Medical intelligent causal decision-making and scheduling system based on improved large language model

Abstract

The invention provides a medical intelligent causal decision-making and scheduling system based on an improved large language model, and the system comprises a pre-training language model module which is used for processing unstructured medical text data, and extracting a text feature vector and a medical entity relationship; the knowledge graph enhancement module is used for constructing a medical knowledge graph based on the medical entity relationship, and reasoning to generate a knowledge enhancement feature vector; the multi-modal feature fusion module is used for generating a fusion feature vector based on the text feature vector and pre-stored structured data; the dynamic time sequence modeling module is used for generating a time sequence prediction vector based on the fusion feature vector and the knowledge enhancement feature vector; the self-adaptive scheduling optimization module is used for making a medical scheduling decision based on the time sequence prediction vector and generating a scheduling scheme; and the scheduling calibration module is used for matching similar case features for the scheduling scheme through a context learning mechanism based on the knowledge enhancement feature vector and calibrating the similar case features to generate an optimized scheduling scheme. And medical resources and scheduling can be allocated more accurately.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a medical intelligent causal decision-making and scheduling system based on an improved large language model. Background Technology

[0002] In the medical field, intelligent decision-making and scheduling optimization systems play a crucial role in improving the efficiency of medical resource utilization, reducing doctors' workload, and enhancing the patient treatment experience. Currently, the application of artificial intelligence technology has shown great potential in medical data processing and decision support.

[0003] However, existing technologies still have many shortcomings, mainly in the following aspects:

[0004] 1. Insufficient medical adaptability of pre-trained language models. Most existing large language models adopt general domain pre-training methods. When processing medical text data (such as medical records, doctor's orders, and surgical records), they often lack the ability to understand medical domain professional knowledge and are prone to erroneous inferences.

[0005] 2. Limitations of Traditional Medical Scheduling Systems. Existing medical scheduling systems are based on structured data and use rule-based or heuristic methods for scheduling decisions. These systems cannot effectively integrate multimodal medical data, are ill-suited to dynamically changing medical environments, and lack the ability to model long-term time-series patterns, resulting in insufficient responsiveness to future changes in medical needs. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application provides a medical intelligent causal decision-making and scheduling system based on an improved large language model, in order to solve the problems existing in the prior art.

[0007] To achieve the above objectives and other advantages, this application adopts the following technical solution:

[0008] This application provides a medical intelligent causal decision-making and scheduling system based on an improved large language model, including:

[0009] The pre-trained language model module is used to process unstructured medical text data and extract text feature vectors and medical entity relationships.

[0010] The knowledge graph enhancement module is used to construct a medical knowledge graph based on the relationships between the medical entities and to infer and generate knowledge-enhanced feature vectors.

[0011] The multimodal feature fusion module is used to generate a fused feature vector based on the text feature vector and pre-stored structured data through a cross-modal alignment mechanism;

[0012] The dynamic temporal modeling module is used to generate a temporal prediction vector based on the fused feature vector and the knowledge-enhanced feature vector through a counterfactual reasoning mechanism.

[0013] An adaptive scheduling optimization module is used to make medical scheduling decisions based on the time-series prediction vector and generate a scheduling plan.

[0014] The scheduling calibration module is used to match similar case features to the scheduling scheme based on the knowledge-enhanced feature vector through a context learning mechanism, and to calibrate the scheduling scheme based on the similar case features to generate an optimized scheduling scheme.

[0015] According to the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application, the pre-trained language model module adopts a prompt learning mechanism that enhances medical domain knowledge. It parses the feature information of the smallest semantic unit in the unstructured medical text data according to the set first prompt information, converts it into the text feature vector, and extracts the core medical entities in the unstructured medical text data according to the set second prompt information to generate the medical entity relationship.

[0016] According to the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application, the pre-trained language model module is further used to: generate a dynamic prompt information template set and a prompt information relevance score matrix through a dynamic prompt information generation mechanism. The dynamic prompt information generation mechanism is obtained by using historical data encoding, medical context encoding, and time encoding as inputs and processing them using a self-attention mechanism. The dynamic prompt information template set is used to optimize the selection of prompt information in the prompt learning mechanism.

[0017] According to the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application, the multimodal feature fusion module uses a cross-modal attention mechanism to calculate the similarity between the text feature vector and the structured data features, and maps them to a unified feature space for feature alignment calculation to generate the fused feature vector and alignment score matrix.

[0018] According to the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application, the dynamic temporal modeling module constructs a causal reasoning graph through a counterfactual reasoning mechanism, receives the current system state vector, and uses the current system state vector, the fused feature vector, and the knowledge-enhanced feature vector as inputs to the causal reasoning graph. The decision value matrix and counterfactual analysis results are calculated through the decision function of the causal reasoning graph, and the temporal prediction vector is obtained based on the decision value matrix, the fused feature vector, and the knowledge-enhanced feature vector.

[0019] According to the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application, the knowledge graph enhancement module is pre-set with a graph neural network and a pre-trained language model. The medical entity relationship includes an entity embedding matrix and a relationship embedding matrix. The adjacency matrix of the connection relationship between entities is obtained from the medical entity relationship. The adjacency matrix, the entity embedding matrix and the relationship embedding matrix are used as inputs to the graph neural network. Feature processing is performed through the graph neural network to extract the structured features of the knowledge graph.

[0020] The knowledge graph enhancement module receives the fused feature vector and the text feature vector, and extracts the semantic information of the fused feature vector based on the pre-trained language model;

[0021] The structured features of the knowledge graph are fused with the semantic information to obtain the knowledge-enhanced feature vector and the relational reasoning path.

[0022] According to the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application, the adaptive scheduling optimization module adopts a multi-objective optimization algorithm. The multi-objective optimization algorithm includes multiple objective optimization functions for calculating resource utilization, professional matching degree, time fluency, and constraint violation penalty. The multi-objective optimization algorithm assigns weights to different objective optimization functions to construct a comprehensive optimization objective function, and solves it in combination with constraint conditions to obtain the scheduling scheme.

[0023] According to the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application, the scheduling calibration module includes: a scheduling scheme reasoning unit, a case matching unit, and a prediction calibration unit;

[0024] The scheduling scheme reasoning unit is used to construct a scheduling optimization reasoning path based on historical reasoning trajectories, the knowledge-enhanced feature vector, and the query vector formed by the scheduling scheme, using a thinking chain mechanism. It also calculates the rationality of the scheduling scheme by combining probability distribution. The reasoning confidence score and the scheduling reasoning step sequence are obtained through the scheduling optimization reasoning path. The reasoning confidence score is used to evaluate the credibility of the scheduling scheme.

