Emergency rescue record and flow chart linkage system

Abstract

The invention discloses an emergency rescue record and flow chart linkage system, which comprises a data acquisition and preprocessing module used for acquiring structured record data and multi-source data and completing preprocessing to generate an event time axis data set; the process modeling rule module is used for generating a node mapping relation and a node rule set; the improved GhostNet feature extraction module is used for carrying out feature extraction and outputting a field feature vector; the extreme random tree judgment module is used for fusing the field feature vector and the structured record data and outputting a judgment result; the flow linkage execution module is used for updating a flow chart node state and outputting next node prompt information; and the sexual verification trace reserving module is used for carrying out consistency verification and trace reserving recording and generating traceable quality control data. According to the invention, through combination of the improved GhostNet network and the extreme random tree algorithm, real-time synchronous traceability of emergency rescue records and process nodes is realized.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to an emergency resuscitation record and flowchart linkage system. Background Technology

[0002] In current emergency resuscitation scenarios, the resuscitation process is characterized by high concurrency, multi-role collaboration, parallel treatment steps, and strict time windows for critical operations. Resuscitation records are typically entered by doctors, nurses, and recorders on different terminals. The recorded content covers vital signs, monitoring waveforms, key events, medication and treatment procedures, examination and test results, medical orders, and important time points. Due to the fast pace of resuscitation, the dispersed information sources, and the high frequency of data generation, medical staff often prioritize completing the treatment and then supplement the records afterward. This leads to frequent problems such as missing record items, inconsistent record granularity, inaccurate time point backfilling, and inconsistencies between the event sequence and the actual treatment sequence. At the same time, parallel entry on multiple terminals can easily result in duplicate records, field conflicts, version overwriting, and inconsistencies between the operator and the timestamp. Consequently, it is difficult to form a complete and continuous event timeline in resuscitation records, affecting the integrity of medical records, cost and quality control verification, resuscitation quality assessment, and the reliability and traceability of evidence in medical disputes.

[0003] Existing systems often treat resuscitation flowcharts, clinical pathways, and resuscitation records as relatively independent functional modules. Flowcharts are typically statically displayed or provided with simple prompts, lacking clear mapping relationships between flowchart nodes and structured record fields, node triggering conditions, and calculable rules for node status. This makes it difficult to achieve record-driven process advancement and process-driven record template generation. For multi-source field data such as vital signs data, monitoring waveform data, equipment operation events, and bedside imaging data, existing systems mostly only display them in a scattered manner or save them as attachments, lacking automatic association with structured records. They cannot provide real-time assessment and prompts regarding the completion status of key nodes, branch path selection, missing items, and timeout risks during resuscitation. They also lack a systematic verification and record-keeping mechanism for the consistency between process node status and records, making it difficult to ensure process execution consistency. Quality control relies more on post-event manual spot checks, which fails to meet the comprehensive requirements of real-time synchronization, traceability, and quality control in emergency resuscitation.

[0004] Therefore, how to provide an emergency rescue record and flowchart linkage system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an emergency resuscitation record and flowchart linkage system. This invention comprehensively utilizes event timeline construction, flowchart node rule modeling, multi-source on-site data and structured record data fusion analysis, and artificial intelligence judgment technology. It details the process of collecting structured record data and multi-source on-site data during emergency resuscitation, generating an event timeline dataset after preprocessing, constructing a resuscitation flowchart and establishing node mapping relationships and triggering conditions, using an improved GhostNet network to output on-site feature vectors and fusing them with structured record data, then outputting judgment results through an extreme random tree algorithm, and finally driving the flowchart node state based on the judgment results. The system updates the status and outputs the next node prompts, and performs consistency verification, record keeping, and traceable quality control data generation for structured record data and flowchart node status. The improved GhostNet network innovatively adopts multi-scale Ghost feature generation, lightweight attention fusion, and dynamic Ghost-ratio strategy. The extreme random tree algorithm innovatively adopts structured shared subtrees and hierarchical shuffling of sample order training strategy to achieve bidirectional linkage between records and flowcharts in high-concurrency and time-sensitive emergency scenarios, reduce omissions and ensure process consistency. At the same time, it achieves real-time synchronization, traceability, and quality control of records and process status under multi-terminal collaboration.

[0006] An emergency resuscitation record and flowchart linkage system according to an embodiment of the present invention includes the following modules:

[0007] The data acquisition and preprocessing module is used to acquire structured recorded data and multi-source field data, and to preprocess and generate an event timeline dataset.

[0008] The process modeling rules module is used to construct the rescue process diagram and generate node mapping relationships and node rule sets;

[0009] An improved GhostNet feature extraction module was developed to extract features from an event timeline dataset and output on-site feature vectors.

[0010] The extreme random tree decision module is used to fuse on-site feature vectors with structured recorded data and output the decision result;

[0011] The process linkage execution module is used to update the status of flowchart nodes and output the next node prompt information based on the judgment result and node rule set;

[0012] The consistency verification and logging module is used to verify the consistency between structured record data and the status of nodes in the rescue flowchart, log the data, and generate traceable quality control data.

