Intelligent triage priority evaluation method and system based on deep time sequence learning

Abstract

The invention provides an intelligent triage priority evaluation method and system based on deep time sequence learning, and relates to the technical field of medical triage, and the method comprises the steps: constructing a feature vector through processing patient registration, symptom text and physiological data, constructing a patient relation graph based on a symptom and department relation, and generating a node enhancement feature through a graph attention network. In combination with the historical treatment sequence of the patient, a time sequence enhancement priority feature is generated through a variational encoder, a priority score is obtained through multiple times of random inactivation deduction, a waiting queue is dynamically reordered according to the priority score, a consulting room distribution instruction is generated, and accurate and efficient evaluation of the triage priority is achieved.

Description

Technical Field

[0001] This invention relates to the field of medical triage technology, and in particular to a method and system for intelligent assessment of triage priority based on deep temporal learning. Background Technology

[0002] In the field of medical triage, traditional methods primarily rely on triage nurses to manually assess and prioritize patients based on standardized triage guidelines, combined with preliminary measurements of vital signs and brief inquiries. The core of this conventional approach is to quickly categorize patients into different levels of urgency using predetermined rules and the experience of healthcare professionals, and then arrange their treatment order accordingly. In addition, some systems integrate basic electronic triage tools. By inputting a limited number of key physiological indicators, the system automatically calculates a preliminary risk score based on preset thresholds, serving as a supplementary reference for manual triage. The entire process emphasizes speed and standardization, aiming to cope with the continuous influx of patients in scenarios such as emergency rooms, ensuring that critically ill patients are prioritized for identification and treatment.

[0003] However, the aforementioned conventional triage model has significant limitations. Its assessment relies on a relatively singular and static information dimension, typically including only a few vital signs collected immediately and keywords from the patient's chief complaint. It lacks a deep semantic understanding of the patient's symptom descriptions and fails to effectively incorporate the temporal health information contained in the patient's rich historical medical records. This results in potentially incomplete and inaccurate assessments of the severity of the patient's condition, especially for patients with complex symptom descriptions or chronic diseases, whose potential risks are easily underestimated or misjudged. Furthermore, existing methods often assess each patient in isolation, ignoring potential correlations in the conditions of patients waiting in the same time period or the overall resource load. This makes it difficult to adjust triage decisions from individual optimality to batch-wide optimality. When dealing with large-scale, high-concurrency waiting scenarios, there is still room for improvement in triage efficiency and fairness. Summary of the Invention

[0004] This invention provides a triage priority intelligent assessment method and system based on deep temporal learning, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a triage priority intelligent assessment method based on deep temporal learning, comprising: The system obtains the registration information, symptom self-report text, and physiological feature data of patients waiting for treatment. It performs word segmentation and word vector mapping on the symptom self-report text to obtain semantic feature vectors. It then concatenates these feature vectors with the registration information and physiological feature data to obtain the waiting feature vector. The feature vectors of each patient in the same waiting batch are used as node features. Edge weights are assigned based on symptom similarity and the department to which the patient belongs. A patient relationship graph is constructed, which is then input into a graph attention network and aggregated with the features of adjacent nodes through a multi-head attention mechanism to generate node enhancement features that are integrated with the whole batch. The historical visit feature sequence of each waiting patient is concatenated with the corresponding node enhancement feature. A variational encoder with learnable positional encoding is introduced to construct a KL divergence sequence to quantify the contribution difference of each historical visit feature and generate temporal enhancement priority features. Multiple random inactivation forward inferences are performed on the waiting time sequence enhancement priority feature to calculate the waiting priority score. The waiting queue is dynamically reordered based on the waiting priority score, and a consultation room allocation instruction is generated and synchronized to the triage workstation.

[0006] The symptom self-report text is segmented and word vector mapped to obtain a semantic feature vector. This semantic feature vector is then concatenated with the registration information and the physiological feature data to obtain a waiting feature vector, including: The symptom self-report text is subjected to medical noise filtering and semantic boundary word segmentation. Each word segmentation unit is mapped to a dedicated medical semantic space according to the word unit type, generating a word vector sequence carrying position encoding. Multi-granularity feature extraction is performed on the word vector sequence. At the word level, the local semantics of a single symptom word are captured. At the phrase level, the semantics of symptom combinations of adjacent words are aggregated. At the sentence level, the global semantics of the entire sentence's symptom description are integrated. Dimensionality is reduced through a linear projection layer, and the resulting semantic feature vectors are generated. Linear embedding mapping is performed on the registration information and the physiological feature data respectively to obtain registration feature vector and physiological feature vector; using the semantic feature vector as the query and the registration feature vector and physiological feature vector as key-value pairs respectively, cross-modal cross-attention calculation is performed to guide the response of feature dimensions that contribute highly to waiting priority in symptom semantic enhancement, thereby obtaining symptom-guided registration features and symptom-guided physiological features. The semantic feature vector, symptom-guided registration feature, and symptom-guided physiological feature are respectively subjected to residual enhancement and then sequentially concatenated to obtain a waiting feature vector that integrates three modal information.

[0007] Each waiting feature vector within the same waiting batch is used as a node feature. Edge weights are assigned based on symptom similarity and department affiliation to construct a waiting patient relationship graph. This graph is then input into a graph attention network, and adjacent node features are aggregated using a multi-head attention mechanism to generate enhanced node features that fuse the entire batch. These features include: Each waiting patient is treated as a node in the waiting patient relationship graph. For any two nodes in the same waiting batch, the symptom similarity and the consistency of the department they belong to are calculated. The symptom similarity and the consistency of the department they belong to are concatenated and then mapped to learnable edge weights through linear projection to obtain the waiting patient relationship graph. Within each attention head of the graph attention network, the change in cosine similarity of the nodes aggregated at the current layer is fed back to update the edge weights, so that the edge weights are iteratively and adaptively refined as the message propagates; the attention coefficients are re-modulated with the updated edge weights and weighted aggregation is performed to obtain the node aggregation features within the head; the aggregation features within each head are concatenated, nonlinearly projected, and superimposed with the residual of the original waiting feature vector to obtain the node enhancement features of a single node. Batch normalization is performed on all single-node augmentation features within a batch, and the mean and variance statistics are retained. The mean and variance are concatenated to form a batch statistical vector. The batch statistical vector is used to generate node-by-node recalibration coefficients through a feedforward network. Recalibration is applied to the augmentation features of each node to obtain the node augmentation features of the entire batch.

[0008] Within each attention head of the graph attention network, the change in cosine similarity between node pairs aggregated from the current layer is fed back to update the edge weights, allowing the edge weights to be iteratively and adaptively refined as message propagation occurs, including: After the message propagation of the current round is completed, the cosine similarity of all node pairs is calculated based on the node aggregation output of the current round, and the difference between the cosine similarity of the current round and the cosine similarity of the previous round is used to obtain the change in cosine similarity of node pairs. Asymmetric learnable gain coefficients are applied to the cosine similarity change of node pairs according to their positive and negative polarities. The positive polarity change is amplified by the enhancement coefficient to strengthen the coupling relationship between node pairs, while the negative polarity change is reduced by the suppression coefficient to weaken the coupling relationship between node pairs. The current edge weights are then multiplicatively updated using the polarity-aware weighted change. After the update, the sum of the weights of the outgoing edges of each node is used as the denominator to perform row normalization on all outgoing edge weights of that node, eliminating the influence of neighborhood size differences on the absolute magnitude of edge weights, while retaining the relative weight relationship between outgoing edges of the same node, thus obtaining the updated edge weights. The updated edge weights are logarithmically summed and added to the corresponding nodes to score the attention. Structural modulation is completed before softmax is executed. The enhancement coefficient, suppression coefficient, and logarithmic scaling coefficient are used as three mutually decoupled learnable scalars for joint training. The differential amplitude of positive enhancement and negative suppression is adaptively controlled, so that the dynamic structure of edge weights and attention coefficients evolve together during the iteration process.