[0025] The case matching unit is used to perform similar case matching based on the historical case database and the scheduling reasoning step sequence, through a retrieval function, to retrieve the historical case most similar to the scheduling scheme, and to calculate similar case features and case relevance scores based on the historical case.

[0026] The prediction calibration unit is used to calculate the scheduling adjustment factor based on the similar case features, the case correlation score and the time-series prediction vector, and optimize the scheduling scheme through the scheduling adjustment factor to obtain the optimized scheduling scheme.

[0027] According to the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application, it further includes: a natural language interpretation generation module, used to receive the counterfactual analysis results output by the dynamic temporal modeling module and the prompt information relevance score matrix of the pre-trained language model module, generate decision interpretation text for the optimized scheduling scheme based on the counterfactual analysis results and the prompt information relevance score matrix, and calculate the credibility of the decision interpretation text through an interpretation evaluation function to obtain an interpretation credibility score.

[0028] According to the present application, a medical intelligent causal decision-making and scheduling system based on an improved large language model is provided. The medical intelligent causal decision-making and scheduling system adopts a multi-agent collaborative language model integration framework. The language model integration framework is used to receive the output of each module, construct the input variables of each agent according to the output of each module, integrate the reasoning results of each agent through the result aggregation function, and obtain the final scheduling decision scheme, comprehensive confidence score and decision explanation report output by the language model integration framework based on the integration result.

[0029] This application provides a medical intelligent causal decision-making and scheduling system based on an improved large language model. It processes unstructured medical text data, extracting text feature vectors and medical entity relationships. The text feature vectors are combined with structured data, and a fused feature vector is generated through a cross-modal alignment mechanism. Based on the fused feature vector and knowledge-enhanced feature vector, a counterfactual reasoning mechanism is used to generate a time-series prediction vector. Scheduling decisions are made based on the time-series prediction vector, and the system utilizes features from similar historical cases for scheme calibration. This application addresses the shortcomings of existing medical scheduling and decision-making systems, enabling more accurate allocation of medical resources and doctor scheduling. It not only improves the intelligence level of medical scheduling but also promotes the transformation of medical decision-making from an "experience-driven" to a "data-driven + causal reasoning" model, providing an innovative solution for the intelligent management of medical institutions. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other implementation methods can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the structure of a medical intelligent causal decision-making and scheduling system based on an improved large language model, provided in an embodiment of this application.

[0032] Figure 2 This is a flowchart of the logical processing of a medical intelligent causal decision-making and scheduling system based on an improved large language model provided in the embodiments of this application;

[0033] Figure 3 This is a flowchart illustrating the logical processing of the pre-trained language model module provided in an embodiment of this application.

[0034] Figure 4 This is a flowchart illustrating the logical processing of the multimodal feature fusion module provided in this application embodiment;

[0035] Figure 5 This is a flowchart illustrating the logical processing of the knowledge graph enhancement module provided in an embodiment of this application.

[0036] Figure 6 This is a flowchart illustrating the logical processing of the dynamic timing modeling module provided in this application embodiment. Detailed Implementation

[0037] The above description is merely an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings.

[0038] It should be noted that those skilled in the art will understand, explicitly and implicitly, that the embodiments described in this application can be combined with other embodiments without conflict. Unless otherwise defined, the technical or scientific terms used in this application should have the ordinary meaning understood by those skilled in the art. The terms "a," "an," "an," "the," and similar words used in this application do not indicate quantity limitation and can represent singular or plural. The terms "comprising," "including," "having," and any variations thereof used in this application are intended to cover non-exclusive inclusion; the terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific ordering of objects.

[0039] To facilitate understanding of the embodiments of this application, the following are explanations of key terms / technical abbreviations in this application:

[0040] Prompt Learning is a technique that optimizes the output of a language model by designing specific prompt templates.

[0041] Cross Entropy (CE): A loss function used to measure the difference between the predicted distribution and the true distribution.

[0042] KL (Kullback-Leibler Divergence): KL divergence is an indicator used to measure the difference between two probability distributions.

[0043] GNN (Graph Neural Network): A deep learning model for processing graph-structured data.

[0044] Chain-of-Thought (CoT): Chain-of-Thought reasoning is a method that allows language models to solve complex problems through step-by-step reasoning.

[0045] In-Context Learning (ICL): Contextual learning is a learning method that uses similar historical cases to improve model performance.

[0046] LM (Language Model): A model used to process and generate natural language.

[0047] Transformer: A neural network architecture based on self-attention mechanism, widely used in natural language processing.

[0048] Retriever: A module used to retrieve relevant information from large-scale data.

[0049] Aggregator: An aggregator is a component used to combine the outputs of multiple models or features.

[0050] softmax: a normalized exponential function used to convert numerical values ​​into a probability distribution.

[0051] QKV (Query-Key-Value): Represents query vector, key vector, and value vector, the three core components of the attention mechanism.

[0052] K (Knowledge Enhancement): Knowledge enhancement feature.

[0053] DV (Decision Value): Decision value matrix.

[0054] CF (Counterfactual): Counterfactual analysis results.

[0055] ET (Explanation Text): Decision Explanation Text.

[0056] EC (Explanation Credibility): The score for the credibility of the explanation.

[0057] RP (Reasoning Path): Relational reasoning path.

[0058] SS (Step Sequence): The sequence of reasoning steps.

[0059] SC (Step Confidence): Inference confidence.

[0060] SF (Similar Features): Similar case features.

[0061] SR (Similarity Rating): Case relevance score.

[0062] FD (Final Decision): Final decision.

[0063] CS (Confidence Score): Overall confidence score.

[0064] DR (Decision Report): Decision Explanation Report.

[0065] Reference Figure 1 , 2 As shown, this application provides a medical intelligent causal decision-making and scheduling system based on an improved large language model, including:

[0066] The pre-trained language model module 10 is used to process unstructured medical text data and extract text feature vectors and medical entity relationships.

[0067] In this embodiment, the pre-trained language model module 10 adopts a prompt learning mechanism that enhances medical domain knowledge. It parses the feature information of the smallest semantic unit in the unstructured medical text data according to the set first prompt information, converts it into a text feature vector, and extracts the core medical entities in the unstructured medical text data according to the set second prompt information to generate medical entity relationships.

