A hospital intelligent recruitment method and system based on multi-source fusion intelligent decision

Abstract

The application provides a hospital intelligent recruitment and adoption multi-source fusion intelligent decision method and system, and relates to the technical field of medical material supply chain management and intelligent decision. The method generates unified data representation through standardized coding, entity disambiguation and space-time alignment based on the multi-source heterogeneous data of the internal system of the hospital and external supply; an associated model of materials, disease types, suppliers and policy constraints is constructed based on a knowledge graph, and the demand and supply changes are predicted by using a space-time graph neural network to form a demand and supply matching degree matrix; a dynamic state vector of the supplier is constructed by combining multi-modal feature fusion, and a cascading failure path is predicted by analyzing the risk propagation of the supply chain; on this basis, a multi-objective optimization algorithm is used to solve the Pareto optimal procurement scheme; and the model is updated online and optimized in a closed loop through execution feedback. The application improves the procurement prediction accuracy and the supply chain risk control capability.

Description

Technical Field

[0001] This invention belongs to the field of medical supplies supply chain management and intelligent decision-making technology, specifically involving a multi-source integrated intelligent decision-making method and system for smart hospital procurement. Background Technology

[0002] In the operation of medical institutions, hospital procurement and bidding involve multiple categories of supplies, including medicines, consumables, and medical equipment. The procurement process typically requires comprehensive consideration of factors such as clinical needs, inventory levels, supplier capabilities, and relevant policy constraints. Existing hospital procurement management systems mostly rely on manual experience or rule-based management methods, which have limitations in demand forecasting, supply matching, and procurement decision-making.

[0003] In existing technologies, hospital procurement data is usually stored in hospital information systems (HIS), enterprise resource management systems (ERP), procurement management systems, and inventory management systems. At the same time, supplier information, market price information, and bidding and procurement information come from different external platforms. The lack of a unified data structure and semantic relationship between the data sources leads to strong heterogeneity and weak correlation among the data, making it difficult to form a unified data analysis foundation.

[0004] Furthermore, traditional procurement decision-making methods typically rely on historical averages, empirical thresholds, or static statistical models for demand forecasting, making it difficult to reflect the dynamic characteristics of clinical demand changes over time, and also failing to characterize the fluctuations in supply capacity within the supply chain. In supply chain management, existing methods often rely on single supplier evaluation indicators, lacking the ability to comprehensively model complex network structures such as equity relationships, partnerships, and logistical dependencies among suppliers, thus hindering the identification of potential cascading supply chain risks. Simultaneously, current technologies for analyzing supply chain risks often remain at the level of local risk identification, lacking systematic analysis methods for the propagation paths and diffusion trends of risks within the supply network, making it difficult to identify cascading failure paths that could lead to supply disruptions in advance. Regarding decision optimization, traditional methods typically consider only cost or a single evaluation indicator, lacking the comprehensive optimization capability under multiple objective constraints such as procurement costs, quality stability, supply continuity, and risk propagation. Additionally, existing procurement systems typically only record simple results after execution, lacking a mechanism for continuously updating predictive and decision models based on execution feedback data, making the system ill-suited to adapt to changes in the market environment and dynamic changes in the supply chain.

[0005] Therefore, how to achieve intelligent decision-making and dynamic adaptive optimization of the entire hospital procurement process by integrating multi-source heterogeneous data, constructing a unified knowledge representation model, and combining spatiotemporal prediction, risk propagation analysis, and multi-objective optimization methods has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This invention provides a multi-source fusion intelligent decision-making method and system for smart procurement in hospitals, aiming to solve the technical problems mentioned in the background.

[0007] In a first aspect, the present invention provides a multi-source fusion intelligent decision-making method for smart procurement in hospitals, comprising: Step S1: Collect data from the hospital's internal systems and external supply chain, and perform standardized coding, entity disambiguation, spatiotemporal alignment, and semantic mapping on the heterogeneous data to generate a unified data representation; Step S2: Construct a hospital procurement knowledge graph based on the unified data representation, model the relationships between materials, diseases, clinical departments, suppliers, inventory nodes and policy constraints as a graph structure, and generate graph embedding vectors; Step S3: Construct a demand and supply dynamic graph based on the knowledge graph, and use a spatiotemporal graph neural network to model the demand and supply dynamic graph to predict the demand, inventory changes and supply capacity of each material in the future cycle, and generate a demand and supply matching degree matrix. Step S4: Integrate supplier performance data, quality inspection data, delivery timeliness data, market fluctuation data, and public opinion data from the unified data representation to construct a dynamic status vector for suppliers; Step S5: Construct a supply chain relationship graph, calculate the probability and vulnerability of node risk propagation, analyze the risk diffusion trend based on risk propagation entropy, predict cascading failure paths, and generate a set of alternative supply chain candidates. Step S6: With procurement cost, quality stability, supply continuity and risk propagation intensity as objectives, and combined with the demand-supply matching degree matrix, supplier dynamic state vector, cascading failure path and alternative supply chain candidate set as constraints, a multi-objective optimization algorithm is used to solve for the Pareto optimal procurement scheme.

[0008] Further, step S1 specifically includes: collecting multi-source heterogeneous data from hospital HIS system, ERP system, inventory management system, procurement management system, supply chain management system, as well as external public resource trading platform, market price database, third-party testing platform, and online public opinion monitoring platform; performing unified coding mapping on material codes, supplier identifiers, and disease information from different data sources in the multi-source heterogeneous data; and performing disambiguation processing on cross-system homonymous entities based on text similarity calculation and attribute feature matching algorithms; and achieving cross-source data entity alignment by combining timestamp consistency constraints, thereby generating a unique entity identifier set.

[0009] Furthermore, step S2 specifically includes: Based on a unified data representation, entity identification and classification are performed on business database records from the hospital's internal systems and external procurement text data. The business database records include structured field data from the HIS system, ERP system, and inventory management system. The structured field data includes at least material codes, supplier identifiers, inventory quantities, and transaction records. The external procurement text data includes tender documents, contract texts, and policy and regulatory texts. The structured field data is standardized and encoded based on field mapping rules, and the procurement text data is extracted into entities and relations through a natural language processing model to generate entity-relation triples. Based on the entity relationship triples, a heterogeneous graph structure is constructed, which includes material nodes, disease nodes, clinical department nodes, supplier nodes, inventory nodes, and policy constraint nodes. Demand relationship edges, supply relationship edges, usage relationship edges, inventory flow relationship edges, and constraint relationship edges are constructed between different nodes. Based on the heterogeneous graph structure, a multi-relationship adjacency matrix and node feature vectors are constructed. A graph neural network is then used to perform representation learning on the heterogeneous graph structure to generate graph embedding vectors corresponding to each node.

[0010] Furthermore, in step S3, the construction of the demand-supply dynamic diagram includes: Based on the hospital procurement knowledge graph, a time dimension is introduced, dividing the nodes into clinical demand nodes, material nodes, inventory nodes, and supplier nodes. Demand consumption relationship edges, supply relationship edges, and inventory flow relationship edges are constructed, with the weight of each edge dynamically updated over time based on historical consumption frequency, supply fulfillment records, and inventory change rate, to form a demand and supply graph structure that evolves over time.

[0011] Further, in step S3, a spatiotemporal graph neural network is used to model the demand-supply dynamic graph to predict the demand, inventory changes, and supply capacity of various materials in future periods, generating a demand-supply matching degree matrix, specifically including: The demand-supply dynamic graph is processed by time slicing to construct a dynamic graph sequence input, and node representation vectors are generated by combining node structural features, time statistical features, and attribute features. A graph neural network is used to model the spatial dependencies between nodes, and a temporal attention mechanism is used to model the temporal dependencies of node states in order to predict the demand, inventory changes, and supply capacity changes of each material in the future cycle. Based on the prediction results, a demand-supply matching degree matrix is ​​calculated to represent the matching relationship between supply and demand.

[0012] Furthermore, step S4 specifically includes: The system acquires supplier-related data from multiple sources, including the hospital's procurement management system, external public resource trading platforms, third-party testing platforms, and online public opinion monitoring platforms. This data includes: performance data from the hospital's procurement management system, consisting of historical supplier performance records; delivery time data from order placement and delivery timestamps; quality testing results generated by third-party testing platforms; market price index data and procurement transaction price fluctuation data collected by public resource trading platforms; and text-based public opinion data related to suppliers obtained from online public opinion monitoring platforms. After aligning the multi-source supplier-related data with time windows, statistical feature vectors of performance data, quality inspection data, and delivery timeliness data are extracted respectively. Time series modeling is performed on the market price index data and procurement transaction price fluctuation data to generate trend feature vectors, and semantic encoding is performed on public opinion data to generate text feature vectors. Feature vectors from different modalities are input into a multimodal fusion network, weighted and fused through an attention mechanism to generate a unified supplier state representation vector. The state representation vector is then dynamically updated based on a time decay mechanism to form a dynamic supplier state vector that evolves over time.

[0013] Furthermore, step S5 specifically includes: Based on the equity, control, cooperation and logistics dependencies among suppliers in the external supply chain data, a multi-type heterogeneous supply chain association graph is constructed, and different relationship edges are assigned corresponding propagation weights. Based on the association graph, the basic risk level of each node is calculated, and the risk propagation probability is calculated based on the weighted adjacency relationship between nodes to characterize the diffusion process of risk in the supply chain network. A risk propagation distribution model is constructed based on the node risk propagation probability, and the risk propagation entropy and its time change rate are calculated to characterize the concentration and evolution trend of risk diffusion. When the node risk level, node vulnerability, or risk propagation entropy change rate exceeds a preset threshold, the potential cascading failure path is predicted based on the graph propagation path simulation method, and an alternative supply chain candidate set is generated based on the neighborhood structure and connectivity of the failed node.

