Infusion detection system based on Internet

Abstract

The invention discloses an internet-based infusion detection system, which relates to the technical field of infusion detection, and comprises an equipment binding module for establishing an electronic file of a patient, binding intelligent infusion equipment, a health bracelet and an infrared thermal image module for the patient, and generating a patient equipment binding mapping table and initial filing information of the patient; the fusion module is used for establishing a data transmission channel of the intelligent infusion equipment, the health bracelet and the infrared thermal image module through an internet platform, and forming a multi-modal original data flow through physical sign collection of the health bracelet, image collection of the infrared thermal image module and dripping speed and pipeline pressure collection of the intelligent infusion equipment. According to the method, the multi-factor identity verification is performed on the electronic archive of the patient, the patient equipment binding mapping table is constructed in combination with the registration equipment information, and the initial archiving information of the patient is generated, so that accurate binding of the identity of the patient and a plurality of monitoring equipment is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of infusion detection, and in particular to an Internet-based infusion detection system. Background Art

[0002] In the clinical nursing process, infusion is one of the common treatment methods, and its safety and real-time monitoring needs are increasingly valued. Traditional infusion monitoring methods mainly rely on nurses' regular inspections or use a single intelligent device within the local area network to collect data. Pressure sensors are used to detect pressure changes in the infusion pipeline or infrared photoelectric detectors are used to monitor the drip rate. Vital sign collection equipment is used to obtain the patient's basic vital sign parameters. Data is collected in real time through sensors and processed locally to determine whether there are any abnormalities in the infusion, such as blood return, leakage or abnormal drip rate. The sensor data is uploaded to the hospital information system or nursing information platform through the local area network to achieve preliminary information integration and alarm prompts, providing auxiliary support for clinical work. It has achieved certain results in actual applications and played an important role in improving infusion monitoring efficiency and ensuring patient safety.

[0003] However, conventional infusion monitoring methods generally have two limitations. At the device level, traditional systems are usually limited to monitoring patients in a single dimension and lack the ability to integrate multi-device and multi-modal information, making it difficult to form a global perception and dynamic prediction of the infusion status. In terms of data processing methods, they often rely on local processing or static threshold judgments, making it difficult to achieve dynamic risk prediction and remote linkage control based on historical trends. Therefore, there is still room for improvement in the timeliness and accuracy of infusion abnormality warnings. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an Internet-based infusion detection system to solve the problem that multimodal device fusion and dynamic risk prediction based on historical trends are difficult to achieve.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions: The present invention provides an Internet-based infusion detection system, which includes a device binding module, which establishes a patient electronic file, binds the patient's smart infusion device, health bracelet and infrared thermal imaging module, and generates a patient-device binding mapping table and the patient's initial file information; The fusion module establishes a data transmission channel between the smart infusion device, the health bracelet, and the infrared thermal imaging module through the Internet platform. By collecting vital signs from the health bracelet, images from the infrared thermal imaging module, and drip rate and pipeline pressure from the smart infusion device, a multimodal raw data stream is formed. The multimodal raw data stream is transmitted to the Internet platform for format fusion processing to form a multimodal time series data packet. The risk prediction module calculates the multimodal time series data packets using a time series neural network model deployed in the cloud to generate risk prediction labels, which are then sent to the smart infusion device for operation and alarm information generation. The log recording module pushes the alarm information to the nurse station terminal interface, obtains prompt information, triggers the nurse to intervene on site based on the prompt information, and forms an intervention processing log record; The archiving module processes the log records based on the intervention and uploads them to the Internet platform to form a closed-loop log data.

[0007] As a preferred solution of the Internet-based infusion detection system of the present invention, the following steps are included: establishing a patient electronic file, binding the patient with a smart infusion device, a health bracelet and an infrared thermal imaging module, generating a patient-device binding mapping table and the patient's initial file information, Verify the patient's electronic file through the multi-factor authentication module to obtain the patient's identity verification result, and upload it to the patient's bound smart infusion device, health bracelet and infrared thermal imaging module to obtain the registered device information; The patient authentication result and the registered device information are mapped through the graph database to obtain the patient-device binding mapping table; The patient identity verification results are integrated with the smart infusion device, health bracelet and infrared thermal imaging module to generate the patient's initial file information.

[0008] As a preferred solution of the Internet-based infusion detection system of the present invention, a data transmission channel among the intelligent infusion device, the health bracelet and the infrared thermal imaging module is established through the Internet platform, including the following steps: Based on the patient-device binding mapping table, the registered device information is triggered through the Internet platform to dynamically register the smart infusion device, health bracelet and infrared thermal imaging module; By configuring the asynchronous message middleware structure on the Internet platform, an asynchronous data transmission channel for the smart infusion device, health bracelet and infrared thermal imaging module is obtained.

[0009] As a preferred solution of the Internet-based infusion detection system described in the present invention, wherein: a multimodal raw data stream is formed by collecting vital signs of a health bracelet, images of an infrared thermal imaging module, and drip rate and pipeline pressure of an intelligent infusion device, including the following steps: The multi-protocol adaptation layer is activated through the Internet platform to access and convert the data of the smart infusion equipment, health bracelet and infrared thermal imaging module protocol into a unified internal format. The unified format data enters the streaming processing engine, performs aggregation and semantic fusion within the time window, and forms a multimodal original data stream.

