CN-121565509-B - Infectious disease tracing method and system based on space-time diagram neural network

CN121565509BCN 121565509 BCN121565509 BCN 121565509BCN-121565509-B

Abstract

The invention relates to the technical field of public health safety, in particular to an infectious disease tracing method and system based on a space-time diagram neural network, which integrates multi-source heterogeneous contact data to construct a topology enhanced space-time diagram, constructs a multi-scale simplex complex by converting the data into space-time point cloud, extracts lasting topological characteristics, integrates the lasting topological characteristics into a diagram representation, and training a multiscale space-time diagram neural network, designing topology feature guided message transmission by using a multi-level time coding and dual-attention mechanism, finally, executing topology guided back propagation tracing, generating a plurality of propagation source point hypotheses by means of persistent homologous information flow back tracing and spectrogram convolution, and outputting probability distribution and propagation paths of potential infectious agents.

Inventors

YAN GANG
RU XIAOLEI
QIN JIAJIE

Assignees

同济大学

Dates

Publication Date: 20260508
Application Date: 20260122

Claims (6)

1. The infectious disease tracing method based on the space-time diagram neural network is characterized by comprising the following steps of: acquiring multi-source heterogeneous contact data, wherein the multi-source heterogeneous contact data comprises a contact object, contact time, a contact position and contact attribute; Based on the multi-source heterogeneous contact data, constructing a topology enhanced space-time diagram, comprising: converting the multi-source heterogeneous contact data into space-time point clouds; Constructing a multi-scale simplex complex sequence based on the space-time point cloud; Calculating a persistence map of the multi-scale simplex complex sequence; Extracting persistent topological features and fusing the persistent topological features into a space-time diagram representation; Training a multiscale space-time diagram neural network, comprising: Constructing a multi-level time code, capturing absolute position, periodicity and relative interval information of time, wherein the constructing the multi-level time code comprises the steps of generating an absolute time position code which represents time of day and time of day marks, generating a periodicity pattern code, capturing periodic behavior patterns of day and week, generating a relative time interval code which describes event interval distribution characteristics, and fusing the absolute time position code, the periodicity pattern code and the relative time interval code to form a multi-dimensional time feature vector; Calculating a dual attention weight based on the physical space distance and the network topology distance; the message transmission mechanism guided by the topological feature is designed, and comprises the steps of distributing message importance weights based on the durable topological feature, endowing higher message transmission priority to a key topological structure comprising stable connected components and key propagation paths, and dynamically adjusting message fusion proportion according to the importance of nodes in the topological structure by adopting a self-adaptive message aggregation mode; based on the multi-scale space-time diagram neural network, performing topology-guided back propagation tracing, including: Performing a persistent homologous information flow reverse trace on the final infection distribution, including analyzing a persistence map of the final infection distribution, identifying key points including birth points and death points, determining key propagation nodes in the network, reversely tracing from the final infection nodes along the key propagation paths, constructing a possible propagation path tree, evaluating the likelihood of potential source points based on topological similarity; reversely learning the structural characteristics of the initial state diagram through spectrogram convolution, wherein the image signal is converted into a frequency domain through the Laplace matrix characteristic decomposition of the image by utilizing the image signal processing theory, and a leachable inverse filter is used for reconstructing the initial state signal; Generating and evaluating a plurality of propagation source point hypotheses including maintaining a plurality of potential source point hypotheses and combinations thereof, calculating topological similarity, propagation timing consistency and kinetic parameter rationality for each hypothesis, ordering the hypotheses based on a composite score, generating suggestions of optimal source point numbers and combinations; The potential infectious agents and their probability distributions and propagation paths are output.
2. The method of claim 1, wherein the acquiring multi-source heterogeneous contact data comprises: Acquiring original contact information from movement track data, public transportation records, position sign-in data and medical system data; performing space-time alignment on the original contact information, and unifying time granularity and a space reference system; evaluating the quality of the original contact information and performing missing data repair; a normalized spatiotemporal contact dataset is generated.
3. The method of claim 1, wherein converting the multi-source heterogeneous contact data into a space-time point cloud comprises: creating a space-time coordinate point for each contact record (i, j, t, d), wherein i and j are contact individuals, t is time, and d is contact duration; attaching a contact intensity and duration attribute to each space-time coordinate point; And constructing a point cloud data structure supporting efficient space-time query.
4. The method of claim 1, wherein said constructing a multi-scale simplex complex sequence comprises: setting a space distance threshold epsilon s and a time interval threshold epsilon t to form a parameterized space-time neighborhood; Constructing a nested simplex complex sequence by systematically changing the spatial distance threshold epsilon s and the time interval threshold epsilon t; the spatial distance threshold epsilon s and the time interval threshold epsilon t are adaptively adjusted according to the characteristics of infectious diseases and the data distribution.
5. The method of claim 1, wherein the calculating a persistence map of the multi-scale simplex complex sequence comprises: Calculating a 0-dimensional persistence map, and representing the appearance and disappearance of connected components; calculating a 1-dimensional persistence map, and representing the formation and damage of a ring structure; Calculating a 2-dimensional persistence map, and representing generation and extinction of a cavity structure; Statistical features of the persistent barcode are extracted, including average lifecycle, maximum persistence, and entropy values.
6. An infectious disease traceability system based on a space-time diagram neural network, configured to implement the method of any of claims 1-5, comprising: The data acquisition and preprocessing module is used for acquiring multi-source heterogeneous contact data, wherein the multi-source heterogeneous contact data comprises a contact object, contact time, a contact position and contact attribute; The space-time diagram construction and enhancement module is used for constructing a topology enhanced space-time diagram based on the multi-source heterogeneous contact data, and comprises the steps of converting the multi-source heterogeneous contact data into space-time point clouds, constructing a multi-scale simplex complex sequence, calculating a persistence map of the multi-scale simplex complex sequence, extracting persistence topological features and fusing the persistence topological features into space-time diagram representation; The space-time diagram neural network module is used for training a multi-scale space-time diagram neural network and comprises a multi-level time code construction, wherein the multi-level time code construction comprises the steps of generating absolute time position codes, generating periodic mode codes and generating relative time interval codes, and fusing the three codes to form a multi-dimensional time feature vector, calculating dual attention weights based on physical space distances and network topology distances, designing a topology feature guided message transmission mechanism, and dynamically adjusting message fusion proportion by adopting an adaptive message aggregation mode, wherein the message importance weights are distributed based on the persistence topology features; The multi-source point tracing reasoning module is used for performing topology-guided back propagation tracing based on the multi-scale space-time diagram neural network and comprises performing persistence homologous information flow back tracing on final infection distribution, wherein the persistence tracing comprises the steps of analyzing persistence maps of final infection distribution and identifying birth points and death points, determining key propagation nodes, reversely tracing from the final infection nodes along key propagation paths to construct a propagation path tree, evaluating the possibility of potential source points based on topology similarity, reversely learning initial state diagram structural features through spectrogram convolution, converting a graph signal into a frequency domain through Laplace matrix feature decomposition of the graph by utilizing graph signal processing theory, applying a leachable back filter to reconstruct initial state signals, and generating and evaluating a plurality of propagation source point hypotheses, wherein the calculation of topology structure similarity, propagation time sequence consistency and dynamics parameter rationality of each hypothesis and the sequencing of the hypotheses based on comprehensive scores are included; and the visualization and decision support module is used for outputting potential infectious agents, probability distribution and propagation paths thereof and providing prevention and control decision suggestions.