[0013] Optionally, modules can be integrated using the following methods:

[0014] Structured record data and multi-source on-site data were collected during the emergency rescue process and preprocessed to generate an event timeline dataset.

[0015] A flowchart is constructed based on the rescue process template, and the mapping relationship between flowchart nodes and structured record fields and node triggering conditions are established to generate a node rule set.

[0016] An improved GhostNet network is constructed. The input event timeline dataset is used, and Ghost feature maps with different receptive fields are generated using multi-scale Ghost features. Selective enhancement is performed by fusing lightweight attention. A dynamic Ghost-ratio strategy is used to generate a ratio based on the input data type and number of channels, and the output is a live feature vector.

[0017] The on-site feature vectors and structured recorded data are fused together, and the judgment is made based on the extreme random tree algorithm. The structure is shared by multiple trees to share the previous split structure. According to the hierarchical random sample order training strategy, the label distribution is randomly shuffled in a hierarchical manner before training each tree, and the judgment result is output.

[0018] Based on the judgment result and the node rule set, the record-driven flowchart linkage processing is executed to update the flowchart node status and trigger the output of the next node prompt information.

[0019] Based on the event timeline dataset, the consistency of structured record data and flowchart node status is verified and recorded, and traceable quality control data is generated.

[0020] Optionally, the multi-source field data includes vital sign data, monitoring waveform data, equipment operation events, and bedside image data.

[0021] Optionally, the generated event timeline dataset includes:

[0022] The system collects structured record data entered by medical staff terminals during emergency resuscitation and multi-source field data output by monitoring equipment, treatment equipment and bedside acquisition equipment, and generates data source identifiers, event type identifiers and event times for the structured record data and the multi-source field data respectively.

[0023] The event time is converted into a timestamp under a unified time base, and the structured record data is processed by unifying the field format, field name, enumeration value encoding and missing field marking. At the same time, the sampling frequency of multi-source field data is unified, noise is filtered and time is aligned.

[0024] The multi-source field data is segmented according to time windows to generate data segments and segment start timestamps and segment end timestamps. Based on the timestamps, the structured record data is associated with the data segments to generate an event timeline dataset containing event sequences, timestamp sequences, data segment indexes, and a set of structured record fields.

[0025] Optionally, the generated node rule set includes:

[0026] Read the rescue process template and parse it to obtain the process node set, node connection relationship set and branch condition set. Generate a node identifier, node type, predecessor node identifier set and successor node identifier set for each process node, and generate a rescue process diagram based on the process node set and node connection relationship set.

[0027] Configure a set of structured record fields for each process node and generate a mapping relationship between the node and the structured record fields. At the same time, configure node trigger conditions for each process node and generate a node rule set. The node trigger conditions include entry trigger conditions, completion trigger conditions, and timeout trigger conditions. The node rule set includes node identifier, mapping field set, trigger condition set, and branch condition set.

[0028] Optionally, the output field feature vector includes:

[0029] An improved GhostNet network is constructed, including a feature extraction backbone network formed by stacking multiple Ghost blocks and an output layer. The event timeline dataset is converted into a network input tensor and then fed into the improved GhostNet network.

[0030] Multi-scale Ghost feature generation is performed in each Ghost block. Two depthwise separable convolutions with different kernel sizes are performed in parallel on the intrinsic feature map output by the main convolution to generate Ghost feature maps with different receptive fields. The intrinsic feature map and the Ghost feature map are then concatenated along the channel dimension to form the module output feature map.

[0031] Lightweight attention fusion is performed in each Ghost block. Global average pooling is performed on the intrinsic feature map to obtain the channel description vector. The channel weight vector is generated through two fully connected transformations. The channel weight vector is then multiplied with the intrinsic feature map channel by channel to obtain the selectively enhanced intrinsic feature map.

[0032] In the improved GhostNet network, a dynamic Ghost-ratio strategy is implemented. The ratio of the number of channels in the intrinsic feature map to the number of channels in the Ghost feature map is determined based on the input data type identifier and the number of input channels in the current layer. The number of output channels of the main convolution and the number of generated channels of the multi-scale depth separable convolution are configured according to the determined ratio.

[0033] Global average pooling is performed on the output feature map of the improved GhostNet network, and the field feature vector is obtained by linear transformation.

[0034] Optionally, the output determination result includes:

[0035] Extract field feature vectors from the structured record data in the event timeline dataset, and concatenate the on-site feature vectors with the field feature vectors according to the feature dimensions to generate a fused feature vector;

[0036] Based on the extreme random tree algorithm, multiple extreme random decision trees are generated, and the number of candidate features, maximum split depth, and stopping conditions are set for each extreme random decision tree.

