Medical waste supervision process risk early warning system based on artificial intelligence

Abstract

The invention relates to a medical waste supervision process risk early warning system based on artificial intelligence. The system comprises a whole process data acquisition module, a data preliminary processing module, a preliminary early warning model construction module, a risk grading model construction module and a supervision process risk early warning module. According to the invention, original data is obtained through data acquisition; a data preliminary processing method of multi-source data cleaning, graph sequence construction, time sequence attribute graph construction, data standardization and data set segmentation is adopted; a clustering model is adopted as a preliminary early warning model, potential risk process nodes are quickly identified through an unsupervised learning mode, high-quality input is provided for subsequent grading, and meanwhile, a graph structure is introduced to facilitate fusion of data attributes, structural relationships and time sequence behaviors; a deep learning model is adopted as a risk grading model, dynamic risk quantification is realized by integrating sub-graph features, time sequence context and environmental factors, and meanwhile, the fine grading precision is improved by utilizing data to drive threshold adjustment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent medical waste supervision, and particularly relates to a medical waste supervision process risk early warning system based on artificial intelligence. BACKGROUND

[0002] The medical waste supervision process risk early warning refers to real-time monitoring and analysis of the whole process of medical waste from generation, collection, transportation to disposal by using information technology, identifying and predicting potential risks, and issuing early warning signals in time to avoid or reduce possible environmental pollution and public health and safety hazards and other problems. The medical waste supervision process risk early warning can effectively prevent illegal disposal or improper disposal of medical waste, ensure compliance and safety in the medical waste disposal process, and reduce risks caused by improper operation or management.

[0003] However, the traditional medical waste supervision process risk early warning system has the technical problems that it relies on fixed rules or simple statistical analysis, cannot effectively process complex process data and time sequence dynamic changes, has insufficient detection capability for hidden abnormalities, is difficult to capture the correlation patterns in multi-dimensional data when the relationship between process nodes is complex, and causes early warning delay or missed reporting. The traditional medical waste supervision process risk early warning system has the technical problems that it is based on static threshold and simple classification model, cannot fully fuse high-dimensional features and time sequence information of process data, lacks adaptability to environmental factors and process evolution, causes rough risk level division and is easily disturbed by noise, and reduces the accuracy and reliability of the early warning system. SUMMARY

[0004] In view of the above, in order to overcome the defects of the prior art, the present application provides an artificial intelligence-based medical waste supervision process risk early warning system, which aims at the technical problems of the traditional medical waste supervision process risk early warning system, such as relying on fixed rules or simple statistical analysis, being unable to effectively process complex process data and time sequence dynamic changes, having insufficient detection capability for hidden abnormalities, being difficult to capture the correlation pattern in multi-dimensional data when the relationship between process nodes is complex, and causing early warning delay or false reporting. The present scheme creatively uses a clustering model as a preliminary early warning model, quickly identifies potential risk process nodes that deviate from the normal pattern through unsupervised learning, and provides high-quality input for subsequent grading, thereby avoiding direct processing of medical waste supervision process data involving multiple nodes, multiple sources of data and complex processes. At the same time, a graph structure is introduced to facilitate the fusion of data attributes, structural relationships and time sequence behaviors, and to enhance the reliability and real-time performance of the overall early warning. For the technical problems of the traditional medical waste supervision process risk early warning system, such as being based on static threshold and simple classification model, being unable to fully integrate the high-dimensional features and time sequence information of process data, and lacking adaptability to environmental factors and process evolution, resulting in rough risk level division and being susceptible to noise interference, reducing the accuracy and reliability of the early warning system. The present scheme creatively uses a deep learning model as a risk grading model, which can capture nonlinear relationships and long-term dependencies by integrating subgraph features, time sequence context and environmental factors, realize dynamic risk quantification, and improve the accuracy of fine-grained grading by using data-driven threshold adjustment, thereby ensuring that the early warning results are more operable.

[0005] The technical scheme adopted by the present application is as follows: The artificial intelligence-based medical waste supervision process risk early warning system provided by the present application comprises a full-process data acquisition module, a data preliminary processing module, a preliminary early warning model construction module, a risk grading model construction module and a supervision process risk early warning module.

[0006] The full-process data acquisition module acquires data from each medical waste supervision process node to obtain a risk early warning original data set, and sends the risk early warning original data set to the data preliminary processing module.

[0007] The data preliminary processing module uses a data preliminary processing method of multi-source data cleaning, graph sequence construction, construction of time sequence attribute graph, data standardization and data set segmentation to convert the collected multi-source data into a unified time sequence attribute graph structure, obtain an early warning data set, an early warning training set and an early warning test set, send the early warning data set to the supervision process risk early warning module, and send the early warning training set and the early warning test set to the preliminary early warning model construction module.

[0008] The preliminary early warning model construction module constructs a clustering model as a preliminary early warning model to provide a process node risk subgraph for subsequent medical waste supervision process risk classification, processes the early warning training set and the early warning test set to obtain a classification training set and a classification test set, and sends the preliminary early warning model to the supervision process risk early warning module and sends the classification training set and the classification test set to the risk classification model construction module.

[0009] The risk classification model construction module is configured to construct a model required for medical waste supervision process risk classification, construct a deep learning model as a risk classification model, and send the risk classification model to the supervision process risk early warning module.

[0010] The supervision process risk early warning module is configured to perform real-time medical waste supervision process risk early warning based on the preliminary early warning model and the risk classification model to obtain a risk early warning result.

[0011] Further, in the whole-process data acquisition module, the risk early warning original data set specifically includes a past risk early warning original data set and a real-time risk early warning original data set, both of which include static data, dynamic time series data, and process relationship data of each medical waste supervision process node, and the past risk early warning original data set further includes risk classification label data.

