AI-based multi-center special disease first-aid cooperation and data closed-loop method and system

Abstract

The invention provides a multi-center special disease first-aid cooperation and data closed-loop method and system based on AI, and relates to the technical field of medical information, and the method comprises the steps: extracting diagnosis and treatment information from multi-center special disease first-aid historical cases, constructing a knowledge pedigree, and formulating diagnosis and treatment standard specifications; calculating an optimal hospital by using federated learning based on the first-aid data and the decentralized node network; feature extraction and correlation analysis are carried out on special disease first-aid data, influence factors are positioned through a traceable chain, and dynamic updating of special disease knowledge pedigree is realized. According to the invention, the cooperative efficiency of special disease first aid is improved, optimal allocation of first aid resources is realized, and a data-driven special disease diagnosis and treatment continuous optimization mechanism is formed.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to an AI-based multi-center disease-specific emergency collaborative and data closed-loop method and system. Background Technology

[0002] With the development of medical technology, emergency medical services are facing increasing challenges in the field of specialized disease treatment. Specialized disease emergency care refers to emergency treatment for specific disease types, such as cardiovascular and cerebrovascular diseases, trauma, and poisoning. These diseases are typically characterized by rapid onset, rapid changes in condition, and a high demand for timely treatment. Currently, my country has established a multi-level, multi-center emergency medical network system, with hospitals across the country undertaking important specialized disease emergency care tasks. With the rapid development of artificial intelligence technology, the application of AI technology in the medical field is becoming increasingly widespread, providing new technical means and methodologies for specialized disease emergency care collaboration and data management.

[0003] While hospitals have accumulated a wealth of specialized emergency care case data, this data is often scattered and unstructured, making it difficult to form a standardized knowledge system. The lack of unified diagnostic and treatment standards for specialized diseases leads to significant differences in treatment protocols among different medical institutions, impacting overall emergency care effectiveness. Traditional patient transfer decisions are often based on experience or simple geographical factors, failing to fully consider dynamic medical resources such as the specialist capabilities, equipment status, and bed availability of each hospital. Furthermore, real-time data sharing and collaboration among multiple centers are impossible, resulting in irrational allocation of emergency resources and affecting patient treatment outcomes. Existing systems often collect specialized emergency care data at a mere recording level, lacking systematic analysis and quantitative assessment of influencing factors during treatment, and failing to establish a complete data traceability chain. This makes it difficult to accurately identify key influencing factors from historical cases and to effectively feed treatment experience and prognostic analysis results back into treatment standards, thus hindering the continuous improvement of specialized emergency care levels. Summary of the Invention

[0004] The embodiments of the present invention provide an AI-based multi-center specialized emergency care collaboration and data closed-loop method and system, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides an AI-based multi-center specialized emergency care collaboration and data closed-loop method, comprising: Information on the entire diagnosis and treatment process is extracted from historical emergency cases of specific diseases from multiple hospitals. This information is then integrated into a specific disease knowledge system based on semantic associations. Based on this specific disease knowledge system, mapping rules between disease development paths and treatment plans are established to form standard norms for the diagnosis and treatment of specific diseases. Based on the specific disease diagnosis and treatment standards and specifications, the hospital generates a specific disease emergency data collection template, collects specific disease emergency data, establishes a decentralized node network among multiple hospitals, and synchronizes the medical resource status of each node and the specific disease emergency data to the decentralized node network in real time; calculates the local score of each node based on the medical resource status, constructs the geographical location of the node within the patient transfer range and the specific disease treatment level as constraints, and uses a federated learning method to combine the local scores and the constraints to determine the optimal hospital; Feature patterns are extracted from the disease-specific emergency data, and influencing factors affecting treatment outcomes are identified from these feature patterns. Correlation analysis is performed between these influencing factors and prognostic results to quantify the contribution of different influencing factors to the prognostic results. A data traceability chain is established for the disease-specific emergency data, and backtracking analysis is performed using the traceability chain to locate the influencing factors. Based on the location results and correlation analysis results, the disease-specific knowledge spectrum is updated.

[0006] Information on the entire diagnosis and treatment process is extracted from historical emergency cases of specific diseases from multiple hospitals. This information is then integrated into a disease-specific knowledge system based on semantic associations. Mapping rules between disease progression paths and treatment plans are established based on this knowledge system, forming standardized guidelines for disease diagnosis and treatment, including: Word segmentation is performed on historical emergency medical records from multiple specialized hospitals. Medical entities in the segmentation results are identified, and each medical entity is labeled with a timestamp. Medical entities with adjacent timestamps are sequentially connected to form a time-series diagnosis and treatment chain. Entity vectors are generated based on the context information of each medical entity. The cosine similarity between the entity vectors is calculated to obtain the semantic association strength between entities. Semantic connections are established between entity pairs whose semantic association strength is greater than a preset semantic threshold to obtain a disease-specific knowledge spectrum. The state transition frequency between any two medical entities in the time-series diagnosis and treatment chain is calculated, and the state transition frequency is integrated into the semantic connection in the disease-specific knowledge spectrum to obtain a disease-specific development network. Clinical intervention nodes are marked in the disease-specific development network, and the predecessor entities, successor entities and their connection relationships of the clinical intervention nodes are extracted to obtain the mapping rules between disease development paths and treatment plans. Based on the mapping rules, corresponding treatment plan selection criteria are configured for each clinical intervention node to form a disease-specific diagnosis and treatment standard specification.

[0007] Based on the aforementioned disease-specific diagnosis and treatment standards and specifications, the hospital generates a disease-specific emergency data collection template, collects disease-specific emergency data, establishes a decentralized node network among multiple hospitals, and synchronizes the medical resource status of each node and the disease-specific emergency data to the decentralized node network in real time, including: The data items in the specific disease diagnosis and treatment standards and specifications are classified, and the frequency of use of each type of data item in the historical specific disease emergency cases is counted to obtain the information importance. Data items with information importance greater than a preset importance threshold are selected to form a specific disease emergency data collection template. In the disease-specific emergency data collection template, numerical range constraint rules and logical relationship verification rules are set, and the diagnosis and treatment data that meet the numerical range constraint rules and logical relationship verification rules are marked as disease-specific emergency data. Digital certificates are created for multiple hospitals, and connections are established between the hospitals based on these digital certificates to form a decentralized node network. Each hospital is treated as a node, and the computing resource status value and network connection quality value of each node are statistically analyzed to generate node priorities. Bed occupancy data, medical equipment status data, and expert resource distribution data of each node are statistically analyzed as medical resource status. According to the node priorities of each node, the medical resource status and the emergency medical data for specific diseases are synchronized to the decentralized node network.

[0008] Based on the status of medical resources, local scores are calculated for each node. The geographical location of nodes within the patient transfer area and the level of specialized disease treatment are used as constraints. A federated learning method is employed, combining the local scores and the constraints, to determine the optimal hospital, which includes: The medical resource status is mapped to a high-dimensional feature space to obtain a feature mapping matrix. The feature mapping matrix is ​​then multiplied by a preset weight vector to obtain the local score of each node. The transfer range is determined based on the patient's location information. The transfer range is used as a distance constraint. The arrival time is calculated based on the geographical location information of each node within the transfer range to obtain a time constraint. The utilization rate of the specialized disease treatment equipment at each node is used as a treatment constraint. Candidate nodes are selected based on the distance constraint, the time constraint, and the treatment constraint. On the candidate nodes, the local scores and constraints are used as node evaluation data. The node evaluation data is encrypted using a homomorphic encryption method and then sent. The encrypted data sent by each candidate node is received. The local encryption gradient of each node is calculated based on the encrypted data and aggregated to obtain the global encryption gradient. The federated comprehensive score is calculated based on the local encryption gradient and the global encryption gradient. The candidate node with the highest federated comprehensive score is determined as the optimal hospital.

[0009] The local encryption gradient of each node is calculated based on the encrypted data and aggregated to obtain the global encryption gradient. The federated comprehensive score is then calculated based on the local encryption gradient and the global encryption gradient, including: The local encryption gradient is obtained by calculating the first derivative of the loss function of each node based on the encrypted data. The local encryption gradient of each candidate node is multiplied by the proportion of the number of samples in the candidate node and then summed to obtain the global encryption gradient. The node contribution is obtained by calculating the cosine similarity between the local encryption gradient of each candidate node and the global encryption gradient. Based on the convergence direction of the global encryption gradient, the directional consistency of the local encryption gradient of each candidate node is calculated to obtain the data quality score. The node contribution and the data quality score are combined to obtain the federated comprehensive score.

[0010] Feature patterns are extracted from the specific emergency medical data. Factors influencing treatment outcomes are identified from these feature patterns. Correlation analysis is performed between these factors and prognostic results to quantify the contribution of different factors to the prognostic outcome, including: The special emergency data is subjected to wavelet transform to obtain time-series features. The feature significance of each scale parameter is calculated based on the rate of change of the time-series features. The scale parameter with the largest feature significance is selected as the optimal scale parameter. A multi-dimensional feature matrix is ​​constructed based on the optimal scale parameter. Feature vectors with a cumulative contribution rate that reaches a preset feature retention threshold are selected from the multi-dimensional feature matrix to construct a feature dimensionality reduction matrix. The multi-dimensional feature matrix and the feature dimensionality reduction matrix are multiplied to obtain the feature pattern. Calculate the regression coefficient between the feature pattern and the prognostic result. Sort the feature components in the feature pattern in descending order according to the absolute value of the regression coefficient. The feature components with a regression coefficient absolute value greater than a preset coefficient screening threshold in the sorting result are identified as influencing factors. Calculate the joint probability distribution and marginal probability distribution of the influencing factors and the prognostic result. Calculate the correlation strength of the influencing factors based on the joint probability distribution and the marginal probability distribution. The influencing factors are combined to generate a feature subset sequence. The prediction bias of each feature subset on the prognostic result is calculated based on the correlation strength of the influencing factors. The contribution value of the influencing factors is calculated based on the prediction bias.

[0011] A data traceability chain is established for the aforementioned disease-specific emergency data. Backtracking analysis is performed using this traceability chain to locate the influencing factors. Based on the location results and correlation analysis results, the disease-specific knowledge spectrum is updated, including: The specialized emergency data is classified and encoded to obtain a data identifier sequence. A data dependency graph is constructed based on the data identifier sequence. Data flow paths are extracted from the data dependency graph to form a data traceability chain. The state vector of the data point to be analyzed is extracted from the data traceability chain. The propagation weight of the state vector and the historical state vector is calculated to obtain the forward propagation score. The tracing weight of the state vector and the target state vector is calculated to obtain the backward propagation score. The forward propagation score and the backward propagation score are combined to obtain the propagation state score. A recursive feature elimination algorithm is used to remove redundant features from the propagation state score. The calculation is iteratively performed until the increase in the prediction accuracy of the remaining features for the prognostic results is lower than a preset feature screening threshold, thereby obtaining the influencing factors corresponding to the data point to be analyzed. The influencing factors are matched with nodes in the disease-specific knowledge spectrum. The importance of the corresponding nodes is updated based on the contribution value of the influencing factors. The knowledge propagation weight is calculated according to the semantic relationship between nodes. The knowledge propagation weight is used to propagate the importance of nodes in a weighted manner to obtain the disease-specific knowledge increment. The disease-specific knowledge spectrum is updated according to the disease-specific knowledge increment.