[0037] Training with shared subtrees is adopted. Within a preset shared depth range, a shared split structure is constructed for multiple extreme random decision trees. After the shared depth is exceeded, a non-shared split structure is constructed for each extreme random decision tree.

[0038] A hierarchical shuffling sample order training strategy is adopted. Before training each extreme random decision tree, the training samples are divided into hierarchical layers according to the label category, and the sample order is randomly shuffled within each layer to generate the training sample sequence of the current tree.

[0039] Based on the training and prediction of fused feature vectors, each extreme random decision tree randomly selects features from candidate features at each split node and randomly generates a split threshold within the feature value range to complete the node split. In the prediction stage, the outputs of multiple extreme random decision trees are integrated to obtain the judgment result.

[0040] Optionally, the step of performing record-driven flowchart linkage processing based on the judgment result and node rule set, updating the flowchart node status and triggering the output of the next node prompt information includes:

[0041] Receive the judgment result and parse it to obtain the node identifier, node status, branch path identifier and missing item prompt set, and read the entry trigger condition, completion trigger condition and timeout trigger condition corresponding to the node identifier from the node rule set;

[0042] The flowchart node status is updated based on the judgment result, completion trigger condition and entry trigger condition, and the node entry timestamp and node completion timestamp are written into the event timeline dataset. At the same time, the successor node is determined according to the branch path identifier and the successor node status is updated.

[0043] Based on the timeout trigger condition, perform timeout judgment on the ongoing node and generate timeout prompt information. Based on the missing item prompt set, generate missing item prompt information and output the node status change result, next node prompt information, timeout prompt information and missing item prompt information to the flowchart interface.

[0044] Optionally, the consistency verification and record keeping of structured record data and flowchart node states based on the event timeline dataset, and the generation of traceable quality control data, include:

[0045] Receive the flowchart node selection operation and obtain the selected node identifier. Read the mapping field set, required field marker, field value range, entry trigger condition, completion trigger condition and timeout trigger condition corresponding to the node identifier from the node rule set.

[0046] A node record template is generated based on the mapping field set. The node record template is then matched with the corresponding structured record field values ​​that already exist in the event timeline dataset to fill in the existing field values ​​and mark the unfilled fields. At the same time, a set of record verification items and a set of time window prompt items are generated based on the entry trigger condition, completion trigger condition, and timeout trigger condition and written into the node record template.

[0047] It receives field entry operations for node record templates and generates structured record entries. It generates operator identifiers and entry timestamps for structured record entries and writes them into the event timeline dataset. After writing, it triggers the calculation of node trigger conditions and outputs the calculation results to the flowchart interface to update the node status and prompt information.

[0048] The beneficial effects of this invention are:

[0049] This invention preprocesses structured emergency rescue records and multi-source on-site data to construct an event timeline dataset. It then combines this with an emergency rescue process template to generate flowcharts and node rule sets, thus establishing a foundation for the linkage between records and flowcharts. An improved GhostNet network is used to perform lightweight feature extraction on the event timeline dataset and output on-site feature vectors. These vectors are then fused with structured record data, and an extreme random tree algorithm outputs the judgment results. This enables the system to perform real-time judgments on the completion status of key nodes, branch paths, and the risk of missing items in high-concurrency, time-sensitive emergency scenarios. Based on the judgment results, the system automatically updates the flowchart node status and outputs prompts for the next node, thereby reducing the probability of missed records and improving the consistency and continuity of process execution.

[0050] This invention uses an event timeline to span the entire process, performing consistency verification and record-keeping on changes in flowchart node status, record entry, and judgment results, forming traceable quality control data and enabling real-time synchronization and audit trails under multi-terminal collaboration. The improved GhostNet network employs multi-scale Ghost feature generation, lightweight attention fusion, and a dynamic Ghost-ratio strategy, enhancing the extraction capability of key on-site features while ensuring real-time performance. The extreme random tree algorithm uses a structure-shared subtree and hierarchical shuffling of sample order training strategies, improving the stability of judgments on samples from different process branches while maintaining inference efficiency. This enhances the completeness, traceability, and quality control efficiency of emergency rescue records, meeting the needs of rescue debriefing, quality management, and compliant record retention. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is a flowchart of an emergency resuscitation record and flowchart linkage system proposed in this invention;

[0053] Figure 2 This is a structural block diagram of an emergency resuscitation record and flowchart linkage method proposed in this invention;

[0054] Figure 3 This is a functional diagram of the improved GhostNet network for the emergency resuscitation record and flowchart linkage method proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figure 1 An emergency resuscitation record and flowchart linkage system includes the following modules:

[0057] The data acquisition and preprocessing module is used to acquire structured recorded data and multi-source field data, and to preprocess and generate an event timeline dataset.

[0058] The process modeling rules module is used to construct the rescue process diagram and generate node mapping relationships and node rule sets;

[0059] An improved GhostNet feature extraction module was developed to extract features from an event timeline dataset and output on-site feature vectors.