[0012] Further, in the data preliminary processing module, the multi-source data cleaning is configured to ensure data quality and consistency, specifically to perform timestamp alignment, missing value interpolation, and cleaning processing on the collected multi-source data, and unify the data format.

[0013] The graph sequence construction is configured to convert a continuous data stream into a discrete graph sequence, specifically by slicing with a fixed time window, constructing a static graph for each slice, representing a process node as a node of the static graph, constructing an edge based on process relationship data, then constructing a virtual node for each static graph, establishing a bidirectional connection with all nodes, and setting the edge weight between the virtual node and each node as a normalized value of the operation frequency of the process node represented by the node within the time slice, and constructing a graph sequence from all slice static graphs.

[0014] The construction of a time series attribute graph is configured to form a complete time series attribute graph, specifically by combining static graphs in chronological order and retaining node attributes, edge attributes, and timestamps to obtain a time series attribute graph.

[0015] The data standardization is configured to normalize data scales to eliminate dimensional effects, specifically by processing node and edge attributes through a Z-score standardization method to obtain a standardized time series attribute graph.

[0016] The data set is divided, and training data and test data are obtained, specifically, the past risk early warning original data set is divided into data sets;

[0017] Through the multi-source data cleaning, the graph sequence construction, the construction of the time sequence attribute graph and the data standardization, the real-time risk early warning original data set is preliminarily processed to obtain an early warning data set. Through the multi-source data cleaning, the graph sequence construction, the construction of the time sequence attribute graph, the data standardization and the data set division, the past risk early warning original data set is preliminarily processed to obtain an early warning training set and an early warning test set.

[0018] Further, in the preliminary early warning model construction module, a model required for constructing a preliminary identification medical waste supervision process risk process node is constructed, specifically, a clustering model is constructed as a preliminary early warning model;

[0019] The preliminary early warning model construction module specifically includes a node embedding generation unit, a similarity calculation unit, a density clustering unit, a risk process output unit and a data set processing unit;

[0020] The node embedding generation unit is configured to represent the nodes as low-dimensional vectors to capture features, specifically, the node attribute and neighbor information are aggregated by a GraphSAGE model to obtain a node embedding set, which includes:

[0021] The initialization node embedding is configured to set the initial state of the node embedding, specifically, the node attribute is used as the initial embedding to obtain the initialization node embedding;

[0022] The GraphSAGE forward propagation is configured to update the node embedding to integrate local structure information, specifically, the neighbor node features are aggregated by a message passing mechanism to obtain the node embedding;

[0023] The similarity calculation unit is configured to measure the similarity between nodes to identify anomalies, specifically, the graph structure similarity is obtained by fusing attribute, structure and time sequence similarity, which includes:

[0024] The node attribute similarity is calculated, which is configured to compare the similarity of static and dynamic attributes of the nodes, specifically, the similarity of the node embedding is quantified by cosine similarity to obtain the node attribute similarity;

[0025] The node structure similarity is calculated, which is configured to compare the structure similarity of the nodes, specifically, the structure embedding of the nodes is generated by a Weisfeiler-Lehman algorithm, and the cosine similarity is calculated to obtain the node structure similarity;

[0026] The node timing behavior similarity is calculated for comparing the behavior patterns of the nodes over time, specifically by dynamic time warping distance, analyzing the operation frequency sequence of the process node represented by the node and converting into similarity to obtain the node timing behavior similarity;

[0027] The graph structure similarity is calculated for generating a comprehensive similarity metric, specifically by weighted fusion of the node attribute similarity, the node structure similarity and the node timing behavior similarity to obtain the graph structure similarity;

[0028] The density clustering unit is configured to identify and group the risk process nodes, specifically by density clustering based on the graph structure similarity to obtain the risk subgraph of the process node, and the content includes:

[0029] The risk node is marked for locating the node not conforming to the historical pattern, specifically by obtaining the graph structure similarity of the node of the current time slice with other nodes of the historical time slice and comparing with the preset risk threshold, and if there are less than d times of the node graph structure similarity greater than the preset risk threshold, the node is marked as a risk node;

[0030] The risk node is clustered for grouping the risk nodes into a continuous area, specifically by aggregating adjacent risk nodes based on the connection relationship of the edges in the standardized timing attribute graph to form a risk subgraph;

[0031] The risk process output unit is configured to map the risk subgraph to the actual supervision process, specifically for each risk subgraph, extracting the nodes and edges contained therein, mapping to the actual supervision process to obtain the preliminary risk label of the process node, and the preliminary risk label specifically includes no risk and risk;

[0032] The dataset processing unit is specifically integrated with the node embedding generation unit, the similarity calculation unit, the density clustering unit and the risk process output unit to construct a clustering model as a preliminary early warning model, and the early warning training set and the early warning test set are processed as inputs of the preliminary early warning model to obtain a hierarchical training set and a hierarchical test set.

[0033] Further, in the risk grading model construction module, a model required for medical waste supervision process risk grading is constructed, specifically a deep learning model is constructed as a risk grading model;

[0034] The risk grading model construction module specifically includes a risk subgraph encoding unit, a timing context encoding unit, an environment context encoding unit, a gating fusion unit, a risk level division unit and a model construction training unit;

[0035] The risk subgraph encoding unit is configured to encode the risk subgraph into a low-dimensional vector, specifically, by performing embedding calculation on the nodes in the risk subgraph through a GraphSAGE model, and fusing all the node embeddings through an average pooling operation to obtain subgraph encoding features.

[0036] The time context encoding unit is configured to capture the time evolution pattern of the risk subgraph, specifically, by processing the operation time sequence of the risk subgraph through a gated recurrent unit network to obtain time context features, which include:

[0037] The operation time sequence is constructed to prepare time series data, specifically, for the risk subgraph, the average of the operation times of the process nodes represented by all the nodes in each time slice of the risk subgraph is calculated, and a sequence is formed in time order to obtain the operation time sequence.

