Infectious disease tracing method and system based on spatio-temporal graph neural network

Abstract

The present application relates to the technical field of public health safety, in particular to an infectious disease tracing method and system based on a space-time graph neural network, which integrates multi-source heterogeneous contact data, constructs a topologically enhanced space-time graph, converts the data into a space-time point cloud, constructs a multi-scale simplex complex, extracts persistent topological features, integrates the features into a graph representation, trains a multi-scale space-time graph neural network, uses multi-level time encoding and a double attention mechanism, designs a topological feature guided message passing, and finally performs topologically guided backpropagation tracing, reversely traces persistent homologous information flow and spectral graph convolution, generates multiple propagation source point hypotheses, and outputs the probability distribution of potential infection sources and the propagation path.

Description

Technical Field

[0001] This invention relates to the field of public health and safety technology, and in particular to a method and system for tracing the source of infectious diseases based on spatiotemporal graph neural networks. This system is used to quickly and accurately identify the source of infection through multi-source heterogeneous data after an outbreak of an infectious disease, and to provide precise prevention and control decision support. Background Technology

[0002] In infectious disease control, tracing the source is a crucial step in cutting off the transmission chain and preventing further spread of the epidemic. Traditional tracing methods are mainly based on epidemiological investigations, involving manual tracing through contact tracing and case analysis. This method relies on expert experience, is time-consuming, and is difficult to handle complex transmission scenarios.

[0003] With the development of computing technology, source tracing methods based on network science and machine learning have gradually emerged. Existing technologies mainly include several categories such as centrality measures, propagation simulation methods, and inference algorithms. Centrality measures calculate feature values ​​based on the position of nodes in the network, such as distance centrality and Jordan centrality, but these methods ignore the randomness and dynamic characteristics of the propagation process; propagation simulation methods, such as soft-boundary Monte Carlo estimators, evaluate the probability of each node being a source point by simulating multiple propagation processes, but they perform poorly when the contact network is dense or the estimated parameters deviate from the actual propagation process; existing deep learning methods usually only consider the static network structure and lack effective modeling of the time dimension.

[0004] Existing source tracing methods have three main shortcomings: First, they rely on high-quality and complete contact network data, which is often difficult to obtain in the actual epidemic; second, they have limited support for multi-source outbreak scenarios and are difficult to identify multiple independent transmission sources; and third, they lack in-depth analysis of network topology characteristics and are difficult to capture key structural information in the transmission network.

[0005] Therefore, there is an urgent need for an infectious disease tracing method that can adapt to incomplete data, support multi-source tracing, and make full use of network topology characteristics. Summary of the Invention

[0006] In view of the above problems, the present invention provides a method and system for tracing the source of infectious diseases based on spatiotemporal graph neural networks, which integrates topological persistence theory and spatiotemporal graph neural networks to achieve high-precision identification and tracking of the source of infectious disease transmission.

[0007] This invention mainly solves the following technical problems: how to obtain reliable tracing results even when the data is incomplete; how to effectively model the complex structure and dynamic evolution characteristics in spatiotemporal networks; how to support accurate tracing in multi-source propagation scenarios; and how to quantify the uncertainty of tracing results and provide reliable decision support.

[0008] This invention achieves the above objectives through three progressively innovative points: first, it proposes a spatiotemporal graph representation method driven by topology persistence; second, it designs a multi-scale spatiotemporal graph neural network architecture; and third, it implements a topology-guided backpropagation tracing mechanism.

[0009] This invention proposes a method for tracing the origin of infectious diseases based on spatiotemporal graph neural networks, including:

[0010] Acquire multi-source heterogeneous contact data, which includes contact object, contact time, contact location, and contact attributes;

[0011] Based on the aforementioned multi-source heterogeneous contact data, a topology-enhanced spatiotemporal graph is constructed, including:

[0012] The multi-source heterogeneous contact data is converted into a spatiotemporal point cloud;

[0013] Based on the spatiotemporal point cloud, a multi-scale simple complex sequence is constructed;

[0014] Calculate the persistence map of the multi-scale simplex sequence;

[0015] Persistent topological features are extracted and fused into the spatiotemporal graph representation;

[0016] Training a multi-scale spatiotemporal graph neural network includes:

[0017] Construct multi-level time coding to capture the absolute position, periodicity, and relative interval information of time;

[0018] Calculate dual attention weights based on physical spatial distance and network topology distance;

[0019] Design a topology-feature-guided message passing mechanism;

[0020] Based on the aforementioned multi-scale spatiotemporal graph neural network, topology-guided backpropagation tracing is performed, including:

[0021] Perform persistent reverse tracing of the homologous information flow for the final infection distribution;

[0022] The initial state graph structure features are learned by inverse learning through spectral graph convolution;

[0023] Generate and evaluate multiple propagation source hypotheses;

[0024] Output potential sources of infection, their probability distribution, and transmission routes.

