, but also the average heart rate, heart rate variability, and trend over the past 5 minutes are calculated. These features are organized into a structured physiological feature vector to facilitate subsequent health status assessment and urgency determination.

[0039] Based on symptom feature vectors and physiological feature vectors, combined with the symptom-disease association network in the emergency rescue knowledge graph, the probability of identifying the user's current symptom is calculated to determine the user's current symptom. In this embodiment, the emergency rescue knowledge graph contains 5,000 common emergency symptoms, each associated with a specific combination of symptoms and physiological feature patterns. A deep neural network model is used, taking the symptom feature vectors and physiological feature vectors as input, and outputting the type of symptom the user has and its probability value. In the specific implementation, the two feature vectors are processed through their respective feature extraction layers, fused, and the probability distribution of each symptom is output through a softmax layer. For example, when the user's symptom features include "severe chest pain, cold sweats, nausea," and the physiological features show "rapid heart rate (110 beats / min), elevated blood pressure (150 / 95 mmHg)," the probability distribution of "acute myocardial infarction: 85.6%, pulmonary embolism: 10.2%, gastric ulcer: 4.2%" is output, determining the user's current symptom as acute myocardial infarction.

[0040] Using the current symptom as the starting state node and the target recovery state as the target state node, initial candidate rescue paths from the starting state node to the target state node are calculated. In the emergency rescue knowledge graph, each symptom state and its corresponding rescue measure form a connection edge. Each edge is labeled with the resources, time cost, and environmental requirements required to execute the rescue measure. An improved search algorithm is used to generate multiple candidate paths from the starting state to the target state in the knowledge graph. Taking acute myocardial infarction as an example, the following candidate paths are generated: Path 1: "Take aspirin - take nitroglycerin sublingually - call emergency services - maintain a semi-recumbent position - continuously monitor vital signs"; Path 2: "Cardiopulmonary resuscitation - use an automated external defibrillator - call emergency services - continuously monitor vital signs"; Path 3: "Take nitroglycerin sublingually - lie flat - take deep breaths to relax - call emergency services - continuously monitor vital signs".

[0041] When pruning the initial candidate rescue routes, three constraints are considered: resource consumption limit, time consumption limit, and environmental adaptability constraint. The resource consumption limit refers to the limited rescue supplies currently available to the user, such as the lack of an automated external defibrillator (AED) in a typical home. The time consumption limit refers to the golden rescue window for certain serious illnesses. The environmental adaptability constraint refers to whether the current environment is suitable for performing specific rescue measures, such as the inability to perform standard CPR in a confined space. Each candidate route is evaluated based on the user's location information, environmental description, and available resource list, eliminating routes that do not meet the constraints. In the example above, assuming the user is at home without a defibrillator, route 2 is pruned due to resource constraints; if the user describes a crowded environment where lying flat is impossible, route 3 is pruned due to environmental constraints; ultimately, only route 1 is retained as the optimized candidate rescue route.

[0042] When calculating the feasibility score of each optimized candidate rescue route, three aspects are comprehensively considered: success probability, execution difficulty, and risk factors. The success probability is calculated based on the success rate of the route in historical rescue cases. The execution difficulty assesses the operational complexity of each step in the route for ordinary people. Risk factors consider the negative impacts generated during the execution process. Each factor is assigned a weight, and the weighted sum is used to obtain the comprehensive feasibility score. The candidate rescue route with the highest feasibility score is selected. The selected rescue route is decomposed according to the execution steps, the target rescue actions are identified, and a self-rescue guidance plan is generated. Taking route 1 as an example, the system decomposes it into the following specific action steps: 1) Take out an aspirin tablet (if available) and chew it; 2) Take out a nitroglycerin tablet and place it under the tongue; 3) Call the emergency number and clearly describe the symptoms; 4) Adjust to a semi-recumbent position with the upper body raised 30 degrees; 5) Maintain steady breathing and record the heart rate every 5 minutes. The level of detail in the guidance plan is adjusted according to the user's actual situation. For example, if the user does not understand what "semi-recumbent position" means, more detailed posture adjustment instructions and illustrations will be provided.

[0043] In practical applications, suppose a user sends a distress message: "I suddenly feel severe chest pain, like being crushed by a heavy object, accompanied by cold sweats and nausea. I'm at home." Simultaneously, the wearable device transmits physiological data: heart rate 112 bpm, blood pressure 152 / 96 mmHg. The system identifies the user as having an acute myocardial infarction and generates a self-rescue guide: "Immediately chew an aspirin (if available); place a nitroglycerin tablet under the tongue (if available); immediately call emergency services; adopt a semi-recumbent position, elevating the upper body approximately 30 degrees with a pillow; remain quiet and avoid any activity; continuously record changes in your heart rate and symptoms, and wait for rescue personnel to arrive." This plan considers the user's home environment, optimizes initial self-rescue measures for acute myocardial infarction, and improves the user's survival rate while waiting for professional assistance.

[0044] In one optional implementation, an initial candidate rescue path is calculated from the starting state node to the target state node. Based on resource consumption limits, time consumption limits, and environmental adaptability constraints, the initial candidate rescue path is pruned to obtain an optimized candidate rescue path, including: Starting from the initial state node, the adjacent state nodes are traversed sequentially through depth search. The node feature vectors of the adjacent state nodes are extracted, and their reachability scores to the target state node are calculated. The adjacent state nodes with reachability scores greater than the reachability threshold are selected as the next search nodes, generating multiple initial candidate rescue paths. For each initial candidate rescue path, calculate the total resource consumption of all its state nodes, and calculate the execution time and transition delay time of each state node to obtain the total execution time. Based on the current environment, calculate the environment adaptability score. When any one of the indicators of the total resource consumption, the total execution time, and the environment adaptability score exceeds the default limit threshold, mark the candidate rescue path as an infeasible path to obtain multiple feasible candidate rescue paths. Calculate the similarity of rescue actions among the feasible candidate rescue paths, and merge the rescue paths with a similarity of rescue actions higher than the path merging threshold to obtain the optimized candidate rescue paths.

