An intelligent pre-examination and triage system for emergency patients

Abstract

The present invention discloses an intelligent pre-examination and triage system for emergency patients, which specifically relates to the field of emergency pre-examination and triage, including: a data processing module that standardizes the patient's chief complaint text and physical sign data to construct a pre-examination data set; a symptom prediction module that inputs the pre-examination data set into a multi-label disease prediction model to output a potential symptom set and its predicted probability; a probability adjustment module that corrects the confidence of the prediction result in combination with the patient's historical medical history; a graph construction module that constructs a semantic adjacency graph of potential symptoms based on semantic similarity; a detection item determination module that performs co-occurrence path compression on the basis of a semantic graph structure, extracts a minimum detection item set, and generates a list of auxiliary detection items; a triage guidance module that schedules and sorts detection items according to the real-time status of detection resources, generates triage detection guidance, and realizes efficient and flexible emergency initial screening and triage planning.

Description

Technical Field

[0001] The present invention relates to the technical field of emergency pre-examination and triage, and more particularly to an intelligent pre-examination and triage system for emergency patients. Background Art

[0002] In the current emergency medical process, patients typically undergo pre-examination and triage, a preliminary consultation with a doctor, and the ordering of auxiliary examinations. However, due to the uncertainty of patient arrivals and the concentrated pressure during peak hours, preliminary consultations often lead to queues, causing many patients to wait for a long time before formal consultations. Under the current process, auxiliary examination items (such as blood routine, urinalysis, and bedside ultrasound) must be ordered after the doctor's consultation, which fails to fully utilize the idle time during the waiting period. At the same time, for patients with clear clinical complaints such as common abdominal pain, fever, and urinary symptoms, preliminary auxiliary examinations are standardized, low-risk, and non-invasive. In fact, they can be pre-arranged based on preliminary symptom screening before consultations, thereby improving overall emergency department turnover efficiency and medical resource utilization.

[0003] In order to solve the above problems, a technical solution is now provided. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an intelligent pre-examination and triage system for emergency patients to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] An intelligent pre-examination and triage system for emergency patients, comprising a data processing module, a symptom prediction module, a probability adjustment module, a semantic adjacency graph construction module, a test item determination module, and a triage guidance module;

[0007] The data processing module performs standardized processing on the patient's chief complaint text and physical sign data collected during the emergency pre-examination registration stage to construct a pre-examination data set;

[0008] The disease prediction module inputs the pre-examination dataset into a multi-label disease prediction model based on supervised learning, and outputs a set of potential diseases and their corresponding prediction probabilities;

[0009] The probability adjustment module adjusts the predicted probability of the potential disease set based on the patient's historical medical history information;

[0010] The graph construction module analyzes the set of potential symptoms after the prediction probability adjustment and constructs a semantic adjacency graph between the potential symptoms based on the semantic similarity calculation results;

[0011] The test item determination module performs co-occurrence path compression processing on the test items associated with each disease node in the semantic adjacency graph, extracts the minimum set of test items that covers all reasoning paths in the potential disease set, and establishes an auxiliary test item list;

[0012] The triage guidance module performs real-time scheduling planning on the auxiliary test item list based on the queuing status of the test resources and generates triage test guidelines.

[0013] In a preferred embodiment, the data processing module performs standardization processing on the patient's chief complaint text and physical sign data collected during the emergency pre-examination registration stage, and constructing the pre-examination data set specifically includes:

[0014] Based on the patient's chief complaint text and physical sign data recorded during the emergency registration stage, the initial data structure is constructed and the corresponding pre-diagnosis session number is generated;

[0015] Perform symbol removal and medical semantic word normalization on the chief complaint text to generate an encoded chief complaint word sequence, and maintain an index mapping relationship with the original chief complaint text;

[0016] Unit conversion and time precision reconstruction are performed on the physical sign data. The main complaint word sequence is used as the main sequence input, and the reconstructed physical sign data are inserted as parallel parameters into the main sequence input to construct a pre-examination data set.

[0017] In a preferred embodiment, the symptom prediction module inputs the pre-examination data set into a multi-label disease prediction model based on supervised learning, and outputs a set of potential symptoms and corresponding prediction probabilities, specifically including:

[0018] Input the pre-examination dataset into the multi-label disease prediction model built based on the supervised learning method to trigger the model inference process;

[0019] During model inference, a forward propagation operation is performed on the input vector, outputting a set of potential symptoms aligned with the dimensions of the preset symptom output space and preliminary probabilities corresponding to the potential symptoms;

[0020] The preliminary probabilities are normalized to generate a standardized potential disease prediction matrix, which is then bound to the current pre-diagnosis session number and written into the disease prediction cache.

[0021] In a preferred embodiment, the multi-label disease prediction model constructed based on the supervised learning method is constructed by extracting the chief complaint text and physical sign data in the historical emergency medical records to construct an initial training sample, and based on the set of doctor diagnosis disease labels corresponding to the historical emergency medical records, the initial training sample is multi-label labeled, and the initial training sample is input into the multi-label classification model based on the supervised learning method, and the model is trained using the multi-label binary cross entropy loss function.

