Pediatric registration intelligent management method and device based on space-time attention mechanism model

Abstract

The application provides a pediatric registration intelligent management method and device based on a space-time attention mechanism model, and relates to the field of artificial intelligence. The method comprises the following steps: first, obtaining multi-source data of sick children, fusing multi-dimensional information such as text, image and audio through a cross-modal attention mechanism to generate disease representation features; then, extracting spatial topology features from the hospital department topology, knowledge graph and expert experience, and extracting disease cycle fluctuation and long-term dependence features from the time dimension; then, dynamically correlating the space-time features through an adaptive gating network, adjusting the fusion weight according to the current treatment demand, and generating comprehensive decision features; finally, outputting a four-dimensional registration suggestion containing recommended departments, doctors, time periods and optimal paths. The application overcomes the defects of the prior art, such as the inability to deeply utilize multi-source data, the neglect of space-time joint modeling and the lack of dynamic adaptation, and significantly improves the accuracy of pediatric registration recommendation and the efficiency of resource scheduling.

Description

Technical Field

[0001] This specification relates to the field of artificial intelligence, specifically to a method and device for intelligent management of pediatric registration based on a spatiotemporal attention mechanism model. Background Technology

[0002] In pediatric registration management, children's conditions are often changeable and their expressions are unclear, leading to subjective judgments by parents that can result in incorrect registrations or referrals. Current technologies lack deep integration of multi-source heterogeneous data such as medical records, images, and audio, making it difficult to accurately characterize the condition; they ignore the spatial topology of hospital departments and the time cycle of diseases, resulting in rigid allocation of appointment slots, ineffective patient movement, and long waiting times; the models are not robust to noise, missing data, and sudden changes in condition, and cannot dynamically adjust registration strategies according to the demand for medical care, which can easily delay the treatment of children with acute illnesses.

[0003] Therefore, how to automatically generate personalized registration suggestions that include department, doctor, time period and optimal route, and realize real-time correction of resource scheduling through dynamic adaptive gating mechanism, has become a technical problem that urgently needs to be solved in the field of smart healthcare and pediatric intelligent triage. Summary of the Invention

[0004] This invention describes a method and device for intelligent management of pediatric registration based on a spatiotemporal attention mechanism model, which can solve the above-mentioned technical problems.

[0005] Firstly, a method for intelligent management of pediatric appointment registration based on a spatiotemporal attention mechanism model is provided, including: Obtain the child's original multi-source data, which should include at least multi-dimensional information reflecting the child's health background, current condition, and living environment. The original multi-source data is input into the trained spatiotemporal attention mechanism model. The model fuses the multi-source data through a cross-modal attention mechanism and dynamically adjusts the attention weights of different data sources according to the characteristics of the data itself, and outputs the fused disease manifestation features. The model extracts disease manifestation features from both spatial and temporal dimensions to obtain spatial topological features and temporal periodic features. Spatial topological features are characteristics that reflect the distribution and correlation of departmental resources after integrating the hospital department topology, medical knowledge graph and clinical expert experience. Temporal periodic features are characteristics that reflect the changes in medical needs over time based on the periodic fluctuations of the disease and the long-term dependence of historical medical data. By using an adaptive gating network, spatial topological features and temporal periodic features are dynamically correlated, and the fusion weights of the two are adjusted in real time according to the current medical needs to generate comprehensive decision features. Based on comprehensive decision-making characteristics, registration suggestions are generated, including recommended departments, recommended doctors, recommended appointment times, and the optimal route from the entrance to the department.

[0006] In this way, by fusing multi-source heterogeneous data through a cross-modal attention mechanism, a precise representation of the child's condition is achieved; by fusing departmental topology and expert experience through a spatial topological graph attention network, and capturing the disease cycle pattern through a time-dimensional hierarchical coding and recursive mechanism, the shortcomings of traditional methods that neglect spatiotemporal joint modeling are overcome; furthermore, by dynamically adjusting the spatiotemporal feature fusion weights through an adaptive gating network, personalized and real-time registration decisions are achieved; finally, four-dimensional registration suggestions are generated, which significantly improves the accuracy of pediatric registration and the efficiency of resource scheduling.

[0007] Among some possible implementations, spatiotemporal attention mechanism models include: The multimodal attention fusion unit is used to perform cross-modal feature fusion on input data from different modalities and dynamically adjust attention weights; The spatial dimension extraction unit includes a topological graph attention network, which is used to perform graph attention calculation on the fused features with the hospital department topology, medical knowledge graph and expert experience, and output spatial topological features. The time dimension extraction unit includes a hierarchical spatiotemporal coding structure and recursive connections. It is used to align the fused features with historical time series data and extract the dependencies of different time spans, outputting time period features. The adaptive gating fusion unit is used to dynamically associate spatial topological features and temporal periodic features according to the current medical needs and adjust the fusion weights to generate comprehensive decision features. The output unit is used to generate registration suggestions based on comprehensive decision features.

[0008] This approach divides the model into functionally independent units, facilitating modular implementation and optimization while ensuring low coupling and high cohesion between the units.

[0009] In some possible implementations, the raw multi-source data is input into a trained spatiotemporal attention mechanism model. The model then fuses the multi-source data through a cross-modal attention mechanism, dynamically adjusts the attention weights of different data sources based on the data's own characteristics, and outputs the fused disease representation features, including: The original multi-source data is divided into text modal data, image modal data, and audio modal data; For text modal data, semantic features are extracted using a text encoder; For image modal data, visual features are extracted by an image encoder, and self-attention weights are applied to the visual features to focus on specific regions in the image; For audio modal data, the audio signal is converted into a spectrogram and then the acoustic features are extracted using an audio encoder. Using textual semantic features as query vectors and visual and acoustic features as key-value pairs, attention weights for each modality feature are calculated through a cross-modal attention mechanism to obtain preliminary fused features; By using an attention head gating network, the activation intensity of each attention head in the cross-modal attention mechanism is dynamically adjusted according to the sequence length or complexity of text semantic features, visual features, and acoustic features, and the dynamically weighted fused features are output as disease representation features.

[0010] In this way, by using text-led cross-modal attention fusion and leveraging gating networks to dynamically adjust the intensity of attention heads, the problems of difficult alignment of multi-source heterogeneous data and easy obscuring of key information are effectively solved, thus improving the accuracy of disease representation.

[0011] In some possible implementations, the model performs in-depth extraction of disease manifestation features from spatial and temporal dimensions to obtain spatial topological features and temporal periodic features, including: inputting the disease manifestation features into spatial dimension extraction units and temporal dimension extraction units respectively; In the spatial dimension extraction unit, a pre-constructed hospital department topology graph structure is obtained. The nodes of the topology graph are departments or medical areas, and the edges represent the spatial associations and business connections between departments. The disease representation features, the topology graph structure, the external medical knowledge graph, and the expert experience embedding vectors are jointly input into the topology graph attention network. The topology graph attention network calculates the attention weights of the neighboring nodes of each node and outputs spatial topology features that aggregate spatial association information. In the time dimension extraction unit, a time series consisting of historical registration data and disease monitoring data is obtained. The disease manifestation features and the time series are input into a hierarchical spatiotemporal coding structure, and the time period features are output.

[0012] In this way, by jointly modeling the individual conditions of children with static spatial topology and dynamic time series, the spatial features can reflect the real collaborative relationships and real-time load between departments, and the temporal features can capture the periodic patterns of disease evolution, providing a high-quality spatiotemporal feature foundation for subsequent dynamic fusion.

[0013] In some possible implementations, disease-related features and time-series sequences are input into a hierarchical spatiotemporal coding structure, and time-periodic features are output, including: The disease manifestation features are copied into vectors with the same number of time steps as the time series, and each copied vector is concatenated with the historical data vector of the corresponding time step to form an aligned input sequence; The aligned input sequence is fed into the bottom layer of the hierarchical spatiotemporal coding structure in chronological order. The bottom layer encodes the position of each time step through timestamp encoding and outputs the bottom layer feature sequence. The low-level feature sequence is input into the high-level layer of the hierarchical spatiotemporal coding structure. The high-level layer extracts features within the local time window through a convolutional neural network, and then calculates the dependency weights between different time steps through a self-attention mechanism to output the high-level feature sequence. In a hierarchical spatiotemporal coding structure, recursive connections are set between encoder layers to pass the previous hidden state of the current layer to the same layer in the next time step and update the hidden state. The final hidden state after recursive connection is output as a time periodic feature.