[0068] Specifically, unstructured medical text data includes medical records, surgical records, and medical orders. The pre-trained language model module 10 is used to process unstructured medical text data, generating text feature vectors T and medical entity relationships R. A Prompt Learning mechanism is employed to guide the large language model in understanding the medical text data and generating structured text feature vectors T by designing appropriate prompts. The Prompt Learning mechanism can also be used to parse unstructured text, extract medical entities, identify relationships between medical entities, and construct medical entity relationships R. Medical entity relationships R refer to the association information between different medical concepts or entities in the medical text data. In medical data processing, "entities" typically include diseases, symptoms, drugs, examination items, treatment plans, doctors, and patients, while "relationships" describe the logical or causal connections between these entities, such as disease-symptom relationships, disease-drug relationships, and examination-diagnosis relationships.

[0069] As an example, the Prompt Learning algorithm, which employs knowledge enhancement in the medical field, can be defined by a loss function L. PLA As shown below:

[0070] L PLA =∑(α) i ×CE(y i f(x) i p i )))+β×KL(q(z|x), p(z))

[0071] Where, x i The input is unstructured medical text data; y i The corresponding true label or target output is used for supervised learning; α i β is the feature weight coefficient, used to balance the importance of different features; β is the weight coefficient of the KL divergence term, used to adjust the regularization strength; CE is the cross-entropy loss function, used to optimize text feature extraction; f is the language model forward function, used for text encoding; p i q(z|x) is a dynamically generated prompt template, generated by a dynamic prompt information generation mechanism; KL is the KL divergence, used to constrain the feature distribution; q(z|x) is the posterior distribution; p(z) is the prior distribution, used for regularization of feature learning.

[0072] The text feature vector T can be generated using the following formula:

[0073] T = Encoder(f(x) i ,p i ))

[0074] Here, Encoder is the encoder function that maps the language model output to the feature space; T has a dimension of d×n, where d is the feature dimension and n is the sequence length.

[0075] The medical entity relationship R can be generated using the following formula:

[0076] R=EntityExtract(T)·RelationClassify(T)

[0077] Wherein, EntityExtract is the entity extraction function, which identifies medical entities from the text feature vector T; RelationClassify is the relation classification function, which determines the relationships between entities; R has a dimension of m×m, where m is the number of medical entities identified.

[0078] The use of a medical knowledge-enhanced Prompt Learning mechanism makes the model more adaptable to specific medical tasks. By embedding medical knowledge, the model can more accurately parse medical-related content such as symptoms, medications, surgical records, and patient conditions, thus improving its medical text understanding capabilities.

[0079] Reference Figure 3 As shown, the pre-trained language model module 10 is also used to: generate a dynamic prompt information template set and a prompt information relevance score matrix through a dynamic prompt information generation mechanism. The dynamic prompt information generation mechanism is obtained by using historical data encoding, medical context encoding and time encoding as input and processing them with a self-attention mechanism. The dynamic prompt information template set is used to optimize the selection of prompt information in the prompt learning mechanism.

[0080] Specifically, the dynamic prompt information generation mechanism receives historical surgical data encoding E(h), medical context encoding M(c), and time information encoding R(t), and its algorithm can be defined as follows:

[0081]

[0082] Wherein, Pr(t) represents the characteristic state of the medical system at time t; historical surgical data encoding E(h) is used to capture historical scheduling patterns; medical context encoding M(c) is used to understand the current medical scenario; time information encoding R(t) is used to grasp temporal features; and Transformer is a transformer encoder structure used for feature fusion and transformation. This is a feature concatenation operation used to combine features from different sources.

[0083] The dynamic prompt message template set PT can be generated using the following formula:

[0084] PT=SplitHead(Pr(t),l)={Pr1,Pr2,...,Pr i}

[0085] SplitHead is a multi-head splitting function that splits Pr(t) into multiple independent prompt template vectors; Pr i Let v be the i-th prompt template vector, with dimension v.

[0086] The dynamic prompt template set PT optimizes the prompt selection of the Prompt Learning algorithm, allowing the model to automatically adapt to different medical tasks, avoiding information loss caused by fixed templates, and further improving the accuracy of medical text parsing.

[0087] The relevance score matrix PS for the prompt information can be generated using the following formula:

[0088] PS = Softmax(Linear(P(t)))

[0089] Among them, Linear is a linear transformation layer that maps Pr(t) to the correlation score space; Softmax is a normalization function that generates the correlation probability distribution of each template.

[0090] The dynamic prompt generation mechanism can adaptively adjust prompts according to different medical scenarios, improve the model's ability to understand medical texts such as medical records, doctor's orders, and surgical records, and enhance causal reasoning and contextual adaptability.

[0091] In this embodiment, the multimodal feature fusion module 20 is used to generate a fused feature vector based on text feature vectors and pre-stored structured data through a cross-modal alignment mechanism.

[0092] Specifically, refer to Figure 4 As shown, the multimodal feature fusion module 20 uses a cross-modal attention mechanism to calculate the similarity between text feature vectors and structured data features, and maps them to a unified feature space for feature alignment calculation to generate fused feature vectors and alignment score matrices.

[0093] As an example, structured data features V refer to data with a fixed format and a clear organization. They are typically stored in databases, spreadsheets, or electronic medical record systems. A cross-modal attention mechanism receives a text feature vector T and structured data features V, and its algorithm can be defined as follows:

[0094]

[0095] Where A(T,V) represents the cross-modal feature representation; the text feature vector T is used to construct the query matrix; the structured data features V are used to construct key-value pairs; σ is the softmax function used to calculate the attention weights; QT i The i-th text query vector is generated from the text feature vector T; KV j The key vector of the j-th structured data is generated from the structured data feature V; VV j d is the value vector of the j-th structured data, used for final feature fusion; d is the feature dimension, used for scaling the attention score; g is the nonlinear transformation function, used for feature space mapping.

[0096] The alignment score matrix Q can be generated using the following formula:

[0097]

[0098] The alignment score matrix Q has a dimension of n×m, where n is the length of the text feature sequence and m is the number of structured data features.

[0099] The fused feature vector F can be generated using the following formula:

[0100] F = Concat(∑(A(T,V)j×VVj))

[0101] Wherein, Concat is the feature concatenation function used to generate the final fused feature vector; A(T,V) j The result of A(T,V) is related to the j-th structured data vector VV. j The corresponding attention weights; the dimension of the fused feature vector F is r×s, where r is the dimension of the fused feature vector and s is the sequence length.

[0102] By calculating the alignment between text feature vector T and structured data feature V through cross-modal feature fusion, the two data types can complement each other to improve medical reasoning capabilities.

