Hospital intelligent operation management system based on large model

Abstract

This invention relates to a large-scale model-based intelligent hospital operation and management system, specifically in the field of intelligent hospital operation and management. This system improves the intelligence and accuracy of medical triage through multimodal alignment, twin matching, feedback optimization, and confidence redirection modules. It combines patients' chief complaint text and physiological data, accurately matching departmental resources through a cross-modal attention mechanism to avoid triage errors and resource waste. The twin neural network dynamically adjusts based on departmental capabilities, optimizing triage decisions and improving the model through a feedback optimization mechanism. When the confidence redirection module detects discrepancies between physiological signs and chief complaints, it automatically adjusts departmental allocation to ensure timely treatment for patients. Overall, this improves the efficiency of emergency resource allocation and reduces misdiagnosis and delays.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and management of hospitals, and more specifically, to an intelligent operation and management system for hospitals based on a large model. Background Technology

[0002] During peak seasons in the emergency department, especially during flu season, the number of patients can exceed 2,000 per day. Patients' needs are extremely urgent and complex. To effectively manage these high-demand patients, hospitals typically use self-service terminals or mobile systems, allowing patients to obtain initial triage information by inputting their self-described complaints. However, due to the large number of patients during peak emergency periods and the diverse range of symptoms, patients often express their specific symptoms using unstructured language. For example, a patient might only provide a vague description such as "headache and fever for three days, difficulty breathing." This unstructured text information increases the difficulty for artificial intelligence systems to accurately understand the patient's condition. In such cases, the patient's complaint may not accurately match the appropriate department or specialist, leading to incorrect triage decisions, misallocation of medical resources, overload of specialties, waste of emergency department resources, and even delays in patient treatment.

[0003] In existing medical triage technologies, most systems still rely on rule engines based on keyword matching for triage. While these systems can handle a certain number of common symptoms, their accuracy typically only reaches around 70% when faced with complex, unstructured text input. Keyword matching triage systems are easily affected by the ambiguity of patients' symptom descriptions, failing to accurately capture disease characteristics from symptoms. For example, the symptom "difficulty breathing" can involve multiple different departments, such as cardiology and pulmonology, but current systems often only assign based on surface descriptions, ignoring patients' historical health data or disease progression, resulting in incorrect department assignments or misallocation of medical resources. Furthermore, existing triage systems fail to synchronize with patients' health records in real time, making it impossible to effectively reference patients' historical medical history and chronic disease information during the triage process. Due to these problems, patients are often misassigned to unsuitable departments, leading to duplicate examinations, ineffective transfers, and waste of medical resources. On average, each misallocation can result in an additional cost of approximately 200 yuan, and further exacerbates the overload of emergency departments.

[0004] Therefore, existing triage technologies urgently need improvement, especially in symptom understanding, correlation of patient historical data, and intelligent decision-making models. To address these issues, future triage systems need to combine clinical knowledge graphs with patients' full-cycle health data to build dynamic triage decision-making models, thereby improving the accuracy and efficiency of the triage system. This will help reduce resource waste caused by incorrect triage, optimize patient flow, reduce the overload pressure on emergency departments, and ultimately improve the quality and efficiency of medical services. Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing a hospital intelligent operation and management system based on a large model. By introducing modules such as multimodal alignment, twin matching, feedback optimization, and confidence redirection, it can effectively improve the accuracy and intelligence level of triage decisions, thereby solving the problems mentioned in the background art.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: it specifically includes: a multimodal alignment module, a twin matching module, a feedback optimization module, and a confidence redirection module; The multimodal alignment module, when receiving patient complaint text and real-time physiological sign data, uses a cross-modal attention mechanism to perform spatial alignment operations on temporal physiological features and complaint semantic vectors, generating a cross-modal feature vector that integrates patient health status information; The twin matching module, based on the cross-modal feature vector and the real-time updated department capability matrix, calculates the decision similarity parameters between patient features and department resource status through a shared architecture twin neural network, and outputs the target department number in combination with a preset triage mismatch loss matrix. The feedback optimization module, after obtaining the final diagnosis result, constructs a multi-dimensional feedback vector from the triage result deviation value, resource waste, and delay time. It dynamically adjusts the weight parameters of the Siamese neural network through a federated aggregation protocol and dynamically updates the modal conflict detection threshold based on the local delay time value of each node. Specifically, it adjusts the scaling ratio of the historical threshold through a nonlinear error function and feeds the updated modal conflict detection threshold back to the confidence redirection module. The confidence redirection module, when it detects that the difference in confidence between physiological signs and the chief complaint text exceeds a set threshold, triggers a graph neural network to traverse the symptom-department knowledge graph and generate a redirection instruction that forcibly covers the target department number; In a preferred embodiment, the specific operation of spatially aligning temporal physiological features and the subject complaint semantic vector using a cross-modal attention mechanism in the multimodal alignment module is as follows: A1. Receive physiological sign data collected by wearable devices to form a time-series physiological sign sequence, extract waveform change patterns of the time-series physiological sign sequence through a time-convolutional network model, and generate a physiological feature vector; the time-convolutional network model adopts a causal convolutional kernel structure.

[0007] A2. Receive patient complaint text data, extract semantic features from the complaint text data using a clinical-specific text encoder, and generate a fixed-dimensional clinical semantic vector; the clinical-specific text encoder is based on a medical knowledge pre-training model to reinforce the encoding of medical terms in the symptom description.

[0008] A3. Using the physiological feature vector as the dynamic query benchmark and the clinical semantic vector as the alignment target, calculate the feature association distribution between each time node in the physiological feature vector and the clinical semantic vector, and perform a weighted fusion operation on the clinical semantic vector according to the feature association distribution to generate a spatially aligned semantic representation vector; the feature association distribution is generated by the vector dot product operation of the trainable projection matrix.

[0009] A4. Perform a ternary combination operation on the physiological feature vector and the spatially aligned semantic representation vector, specifically including: directly concatenating the physiological feature vector and the semantic representation vector as independent feature components, calculating the element-wise product between the two as an interactive feature component, and finally forming a cross-modal feature vector that integrates the patient's multi-dimensional health status information.

[0010] In a preferred embodiment, the calculation and weighted fusion operation of the feature association distribution specifically involves: B1. The physiological feature vector is mapped to the attention query vector through a linear transformation matrix, and the clinical semantic vector is mapped to the attention key vector. B2. Calculate the cosine similarity between the attention query vector and the attention key vector, and transform it into an association weight distribution using an exponential normalization function; B3. Based on the aforementioned associated weight distribution, perform weight allocation operations on the original components of the clinical semantic vector to generate a semantic representation vector that reflects temporal dependence. The associated weight allocation operation satisfies the following: when abnormal fluctuations occur in the waveform of physiological signs, the weight of related symptom description segments is automatically strengthened; when patients describe their past medical history, text features unrelated to the current signs are weakened.