[0010] As a preferred solution of the Internet-based infusion detection system of the present invention, the multimodal raw data stream is transmitted to the Internet platform for format fusion processing to form a multimodal time series data packet, including the following steps: Use the multimodal timing alignment module in the Internet platform to perform unified timestamp standardization on the multimodal raw data stream to obtain time-aligned multimodal data; Use the cross-attention feature fusion engine to perform deep feature cross-fusion on the time-aligned multimodal data to obtain the fusion result; The fusion result is input into the self-attention temporal encoder to perform temporal modeling on the multimodal fusion features to form a structured sequence. The data packet builder is used to encapsulate the structured sequence into a multimodal time series data packet according to the set time window.

[0011] As a preferred solution of the Internet-based infusion detection system of the present invention, wherein: deploying a multimodal time series data packet in a time series neural network model on the cloud to perform calculations and generate risk prediction labels includes the following steps: Input the multimodal time series data packet into the graph modeling module to generate the relationship graph structure between the modalities; The relationship graph structure between modalities is processed by the graph convolution processing module, and the node structural features are extracted and passed to the temporal attention module. The temporal attention module performs temporal modeling on the structural features to form a temporal embedding table of the modal nodes; The temporal embedding table of the modal node is input into the fusion layer for unified encoding to generate the overall fusion feature; The fused features are passed to the prediction layer to generate risk prediction labels.

[0012] As a preferred solution of the Internet-based infusion detection system of the present invention, the risk prediction tag is sent to the intelligent infusion device to generate an alarm message, including the following steps: Input the risk prediction label into the label interpretation engine for decoding processing to obtain the control intention set representing the operation purpose; The control intention set representing the operation purpose is mapped into a behavior configuration table of the target intelligent infusion device according to the control intention, and the behavior configuration table is input into the execution engine of the intelligent infusion device to form a control instruction queue; Based on the control instruction queue, the status information and execution feedback information of the intelligent infusion device are collected and transmitted to the alarm judgment module; The alarm information is generated by comparing the alarm determination module state information and the execution feedback information.

[0013] As a preferred solution of the Internet-based infusion detection system of the present invention, wherein: pushing the alarm information to the terminal interface of the nurse station to obtain the prompt information includes the following steps: Encapsulate the alarm information into an asynchronous event object and input it into the asynchronous event bus; Match the corresponding nurse station terminal monitoring node on the event bus, and push the encapsulated alarm event to the corresponding nurse station terminal asynchronously through the event bus; The nurse station terminal receives the alarm event, extracts the alarm information through the local event decoding module, and obtains the prompt information.

[0014] As a preferred solution of the Internet-based infusion detection system of the present invention, wherein: triggering a nurse to perform on-site intervention based on prompt information and forming an intervention processing log record includes the following steps: The nurse station terminal receives the prompt information, triggering the intervention process initialization module to generate a unique intervention session identifier; Push the intervention session identifier and prompt information to the nurse's mobile terminal, activate the corresponding intervention task module in the mobile terminal, start the intervention state and enter the site to perform intervention operations, and obtain intervention process data; Mapping intervention process data to the intervention behavior model to generate a standardized intervention operation path; Generate intervention processing log records based on the intervention operation path and on-site intervention results.

[0015] As a preferred solution of the Internet-based infusion detection system of the present invention, wherein: based on the intervention processing log record, uploading to the Internet platform to form a closed-loop log data, including the following steps, Encapsulating intervention processing log records into log entries, performing compression encoding processing on the log entries, and generating identity-identified upload data units; The uploaded data unit of the identity identifier is uploaded to the log receiving interface of the Internet platform through a bandwidth-adaptive network protocol. The platform receives the uploaded data unit of the identity identifier, decodes and restores it, and generates the original log data; The restored log data is written into the closed-loop log database to form closed-loop log data.

[0016] The beneficial effects of the present invention are: by performing multi-factor identity authentication on the patient's electronic file, combining the registered device information to build a patient-device binding mapping table, and generating the patient's initial file information, the patient's identity is accurately bound to multiple monitoring devices, ensuring the accuracy of data attribution and the standardization of device management, thereby improving the overall collaborative efficiency of the system. The risk prediction module uses the time series neural network model deployed in the cloud to perform graph modeling, time series embedding and fusion coding processing on the multimodal time series data packets after format fusion, and finally generates a risk prediction label, realizing forward-looking identification of infusion abnormalities, and driving the downstream intelligent infusion equipment to perform operations through the label to complete automated response. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 Schematic diagram of the Internet-based infusion detection system.

[0019] Figure 2 Schematic diagram of the medium risk prediction label. DETAILED DESCRIPTION

[0020] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0021] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0022] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0023] Reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides an Internet-based infusion detection system, comprising the following steps: The device binding module establishes the patient's electronic file and binds the patient's smart infusion device, health bracelet and infrared thermal imaging module to generate the patient device binding mapping table and the patient's initial file information.

[0024] The patient's electronic file is verified through the multi-factor authentication module to obtain the patient's authentication result, which is then uploaded to the patient's bound smart infusion device, health bracelet and infrared thermal imaging module to obtain the registered device information.

[0025] Furthermore, in the process of verifying the patient's electronic file through the multi-factor authentication module, the patient's identity information and verification factors are collected. The verification factors may include facial feature information, biometric information, ID card number or medical record number multi-dimensional data. The multi-factor authentication module is used to perform a comprehensive comparison and analysis of the verification factors to confirm the authenticity and uniqueness of the patient's identity, generate the patient's identity verification results, and use the patient's identity verification results as input parameters to upload them to the patient-bound smart infusion device, health bracelet and infrared thermal imaging module to trigger each device to complete the binding operation for the patient, and return their respective registration status and identification information from the device side, including the device's unique identification, communication parameters and initialization status, etc., to obtain the registered device information.