[0103] In this embodiment, the knowledge graph enhancement module 30 is used to construct a medical knowledge graph based on medical entity relationships and to infer and generate knowledge enhancement feature vectors.

[0104] Specifically, refer to Figure 5 As shown, the knowledge graph enhancement module 30 is pre-set with a graph neural network and a pre-trained language model. The medical entity relationship includes an entity embedding matrix and a relationship embedding matrix. The adjacency matrix of the connection relationship between entities is obtained from the medical entity relationship. The adjacency matrix, entity embedding matrix and relationship embedding matrix are used as inputs to the graph neural network. The graph neural network performs feature processing to extract the structured features of the knowledge graph.

[0105] The knowledge graph enhancement module 30 receives the fused feature vector F and the text feature vector T, and extracts the semantic information of the fused feature vector based on the pre-trained language model;

[0106] By fusing the structured features of the knowledge graph with semantic information, we obtain knowledge-enhanced feature vectors and relational reasoning paths.

[0107] As an example, the knowledge graph enhancement module 30 employs a knowledge graph-enhanced reasoning mechanism, the algorithm of which can be defined as:

[0108]

[0109] Where Z is the knowledge-enhanced feature representation; E is the entity embedding matrix, used to represent medical entities; R is the medical entity relationship, used to represent the relationship between entities; A is the adjacency matrix, with a dimension of m×m, representing the connection relationship between entities, calculated from the medical entity relationship R: A=AdjMatrix(R), which is the calculation function of the adjacency matrix; GNN is the graph neural network, specifically implemented as: GNN(E,R,A)=σ(A·E·Wg+bg), where Wg is the weight matrix, bg is the bias vector, and σ is the activation function, used for graph structure feature extraction; LM is the language model, integrating and fusing the feature vector F for feature enhancement: LM(T,F)=Attention(T,F), where Attention is the attention mechanism, used to fuse the text feature vector T and the fused feature vector F; This is a feature concatenation operation used for feature fusion.

[0110] The knowledge-enhanced feature vector K can be generated using the following formula:

[0111] K = LinearProj(Z)

[0112] Where LinearProj is the linear projection function: LinearProj(Z) = Z·w k +b k , where w k Let b be the projection weight matrix. k Let K be the bias vector; K has a dimension of g×h, where g is the length of the knowledge feature sequence and h is the dimension of the knowledge feature.

[0113] The relational reasoning path RP can be generated from Z using the following formula:

[0114] RP=PathFinder(Z,E,R,TopK(A))

[0115] Among them, PathFinder is a path discovery function used to identify relationship paths between key entities; TopK is a function to select the top K strongest connections used to filter important relationships; the dimension of RP is p×q, where p is the number of paths and q is the dimension of the path description vector; the path description of each relationship reasoning path RP includes: starting entity, ending entity, intermediate nodes, relationship type and path importance score.

[0116] Large language models extract medical text information, knowledge graphs provide disease-symptom-drug-treatment relationship, and knowledge graphs enhance the reasoning mechanism to infer the disease development path and assist doctors in etiological analysis, thus achieving deep integration of medical knowledge.

[0117] The dynamic temporal modeling module 40 is used to generate temporal prediction vectors based on fused feature vectors and knowledge-enhanced feature vectors through a counterfactual reasoning mechanism.

[0118] Specifically, refer to Figure 6 As shown, the dynamic time series modeling module 40 constructs a causal reasoning graph through a counterfactual reasoning mechanism. It receives the current system state vector and uses the current system state vector, fused feature vector, and knowledge-enhanced feature vector as inputs to the causal reasoning graph. It calculates the decision value matrix and counterfactual analysis results through the decision function of the causal reasoning graph, and obtains the time series prediction vector based on the decision value matrix, fused feature vector, and knowledge-enhanced feature vector.

[0119] As an example, the dynamic time series modeling module 40 employs a counterfactual reasoning mechanism, the algorithm of which can be defined as:

[0120] D(s,a)=λ1V(s,a)+λ2C(s,a′)+λ3I(s,a)+λ4K(s,a)

[0121] Where D(s,a) is the comprehensive decision evaluation; s is the current system state vector, represented as s = Transform(F,Q), used for state representation; a is the current decision action; a' is the counterfactual decision action; V(s,a) is the basic value function, used to evaluate the value of the action; C(s,a') is the counterfactual outcome evaluation, used to simulate the consequences of the decision; I(s,a) is the causal impact measure, used to evaluate the impact of the decision; K(s,a) is the knowledge enhancement evaluation function, which integrates the knowledge enhancement feature vector K into the decision evaluation: K(s,a) = K T MLP(s, a), where MLP is a multilayer perceptron; λ1, λ2, λ3, and λ4 are adaptive weights used to balance various indicators and satisfy λ1+λ2+λ3+λ4=1.

[0122] Based on the current system state vector *s* and the current decision action *a*, counterfactual reasoning is used to calculate the possible consequences *C(s,a')* under different decision actions *a'*. This is used to simulate the potential impact of alternative decisions and quantify the differences in the benefits of different decisions to the healthcare system. By calculating the causal impact metric *I(s,a)*, the long-term impact of decision action *a* on the current system state vector *s* is measured, including key indicators such as physician workload, patient treatment outcomes, and healthcare resource consumption. A knowledge-enhanced evaluation function *K(s,a)* is combined to improve the rationality of decisions, enabling intelligent healthcare decision-making to optimize and adjust based on historical experience and knowledge graph reasoning results. Counterfactual reasoning is used to evaluate the expected effects of different treatment plans and recommend the optimal treatment decision. Different scheduling strategies are quantified based on physician workload balancing, providing a scientific basis for healthcare resource allocation, treatment recommendations, and scheduling optimization.

[0123] The decision value matrix (DV) and counterfactual analysis result (CF) of the counterfactual reasoning mechanism are generated through D(s,a):

[0124] DV = Reshape(D(s, a))

[0125] The decision value matrix DV has a dimension of b×c, where b is the decision batch size and c is the number of decision categories.

[0126] CF=Compare(D(s,a),D(s,a'))

[0127] The counterfactual analysis result (CF) has a dimension of b×2, representing the comparison between the original decision and the counterfactual decision.

[0128] The generation process of the time series prediction vector P can be defined as follows:

[0129] P = TimeSeq(F, K, DV)

[0130] Where F is the fusion feature vector; K is the knowledge enhancement feature vector; and DV is the decision value matrix.