[0045] In this specific embodiment, starting from the initial state node, a depth-first search algorithm is used to traverse adjacent state nodes. For each adjacent state node, its node feature vector is extracted, including key features such as symptom relief degree, vital sign improvement degree, patient self-rescue difficulty, and treatment response time. The reachability score from this node to the target state node is calculated, and this score is jointly determined by the symptom improvement speed, vital sign stability, and treatment effectiveness. For example, in a self-rescue scenario for an acute asthma attack, node A represents the "using bronchodilator inhalation" state. Its symptom relief degree is 0.7 (out of 1), vital sign improvement degree is 0.6 (out of 1), and treatment response time is 3 minutes, so its reachability score is 0.75. The reachability score is compared with a preset reachability threshold (e.g., 0.6), and nodes with scores higher than the threshold are selected as the next search nodes. This process is repeated to generate multiple initial candidate self-rescue guidance paths.

[0046] For each generated initial candidate self-rescue guidance path, a three-pronged constraint assessment is performed to calculate the total resource consumption, including drug consumption, medical device usage, and physiological energy consumption. Taking an asthma self-rescue path as an example, this path contains 5 nodes: "use bronchodilator inhalation", "adopt a sitting position and lean forward", "perform pursed-lip breathing", "use bronchodilator inhalation again", and "remain calm and monitor symptoms". The drug consumption of each node is [1 dose, 0, 0, 1 dose, 0] units, so the total bronchodilator consumption of this path is 2 doses. If the preset upper limit for bronchodilator use is 3 doses, then this path is feasible in terms of drug consumption constraints. The total execution time is calculated, including the execution time at each state node and the transition delay between adjacent nodes. Taking the above path as an example, assuming the execution times of the 5 nodes are [2, 5, 8, 2, 10] minutes and the node transition delays are [1, 1, 2, 1] minutes, the total execution time is 27 + 5 = 32 minutes. The maximum time consumption is 45 minutes, so the path is feasible in terms of time constraints. An environmental adaptability score is calculated based on the current environment, including a comprehensive assessment of factors such as whether there is assistance from others, whether medication is available, and whether the spatial conditions are suitable. For example, if the current environment is a solitary state, bronchodilators are carried, and the location is a well-ventilated indoor space, the environmental adaptability score of this path is 0.85 (out of 1). The environmental adaptability threshold is set to 0.6, so the path is also feasible in terms of environmental constraints.

[0047] Through the above triple constraint assessment, candidate paths exceeding the default limit threshold are marked as infeasible paths, thus obtaining multiple feasible candidate self-rescue guidance paths. To further optimize path selection, the similarity of rescue actions among these feasible candidate paths is calculated. Action similarity is determined by comparing the proportion of similar self-rescue action sequences in different paths. For example, path A contains the action sequence [inhaling a bronchodilator, sitting and leaning forward, performing pursed-lip breathing, inhaling a bronchodilator again, remaining calm and monitoring symptoms], and path B contains the action sequence [inhaling a bronchodilator, sitting and leaning forward, performing abdominal breathing, inhaling a bronchodilator again, remaining calm and monitoring symptoms]. Both have 4 similar actions, with a similarity of 4 / 5 = 0.8. If the path merging threshold is set to 0.7, these two paths can be merged. The merging process will retain action choices that provide better symptom relief, are easier to execute, or consume fewer resources. For example, in the example above, the system compares various indicators of the two different actions, "pursed-lip breathing" and "diaphragmatic breathing". It assumes that "pursed-lip breathing" is more complicated to perform but more effective in relieving asthma symptoms, and selects the "pursed-lip breathing" action to form the optimal combined path.

[0048] Through path merging, several optimized candidate self-rescue guidance paths are obtained. These paths not only meet the upper limits of resource consumption, time consumption, and environmental adaptability, but also undergo optimization by merging similar paths, reducing redundant choices. For example, in the process of generating a self-rescue guidance plan for acute hypoglycemia, the system pruned the initial 15 candidate paths into 8 feasible paths after triple constraint evaluation. After similarity calculation and path merging, three optimized self-rescue guidance paths with significant differences and advantages were finally obtained: The first path focuses on "rapid glucose replenishment → sitting rest → rechecking blood glucose → eating complex carbohydrates", suitable for patients with a blood glucose meter and who are conscious; the second path focuses on "oral glucose tablets or sugary drinks → supine position → seeking help from others → eating complex carbohydrates", suitable for situations where there is no blood glucose meter but others are present; the third path focuses on "using a pre-filled glucose injector → lateral position → calling emergency services → waiting for professional rescue", suitable for critically ill patients who are confused and alone. This differentiated approach greatly enhances the relevance and effectiveness of self-rescue guidance programs, enabling the provision of the most suitable self-rescue guidance based on the patient's specific situation and environmental conditions.

[0049] In practical applications, various thresholds are dynamically adjusted according to different disease scenarios. For example, in emergency situations, the upper limit of time consumption can be appropriately relaxed; in situations of resource scarcity, the resource consumption threshold can be lowered; and in extreme environments, the environmental adaptability requirements can be increased. This dynamic adjustment mechanism makes rescue path calculation more flexible and practical, capable of handling various complex rescue scenarios.