[0025] Preferably, the acquisition of multi-source heterogeneous contact data includes:

[0026] Raw contact information is obtained from movement trajectory data, public transportation records, location check-in data, and medical system data;

[0027] The original contact information is spatiotemporally aligned to unify the temporal granularity and spatial reference system;

[0028] Assess the quality of the original contact information and perform missing data repair;

[0029] Generate a standardized spatiotemporal contact dataset.

[0030] Preferably, the conversion of multi-source heterogeneous contact data into spatiotemporal point clouds includes:

[0031] Create spatiotemporal coordinates for each contact record (i,j,t,d), where i and j are the contacting individuals, t is the time, and d is the contact duration;

[0032] Attributes such as contact intensity and duration are added to each spatiotemporal coordinate point;

[0033] Construct a point cloud data structure that supports efficient spatiotemporal queries.

[0034] Preferably, the construction of the multi-scale simple complex sequence includes:

[0035] Set spatial distance threshold εs and time interval threshold εt to form a parameterized spatiotemporal neighborhood;

[0036] By systematically varying the spatial distance threshold εs and the time interval threshold εt, nested simple complex sequences are constructed.

[0037] The spatial distance threshold εs and time interval threshold εt are adaptively adjusted based on the characteristics of infectious diseases and data distribution.

[0038] Preferably, the calculation of the persistence map of the multi-scale simplex sequence includes:

[0039] Calculate the 0-persistence graph to characterize the appearance and disappearance of connected components;

[0040] Calculate the durability spectrum of ring structure 1 to characterize the formation and destruction of ring structures;

[0041] Calculate the durability spectrum to characterize the formation and disappearance of cavity structures;

[0042] Extract statistical features of persistent barcodes, including average lifetime, maximum persistence, and entropy value.

[0043] Preferably, the construction of the multi-level time coding includes:

[0044] Generate absolute time location codes to represent time markers such as time of day and date;

[0045] Generate periodic pattern codes to capture periodic behavioral patterns such as intraday and intraweek patterns;

[0046] Generate relative time interval codes to describe the distribution characteristics of event intervals;

[0047] The absolute time position encoding, periodic pattern encoding, and relative time interval encoding are fused to form a multidimensional time feature vector.

[0048] Preferably, the message passing mechanism guided by the design topology features includes:

[0049] Message importance weights are assigned based on persistent topological features;

[0050] For critical topologies, including stable connected components and critical propagation paths, assign higher message passing priority;

[0051] An adaptive message aggregation method is adopted, which dynamically adjusts the message fusion ratio according to the importance of nodes in the topology.

[0052] Preferably, the process of performing persistent same-origin information flow reverse tracing includes:

[0053] Analyze the persistence profile of terminal infection distribution and identify key points including birth and death points;

[0054] Identify key propagation nodes in the network;

[0055] Tracing back along key transmission paths from the terminal infection node, a possible transmission path tree is constructed;

[0056] The probability of potential source points is assessed based on topological similarity.

[0057] Preferably, the generation and evaluation of multiple propagation source hypotheses includes:

[0058] Maintain multiple potential source point hypotheses and their combinations;

[0059] Calculate the topological similarity, propagation timing consistency, and rationality of dynamic parameters for each hypothesis;

[0060] The hypotheses are ranked based on the overall score;

[0061] Propose the optimal number and combination of source points.

[0062] Infectious disease tracing systems based on spatiotemporal graph neural networks include:

[0063] The data acquisition and preprocessing module is used to acquire multi-source heterogeneous contact data, which includes contact objects, contact time, contact location, and contact attributes.

[0064] The spatiotemporal graph construction and enhancement module is used to construct a topology-enhanced spatiotemporal graph based on the multi-source heterogeneous contact data, including converting the multi-source heterogeneous contact data into a spatiotemporal point cloud, constructing a multi-scale simple complex sequence, calculating the persistent map of the multi-scale simple complex sequence, extracting persistent topological features and fusing them into the spatiotemporal graph representation;

[0065] The spatiotemporal graph neural network module is used to train multi-scale spatiotemporal graph neural networks, including constructing multi-level temporal encoding, calculating dual attention weights based on physical spatial distance and network topology distance, and designing a message passing mechanism guided by topology features.

[0066] The multi-source tracing reasoning module is used to perform topology-guided backpropagation tracing based on the multi-scale spatiotemporal graph neural network, including performing persistent homologous information flow reverse tracing on the final state infection distribution, learning the initial state graph structure features in reverse through spectral graph convolution, and generating and evaluating multiple propagation source hypotheses.