[0011] In a preferred embodiment, the specific steps in the twin matching module for calculating the decision similarity parameters between patient characteristics and departmental resource status using a shared architecture twin neural network include: C1. Input the cross-modal feature vector into an independent patient feature encoder, extract the patient state embedding vector through a depthwise separable convolutional layer and a self-attention mechanism, and output a dense vector of fixed dimension to represent the patient's overall health status. C2. Input the real-time department capability matrix into the department feature encoder with shared parameters. This encoder reuses the same set of convolutional kernel weights for all departments to generate the resource state embedding vector for each department. C3. Dynamically integrate department saturation and expert resource status: Construct a three-dimensional weight tensor. The first dimension is a constant bias term, the second dimension is the negative logarithmic scaling value of the real-time occupancy rate of the department, and the third dimension is the Boolean amplification coefficient of the on-duty status of experts for the target disease. Perform element-wise multiplication operations on the weight tensor and the department resource status embedding vector to generate a weighted department feature vector. C4. Calculate the hyperbolic geometric space similarity between the patient state embedding vector and the weighted department feature vector: First, perform hyperbolic secant normalization on the two vectors, then calculate the reciprocal of their hyperbolic cosine distance as the basic similarity value, and finally multiply by the exponential decay factor of the angle difference between the two in hyperbolic space to obtain the nonlinear decision similarity parameter.

[0012] In a preferred embodiment, the negative logarithmic scaling value of the department's real-time occupancy rate is calculated as follows: The average of the department's bed occupancy rate and equipment working time within the past ten minutes is used as the real-time occupancy rate benchmark. The natural logarithm of this benchmark is multiplied by a preset penalty coefficient, and the resulting negative value is used as the weight of the second dimension. This calculation significantly reduces the similarity of departments with occupancy rates exceeding the threshold. For every certain percentage increase in occupancy rate, a curvature increase effect of the logarithmic function is triggered, which accelerates the decline in the matching priority of high-load departments.

[0013] In a preferred embodiment, the specific operation of outputting the target department number in conjunction with the preset triage mismatch loss matrix is ​​as follows: The decision similarity parameters for each department are scaled by a temperature coefficient, and the department matching probability distribution is generated by an exponential normalization function. The probability distribution matrix and the triage mismatch loss matrix are subjected to Hadamard product operation to obtain the expected loss tensor. Each element of the loss tensor represents the expected risk value of triaging the patient to the corresponding department. The department number with the smallest value in the expected loss tensor is selected as the final output target. The principle for setting the triage mismatch loss matrix is ​​that the loss value of triaging critical illnesses to non-emergency departments is at least five times the loss value of ordinary triage errors.

[0014] In a preferred embodiment, the specific operation of constructing a multidimensional feedback vector in the feedback optimization module is as follows: First, the triage result deviation value is defined as the mismatch indicator between the actual diagnostic department number and the triage decision department number. If the two are inconsistent, an integer value of 1 is assigned; if they are consistent, an integer value of 0 is assigned. Second, the logarithmic compression value of resource waste is calculated. The total value of resource waste is proportionally converted using the cost of a single standard examination as the benchmark unit, and the natural logarithm is taken after adding one to the proportional value. Finally, the exponential saturation function value of the delay time is calculated. The actual delay time is input into the negative exponential function of the preset decay factor, and the delay risk saturation is obtained by subtracting the function output value from the constant 1. The above three values ​​are combined into a three-dimensional column vector in the order of diagnostic difference, resource loss, and delay risk to form a multi-dimensional feedback vector data structure. The resource loss dimension of the multidimensional feedback vector satisfies the following condition: when the total value of resource waste exceeds twenty times the benchmark unit, the growth rate of the logarithmic compression value decreases to less than one-fifth of the original rate.

[0015] In a preferred embodiment, the specific operation of the federated aggregation protocol dynamically adjusting the weights is as follows: D1. Calculate the node contribution weight coefficient: Based on the Euclidean norm value of the multidimensional feedback vector uploaded by each medical institution node, compress it to the zero-to-one interval through the hyperbolic tangent function as the basic weight component; at the same time, obtain the model timeliness deviation duration of the node, multiply the duration by the decay factor and take the negative value of the natural exponent as the timeliness decay component; multiply the basic weight component and the timeliness decay component to obtain the final node contribution weight. D2. Perform secure aggregation operation: Each node calculates the gradient tensor of the Siamese neural network parameters in an encrypted environment, performs tensor product operation on the coupling matrix constructed by the node's own multidimensional feedback vector, and then multiplies it with the node's contribution weight coefficient to form a weighted encrypted gradient; all nodes decrypt and aggregate the weighted encrypted gradient through a secure multi-party computation protocol, and calculate the arithmetic mean as the global gradient update amount. D3. Apply dynamic parameter constraints: Calculate the dynamic pruning threshold based on the Manhattan norm of the node's multidimensional feedback vector. This threshold is negatively correlated with the feedback strength. When the Manhattan norm exceeds two, the pruning threshold is linearly reduced to less than 40% of the standard threshold. Use this threshold to truncate the global gradient update and add the truncated gradient to the original weights of the Siamese neural network to complete the parameter update. The specific operation of dynamically updating the modal conflict detection threshold is as follows: The final diagnostic delay duration before the current time of this node is obtained as the core input. The delay duration is divided by a preset time scale parameter to obtain a scaling factor. The scaling factor is input into a Gaussian error function for nonlinear mapping. The mapping result is multiplied by a sensitivity coefficient and then incremented by one to form a scaling factor. The scaling factor is multiplied by the historical modal conflict detection threshold to generate an updated detection threshold.

[0016] In a preferred embodiment, the specific operation of detecting the confidence difference between physiological signs and the chief complaint text in the confidence redirection module includes: First, the temporal entropy value corresponding to the physiological sign temporal feature vector generated by the multimodal alignment module is obtained. This temporal entropy value is obtained by calculating the information entropy of the probability distribution of each dimension of the feature vector. At the same time, the semantic focus degree corresponding to the semantic vector of the complaint text is obtained. This semantic focus degree is determined by extracting the maximum value component in the semantic vector. The temporal entropy value and the semantic focus degree are subtracted to obtain the preliminary difference value. The absolute value of the preliminary difference value is processed to obtain the absolute difference amount. Finally, the absolute difference amount is compared with the dynamically updated modal conflict detection threshold. When the absolute difference amount exceeds the threshold, the graph traversal operation is triggered.

[0017] In a preferred embodiment, the operation of triggering the graph neural network to traverse the symptom-department knowledge graph to generate redirection instructions includes: A patient query node is constructed with physiological signs and chief complaint semantic vectors as attributes. This patient query node is injected into a symptom-department knowledge graph architecture. Vector state propagation is performed along the medical relationship path through a multi-layer graph convolutional network. During each layer of propagation, the association weight between the current department node and adjacent symptom nodes is calculated. This weight combines the relationship type embedding factor and the patient feature similarity. After completing the specified number of propagation layers, the final state vectors of all department nodes are summarized. The matching probability distribution of each department is calculated based on the state vectors. The department number with the highest value in the probability distribution is identified. When the probability ratio between the highest value department and the original target department exceeds the critical coefficient, a triple instruction containing the new department number, coverage strength, and medical evidence chain is generated to forcibly cover the original triage decision result.