[0050] In one optional implementation, based on the user's current symptoms and location, matching medical resource needs are retrieved, multiple candidate medical institutions are identified, and a multi-objective optimization function is constructed by combining the current traffic status of each road segment. This function is then solved using an ant colony algorithm to obtain the optimal transfer path, including: Based on the user's current symptoms, the importance weights of various medical resources are calculated, and combined with the standard configuration requirements of various medical resources, candidate medical institutions that meet the basic requirements are selected. The system obtains information on bed occupancy, expert availability, and equipment usage of the candidate medical institutions, calculates resource availability scores, and obtains a comprehensive score for each candidate medical institution based on the resource availability scores and transport distance. The congestion intensity of road segments is calculated based on road traffic flow data and average vehicle speed data, and the real-time travel time of each road segment is predicted by combining traffic light timing data; based on the comprehensive score and the real-time travel time, a multi-objective optimization function is constructed, and the maximum allowable transfer time is set. Multiple search units are set in the road grid. The transfer path is determined based on the path pheromone concentration and dynamic heuristic information. The pheromone update amount is adjusted according to the multi-objective optimization function, and the optimal transfer path is obtained through iterative optimization.

[0051] In this specific embodiment, the system acquires the user's current symptom information, such as "acute ST-segment elevation myocardial infarction," and simultaneously obtains the user's current geographical coordinates via GPS or mobile network positioning. For different symptoms, a medical resource weight database is constructed, recording the degree of dependence of various diseases on different medical resources. Medical resources are divided into four main categories: equipment resources, drug resources, medical personnel resources, and specialist resources. Each category is further subdivided into several specific items. For example, equipment resources include electrocardiogram monitors, defibrillators, and ventilators; drug resources include thrombolytic drugs, antiplatelet drugs, and nitrates; medical personnel resources include cardiologists, critical care physicians, and interventional cardiology specialists; and specialist resources include cardiac catheterization labs and coronary care wards.

[0052] For the aforementioned "acute ST-segment elevation myocardial infarction," relevant medical resources and their initial importance values ​​were retrieved from the database. For example, the initial importance of the cardiac catheterization lab was 0.95, coronary care unit 0.90, electrocardiogram monitor 0.85, defibrillator 0.80, thrombolytic drugs 0.75, antiplatelet drugs 0.70, cardiologist 0.90, and interventional cardiologist 0.95. These initial importance values ​​were set based on extensive clinical data and expert consensus, reflecting the fundamental importance of different medical resources in treating this disease. These initial importance values ​​were dynamically adjusted according to the user's specific circumstances. Adjustment factors included the user's age, comorbidities, past medical history, and duration of illness. The adjustment mechanism employed a weighted correction method, implemented through a series of pre-defined conditional judgment rules. For example, when the user is over 75 years old, the importance of thrombolytic drugs was reduced by 0.15, and when the duration of illness exceeded 12 hours, the importance of the cardiac catheterization lab was reduced by 0.25. In this example, assuming the user is 62 years old, has a history of hypertension, and the onset time is approximately 2 hours, the adjusted importance weights of various medical resources are calculated as follows: cardiac catheterization lab 0.98, coronary care unit 0.92, electrocardiogram monitor 0.85, defibrillator 0.82, thrombolytic drugs 0.72, antiplatelet drugs 0.75, cardiologist 0.90, interventional cardiologist 0.98, etc.

[0053] After calculating the importance weights of medical resources, the next step is to determine the standard configuration requirements of medical resources. This involves retrieving the standard configuration requirements for various medical resources needed to treat "acute ST-segment elevation myocardial infarction" from the medical guidelines database. These standard configuration requirements represent the minimum resource allocation standards required to treat this disease and are based on international medical guidelines and clinical practice experience. In this example, the retrieved standard configuration requirements include: at least one cardiac catheterization lab, at least one coronary care unit bed, at least one cardiac monitor, at least one defibrillator, thrombolytic drugs covering at least three common types, antiplatelet drugs covering at least two common types, at least two cardiologists, and at least one interventional cardiologist.

[0054] By combining the importance weight of medical resources with the standard configuration requirements, and setting a 30-kilometer search radius centered on the user's location, the resource configuration information of surrounding medical institutions is extracted from the medical institution resource database. Medical institutions that meet the basic resource configuration requirements are retrieved to form a list of candidate medical institutions.

[0055] Resource status data of candidate medical institutions is obtained in real time through the hospital information system interface, including specialist bed occupancy rate, on-duty status of relevant experts, and usage status of key medical equipment. Taking Hospital A as an example, the data shows that the cardiology bed occupancy rate is 78%, there is 1 vacant cardiac catheterization lab, 3 cardiology experts on duty, and 3 vacant ICU beds. A weighted calculation method is used to generate a resource availability score, such as Hospital A's resource availability score of 0.82. The straight-line distance from the user's location to each candidate medical institution is calculated and standardized. For example, the distance from the user to Hospital A is 8.5 kilometers, and the standardized distance score is 0.72. The resource availability score and distance score are combined to calculate a comprehensive score, with resource availability weighted at 0.7 and distance weighted at 0.3, resulting in a comprehensive score of 0.79 for Hospital A.

[0056] Real-time road traffic flow data is obtained through traffic monitoring networks and road condition information services. Taking road segment "R1" as an example, the current traffic flow is 320 vehicles / hour, and the average speed of this road segment is 35 km / h. According to the preset congestion assessment model, the congestion intensity is calculated to be 0.65. At the same time, traffic signal timing data is obtained, such as the red light duration of intersection "J1" being 60 seconds and the green light duration being 40 seconds. Combining traffic flow and average speed, the real-time travel time through this road segment is predicted to be 12 minutes. Similar calculations are performed on all major road segments in the area to construct a real-time road condition network model.