[0014] Furthermore, step S6 specifically includes: An optimization model is constructed with the allocation relationship between suppliers and demand nodes as decision variables. The objective functions include minimizing procurement costs, maximizing quality stability, maximizing supply continuity, and minimizing risk propagation intensity. The demand-supply matching degree matrix is ​​used as the allocation constraint, the supplier dynamic state vector is used as the supply capacity constraint, the cascading failure path is used as the non-selectable constraint, and the alternative supply chain candidate set is limited to the feasible solution space. The optimization model is solved using a multi-objective evolutionary algorithm based on non-dominated sorting to obtain a Pareto optimal solution set. Each Pareto optimal solution corresponds to a set of structured procurement resource allocation schemes. The procurement resource allocation schemes include: the material allocation ratio or procurement quantity corresponding to each supplier, the allocation path relationship of each material among different suppliers, the risk level identifier of the corresponding supplier, the supply chain alternative path selection results, and the cost indicators, quality indicators, supply continuity indicators, and risk propagation indicators of the corresponding schemes. Select a procurement resource allocation scheme that satisfies the constraints from the Pareto optimal solution set.

[0015] Furthermore, after step S6, the process further includes: collecting execution feedback data, constructing prediction error feedback, and performing online incremental updates on the spatiotemporal graph neural network, supplier dynamic state vector, and multi-objective optimization algorithm to achieve adaptive closed-loop optimization; wherein, the execution feedback data includes actual execution data from the hospital procurement management system and inventory management system, and the actual execution data includes purchase order execution records, actual material consumption data, inventory change data, and supplier fulfillment data; the prediction error feedback is constructed based on the deviation between the execution feedback data and the prediction output result of step S3, including demand prediction error, inventory change error, and supply capacity error.

[0016] Secondly, the present invention provides a multi-source fusion intelligent decision-making system for hospital intelligent procurement, comprising: The data acquisition and processing module is used to collect data from the hospital's internal systems and external supply chain data, and to perform standardized encoding, entity disambiguation, spatiotemporal alignment and semantic mapping on heterogeneous data to generate a unified data representation. The knowledge graph construction module is used to construct a hospital procurement knowledge graph based on the unified data representation, model the relationships between materials, diseases, clinical departments, suppliers, inventory nodes and policy constraints as a graph structure, and generate graph embedding vectors; The demand and supply forecasting module is used to construct a dynamic demand and supply graph based on the knowledge graph, and to model the dynamic demand and supply graph using a spatiotemporal graph neural network to predict the demand, inventory changes and supply capacity of each material in the future cycle, and to generate a demand and supply matching degree matrix. The supplier status modeling module is used to integrate supplier performance data, quality inspection data, delivery timeliness data, market fluctuation data, and public opinion data in the unified data representation to construct a dynamic supplier status vector. The supply chain risk analysis module is used to construct a supply chain relationship graph, calculate the probability of risk propagation at nodes and the vulnerability of nodes, analyze the risk diffusion trend based on risk propagation entropy, predict cascading failure paths, and generate a set of alternative supply chain candidates. The multi-objective optimization decision module is used to solve the Pareto optimal procurement scheme by taking procurement cost, quality stability, supply continuity and risk propagation intensity as objective functions, and combining the demand-supply matching degree matrix, supplier dynamic state vector and cascading failure path and alternative supply chain candidate set as constraints. The adaptive closed-loop update module is used to collect execution feedback data, construct prediction error feedback, and perform online incremental updates on the spatiotemporal graph neural network, supplier dynamic state vector, and multi-objective optimization algorithm to achieve adaptive closed-loop optimization of the system.

[0017] This invention offers the following advantages: First, by constructing a hospital procurement knowledge graph based on multi-source heterogeneous data fusion and combining it with a spatiotemporal graph neural network to model the dynamic graph of demand and supply, it achieves dynamic prediction of hospital material demand, inventory changes, and supply capacity, thereby improving the accuracy and time-series response capability of procurement demand prediction. Second, by introducing a multimodal feature fusion mechanism for suppliers, it maps supplier performance data, quality inspection data, delivery timeliness data, and market and public opinion information into a unified dynamic state vector for suppliers, achieving continuous characterization of the comprehensive state of suppliers and improving the comprehensiveness and stability of supplier evaluation. Third, by constructing a supply chain relationship graph based on equity relationships, cooperative relationships, and logistics dependencies, and introducing risk propagation probability... The system employs risk propagation entropy analysis to identify the risk diffusion process and cascading failure paths in the supply chain in advance, thereby enhancing the supply chain risk early warning capability and system robustness. Fourth, by introducing a Pareto optimal solution mechanism under multi-objective constraints, a comprehensive balance is achieved between procurement cost, quality stability, supply continuity, and risk propagation intensity, improving the global optimization capability of procurement decisions. Fifth, by performing online incremental updates to the spatiotemporal graph neural network, supplier state modeling model, and multi-objective optimization model based on procurement execution feedback data, the system can continuously correct the prediction model and optimization strategy according to the actual execution results, thereby realizing the adaptive closed-loop optimization capability of the procurement decision system and improving the long-term stability and adaptability of the system in dynamic environments. Attached Figure Description

[0018] Figure 1 A flowchart of a multi-source fusion intelligent decision-making method for smart hospital procurement provided by the present invention; Figure 2 A comparison diagram showing the effects before and after implementation of the multi-source fusion intelligent decision-making method for hospital intelligent procurement provided by this invention; Figure 3 The present invention provides a structural block diagram of a multi-source fusion intelligent decision-making system for hospital intelligent procurement. Detailed Implementation

[0019] To further understand the content of this invention, a detailed description of the invention is provided in conjunction with the accompanying drawings and embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0020] This invention is specifically applied to the bidding and procurement management of hospital supplies such as medicines, consumables and medical equipment. It can be used in large general hospitals and medical groups to realize intelligent applications for procurement demand forecasting, supply chain risk management and procurement decision optimization.

[0021] Firstly, such as Figure 1 This embodiment provides a multi-source fusion intelligent decision-making method for smart procurement in hospitals, including: Step S1: Collect data from the hospital's internal systems and external supply chain, and perform standardized coding, entity disambiguation, spatiotemporal alignment, and semantic mapping on the heterogeneous data to generate a unified data representation.

[0022] Specifically, step S1 includes: collecting multi-source heterogeneous data from hospital HIS system, ERP system, inventory management system, procurement management system, supply chain management system, as well as external public resource trading platform, market price database, third-party testing platform, and online public opinion monitoring platform; uniformly encoding and mapping the material codes, supplier identifiers, and disease information from different data sources in the multi-source heterogeneous data; disambiguating cross-system homonymous entities based on text similarity calculation and attribute feature matching algorithms; and aligning cross-source data entities by combining timestamp consistency constraints, thereby generating a unique set of entity identifiers.

[0023] More specifically, the process begins with collecting heterogeneous data from multiple sources, including the hospital's HIS, ERP, inventory management, procurement management, and supply chain management systems, as well as external public resource trading platforms, market price databases, third-party testing platforms, and online public opinion monitoring platforms. Within this data, material codes may manifest as hospital-wide codes, national medical insurance codes, or product barcodes in different systems; supplier identifiers may exist as unified social credit codes, supplier names, or platform registration numbers; and disease information involves inconsistencies between clinical diagnosis names and ICD-10 codes. To address these coding differences, a mapping table is established to uniformly convert material codes from different sources into standard hospital-wide material codes, map various supplier identifiers to unified social credit codes, and map clinical disease names to ICD-10 codes. For newly added materials or suppliers not covered in the mapping table, a similarity matching algorithm based on string edit distance is used to automatically recommend candidate standard codes, which are then manually confirmed and added to the mapping table.

[0024] After completing the encoding mapping, disambiguation is performed on entities with the same name appearing across systems. The same supplier might be recorded as "Sinopharm Holding Co., Ltd.", "Sinopharm Holding", and "Sinopharm Shares" in the HIS system, procurement management system, and third-party inspection platform, respectively. The name text of each entity is extracted, and its cosine similarity based on the BERT model is calculated. Attributes such as registered address, legal representative, and contact number are also extracted. A weighted formula is used, for example, name similarity weight is 0.5, registered address Jaccard similarity weight is 0.2, and exact matching weights for legal representative and contact number are each 0.15, to calculate the comprehensive matching score. Entities with scores higher than 0.85 are automatically merged, scores between 0.60 and 0.85 are sent for manual review, and scores below 0.60 are determined to be different entities. The merged entities receive a globally unique entity identifier, which is used throughout the entire system's data processing and decision-making process.

[0025] Timestamp consistency constraints are used for cross-source data entity alignment. The order placement time recorded in the ERP system, the outbound time recorded in the logistics system, and the payment time recorded in the financial system for the same purchase order often differ by several seconds to several minutes. A time window of 5 seconds is set; multiple records within the window are considered the same event, and alignment is performed based on the earliest timestamp. Records exceeding the window trigger data quality alarms, prompting manual inspection. For GPS coordinates in supplier logistics data, reverse geocoding is used to obtain administrative divisions; points more than 1000 kilometers away from hospital addresses are considered anomalies and removed.

[0026] Semantic mapping resolves inconsistencies in field names and value ranges. Taking a purchase requisition as an example, the field name is "requisition quantity" in the HIS system and "order_qty" in the ERP system. Through a pre-built field mapping rule base (based on regular expressions matching the keywords "quantity" or "qty"), it is uniformly mapped to the standard field "requisition quantity". Regarding value ranges, the urgency level is "high / medium / low" in HIS and "1 / 2 / 3" in ERP, which is converted into a unified enumeration value through a configuration table. After the above standardization coding, entity disambiguation, spatiotemporal alignment, and semantic mapping, a unified data representation is generated. This representation includes a globally unique identifier for each entity, standardized field values, data source, and lineage information. For example, a supplier's unified data representation is: Entity ID SUP_20230615_A1B2C3, Standard Name "A Certain Holding Co., Ltd.", Unified Social Credit Code 9131000010000XXXXX, accompanied by performance data from the procurement management system (on-time delivery rate 0.982), quality pass rate from a third-party inspection platform (0.996), price index from a market price database (118.5), and negative labels from an online public opinion monitoring platform (none). This unified data representation provides consistent and reliable input for subsequent knowledge graph construction and intelligent decision-making.