[0067] The visualization and decision support module is used to output potential sources of infection, their probability distribution, and transmission routes, and to provide suggestions for prevention and control decisions.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] 1. Significantly improved source tracing accuracy: On real spatiotemporal network datasets, Top-1 accuracy is improved by about 30%, Top-5 accuracy is improved by about 15%, and the recognition rate reaches 65% in multi-source scenarios.

[0070] 2. Reduced data requirements: Maintains high accuracy even with 50% network structure missing, significantly lowering the data collection threshold.

[0071] 3. Improved computational efficiency: Large-scale network processing time is reduced by 75%, supporting real-time analysis.

[0072] 4. Enhanced early warning: It enables effective source tracing in the early stages of an epidemic (when the number of infected people accounts for 0.5% of the total population), providing a basis for early intervention.

[0073] 5. Enhanced source tracing interpretability: Provides intuitive and interpretable source tracing results through topological structure analysis, enhancing decision support capabilities. Attached Figure Description

[0074] Figure 1 A flowchart of an infectious disease tracing method based on a spatiotemporal graph neural network provided in an embodiment of the present invention;

[0075] Figure 2 This is a flowchart of the spatiotemporal graph representation driven by topology persistence in an embodiment of the present invention;

[0076] Figure 3This is a flowchart of the topology-guided backpropagation tracing mechanism in an embodiment of the present invention. Detailed Implementation

[0077] Please refer to Figures 1-3 The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0078] Reference Figure 1 This invention provides a method for tracing the source of infectious diseases based on a spatiotemporal graph neural network, comprising the following steps:

[0079] In one embodiment of the present invention, acquiring multi-source heterogeneous contact data includes acquiring contact object, contact time, contact location, and contact attributes. This multi-source heterogeneous data may include movement trajectory data, public transportation records, location check-in data, and medical system data, etc.

[0080] Preferably, acquiring multi-source heterogeneous contact data includes: obtaining raw contact information from mobile trajectory data, public transportation records, location check-in data, and medical system data; performing spatiotemporal alignment on the raw contact information to unify the temporal granularity and spatial reference system; evaluating the quality of the raw contact information and performing missing data repair; and generating a standardized spatiotemporal contact dataset.

[0081] In practice, spatiotemporal alignment is achieved by converting timestamps from different data sources into a unified time format (such as UTC time) and spatial location information into a unified coordinate system (such as the WGS84 coordinate system). Missing data repair employs interpolation methods based on spatiotemporal correlation. For example, in high temporal resolution data, linear interpolation can be used to supplement the location information at intermediate times; in high spatial density regions, nearest neighbor interpolation can be used to supplement missing location information.

[0082] Construct a topology-enhanced spatiotemporal graph, referring to Figure 2 In one embodiment of the present invention, constructing a topology-enhanced spatiotemporal graph based on the multi-source heterogeneous contact data includes the following steps:

[0083] Create spatiotemporal coordinate points for each contact record (i,j,t,d), where i and j are the contacting individuals, t is the time, and d is the contact duration; attach attributes such as contact intensity and duration to each spatiotemporal coordinate point; and construct a point cloud data structure that supports efficient spatiotemporal queries.

[0084] In practical applications, contact strength can be calculated based on a combination of contact distance and duration. For example, for face-to-face contact, the following formula can be used to calculate contact strength:

[0085] ,

[0086] in: Contact strength, in dimensionless normalized values; The physical distance between individuals i and j is expressed in meters. This is a smoothing parameter, in meters, usually taken as 0.1 meters, to prevent division overflow when the distance is zero; The duration of contact is in minutes; This is a normalization factor, expressed in meters per minute (m / min), used to convert contact intensity into a dimensionless value. It can range from 0.5 to 2.0 m / min, adjusted according to the transmission characteristics of infectious diseases. For respiratory infectious diseases, a value of 1.5 is recommended; for contact-transmitted diseases, a value of 0.8 is recommended.

[0087] Set a spatial distance threshold εs and a time interval threshold εt to form a parameterized spatiotemporal neighborhood; construct a nested simple complex sequence by systematically changing the spatial distance threshold εs and the time interval threshold εt.

[0088] Preferably, the spatial distance threshold εs and the time interval threshold εt are adaptively adjusted according to the characteristics of the infectious disease and the data distribution. For example, for diseases transmitted through droplets, the spatial distance threshold εs can be set to 1-2 meters; for diseases transmitted through contact, the spatial distance threshold εs can be set to within 0.5 meters. The time interval threshold εt can be set according to the incubation period of the disease.