[0022] In a preferred embodiment, the probability adjustment module adjusts the predicted probability of the potential disease set based on the patient's historical medical history information, specifically including:

[0023] Call the patient's historical medical history information, filter the historical disease label set that is the same as the potential disease in the disease prediction cache, and build a diagonal weighted matrix of disease prediction labels based on the hit frequency of historical disease labels;

[0024] Based on the diagonal weighted matrix, the predicted probability distribution of the potential disease set is multiplicatively weighted to generate a calibrated potential disease probability matrix, and the calibrated potential disease probability matrix is ​​updated into the disease prediction cache.

[0025] In a preferred embodiment, the graph construction module analyzes the potential disease set after the prediction probability adjustment and constructs a semantic adjacency graph between the potential diseases based on the semantic similarity calculation results, specifically including:

[0026] Extract the potential disease set in the disease prediction cache and convert it into disease semantic nodes. According to the preset disease semantic relationship dictionary, calculate the association strength between disease node pairs through semantic similarity;

[0027] Based on the association strength between pairs of disease nodes, a semantic adjacency graph between potential diseases is constructed, and the graph structure is initialized with nodes as vertices and association strength as edge weights.

[0028] The semantic adjacency graph between potential diseases is subjected to threshold pruning, and edge connections with edge weights lower than the set association threshold are removed.

[0029] In a preferred embodiment, the detection item determination module performs co-occurrence path compression processing on the detection items associated with each disease node in the semantic adjacency graph, extracts the minimum detection item set that covers all reasoning paths in the potential disease set, and establishes the auxiliary detection item list, specifically including:

[0030] In the semantic adjacency graph, adjacent disease node pairs with prediction probabilities greater than or equal to a set threshold are extracted, and at least one reasoning path is constructed based on the continuous adjacent disease node pairs;

[0031] Establish a set of auxiliary detection items corresponding to each disease node, and calculate the ratio of the number of reasoning path nodes covered by each detection item to the total number of nodes in the corresponding reasoning path, which is used as the coverage completeness of the path;

[0032] Construct a detection-reasoning path coverage matrix, where the matrix elements are the path coverage completeness of the reasoning path;

[0033] Using the predicted probability of potential disease nodes as weights, a weighted coverage compression algorithm is applied to the detection-reasoning path coverage matrix to extract the minimum set of detection items that covers all reasoning paths. Based on this minimum set of detection items, an auxiliary detection item list is established.

[0034] Eliminate the test items in the list of auxiliary test items that cannot be tested without medical advice.

[0035] In a preferred embodiment, the triage guidance module performs real-time scheduling planning on the auxiliary test item list based on the test resource queue status, and generates triage test guidance specifically including:

[0036] Obtain the current queue waiting time and equipment availability status information for each item in the auxiliary detection item list;

[0037] Execute a multi-objective sorting algorithm based on detection priority sorting to generate a detection dispatch order that maximizes path coverage completeness and resource load balance;

[0038] Bind the test dispatch order to the pre-diagnosis session number to provide triage testing guidance for emergency patients corresponding to the pre-diagnosis session number.

[0039] The technical effects and advantages of the intelligent pre-examination and triage system for emergency patients of the present invention are as follows:

[0040] The symptom prediction module performs multi-label predictions on the chief complaint and physical sign information, and combines historical medical history to make probability corrections, thereby improving the individual accuracy of initial diagnosis and identification. The graph construction module constructs the semantic structure between symptoms, avoiding the redundancy of detection paths caused by isolated symptom labels. The detection item determination module extracts detection combinations based on the co-occurrence path compression method, effectively compressing redundant detections and covering key diagnostic pathways. The triage guidance module optimizes scheduling based on the status of detection resources, improving detection execution efficiency and resource utilization. The overall solution not only improves the efficiency of emergency patient triage, but also enhances the stability of initial detection decisions and the rationality of path recommendations.

[0041] The system achieves a closed-loop process, from driving the chief complaint data to guiding the testing path. Compared with the traditional method of relying on doctors' experience for preliminary triage, it significantly improves the intelligence and scheduling flexibility of emergency pre-examination and triage. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The present invention is a structural diagram of an intelligent pre-examination and triage system for emergency patients. DETAILED DESCRIPTION

[0043] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0044] Example 1

[0045] Figure 1 The present invention provides a schematic structural diagram of an intelligent pre-examination and triage system for emergency patients, which includes a data processing module, a symptom prediction module, a probability adjustment module, a semantic adjacency graph construction module, a test item determination module, and a triage guidance module.