[0012] A second aspect of this invention provides an AI-based multi-center specialized emergency care collaboration and data closed-loop system, comprising: The first unit is used to extract the entire process information of diagnosis and treatment from historical emergency cases of specific diseases in multiple hospitals, integrate the entire process information of diagnosis and treatment into a specific disease knowledge spectrum according to semantic association, establish mapping rules between disease development path and treatment plan based on the specific disease knowledge spectrum, and form a standard specification for the diagnosis and treatment of specific diseases. The second unit is used by hospitals to generate disease-specific emergency data collection templates based on the disease-specific diagnosis and treatment standards and specifications, collect disease-specific emergency data, establish a decentralized node network among multiple hospitals, and synchronize the medical resource status of each node and the disease-specific emergency data to the decentralized node network in real time; calculate the local score of each node based on the medical resource status, construct the geographical location of the node within the patient transfer range and the level of disease-specific treatment as constraints, and use a federated learning method to combine the local scores and the constraints to determine the optimal hospital; The third unit is used to extract feature patterns from the disease-specific emergency data, identify influencing factors affecting treatment outcomes from the feature patterns, perform correlation analysis between the influencing factors and prognostic results, quantify the contribution of different influencing factors to the prognostic results, establish a data traceability chain for the disease-specific emergency data, use the traceability chain to perform backtracking analysis, locate the influencing factors, and update the disease-specific knowledge spectrum based on the location results and correlation analysis results.

[0013] 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.

[0014] 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.

[0015] The beneficial effects of this application are as follows: This invention establishes a disease-specific knowledge system and standardized specifications, enabling structured collection and multi-center collaborative sharing of emergency data, effectively improving the quality of diagnosis and treatment and the efficiency of resource allocation in disease-specific emergency care.

[0016] This invention employs a decentralized node network combined with federated learning to achieve real-time sharing of medical resource information and intelligent recommendation of optimal hospitals while ensuring patient privacy and security. This significantly shortens patient transfer time and improves emergency care efficiency.

[0017] This invention constructs a complete data closed-loop system. Through quantitative analysis of various influencing factors in the emergency rescue process and backtracking of the traceable chain, it continuously optimizes the disease-specific knowledge spectrum, forming a data-driven emergency rescue optimization mechanism, which effectively improves the overall success rate of disease-specific emergency rescue and the prognosis of patients. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the AI-based multi-center specialized emergency care collaboration and data closed-loop method according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the optimal hospital selection process based on federated learning. Detailed Implementation

[0019] 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.

[0020] 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.

[0021] Figure 1 This is a flowchart illustrating the AI-based multi-center specialized emergency care collaboration and data closed-loop method of this invention, as shown in the embodiment of the invention. Figure 1As shown, the method includes: Information on the entire diagnosis and treatment process is extracted from historical emergency cases of specific diseases from multiple hospitals. This information is then integrated into a specific disease knowledge system based on semantic associations. Based on this specific disease knowledge system, mapping rules between disease development paths and treatment plans are established to form standard norms for the diagnosis and treatment of specific diseases. Based on the specific disease diagnosis and treatment standards and specifications, the hospital generates a specific disease emergency data collection template, collects specific disease emergency data, establishes a decentralized node network among multiple hospitals, and synchronizes the medical resource status of each node and the specific disease emergency data to the decentralized node network in real time; calculates the local score of each node based on the medical resource status, constructs the geographical location of the node within the patient transfer range and the specific disease treatment level as constraints, and uses a federated learning method to combine the local scores and the constraints to determine the optimal hospital; Feature patterns are extracted from the disease-specific emergency data, and influencing factors affecting treatment outcomes are identified from these feature patterns. Correlation analysis is performed between these influencing factors and prognostic results to quantify the contribution of different influencing factors to the prognostic results. A data traceability chain is established for the disease-specific emergency data, and backtracking analysis is performed using the traceability chain to locate the influencing factors. Based on the location results and correlation analysis results, the disease-specific knowledge spectrum is updated.

[0022] In one optional implementation, comprehensive diagnosis and treatment information is extracted from historical emergency cases of specific diseases from multiple hospitals. This comprehensive information is then integrated into a disease-specific knowledge system based on semantic associations. Mapping rules between disease progression paths and treatment plans are established based on this knowledge system, forming standardized guidelines for disease diagnosis and treatment, including: Word segmentation is performed on historical emergency medical records from multiple specialized hospitals. Medical entities in the segmentation results are identified, and each medical entity is labeled with a timestamp. Medical entities with adjacent timestamps are sequentially connected to form a time-series diagnosis and treatment chain. Entity vectors are generated based on the context information of each medical entity. The cosine similarity between the entity vectors is calculated to obtain the semantic association strength between entities. Semantic connections are established between entity pairs whose semantic association strength is greater than a preset semantic threshold to obtain a disease-specific knowledge spectrum. The state transition frequency between any two medical entities in the time-series diagnosis and treatment chain is calculated, and the state transition frequency is integrated into the semantic connection in the disease-specific knowledge spectrum to obtain a disease-specific development network. Clinical intervention nodes are marked in the disease-specific development network, and the predecessor entities, successor entities and their connection relationships of the clinical intervention nodes are extracted to obtain the mapping rules between disease development paths and treatment plans. Based on the mapping rules, corresponding treatment plan selection criteria are configured for each clinical intervention node to form a disease-specific diagnosis and treatment standard specification.

[0023] The text segmentation processing for emergency medical records employs deep learning-based Chinese medical text segmentation technology. It receives electronic medical record data from hospitals, including structured and unstructured text information such as admission records, diagnostic reports, treatment plans, and surgical records. The segmentation algorithm first preprocesses the text, removing special symbols and standardizing the encoding format. Then, it uses a medical-specific dictionary for word segmentation. Taking cardiovascular emergency medical records as an example, the original text "The patient suddenly experienced chest pain accompanied by difficulty breathing; the electrocardiogram showed ST-segment elevation, diagnosed as acute myocardial infarction, and immediately underwent coronary intervention" is segmented into the following word sequence: "patient," "sudden," "chest pain," "difficulty breathing," "electrocardiogram," "ST-segment elevation," "acute myocardial infarction," and "coronary intervention," among other medical terms.

[0024] Medical entity recognition is based on a pre-trained biomedical named entity recognition model, which can accurately identify different types of medical entities such as symptoms, diseases, drugs, treatment plans, and examination items. During the recognition process, a standardized medical ontology identifier is assigned to each entity to ensure that the same concept expressed in different ways can be uniformly classified. For example, different expressions such as "acute myocardial infarction," "acute MI," and "STEMI" are all identified as the same disease entity. A timestamp annotation mechanism extracts the occurrence time of each medical entity from the time information of the medical record, including both absolute and relative time. Absolute time records the specific date and moment, while relative time indicates the chronological order of events in the diagnosis and treatment process.

[0025] By analyzing the timestamp information of medical entities, adjacent or related entities in the time dimension are connected according to their occurrence sequence. A time window threshold is set, typically six to twenty-four hours. Medical events occurring within this time frame are considered to have potential causal relationships. In a specific case, the temporal diagnosis and treatment chain is represented as a complete sequence of diagnosis and treatment: "chest pain → electrocardiogram → ST-segment elevation → acute myocardial infarction diagnosis → medication → coronary angiography → stent implantation → symptom relief". The time intervals and medical entity types between each node are recorded in detail, providing basic data for subsequent semantic association analysis.

[0026] Entity vector generation employs a pre-trained medical text model based on the Transformer architecture. This model, trained on large-scale medical literature and clinical data, captures deep semantic information of medical concepts. The contextual information of each medical entity includes the context of several words preceding and following it in the original text, other co-occurring medical entities, and the entity's position throughout the diagnosis and treatment process. An encoder converts this contextual information into a high-dimensional vector representation, typically with a dimension of 512 or 768. For example, the vector representation of the entity "acute myocardial infarction" not only includes the semantic information of the disease itself but also incorporates semantic features of related concepts such as common symptoms, diagnostic methods, and treatment plans.

[0027] Semantic association strength is calculated by measuring the semantic distance between entity vectors using cosine similarity. Cosine similarity values ​​range from -1 to +1; the closer the value is to +1, the more semantically similar the two entities are. The calculation iterates through all entity pairs, calculating the cosine similarity for each pair, and then uses a semantic threshold for filtering. In practice, the semantic threshold is typically set between 0.7 and 0.85; entity pairs exceeding this threshold are considered to have strong semantic association. For example, the cosine similarity between "chest pain" and "myocardial infarction" reaches 0.82, indicating a high semantic correlation.

[0028] In the process of constructing a disease-specific knowledge spectrum, directed or undirected semantic connections are established between entity pairs that meet semantic threshold conditions, forming a graph-structured knowledge network. The weight of each edge corresponds to the semantic association strength, and the directionality of the edges is determined based on medical causal relationships. The knowledge spectrum contains information at multiple levels, from the symptom layer, diagnosis layer, treatment layer to the prognosis layer, constituting a complete medical knowledge system. In addition to basic identification information, entity nodes also contain attribute information such as frequency statistics and confidence scores, which are used for subsequent path analysis and rule extraction.

[0029] By analyzing the transition probabilities between any two medical entities in a time-series medical treatment chain from a large amount of historical case data, the state transition frequency is obtained. The number of times the transition from entity A to entity B occurs in all cases is counted, and the relative frequency is calculated as a probability indicator of state transition. For example, the frequency of transitioning from the "chest pain" state to the "ECG abnormality" state is 0.75, indicating that an ECG examination will be performed and an abnormality will be found in 75% of chest pain cases. This frequency statistics not only consider direct sequential relationships but also analyze indirect association patterns across multiple time points.

[0030] Disease-specific progression networks are formed by integrating state transition frequency information into established semantic connections. Each edge contains information in two dimensions: semantic association strength and state transition frequency. The network structure presents a complete disease progression trajectory from symptom onset, diagnosis confirmation, treatment implementation to prognostic assessment. Critical paths in the network correspond to common clinical disease progression patterns, while anomalous paths indicate the management of rare complications or special cases. The importance score of each path is calculated, and a comprehensive assessment is performed based on factors such as path frequency, clinical significance, and prognostic impact.

[0031] Clinical intervention node labeling is based on medical behavior classification standards. It involves specially labeling entity nodes in the network that represent proactive medical behaviors such as diagnosis, treatment, and nursing. These nodes typically correspond to controllable medical intervention behaviors such as drug therapy, surgical procedures, examinations, and nursing interventions. The identification criteria for intervention nodes include the entity's medical classification attributes, its role in the diagnosis and treatment process, and the degree of its impact on the patient's condition. For example, "coronary stent implantation," "thrombolytic therapy," and "anticoagulant drug use" are all labeled as important clinical intervention nodes.