[0057] A multi-objective optimization function is constructed, which includes two objectives: maximizing the comprehensive score of medical institutions and minimizing the transfer time. The maximum allowable transfer time is set to 30 minutes. In the optimization function, the comprehensive score of medical institutions has a weight of 0.6, and the transfer time has a weight of 0.4.

[0058] The road is divided into multiple grids, and search units are set in the road grids. Each unit represents an intersection or road segment node. The ant colony algorithm parameters are initialized, including the number of ants set to 50, the pheromone evaporation coefficient set to 0.5, and the number of iterations set to 100. Each ant starts from the user's position and selects the next search unit based on the path pheromone concentration and dynamic heuristic information. The heuristic information is determined by the reciprocal of the real-time passage time of the road segment.

[0059] After each ant completes its path search, the path score is calculated based on the multi-objective optimization function, and the pheromone update amount is adjusted accordingly. Paths with high scores receive more pheromone enhancements. For example, a path to Hospital A has a score of 0.85 and a pheromone increment of 0.17. Through multiple iterations, the optimal transport path scheme is gradually converged.

[0060] To illustrate with a real-world example, a user reports "severe chest pain" at coordinates (119.9876°E, 30.1234°N). The current condition is acute myocardial infarction. The highest priority medical resources are cardiac catheterization labs and cardiology specialists. Five candidate medical institutions were found, with resource availability scores of 0.82, 0.75, 0.68, 0.79, and 0.71, respectively. Considering distance, the comprehensive scores are 0.79, 0.80, 0.65, 0.82, and 0.66. Multiple transfer routes were calculated. After 50 iterations using the ant colony algorithm, the optimal route was determined to be "starting point → road R5 → road R8 → road R12 → hospital A". The estimated transfer time is 18 minutes, passing through 3 traffic lights, with a total distance of 9.7 kilometers.

[0061] The final transfer plan is displayed to users or ambulance drivers via a mobile application, including the transfer route, estimated arrival time, basic information about the target hospital, and contact information for specialists. At the same time, a notification of the patient's expected arrival is sent to the target hospital so that the medical team can prepare for the admission work in advance.

[0062] In one optional implementation, multiple search units are set in the road grid, a transport path is determined based on the path pheromone concentration and heuristic strength, the pheromone update amount is adjusted according to the multi-objective optimization function, and the optimal transport path is obtained through iterative optimization, including: Multiple search units are randomly deployed in the road grid. Each search unit contains the current location, direction of movement, and visited paths. The heuristic intensity of each search unit is calculated by combining the historical heuristic intensity and the real-time travel time of each road segment. An initial pheromone concentration is randomly distributed in the road grid. The product of the heuristic intensity and the initial pheromone concentration is used as the path selection probability of the search unit. The next transfer direction is determined based on the path selection probability, and the transfer path formed by each search unit in the current iteration cycle is obtained. Calculate the multi-objective optimization function value for each transport path, map it to the pheromone increment, and update the pheromone concentration in the road grid according to the pheromone evaporation rate; Calculate the distribution entropy of the multi-objective optimization function value. When the change in distribution entropy over multiple consecutive iterations is less than the entropy change threshold, select the transfer path with the largest multi-objective function value from the transfer paths in the current iteration cycle as the optimal transfer path.

[0063] In a road grid, multiple search units are randomly deployed. For example, 50 search units are deployed in a 100×100 urban road grid. Each search unit contains three basic attributes: current location coordinates (x, y), selectable movement direction (up, down, left, right), and a set P of visited path records. The standard deviation of historical travel time is calculated for each road segment as the base value of the heuristic intensity. For example, if the average travel time of a road segment during peak hours is 15 minutes with a standard deviation of 3 minutes, and the average travel time during off-peak hours is 5 minutes with a standard deviation of 1 minute, then the base heuristic intensity of that road segment can be set to 3 / 15 = 0.2. Simultaneously, the current traffic conditions of each road segment are collected in real time. For example, if the current travel time of a road segment is 10 minutes, then the real-time heuristic intensity of that road segment is 15 - 10 = 5 minutes. The historical heuristic intensity and the real-time heuristic intensity are combined with a weight of 0.4:0.6 to obtain the comprehensive heuristic intensity value of that road segment.

[0064] Initial pheromone concentrations are randomly distributed across road grid segments, with the initial concentration range set to a random value between 0.1 and 0.5. Taking a search unit currently positioned at (50, 50) as an example, the road segments in its four selectable directions have the following pheromone concentrations: up (0.2), down (0.3), left (0.4), and right (0.1), with corresponding heuristic intensities of up (0.6), down (0.4), left (0.5), and right (0.8), respectively. The path selection probability for each direction is calculated as the product of the pheromone concentration and the heuristic intensity, resulting in: up (0.12), down (0.12), left (0.2), and right (0.08). These product values ​​are normalized to a probability distribution: [0.23, 0.23, 0.38, 0.16]. Using the roulette wheel selection method, a random number between 0 and 1 is generated, such as 0.3. If this number is 0.3, the search unit selects the left direction as its next movement direction. Repeat this process until each search unit reaches the destination from the starting point, forming a complete transfer path.

[0065] The transfer paths formed by all search units in the current iteration cycle are evaluated, and the multi-objective optimization function value of each path is calculated. Taking a certain path as an example, the comprehensive score of its destination hospital is 0.82, and the estimated travel time is 40 minutes. After weighted summation, the multi-objective optimization function value of the path is 0.75.