[0026] The patient identity authentication result and the registered device information are mapped through the graph database to obtain the patient device binding mapping table.

[0027] Furthermore, after obtaining the patient identity authentication results and registered device information, the patient identity authentication results are first used as the patient node, and the device unique identifier in the registration information of each device is used as the smart infusion device node, health bracelet node and infrared thermal imaging module node respectively. The patient identity authentication results and the registered device information are corresponding one by one to establish structural nodes, and edge connections between the nodes are created in the graph database. The attributes of the edges include information such as device type, binding timestamp and communication parameters, completing the construction of the graph structure based on the graph database. Through the relationship query function of the graph database, the association relationship between the patient node and its corresponding device nodes is extracted from the node and its edge connection, and these one-to-many mapping relationships are parsed into table structure data to generate a patient-device binding mapping table, in which each record includes the patient identity authentication result and the corresponding device unique identifier, device type, binding time and other field information, which is used for subsequent management and call of the patient-device binding relationship. The patient identity verification results are integrated with the smart infusion device, health bracelet and infrared thermal imaging module to generate the patient's initial file information.

[0028] Furthermore, after the patient-device binding mapping table built based on the graph database is generated, the patient identity verification result is used as the core identity identification information, and the smart infusion device registration information, health bracelet registration information and infrared thermal imaging module registration information in the corresponding mapping relationship are successively integrated. During the fusion process, the identity identification field in the patient identity verification result is used as the primary key field, and the device unique identification, device type, and device registration time in the smart infusion device registration information, the device unique identification, physiological parameter collection capability, and registration time in the health bracelet registration information, and the device unique identification, temperature image resolution, and registration time in the infrared thermal imaging module registration information are matched and aggregated at the field level to form a structured information set, and the structured information set is organized into the patient's initial file information.

[0029] The fusion module establishes a data transmission channel for the smart infusion device, health bracelet and infrared thermal imaging module through the Internet platform. Through the vital sign collection of the health bracelet, the image collection of the infrared thermal imaging module, and the drip rate and pipeline pressure collection of the smart infusion device, a multimodal original data stream is formed. The multimodal original data stream is transmitted to the Internet platform for format fusion processing to form a multimodal time series data packet.

[0030] Establish data transmission channels for smart infusion equipment, health bracelets and infrared thermal imaging modules through the Internet platform.

[0031] Based on the patient-device binding mapping table, the dynamic registration of smart infusion devices, health bracelets and infrared thermal imaging modules is triggered through the Internet platform to register device information.

[0032] Furthermore, after the initial patient file information is generated, based on the patient device binding mapping table, the Internet platform uses the identity identification field in the patient identity verification result as the trigger index to initiate dynamic registration requests for the corresponding smart infusion device registration information, health bracelet registration information, and infrared thermal imaging module registration information in the mapping table. During the dynamic registration process, the Internet platform calls the registration interface of the smart infusion device, health bracelet, and infrared thermal imaging module one by one according to the device unique identification field in the patient device binding mapping table. During the call process, the registration parameters including the patient identity verification result, device type, and registration time are submitted. Each registration call completes the binding confirmation between the device and the patient's identity. The registration response status and confirmation time will be returned to the Internet platform in real time and will be synchronously updated to the patient device binding mapping table to ensure the uniqueness, validity, and real-time nature of the device registration process. Finally, the dynamic registration process of all devices is completed through the Internet platform, and the activation status of the device in the patient's file information is initialized.

[0033] By configuring the asynchronous message middleware structure on the Internet platform, an asynchronous data transmission channel for the smart infusion device, health bracelet and infrared thermal imaging module is obtained.

[0034] Furthermore, after completing the dynamic registration of the smart infusion device, health bracelet and infrared thermal imaging module, the asynchronous message middleware is configured through the Internet platform, and the data topics are defined with the smart infusion device, health bracelet and infrared thermal imaging module as data sources respectively. The Internet platform creates independent data topics corresponding to the smart infusion device, health bracelet and infrared thermal imaging module according to the device type field of each type of registered device, and constructs a logical channel for receiving and publishing device data. Each data topic is configured with a unique identifier and subscription strategy to ensure source traceability when data is published and consumption certainty when data is received. The unique identification field of each type of device is bound to the corresponding data topic to achieve a one-to-one correspondence between the device and the data channel. The asynchronous message middleware is used to establish an asynchronous transmission channel for smart infusion device data, an asynchronous transmission channel for health bracelet data and an asynchronous transmission channel for infrared thermal imaging module data.

[0035] A multimodal raw data stream is formed by collecting vital signs from the health bracelet, images from the infrared thermal imaging module, and drip rate and pipeline pressure from the smart infusion device.