[0032] By analyzing the preceding and succeeding entities at clinical intervention points, the correspondence between disease states and treatment plans is identified. Precursor entities typically represent the disease state or symptom that triggers the treatment decision, while succeeding entities reflect the treatment effect or changes in the patient's condition. These association patterns are extracted to form "IF-THEN" mapping rules. Specifically, the rule is expressed as the decision logic: "If a patient experiences acute ST-segment elevation and the onset time is less than twelve hours, emergency coronary intervention is the preferred treatment." Each rule includes a confidence score and applicable condition restrictions.

[0033] The treatment plan selection criteria are based on mapping rules, configuring specific implementation conditions and evaluation indicators for each clinical intervention node. The criteria include multiple dimensions such as indication definition, contraindication exclusion, time window limitations, equipment requirements, and personnel qualifications. Taking coronary intervention as an example, the selection criteria include specific conditions such as "ST-segment elevation greater than one millimeter," "onset time less than twelve hours," "no severe bleeding tendency," and "access to a catheterization lab and interventional physician." These criteria are expressed through a quantitative scoring system, facilitating rapid decision-making during emergency procedures.

[0034] The standardized guidelines for disease diagnosis and treatment integrate all mapping rules and selection criteria to form a structured treatment guideline document. The content of the guidelines is categorized and organized according to dimensions such as disease type, severity, and stage of disease. Each branch contains a complete treatment decision tree and operational procedures. The guidelines also include supporting content such as quality control indicators, effectiveness evaluation standards, and contingency plans for abnormal situations to ensure that comprehensive technical support and operational guidance can be provided in practical applications.

[0035] In one optional implementation, the hospital generates a disease-specific emergency data collection template based on the disease-specific diagnosis and treatment standards and specifications, collects disease-specific emergency data, establishes a decentralized node network among multiple hospitals, and synchronizes the medical resource status of each node and the disease-specific emergency data to the decentralized node network in real time, including: The data items in the specific disease diagnosis and treatment standards and specifications are classified, and the frequency of use of each type of data item in the historical specific disease emergency cases is counted to obtain the information importance. Data items with information importance greater than a preset importance threshold are selected to form a specific disease emergency data collection template. In the disease-specific emergency data collection template, numerical range constraint rules and logical relationship verification rules are set, and the diagnosis and treatment data that meet the numerical range constraint rules and logical relationship verification rules are marked as disease-specific emergency data. Digital certificates are created for multiple hospitals, and connections are established between the hospitals based on these digital certificates to form a decentralized node network. Each hospital is treated as a node, and the computing resource status value and network connection quality value of each node are statistically analyzed to generate node priorities. Bed occupancy data, medical equipment status data, and expert resource distribution data of each node are statistically analyzed as medical resource status. According to the node priorities of each node, the medical resource status and the emergency medical data for specific diseases are synchronized to the decentralized node network.

[0036] The data items in the standard guidelines for the diagnosis and treatment of specific diseases are organized in a hierarchical structure. All data items are divided into seven main categories according to medical information type: patient basic information, symptoms and signs, diagnostic examinations, treatment interventions, medication use, monitoring indicators, and prognostic assessment. The patient basic information category includes basic data fields such as age, gender, past medical history, and allergy history. The symptoms and signs category covers clinical manifestations such as chief complaints, physical examination results, and vital sign monitoring data. The diagnostic examinations category includes auxiliary diagnostic information such as laboratory test items, imaging examination results, and special examination data. The treatment interventions category records proactive medical behaviors such as surgical procedures, interventional treatments, and physical therapy. The medication use category records detailed drug treatment information such as the type of medication, dosage, route of administration, and duration of administration. The monitoring indicators category tracks changes in various physiological parameters during treatment. The prognostic assessment category includes prognostic data such as treatment effect evaluation, complication records, and follow-up results.

[0037] The data items in the standard guidelines for the diagnosis and treatment of specific diseases are organized in a hierarchical structure. All data items are divided into seven main categories based on medical information type: patient basic information, symptoms and signs, diagnostic examinations, treatment interventions, medication use, monitoring indicators, and prognostic assessment. The patient basic information category includes basic data fields such as age, gender, past medical history, and allergy history. The symptoms and signs category covers clinical manifestations such as chief complaints, physical examination results, and vital sign monitoring data. The diagnostic examinations category includes auxiliary diagnostic information such as laboratory test items, imaging examination results, and special examination data. The treatment interventions category records proactive medical behaviors such as surgical procedures, interventional treatments, and physical therapy. The medication use category records detailed drug treatment information such as the type of medication, dosage, route of administration, and duration of administration. The monitoring indicators category tracks changes in various physiological parameters during treatment. The prognostic assessment category includes prognostic data such as treatment effect evaluation, complication records, and follow-up results. Information importance statistics are calculated based on a large-scale historical emergency medical records database. The frequency and completeness of each data item across all cases are scanned. For example, in cardiovascular emergency care, the patient age field appeared 9,800 times in 10,000 cases (98% completeness rate); chest pain symptom records appeared 9,500 times; and electrocardiogram (ECG) results appeared 9,200 times. Some rare and specialized examinations appeared in fewer than 1,000 cases. The correlation between data items and clinical decisions is analyzed. By statistically analyzing the differences in diagnostic accuracy and treatment outcomes among cases containing specific data items, the impact weight of each data item on emergency care quality is quantified. For instance, the onset time record has a decisive influence on the choice of treatment strategy for patients with acute myocardial infarction, achieving a high importance score of 0.95.

[0038] During the data item screening process, an importance threshold of 0.7 was set. Only data items with an importance score exceeding this threshold were included in the emergency data collection template. The screening results showed that the cardiovascular emergency template ultimately contained 48 core data items, including 8 items of basic patient information, 12 items of symptoms and signs, 15 items of diagnostic examinations, 9 items of treatment interventions, and 4 items of medication use. Each data item was accompanied by detailed field definitions, data type descriptions, value range limitations, and filling requirements to ensure consistency and standardization among different hospitals during the data collection process.

[0039] Numerical range constraints are set based on medical common sense and clinical experience to establish reasonable numerical boundaries, preventing the entry and transmission of abnormal data. The age field is constrained to 0 to 120 years old; body temperature is limited to 30 to 45 degrees Celsius; systolic blood pressure is limited to 50 to 300 mmHg; and diastolic blood pressure is limited to 30 to 200 mmHg. Heart rate is constrained to 20 to 200 beats per minute, and blood oxygen saturation is limited to 70% to 100%. Drug dosage constraints are set according to pharmacopoeia standards and clinical medication guidelines.

[0040] Logical relationship verification rules ensure the medical logical consistency and clinical rationality between data items. Various types of verification rules are set, including mutual exclusion verification, dependency verification, time series verification, and dose-matching verification. Mutual exclusion verification ensures that conflicting diagnoses or treatment plans do not occur simultaneously; for example, the diagnoses of acute myocardial infarction and angina pectoris are mutually exclusive, as are the time windows for thrombolytic therapy and anticoagulation therapy. Dependency verification verifies the preconditions for certain data items, such as requiring coronary angiography results to be entered only after the coronary angiography examination, and requiring interventional treatment to be later than the diagnosis time. Time series verification ensures that the chronological order of various medical actions conforms to the clinical workflow; the onset time must be earlier than the consultation time, and the consultation time must be earlier than the treatment start time.

[0041] The specialized emergency medical data labeling mechanism automatically detects whether the entered medical data simultaneously meets two conditions: numerical range constraints and logical relationship verification. It employs real-time verification, performing constraint checks and logical validations immediately during data entry. Data that does not meet the requirements is immediately marked as abnormal and prompted for correction. For example, if the entered systolic blood pressure value is 400 mmHg, it will indicate that the value exceeds the normal range and needs to be reviewed; if the recorded treatment start time is earlier than the consultation time, it will indicate a time logic error requiring adjustment. Only data that passes all verification rules is marked as valid specialized emergency medical data and allowed to enter the subsequent network synchronization process.

[0042] Digital certificate creation utilizes Public Key Infrastructure (PKI) technology to generate a unique digital identity for each hospital. The certificate contains key elements such as the hospital's institution identifier, public key information, validity period, and issuing authority signature. The institution identifier uses a hierarchical coding system: the first three digits represent the regional code, the middle four digits represent the institution type, and the last five digits represent the specific institution serial number. The public key is generated using an elliptic curve cryptography algorithm with a key length of 256 bits to ensure sufficient security strength. The certificate validity period is set at three years, with automatic renewal reminders before expiration. Each certificate also contains basic hospital information, such as institution name, contact information, service scope, and technical level.

[0043] Point-to-point connections are established based on digital certificate-based authentication and key exchange protocols. Any two hospitals can verify each other's identities and establish an encrypted communication channel through certificates. The connection establishment process includes steps such as certificate validity verification, identity legitimacy confirmation, session key negotiation, and communication parameter configuration. It also supports dynamic network topology; newly joining hospitals can request connections from existing nodes in the network via broadcast. Existing nodes verify the new node's certificate and then establish a connection. The network adopts a full graph connectivity model, maintaining a direct connection between any two nodes, thus avoiding single points of failure and network segmentation issues.

[0044] The decentralized node network employs distributed hash table technology for data storage and routing. Each hospital node maintains complete network topology information and a partial copy of shared data. There is no central control node in the network; all decisions are reached through a consensus mechanism among nodes. Inter-node communication uses asynchronous message passing, supporting various communication modes such as data broadcast, on-demand, and multicast. The network has self-healing capabilities; when a node fails, other nodes automatically reorganize the network topology and reallocate data storage responsibilities.

[0045] The system calculates resource status values ​​by tracking hardware resource metrics such as processor utilization, memory usage, storage space utilization, and network bandwidth utilization for each node. Resource status data is collected every minute, and the average value over the past 10 minutes is used as the current status assessment benchmark. Processor utilization is calculated by monitoring the workload of each core of the central processing unit. Memory usage reflects the actual memory usage of the system. Storage space is calculated as the ratio of available disk capacity to total capacity. Network connection quality values ​​include network performance metrics such as connection stability, transmission latency, packet loss rate, and throughput, which are measured by periodically sending test data packets and analyzing the response results.

[0046] Node priority generation considers two dimensions: computing resource status and network connection quality. Priority calculation uses a weighted average method, with computing resource status weighted at 0.6 and network connection quality weighted at 0.4. Specifically, processor utilization has a weight of 0.3, memory utilization at 0.2, storage space at 0.1, network latency at 0.2, and network stability at 0.2. The final priority score ranges from 0 to 1; a higher score indicates superior overall node performance and a higher priority during data synchronization.

[0047] The medical resource status statistics cover three core dimensions: bed occupancy, medical equipment operation status, and expert resource distribution. Bed occupancy data includes information such as the total number of beds, the number of occupied beds, the number of available beds, the status of intensive care unit beds, and the use of operating rooms, reflecting the hospital's capacity to treat patients in real time. Medical equipment status data records information such as the operation status, maintenance plans, and fault records of key equipment, such as the number and availability of electrocardiogram monitors, the operational status of defibrillators, the use of ventilators, and the operational status of imaging equipment. Expert resource distribution statistics include information such as the on-duty status, professional expertise, and workload of experts in each department, including the real-time status of cardiovascular experts, neurology experts, and emergency physicians.