[0014] In this way, by combining fine-grained recording at the bottom level with abstract extraction at the top level, and introducing cross-layer recursive connections, the model can capture both short-term fluctuation details and learn long-term dependent trends, significantly improving its ability to express time cycle features.

[0015] In some possible implementations, an adaptive gating network dynamically correlates spatial topological features with temporal periodic features, and adjusts the fusion weights of the two in real time according to the current medical needs to generate comprehensive decision features, including: The characteristics of the current medical needs are obtained, including the urgency of the child's condition and the real-time busy status of the selected department. Spatial topological features and temporal periodic features are respectively input into the feature fusion gating network. The feature fusion gating network calculates spatial fusion weights and temporal fusion weights based on the feature representation of the current medical needs. The spatial topological features and temporal periodic features are weighted and summed according to spatial fusion weights and temporal fusion weights to generate comprehensive decision features.

[0016] In this way, by using a gating network to dynamically adjust the fusion ratio of spatial and temporal characteristics according to real-time consultation needs, critically ill children can get priority access to the nearest spatial resources, while regular children pay more attention to time patterns, thus realizing personalized and adaptive registration decisions.

[0017] In some possible implementations, registration suggestions are generated based on comprehensive decision-making characteristics, including: The comprehensive decision-making features are input into the output unit, and the output unit includes department recommendation, doctor recommendation, time period recommendation and route planning. The department recommendation section outputs a department recommendation score based on the department's corresponding specialty and current patient saturation. The doctor recommendation section is based on the doctor's area of ​​expertise, the degree of match between the patient's symptoms and the patient's area of ​​expertise, patient reviews, and the doctor recommendation score generated by the daily schedule. The time slot recommendation section is based on the busyness score of each future time slot and the sorting of time slots output by checking the idle status of the equipment; The path planning part is based on the hospital's three-dimensional spatial layout data. Starting from the current entrance and ending at the recommended department location, it generates the optimal path through a path search algorithm. The outputs from the department recommendation section, doctor recommendation section, time slot recommendation section, and route planning section are combined into registration suggestions.

[0018] In this way, the four-dimensional recommendation results are output in parallel, providing parents with one-stop information for registration decisions. This avoids parents having to repeatedly compare information from multiple sources, significantly improving registration efficiency and the medical experience.

[0019] In some possible implementations, the trained spatiotemporal attention mechanism model is obtained by training in the following way: Acquire unlabeled spatiotemporal data, which includes historical registration data, hospital department topology data, and disease monitoring data; A self-supervised learning strategy is employed to pre-train the initial spatiotemporal attention mechanism model. Specifically, the hospital department topology data is masked, randomly masking some department nodes or edge relationships. The masked topology data is then input into the spatial dimension extraction unit. The topology graph attention network in the spatial dimension extraction unit predicts the masked node or edge relationships based on the remaining visible node and edge information. The first loss is calculated based on the difference between the prediction result and the original data, and the parameters of the spatial dimension extraction unit are updated, enabling the model to learn the potential spatial associations and collaboration patterns between hospital departments. Similarly, historical registration data is masked, randomly masking registration data for future time periods. The masked historical registration data is then input into the temporal dimension extraction unit. The hierarchical spatiotemporal encoding structure and recursive connections in the temporal dimension extraction unit predict the registration data for the masked time periods based on the historical visible data. The second loss is calculated based on the difference between the prediction result and the original data, and the parameters of the temporal dimension extraction unit are updated, enabling the model to learn the temporal evolution patterns of disease cycle fluctuations and patient demand. The parameters of the initial spatiotemporal attention mechanism model are jointly optimized based on the first and second losses to obtain the pre-trained model.

[0020] Through self-supervised learning strategies, the model can extract general spatiotemporal feature representations from massive amounts of unlabeled data, thereby enabling it to have stronger generalization and feature understanding capabilities in subsequent actual registration tasks. An adversarial training mechanism is employed to enhance the pre-trained model. Specifically, a generator and a discriminator are constructed. Unlabeled spatiotemporal data is input into the generator, which generates anomalous spatiotemporal data samples containing noise interference or missing data. These anomalous spatiotemporal data samples are then input into the pre-trained model, which extracts spatiotemporal features. The spatiotemporal features extracted by the pre-trained model and those extracted from real data are then input into the discriminator, which distinguishes the source of the input features. An adversarial loss is calculated based on the discriminator's judgment results. A task loss is calculated based on the pre-trained model's performance on the registration suggestion generation task. The pre-trained model and the generator are jointly optimized based on the adversarial loss and the task loss to obtain the trained spatiotemporal attention mechanism model. Through the adversarial training mechanism, the model's robustness and feature discrimination ability under complex scenarios such as data anomalies and missing information are significantly enhanced.

[0021] In this way, by using a self-supervised learning strategy to uncover the hidden spatiotemporal features in unlabeled spatiotemporal data, the model can be provided with sufficient pre-training knowledge reserves. Combined with an adversarial training mechanism, the model's ability to resist abnormal data and noise interference can be improved. This allows the spatiotemporal attention mechanism model trained to maintain stable feature extraction and decision-making capabilities in the context of scarce data and dynamic and ever-changing pediatric registration scenarios, thereby ensuring the accuracy and reliability of registration suggestions.

[0022] Secondly, a smart management device for pediatric registration based on a spatiotemporal attention mechanism model is provided, including: The first processing module is used to acquire the child's original multi-source data, which includes at least multi-dimensional information reflecting the child's health background, current condition and living environment. The second processing module is used to input the raw multi-source data into the trained spatiotemporal attention mechanism model, fuse the multi-source data through the cross-modal attention mechanism, dynamically adjust the attention weights of different data sources according to the data's own characteristics, and output the fused disease characterization features. The third processing module is used to extract the disease manifestation features from the spatial and temporal dimensions by the model to obtain spatial topological features and temporal periodic features. The spatial topological features are the characteristics that reflect the distribution and correlation of departmental resources after integrating the hospital department topology, medical knowledge graph and clinical expert experience. The temporal periodic features are the characteristics that reflect the changes in medical needs over time based on the periodic fluctuation pattern of the disease and the long-term dependence of historical medical data. The fourth processing module is used to dynamically associate spatial topological features with temporal periodic features through an adaptive gating network, and adjust the fusion weights of the two in real time according to the current medical needs to generate comprehensive decision features. Based on the comprehensive decision features, registration suggestions are generated, including recommended departments, recommended doctors, recommended appointment time slots, and the optimal path information from the entrance to the department.

[0023] Thirdly, an electronic device is provided, comprising: one or more processors; one or more memories; and one or more programs, wherein the one or more programs are stored in the one or more memories, and the one or more programs include instructions that, when executed by the one or more processors, cause the configured device to perform the method as described in the first aspect.

[0024] Fourthly, a computer storage medium is provided, the computer storage medium storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect or the second aspect.

[0025] Fifthly, a computer program product is provided, the computer program product storing instructions that, when executed by a computer, cause the computer to perform the method described in the first aspect or the second aspect.

[0026] The beneficial effects of the second to fifth aspects can be referred to the introduction of the beneficial effects of the first aspect above, and will not be repeated here. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall process of an intelligent management method for pediatric registration based on a spatiotemporal attention mechanism model, as provided in an embodiment of the present invention. Figure 2 This is a flowchart illustrating an intelligent management method for pediatric registration based on a spatiotemporal attention mechanism model, as provided in an embodiment of the present invention. Figure 3 This is a schematic diagram of a pediatric registration intelligent management device based on a spatiotemporal attention mechanism model that can be used to implement the method of the present invention; Figure 4 This is a schematic diagram of an electronic device provided by the present invention. Detailed Implementation

[0028] The solutions provided by the embodiments of the present invention will now be described with reference to the accompanying drawings. In the embodiments of the present invention, "multiple" refers to two or more objects, and "various kinds" refers to two or more types. Terms such as "first," "second," etc., are only used to distinguish similar objects and are not necessarily used to describe a specific order or number of objects.