[0009] The historical visit feature sequences of each waiting patient are concatenated with the corresponding node enhancement features. A variational encoder with learnable positional encoding is introduced to construct a KL divergence sequence to quantify the contribution differences of each historical visit feature, generating temporal enhancement priority features, including: Each historical visit feature in the historical visit feature sequence of each waiting patient is concatenated with the corresponding node augmented feature according to the feature dimension, and a learnable location code is superimposed to obtain a location-aware concatenated feature sequence. Based on the variational encoder and the location-aware spliced ​​feature sequence, the random latent representation of each historical visit feature is obtained by reparameterized sampling, and the mean and logarithmic variance of the latent distribution are calculated. At the same time, the target distribution is generated by learningable projection of the corresponding node enhancement features, and the KL divergence sequence between the latent distribution and the target distribution of each historical visit feature is constructed. The contribution weights of each historical visit feature are obtained by normalizing the KL divergence sequence after negative value transformation. KL divergence is the similarity of the information compression direction between each historical visit feature and the current patient status. The contribution weights quantitatively characterize the correlation differences between each historical visit feature and the priority feature model from the perspective of information theory. The random latent representations of each historical visit time are weighted and aggregated using contribution weights, and then mapped to temporal enhanced priority features of each waiting patient by a learnable decoder.

[0010] The target distribution parameters are generated by using the corresponding node enhancement features through learnable projection, and the KL divergence sequence between the latent distribution and the target distribution of each historical visit feature is constructed, including: The node enhancement features of each waiting patient are decomposed in a dual-path parallel manner along the feature dimension. They are mapped to semantic target distribution parameters and state target distribution parameters through two sets of parameter-independent learnable projection networks. The corresponding mean vector and log variance vector are linearly combined by learnable fusion weights to integrate them into a unified target distribution parameter, which represents the current patient's comprehensive prior state information in the form of a parameterized probability distribution. Using the mean vector and log-variance vector of the unified target distribution parameters, combined with the mean and log-variance of the latent distribution independently inferred by the variational encoder for each historical consultation time, the forward KL divergence from the latent distribution to the target distribution and the backward KL divergence from the target distribution to the latent distribution are calculated for each historical consultation time. The symmetric bidirectional information bias metric is obtained by weighted superposition of learnable coefficients. The bidirectional information deviation metric values ​​corresponding to each historical visit time are arranged in the order of visit time. The amplitude of each divergence value is normalized by a learnable nonlinear correction network to eliminate the dimensional bias introduced by the difference in variance of the potential distribution at each time. This forms a KL divergence sequence that quantitatively measures the degree of information deviation between the potential distribution of each historical visit feature and the target distribution of the current patient status.

[0011] Multiple random inactivation forward inferences are performed on the enhanced priority features of the waiting time sequence to calculate the waiting priority score. Based on the waiting priority score, the waiting queue is dynamically reordered, and a consultation room allocation instruction is generated and synchronized to the triage workstation, including: Before each random inactivation forward inference, a random perturbation with controlled amplitude is applied to the temporal position code in the waiting time-series enhanced priority feature, so that random inactivation and position code perturbation work together to expand the output diversity of forward inference. Multiple collaborative double randomization forward inferences are performed, and the stability of the relative priority ranking of each waiting patient within the batch in all inference times is calculated. The weighted product of the inference output mean and the ranking stability coefficient is used as the waiting priority score. Based on the waiting priority score, a ranking weight is calculated for each patient in the waiting queue. The ranking stability coefficient of each patient is used as a weight reliability factor to further correct the ranking weight. Based on the corrected ranking weight, the waiting queue is dynamically reordered to generate a dynamic ranking result. Based on the dynamic sorting results, the real-time load of each clinic is matched to generate a clinic allocation instruction, which is then synchronized to the triage workstation.

[0012] A second aspect of the present invention provides a triage priority intelligent assessment system based on deep temporal learning, comprising: The feature extraction unit is used to acquire the registration information, symptom self-report text and physiological feature data of the waiting patients, perform word segmentation and word vector mapping on the symptom self-report text to obtain semantic feature vectors, and concatenate the feature vectors with the registration information and the physiological feature data to obtain the waiting feature vector. The graph aggregation unit is used to take the feature vectors of each patient in the same waiting batch as node features, assign edge weights according to symptom similarity and the department to which the patient belongs, construct a patient relationship graph, input it into the graph attention network and aggregate the features of adjacent nodes through a multi-head attention mechanism to generate node enhancement features that are integrated into the whole batch. The temporal enhancement unit is used to concatenate the historical visit feature sequence of each waiting patient with the corresponding node enhancement feature, introduce a variational encoder with learnable positional encoding, construct a KL divergence sequence to quantify the contribution difference of each historical visit feature, and generate temporal enhancement priority features. The priority sorting unit is used to perform multiple random inactivation forward inferences on the enhanced priority features of the waiting time sequence, calculate the waiting priority score, perform dynamic reordering of the waiting queue based on the waiting priority score, generate a consultation room allocation instruction and synchronize it to the triage workstation.

[0013] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0014] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0015] This method can achieve accurate and intelligent assessment of triage priority, significantly improving triage efficiency and the rationality of medical resource allocation in emergency or outpatient scenarios.

[0016] By fusing multimodal patient data and constructing a relationship graph of waiting patients, we can effectively capture individual characteristics and group correlations, enhancing the comprehensiveness and discriminative power of feature representation. Utilizing a graph attention network to aggregate features of adjacent nodes allows priority assessment to be based not only on the patient's own condition but also on the overall situation of patients with similar symptoms in the same batch or within the same department, reducing assessment bias.

[0017] A variational encoder with learnable location encoding is introduced to process historical visit sequences. KL divergence is used to quantify the differences in historical feature contributions, enhancing the ability to model temporal dependencies and making the assessment results more closely reflect the evolution of the patient's condition. Priority scores are calculated using multiple random inactivation forward inferences to improve model robustness and assessment stability, and reduce the risk of overfitting.

[0018] Based on dynamic reordering, consultation room allocation instructions are generated, enabling real-time optimization and adjustment of the waiting queue, shortening waiting times for critically ill patients, and balancing the workload of each consultation room. Ultimately, through automated triage decision support, the workload of medical staff is reduced, and the speed of diagnosis and treatment response and patient satisfaction are improved. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the intelligent triage priority assessment method based on deep temporal learning, as described in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the method for generating node enhancement features of the entire fused batch in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 This is a flowchart illustrating the intelligent triage priority assessment method based on deep temporal learning, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: The system obtains the registration information, symptom self-report text, and physiological feature data of patients waiting for treatment. It performs word segmentation and word vector mapping on the symptom self-report text to obtain semantic feature vectors. It then concatenates these feature vectors with the registration information and physiological feature data to obtain the waiting feature vector. The feature vectors of each patient in the same waiting batch are used as node features. Edge weights are assigned based on symptom similarity and the department to which the patient belongs. A patient relationship graph is constructed, which is then input into a graph attention network and aggregated with the features of adjacent nodes through a multi-head attention mechanism to generate node enhancement features that are integrated with the whole batch. The historical visit feature sequence of each waiting patient is concatenated with the corresponding node enhancement feature. A variational encoder with learnable positional encoding is introduced to construct a KL divergence sequence to quantify the contribution difference of each historical visit feature and generate temporal enhancement priority features. Multiple random inactivation forward inferences are performed on the waiting time sequence enhancement priority feature to calculate the waiting priority score. The waiting queue is dynamically reordered based on the waiting priority score, and a consultation room allocation instruction is generated and synchronized to the triage workstation.