[0089] During the construction process, we can define the spatiotemporal distance function as follows:

[0090] ,

[0091] in: It represents spatiotemporal distance, and the unit depends on the combination of spatial and temporal units. and For two spacetime points, and This is a spatial coordinate vector, with the unit being meters. and This is a timestamp, in hours. Represents spatial Euclidean distance, in meters; Represents the absolute value of a time interval, in hours; and This is a dimensionless weighting parameter used to balance the importance of spatial and temporal distances. For rapidly spreading diseases, it can be increased. Value; for diseases with long incubation periods, it increases. Value. In practical applications, and It can be set according to the characteristics of the epidemic.

[0092] Calculate the 0-persistence graph to characterize the appearance and disappearance of connected components; calculate the 1-persistence graph to characterize the formation and destruction of ring structures; calculate the 2-persistence graph to characterize the generation and disappearance of cavity structures; extract the statistical features of persistent barcodes, including average lifetime, maximum persistence, and entropy value.

[0093] In this invention, the persistence map is obtained by calculating the homology group changes of nested simplex sequences. For each dimension k, the persistence map records the "birth" and "death" times of k-dimensional topological features (such as connected components, loops, cavities). This information can be represented as persistence pairs ( , The set of ), where The scale parameter indicating the occurrence of features. The scale parameter that indicates feature vanishing.

[0094] The statistics for persistence characteristics are calculated as follows:

[0095] ,

[0096] ,

[0097] ,

[0098] in: The average lifespan is expressed in units and scale parameters. For maximum durability, the unit and scale parameters are the same; Entropy value of persistent distribution, in bits; The number of persistent pairs, unitless; and These are the birth and death scale parameters for the i-th persistence pair, respectively, with units depending on the scale used; Let be the normalized lifetime of the i-th persistence pair, which is dimensionless; This represents a logarithmic function to the base 2. These statistics reflect the stability and complexity of the network structure. A larger value indicates more stable topological features. The larger the value, the more complex and diverse the topology.

[0099] By mapping the statistical features of persistent barcodes to feature vectors for graph nodes and edges, the topological awareness of graph representation is enhanced. Specifically, for each node v in the graph, we can extract the following topological features:

[0100] ,

[0101] in: Let v be the topological feature vector of node v, with dimensions 9×1; , , These represent the average lifetime, maximum persistence, and entropy of the k-dimensional topological features of node v, respectively, where k takes values ​​of 0, 1, and 2, corresponding to connected components, ring structures, and cavity structures, respectively. These features are then concatenated with the node's original features (such as degree and centrality) to form an enhanced feature vector. In practical applications, this topological feature can capture the importance of nodes in the network structure, rather than being limited to local connectivity patterns.

[0102] In an embodiment of the present invention, training a multi-scale spatiotemporal graph neural network includes the following steps:

[0103] Generate absolute time location codes to represent time markers such as time and date within a day; generate periodic pattern codes to capture periodic behavioral patterns such as within a day and week; generate relative time interval codes to describe the distribution characteristics of event intervals; and fuse the absolute time location codes, periodic pattern codes, and relative time interval codes to form a multidimensional time feature vector.

[0104] In this embodiment, the absolute time position encoding is implemented using sine-cosine encoding:

[0105] ,

[0106] ,

[0107] in: and , where are the encoding values ​​of position pos in the 2i and 2i+1 dimensions, respectively, dimensionless, ranging from [-1, 1]; pos is the position index of the time point, unitless; i is the index of the feature dimension, ranging from [0, ]. / 2-1]; This is the total dimension of the encoding, usually an even number, such as 128 or 256. For a time of day, pos can be an hour value from 0 to 23; for a date, pos can be the number of days since the reference date. This encoding method captures the absolute location information of time while maintaining the relative relationship between different points in time.

[0108] Periodic pattern coding extracts the periodic features of the time series through Fourier transform, while relative time interval coding is obtained by modeling the statistical distribution of event intervals. Finally, these codes are concatenated into a multidimensional vector. :

[0109] ,

[0110] in: The final time-encoded vector has dimensions ( + + )×1; Encoding for absolute time location, with dimensions of ×1; Encoding for periodic patterns, with dimensions of ×1; Encoded for relative time intervals, with dimensions of ×1; the semicolon indicates a vector concatenation operation. In practical applications, it can be set... =128, =64, =32, resulting in a 224-dimensional time encoding vector.

[0111] In one embodiment of the present invention, the dual attention weights are jointly determined by the physical spatial distance matrix and the network topology distance matrix. For nodes i and j, the attention weights are calculated as follows:

[0112] ,

[0113] in: Let be the attention weight of node i to node j, which is dimensionless and ranges from [0,1]. and These are the query vector and key vector for nodes i and j, respectively, obtained by linear transformation of node features; It is the vector dimension, usually 64 or 128; It refers to physical spatial distance, measured in meters. It is the network topological distance, i.e., the shortest path length, and has no unit. and These are the attenuation coefficients, with units of [missing information]. And dimensionless, controlling the degree of influence of distance on attention; softmax is the softmax function used to normalize the attention score; express The transpose of . Physical spatial distance reflects the physical possibility of propagation, while network topological distance reflects connectivity in the network structure. In practical applications, and It can be set according to the characteristics of disease transmission. For example, for diseases transmitted through the air, it can be set... For diseases that require direct contact for transmission, settings can be configured. .