[0029] In the pediatric healthcare system, the registration process is the first hurdle for children to access medical resources, and its efficiency and accuracy directly affect the timeliness and effectiveness of subsequent diagnosis and treatment. Especially in pediatrics, children's conditions can change rapidly, and their limited verbal communication skills often make it difficult for parents to accurately describe their symptoms, leading to a high reliance on subjective judgment in registration decisions. While technologies such as multi-channel registration, intelligent recommendations, real-time information updates, and health record management have improved the registration experience to some extent, they still cannot fundamentally solve the specific pain points in pediatric registration.

[0030] However, existing registration management methods and systems generally face serious constraints such as insufficient information utilization, lack of spatiotemporal dynamic perception, and inadequate decision-making generalization ability. On the one hand, traditional registration methods mainly rely on parents' subjective descriptions or simple rule recommendations, failing to effectively integrate multimodal data such as children's medical records, growth and development indicators, and audio-visual recordings. This results in low accuracy in matching departments with doctors, leading to frequent instances of mis-registration, incorrect registration, and repeated referrals. On the other hand, existing systems lack the ability to jointly model the spatial topology of hospital departments with the temporal cycle patterns of diseases. The allocation of appointment slots is rigid and unable to respond in real time to dynamic changes in patient conditions and fluctuations in demand. This results in resource shortages in popular departments during peak hours while other departments remain idle, leading to long, ineffective waiting times for children and parents and a high risk of cross-infection within the hospital. Even more challenging is that existing intelligent recommendation models are mostly based on shallow statistical analysis or simple neural networks, which are not robust to complex scenarios such as non-Gaussian noise and missing data. They also have difficulty capturing strong nonlinear and long-term dependent spatiotemporal features. When dealing with public health emergencies or rapid changes in the condition of children, they cannot quickly adjust the registration strategy, which can easily delay the treatment of critically ill children.

[0031] Therefore, given that existing methods struggle to simultaneously achieve in-depth utilization of multi-source heterogeneous data, dynamic perception of departmental space and disease cycles, nonlinear spatiotemporal feature modeling, and highly robust decision-making in complex scenarios, designing a non-destructive, highly accurate matching, spatiotemporally dynamic adaptive, and effective response to data anomalies and sudden changes in disease progression has become a pressing technical challenge in the field of smart healthcare and children's health services.

[0032] To address the aforementioned technical challenges, this invention proposes an intelligent management method for pediatric registration based on a spatiotemporal attention mechanism model. This method first designs a multimodal attention fusion and dynamic adaptive attention mechanism to fuse and process the original collected basic health data, disease-related data, and environmental behavior data. The multimodal attention fusion employs a hierarchical structure, extracting text, image, and audio features using encoders such as BERT and CNN, which are then integrated by a top-level attention layer. The dynamic adaptive attention mechanism uses a gating network to adjust the weights of each attention head in real time based on the input data features, quickly focusing on key information about the child's condition and avoiding information omissions and subjective biases.

[0033] Secondly, a topological graph attention network in the spatial dimension and a disease cycle Transformer in the temporal dimension are constructed: In the spatial dimension, the topological structure of the hospital's departments is modeled as a graph structure, integrating external knowledge graphs and expert experience, and using graph attention mechanisms to focus on key nodes and edges to achieve dynamic perception of departmental reception capacity and patient flow paths; In the temporal dimension, disease cycle fluctuation characteristics are extracted, a hierarchical spatiotemporal coding structure is designed, and a recursive mechanism is introduced between Transformer encoders to capture seasonal high incidence, holiday peaks, and long-term time dependencies, forming a spatiotemporal joint model of registration resources.

[0034] Subsequently, a self-supervised learning strategy and an adversarial training mechanism were introduced to enhance the aforementioned spatiotemporal attention mechanism Transformer model: self-supervised learning uses a masked language model and sequence prediction to pre-train on unlabeled spatiotemporal data to uncover hidden departmental relationships and patterns of changing patient needs; adversarial training simulates complex scenarios such as data anomalies and missing data by generating adversarial networks, thereby improving the model's robustness and discriminative ability in real-world environments.

[0035] Furthermore, the aforementioned spatial feature extractor, temporal feature extractor, self-supervised learning and adversarial training module are integrated into an end-to-end reverse feedback loop containing a feature fusion layer and a four-dimensional registration suggestion output. The feature fusion layer employs an attention mechanism and a gated recurrent unit to dynamically associate spatial topological features and temporal periodic features. The gating mechanism adaptively adjusts the fusion ratio between the two and suppresses interference from irrelevant information.

[0036] Ultimately, the system outputs a four-dimensional registration suggestion that includes department, doctor, time slot, and optimal route: the department recommendation takes into account the relevant specialty of the disease and the current saturation of patients; the doctor recommendation integrates the professional field, the diseases they are good at, patient reviews and scheduling; the time slot recommendation is based on the busyness of each time slot and the equipment usage; and the optimal route is planned in combination with the three-dimensional spatial layout of the hospital.

[0037] The entire method integrates a spatiotemporal attention mechanism model with multimodal and multidimensional features to achieve continuous, non-destructive, and highly accurate intelligent matching and dynamic resource scheduling for pediatric registration needs, significantly improving the accuracy and generalization ability of registration decisions and the medical experience of children and their parents.

[0038] Figure 1 This is a flowchart illustrating an intelligent management method for pediatric registration based on a spatiotemporal attention mechanism model, as shown in an embodiment of the present invention. Figure 1 As shown, it includes the following steps: In step S1, the patient's original multi-source data is collected and preprocessed.

[0039] In the embodiments of this application, the raw data includes basic health data, disease-related data, and environmental and behavioral data.

[0040] The basic health data includes personal information such as name, gender, age, date of birth, and contact information. It also includes past medical history, such as major illnesses, surgeries, and allergies. Furthermore, it includes family medical history, vaccination history, such as vaccination records, adverse reactions, and growth and development indicators, such as height, weight, and head circumference, along with comparative data. This data forms the foundation for understanding a child's health background, genetic predisposition, immune status, and growth and development, providing crucial information for preliminary disease diagnosis.

[0041] Disease-related data includes symptom descriptions, such as specific symptoms like fever and cough, as well as their duration and severity. It also includes medication information. This information directly reflects the child's current physical condition, helping doctors quickly pinpoint potential illnesses and clarifying medication use to avoid diagnostic interference.

[0042] Environmental and behavioral data includes information on lifestyle habits, living environment, recent travel history, daily activities, and psychological and emotional state. This type of data reflects a child's living environment, behavioral patterns, and psychological state, and helps analyze the causes of illness and determine whether environmental, psychological, or other factors are causing physical discomfort.

[0043] Data collected through various channels is diverse, and the data types from different sources vary significantly, requiring targeted collection, processing, and fusion.

[0044] The main methods for collecting structured data include electronic medical record (EMR) interfaces and registration form filling. The EMR interface utilizes HIS system integration technology to automate the collection and structured integration of internal hospital medical data. Registration form filling leverages the form function of an online registration mini-program, allowing parents to proactively enter their basic personal information, medical history, and other data when making an appointment.

[0045] The collection of unstructured data primarily relies on symptom text descriptions and OCR recognition of doctors' handwritten medical records. The detailed symptoms and disease progression described by patients during their visits can be extracted and structured using NLP (Natural Language Processing) technology. Doctors' handwritten medical records utilize image recognition technology to digitally identify and analyze handwriting, transforming illegible handwriting into searchable text data.

[0046] Real-time dynamic data is primarily collected through real-time synchronization of appointment availability and updates to the waiting queue. Utilizing the Hospital Information System (HIS) API, dynamic information such as appointment availability and doctor's appointment times can be retrieved in real time, ensuring synchronized updates between the registration platform and the hospital's backend data. Real-time changes in the waiting queue are pushed via WebSocket real-time communication technology, transmitting patient queue positions, estimated appointment times, and other data to the front-end interface in real time.