[0023] The symptom self-report text is segmented and word vector mapped to obtain a semantic feature vector. This semantic feature vector is then concatenated with the registration information and the physiological feature data to obtain a waiting feature vector, including: The symptom self-report text is subjected to medical noise filtering and semantic boundary word segmentation. Each word segmentation unit is mapped to a dedicated medical semantic space according to the word unit type, generating a word vector sequence carrying position encoding. Multi-granularity feature extraction is performed on the word vector sequence. At the word level, the local semantics of a single symptom word are captured. At the phrase level, the semantics of symptom combinations of adjacent words are aggregated. At the sentence level, the global semantics of the entire sentence's symptom description are integrated. Dimensionality is reduced through a linear projection layer, and the resulting semantic feature vectors are generated. Perform linear embedding mapping on the registration information and the physiological feature data respectively to obtain the registration feature vector and the physiological feature vector; Using the semantic feature vector as the query and the registration feature vector and the physiological feature vector as key-value pairs, cross-modal cross-attention calculation is performed to guide the response of feature dimensions that contribute highly to waiting priority in symptom semantic enhancement, thereby obtaining symptom-guided registration features and symptom-guided physiological features. The semantic feature vector, symptom-guided registration feature, and symptom-guided physiological feature are respectively subjected to residual enhancement and then sequentially concatenated to obtain a waiting feature vector that integrates three modal information.

[0024] In this specific embodiment, the patient's symptom self-report text needs to undergo medical noise filtering. By constructing a medical professional vocabulary list and a stop word list, irrelevant information in the text, such as interjections, colloquial expressions, and non-medical descriptive words, is identified and removed. At the same time, regular expression matching technology is used to standardize the expression of medical terminology, converting colloquial symptom descriptions into standardized medical terms.

[0025] In the semantic boundary segmentation stage, a maximum matching algorithm based on a medical dictionary combined with a conditional random field model is used for segmentation. The medical dictionary contains professional terms such as disease names, symptom terms, and anatomical structure names to ensure that the integrity of medical terms is not compromised. The conditional random field model learns the contextual features of medical text to accurately identify word boundaries and avoids incorrect segmentation of compound medical terms.

[0026] Each word segmentation word is classified and mapped according to its semantic type to establish a medical semantic space, including different semantic dimensions such as symptom, location, degree, and time. Each word is mapped to the corresponding semantic subspace according to its medical attributes, forming a vector representation with specific meaning in the medical field. Through positional encoding technology, the positional information of each word in the original text is added. The positional code is generated using sine and cosine functions, enabling the model to capture the relative positional relationship between words.

[0027] In the multi-granularity feature extraction process, word-level granularity is processed by a single convolutional kernel of a convolutional neural network with a kernel size of 1, specifically capturing the local semantic features of a single symptom word. For single symptom words such as "headache" and "fever", their independent semantic representations are extracted. Phrase-level granularity uses convolutional kernels of size 3 to 5, with a sliding window covering multiple adjacent words, aggregating the composite semantics of symptom combinations such as "persistent chest pain" and "intermittent diarrhea". Sentence-level granularity processes the entire sentence sequence through a long short-term memory network, integrating the global semantic information of all symptom descriptions in the whole sentence.

[0028] The features at three granular levels are dimensionality-reduced through their respective independent linear projection layers, mapping the high-dimensional features to a unified low-dimensional space. The projection layers are implemented using fully connected neural networks, which reduce the feature vectors of different granularities to the same dimension through weight matrix transformation. Then, the feature vectors of the three granularities are weighted and combined. The weight parameters are automatically learned through the training process, and finally, feature vectors containing multi-level semantic information are generated.

[0029] For registration information, including structured data such as the department, appointment time, and doctor's level, one-hot encoding is used to convert categorical variables into vector form, and continuous variables are standardized. The processed features are then mapped to the vector space through a linear embedding layer containing a trainable weight matrix, transforming the original registration features into vector representations with the same dimensions as the symptom semantic features. Physiological feature data covers vital signs such as blood pressure, heart rate, body temperature, and blood oxygen saturation. Normalization preprocessing is performed on each indicator to eliminate differences in numerical magnitude between different indicators. A linear transformation is used to map the multidimensional physiological data into a unified feature vector, and the parameters of the transformation matrix are optimized and adjusted during training.

[0030] In cross-modal attention computation, the symptom semantic feature vector is used as the query vector, and the registration feature vector and physiological feature vector are used as the key vector and value vector, respectively. The attention weight calculation adopts the scaling dot product attention mechanism. The attention distribution is obtained by performing the dot product operation between the query vector and the key vector and then normalizing it through the softmax function. This mechanism can automatically identify the registration and physiological feature dimensions that are most relevant to the current symptom.

[0031] By weighting and summing the value vectors using attention weights, a symptom-guided feature representation is generated. For registration information, the attention mechanism focuses on department selection and doctor's expertise that match the symptoms. For physiological characteristics, it emphasizes abnormal vital signs related to the symptoms. This guidance mechanism ensures that the priority of waiting is determined by paying more attention to key information directly related to the patient's symptoms.

[0032] To enhance feature representation capabilities, residual connection operations are performed on the original semantic feature vector, symptom-guided registration features, and symptom-guided physiological features, respectively. The residual connection adds the input features to the output features after attention processing, preserving the original information while incorporating enhanced semantic associations. This approach effectively alleviates the gradient vanishing problem in deep network training and improves feature learning performance.

[0033] The three residual-enhanced feature vectors are concatenated in the order of symptom semantics, registration information, and physiological features, while maintaining the relative positions of the features in each modality. This results in a high-dimensional waiting feature vector that integrates multimodal information. This vector comprehensively incorporates semantic understanding of patient symptoms, intention analysis of registration behavior, and objective assessment of physiological state, providing comprehensive feature support for subsequent waiting priority determination.

[0034] Figure 2 This is a flowchart illustrating the method for generating node enhancement features for the entire fused batch according to an embodiment of the present invention. Each waiting feature vector within the same waiting batch is used as a node feature. Edge weights are assigned based on symptom similarity and department affiliation to construct a waiting patient relationship graph. This graph is then input into a graph attention network, and adjacent node features are aggregated using a multi-head attention mechanism to generate node enhancement features for the entire fused batch, including: Each waiting patient is treated as a node in the waiting patient relationship graph. For any two nodes in the same waiting batch, the symptom similarity and the consistency of the department they belong to are calculated. The symptom similarity and the consistency of the department they belong to are concatenated and then mapped to learnable edge weights through linear projection to obtain the waiting patient relationship graph. Within each attention head of the graph attention network, the change in cosine similarity of the nodes aggregated at the current layer is fed back to update the edge weights, so that the edge weights are iteratively and adaptively refined as the message propagates; the attention coefficients are re-modulated with the updated edge weights and weighted aggregation is performed to obtain the node aggregation features within the head; the aggregation features within each head are concatenated, nonlinearly projected, and superimposed with the residual of the original waiting feature vector to obtain the node enhancement features of a single node. Batch normalization is performed on all single-node augmentation features within a batch, and the mean and variance statistics are retained. The mean and variance are concatenated to form a batch statistical vector. The batch statistical vector is used to generate node-by-node recalibration coefficients through a feedforward network. Recalibration is applied to the augmentation features of each node to obtain the node augmentation features of the entire batch.