[0060] The extreme random tree decision module is used to fuse on-site feature vectors with structured recorded data and output the decision result;

[0061] The process linkage execution module is used to update the status of flowchart nodes and output the next node prompt information based on the judgment result and node rule set;

[0062] The consistency verification and logging module is used to verify the consistency between structured record data and the status of nodes in the rescue flowchart, log the data, and generate traceable quality control data.

[0063] refer to Figure 2 and Figure 3 A method for linking emergency resuscitation records and flowcharts, including:

[0064] Structured record data and multi-source on-site data were collected during the emergency rescue process and preprocessed to generate an event timeline dataset.

[0065] A flowchart is constructed based on the rescue process template, and the mapping relationship between flowchart nodes and structured record fields and node triggering conditions are established to generate a node rule set.

[0066] An improved GhostNet network is constructed. The input event timeline dataset is used, and Ghost feature maps with different receptive fields are generated using multi-scale Ghost features. Selective enhancement is performed by fusing lightweight attention. A dynamic Ghost-ratio strategy is used to generate a ratio based on the input data type and number of channels, and the output is a live feature vector.

[0067] The on-site feature vectors and structured recorded data are fused together, and the judgment is made based on the extreme random tree algorithm. The structure is shared by multiple trees to share the previous split structure. According to the hierarchical random sample order training strategy, the label distribution is randomly shuffled in a hierarchical manner before training each tree, and the judgment result is output.

[0068] Based on the judgment result and the node rule set, the record-driven flowchart linkage processing is executed to update the flowchart node status and trigger the output of the next node prompt information.

[0069] Based on the event timeline dataset, the consistency of structured record data and flowchart node status is verified and recorded, and traceable quality control data is generated.

[0070] In this embodiment, the multi-source field data includes vital sign data, monitoring waveform data, equipment operation events, and bedside image data.

[0071] In this embodiment, the generation of the event timeline dataset includes:

[0072] The system collects structured record data entered by medical staff terminals during emergency resuscitation and multi-source field data output by monitoring equipment, treatment equipment and bedside acquisition equipment, and generates data source identifiers, event type identifiers and event times for the structured record data and the multi-source field data respectively.

[0073] The event time is converted into a timestamp under a unified time base, and the structured record data is processed by unifying the field format, field name, enumeration value encoding and missing field marking. At the same time, the sampling frequency of multi-source field data is unified, noise is filtered and time is aligned.

[0074] The multi-source field data is segmented according to time windows to generate data segments and segment start timestamps and segment end timestamps. Based on the timestamps, the structured record data is associated with the data segments to generate an event timeline dataset containing event sequences, timestamp sequences, data segment indexes, and a set of structured record fields.

[0075] In this embodiment, the generated node rule set includes:

[0076] Read the rescue process template and parse it to obtain the process node set, node connection relationship set and branch condition set. Generate a node identifier, node type, predecessor node identifier set and successor node identifier set for each process node, and generate a rescue process diagram based on the process node set and node connection relationship set.

[0077] Configure a set of structured record fields for each process node and generate a mapping relationship between the node and the structured record fields. At the same time, configure node trigger conditions for each process node and generate a node rule set. The node trigger conditions include entry trigger conditions, completion trigger conditions, and timeout trigger conditions. The node rule set includes node identifier, mapping field set, trigger condition set, and branch condition set.

[0078] In this embodiment, the output field feature vector includes:

[0079] An improved GhostNet network is constructed, comprising a feature extraction backbone network formed by stacking multiple Ghost blocks and an output layer. The event timeline dataset is converted into a network input tensor and then fed into the improved GhostNet network, where:

[0080] Two depthwise separable convolutions with different kernel sizes are set up in parallel computation in the low-cost feature generation branch of the Ghost block to form a multi-scale Ghost feature generation structure, which is used to extract local mutations and large-scale trend features in the same layer. A lightweight attention fusion structure is connected after the intrinsic feature map output by the main convolution branch to selectively enhance the intrinsic feature channels, which is used to highlight key physiological changes or operational features related to the rescue event and suppress irrelevant channel responses.

[0081] A dynamic Ghost-ratio strategy is configured layer by layer in the network. The generation ratio of the number of intrinsic feature channels to the number of Ghost feature channels is dynamically determined based on the input data type identifier and the number of input channels in the current layer. This is used to avoid information bottlenecks in early layers and improve the overall feature representation capability while ensuring real-time inference overhead.

[0082] Multi-scale Ghost feature generation is performed in each Ghost block. Two depthwise separable convolutions with different kernel sizes are performed in parallel on the intrinsic feature map output by the main convolution to generate Ghost feature maps with different receptive fields. The intrinsic feature map and the Ghost feature map are then concatenated along the channel dimension to form the module output feature map.

[0083] Lightweight attention fusion is performed in each Ghost block. Global average pooling is performed on the intrinsic feature map to obtain the channel description vector. The channel weight vector is generated through two fully connected transformations. The channel weight vector is then multiplied with the intrinsic feature map channel by channel to obtain the selectively enhanced intrinsic feature map.

