Medical knowledge graph construction method for multi-source heterogeneous data fusion and incremental updating

Abstract

The invention discloses a medical knowledge graph construction method and device for multi-source heterogeneous data fusion and incremental updating, and relates to the technical field of medical knowledge graphs. The method comprises the steps that multi-source medical data such as outpatient records, hospitalization medical records, inspection reports and image data of a patient in the whole life cycle are collected; performing time sequence alignment by adopting a space-time alignment algorithm to obtain standardized medical data, calculating a data fusion weight and constructing an incidence matrix; performing singular value decomposition on the incidence matrix to extract a feature vector, and obtaining a dynamic update coefficient through standardization, cosine similarity calculation, convolutional neural network processing and exponential weighted moving average; adjusting the reference map according to the dynamic updating coefficient increment to generate an updated medical knowledge map, and outputting diagnosis and treatment decision support information; and multi-dimensional medical data space monitoring abnormity can be constructed, a joint optimization model is established to generate a personalized diagnosis and treatment scheme, and edge deployment is realized through a knowledge distillation compression map.

Description

Technical Field

[0001] This invention relates to the field of medical knowledge graph construction technology, and in particular to a method and apparatus for constructing a medical knowledge graph through multi-source heterogeneous data fusion and incremental updating. Background Technology

[0002] In the current era of rapid development in medical informatization, medical data is experiencing explosive growth, and this data is characterized by its diverse sources and heterogeneity, including various types such as outpatient records, inpatient medical records, laboratory reports, and imaging data. How to effectively integrate this massive amount of multi-source, heterogeneous medical data to construct a medical knowledge graph that accurately reflects the medical knowledge system and can be continuously updated and improved based on new data has become a key challenge facing the informatization construction of the medical field.

[0003] Traditional methods for constructing medical knowledge graphs have many shortcomings when dealing with multi-source heterogeneous data. For data from different sources and in different formats, there is a lack of effective spatiotemporal alignment methods, making it difficult to unify the data in both time and space dimensions, and thus failing to accurately reflect the patient's medical status throughout their entire life cycle.

[0004] Traditional medical knowledge graph update mechanisms are relatively lagging, typically involving periodic batch updates and failing to achieve real-time incremental updates to new data. This prevents the knowledge graph from reflecting the latest medical advancements and changes in patients' conditions in a timely manner, potentially leading to information lag when supporting diagnostic and treatment decisions and impacting the accuracy and timeliness of those decisions.

[0005] With the continuous increase in medical data, traditional knowledge graph construction methods suffer from high computational complexity and low efficiency when processing large-scale data, making it difficult to meet the requirements of real-time performance and accuracy. Moreover, traditional methods lack effective monitoring and early warning mechanisms for abnormal data states when constructing knowledge graphs, failing to promptly detect and handle anomalies in the data, thus affecting the quality and reliability of the knowledge graph.

[0006] When applying knowledge graphs to medical decision support, traditional methods often lack an effective joint optimization model between knowledge graphs and medical decisions. This results in the knowledge in the knowledge graph not being well translated into actual medical decisions, failing to provide doctors with personalized treatment plans, and limiting the application value of knowledge graphs in clinical practice. Traditional medical knowledge graphs are usually large in size, making them difficult to deploy and apply in resource-constrained environments such as edge computing devices, thus failing to achieve real-time medical decision support. Furthermore, in the feature extraction and dynamic update stages, there is a lack of methods for screening orthogonal feature vectors and optimizing coefficients by combining cosine similarity and convolutional neural networks, leading to low update efficiency and insufficient accuracy. At the same time, the lack of spectral clustering space partitioning and joint optimization models based on Laplacian matrices makes it difficult to achieve data anomaly monitoring and personalized treatment plan generation, hindering the further development and popularization of medical informatization. Summary of the Invention

[0007] To address the technical problem in existing technologies where there is a lack of effective spatiotemporal alignment methods for data from different sources and in different formats, leading to difficulties in unifying data across time and space dimensions and failing to accurately reflect a patient's medical status throughout their entire life cycle, this invention provides a method and apparatus for constructing a medical knowledge graph through multi-source heterogeneous data fusion and incremental updates. The technical solution is as follows:

[0008] On the one hand, a method for constructing a medical knowledge graph through multi-source heterogeneous data fusion and incremental updates is provided. This method is implemented by a medical knowledge graph construction device and includes:

[0009] Collect multi-source medical data throughout the patient's entire life cycle; the multi-source medical data includes outpatient records, inpatient medical records, laboratory reports, and imaging data;

[0010] A spatiotemporal alignment algorithm is used to align multi-source medical data in time to obtain standardized medical data; based on the distribution feature calculation method, the data fusion weight is calculated according to the standardized medical data, and the corresponding correlation matrix is ​​constructed.

[0011] Based on the singular value decomposition method, feature vectors are extracted from the correlation matrix; cosine similarity is calculated based on the feature vectors and the preset benchmark vector, and the initial coefficients are input into the convolutional neural network. Then, the coefficients are smoothed by the exponential weighted moving average method to obtain the dynamic updated coefficients of each medical event.

[0012] A baseline knowledge graph is constructed based on standardized medical data; the baseline knowledge graph is incrementally adjusted based on dynamic update coefficients to generate an updated medical knowledge graph; and knowledge distillation technology is used to output diagnostic and treatment decision support information based on the updated medical knowledge graph.

[0013] On the other hand, a medical knowledge graph construction device based on multi-source heterogeneous data fusion and incremental update is provided. This device is applied to the medical knowledge graph construction method based on multi-source heterogeneous data fusion and incremental update. The device includes:

[0014] The data acquisition module is used to collect multi-source medical data throughout the patient's entire life cycle; the multi-source medical data includes outpatient records, inpatient medical records, laboratory reports, and imaging data;

[0015] The association matrix construction module is used to perform temporal alignment of multi-source medical data using a spatiotemporal alignment algorithm to obtain standardized medical data; based on the distribution feature calculation method, it calculates the data fusion weights according to the standardized medical data and constructs the corresponding association matrix.

[0016] The coefficient update module is used to extract feature vectors based on the correlation matrix using the singular value decomposition method; calculate cosine similarity based on the feature vectors and preset benchmark vectors, input the cosine similarity into the convolutional neural network to obtain initial coefficients, and then smooth them using the exponential weighted moving average method to obtain the dynamic updated coefficients for each medical event.

[0017] The system supports an information acquisition module for constructing a baseline knowledge graph based on standardized medical data; incrementally adjusting the baseline knowledge graph based on dynamic update coefficients to generate an updated medical knowledge graph; and using knowledge distillation technology, outputting diagnostic and treatment decision support information based on the updated medical knowledge graph.

[0018] On the other hand, a medical knowledge graph construction device is provided, the medical knowledge graph construction device comprising: a processor; a memory, the memory storing computer-readable instructions, which, when executed by the processor, implement any one of the above-described methods for constructing a medical knowledge graph through multi-source heterogeneous data fusion and incremental updates.

[0019] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described methods for constructing a medical knowledge graph with multi-source heterogeneous data fusion and incremental updates.

[0020] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0021] The proposed method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates offers several significant advantages. In terms of data processing, by collecting multi-source medical data from the entire patient lifecycle and applying a spatiotemporal alignment algorithm for temporal alignment, the method effectively addresses the issue of unifying multi-source heterogeneous data across time and space dimensions. The algorithm extracts data within a preset time window, constructs tensors for normalization, and then utilizes bidirectional long short-term memory networks and graph neural networks to extract temporal features and cross-modal association features. This ensures accurate alignment of data from different sources and in different formats, laying a solid foundation for subsequent data fusion and knowledge graph construction.

[0022] In calculating the weights for data fusion, a covariance matrix of the feature vectors of adjacent events is calculated using benchmark data. The weights are then determined by combining the geometric mean of the fusion amounts within and between events. This dynamic weight calculation method fully considers the distribution characteristics of the data and the degree of correlation between events. Compared to traditional fixed weight allocation, this method more accurately reflects the importance of each medical event's data, making the fused data more representative of the true medical knowledge system.

[0023] During feature vector extraction, singular value decomposition is performed on the association matrix, and feature vectors with cumulative contribution rates exceeding a threshold are selected to form a subspace and projected, achieving effective dimensionality reduction of the data. This not only reduces computational complexity and improves processing efficiency but also preserves key feature information, ensuring the accuracy and reliability of the knowledge graph.