[0114] Message importance weights are assigned based on persistent topological features; key topological structures, including stable connected components and key propagation paths, are given higher message delivery priority; and an adaptive message aggregation method is adopted to dynamically adjust the message fusion ratio according to the importance of nodes in the topological structure.

[0115] In the message passing process of a graph neural network, the representation update of node v at layer l can be represented as:

[0116] ,

[0117] in: This represents the feature representation of node v at layer l, with dimension 1. ; This represents the feature representation of node v at layer l+1, with the same dimension. ; Represents the set of neighbors of node v; It is the attention weight of node v to its neighbor u, which is dimensionless and ranges from [0,1]. and It is a learnable weight matrix; It is a non-linear activation function (such as ReLU). The key is... The function assigns message importance weights based on the topological characteristics of node u:

[0118] ,

[0119] in: is the message importance weight based on topological features, dimensionless, and its value range is [0,1]. It is a learnable parameter vector used to evaluate the importance of different topological features; It is the topological feature vector of node u; The function is a natural exponential function. This design prioritizes message passing on key topological structures (such as stable connected components and important propagation paths), enhancing the model's awareness of network structure. In practical applications, the optimal function can be automatically learned through multiple rounds of training. The values ​​can also be initialized based on prior knowledge, for example, assigning higher initial weights to features that maintain durability at 0 (such as [0.5, 0.3, 0.2, 0.3, 0.2, 0.1, 0.2, 0.1, 0.1]).

[0120] Reference Figure 3 In an embodiment of the present invention, performing topology-guided backpropagation tracing based on the multi-scale spatiotemporal graph neural network includes the following steps:

[0121] Analyze the persistence map of the terminal infection distribution to identify key points, including birth points and death points; determine the key propagation nodes in the network; trace backward from the terminal infection nodes along the key propagation paths to construct a possible propagation path tree; and assess the probability of potential source points based on topological similarity.

[0122] In the specific implementation, key propagation nodes are identified through key points in the persistence graph. For a 0-persistence persistence graph, the birth point corresponds to a possible propagation source point, and the death point corresponds to the merging of connected components, usually the intersection of propagation paths. Based on these key points, starting from the final infected node, the possible propagation path tree is constructed by tracing backward along the key propagation paths.

[0123] Topological similarity assessment is achieved by comparing persistent graphs under different propagation source assumptions. For a candidate source point *s*, its similarity score can be calculated as:

[0124] ,

[0125] in: is the topological similarity score of candidate source point s, which is dimensionless and ranges from (0,1]. The larger the value, the more likely it is to be a true source point. and These are the observed final state distribution and the persistence maps of the final state distribution simulated with s as the source point, respectively. It is the Wasserstein distance between two persistent maps, without units; This is a scaling factor, dimensionless, typically ranging from 1 to 5. Higher values ​​indicate stricter similarity requirements. In practical applications, This can be adjusted based on data quality and noise levels; for example, for high-quality data, it can be adjusted accordingly. For data with high noise, it is advisable to take The Wasserstein distance measures the difference between two probability distributions and is suitable for comparing differences in persistence maps; the smaller the value, the more similar the two distributions are.

[0126] Spectral graph convolution inverse learning utilizes graph signal processing theory to map from the final state to the initial state through frequency domain transformation. In this embodiment, the graph signal is transformed to the frequency domain through eigenvalue decomposition of the graph's Laplacian matrix, and then a learnable inverse filter is applied to reconstruct the initial state signal:

[0127] ,

[0128] in: This is the initial state graph signal, where n is the number of nodes and c is the feature dimension; It is a final state graph signal (such as node infection status); It is the eigenvector matrix of the normalized Laplacian matrix, with each column being an eigenvector; It is an eigenvalue diagonal matrix, where the elements on the diagonal are the eigenvalues ​​of the Laplace matrix; It is determined by parameters Defined frequency domain inverse filter function; This represents the transpose of matrix U. This function can be designed in polynomial form:

[0129] ,

[0130] in: This is the inverse filter function in the frequency domain, and it is dimensionless. These are the eigenvalues ​​of the Laplace matrix; It is the polynomial order, which is unitless and usually takes the value of 3-5. These are learnable parameters, dimensionless. In practical applications, one can start with simple low-order polynomials, such as... The initial parameters can be set to , , , Then, these parameters are optimized using training data. By optimizing these parameters using training data, the model can learn the optimal mapping from the final state to the initial state.