[0035] When constructing the patient relationship graph, the waiting feature vector corresponding to each patient in the same waiting batch should be used as a node in the graph. For any two nodes in the batch, extract the symptom-related dimension subset from their waiting feature vectors, and calculate the symptom similarity using a cosine similarity metric, with a value ranging from zero to one. Simultaneously, extract the department identifier from the registration information. If the department codes of two nodes are the same, assign a consistency metric value of one; otherwise, assign a value of zero. Concatenate the symptom similarity scalar and the consistency metric scalar to form a two-dimensional vector. Map this two-dimensional vector to a one-dimensional scalar using a linear projection network with a single hidden layer. This scalar serves as the initial edge weight connecting the two nodes. After traversing all nodes in the batch and initializing the edge weights, the complete patient relationship graph is obtained.

[0036] The constructed patient relationship graph is input into the graph attention network. This network employs a multi-head attention mechanism, with four attention heads. Within each attention head, the current layer receives the node features output from the previous layer as input. For any node and its neighboring nodes, the cosine similarity between the current layer's input features is calculated. Simultaneously, the cosine similarity value of that node pair when inputting in the previous layer is extracted, and the difference between the two yields the change in cosine similarity. This change reflects the dynamic evolution trend of the relationships between nodes. It is then weighted and summed with the current edge weights. The weighting coefficients are generated using a learnable parameter matrix, enabling iterative adaptive refinement and updating of the edge weights.

[0037] The attention coefficient is recalculated using the updated edge weights. The node features are transformed linearly to obtain the query vector and key vector. The dot product of the query vector and key vector is calculated, multiplied by the updated edge weights, and then normalized using softmax to obtain the new attention coefficient. This coefficient reflects the degree of attention the current node pays to the features of its neighboring nodes.

[0038] A weighted aggregation operation is performed based on the recalculated attention coefficients. The value vectors of neighboring nodes are multiplied by their corresponding attention coefficients and then summed to obtain the aggregated features of the current node within that attention head. The aggregated features integrate information from neighboring nodes while maintaining the relevance weight allocation to the current node.

[0039] The aggregated features of each attention head are combined into a high-dimensional feature vector through a concatenation operation. This vector is processed by a nonlinear projection layer, which employs a multilayer perceptron structure and includes ReLU activation and Dropout regularization. The projected features are residually concatenated with the original candidate feature vector, and element-wise addition is used to obtain the node-enhanced features for each node. Batch normalization is performed on the enhanced features of all nodes within a batch. The mean and variance statistics of all node features within a batch are calculated, and feature standardization is achieved by subtracting the mean and dividing by the standard deviation of the variance. The normalization process eliminates the differences in feature distribution between different nodes, improving the stability of model training.

[0040] The mean and variance statistics calculated during batch normalization are retained, and these two statistics are concatenated into a batch statistical vector. This vector contains the overall characteristic distribution information of the waiting patient group in the current batch, reflecting the overall distribution of the severity and urgency of the patients' conditions within the batch.

[0041] Batch statistics vectors are input into a feedforward network to generate node-by-node recalibration coefficients. The feedforward network employs a multi-layer fully connected structure, using a nonlinear transformation to map batch statistics into a recalibration coefficient vector equal to the number of nodes. Each coefficient corresponds to a node and is used to adjust the importance weight of the node's features.

[0042] The generated recalibration coefficients are applied to the enhanced features of each node. Through element-wise multiplication, the recalibration coefficients scale and adjust each dimension of the node features. The recalibration process enhances relatively important node features within the overall batch context, while moderately suppressing relatively minor features. The recalibrated node features incorporate the overall batch statistical information, forming a node enhanced feature representation that retains individual characteristics while reflecting the group context.

[0043] Within each attention head of the graph attention network, the change in cosine similarity between node pairs aggregated from the current layer is fed back to update the edge weights, allowing the edge weights to be iteratively and adaptively refined as message propagation occurs, including: After the message propagation of the current round is completed, the cosine similarity of all node pairs is calculated based on the node aggregation output of the current round, and the difference between the cosine similarity of the current round and the cosine similarity of the previous round is used to obtain the change in cosine similarity of node pairs. Asymmetric learnable gain coefficients are applied to the cosine similarity change of node pairs according to their positive and negative polarities. The positive polarity change is amplified by the enhancement coefficient to strengthen the coupling relationship of node pairs, while the negative polarity change is reduced by the suppression coefficient to weaken the coupling relationship of node pairs. The polarity-aware weighted change is used to perform multiplicative updates on the current edge weights. After the update, the sum of the outgoing edge weights of each node is used as the denominator to perform row normalization on all outgoing edge weights of the node, eliminating the influence of neighborhood size differences on the absolute magnitude of edge weights, while retaining the relative weight relationship between outgoing edges of the same node, thus obtaining the updated edge weights. The updated edge weights are logarithmically summed and added to the corresponding nodes to score the attention. Structural modulation is completed before softmax is executed. The enhancement coefficient, suppression coefficient, and logarithmic scaling coefficient are used as three mutually decoupled learnable scalars for joint training. The differential amplitude of positive enhancement and negative suppression is adaptively controlled, so that the dynamic structure of edge weights and attention coefficients evolve together during the iteration process.

[0044] In this specific embodiment, when the graph attention network performs multi-layer message propagation, the edge weights are not fixed, but are iteratively and adaptively refined based on the spatial relationship changes after node feature aggregation. Within each attention head, after the current layer completes one round of message propagation, all node representation vectors output by the current layer are extracted, and cosine similarity is calculated for each pair of nodes in the waiting patient relationship graph. Specifically, the current layer output feature vectors of nodes i and j are... and The cosine similarity is calculated by dividing the dot product of the two vectors by the product of their respective magnitudes. The change in the cosine similarity of the node pairs is obtained by subtracting the corresponding value stored in the previous round from the cosine similarity calculated in the current round. This change reflects the trend of increasing or decreasing proximity between node pairs in the semantic space after one layer of message propagation.

[0045] To address the positive or negative polarity of the cosine similarity change, an asymmetric learnable gain coefficient is applied when... A value greater than zero indicates that the node pairs are more similar in features after information exchange, and in this case, the enhancement coefficient is applied. Multiplicative amplification is performed to strengthen the coupling relationship between the node pairs; when When the value is less than zero, it indicates that the feature differences between nodes are amplified, and the suppression coefficient is applied. Multiplicative reduction weakens the coupling relationship between the node pairs. The change after polarity-aware weighting is expressed as follows: or The polarity-aware weighted change is compared with the current edge weight w. ij Perform a multiplicative update, that is ,in Select the corresponding coefficient based on the polarity.

[0046] Because the neighborhood sizes of different nodes in the waiting patient relationship graph vary, the absolute magnitude of the edge weights after direct updates will be affected by the number of neighbors. To eliminate this effect, row normalization is performed on all outgoing edge weights of each node. Specifically, for node i, the sum of all its outgoing edge weights is calculated as the normalization denominator, and the edge weights of each node pointing to its neighboring nodes are divided by this denominator to obtain the normalized edge weights. Row normalization ensures that the total sum of outgoing edge weights of the same node remains constant, eliminating the influence of neighborhood size on the absolute magnitude of the weights, while preserving the relative weight relationships between outgoing edges of the same node.

[0047] The updated edge weights need to be applied to the attention calculation process. Within the attention head, the attention score for a node is obtained by the dot product of the query vector and the key vector. The natural logarithm of the updated edge weights is taken, and the logarithmic value is used to additively modulate the attention score; that is, the logarithmic edge weights are added to the original attention score before softmax normalization. The logarithmic transformation converts the multiplicative relationship into an additive one, avoiding gradient instability caused by the direct multiplicative effect of edge weights, and introducing a logarithmic scaling factor. The overall strength of the structural modulation of the edge weights is controlled. This coefficient, together with the enhancement coefficient and the inhibition coefficient, serves as three mutually decoupled learnable scalar parameters. During model training, the three scalars are jointly optimized through backpropagation, adaptively learning the differential amplitude of positive reinforcement and negative inhibition, as well as the contribution ratio of structural modulation to the original attention. This allows the dynamic structure of the edge weights and the attention coefficient to co-evolve in multi-layer propagation, ultimately improving the ability of node enhancement features to represent the differences in the urgency of waiting patients.