[0046] The data processing module performs standardized processing on the patient's chief complaint text and physical sign data collected during the emergency pre-examination registration stage to construct a pre-examination data set;

[0047] The disease prediction module inputs the pre-examination dataset into a multi-label disease prediction model based on supervised learning, and outputs a set of potential diseases and their corresponding prediction probabilities;

[0048] The probability adjustment module adjusts the predicted probability of the potential disease set based on the patient's historical medical history information;

[0049] The graph construction module analyzes the set of potential symptoms after the prediction probability adjustment and constructs a semantic adjacency graph between the potential symptoms based on the semantic similarity calculation results;

[0050] The test item determination module performs co-occurrence path compression processing on the test items associated with each disease node in the semantic adjacency graph, extracts the minimum set of test items that covers all reasoning paths in the potential disease set, and establishes an auxiliary test item list;

[0051] The triage guidance module performs real-time scheduling planning on the auxiliary test item list based on the queuing status of the test resources and generates triage test guidelines.

[0052] The data processing module performs standardization processing on the patient's chief complaint text and physical sign data collected during the emergency pre-examination registration stage to construct a pre-examination data set.

[0053] After the patient completes the emergency pre-examination registration process, the system automatically pulls the patient's main complaint text entered in the current consultation session and the vital sign parameter fields collected at the registration node from the front-end information platform, and uses the two as the original input data for the construction of the pre-examination data structure. First, a structured data encapsulation unit is established, using the patient's ID number, medical card number or HIS system internal identification code as the primary key identifier, and the registration timestamp field is connected to generate a set of traceable pre-examination data object IDs. On this basis, a pre-examination session number is generated for the data object through time precision control rules and platform scheduling identifier generation strategy. The pre-examination session number can be generated using time series coding, institution ID plus serial number embedding, or UUID global unique numbering. This session number is embedded as a unique identifier in the system into the disease prediction results, path structure construction, test recommendation sequence and triage instruction scheduling instruction set, realizing full process tracking from pre-examination access to test completion.

[0054] The initial data structure contains three dimensions of information: the first is the original chief complaint text string and input channel information (such as voice transcription, manual input, and mobile text upload); the second is structured vital sign parameter fields, covering basic vital sign parameters (such as temperature, heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate, and blood oxygen saturation) as well as platform-defined extended fields (such as level of consciousness, pain score, and body position score); and the third is time field information, including registration time, reporting time, pre-diagnosis start time, and historical concurrent session indexes. Once this structure is encapsulated, it is submitted to the main data channel to enter the chief complaint preprocessing and vital sign normalization phase.

[0055] Standardized preprocessing and medical semantic structure building are performed on the original text fields of the chief complaint contained in the initial data structure to generate a chief complaint encoding sequence vector for input into the subsequent disease prediction model. The text processing module is called to perform symbol cleaning operations on the chief complaint text to remove non-standard characters, format control symbols, punctuation combinations, redundant spaces, emoticons, or colloquial affixes, and other unstructured semantic content. Through rule-based regular expression templates and language filtering dictionaries, the cleaned text is ensured to retain entity tokens related to the disease. The system then initiates the semantic segmentation process, using a token segmenter trained on the BioBERT semantic space or other medical-specific dictionaries to divide the cleaned chief complaint text into tokens, identify symptom descriptions, organ locations, time descriptions, and qualitative word modification structures, and mark each token with its part of speech and medical semantic category.

[0056] The system performs semantic normalization on the words, uniformly mapping words such as patient terms, common abbreviations, and colloquial names to standard medical terms. For example, "fever", "high body temperature", and "fever to 38 degrees" are uniformly classified as "high fever" standard words, and "chest tightness", "shortness of breath", and "holding breath" are classified as "difficulty breathing". After the word normalization, the system attaches a word number to each standard word, and constructs a sequence of coded chief complaint terms. At the same time, it establishes a mapping table with the character position information in the original text, recording the start and end character positions of each standard word in the original text. This mapping relationship table will be used as the basis for chief complaint reconstruction in subsequent patient-side visualization, doctor review interface, or multi-channel output, so that the model processing structure maintains a bidirectional locatable relationship with the original input information.

[0057] The sequence of chief complaint terms and vital sign data are structurally merged to generate a pre-screening data set for the disease prediction model. Vital sign data comes from various sources during the emergency registration stage. They may be automatically reported by electronic vital sign collection devices or manually entered by nurses. There are problems such as inconsistent units, inconsistent accuracy, and chaotic time alignment. Therefore, the system first performs unit conversion and format standardization on the original vital sign parameter fields. For example, the temperature fields reported by different devices are uniformly converted to the platform standard unit (Celsius), and parameters such as blood pressure, blood sugar, and heart rate are uniformly retained to one decimal place or converted to discrete graded interval format, and missing values ​​are filled and outliers are identified.