[0036] Furthermore, after establishing the asynchronous data transmission channels for the smart infusion device, the health wristband, and the infrared thermal imaging module, the health wristband is activated to collect vital signs from the patient, including basic vital signs such as heart rate, blood oxygen saturation, and body temperature. Simultaneously, the infrared thermal imaging module collects images of the patient, obtaining a thermal imaging image sequence containing temperature distribution information. The smart infusion device collects the drip rate and pressure in the infusion pipeline in real time during the infusion process. The health wristband collects vital sign information through its configured sensor unit, refreshing the data once within a preset time period. The infrared thermal imaging module collects images by outputting an image frame sequence using temperature image matrix encoding. The smart infusion device collects drip rate and pipeline pressure through its flow detector and pressure sensor, outputting corresponding real-time values ​​in parallel. The three types of devices transmit the collected content to the data receiving end of the internet platform through their corresponding asynchronous data transmission channels, forming a multimodal raw data stream containing vital sign information, infrared thermal imaging image sequences, and infusion parameters for unified fusion analysis in subsequent processing steps.

[0037] The multi-protocol adaptation layer is activated through the Internet platform to access and convert the data of the smart infusion equipment, health bracelet and infrared thermal imaging module protocol into a unified internal format. The unified format data enters the streaming processing engine, performs aggregation and semantic fusion within the time window, and forms a multimodal original data stream.

[0038] Furthermore, after completing the collection of the multimodal raw data streams of the smart infusion device, health bracelet, and infrared thermal imaging module and uploading them to the Internet platform via the data asynchronous transmission channel, the multi-protocol adaptation layer in the Internet platform is activated to receive the raw protocol data from the smart infusion device, health bracelet, and infrared thermal imaging module. The multi-protocol adaptation layer uses the protocol parsing component to identify the communication protocols used by the three types of devices one by one, including parameters such as the data frame structure, field encoding method, and timestamp identification method. The multi-protocol adaptation layer converts the fields in the protocol raw format into a unified internal format according to the predefined field mapping rules, ensuring that all raw data from the smart infusion device, health bracelet, and infrared thermal imaging module are organized according to the unified field structure. Data in a unified format enters the streaming processing engine on the Internet platform. The streaming processing engine aggregates data from different sources along the time axis based on the set sliding time window, ensuring that vital signs information, thermal imaging image data and infusion process parameters within the same time period are logically aligned. The aggregated data is further semantically fused in the streaming processing engine, that is, based on the field meaning and collection source, the semantic correspondence between heart rate, body temperature, image temperature matrix, drip rate and pressure value and other contents is established between cross-source data, forming a multi-modal original data stream with complete structure, synchronized time sequence and consistent semantics.

[0039] The multimodal raw data stream is transmitted to the Internet platform for format fusion processing to form a multimodal time series data packet.

[0040] The multimodal timing alignment module in the Internet platform is used to perform unified timestamp standardization on the multimodal original data stream to obtain time-aligned multimodal data.

[0041] Furthermore, after completing the conversion of the data from the smart infusion device, health bracelet, and infrared thermal imaging module protocols into a unified internal format through the multi-protocol adaptation layer in the Internet platform, and performing time window aggregation and semantic fusion through the stream processing engine to form a multimodal raw data stream, the multimodal timing alignment module in the Internet platform is used to perform unified timestamp standardization on this multimodal raw data stream; The specific operation is that the multimodal time series alignment module extracts the local timestamp fields contained in various raw data of smart infusion devices, health bracelets and infrared thermal imaging modules, and uses the Internet platform server time as the reference time source to uniformly convert all timestamps through linear mapping or interpolation, so that the time information of different data sources has the same time zone, the same time format, and the same time granularity. Then, the records in the multimodal raw data stream are reordered according to the converted standard timestamps, and aligned to the same time node according to the set time step. All records from smart infusion devices, health bracelets and infrared thermal imaging modules must be aligned to the time node of every second, and the missing time node values ​​are supplemented by forward filling in the absence of data points. Finally, the multimodal data output after time alignment is completed to ensure that the records from smart infusion devices, health bracelets and infrared thermal imaging modules are synchronized in the same time dimension. The cross-attention feature fusion engine is used to perform deep feature cross-fusion on the time-aligned multimodal data to obtain the fusion result.

[0042] Furthermore, after completing the unified timestamp standardization processing of the multimodal raw data stream using the multimodal time series alignment module in the Internet platform and forming the time-aligned multimodal data, the cross-attention feature fusion engine is used to perform deep feature cross-fusion on the time-aligned multimodal data; The specific operation is to map the time-aligned multimodal data from smart infusion devices, health bracelets and infrared thermal imaging modules into equal-dimensional embedded feature vectors, for example, through convolutional neural networks or fully connected networks for feature encoding, to ensure that different modal inputs are comparable in the same dimensional space; then in the cross-attention feature fusion engine, the embedding vector of a certain modality is used as the query vector, and the embedding vectors of other modalities are used as keys and values, and the attention weight is calculated based on the standard attention mechanism, that is, the correlation between the features of each modality is measured by dot product, and the attention weight is applied to the weighted combination of feature information from different sources. The attention fusion operation as a query vector is performed on each modality embedding vector to complete the two-way cross fusion between all modalities, and the fused feature vectors of each modality are spliced ​​or weighted summed to output the fusion result.

[0043] The fusion result is input into the self-attention temporal encoder to perform temporal modeling on the multimodal fusion features to form a structured sequence. The data packet builder is used to encapsulate the structured sequence into a multimodal time series data packet according to the set time window.

[0044] Furthermore, the fusion results are arranged in chronological order to form a feature sequence and input into the self-attention temporal encoder; the self-attention temporal encoder calculates the correlation between the features of each time step in the sequence and the features of all time steps, dynamically adjusts the weights between time steps, captures long-term and short-term dependencies, and thus extracts the temporal structure information of the multimodal fusion features to form a structured sequence. Using a data packet builder, the structured sequence is divided into continuous time periods according to the preset time window, and the corresponding multimodal fusion features are encapsulated in each time period to generate a multimodal time series data packet.