[0131] The specific implementation of the time series prediction process TimeSeq is as follows:

[0132] TimeSeq(F,K,DV)=LSTM(Concat(F·W F F·W K DV·W D ))

[0133] Among them, W F W K W DThese are the mapping weight matrices for the fused feature vector F, the knowledge-enhanced feature vector K, and the decision value matrix DV, used for feature space alignment; Concat is the feature concatenation function, concatenating the three features along the feature dimension; LSTM is a Long Short-Term Memory network used to capture temporal dependencies, and its recursive calculation process is as follows:

[0134] h t =LSTM(Concat(F·W) F F·W K DV·W D )t,h{t-1})

[0135] P t =Linear(h t )

[0136] The temporal prediction vector P has a dimension of t×f, where t is the time step size and f is the feature dimension; P at each time step... t Specifically, it includes the following information: medical resource demand forecast, medical staff scheduling recommendations, surgical scheduling priorities, and resource utilization forecast.

[0137] In this embodiment, the adaptive scheduling optimization module 50 is used to make medical scheduling decisions based on time-series prediction vectors and generate scheduling schemes.

[0138] Specifically, the adaptive scheduling optimization module 50 adopts a multi-objective optimization algorithm, which includes multiple objective optimization functions for calculating resource utilization, professional matching degree, time smoothness and constraint violation penalty. The multi-objective optimization algorithm assigns weights to different objective optimization functions, constructs a comprehensive optimization objective function, and solves it in combination with constraint conditions to obtain the scheduling scheme.

[0139] As an example, a multi-objective optimization algorithm can be defined as:

[0140] OBJ(ScheP,P,DV,C)=α1·U(ScheP)+α2·E(ScheP)+α3·F(ScheP)-β·Const(ScheP,C)

[0141] Wherein, OBJ(ScheP, P, DV, C) is the objective function for evaluating the scheduling scheme; ScheP is the candidate scheduling scheme; P is the time series prediction vector; DV is the decision value matrix; C is the set of medical environment constraints, including constraints such as personnel availability and equipment limitations; α1, α2, α3 and β are weight coefficients used to balance the importance of each optimization objective, and satisfy α1+α2+α3=1.

[0142] In multi-objective optimization algorithms, U(ScheP) is a resource utilization function that evaluates the efficiency of scheduling scheme s in utilizing medical resources.

[0143] U(ScheP)=∑(w i .Util(r i ,ScheP))

[0144] Where, r i For the i-th type of medical resource; Util is the utilization rate calculation function; w i This refers to the resource weighting coefficient.

[0145] E(ScheP) is a professional fit function that assesses the degree of match between the healthcare professional and the task:

[0146] E(ScheP)=∑(Exp(p j , t j ))

[0147] Where, p j For medical staff j;t j The task assigned to p_j; Exp is the function for calculating the professional matching degree;

[0148] F(ScheP) is the time fluency function, which evaluates the time continuity of the scheduling:

[0149] F(ScheP)=∑(Flow(t k , t k+1 ))

[0150] Among them, t k and t k+1 For task scheduling over consecutive time periods, Flow is the function for calculating time smoothness.

[0151] Const(ScheP,C) is the constraint violation penalty function, which penalizes scheduling schemes that do not meet the constraints.

[0152] Const(ScheP,C)=∑(Violation(ScheP,c i )·ρ i )

[0153] Among them, c i Let ρ be the i-th constraint, and let Violation be the function for calculating the degree of violation. i This is the penalty coefficient.

[0154] This multi-objective optimization algorithm combines heuristic search and reinforcement learning, and its implementation process is as follows:

[0155] Step 1: Initialize the candidate scheduling scheme set Ω = {ScheP1,ScheP2,...,ScheP} m}, generated based on the time-series prediction vector P.

[0156] Step 2: For each candidate scheduling scheme, Schep i Calculate the score OBJ (ScheP) i (P, DV, C).

[0157] Step 3: Optimize the solution using a genetic algorithm:

[0158] Selection(Ω): Based on OBJ(ScheP) i (P, DV, C) Select the highest-scoring option;

[0159] Crossover(Ω): Generates a new scheduling scheme through crossover;

[0160] Mutation (Ω): Mutation to explore new solution spaces;

[0161] Iterate and optimize until convergence or the maximum number of iterations is reached.

[0162] Step 4: Dynamically adjust using reinforcement learning:

[0163] The weights α1, α2, α3, and β are updated based on historical scheduling performance.

[0164]

[0165] Where Q(ScheP,a) is the state-action value function, representing the long-term cumulative reward that can be obtained based on the optimal scheduling strategy; R(ScheP,a) is the reward for the current scheduling action; and γ is the discount factor.

[0166] ScheP' represents the new state reached after executing a; Q is the optimal action for the future, representing the maximum long-term benefit that can be obtained by taking the optimal action a' starting from the new state s'.

[0167] Based on the Q-learning update strategy in reinforcement learning, the system can dynamically optimize medical scheduling, resource allocation, and intelligent diagnosis and treatment decisions by calculating the immediate reward R(s,a) of the current decision and the optimal Q value in the future.

[0168] Scheduling plan O is generated in the following way:

[0169] O = PostProcess(s * )

[0170] Among them, s *The scheduling scheme obtained by the multi-objective optimization algorithm:

[0171] This indicates that among all possible candidate scheduling schemes, ScheP i In the process, we find the candidate scheduling scheme ScheP that maximizes the OBJ(ScheP,P,DV,C) function; PostProcess is a post-processing function used to refine and format the scheduling scheme; the data structure of the scheduling scheme O is a three-dimensional matrix of k×m×n, where k is the time dimension, representing the number of time periods in the scheduling; m is the personnel dimension, representing the number of medical staff; and n is the task dimension, representing the number of task types. O[k,m,n]=1 means that personnel m are assigned to execute task n in time period k, otherwise it is 0.

[0172] The scheduling plan also includes core information such as: work schedules for medical staff, operating room and equipment allocation plans, task priorities and execution sequences, emergency resource reservation plans, and resource utilization forecasts for each time period.

[0173] The adaptive scheduling optimization module integrates multimodal medical data features, predicts future scheduling needs based on dynamic time series modeling, and combines reinforcement learning and multi-objective optimization methods to achieve adaptive optimization of medical scheduling schemes. This improves the utilization rate of medical resources, ensures a balanced workload for doctors, and enhances the rationality and flexibility of scheduling schemes.