[0048] The historical visit feature sequences of each waiting patient are concatenated with the corresponding node enhancement features. A variational encoder with learnable positional encoding is introduced to construct a KL divergence sequence to quantify the contribution differences of each historical visit feature, generating temporal enhancement priority features, including: Each historical visit feature in the historical visit feature sequence of each waiting patient is concatenated with the corresponding node augmented feature according to the feature dimension, and a learnable location code is superimposed to obtain a location-aware concatenated feature sequence. Based on the variational encoder and the location-aware spliced ​​feature sequence, the random latent representation of each historical visit feature is obtained by reparameterized sampling, and the mean and logarithmic variance of the latent distribution are calculated. At the same time, the target distribution is generated by learningable projection of the corresponding node enhancement features, and the KL divergence sequence between the latent distribution and the target distribution of each historical visit feature is constructed. The contribution weights of each historical visit feature are obtained by normalizing the KL divergence sequence after negative value transformation. KL divergence is the similarity of the information compression direction between each historical visit feature and the current patient status. The contribution weights quantitatively characterize the correlation differences between each historical visit feature and the priority feature model from the perspective of information theory. The random latent representations of each historical visit time are weighted and aggregated using contribution weights, and then mapped to temporal enhanced priority features of each waiting patient by a learnable decoder.

[0049] In generating temporal enhancement priority features, the historical medical records of each waiting patient need to be arranged into a sequence by timestamp. Each historical record includes a 768-dimensional semantic vector of the chief complaint text encoded by BERT, a 12-dimensional vector consisting of physiological parameters such as body temperature and blood pressure measured at the time, and a unique-hot encoding of the registered department. These three are concatenated to form a single historical medical record feature. Assuming a patient has 5 historical medical records, a feature sequence of length 5 is formed, with each time step feature dimension of 792. The features of each time step in this sequence are directly concatenated with the node enhancement features output by the aforementioned graph attention network in terms of feature dimension. The node enhancement feature dimension is 256, and after concatenation, each time step feature is expanded to 1048 dimensions. To enable the model to perceive the temporal positional relationship of the historical records, a learnable positional encoding vector is superimposed on the concatenated sequence. This positional encoding adopts a sine-cosine combination form with the same dimension as the Transformer structure, but the encoding coefficient matrix is ​​obtained through backpropagation learning, so that the positional representation can adapt to the uneven time intervals in the medical scenario, resulting in a position-aware concatenated feature sequence.

[0050] The variational encoder employs a two-layer LSTM architecture with 512 hidden units. The location-aware concatenated feature sequence is input into the encoder, and the hidden states at the final time step are mapped to 128-dimensional latent mean and log-variance vectors via two parallel fully connected layers. Using a reparameterization technique, noise vectors are sampled from a standard normal distribution, multiplied by the standard deviation (after taking the exponent of the log-variance), and then added to the mean to generate random latent representations corresponding to each historical visit feature. This process is performed at each time step, forming a latent sequence consisting of five 128-dimensional latent representations. Simultaneously, the enhanced features of the current node are input into a learnable projection layer, which consists of two fully connected layers with ReLU activation function in the middle layer. The output dimensions are also 128-dimensional mean and log-variance, defined as the target distribution parameters.

[0051] Calculate the KL divergence between the potential distribution and the target distribution at each historical time step. Specifically, let the mean of the potential distribution at the $t$-th historical time be... variance is The target distribution mean is variance is The KL divergence is calculated as a combination of the difference between the logarithmic terms of the two distributions and the mean and variance terms. This divergence value quantifies the information loss required to compress historical medical visit states to the current state; the smaller the divergence, the more relevant the historical feature is to the current state. A divergence sequence is constructed from the KL divergence values ​​of five historical time points. After negativeing ​​each element and normalizing using the softmax function, five contribution weights are obtained, with a sum of 1. Historical time points with higher contribution weights have a more significant impact on the current priority assessment.

[0052] Five random latent representations are weighted and summed using contribution weights to obtain a 128-dimensional fused latent vector. This vector is input to a learnable decoder, which employs a three-layer fully connected network with intermediate layers of dimensions 256 and 512, using ReLU and Tanh activation functions respectively. The final mapping is a 256-dimensional temporal enhanced priority feature. This feature simultaneously integrates the temporal dependence of patients' historical medical history, the correlation between patients in the same batch, and the differences in historical contributions based on information theory. This provides a high-quality feature representation for subsequent priority scoring, enabling urgency assessment to consider both individual historical patterns and the current group status.

[0053] The target distribution parameters are generated by using the corresponding node enhancement features through learnable projection, and the KL divergence sequence between the latent distribution and the target distribution of each historical visit feature is constructed, including: The node enhancement features of each waiting patient are decomposed in a dual-path parallel manner along the feature dimension. They are mapped to semantic target distribution parameters and state target distribution parameters through two sets of parameter-independent learnable projection networks. The corresponding mean vector and log variance vector are linearly combined by learnable fusion weights to integrate them into a unified target distribution parameter, which represents the current patient's comprehensive prior state information in the form of a parameterized probability distribution. Using the mean vector and log-variance vector of the unified target distribution parameters, combined with the mean and log-variance of the latent distribution independently inferred by the variational encoder for each historical consultation time, the forward KL divergence from the latent distribution to the target distribution and the backward KL divergence from the target distribution to the latent distribution are calculated for each historical consultation time. The symmetric bidirectional information bias metric is obtained by weighted superposition of learnable coefficients. The bidirectional information deviation metric values ​​corresponding to each historical visit time are arranged in the order of visit time. The amplitude of each divergence value is normalized by a learnable nonlinear correction network to eliminate the dimensional bias introduced by the difference in variance of the potential distribution at each time. This forms a KL divergence sequence that quantitatively measures the degree of information deviation between the potential distribution of each historical visit feature and the target distribution of the current patient status.

[0054] For the node enhancement features of waiting patients, a dual-path parallel decomposition operation is performed along their feature dimensions. Specifically, the node enhancement feature tensor has the following dimensions: Divided into front through slicing operation Wei and Hou Two sub-feature vectors are used. The first sub-feature vector is input into a semantic projection network, which consists of three fully connected layers with 512, 256, and 128 neurons respectively, using GELU as the activation function. The second sub-feature vector is input into a state projection network, which has the same structure as the semantic projection network but its parameters are trained independently. The two networks output a 64-dimensional mean vector and a 64-dimensional log-variance vector, respectively. The semantic target distribution parameters are denoted as... , The state target distribution parameters are denoted as , Introducing learnable fusion weights and The mean vector of the uniform target distribution is calculated through linear combination. With log-variance vector ,in and During training, the target distribution is automatically updated via gradient descent to conform to a normal distribution. It represents the patient's current prior state.

[0055] The variational encoder encodes the historical visit feature sequence of the waiting patient, which includes the patient's medical records over the past 12 months. Each record includes the department code, embedded chief complaint keywords, and physiological indicator values, forming a temporal feature matrix. The variational encoder uses a two-layer LSTM structure with a hidden layer dimension of 256, encoding each historical visit time t. i Output the corresponding latent distribution parameters and For the i-th historical visit time, calculate the forward KL divergence. with reverse KL divergence Forward divergence measures the degree to which historical features deviate from the current state, while backward divergence measures the sufficiency of the current state's coverage of historical features. Learnable coefficients are introduced. and Calculate the symmetric two-way information deviation metric. The initial value is set to Adaptive adjustment during training.