[0058] Subsequently, the system accurately reconstructs the time field of the physical sign parameters, maps the physical sign data collected at multiple time points to the time range of the chief complaint word generation according to the nearest principle, or generates a unified physical sign snapshot at the time point corresponding to the current chief complaint through weighted interpolation. After the reconstruction is completed, the system constructs a text nested structure with the chief complaint word sequence as the main sequence, and inserts the reconstructed physical sign parameters as the parallel input structure into the main sequence structure. The embedding strategy adopts a nested splicing form: the chief complaint word sequence constitutes the time series input dimension, and the physical sign parameters are embedded in parallel as the channel external input vector according to the field category position, and are uniformly entered into the joint encoding structure of the multi-label disease prediction model, and the final output is the pre-examination data set.

[0059] The disease prediction module inputs the pre-examination data set into a multi-label disease prediction model based on supervised learning, and outputs a set of potential diseases and corresponding prediction probabilities.

[0060] After completing the pre-examination dataset, we entered the inference phase of a multi-label disease prediction model built using supervised learning methods. This model had already achieved structural convergence during the initial training process. The training data was sourced from the hospital's historical emergency case database. By extracting sequences of chief complaint terms and standardized physical sign parameter vectors as input features, and using the set of disease labels diagnosed by doctors as the multi-label target, the model parameters were trained and the weights were frozen.

[0061] When performing a prediction task, the system loads the pre-examination dataset for the current patient into the model inference interface, triggering the forward reasoning process. The model input interface receives the nested semantic complaint encoding sequence and parallel sign vectors according to the preset data vector specification and inputs them into the joint encoder structure for feature integration. The joint encoder can adopt a multi-channel embedding strategy or a cross-attention mechanism to fuse text and numerical features, improving multimodal representation capabilities through shared hidden layers.

[0062] During the inference task initialization phase, the system assigns a unique model task number to the current patient input data and establishes a logical reference with the current pre-diagnosis session number to support the mapping and binding of subsequent prediction output results to the main process status. The prediction model structure is based on a multi-label neural network. After the inference is started, the system executes the forward propagation path starting from the current input feature, completes the output of the potential disease label score to each output label node, and pauses at the model terminal interface to wait for post-processing. The entire inference call process is hosted by the disease prediction module. No backpropagation or gradient update operations are performed within the inference process to ensure that the prediction output runs stably under the weight freeze condition after training is completed, and the output result structure is aligned with the model output dimension during the training period.

[0063] After the pre-inspection data set is input into the model inference entrance, the model's internal forward propagation process officially begins. The input vector is first encoded into a feature representation vector group. The main complaint word sequence is processed by the text embedding layer and mapped into a high-dimensional semantic vector. Medical pre-trained word vectors (such as BioBERT embedding) can be used to complete the initial semantic loading, followed by context fusion through one-dimensional convolution, BiLSTM structure or multi-head attention mechanism. The parallel sign parameter vector is directly projected into a unified latent space through a numerical normalization layer and a structural mapping layer. The two types of vectors complete feature fusion in the joint encoding module, forming a unified representation vector and then entering the hidden layer propagation process.

[0064] In a multi-label prediction model, the output space is a fixed-dimensional set of disease labels, with each output node corresponding to a candidate disease label. Sigmoid activation units can be used structurally to allow each node to independently output a probability value. The forward propagation process executes a forward computation graph without gradient calculations based on the trained frozen weights, unfolding layer by layer along the model structure, ultimately obtaining a probability distribution vector aligned with the label dimension at the output layer. The system performs confidence analysis on the output probabilities of each dimension and extracts a set of non-zero label indices, corresponding to the set of potential diseases identified in this pre-screening task. The label number of this set aligns with the index in the model's internal label space, and records the original model output value (i.e., preliminary predicted probability) corresponding to each disease label.

[0065] After obtaining the set of potential symptoms and their corresponding preliminary predicted probabilities, the system enters the standardization phase of this step. This process aims to convert the original model output values ​​into a structurally consistent and horizontally comparable symptom prediction matrix for subsequent probability adjustment, graph construction, and detection path compression. The system first initializes a two-dimensional symptom prediction matrix structure that matches the spatial dimensions of the model output (depending on system performance, the default setting is the maximum number of symptom labels after deduplication). This matrix is ​​indexed by the symptom label number as a column vector and the current pre-screening task as a single row. Finally, the data is written to the disease prediction cache as key-value pairs. The cache can be managed using databases such as Redis and MongoDB.

[0066] The multi-label disease prediction model training process involves batch extraction of real-world case samples from emergency department visits. The sample data sources cover different time periods, patient populations, and disease types, enhancing the model's adaptability to actual clinical complexity. Each sample is composed of two main input dimensions: the first is the textual data of the chief complaint, derived from the patient's self-report during the emergency department visit or the chief complaint recorded by medical staff. The field is in the form of a short, free-to-describe natural language sentence that may include a variety of clinical semantic elements such as symptoms, location, duration, and triggers; the second is basic vital sign parameters, derived from the initial vital sign monitoring and manual entry, including fields such as body temperature, heart rate, blood pressure, respiratory rate, and blood oxygen saturation.