[0047] External data collection primarily involves data from regional healthcare platforms. This data can be securely acquired and integrated across different institutions through data platform integration technology.

[0048] In step S2, the spatiotemporal attention mechanism model Transformer is constructed and initialized.

[0049] In this embodiment, the attention mechanism Transformer model is improved as follows: multi-source data collected are integrated through multimodal attention fusion, and attention allocation is adjusted according to data characteristics through a dynamic adaptive attention gating network. The improved model is referred to as the spatiotemporal attention mechanism Transformer model.

[0050] Specifically, multimodal attention fusion adopts a hierarchical attention structure, with each layer corresponding to a specific modality. At the bottom layer, feature extraction is performed using a corresponding modality-specific encoder for different modalities, converting the data of each modality into a unified feature representation. For text modality, symptom descriptions and medical histories use BERT-like models, including pediatric-specific pre-trained models, to extract semantic features. For image modality, convolutional neural networks are used to extract visual features, and medical images or ordinary images are converted into feature maps with a self-attention mechanism. For audio modality, audio signals are converted into spectrograms using Mel-frequency cepstral coefficients before using convolutional neural networks or recurrent neural networks to extract acoustic features.

[0051] Within each modality, an adaptive attention mechanism is used to process detailed information. In the image modality, the adaptive attention mechanism focuses on specific regions in the image, including abnormal shadow areas, while in the text modality, the adaptive attention mechanism focuses on specific words or phrases, including "cough" and "fever". An attention mechanism is set at the top level to fuse information from different modalities.

[0052] The dynamic adaptive attention gating network utilizes a learnable gating network to control the weight allocation of attention heads: for a given input, each attention head transforms the input through its own linear transformation matrix to obtain a query vector, a key vector, and a value vector, respectively. The attention weight of each position is determined by calculating the similarity between the query vector and the key vector, and this weight represents the degree of association between the current position and other positions. The gating network dynamically adjusts the activation intensity of each attention head in the attention layer based on the features of the input data, including sequence length, complexity, and the presence of specific labels. A time-aware element is introduced into the attention weight calculation, embedding time information into the attention weight calculation process.

[0053] In step S3, a topological graph attention network is constructed in the spatial dimension, embedding the hospital's departmental topology, external knowledge graph, and expert experience.

[0054] In this embodiment, the various departments and key areas of the hospital campus are used as nodes, and the spatial relationships and business connections between departments are used as edges to construct a graph structure. A graph neural network (GNN) combined with knowledge graph construction and embedding technology is used to import prior knowledge such as pediatric medical knowledge, departmental resource distribution, and disease diagnosis and treatment processes.

[0055] Simultaneously, through knowledge distillation, expert experience, such as optimal departments for consultation and priority examinations, is transformed into soft labels. Parameter embedding transforms information such as department capacity limitations and saturation levels into model parameters. The resulting topological graph attention network utilizes graph attention mechanisms to enable the model to focus on key nodes and edges in the topological structure, such as the strong association between the pediatric emergency department, the laboratory, and the pharmacy. This dynamically optimizes the registration allocation strategy and provides parents with guidance on departmental specialties, current waiting numbers, and the optimal path to the department when registering.

[0056] Specifically, in the field of intelligent pediatric appointment booking, the spatiotemporal attention mechanism model constructs a resource scheduling system through processing spatial and temporal dimensions. In the spatial dimension, a spatial feature extractor is built: the hospital's departmental topology is modeled as a graph structure, and graph neural network technology is used to treat each department or area as a node, with edges between nodes representing spatial relationships and business connections between departments.

[0057] An external knowledge graph is introduced, and knowledge graph construction and embedding technology is adopted. The knowledge graph includes pediatric medical knowledge, hospital department resource distribution, and disease diagnosis and treatment process information. Domain expert knowledge is integrated into the model. Knowledge distillation is used to transform expert experience into soft tags, and parameter embedding is used to transform the department's capacity limit and the saturation of patients at different time periods into model parameters.

[0058] Based on the above technologies, a topological graph attention network is formed. The graph attention mechanism is used to focus on the nodes and edges in the topological structure. The registration allocation strategy is adjusted by analyzing historical registration data and the real-time busyness of departments, and combining the positional relationship between departments.

[0059] In step S4, disease cycle fluctuation features are extracted in the time dimension, a hierarchical spatiotemporal coding structure is designed, and a recursive mechanism is introduced.

[0060] In this embodiment, the temporal feature extractor takes historical registration data, disease monitoring data, and other time-series data as input, and uses time-series prediction models such as Prophet to identify the cyclical fluctuations of diseases such as seasonal influenza and hand-foot-mouth disease. The hierarchical spatiotemporal coding structure at the bottom layer accurately records micro-information such as the number of registrations, disease types, and age distribution of children in each department at different times of the day. The higher layers use CNN to extract features from local time windows and capture dependencies across different spans through self-attention, extracting macro-patterns such as weekly peak visits and monthly epidemic trends. Recursive connections are added between Transformer encoder layers, allowing the calculation of each time step to reference the hidden states of previous time steps, effectively capturing long-term dependencies over scales of months or even years, thereby accurately predicting future changes in the demand for medical services in various departments.

[0061] Specifically, in the field of intelligent pediatric appointment registration, in order to further improve the performance of the Transformer model with spatiotemporal attention mechanism and effectively cope with complex and ever-changing real-world scenarios, a self-supervised learning strategy and adversarial training mechanism are introduced, and the feature fusion process is deeply optimized to achieve more accurate and intelligent appointment registration decisions.

[0062] At the level of self-supervised learning strategies, a pre-training approach based on masked language modeling (MLM) and sequence prediction is employed to process large-scale unlabeled spatiotemporal data, accurately addressing the complex issues in children's registration scenarios. For spatial topological data, a graph neural network (GNN) combined with MLM technology is used.

[0063] Taking a children's hospital as an example, the spatial relationships between departments change due to the opening of a new children's health center. At this point, some departmental nodes or edge relationships are randomly masked. For example, the edges connecting the newly built children's health center with the existing pediatric outpatient clinic and laboratory are masked, allowing the model to predict the masked spatial information based on the remaining visible information. By continuously learning these masked spatial relationships, the model can quickly capture the potential collaboration patterns and patient flow paths between departments throughout the hospital after the addition of the new department. This helps parents obtain more accurate departmental location guidance and association suggestions when registering, reducing unnecessary movements caused by unclear spatial information.

[0064] For time series data, we leverage the sequence modeling capabilities of Transformer combined with time series prediction algorithms. In practical applications, the demand for children's medical care fluctuates significantly during the winter and summer vacations each year due to changes in daily routines, travel, and other factors. In such cases, we mask future registration data, such as randomly hiding the number of pediatric appointments for specific periods within a week during the winter or summer vacation, and train the model to predict missing values ​​based on historical data.

[0065] Through repeated training, the model not only masters the regular cyclical fluctuations of diseases but also learns the changing patterns of medical needs during special periods. This allows for more accurate prediction of peak visit times for various departments before summer and winter vacations, enabling the model to send off-peak appointment suggestions to parents and assist hospitals in pre-allocating medical staff and equipment to avoid resource shortages or underutilization. This pre-training method fully utilizes massive amounts of unlabeled data to uncover hidden spatiotemporal features, greatly enhancing the model's understanding of these features and providing a more reliable decision-making basis for subsequent actual appointment booking tasks.

[0066] The introduction of adversarial training mechanisms, leveraging Generative Adversarial Networks (GANs) or adversarial example generation techniques, effectively enhances the robustness of models in complex scenarios, addressing practical pain points in children's registration. Taking GANs as an example, a generator and discriminator specifically designed for pediatric registration scenarios are constructed. The generator simulates abnormal or complex spatiotemporal data distributions, generating registration data samples with noise interference and missing data.