[0024] The calculation of dynamically updated coefficients combines techniques such as cosine similarity, convolutional neural networks, and exponentially weighted moving averages, enabling real-time adjustment of the updated coefficients based on data changes. Based on this, when incrementally adjusting the baseline graph, operations such as tensor multiplication and Kalman filter correction of discrepancies are used to achieve real-time incremental updates of the knowledge graph. This allows the graph to promptly reflect the latest medical data and knowledge, providing the most up-to-date and accurate support for diagnostic and treatment decisions.

[0025] In terms of data anomaly monitoring, a multi-dimensional medical data space is constructed, and cluster centers and outliers are extracted. By setting thresholds and density limits, data anomalies are identified and early warning signals are generated. This mechanism can promptly detect anomalies in the data, ensure the data quality of the knowledge graph, and avoid diagnostic and treatment decision-making errors caused by data anomalies.

[0026] A joint optimization model for medical knowledge graphs and treatment decisions was established. Using the weighted harmonic mean of medical data coverage and decision accuracy as the objective function, the optimal decision path was solved using gradient descent to generate personalized treatment plans. This model achieves deep integration of knowledge graphs and treatment decisions, enabling knowledge to be better transformed into practical treatment plans, improving the accuracy and personalization of treatment decisions, and providing patients with more precise medical services.

[0027] By employing knowledge distillation technology to compress knowledge graphs and generate lightweight sub-graphs, which are then deployed via edge computing devices, the problem of traditional knowledge graphs being large in size and difficult to apply in resource-constrained environments is solved. This enables real-time diagnostic and treatment decision support, promotes the application and popularization of medical informatics in primary care and edge computing scenarios, and has significant clinical application value and social significance. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of a method for constructing a medical knowledge graph through multi-source heterogeneous data fusion and incremental updating, provided by an embodiment of the present invention.

[0030] Figure 2 This is a block diagram of a medical knowledge graph construction device for multi-source heterogeneous data fusion and incremental updating provided in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the structure of a medical knowledge graph construction device provided in an embodiment of the present invention. Detailed Implementation

[0032] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0033] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0034] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0035] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0036] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0037] This invention provides a method for constructing a medical knowledge graph through multi-source heterogeneous data fusion and incremental updates. This method can be implemented using a medical knowledge graph construction device, which can be a terminal or a server. Figure 1 The flowchart shown illustrates a method for constructing a medical knowledge graph through multi-source heterogeneous data fusion and incremental updates. This method's processing flow may include the following steps:

[0038] S1. Collect multi-source medical data throughout the patient's entire life cycle; multi-source medical data includes outpatient records, inpatient medical records, laboratory reports, and imaging data;

[0039] One feasible implementation involves collecting multi-source medical data throughout the patient's entire lifecycle. This multi-source medical data includes outpatient records, inpatient medical records, laboratory reports, and imaging data. Specifically, through multiple data source interfaces such as the Hospital Information System (HIS), Laboratory Information Management System (LIS), and Picture Archiving and Communication System (PACS), medical data from each stage of the patient's lifecycle, from the initial outpatient visit and inpatient treatment to subsequent follow-up examinations, is acquired in real time or periodically. The acquired data undergoes preliminary cleaning and deduplication to ensure its integrity and accuracy.

[0040] S2. A spatiotemporal alignment algorithm is used to align multi-source medical data in time to obtain standardized medical data; based on the distribution feature calculation method, the data fusion weight is calculated according to the standardized medical data, and the corresponding correlation matrix is ​​constructed.

[0041] Optionally, a spatiotemporal alignment algorithm is used to perform temporal alignment on multi-source medical data to obtain standardized medical data, including:

[0042] Multi-source medical data is captured within a preset time window, a medical data tensor is constructed, and normalization processing is performed.

[0043] A bidirectional long short-term memory network is used to extract temporal features from medical data tensors, and outputs temporal feature vectors and attention weights.

[0044] Based on attention weights, temporal feature vectors are input into a graph neural network, cross-modal correlation features are extracted through graph convolution operations, and the feature channels are filtered through gated recurrent units to generate standardized medical data.

[0045] In one feasible implementation, the process of spatiotemporally aligning multi-source medical data to generate standardized medical data involves the following steps: A preset time window needs to be determined, the size of which must be set based on the type and characteristics of the medical data. After determining the time window, multi-source medical data is extracted, encompassing various types such as outpatient records, inpatient medical records, laboratory reports, and imaging data. Taking a patient with coronary heart disease as an example, the patient's symptom description during outpatient visits, electrocardiogram records during hospitalization, myocardial enzyme levels from blood test reports, and coronary artery CT imaging data are extracted. These data from different modalities and time points are integrated to construct a medical data tensor. The dimensions of this tensor need to be designed according to the characteristics of the data; for example, time can be used as one dimension, and different types of medical data can be used as other dimensions, thus forming a multi-dimensional tensor structure.

[0046] After constructing the medical data tensor, it needs to be normalized. Since different modalities have different data units (e.g., the voltage value in an electrocardiogram is in millivolts, while the blood glucose value in a blood test report is in millimoles per liter), without normalization, the model will be biased towards features with larger values. The normalization process uses a min-max normalization method to linearly transform the data of each modality to the [0,1] interval. Specifically, for data of a certain feature dimension, its maximum and minimum values ​​are found, and then expressed using a formula...

[0047] The data is transformed to ensure that the processed data are within the same numerical range, thus ensuring fairness in subsequent model processing.

[0048] After normalization, a Bidirectional Long Short-Term Memory (Bi-LSTM) network is used to extract temporal features from the medical data tensor. The Bi-LSTM network consists of a forward Long Short-Term Memory (LSTM) network and a backward LSTM network. The forward LSTM processes the data sequentially, while the backward LSTM processes it in reverse chronological order. The outputs of both networks are merged in the hidden layer, allowing for the simultaneous capture of past and future information and a more comprehensive understanding of the temporal dependencies within the data. Taking patient temperature monitoring data as an example, the forward LSTM can learn the trend of increasing body temperature over time, while the backward LSTM can learn the trend of decreasing body temperature over time. The merged output more accurately reflects the overall pattern of temperature changes. During Bi-LSTM processing, a temporal feature vector is generated for each time point, along with corresponding attention weights. These attention weights characterize the importance of the data at that time point within the overall temporal sequence.

[0049] After obtaining the temporal feature vector, it is input into a Graph Neural Network (GNN) to extract cross-modal correlation features through graph convolution operations. In a graph neural network, data from different modalities are treated as nodes in a graph. For example, symptom descriptions in outpatient records are one node, and electrocardiogram data are another. Edges between nodes represent the correlations between different modalities, and the weights of the edges are determined by data fusion weights. Through graph convolution operations, each node can aggregate information from its neighboring nodes, thereby extracting cross-modal correlation features. To filter out key feature channels and improve feature quality and efficiency, a Gated Recurrent Unit (GRU) is used to filter feature channels. A GRU is a simplified recurrent neural network unit that uses a gating mechanism to control the flow of information. It can automatically determine which feature channels are important and which are not, thus retaining key features, removing redundant features, and ultimately generating standardized medical data.

[0050] Optionally, based on the distribution characteristic calculation method, the data fusion weights are calculated according to the standardized medical data, and the corresponding correlation matrix is ​​constructed, including:

[0051] Standardized medical data for any medical event is selected as the baseline data, and the covariance matrix of the feature vectors of its n neighboring medical events is calculated.

[0052] The ratio of the trace of the covariance matrix to the Euclidean distance of the benchmark data is used as the intra-event fusion measure;

[0053] Calculate the mutual information between the baseline data and the preset associated event features, and take its logarithm as the inter-event fusion quantity;

[0054] The geometric mean of the fusion amount within an event and the fusion amount between events is used as the data fusion weight;

[0055] Based on standardized medical data, an association matrix is ​​constructed according to the data fusion weights;

[0056] The rows and columns of the correlation matrix correspond to medical events; the values ​​of the elements of the correlation matrix correspond to the data fusion weights of two medical events; the correlation matrix is ​​used to represent the network of relationships between medical events.