[0131] Maintain multiple potential source point hypotheses and their combinations; calculate the topological similarity, propagation timing consistency, and dynamic parameter rationality of each hypothesis; rank the hypotheses based on a comprehensive score; and generate suggestions for the optimal number and combination of source points.

[0132] In multi-source hypothesis evaluation, for the hypothesis set Each of the hypotheses Given a set of potential source points, the comprehensive scoring function can be designed as follows:

[0133] ,

[0134] in: Assumption The overall score is dimensionless, and its range depends on the normalization method of each score. , and They are respectively assumptions The topological similarity, propagation timing consistency, and dynamic parameter rationality scores are all dimensionless quantities, and their values ​​are usually in the range of [0,1]. , and These are weighting coefficients, dimensionless, used to balance the importance of different factors, and satisfy the following conditions: In practical applications, these weights can be adjusted based on specific disease characteristics and data quality. For example, for cases with high data quality, weights can be set to... , , For diseases with clearly defined transmission mechanisms, a system can be set up. , , .

[0135] Finally, the system outputs potential sources of infection and their probability distribution, along with the most likely transmission path. The probability distribution can be obtained by normalizing the scores of each candidate source:

[0136] ,

[0137] in Indicates candidate source points The probability that a given value is a true source of infection is dimensionless, takes values ​​in the range [0,1], and satisfies the following condition: It is the i-th candidate source point; It is the total number of candidate source points; It is the source point The system provides a comprehensive score. It also outputs possible transmission paths from the source to each infected node, forming a complete transmission tree, providing an intuitive reference for prevention and control decisions.

[0138] This invention also provides an infectious disease tracing system based on spatiotemporal graph neural networks, comprising the following modules:

[0139] The data acquisition and preprocessing module is used to acquire multi-source heterogeneous contact data, including contact objects, contact time, contact location, and contact attributes. Specific functions include multi-source data interface adaptation, spatiotemporal alignment, quality assessment, and repair.

[0140] The multi-source data interface adapter supports data acquisition from various data sources, including mobile trajectory APIs, transportation card systems, location check-in platforms, and medical information systems. The spatiotemporal alignment engine unifies data from different sources into a standard spatiotemporal reference frame, while the quality assessment and repair component is responsible for detecting and repairing issues such as missing data and outliers.

[0141] The spatiotemporal graph construction and enhancement module is used to construct a topology-enhanced spatiotemporal graph based on the multi-source heterogeneous contact data. This includes converting the multi-source heterogeneous contact data into a spatiotemporal point cloud, constructing a multi-scale simple complex sequence, calculating the persistent map of the multi-scale simple complex sequence, extracting persistent topological features, and fusing them into the spatiotemporal graph representation.

[0142] This module first converts standardized contact data into a spatiotemporal point cloud representation, and then constructs a multi-layered spatiotemporal graph structure, including a physical contact layer, a region co-occurrence layer, and a social relationship layer. Subsequently, it extracts topological features by computing persistent graphs and integrates these features into the graph representation to enhance the model's ability to perceive network structure.

[0143] The Spatiotemporal Graph Neural Network module is used to train multi-scale spatiotemporal graph neural networks. This includes constructing multi-level temporal encoding, calculating dual attention weights based on physical spatial distance and network topological distance, and designing a message passing mechanism guided by topological features.

[0144] This module employs a multi-scale design, enabling it to simultaneously capture microscopic contact patterns and macroscopic propagation trends. Multi-level temporal encoding captures different aspects of temporal information, a dual attention mechanism balances the influence of physical space and network topology, and topology-guided message passing enhances the model's perception of network structure.

[0145] The multi-source tracing reasoning module is used to perform topology-guided backpropagation tracing based on the multi-scale spatiotemporal graph neural network, including performing persistent homologous information flow reverse tracing on the final state infection distribution, learning the initial state graph structure features through spectral graph convolution, and generating and evaluating multiple propagation source hypotheses.

[0146] This module first identifies key propagation nodes by analyzing the topological characteristics of the final state distribution through persistent homology tracing. Then, it uses spectral graph convolution to inversely learn the mapping relationship from the final state to the initial state, inferring possible initial states. Finally, by comprehensively evaluating multiple propagation source hypotheses, it determines the most likely source of infection and its combination.

[0147] The visualization and decision support module outputs potential sources of infection, their probability distribution, and transmission routes, and provides suggestions for prevention and control decisions.

[0148] This module visualizes the source tracing results in an intuitive way, including the location of the source, probability distribution, and transmission path. Simultaneously, it provides targeted prevention and control recommendations based on the source tracing results, such as key monitoring areas, priority vaccination populations, and precise isolation measures, providing a scientific basis for epidemic prevention and control decisions.