[0027] Through the above methods, encoding mapping and entity disambiguation resolve the issue of inconsistent material, supplier, and disease identification across different systems, avoiding procurement plan mismatches or compliance risks caused by data inconsistencies. Spatiotemporal alignment eliminates process traceability obstacles caused by cross-system time deviations, enabling accurate reconstruction of the procurement lifecycle timeline. Semantic mapping ensures that business fields from systems such as HIS, ERP, and finance can be uniformly understood and calculated, laying a reliable foundation for demand forecasting and cost analysis. Through the above processing, the data acquisition stage achieves a transformation from "heterogeneous fragments" to "homogeneous assets," and subsequent spatiotemporal graph neural network prediction, supplier state modeling, and multi-objective optimization all directly rely on the integrity and accuracy of this unified data representation. Compared to existing preprocessing methods that only perform simple format conversion, the synergistic effect of standardized encoding, entity disambiguation, spatiotemporal alignment, and semantic mapping in this embodiment significantly improves the fusion quality of multi-source procurement data, fundamentally solving the technical problems of "data silos" and "lack of a unified knowledge base" mentioned in the background technology.

[0028] Step S2: Construct a hospital procurement knowledge graph based on the unified data representation, model the relationships between materials, diseases, clinical departments, suppliers, inventory nodes and policy constraints as a graph structure, and generate graph embedding vectors.

[0029] Step S2 specifically includes: based on unified data representation, performing entity recognition and classification on business database records from the hospital's internal system and external procurement text data. The business database records include structured field data from the HIS system, ERP system, and inventory management system. The structured field data includes at least material codes, supplier identifiers, inventory quantities, and transaction records. The external procurement text data includes tender documents, contract texts, and policy and regulatory texts. The structured field data is standardized and encoded based on field mapping rules, and the procurement text data is extracted into entities and relationships using a natural language processing model to generate entity-relationship triples. Based on the entity-relationship triples, a heterogeneous graph structure is constructed, including material nodes, disease nodes, clinical department nodes, supplier nodes, inventory nodes, and policy constraint nodes. Demand relationship edges, supply relationship edges, usage relationship edges, inventory flow relationship edges, and constraint relationship edges are constructed between different nodes. Based on the heterogeneous graph structure, a multi-relationship adjacency matrix and node feature vectors are constructed. A graph neural network is used to learn the representation of the heterogeneous graph structure to generate graph embedding vectors corresponding to each node.

[0030] More specifically, step S2 constructs a hospital procurement knowledge graph based on a unified data representation. Material codes, supplier identifiers, inventory quantities, and transaction records are extracted from structured field data from systems such as HIS, ERP, and inventory management. Standardized coding is then performed according to the field mapping rules established in step S1, ensuring that the material codes output by each system are unified as hospital standard codes and the supplier identifiers are unified as social credit codes. Simultaneously, external procurement text data, such as tender documents, contract texts, and policy and regulatory texts, are extracted using a natural language processing model for entity and relation extraction. Taking an antibiotic tender document as an example, the model extracts "cefotaxime sodium" (material entity), "Sinopharm Holdings" (supplier entity), "lowest price in volume-based procurement" (policy entity), and relationships such as "supply" and "constraint," generating entity-relationship triples.

[0031] The entity instances obtained from the above structured field transformation are merged with the triples obtained from text extraction to construct a heterogeneous graph structure. This graph contains six types of nodes: material nodes (e.g., "syringe," "cephalosporin antibiotics"), disease nodes (e.g., "pneumonia," "diabetes," identified by ICD-10 codes), clinical department nodes (e.g., "respiratory medicine"), supplier nodes (e.g., "Sinopharm Holdings"), inventory nodes (e.g., "central warehouse"), and policy constraint nodes (e.g., "centralized procurement catalog"). Five types of directed edges are constructed between different nodes based on business logic: demand relationship edges represent the historical requisition behavior of a department for a certain material; supply relationship edges represent the availability of a certain material by a supplier; usage relationship edges represent the common use of a certain material in treating a certain disease; inventory flow relationship edges represent the allocation path of materials between different warehouses; and constraint relationship edges represent the price limits or mandatory procurement rules of a certain policy that restrict the purchase of a certain material. Taking the requisition of nebulized inhalers by the Department of Respiratory Medicine as an example, a demand relationship edge is established between the "Department of Respiratory Medicine" node and the "Nebulized Inhaler" node, with an attached monthly average requisition quantity attribute; if the inhaler is within the scope of the "Centralized Procurement Catalog", a constraint relationship edge is established between the material node and the policy node, with an attached maximum price limit attribute.

[0032] Based on the aforementioned heterogeneous graph structure, a multi-relationship adjacency matrix is ​​constructed. Each type of relationship corresponds to a sub-adjacency matrix, and the matrix elements represent whether a relationship of that type exists between nodes and its weight (such as consumption frequency). Simultaneously, a feature vector is constructed for each node. The features of material nodes may include historical average purchase price and material classification code; the features of supplier nodes may include performance rating and on-time delivery rate; the features of department nodes may include annual total consumption; the features of disease nodes may include incidence rate or surgical volume; the features of inventory nodes may include safety stock threshold and current inventory level; and the features of policy nodes may include constraint type code. This multi-relationship adjacency matrix and the node feature vectors are input into a graph neural network for representation learning. The graph neural network updates the current node representation by aggregating information from neighboring nodes, with different weight transformation matrices corresponding to different relationship types. After multiple propagations, the embedding vector of each node integrates its own attributes, neighbor attributes, and multi-hop structure information. For example, the embedding vector of a supplier node not only reflects its own performance rating but also implicitly contains multiple semantics such as the types of materials it supplies, the associated clinical departments and diseases, and the policy constraints it is subject to. The generated node graph embedding vectors will serve as the initial node features of the demand-supply dynamic graph in step S3, enabling the spatiotemporal graph neural network to directly utilize the structural knowledge in the embedding vectors for more accurate demand prediction.

[0033] The aforementioned knowledge graph construction process mutually supports the functions of hospital procurement scenarios. Standardized encoding of structured field data resolves the inconsistency in material and supplier identification across multiple systems, ensuring that the same entity corresponds to only one node in the graph. Entity relationships extracted from tender documents and contract texts by the natural language processing model supplement constraints (such as policy price limits) and usage relationships (such as the association between materials and diseases) not explicitly recorded in the structured data. The heterogeneous graph structure unifies the dispersed entity types in procurement operations. Five types of edges—demand, supply, usage, circulation, and constraints—correspond to five dimensions in the procurement process: demand initiation, supply capacity, clinical adaptation, inventory management, and compliance supervision. This structure provides a complete topological foundation for subsequent risk transmission analysis and multi-objective optimization. The embedding vectors generated by the graph neural network fuse numerical attributes in structured data with semantic information in unstructured text into low-dimensional vectors, solving the problem of "lack of a unified knowledge base" and making cross-entity and cross-relationship reasoning computable in the vector space. Compared to existing procurement management systems based solely on relational databases, the knowledge graph and graph embedding method in this embodiment significantly improve the ability of procurement decisions to model complex business relationships.

[0034] Step S3: Construct a demand and supply dynamic graph based on the knowledge graph, and use a spatiotemporal graph neural network to model the demand and supply dynamic graph to predict the demand, inventory changes and supply capacity of each material in the future cycle, and generate a demand and supply matching degree matrix.

[0035] Specifically, step S3 includes: introducing a time dimension into the hospital procurement knowledge graph, dividing nodes into clinical demand nodes, material nodes, inventory nodes, and supplier nodes, and constructing demand consumption relationship edges, supply relationship edges, and inventory flow relationship edges. The weights of each edge are dynamically updated over time based on historical consumption frequency, supply fulfillment records, and inventory change rate to form a demand-supply graph structure that evolves over time; performing time slicing on the dynamic demand-supply graph to construct a dynamic graph sequence input, and generating node representation vectors by combining node structural features, time statistical features, and attribute features; using a graph neural network to model the spatial dependencies between nodes, and using a temporal attention mechanism to model the temporal dependencies of node states to predict the demand for each material, inventory changes, and supply capacity changes in the future period, and calculating a demand-supply matching degree matrix based on the prediction results to characterize the matching relationship between supply and demand.

[0036] More specifically, step S3, based on the hospital procurement knowledge graph constructed in step S2, introduces a time dimension to build a dynamic demand-supply graph. The nodes in the knowledge graph are divided into four categories: clinical demand nodes (corresponding to material requests from specific departments), material nodes, inventory nodes, and supplier nodes. Three types of dynamic relationship edges are retained: demand consumption relationship edges connect clinical demand nodes and material nodes, representing the department's historical consumption behavior of that material; supply relationship edges connect supplier nodes and material nodes, representing the supplier's supply capacity for the material; and inventory flow relationship edges connect material nodes and inventory nodes, representing the inbound, outbound, and transfer records of materials between different warehouses. The weights of each type of edge are dynamically updated over time: the weight of the demand consumption relationship edge is updated monthly to the average consumption of the past three months; the weight of the supply relationship edge is calculated based on the supplier's on-time performance and delivery integrity rates over the past six months; and the weight of the inventory flow relationship edge is dynamically adjusted based on the moving average of the inventory change rate. Taking the demand-consumption relationship between the respiratory medicine department and syringes as an example, the system records the monthly usage of the department over the past 12 months, and updates the weight monthly with the moving average of the most recent three months; when the usage increases during the flu season, the edge weight increases accordingly, reflecting the spatiotemporal fluctuations in demand.

[0037] The demand-supply dynamic graph is processed by time slicing, using weeks or months as the time granularity, dividing the continuous time axis graph sequence into discrete dynamic graph sequences. The graph structure within each time slice includes all nodes and their edge weights within that time period. When constructing the representation vector of each node, three types of features are integrated: node structural features (such as the graph embedding vector obtained from pre-training the neural network, reflecting the node's role and association in the static knowledge graph), time statistical features (such as the historical consumption mean, variance, and trend of the node over several time slices), and attribute features (such as the price classification code of materials, supplier performance rating, and inventory node safety threshold). These three features are concatenated to form the initial representation of the node in the dynamic graph sequence.