[0084] In the improved GhostNet network, a dynamic Ghost-ratio strategy is implemented. This strategy determines the ratio of intrinsic feature map channels to Ghost feature map channels based on the input data type identifier and the number of input channels in the current layer. The number of main convolution output channels and multi-scale depthwise separable convolution generation channels are then configured according to this determined ratio. Specifically, the dynamic Ghost-ratio strategy involves:

[0085] Before each Ghost block, the input data type identifier is read and the number of input channels for the current layer is obtained. The ghost-ratio value of the current layer is determined according to the type-channel mapping table. The number of target output channels of the current layer is split into the number of intrinsic feature map channels and the number of Ghost feature map channels according to the ghost-ratio. The number of output channels of the main convolution is configured as the number of intrinsic feature map channels. At the same time, the number of generated channels of the multi-scale depth separable convolution is configured as the number of Ghost feature map channels, and the channels are allocated among the branches of each scale according to the allocation ratio, so that the sum of the number of Ghost feature map channels generated by each scale branch is equal to the number of Ghost feature map channels.

[0086] Global average pooling is performed on the output feature map of the improved GhostNet network, and the field feature vector is obtained by linear transformation.

[0087] In this embodiment, the output determination result includes:

[0088] Extract field feature vectors from the structured record data in the event timeline dataset, and concatenate the on-site feature vectors with the field feature vectors according to the feature dimensions to generate a fused feature vector;

[0089] Based on the extreme random tree algorithm, multiple extreme random decision trees are generated, and the number of candidate features, maximum split depth, and stopping conditions are set for each extreme random decision tree. Specifically, setting the number of candidate features, maximum split depth, and stopping conditions for each extreme random decision tree involves:

[0090] The number of candidate features is determined as the output value of a preset function based on the total feature dimension of the fused feature vector. The preset function includes one of three methods: taking the square root of the total feature dimension, taking half of the total feature dimension, or taking a fixed upper limit value. The number of candidate features is limited to not exceeding the total feature dimension. The maximum split depth is determined as a preset depth value based on the sample size of the event timeline dataset and the number of process node categories. The maximum split depth is limited to not exceeding the preset upper limit depth.

[0091] The stopping condition is set to include two conditions simultaneously: the number of node samples is less than the preset minimum number of samples threshold, the node purity reaches the preset purity threshold, the improvement of the split evaluation index is less than the preset minimum improvement threshold, and the tree depth reaches the maximum split depth. When any one of the stopping conditions is met, the node will stop splitting.

[0092] Training using a shared subtree structure involves constructing a shared split structure for multiple extreme random decision trees within a preset shared depth range, and constructing a non-shared split structure for each extreme random decision tree after exceeding the shared depth. Specifically, constructing a shared split structure for multiple extreme random decision trees within the preset shared depth range involves:

[0093] First, set the shared depth to three layers starting from the root node, and initialize a set of shared nodes. Starting from the root node, for each node with a current depth less than or equal to the shared depth, generate a splitting candidate set using the same candidate feature subset generation method and the same random threshold generation method. Based on the same splitting evaluation index, select splitting features and splitting thresholds from the splitting candidate set to generate left and right child nodes. Then, write the selected splitting features, splitting thresholds, and child node indices into the shared node set.

[0094] For each extremely random decision tree in the forest, the shared node set is copied as the front-layer structure of the tree within the corresponding depth range, so that each tree has consistent splitting characteristics and splitting threshold within the shared depth range, and continues to grow according to the independent random splitting process of each tree after exceeding the shared depth.

[0095] A hierarchical shuffling sample order training strategy is adopted. Before training each extreme random decision tree, the training samples are divided into hierarchical levels according to their label categories, and the sample order within each level is randomly shuffled to generate the training sample sequence for the current tree. Specifically, the hierarchical shuffling sample order training strategy is as follows:

[0096] Before training each extreme random decision tree, the training samples are grouped according to the label category to obtain multiple category layers. Within each category layer, random permutations are generated independently to obtain the shuffled sequence of the current category layer. The shuffled sequences of each category layer are alternately spliced ​​according to the category traversal order to form the training sample sequence. The alternate splicing includes polling to take one sample from each category layer in turn and appending it to the training sample sequence until the samples of each category layer are exhausted, or taking a number of samples from each category layer in turn and appending them to the training sample sequence in a fixed batch size until the samples of each category layer are exhausted. The generated training sample sequence is used as the training input order of the current extreme random decision tree for node splitting and stopping condition judgment.

[0097] Based on the training and prediction of fused feature vectors, each extreme random decision tree randomly selects features from candidate features at each split node and randomly generates a split threshold within the feature value range to complete the node split. In the prediction stage, the outputs of multiple extreme random decision trees are integrated to obtain the judgment result.