[0038] The time context features are obtained to encode the time dependence, specifically, by processing the operation time sequence of the risk subgraph through a gated recurrent unit network and outputting the final hidden state as the time context features.

[0039] The environmental context encoding unit is configured to incorporate the influence of external environmental factors on the risk, specifically, by encoding the environment-related attributes of all the nodes in the risk subgraph through an independent multi-layer perceptron to obtain environmental context features.

[0040] The gated fusion unit is configured to dynamically weight and fuse multi-source features, specifically, by controlling the contribution degree of the time context features and the environmental context features through a gating weight, and weighting and fusing the subgraph encoding features and the time context features to obtain final fusion features.

[0041] The risk level division unit is configured to generate a dynamic risk classification threshold and divide the risk level, specifically, by calculating the degree of deviation of the final fusion features from the normal reference through Mahalanobis distance, and comparing it with the dynamic risk classification threshold to divide the risk level to obtain a risk classification label, which includes:

[0042] The historical normal reference set is constructed to establish a baseline distribution of the normal process, specifically, subgraph instances isomorphic to the current risk subgraph are extracted from training data with a risk-free label, and the subgraph instances are input into the risk subgraph encoding unit, the time context encoding unit, the environmental context encoding unit, and the gated fusion unit in turn to obtain the final fusion features of each subgraph instance, which constitute the historical normal reference set.

[0043] The mean and covariance are calculated to describe the statistical distribution of the normal data, specifically, the mean and covariance matrix of the final fusion features of each subgraph instance in the historical normal reference set are calculated to obtain the parameters of the normal distribution.

[0044] The Mahalanobis distance is calculated to measure the deviation of the current risk from the normal state, specifically, the Mahalanobis distance is calculated based on the mean and covariance matrix of the final fusion features of each subgraph instance in the historical normal reference set, and the Mahalanobis distance is taken as a risk score;

[0045] The dynamic threshold division is used to determine the risk level boundary, specifically, the training data labeled as low risk is taken as input and sequentially processed by the risk subgraph encoding unit, the time sequence context encoding unit, the environment context encoding unit and the gate fusion unit, and the risk score is calculated to obtain a low risk score set, and the 60th percentile of the low risk score set is taken as a first dynamic risk grading threshold, and the training data labeled as medium risk is processed in the same way to obtain a medium risk score set, and the 85th percentile of the medium risk score set is taken as a second dynamic risk grading threshold;

[0046] The current risk score is compared with the first dynamic risk grading threshold and the second dynamic risk grading threshold to divide the risk level.

[0047] The model construction training unit is specifically integrated with the risk subgraph encoding unit, the time sequence context encoding unit, the environment context encoding unit, the gate fusion unit and the risk level division unit to construct a deep learning model, and the model is trained based on the grading training set and the grading test set and the performance is verified to obtain a deep learning model as a risk grading model.

[0048] Further, in the regulatory process risk early warning module, specifically, first, the early warning data set is taken as input of the preliminary early warning model to preliminarily identify the medical waste regulatory process risk process node, and output the process node name with a preliminary risk label of risk and a risk subgraph, then the early warning data set and the risk subgraph output by the preliminary early warning model are input into the risk grading model to output a risk grading label, and finally, the process node name with a preliminary risk label of risk and the risk grading label are taken as a risk early warning result to issue an early warning to relevant management personnel. Figure One Further, in the regulatory process risk early warning module, specifically, first, the early warning data set is taken as input of the preliminary early warning model to preliminarily identify the medical waste regulatory process risk process node, and output the process node name with a preliminary risk label of risk and a risk subgraph, then the early warning data set and the risk subgraph output by the preliminary early warning model are input into the risk grading model to output a risk grading label, and finally, the process node name with a preliminary risk label of risk and the risk grading label are taken as a risk early warning result to issue an early warning to relevant management personnel.

[0049] The above scheme has the following beneficial effects:

[0050] (1) In view of the technical problems of the traditional medical waste supervision process risk early warning system, such as relying on fixed rules or simple statistical analysis, being unable to effectively process complex process data and time sequence dynamic changes, having insufficient detection capability for hidden abnormalities, being difficult to capture the correlation pattern in multi-dimensional data when the relationship between process nodes is complex, and causing early warning delay or missed report, the scheme creatively adopts a clustering model as a preliminary early warning model, quickly identifies potential risk process nodes deviating from the normal pattern through unsupervised learning, provides high-quality input for subsequent grading, thereby avoiding direct processing of medical waste supervision process data involving multi-node, multi-source data and complex process, and introducing a graph structure to facilitate the fusion of data attributes, structural relationships and time sequence behaviors, and enhance the reliability and real-time performance of the overall early warning.

[0051] (2) In view of the technical problems of the traditional medical waste supervision process risk early warning system, such as being based on static threshold and simple classification model, being unable to fully fuse high-dimensional features and time sequence information of process data, and lacking adaptability to environmental factors and process evolution, resulting in rough risk level division and being susceptible to noise interference, and reducing the accuracy and reliability of the early warning system, the scheme creatively adopts a deep learning model as a risk grading model, integrates sub-graph features, time sequence context and environmental factors, can capture nonlinear relationships and long-term dependencies, realizes dynamic risk quantification, and uses data-driven threshold adjustment to improve fine-grained grading accuracy, thereby ensuring that the early warning result is more operable. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 A module schematic diagram of the medical waste supervision process risk early warning system based on artificial intelligence is provided.

[0053] Figure 2 A flowchart of the data preliminary processing module.

[0054] Figure 3 A flowchart of the preliminary early warning model construction module.