[0048] The data synchronization mechanism executes according to node priority, with the highest-priority node acting as the lead node for this round of synchronization, responsible for initiating data update requests and coordinating the responses of other nodes. The synchronization process employs a distributed consensus algorithm to ensure that all nodes' data remains synchronized. Each synchronization includes incremental data and timestamp information. A synchronization cycle of 5 minutes is set, with real-time synchronization possible in emergencies. Data transmission utilizes compression and encryption technologies to reduce network load and ensure data security. Upon completion of synchronization, each node generates an acknowledgment response and updates its local data version number.

[0049] In one optional implementation, a local score for each node is calculated based on the status of medical resources. The geographical location of nodes within the patient transfer range and the level of specialized disease treatment are constructed as constraints. A federated learning method is used to combine the local scores and the constraints to determine the optimal hospital, which includes: The medical resource status is mapped to a high-dimensional feature space to obtain a feature mapping matrix. The feature mapping matrix is ​​then multiplied by a preset weight vector to obtain the local score of each node. The transfer range is determined based on the patient's location information. The transfer range is used as a distance constraint. The arrival time is calculated based on the geographical location information of each node within the transfer range to obtain a time constraint. The utilization rate of the specialized disease treatment equipment at each node is used as a treatment constraint. Candidate nodes are selected based on the distance constraint, the time constraint, and the treatment constraint. On the candidate nodes, the local scores and constraints are used as node evaluation data. The node evaluation data is encrypted using a homomorphic encryption method and then sent. The encrypted data sent by each candidate node is received. The local encryption gradient of each node is calculated based on the encrypted data and aggregated to obtain the global encryption gradient. The federated comprehensive score is calculated based on the local encryption gradient and the global encryption gradient. The candidate node with the highest federated comprehensive score is determined as the optimal hospital.

[0050] like Figure 2 As shown, the method includes: The medical resource status mapping employs a multi-layer neural network to transform raw medical resource data into a high-dimensional feature vector representation. This represents the medical resource status of each hospital as a raw feature vector encompassing multiple dimensions, including bed occupancy rate, equipment availability, specialist on-duty status, drug inventory, and blood bank reserves. Taking a specific hospital as an example, its raw feature vector includes 18 basic indicators, such as an ICU bed occupancy rate of 0.75, an emergency department bed occupancy rate of 0.82, 3 available cardiac interventional devices, 2 neurosurgical specialists on duty, and a thrombolytic drug inventory adequacy level of 0.9. The mapping network uses a 3-layer fully connected structure. The input layer receives the 18-dimensional raw features, the hidden layer contains 64 neurons, and the output layer generates a 128-dimensional high-dimensional feature representation.

[0051] During the feature mapping matrix construction process, a corresponding high-dimensional feature vector is generated for each candidate hospital in the network. These vectors are then arranged in rows to form the feature mapping matrix. Assuming there are 12 candidate hospitals in the network, the feature mapping matrix has a dimension of 12 rows and 128 columns, with each row representing the location coordinates of a hospital in the high-dimensional feature space. Each element value in the matrix is ​​standardized to ensure that feature values ​​of different dimensions can be compared and calculated within a uniform numerical range. The feature mapping process also introduces a time decay factor, giving more weight to newer medical resource status data in the feature representation.

[0052] The preset weight vector is designed based on historical emergency response data and clinical expert experience. The vector dimension matches the number of columns in the feature mapping matrix, i.e., 128 dimensions. Each component of the weight vector corresponds to a dimension in the high-dimensional feature space, and its value reflects the importance of that dimension to the hospital assessment. For example, the weight of the feature dimension corresponding to intensive care capacity is set to 0.15, the weight of the feature dimension corresponding to the adequacy of expert resources is 0.12, and the weight of the feature dimension corresponding to the advancement of equipment is 0.10. The weight vector will be adjusted according to the specific emergency needs of different types of diseases; the weight vector for cardiovascular emergency care will place greater emphasis on interventional treatment equipment and cardiovascular expert resources.

[0053] Local score calculation is achieved through the inner product operation of the feature mapping matrix and the weight vector. The local score of each hospital is equal to the sum of the products of its high-dimensional feature vector and the corresponding components of the weight vector. The 128-dimensional feature vector of each hospital is multiplied element-wise with the weight vector, and then all the product results are added together to obtain the local score of the center. For example, one hospital has a local score of 83.7 points, and another center has a score of 76.2 points. The higher the score, the stronger the comprehensive service capability of the center under the current medical resource allocation.

[0054] Patient location information is obtained through ambulance-mounted positioning devices or the patient's mobile phone location service. This receives location data including latitude and longitude coordinates, altitude, and positioning accuracy. Based on the patient's current location, Geographic Information System (GIS) technology is used to calculate a reasonable transport range, typically set as a circular area with a radius of 50 kilometers centered on the patient's location. In areas with complex terrain or limited transportation, the transport range is dynamically adjusted according to the actual road network and traffic conditions. Considering the time window requirements of different emergency types, for patients with acute myocardial infarction requiring immediate treatment, the transport range is reduced to 30 kilometers to ensure timely medical care.

[0055] Distance constraints are set based on the geographical distance between the patient's location and each hospital. Two indicators are calculated: the straight-line distance from the patient to each hospital and the actual road distance. The straight-line distance is calculated using spherical distances of latitude and longitude coordinates, while the actual road distance is calculated using an electronic map service interface, taking into account factors such as road network, traffic control, and construction zones. Distance constraints are typically set so that the actual road distance does not exceed 60 kilometers and the straight-line distance does not exceed 45 kilometers. The distance constraints are adjusted in real time according to current traffic conditions, and the constraint range is appropriately narrowed during periods of severe traffic congestion.

[0056] The time constraints calculate the expected arrival time for each node, including multiple components such as travel time, traffic delay time, and intra-hospital transfer time. Travel time is calculated based on actual road distance and the average speed of the ambulance. The average speed of the ambulance on urban roads is set at 40 km / h, and on highways at 80 km / h. Traffic delay time is estimated by analyzing historical traffic data and real-time road condition information, maintaining statistical models of delays for different time periods and road segments. Intra-hospital transfer time includes the time taken for the ambulance to enter the hospital, patient handover, and initial assessment, typically set at 10 to 15 minutes. The total arrival time constraint requires that patients with acute myocardial infarction arrive within 90 minutes and stroke patients arrive within 60 minutes.

[0057] The assessment of treatment constraints evaluates the utilization rate of specialized treatment equipment and the level of professional competence at each stage. Equipment utilization rate statistics include the current usage status of key medical equipment, such as the utilization rate of cardiac catheterization labs, the occupancy rate of computed tomography (CT) scanners, and the availability of MRI equipment. Taking cardiovascular emergency care as an example, the target hospital is required to have at least one usable cardiac catheterization lab equipped with complete coronary intervention equipment, with a current utilization rate of less than 80%. Professional competence is assessed through indicators such as expert resource allocation, historical treatment success rate, and technical level certification. The target center is required to have at least two cardiovascular interventional specialists on duty, with an acute myocardial infarction treatment success rate of over 95% in the past three months.

[0058] The candidate node selection process comprehensively considered three constraints: distance, time, and treatment. Only hospitals that simultaneously met all constraints could enter the candidate node list. A step-by-step selection method was adopted: first, nodes outside the transport range were eliminated based on distance constraints; then, nodes unable to arrive within the specified time were eliminated based on time constraints; and finally, nodes lacking sufficient professional capabilities were eliminated based on treatment constraints. After the selection, only 5 candidate nodes remained from the original 12 hospitals. These nodes all possessed the basic conditions for receiving and treating patients.

[0059] The node evaluation data includes local scores and quantitative indicators of various constraints, forming a multi-dimensional evaluation vector. The evaluation data for each candidate node includes 10 key indicators such as local score, arrival time, equipment utilization rate, expert resource index, and historical success rate. For example, the evaluation data for a candidate node might include a local score of 83.7, expected arrival time of 45 minutes, cardiac catheterization lab utilization rate of 60%, number of on-duty experts of 3, and recent success rate of 97%. This data constitutes a 10-dimensional evaluation vector, providing input for the subsequent federated learning process.

[0060] The homomorphic encryption method employs a lattice-based fully homomorphic encryption algorithm, supporting addition and multiplication operations of arbitrary depth in the encrypted state. It generates independent public-private key pairs for each candidate node; the public key is used to encrypt the evaluation data, and the private key is used to decrypt the computation results. During encryption, each component of the 10-dimensional evaluation vector is independently encrypted into a ciphertext vector. The encrypted data is approximately 1000 times longer than the original data, ensuring sufficient security strength. The encrypted data is sent to other candidate nodes through a secure channel. Recipients cannot directly access the original evaluation data content but can participate in the federated computation process in the ciphertext state.

[0061] In one optional implementation, the local encryption gradient of each node is calculated based on the encrypted data and aggregated to obtain the global encryption gradient. The federated comprehensive score is then calculated based on the local encryption gradient and the global encryption gradient, including: The local encryption gradient is obtained by calculating the first derivative of the loss function of each node based on the encrypted data. The local encryption gradient of each candidate node is multiplied by the proportion of the number of samples in the candidate node and then summed to obtain the global encryption gradient. The node contribution is obtained by calculating the cosine similarity between the local encryption gradient of each candidate node and the global encryption gradient. Based on the convergence direction of the global encryption gradient, the directional consistency of the local encryption gradient of each candidate node is calculated to obtain the data quality score. The node contribution and the data quality score are combined to obtain the federated comprehensive score.

[0062] The first derivative of the loss function is calculated based on the encrypted evaluation data of each node and a predefined optimization objective function. A loss function is constructed for each candidate node, which measures the degree of deviation between the current node's evaluation result and the ideal hospital selection criteria. The loss function includes a weighted combination of multiple evaluation indicators, such as arrival time error, equipment utilization deviation, and expert resource matching degree. Taking a candidate node as an example, its loss function includes five main components: arrival time deviation (weight 0.3), equipment availability deviation (weight 0.25), expert resource deviation (weight 0.2), historical success rate deviation (weight 0.15), and bed availability deviation (weight 0.1).

[0063] The first derivative is calculated using numerical differentiation under encrypted conditions. The loss function is slightly perturbed near the current evaluation point, and the change in function value is observed to estimate the derivative. Each evaluation metric is perturbed both positively and negatively, with the perturbation magnitude set to 0.01 times the current value. The partial derivative is calculated by comparing the difference in loss function values ​​before and after the perturbation. For example, for the arrival time metric, the current value of 45 minutes is adjusted to 44.55 minutes and 45.45 minutes respectively, and the corresponding loss function values ​​are calculated. The partial derivative of the loss function with respect to arrival time is obtained as 0.12 using the difference method. Each node ultimately yields a 10-dimensional local encrypted gradient vector, where each component corresponds to the partial derivative value of an evaluation metric.