[0098] In this embodiment, the step of performing record-driven flowchart linkage processing based on the judgment result and node rule set, updating the flowchart node status, and triggering the output of the next node prompt information includes:

[0099] Receive the judgment result and parse it to obtain the node identifier, node status, branch path identifier and missing item prompt set, and read the entry trigger condition, completion trigger condition and timeout trigger condition corresponding to the node identifier from the node rule set;

[0100] The flowchart node status is updated based on the judgment result, completion trigger condition and entry trigger condition, and the node entry timestamp and node completion timestamp are written into the event timeline dataset. At the same time, the successor node is determined according to the branch path identifier and the successor node status is updated.

[0101] Based on the timeout trigger condition, perform timeout judgment on the ongoing node and generate timeout prompt information. Based on the missing item prompt set, generate missing item prompt information and output the node status change result, next node prompt information, timeout prompt information and missing item prompt information to the flowchart interface.

[0102] In this embodiment, the process of verifying and recording the consistency between structured record data and flowchart node states based on the event timeline dataset, and generating traceable quality control data, includes:

[0103] Receive the flowchart node selection operation and obtain the selected node identifier. Read the mapping field set, required field marker, field value range, entry trigger condition, completion trigger condition and timeout trigger condition corresponding to the node identifier from the node rule set.

[0104] A node record template is generated based on the mapping field set. The node record template is then matched with the corresponding structured record field values ​​that already exist in the event timeline dataset to fill in the existing field values ​​and mark the unfilled fields. At the same time, a set of record verification items and a set of time window prompt items are generated based on the entry trigger condition, completion trigger condition, and timeout trigger condition and written into the node record template.

[0105] It receives field entry operations for node record templates and generates structured record entries. It generates operator identifiers and entry timestamps for structured record entries and writes them into the event timeline dataset. After writing, it triggers the calculation of node trigger conditions and outputs the calculation results to the flowchart interface to update the node status and prompt information.

[0106] Example 1:

[0107] To verify the feasibility of this invention in practice, it was applied to a peak emergency resuscitation scenario. The resuscitation room is typically staffed with 2 attending physicians, 3 nurses, and 1 recorder. During peak periods, 2-4 resuscitation beds are used simultaneously. Data sources include structured records entered through medical staff terminals, vital signs and waveforms from monitors, operational events from defibrillators, ventilators, and infusion pumps, and images captured by bedside ultrasound and monitoring screens. In this scenario, issues frequently arise such as treatment preceding recording, repeated data entry by multiple personnel and terminals, inconsistencies in backfilling at key time points, and mismatches between process branch selections and record fields. This leads to missing items in resuscitation records and inconsistencies in process execution, making post-event quality control reliant on manual traceability, which is time-consuming.

[0108] The system is deployed within the emergency room's local area network. One edge inference host connects to the monitoring equipment's data gateway and medical terminals, which include two doctor stations, three nurse stations, and one mobile tablet. The system first unifies and aligns the timestamps of structured records and multi-source on-site data to generate an event timeline dataset. Then, it constructs a flowchart based on the emergency procedure template and generates a set of node rules. Improved GhostNet is used to extract features from waveform segments, key image frames, and event sequences in the event timeline, outputting on-site feature vectors. These feature vectors are fused with structured record fields and then processed using an extreme random tree for evaluation, outputting node status, branch paths, and missing item prompts. This drives automatic updates to the flowchart nodes and provides prompts for the next node and a list of required fields on the interface. During parallel data entry by multiple users, the system automatically writes the operator and entry timestamp for each record. All node status changes, prompt triggers, and verification results are synchronously written to the log data. The system supports timeline playback and generates quality control data packages for quality control personnel to review.

[0109] To demonstrate its beneficial effects, a total of 126 emergency cases were covered during the trial operation period, accumulating approximately 190 million vital sign sampling points, approximately 384,000 monitoring waveform segments, approximately 26,000 equipment operation events, and approximately 78,000 bedside image keyframes. The median message transmission latency of the end-to-end linkage link in the emergency room LAN was 120 milliseconds, and the median total inference latency for a single feature extraction and judgment on the edge inference host was 32 milliseconds. When four beds were being treated simultaneously during peak concurrent operation, the interface refresh interval remained stable within 1 second. During the trial operation period, quality control personnel were able to generate quality control data packages for each case, including timelines, node status changes, and judgment criteria, significantly reducing the need to rely on recalling records and manually flipping through equipment logs for time points.

[0110] Table 1 Comparison of the linkage effect of emergency resuscitation records

[0111] As shown in Table 1, the solution of this invention is optimal overall in three core aspects: omission rate, node consistency rate, and recording completion delay. The omission rate is reduced to 1.8%, a significant decrease compared to the 9.6% of traditional manual recording; the node consistency rate reaches 97.9%, higher than the 89.5% of flowchart-only prompts and the 92.3% of rule-based linkage; the recording completion delay is only 3.2 minutes, significantly shorter than the 14.8 minutes of traditional manual recording and the 12.1 minutes of flowchart-only prompts, indicating that bidirectional linkage and automatic judgment are highly effective in reducing the time for supplementary recording and error correction.