[0055] Figure 4 A flowchart of the risk grading model construction module.

[0056] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation of the present application. DETAILED DESCRIPTION

[0057] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below, obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments; based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0058] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0059] Embodiment one, refer to Figure 1 The present application provides a medical waste supervision process risk early warning system based on artificial intelligence, comprising a whole-process data acquisition module, a data preliminary processing module, a preliminary early warning model construction module, a risk grading model construction module and a supervision process risk early warning module.

[0060] The whole-process data acquisition module acquires data from each medical waste supervision process node to obtain a risk early warning original data set, and sends the risk early warning original data set to the data preliminary processing module.

[0061] The data preliminary processing module adopts a data preliminary processing method of multi-source data cleaning, graph sequence construction, construction of time series attribute graph, data standardization and data set segmentation, converts the collected multi-source data into a unified time series attribute graph structure, obtains an early warning data set, an early warning training set and an early warning test set, sends the early warning data set to the supervision process risk early warning module, and sends the early warning training set and the early warning test set to the preliminary early warning model construction module.

[0062] The preliminary early warning model construction module constructs a clustering model as a preliminary early warning model, provides a process node risk subgraph for subsequent medical waste supervision process risk grading, processes the early warning training set and the early warning test set to obtain a grading training set and a grading test set, and sends the preliminary early warning model to the supervision process risk early warning module and sends the grading training set and the grading test set to the risk grading model construction module.

[0063] The risk grading model construction module is used to construct a model required for medical waste supervision process risk grading, constructs a deep learning model as a risk grading model, and sends the risk grading model to the supervision process risk early warning module.

[0064] The monitoring process risk early warning module, specifically based on the preliminary early warning model and the risk classification model, performs real-time medical waste monitoring process risk early warning to obtain a risk early warning result.

[0065] Embodiment Two, see Figure 1 This embodiment is based on the above embodiment. In the full-process data acquisition module, the risk early warning original data set specifically includes a past risk early warning original data set and a real-time risk early warning original data set. Both the past risk early warning original data set and the real-time risk early warning original data set include static data, dynamic time series data, and process relationship data of each medical waste monitoring process node. The past risk early warning original data set further includes risk classification label data.

[0066] The medical waste monitoring process node specifically includes a generation and classification node, a collection and packaging node, an in-hospital temporary storage node, an in-hospital and out-hospital transportation node, a centralized storage node, and a final disposal node.

[0067] The generation and classification node collects static data including generation source type data, responsible personnel ID data, department affiliation data, and in-hospital location data, and collects dynamic time series data including waste deposit timestamp data, deposit and classification operation personnel ID data, deposit and classification operation time consumption data, waste type data, waste weight data, and waste package integrity data.

[0068] The collection and packaging node collects static data including collection node type data, responsible personnel ID data, unit affiliation data, and in-hospital location data, and collects dynamic time series data including collection and packaging operation timestamp data, collection and packaging operation personnel ID data, collection and packaging operation time consumption data, waste weight data, and waste package integrity data.

[0069] The in-hospital temporary storage node collects static data including temporary storage node type data, responsible personnel ID data, unit affiliation data, and in-hospital location data, and collects dynamic time series data including waste entry into temporary storage timestamp data, temporary storage operation personnel ID data, waste temporary storage duration data, temporary storage area environmental temperature and humidity data, temporary storage area odor gas concentration data, waste weight data, and waste package integrity data.

[0070] The in-hospital and out-hospital transportation node collects static data including transportation tool type data, responsible personnel ID data, unit affiliation data, and transportation location coordinate data, and collects dynamic time series data including transportation operation timestamp data, transportation operation personnel ID data, transportation operation time consumption data, transportation tool load data, and transportation tool vibration sensor data.

[0071] The centralized storage node collects static data including storage node type data, responsible personnel ID data, unit data, and storage location coordinate data, and collects dynamic time series data including timestamp data, storage operation personnel ID data, waste storage time length data, storage area environment temperature and humidity data, storage area odor gas concentration data, waste weight data, and waste package integrity data.

[0072] The final disposal node collects static data including disposal tool type data, responsible personnel ID data, unit data, and disposal location coordinate data, and collects dynamic time series data including disposal start and end timestamp data, disposal operation type data, disposal operation personnel ID data, disposal operation time consumption data, disposal equipment operation data, and disposal area odor gas concentration data.

[0073] The risk classification label data specifically includes no risk, low risk, medium risk, and high risk.

[0074] Embodiment three, refer to Figure 1 and Figure 2 This embodiment is based on the above-mentioned embodiment, in the data preliminary processing module, the multi-source data cleaning is used to ensure data quality and consistency, specifically for the collected multi-source data to carry out timestamp alignment, missing value interpolation and cleaning processing, and unify the data format;

[0075] The graph sequence construction is used to convert continuous data stream into discrete graph sequence, specifically by fixed time window slicing, each slice constructs a static graph, the node of the static graph represents the process node, the edge is constructed based on the process relationship data, then a virtual node is constructed for each static graph, and a bidirectional connection is established with all nodes, the edge weight between the virtual node and each node is the normalized value of the operation frequency of the process node represented by the node within the time slice, and all slice static graphs constitute a graph sequence.

[0076] The construction of time series attribute graph is used to form a complete time series attribute graph, specifically by combining static graphs in time sequence and retaining node attributes, edge attributes, and timestamps to obtain a time series attribute graph.

[0077] The data standardization is used to normalize the data scale to eliminate the influence of dimension, specifically by processing node and edge attributes through Z-score standardization method to obtain a standardized time series attribute graph.

[0078] The data set segmentation is used to obtain training data and test data, specifically by performing data set segmentation on the past risk early warning original data set.