[0038] A spatiotemporal graph neural network is used to model the dynamic graph sequence. Spatially, the graph neural network aggregates the neighbor node information of each node, with different edge types corresponding to different aggregation weight matrices, outputting an updated spatial state vector. Temporally, a temporal attention mechanism is used to model the state sequence of each node along the time axis, with attention weights reflecting the importance of different historical time slices to the current prediction. Taking a certain material node as an example, the temporal attention mechanism can automatically assign higher attention weights to recent time slices (such as last month) while retaining lower weights for time slices from the same period last year, thus capturing seasonal demand patterns. After multi-layer propagation and temporal attention calculations by the spatiotemporal graph neural network, the system outputs predicted values ​​for the demand of each material in the future period, predicted values ​​for inventory changes at each inventory node, and predicted values ​​for supply capacity changes at each supplier node. For example, for cephalosporin antibiotics, the system predicts that the demand in respiratory medicine will increase by 15% in the next four weeks compared to last week, the central warehouse inventory will drop below the safety threshold, and supplier A's production capacity will decrease by 5% due to raw material shortages.

[0039] Based on the above forecast results, a demand-supply matching degree matrix is ​​calculated. The rows of this matrix correspond to demand nodes (clinical demand nodes or material nodes), and the columns correspond to supply nodes (supplier nodes or inventory nodes). The values ​​of the matrix elements represent the degree of matching between a demand node and a supply node. The matching degree calculation comprehensively considers three factors: the ratio of the demand forecast to the supply capacity forecast (supply-demand ratio), the availability coefficient after normalizing logistics distance or allocation costs, and the compliance coefficient under policy constraints. When the supply-demand ratio is close to 1, availability is high, and compliance is achieved, the matching degree is close to 1; when supply capacity is insufficient or there are policy restrictions, the matching degree approaches 0. This matrix is ​​output to step S6 as an allocation constraint for optimizing the procurement plan.

[0040] Through the above methods, the dynamic updating of edge weights enables the model to respond to fluctuations in demand caused by seasonal influenza and public health emergencies, solving the problem of "disconnect between demand forecasting and supply capacity assessment." The temporal attention mechanism automatically identifies periodic patterns (such as the high incidence of respiratory diseases in winter and spring) and sudden changes in historical data, significantly improving the adaptability of demand forecasting compared to time-series models with fixed weights. The spatiotemporal graph neural network jointly models supply relationships (supplier capacity, inventory reserves) and demand relationships (departmental consumption, disease structure) in a graph structure, ensuring that the demand-supply matching degree matrix not only reflects quantitative supply-demand balance but also contains multi-dimensional information such as spatial accessibility and policy compliance, providing rich constraints for multi-objective optimization. Compared to existing procurement planning methods based solely on historical consumption statistics, this method achieves spatiotemporally dynamically coupled forecasting and matching analysis, effectively overcoming the problems of data silos and information lag in procurement decisions.

[0041] Step S4: Integrate supplier performance data, quality inspection data, delivery timeliness data, market fluctuation data, and public opinion data from the unified data representation to construct a dynamic status vector for suppliers.

[0042] Step S4 specifically includes: acquiring multi-source supplier-related data from the hospital procurement management system, external public resource trading platform, third-party testing platform, and online public opinion monitoring platform. This related data includes: performance data formed from the supplier's historical performance records recorded by the hospital procurement management system; delivery time data formed from order placement time and order delivery timestamps; quality inspection result data generated by the third-party testing platform; market price index data and procurement transaction price fluctuation data collected by the public resource trading platform; and text-based public opinion data related to suppliers obtained by the online public opinion monitoring platform. After aligning the multi-source supplier-related data with time windows, statistical feature vectors are extracted from the performance data, quality inspection data, and delivery time data. Time-series modeling is performed on the market price index data and procurement transaction price fluctuation data to generate trend feature vectors, and semantic encoding is performed on the public opinion data to generate text feature vectors. The feature vectors of different modalities are input into a multimodal fusion network, weighted and fused through an attention mechanism to generate a unified supplier state representation vector. The state representation vector is then dynamically updated based on a time decay mechanism to form a dynamic supplier state vector that evolves over time.

[0043] More specifically, step S4 constructs a dynamic state vector for the supplier. Historical supplier performance records are obtained from the hospital's procurement management system, extracting on-time delivery rate, complete order delivery rate, and number of returns / exchanges as performance data. The average delivery cycle and its standard deviation are calculated from the order placement time and delivery timestamp recorded in the same system, forming delivery timeliness data. Quality inspection results for each batch of materials are obtained from a third-party inspection platform, with batch pass rate and fluctuations in key quality indicators (such as sterility test results and effective ingredient content) constituting quality inspection data. Market price index data and procurement transaction price fluctuation data are collected from the public resource trading platform to obtain the supplier's price deviation from the market average and price change trends. Textual public opinion data related to the supplier, such as news reports, administrative penalty announcements, and complaint information, are obtained from the online public opinion monitoring platform.

[0044] The aforementioned multi-source data undergoes time window alignment processing, using months as the basic unit to group data from different data sources within the same month into the same time window. For fulfillment data, the on-time delivery rate sequence for the most recent six months is extracted, and the mean, standard deviation, and trend slope are calculated to form a statistical feature vector. For example, a supplier's on-time delivery rates for the past six months are 0.94, 0.96, 0.95, 0.97, 0.98, and 0.98, with a mean of 0.96, a standard deviation of 0.015, and a positive trend slope, indicating a stable improvement in its fulfillment capabilities. For delivery timeliness data, the average delivery cycle and cycle fluctuation amplitude are calculated, and statistical features are extracted accordingly. For quality inspection data, the mean and standard deviation of batch pass rates, as well as the deviation of key quality indicators from the pass / fail boundaries, are calculated. For market price index data and procurement transaction price fluctuation data, time series models (such as exponential smoothing or autoregressive moving average) are used to extract trend and cycle components, generating a trend feature vector that reflects the long-term deviation direction and fluctuation intensity of the supplier's price relative to the market benchmark. For text-based public opinion data, a pre-trained language model (such as BERT) is used to semantically encode each text to obtain a semantic vector. Then, attention pooling is used to fuse the semantic vectors of multiple public opinion data into a single text feature vector. Negative public opinion (such as "penalty from the drug regulatory authority" and "supply interruption") will be amplified in the attention weights.

[0045] The five types of feature vectors mentioned above—performance statistics feature vector, delivery timeliness statistics feature vector, quality inspection statistics feature vector, price trend feature vector, and public opinion text feature vector—are input into a multimodal fusion network. This network uses an attention mechanism to calculate the weights of each modality feature vector. The weight value depends on the information entropy of that modality within the current time window and its relevance to the supplier's status. For example, when a supplier frequently experiences negative public opinion recently, the attention weight of the public opinion modality increases; when market prices fluctuate sharply, the weight of the price trend modality increases. The feature vectors of each modality are weighted and summed according to their attention weights to generate a unified supplier status representation vector. This vector is a 128-dimensional or 256-dimensional real-number vector, with each dimension comprehensively encoding the supplier's status across multiple dimensions, including performance, quality, timeliness, price, and reputation.

[0046] The system dynamically updates the aforementioned state representation vectors based on a time decay mechanism. For each new month, the system calculates the supplier's state representation vector for that month and performs an exponentially weighted moving average with the previous month's state vector. The decay factor λ ranges from 0.7 to 0.9, giving newer state data higher weights while preserving the smoothing effect of historical states. This update mechanism allows the supplier's dynamic state vector to reflect state declines caused by short-term unforeseen events (such as a batch of substandard products), while avoiding drastic state oscillations caused by occasional noise. Taking a supplier as an example, its state vector remains stable in normal months. However, when a month experiences severe delivery delays and negative online public opinion events, the delivery timeliness feature vector and the public opinion text feature vector change significantly. After attention fusion and time decay updates, the values ​​of relevant dimensions in the new state vector decrease. This decrease signal is captured by the multi-objective optimization model in step S6, reducing the supplier's allocation weight in the procurement plan or removing it from the feasible solution space.

[0047] The aforementioned technical means mutually support the functions of hospital procurement scenarios. The fusion of multi-source data solves the problem of "lagging supplier risk identification" in the background technology, unifying information across five dimensions: performance, quality, timeliness, price, and public opinion, avoiding the one-sidedness of relying on only a single dimension to evaluate suppliers. The attention mechanism enables the model to automatically focus on the most informative data modality at present. For example, when public opinion data reflects operational anomalies in a supplier, the system increases the weight of the public opinion modality and correspondingly lowers the supplier status score, achieving early warning of risks. The time decay mechanism ensures that the state vector reflects both long-term performance trends and is sensitive to recent abnormal events, allowing supplier evaluation results to be updated in real time as business changes occur. Compared to existing supplier management systems that only use static scoring or simple weighted averaging methods, the dynamic state vector in this embodiment provides time-varying, multi-dimensional, and quantifiable supply capacity constraints for multi-objective optimization of procurement schemes, significantly improving the scientific nature of decision-making and risk response capabilities.

[0048] Step S5: Construct a supply chain relationship graph, calculate the probability and vulnerability of node risk propagation, analyze the risk diffusion trend based on risk propagation entropy, predict cascading failure paths, and generate a set of alternative supply chain candidates.

[0049] Specifically, step S5 includes: constructing a multi-type heterogeneous supply chain association graph based on the equity relationships, control relationships, cooperation relationships, and logistics dependencies among suppliers in the external supply chain data, and assigning corresponding propagation weights to different relationship edges; calculating the basic risk level of each node based on the association graph, and calculating the risk propagation probability based on the weighted adjacency relationship between nodes to characterize the risk diffusion process in the supply chain network; constructing a risk propagation distribution model based on the node risk propagation probability, and calculating the risk propagation entropy and its time change rate to characterize the concentration and evolution trend of risk diffusion; when the node risk level, node vulnerability, or risk propagation entropy change rate exceeds a preset threshold, predicting potential cascading failure paths based on the graph propagation path simulation method, and generating a candidate set of alternative supply chains based on the neighborhood structure and connectivity of the failed nodes.