[0067] In practical applications, when a hospital information system malfunctions, resulting in the loss or incorrect entry of some registration data, the generator can simulate similar data situations, such as generating samples containing incorrect department association information or abnormal registration times.

[0068] The discriminator is responsible for distinguishing between real and generated samples, while the Transformer model is trained under the supervision of the discriminator. Through this adversarial process, the model is forced to learn how to accurately identify key features in complex environments with anomalous data and chaotic information.

[0069] For example, during a major flu outbreak, a large influx of patients registering for appointments can lead to data congestion and anomalies. In such situations, the adversarially trained model can quickly identify genuine and valid registration needs, preventing the misallocation of registration resources due to data anomalies and reducing the likelihood of parents repeatedly queuing and re-registering due to poor decision-making. Simultaneously, the model's ability to handle abnormal data and its feature discrimination capabilities are significantly enhanced. Even in the face of sudden surges in patient visits or abnormal data fluctuations, it can make reasonable registration decisions, ensuring the rationality and efficiency of hospital resource allocation and effectively improving the stability and reliability of the pediatric intelligent registration system in complex and ever-changing scenarios.

[0070] Furthermore, in the embodiments of this application, a self-supervised learning strategy is used to pre-train the model, and an adversarial training mechanism is introduced to enhance the robustness of the model.

[0071] In this embodiment, self-supervised learning employs a masked language model and sequence prediction approach: for spatial topological data, some departmental nodes or edge relationships are randomly masked, and the model is trained to predict the masked information; for time-series data, registration data for future periods is masked, and the model is trained to predict missing values ​​based on historical data. An adversarial training mechanism constructs a generative adversarial network: the generator simulates abnormal or complex spatiotemporal data distributions, the discriminator distinguishes between real and generated samples, and the Transformer model undergoes adversarial training under the supervision of the discriminator, thereby improving its ability to identify complex scenarios such as data anomalies and information chaos, and enhancing decision-making stability.

[0072] In step S5, spatial topological features and time periodic features are dynamically associated through the feature fusion layer, and the fusion ratio is adaptively adjusted by an attention mechanism and a gated loop unit, ultimately outputting a four-dimensional registration suggestion.

[0073] In this embodiment, the feature fusion layer uses an attention mechanism to identify spatiotemporal feature combinations that play a key role in the current registration decision. The gating loop unit flexibly adjusts the interaction mode of spatiotemporal features according to the urgency of the child's condition. By dynamically controlling the fusion ratio of spatial and temporal features through the gating mechanism, irrelevant interference information is effectively suppressed. Finally, a four-dimensional registration suggestion containing department, doctor, time period, and optimal route is generated.

[0074] Specifically, the feature fusion layer adopts a technical architecture combining an attention mechanism and a gating loop unit: the attention mechanism assigns weights to the spatial relationships between the pediatric fever clinic, the laboratory, and the pharmacy, as well as the high-load periods of each department, based on the spatiotemporal characteristics of the current registration demand; the gating loop unit adjusts the interaction mode of spatiotemporal features based on the child's condition information, prioritizing the spatial distance of the emergency department and the current available registration time, calculating the nearest emergency department with available appointment slots, and selecting doctors based on the professional qualifications and consultation efficiency of the doctors in that department; the gating mechanism adjusts the fusion ratio of spatial features and temporal features.

[0075] After comprehensively analyzing the spatiotemporal characteristics, a four-dimensional registration suggestion is generated, including department, doctor, time period, and optimal route: department recommendation is based on the corresponding specialty of the disease and the current saturation of the department and the configuration of medical staff; doctor recommendation is based on the doctor's professional field, diseases of expertise, patient reviews and daily schedule; time period recommendation is based on the busy level of each department at different times and the usage of examination equipment; the optimal route is based on the hospital's three-dimensional spatial layout data, and a route from the hospital entrance to the target department is generated through a path planning algorithm.

[0076] In summary, this application embodiment achieves accurate characterization of children's conditions by fusing multi-source heterogeneous data through a cross-modal attention mechanism; it overcomes the shortcomings of traditional methods that neglect spatiotemporal joint modeling by fusing departmental topology and expert experience through a spatial topological graph attention network and capturing disease cycle patterns through a temporal hierarchical coding and recursive mechanism; furthermore, it achieves personalized and real-time registration decisions by dynamically adjusting the spatiotemporal feature fusion weights through an adaptive gating network; and finally generates four-dimensional registration suggestions, significantly improving the accuracy of pediatric registration and the efficiency of resource scheduling.

[0077] Furthermore, the spatiotemporal attention mechanism model is obtained through the following training process: Acquire unlabeled spatiotemporal data, which includes historical registration data, hospital department topology data, and disease surveillance data. Historical registration data includes the number of registrations, types of diseases treated, and age distribution of patients in each department during different historical periods. Hospital department topology data includes the spatial location of each hospital department, the physical distance between departments, and the workflow relationships. Disease surveillance data includes disease epidemic trend monitoring data released by the health department.

[0078] A self-supervised learning strategy was used to pre-train the initial spatiotemporal attention mechanism model.

[0079] Specifically, the hospital department topology data from unlabeled spatiotemporal data is input into the spatial dimension extraction unit, and partial node or edge relationships in the department topology graph are masked. A topology graph is constructed using departments as nodes and spatial and business connections between departments as edges. Partial nodes and their associated edges in this topology graph are randomly masked. The masked topology graph is then input into a topology graph attention network. This network predicts the attributes of the masked nodes or edges based on information from the remaining visible nodes and edges in the graph. A first loss is calculated based on the difference between the prediction results and the original data. Through this masking prediction task, the spatial dimension extraction unit learns the potential spatial connections between hospital departments and patient flow paths, enabling the model to quickly capture new spatial topological relationships when new departments are added or departmental layouts change.

[0080] Historical registration data from unlabeled spatiotemporal datasets is input into the time dimension extraction unit, where future time period data within the historical registration data is masked. Using the number of registrations and the types of diseases treated in each historical time period as the time series input, data points for specific future time periods are randomly masked within this time series. The masked time series is then input into a hierarchical spatiotemporal coding structure. This structure predicts the data for the masked time period based on the historically visible time periods, and a second loss is calculated based on the difference between the prediction results and the original data. Through this masking prediction task, the time dimension extraction unit learns the cyclical fluctuations of pediatric diseases with seasonal changes and the changing patterns of medical needs during special periods such as holidays and summer / winter breaks.

[0081] The parameters of the initial spatiotemporal attention mechanism model are jointly optimized using the first and second losses to obtain a pre-trained model. Through the above self-supervised pre-training process, the model learns general spatial topological features and temporal periodic feature representations from massive spatiotemporal data without manual annotation, providing a reliable decision-making basis for subsequent actual registration tasks.

[0082] An adversarial training mechanism is used to enhance the training of the pre-trained model.

[0083] Specifically, a generator and a discriminator are constructed. Unlabeled spatiotemporal data is input into the generator, which generates anomalous spatiotemporal data samples containing noise interference or missing data. In specific application scenarios, when a hospital information system malfunctions, resulting in the loss or incorrect entry of some registration data, the generator can simulate samples containing incorrect department association information or abnormal registration times; during a major influenza outbreak, when a large number of patients register simultaneously, causing data congestion, the generator can simulate samples of abnormal data fluctuations.

[0084] Abnormal spatiotemporal data samples are input into a pre-trained model, where spatial and temporal dimension extraction units extract spatial topological and temporal periodic features from the abnormal samples, respectively. The spatiotemporal features extracted by the pre-trained model and those extracted from real data are then input into a discriminator. The discriminator distinguishes the source of the input features—that is, it determines whether the current input feature comes from the feature extraction result of the pre-trained model on real data or from the feature extraction result of the pre-trained model on the abnormal data generated by the generator.