[0056] The $D_{sym}(t_i)$ corresponding to each historical visit time is arranged chronologically to form the original divergence sequence. Since the variance differences in the latent distribution at different times can lead to inconsistent dimensions of the divergence values, a nonlinear correction network is introduced for normalization transformation. This network adopts a residual gating unit structure, taking the original divergence sequence as input and dynamically adjusting the amplitude of each divergence value through a gating mechanism. Specifically, a sliding window statistic is calculated on the original divergence sequence with a window size of 5 to obtain the local mean and standard deviation. Each divergence value is subtracted from the local mean and divided by the local standard deviation, then mapped to the 0-1 interval using the Sigmoid function to eliminate the dimensional bias introduced by variance differences. The corrected KL divergence sequence quantitatively reflects the degree of information deviation between each historical visit feature and the current patient status; a larger divergence value indicates a lower contribution weight of the corresponding historical visit feature to the current priority assessment.

[0057] Multiple random inactivation forward inferences are performed on the enhanced priority features of the waiting time sequence to calculate the waiting priority score. Based on the waiting priority score, the waiting queue is dynamically reordered, and a consultation room allocation instruction is generated and synchronized to the triage workstation, including: Before each random inactivation forward inference, a random perturbation with controlled amplitude is applied to the temporal position code in the waiting time-series enhanced priority feature, so that random inactivation and position code perturbation work together to expand the output diversity of forward inference. Multiple collaborative double randomization forward inferences are performed, and the stability of the relative priority ranking of each waiting patient within the batch in all inference times is calculated. The weighted product of the inference output mean and the ranking stability coefficient is used as the waiting priority score. Based on the waiting priority score, a ranking weight is calculated for each patient in the waiting queue. The ranking stability coefficient of each patient is used as a weight reliability factor to further correct the ranking weight. Based on the corrected ranking weight, the waiting queue is dynamically reordered to generate a dynamic ranking result. Based on the dynamic sorting results, the real-time load of each clinic is matched to generate a clinic allocation instruction, which is then synchronized to the triage workstation.

[0058] In practical deployment, a dual randomization mechanism is introduced during the forward inference process of the waiting time-series enhanced priority features to improve scoring robustness. Specifically, the waiting time-series enhanced priority features are denoted as a vector f of dimension d, which contains the temporal position encoding component p. Before performing the k-th forward inference, a perturbation is applied to the position encoding: ,in Follows a standard normal distribution. The amplitude control factor is set to 0.05 to 0.15 to ensure that the perturbation does not disrupt the temporal structure. Simultaneously, random deactivation is activated in the fully connected layers of the inference network, with a deactivation rate set to 0.3 to 0.5. Through this double randomization operation, the output s of the k-th inference is optimized. k While retaining core features, a moderate degree of randomness is introduced.

[0059] Perform 20 to 50 rounds of collaborative double-randomized forward inference and collect all inference outputs. For N patients waiting in the same batch, N scores are generated for each inference. The ranking of each patient within the batch in each inference is statistically analyzed, and the stability of the ranking is measured by calculating the standard deviation of the ranking. A ranking stability coefficient is defined. ,in This represents the standard deviation of the ranking; the closer this coefficient is to 1, the smaller the ranking fluctuation. Calculate the mean of all inferred outputs. The waiting priority score is defined as follows: The weighting coefficients The threshold is set at 1.2 to 1.5 based on clinical needs, so that patients with stable rankings and high scores can be given higher priority.

[0060] Based on the calculated waiting priority score, a preliminary ranking weight is calculated for each patient in the waiting queue. The ranking weight is in an exponentially normalized form. , To ensure that all weights sum to 1, and to further improve the reliability of the ranking, a ranking stability coefficient is introduced as a weight credibility factor for secondary correction. The corrected ranking weights are: The coefficient design suppresses the weighting of patients with low stability coefficients, avoiding erroneous rankings caused by inference fluctuations. This is based on the corrected ranking weights. The waiting queue is sorted in descending order, generating a dynamic ranking result. This result is refreshed every 5 to 10 minutes, and the queue order is adjusted in real time based on newly arriving patients or changes in their conditions.

[0061] After generating the dynamic sorting results, waiting patients need to be rationally allocated to various consultation rooms. Real-time load indicators for each consultation room are obtained, including the current number of waiting patients, average consultation time, and remaining doctor workload. A load balancing strategy is adopted to prioritize allocating high-priority patients to consultation rooms with lower current load and matching departments. Specifically, a matching score is calculated for each patient and each consultation room. Considering departmental consistency, inverse ratio of consultation room load, and patient priority weight, the consultation room with the highest matching score is selected to generate an allocation instruction. The instruction format includes the patient identifier, target consultation room number, and estimated consultation time window. The generated consultation room allocation instruction is synchronized in real-time to the triage workstation display terminal via a network interface. The workstation updates the call number screen and voice broadcast system according to the instruction, guiding patients to the designated consultation room according to the allocation results, thus achieving intelligent closed-loop management of the triage process.

[0062] This invention also provides an optional specific embodiment in which a handheld mobile triage device acquires the patient's registration information, symptom self-report text, and physiological characteristic data through an integrated multimodal sensor. The registration information includes the appointment department, doctor's number, appointment time, and basic patient information, which is obtained by scanning the QR code on the medical card or by directly retrieving it from the hospital information system. The symptom self-report text is collected by the patient's oral description of symptoms through voice recognition technology, or by medical staff directly inputting it through a touch screen interface. The physiological characteristic data includes key indicators such as blood pressure, heart rate, and body temperature, which are measured in real time by built-in physiological parameter sensors. The handheld device directly uploads the collected raw data to the backend server for processing.

[0063] The backend server performs preprocessing on the received symptom description text, removing punctuation and meaningless interjections. A dictionary-based word segmentation method is used to divide the text into lexical units. A dedicated dictionary for medical terminology is established to improve segmentation accuracy. After segmentation, each word is mapped to a corresponding word vector representation. The word vectors are obtained through a pre-trained medical language model with a 512-dimensional dimension. Average pooling is then performed on the entire sequence of word vectors describing the symptoms to obtain a fixed-dimensional semantic feature vector.

[0064] Registration information is converted into numerical feature vectors through encoding, department information is represented using one-hot encoding, doctor IDs are converted into embedded vectors, and consultation time is converted into timestamp numerical features. Physiological feature data is standardized, converting each indicator value into a standardized value between 0 and 1. The semantic feature vector, registration information feature vector, and physiological feature vector are combined into a complete waiting feature vector through vector concatenation, with a total dimension of 768. Each waiting feature vector within the same waiting batch serves as a node feature in the waiting patient relationship graph. The cosine similarity of the symptom semantic components of the feature vectors corresponding to any two nodes is calculated to obtain the symptom similarity value. A consistency measure of department affiliation is calculated based on the patient's registered department information, with the same department assigned 1, different departments assigned 0, and departments with collaborative relationships assigned an intermediate value. After concatenating the symptom similarity with the department affiliation consistency, a linear projection layer is used to map it into learnable edge weights, constructing a complete waiting patient relationship graph.