[0018] The beneficial effects of this invention are as follows: By introducing modules such as multimodal alignment, twin matching, feedback optimization, and confidence redirection, the system can effectively improve the accuracy and intelligence of triage decisions. Through a cross-modal attention mechanism, the system can integrate patient complaint text with real-time physiological sign data to achieve accurate symptom understanding and department matching, thereby avoiding triage errors caused by ambiguous descriptions. The twin neural network optimizes resources based on the department capability matrix, reducing resource misallocation and medical waste. At the same time, the feedback optimization module further improves the efficiency and flexibility of the triage system by dynamically adjusting network parameters and modal conflict detection thresholds. The confidence redirection module can automatically trigger the graph neural network for redirection when encountering inconsistencies between physiological signs and complaint text, ensuring that patients receive the correct departmental diagnosis and treatment, avoiding overload or delayed treatment in the emergency department. Overall, the system can optimize resource allocation, reduce medical waste, and improve the efficiency and accuracy of emergency treatment. Attached Figure Description

[0019] Figure 1This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0022] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0023] Example 1 This embodiment provides, for example Figure 1-2 The present invention relates to an intelligent hospital operation and management system based on a large model, which specifically includes: a multimodal alignment module, a twin matching module, a feedback optimization module, and a confidence redirection module; The multimodal alignment module, when receiving patient complaint text and real-time physiological sign data, uses a cross-modal attention mechanism to perform spatial alignment operations on temporal physiological features and complaint semantic vectors, generating a cross-modal feature vector that integrates patient health status information; The twin matching module, based on cross-modal feature vectors and real-time updated department capability matrix, calculates decision similarity parameters between patient features and department resource status through a shared architecture twin neural network, and outputs the target department number by combining a preset triage mismatch loss matrix. The feedback optimization module, after obtaining the final diagnosis result, constructs a multi-dimensional feedback vector from the triage result deviation value, resource waste, and delay time. It dynamically adjusts the weight parameters of the Siamese neural network through a federated aggregation protocol and dynamically updates the modal conflict detection threshold based on the local delay time value of each node. Specifically, it adjusts the scaling ratio of the historical threshold through a nonlinear error function and feeds the updated modal conflict detection threshold back to the confidence redirection module. In the medical triage closed loop, after the patient completes the final diagnosis, dynamic feedback needs to be constructed based on the deviation between the triage decision and the actual department, resource consumption, and delay time. The parameters of the Siamese neural network are optimized through a federated learning framework. Traditional methods only correct classification labels and cannot integrate heterogeneous indicators such as resource consumption and temporal risk, resulting in the model emphasizing accuracy while ignoring clinical costs. The confidence redirection module triggers a graph neural network to traverse the symptom-department knowledge graph when the confidence difference between physiological signs and the patient's complaint exceeds a set threshold. This generates a redirection instruction that forcibly covers the target department number. In medical triage systems, when there is a significant confidence difference between physiological sign monitoring data and the patient's complaint (e.g., signs indicate acute myocardial infarction but the complaint only describes "stomach discomfort"), traditional rule engines cannot effectively handle such modal conflicts. This solution addresses the clinical challenge of cross-modal decision-making conflicts through a knowledge graph reasoning mechanism with dynamic threshold control.

[0024] In this embodiment, the multimodal alignment module, which uses a cross-modal attention mechanism to spatially align temporal physiological features with the subject complaint semantic vector, is described as follows: A1. Receive physiological data collected by wearable devices. ( Indicates an adjustable time window. Including heart rate, blood oxygen, etc. The time-series physiological signs sequence is formed by using dimensional indicators. The waveform variation patterns of this sequence are extracted using a temporal convolutional network model. The feature extraction formula is as follows: ; in, This indicates that the convolutional kernel is trainable, and ( Indicates the kernel width. (representing feature dimension) Indicates causal convolution operation (ensuring) The output at any time depends only on data), Represents the convolution bias term. This represents the output physiological feature vector, and its dimension is... ( Generate physiological feature vectors (characterization) (Evolutionary patterns of physiological states within the window); the temporal convolutional network model adopts a causal convolutional kernel structure to ensure that the output features depend only on historical time window data and do not leak future information.

[0025] A2. Receive patient complaint text data (Unstructured string sequence), semantic features of the chief complaint text data are extracted using a clinically-specific text encoder. The extraction formula is as follows: ; in, This represents the patient's chief complaint text (the original symptom description input source), whose value is a string sequence (such as "headache for three days"). This indicates the classification marker (the core identifier that aggregates the semantics of the entire sentence). Represents clinical semantic vectors (encoding the medical meaning features of symptoms). This indicates a clinically specific text encoder that generates fixed-dimensional clinical semantic vectors. ( (This represents the hidden layer dimension, set to 768); the clinical text encoder is based on a medical knowledge pre-trained model to reinforce the encoding of medical terms in symptom descriptions.

[0026] A3. Using the physiological feature vector as the dynamic query benchmark and the clinical semantic vector as the alignment target, calculate the feature association distribution between each time point in the physiological feature vector and the clinical semantic vector. The formula is as follows: ; in, Represents the temporal feature vector of physiological signs (dimension h). This represents a trainable projection matrix (which maps vectors to the attention space). This represents the scaling factor (to prevent the gradient from vanishing due to excessively large dot product values). The physiological-semantic association weight matrix (after Softmax normalization) is represented by this matrix. A weighted fusion operation is then performed on the clinical semantic vectors based on the feature association distribution to generate a spatially aligned semantic representation vector. Its expression is as follows: ; in, This represents the trainable value projection matrix (reconstructing semantic vectors). This represents an aligned semantic vector, i.e., a symptom representation that integrates temporal information of vital signs. This formula breaks through the traditional splicing / addition fusion method and uses a trainable correlation matrix. accomplish: Feature decoupling: When the chief complaint is vague (e.g., "chest discomfort"), Automatically enhance semantic segments related to current abnormal signs (such as ST segment elevation); Status tracing: If the patient mentions "vomiting yesterday," but their current physical signs are normal, This design weakens the semantic vector weights, enabling the system to be interpretable for clinical decisions and providing quantitative evidence for confidence conflict detection; the feature association distribution is generated through vector dot product operations of trainable projection matrices.