[0149] This invention has been comprehensively evaluated on real-world datasets. The experiments used five real-world spatiotemporal network datasets, including Sexualcontact, Hospital, Europeananemail, Bitcoin, and Message.

[0150] Regarding source tracing accuracy, this invention was compared with existing methods, including Jordan centrality, modified betweenness centrality, dynamic age, and soft-boundary Monte Carlo estimators. Experimental results show that this invention significantly outperforms existing methods in all three evaluation metrics: Top-1, Top-5, and Hop-1. For example, on sexual contact networks, the invention achieves a Top-1 accuracy of 50% and a Top-5 accuracy of 92%, while the best baseline methods only achieve 12% and 28%, respectively.

[0151] To verify the robustness of the system under data missing conditions, we designed two missing scenarios: missing contact network structure and missing final state observations. The results show that even with 50% missing network structure, the present invention still maintains approximately 70% Top-5 accuracy; and with 80% missing final state observations, it still achieves approximately 60% Top-5 accuracy, significantly outperforming existing methods.

[0152] Furthermore, this invention reveals the characteristics and patterns of propagation sources through topological feature analysis. We found that nodes assigned high probabilities by the model are closer to the real propagation sources in terms of activity sequence similarity and core number, while this relationship is not obvious in terms of node degree. This indicates that this invention can capture deeper topological features, and is not limited to simple connectivity metrics.

[0153] This invention can be applied to various infectious disease prevention and control scenarios, such as rapid source tracing of outbreaks, accurate identification of key populations, and early warning of transmission risks. Furthermore, the technical framework of this invention can also be extended to areas such as information dissemination tracking in social networks, attack source identification in cybersecurity, and invasive species tracing in ecosystems.

[0154] In future expansions, this invention can be further enhanced in the following ways: First, by fusing genomic data for cross-scale source tracing and combining it with viral mutation information to improve the accuracy of source tracing; second, by integrating more heterogeneous data sources, such as meteorological data and population flow data, to enrich the modeling of the transmission environment; and third, by designing an incremental learning mechanism to support the continuous updating and optimization of the source tracing model as new data arrives, adapting to the dynamically changing transmission environment.

[0155] This invention proposes a method and system for tracing the source of infectious diseases based on spatiotemporal graph neural networks. By integrating topological persistence theory with spatiotemporal graph neural networks, it achieves high-precision identification and tracking of the source of infectious disease transmission. Core innovations include topological persistence-driven spatiotemporal graph representation, multi-scale spatiotemporal graph neural networks, and a topology-guided backpropagation tracing mechanism. These innovations significantly outperform existing technologies in terms of tracing accuracy, data requirements, computational efficiency, and result interpretability.

[0156] Experimental results demonstrate that this invention exhibits excellent performance on various real-world network datasets, maintaining high accuracy even in challenging scenarios with missing data, thus providing reliable technical support for practical epidemic prevention and control. By accurately identifying the source of transmission and reconstructing the transmission path, this invention can help health departments implement more precise prevention and control measures, minimizing the impact of epidemic spread and providing strong technical protection for public health security.

[0157] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for tracing the source of infectious diseases based on spatiotemporal graph neural networks, characterized in that, include: Acquire multi-source heterogeneous contact data, which includes contact object, contact time, contact location, and contact attributes; Based on the aforementioned multi-source heterogeneous contact data, a topology-enhanced spatiotemporal graph is constructed, including: The multi-source heterogeneous contact data is converted into a spatiotemporal point cloud; Based on the spatiotemporal point cloud, a multi-scale simple complex sequence is constructed; Calculate the persistence map of the multi-scale simplex sequence; Persistent topological features are extracted and fused into the spatiotemporal graph representation; Training a multi-scale spatiotemporal graph neural network includes: A multi-level time coding system is constructed to capture the absolute location, periodicity, and relative interval information of time. The construction of the multi-level time coding system includes: generating an absolute time location code to represent intraday time and date / time markers; generating a periodic pattern code to capture intraday and intraweek periodic behavior patterns; generating a relative time interval code to describe the distribution characteristics of event intervals; and fusing the absolute time location code, periodic pattern code, and relative time interval code to form a multi-dimensional time feature vector. Calculate dual attention weights based on physical spatial distance and network topology distance; The design incorporates a topology-feature-guided message passing mechanism, including: assigning message importance weights based on persistent topology features; assigning higher message passing priority to key topology structures, including stable connected components and key propagation paths; and employing an adaptive message aggregation method to dynamically adjust the message fusion ratio according to the importance of nodes in the topology structure. Based on the aforementioned multi-scale spatiotemporal graph neural network, topology-guided backpropagation tracing is performed, including: Perform persistent reverse tracing of the same source information flow on the final state infection distribution, including: analyzing the persistent map of the final state infection distribution and identifying key points including birth points and death points; determining key propagation nodes in the network; tracing back from the final state infection nodes along key propagation paths to construct a possible propagation path tree; and assessing the probability of potential source points based on topological similarity. The initial state graph structure features are learned by inverse learning through spectral convolution. Specifically, graph signal processing theory is used to transform the graph signal into the frequency domain through the Laplacian matrix eigenvalue decomposition of the graph, and the initial state signal is reconstructed by applying a learnable inverse filter. Generate and evaluate multiple propagation source hypotheses, including: maintaining multiple potential source hypotheses and their combinations; calculating the topological similarity, propagation timing consistency, and dynamic parameter rationality of each hypothesis; ranking the hypotheses based on a comprehensive score; and generating suggestions for the optimal number and combination of sources. Output potential sources of infection, their probability distribution, and transmission routes.