[0050] More specifically, based on the equity, control, cooperation, and logistics dependencies among suppliers in external supply chain data, multi-type heterogeneous supply chain relationship graphs are constructed. Taking equity relationships as an example, if a parent company holds more than 50% of the shares in a subsidiary, an equity control edge is established between the two nodes, with a propagation weight of 0.8; for shareholding relationships (shareholding ratio between 20% and 50%), the propagation weight is 0.4. Control relationships refer to the relationship edges established between sister companies under the same actual controller, with a propagation weight of 0.6. Cooperation relationships are constructed based on historical joint bidding or joint supply records; if two suppliers jointly win bids in more than three projects, a cooperation edge is established, with a propagation weight of 0.3. Logistics dependencies refer to the establishment of a logistics dependency edge when supplier A purchases core raw materials or semi-finished products from supplier B, with a propagation weight of 0.7. Taking a medical consumables supply chain as an example, if supplier X holds 60% of the shares in supplier Y, a high-weight equity control edge is established between X and Y; if Y purchases sterilization packaging materials from supplier Z, a logistics dependency edge is established between Y and Z. The aforementioned edge weights are used to quantify the strength of risk propagation along this relationship.

[0051] The basic risk level of each node is calculated based on this association graph. The basic risk level consists of two parts: the node's inherent risk (such as debt-to-equity ratio, current ratio, and quality inspection failure rate, administrative penalty records in financial indicators) and the node's betweenness centrality in the association graph. Nodes with higher inherent risk have a higher basic risk level; nodes located at network hubs have a higher basic risk level because their risk is more easily propagated to more neighbors. For each node, its initial basic risk level is set to a real number between 0 and 1. For example, a supplier with a recent batch of substandard products has a basic risk level of 0.6.

[0052] The risk propagation probability is calculated based on the weighted adjacency relationships between nodes. Starting from any risk source node, the probability of risk propagating along an edge to neighboring nodes is equal to the risk level of the source node multiplied by the propagation weight of that edge. Propagation is iterated repeatedly, and the risk level of the target node is updated after each propagation. The new risk level is the maximum (or weighted sum) of its original risk level and the incoming risk level. The propagation process continues until the change in the risk level of all nodes is less than a preset threshold. This calculation characterizes the risk diffusion process in the supply chain network. An operational anomaly of a single supplier may be transmitted step-by-step through equity, cooperation, or logistical dependencies, ultimately affecting multiple indirect suppliers.

[0053] A risk propagation distribution model is constructed based on the node risk propagation probability. For each node, its final risk level is taken as its value in the risk propagation distribution. The risk propagation distribution of the entire network is the set of risk levels of all nodes. The Shannon entropy of this distribution is calculated to obtain the risk propagation entropy, which reflects the degree of risk concentration in the network: a low entropy value indicates that the risk is concentrated in a few nodes, while a high entropy value indicates that the risk has spread evenly to most nodes. Simultaneously, the time change rate of entropy is calculated by comparing the difference in entropy values ​​between adjacent time windows (e.g., last month and this month) and dividing by the time interval. A positive entropy change rate indicates that the risk is spreading further, while a negative rate indicates that the risk is converging.

[0054] The system is considered to be in a high-risk state when the risk level of any node exceeds the first threshold (e.g., 0.7), the node vulnerability (i.e., the probability that the node can still maintain normal operation after its own risk is removed) is below the second threshold (e.g., 0.4), or the risk propagation entropy change rate exceeds the third threshold (e.g., a positive value and greater than 0.1). At this time, a graph propagation path simulation method is used to predict potential cascading failure paths: starting from the high-risk node, the risk diffusion direction is simulated along the edge with the highest propagation weight, and the sequence of nodes passed through is recorded, forming one or more failure paths. Taking supplier X as the risk source, its risk propagates along the equity control edge to Y, and then propagates along the logistics dependence edge to Z; the failure path is X→Y→Z. For the end nodes or key hub nodes on the failure path, a candidate set of alternative supply chains is generated based on their neighborhood structure (i.e., other suppliers directly connected to them) and connectivity (i.e., the existence of alternative paths) in the association graph. For example, when Z faces a supply disruption due to the risk transmission from Y, the system searches for other suppliers W and U that are in the same material category as Z and have a direct supply relationship with the hospital. Simultaneously, it checks whether W and U are affected by the same risk source (equity penetration, cooperative relationship). After excluding suppliers with similar risks, the remaining suppliers are added to the alternative candidate set. This set is output to step S6 as a constraint on the feasible solution space in multi-objective optimization.

[0055] The aforementioned technical means mutually support the functions of hospital procurement scenarios. The multi-type edge weights of equity, control, cooperation, and logistics dependencies reflect the true path and intensity of risk propagation in the hospital supply chain, enabling risk analysis to move beyond isolated assessments of single suppliers. Iterative calculations of risk propagation probability simulate the gradual process of cascading failures, solving the problems of "lagging supplier risk identification and lack of full-chain transmission analysis." Risk propagation entropy and its rate of change provide quantitative indicators for judging risk diffusion trends, allowing for dynamic setting of warning thresholds. When a threshold is triggered, the system automatically generates an alternative candidate set, avoiding the predicament of passively switching suppliers after a risk outbreak, as is common in traditional methods. Compared to existing supplier management systems that rely solely on static blacklists and whitelists, this embodiment's supply chain risk analysis integrates proactive warning, path prediction, and alternative recommendation, providing structured input for flexible procurement design.

[0056] Step S6: With procurement cost, quality stability, supply continuity and risk propagation intensity as objectives, and combined with the demand-supply matching degree matrix, supplier dynamic state vector, cascading failure path and alternative supply chain candidate set as constraints, a multi-objective optimization algorithm is used to solve for the Pareto optimal procurement scheme.

[0057] Step S6 specifically includes: constructing an optimization model with the allocation relationship between suppliers and demand nodes as decision variables, wherein the objective functions include minimizing procurement costs, maximizing quality stability, maximizing supply continuity, and minimizing risk propagation intensity; using the demand-supply matching degree matrix as allocation constraints, the supplier dynamic state vector as supply capacity constraints, and cascading failure paths as non-selectable constraints, while limiting the alternative supply chain candidate set to the feasible solution space; solving the optimization model using a multi-objective evolutionary algorithm based on non-dominated sorting to obtain a Pareto optimal solution set, wherein each Pareto optimal solution corresponds to a set of structured procurement resource allocation schemes, the procurement resource allocation schemes include: the material allocation ratio or procurement quantity corresponding to each supplier, the allocation path relationship of each material among different suppliers, the risk level identifier of the corresponding supplier, the supply chain alternative path selection results, and the corresponding scheme's cost indicators, quality indicators, supply continuity indicators, and risk propagation indicators; selecting procurement resource allocation schemes that meet the constraints from the Pareto optimal solution set.

[0058] More specifically, for each material, the decision variable represents the quantity or proportion purchased from each supplier. Four objective functions are defined: minimizing procurement costs (minimizing the sum of the quantities of all purchased items multiplied by their corresponding supplier quotes); maximizing quality stability (maximizing the quantity after weighted allocation based on the quality inspection dimension scores in the supplier dynamic state vector); maximizing supply continuity (maximizing the quantity after weighted allocation based on the fulfillment and delivery timeliness dimension scores in the supplier dynamic state vector); and minimizing risk propagation intensity (minimizing the risk level of selected suppliers after weighted summation based on the risk propagation probabilities of each node in the supply chain diagram).

[0059] The constraints include the following: The demand-supply matching matrix serves as an allocation constraint. For each material, the total procurement amount allocated to each supplier must match the demand predicted in step S3, with an allowable deviation within a preset elasticity range (e.g., ±5%). The supplier dynamic state vector serves as a supply capacity constraint. The allocated procurement amount for each supplier must not exceed its supply capacity limit, which is determined by the capacity dimension in the supplier dynamic state vector combined with the historical maximum supply volume. Cascading failure paths serve as a non-selectable constraint. Risk nodes on the cascading failure paths predicted in step S5 are marked as unavailable, and their corresponding decision variables are forcibly set to zero. An alternative supply chain candidate set limits the feasible solution space. Suppliers used in the procurement plan must belong to this candidate set or meet the alternative screening conditions.

[0060] A multi-objective evolutionary algorithm based on non-dominated sorting (such as NSGA-II) is used to solve the above optimization model. The algorithm initializes a population containing several individuals (each corresponding to a set of allocation schemes). Each individual's chromosome is encoded as a real-number vector with a length equal to the total number of supplier-material pairings. The population is then non-dominatedly sorted, dividing individuals into different levels of Pareto fronts. The crowding distance for each individual is calculated to maintain solution diversity. Offspring are generated through tournament selection, simulated binary crossover, and polynomial mutation. Parents and offspring are merged, and non-dominated sorting and crowding distance calculations are performed again to select the next generation. This process is repeated iteratively until the preset number of generations or convergence criteria are reached.

[0061] After solving, a Pareto optimal solution set is obtained, where each solution is not completely dominated by other solutions in the four objective functions. Each Pareto optimal solution corresponds to a set of structured procurement resource allocation schemes. Taking the procurement of antibiotics by a hospital as an example, the Pareto optimal solution set contains three schemes: Scheme A has the lowest procurement cost (1 million yuan), but its quality stability score is 0.92, supply continuity score is 0.85, and risk propagation intensity is 0.4; Scheme B has the highest quality stability (0.98), but the cost rises to 1.2 million yuan, with a supply continuity score of 0.90 and a risk propagation intensity of 0.3; Scheme C achieves a balance among the four objectives, with a cost of 1.1 million yuan, a quality stability score of 0.95, a supply continuity score of 0.92, and a risk propagation intensity of 0.2. Each plan also includes the material allocation ratio for each supplier (e.g., supplier A bears 60% and supplier B bears 40%), the allocation path (direct delivery from the central warehouse or transshipment through regional distribution centers), supplier risk level indicators (high risk, medium risk, low risk), and recommended supply chain alternative paths (automatically switching to alternative suppliers when the primary supplier fails). The procurement resource allocation plan that meets the hospital's actual constraints (such as budget limits and minimum quality requirements) is selected from the Pareto optimal solution set and output as the final execution plan to the procurement management system.