Description

Infectious disease tracing method and system based on space-time diagram neural network Technical Field The invention relates to the technical field of public health safety, in particular to an infectious disease tracing method and system based on a space-time diagram neural network, which are used for rapidly and accurately identifying an infectious disease source through multi-source heterogeneous data after an infectious disease outbreak and providing accurate prevention and control decision support. Background In the prevention and control of infectious diseases, tracing is a key link for cutting off a transmission chain and blocking further spread of epidemic situations. Traditional tracing methods are mainly based on epidemiological investigation, manual tracing is performed through contactor tracing and case analysis, and the methods rely on expert experience, are long in time consumption and are difficult to deal with complex propagation scenes. With the development of computing technology, a tracing method based on network science and machine learning is gradually rising. The prior art mainly comprises centrality measurement method, propagation simulation method, reasoning algorithm and the like. The centrality measurement method calculates characteristic values according to the positions of nodes in a network, such as distance centrality, jordan centrality and the like, but the method ignores randomness and dynamic characteristics of a propagation process, a propagation simulation method such as a soft boundary Monte Carlo estimator evaluates the possibility of each node as a source point by simulating a plurality of propagation processes, but has poor effect when the contact network is dense or estimated parameters deviate from a real propagation process, and the existing deep learning method generally only considers a static network structure and lacks effective modeling on a time dimension. The existing tracing method has three main defects that firstly, high-quality complete contact network data are relied, the data are difficult to obtain completely in actual epidemic situations, secondly, the multi-source point explosion scene is supported to be limited, a plurality of independent propagation sources are difficult to identify, thirdly, deep mining of network topological structure characteristics is lacking, and key structure information in a propagation network is difficult to capture. Therefore, there is a need for an infectious disease tracing method that can accommodate incomplete data, support multi-source tracing, and make full use of network topology characteristics. Disclosure of Invention In view of the problems, the invention provides an infectious disease tracing method and system based on a space-time diagram neural network, which integrate topology persistence theory and the space-time diagram neural network to realize high-precision identification and tracking of infectious disease transmission sources. The method mainly solves the following technical problems of how to obtain reliable tracing results under the condition of incomplete data, how to effectively model complex structures and dynamic evolution characteristics in a space-time network, how to support accurate tracing of multi-source point propagation scenes, how to quantify uncertainty of tracing results and provide reliable decision support. The invention achieves the above-mentioned aim through three innovation points of progressive layer by layer, firstly, a topology durability driven space-time diagram representation method is provided, secondly, a multi-scale space-time diagram neural network architecture is designed, and thirdly, a topology guided back propagation tracing mechanism is achieved. The invention provides an infectious disease tracing method based on a space-time diagram neural network, which comprises the following steps: acquiring multi-source heterogeneous contact data, wherein the multi-source heterogeneous contact data comprises a contact object, contact time, a contact position and contact attribute; Based on the multi-source heterogeneous contact data, constructing a topology enhanced space-time diagram, comprising: converting the multi-source heterogeneous contact data into space-time point clouds; Constructing a multi-scale simplex complex sequence based on the space-time point cloud; Calculating a persistence map of the multi-scale simplex complex sequence; Extracting persistent topological features and fusing the persistent topological features into a space-time diagram representation; Training a multiscale space-time diagram neural network, comprising: constructing multi-level time codes, and capturing absolute position, periodicity and relative interval information of time; Calculating a dual attention weight based on the physical space distance and the network topology distance; Designing a topology feature guided message passing mechanism; based on the multi-scale space-time diagram neu