[0045] It should be noted that the training of the self-attention temporal encoder uses the time-aligned multimodal fusion feature sequence as input during the training process, combined with the corresponding sequence annotation or prediction target as the supervision signal; adopts the standard sequence learning loss function, such as sequence cross entropy or mean square error, and updates the self-attention temporal encoder parameters through the backpropagation algorithm to optimize the temporal feature extraction capability; uses the validation set to monitor the model performance during the training iteration to ensure the accuracy and generalization ability of temporal modeling.

[0046] The risk prediction module deploys the multimodal time series data packets in the cloud-based time series neural network model for calculation, generates risk prediction labels, sends the risk prediction labels to the smart infusion device for operation, and generates alarm information.

[0047] The multimodal time series data packets are deployed in the cloud-based time series neural network model for operation to generate risk prediction labels.

[0048] The multimodal time series data packets are input into the graph modeling module to generate the relationship graph structure between the modalities.

[0049] Furthermore, the different modal features in the multimodal time series data packets are extracted as nodes in the graph structure. The edge weights between the nodes are calculated based on the correlation between the modal features, and the edge connection relationship is constructed to form a relational graph structure between the modalities. The graph modeling module reflects the temporal and semantic dependencies between the modalities by representing the attributes of the nodes and edges, and generates a relational graph structure that reflects the inherent connections between the multimodalities, providing a foundation for subsequent graph-based analysis and processing. It should be noted that the graph neural network training in the graph modeling module uses the constructed inter-modal relationship graph structure as input during training, combines the label information of the specific task as the supervision signal, adopts graph convolution or graph attention mechanism to extract graph structure features, optimizes the network parameters through gradient descent, improves the representation ability of nodes and edges in the relationship graph structure, and uses verification indicators to monitor model performance during training to ensure that the graph structure accurately reflects the correlation between multiple modalities.

[0050] The relationship graph structure between modalities is processed by the graph convolution module, and the node structural features are extracted and passed to the temporal attention module. The temporal attention module performs temporal modeling on the structural features to form a temporal embedding table of the modal nodes.

[0051] Furthermore, the relationship graph structure between modalities first passes through the graph convolution processing module, which performs weighted aggregation on the features of each node and its adjacent nodes in the graph, extracts the structural features of each node, and forms the contextual representation of the node. The structural features are then passed to the temporal attention module. The temporal attention module calculates the importance weights of features at different time steps based on the attention mechanism, dynamically models the node structural features in the time dimension, and thus generates a modal node temporal embedding table that reflects the time-varying characteristics of the node. It should be noted that the training of the graph convolution processing module adopts a graph convolutional network, using the features of nodes and edges in the multimodal relationship graph structure as input, and optimizing the network parameters through backpropagation to improve the expressiveness of node features. The training of the temporal attention module is based on the attention mechanism, using temporal features and corresponding labels for supervised learning, and enhancing the time series modeling effect by optimizing the attention weight distribution. During the training process, the model performance is monitored through cross-validation and loss function to ensure that the modal node temporal embedding table accurately captures the node temporal dynamic information. Specifically, the expression is, ; in, For the The output node structure feature matrix of the layer graph convolution processing module, For the The input node feature matrix of the layer graph convolution processing module, is the normalized adjacency matrix of the graph structure, For the The trainable weight matrix of the layer graph convolution processing module.

[0052] The temporal embedding table of the modal node is input into the fusion layer for unified encoding to generate the overall fusion features.

[0053] Furthermore, after the temporal embedding table of the modal node is input into the fusion layer, the fusion layer uniformly encodes the temporal features of each modal node. Through multi-layer nonlinear transformation and feature weighted integration, it extracts the correlation features of cross-modal temporal information, eliminates the differences between modalities, and fuses the node temporal features, ultimately generating an overall fusion feature containing multi-modal temporal information. It should be noted that the fusion layer adopts a multi-layer perception network, performs forward propagation through the input modal node time series embedding table, uses labeled training samples for supervised learning, optimizes the network weight parameters to improve the cross-modal fusion effect, and adopts the backpropagation algorithm in the training process, combined with the loss function to monitor the expression accuracy and generalization ability of the overall fusion features, to ensure that the overall fusion features output by the fusion layer can effectively reflect the multimodal time series relationship.

[0054] The fused features are passed to the prediction layer to generate risk prediction labels.

[0055] Furthermore, risk-related features are extracted through multi-layer nonlinear transformations, the probability distribution or risk value of each risk category is calculated, and finally a risk prediction label is output to complete the risk assessment of multimodal fusion information. The prediction layer processing includes feature mapping, activation function application, and probability normalization to ensure that the risk prediction label has high discrimination accuracy and stability. It should be noted that the prediction layer uses a feedforward neural network structure, employing supervised learning using labeled training samples, and optimizing network parameters using either a cross-entropy loss function or a mean squared error loss function. A backpropagation algorithm is used during training to adjust weights, ensuring the prediction layer accurately maps fused features to risk prediction labels, ensuring effective and reliable risk assessment.

[0056] Specifically, the expression is, ; in, is the risk prediction label, is the fusion feature vector, is the bias vector of the prediction layer,

[0057] is the weight matrix of the prediction layer.