[0064] The sample size ratio is determined based on the historical case volume and current medical resource scale of each candidate node. The total number of emergency cases handled by each candidate node in the past 12 months is calculated, and the overall sample weight is calculated by combining resource indicators such as the number of beds, doctors, and equipment of the node. For example, node A handled 1200 emergency cases, has 80 beds and 15 specialist doctors, and its sample weight is calculated as 0.28. Node B handled 800 cases, has 60 beds and 12 doctors, and its sample weight is 0.19. The sample weights of nodes C, D, and E are 0.22, 0.16, and 0.15, respectively. The sum of the weights of the five nodes is 1.0, ensuring mathematical consistency in the aggregation process.

[0065] The global encryption gradient aggregation process calculates a weighted average of the local encryption gradients of each candidate node according to the corresponding sample size ratio. The aggregation algorithm performs vector multiplication and addition operations while maintaining data encryption. The 10-dimensional local gradient vector of each node is first multiplied by the sample weights of that node, and then the weighted gradient vectors of all nodes are accumulated component by component. For example, the local gradient vector of node A is [0.12, -0.08, 0.15, 0.03, -0.06, 0.09, 0.11, -0.04, 0.07, 0.02], which, after multiplying by the sample weights of 0.28, yields the weighted gradient vector. Similarly, the gradient vectors of the other four nodes are processed, and finally accumulated to obtain the global encrypted gradient vector [0.095, -0.063, 0.118, 0.025, -0.048, 0.073, 0.087, -0.032, 0.056, 0.018].

[0066] Cosine similarity is calculated by taking the cosine of the angle between the local and global encryption gradient vectors of each candidate node. This quantifies the consistency between the node's gradient direction and the global optimization direction. The calculation involves two steps: vector dot product and vector magnitude calculation. The cosine similarity is equal to the dot product of the two vectors divided by the product of their respective magnitudes. For example, node A has a dot product of 0.0847, a local gradient vector magnitude of 0.235, and a global gradient vector magnitude of 0.189. The cosine similarity is calculated as 0.0847 divided by the product of 0.235 and 0.189, resulting in 0.191. Similarly, the cosine similarities for nodes B, C, D, and E are 0.267, 0.223, 0.156, and 0.134, respectively.

[0067] Node contribution evaluation is based on cosine similarity results, which are then standardized to map the original similarity values ​​to a scoring range of 0 to 1. The standardization process uses a min-max normalization method, with the minimum similarity value as the lower bound and the maximum similarity value as the upper bound, while other values ​​are scaled proportionally to the range. In the example above, a minimum similarity of 0.134 corresponds to a contribution of 0, and a maximum similarity of 0.267 corresponds to a contribution of 1. Node A's contribution is (0.191 - 0.134) divided by (0.267 - 0.134), resulting in 0.429. The contributions of nodes B, C, D, and E are 1.0, 0.669, 0.165, and 0, respectively. A higher contribution indicates that the node's local optimization direction is more consistent with the global optimization objective, playing a more important role in the federated learning process.

[0068] Convergence direction analysis is based on the numerical characteristics and trends of the global encrypted gradient vector. A historical global gradient sequence is maintained, and convergence performance is evaluated by comparing the gradient vector changes over several consecutive iterations. Convergence direction indicators include the decay rate of the gradient vector magnitude, the stability of the vector direction, and the oscillation amplitude of the component values. For example, if the magnitude of the global gradient vector in the current iteration is 0.189, compared to 0.201 in the previous iteration and 0.218 two iterations prior, showing a clear decreasing trend, it indicates that the optimization process is moving towards convergence. The sign stability of each component of the gradient vector is analyzed; components that maintain the same sign over multiple iterations are considered to have good directional consistency.

[0069] Directional consistency is calculated to determine the degree of matching between the local gradient vectors of each candidate node and the convergence direction. This is quantified by analyzing the projection components of the local gradients onto the global convergence direction, representing the convergence direction as a unit vector, and calculating the projection length of each node's local gradient vector along that direction. A positive projection length indicates that the node's gradient contributes to convergence, while a negative projection length indicates that the node's gradient hinders convergence. For example, node A has a projection length of 0.156 on the convergence direction, which, after standardization, yields a directional consistency score of 0.74. The directional consistency scores for nodes B, C, D, and E are 0.89, 0.81, 0.52, and 0.38, respectively. Higher scores indicate better data quality for that node, enabling it to more effectively drive the optimization process towards convergence.

[0070] The data quality score is calculated by considering both directional consistency and gradient stability. Gradient stability is evaluated by analyzing the changes in the local gradient of a node over several iterations; nodes with smaller changes are considered to have higher data stability. The standard deviation of the local gradient vector of each node in the most recent 5 iterations is calculated; the smaller the standard deviation, the more stable the gradient. The data quality score is equal to the directional consistency score multiplied by a weight of 0.7, plus the gradient stability score multiplied by a weight of 0.3. The data quality score of node A is 0.722 (0.74 multiplied by 0.7, plus 0.68 multiplied by 0.3). The data quality scores of nodes B, C, D, and E are 0.845, 0.763, 0.486, and 0.341, respectively.

[0071] The federated comprehensive score combines the evaluation results of node contribution and data quality score, and is calculated using a weighted geometric mean method. The geometric mean effectively balances the influence of the two scoring dimensions, preventing any single dimension from dominating the final result. The weighting is set to consider the actual needs of emergency rescue scenarios: node contribution is weighted at 0.6, and data quality score is weighted at 0.4, reflecting the emphasis on node collaboration capabilities and data reliability. Node A's federated comprehensive score is 0.429^0.6 multiplied by 0.722^0.4, resulting in 0.563. The federated comprehensive scores for nodes B, C, D, and E are 0.916, 0.710, 0.294, and 0.208, respectively.

[0072] The scoring results were validated through cross-validation and sensitivity analysis to ensure the reliability and stability of the calculation results. The cross-validation process randomly divided candidate nodes into multiple subsets, repeatedly performing the federated scoring calculation on each subset and comparing the consistency of the scoring results across different subsets. Sensitivity analysis observed the magnitude of changes in the scoring results by adjusting weight parameters and perturbing the input data, ensuring that the scoring algorithm has appropriate robustness to parameter changes and data noise. The validation results show that node B maintained the highest federated comprehensive score under various test conditions, confirming its rationality as the optimal hospital selection. A detailed scoring decomposition report was generated, demonstrating the performance differences of each node across different evaluation dimensions, providing comprehensive data support for emergency medical decision-making.

[0073] In one optional implementation, feature patterns are extracted from the disease-specific emergency data; influencing factors affecting treatment outcomes are identified from these feature patterns; and correlation analysis is performed between these influencing factors and prognostic results to quantify the contribution of different influencing factors to the prognostic results, including: The special emergency data is subjected to wavelet transform to obtain time-series features. The feature significance of each scale parameter is calculated based on the rate of change of the time-series features. The scale parameter with the largest feature significance is selected as the optimal scale parameter. A multi-dimensional feature matrix is ​​constructed based on the optimal scale parameter. Feature vectors with a cumulative contribution rate that reaches a preset feature retention threshold are selected from the multi-dimensional feature matrix to construct a feature dimensionality reduction matrix. The multi-dimensional feature matrix and the feature dimensionality reduction matrix are multiplied to obtain the feature pattern. Calculate the regression coefficient between the feature pattern and the prognostic result. Sort the feature components in the feature pattern in descending order according to the absolute value of the regression coefficient. The feature components with a regression coefficient absolute value greater than a preset coefficient screening threshold in the sorting result are identified as influencing factors. Calculate the joint probability distribution and marginal probability distribution of the influencing factors and the prognostic result. Calculate the correlation strength of the influencing factors based on the joint probability distribution and the marginal probability distribution. The influencing factors are combined to generate a feature subset sequence. The prediction bias of each feature subset on the prognostic result is calculated based on the correlation strength of the influencing factors. The contribution value of the influencing factors is calculated based on the prediction bias.

[0074] Wavelet transform processing of emergency medical data employs continuous wavelet transform technology to decompose the time-domain signal into components with different frequencies and time scales. The Morlet wavelet is selected as the mother wavelet function, as it possesses excellent time-frequency localization characteristics, making it suitable for analyzing transient changes in medical indicators during emergency care. For cardiovascular emergency data, the patient's vital signs monitoring sequence is used as the input signal, including continuous monitoring data such as heart rate, blood pressure, blood oxygen saturation, and ST segment changes on electrocardiogram. Taking a patient with acute myocardial infarction as an example, their heart rate data was collected every minute during a 120-minute emergency treatment, forming a time series of 120 data points. Wavelet transform decomposes this sequence into 32 frequency components at different scales, with scale parameters ranging from 1 to 32, each corresponding to different time-frequency characteristics.

[0075] Temporal feature extraction is based on the amplitude and phase information of wavelet transform coefficients. The local energy distribution of wavelet coefficients at each scale is calculated, and the time-varying characteristics of the coefficient sequence are analyzed using a sliding window method. The window length is set to 15 minutes, and the sliding step is 5 minutes. Within each window, statistical characteristics such as the mean, variance, skewness, and kurtosis of the wavelet coefficients are calculated. For example, in the frequency components corresponding to scale 8, feature vectors for 24 time windows are extracted, each containing 4 statistical feature values. For heart rate variability analysis, particular attention is paid to the energy changes of low-frequency components (scales 16-24) and high-frequency components (scales 4-8), as these frequency bands reflect the regulatory state of the autonomic nervous system and the stability of cardiovascular function.

[0076] The rate of change is calculated based on the characteristic differences between adjacent time windows. A first-order difference method is used to calculate the rate of change of the feature vector over time. The rate of change equals the current window's feature value minus the previous window's feature value, divided by the time interval. Taking blood pressure changes as an example, a patient's systolic blood pressure averages 145 mmHg in the third time window and 152 mmHg in the fourth window, corresponding to a rate of change of 1.4 mmHg every 5 minutes. A second-order rate of change is calculated to capture acceleration information, reflecting the abruptness of changes in physiological indicators. The statistical characteristics of the rate of change sequence, such as the frequency of extreme values, the distribution of change amplitude, and the consistency of the direction of change, constitute important indicators describing the dynamic characteristics of the emergency treatment process.

[0077] Feature significance assesses the discriminative ability of each scale parameter in distinguishing different prognostic outcomes. Analysis of variance (ANOVA) is used to compare the distribution differences of features across different prognostic groups at each scale, and the ratio of between-group variance to within-group variance is calculated as the significance index. For example, emergency patients are divided into survival and death groups based on prognostic outcomes, and the feature distributions of the two groups across 32 scale parameters are analyzed. The feature corresponding to scale 12 has a mean of 0.82 and a standard deviation of 0.15 in the survival group, and a mean of 0.34 and a standard deviation of 0.18 in the death group. The calculated feature significance for this scale is 3.67; a higher significance value indicates a more significant contribution of the feature at that scale to prognostic prediction. Information gain and Gini impurity are used to verify the reliability of feature significance from different perspectives.