[0079] The real-time risk early warning original data set is preliminarily processed through the multi-source data cleaning, the graph sequence construction, the constructed time sequence attribute graph and the data standardization to obtain an early warning data set, and the past risk early warning original data set is preliminarily processed through the multi-source data cleaning, the graph sequence construction, the constructed time sequence attribute graph, the data standardization and the data set segmentation to obtain an early warning training set and an early warning test set.

[0080] Embodiment four, refer to Figure 1 and Figure 3 This embodiment is based on the above-mentioned embodiments, and in the preliminary early warning model construction module, a model required for constructing a preliminary identification medical waste supervision process risk process node is constructed, specifically a clustering model is constructed as a preliminary early warning model.

[0081] The preliminary early warning model construction module specifically includes a node embedding generation unit, a similarity calculation unit, a density clustering unit, a risk process output unit and a data set processing unit.

[0082] The node embedding generation unit is used to represent the node as a low-dimensional vector to capture the characteristics, specifically by aggregating node attributes and neighbor information through the GraphSAGE model to obtain a node embedding set, which includes:

[0083] The initialization node embedding is used to set the initial state of the node embedding, specifically using the node attribute as the initial embedding to obtain the initialization node embedding.

[0084] The GraphSAGE forward propagation is used to update the node embedding to integrate local structure information, specifically by aggregating neighbor node features through a message passing mechanism to obtain the node embedding, and the formula used is as follows:

[0085] ;

[0086] In the formula, represents the node embedding of the ath node in the time slice t, represents the ReLU activation function, represents the node embedding generation learnable weight, represents the connection operation function, represents the node embedding of the ath node in the time slice t-1, represents the neighbor set of the ath node, represents the node embedding of the bth node in the time slice t-1.

[0087] The similarity calculation unit is used to measure the similarity between nodes to identify abnormalities, specifically by fusing attribute, structure and time sequence similarity to obtain a graph structure similarity, which includes:

[0088] The node attribute similarity is calculated to compare the similarity of the static and dynamic attributes of the nodes, specifically, the similarity of the node embedding is quantified by the cosine similarity, to obtain the node attribute similarity, and the formula used is as follows:

[0089] ;

[0090] In the formula, denotes the node attribute similarity between the a th node and the c th node, denotes the cosine similarity calculation function, denotes the node embedding of the a th node, denotes the node embedding of the c th node;

[0091] The node structure similarity is calculated to compare the structural similarity of the nodes, specifically, the structural embedding of the nodes is generated by the Weisfeiler-Lehman algorithm, and the cosine similarity is calculated to obtain the node structure similarity, and the formula used is as follows:

[0092] ;

[0093] In the formula, denotes the node structure similarity between the a th node and the c th node, denotes the Weisfeiler-Lehman algorithm function;

[0094] The node timing behavior similarity is calculated to compare the behavior pattern of the nodes over time, specifically, the operation frequency sequence of the flow node represented by the node is analyzed by the dynamic time warping distance, and is converted into similarity to obtain the node timing behavior similarity, and the formula used is as follows:

[0095] ;

[0096] In the formula, denotes the node timing behavior similarity between the a th node and the c th node, denotes the node timing behavior similarity scaling parameter, denotes the dynamic time warping distance calculation function, denotes the operation frequency sequence of the flow node represented by the a th node, denotes the operation frequency sequence of the flow node represented by the c th node;

[0097] The graph structure similarity is calculated to generate a comprehensive similarity measure, specifically, the node attribute similarity, the node structure similarity and the node timing behavior similarity are fused by weighting to obtain the graph structure similarity, and the formula used is as follows:

[0098] ;

[0099] wherein, denotes the graph structure similarity between the a-th node and the c-th node, denotes the attribute similarity weight, denotes the structure similarity weight, denotes the time series similarity weight;

[0100] The density clustering unit is configured to identify and group the risk process nodes, specifically, to perform density clustering based on the graph structure similarity, to obtain a risk subgraph of the process nodes, and the content includes:

[0101] The risk node marking unit is configured to locate the nodes that do not conform to the historical pattern, specifically, to obtain the graph structure similarity between the nodes of the current time slice and the other nodes of the historical time slice, and compare the graph structure similarity with a preset risk threshold, if there are less than d times of nodes with a graph structure similarity greater than the preset risk threshold, the node is marked as a risk node;

[0102] The clustering risk node unit is configured to group the risk nodes into continuous regions, specifically, to gather adjacent risk nodes based on the connection relationship of the edges in the standardized time series attribute graph, to form a risk subgraph;

[0103] The risk process output unit is configured to map the risk subgraph to the actual supervision process, specifically, for each risk subgraph, to extract the nodes and edges contained therein, to map to the actual supervision process, to obtain a preliminary risk label of the process node, and the preliminary risk label specifically includes no risk and risk;

[0104] The data set processing unit is specifically configured to integrate the node embedding generation unit, the similarity calculation unit, the density clustering unit and the risk process output unit, to construct a clustering model as a preliminary early warning model, and to process the preliminary early warning model by taking the early warning training set and the early warning test set as inputs, to obtain a hierarchical training set and a hierarchical test set.

[0105] By performing the above operations, the technical problem of the conventional medical waste supervision process risk early warning system, which relies on fixed rules or simple statistical analysis, cannot effectively process complex process data and time series dynamic changes, has insufficient detection capability for hidden abnormalities, when the relationship between process nodes is complex, it is difficult to capture the correlation pattern in multi-dimensional data, causing early warning delay or false reporting, is solved. The scheme creatively uses a clustering model as a preliminary early warning model, quickly identifies potential risk process nodes that deviate from the normal pattern through unsupervised learning, provides high-quality input for subsequent hierarchical analysis, thereby avoiding direct processing of medical waste supervision process data involving multi-node, multi-source data and complex processes, and introducing a graph structure facilitates the fusion of data attributes, structural relationships and time series behavior, and enhances the reliability and real-time performance of the overall early warning.