[0067] The chief complaint text field is first cleaned and standardized. After removing symbolic noise, the medical semantic segmenter is called to perform phrase-level segmentation, and the segmentation results are normalized and mapped in combination with the medical ontology library to output a standardized word sequence; the physical sign parameter field is simultaneously subjected to unit normalization, missing value interpolation and numerical discretization processing, and a physical sign vector with a unified structure is output. The above-mentioned text semantic coding sequence and physical sign vector are combined to form a joint feature vector of the sample and form a sample input matrix. Subsequently, the system extracts the corresponding symptom label set based on the main diagnosis label finally confirmed by the professional doctor in each case record. The set is usually composed of multiple labels, covering comorbidities, multiple diseases or inferential symptom structures in emergency scenarios. In order to adapt to the supervised learning model structure, the label set will be converted into a multi-label binary vector. The vector length is consistent with the dimension of the full label space. The corresponding symptom position is assigned a value of 1, and the rest are set to 0, forming the supervised label output matrix of the sample.

[0068] During the training phase, the system batches the constructed sample input matrix and multi-label output matrix into the training model structure, which is a multi-label neural network. The training process uses the binary cross-entropy loss function commonly used in multi-label tasks. This function calculates the cross-entropy loss between the output probability and the true label for each label node, and then performs batch averaging of the full-label loss before performing error propagation. The Adam optimizer is used for optimization until the model weights converge after multiple training iterations, and the output structure can be used for subsequent inference processes.

[0069] The probability adjustment module performs probability adjustment on the predicted probability of the potential disease set according to the patient's historical medical history information.

[0070] The system calls the patient's historical medical record interface to retrieve all historical medical records from the hospital information system, covering structured diagnosis tag fields for outpatient, inpatient, and emergency departments. The system standardizes the structure of the primary and secondary diagnosis symptom labels recorded in the historical medical records, mapping the label text content to the unified label space used in the current symptom prediction model. Using a label standardization dictionary or ICD-10 code conversion, the system constructs a historical symptom label set.

[0071] The system then selects a subset of labels that precisely match the historical label set based on the set of potential symptoms in the disease prediction cache for the current pre-diagnosis task, extracting their hit frequency information from the historical records. To enhance expressiveness and avoid sparsity bias, the system performs a log transformation on the hit frequencies, converting them into confidence weighting factors. This constructs a diagonal weighted matrix with the same dimensions as the set of potential symptom labels. The diagonal elements of the matrix represent the historical frequency weights of the corresponding symptom labels, and the off-diagonal positions are filled with zeros.

[0072] This diagonal weighted matrix is ​​applied to the current potential disease prediction probability matrix in label index order, performing a multiplicative weighting operation. This multiplication of each label prediction probability value output by the original model is multiplied by the corresponding diagonal weight to form a weighted, revised prediction probability. After weighting, the system normalizes the product to preserve the original probability distribution proportional characteristics. The calibrated disease probability matrix is ​​then rewritten as a separate version to the disease prediction cache, maintaining the same session number for the current patient and marking it as a historical, calibrated version.

[0073] The graph construction module analyzes the potential disease set after the prediction probability adjustment, and constructs a semantic adjacency graph between the potential diseases according to the semantic similarity calculation results.

[0074] First, a set of potential conditions with adjusted prediction probabilities is extracted from the disease prediction cache. This set typically contains multiple condition labels that match the current patient's chief complaint and physical sign data. Based on this set of conditions, the system constructs a semantic node structure for each condition in turn. The node stores the condition's code, the corresponding predicted probability value, and a future scalable test recommendation index structure. All semantic nodes are integrated into the graph construction environment of the current pre-diagnosis task as a set of vertices in the graph structure.

[0075] After completing the node initialization, the system loads the preset disease semantic relationship dictionary, which can be based on the international standard disease classification library, knowledge graph construction results, or trained based on clinical case data to obtain a set of semantic representations of disease labels. The dictionary records the semantic similarity reference information between common diseases in terms of clinical manifestations, etiology, anatomical location or comorbidity trends. The system processes all disease nodes under the current task in pairs, and retrieves the semantic relevance value between each pair of diseases through the dictionary. The value can be calculated based on the word vector distance between the diseases, the classification path overlap depth or the frequency of common diagnosis. For disease pairs that do not appear explicitly in the dictionary, the system estimates semantic similarity through synonym matching, stem restoration or model extrapolation to ensure that valid similarity scores can be obtained between all node pairs.

[0076] Based on the generated set of node-based semantic representations of symptoms and the semantic association strength scores between pairs of symptom nodes, a semantic adjacency graph structure is constructed between potential symptoms, serving as the underlying graph data for subsequent inference and detection path compression. The graph structure is composed of symptom semantic nodes, each of which is initialized with structured diagnostic label information, predicted probability, standard description, and test item index. The system uses the similarity scores of symptom node pairs output from the previous stage as the basis for edge connections. Node pairs with scores exceeding the system's minimum association threshold are sequentially registered as edge relationships in the graph structure. Each edge is represented as an undirected structure and has its semantic association value attached as the edge weight. When constructing edge connections, the system automatically maintains an adjacency list structure for each node. The constructed graph structure is persistently stored in an adjacency list format and registered in the current pre-diagnosis task session context. This semantic adjacency graph structure preserves the static co-occurring semantic connections between all potential symptoms.