[0065] The input graph of the waiting patient relationship graph is processed by an attention network employing a multi-head attention mechanism with eight attention heads. Each attention head independently calculates the attention coefficients between nodes through linear transformations of query vectors, key vectors, and value vectors. Edge weight information is incorporated into the attention calculation process, and aggregated features of each head are generated by weighted aggregation of features from adjacent nodes. The multi-head aggregated features are concatenated and nonlinearly projected, and then residually connected to the original features to obtain the enhanced features of a single node. Batch normalization is performed on the enhanced features of all nodes within a batch, and the mean and variance statistics are calculated and retained. The mean and variance are concatenated to form a batch statistical vector, reflecting the overall feature distribution of the current batch of patients. The batch statistical vector is converted into node-by-node recalibration coefficients through a feedforward network, and the enhanced features of each node are recalibrated to obtain node enhanced features that integrate the overall information of the batch.

[0066] Historical visit feature sequences of each waiting patient are extracted from the electronic medical record database. Each historical visit record includes information such as symptom description, diagnosis, treatment plan, and visit time. These historical visit features are processed using the same feature extraction method as the waiting patient feature vector, generating a fixed-dimensional historical feature vector. The historical visit feature sequence is concatenated with the corresponding node augmentation features to form a comprehensive feature representation containing both current state and historical information. A learnable positional encoding technique is introduced to process temporal information; the positional encoding vector has the same dimension as the feature vector. The positional encoding is generated using sine and cosine functions, representing the temporal order of historical visit records. Incorporating the positional encoding into the historical visit feature vector enables the network to perceive the temporal location information of different historical records.

[0067] A variational encoder is constructed to process historical medical visit sequences containing location information. The encoder consists of two parts: an encoding network and a decoding network. The encoding network maps the input sequence to a mean vector and a variance vector, while the decoding network samples and reconstructs the original sequence from the latent space. The contribution difference of each historical medical visit feature to the current condition assessment is quantified by calculating the KL divergence between the reconstructed sequence and the original sequence. The KL divergence value reflects the correlation between historical records and the current state; the smaller the divergence, the higher the reference value of the historical records. Attention weights are constructed based on the KL divergence sequence to perform weighted fusion of historical medical visit features. The weight allocation gives higher attention to historical records more relevant to the current symptoms, while irrelevant historical information is moderately suppressed. The weighted fused historical features are further concatenated with node-enhanced features to generate temporal-enhanced priority features, which simultaneously include current symptom information, patient group background, and individual historical patterns. Multiple random inactivation forward inferences are performed on the temporal-enhanced priority features, with some neuron connections randomly discarded during each inference to increase the robustness of the prediction. Ten waiting priority scores were obtained through ten independent forward inferences. The mean of the scores was calculated as the final priority value, and the standard deviation of the scores was calculated as a measure of predictive uncertainty. The priority score ranges from 0 to 100, with higher values ​​indicating greater urgency.

[0068] The backend server pushes the priority scoring results to handheld devices and triage workstations. After receiving the scores, the handheld devices display the triage results on a touchscreen interface, visually indicating the priority level through color coding. The triage workstation dynamically reorders the current waiting queue based on the received priority scores, automatically moving higher-priority patients to the front of the queue. The sorting algorithm considers multiple factors such as priority scores, waiting time, and doctor expertise to ensure the rationality of the queue arrangement.

[0069] The triage workstation generates consultation room allocation instructions, which assign the reordered patients to the corresponding consultation rooms. The allocation strategy comprehensively considers the doctor's professional field, current workload, and the patient's symptom characteristics. The allocation instructions are synchronized to the display terminals and queuing devices in each consultation room through the hospital network, ensuring the smooth progress of the consultation process. The portability of handheld devices allows medical staff to complete data collection directly at the patient's bedside without having to travel back and forth between the triage station and computer terminal, improving triage efficiency and patient service quality.

[0070] This invention provides a triage priority intelligent assessment system based on deep temporal learning, comprising: The feature extraction unit is used to acquire the registration information, symptom self-report text and physiological feature data of the waiting patients, perform word segmentation and word vector mapping on the symptom self-report text to obtain semantic feature vectors, and concatenate the feature vectors with the registration information and the physiological feature data to obtain the waiting feature vector. The graph aggregation unit is used to take the feature vectors of each patient in the same waiting batch as node features, assign edge weights according to symptom similarity and the department to which the patient belongs, construct a patient relationship graph, input it into the graph attention network and aggregate the features of adjacent nodes through a multi-head attention mechanism to generate node enhancement features that are integrated into the whole batch. The temporal enhancement unit is used to concatenate the historical visit feature sequence of each waiting patient with the corresponding node enhancement feature, introduce a variational encoder with learnable positional encoding, construct a KL divergence sequence to quantify the contribution difference of each historical visit feature, and generate temporal enhancement priority features. The priority sorting unit is used to perform multiple random inactivation forward inferences on the enhanced priority features of the waiting time sequence, calculate the waiting priority score, perform dynamic reordering of the waiting queue based on the waiting priority score, generate a consultation room allocation instruction and synchronize it to the triage workstation.

[0071] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0072] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0073] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A triage priority intelligent assessment method based on deep temporal learning, characterized in that, include: The system obtains the registration information, symptom self-report text, and physiological feature data of patients waiting for treatment. It performs word segmentation and word vector mapping on the symptom self-report text to obtain semantic feature vectors. It then concatenates these feature vectors with the registration information and physiological feature data to obtain the waiting feature vector. The feature vectors of each patient in the same waiting batch are used as node features. Edge weights are assigned based on symptom similarity and the department to which the patient belongs. A patient relationship graph is constructed, which is then input into a graph attention network and aggregated with the features of adjacent nodes through a multi-head attention mechanism to generate node enhancement features that are integrated with the whole batch. The historical visit feature sequence of each waiting patient is concatenated with the corresponding node enhancement feature. A variational encoder with learnable positional encoding is introduced to construct a KL divergence sequence to quantify the contribution difference of each historical visit feature and generate temporal enhancement priority features. Multiple random inactivation forward inferences are performed on the waiting time sequence enhancement priority feature to calculate the waiting priority score. The waiting queue is dynamically reordered based on the waiting priority score, and a consultation room allocation instruction is generated and synchronized to the triage workstation.

2. The method according to claim 1, characterized in that, The symptom self-report text is segmented and word vector mapped to obtain a semantic feature vector. This semantic feature vector is then concatenated with the registration information and the physiological feature data to obtain a waiting feature vector, including: The symptom self-report text is subjected to medical noise filtering and semantic boundary word segmentation. Each word segmentation unit is mapped to a dedicated medical semantic space according to the word unit type, generating a word vector sequence carrying position encoding. Multi-granularity feature extraction is performed on the word vector sequence. At the word level, the local semantics of a single symptom word are captured. At the phrase level, the semantics of symptom combinations of adjacent words are aggregated. At the sentence level, the global semantics of the entire sentence's symptom description are integrated. Dimensionality is reduced through a linear projection layer, and the resulting semantic feature vectors are generated. Linear embedding mapping is performed on the registration information and the physiological feature data respectively to obtain registration feature vector and physiological feature vector; using the semantic feature vector as the query and the registration feature vector and physiological feature vector as key-value pairs respectively, cross-modal cross-attention calculation is performed to guide the response of feature dimensions that contribute highly to waiting priority in symptom semantic enhancement, thereby obtaining symptom-guided registration features and symptom-guided physiological features. The semantic feature vector, symptom-guided registration feature, and symptom-guided physiological feature are respectively subjected to residual enhancement and then sequentially concatenated to obtain a waiting feature vector that integrates three modal information.