[0057] Medical events refer to all kinds of diagnosis and treatment-related events throughout a patient's entire life cycle; medical events include outpatient visits, laboratory tests, imaging diagnoses and medication records, as well as follow-up examinations after discharge.

[0058] In one feasible implementation, during the generation of standardized medical data, it is also necessary to calculate data fusion weights based on its distribution characteristics. Standardized medical data for any given medical event is selected as the baseline data.

[0059] The ratio of the trace of the covariance matrix to the Euclidean distance of the benchmark data is used as the intra-event fusion measure. The trace of the covariance matrix is ​​the sum of its main diagonal elements, representing the sum of the variances of the feature vectors of these adjacent medical events, reflecting the degree of variation in the data within the event. The Euclidean distance of the benchmark data represents the distance between the benchmark data and the data of adjacent medical events, reflecting the degree of difference between them. By comparing the two, the degree of fusion of the intra-event data can be measured. A larger ratio indicates that the degree of variation in the intra-event data is relatively large, while the degree of difference from the benchmark data is relatively small, meaning that the degree of fusion of the intra-event data is high.

[0060] Next, the mutual information between the baseline data and the features of the associated events is calculated, and its logarithm is taken as the fusion measure between the events. Mutual information is used to measure the dependency between two random variables; the larger the mutual information between the baseline data and the features of the associated events, the stronger the correlation between them. Taking the logarithm allows the value of the mutual information to be adjusted to a suitable range for subsequent calculations.

[0061] Finally, the geometric mean of the intra-event fusion amount and the inter-event fusion amount is used as the data fusion weight. The geometric mean can comprehensively consider the fusion situation within and between events, avoid the one-sidedness of a single indicator, and make the data fusion weight more reasonably reflect the degree of fusion between different medical event data.

[0062] When calculating the data fusion weights, the specific implementation method is as follows: First, it is necessary to clarify the selection rules for the benchmark data. The selection of benchmark data must be representative and can be determined according to the type or time sequence of medical events.

[0063] In calculating the covariance matrix, the feature vectors of n adjacent medical events must first be preprocessed to ensure data consistency and comparability. Taking electrocardiogram feature vectors as an example, parameters such as time span and sampling frequency need to be standardized. For feature vectors such as blood pressure monitoring values, outliers and noise interference need to be removed. After preprocessing, the covariance between these feature vectors is calculated using statistical methods, forming an n×n covariance matrix. Each element in this matrix represents the covariance of the feature vectors of two corresponding medical events, reflecting the correlation between them.

[0064] The ratio of the trace of the covariance matrix to the Euclidean distance of the benchmark data is used as the intra-event fusion measure. The trace of the covariance matrix is ​​the sum of its main diagonal elements, representing the sum of the variances of the feature vectors of the n adjacent medical events, reflecting the degree of variation in the data within the event. A larger trace indicates that the feature vectors of these adjacent medical events vary more drastically in their respective dimensions, while a smaller trace indicates more gradual variation. The Euclidean distance of the benchmark data is calculated by comparing the benchmark data with the data of the n adjacent medical events. Specifically, the benchmark data and each adjacent medical event are considered as points in a multi-dimensional space, and the degree of difference between them is measured by calculating their Euclidean distance in multi-dimensional space. The intra-event fusion measure is the ratio obtained by dividing the trace of the covariance matrix by the Euclidean distance of the benchmark data.

[0065] The mutual information between the baseline data and related event features is calculated, and its logarithm is taken as the inter-event fusion measure. Related event features refer to the features of other events that are logically related to the baseline data in terms of medical relevance. The mutual information calculation measures the dependency between the baseline data and related event features. Its principle is based on the concept of entropy in information theory, determining the mutual information value by calculating the joint entropy and marginal entropy of the two. A larger mutual information value between the baseline data and related event features indicates a higher degree of information sharing and stronger correlation between them. To adjust the mutual information value to a suitable range for subsequent comprehensive calculation with the intra-event fusion measure, the natural logarithm of the mutual information value is taken as the inter-event fusion measure.

[0066] The geometric mean of the intra-event fusion amount and the inter-event fusion amount is used as the data fusion weight. The geometric mean is calculated by multiplying the two values ​​and then taking the square root. This method comprehensively considers the fusion situation within and between events, avoiding bias in the fusion weight due to extreme values ​​of a single indicator. The data fusion weight is used to characterize the degree of fusion between different medical event data. In the subsequent construction of the medical knowledge graph, this weight will be used to adjust the importance and correlation strength of different data within the graph.

[0067] Throughout the calculation process, the timeliness and accuracy of the data must be carefully considered. Different feature extraction methods should be used to generate feature vectors for different types of medical data, such as imaging data and laboratory report data, ensuring that the feature vectors accurately reflect the essential characteristics of the data. Simultaneously, when processing multi-source heterogeneous data, unified data preprocessing is necessary, including data cleaning, format conversion, and standardization, to eliminate the impact of data differences on the calculation results. Furthermore, the value of n and the parameter settings in the mutual information calculation should be dynamically adjusted according to actual clinical needs and data characteristics to make the calculated data fusion weights more consistent with actual medical scenarios, laying the foundation for the subsequent construction of an accurate and effective medical knowledge graph. The data fusion weights calculated in this way can reasonably integrate information from multi-source medical data, improving the quality and practicality of the medical knowledge graph and providing a more reliable basis for diagnostic and treatment decision support.

[0068] S3. Based on the singular value decomposition method, feature vectors are extracted from the correlation matrix; cosine similarity is calculated based on the feature vectors and the preset benchmark vector, and the initial coefficients are input into the convolutional neural network. Then, the coefficients are smoothed by the exponential weighted moving average method to obtain the dynamic update coefficients of each medical event.

[0069] Optionally, based on the singular value decomposition method, feature vectors are extracted from the correlation matrix, including:

[0070] Perform singular value decomposition on the incidence matrix to obtain the right singular matrix;

[0071] The column vectors of the right singular matrix are determined as orthogonal eigenvectors; orthogonal eigenvectors are used to reveal the latent structure and main characteristic directions in the incidence matrix;

[0072] Select orthogonal eigenvectors whose singular value cumulative contribution rate exceeds a set threshold to form an eigenspace;

[0073] The correlation matrix is ​​projected onto the feature subspace to obtain the dimensionality-reduced feature vectors.

[0074] In one feasible implementation, the specific implementation method for extracting feature vectors and calculating dynamically updated coefficients is as follows: First, construct an association matrix with different medical event data fusion weights (the rows and columns of the association matrix correspond to medical events, and the element values ​​correspond to the data fusion weights of two medical events, representing the association network between medical events); perform singular value decomposition on the association matrix to obtain a right singular matrix and determine its column vectors as orthogonal feature vectors (used to reveal the potential structure and main feature directions in the association matrix); select orthogonal feature vectors with a singular value cumulative contribution rate exceeding a set threshold to form a feature subspace, and project the association matrix onto the feature subspace to obtain the dimensionality-reduced feature vectors;

[0075] Optionally, cosine similarity is calculated based on feature vectors and preset baseline vectors, input into a convolutional neural network to obtain initial coefficients, and then smoothed using an exponentially weighted moving average method to obtain dynamic updated coefficients for each medical event, including:

[0076] The feature vectors are standardized, and cosine similarity is calculated based on a preset benchmark vector. The preset benchmark vector is obtained based on prior knowledge or historical data in the medical field. The preset benchmark vector is used to represent the feature vectors of medical event correlation under ideal conditions.

[0077] The cosine similarity is input into the convolutional neural network, and the initial coefficients are generated after the convolutional layer-pooling layer transformation.

[0078] The convolutional layer extracts local features by performing convolution operations with cosine similarity data using different convolutional kernels; the pooling layer is used to downsample local features; and the convolutional neural network extracts deep-level features of cosine similarity through multiple convolution and pooling operations.

[0079] The initial coefficients are smoothed using an exponentially weighted moving average method, and the dynamically updated coefficients for each medical event are output.

