CN113517076A - Disease case number prediction method and system based on graph neural network and transfer learning - Google Patents
Disease case number prediction method and system based on graph neural network and transfer learning Download PDFInfo
- Publication number
- CN113517076A CN113517076A CN202110877916.5A CN202110877916A CN113517076A CN 113517076 A CN113517076 A CN 113517076A CN 202110877916 A CN202110877916 A CN 202110877916A CN 113517076 A CN113517076 A CN 113517076A
- Authority
- CN
- China
- Prior art keywords
- case number
- information aggregation
- node
- network
- number prediction
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/80—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for detecting, monitoring or modelling epidemics or pandemics, e.g. flu
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
- G16H10/60—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for patient-specific data, e.g. for electronic patient records
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Public Health (AREA)
- Medical Informatics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Primary Health Care (AREA)
- Epidemiology (AREA)
- Databases & Information Systems (AREA)
- Pathology (AREA)
- Biophysics (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The disease case number prediction method and system based on the graph neural network and the transfer learning, disclosed by the disclosure, are used for obtaining the case number of the historical time of each region of a country to be predicted and the moving number of each region in corresponding time; inputting the historical time case number of each region and the moving number of each region in corresponding time into the constructed digraph to obtain the characteristic vector of each node of the digraph; inputting the feature vector of each node into a trained case number prediction model to obtain a case number prediction result; the case number prediction model adopts a multi-layer information aggregation network, the output of the previous layer of information aggregation network is used as the input of the current information aggregation network, the input of the first layer of information aggregation network is the feature vector of each node of the directed graph, and the feature vector of each node of the directed graph and the output of each layer of information aggregation network are connected through a full connection layer to obtain a case number prediction result. The accuracy of case number prediction is improved.
Description
Technical Field
The invention relates to the technical field of disease case number, in particular to a disease case number prediction method and system based on a graph neural network and transfer learning.
Background
The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art.
The highly contagious and serious harmfulness of epidemic etiology has brought about a number of disasters. Under the condition of epidemic disease abuse, accurate prediction of the number of cases can help a decision maker to make a reasonable scheme, so that economic loss caused by epidemic disease is reduced, and physical health and safety of people are guaranteed.
In the early stages of the study, the number of cases of epidemic disease was predicted by using a mathematical model. Such methods are commonly used to study the way viruses are transmitted between individuals, including SIR models (susceptible-infection-convalescent model) and SEIR models (susceptible-latent-infection-convalescent model), among others. However, the preventive action taken by human beings to prevent the spread of disease has a certain influence on the above method, resulting in inaccurate prediction results. Thus, some research work has also incorporated factors such as preventative behavior into mathematical models, and the ways to incorporate preventative behavior can be divided into two categories: the first is to use preventative behavior as a parameter for disease models; the second is to introduce a new dynamic state, in order to distinguish the person taking preventive measures from the person not taking preventive measures and to study them separately, the mathematical model principle used above is to fit the whole trend of epidemic development by a set of differential equations, but these models consider too few or too many parameters, and are prone to under-fitting or over-fitting problems.
In recent years, with the rapid rise of deep learning, a method of predicting the number of cases based on time series has also come to be used. The autoregressive moving average model ARIMA is widely used in the prediction of the number of cases. The method requires that time sequence data is stable or data after difference is stable, and the input is the whole time sequence data before the prediction of the region to be predicted; the input of the PROPHET method recently used by Mahmud et al is similar to ARIMA, and the method can process the condition that some abnormal values exist in the time series and can also process the condition that partial values are missing, thereby effectively predicting the future trend of the time series; meanwhile, the recurrent neural network is also applied to the field of case number prediction, and the method generally establishes a separate model for each region (for example, a model is established for each city), each model updates characteristics by combining the previous case information, and Chimmula and the like predict the number of new coronary pneumonia diagnoses in various regions of Canada by using a long-short term memory network LSTM.
However, in the case number prediction of these time series methods, geographical neighborhood and migration with nearby areas are not considered, and no model is added for interaction between areas, and because epidemic diseases are mainly propagated by contact between people, the case number prediction is inaccurate when the mobility of people is not considered.