[0058] The predicted tags are sent to the smart infusion device for operation and generate alarm information.

[0059] The risk prediction label is input into the label interpretation engine for decoding processing to obtain the control intention set that represents the operation purpose.

[0060] Furthermore, after the risk prediction label is input into the label interpretation engine, the label interpretation engine decodes the risk prediction label, converts the label information through mapping rules or pre-trained decoding algorithms, parses the corresponding operation purpose and control intention, and forms a control intention set that represents the operation purpose, thereby achieving semantic understanding of the risk prediction results and expression of target instructions, ensuring that the decoding results accurately reflect the meaning of the risk assessment and are executable.

[0061] It should be noted that the label interpretation engine adopts a rule-based decoding method or combines it with a sequence decoding network, uses annotated risk labels and corresponding control intention samples for training, and applies a sequence-to-sequence learning algorithm to optimize the decoding accuracy. The training process uses a cross-entropy loss function and adjusts parameters through backpropagation to improve the label interpretation engine's decoding performance for risk prediction labels.

[0062] According to the control intention, the control intention set representing the operation purpose is mapped into the behavior configuration table of the target intelligent infusion device, and the behavior configuration table is input into the execution engine of the intelligent infusion device to form a control instruction queue.

[0063] Furthermore, based on the control intent set representing the operation purpose, it is mapped into a behavior configuration table of the target intelligent infusion device. Through preset mapping rules or configuration files, the control intent is converted into corresponding behavior instruction parameters, and the execution actions and sequence are clearly defined to ensure that the behavior configuration table is complete and meets the functional requirements of the intelligent infusion device. The behavior configuration table is passed as input to the execution engine of the intelligent infusion device. The execution engine parses the various instruction parameters in the behavior configuration table and generates a control instruction queue in sequence. It should be noted that the mapping process uses rule-based matching or machine learning classification algorithms, using labeled control intent and corresponding behavior configuration table datasets for training to optimize mapping accuracy. The execution engine's instruction parsing module generates instruction queues through state machine logic.

[0064] Based on the control instruction queue, the status information and execution feedback information of the intelligent infusion device are collected and transmitted to the alarm judgment module.

[0065] Furthermore, based on the control instruction queue, the intelligent infusion device collects operating status information and execution feedback information in real time. The status information includes parameters such as the device working status, liquid infusion rate, and remaining liquid volume. The execution feedback information reflects the response results and abnormal conditions of the device's execution instructions. After the collection is completed, the status information and execution feedback information are transmitted to the alarm judgment module through the communication interface.

[0066] The alarm information is generated by comparing the alarm determination module state information and the execution feedback information.

[0067] Furthermore, the alarm determination module compares the status information collected by the intelligent infusion device with the execution feedback information item by item. Based on preset thresholds or rules, it determines whether there are any abnormalities in the device operation. Specifically, it determines whether the infusion rate deviates from the set range, whether the device response is delayed or erroneous. During this comparison process, the alarm determination module generates corresponding alarm information through logical judgment and conditional screening.

[0068] The log recording module pushes the alarm information to the nurse station terminal interface, obtains prompt information, triggers the nurse to perform on-site intervention based on the prompt information, and forms an intervention processing log record.

[0069] Push the alarm information to the nurse station terminal interface to obtain prompt information.

[0070] Furthermore, alarm information is sent to the nurse station terminal interface via the communication module, which converts the alarm format and processes the transmission protocol to ensure complete and real-time information delivery. Upon receiving the alarm information, the nurse station terminal interface generates corresponding prompts based on the alarm type and priority. These prompts include audio prompts, icon displays, and text descriptions. These prompts accurately reflect the alarm content and prompt nursing staff to respond promptly.

[0071] Encapsulate the alarm information into an asynchronous event object and input it into the asynchronous event bus.

[0072] Furthermore, the alarm information is encapsulated into an asynchronous event object through the encapsulation module. The encapsulated content includes key fields such as the type, timestamp and detailed description of the alarm information. The asynchronous event object is sent to the asynchronous event bus. The asynchronous event bus is responsible for managing the queuing and scheduling of events, ensuring that the asynchronous event object can be delivered to the corresponding processing unit according to the set priority and order, thereby realizing asynchronous processing and response of the alarm information.

[0073] The corresponding nurse station terminal monitoring node is matched on the event bus, and the encapsulated alarm event is asynchronously pushed to the corresponding nurse station terminal through the event bus.

[0074] Furthermore, the asynchronous event bus matches the corresponding nurse station terminal listening node based on the event type and target identifier in the alarm event object, confirming the target terminal's subscription relationship. Once matched, the asynchronous event bus asynchronously pushes the encapsulated alarm event to the corresponding nurse station terminal listening node, enabling real-time transmission and notification of the alarm event. During the push process, the asynchronous event bus ensures the reliability and sequentiality of event transmission, ensuring that the nurse station terminal can receive and process the alarm event in a timely manner.

[0075] The nurse station terminal receives the alarm event, extracts the alarm information through the local event decoding module, and obtains the prompt information.

[0076] Furthermore, after the nurse station terminal receives the alarm event pushed by the asynchronous event bus, it parses the alarm event through the local event decoding module and extracts the specific content and related parameters in the alarm information. The event decoding module structures the alarm event according to the predefined event format to ensure the integrity and accuracy of the alarm information. The extracted alarm information is converted into prompt information for display on the nurse station terminal interface and subsequent processing.