[0027] A4. Perform a ternary combination operation on the physiological feature vector and the spatially aligned semantic representation vector, specifically including: directly concatenating the physiological feature vector. and semantic representation vector As independent feature components, the element-wise product between the two is calculated as the interactive feature component, and the formula is as follows: ; in, It represents the physiological feature vector (temporal physiological features extracted by a temporal convolutional network (TCN)) and characterizes the evolution of the patient's physiological state during the window period (such as the past 1 hour) (e.g., heart rate and blood oxygen trends). The spatially aligned semantic representation vector (the chief complaint semantic features adjusted by attention alignment) is used to resolve temporal conflicts between symptom description and real-time signs (such as "past chest pain" vs. current normal ECG). This represents a cross-modal feature vector (a fused representation of patient health status, dimension 3h), used as the core input for downstream triage decisions, while preserving modal independence and interaction features. This represents a vector concatenation operation (connecting multiple vectors along the column direction), used to merge independent feature components (original physiological and semantic information) and avoid information dilution. This represents the Hadamard product (element-wise vector multiplication), used to generate interactive feature components (such as the enhanced association between a sudden drop in blood oxygen and the text "dyspnea"), ultimately forming a cross-modal feature vector that integrates multi-dimensional health status information of the patient. h represents the cardinality of the feature dimensions, i.e., the basic dimension of all vectors (set to 128 or 256), used to control model complexity and ensure feature dimension compatibility at each stage. Furthermore, by preserving independent features (concatenation) and enhancing interactive features (Hadamard product), the shortcomings of traditional single-path fusion are addressed, enabling the fusion to work even when the patient's description is vague (e.g., "uncomfortable") but blood oxygen levels suddenly drop. It automatically amplifies the correlation of key vital signs, enabling the subsequent triage module to accurately locate the ICU (rather than the general internal medicine department). In summary, this module addresses the following key issues through four-stage operations: temporal convolutional compression, clinical semantic encoding, dynamic attention alignment, and interaction feature fusion: Temporal inconsistency: Through causal convolution and the correlation matrix Align the patient's described "past symptoms" with the "current signs" collected by the wearable device on the timeline; Modal heterogeneity: using Hadamard product Explicit modeling of the interaction effect between physiological indicators and symptom description; Downstream compatibility: Output vector It also includes independent features ( ) and joint features ( The results can be flexibly called by subsequent modules; The calculation and weighted fusion operation of the feature association distribution are as follows: B1. The physiological feature vector is mapped to the attention query vector through a linear transformation matrix, and the clinical semantic vector is mapped to the attention key vector. B2. Calculate the cosine similarity between the attention query vector and the attention key vector, and transform it into an association weight distribution using an exponential normalization function; B3. Based on the correlation weight distribution, perform weight allocation operation on the original components of the clinical semantic vector to generate a semantic representation vector that reflects the temporal dependency. The associated weight allocation operation satisfies the following: when abnormal fluctuations occur in the waveform of physiological signs, the weight of related symptom description segments is automatically strengthened; when patients describe their past medical history, text features unrelated to the current signs are weakened.

[0028] In this embodiment, the twin matching module, specifically the steps of which use a shared architecture twin neural network to calculate the decision similarity parameters between patient characteristics and departmental resource status, include: C1, cross-modal feature vectors (Including physiological temporal features, aligned semantic features, and interaction features) Input independent patient feature encoder ,in This represents the cross-modal feature vector (multimodal data of fused patient health status generated by the preceding module), used as input to the patient feature encoder to characterize the overall health status. This represents a patient feature encoder (a standalone Transformer encoder) used to extract patient-specific features while preserving individual patient differences. The patient state embedding vector (a dense vector of dimension h) is used as a benchmark for downstream similarity calculation. It represents the projection of the patient's health status into the embedding space. The patient state embedding vector is extracted through a depthwise separable convolutional layer and a self-attention mechanism. The encoder outputs a dense vector of fixed dimension that represents the patient's overall health status. C2, Real-time Departmental Capability Matrix ( The department feature encoder represents the total number of departments (a three-dimensional tensor containing saturation, expert specialties, and equipment availability). It uses shared input parameters. ,in Indicates the first The department capability matrix (a three-dimensional tensor containing real-time resource status such as saturation and expert specialties) is used as input to the department feature encoder to characterize the department. Resource status, This represents a departmental feature encoder, specifically a Transformer encoder with shared weights (all departments reuse the same parameters), used to map resources from different departments to the same semantic space, achieving cross-departmental comparability. Indicates the first Departmental resource embedding vector (a dense vector of dimension h), representing the department The projection of the resource state into the embedding space is used for matching with the patient vector. This represents the department number index, traversing all departments ( This encoder reuses the same set of convolutional kernel weights for all departments to generate resource state embedding vectors for each department, ensuring that the resource features of different departments are comparable in the same semantic space. C3. Dynamically integrating departmental saturation and expert resource status: Constructing a three-dimensional weight tensor, the expression of which is: ; in, This represents the three-dimensional decision weight tensor (feature scaling coefficients that integrate departmental resource constraints), used to transform business rules (saturation, expert resources) into mathematical operation terms. Indicates department Real-time saturation (average bed / equipment occupancy rate over the past 10 minutes) Used to quantify resource load. This indicates that the department is overloaded. This represents the saturation penalty coefficient (set to 0.2), used to control the intensity of the effect of saturation on similarity. , Indicates the on-duty status of the expert for the target disease ( (), used to identify the availability of critical medical resources, This represents the expert resource gain coefficient (set to 1.5), used to amplify the matching weight of departments with experts on staff. This represents the expert state indicator function, when The value is 1 if the condition is met and 0 otherwise, used to map Boolean conditions to numerical values. The first dimension is a constant bias term, the second dimension is the negative logarithmic scaling value of the department's real-time occupancy rate, and the third dimension is the Boolean amplification factor of the on-duty status of the target disease specialist. This weight tensor is multiplied element-wise with the department resource status embedding vector to generate a weighted department feature vector, calculated using the following formula: ; in, This represents the Hadamard product (vector element-wise multiplication), used to perform targeted scaling (not simple concatenation or weighting) on ​​department feature vectors. This represents the weighted department feature vector (a representation of the department's state after incorporating resource constraints), which is used as the input for downstream similarity calculation. At the same time, the saturation penalty and expert resources are treated as multiplicative terms rather than additive terms, so that the feature vector of high-load departments is directly shrunk. C4. Calculate the hyperbolic geometric similarity between the patient state embedding vector and the weighted department feature vector: First, perform hyperbolic secant normalization on the two vectors, then calculate the reciprocal of their hyperbolic cosine distance as the basic similarity value, and finally multiply by the exponential decay factor of the angle difference between the two in hyperbolic space to obtain the nonlinear decision similarity parameter. The calculation formula is as follows: ; in, This represents the decision similarity parameter, i.e., patient and department. The overall matching degree (the higher the value, the higher the triage suitability). The patient state embedding vector represents a high-dimensional representation of the patient's health status (generated by cross-modal feature vector encoding). This represents the weighted department feature vector, which is the dynamic feature vector of departments that integrates saturation penalty and expert resources. This represents the hyperbolic inner product operation (a geometric operation that measures the consistency of vector direction in the Poincaré disk model). It represents the hyperbolic space norm, which is a measure of the scale of vectors in hyperbolic space (as opposed to the magnitude in Euclidean space). This represents the inverse hyperbolic cosine function, used to map the angle between vectors to a similarity value. (when the output value is at its maximum) Represents the angular characteristics of the hyperbolic space for patients / departments, i.e., the directional properties of vectors in hyperbolic space (e.g., ... (Indicates orthogonal to the characteristics of the emergency department). This represents the angle difference penalty coefficient, used to control the attenuation of similarity due to angle deviation. Its value range is The effect of this formula is that traditional cosine similarity fails with orthogonal features, while the hyperbolic kernel function... It provides a non-linear decay gradient; the exponential decay term penalizes directional differences (such as the angular deviation between cardiology features and "stroke" patients). Furthermore, when patient features... with cardiology vector Angle deviation Time (e.g., if it is actually a stroke). Forced This mathematically blocks the path to misdiagnosis (superior to traditional threshold judgment). The calculation rule for the negative logarithmic scaling value of the department's real-time occupancy rate is as follows: The average of the department's bed occupancy rate and equipment working time within the past ten minutes is used as the real-time occupancy rate benchmark. The natural logarithm of this benchmark is multiplied by a preset penalty coefficient, and the resulting negative value is used as the weight of the second dimension. This calculation significantly reduces the similarity of departments with occupancy rates exceeding the threshold. Specifically, every percentage increase in occupancy rate will trigger a sharp increase in the curvature of the logarithmic function, causing the matching priority of high-load departments to decrease rapidly. The specific steps for outputting the target department number based on the preset triage mismatch loss matrix are as follows: The decision similarity parameters for each department are scaled using a temperature coefficient, and a department matching probability distribution is generated using an exponential normalization function. The probability distribution matrix and the triage mismatch loss matrix are then subjected to a Hadamard product to obtain the expected loss tensor. Each element of this loss tensor represents the expected risk value of triaging the patient to the corresponding department. The department number with the smallest value in the expected loss tensor is selected as the final output target, and its expression is: ; in, This indicates the target department number, i.e., the final decision result (e.g., "Cardiology = 3"). Indicate patient characteristics and department The similarity parameter is used to quantify the matching degree between patients and departments (the larger the value, the higher the matching degree). This represents the mismatch loss value, i.e., the actual department. Patient triage to The risk cost (the loss value of triaging a myocardial infarction patient to the gastroenterology department is 5). Indicates department The predicted probability is obtained by transforming similarity into a probability distribution using the Softmax function. Represents the temperature coefficient, used to control the steepness of the probability distribution. The smaller the value, the closer the probability of highly similar departments is to 1). This indicates the total number of departments in the hospital, which determines the scope of the calculation (e.g., a tertiary hospital). ), Indicates department The expected loss (a decision indicator that combines clinical risk and probability) is calculated as follows: when a patient with myocardial infarction ( Similarity High time, If the number is close to 1, and the patient is triaged to the gastroenterology department ( ), then high loss value This is significantly amplified, and the system automatically avoids this decision; The principle for setting the triage mismatch loss matrix is ​​that the loss value of triaging critical illnesses to non-emergency departments is at least five times the loss value of ordinary triage errors.