[0077] Based on a system-defined minimum relevance threshold (specifically set based on the similarity distribution of adjacent nodes, with a default setting of 0.4), all edge connections are traversed and analyzed. The system evaluates the semantic relevance score of each edge. If this score falls below the set pruning threshold, the edge is deemed insufficiently clinically semantically strong and unsuitable as a component of the path for subsequent structural reasoning. Therefore, the edge is removed from the graph. After the edge removal operation is complete, the system resynchronizes and updates the adjacency table to ensure that the structural relationships between nodes remain consistent.

[0078] The detection item determination module performs co-occurrence path compression processing on the detection items associated with each disease node in the semantic adjacency graph, extracts the minimum detection item set covering all reasoning paths in the potential disease set, and establishes an auxiliary detection item list.

[0079] Based on the pruned semantic adjacency graph, a structural path extraction operation is performed to identify candidate disease combination paths that can form semantic connectivity in the current task. The system first screens all potential disease nodes for predicted probabilities, extracting the set of disease nodes with predicted probabilities greater than or equal to a system-defined threshold. This threshold is statistically derived from historical model training data and is set to 0.7 by default. It is typically used to calibrate the set of disease labels with diagnostic confidence in the current model output.

[0080] The system traverses all retained edge connections in the disease semantic adjacency graph, identifies pairs of adjacent disease nodes consisting of two high-confidence disease nodes, and records their connectivity. On this basis, the system executes a path construction strategy based on depth-first or breadth-first, identifying all path chains consisting of continuous high-confidence nodes in the graph. Each path is composed of two or more disease nodes connected end to end. The connection between the nodes must be an actual edge structure in the graph, and the edge weights have passed the pre-order pruning check. These continuous disease paths are marked as candidate reasoning paths by the system. Each path represents a set of diagnostic directions with semantic coherence and high model prediction probability. Among them, if a single disease node without edge connection appears, the single disease node is used as a candidate reasoning path.

[0081] Establish an association mapping relationship between symptom labels and test items. The system calls the symptom-test standard mapping database to retrieve the set of auxiliary test items recommended for each symptom node. This set contains test items that can be used for preliminary diagnosis or screening of the symptom, giving priority to screening test items that can be performed independently without a doctor's order (such as blood routine, urine analysis, bedside B-ultrasound, etc.), and eliminating test items that require a doctor's order (such as CT, electrocardiogram, endoscopic examination, etc.). Each symptom node corresponds to one or more test items, and the test items can be reused across multiple symptoms. The system traverses the symptom nodes in each reasoning path in turn and counts the test items that appear in the path.

[0082] For each test item, the system counts the number of disease nodes that the item can cover in a certain reasoning path, and also records the total number of nodes in that path. The system then calculates the coverage ratio for each test item on each path. This ratio represents the completeness of the test item's coverage of that path. This metric is a key parameter in this system for measuring the information coverage capability of a test item in a structural path, reflecting its importance in the path semantic chain. For example, if a test item can detect two out of three diseases in a certain path, its coverage completeness is 0.67.

[0083] Construct a detection-reasoning path coverage matrix, where the matrix elements are the path coverage completeness of the reasoning path. Call the compression algorithm to perform weighted coverage compression on the matrix to extract the minimum set of detection items covering all reasoning paths. The system first aggregates these probability values ​​into a path confidence score based on the predicted probability of each disease node in each reasoning path to characterize the importance of the path from the model perspective. Subsequently, the system weights the path columns in the coverage matrix according to the path confidence, and assigns different compression priorities to different paths. The system calls the weighted set coverage algorithm to give priority to detection items that cover high-weight disease nodes in high-confidence paths, and covers the semantic connectivity structure in all reasoning paths as much as possible with the least number of detection items. During the execution of the algorithm, the system continuously tracks the proportion of covered paths and the cumulative number of selected detection items until all reasoning paths are completely or within the set proportion range. The final minimum set of detection items is the set of detection recommendations under the current patient's individualized conditions. The following is a specific example:

[0084] Suppose the system currently predicts the following set of potential symptoms: Symptom A (predicted probability 0.8), Symptom B (predicted probability 0.7), Symptom C (predicted probability 0.9), and Symptom D (predicted probability 0.9). The graph structure forms the following two inference paths: Path P1 (ABC) and Path P2 (BD).

[0085] Each symptom node is associated with the following test items through the symptom-test mapping table: symptom A is associated with blood routine and C-reactive protein; symptom B is associated with blood routine and chest X-ray; symptom C is associated with blood gas analysis; symptom D is associated with chest X-ray and urine routine.