[0106] Embodiment five, refer to Figure 1 and Figure 4 , based on the above embodiment, in the risk grading model construction module, for constructing the model required for constructing the medical waste supervision process risk grading, by constructing a deep learning model as a risk grading model;

[0107] The risk grading model construction module, the specific content includes risk subgraph encoding unit, time sequence context encoding unit, environment context encoding unit, gate fusion unit, risk grade division unit and model construction training unit;

[0108] The risk subgraph encoding unit is used to represent the risk subgraph as a low-dimensional vector, specifically by embedding the nodes in the risk subgraph through the GraphSAGE model, and fusing all node embeddings through average pooling operation to obtain subgraph encoding features;

[0109] The time sequence context encoding unit is used to capture the time evolution pattern of the risk subgraph, specifically by processing the operation time sequence of the risk subgraph through the gated recurrent unit network to obtain the time sequence context features, which include:

[0110] The operation time sequence is constructed to prepare the time series data, specifically for the risk subgraph, the average of the operation times of the process nodes represented by all nodes of each time slice is calculated, and the sequence is formed in time order to obtain the operation time sequence, the formula is as follows:

[0111] ;

[0112] In the formula, represents the average of the operation times of the process nodes represented by all nodes of the risk subgraph of time slice t, represents the node set of the risk subgraph of time slice t, represents the operation times of the process nodes represented by the mth node of the risk subgraph, and Os represents the operation time sequence, represents the average of the operation times of the process nodes represented by all nodes of the risk subgraph of the first time slice, represents the average of the operation times of the process nodes represented by all nodes of the risk subgraph of the second time slice;

[0113] The time sequence context feature is obtained to encode the time sequence dependency, specifically by processing the operation time sequence of the risk subgraph through the gated recurrent unit network and outputting the last hidden state as the time sequence context feature, the formula is as follows:

[0114] ;

[0115] In the formula, representing a time sequence context feature, representing a gating recurrent unit network function;

[0116] The environmental context encoding unit is used to integrate the influence of external environmental factors on risk, specifically by using an independent multi-layer perception machine to encode the environmental-related attributes of all nodes in the risk subgraph to obtain the environmental context feature.

[0117] The gating fusion unit is used to dynamically weight and fuse multi-source features, specifically by using a gating weight to control the contribution of the time sequence context feature and the environmental context feature, and weighting and fusing the subgraph encoding features and the time sequence context feature to obtain the final fusion feature, using the following formula:

[0118] ;

[0119] In the formula, Gw represents the gating weight, representing a sigmoid activation function, representing a gating weight linear mapping weight, representing a gating weight linear mapping bias term, representing an environmental context feature, representing a final fusion feature, representing a subgraph encoding feature, representing an element-wise multiplication operation;

[0120] The risk level division unit is used to generate a dynamic risk classification threshold and divide the risk level, specifically by calculating the degree of deviation of the final fusion feature from the normal reference, and comparing it with the dynamic risk classification threshold to divide the risk level and obtain the risk classification label, which includes:

[0121] Historical normal reference set construction is used to establish a baseline distribution of normal processes, specifically by extracting subgraph instances isomorphic to the current risk subgraph from training data labeled as risk-free, and inputting the subgraph instances into the risk subgraph encoding unit, the time sequence context encoding unit, the environmental context encoding unit, and the gating fusion unit in turn to obtain the final fusion features of each subgraph instance, which constitute the historical normal reference set.

[0122] Calculating the mean and covariance is used to describe the statistical distribution of normal data, specifically by calculating the mean and covariance matrix of the final fusion features of each subgraph instance in the historical normal reference set to obtain the normal distribution parameters.

[0123] Calculating the Mahalanobis distance is used to measure the degree of deviation of the current risk from the normal state, specifically by calculating the Mahalanobis distance based on the mean and covariance matrix of the final fusion features of each subgraph instance in the historical normal reference set, and using it as the risk score, using the following formula:

[0124] ;

[0125] wherein, denotes the risk score, denotes the final fusion feature of the risk sub-graph, denotes the mean of the final fusion feature of each sub-graph instance in the historical normal reference set, denotes the covariance matrix of the final fusion feature of each sub-graph instance in the historical normal reference set, and T denotes the transpose operation;

[0126] Dynamic threshold division is used to determine the risk level boundary, specifically, the training data labeled as low risk is taken as input to sequentially pass through the risk sub-graph encoding unit, the time series context encoding unit, the environmental context encoding unit, and the gate fusion unit, and the risk score is calculated to obtain a low risk score set, and the 60th percentile of the low risk score set is taken as the first dynamic risk grading threshold, and the training data labeled as medium risk is processed in the same way to obtain a medium risk score set, and the 85th percentile of the medium risk score set is taken as the second dynamic risk grading threshold;

[0127] The current risk score is compared with the first dynamic risk grading threshold and the second dynamic risk grading threshold to divide the risk level, if , the risk level is divided into low risk, if , the risk level is divided into medium risk, if , the risk level is divided into high risk, wherein, denotes the first dynamic risk grading threshold, denotes the second dynamic risk grading threshold;

[0128] The model construction training unit is specifically integrated with the risk sub-graph encoding unit, the time series context encoding unit, the environmental context encoding unit, the gate fusion unit, and the risk level division unit to construct a deep learning model, train the model based on the grading training set and the grading test set, and verify the performance to obtain a deep learning model as a risk grading model.