[0085] The adversarial loss is calculated based on the discriminator's judgment result, and the task loss is calculated based on the performance of the pre-trained model on the registration suggestion generation task. The adversarial loss encourages the generator to generate anomalous samples that more closely resemble the real data distribution to "deceive" the discriminator, while simultaneously encouraging the pre-trained model to extract spatiotemporal features from the anomalous samples that are difficult to distinguish from real samples. The task loss measures the difference between the registration suggestions output by the pre-trained model and the optimal registration decision. The pre-trained model and the generator are jointly optimized based on the adversarial loss and the task loss. When both the adversarial loss and the task loss converge, the trained spatiotemporal attention mechanism model is obtained. Figure 2 This is a flowchart illustrating an intelligent management method for pediatric registration based on a spatiotemporal attention mechanism model, as shown in an embodiment of the present invention. Figure 2 As shown, it includes the following steps: 210: Obtain the original multi-source data of the child. The original multi-source data should include at least multi-dimensional information reflecting the child's health background, current condition and living environment.

[0086] 220: Input the original multi-source data into the trained spatiotemporal attention mechanism model, fuse the multi-source data through the cross-modal attention mechanism, dynamically adjust the attention weights of different data sources according to the data's own characteristics, and output the fused disease representation features.

[0087] 230: Deep extraction of disease manifestation features from spatial and temporal dimensions yields spatial topological features and temporal periodic features. Spatial topological features are characteristics reflecting the distribution and correlation of departmental resources after integrating the hospital department topology, medical knowledge graph, and clinical expert experience. Temporal periodic features are characteristics reflecting the changes in medical needs over time based on the periodic fluctuations of the disease and the long-term dependence of historical medical data.

[0088] 240: The spatial topological features and time periodic features are dynamically correlated through an adaptive gating network, and the fusion weights of the two are adjusted in real time according to the current medical needs to generate comprehensive decision features; based on the comprehensive decision features, registration suggestions are generated, including recommended departments, recommended doctors, recommended appointment time slots, and the optimal path information from the entrance to the department.

[0089] In this embodiment, a cross-modal attention mechanism is used to fuse multi-source heterogeneous data, achieving accurate characterization of the child's condition. A spatial-dimensional topological graph attention network is used to fuse departmental topology and expert experience, while a temporal-dimensional hierarchical coding and recursive mechanism captures the disease cycle pattern, overcoming the shortcomings of traditional methods that neglect spatiotemporal joint modeling. Furthermore, an adaptive gating network is used to dynamically adjust the spatiotemporal feature fusion weights, achieving personalized and real-time registration decisions. Finally, a four-dimensional registration suggestion is generated, significantly improving the accuracy of pediatric registration and the efficiency of resource scheduling.

[0090] In some embodiments, the spatiotemporal attention mechanism model includes: The multimodal attention fusion unit is used to perform cross-modal feature fusion on input data from different modalities and dynamically adjust attention weights.

[0091] The spatial dimension extraction unit includes a topological graph attention network, which is used to perform graph attention calculations on the fused features with the hospital department topology, medical knowledge graph, and expert experience, and output spatial topological features.

[0092] The time dimension extraction unit includes a hierarchical spatiotemporal coding structure and recursive connections. It is used to align the fused features with historical time series data and extract the dependencies of different time spans, outputting time periodic features.

[0093] The adaptive gating fusion unit is used to dynamically associate spatial topological features and time periodic features according to the current medical needs and adjust the fusion weights to generate comprehensive decision features.

[0094] The output unit is used to generate registration suggestions based on comprehensive decision features.

[0095] In some embodiments, step 220 includes: The original multi-source data is divided into text modal data, image modal data, and audio modal data.

[0096] For text modal data, semantic features are extracted using a text encoder.

[0097] For image modal data, visual features are extracted by an image encoder, and self-attention weights are applied to the visual features to focus on specific regions in the image.

[0098] For audio modal data, the audio signal is converted into a spectrogram and then the acoustic features are extracted using an audio encoder.

[0099] Using textual semantic features as query vectors and visual and acoustic features as key-value pairs, attention weights for each modality feature are calculated through a cross-modal attention mechanism to obtain preliminary fused features.

[0100] By using an attention head gating network, the activation intensity of each attention head in the cross-modal attention mechanism is dynamically adjusted according to the sequence length or complexity of text semantic features, visual features, and acoustic features, and the dynamically weighted fused features are output as disease representation features.

[0101] In some embodiments, step 230 includes: The disease manifestation features are input into the spatial dimension extraction unit and the temporal dimension extraction unit, respectively.

[0102] In the spatial dimension extraction unit, a pre-constructed hospital department topology graph structure is obtained. The nodes of the topology graph are departments or medical areas, and the edges represent the spatial associations and business connections between departments. The disease representation features, the topology graph structure, the external medical knowledge graph, and the expert experience embedding vectors are jointly input into the topology graph attention network. The topology graph attention network calculates the attention weights of the neighboring nodes of each node and outputs spatial topology features that aggregate spatial association information.

[0103] In the time dimension extraction unit, a time series consisting of historical registration data and disease monitoring data is obtained. The disease manifestation features and the time series are input into a hierarchical spatiotemporal coding structure, and the time period features are output.

[0104] In some embodiments, step 230 inputs the disease manifestation features and time sequence into a hierarchical spatiotemporal coding structure and outputs time periodic features, including: The disease characteristics are copied into vectors with the same number of time steps as the time series, and each copied vector is concatenated with the historical data vector of the corresponding time step to form an aligned input sequence.

[0105] The aligned input sequence is fed into the bottom layer of the hierarchical spatiotemporal coding structure in chronological order. The bottom layer encodes the position of each time step through timestamp encoding and outputs the bottom layer feature sequence.

[0106] The low-level feature sequence is input into the high-level layer of the hierarchical spatiotemporal coding structure. The high-level layer extracts features within a local time window through a convolutional neural network, and then calculates the dependency weights between different time steps through a self-attention mechanism, outputting the high-level feature sequence.

[0107] In a hierarchical spatiotemporal coding structure, recursive connections are set between encoder layers to pass the previous hidden state of the current layer to the same layer at the next time step, thus updating the hidden state.

[0108] The final hidden state after recursive connection is output as a time periodic feature.

[0109] In some embodiments, step 240 includes: The characteristics of current medical needs are represented, including the urgency of the child's condition and the real-time busy status of the selected department.

[0110] Spatial topological features and temporal periodic features are respectively input into the feature fusion gating network. The feature fusion gating network calculates spatial fusion weights and temporal fusion weights based on the feature representation of the current medical needs.

[0111] The spatial topological features and temporal periodic features are weighted and summed according to spatial fusion weights and temporal fusion weights to generate comprehensive decision features.

[0112] In some embodiments, step 240 includes: The comprehensive decision-making features are input into the output unit, and the output unit includes department recommendation, doctor recommendation, time period recommendation and route planning.

[0113] The department recommendation section outputs a score based on the department's relevant specialty and current patient saturation.

[0114] The doctor recommendation section is based on the doctor's area of ​​expertise, the degree of match between the doctor's expertise and the child's symptoms, patient reviews, and the doctor recommendation score generated by the daily schedule.

[0115] The time slot recommendation section is based on the busyness score of each future time slot and the sorting of time slots output by checking the idle status of the equipment.

[0116] The path planning component is based on the hospital's 3D spatial layout data. Starting from the current entrance and ending at the recommended department location, it generates the optimal path through a path search algorithm.

[0117] The outputs from the department recommendation section, doctor recommendation section, time slot recommendation section, and route planning section are combined into registration suggestions.

[0118] Figure 3 This is a schematic diagram of a pediatric registration intelligent management device based on a spatiotemporal attention mechanism model, as illustrated in an embodiment of the present invention. Figure 3 As shown, it includes: The first processing module is used to acquire the child's original multi-source data, which includes at least multi-dimensional information reflecting the child's health background, current condition, and living environment.

[0119] The second processing module is used to input the raw multi-source data into the trained spatiotemporal attention mechanism model, fuse the multi-source data through the cross-modal attention mechanism, dynamically adjust the attention weights of different data sources according to the data's own characteristics, and output the fused disease representation features.