[0086] At this time, the proportion of disease nodes covered by each test item in each path should be: the corresponding P1 and P2 coverage of routine blood test are 0.67 and 0.5 respectively; the corresponding P1 and P2 coverage of C-reactive protein and blood gas analysis are 0.33 and 0 respectively; the corresponding P1 and P2 coverage of chest X-ray are 0.33 and 0.5 respectively; the corresponding P1 and P2 coverage of routine urine test are 0 and 0.5 respectively.

[0087] Confidence weights are assigned to the paths based on the average predicted probability of the nodes, with P1 and P2 weights of 0.8 and 0.6, respectively. The weighted coverage score (weight × coverage) of each test item is calculated. Based on the premise of selecting the minimum number of test items so that each path has at least all disease nodes covered, a greedy heuristic is used for selection:

[0088] Routine blood test (covering symptoms A and B) has the largest contribution of 0.833 (0.8×0.67+0.6×0.5), which has covered A and B; blood gas analysis (covering symptoms C) completes symptoms C, and now P1 is fully covered; routine urine test (covering symptoms D) completes symptoms D, and now P2 is fully covered.

[0089] The final minimum test set is [blood routine, blood gas analysis, urine routine].

[0090] The triage guidance module performs real-time scheduling planning on the auxiliary detection item list based on the detection resource queuing status and generates triage detection guidance.

[0091] The system sequentially establishes a test resource index structure for each test item in the auxiliary test item list, recording all device nodes that the test item can access at the current moment and calculating the expected wait time for each device node. This wait time is calculated by combining the queue length and the task processing rate. For test items that support multiple test devices simultaneously, the system constructs multiple candidate execution paths and evaluates their resource scheduling costs in both the temporal and spatial dimensions.

[0092] The path coverage role of the test items in the semantic structure graph is combined with their real-time resource usage status, and a multi-objective sorting algorithm is executed to generate an execution priority queue for the test items. Without sacrificing the integrity of the path structure, the equipment pressure is alleviated as much as possible, the patient waiting time is reduced, and the rapid execution of key test items is achieved.

[0093] First, the system calculates a structural coverage priority for each test item. This priority is based on a comprehensive score of three metrics: the number of reasoning paths covered by the test item in the semantic graph, the completeness of coverage, and the average symptom prediction confidence of the covered paths. This creates a structural importance score. Next, the system calculates the current resource pressure indicator for each test item based on queue length, estimated wait time, and device load obtained in the previous stage. This resource pressure score uses an inverse weighting strategy, meaning that shorter wait times and lower loads are associated with higher resource scores, indicating greater suitability for the current execution. A multi-objective ranking model is then invoked to combine the structural priority and resource pressure scores for ranking. The ranking strategy can employ linear weighting, primary-objective-slave priority, or a constrained optimization model. Different scheduling preferences can be configured, such as prioritizing device load balancing during peak periods and symptom structural integrity during risk periods. The ranking model outputs a set of ordered test execution plans. Each test item is accompanied by a ranking weight, a recommended execution window, and a list of target device candidates to guide patient triage.

[0094] Testing instructions can be delivered through two channels: patient-side guidance, which pushes test information to patients via the hospital's mobile app, a patient guidance terminal, or an electronic call center, directing them to their designated location to queue. Second, a backend dispatch channel allows the system to directly send task requests to relevant equipment or the testing task system, automatically dispatching tasks. Each dispatch instruction record is associated with the current pre-diagnosis session ID and updated in the pre-diagnosis status manager, which is used to track patient testing completion progress, identify task failures, and initiate rescheduling operations.

[0095] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0096] The above embodiments can be implemented in whole or in part via software, hardware, firmware, or any other combination. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product comprises one or more computer instructions or computer programs. When loaded or executed on a computer, the processes or functions described in the embodiments of this application are fully or partially performed. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server or data center that contains a collection of one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, tapes), optical media (e.g., DVDs), or semiconductor media. The semiconductor media can be a solid-state drive.

[0097] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0098] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0099] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0100] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0101] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0102] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0103] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