[0112] In terms of timeout alarm accuracy, concurrency conflict rate, and terminal response latency, the solution of this invention has more obvious advantages in real-time performance and collaborative stability. The timeout alarm accuracy rate is 95.6%, which is significantly better than the 61.3% of traditional manual recording and the 68.5% of flowchart prompts only, and also higher than the 82.4% of rule linkage; the concurrency conflict rate is only 0.7%, which is lower than the 4.2% of traditional manual recording and the 3.9% of flowchart prompts only, and is more stable under multi-terminal concurrent input; the terminal response latency is 165 milliseconds, which is better than the comparison method, indicating that it can still maintain a relatively smooth interactive experience in high-concurrency rescue scenarios.

[0113] This invention also leads in two key metrics reflecting post-management capabilities: the time required for quality control generation and the time required for traceability and location. Quality control generation takes 1.6 minutes, significantly lower than traditional manual recording (10.4 minutes) and flowchart-only prompts (8.9 minutes), and also faster than the two original GhostNet methods (2.9 minutes and 2.6 minutes). Traceability and location takes 2.3 minutes, a substantial reduction compared to traditional manual recording (18.6 minutes) and rule-based linkage (7.4 minutes). This demonstrates that this invention is more efficient in terms of record keeping, consistency verification, and the formation of traceable quality control data.

[0114] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An emergency resuscitation record and flowchart linkage system, characterized in that, include: The data acquisition and preprocessing module is used to acquire structured recorded data and multi-source field data, and to preprocess and generate an event timeline dataset. The process modeling rules module is used to construct the rescue process diagram and generate node mapping relationships and node rule sets; An improved GhostNet feature extraction module was developed to extract features from an event timeline dataset and output on-site feature vectors. The extreme random tree decision module is used to fuse on-site feature vectors with structured recorded data and output the decision result; The process linkage execution module is used to update the status of flowchart nodes and output the next node prompt information based on the judgment result and node rule set; The consistency verification and logging module is used to verify the consistency between structured record data and the status of nodes in the rescue flowchart, log the data, and generate traceable quality control data.

2. The method for linking emergency resuscitation records and flowcharts according to claim 1, applied to the emergency resuscitation record and flowchart linkage system according to claim 1, is characterized in that, include: Structured record data and multi-source on-site data were collected during the emergency rescue process and preprocessed to generate an event timeline dataset. A flowchart is constructed based on the rescue process template, and the mapping relationship between flowchart nodes and structured record fields and node triggering conditions are established to generate a node rule set. An improved GhostNet network is constructed. The input event timeline dataset is used, and Ghost feature maps with different receptive fields are generated using multi-scale Ghost features. Selective enhancement is performed by fusing lightweight attention. A dynamic Ghost-ratio strategy is used to generate a ratio based on the input data type and number of channels, and the output is a live feature vector. The on-site feature vectors and structured recorded data are fused together, and the judgment is made based on the extreme random tree algorithm. The structure is shared by multiple trees to share the previous split structure. According to the hierarchical random sample order training strategy, the label distribution is randomly shuffled in a hierarchical manner before training each tree, and the judgment result is output. Based on the judgment result and the node rule set, the record-driven flowchart linkage processing is executed to update the flowchart node status and trigger the output of the next node prompt information. Based on the event timeline dataset, the consistency of structured record data and flowchart node status is verified and recorded, and traceable quality control data is generated.

3. The method for linking emergency resuscitation records and flowcharts according to claim 2, characterized in that, The multi-source field data includes vital signs data, monitoring waveform data, equipment operation events, and bedside image data.

4. The method for linking emergency resuscitation records and flowcharts according to claim 2, characterized in that, The generated event timeline dataset includes: The system collects structured record data entered by medical staff terminals during emergency resuscitation and multi-source field data output by monitoring equipment, treatment equipment and bedside acquisition equipment, and generates data source identifiers, event type identifiers and event times for the structured record data and the multi-source field data respectively. The event time is converted into a timestamp under a unified time base, and the structured record data is processed by unifying the field format, field name, enumeration value encoding and missing field marking. At the same time, the sampling frequency of multi-source field data is unified, noise is filtered and time is aligned. The multi-source field data is segmented according to time windows to generate data segments and segment start timestamps and segment end timestamps. Based on the timestamps, the structured record data is associated with the data segments to generate an event timeline dataset containing event sequences, timestamp sequences, data segment indexes, and a set of structured record fields.