[0174] In this embodiment, the scheduling calibration module 60 is used to match similar case features to the scheduling scheme based on the knowledge-enhanced feature vector through a context learning mechanism, and calibrate the scheduling scheme based on the similar case features to generate an optimized scheduling scheme.

[0175] Specifically, the scheduling calibration module 60 includes: a scheduling scheme reasoning unit, a case matching unit, and a prediction calibration unit;

[0176] The scheduling scheme reasoning unit is used to construct a scheduling optimization reasoning path based on historical reasoning trajectories, knowledge-enhanced feature vectors, and query vectors formed by scheduling schemes. It also uses a thinking chain mechanism to construct the scheduling optimization reasoning path and calculates the rationality of the scheduling scheme by combining probability distribution. The reasoning confidence score and scheduling reasoning step sequence are obtained through the scheduling optimization reasoning path. The reasoning confidence score is used to evaluate the credibility of the scheduling scheme.

[0177] As an example, the scheduling scheme reasoning unit uses a chain-of-thought mechanism for reasoning. The algorithm of the chain-of-thought mechanism can be defined as follows:

[0178] CoT(q,K)=Reasoning(q,K,n)

[0179] Where CoT(q,K) represents the reasoning result; q is the current query vector, constructed based on the scheduling scheme O: q = Query(O), used for task definition; Reasoning is the recursive reasoning function, which generates n reasoning steps step by step, defined as:

[0180] Reasoning(q,K,n)={s1,s2,...,s n}

[0181] s i =argmax[P(s|s ＜i ,q,K)×V(s,K)]

[0182] Wherein, P(s|s) ＜i Let ,q,K) be the conditional generation probability, representing the probability generated given historical reasoning s. ＜i Given the query q and the knowledge-enhanced feature vector K, the probability of generating the current reasoning step s is given; V(s,K) is the effectiveness scoring function based on knowledge features: V(s,K)=cosine(Embed(s),K), where cosine is the cosine similarity and Embed is the text embedding function.

[0183] The reasoning step sequence SS is generated using the following formula:

[0184] SS = CoT(q,K) = {s1,s2,...,s} n}

[0185] The structure of the reasoning step sequence SS includes: initial analysis: preliminary reasoning based on query q and time-series prediction vector P; knowledge integration: supplementing medical expertise based on knowledge-enhancing feature vector K; scheme demonstration: step-by-step reasoning on the rationality of scheduling scheme O; risk assessment: analyzing the potential risks of the scheme based on the reasoning results.

[0186] The inference confidence score (SC) is calculated using the following formula:

[0187]

[0188] Where ∏ represents a multiplication operation, taking into account the credibility of all reasoning steps; max represents selecting the reasoning step with the highest probability; and the reasoning confidence SC is a scalar value between 0 and 1, representing the overall credibility of the entire reasoning chain.

[0189] The chain-of-thought reasoning mechanism is used to deduce the decision-making process step by step, generating an interpretable sequence of reasoning steps to ensure the traceability of the system's reasoning path. This ensures a clear logical chain for the scheduling plan.

[0190] The case matching unit is used to perform similar case matching based on the historical case database and the scheduling reasoning step sequence, using a retrieval function to retrieve the historical cases most similar to the scheduling optimization scheme, and calculate similar case features and case relevance scores based on the historical cases.

[0191] As an example, the case matching unit employs an in-context learning mechanism for case retrieval and similarity matching, and its algorithm can be defined as follows:

[0192]

[0193] Where M is the historical case library, used for knowledge retrieval; x is the current input feature, consisting of the scheduling scheme O, the time-series prediction vector P, and the decision confidence DO: The first part is the feature concatenation operation, used to combine the original features and the retrieval results; f is the feature fusion function, defined as: f(z) = MLP(z), where MLP is a multilayer perceptron; Retriever is the similar case retrieval function, which combines the inference step sequence to improve the retrieval quality.

[0194] Retriever(x,SS,M) = TopK(Similarity(x,M,SS)), where Similarity is the similarity calculation function:

[0195] Similarity(x,M,SS)=α·sim(x,M)+β·sim(SS,M reasoning ), sim is the vector similarity function, M reasoning The reasoning steps of historical cases are recorded, and α and β are weighting coefficients used to balance feature similarity and reasoning similarity.

[0196] Similar case features (SF) are generated using the following formula:

[0197] SF=ExtractFeatures((x,SS,M))

[0198] Among them, ExtractFeatures is the feature extraction function, which extracts key features from the retrieved cases; the dimension of the similar case feature SF is k×d, where k is the number of retrieved cases and d is the feature dimension; each similar case feature SF includes: historical scheduling pattern, resource allocation strategy, time arrangement plan and performance evaluation indicators.

[0199] The case relevance score (SR) is generated using the following formula:

[0200] SR=SimilarityScore(x,Retriever(x,SS,M),SS)

[0201] Where SimilarityScore is the detailed similarity scoring function:

[0202]

[0203] μ is the weighting coefficient for reasoning similarity; the dimension of the case relevance score SR is k×1, representing the relevance score of each retrieved case to the current situation.

[0204] In-Context Learning mechanisms are used to retrieve historical similar cases and make decision adjustments based on the characteristics of similar cases, thereby optimizing scheduling plans or medical diagnosis recommendations.

[0205] The prediction calibration unit is used to calculate the scheduling adjustment factor based on similar case features, case relevance scores, and time-series prediction vectors, and then optimize the scheduling scheme using the scheduling adjustment factor to obtain the optimized scheduling scheme.

[0206] As an example, the predictive calibration process employs an uncertainty quantification mechanism for calculation, and its algorithm can be defined as follows:

[0207] O′=O+Calibrate(SF,SR,P,DO)

[0208] Wherein, Calibrate is the calibration function, which is adjusted by combining similar case features SF, case relevance score SR, time series prediction vector P, original prediction and decision confidence DO:

[0209] Calibrate(SF,SR,P,DO)=∑(SR i ×SF i )×λ(DO)

[0210] Where λ(DO) is the adjustment coefficient based on decision confidence, which reflects the reliability and credibility of the scheduling plan. It can be set according to the actual situation, typically ranging from 0.55 to 0.8. When the DO value is high, the adjustment range is reduced; when the DO value is low, the adjustment range is increased. ' This is the final optimized scheduling plan after calibration.

[0211] The uncertainty quantification mechanism calculates decision credibility by integrating features of similar cases and the confidence level of thought chain reasoning, and uses this credibility to calibrate scheduling plans. Simultaneously, the scheduling plan is dynamically adjusted and optimized based on time-series prediction vectors to identify potential scheduling conflicts, underutilization of resources, or overload, and adaptively corrects the plan. This allows for an assessment of whether predictive information was appropriately utilized in the decision-making process, leading to more accurate calibration.