[0066] Based on the multi-objective optimization function value, the pheromone increment for each path is calculated. For example, for a path with an optimization function value of 0.75, the pheromone increment is set to 0.75 multiplied by the pheromone enhancement factor of 0.1, which is 0.075. The system updates the pheromone concentration in the road grid according to a preset pheromone evaporation rate (e.g., 0.2). For example, if the current pheromone concentration of a certain road segment is 0.4, and the pheromone increment brought by the transit path passing through this segment is 0.075, then the updated pheromone concentration of this road segment is 0.4 × (1 - 0.2) + 0.075 = 0.395.

[0067] After each iteration, the distribution entropy of the multi-objective optimization function values ​​for all transit paths within the current iteration period is calculated. Assuming 50 transit paths were generated in the current iteration period, their multi-objective optimization function values ​​are divided into 10 intervals, and the number of paths within each interval is counted, such as [3, 5, 8, 12, 10, 7, 3, 1, 1, 0]. Based on this distribution, the information entropy is calculated, resulting in a distribution entropy value of 2.05 for the current iteration period. The distribution entropy values ​​for consecutive iterations are recorded. For example, if the distribution entropies for the first three iterations are 2.15, 2.08, and 2.05 respectively, then the entropy changes for consecutive iterations are 0.07 and 0.03. When the entropy changes for 10 consecutive iterations are all less than the preset entropy change threshold of 0.05, the search process is considered converged, and the iteration stops.

[0068] After the algorithm converges, the transfer path with the largest multi-objective function value is selected from the transfer path set of the current iteration cycle as the optimal transfer path. For example, if the path with the largest multi-objective function value in the current iteration cycle is 0.82, the specific path is from the starting point (10, 10) through nodes (12, 10), (12, 15), (20, 15) and finally to the destination hospital A (25, 25), with a total length of 23 kilometers and an estimated travel time of 35 minutes. This path is taken as the final output optimal transfer path and provided to the scheduling system for execution.

[0069] In practical applications, this method can quickly calculate the optimal transfer route that balances time, distance, and hospital matching based on dynamic changes in the road network and real-time adjustments to transfer needs. Tests show that compared to traditional single-objective route planning algorithms, this method can save an average of 15% of transfer time and reduce transfer distance by 20%, making it particularly suitable for scenarios such as emergency rescue vehicle dispatching that require consideration of multiple optimization objectives.

[0070] This invention also provides a specific embodiment in which feature extraction and format conversion are performed on the optimal transfer path and the self-rescue guidance plan to generate a data format suitable for input to a federated learning model. For the optimal transfer path, key features such as path length, transfer time, road conditions, and transfer risks are extracted; for the self-rescue guidance plan, key features such as the execution difficulty, resource requirements, and expected effects of each step are extracted.

[0071] Federated learning models consist of multiple local models distributed across different stakeholders. They are collaboratively trained and inferred through a secure aggregation mechanism. In medical rescue scenarios, typical participating institutions include hospitals, emergency centers, traffic management departments, and meteorological service centers. Each stakeholder trains a domain-specific rescue decision-making model based on its own expertise and data resources. For example, a hospital's local model focuses on medical resource allocation and patient assessment; an emergency center's local model focuses on ambulance dispatch and on-site emergency care; a traffic management department's local model focuses on road condition analysis and transport time prediction; and a meteorological service center's local model focuses on assessing the impact of weather on rescue operations.

[0072] The federated learning system adopts a horizontal federated learning architecture. Each participant has the same local model structure, but different training data. In this embodiment, each local model uses a deep neural network structure, including an input layer, multiple hidden layers, and an output layer. The input layer receives feature vectors of the optimal transfer path and self-rescue guidance plan; the hidden layers extract high-level features through nonlinear transformations; and the output layer generates a comprehensive decision score and decision suggestions. During training, each participant uses local data to train the model, exchanging only model parameter gradients without sharing the original data, thus protecting data security while achieving collaborative learning. Model training employs a federated averaging algorithm. The central server periodically collects the model parameter gradients uploaded by each participant, performs a weighted average, updates the global model, and distributes the updated global model parameters to each participant for iterative model optimization.

[0073] To ensure the security of the federated learning process, several technical measures are employed. These include adding random noise to the model parameter gradients to prevent the leakage of sensitive information; secure multi-party computation enabling participants to jointly compute the function without revealing their individual inputs; homomorphic encryption allowing computation on encrypted data to ensure data security during transmission; and blockchain technology recording model updates and participation to ensure transparency and traceability. In this example, the system adds random noise conforming to a Laplace distribution to the model parameter gradients, with the noise intensity parameter ε set to 0.5, ensuring model accuracy while protecting data security.

[0074] The transformed optimal transfer path feature vector [23, 0.85, 0.15, 4] and self-rescue guidance plan feature vector [0.3, 0.2, 0.7, 0.8] are input into the trained federated learning model. Forward propagation calculations are performed locally at each participant to generate local inference results. The hospital model generates the result [0.75, "It is recommended to prioritize self-rescue measures and wait for professional rescue"]; the emergency center model generates the result [0.82, "It is recommended to immediately initiate the transfer procedure"]; the transportation department model generates the result [0.78, "The transfer path is clear, it is recommended to transfer as soon as possible"]; and the meteorological center model generates the result [0.7, "The weather conditions are good, no need to adjust the transfer plan"]. These local inference results are securely transmitted to the central server for comprehensive analysis.