[0129] By performing the above operation, the technical problem that the traditional medical waste supervision process risk early warning system exists based on a static threshold and a simple classification model, cannot fully fuse high-dimensional features and time sequence information of process data, and lacks adaptability to environmental factors and process evolution, leading to rough risk level division and easy noise interference, reducing the accuracy and reliability of the early warning system is solved. The scheme creatively uses a deep learning model as a risk grading model, which can capture nonlinear relationships and long-term dependencies by integrating subgraph features, time sequence context and environmental factors, realize dynamic risk quantification, and improve fine grading accuracy by using data-driven threshold adjustment, so as to ensure that the early warning result is more operable.

[0130] Embodiment six, see Figure 1 The embodiment is based on the above-mentioned embodiment. In the supervision process risk early warning module, specifically, first, the early warning data set is taken as the input of the preliminary early warning model, the medical waste supervision process risk process node is preliminarily identified, and the preliminary risk label is output as the process node name and risk subgraph of the process node with risk. Then, the early warning data set and the risk subgraph output by the preliminary early warning model are input into the risk grading model, and the risk grading label is output. Finally, the preliminary risk label of the process node with risk and the risk grading label are taken as the risk early warning result, and the early warning is sent to the relevant management personnel. Figure One

[0131] It should be noted that in this paper, relationship terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment.

[0132] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application.

[0133] The above describes the present application and its embodiments, which are not limited, and the embodiments shown in the drawings are only one of the embodiments of the present application, and the actual structure is not limited thereto. In general, if those skilled in the art are inspired, without departing from the purpose of the present application, without creative design, similar structure and embodiments of the technical solution can be designed, which should belong to the protection scope of the present application.​

Claims

1. An artificial intelligence-based medical waste supervision process risk early warning system, characterized in that: The system comprises a full-process data acquisition module, a data preliminary processing module, a preliminary early warning model construction module, a risk grading model construction module and a supervision process risk early warning module. The full-process data acquisition module acquires data from each medical waste supervision process node to obtain a risk early warning original data set, which specifically comprises a past risk early warning original data set and a real-time risk early warning original data set. The data preliminary processing module adopts a data preliminary processing method of multi-source data cleaning, graph sequence construction, construction of time series attribute graph, data standardization and data set segmentation to obtain an early warning data set, an early warning training set and an early warning test set. The preliminary early warning model construction module is used for constructing a model required for preliminary identification of a medical waste supervision process risk process node, specifically a clustering model as a preliminary early warning model, and processing the early warning training set and the early warning test set to obtain a grading training set and a grading test set. The risk grading model construction module is used for constructing a model required for medical waste supervision process risk grading, specifically a deep learning model as a risk grading model. The supervision process risk early warning module specifically comprises the following steps: firstly, inputting the early warning data set into the preliminary early warning model to preliminarily identify a medical waste supervision process risk process node and outputting a preliminary risk label of a process node name and a risk subgraph; secondly, inputting the early warning data set and the risk subgraph output by the preliminary early warning model into the risk grading model to output a risk grading label; and finally, outputting the preliminary risk label of the process node name and the risk grading label as a risk early warning result to relevant management personnel for early warning.

2. The medical waste supervision process risk early warning system based on artificial intelligence according to claim 1, characterized in that: The preliminary early warning model construction module specifically comprises a node embedding generation unit, a similarity calculation unit, a density clustering unit, a risk process output unit and a data set processing unit. 3.The medical waste supervision process risk early warning system based on artificial intelligence according to claim 2, characterized in that: The node embedding generation unit is used for representing a node as a low-dimensional vector to capture features, specifically aggregating node attributes and neighbor information through a GraphSAGE model to obtain a node embedding set, which comprises the following steps: initializing node embedding, which is used for setting an initial state of node embedding, specifically using node attributes as initial embedding to obtain initialized node embedding; GraphSAGE forward propagation, which is used for updating node embedding to incorporate local structure information, specifically aggregating neighbor node features through a message passing mechanism to obtain node embedding; The similarity calculation unit is used for measuring the similarity between nodes to identify abnormalities, specifically obtaining graph structure similarity by fusing attribute, structure and time series similarity, which comprises the following steps: calculating node attribute similarity, which is used for comparing the similarity of node static and dynamic attributes, specifically quantifying the similarity of node embedding through cosine similarity to obtain node attribute similarity; calculating node structure similarity, which is used for comparing the structure similarity of nodes, specifically generating node structure embedding through a Weisfeiler-Lehman algorithm and calculating cosine similarity to obtain node structure similarity; and The node timing behavior similarity is calculated to compare the behavior patterns of the nodes over time, specifically by dynamic time warping distance, analyzing the operation frequency sequence of the process node represented by the node and converting into similarity to obtain the node timing behavior similarity; The graph structure similarity is calculated to generate a comprehensive similarity metric, specifically by weighted fusion of the node attribute similarity, the node structure similarity and the node timing behavior similarity to obtain the graph structure similarity; The density clustering unit is configured to identify and group the risk process nodes, specifically by density clustering based on the graph structure similarity to obtain the risk subgraph of the process node, and the content includes: The risk node is marked to locate the node that does not conform to the historical pattern, specifically by obtaining the graph structure similarity of the node of the current time slice and other nodes of the historical time slice, and comparing with the preset risk threshold, if there are less than d times of node graph structure similarity greater than the preset risk threshold, the node is marked as a risk node; The risk node is clustered to group the risk nodes into a continuous area, specifically by aggregating adjacent risk nodes based on the connection relationship of the edges in the standardized timing attribute graph to form a risk subgraph; The risk process output unit is configured to map the risk subgraph to the actual supervision process, specifically for each risk subgraph, extracting the nodes and edges contained therein, and mapping to the actual supervision process to obtain a preliminary risk label of the process node, and the preliminary risk label specifically includes no risk and risk; The dataset processing unit is configured to integrate the node embedding generation unit, the similarity calculation unit, the density clustering unit and the risk process output unit to construct a clustering model as a preliminary warning model, and process the preliminary warning training set and the preliminary warning test set as inputs of the preliminary warning model to obtain a hierarchical training set and a hierarchical test set.