[0062] This paper adopts four objective functions corresponding to the core performance indicators in hospital procurement management—economy, quality and safety, supply resilience, and risk control—transforming the problem of "procurement decisions relying on a single objective or subjective experience" in existing technologies into a quantifiable multi-objective optimization problem. The constraints organically combine the demand-supply matching degree in step S3, the supplier dynamic state vector in step S4, and the cascading failure paths and alternative candidate set in step S5, allowing the optimization model to simultaneously incorporate demand forecasting, real-time supplier status, and risk analysis results. The output method of a Pareto optimal solution set rather than a single solution respects the hospital's autonomy in adjusting the weights of different objectives in actual decision-making (e.g., prioritizing supply continuity over cost during an influenza outbreak). Compared to existing procurement systems that use fixed weighted summation or experience-based decision-making, the multi-objective optimization method in this embodiment can automatically discover multiple non-dominated solutions, providing procurement decision-makers with quantifiable comparative evidence and significantly improving the scientific rigor and interpretability of procurement plans.

[0063] Furthermore, after step S6, the process further includes: collecting execution feedback data, constructing prediction error feedback, and performing online incremental updates on the spatiotemporal graph neural network, supplier dynamic state vector, and multi-objective optimization algorithm to achieve adaptive closed-loop optimization; wherein, the execution feedback data includes actual execution data from the hospital procurement management system and inventory management system, and the actual execution data includes purchase order execution records, actual material consumption data, inventory change data, and supplier fulfillment data; the prediction error feedback is constructed based on the deviation between the execution feedback data and the prediction output result of step S3, including demand prediction error, inventory change error, and supply capacity error.

[0064] Specifically, after step S6, the system collects execution feedback data and constructs prediction error feedback. It then performs online incremental updates to the spatiotemporal graph neural network, the supplier dynamic state vector generation model, and the multi-objective optimization algorithm, forming an adaptive closed-loop optimization. The execution feedback data comes from the hospital's procurement management system and inventory management system, including purchase order execution records (actual order quantity, actual purchase price, order completion status), actual material consumption data (requisition quantity by each department, consumption time distribution), inventory change data (inbound quantity, outbound quantity, inventory discrepancies), and supplier fulfillment data (actual delivery quantity, delivery time, quality acceptance results). This data is automatically collected after the procurement plan is executed and summarized by time window (e.g., weekly or monthly).

[0065] The prediction error feedback is constructed based on the deviation between the execution feedback data and the prediction output of step S3. Step S3 outputs the predicted values ​​of demand, inventory changes, and supply capacity for each material in the future period, while the execution feedback data provides the actual values ​​for the corresponding period. The error feedback specifically includes three components: demand prediction error, which is the absolute or relative deviation between actual consumption and predicted demand. For example, if the system predicts that the respiratory medicine department needs 1200 vials of cephalosporin antibiotics in a certain month, but actually consumes 1350 vials, a positive error of +150 vials occurs. Inventory change error, which is the deviation between actual inventory changes and predicted inventory changes, reflecting the inaccuracy of the inventory turnover model. Supply capacity error, which is the deviation between the actual supply available from suppliers and the predicted supply capacity. For example, if the system predicts that supplier A can stably supply 800 vials, but only delivers 600 vials due to capacity issues, a negative error of -200 vials occurs. Each error component is accompanied by a timestamp, material identifier, and supplier identifier to facilitate locating the specific source of model deviation.

[0066] For the spatiotemporal graph neural network, an online incremental learning approach is used for updates. The prediction error feedback is used as the increment of the loss function, and gradient descent is applied to update the network parameters. Each iteration uses only the error data of the current period, eliminating the need to retrain on the entire historical dataset. Specifically, the original network parameters are retained as a foundation, and the gradient is calculated using backpropagation of the new error. The network weights are adjusted with a small learning rate (e.g., 0.001) to gradually adapt the model to changes in demand patterns (e.g., peak seasons for seasonal diseases). For the supplier dynamic state vector generation model, supplier fulfillment data (actual on-time delivery rate, actual quality pass rate) from the execution feedback data is used to correct the state vector. The actual fulfillment indicators are compared with the corresponding dimensions in the supplier state vector predicted in step S4. If a systematic deviation is found (e.g., actual delivery delays are higher than predicted values ​​for three consecutive months), the attention weights of the multimodal fusion network are adjusted or the decay factor in the time decay mechanism is updated to make the supplier state assessment more consistent with reality. For the multi-objective optimization algorithm, an adaptive parameter adjustment strategy is adopted. Based on the prediction error feedback analysis of the prediction accuracy of each objective function, if the prediction error of a certain objective (such as supply continuity) is consistently high, the weight of the crowding distance calculation in the non-dominated ranking of that objective will be adjusted accordingly, or the relaxation of the constraint conditions will be adjusted, so that the optimization algorithm can handle high error objectives more robustly in subsequent solutions.

[0067] The aforementioned closed-loop update mechanism enables the system's adaptive evolution. Taking a hospital's peak respiratory disease season as an example, the initial demand forecasting model, trained on three years of historical data, underestimated the surge in demand caused by a sudden flu outbreak, resulting in a significant positive demand forecasting error. After the system collected this error, the spatiotemporal graph neural network was updated online, enhancing the time attention weight of recent data and improving the accuracy of subsequent predictions. Simultaneously, supplier A's supply capacity forecast suffered a continuous negative deviation due to logistical disruptions. The supplier state vector generation model accordingly lowered its state score, and the multi-objective optimization algorithm automatically reduced its allocated quota and increased the weight of alternative suppliers in subsequent procurement plans. After several cycles of online updates, the entire model converged to a new steady-state, with the prediction error controlled within an acceptable range.

[0068] The aforementioned prediction error feedback quantifies the deviation between the actual execution result and the predicted output, providing a supervisory signal for model updates and solving the problems of missing closed-loop feedback mechanisms and the inability of the system to adaptively iteratively optimize in existing technologies. Online incremental updates avoid the high computational cost of full retraining, enabling the model to quickly respond to changes in the procurement environment (such as fluctuations in supplier capacity and changes in clinical demand patterns). The supplier dynamic state vector generation model adjusts its parameters based on actual performance feedback, transforming the supplier profile from static evaluation to dynamic evolution. The multi-objective optimization algorithm adjusts the optimization strategy based on prediction errors, improving the robustness of the decision-making scheme. Compared to the shortcomings of existing procurement systems where models are fixed and unable to adapt, the closed-loop optimization mechanism in this embodiment achieves a complete closed loop from data collection, prediction, decision-making to execution feedback and model updates. The system continuously iterates and evolves with business operations, maintaining prediction accuracy and decision quality over the long term.

[0069] like Figure 2The diagram shown in this embodiment illustrates the before-and-after effects of a multi-source fusion intelligent decision-making method for smart hospital procurement. A tertiary-level hospital achieved significant technical results after implementing this system. According to semi-annual statistics, the accuracy of demand forecasting increased from 64% to 92%, an increase of 28 percentage points; the number of emergency purchases decreased from 15 to 4, a reduction of 73%; the number of supplier risk events decreased from 8 to 2, a reduction of 75%; the inventory backlog rate decreased from 32% to 19%, a reduction of 40%; the supplier risk warning lead time increased from 0 days (post-event discovery) to 7 days, achieving pre-event warning; the alternative supplier response time was shortened from 48 hours to 2 hours, improving response efficiency by 96%; the relative value of procurement costs decreased from 100% to 92%, saving 8%; the supply chain continuity score increased from 72 to 90, an improvement of 25%; and the deviation between procurement plans and execution decreased from 12% to 3%, a reduction of 75%. The above data shows that the present invention has achieved substantial breakthroughs in terms of demand forecast accuracy, risk warning capability, supply chain response speed, procurement cost control and execution reliability, and effectively solved technical problems such as forecasting disconnect, risk lag and subjective decision-making in hospital procurement management.

[0070] Secondly, such as Figure 3As shown, this embodiment provides a hospital intelligent procurement multi-source fusion intelligent decision-making system, including: a data acquisition and processing module, used to collect data from the hospital's internal system and external supply chain data, and to perform standardized encoding, entity disambiguation, spatiotemporal alignment, and semantic mapping processing on heterogeneous data to generate a unified data representation; a knowledge graph construction module, used to construct a hospital procurement knowledge graph based on the unified data representation, modeling the relationships between materials, diseases, clinical departments, suppliers, inventory nodes, and policy constraints as a graph structure, and generating graph embedding vectors; a demand and supply forecasting module, used to construct a demand and supply dynamic graph based on the knowledge graph, and to model the demand and supply dynamic graph using a spatiotemporal graph neural network to predict the demand, inventory changes, and supply capacity of each material in the future period, and generate a demand and supply matching degree matrix; and a supplier status modeling module, used to integrate supplier performance data from the unified data representation. The system employs a multi-objective optimization decision-making module. This module uses quality inspection data, delivery timeliness data, market fluctuation data, and public opinion data to construct a supplier dynamic state vector. A supply chain risk analysis module constructs a supply chain relationship graph, calculates the probability of risk propagation at each node and the node's vulnerability, analyzes risk diffusion trends based on risk propagation entropy, predicts cascading failure paths, and generates a set of alternative supply chain candidates. A multi-objective optimization decision-making module uses procurement cost, quality stability, supply continuity, and risk propagation intensity as objective functions, combined with the demand-supply matching degree matrix, supplier dynamic state vector, cascading failure paths, and the set of alternative supply chain candidates as constraints, to solve for the Pareto optimal procurement plan using a multi-objective optimization algorithm. An adaptive closed-loop update module collects execution feedback data, constructs prediction error feedback, and performs online incremental updates to the spatiotemporal graph neural network, supplier dynamic state vector, and multi-objective optimization algorithm to achieve adaptive closed-loop optimization of the system.