[0212] In this embodiment, the medical intelligent causal decision-making and scheduling system further includes: receiving the counterfactual analysis results output by the dynamic time series modeling module and the prompt information relevance score matrix from the pre-trained language model module; generating decision explanation text for the optimized scheduling scheme based on the counterfactual analysis results and the prompt information relevance score matrix; and calculating the credibility of the decision explanation text through an explanation evaluation function to obtain an explanation credibility score.

[0213] As an example, the algorithm used by the natural language interpretation generation module can be defined as follows:

[0214] G(x,y)=argmax[P(s|x,y)×R(s)]

[0215] Where x is the input feature, which consists of the counterfactual analysis result (CF) and contextual information: y represents the prediction result, constructed based on the optimized scheduling decision: y = Decoder(O'); s represents the generated decision explanation text.

[0216] The decision explanation text (ET) can be generated using the following formula:

[0217] ET = Template(top) k (G(x,y)), PS,RP)

[0218] Among them, top k The function selects the k candidate explanatory texts with the highest probabilities; PS is the cue information relevance score matrix; RP is the relational reasoning path, used to assist in the generation of explanatory text; Template is a template-based text generation function that uses the template in the cue information relevance score matrix PS to optimize the output text structure, and its definition is:

[0219] Template(S,PS)=∑(PS i ×Pattern i (S))

[0220] Among them, PS i For the relevance score of the i-th template, Pattern i Let i be the i-th text pattern function.

[0221] The structure of the decision explanation text ET includes: a decision summary section (outlining the key points of the scheduling decision); a causal analysis section (based on the causal relationship of the decision); a comparative analysis section (showing the advantages and disadvantages of different decision options); and a suggestion section (providing possible directions for further optimization).

[0222] The Explanation Credibility Score (EC) can be generated using the following formula:

[0223] EC = Evaluate(ET, OF)

[0224] Where Evaluate is the interpretation evaluation function, which evaluates the generated interpretation text, and is defined as follows:

[0225] Evaluate(ET,CF)=α·Consistency(ET,CF)+β·Completeness(ET)+γ·Clarity(ET)

[0226] Among them, Consistency is a measure of consistency between the explanation and the counterfactual analysis results; Completeness is a measure of the completeness of the explanation; Clarity is a measure of the clarity of the explanation; α, β, and γ are weighting coefficients used to balance the importance of each evaluation indicator.

[0227] The credibility score (EC) is calculated using a 1×3 scale, which includes scores for consistency, completeness, and clarity.

[0228] Compared to existing medical decision-making systems, which are generally characterized by a "black box" problem and lack of explainability, the natural language interpretation generation module combines counterfactual analysis results (CF) and a relevance score matrix (PS) of prompt information to generate an explainable decision reasoning path. Through the natural language generation mechanism G(x,y), it outputs the decision explanation text ET, ensuring the transparency of the medical decision-making process and increasing the trust of doctors and administrators in intelligent decision-making.

[0229] In this embodiment, the medical intelligent causal decision-making and scheduling system adopts a multi-agent collaborative language model integration framework. The language model integration framework is used to receive the output of each module, construct the input variables of each agent based on the output of each module, integrate the reasoning results of each agent through the result aggregation function, and obtain the final scheduling decision scheme, comprehensive confidence score and decision explanation report output by the language model integration framework based on the integration result.

[0230] As an example, the algorithm for a multi-agent collaborative language model ensemble framework can be defined as:

[0231] Y = Aggregator(Σw i ×LM i (x i )×Trust i )

[0232] Among them, the multi-agent collaborative language model integration framework receives the outputs of each module and constructs a specific input x. ix1 = Transform1(T,F), where T is the text feature vector and F is the fused feature vector; x2 = Transform2(DV,P,DO,O) ' ), where DV is the decision value matrix, P is the time series prediction vector, DO is the decision confidence level, and O ' To optimize the scheduling scheme; x3 = Transform3(EC,SC), where EC is the explanation confidence score and SC is the reasoning confidence; x4 = Transform4(SR,K), where SR is the case relevance score and K is the knowledge enhancement feature vector.

[0233] LM i For the i-th language model, optimizations are performed for specific tasks: LM1 focuses on text understanding and feature representation; LM2 focuses on decision evaluation and predictive analysis; LM3 focuses on credibility assessment and uncertainty quantification; and LM4 focuses on knowledge integration and case reasoning.

[0234] w i The weights are dynamic and adjusted based on the output of each module.

[0235] w i =Softmax(MLP([EC,SC,SR,DO]))

[0236] Trust i Credibility scoring is calculated based on historical performance:

[0237] Trust i =α·Accuracy i +β·Consistency i +γ·Calibration i

[0238] Where α, β, and γ are weighting coefficients.

[0239] Aggregator is the result aggregation function, defined as:

[0240] Aggregator(z)=WeightedSum(z, ConfidenceScores(z))

[0241] Where WeightedSum is a weighted summation function and ConfidenceScores is a confidence score function.

[0242] The final scheduling decision (FD) is generated using the following formula:

[0243] FD = ExtractDecision(Y)

[0244] ExtractDecision is the decision extraction function that extracts scheduling decisions from the aggregation results; the dimension of FD is the same as that of the scheduling scheme O'.

[0245] The overall confidence score (CS) is generated using the following formula:

[0246] CS=CalculateConfidence(Y,{w i}, {Trust i})

[0247] CalculateConfidence is the confidence calculation function, which calculates the overall confidence based on the aggregation results and the weights of each model; CS has a dimension of 1×3, including scores for three dimensions: accuracy, consistency and calibration.

[0248] The Decision Explanation Report (DR) is generated using the following formula:

[0249] DR=GenerateReport(Y, EC, CF, SS)

[0250] Among them, GenerateReport is the report generation function, which combines the aggregation results and the explanatory outputs of each module to generate a comprehensive report; the structure of DR includes four parts: decision summary, causal analysis, counterfactual comparison and credibility assessment.

[0251] The multi-agent collaborative language model integration framework includes multiple agents, each responsible for different tasks. Through the division of labor and cooperation among multiple agents, scheduling, diagnosis and treatment, and resource allocation can be optimized in complex medical decision-making environments, thereby improving the robustness and adaptability of the system.