Disclosure of Invention
In order to solve the above problems, the present disclosure provides a disease case number prediction method and system based on a graph neural network and transfer learning, and when predicting the disease number, the number of historical cases and the mobility of people between different regions are fully considered, so that the accuracy of predicting the disease number is ensured.
In order to achieve the purpose, the following technical scheme is adopted in the disclosure:
in a first aspect, a disease case number prediction method based on a graph neural network and transfer learning is provided, which includes:
acquiring the historical time case number of each region of a country to be predicted and the moving number of each region within corresponding time;
inputting the historical time case number of each region and the moving number of each region in corresponding time into the constructed digraph to obtain the characteristic vector of each node of the digraph;
inputting the feature vector of each node into a trained case number prediction model to obtain a case number prediction result;
the case number prediction model adopts a multi-layer information aggregation network, the output of the previous layer of information aggregation network is used as the input of the current information aggregation network, the input of the first layer of information aggregation network is the feature vector of each node of the directed graph, and the feature vector of each node of the directed graph and the output of each layer of information aggregation network are connected through a full connection layer to obtain a case number prediction result.
In a second aspect, a disease case number prediction system based on graph neural network and transfer learning is provided, which includes:
the data acquisition module is used for acquiring the historical time case number of each region of the country to be predicted and the moving number of people of each region in corresponding time;
the characteristic vector acquisition module of the node is used for inputting the historical time case number of each region and the moving number of each region in corresponding time into the constructed digraph to obtain the characteristic vector of each node of the digraph;
the case number prediction module is used for inputting the feature vectors of all the nodes into a trained case number prediction model to obtain a case number prediction result;
the case number prediction model adopts a multi-layer information aggregation network, the output of the previous layer of information aggregation network is used as the input of the current information aggregation network, the input of the first layer of information aggregation network is the feature vector of each node of the directed graph, and the feature vector of each node of the directed graph and the output of each layer of information aggregation network are connected through a full connection layer to obtain a case number prediction result.
In a third aspect, an electronic device is provided, which includes a memory, a processor, and computer instructions stored in the memory and executed on the processor, where the computer instructions, when executed by the processor, perform the steps of the disease case number prediction method based on graph neural network and transfer learning.
In a fourth aspect, a computer-readable storage medium is provided for storing computer instructions, which when executed by a processor, perform the steps of the disease case number prediction method based on graph neural network and transfer learning.
Compared with the prior art, the beneficial effect of this disclosure is:
1. according to the method, when the number of disease cases is predicted, the number of historical cases in a region is considered, and the mobility of people in each region is also considered fully, so that the prediction result of the number of cases is more accurate.
2. In case number prediction, the present disclosure uses not only case data of the previous day but also case data of a plurality of days, thereby further improving the accuracy of case number prediction.
3. According to the method, the neighborhood information of each node is combined through the information aggregation network, the output characteristics of each node are obtained, the influence degree of the neighborhood around a central node is not always the same, so that in the process of combining the neighborhood information, the information aggregation network uses a graph convolution network and a graph attention machine mechanism, the difference of the importance of each neighborhood and the central node is evaluated through the graph attention machine mechanism, the combination of the neighborhood information is carried out according to the difference, the output characteristics of each node are more accurate, and the accuracy of case number prediction is further guaranteed.
4. The case number prediction model provided by the disclosure comprises a plurality of layers of information aggregation networks, the characteristics of each node are updated through the plurality of information aggregation networks in combination with the information of time layers, and with the increase of the times of the time layers, the final node characteristics can obtain more and more global information, so that the accuracy of case number prediction is improved.
5. According to the method, when the case number prediction model is trained, a transfer learning method is introduced, so that the case number prediction model of the country in the early stage of the disease can utilize the trained parameters of the country in the relatively stable stage of the disease, and the accuracy of case number prediction of the country in the early stage of the disease is ensured.
Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this application, illustrate embodiments of the application and, together with the description, serve to explain the application and are not intended to limit the application.
Fig. 1 is a flow chart of case number prediction when two IAN frameworks are used, as disclosed in example 1 of the present disclosure;
fig. 2 is a structural diagram of an IAN framework disclosed in embodiment 1 of the present disclosure;
fig. 3 is a flowchart of model training by adding a transfer learning method disclosed in embodiment 1 of the present disclosure.