4. The artificial intelligence-based medical waste supervision process risk early warning system according to claim 1, characterized in that: The risk grading model construction module specifically includes a risk subgraph encoding unit, a timing context encoding unit, an environment context encoding unit, a gated fusion unit, a risk level division unit and a model construction training unit.

5. The artificial intelligence-based medical waste supervision process risk early warning system according to claim 4, characterized in that: The risk subgraph encoding unit is configured to represent the risk subgraph as a low-dimensional vector, specifically by embedding calculation of the nodes in the risk subgraph through the GraphSAGE model, and fusing all node embeddings through average pooling operation to obtain subgraph encoding features; The timing context encoding unit is configured to capture the time evolution pattern of the risk subgraph, specifically by processing the operation timing sequence of the risk subgraph through the gated recurrent unit network to obtain timing context features, and the content includes: The operation timing sequence is constructed to prepare time series data, specifically for the risk subgraph, calculating the mean of the operation frequency of the process node represented by all nodes of each time slice of the risk subgraph, and forming a sequence in chronological order to obtain the operation timing sequence; The timing context feature is obtained to encode the timing dependency, specifically by processing the operation timing sequence of the risk subgraph through the gated recurrent unit network and outputting the last hidden state as the timing context feature; The environmental context coding unit is configured to integrate the influence of external environmental factors on the risk, specifically by coding the environmental-related attributes of all nodes in the risk subgraph through an independent multi-layer perception machine to obtain environmental context features. The gated fusion unit is configured to dynamically weight and fuse multi-source features, specifically by controlling the contribution degree of the temporal context features and the environmental context features through a gating weight, weighting and fusing the subgraph coding features and the temporal context features to obtain final fusion features. The risk level division unit is configured to generate a dynamic risk grading threshold and divide the risk level, specifically by calculating the degree of deviation of the final fusion features from the normal benchmark through Mahalanobis distance, comparing it with the dynamic risk grading threshold to divide the risk level, and obtaining a risk grading label, including: Historical normal reference set construction is configured to establish a benchmark distribution of normal processes, specifically by extracting subgraph instances isomorphic to the current risk subgraph from training data with a risk-free label, and inputting the subgraph instances into the risk subgraph coding unit, the temporal context coding unit, the environmental context coding unit and the gated fusion unit in turn to obtain the final fusion features of each subgraph instance, which constitutes the historical normal reference set. Calculating the mean and covariance is configured to describe the statistical distribution of normal data, specifically by calculating the mean and covariance matrix of the final fusion features of each subgraph instance in the historical normal reference set to obtain the parameters of the normal distribution. Calculating the Mahalanobis distance is configured to measure the degree of deviation of the current risk from the normal state, specifically by calculating the Mahalanobis distance based on the mean and covariance matrix of the final fusion features of each subgraph instance in the historical normal reference set, and taking it as the risk score. Dynamic threshold division is configured to determine the risk level boundary, specifically by inputting training data with a low-risk label into the risk subgraph coding unit, the temporal context coding unit, the environmental context coding unit and the gated fusion unit in turn, calculating the risk score, obtaining a low-risk score set, and taking the 60th percentile of the low-risk score set as the first dynamic risk grading threshold, and doing the same for training data with a medium-risk label to obtain a medium-risk score set, and taking the 85th percentile of the medium-risk score set as the second dynamic risk grading threshold. The current risk score is compared with the first dynamic risk grading threshold and the second dynamic risk grading threshold to divide the risk level. The model construction training unit is configured to integrate the risk subgraph coding unit, the temporal context coding unit, the environmental context coding unit, the gated fusion unit and the risk level division unit to construct a deep learning model, train the model based on the grading training set and the grading test set and verify the performance, and obtain a deep learning model as a risk grading model.

6. The artificial intelligence-based medical waste supervision process risk early warning system according to claim 1, characterized in that: The past risk warning original data set and the real-time risk warning original data set both include static data, dynamic time series data and process relationship data of each medical waste supervision process node, and the past risk warning original data set further includes risk grading label data.

7. The artificial intelligence-based medical waste supervision process risk early warning system according to claim 1, characterized in that: In the data preliminary processing module, the multi-source data cleaning is used to ensure data quality and consistency, specifically for time stamp alignment, missing value interpolation and cleaning processing on the collected multi-source data, and unifying data format; The graph sequence construction is used to convert continuous data stream into discrete graph sequence, specifically for constructing static graph through fixed time window slicing, each slice constructing a static graph, the node of the static graph representing the process node, the edge being constructed based on process relationship data, then a virtual node being constructed for each static graph, and a bidirectional connection being established between the virtual node and all nodes, and the edge weight between the virtual node and each node being a normalized value of the operation times of the process node represented by the node within the time slice, and all slices of the static graph constituting a graph sequence; The construction of timing attribute graph is used to form a complete timing attribute graph, specifically for combining static graphs in time sequence, and retaining node attributes, edge attributes and time stamps to obtain a timing attribute graph; The data standardization is used to normalize data scale to eliminate dimension influence, specifically for processing node and edge attributes through Z-score standardization method to obtain a standardized timing attribute graph; The data set segmentation is used to obtain training data and test data, specifically for data set segmentation on the past risk early warning original data set; Through the multi-source data cleaning, the graph sequence construction, the construction of timing attribute graph and the data standardization, the real-time risk early warning original data set is subjected to data preliminary processing to obtain an early warning data set, and through the multi-source data cleaning, the graph sequence construction, the construction of timing attribute graph, the data standardization and the data set segmentation, the past risk early warning original data set is subjected to data preliminary processing to obtain an early warning training set and an early warning test set.

Images