[0077] The prompt information triggers the nurse to perform on-site intervention and form an intervention processing log record.

[0078] The nurse station terminal receives the prompt information, triggering the intervention process initialization module to generate a unique intervention session identifier.

[0079] Furthermore, upon receiving the prompt, the nurse station terminal triggers the intervention process initialization module to generate a unique intervention session identifier. By invoking a unique identifier generation algorithm and combining it with information such as the current timestamp and the terminal's unique identifier, the generated intervention session identifier is guaranteed to be unique and traceable system-wide. This unique intervention session identifier serves as a key identifier for subsequent intervention processes.

[0080] The intervention session identifier and prompt information are pushed to the nurse's mobile terminal, the corresponding intervention task module is activated in the mobile terminal, the intervention state is started, the intervention operation is performed on site, and the intervention process data is obtained.

[0081] Furthermore, the intervention session identifier and prompt information are pushed to the nurse's mobile terminal through the communication interface. After receiving the information, the nurse's mobile terminal activates the corresponding intervention task module. The intervention task module starts the intervention state according to the intervention session identifier. After the intervention state is started, the nursing staff enters the scene to perform the intervention operation, and the intervention task module collects and records the intervention process data in real time.

[0082] Map the intervention process data to the intervention behavior model to generate a standardized intervention operation path.

[0083] Furthermore, intervention process data is mapped one-to-one to standard action nodes in the intervention behavior model using predefined mapping rules, transforming the actual execution sequence and status into a standardized intervention operation path. This mapping process includes data format conversion, time series sorting, and key action matching, ensuring the continuity and integrity of the intervention process is accurately reflected. The resulting standardized intervention operation path can be used for subsequent evaluation and analysis.

[0084] Generate intervention processing log records based on the intervention operation path and on-site intervention results.

[0085] Furthermore, based on the intervention operation path, the specific content, timestamp and executor information of each operation step are summarized with the on-site intervention results, and recorded one by one according to the predefined log format to generate an intervention processing log record.

[0086] The archiving module processes the log records based on the intervention and uploads them to the Internet platform to form a closed-loop log data.

[0087] Based on the intervention processing log records, upload them to the Internet platform to form closed-loop log data.

[0088] The intervention processing log record is encapsulated as a log entry, compression encoding processing is performed on the log entry, and an upload data unit with an identity identifier is generated.

[0089] Furthermore, intervention log records are encapsulated as log entries, and the log content is structured according to a predetermined format. Then, the log entries are processed using a compression encoding algorithm to reduce data size and improve transmission efficiency. After compression encoding, the log entries are bound to a unique identifier to generate an uploaded data unit containing the identifier.

[0090] The uploaded data unit of the identity identifier is uploaded to the log receiving interface of the Internet platform through a bandwidth-adaptive network protocol. The platform receives the uploaded data unit of the identity identifier, decodes and restores it, and generates original log data.

[0091] Furthermore, the uploaded data unit of the identity identification is transmitted through a bandwidth-adaptive network protocol. The network protocol adjusts the transmission rate and data packetization method according to the current bandwidth conditions to ensure that the upload process is stable and efficient. After the log receiving interface of the Internet platform receives the uploaded data unit, it performs decoding processing to restore the compressed encoded data to the original log content. The restored log content is then checked for consistency to verify the integrity and accuracy of the log content, ensuring that the log information processed subsequently is true and reliable.

[0092] The restored log data is written into the closed-loop log database to form closed-loop log data.

[0093] Furthermore, after the restored log data is parsed and formatted, it is written into the closed-loop log database one by one according to the preset data structure and field requirements, ensuring that the log content is stored completely and accurately in the database, forming structured closed-loop log data for subsequent retrieval, analysis and processing, while supporting the persistent storage and management of log data.

[0094] In summary, the present invention performs multi-factor identity authentication on the patient's electronic file, builds a patient-device binding mapping table based on the registered device information, and generates the patient's initial file information, thereby achieving accurate binding of the patient's identity with multiple monitoring devices, ensuring the accuracy of data attribution and the standardization of device management, and thus improving the overall collaborative efficiency of the system. The risk prediction module uses the time series neural network model deployed in the cloud to perform graph modeling, time series embedding and fusion coding processing on the multimodal time series data packets after format fusion, and finally generates risk prediction labels, realizing forward-looking identification of infusion abnormalities, and driving downstream intelligent infusion devices to perform operations through labels to complete automated response.

[0095] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An Internet-based infusion detection system, characterized by: include, The device binding module establishes the patient's electronic file and binds the patient's smart infusion device, health bracelet and infrared thermal imaging module to generate the patient device binding mapping table and the patient's initial file information; The fusion module establishes a data transmission channel between the smart infusion device, the health bracelet, and the infrared thermal imaging module through the Internet platform. By collecting vital signs from the health bracelet, images from the infrared thermal imaging module, and drip rate and pipeline pressure from the smart infusion device, a multimodal raw data stream is formed. The multimodal raw data stream is transmitted to the Internet platform for format fusion processing to form a multimodal time series data packet. The risk prediction module calculates the multimodal time series data packets using a time series neural network model deployed in the cloud to generate risk prediction labels, which are then sent to the smart infusion device for operation and alarm information generation. The log recording module pushes the alarm information to the nurse station terminal interface, obtains prompt information, triggers the nurse to intervene on site based on the prompt information, and forms an intervention processing log record; The archiving module processes the log records based on the intervention and uploads them to the Internet platform to form a closed-loop log data.