[0104] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent pre-examination and triage system for emergency patients, characterized by: It includes data processing module, disease prediction module, probability adjustment module, semantic adjacency graph construction module, test item determination module and triage guidance module; The data processing module performs standardized processing on the patient's chief complaint text and physical sign data collected during the emergency pre-examination registration stage to construct a pre-examination data set; The disease prediction module inputs the pre-examination dataset into a multi-label disease prediction model based on supervised learning, and outputs a set of potential diseases and their corresponding prediction probabilities; The probability adjustment module adjusts the predicted probability of the potential disease set based on the patient's historical medical history information; The graph construction module analyzes the set of potential symptoms after the prediction probability adjustment and constructs a semantic adjacency graph between the potential symptoms based on the semantic similarity calculation results; The test item determination module performs co-occurrence path compression processing on the test items associated with each disease node in the semantic adjacency graph, extracts the minimum set of test items that covers all reasoning paths in the potential disease set, and establishes an auxiliary test item list; The triage guidance module performs real-time scheduling planning on the auxiliary test item list based on the test resource queue status and generates triage test guidelines; The disease prediction module inputs the pre-examination data set into the multi-label disease prediction model based on supervised learning, and outputs a set of potential diseases and corresponding prediction probabilities, specifically including: Input the pre-examination dataset into the multi-label disease prediction model built based on the supervised learning method to trigger the model inference process; During model inference, a forward propagation operation is performed on the input vector, outputting a set of potential symptoms aligned with the dimensions of the preset symptom output space and preliminary probabilities corresponding to the potential symptoms; Normalize the preliminary probabilities to generate a standardized potential disease prediction matrix, bind the potential disease prediction matrix to the current pre-diagnosis session number, and write it into the disease prediction cache; The graph construction module analyzes the potential disease set after the prediction probability adjustment and constructs a semantic adjacency graph between the potential diseases based on the semantic similarity calculation results, specifically including: Extract the potential disease set in the disease prediction cache and convert it into disease semantic nodes. According to the preset disease semantic relationship dictionary, calculate the association strength between disease node pairs through semantic similarity; Based on the association strength between pairs of disease nodes, a semantic adjacency graph between potential diseases is constructed, and the graph structure is initialized with nodes as vertices and association strength as edge weights. Threshold pruning is performed on the semantic adjacency graph between potential diseases to remove edge connections with edge weights lower than the set correlation threshold; The detection item determination module performs co-occurrence path compression processing on the detection items associated with each disease node in the semantic adjacency graph, extracts the minimum detection item set covering all reasoning paths in the potential disease set, and establishes the auxiliary detection item list, specifically including: In the semantic adjacency graph, adjacent disease node pairs with prediction probabilities greater than or equal to a set threshold are extracted, and at least one reasoning path is constructed based on the continuous adjacent disease node pairs; Establish a set of auxiliary detection items corresponding to each disease node, and calculate the ratio of the number of reasoning path nodes covered by each detection item to the total number of nodes in the corresponding reasoning path, which is used as the coverage completeness of the path; Construct a detection-reasoning path coverage matrix, where the matrix elements are the path coverage completeness of the reasoning path; Using the predicted probability of potential disease nodes as weights, a weighted coverage compression algorithm is applied to the detection-reasoning path coverage matrix to extract the minimum set of detection items that covers all reasoning paths. Based on this minimum set of detection items, an auxiliary detection item list is established. Eliminate the test items in the list of auxiliary test items that cannot be tested without medical advice.

2. The intelligent pre-examination and triage system for emergency patients according to claim 1, characterized in that: The data processing module performs standardization processing on the patient's chief complaint text and physical sign data collected during the emergency pre-examination registration stage, and constructs a pre-examination data set, specifically including: Based on the patient's chief complaint text and physical sign data recorded during the emergency registration stage, the initial data structure is constructed and the corresponding pre-diagnosis session number is generated; Perform symbol removal and medical semantic word normalization on the chief complaint text to generate an encoded chief complaint word sequence, and maintain an index mapping relationship with the original chief complaint text; Unit conversion and time precision reconstruction are performed on the physical sign data. The main complaint word sequence is used as the main sequence input, and the reconstructed physical sign data are inserted as parallel parameters into the main sequence input to construct a pre-examination data set.

3. The intelligent pre-examination and triage system for emergency patients according to claim 1, characterized in that: The multi-label disease prediction model constructed based on the supervised learning method is constructed by extracting the chief complaint text and physical sign data in the historical emergency medical records to construct an initial training sample, and based on the set of doctor diagnosis disease labels corresponding to the historical emergency medical records, the initial training sample is multi-label labeled. The initial training sample is input into the multi-label classification model based on the supervised learning method, and the model is trained using the multi-label binary cross entropy loss function.

4. The intelligent pre-examination and triage system for emergency patients according to claim 1 is characterized in that The probability adjustment module adjusts the predicted probability of the potential disease set based on the patient's historical medical history information, specifically including: Call the patient's historical medical history information, filter the historical disease label set that is the same as the potential disease in the disease prediction cache, and build a diagonal weighted matrix of disease prediction labels based on the hit frequency of historical disease labels; Based on the diagonal weighted matrix, the predicted probability distribution of the potential disease set is multiplicatively weighted to generate a calibrated potential disease probability matrix, and the calibrated potential disease probability matrix is ​​updated into the disease prediction cache.

5. The intelligent pre-examination and triage system for emergency patients according to claim 1, characterized in that: The triage guidance module performs real-time scheduling planning on the auxiliary test item list based on the test resource queue status, and generates triage test guidance specifically including: Obtain the current queue waiting time and equipment availability status information for each item in the auxiliary detection item list; Execute a multi-objective sorting algorithm based on detection priority sorting to generate a detection dispatch order that maximizes path coverage completeness and resource load balance; Bind the test dispatch order to the pre-diagnosis session number to provide triage testing guidance for emergency patients corresponding to the pre-diagnosis session number.

Images