[0075] The central server integrates the inference results from various local models using a weighted voting mechanism. Weight allocation is based on the professional relevance and historical accuracy of each participant. In this example, the hospital model has a weight of 0.4, the emergency center model 0.3, the transportation department model 0.2, and the meteorological center model 0.1. Through weighted voting, a comprehensive decision score of 0.772 is obtained, along with the comprehensive decision recommendation: "Considering the patient's condition and transport conditions, it is recommended to immediately administer nitroglycerin tablets and place the patient in a supine position, while simultaneously initiating the transport procedure. Continuously monitor symptom changes during transport and maintain uninterrupted communication."

[0076] Based on the comprehensive decision-making results, a specific execution schedule is generated, clearly indicating the execution time points for each self-rescue measure and transfer step. In this example, the execution schedule is: "0 minutes: take nitroglycerin tablets sublingually; 1 minute: lie flat and loosen tight clothing; 2 minutes: call emergency services; 5 minutes: ambulance arrives and transfer begins; 28 minutes: expected arrival at medical facility." This time arrangement takes into account the urgency of self-rescue measures and the time prediction of the transfer route, ensuring that the rescue operation is carried out in an orderly manner.

[0077] The generated comprehensive decision-making content is formatted and converted into a format suitable for user terminals. For mobile terminals such as smartphones, multimedia content including text descriptions, icon indicators, and voice prompts is generated; for wearable devices such as smartwatches, simplified content is generated, highlighting key steps and time points; and for in-vehicle devices, content focusing on transfer routes and navigation is generated. In this example, for the user's smartphone, a comprehensive content package including self-rescue step diagrams, transfer routes, time reminders, and voice guidance is generated.

[0078] Comprehensive decision-making information is transmitted to the user terminal via an encrypted communication channel. The system automatically selects the most suitable transmission method based on network conditions, such as 4G / 5G network, Wi-Fi, or SMS. To ensure transmission reliability, breakpoint resumption and data compression technologies are employed, and a degradation scheme for weak network environments is designed to ensure that core rescue information can still be transmitted even under poor network conditions. In this example, the comprehensive decision-making information was successfully transmitted to the user's smartphone via the 4G network in 0.8 seconds. The user terminal immediately displayed a self-rescue guidance interface and played voice prompts.

[0079] Through the above methods, this invention successfully inputs the optimal transfer route and self-rescue guidance plan into the federated learning model, obtains a comprehensive decision that takes into account medical professionalism, emergency response timeliness, transportation practicality and environmental adaptability, and sends it safely and efficiently to the user terminal, providing an effective solution for intelligent rescue decision-making in emergency situations.

[0080] This invention also provides a specific embodiment: User A, driving a vehicle, is involved in a traffic accident on a highway. The vehicle collides with the guardrail, and User A suffers chest injuries, suspected rib fractures, and difficulty breathing. The accident occurs at kilometer marker K127+500 on the highway, with no other vehicles or pedestrians nearby. User A sends an emergency assistance message via the vehicle's smart terminal: "My car has hit the guardrail. I have chest pain and difficulty breathing. My location is K127+500. Requesting assistance." Upon receiving this emergency assistance message, the rescue process is immediately initiated.

[0081] Based on User A's emergency request for help, the keywords "chest pain" and "difficulty breathing" were extracted. Combined with physiological data uploaded from the vehicle's intelligent terminal, including heart rate of 95 bpm, blood pressure of 135 / 85 mmHg, respiratory rate of 24 breaths / min, and blood oxygen saturation of 94%, this data was analyzed using a symptom recognition algorithm. The algorithm identified User A's current condition as "closed rib fracture with mild pneumothorax," with a confidence level of 0.87. Using "closed rib fracture with mild pneumothorax" as the starting point and "stabilizing vital signs" as the endpoint in the emergency rescue knowledge graph, a path search was performed to generate a potential rescue action sequence. Combined with on-site environmental information, including the user being alone, the highway environment, the temperature of 28°C, and the straight-line distance to the nearest medical facility being 15 kilometers, a self-rescue guidance plan was generated: maintain a semi-recumbent position and loosen tight clothing; perform slow, shallow breathing, avoiding deep breathing and violent coughing; gently press and stabilize the injured side of the chest with a soft pad or clothing; maintain ventilation inside the vehicle to avoid high temperatures; wait for professional rescue to arrive and avoid moving on your own. The self-rescue plan has an execution difficulty score of 0.35 (out of 1), an expected stability effect score of 0.82 (out of 1), and a resource requirement score of 0.2 (out of 1).

[0082] Based on User A's current symptom "closed rib fracture with mild pneumothorax" and location information "highway K127+500", the system retrieves matching medical resource needs. For this symptom, the importance weights of various medical resources are calculated as follows: thoracic surgeon 0.85, chest CT equipment 0.90, thoracentesis equipment 0.65, respiratory support equipment 0.75, and pain management medication 0.70. Four medical institutions within a 30km radius of User A are retrieved from the medical institution database: Medical institution M1 (18km away) has a complete thoracic surgery team and equipment; Medical institution M2 (12km away) has basic equipment but no specialized thoracic surgery team; Medical institution M3 (25km away) has a thoracic surgery team but its CT equipment is under maintenance; and Medical institution M4 (15km away) is a primary healthcare center with limited equipment and specialist resources. The system calculates the resource matching degrees for each medical institution as 0.92, 0.65, 0.78, and 0.45, respectively, and identifies M1, M2, and M3 as candidate medical institutions.

[0083] Based on real-time traffic information, a transfer path network from user A's location to various candidate medical institutions was constructed. Traffic information showed minor traffic congestion near the highway exit leading to M1, with average speeds dropping to 35 km / h; the road to M2 was clear but involved multiple turns; the highway section leading to M3 was under road maintenance, allowing only one-way traffic. A multi-objective optimization function was constructed, comprehensively considering factors such as transfer time, road risk, medical resource matching, and treatment timeliness. Ant colony optimization was used, with 100 iterations, 20 ants, a pheromone importance factor of 1.0, a heuristic factor of 2.0, and a pheromone evaporation coefficient of 0.5. After solving, the system determined the optimal transfer path to be "current location → highway exit K132 → county road X103 → medical institution M1," with an estimated transfer time of 24 minutes, a road risk score of 0.28 (out of 1), and a comprehensive score of 0.85 (out of 1).