[0080] In one feasible implementation, the dimensionality-reduced feature vectors are standardized, and their cosine similarity with a preset benchmark vector (a feature vector representing the correlation of medical events under ideal conditions, obtained based on prior knowledge or historical data in the medical field) is calculated. The cosine similarity is then input into a convolutional neural network (CNN), where it undergoes transformation through convolutional layers (extracting local features through convolution operations with different kernels and cosine similarity data) and pooling layers (downsampling local features) to generate initial coefficients (the CNN extracts deep-level features of cosine similarity through multiple convolutional and pooling operations). These initial coefficients are then smoothed using an exponentially weighted moving average to output dynamically updated coefficients for each medical event. The exponentially weighted moving average is a time-series smoothing method that assigns greater weight to recent observations and less weight to distant observations, thus reflecting data trends promptly while reducing the impact of random noise. During processing, a smoothing coefficient needs to be set, which determines the weight allocation between recent and distant data.

[0081] Throughout the implementation process, attention must be paid to data quality and model parameter tuning. For different types of medical data and different clinical scenarios, the construction method of the correlation matrix, the threshold setting of singular value decomposition, the structure of the convolutional neural network (CNN) model, and the smoothing coefficient of the exponentially weighted moving average (EWMA) method all need to be optimized according to the actual situation.

[0082] S4. Construct a baseline knowledge graph based on standardized medical data; make incremental adjustments to the baseline knowledge graph based on the dynamic update coefficient to generate an updated medical knowledge graph; use knowledge distillation technology to output diagnostic and treatment decision support information based on the updated medical knowledge graph.

[0083] Optionally, based on the dynamic update coefficients, the baseline knowledge graph is incrementally adjusted to generate an updated medical knowledge graph, including:

[0084] Based on the dynamic update coefficients, the weights of motion-related nodes and edges in the baseline graph are adjusted to obtain the weight-adjusted graph.

[0085] The difference is calculated based on the weighted adjusted graph and the baseline graph; the difference includes changes in node weights, changes in edge weights, and changes in topology.

[0086] A Kalman filter is used to dynamically correct the difference, thus obtaining the corrected difference.

[0087] The Kalman filter corrects for discrepancies based on a pre-defined dynamic model and measurement noise; during the correction process, the discrepancies are treated as observed values; the dynamic model is built based on the update patterns of the medical knowledge graph and historical update data.

[0088] The corrected discrepancies are superimposed onto the baseline graph to generate an updated medical knowledge graph.

[0089] In one feasible implementation, based on prior knowledge in the medical field or historical clinical data, the core nodes and initial associations of the baseline graph are determined. The core nodes encompass various medical events throughout a patient's life cycle (such as outpatient visits, laboratory tests, imaging diagnoses, medication records, and post-discharge follow-up tests) and diagnostic entities (such as diseases, symptoms, laboratory indicators, medications, and diagnostic procedures). The initial associations are set based on clinical diagnostic logic and the association patterns of medical events in historical data. Subsequently, standardized medical data processed by a spatiotemporal alignment algorithm is mapped to the core nodes of the baseline graph, supplementing node attribute information (such as reference ranges, units, and testing methods for laboratory indicators). The weights of the initial associations are adjusted according to the distribution characteristics of the standardized medical data, ensuring that the baseline graph reflects the core association patterns within the standardized medical data. Finally, medical professionals verify the rationality of the node system, associations, and weights of the baseline graph, deleting contradictory associations and supplementing missing nodes and attributes to form a baseline graph suitable for incremental adjustments.

[0090] When incrementally adjusting the baseline graph of standardized medical data and generating an updated medical knowledge graph, the specific implementation method is as follows: Taking the diagnosis and treatment process of a suspected coronary heart disease patient as an example, the patient's multi-source medical data is acquired, including chest pain symptom records during outpatient visits, resting electrocardiogram reports, blood myocardial enzyme test results, and imaging data from exercise stress tests, etc., and a baseline graph is constructed through preprocessing. In this baseline graph, each medical event node, such as "chest pain attack," "abnormal resting electrocardiogram," and "elevated myocardial enzymes," is connected by edges, and the weights of the edges are determined by data fusion weights, reflecting the correlation strength between each event.

[0091] The dynamically updated coefficients are multiplied by a tensor with the baseline graph to obtain the weighted graph. The difference between the weighted graph and the baseline graph is calculated. The difference calculation includes changes in node weights, changes in edge weights, and any newly added or deleted nodes and edges. Taking this patient as an example, the edge weight between the "ST-segment depression during motion" node and the "coronary heart disease diagnosis" node in the weighted graph increased by 0.3 compared to the baseline graph; this change is part of the difference. To dynamically correct the difference, a Kalman filter is used. The Kalman filter can optimally estimate and correct the difference based on the system's dynamic model and measurement noise. In this process, the difference between the weighted graph and the baseline graph is considered as the system's observation value, while the system's dynamic model is built based on the update patterns and historical update data of the medical knowledge graph. Through recursive calculation using the Kalman filter, the estimate of the difference is continuously updated, eliminating noise and uncertainty, making the difference more accurate and reliable.

[0092] Based on the dynamic update coefficients, the node weights and edge weights related to medical events in the baseline graph are adjusted to obtain a weighted graph. The difference between the weighted graph and the baseline graph (including changes in node weights, edge weights, and topology) is calculated. A Kalman filter is used to dynamically correct the difference (the Kalman filter is established based on the dynamic model of the incremental update system of the medical knowledge graph—based on the update rules and historical update data of the medical knowledge graph—and measures noise to correct the difference; the difference is treated as an observation during the correction process). The corrected difference is then superimposed onto the baseline graph to generate an updated medical knowledge graph. For this patient with coronary artery disease, the difference corrected by the Kalman filter is accurately applied to the baseline graph. The edge weights of the "ST segment depression during exercise" node and other related nodes are adjusted, and a new connection between this node and the "positive exercise stress test" node may be added, thus forming an updated medical knowledge graph. The updated graph more comprehensively reflects the changes in the patient's condition and diagnostic basis after the exercise stress test, providing more accurate information for subsequent treatment decisions.

[0093] Throughout the incremental adjustment and construction of the multidimensional medical data space, attention must be paid to the real-time nature and accuracy of the data. Newly generated medical data, such as subsequent patient follow-up results or new treatment responses, needs to be processed and updated promptly to ensure that the medical knowledge graph and multidimensional data space can reflect changes in the patient's condition in real time. Simultaneously, the system needs regular maintenance and optimization, adjusting relevant parameters based on actual clinical applications, such as the calculation parameters for dynamic update coefficients and threshold settings for spectral clustering algorithms. This allows the system to better adapt to the needs of different patients and different clinical scenarios, providing more reliable support for medical diagnosis and treatment. In this way, incremental updates to the medical knowledge graph and the construction of the multidimensional data space are achieved, improving the efficiency of medical data utilization and the accuracy of diagnostic and treatment decisions, providing strong technical support for the development of precision medicine.

[0094] Optionally, knowledge distillation technology is employed to output diagnostic and treatment decision support information based on the updated medical knowledge graph, including:

[0095] The teacher model is defined as the updated medical knowledge graph, and the student model is a lightweight subgraph.

[0096] KL divergence is used to measure the distribution difference between the teacher model and the student model, and the student model parameters are optimized through backpropagation.

[0097] When the distribution difference is less than a set threshold, a lightweight sub-map is obtained;

[0098] Lightweight sub-maps are deployed through edge computing devices, and real-time diagnostic and treatment decision support information is provided based on the deployed lightweight sub-maps.

[0099] In one feasible implementation, knowledge distillation is used to compress the medical knowledge graph, generating a lightweight subgraph. The teacher model is defined as the complete medical knowledge graph containing all nodes and edges; for example, the complete graph for a heart failure patient includes over 50 nodes such as symptoms, examinations, diagnoses, treatments, and complications. The student model is the lightweight subgraph, initially set to retain only key diagnostic and treatment nodes. KL divergence is used to measure the distribution difference between the teacher and student models, specifically calculating the difference between the edge weight distribution of "elevated BNP → heart failure diagnosis" in the complete graph and the corresponding edge weight distribution in the lightweight subgraph. The student model parameters are optimized through backpropagation.

[0100] When the distribution difference is less than a set threshold (e.g., 0.1), the lightweight sub-map generation is complete. The lightweight sub-map retains core nodes such as "shortness of breath", "elevated BNP", "diuretic treatment", and "serum potassium monitoring", reducing the number of nodes to 20, and keeping the difference between the edge weight distribution and the complete map within an acceptable range.