2. The Internet-based infusion detection system according to claim 1, wherein: Establishing a patient electronic file, binding the patient with a smart infusion device, health bracelet, and infrared thermal imaging module, generating a patient device binding mapping table and initial patient file information, includes the following steps: Verify the patient's electronic file through the multi-factor authentication module to obtain the patient's identity verification result, and upload it to the patient's bound smart infusion device, health bracelet and infrared thermal imaging module to obtain the registered device information; The patient authentication result and the registered device information are mapped through the graph database to obtain the patient-device binding mapping table; The patient identity verification results are integrated with the smart infusion device, health bracelet and infrared thermal imaging module to generate the patient's initial file information.

3. The Internet-based infusion detection system according to claim 2, wherein: Establish data transmission channels for smart infusion equipment, health bracelets and infrared thermal imaging modules through the Internet platform. The following steps are included: Based on the patient-device binding mapping table, the registered device information is triggered through the Internet platform to dynamically register the smart infusion device, health bracelet and infrared thermal imaging module; By configuring the asynchronous message middleware structure on the Internet platform, an asynchronous data transmission channel for the smart infusion device, health bracelet and infrared thermal imaging module is obtained.

4. The Internet-based infusion detection system according to claim 3, wherein: Through the collection of vital signs from the health bracelet, image collection from the infrared thermal imaging module, and drip rate and pipeline pressure from the smart infusion device, a multi-modal raw data stream is formed. The following steps are included: The multi-protocol adaptation layer is activated through the Internet platform to access and convert the data of the smart infusion equipment, health bracelet and infrared thermal imaging module protocol into a unified internal format. The unified format data enters the streaming processing engine, performs aggregation and semantic fusion within the time window, and forms a multimodal original data stream.

5. The Internet-based infusion detection system according to claim 4, wherein: The multimodal raw data stream is transmitted to the Internet platform for format fusion processing to form a multimodal time series data packet, including the following steps: Use the multimodal timing alignment module in the Internet platform to perform unified timestamp standardization on the multimodal raw data stream to obtain time-aligned multimodal data; Use the cross-attention feature fusion engine to perform deep feature cross-fusion on the time-aligned multimodal data to obtain the fusion result; The fusion result is input into the self-attention temporal encoder to perform temporal modeling on the multimodal fusion features to form a structured sequence. The data packet builder is used to encapsulate the structured sequence into a multimodal time series data packet according to the set time window.

6. The Internet-based infusion detection system according to claim 5, wherein: The multimodal time series data packets are deployed in the cloud-based time series neural network model for operation to generate risk prediction labels, including the following steps: Input the multimodal time series data packet into the graph modeling module to generate the relationship graph structure between the modalities; The relationship graph structure between modalities is processed by the graph convolution processing module, and the node structural features are extracted and passed to the temporal attention module. The temporal attention module performs temporal modeling on the structural features to form a temporal embedding table of the modal nodes; The temporal embedding table of the modal node is input into the fusion layer for unified encoding to generate the overall fusion feature; The fused features are passed to the prediction layer to generate risk prediction labels.

7. The Internet-based infusion detection system according to claim 6, wherein: Send risk prediction tags to smart infusion equipment to operate and generate alarm information. The following steps are included: Input the risk prediction label into the label interpretation engine for decoding processing to obtain the control intention set representing the operation purpose; The control intention set representing the operation purpose is mapped into a behavior configuration table of the target intelligent infusion device according to the control intention, and the behavior configuration table is input into the execution engine of the intelligent infusion device to form a control instruction queue; Based on the control instruction queue, the status information and execution feedback information of the intelligent infusion device are collected and transmitted to the alarm judgment module; The alarm information is generated by comparing the alarm determination module state information and the execution feedback information.

8. The Internet-based infusion detection system according to claim 7, wherein: Push the alarm information to the nurse station terminal interface and get the prompt information, including the following steps: Encapsulate the alarm information into an asynchronous event object and input it into the asynchronous event bus; Match the corresponding nurse station terminal monitoring node on the event bus, and push the encapsulated alarm event to the corresponding nurse station terminal asynchronously through the event bus; The nurse station terminal receives the alarm event, extracts the alarm information through the local event decoding module, and obtains the prompt information.

9. The Internet-based infusion detection system according to claim 8, wherein: Based on the prompt information, the nurse is triggered to perform on-site intervention and form an intervention processing log record, which includes the following steps: The nurse station terminal receives the prompt information, triggering the intervention process initialization module to generate a unique intervention session identifier; Push the intervention session identifier and prompt information to the nurse's mobile terminal, activate the corresponding intervention task module in the mobile terminal, start the intervention state and enter the site to perform intervention operations, and obtain intervention process data; Mapping intervention process data to the intervention behavior model to generate a standardized intervention operation path; Generate intervention processing log records based on the intervention operation path and on-site intervention results.

10. The Internet-based infusion detection system according to claim 9, wherein: Based on the intervention processing log records, upload them to the Internet platform to form closed-loop log data, including the following steps: Encapsulating intervention processing log records into log entries, performing compression encoding processing on the log entries, and generating identity-identified upload data units; The uploaded data unit of the identity identifier is uploaded to the log receiving interface of the Internet platform through a bandwidth-adaptive network protocol. The platform receives the uploaded data unit of the identity identifier, decodes and restores it, and generates the original log data; The restored log data is written into the closed-loop log database to form closed-loop log data.

Images