[0071] Specifically, the system is deployed in the hospital's data center and includes several high-performance servers, storage arrays, and network communication equipment. The data acquisition and processing module runs on a data acquisition server configured with multiple network interfaces. This server establishes data connections with the HIS system server, ERP system server, inventory management system server, procurement management system server, and supply chain management system server through the hospital's internal LAN. Simultaneously, it interfaces with external public resource trading platforms, market price database servers, third-party testing platform interface machines, and online public opinion monitoring platforms through firewalls and dedicated lines. The data acquisition server periodically extracts data from the aforementioned internal and external data sources. Each type of data source is configured with an independent acquisition thread, and the acquisition frequency is set according to the data update characteristics. For example, clinical consumption data from the HIS system is synchronized hourly, while external market price data is updated daily. The collected multi-source heterogeneous data is temporarily stored in the local cache of the data acquisition server and then handed over to the module's data processing unit for standardized encoding, entity disambiguation, spatiotemporal alignment, and semantic mapping processing. The data processing unit first identifies the format differences in material codes across various data sources, then uses a pre-installed encoding mapping table on the server to uniformly convert hospital codes, national medical insurance codes, and product barcodes into standard hospital material codes. For supplier identifiers, the unified social credit code is used as the primary key for merging; if the credit code is missing, a unique identifier is generated using a combination of the supplier name and registered address. Disease information is standardized using an ICD-10 encoding mapping table. The entity disambiguation unit uses the BERT model to calculate the vector similarity of entity names, combining weighted matching of attributes such as registered address and legal representative, to merge cross-system entities with the same name but different entities and assign globally unique entity identifiers. The spatiotemporal alignment unit uniformly converts the timestamps from various data sources to ISO 8601 format, sets a 5-second time window to align cross-system records of the same event, and triggers an alarm for records exceeding the window. The semantic mapping unit maps source system fields to standard fields uniformly based on a pre-built field mapping rule base and value domain conversion table. After processing, a unified data representation is generated and stored in the result database of the data acquisition server in the form of a structured data table. Each record contains a globally unique entity identifier, standardized field values, data source identifier, and timestamp.

[0072] The knowledge graph construction module runs on a graph database server equipped with large-capacity memory and solid-state drives to store the Neo4j graph database. The graph database server reads a unified data representation from the result database of the data acquisition server over the network, while simultaneously loading tender documents, contract texts, and policy and regulatory text files. This module first extracts six types of entity instances from the structured field data: materials, diseases, clinical departments, suppliers, inventory nodes, and policy constraints, ensuring that entities are not duplicated based on globally unique entity identifiers. For external procurement text data, it calls a natural language processing model (a named entity recognition and relation extraction model based on pre-trained BERT) deployed on the same server to parse the data and extract entity relation triples. After fusing structured entities with text-extracted triples, nodes and edges are created in the graph database: Material nodes include attributes such as material code, specifications, and unit price; disease nodes include ICD-10 code and name; clinical department nodes include department code and annual total consumption; supplier nodes include unified credit code and performance score; inventory nodes include inventory quantity and safety threshold; and policy constraint nodes include constraint type and effective date. Demand consumption edges (with average monthly usage), supply edges (with supply ratio), usage edges, inventory circulation edges (with allocation cycle), and constraint edges (with price limit parameters) are established between nodes. After construction, this module calls a graph neural network model (R-GCN) to perform representation learning on the graph, generating a 128-dimensional embedding vector for each node. The correspondence between the embedding vector and the node ID is stored in the graph database index.

[0073] The demand and supply forecasting module runs on a computing server equipped with a graphics processing unit (GPU). This server reads knowledge graph data and node embedding vectors from a graph database server and connects to the real-time inventory data stream of the inventory management system. This module introduces a time dimension into the knowledge graph, constructing a dynamic demand and supply graph: nodes are divided into clinical demand nodes (corresponding departments), material nodes, inventory nodes, and supplier nodes. Edge weights are dynamically updated according to time windows. The weight of clinical demand edges is based on the moving average consumption of the past three months, and the weight of supply edges is based on the moving average of on-time fulfillment rates of the past six months. The module slices the dynamic graph over time (on a weekly basis), generating a comprehensive representation vector for each node that combines graph embedding vectors, temporal statistical features (mean and slope of historical consumption), and attribute features (safety threshold, classification coding). The spatiotemporal graph neural network uses graph convolutional layers to capture the spatial dependencies between nodes and a temporal attention layer (based on a Transformer architecture) to capture the evolution of node states along the time axis, outputting the predicted demand for each material, the predicted inventory change for each inventory node, and the predicted supply capacity for each supplier over the next four weeks. Based on the prediction results, a demand-supply matching degree matrix is ​​calculated. Each element in the matrix represents the degree of matching between a specific demand node and a supply node. This value is obtained by combining the supply-demand ratio, logistics distance coefficient, and compliance coefficient. The matrix data is stored in the in-memory database of the computing server for subsequent modules to access.

[0074] The supplier status modeling module runs on the same computing server. This module reads supplier performance data, quality inspection data, delivery timeliness data, market fluctuation data, and public opinion data from the unified data representation database of the data acquisition server. Performance data includes on-time delivery rate and complete delivery rate of orders recorded in the hospital procurement management system; quality inspection data is pushed to the system by a third-party inspection platform via API; delivery timeliness data is extracted from the logistics management system to calculate the average delivery cycle by combining the order time and delivery timestamp; market fluctuation data is collected from the public resource trading platform to obtain price index and quotation deviation; and public opinion data is obtained from the RESTful interface of the network public opinion monitoring platform to obtain text information. The module aligns the above data with a time window (monthly), extracts statistical feature vectors, trend feature vectors, and text semantic vectors, and inputs the five types of features into a multimodal fusion network. This network uses an attention mechanism to calculate the weights of each modality, and the weighted sum generates a supplier status representation vector (128 dimensions). This vector is then subjected to an exponentially weighted moving average with the previous month's status vector using a time decay mechanism (decay factor 0.8) to form a dynamic status vector that evolves over time and is stored in the vector database of the computing server.

[0075] The supply chain risk analysis module runs on the same computing server. This module retrieves equity relationships, control relationships, cooperative relationships, and logistics dependencies among suppliers from an external supply chain database, constructing a multi-type heterogeneous supply chain relationship graph. Equity relationship data is obtained from a corporate credit information query interface; control relationships are obtained through a thorough analysis of the actual controller; cooperative relationships are extracted from historical joint bidding records; and logistics dependencies are identified from supplier purchase orders. The edge weights of the relationship graph are set according to the relationship type (equity control edge 0.8, equity participation edge 0.4, control edge 0.6, cooperative edge 0.3, logistics dependency edge 0.7). The module calculates the basic risk level of each node (comprehensive financial indicators, quality non-compliance rate, and centrality), and iteratively calculates the risk propagation probability based on weighted adjacency relationships until convergence. A risk propagation distribution is constructed based on the node risk probability, and Shannon entropy and its time change rate are calculated. When a node's risk level exceeds 0.7, its vulnerability is below 0.4, or its entropy change rate exceeds 0.1, an early warning is triggered, and a cascading failure path is simulated, spreading step-by-step from the high-risk node along the edge with the highest propagation weight, recording the sequence of failed nodes. Based on the neighborhood structure of the failed node, other suppliers with the same material supply capabilities and no risk association with the failed node are selected to generate a candidate set of alternative supply chains. This set is stored in the computing server as a list.

[0076] The multi-objective optimization decision-making module runs on the same computing server. This module receives the demand-supply matching degree matrix from the demand-supply forecasting module, the dynamic state vector from the supplier state modeling module, and the cascading failure paths and alternative candidate set from the supply chain risk analysis module. The module constructs an optimization model with the allocation relationship between suppliers and demand nodes as the decision variables. The objective functions include minimizing procurement costs, maximizing quality stability, maximizing supply continuity, and minimizing risk propagation intensity. Constraints include: the total allocation must match the demand in the matching degree matrix (allowing ±5% deviation); the allocation amount of each supplier cannot exceed the upper limit of the supply capacity implied by its dynamic state vector; suppliers on cascading failure paths are forcibly excluded; and the feasible solution space is limited to the alternative candidate set. The module runs a multi-objective evolutionary algorithm based on non-dominated sorting (NSGA-II), initializing the population size to 200, iterating for 100 generations, and finally outputting a Pareto optimal solution set. Each solution includes the material allocation ratio, allocation path, risk level, and four indicators: cost, quality, continuity, and risk for each supplier. System administrators can select solutions from the solution set that meet the hospital's actual constraints (such as budget limits and minimum quality requirements) and automatically generate purchase orders by calling the API interface of the procurement management system.

[0077] The adaptive closed-loop update module runs on the computing server. This module periodically (once a month) collects execution feedback data from the hospital's procurement management system and inventory management system, including purchase order execution records, actual material consumption data, inventory change data, and supplier fulfillment data. The module compares the execution feedback data with the prediction outputs from step S3 (demand forecast, inventory change forecast, and supply capacity forecast), calculating the demand forecast error, inventory change error, and supply capacity error. Based on these error values, the module performs online incremental updates to the parameters of the spatiotemporal graph neural network, adjusting network weights using mini-batch gradient descent (learning rate 0.001); it corrects the attention weights and time decay factors in the supplier dynamic state vector generation model to minimize the deviation between state estimation and actual fulfillment results; and it adaptively adjusts the objective function weights and constraint slackness in the multi-objective optimization algorithm to better reflect actual execution conditions. The updated model parameters are persisted to the server disk for use in the next cycle's prediction and decision-making, thus achieving continuous iterative adaptive closed-loop optimization as the system operates.

[0078] The above description is merely a specific embodiment of this specification. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the scope of protection of this specification is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this specification, and these modifications or substitutions should all be covered within the scope of protection of this specification.