5. The method for linking emergency resuscitation records and flowcharts according to claim 2, characterized in that, The generated node rule set includes: Read the rescue process template and parse it to obtain the process node set, node connection relationship set and branch condition set. Generate a node identifier, node type, predecessor node identifier set and successor node identifier set for each process node, and generate a rescue process diagram based on the process node set and node connection relationship set. Configure a set of structured record fields for each process node and generate a mapping relationship between the node and the structured record fields. At the same time, configure node trigger conditions for each process node and generate a node rule set. The node trigger conditions include entry trigger conditions, completion trigger conditions, and timeout trigger conditions. The node rule set includes node identifier, mapping field set, trigger condition set, and branch condition set.

6. The method for linking emergency resuscitation records and flowcharts according to claim 2, characterized in that, The output site feature vector includes: An improved GhostNet network is constructed, including a feature extraction backbone network formed by stacking multiple Ghost blocks and an output layer. The event timeline dataset is converted into a network input tensor and then fed into the improved GhostNet network. Multi-scale Ghost feature generation is performed in each Ghost block. Two depthwise separable convolutions with different kernel sizes are performed in parallel on the intrinsic feature map output by the main convolution to generate Ghost feature maps with different receptive fields. The intrinsic feature map and the Ghost feature map are then concatenated along the channel dimension to form the module output feature map. Lightweight attention fusion is performed in each Ghost block. Global average pooling is performed on the intrinsic feature map to obtain the channel description vector. The channel weight vector is generated through two fully connected transformations. The channel weight vector is then multiplied with the intrinsic feature map channel by channel to obtain the selectively enhanced intrinsic feature map. In the improved GhostNet network, a dynamic Ghost-ratio strategy is implemented. The ratio of the number of channels in the intrinsic feature map to the number of channels in the Ghost feature map is determined based on the input data type identifier and the number of input channels in the current layer. The number of output channels of the main convolution and the number of generated channels of the multi-scale depth separable convolution are configured according to the determined ratio. Global average pooling is performed on the output feature map of the improved GhostNet network, and the field feature vector is obtained by linear transformation.

7. The method for linking emergency resuscitation records and flowcharts according to claim 2, characterized in that, The output determination result includes: Extract field feature vectors from the structured record data in the event timeline dataset, and concatenate the on-site feature vectors with the field feature vectors according to the feature dimensions to generate a fused feature vector; Based on the extreme random tree algorithm, multiple extreme random decision trees are generated, and the number of candidate features, maximum split depth, and stopping conditions are set for each extreme random decision tree. Training with shared subtrees is adopted. Within a preset shared depth range, a shared split structure is constructed for multiple extreme random decision trees. After the shared depth is exceeded, a non-shared split structure is constructed for each extreme random decision tree. A hierarchical shuffling sample order training strategy is adopted. Before training each extreme random decision tree, the training samples are divided into hierarchical layers according to the label category, and the sample order is randomly shuffled within each layer to generate the training sample sequence of the current tree. Based on the training and prediction of fused feature vectors, each extreme random decision tree randomly selects features from candidate features at each split node and randomly generates a split threshold within the feature value range to complete the node split. In the prediction stage, the outputs of multiple extreme random decision trees are integrated to obtain the judgment result.

8. The method for linking emergency resuscitation records and flowcharts according to claim 2, characterized in that, The process of executing record-driven flowchart linkage based on the judgment result and node rule set, updating the flowchart node status and triggering the output of the next node prompt information includes: Receive the judgment result and parse it to obtain the node identifier, node status, branch path identifier and missing item prompt set, and read the entry trigger condition, completion trigger condition and timeout trigger condition corresponding to the node identifier from the node rule set; The flowchart node status is updated based on the judgment result, completion trigger condition and entry trigger condition, and the node entry timestamp and node completion timestamp are written into the event timeline dataset. At the same time, the successor node is determined according to the branch path identifier and the successor node status is updated. Based on the timeout trigger condition, perform timeout judgment on the ongoing node and generate timeout prompt information. Based on the missing item prompt set, generate missing item prompt information and output the node status change result, next node prompt information, timeout prompt information and missing item prompt information to the flowchart interface.

9. The method for linking emergency resuscitation records and flowcharts according to claim 2, characterized in that, The consistency verification and logging of structured record data and flowchart node states based on the event timeline dataset, and the generation of traceable quality control data, include: Receive the flowchart node selection operation and obtain the selected node identifier. Read the mapping field set, required field marker, field value range, entry trigger condition, completion trigger condition and timeout trigger condition corresponding to the node identifier from the node rule set. A node record template is generated based on the mapping field set. The node record template is then matched with the corresponding structured record field values ​​that already exist in the event timeline dataset to fill in the existing field values ​​and mark the unfilled fields. At the same time, a set of record verification items and a set of time window prompt items are generated based on the entry trigger condition, completion trigger condition, and timeout trigger condition and written into the node record template. It receives field entry operations for node record templates and generates structured record entries. It generates operator identifiers and entry timestamps for structured record entries and writes them into the event timeline dataset. After writing, it triggers the calculation of node trigger conditions and outputs the calculation results to the flowchart interface to update the node status and prompt information.

Images