[0029] In this embodiment, the feedback optimization module, specifically the operation of constructing a multi-dimensional feedback vector, needs to be explained as follows: First, the deviation value of the triage result is defined as the actual diagnostic department number. Department number for triage decision-making The mismatch indicator is assigned an integer value of 1 if the two values ​​are inconsistent, and an integer value of 0 if they are consistent; secondly, the amount of resource waste is calculated. The logarithmic compression value of (costs of duplicate inspections, ineffective transfers, etc.) is calculated by proportionally converting the total value of resource waste to the cost of a single standard inspection, adding one to this proportional value, and then taking the natural logarithm. Finally, the delay time is calculated. The exponential saturation function value is obtained by inputting the actual delay time into a negative exponential function with a preset attenuation factor, and subtracting the function output value from a constant to obtain the delay risk saturation. These three values ​​are combined into a three-dimensional column vector in the order of diagnostic difference, resource loss, and delay risk to form a multi-dimensional feedback vector data structure, the expression of which is: ; in, This represents a three-dimensional feedback vector (a heterogeneous index vector fusing diagnostic differences, resource consumption, and delay risk) used as input for downstream federated learning, with dimensions of [dimensional value missing]. , This represents a diagnostic difference indicator function, when Output 0 when the time is right; The system outputs 1 at certain times to quantify the triage accuracy Boolean value into a computable numerical value. This represents the baseline value for resource costs, i.e., the cost reference for a single standard medical procedure (such as the cost of a single CT scan). Its function is to: compress the numerical range of high resource waste (such as...) ),prevent Temporal gradient explosion Represents the natural logarithm function (mathematical function) ), used to compress the growth rate of resource waste value ( (The growth rate has slowed to below 5%) Indicates the time decay factor ( ,control The saturation rate), mapping the delay time to a risk value of [0,1) (e.g., λ=0.1: Time risk value ), Represents the exponential decay function (mathematical function) This is used to convert delay time into risk saturation. ; The resource loss dimension of the multidimensional feedback vector satisfies the following: when the total value of resource waste exceeds twenty times the benchmark unit, the growth rate of the logarithmic compression value decreases to less than one-fifth of the original rate; The specific operation of dynamically adjusting weights in the federated aggregation protocol is as follows: D1. Calculate the node contribution weight coefficient: Based on the Euclidean norm value of the multidimensional feedback vector uploaded by each medical institution node, compress it to the zero-to-one interval using the hyperbolic tangent function as the basic weight component; simultaneously, obtain the model timeliness deviation duration of the node, multiply this duration by the decay factor, and take the negative value of the natural exponent as the timeliness decay component; multiply the basic weight component and the timeliness decay component to obtain the final node contribution weight, the calculation formula is as follows: ; in, Represents a node The contribution weight is used to quantify the value of node data; the higher the value, the greater the influence in the aggregation. Represents a node The three-dimensional feedback vector (carrying information on triage quality, resource consumption, and delay risk). The Euclidean norm is used to calculate the overall feedback intensity value. This represents the feedback intensity normalization factor (set to 5), used to control the magnitude of the norm value and prevent numerical explosion. Represents a node Data timeliness deviation (hours since the latest model) is used to measure data freshness. This indicates a lag behind the latest model. This represents the aging decay coefficient (set to 0.3), used to control the intensity of aging penalty. A larger value indicates a lower weight for older data. In this formula, Compressing any feedback to [0,1) avoids outlier dominance, exponential term Achieve nonlinear time-dependent decay; D2. Perform secure aggregation operation: Each node calculates the gradient tensor of the Siamese neural network parameters in an encrypted environment. It then performs a tensor product operation on the coupling matrix constructed from the node's own multidimensional feedback vector, and multiplies this by the node's contribution weight coefficient to form a weighted encrypted gradient. All nodes decrypt and aggregate the weighted encrypted gradient using a secure multi-party computation protocol, calculating the arithmetic mean as the global gradient update. The expression for this global gradient update is: ; in, This indicates the global gradient update amount (the direction of parameter adjustment shared across the entire network after federated aggregation). This refers to a secure multi-party computation decryption protocol, whose function is to aggregate gradients in an encrypted state to prevent the leakage of the original data. Represents a node The local gradient tensor is used to carry the parameter update requirements of the twin network of this node. Representing the feedback coupling matrix: Its function is to inject feedback values ​​into the gradient. (Resource depletion) Amplify the gradient of relevant parameters; (Delay risk) Targeted enhancement of sensitive feature weights, This represents the tensor product operation, used to achieve high-dimensional coupling between matrices and gradients. (Dimension expanded to 3 × original size), in which resource consumption... The gradient is linearly scaled using the position in the first row and second column of the matrix: ; The more severe the resource waste, the greater the magnitude of the relevant parameter updates; D3. Apply dynamic parameter constraints: Calculate the dynamic pruning threshold based on the Manhattan norm of the node's multidimensional feedback vector. The formula for calculating the dynamic pruning threshold is: ; in, This represents the baseline clipping threshold, which controls the standard update magnitude (its value is set to 0.01). The Manhattan norm of the feedback vector (used to calculate the overall defect strength) has the following effects when the overall defect is... When the threshold is large, reduce the clipping threshold. 1. Suppress parameter mutations; 2. Prevent extreme cases (such as...) The threshold (in yuan) causes model oscillations. This threshold is negatively correlated with the feedback strength. When the Manhattan norm exceeds two, the pruning threshold linearly decreases to below 40% of the standard threshold. This threshold is used to truncate the global gradient update, and the truncated gradient is superimposed on the original weights of the Siamese neural network to complete the parameter update. Its expression is: ; in, This represents the optimized twin network parameters. This represents the twin network parameters before the update (the original parameters that need to be adjusted). This represents the global gradient update amount (the total parameter update direction generated by federated aggregation). This represents the dynamic clipping threshold, used to adjust based on feedback intensity. Adjust the update range in real time; The specific steps for dynamically updating the modal conflict detection threshold are as follows: The final diagnostic delay duration before the current time of this node is obtained as the core input. This delay duration is divided by a preset time scale parameter to obtain a scaling factor. The scaling factor is then input into a Gaussian error function for nonlinear mapping. The mapping result is multiplied by a sensitivity coefficient and then incremented by one to form a scaling factor. This scaling factor is multiplied by the historical modal conflict detection threshold to generate an updated detection threshold, the expression of which is: ; in, This represents the updated confidence detection threshold, which is used for modal conflict determination in the subsequent confidence redirection module (the higher the threshold, the easier it is to trigger redirection). This represents the historical confidence detection threshold, which is used as the updated baseline value (its value is set to 0.3). This indicates the time delay in final diagnosis, i.e., the time difference between incorrect triage and confirmed diagnosis (unit: minutes). This represents the time dimension parameter (set to 30 minutes), which is the sensitive period for controlling the impact of delays. Mild delay zone Severe delay zone; This represents the sensitivity coefficient (set to a value of 0.2), indicating the degree of conservatism in adjusting the threshold increase. This represents the Gaussian error function, used to implement a nonlinear mapping from delay duration to scaling factor: 1. Time approximates linear (slow growth rate) 2. The acceleration saturation is calculated as follows: ; The calculation process of the Gaussian error function is defined as doubly normalizing the integral value of the negative exponential function in the interval from zero to the scaling factor. This operation satisfies the time-sensitive characteristic: when the actual delay duration does not reach the time scale parameter, the threshold growth rate is lower than linear growth; when the delay duration exceeds the time scale parameter, the threshold growth rate increases to more than 1.5 times the linear growth rate, and its growth rate change is determined by the error function. The derivative determines: 1. Mild delay zone ( Derivative: ,when hour, The growth rate is lower than linear growth (the linear derivative is 1). 2. Severe Delay Zone ( ):when hour, It decayed rapidly from 0.367, but after... After modulation: Actual scaling factor growth rate: ; Comparison with linear growth rate: If a linear function is used, the growth rate is always 0. The error function scheme reduces speed by 63.5%.