[0078] The optimal scale parameters were selected based on the significance ranking of features. The 32 scale parameters were ranked from highest to lowest significance, and the top 8 scales with the highest significance were selected as the optimal set of scale parameters. In cardiovascular emergency cases, the optimal scale parameters included scales 3, 7, 12, 18, 24, 28, 31, and 32, corresponding to different ranges of physiological change frequencies. Scales 3 and 7 primarily capture rapid physiological responses in the acute phase, scales 12 and 18 reflect hemodynamic regulation in the intermediate phase, scales 24 and 28 correspond to slow metabolic adaptation processes, and scales 31 and 32 contain long-term trend information. The selected optimal scale parameters needed to be cross-validated to ensure their generalization ability across different patient populations.

[0079] The temporal features corresponding to the optimal scale parameters are combined to form a high-dimensional feature space, constructing a multi-dimensional feature matrix. The number of rows in the matrix equals the number of training samples, and the number of columns equals the feature dimension. For a dataset containing 500 emergency cases, 96-dimensional features (8 scales × 12 time windows × 1 mean feature) are extracted for each case, forming a 500×96 feature matrix. Each element in the matrix represents the feature value of a specific patient at a specific time window and scale. All feature values ​​are standardized to eliminate the influence of different units and numerical ranges. Standardization uses the Z-score method, adjusting the mean of each feature column to 0 and the standard deviation to 1, ensuring that all features are compared and analyzed within the same numerical range.

[0080] Feature dimensionality reduction analysis employs principal component analysis (PCA) to identify the main directions of change in the feature space. A 96×96 covariance matrix is ​​calculated, and eigenvalues ​​and eigenvectors are obtained. The principal components are then sorted according to their eigenvalues. The cumulative variance contribution of the top 20 principal components reaches 85%, meeting the preset feature retention threshold. The feature dimensionality reduction matrix consists of the eigenvectors of these 20 principal components, with a dimension of 96×20. Each principal component represents a linear combination direction in the original feature space. The first principal component primarily reflects the stability of the overall physiological state, the second principal component corresponds to the intensity of the acute stress response, and the third principal component represents the sensitivity to treatment response.

[0081] Feature pattern generation achieves dimensionality reduction projection through matrix multiplication of the multidimensional feature matrix and the dimensionality reduction matrix. The calculation process multiplies the original 500×96 feature matrix with a 96×20 dimensionality reduction matrix to obtain a 500×20 feature pattern matrix. Each row in this matrix represents a patient's feature vector in the 20-dimensional principal component space, and each column represents the projection distribution of all patients on a certain principal component. The feature patterns retain the most important variation information in the original data while significantly reducing the data dimensionality, improving the computational efficiency of subsequent analysis. The reconstruction error is calculated to verify the dimensionality reduction quality; a reconstruction error of less than 5% indicates that the feature patterns have well preserved the main information of the original data.

[0082] A linear regression model was established to calculate regression coefficients between the feature patterns and prognostic outcomes. Prognostic outcomes were numerically coded, with 1 for survival and 0 for death. Ridge regression was used to address multicollinearity. The regularization parameter was set to 0.01 to balance model fitting accuracy and generalization ability. For the 20-dimensional feature patterns, regression analysis yielded 20 regression coefficients, each corresponding to the weight of each principal component's influence on the prognostic outcome. For example, the regression coefficient for the first principal component was 0.24, the second principal component was -0.18, the third principal component was 0.31, the fourth principal component was 0.09, and the coefficients of the other principal components decreased sequentially. Positive coefficients indicate a positive correlation between the principal component and a favorable prognosis, while negative coefficients indicate a correlation with a poor prognosis.

[0083] Feature components were ranked in descending order based on the absolute value of their regression coefficients. The absolute value of the regression coefficient for each principal component was calculated, and the 20 principal components were ranked from largest to smallest. The ranking results were: Principal Component 3 (0.31), Principal Component 1 (0.24), Principal Component 2 (0.18), Principal Component 7 (0.15), Principal Component 5 (0.12), etc. A preset coefficient screening threshold of 0.10 was set; only principal components with absolute values ​​greater than this threshold were considered significant influencing factors. After screening, the top 8 principal components were identified as key influencing factors. These factors explained the main part of the variation in prognostic outcomes, providing important reference for clinical decision-making.

[0084] The value range of each influencing factor was divided into 10 intervals, and the prognostic results were divided into two categories, constructing a 10×2 joint frequency matrix. Taking the first principal component as an example, its value range is -2.5 to 3.2, divided into 10 equal-width intervals, and the number of surviving and deceased patients in each interval was counted. The joint probability is equal to the frequency of the corresponding cell divided by the total number of samples. For example, in the third interval, there were 45 surviving patients and 12 deceased patients, with a total sample size of 500, so the joint probability is 57 / 500, which equals 0.114. Laplace smoothing was used to process cells with a frequency of 0 to avoid the singularity problem in probability estimation.

[0085] Marginal probability distributions are obtained by summing the joint probability distributions. The marginal probability of an influencing factor equals the sum of the joint probabilities of all its value intervals. The marginal probability of a prognostic outcome equals the sum of the joint probabilities of all influencing factors under that outcome. Taking the third interval of the first principal component as an example, its marginal probability is the number of all patients in that interval divided by the total sample size, i.e., 57 / 500 equals 0.114. The marginal probability of surviving patients is the number of all surviving patients divided by the total sample size, assumed to be 380 / 500 equals 0.76. Marginal probability distributions reflect the individual distribution characteristics of each variable, providing basic data for subsequent association strength calculations.

[0086] The correlation strength between influencing factors was quantified using the mutual information method. Mutual information measures the statistical dependence between two random variables; a higher value indicates a stronger correlation. The calculation process requires logarithmic operations on the joint probability distribution and marginal probability distributions, using the natural logarithm and addressing numerical stability issues. The mutual information value between the first principal component and the prognostic result was 0.187, the third principal component was 0.243, and the second principal component was 0.156. The correlation strength was further validated using Pearson correlation coefficient and Spearman rank correlation coefficient to ensure the reliability and consistency of the results.

[0087] Feature subset sequence generation employs a combination of exhaustive and heuristic search strategies. For the eight key influencing factors, all possible non-empty subsets are generated, totaling 255. Subsets are grouped according to the number of factors they contain: 8 single-factor subsets, 28 two-factor subsets, 56 three-factor subsets, and so on. Each subset represents a possible combination of influencing factors, and the differences in the effectiveness of different combinations on prognostic prediction are evaluated. To control computational complexity, lower-order subsets are evaluated first, and a pruning strategy is used to exclude obviously poor combinations.

[0088] Prediction bias is calculated based on the difference between the prediction model built for each feature subset and the actual prognostic results. Logistic regression is used as the base prediction model, and 10-fold cross-validation is employed to evaluate model performance. For a two-factor subset containing the first and third principal components, the trained model achieved an accuracy of 78%, a recall of 72%, and an F1 score of 0.75 on the test set. Prediction bias is defined as 1 minus the F1 score, which is 0.25. The prediction bias is calculated for all 255 subsets; a smaller bias indicates stronger predictive ability for that subset and a more important combination of influencing factors.

[0089] The contribution value calculation uses the Shapley method to quantify the marginal contribution of a single influencing factor to the overall predictive performance. The calculation process considers the marginal effect of the factor across all possible subsets, and the final contribution value is obtained through weighted averaging. The contribution value calculation of the first principal component requires comparison of all subsets including and excluding the factor, with the weights determined based on the subset size and number of combinations. After calculation, the contribution value of the first principal component is 0.089, the third principal component is 0.124, the second principal component is 0.067, and the contribution values ​​of other factors decrease sequentially. The sum of the contribution values ​​equals the overall predictive performance of the optimal subset, ensuring fairness and completeness in the allocation.

[0090] In one optional implementation, a data traceability chain is established for the specialized emergency data, and the traceability chain is used for backtracking analysis to locate the influencing factors. Based on the location results and correlation analysis results, the specialized disease knowledge spectrum is updated, including: The specialized emergency data is classified and encoded to obtain a data identifier sequence. A data dependency graph is constructed based on the data identifier sequence. Data flow paths are extracted from the data dependency graph to form a data traceability chain. The state vector of the data point to be analyzed is extracted from the data traceability chain. The propagation weight of the state vector and the historical state vector is calculated to obtain the forward propagation score. The tracing weight of the state vector and the target state vector is calculated to obtain the backward propagation score. The forward propagation score and the backward propagation score are combined to obtain the propagation state score. A recursive feature elimination algorithm is used to remove redundant features from the propagation state score. The calculation is iteratively performed until the increase in the prediction accuracy of the remaining features for the prognostic results is lower than a preset feature screening threshold, thereby obtaining the influencing factors corresponding to the data point to be analyzed. The influencing factors are matched with nodes in the disease-specific knowledge spectrum. The importance of the corresponding nodes is updated based on the contribution value of the influencing factors. The knowledge propagation weight is calculated according to the semantic relationship between nodes. The knowledge propagation weight is used to propagate the importance of nodes in a weighted manner to obtain the disease-specific knowledge increment. The disease-specific knowledge spectrum is updated according to the disease-specific knowledge increment.

[0091] The classification and coding process employs a hierarchical coding scheme to identify emergency medical data. The data is categorized according to dimensions such as medical information type, time attribute, source channel, and processing stage, with each dimension assigned a specific number of bits. Medical information type occupies 2 bits: 01 for basic patient information, 02 for symptoms and signs, 03 for diagnostic examinations, 04 for treatment interventions, 05 for medication use, 06 for monitoring indicators, and 07 for prognostic assessment. Time attributes use an 8-bit timestamp, accurate to the minute. The source channel uses a 2-bit code to distinguish different data acquisition devices and systems. The processing stage uses a 1-bit code to identify the data processing status: 0 for raw data, 1 for cleaned data, and 2 for data after feature extraction.

[0092] The data identifier sequence generation method connects the codes of each dimension in a predetermined order to form a unique identifier. Taking the electrocardiogram data of a cardiovascular emergency patient as an example, its complete identifier sequence is 03202401151045A112, indicating that the data is of the diagnostic examination type, collected at 10:45 AM on January 15, 2024, originating from the A1 monitoring device, and is the second data item in the first batch of data after two processing steps. An independent identifier sequence is assigned to each piece of emergency data to ensure data uniqueness and traceability. The identifier sequence also includes a check bit, using a cyclic redundancy check algorithm to verify the integrity and accuracy of the code, preventing errors during data transmission.

[0093] Data dependency graphs are constructed by modeling the processing and referencing relationships of data identifier sequences. The generation process of each data item is analyzed, identifying its dependent input data and generated output data, and a directed graph structure is established to represent these dependencies. Nodes in the graph represent specific data items, edges represent dependencies between data, and the weight of the edges reflects the strength of the dependency. For example, a patient's diagnosis result node depends on multiple input nodes such as an electrocardiogram (ECG) examination node, a blood test node, and a symptom description node. The weight of each dependency edge is determined based on the degree of influence of that input data on the diagnostic decision. The weight of an ECG examination for the diagnosis of acute myocardial infarction is 0.4, the weight of a blood troponin test is 0.35, and the weight of a symptom description is 0.25.