[0120] The third processing module is used to extract the disease manifestation features from the spatial and temporal dimensions by the model, respectively, to obtain spatial topological features and temporal periodic features. The spatial topological features are the characteristics that reflect the distribution and correlation of departmental resources after integrating the hospital department topology, medical knowledge graph and clinical expert experience. The temporal periodic features are the characteristics that reflect the changes in medical needs over time based on the periodic fluctuation pattern of the disease and the long-term dependence of historical medical data.

[0121] The fourth processing module is used to dynamically associate spatial topological features with temporal periodic features through an adaptive gating network, and adjust the fusion weights of the two in real time according to the current medical needs to generate comprehensive decision features. Based on the comprehensive decision features, registration suggestions are generated, including recommended departments, recommended doctors, recommended appointment time slots, and the optimal path information from the entrance to the department.

[0122] In some possible implementations, the second processing module is specifically used to: divide the original multi-source data into text modal data, image modal data, and audio modal data; For text modal data, semantic features are extracted using a text encoder; For image modal data, visual features are extracted by an image encoder, and self-attention weights are applied to the visual features to focus on specific regions in the image; For audio modal data, the audio signal is converted into a spectrogram and then the acoustic features are extracted using an audio encoder. Using textual semantic features as query vectors and visual and acoustic features as key-value pairs, attention weights for each modality feature are calculated through a cross-modal attention mechanism to obtain preliminary fused features; By using an attention head gating network, the activation intensity of each attention head in the cross-modal attention mechanism is dynamically adjusted according to the sequence length or complexity of text semantic features, visual features, and acoustic features, and the dynamically weighted fused features are output as disease representation features.

[0123] In some possible implementations, the third processing module is specifically used to: input the disease manifestation features into the spatial dimension extraction unit and the temporal dimension extraction unit, respectively; In the spatial dimension extraction unit, a pre-constructed hospital department topology graph structure is obtained. The nodes of the topology graph are departments or medical areas, and the edges represent the spatial associations and business connections between departments. The disease representation features, the topology graph structure, the external medical knowledge graph, and the expert experience embedding vectors are jointly input into the topology graph attention network. The topology graph attention network calculates the attention weights of the neighboring nodes of each node and outputs spatial topology features that aggregate spatial association information. In the time dimension extraction unit, a time series consisting of historical registration data and disease monitoring data is obtained. The disease manifestation features and the time series are input into a hierarchical spatiotemporal coding structure, and the time period features are output.

[0124] In some possible implementations, the third processing module is also used to: copy the disease manifestation features into vectors with the same number of time steps as the time series, and concatenate each copied vector with the historical data vector of the corresponding time step to form an aligned input sequence; The aligned input sequence is fed into the bottom layer of the hierarchical spatiotemporal coding structure in chronological order. The bottom layer encodes the position of each time step through timestamp encoding and outputs the bottom layer feature sequence. The low-level feature sequence is input into the high-level layer of the hierarchical spatiotemporal coding structure. The high-level layer extracts features within the local time window through a convolutional neural network, and then calculates the dependency weights between different time steps through a self-attention mechanism to output the high-level feature sequence. In a hierarchical spatiotemporal coding structure, recursive connections are set between encoder layers to pass the previous hidden state of the current layer to the same layer in the next time step and update the hidden state. The final hidden state after recursive connection is output as a time periodic feature.

[0125] In some possible implementations, the fourth processing module is specifically used to: obtain the feature representation of the current medical needs, which includes the urgency of the child's condition and the real-time busy status of the selected department; input the spatial topological features and the temporal periodic features into the feature fusion gating network respectively, and the feature fusion gating network calculates the spatial fusion weight and the temporal fusion weight based on the feature representation of the current medical needs; The spatial topological features and temporal periodic features are weighted and summed according to spatial fusion weights and temporal fusion weights to generate comprehensive decision features.

[0126] In some possible implementations, the fourth processing module is also used to: input the comprehensive decision features into the output unit, the output unit including the department recommendation part, the doctor recommendation part, the time period recommendation part and the route planning part; The department recommendation section outputs a department recommendation score based on the department's corresponding specialty and current patient saturation; the doctor recommendation section outputs a doctor recommendation score based on the doctor's professional field, the degree of matching between the doctor's expertise and the child's symptoms, patient evaluations, and the daily schedule; the time slot recommendation section outputs a time slot ranking based on the busyness score of each time slot in the future and the idle status of examination equipment; and the path planning section generates the optimal path based on the hospital's three-dimensional spatial layout data, starting from the current entrance and ending at the recommended department location, through a path search algorithm. The outputs from the department recommendation section, doctor recommendation section, time slot recommendation section, and route planning section are combined into registration suggestions.

[0127] In some possible implementations, a fifth processing module is also included, which is used to acquire unlabeled spatiotemporal data, including historical registration data, hospital department topology data, and disease monitoring data. A self-supervised learning strategy is employed to pre-train the initial spatiotemporal attention mechanism model. Specifically, the hospital department topology data is masked, randomly masking some department nodes or edge relationships. The masked topology data is then input into the spatial dimension extraction unit. The topology graph attention network in the spatial dimension extraction unit predicts the masked node or edge relationships based on the remaining visible node and edge information. The first loss is calculated based on the difference between the prediction result and the original data, and the parameters of the spatial dimension extraction unit are updated, enabling the model to learn the potential spatial associations and collaboration patterns between hospital departments. Similarly, historical registration data is masked, randomly masking registration data for future time periods. The masked historical registration data is then input into the temporal dimension extraction unit. The hierarchical spatiotemporal encoding structure and recursive connections in the temporal dimension extraction unit predict the registration data for the masked time periods based on the historical visible data. The second loss is calculated based on the difference between the prediction result and the original data, and the parameters of the temporal dimension extraction unit are updated, enabling the model to learn the temporal evolution patterns of disease cycle fluctuations and patient demand. The parameters of the initial spatiotemporal attention mechanism model are jointly optimized based on the first and second losses to obtain the pre-trained model.

[0128] An adversarial training mechanism is employed to enhance the pre-trained model. Specifically, a generator and a discriminator are constructed. Unlabeled spatiotemporal data is input into the generator, which generates anomalous spatiotemporal data samples containing noise interference or missing data. These anomalous spatiotemporal data samples are then input into the pre-trained model, which extracts spatiotemporal features. The spatiotemporal features extracted by the pre-trained model and those extracted from real data are then input into the discriminator, which distinguishes the source of the input features. An adversarial loss is calculated based on the discriminator's judgment results. A task loss is calculated based on the pre-trained model's performance on the registration suggestion generation task. The pre-trained model and the generator are jointly optimized based on the adversarial loss and the task loss to obtain the trained spatiotemporal attention mechanism model. Through the adversarial training mechanism, the model's robustness and feature discrimination ability under complex scenarios such as data anomalies and missing information are significantly enhanced.

[0129] The outputs of the department recommendation, doctor recommendation, time slot recommendation, and route planning sections are combined to form an appointment suggestion. Those skilled in the art will readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, the present invention can be implemented in hardware or a combination of hardware and computer software. Whether a function is implemented in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the present invention.

[0130] It should be noted that, Figure 3The division of modules / units is illustrative and represents only one logical functional division; in actual implementation, other division methods are possible. For example, two or more functions can be integrated into a single data acquisition module. The integrated modules described above can be implemented either in hardware or as software functional modules.

[0131] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned intelligent management methods for pediatric registration based on a spatiotemporal attention mechanism model. That is, an electronic device according to an embodiment of the present invention may include, but is not limited to: a processor and a memory; the memory is used to store the computer program; the processor is used to execute the intelligent management method for pediatric registration based on a spatiotemporal attention mechanism model shown in any embodiment of the present invention by calling the computer program.

[0132] In one alternative embodiment, an electronic device is provided, such as Figure 4 As shown, Figure 4 The illustrated electronic device 400 includes a processor 401 and a memory 403. The processor 401 and the memory 403 are connected, for example, via a bus 402. Optionally, the electronic device 400 may further include a transceiver 404, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 404 is not limited to one type, and the structure of the electronic device 400 does not constitute a limitation on the embodiments of the present invention.