[0030] In this embodiment, the confidence redirection module specifically describes the following operations for detecting the difference in confidence between physiological signs and the chief complaint text: First, the temporal entropy value corresponding to the physiological sign temporal feature vector generated by the multimodal alignment module is obtained. This temporal entropy value is obtained by calculating the information entropy of the probability distribution of each dimension of the feature vector. The calculation formula is as follows: ; in, Represents the entropy value of physiological characteristics ( A higher value indicates more complex and abnormal vital signs, while a lower value indicates more stable vital signs. The feature dimension is used to control computational complexity. Represents the vector of the first The dimension component is used to represent the intensity of physiological indicators within a specific time window; simultaneously, the semantic focus degree corresponding to the semantic vector of the chief complaint text is obtained. This semantic focus degree is determined by extracting the maximum value component in the semantic vector, and its calculation formula is as follows: ; in, Indicates the semantic focus of the chief complaint ( A higher value indicates a more specific symptom description, while a lower value indicates a more vague description. This represents a semantic vector of the chief complaint text, used to encode the clinical meaning of the symptom description. The normalization exponential function is used to transform semantic vectors into probability distributions; the temporal entropy value is subtracted from the semantic focus to obtain the initial difference value; the absolute value of the initial difference value is processed to obtain the absolute difference amount; finally, the absolute difference amount is compared with the dynamically updated modal conflict detection threshold, and when the absolute difference amount exceeds the threshold, the graph traversal operation is triggered. The dynamic threshold determination formula is: ; when When this is triggered, graph traversal is traversed (e.g.: =0.92, a high anomaly. =0.35 (low clarity, β=0.5). Redirection will be triggered in the following scenarios: ; For example: A patient with acute myocardial infarction complains of "stomach discomfort," and physical examination reveals ST-segment elevation and ventricular fibrillation. =0.92), chief complaint focus ( =0.35); The operations for triggering graph neural networks to traverse symptom-department knowledge graphs and generate redirection instructions include: Construct patient query nodes with physiological signs and chief complaint semantic vectors as attributes. ,in It represents the time-series feature vector of physiological signs, which is used to quantify the patient's condition (such as ST segment elevation on electrocardiogram). This represents the semantic vector of the chief complaint, used to encode symptom descriptions (e.g., "chest pain lasting 2 hours"). This represents the gradient vector of the top-3 symptom words in the main complaint text, used to explicitly identify key symptoms (such as semantic weight enhancement for "chest pain", "difficulty breathing", and "radiating pain"). This represents a vector concatenation operation used to fuse vital signs, semantics, and key symptom features. This represents a multilayer perceptron model that generates patient feature nodes with medical semantics. The patient query node is injected into the symptom-department knowledge graph architecture, and vector state propagation is performed along the medical relationship path through a multi-layer graph convolutional network. The formula is as follows: ; in, Represents a node In the The layer's state vector represents the dynamic semantic state of a node in the medical knowledge graph. Represents a node With adjacent nodes The association weights are used to quantify the strength of medical relationships (such as the association between "myocardial infarction → radiation pain"). A weight matrix specifically designed for representing relationships, used to independently model different medical relationships (inducement / demand / mutual exclusion). This represents the activation function, used to introduce non-linear decision-making capabilities. It calculates the association weight between the current department node and adjacent symptom nodes during each layer of propagation. Its formula is: ; in, This represents a medical relation type embedding vector, encoding three types of medical logic: 1. [1,0,0]: induced relation (e.g., smoking → lung cancer) 2. [0,1,0]: demand relation (e.g., surgery → anesthesiology) 3. [0,0,1]: mutually exclusive relation (e.g., pregnancy → CT scan). This represents a scaling factor used to prevent attention score gradient explosion. This weight combines the relationship type embedding factor with patient feature similarity. After propagation through a specified number of layers, the final state vectors of all department nodes are summarized. Based on the state vectors, the matching probability distribution of each department is calculated using the following formula: ; in, Indicates the first The department is at the bottom floor. The state vector is used to aggregate the departmental decision-making characteristics (such as resource capabilities and disease expertise) after propagation. This represents a trainable input weight matrix used to map departmental states to a probability distribution. Indicates department The redirection probability value is used to generate interpretable triage criteria (e.g., cardiology probability = 0.92, gastroenterology probability = 0.08); the department number with the highest probability value in the probability distribution is identified; when the probability ratio of the department with the highest probability value to the original target department exceeds the critical coefficient, its expression is: ; in, This represents the maximum value (i.e., the optimal department probability) among all department probability distributions, used to identify the department that best matches the graph inference results, such as cardiology (probability 0.92). This represents the probability of the target department originally output by the twin matching module, indicating the credibility of the initial triage decision (e.g., a probability of 0.35 for the gastroenterology department). This represents the coverage intensity coefficient, used to quantify the degree of difference between new and old departments. (Forcible coverage at the time), generating a triplet instruction containing the new department number, coverage strength, and medical evidence chain, the expression of which is: ; ; ; in, This indicates selecting the department number with the highest probability (e.g., ...). correspond (cardiology)), Represents medical relationship edges in a knowledge graph (such as the precipitating relationship of "myocardial infarction → upper abdominal radiating pain"). Represents patient node With Relationship Attention weights are used to quantify the importance of the relationship to the decision (the larger the value, the more critical it is). Indicates attention threshold ( ), used to screen for strongly relevant medical evidence (such as Only then is the relationship included This will force the original triage decision results to be overridden.