[0101] During the solution process of the joint optimization model, the weight parameters of the objective function need to be adjusted according to the actual clinical situation.

[0102] Taking the diagnosis and treatment of this heart failure patient as an example, after the lightweight sub-map is deployed on the edge device, when nurses enter the patient's daily weight data, the device will automatically calculate the adjustment range of weight change and diuretic dosage based on the correlation path of "weight gain → fluid retention → diuretic dosage adjustment" in the sub-map. For example, if the weight increases by 1 kg, it will suggest increasing the furosemide dose by 10 mg. At the same time, when the patient develops new symptoms such as "paroxysmal nocturnal dyspnea", the edge device will quickly match the path of "dyspnea → worsening left heart failure → need to assess pulmonary congestion" in the sub-map, prompting the doctor to perform a follow-up chest X-ray, realizing real-time push and dynamic optimization of diagnosis and treatment decisions.

[0103] Throughout the implementation process, the teacher model needs to be updated regularly to incorporate the latest clinical guidelines and research evidence. This approach enables the joint optimization and lightweight deployment of medical knowledge graphs and treatment decisions. While ensuring the accuracy of treatment decisions, it meets the real-time computing needs of edge devices, providing frontline clinicians with an efficient and convenient decision support tool. It is particularly suitable for primary care hospitals and emergency settings, helping medical staff quickly and accurately manage patient conditions and improve the quality and efficiency of medical services.

[0104] Optionally, after incrementally adjusting the baseline graph based on the dynamic update coefficients to generate the updated medical knowledge graph, the method further includes:

[0105] A multidimensional medical data space is constructed based on the updated medical knowledge graph, and cluster anomaly detection is performed using spectral clustering algorithm and median filtering algorithm to extract cluster centers and outliers.

[0106] Based on the threshold set for the distance to cluster centers and the limit value for the density of outliers, the abnormal state of the data is determined according to the cluster centers and outliers, and multi-level early warning signals are generated.

[0107] A joint optimization model for medical knowledge graph and auxiliary diagnosis and treatment decision-making is established based on a multidimensional medical data space. Based on the updated medical knowledge graph, the optimal decision path is solved using the gradient descent method according to the joint optimization model, and a personalized diagnosis and treatment plan is output.

[0108] In one feasible implementation, when constructing a multidimensional medical data space based on the updated medical knowledge graph, the updated medical knowledge graph is first mapped to a multidimensional coordinate system according to medical event types. For this patient with coronary heart disease, the dimensions of the multidimensional coordinate system may include symptom type, examination indicators, diagnostic results, treatment plan, etc.

[0109] After constructing a multidimensional medical data space, cluster centers and outliers are extracted. For this coronary artery disease patient, the cluster centers of the "high-risk patient" region may correspond to specific combinations of symptoms, examination indicators, and diagnostic results, such as chest pain lasting more than 30 minutes, ST segment depression exceeding 0.2 mV during exercise, and myocardial enzyme values ​​exceeding three times the upper limit of normal. Outliers are data points that deviate from the main cluster centers. For example, a patient may have chest pain symptoms but no significant ST segment changes during exercise, and normal myocardial enzyme values. Such data points may indicate special circumstances or the possibility of misdiagnosis.

[0110] When the distance between cluster centers exceeds a set threshold or the density of outliers exceeds a limit, the data is deemed to be in an abnormal state, and multi-level warning signals are generated. The set threshold and limit need to be determined based on professional knowledge in the medical field and historical data statistics.

[0111] When constructing a joint optimization model for medical knowledge graphs and diagnostic decisions and achieving lightweight deployment, the specific implementation method is as follows: Taking the diagnosis and treatment process of a 65-year-old patient suspected of having heart failure as an example, the initial medical data includes the patient's symptoms of shortness of breath after activity and edema in both lower extremities, an electrocardiogram showing high voltage in the left ventricle, a chest X-ray indicating pulmonary congestion, and elevated brain natriuretic peptide (BNP) levels in blood tests. Based on this data, an initial medical knowledge graph is constructed, where nodes cover medical events such as "shortness of breath," "edema," "high voltage in the left ventricle," "pulmonary congestion," and "elevated BNP." The weights of the edges are determined by data fusion weights, reflecting the strength of the association between each event. For example, the edge weight between the "elevated BNP" and the diagnostic node for "heart failure" is 0.8, indicating that this indicator is of high importance to the diagnosis.

[0112] Optionally, a multidimensional medical data space is constructed based on the updated medical knowledge graph, and cluster anomaly detection is performed using spectral clustering and median filtering algorithms to extract cluster centers and outliers, including:

[0113] The updated medical knowledge graph is mapped to a multi-dimensional coordinate system according to the type of medical event. Each medical event is projected as a data point into the multi-dimensional coordinate system to generate a data distribution point set.

[0114] The spectral clustering algorithm is used to spatially partition the data distribution point set: the data point set is regarded as the nodes of a graph, and the similarity between nodes is regarded as the weight of the edge. By constructing the Laplacian matrix and performing feature decomposition, the data points are divided into different clusters, and the density gradient and silhouette coefficient of the partition space are calculated.

[0115] Based on the silhouette coefficient, the density gradient is smoothed by the median filtering algorithm to eliminate noise interference and extract cluster centers and outliers.

[0116] In one feasible implementation, a spectral clustering algorithm is used to spatially partition the data distribution point set. Spectral clustering is a graph-based clustering method that treats the data point set as nodes in a graph, with the similarity between nodes considered as edge weights. By constructing a Laplacian matrix and performing eigenvalue decomposition, the data points are divided into different clusters. For the patient's data distribution point set, the spectral clustering algorithm can divide them into different spatial regions, such as "low-risk patient," "intermediate-risk patient," and "high-risk patient" regions, based on the similarity between the data points of each medical event. During the spatial partitioning process, the density gradient and silhouette coefficient of the partitioned space are calculated. The density gradient reflects the variation in the distribution density of data points in the space.

[0117] Median filtering is used to smooth the density gradient and eliminate noise interference. Median filtering is a non-linear filtering method that sorts pixel values ​​within a window and takes the median as the output value. It effectively removes noise while preserving edge information. For this patient's data distribution space, there may be fluctuations in the density gradient due to individual outliers. Median filtering smooths these fluctuations, making the density gradient curve smoother and more accurate, thus more clearly reflecting the true data distribution.

[0118] Optionally, based on the updated medical knowledge graph, the optimal decision path is solved using gradient descent according to the joint optimization model, and a personalized treatment plan is output, including:

[0119] The objective function is defined as the weighted harmonic mean of medical data coverage and decision accuracy; medical data coverage refers to the proportion of patient data dimensions to total data dimensions; decision accuracy refers to the degree to which treatment recommendations conform to clinical guidelines.

[0120] Define the constraint as compliance with the diagnostic and treatment guidelines;

[0121] The objective function is decomposed into a data optimization subproblem and a decision optimization subproblem by performing a Laplace relaxation transformation.

[0122] The data optimization subproblem and the decision optimization subproblem are solved alternately and iteratively until convergence, and the optimal decision path that satisfies the constraints is output.

[0123] The optimal decision path is matched with the updated medical knowledge graph in real time to generate personalized treatment plans.

[0124] In one feasible implementation, the objective function is defined as the weighted harmonic mean of medical data coverage and decision accuracy. Medical data coverage refers to the proportion of patient data dimensions considered by the model out of the total data dimensions.

[0125] The objective function is transformed using a Lagrange relaxation method, decomposing it into a data optimization subproblem and a decision optimization subproblem. The data optimization subproblem focuses on how to extract key information from patients' multi-source medical data to improve data coverage.

[0126] The two subproblems are solved iteratively, alternating between them, until convergence. In the first iteration, the data optimization subproblem is solved first. By analyzing the patient's medical data, the weights of each indicator are adjusted to increase the data coverage to 75%. Then, the adjusted weights are input into the decision optimization subproblem to generate a preliminary treatment decision path, such as recommending intravenous furosemide injection and monitoring electrolytes. Next, based on the output of the decision optimization subproblem, the data optimization subproblem is returned to further adjust the data weights, and this process is repeated iteratively. When the value of the objective function changes by less than 5% over five consecutive iterations, the model is considered converged, and the optimal decision path that satisfies the constraints is output.