[0084] The optimal transfer route and self-rescue guidance plan were input into a federated learning model, which was jointly trained by five participants: a general hospital model, an emergency center model, a traffic management department model, a meteorological service center model, and a vehicle rescue service model. Each participant trained a specialized model for traffic accident rescue based on its own professional domain and data resources, and performed collaborative reasoning through the federated learning framework. The transfer route feature vector [24, 0.28, 0.92, 18] and the self-rescue plan feature vector [0.35, 0.2, 0.82, 0.75] were input into the federated learning model. Each participant's local model generated local inference results and securely transmitted them to the central server. The general hospital model suggested "prioritize professional transfer while performing chest immobilization"; the emergency center model suggested "immediately initiate ambulance transfer, maintaining a semi-recumbent position"; the traffic management department model suggested "choose the K132 exit transfer route, avoiding congested sections"; the meteorological service center model indicated "good weather, no precipitation, suitable for air rescue"; and the vehicle rescue service model suggested "wait for professional rescue to arrive, avoid unauthorized movement of the injured person." The central server, through a weighted voting mechanism, generates a comprehensive decision: "Considering the risk of rib fracture with mild pneumothorax and the specialized thoracic surgery resources of M1 Hospital, it is recommended to immediately transfer the patient to M1 Hospital via Exit K132. Before transfer, maintain the patient in a semi-recumbent position and perform simple chest immobilization. Simultaneously, consider initiating an air rescue assessment. It is estimated that a rescue helicopter can arrive at the scene within 15 minutes, or an ambulance can arrive within 20 minutes and complete the transfer within 24 minutes." This comprehensive decision is sent to the user's terminal through multiple channels. User A's in-vehicle smart terminal immediately displays a self-rescue guidance interface, including a diagram of the semi-recumbent position, an illustration of chest fixation methods, and voice prompts. Simultaneously, the transfer route is sent to the in-vehicle navigation system, and the user's situation and location information are pushed to the nearest rescue unit via the emergency communication network. A rescue progress message is sent to User A: "The ambulance has been dispatched and is expected to arrive in 20 minutes; the rescue helicopter is assessing the situation and may arrive in 15 minutes; please keep communication open and follow the self-rescue instructions." Through the above methods, this invention realizes intelligent processing of the entire process from emergency assistance to comprehensive rescue decision-making in traffic accident scenarios. It effectively integrates medical knowledge, route planning, and multi-party collaborative decision-making, providing users with timely and professional rescue services and improving the timeliness and success rate of emergency rescue.

[0085] This invention provides an intelligent transfer and self-rescue guidance and decision support system based on federated learning, comprising: The first unit is used to obtain emergency help information sent by user terminals; The second unit is used to extract knowledge entities and relationships from the medical emergency knowledge base and the historical rescue case database to construct an emergency rescue knowledge graph; based on the emergency help information, extract physiological characteristic parameters and identify the user's current symptoms, perform path search in the emergency rescue knowledge graph, and generate a self-rescue guidance plan in combination with the on-site environment; The third unit is used to retrieve matching medical resource needs based on the user's current symptoms and location, determine multiple candidate medical institutions, and construct a multi-objective optimization function by combining the current traffic status of each road segment, and solve it using the ant colony algorithm to obtain the optimal transfer path. The fourth unit is used to input the optimal transfer path and the self-rescue guidance plan into the federated learning model to obtain a comprehensive decision, and then send the comprehensive decision to the user terminal.

[0086] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0087] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0088] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent transportation and self-rescue guidance and decision support based on federated learning, characterized in that, include: Obtain emergency help requests sent by user terminals; Knowledge entities and relationships are extracted from the medical emergency knowledge base and historical rescue case database to construct an emergency rescue knowledge graph; Based on the emergency help information, physiological feature parameters are extracted and the user's current symptoms are identified. Path search is performed in the emergency rescue knowledge graph, and a self-rescue guidance plan is generated in combination with the on-site environment. Based on the user's current symptoms and location, matching medical resource needs are retrieved, multiple candidate medical institutions are identified, and a multi-objective optimization function is constructed by combining the current traffic status of each road segment. The ant colony algorithm is used to solve the problem and obtain the optimal transfer path. The optimal transfer route and the self-rescue guidance plan are input into the federated learning model to obtain a comprehensive decision, which is then sent to the user terminal.

2. The method according to claim 1, characterized in that, Knowledge entities and relationships are extracted from the medical emergency knowledge base and historical rescue case database to construct an emergency rescue knowledge graph, including: The medical emergency knowledge base retrieves treatment guidelines and emergency care instructions, generates structured medical knowledge text, constructs a medical domain word vector space, maps medical terms in the medical knowledge text to semantic vectors, extracts textual features from the semantic vectors, and identifies medical knowledge entities based on the textual features. The medical knowledge entities are divided into multiple entity categories according to a medical professional tree classification system. The correlation degree between the entity categories is calculated. Entity pairs with a correlation degree greater than the correlation threshold are marked as candidate relationship pairs. The candidate relationship pairs are screened by confidence evaluation to obtain the initial knowledge graph. New types of knowledge entities and relationships are identified from the historical rescue case database, and the semantic compatibility of the new entities and relationships with the existing knowledge in the initial knowledge graph is verified. The verified new entities and relationships are then integrated into the initial knowledge graph to obtain the emergency rescue knowledge graph.