[0031] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0032] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0033] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0034] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0035] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0036] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0037] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A hospital intelligent operation management system based on a large model, characterized in that, Specifically, it includes: Multimodal alignment module, twin matching module, feedback optimization module, and confidence redirection module; The multimodal alignment module, when receiving patient complaint text and real-time physiological sign data, uses a cross-modal attention mechanism to perform spatial alignment operations on temporal physiological features and complaint semantic vectors, generating a cross-modal feature vector that integrates patient health status information; The twin matching module, based on the cross-modal feature vector and the real-time updated department capability matrix, calculates the decision similarity parameters between patient features and department resource status through a shared architecture twin neural network, and outputs the target department number in combination with a preset triage mismatch loss matrix. The feedback optimization module, after obtaining the final diagnosis result, constructs a multi-dimensional feedback vector from the triage result deviation value, resource waste, and delay time. It dynamically adjusts the weight parameters of the Siamese neural network through a federated aggregation protocol and dynamically updates the modal conflict detection threshold based on the local delay time value of each node. Specifically, it adjusts the scaling ratio of the historical threshold through a nonlinear error function and feeds the updated modal conflict detection threshold back to the confidence redirection module. The confidence redirection module, when it detects that the difference in confidence between physiological signs and the chief complaint exceeds a set threshold, triggers a graph neural network to traverse the symptom-department knowledge graph and generate a redirection instruction that forcibly covers the target department number.

2. The intelligent hospital operation management system based on a large model according to claim 1, characterized in that: In the multimodal alignment module, the specific operation of spatial alignment between temporal physiological features and the subject complaint semantic vector using a cross-modal attention mechanism is as follows: A1. Receive physiological sign data collected by wearable devices to form a time-series physiological sign sequence, extract waveform change patterns of the time-series physiological sign sequence through a temporal convolutional network model, and generate a physiological feature vector; the temporal convolutional network model adopts a causal convolutional kernel structure; A2. Receive patient complaint text data, extract semantic features from the complaint text data using a clinical-specific text encoder, and generate a fixed-dimensional clinical semantic vector; the clinical-specific text encoder is based on a medical knowledge pre-training model to reinforce the encoding of medical terms in the symptom description. A3. Using the physiological feature vector as the dynamic query benchmark and the clinical semantic vector as the alignment target, calculate the feature association distribution between each time node in the physiological feature vector and the clinical semantic vector, and perform a weighted fusion operation on the clinical semantic vector according to the feature association distribution to generate a spatially aligned semantic representation vector; the feature association distribution is generated by the vector dot product operation of a trainable projection matrix. A4. Perform a ternary combination operation on the physiological feature vector and the spatially aligned semantic representation vector, specifically including: directly concatenating the physiological feature vector and the semantic representation vector as independent feature components, calculating the element-wise product between the two as an interactive feature component, and finally forming a cross-modal feature vector that integrates the patient's multi-dimensional health status information.

3. The intelligent hospital operation management system based on a large model according to claim 2, characterized in that: The calculation and weighted fusion operation of the feature association distribution are specifically as follows: B1. The physiological feature vector is mapped to the attention query vector through a linear transformation matrix, and the clinical semantic vector is mapped to the attention key vector. B2. Calculate the cosine similarity between the attention query vector and the attention key vector, and transform it into an association weight distribution using an exponential normalization function; B3. Based on the aforementioned associated weight distribution, perform weight allocation operations on the original components of the clinical semantic vector to generate a semantic representation vector that reflects temporal dependence. The associated weight allocation operation satisfies the following: when abnormal fluctuations occur in the waveform of physiological signs, the weight of related symptom description segments is automatically strengthened; when patients describe their past medical history, text features unrelated to the current signs are weakened.