[0133] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0134] It should be noted that, Figure 4 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0135] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any one of the above-mentioned intelligent management methods for pediatric registration based on a spatiotemporal attention mechanism model.

[0136] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0137] It should be noted that the terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of the invention described herein can be implemented in an order other than that shown or described.

[0138] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0139] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for intelligent management of pediatric registration based on a spatiotemporal attention mechanism model, characterized in that, include: Acquire multi-source data of the child, wherein the multi-source data includes multi-dimensional information reflecting the child’s health background, current condition and living environment; The multi-source data is input into a trained spatiotemporal attention mechanism model, which then fuses the multi-source data through a cross-modal attention mechanism to output disease characterization features. Using the spatiotemporal attention mechanism model and the disease representation features, spatial topological features and temporal periodic features are obtained, wherein the spatial topological features reflect the distribution and association of departmental resources; and the temporal periodic features reflect the changes in medical needs over time. The spatiotemporal attention mechanism model dynamically associates the spatial topological features with the temporal periodic features, and adjusts the fusion weights of the two in real time according to the current medical needs to generate comprehensive decision features. Based on the comprehensive decision-making characteristics, registration suggestions are generated.

2. The method according to claim 1, characterized in that, The step of inputting the multi-source data into a trained spatiotemporal attention mechanism model, and having the spatiotemporal attention mechanism model fuse the multi-source data through a cross-modal attention mechanism to output disease characterization features, includes: The multi-source data is divided into text modal data, image modal data, and audio modal data; For the aforementioned text modal data, text semantic features are extracted using a text encoder; For the aforementioned image modal data, visual features are extracted using an image encoder; For the aforementioned audio modal data, the audio signal is converted into a spectrogram, and then acoustic features are extracted using an audio encoder. Using the text semantic features as query vectors and the visual and acoustic features as key-value pairs, the attention weights of each modality feature are calculated through a cross-modal attention mechanism to obtain preliminary fused features; The attention head gating network adjusts the activation intensity of each attention head in the cross-modal attention mechanism according to the sequence length of the text semantic features, the visual features, and the acoustic features, and outputs the dynamically weighted fusion features as the disease characterization features.

3. The method according to claim 1, characterized in that, The spatiotemporal attention mechanism model includes a spatial dimension extraction unit and a temporal dimension extraction unit; The process of utilizing the spatiotemporal attention mechanism model and the disease representation features to obtain spatial topological features and temporal periodic features includes: The disease manifestation features are respectively input into the spatial dimension extraction unit and the temporal dimension extraction unit; In the spatial dimension extraction unit, a pre-constructed hospital department topology graph structure is obtained. The nodes of the topology graph are departments or medical areas, and the edges represent the spatial associations and business connections between departments. The disease characterization features, the topology graph structure, the external medical knowledge graph, and the expert experience embedding vector are jointly input into the topology graph attention network. The topology graph attention network calculates the attention weights of the neighboring nodes of each node and outputs the spatial topology features. In the time dimension extraction unit, a time series consisting of historical registration data and disease monitoring data is obtained. The disease condition characteristics and the time series are input into a hierarchical spatiotemporal coding structure, and the time period characteristics are output.

4. The method according to claim 3, characterized in that, The step of inputting the disease manifestation features and the time sequence into a hierarchical spatiotemporal coding structure and outputting the time periodic features includes: The disease manifestation features are copied into vectors with the same number of time steps as the time series, and each copied vector is concatenated with the historical data vector of the corresponding time step to form an aligned input sequence; The aligned input sequence is fed into the bottom layer of the hierarchical spatiotemporal coding structure in chronological order. The position of each time step is encoded by timestamp encoding, and the bottom layer feature sequence is output. The low-level feature sequence is input into the high-level structure of the hierarchical spatiotemporal coding structure. Features within the local time window are extracted by a convolutional neural network. Then, the dependency weights between different time steps are calculated by a self-attention mechanism, and the high-level feature sequence is output. A recursive connection is set between the encoder layers of the hierarchical spatiotemporal coding structure to pass the previous hidden state of the current layer to the same layer in the next time step and update the hidden state. The final hidden state after recursive connection is used as the time period feature.

5. The method according to claim 1, characterized in that, The process of dynamically associating spatial topological features with temporal periodic features through the spatiotemporal attention mechanism model, and adjusting the fusion weights of the two in real time according to the current medical needs to generate comprehensive decision features, includes: The feature representation of the current medical needs is obtained, which includes the urgency of the child's condition and the real-time busy status of the selected department; The spatial topological features and the temporal periodic features are respectively input into the feature fusion gating network, and the spatial fusion weights and temporal fusion weights are calculated based on the feature representation of the current medical needs. The spatial topological features and the temporal periodic features are weighted and summed according to the spatial fusion weight and the temporal fusion weight to generate the comprehensive decision features.

6. The method according to claim 1, characterized in that, The trained spatiotemporal attention mechanism model was obtained through the following method: Acquire unlabeled spatiotemporal data, which includes historical registration data, hospital department topology data, and disease monitoring data; A self-supervised learning strategy was used to pre-train the initial spatiotemporal attention mechanism model to obtain a pre-trained model. The pre-trained model is enhanced by an adversarial training mechanism to obtain the trained spatiotemporal attention mechanism model.

7. The method according to claim 6, characterized in that, The pre-training of the initial spatiotemporal attention mechanism model using a self-supervised learning strategy includes: The hospital department topology data is masked, randomly masking some department nodes or edge relationships. The masked topology data is then input into the spatial dimension extraction unit, which predicts the masked node or edge relationships. The first loss is calculated based on the difference between the prediction result and the original data, and the parameters of the spatial dimension extraction unit are updated. The historical registration data is masked to randomly mask registration data for future periods. The masked historical registration data is then input into the time dimension extraction unit, which predicts the registration data for the masked period. A second loss is calculated based on the difference between the prediction result and the original data, and the parameters of the time dimension extraction unit are updated. The parameters of the initial spatiotemporal attention mechanism model are jointly optimized based on the first loss and the second loss; The step of using an adversarial training mechanism to enhance the pre-trained model includes: Construct generators and discriminators; The unlabeled spatiotemporal data is input into the generator, which generates anomalous spatiotemporal data samples containing noise interference or missing data. The abnormal spatiotemporal data samples are input into the pre-trained model, and the pre-trained model extracts spatiotemporal features. The spatiotemporal features extracted from the pre-trained model and the spatiotemporal features extracted from real data are respectively input into the discriminator, which distinguishes the source of the input features. The adversarial loss is calculated based on the discrimination result of the discriminator, and the task loss is calculated based on the performance of the pre-trained model on the registration suggestion generation task. The pre-trained model and the generator are jointly optimized based on the adversarial loss and the task loss to obtain the trained spatiotemporal attention mechanism model.

8. A pediatric registration intelligent management device based on a spatiotemporal attention mechanism model, characterized in that, include: The first processing module is used to acquire multi-source data of the child, wherein the multi-source data includes multi-dimensional information reflecting the child's health background, current condition and living environment; The second processing module is used to input the multi-source data into a trained spatiotemporal attention mechanism model, and the spatiotemporal attention mechanism model fuses the multi-source data through a cross-modal attention mechanism to output disease characterization features. The third processing module is used to obtain spatial topological features and temporal periodic features by utilizing the spatiotemporal attention mechanism model and the disease manifestation features, wherein the spatial topological features are features reflecting the distribution and association of departmental resources; and the temporal periodic features are features reflecting the changes in medical needs over time. The fourth processing module is used to dynamically associate the spatial topological features with the temporal periodic features through the spatiotemporal attention mechanism model, and adjust the fusion weight of the two in real time according to the current medical needs to generate comprehensive decision features. Based on the comprehensive decision-making characteristics, registration suggestions are generated.

9. An electronic device, characterized in that, The device includes a memory and one or more processors, wherein the memory stores a computer program that, when executed by the one or more processors, implements the intelligent management method for pediatric registration based on a spatiotemporal attention mechanism model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to enable the computer to implement the intelligent management method for pediatric registration based on the spatiotemporal attention mechanism model as described in any one of claims 1 to 7.

Images