3. The method according to claim 1, characterized in that, Based on the emergency assistance information, physiological characteristic parameters are extracted and the user's current symptoms are identified. A path search is performed within the emergency rescue knowledge graph, and a self-rescue guidance plan is generated based on the on-site environment, including: From the emergency help information, extract the user's symptom feature vector and physiological feature data, perform time alignment and synchronization calibration on the physiological feature data, and generate physiological feature vector; Based on the symptom feature vector and the physiological feature vector, and combined with the symptom-disease association network in the emergency rescue knowledge graph, the probability of symptom identification of the user at the current moment is calculated, and the user's current symptom is determined. The current symptom is taken as the starting state node, the target recovery state is determined as the target state node, the initial candidate rescue path from the starting state node to the target state node is calculated, and the initial candidate rescue path is pruned based on the upper limit of resource consumption, the upper limit of time consumption and environmental adaptability constraints to obtain the optimized candidate rescue path. Calculate the feasibility score for each optimized candidate rescue path, select the candidate rescue path with the highest feasibility score, decompose it according to the execution steps, identify the target rescue action, and obtain a self-rescue guidance plan.

4. The method according to claim 3, characterized in that, Calculate the initial candidate rescue path from the starting state node to the target state node. Prune the initial candidate rescue path based on resource consumption limits, time consumption limits, and environmental adaptability constraints to obtain an optimized candidate rescue path, including: Starting from the initial state node, the adjacent state nodes are traversed sequentially through depth search. The node feature vectors of the adjacent state nodes are extracted, and their reachability scores to the target state node are calculated. The adjacent state nodes with reachability scores greater than the reachability threshold are selected as the next search nodes, generating multiple initial candidate rescue paths. For each initial candidate rescue path, calculate the total resource consumption of all its state nodes, and calculate the execution time and transition delay time of each state node to obtain the total execution time. Based on the current environment, calculate the environment adaptability score. When any one of the indicators of the total resource consumption, the total execution time, and the environment adaptability score exceeds the default limit threshold, mark the candidate rescue path as an infeasible path to obtain multiple feasible candidate rescue paths. Calculate the similarity of rescue actions among the feasible candidate rescue paths, and merge the rescue paths with a similarity of rescue actions higher than the path merging threshold to obtain the optimized candidate rescue paths.

5. The method according to claim 1, characterized in that, Based on the user's current symptoms and location, matching medical resource needs are retrieved, multiple candidate medical institutions are identified, and a multi-objective optimization function is constructed by combining the current traffic status of each road segment. This function is then solved using an ant colony optimization algorithm to obtain the optimal transfer path, including: Based on the user's current symptoms, the importance weights of various medical resources are calculated, and candidate medical institutions that meet the basic requirements are selected by combining the standard configuration requirements of various medical resources. The system obtains information on bed occupancy, expert availability, and equipment usage of the candidate medical institutions, calculates resource availability scores, and obtains a comprehensive score for each candidate medical institution based on the resource availability scores and transport distance. The congestion intensity of road segments is calculated based on road traffic flow data and average vehicle speed data, and the real-time travel time of each road segment is predicted by combining traffic light timing data; based on the comprehensive score and the real-time travel time, a multi-objective optimization function is constructed, and the maximum allowable transfer time is set. Multiple search units are set in the road grid. The transfer path is determined based on the path pheromone concentration and dynamic heuristic information. The pheromone update amount is adjusted according to the multi-objective optimization function, and the optimal transfer path is obtained through iterative optimization.

6. The method according to claim 5, characterized in that, Multiple search units are set up in the road grid. The transport path is determined based on the path pheromone concentration and heuristic strength. The pheromone update amount is adjusted according to the multi-objective optimization function, and the optimal transport path is obtained through iterative optimization, including: Multiple search units are randomly deployed in the road grid. Each search unit contains the current location, direction of movement, and visited paths. The heuristic intensity of each search unit is calculated by combining the historical heuristic intensity and the real-time travel time of each road segment. An initial pheromone concentration is randomly distributed in the road grid. The product of the heuristic intensity and the initial pheromone concentration is used as the path selection probability of the search unit. The next transfer direction is determined based on the path selection probability, and the transfer path formed by each search unit in the current iteration cycle is obtained. Calculate the multi-objective optimization function value for each transport path, map it to the pheromone increment, and update the pheromone concentration in the road grid according to the pheromone evaporation rate; Calculate the distribution entropy of the multi-objective optimization function value. When the change in distribution entropy over multiple consecutive iterations is less than the entropy change threshold, select the transfer path with the largest multi-objective function value from the transfer paths in the current iteration cycle as the optimal transfer path.

7. A federated learning-based intelligent transfer and self-rescue guidance and decision support system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to obtain emergency help information sent by user terminals; The second unit is used to extract knowledge entities and relationships from the medical emergency knowledge base and historical rescue case database to construct an emergency rescue knowledge graph. Based on the emergency help information, physiological feature parameters are extracted and the user's current symptoms are identified. Path search is performed in the emergency rescue knowledge graph, and a self-rescue guidance plan is generated in combination with the on-site environment. The third unit is used to retrieve matching medical resource needs based on the user's current symptoms and location, determine multiple candidate medical institutions, and construct a multi-objective optimization function by combining the current traffic status of each road segment, and solve it using the ant colony algorithm to obtain the optimal transfer path. The fourth unit is used to input the optimal transfer path and the self-rescue guidance plan into the federated learning model to obtain a comprehensive decision, and then send the comprehensive decision to the user terminal.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images