4. The intelligent hospital operation management system based on a large model according to claim 3, characterized in that: In the twin matching module, the specific steps for calculating the decision similarity parameters between patient characteristics and departmental resource status using a shared architecture twin neural network include: C1. Input the cross-modal feature vector into an independent patient feature encoder, extract the patient state embedding vector through a depthwise separable convolutional layer and a self-attention mechanism, and output a dense vector of fixed dimension to represent the patient's overall health status. C2. Input the real-time department capability matrix into the department feature encoder with shared parameters. This encoder reuses the same set of convolutional kernel weights for all departments to generate the resource state embedding vector for each department. C3. Dynamically integrate department saturation and expert resource status: Construct a three-dimensional weight tensor. The first dimension is a constant bias term, the second dimension is the negative logarithmic scaling value of the real-time occupancy rate of the department, and the third dimension is the Boolean amplification coefficient of the on-duty status of experts for the target disease. Perform element-wise multiplication operations on the weight tensor and the department resource status embedding vector to generate a weighted department feature vector. C4. Calculate the hyperbolic geometric space similarity between the patient state embedding vector and the weighted department feature vector: First, perform hyperbolic secant normalization on the two vectors, then calculate the reciprocal of their hyperbolic cosine distance as the basic similarity value, and finally multiply by the exponential decay factor of the angle difference between the two in hyperbolic space to obtain the nonlinear decision similarity parameter.

5. The intelligent hospital operation management system based on a large model according to claim 4, characterized in that: The calculation rule for the negative logarithmic scaling value of the department's real-time occupancy rate is as follows: The average of the department's bed occupancy rate and equipment working time within the past ten minutes is used as the real-time occupancy rate benchmark. The natural logarithm of this benchmark is multiplied by a preset penalty coefficient, and the resulting negative value is used as the weight of the second dimension. This calculation significantly reduces the similarity of departments with occupancy rates exceeding the threshold. For every certain percentage increase in occupancy rate, a curvature increase effect of the logarithmic function is triggered, which accelerates the decline in the matching priority of high-load departments.

6. The intelligent hospital operation management system based on a large model according to claim 5, characterized in that: The specific operation of outputting the target department number by combining the preset triage mismatch loss matrix is ​​as follows: The decision similarity parameters for each department are scaled by a temperature coefficient, and the department matching probability distribution is generated by an exponential normalization function. The probability distribution matrix and the triage mismatch loss matrix are subjected to Hadamard product operation to obtain the expected loss tensor. Each element of the loss tensor represents the expected risk value of triaging the patient to the corresponding department. The department number with the smallest value in the expected loss tensor is selected as the final output target. The principle for setting the triage mismatch loss matrix is ​​that the loss value of triaging critical illnesses to non-emergency departments is at least five times the loss value of ordinary triage errors.

7. The intelligent hospital operation management system based on a large model according to claim 6, characterized in that: In the feedback optimization module, the specific operation of constructing a multi-dimensional feedback vector is as follows: First, the deviation value of the triage result is defined as the mismatch indicator between the actual diagnosis department number and the triage decision department number. If the two are inconsistent, the value is assigned to the integer one; if they are consistent, the value is assigned to the integer zero. Secondly, the logarithmic compression value of resource waste is calculated. The total value of resource waste is proportionally converted using the cost of a single standard inspection as the benchmark unit, and the natural logarithm is taken after adding one to the proportional value. Finally, the exponential saturation function value of delay time is calculated. The actual delay time is input into the negative exponential function of the preset decay factor, and the delay risk saturation is obtained by subtracting the output value of the function from the constant. The above three values ​​are combined into a three-dimensional column vector in the order of diagnostic difference, resource loss, and delay risk to form a multi-dimensional feedback vector data structure. The resource loss dimension of the multidimensional feedback vector satisfies the following condition: when the total value of resource waste exceeds twenty times the benchmark unit, the growth rate of the logarithmic compression value decreases to less than one-fifth of the original rate.

8. The intelligent hospital operation management system based on a large model according to claim 7, characterized in that: The specific operation of the federated aggregation protocol to dynamically adjust weights is as follows: D1. Calculate the node contribution weight coefficient: Based on the Euclidean norm value of the multidimensional feedback vector uploaded by each medical institution node, compress it to the zero-to-one interval through the hyperbolic tangent function as the basic weight component; at the same time, obtain the model timeliness deviation duration of the node, multiply the duration by the decay factor and take the negative value of the natural exponent as the timeliness decay component. The final node contribution weight is obtained by multiplying the basic weight component by the time-decrease component. D2. Perform secure aggregation operation: Each node calculates the gradient tensor of the Siamese neural network parameters in an encrypted environment, performs tensor product operation on the coupling matrix constructed by the node's own multidimensional feedback vector, and then multiplies it with the node's contribution weight coefficient to form a weighted encrypted gradient; all nodes decrypt and aggregate the weighted encrypted gradient through a secure multi-party computation protocol, and calculate the arithmetic mean as the global gradient update amount. D3. Apply dynamic parameter constraints: Calculate the dynamic pruning threshold based on the Manhattan norm of the node's multidimensional feedback vector. This threshold is negatively correlated with the feedback strength. When the Manhattan norm exceeds two, the pruning threshold is linearly reduced to less than 40% of the standard threshold. Use this threshold to truncate the global gradient update and add the truncated gradient to the original weights of the Siamese neural network to complete the parameter update. The specific operation of dynamically updating the modal conflict detection threshold is as follows: The final diagnostic delay duration before the current time of this node is obtained as the core input. The delay duration is divided by the preset time scale parameter to obtain the scaling factor. The scaling factor is input into the Gaussian error function for nonlinear mapping. The mapping result is multiplied by the sensitivity coefficient and then one is added to form the scaling factor. Multiply the scaling factor by the historical modal conflict detection threshold to generate the updated detection threshold.

9. The intelligent hospital operation management system based on a large model according to claim 8, characterized in that: In the confidence redirection module, the specific operations for detecting the difference in confidence between physiological signs and the chief complaint text include: First, the temporal entropy value corresponding to the physiological sign temporal feature vector generated by the multimodal alignment module is obtained. This temporal entropy value is obtained by calculating the information entropy of the probability distribution of each dimension of the feature vector. At the same time, the semantic focus degree corresponding to the semantic vector of the complaint text is obtained. This semantic focus degree is determined by extracting the maximum value component in the semantic vector. The temporal entropy value and the semantic focus degree are subtracted to obtain the preliminary difference value. The absolute value of the preliminary difference value is processed to obtain the absolute difference amount. Finally, the absolute difference amount is compared with the dynamically updated modal conflict detection threshold. When the absolute difference amount exceeds the threshold, the graph traversal operation is triggered.

10. The intelligent hospital operation management system based on a large model according to claim 9, characterized in that: The operation of the trigger graph neural network traversing the symptom-department knowledge graph to generate redirection instructions includes: A patient query node is constructed with physiological signs and chief complaint semantic vectors as attributes. This patient query node is injected into a symptom-department knowledge graph architecture. Vector state propagation is performed along the medical relationship path through a multi-layer graph convolutional network. During each layer of propagation, the association weight between the current department node and adjacent symptom nodes is calculated. This weight combines the relationship type embedding factor and the patient feature similarity. After completing the specified number of propagation layers, the final state vectors of all department nodes are summarized. The matching probability distribution of each department is calculated based on the state vectors. The department number with the highest value in the probability distribution is identified. When the probability ratio between the highest value department and the original target department exceeds the critical coefficient, a triple instruction containing the new department number, coverage strength, and medical evidence chain is generated to forcibly cover the original triage decision result.

Images