[0127] The optimal decision path is matched in real time with the updated medical knowledge graph to generate personalized treatment plans. If a patient's blood potassium level drops to 3.2 mmol / L 24 hours after treatment, the updated medical knowledge graph will add a new node for this data and dynamically adjust the edge weights of the "hypokalemia" and "diuretic use" nodes to 0.9. At this point, the optimal decision path automatically matches the new data and generates an adjusted plan: "Discontinue furosemide, switch to torasemide and combine with oral potassium supplementation, monitor blood potassium levels every 2 hours," thus achieving dynamic adjustment of the treatment plan.

[0128] The proposed method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates offers several significant advantages. In terms of data processing, by collecting multi-source medical data from the entire patient lifecycle and applying a spatiotemporal alignment algorithm for temporal alignment, the method effectively addresses the issue of unifying multi-source heterogeneous data across time and space dimensions. The algorithm extracts data within a preset time window, constructs tensors for normalization, and then utilizes bidirectional long short-term memory networks and graph neural networks to extract temporal features and cross-modal association features. This ensures accurate alignment of data from different sources and in different formats, laying a solid foundation for subsequent data fusion and knowledge graph construction.

[0129] In calculating the weights for data fusion, a covariance matrix of the feature vectors of adjacent events is calculated using benchmark data. The weights are then determined by combining the geometric mean of the fusion amounts within and between events. This dynamic weight calculation method fully considers the distribution characteristics of the data and the degree of correlation between events. Compared to traditional fixed weight allocation, this method more accurately reflects the importance of each medical event's data, making the fused data more representative of the true medical knowledge system.

[0130] During feature vector extraction, singular value decomposition is performed on the association matrix, and feature vectors with cumulative contribution rates exceeding a threshold are selected to form a subspace and projected, achieving effective dimensionality reduction of the data. This not only reduces computational complexity and improves processing efficiency but also preserves key feature information, ensuring the accuracy and reliability of the knowledge graph.

[0131] The calculation of dynamically updated coefficients combines techniques such as cosine similarity, convolutional neural networks, and exponentially weighted moving averages, enabling real-time adjustment of the updated coefficients based on data changes. Based on this, when incrementally adjusting the baseline graph, operations such as tensor multiplication and Kalman filter correction of discrepancies are used to achieve real-time incremental updates of the knowledge graph. This allows the graph to promptly reflect the latest medical data and knowledge, providing the most up-to-date and accurate support for diagnostic and treatment decisions.

[0132] In terms of data anomaly monitoring, a multi-dimensional medical data space is constructed, and cluster centers and outliers are extracted. By setting thresholds and density limits, data anomalies are identified and early warning signals are generated. This mechanism can promptly detect anomalies in the data, ensure the data quality of the knowledge graph, and avoid diagnostic and treatment decision-making errors caused by data anomalies.

[0133] A joint optimization model for medical knowledge graphs and treatment decisions was established. Using the weighted harmonic mean of medical data coverage and decision accuracy as the objective function, the optimal decision path was solved using gradient descent to generate personalized treatment plans. This model achieves deep integration of knowledge graphs and treatment decisions, enabling knowledge to be better transformed into practical treatment plans, improving the accuracy and personalization of treatment decisions, and providing patients with more precise medical services.

[0134] By employing knowledge distillation technology to compress knowledge graphs and generate lightweight sub-graphs, which are then deployed via edge computing devices, the problem of traditional knowledge graphs being large in size and difficult to apply in resource-constrained environments is solved. This enables real-time diagnostic and treatment decision support, promotes the application and popularization of medical informatics in primary care and edge computing scenarios, and has significant clinical application value and social significance.

[0135] Figure 2 This is a block diagram of a medical knowledge graph construction device for multi-source heterogeneous data fusion and incremental updating provided in an embodiment of the present invention. This device is used for a method of constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updating. (Refer to...) Figure 2 The device includes a data acquisition module 210, an association matrix construction module 220, a coefficient update module 230, and a support information acquisition module 240. Among them:

[0136] Data acquisition module 210 is used to collect multi-source medical data throughout the patient's life cycle; multi-source medical data includes outpatient records, inpatient medical records, laboratory reports and imaging data;

[0137] The association matrix construction module 220 is used to perform temporal alignment of multi-source medical data using a spatiotemporal alignment algorithm to obtain standardized medical data; based on the distribution feature calculation method, it calculates the data fusion weights according to the standardized medical data and constructs the corresponding association matrix.

[0138] The coefficient update module 230 is used to extract feature vectors based on the correlation matrix using the singular value decomposition method; calculate cosine similarity based on the feature vectors and preset benchmark vectors, input the cosine similarity into the convolutional neural network to obtain initial coefficients, and then smooth them using the exponential weighted moving average method to obtain the dynamic updated coefficients of each medical event.

[0139] The system supports an information acquisition module 240, which is used to construct a baseline graph based on standardized medical data; incrementally adjust the baseline graph according to the dynamic update coefficient to generate an updated medical knowledge graph; and output diagnostic and treatment decision support information based on the updated medical knowledge graph using knowledge distillation technology.

[0140] Optionally, the association matrix construction module 220 is further used for:

[0141] Multi-source medical data is captured within a preset time window, a medical data tensor is constructed, and normalization processing is performed.

[0142] A bidirectional long short-term memory network is used to extract temporal features from medical data tensors, and outputs temporal feature vectors and attention weights.

[0143] Based on attention weights, temporal feature vectors are input into a graph neural network, cross-modal correlation features are extracted through graph convolution operations, and the feature channels are filtered through gated recurrent units to generate standardized medical data.

[0144] Optionally, the association matrix construction module 220 is further used for:

[0145] Standardized medical data for any medical event is selected as the baseline data, and the covariance matrix of the feature vectors of its n neighboring medical events is calculated.

[0146] The ratio of the trace of the covariance matrix to the Euclidean distance of the benchmark data is used as the intra-event fusion measure;

[0147] Calculate the mutual information between the baseline data and the preset associated event features, and take its logarithm as the inter-event fusion quantity;

[0148] The geometric mean of the fusion amount within an event and the fusion amount between events is used as the data fusion weight;

[0149] Based on standardized medical data, an association matrix is ​​constructed according to the data fusion weights;

[0150] The rows and columns of the correlation matrix correspond to medical events; the values ​​of the elements of the correlation matrix correspond to the data fusion weights of two medical events; the correlation matrix is ​​used to represent the network of relationships between medical events.

[0151] Medical events refer to all kinds of diagnosis and treatment-related events throughout a patient's entire life cycle; medical events include outpatient visits, laboratory tests, imaging diagnoses and medication records, as well as follow-up examinations after discharge.

[0152] Optionally, the coefficient update module 230 is further used for:

[0153] Perform singular value decomposition on the incidence matrix to obtain the right singular matrix;

[0154] The column vectors of the right singular matrix are determined as orthogonal eigenvectors; orthogonal eigenvectors are used to reveal the latent structure and main characteristic directions in the incidence matrix;

[0155] Select orthogonal eigenvectors whose singular value cumulative contribution rate exceeds a set threshold to form an eigenspace;

[0156] The correlation matrix is ​​projected onto the feature subspace to obtain the dimensionality-reduced feature vectors.

[0157] Optionally, the coefficient update module 230 is further used for:

[0158] The feature vectors are standardized, and cosine similarity is calculated based on a preset benchmark vector. The preset benchmark vector is obtained based on prior knowledge or historical data in the medical field. The preset benchmark vector is used to represent the feature vectors of medical event correlation under ideal conditions.

[0159] The cosine similarity is input into the convolutional neural network, and the initial coefficients are generated after the convolutional layer-pooling layer transformation.

[0160] The convolutional layer extracts local features by performing convolution operations with cosine similarity data using different convolutional kernels; the pooling layer is used to downsample local features; and the convolutional neural network extracts deep-level features of cosine similarity through multiple convolution and pooling operations.

[0161] The initial coefficients are smoothed using an exponentially weighted moving average method, and the dynamically updated coefficients for each medical event are output.

[0162] Optionally, the information acquisition module 240 is further used for:

[0163] Based on the dynamic update coefficients, the weights of motion-related nodes and edges in the baseline graph are adjusted to obtain the weight-adjusted graph.