The specific implementation mode is as follows:
the present disclosure is further described with reference to the following drawings and examples.
It should be noted that the following detailed description is exemplary and is intended to provide further explanation of the disclosure. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.
It is noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of example embodiments according to the present application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, and it should be understood that when the terms "comprises" and/or "comprising" are used in this specification, they specify the presence of stated features, steps, operations, devices, components, and/or combinations thereof, unless the context clearly indicates otherwise.
Example 1
In this embodiment, a disease case number prediction method based on a graph neural network and transfer learning is disclosed, which includes:
acquiring the historical time case number of each region of a country to be predicted and the moving number of each region within corresponding time;
inputting the historical time case number of each region and the moving number of each region in corresponding time into the constructed digraph to obtain the characteristic vector of each node of the digraph;
inputting the feature vector of each node into a trained case number prediction model to obtain a case number prediction result;
the case number prediction model adopts a multi-layer information aggregation network, the output of the previous layer of information aggregation network is used as the input of the current information aggregation network, the input of the first layer of information aggregation network is the feature vector of each node of the directed graph, and the feature vector of each node of the directed graph and the output of each layer of information aggregation network are connected through a full connection layer to obtain a case number prediction result.
Furthermore, a series of directed graphs are created for each country according to different time, nodes of the directed graphs represent regions in the country, and edges of the directed graphs represent the number of people moving between the regions in the country.
Further, the information aggregation network adopts a graph convolution network, and a graph attention mechanism is added in the graph convolution network.
Further, the process of obtaining the output characteristics of the information aggregation network node is as follows:
the method comprises the steps of obtaining the depth characteristics of each node of the graph convolution network, inputting the depth characteristics of each node into a graph attention mechanism, obtaining attention weight coefficients among the nodes, and obtaining the output characteristics of each node of the information aggregation network according to the depth characteristics and the attention weight coefficients of the nodes.
Further, when calculating the attention weight coefficient, the attention mechanism of the adjacent node of a certain node is normalized by using a softmax function.
Furthermore, a transfer learning method is added into the information aggregation network, and the information aggregation network is trained by using disease data of different countries.
Further, the specific process of training the information aggregation network is as follows:
selecting the disease data of one country as a meta-test set, selecting the disease data of the rest countries as meta-training sets, inputting the disease data of the meta-training sets into a group of information aggregation networks for training, initializing the information aggregation networks of the meta-test sets by using the parameters of the trained information aggregation networks, inputting the disease data of the meta-test sets into the initialized information aggregation networks, and training the information aggregation networks of the meta-test sets.
The disease case number prediction method based on the graph neural network and the transfer learning disclosed in this embodiment will be described in detail.
The disease case number prediction method based on the graph neural network and the transfer learning comprises the following steps:
s1: and acquiring the historical time case number of each region of the country to be predicted and the moving number of each region in corresponding time.
Since the daily case data in each region of each country is closely related to not only the mobility between the current day and its neighborhood but also the case data before the region, in order to improve the accuracy of case number prediction, case number prediction is performed by using case data of a plurality of days in combination with case data of not only the previous day.
It should be noted that: the sources of the data acquired by the embodiment are legal and do not relate to privacy, and all the data are acquired according with laws and regulations.
S2: and inputting the historical time case number of each region and the moving number of each region in corresponding time into the constructed digraph to obtain the characteristic vector of each node of the digraph.
In specific implementation, a directed graph G ═ (V, E) is created for each country, and the situation within T days of one country is represented as a series of graphs G(1),…,G(t),…,G(T)The edges of the graph are created according to the mobility between regions, the mobility refers to the number of moving people between the regions in each country, and whether the edges exist between the nodes in the created directed graph depends on the number of moving people between the two regions.
V in G ═ V, E represents a node of the graph, and the node represents each region of the country. E represents the edge of the graph that captures the number of people moving between the two regions. Use ofRepresents the weight from vertex V to vertex U of the t-th day, and the weight represents the weight between two regionsThe number of moving persons in the room. For the accuracy of case number prediction, case data of a plurality of days is used in combination with case data of a day before prediction, instead of case data of a day before prediction, before case number prediction is performedThe vector represents the attribute of the node, and the attribute comprises confirmed cases of each day of past d days of the region u, and the characteristics of the node are jointly constructed by combining the daily case data of the region and the movement data between the regions. For a set time T, the characteristics of the nodes are shown in formula (1):
in the formula: a. the(t)Is a drawing G(t)The matrix elements are determined by the weight between two regionsAnd (4) forming. If the number of people moving between the two regions is 0, no edge is established, the element value is 0, X(t)Is a feature vector of the number of cases in different regions, F(t)A total feature vector is represented that combines the number of case cases in the region and the number of people moving between regions.