2. The method according to claim 1, characterized in that, The acquisition of multi-source heterogeneous contact data includes: Raw contact information is obtained from movement trajectory data, public transportation records, location check-in data, and medical system data; The original contact information is spatiotemporally aligned to unify the temporal granularity and spatial reference system; Assess the quality of the original contact information and perform missing data repair; Generate a standardized spatiotemporal contact dataset.

3. The method according to claim 1, characterized in that, The process of converting multi-source heterogeneous contact data into spatiotemporal point clouds includes: Create spatiotemporal coordinates for each contact record (i,j,t,d), where i and j are the contacting individuals, t is the time, and d is the contact duration; Add contact intensity and duration attributes to each spatiotemporal coordinate point; Construct a point cloud data structure that supports efficient spatiotemporal queries.

4. The method according to claim 1, characterized in that, The construction of the multi-scale simple complex sequence includes: Set spatial distance threshold εs and time interval threshold εt to form a parameterized spatiotemporal neighborhood; By systematically varying the spatial distance threshold εs and the time interval threshold εt, nested simple complex sequences are constructed. The spatial distance threshold εs and time interval threshold εt are adaptively adjusted based on the characteristics of infectious diseases and data distribution.

5. The method according to claim 1, characterized in that, The calculation of the persistence map of the multi-scale simplex sequence includes: Calculate the 0-persistence graph to characterize the appearance and disappearance of connected components; Calculate the durability spectrum of ring structure 1 to characterize the formation and destruction of ring structures; Calculate the durability spectrum to characterize the formation and disappearance of cavity structures; Extract statistical features of persistent barcodes, including average lifetime, maximum persistence, and entropy value.

6. An infectious disease tracing system based on spatiotemporal graph neural networks, used to implement the method described in any one of claims 1-5, characterized in that, include: The data acquisition and preprocessing module is used to acquire multi-source heterogeneous contact data, which includes contact objects, contact time, contact location, and contact attributes. The spatiotemporal graph construction and enhancement module is used to construct a topology-enhanced spatiotemporal graph based on the multi-source heterogeneous contact data, including converting the multi-source heterogeneous contact data into a spatiotemporal point cloud, constructing a multi-scale simple complex sequence, calculating the persistent map of the multi-scale simple complex sequence, extracting persistent topological features and fusing them into the spatiotemporal graph representation; The spatiotemporal graph neural network module is used to train a multi-scale spatiotemporal graph neural network. It includes constructing a multi-level temporal encoding, which involves generating absolute time position encoding, generating periodic pattern encoding, and generating relative time interval encoding, and fusing the three to form a multi-dimensional temporal feature vector. It calculates dual attention weights based on physical spatial distance and network topology distance, and designs a topology feature-guided message passing mechanism, including assigning message importance weights based on persistent topology features, giving higher message passing priority to key topology structures, and dynamically adjusting the message fusion ratio using an adaptive message aggregation method. The multi-source tracing reasoning module is used to perform topology-guided backpropagation tracing based on the multi-scale spatiotemporal graph neural network. This includes performing persistent homologous information flow reverse tracing on the final-state infection distribution, which includes analyzing the persistent graph of the final-state infection distribution and identifying birth and death points, determining key propagation nodes, tracing back from the final-state infection nodes along key propagation paths to construct a propagation path tree, evaluating the probability of potential source points based on topological similarity, and learning the initial-state graph structure features through spectral graph convolution. Specifically, graph signal processing theory is used to convert the graph signal to the frequency domain through the Laplacian matrix eigenvalue decomposition of the graph and apply a learnable inverse filter to reconstruct the initial-state signal. Multiple propagation source point hypotheses are generated and evaluated, including calculating the topological similarity, propagation timing consistency, and dynamic parameter rationality of each hypothesis and ranking the hypotheses based on a comprehensive score. The visualization and decision support module is used to output potential sources of infection, their probability distribution, and transmission routes, and to provide suggestions for prevention and control decisions.

Images