Claims

1. A multi-source fusion intelligent decision-making method for smart hospital procurement, characterized in that, include: Step S1: Collect data from the hospital's internal systems and external supply chain, and perform standardized coding, entity disambiguation, spatiotemporal alignment, and semantic mapping on the heterogeneous data to generate a unified data representation; Step S2: Construct a hospital procurement knowledge graph based on the unified data representation, model the relationships between materials, diseases, clinical departments, suppliers, inventory nodes and policy constraints as a graph structure, and generate graph embedding vectors; Step S3: Construct a demand and supply dynamic graph based on the knowledge graph, and use a spatiotemporal graph neural network to model the demand and supply dynamic graph to predict the demand, inventory changes and supply capacity of each material in the future cycle, and generate a demand and supply matching degree matrix. Step S4: Integrate supplier performance data, quality inspection data, delivery timeliness data, market fluctuation data, and public opinion data from the unified data representation to construct a dynamic status vector for suppliers; Step S5: Construct a supply chain relationship graph, calculate the probability and vulnerability of risk propagation at nodes, analyze the risk diffusion trend based on risk propagation entropy, predict cascading failure paths, and generate a set of alternative supply chain candidates. Step S6: With procurement cost, quality stability, supply continuity and risk propagation intensity as objectives, and combined with the demand-supply matching degree matrix, supplier dynamic state vector, cascading failure path and alternative supply chain candidate set as constraints, a multi-objective optimization algorithm is used to solve for the Pareto optimal procurement scheme.

2. The intelligent decision-making method for hospital procurement through multi-source fusion as described in claim 1, characterized in that, Step S1 specifically includes: collecting multi-source heterogeneous data from hospital HIS system, ERP system, inventory management system, procurement management system, supply chain management system, as well as external public resource trading platform, market price database, third-party testing platform, and online public opinion monitoring platform; uniformly encoding and mapping the material codes, supplier identifiers, and disease information from different data sources in the multi-source heterogeneous data; and disambiguating cross-system homonymous entities based on text similarity calculation and attribute feature matching algorithms. Combined with timestamp consistency constraints, cross-source data entity alignment is achieved, thereby generating a unique entity identifier set.

3. The intelligent decision-making method for hospital procurement through multi-source fusion as described in claim 1, characterized in that, Step S2 specifically includes: Based on a unified data representation, entity identification and classification are performed on business database records from the hospital's internal systems and external procurement text data. The business database records include structured field data from the HIS system, ERP system, and inventory management system. The structured field data includes at least material codes, supplier identifiers, inventory quantities, and transaction records. The external procurement text data includes tender documents, contract texts, and policy and regulatory texts. The structured field data is standardized and encoded based on field mapping rules, and the procurement text data is extracted into entities and relationships using a natural language processing model to generate entity-relation triples. Based on the entity relationship triples, a heterogeneous graph structure is constructed, which includes material nodes, disease nodes, clinical department nodes, supplier nodes, inventory nodes, and policy constraint nodes. Demand relationship edges, supply relationship edges, usage relationship edges, inventory flow relationship edges, and constraint relationship edges are constructed between different nodes. Based on the heterogeneous graph structure, a multi-relationship adjacency matrix and node feature vectors are constructed. A graph neural network is then used to perform representation learning on the heterogeneous graph structure to generate graph embedding vectors corresponding to each node.

4. The intelligent decision-making method for hospital procurement through multi-source fusion as described in claim 1, characterized in that, In step S3, the construction of the demand-supply dynamic diagram includes: Based on the hospital procurement knowledge graph, a time dimension is introduced, dividing the nodes into clinical demand nodes, material nodes, inventory nodes, and supplier nodes. Demand consumption relationship edges, supply relationship edges, and inventory flow relationship edges are constructed, with the weight of each edge dynamically updated over time based on historical consumption frequency, supply fulfillment records, and inventory change rate, to form a demand and supply graph structure that evolves over time.

5. The intelligent decision-making method for hospital procurement through multi-source fusion as described in claim 1, characterized in that, In step S3, a spatiotemporal graph neural network is used to model the dynamic demand-supply graph to predict the demand, inventory changes, and supply capacity of various materials in future periods, generating a demand-supply matching degree matrix. Specifically, this includes: The demand-supply dynamic graph is processed by time slicing to construct a dynamic graph sequence input, and node representation vectors are generated by combining node structural features, time statistical features, and attribute features. A graph neural network is used to model the spatial dependencies between nodes, and a temporal attention mechanism is used to model the temporal dependencies of node states in order to predict the demand, inventory changes, and supply capacity changes of each material in the future cycle. Based on the prediction results, a demand-supply matching degree matrix is ​​calculated to represent the matching relationship between supply and demand.

6. The intelligent decision-making method for multi-source fusion in hospital procurement as described in claim 1, characterized in that, Step S4 specifically includes: The system acquires supplier-related data from multiple sources, including the hospital's procurement management system, external public resource trading platforms, third-party testing platforms, and online public opinion monitoring platforms. This data includes: performance data from the hospital's procurement management system, consisting of historical supplier performance records; delivery time data from order placement and delivery timestamps; quality testing results generated by third-party testing platforms; market price index data and procurement transaction price fluctuation data collected by public resource trading platforms; and text-based public opinion data related to suppliers obtained from online public opinion monitoring platforms. After aligning the multi-source supplier-related data with time windows, statistical feature vectors of performance data, quality inspection data, and delivery timeliness data are extracted respectively. Time series modeling is performed on the market price index data and procurement transaction price fluctuation data to generate trend feature vectors, and semantic encoding is performed on public opinion data to generate text feature vectors. Feature vectors from different modalities are input into a multimodal fusion network, weighted and fused through an attention mechanism to generate a unified supplier state representation vector. The state representation vector is then dynamically updated based on a time decay mechanism to form a dynamic supplier state vector that evolves over time.

7. The intelligent decision-making method for hospital procurement through multi-source fusion as described in claim 1, characterized in that, Step S5 specifically includes: Based on the equity, control, cooperation and logistics dependencies among suppliers in the external supply chain data, a multi-type heterogeneous supply chain association graph is constructed, and different relationship edges are assigned corresponding propagation weights. Based on the association graph, the basic risk level of each node is calculated, and the risk propagation probability is calculated based on the weighted adjacency relationship between nodes to characterize the diffusion process of risk in the supply chain network. A risk propagation distribution model is constructed based on the node risk propagation probability, and the risk propagation entropy and its time change rate are calculated to characterize the concentration and evolution trend of risk diffusion. When the node risk level, node vulnerability, or risk propagation entropy change rate exceeds a preset threshold, the potential cascading failure path is predicted based on the graph propagation path simulation method, and an alternative supply chain candidate set is generated based on the neighborhood structure and connectivity of the failed node.

8. The intelligent decision-making method for hospital procurement through multi-source fusion as described in claim 1, characterized in that, Step S6 specifically includes: An optimization model is constructed with the allocation relationship between suppliers and demand nodes as decision variables. The objective functions include minimizing procurement costs, maximizing quality stability, maximizing supply continuity, and minimizing risk propagation intensity. The demand-supply matching degree matrix is ​​used as the allocation constraint, the supplier dynamic state vector is used as the supply capacity constraint, the cascading failure path is used as the non-selectable constraint, and the alternative supply chain candidate set is limited to the feasible solution space. The optimization model is solved using a multi-objective evolutionary algorithm based on non-dominated sorting to obtain a Pareto optimal solution set. Each Pareto optimal solution corresponds to a set of structured procurement resource allocation schemes. The procurement resource allocation schemes include: the material allocation ratio or procurement quantity corresponding to each supplier, the allocation path relationship of each material among different suppliers, the risk level identifier of the corresponding supplier, the supply chain alternative path selection results, and the cost indicators, quality indicators, supply continuity indicators, and risk propagation indicators of the corresponding schemes. Select a procurement resource allocation scheme that satisfies the constraints from the Pareto optimal solution set.

9. The intelligent decision-making method for multi-source fusion in hospital procurement as described in claim 1, characterized in that, Following step S6, the process further includes: collecting execution feedback data, constructing prediction error feedback, and performing online incremental updates on the spatiotemporal graph neural network, supplier dynamic state vector, and multi-objective optimization algorithm to achieve adaptive closed-loop optimization; wherein, the execution feedback data includes actual execution data from the hospital procurement management system and inventory management system, and the actual execution data includes purchase order execution records, actual material consumption data, inventory change data, and supplier fulfillment data; the prediction error feedback is constructed based on the deviation between the execution feedback data and the prediction output result of step S3, including demand prediction error, inventory change error, and supply capacity error.

10. A multi-source fusion intelligent decision-making system for hospital intelligent procurement, characterized in that, include: The data acquisition and processing module is used to collect data from the hospital's internal systems and external supply chain data, and to perform standardized encoding, entity disambiguation, spatiotemporal alignment and semantic mapping on heterogeneous data to generate a unified data representation. The knowledge graph construction module is used to construct a hospital procurement knowledge graph based on the unified data representation, model the relationships between materials, diseases, clinical departments, suppliers, inventory nodes and policy constraints as a graph structure, and generate graph embedding vectors; The demand and supply forecasting module is used to construct a dynamic demand and supply graph based on the knowledge graph, and to model the dynamic demand and supply graph using a spatiotemporal graph neural network to predict the demand, inventory changes and supply capacity of each material in the future cycle, and to generate a demand and supply matching degree matrix. The supplier status modeling module is used to integrate supplier performance data, quality inspection data, delivery timeliness data, market fluctuation data, and public opinion data in the unified data representation to construct a dynamic supplier status vector. The supply chain risk analysis module is used to construct a supply chain relationship graph, calculate the probability of risk propagation at nodes and the vulnerability of nodes, analyze the risk diffusion trend based on risk propagation entropy, predict cascading failure paths, and generate a set of alternative supply chain candidates. The multi-objective optimization decision module is used to solve the Pareto optimal procurement scheme by taking procurement cost, quality stability, supply continuity and risk propagation intensity as objective functions, and combining the demand-supply matching degree matrix, supplier dynamic state vector and cascading failure path and alternative supply chain candidate set as constraints. The adaptive closed-loop update module is used to collect execution feedback data, construct prediction error feedback, and perform online incremental updates on the spatiotemporal graph neural network, supplier dynamic state vector, and multi-objective optimization algorithm to achieve adaptive closed-loop optimization of the system.

Images