[0252] In summary, the medical intelligent causal decision-making and scheduling system based on an improved large language model provided in this application processes unstructured medical text data to extract text feature vectors and medical entity relationships; combines the text feature vectors with structured data and generates fused feature vectors through a cross-modal alignment mechanism; generates time-series prediction vectors based on the fused feature vectors and knowledge-enhanced feature vectors using a counterfactual reasoning mechanism; makes scheduling decisions based on the time-series prediction vectors and uses features from similar historical cases for scheme calibration. This application overcomes the shortcomings of existing medical scheduling and decision-making systems, enabling more accurate allocation of medical resources and doctor scheduling. It not only improves the intelligence level of medical scheduling but also promotes the transformation of medical decision-making from an "experience-driven" to a "data-driven + causal reasoning" model, providing an innovative solution for the intelligent management of medical institutions.

[0253] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.

[0254] The flowcharts or block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of devices, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-specific system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0255] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily made by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims, and the above embodiments should be regarded as exemplary and non-limiting.

Claims

1. A medical intelligent causal decision-making and scheduling system based on an improved large language model, characterized in that, include: The pre-trained language model module is used to process unstructured medical text data and extract text feature vectors and medical entity relationships. The knowledge graph enhancement module is used to construct a medical knowledge graph based on the relationships between the medical entities and to infer and generate knowledge-enhanced feature vectors. The multimodal feature fusion module is used to generate a fused feature vector based on the text feature vector and pre-stored structured data through a cross-modal alignment mechanism; The dynamic temporal modeling module is used to generate a temporal prediction vector based on the fused feature vector and the knowledge-enhanced feature vector through a counterfactual reasoning mechanism. An adaptive scheduling optimization module is used to make medical scheduling decisions based on the time-series prediction vector and generate a scheduling plan. The scheduling calibration module is used to match similar case features to the scheduling scheme based on the knowledge-enhanced feature vector through a context learning mechanism, and to calibrate the scheduling scheme based on the similar case features to generate an optimized scheduling scheme.

2. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 1, characterized in that, The pre-trained language model module employs a medical domain knowledge-enhanced prompting learning mechanism. It parses the feature information of the smallest semantic unit in the unstructured medical text data according to the set first prompt information, converts it into the text feature vector, and extracts the core medical entities in the unstructured medical text data according to the set second prompt information, generating the medical entity relationship.

3. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 2, characterized in that, The pre-trained language model module is further used to: generate a dynamic prompt information template set and a prompt information relevance score matrix through a dynamic prompt information generation mechanism. The dynamic prompt information generation mechanism is obtained by using historical data encoding, medical context encoding, and time encoding as inputs and processing them using a self-attention mechanism. The dynamic prompt information template set is used to optimize the selection of prompt information in the prompt learning mechanism.

4. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 1, characterized in that, The multimodal feature fusion module uses a cross-modal attention mechanism to calculate the similarity between the text feature vector and the structured data features, and maps them to a unified feature space for feature alignment calculation to generate the fused feature vector and alignment score matrix.

5. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 1, characterized in that, The dynamic time series modeling module constructs a causal reasoning graph through a counterfactual reasoning mechanism. It receives the current system state vector and uses the current system state vector, the fused feature vector, and the knowledge-enhanced feature vector as inputs to the causal reasoning graph. It calculates the decision value matrix and counterfactual analysis results through the decision function of the causal reasoning graph, and obtains the time series prediction vector based on the decision value matrix, the fused feature vector, and the knowledge-enhanced feature vector.

6. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 4, characterized in that, The knowledge graph enhancement module is pre-set with a graph neural network and a pre-trained language model. The medical entity relationship includes an entity embedding matrix and a relationship embedding matrix. The adjacency matrix of the connection relationship between entities is obtained from the medical entity relationship. The adjacency matrix, the entity embedding matrix and the relationship embedding matrix are used as inputs to the graph neural network. Feature processing is performed through the graph neural network to extract the structured features of the knowledge graph. The knowledge graph enhancement module receives the fused feature vector and the text feature vector, and extracts the semantic information of the fused feature vector based on the pre-trained language model; The structured features of the knowledge graph are fused with the semantic information to obtain the knowledge-enhanced feature vector and the relational reasoning path.

7. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 1, characterized in that, The adaptive scheduling optimization module employs a multi-objective optimization algorithm, which includes multiple objective optimization functions for calculating resource utilization, professional matching degree, time fluency, and constraint violation penalties. The multi-objective optimization algorithm assigns weights to different objective optimization functions to construct a comprehensive optimization objective function, which is then solved in conjunction with the constraint conditions to obtain the scheduling scheme.

8. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 1, characterized in that, The scheduling calibration module includes: a scheduling scheme reasoning unit, a case matching unit, and a prediction calibration unit; The scheduling scheme reasoning unit is used to construct a scheduling optimization reasoning path based on historical reasoning trajectories, the knowledge-enhanced feature vector, and the query vector formed by the scheduling scheme, using a thinking chain mechanism. It also calculates the rationality of the scheduling scheme by combining probability distribution. The reasoning confidence score and the scheduling reasoning step sequence are obtained through the scheduling optimization reasoning path. The reasoning confidence score is used to evaluate the credibility of the scheduling scheme. The case matching unit is used to perform similar case matching based on the historical case database and the scheduling reasoning step sequence, through a retrieval function, to retrieve the historical case most similar to the scheduling optimization scheme, and to calculate similar case features and case relevance scores based on the historical case. The prediction calibration unit is used to calculate the scheduling adjustment factor based on the similar case features, the case correlation score and the time-series prediction vector, and optimize the scheduling scheme through the scheduling adjustment factor to obtain the optimized scheduling scheme.

9. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 1, characterized in that, Also includes: The natural language interpretation generation module receives the counterfactual analysis results output by the dynamic temporal modeling module and the prompt information relevance score matrix from the pre-trained language model module. Based on the counterfactual analysis results and the prompt information relevance score matrix, it generates decision interpretation text for the optimized scheduling scheme and calculates the credibility of the decision interpretation text through an interpretation evaluation function to obtain an interpretation credibility score.

10. The medical intelligent causal decision-making and scheduling system based on an improved large language model according to claim 1, characterized in that, The medical intelligent causal decision-making and scheduling system adopts a multi-agent collaborative language model integration framework. The language model integration framework is used to receive the output of each module, construct the input variables of each agent based on the output of each module, integrate the reasoning results of each agent through the result aggregation function, and obtain the final scheduling decision scheme, comprehensive confidence score and decision explanation report output by the language model integration framework based on the integration result.

Images