3. The method according to claim 1, characterized in that, Each waiting feature vector within the same waiting batch is used as a node feature. Edge weights are assigned based on symptom similarity and department affiliation to construct a waiting patient relationship graph. This graph is then input into a graph attention network, and adjacent node features are aggregated using a multi-head attention mechanism to generate enhanced node features that fuse the entire batch. These features include: Each waiting patient is treated as a node in the waiting patient relationship graph. For any two nodes in the same waiting batch, the symptom similarity and the consistency of the department they belong to are calculated. The symptom similarity and the consistency of the department they belong to are concatenated and then mapped to learnable edge weights through linear projection to obtain the waiting patient relationship graph. Within each attention head of the graph attention network, the change in cosine similarity of the nodes aggregated at the current layer is fed back to update the edge weights, so that the edge weights are iteratively and adaptively refined as the message propagates; the attention coefficients are re-modulated with the updated edge weights and weighted aggregation is performed to obtain the node aggregation features within the head; the aggregation features within each head are concatenated, nonlinearly projected, and superimposed with the residual of the original waiting feature vector to obtain the node enhancement features of a single node. Batch normalization is performed on all single-node augmentation features within a batch, and the mean and variance statistics are retained. The mean and variance are concatenated to form a batch statistical vector. The batch statistical vector is used to generate node-by-node recalibration coefficients through a feedforward network. Recalibration is applied to the augmentation features of each node to obtain the node augmentation features of the entire batch.

4. The method according to claim 3, characterized in that, Within each attention head of the graph attention network, the change in cosine similarity between node pairs aggregated from the current layer is fed back to update the edge weights, allowing the edge weights to be iteratively and adaptively refined as message propagation occurs, including: After the message propagation of the current round is completed, the cosine similarity of all node pairs is calculated based on the node aggregation output of the current round, and the difference between the cosine similarity of the current round and the cosine similarity of the previous round is used to obtain the change in cosine similarity of node pairs. Asymmetric learnable gain coefficients are applied to the cosine similarity change of node pairs according to their positive and negative polarities. The positive polarity change is amplified by the enhancement coefficient to strengthen the coupling relationship between node pairs, while the negative polarity change is reduced by the suppression coefficient to weaken the coupling relationship between node pairs. The current edge weights are then multiplicatively updated using the polarity-aware weighted change. After the update, the sum of the weights of the outgoing edges of each node is used as the denominator to perform row normalization on all outgoing edge weights of that node, eliminating the influence of neighborhood size differences on the absolute magnitude of edge weights, while retaining the relative weight relationship between outgoing edges of the same node, thus obtaining the updated edge weights. The updated edge weights are logarithmically summed and added to the corresponding nodes to score the attention. Structural modulation is completed before softmax is executed. The enhancement coefficient, suppression coefficient, and logarithmic scaling coefficient are used as three mutually decoupled learnable scalars for joint training. The differential amplitude of positive enhancement and negative suppression is adaptively controlled, so that the dynamic structure of edge weights and attention coefficients evolve together during the iteration process.

5. The method according to claim 1, characterized in that, The historical visit feature sequences of each waiting patient are concatenated with the corresponding node enhancement features. A variational encoder with learnable positional encoding is introduced to construct a KL divergence sequence to quantify the contribution differences of each historical visit feature, generating temporal enhancement priority features, including: Each historical visit feature in the historical visit feature sequence of each waiting patient is concatenated with the corresponding node augmented feature according to the feature dimension, and a learnable location code is superimposed to obtain a location-aware concatenated feature sequence. Based on the variational encoder and the location-aware spliced ​​feature sequence, the random latent representation of each historical visit feature is obtained by reparameterized sampling, and the mean and logarithmic variance of the latent distribution are calculated. At the same time, the target distribution is generated by learningable projection of the corresponding node enhancement features, and the KL divergence sequence between the latent distribution and the target distribution of each historical visit feature is constructed. The contribution weights of each historical visit feature are obtained by normalizing the KL divergence sequence after negative value transformation. KL divergence is the similarity of the information compression direction between each historical visit feature and the current patient status. The contribution weights quantitatively characterize the correlation differences between each historical visit feature and the priority feature model from the perspective of information theory. The random latent representations of each historical visit time are weighted and aggregated using contribution weights, and then mapped to temporal enhanced priority features of each waiting patient by a learnable decoder.

6. The method according to claim 5, characterized in that, The target distribution parameters are generated by using the corresponding node enhancement features through learnable projection, and the KL divergence sequence between the latent distribution and the target distribution of each historical visit feature is constructed, including: The node enhancement features of each waiting patient are decomposed in a dual-path parallel manner along the feature dimension. They are mapped to semantic target distribution parameters and state target distribution parameters through two sets of parameter-independent learnable projection networks. The corresponding mean vector and log variance vector are linearly combined by learnable fusion weights to integrate them into a unified target distribution parameter, which represents the current patient's comprehensive prior state information in the form of a parameterized probability distribution. Using the mean vector and log-variance vector of the unified target distribution parameters, combined with the mean and log-variance of the latent distribution independently inferred by the variational encoder for each historical consultation time, the forward KL divergence from the latent distribution to the target distribution and the backward KL divergence from the target distribution to the latent distribution are calculated for each historical consultation time. The symmetric bidirectional information bias metric is obtained by weighted superposition of learnable coefficients. The bidirectional information deviation metric values ​​corresponding to each historical visit time are arranged in the order of visit time. The amplitude of each divergence value is normalized by a learnable nonlinear correction network to eliminate the dimensional bias introduced by the difference in variance of the potential distribution at each time. This forms a KL divergence sequence that quantitatively measures the degree of information deviation between the potential distribution of each historical visit feature and the target distribution of the current patient status.

7. The method according to claim 1, characterized in that, Multiple random inactivation forward inferences are performed on the enhanced priority features of the waiting time sequence to calculate the waiting priority score. Based on the waiting priority score, the waiting queue is dynamically reordered, and a consultation room allocation instruction is generated and synchronized to the triage workstation, including: Before each random inactivation forward inference, a random perturbation with controlled amplitude is applied to the temporal position code in the waiting time-series enhanced priority feature, so that random inactivation and position code perturbation work together to expand the output diversity of forward inference. Multiple collaborative double randomization forward inferences are performed, and the stability of the relative priority ranking of each waiting patient within the batch in all inference times is calculated. The weighted product of the inference output mean and the ranking stability coefficient is used as the waiting priority score. Based on the waiting priority score, a ranking weight is calculated for each patient in the waiting queue. The ranking stability coefficient of each patient is used as a weight reliability factor to further correct the ranking weight. Based on the corrected ranking weight, the waiting queue is dynamically reordered to generate a dynamic ranking result. Based on the dynamic sorting results, the real-time load of each clinic is matched to generate a clinic allocation instruction, which is then synchronized to the triage workstation.

8. A triage priority intelligent assessment system based on deep temporal learning, used to implement the method as described in any one of claims 1-7, characterized in that, include: The information acquisition unit is used to acquire the registration information, symptom self-report text and physiological feature data of the waiting patients, perform word segmentation and word vector mapping on the symptom self-report text to obtain semantic feature vectors, and concatenate the feature vectors with the registration information and the physiological feature data to obtain the waiting feature vector. The graph construction unit is used to take the feature vectors of each patient in the same waiting batch as node features, assign edge weights according to symptom similarity and the department to which the patient belongs, construct a patient relationship graph, input it into the graph attention network and aggregate the features of adjacent nodes through a multi-head attention mechanism to generate node enhancement features that are integrated into the whole batch. The temporal enhancement unit is used to concatenate the historical visit feature sequence of each waiting patient with the corresponding node enhancement feature, introduce a variational encoder with learnable positional encoding, construct a KL divergence sequence to quantify the contribution difference of each historical visit feature, and generate temporal enhancement priority features. The priority calculation unit is used to perform multiple random inactivation forward inferences on the waiting time sequence enhanced priority features, calculate the waiting priority score, perform dynamic reordering of the waiting queue based on the waiting priority score, generate a consultation room allocation instruction and synchronize it to the triage workstation.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images