[0164] The difference is calculated based on the weighted adjusted graph and the baseline graph; the difference includes changes in node weights, changes in edge weights, and changes in topology.

[0165] A Kalman filter is used to dynamically correct the difference, thus obtaining the corrected difference.

[0166] The Kalman filter corrects for discrepancies based on a pre-defined dynamic model and measurement noise; during the correction process, the discrepancies are treated as observed values; the dynamic model is built based on the update patterns of the medical knowledge graph and historical update data.

[0167] The corrected discrepancies are superimposed onto the baseline graph to generate an updated medical knowledge graph.

[0168] Optionally, information acquisition module 240 is supported:

[0169] A multidimensional medical data space is constructed based on the updated medical knowledge graph, and cluster anomaly detection is performed using spectral clustering algorithm and median filtering algorithm to extract cluster centers and outliers.

[0170] Based on the threshold set for the distance to cluster centers and the limit value for the density of outliers, the abnormal state of the data is determined according to the cluster centers and outliers, and multi-level early warning signals are generated.

[0171] A joint optimization model for medical knowledge graph and auxiliary diagnosis and treatment decision-making is established based on a multidimensional medical data space. Based on the updated medical knowledge graph, the optimal decision path is solved using the gradient descent method according to the joint optimization model, and a personalized diagnosis and treatment plan is output.

[0172] Optionally, information acquisition module 240 is supported:

[0173] The updated medical knowledge graph is mapped to a multi-dimensional coordinate system according to the type of medical event. Each medical event is projected as a data point into the multi-dimensional coordinate system to generate a data distribution point set.

[0174] The spectral clustering algorithm is used to spatially partition the data distribution point set: the data point set is regarded as the nodes of a graph, and the similarity between nodes is regarded as the weight of the edge. By constructing the Laplacian matrix and performing feature decomposition, the data points are divided into different clusters, and the density gradient and silhouette coefficient of the partition space are calculated.

[0175] Based on the silhouette coefficient, the density gradient is smoothed by the median filtering algorithm to eliminate noise interference and extract cluster centers and outliers.

[0176] Optionally, information acquisition module 240 is supported:

[0177] The objective function is defined as the weighted harmonic mean of medical data coverage and decision accuracy; medical data coverage refers to the proportion of patient data dimensions to total data dimensions; decision accuracy refers to the degree to which treatment recommendations conform to clinical guidelines.

[0178] Define the constraint as compliance with the diagnostic and treatment guidelines;

[0179] The objective function is decomposed into a data optimization subproblem and a decision optimization subproblem by performing a Laplace relaxation transformation.

[0180] The data optimization subproblem and the decision optimization subproblem are solved alternately and iteratively until convergence, and the optimal decision path that satisfies the constraints is output.

[0181] The optimal decision path is matched with the updated medical knowledge graph in real time to generate personalized treatment plans.

[0182] Optionally, information acquisition module 240 is supported:

[0183] The teacher model is defined as the updated medical knowledge graph, and the student model is a lightweight subgraph.

[0184] KL divergence is used to measure the distribution difference between the teacher model and the student model, and the student model parameters are optimized through backpropagation.

[0185] When the distribution difference is less than a set threshold, a lightweight sub-map is obtained;

[0186] Lightweight sub-maps are deployed through edge computing devices, and real-time diagnostic and treatment decision support information is provided based on the deployed lightweight sub-maps.

[0187] The proposed method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates offers several significant advantages. In terms of data processing, by collecting multi-source medical data from the entire patient lifecycle and applying a spatiotemporal alignment algorithm for temporal alignment, the method effectively addresses the issue of unifying multi-source heterogeneous data across time and space dimensions. The algorithm extracts data within a preset time window, constructs tensors for normalization, and then utilizes bidirectional long short-term memory networks and graph neural networks to extract temporal features and cross-modal association features. This ensures accurate alignment of data from different sources and in different formats, laying a solid foundation for subsequent data fusion and knowledge graph construction.

[0188] In calculating the weights for data fusion, a covariance matrix of the feature vectors of adjacent events is calculated using benchmark data. The weights are then determined by combining the geometric mean of the fusion amounts within and between events. This dynamic weight calculation method fully considers the distribution characteristics of the data and the degree of correlation between events. Compared to traditional fixed weight allocation, this method more accurately reflects the importance of each medical event's data, making the fused data more representative of the true medical knowledge system.

[0189] During feature vector extraction, singular value decomposition is performed on the association matrix, and feature vectors with cumulative contribution rates exceeding a threshold are selected to form a subspace and projected, achieving effective dimensionality reduction of the data. This not only reduces computational complexity and improves processing efficiency but also preserves key feature information, ensuring the accuracy and reliability of the knowledge graph.

[0190] The calculation of dynamically updated coefficients combines techniques such as cosine similarity, convolutional neural networks, and exponentially weighted moving averages, enabling real-time adjustment of the updated coefficients based on data changes. Based on this, when incrementally adjusting the baseline graph, operations such as tensor multiplication and Kalman filter correction of discrepancies are used to achieve real-time incremental updates of the knowledge graph. This allows the graph to promptly reflect the latest medical data and knowledge, providing the most up-to-date and accurate support for diagnostic and treatment decisions.

[0191] In terms of data anomaly monitoring, a multi-dimensional medical data space is constructed, and cluster centers and outliers are extracted. By setting thresholds and density limits, data anomalies are identified and early warning signals are generated. This mechanism can promptly detect anomalies in the data, ensure the data quality of the knowledge graph, and avoid diagnostic and treatment decision-making errors caused by data anomalies.

[0192] A joint optimization model for medical knowledge graphs and treatment decisions was established. Using the weighted harmonic mean of medical data coverage and decision accuracy as the objective function, the optimal decision path was solved using gradient descent to generate personalized treatment plans. This model achieves deep integration of knowledge graphs and treatment decisions, enabling knowledge to be better transformed into practical treatment plans, improving the accuracy and personalization of treatment decisions, and providing patients with more precise medical services.

[0193] By employing knowledge distillation technology to compress knowledge graphs and generate lightweight sub-graphs, which are then deployed via edge computing devices, the problem of traditional knowledge graphs being large in size and difficult to apply in resource-constrained environments is solved. This enables real-time diagnostic and treatment decision support, promotes the application and popularization of medical informatics in primary care and edge computing scenarios, and has significant clinical application value and social significance.

[0194] Figure 3 This is a schematic diagram of the structure of a medical knowledge graph construction device provided in an embodiment of the present invention, such as... Figure 3 As shown, the medical knowledge graph construction device may include the above-mentioned Figure 2 The illustrated medical knowledge graph construction apparatus involves multi-source heterogeneous data fusion and incremental updates. Optionally, the medical knowledge graph construction apparatus 310 may include a first processor 2001.

[0195] Optionally, the medical knowledge graph construction device 310 may also include a memory 2002 and a transceiver 2003.

[0196] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0197] The following is combined with Figure 3 A detailed introduction to each component of the medical knowledge graph construction device 310:

[0198] The first processor 2001 is the control center of the medical knowledge graph construction device 310. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0199] Optionally, the first processor 2001 can perform various functions of the medical knowledge graph construction device 310 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0200] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 3 CPU0 and CPU1 are shown in the diagram.

[0201] In a specific implementation, as one example, the medical knowledge graph construction device 310 may also include multiple processors, for example... Figure 3 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0202] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0203] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or exist independently, and may be connected to the interface circuit of the device 310 via a medical knowledge graph. Figure 3 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0204] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0205] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 3 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0206] Optionally, the transceiver 2003 can be integrated with the first processor 2001 or exist independently, and the interface circuit of the device 310 can be constructed through a medical knowledge graph. Figure 3 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0207] It should be noted that, Figure 3 The structure of the medical knowledge graph construction device 310 shown in the diagram does not constitute a limitation on the router. Actual medical knowledge graph construction devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0208] Furthermore, the technical effects of the medical knowledge graph construction device 310 can be referenced from the technical effects of the medical knowledge graph construction method of multi-source heterogeneous data fusion and incremental update described in the above method embodiments, and will not be repeated here.

[0209] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor, or it may be any conventional processor.

[0210] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0211] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0212] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0213] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0214] It should be understood that, in various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0215] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software 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 this invention.