[0094] Data flow path extraction employs a graph traversal algorithm combining depth-first search and breadth-first search. Starting from any data node, it traces the source and influence paths of data along the directions of dependencies, forming a complete data flow chain. For a specific treatment decision data point, its backtracking path includes multiple stages such as symptom observation → physical examination → auxiliary examinations → diagnosis confirmation → treatment plan formulation. The data identifier sequence corresponding to each stage constitutes the node sequence on the path, and the time intervals and processing relationships between nodes constitute the path's attribute information. It identifies branch paths and convergence paths, handling the complex network structure in data dependencies.

[0095] The data traceability chain establishes a chronological and logically ordered data flow path, encompassing the entire lifecycle of data generation from initial data collection to final output. Each chain node includes detailed attributes such as data content, processing operations, timestamps, responsible personnel, and equipment information. Taking the treatment process of an acute myocardial infarction patient as an example, the traceability chain contains 120 data nodes, covering the entire data flow from the patient's emergency call to discharge follow-up. The chain supports both forward tracing and reverse backtracking query modes. Forward tracing analyzes the subsequent impact of data, while reverse backtracking identifies the original source of the data.

[0096] State vector extraction provides multi-dimensional feature descriptions for specific data points in a traceable chain. It encodes the attribute information of each data point into a fixed-length numerical vector. The vector dimensions include features such as data type, numerical range, time attribute, quality score, and processing complexity. For example, the state vector for a blood pressure monitoring data point is [0.6, 0.8, 0.3, 0.9, 0.4, 0.7, 0.5, 0.2], where each component represents the normalized value of the data type, the position of the value relative to the normal range, the urgency of the situation, the data quality score, processing complexity, equipment reliability, operator experience, and environmental interference. The state vector is set to 8 dimensions, and principal component analysis is used to ensure the independence and representativeness of each dimension.

[0097] Historical state vectors are constructed based on the historical statistical characteristics of data points of the same type. A historical state database is maintained for different data types, recording the distribution of state vectors for similar data points over the past six months. The historical state vector is calculated by weighted averaging of historical data, with weights determined by time distance and similarity. Recent data has higher weights, as do data with high similarity. For blood pressure monitoring data, the historical state vector is [0.55, 0.72, 0.35, 0.85, 0.45, 0.68, 0.52, 0.28], representing typical state characteristics of historical blood pressure monitoring data.

[0098] The forward propagation weight calculation employs a combination of cosine similarity and Euclidean distance. The forward propagation weight measures the similarity between the current state vector and historical state vectors; a higher similarity indicates that the current data point's state more closely matches historical experience, resulting in a larger propagation weight. The calculation process includes vector normalization, inner product calculation, and distance measurement. For example, a data point has a cosine similarity of 0.87 and an Euclidean distance of 0.34 with historical states, resulting in a forward propagation weight of 0.76. The backward propagation weight similarly calculates the similarity between the current state vector and the target state vector, which represents the ideal data processing result.

[0099] The forward propagation score reflects the reasonableness of a data point within a historical experience framework. The score is calculated by multiplying the forward propagation weight by the data point's intrinsic quality index to obtain a comprehensive forward evaluation result. Intrinsic quality indicators include scores for data completeness, consistency, and timeliness. For example, if a data point has a forward propagation weight of 0.76 and an intrinsic quality index of 0.92, its forward propagation score is 0.76 multiplied by 0.92, which equals 0.699. The score ranges from 0 to 1; a higher score indicates that the data point better conforms to the expected processing pattern and has higher credibility in traceability analysis.

[0100] Backpropagation scores assess the contribution of data points to the final goal. The calculation process considers the matching degree between the data point and the target state, as well as the influence weight of the data point in the decision-making process. The target state vector represents the ideal data characteristics for successful treatment. If a data point has a backpropagation weight of 0.83 with respect to the target state, its influence weight in the decision tree is 0.68, and its backpropagation score is 0.83 multiplied by 0.68, which equals 0.564. Data points with higher backpropagation scores have a greater positive impact on the final treatment outcome and have higher priority in the influencing factor analysis.

[0101] The propagation status score is combined with the forward propagation score and the backward propagation score, and a weighted average method is used to calculate the comprehensive evaluation result. The weight allocation takes into account the specific needs of the analysis task. For the quality assessment task, the forward propagation score weight is set to 0.6, and the backward propagation score weight is set to 0.4. For the effect prediction task, the weight ratio is adjusted to 0.3 and 0.7. For a certain data point, the forward propagation score is 0.699, and the backward propagation score is 0.564. Using the quality assessment weight, the propagation status score is calculated as 0.699 multiplied by 0.6 plus 0.564 multiplied by 0.4, which equals 0.645.

[0102] The recursive feature elimination algorithm iteratively removes the feature dimensions that contribute the least to the prediction performance from the propagation state scores. The algorithm builds a prediction model based on the propagation state scores and evaluates the contribution of each feature dimension to the prediction accuracy. The model uses a random forest algorithm to identify redundant features by ranking their importance. The initial state contains 8 feature dimensions. In the first iteration, the feature with the lowest importance is removed, leaving 7 features. The change in prediction accuracy is then evaluated. If the accuracy decreases by less than 1%, the next feature is removed; if the decrease exceeds a threshold, the iteration stops and the current feature set is retained.

[0103] The feature selection process monitored the trend of the increase in prediction accuracy. A feature selection threshold of 0.5% was set. If the improvement in prediction accuracy of the remaining features was less than this threshold, further feature elimination was considered not to bring significant model improvement. After five iterations, the number of feature dimensions was reduced from eight to three, and the prediction accuracy improved from the initial 78.5% to 82.3%. The improvement in the final round was only 0.4%, which was below the threshold requirement. The three feature dimensions that were ultimately retained corresponded to data quality score, time urgency, and processing complexity. These features constituted the key influencing factors for the data points to be analyzed.

[0104] The node matching process semantically maps identified influencing factors to concept nodes in a disease-specific knowledge spectrum. This knowledge spectrum includes various node types, such as disease concepts, symptom concepts, examination concepts, and treatment concepts. Each node has standardized medical terminology and semantic encoding. Influencing factors are converted into standard medical concepts using natural language processing (NLP). For example, "data quality score" corresponds to the "diagnostic reliability" node in the knowledge spectrum, "urgency" corresponds to the "emergency response timeliness" node, and "processing complexity" corresponds to the "treatment difficulty" node. The matching process employs a combination of semantic similarity calculation and expert rule base verification to ensure the accuracy of the mapping relationships.

[0105] Importance updates adjust the weights of knowledge hierarchy nodes based on the contribution values ​​of influencing factors, maintaining a base importance score for each node and dynamically updating it according to the contribution values ​​of newly discovered influencing factors. The update strategy employs an exponential smoothing method; the new importance score equals the original importance score multiplied by a decay factor, plus the contribution value multiplied by the learning rate. For example, the original importance score of the "Diagnostic Reliability" node was 0.72, with a corresponding influencing factor contribution value of 0.35. With a decay factor set to 0.9 and a learning rate of 0.1, the updated importance score is 0.72 multiplied by 0.9 plus 0.35 multiplied by 0.1, equaling 0.683.

[0106] Semantic association calculation is based on the quantification of ontological and statistical association relationships between nodes in the knowledge spectrum. Ontological relationships include structured relationships such as parent-child, sibling, and association relationships, each with a predefined weight value. Statistical association relationships are calculated by analyzing the co-occurrence frequency and relevance of node concepts in historical data. A strong association exists between the "Diagnostic Reliability" node and the "Treatment Effectiveness" node, with a weight of 0.78; the association weight with the "Emergency Response Timeliness" node is 0.65; and the association weight with the "Treatment Difficulty" node is 0.43.

[0107] The knowledge propagation weight comprehensively considers the semantic association strength and network topology characteristics. The calculation process adopts the message passing mechanism of graph neural networks, and the propagation weight of a node is jointly affected by the importance of its neighboring nodes and the edge weights. The propagation weight is equal to the weighted sum of the products of the importance of all neighboring nodes and their corresponding edge weights. Weight normalization ensures numerical stability. The propagation weight of the "treatment effect" node considers its 12 connected neighboring nodes, and the weighted sum is 2.47. After normalization, the propagation weight is 0.206.

[0108] The weighted propagation process simulates the diffusion mechanism of knowledge in a phylogenetic network, executing multiple rounds of propagation iterations. In each iteration, the importance of each node is influenced by its neighbors. The propagation rule adopts a linear combination approach: the new importance of a node equals its own importance multiplied by a retention factor, plus the sum of the products of the importance of all neighboring nodes and their propagation weights. The retention factor is set to 0.7 to ensure that nodes retain some of their original importance. After 5 rounds of propagation iterations, the distribution of node importance in the network tends to stabilize, forming a new knowledge weight distribution.

[0109] The calculation of knowledge increment for specific diseases involves quantitative analysis by comparing the changes in the importance of nodes before and after dissemination. The increment value equals the difference between the importance after dissemination and the importance before dissemination; a positive value indicates an increase in node importance, while a negative value indicates a decrease. The importance of the "Emergency Timeliness" node increased from 0.64 to 0.71, with a knowledge increment of 0.07. The importance of the "Drug Dosage Control" node decreased from 0.52 to 0.48, with a knowledge increment of -0.04. The distribution of knowledge increments for all nodes was statistically analyzed to identify conceptual domains with significant changes in importance.

[0110] The knowledge system update process applies the calculated disease-specific knowledge increments to the existing knowledge structure. The update strategy includes multiple aspects such as node weight adjustment, edge weight correction, and discovery of new relationships. For nodes with an absolute knowledge increment greater than 0.05, their weight configuration in knowledge reasoning and decision support is adjusted. For newly discovered high-strength relationships, corresponding edge connections are added to the knowledge system. The updated knowledge system can provide more accurate knowledge support in the next round of emergency response decision-making, forming a continuously optimizing knowledge evolution mechanism.

[0111] This invention relates to an AI-based multi-center specialized emergency care collaboration and data closed-loop system, the system comprising: The first unit is used to extract the entire process information of diagnosis and treatment from historical emergency cases of specific diseases in multiple hospitals, integrate the entire process information of diagnosis and treatment into a specific disease knowledge spectrum according to semantic association, establish mapping rules between disease development path and treatment plan based on the specific disease knowledge spectrum, and form a standard specification for the diagnosis and treatment of specific diseases. The second unit is used by hospitals to generate disease-specific emergency data collection templates based on the disease-specific diagnosis and treatment standards and specifications, collect disease-specific emergency data, establish a decentralized node network among multiple hospitals, and synchronize the medical resource status of each node and the disease-specific emergency data to the decentralized node network in real time; calculate the local score of each node based on the medical resource status, construct the geographical location of the node within the patient transfer range and the level of disease-specific treatment as constraints, and use a federated learning method to combine the local scores and the constraints to determine the optimal hospital; The third unit is used to extract feature patterns from the disease-specific emergency data, identify influencing factors affecting treatment outcomes from the feature patterns, perform correlation analysis between the influencing factors and prognostic results, quantify the contribution of different influencing factors to the prognostic results, establish a data traceability chain for the disease-specific emergency data, use the traceability chain to perform backtracking analysis, locate the influencing factors, and update the disease-specific knowledge spectrum based on the location results and correlation analysis results.