In summary, for a specific region u, vector product operation is performed on the case number feature of the neighborhood region and the moving person number feature between the specific region u and the neighborhood region, and the obtained result is the final feature vector of the region u. Note that, here, the movement inside the area u is also considered, and is expressed as equation (1)
S3: inputting the feature vector of each node into a trained case number prediction model to obtain a case number prediction result;
the case number prediction model adopts a multi-layer information aggregation network, the multi-layer information aggregation network is connected in an input-output relationship, the output of the upper layer information aggregation network is used as the input of the current information aggregation network, the input of the first layer information aggregation network is the feature vector of each node of the directed graph, and the feature vector of each node of the directed graph and the output of each layer information aggregation network are connected through a full connection layer to obtain a case number prediction result.
In specific implementation, the information aggregation network adopts a graph convolution network, a graph attention mechanism is added in the graph convolution network, when the graph attention mechanism is added, the depth characteristics of each node of the graph convolution network are obtained, the depth characteristics of each node are input into the graph attention mechanism, the attention weight coefficients among the nodes are obtained, and the output characteristics of each node are obtained according to the depth characteristics of the nodes and the attention weight coefficients.
The framework of the Information Aggregation Network (IAN) is shown in fig. 2, and includes a Graph Convolution Network (GCN) and a graph attention mechanism (GAT), and the information aggregation network performs neighborhood information aggregation on the features of each node input thereto, so as to obtain the output features of each node.
The Graph Convolution Network (GCN) is a popularization of a standard convolution network (CNN) on a graph domain, and can process generalized graph structure data through methods such as Fourier transform and the like, so that the feature information of nodes is deeply captured, and the depth feature information of the nodes is obtained.
The propagation mode between layers of the GCN is as follows:
in the formula: a is an adjacent matrix, and A is an adjacent matrix,is a matrix of the units,is composed ofDegree matrix of (H)(l)Is a feature matrix of the l-th layer, W(l)Is a trainable matrix at the l-th level, and σ is an activation function.
And inputting the feature vector of each node into the GCN to obtain the depth feature H of each node.
Considering that the influence degrees of surrounding neighborhoods of a central node are not always the same, in the process of combining neighborhood information, the difference of the importance of each node relative to the central node must be considered, so a graph attention force mechanism (GAT) more suitable for a directed graph is added into an information aggregation network, unequal weights can be made for each node in the neighborhood to represent the importance degree of each node to the central node, a feature vector of each node acquired through GCN is input into the GAT network, an attention weight coefficient among each node is acquired, and the output feature of each node is acquired according to the depth feature and the attention weight coefficient of the node.
The method specifically comprises the following steps:
first, a self-attention mechanism is implemented for each node, denoted by a, which has different parameters for different propagation rounds. The self-attention mechanism a is used for calculating the attention cross-correlation coefficient e between the node i and the node ji,jI.e. the degree of importance between node i and node j to each other.
ei,j=a(WHi,WHj) (3)
Wherein a is self-attitude mechanism, W is weight matrix, Hi、HjDepth features of nodes i, j, respectively.
Then, the self-attribute mechanism a is introduced into the structure of the graph, taking the node i as an example, and a softmax function is added to perform normalization processing on all the adjacent nodes j of i.
In the formula: alpha is alphai,jTo pay attention to the weight coefficient, NiAll neighborhood nodes of node i.
Combining the above formulas (3) and (4), the complete attention mechanism can be obtained by arrangement as follows
In the formula: and | is a join operation. LeakyRelu is a non-zero slope assigned to all negative values and is mathematically represented as:
the final attention weight coefficient a among different nodes is obtained through the operationi,jAnd thus the output characteristics of each node can be obtained.