[0216] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0217] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0218] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0219] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0220] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0221] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for constructing a medical knowledge graph through multi-source heterogeneous data fusion and incremental updates, characterized in that, The method includes: Collect multi-source medical data throughout the patient's entire life cycle; the multi-source medical data includes outpatient records, inpatient medical records, laboratory reports, and imaging data; A spatiotemporal alignment algorithm is used to align multi-source medical data in time to obtain standardized medical data; based on the distribution feature calculation method, the data fusion weight is calculated according to the standardized medical data, and the corresponding correlation matrix is ​​constructed. Based on the singular value decomposition method, feature vectors are extracted from the correlation matrix; cosine similarity is calculated based on the feature vectors and the preset benchmark vector, and the initial coefficients are input into the convolutional neural network. Then, the coefficients are smoothed by the exponential weighted moving average method to obtain the dynamic updated coefficients of each medical event. A baseline knowledge graph is constructed based on standardized medical data; the baseline knowledge graph is incrementally adjusted based on dynamic update coefficients to generate an updated medical knowledge graph; and knowledge distillation technology is used to output diagnostic and treatment decision support information based on the updated medical knowledge graph.

2. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 1, characterized in that, The method employs a spatiotemporal alignment algorithm to perform temporal alignment on multi-source medical data, resulting in standardized medical data, including: Multi-source medical data is captured within a preset time window, a medical data tensor is constructed, and normalization processing is performed. A bidirectional long short-term memory network is used to extract temporal features from medical data tensors, and outputs temporal feature vectors and attention weights. Based on attention weights, temporal feature vectors are input into a graph neural network, cross-modal correlation features are extracted through graph convolution operations, and the feature channels are filtered through gated recurrent units to generate standardized medical data.

3. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 1, characterized in that, The distribution feature-based calculation method calculates data fusion weights based on standardized medical data and constructs a corresponding correlation matrix, including: Standardized medical data for any medical event is selected as the baseline data, and the covariance matrix of the feature vectors of its n neighboring medical events is calculated. The ratio of the trace of the covariance matrix to the Euclidean distance of the benchmark data is used as the intra-event fusion measure; Calculate the mutual information between the baseline data and the preset associated event features, and take its logarithm as the inter-event fusion quantity; The geometric mean of the fusion amount within an event and the fusion amount between events is used as the data fusion weight; Based on standardized medical data, an association matrix is ​​constructed according to the data fusion weights; The rows and columns of the correlation matrix correspond to medical events; the values ​​of the elements of the correlation matrix correspond to the data fusion weights of two medical events; the correlation matrix is ​​used to characterize the network of relationships between medical events. The medical events refer to various diagnosis and treatment-related events throughout a patient's entire life cycle; these medical events include outpatient visits, laboratory tests, imaging diagnoses, medication records, and follow-up tests after discharge.

4. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 1, characterized in that, The singular value decomposition-based method extracts feature vectors from the correlation matrix, including: Perform singular value decomposition on the incidence matrix to obtain the right singular matrix; The column vectors of the right singular matrix are determined as orthogonal eigenvectors; these orthogonal eigenvectors are used to reveal the latent structure and main characteristic directions in the incidence matrix. Select orthogonal eigenvectors whose singular value cumulative contribution rate exceeds a set threshold to form an eigenspace; The correlation matrix is ​​projected onto the feature subspace to obtain the dimensionality-reduced feature vectors.

5. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 1, characterized in that, The process of calculating cosine similarity based on feature vectors and preset baseline vectors, inputting the initial coefficients into a convolutional neural network, and then smoothing them using an exponentially weighted moving average method to obtain the dynamically updated coefficients for each medical event includes: The feature vectors are standardized, and cosine similarity is calculated based on a preset benchmark vector. The preset benchmark vector is obtained based on prior knowledge or historical data in the medical field. The preset benchmark vector is used to represent the feature vectors of medical event correlation under ideal conditions. The cosine similarity is input into the convolutional neural network, and the initial coefficients are generated after the convolutional layer-pooling layer transformation. The convolutional layer extracts local features by performing convolution operations with cosine similarity data using different convolutional kernels; the pooling layer is used to downsample local features; and the convolutional neural network extracts deep-level features of cosine similarity through multiple convolution and pooling operations. The initial coefficients are smoothed using an exponentially weighted moving average method, and the dynamically updated coefficients for each medical event are output.

6. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 1, characterized in that, The step of incrementally adjusting the baseline knowledge graph based on the dynamic update coefficient to generate an updated medical knowledge graph includes: Based on the dynamic update coefficients, the weights of motion-related nodes and edges in the baseline graph are adjusted to obtain the weight-adjusted graph. The difference is calculated based on the weighted adjusted graph and the baseline graph; the difference includes changes in node weights, changes in edge weights, and changes in topology. A Kalman filter is used to dynamically correct the difference, thus obtaining the corrected difference. The Kalman filter corrects the difference based on a preset dynamic model and measurement noise; during the correction process, the difference is treated as an observed value; the dynamic model is established based on the update rules of the medical knowledge graph and historical update data. The corrected discrepancies are superimposed onto the baseline graph to generate an updated medical knowledge graph.

7. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 1, characterized in that, After incrementally adjusting the baseline knowledge graph based on the dynamic update coefficient to generate the updated medical knowledge graph, the method further includes: A multidimensional medical data space is constructed based on the updated medical knowledge graph, and cluster anomaly detection is performed using spectral clustering algorithm and median filtering algorithm to extract cluster centers and outliers. Based on the threshold set for the distance to cluster centers and the limit value for the density of outliers, the abnormal state of the data is determined according to the cluster centers and outliers, and multi-level early warning signals are generated. A joint optimization model for medical knowledge graph and auxiliary diagnosis and treatment decision-making is established based on a multidimensional medical data space. Based on the updated medical knowledge graph, the optimal decision path is solved using the gradient descent method according to the joint optimization model, and a personalized diagnosis and treatment plan is output.

8. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 7, characterized in that, The process involves constructing a multidimensional medical data space based on the updated medical knowledge graph, and using spectral clustering and median filtering algorithms for cluster anomaly detection to extract cluster centers and outliers, including: The updated medical knowledge graph is mapped to a multi-dimensional coordinate system according to the type of medical event. Each medical event is projected as a data point into the multi-dimensional coordinate system to generate a data distribution point set. The data distribution point set is spatially partitioned using a spectral clustering algorithm: the data point set is regarded as nodes of a graph, and the similarity between nodes is regarded as the weight of the edge. By constructing a Laplacian matrix and performing feature decomposition, the data points are divided into different clusters, and the density gradient and silhouette coefficient of the partition space are calculated. Based on the silhouette coefficient, the density gradient is smoothed by the median filtering algorithm to eliminate noise interference and extract cluster centers and outliers.

9. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 7, characterized in that, Based on the updated medical knowledge graph, the optimal decision path is solved using gradient descent according to the joint optimization model, and a personalized treatment plan is output, including: The objective function is defined as the weighted harmonic mean of medical data coverage and decision accuracy; the medical data coverage refers to the proportion of patient data dimensions to total data dimensions; the decision accuracy refers to the degree to which treatment recommendations conform to clinical guidelines. Define the constraint as compliance with the diagnostic and treatment guidelines; The objective function is decomposed into a data optimization subproblem and a decision optimization subproblem by performing a Laplace relaxation transformation. The data optimization subproblem and the decision optimization subproblem are solved alternately and iteratively until convergence, and the optimal decision path that satisfies the constraints is output. The optimal decision path is matched with the updated medical knowledge graph in real time to generate personalized treatment plans.

10. The method for constructing a medical knowledge graph based on multi-source heterogeneous data fusion and incremental updates according to claim 1, characterized in that, The method employs knowledge distillation technology to output diagnostic and treatment decision support information based on the updated medical knowledge graph, including: The teacher model is defined as the updated medical knowledge graph, and the student model is a lightweight subgraph. KL divergence is used to measure the distribution difference between the teacher model and the student model, and the student model parameters are optimized through backpropagation. When the distribution difference is less than a set threshold, a lightweight sub-map is obtained; Lightweight sub-maps are deployed through edge computing devices, and real-time diagnostic and treatment decision support information is provided based on the deployed lightweight sub-maps.

Images