[0112] 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.

[0113] 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.

[0114] 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.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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. An AI-based multi-center specialized emergency care collaboration and data closed-loop method, characterized in that, include: Information on the entire diagnosis and treatment process is extracted from historical emergency cases of specific diseases from multiple hospitals. This information is then integrated into a specific disease knowledge system based on semantic associations. Based on this specific disease knowledge system, mapping rules between disease development paths and treatment plans are established to form standard norms for the diagnosis and treatment of specific diseases. Based on the aforementioned disease diagnosis and treatment standards and specifications, the hospital generates a disease emergency data collection template, collects disease emergency data, establishes a decentralized node network among multiple hospitals, and synchronizes the medical resource status of each node and the disease emergency data to the decentralized node network in real time. Based on the status of medical resources, the local scores of each node are calculated. The geographical location of the nodes within the patient transfer range and the level of specialized disease treatment are constructed as constraints. The federated learning method is used to combine the local scores and the constraints to obtain the optimal hospital. Feature patterns are extracted from the disease-specific emergency data, and influencing factors affecting treatment outcomes are identified from these feature patterns. Correlation analysis is performed between these influencing factors and prognostic results to quantify the contribution of different influencing factors to the prognostic results. A data traceability chain is established for the disease-specific emergency data, and backtracking analysis is performed using the traceability chain to locate the influencing factors. Based on the location results and correlation analysis results, the disease-specific knowledge spectrum is updated.

2. The method according to claim 1, characterized in that, Information on the entire diagnosis and treatment process is extracted from historical emergency cases of specific diseases from multiple hospitals. This information is then integrated into a disease-specific knowledge system based on semantic associations. Mapping rules between disease progression paths and treatment plans are established based on this knowledge system, forming standardized guidelines for disease diagnosis and treatment, including: Word segmentation is performed on historical emergency medical records from multiple specialized hospitals. Medical entities in the segmentation results are identified, and each medical entity is labeled with a timestamp. Medical entities with adjacent timestamps are sequentially connected to form a time-series diagnosis and treatment chain. Entity vectors are generated based on the context information of each medical entity. The cosine similarity between the entity vectors is calculated to obtain the semantic association strength between entities. Semantic connections are established between entity pairs whose semantic association strength is greater than a preset semantic threshold to obtain a disease-specific knowledge spectrum. The state transition frequency between any two medical entities in the time-series diagnosis and treatment chain is calculated, and the state transition frequency is integrated into the semantic connection in the disease-specific knowledge spectrum to obtain a disease-specific development network. Clinical intervention nodes are marked in the disease-specific development network, and the predecessor entities, successor entities and their connection relationships of the clinical intervention nodes are extracted to obtain the mapping rules between disease development paths and treatment plans. Based on the mapping rules, corresponding treatment plan selection criteria are configured for each clinical intervention node to form a disease-specific diagnosis and treatment standard specification.

3. The method according to claim 1, characterized in that, Based on the aforementioned disease-specific diagnosis and treatment standards and specifications, the hospital generates a disease-specific emergency data collection template, collects disease-specific emergency data, establishes a decentralized node network among multiple hospitals, and synchronizes the medical resource status of each node and the disease-specific emergency data to the decentralized node network in real time, including: The data items in the specific disease diagnosis and treatment standards and specifications are classified, and the frequency of use of each type of data item in the historical specific disease emergency cases is counted to obtain the information importance. Data items with information importance greater than a preset importance threshold are selected to form a specific disease emergency data collection template. In the disease-specific emergency data collection template, numerical range constraint rules and logical relationship verification rules are set, and the diagnosis and treatment data that meet the numerical range constraint rules and logical relationship verification rules are marked as disease-specific emergency data. Digital certificates are created for multiple hospitals, and connections are established between the hospitals based on these digital certificates to form a decentralized node network. Each hospital is treated as a node, and the computing resource status value and network connection quality value of each node are statistically analyzed to generate node priorities. Bed occupancy data, medical equipment status data, and expert resource distribution data of each node are statistically analyzed as medical resource status. According to the node priorities of each node, the medical resource status and the emergency medical data for specific diseases are synchronized to the decentralized node network.

4. The method according to claim 1, characterized in that, Based on the status of medical resources, local scores are calculated for each node. The geographical location of nodes within the patient transfer area and the level of specialized disease treatment are used as constraints. A federated learning method is employed, combining the local scores and the constraints, to determine the optimal hospital, which includes: The medical resource status is mapped to a high-dimensional feature space to obtain a feature mapping matrix. The feature mapping matrix is ​​then multiplied by a preset weight vector to obtain the local score of each node. The transfer range is determined based on the patient's location information. The transfer range is used as a distance constraint. The arrival time is calculated based on the geographical location information of each node within the transfer range to obtain a time constraint. The utilization rate of the specialized disease treatment equipment at each node is used as a treatment constraint. Candidate nodes are selected based on the distance constraint, the time constraint, and the treatment constraint. On the candidate nodes, the local scores and constraints are used as node evaluation data. The node evaluation data is encrypted using a homomorphic encryption method and then sent. The encrypted data sent by each candidate node is received. The local encryption gradient of each node is calculated based on the encrypted data and aggregated to obtain the global encryption gradient. The federated comprehensive score is calculated based on the local encryption gradient and the global encryption gradient. The candidate node with the highest federated comprehensive score is determined as the optimal hospital.

5. The method according to claim 4, characterized in that, The local encryption gradient of each node is calculated based on the encrypted data and aggregated to obtain the global encryption gradient. The federated comprehensive score is then calculated based on the local encryption gradient and the global encryption gradient, including: The local encryption gradient is obtained by calculating the first derivative of the loss function of each node based on the encrypted data. The local encryption gradient of each candidate node is multiplied by the proportion of the number of samples in the candidate node and then summed to obtain the global encryption gradient. The node contribution is obtained by calculating the cosine similarity between the local encryption gradient of each candidate node and the global encryption gradient. Based on the convergence direction of the global encryption gradient, the directional consistency of the local encryption gradient of each candidate node is calculated to obtain the data quality score. The node contribution and the data quality score are combined to obtain the federated comprehensive score.

6. The method according to claim 1, characterized in that, Feature patterns are extracted from the specific emergency medical data. Factors influencing treatment outcomes are identified from these feature patterns. Correlation analysis is performed between these factors and prognostic results to quantify the contribution of different factors to the prognostic outcome, including: The special emergency data is subjected to wavelet transform to obtain time-series features. The feature significance of each scale parameter is calculated based on the rate of change of the time-series features. The scale parameter with the largest feature significance is selected as the optimal scale parameter. A multi-dimensional feature matrix is ​​constructed based on the optimal scale parameter. Feature vectors with a cumulative contribution rate that reaches a preset feature retention threshold are selected from the multi-dimensional feature matrix to construct a feature dimensionality reduction matrix. The multi-dimensional feature matrix and the feature dimensionality reduction matrix are multiplied to obtain the feature pattern. Calculate the regression coefficient between the feature pattern and the prognostic result. Sort the feature components in the feature pattern in descending order according to the absolute value of the regression coefficient. The feature components with a regression coefficient absolute value greater than a preset coefficient screening threshold in the sorting result are identified as influencing factors. Calculate the joint probability distribution and marginal probability distribution of the influencing factors and the prognostic result. Calculate the correlation strength of the influencing factors based on the joint probability distribution and the marginal probability distribution. The influencing factors are combined to generate a feature subset sequence. The prediction bias of each feature subset on the prognostic result is calculated based on the correlation strength of the influencing factors. The contribution value of the influencing factors is calculated based on the prediction bias.

7. The method according to claim 1, characterized in that, A data traceability chain is established for the aforementioned disease-specific emergency data. Backtracking analysis is performed using this traceability chain to locate the influencing factors. Based on the location results and correlation analysis results, the disease-specific knowledge spectrum is updated, including: The specialized emergency data is classified and encoded to obtain a data identifier sequence. A data dependency graph is constructed based on the data identifier sequence. Data flow paths are extracted from the data dependency graph to form a data traceability chain. The state vector of the data point to be analyzed is extracted from the data traceability chain. The propagation weight of the state vector and the historical state vector is calculated to obtain the forward propagation score. The tracing weight of the state vector and the target state vector is calculated to obtain the backward propagation score. The forward propagation score and the backward propagation score are combined to obtain the propagation state score. A recursive feature elimination algorithm is used to remove redundant features from the propagation state score. The calculation is iteratively performed until the increase in the prediction accuracy of the remaining features for the prognostic results is lower than a preset feature screening threshold, thereby obtaining the influencing factors corresponding to the data point to be analyzed. The influencing factors are matched with nodes in the disease-specific knowledge spectrum. The importance of the corresponding nodes is updated based on the contribution value of the influencing factors. The knowledge propagation weight is calculated according to the semantic relationship between nodes. The knowledge propagation weight is used to propagate the importance of nodes in a weighted manner to obtain the disease-specific knowledge increment. The disease-specific knowledge spectrum is updated according to the disease-specific knowledge increment.

8. An AI-based multi-center specialized emergency care collaboration and data closed-loop system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The first unit is used to extract the entire process information of diagnosis and treatment from historical emergency cases of specific diseases in multiple hospitals, integrate the entire process information of diagnosis and treatment into a specific disease knowledge spectrum according to semantic association, establish mapping rules between disease development path and treatment plan based on the specific disease knowledge spectrum, and form a standard specification for the diagnosis and treatment of specific diseases. The second unit is used by hospitals to generate disease-specific emergency data collection templates based on the disease-specific diagnosis and treatment standards and specifications, collect disease-specific emergency data, establish a decentralized node network among multiple hospitals, and synchronize the medical resource status of each node and the disease-specific emergency data to the decentralized node network in real time. Based on the status of medical resources, the local scores of each node are calculated. The geographical location of the nodes within the patient transfer range and the level of specialized disease treatment are constructed as constraints. The federated learning method is used to combine the local scores and the constraints to obtain the optimal hospital. The third unit is used to extract feature patterns from the disease-specific emergency data, identify influencing factors affecting treatment outcomes from the feature patterns, perform correlation analysis between the influencing factors and prognostic results, quantify the contribution of different influencing factors to the prognostic results, establish a data traceability chain for the disease-specific emergency data, use the traceability chain to perform backtracking analysis, locate the influencing factors, and update the disease-specific knowledge spectrum based on the location results and correlation analysis results.

9. 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 7.

10. 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 7.

Images