The characteristics of each node are updated by combining a plurality of IAN frameworks and a plurality of time levels of information, and as the number of time levels increases, the final node characteristics can obtain more and more global information.
And inputting the output characteristics of the nodes obtained by carrying out neighborhood information aggregation on the characteristics of the nodes by the IAN frame of the previous layer into a formula (8) to obtain the output characteristics of the nodes output by the IAN of the current layer.
Ht+1=f(A~HtMt+1) (8)
In the formula: htFeature vector representing the last temporal level, where H0The input features X, f to start with are nonlinear activation functions, e.g. tanh functions, Relu functions, etc., Mt+1Is a trainable matrix at level t + 1.
For ease of calculation, the adjacency matrix a is changed so that the sum of the weights of all edges coming in from each node is equal to 1.
And connecting the characteristic X input into the first layer IAN with the output characteristic of the node output by each layer IAN by using a function (9), and obtaining and outputting the case number prediction result through a Dropout discarding layer, a Relu function and a full connection layer.
H=Concatenate(H0,…,Hn) (9)
Finally, the Loss of prediction is calculated using the mean square error function Loss.
In the formula: n represents the total number of nodes, yvThe real value of the number of cases of the region v is represented,representing the predicted value of the number of regional v cases.
The case number prediction model disclosed in this embodiment includes a plurality of connected IAN, and according to the information of the time hierarchy, the output characteristics of the nodes output by the first layer IAN are input into the second layer IAN, the output characteristics of the nodes output by the second layer IAN are input into the third layer IAN, and sequentially reach the last layer IAN, so that the characteristics of each node are updated according to the information of the time hierarchy, and after the characteristics input into the first layer IAN and the output characteristics of the nodes output by each layer IAN are fully connected, the case number prediction result is obtained. When two IAN are included in the case number prediction model, the structure is shown in fig. 1, where Lin represents the fully connected layer.
Since different countries are affected by epidemic viruses at different times, it is likely that viral infections in other countries have already entered a relatively stable stage when some countries begin to infect the virus. Because the epidemic diseases belong to the same virus, the epidemic disease infection modes in the early stages of different countries have similar basic characteristics to a great extent, and moreover, a model trained for the whole period of the epidemic disease can better fit the future development trend of the epidemic disease, which cannot be achieved by a new model trained by the countries in the early stage of infection. Therefore, in the present embodiment, the parameter data in the trained case count prediction model of the other countries in the stable period of the epidemic disease is used for the case count prediction model of the country in the early stage of the epidemic disease.
Adding a migration learning method (TL) into the IAN, and training the IAN by using case data of different countries, wherein the method specifically comprises the following steps:
set-up meta training set Mxl={O(1),…,O(k),…,O(n)In which O is(k)The data set representing the kth country will also be scaled up for each part of the training as the number of predicted days increases. Taking the data of four countries as an example for explanation, the training of the IAN is divided into 4 times in total, each time the data set of three of the four countries is used as a meta-training set, and the data set of the remaining one country is used as a meta-test set, which is denoted as Mcs. Specific prediction goals will also be set, such as day 3, day 7, and day 14, etc. The task of this method is to pass through the meta-training set MxlMigration of learned parameters to meta-test set McsAmong the IAN models of (a).
Different training sets and different test sets are set as a single model. So, for country K, its set of tasks is:
in the formula:the training set and the test set for country k, respectively, are the first i days (using a minimum of two weeks of training days) and the test days are the j th day after the training days.
Will MxlMinimizing the loss of each partial training set to a specific task thetaSAs shown in formula (12).
In the formula: theta is the overall weight parameter and the bias parameter derived by the IAN framework,it is initially randomly initialized. Then will be atTheta after middle trainingSIs used for updatingA gradient of middle theta. As shown in equation (13).
Will pass through MxlThe learned parameter θ is applied to McsAnd (4) performing training.
Finally, at McsThe error is obtained by testing the test set, and fig. 3 shows the process of performing IAN training by using the transfer learning method in this embodiment. Firstly, sequentially inputting three national data of a meta-training set into the same group of IAN frames for training; all learned weights and bias parameters are then migrated into the model of the meta-test set. In other words, the learned parameters are initialized to the IAN framework of the meta-test set; and finally, inputting the national data of the meta-test set into the initialized IAN framework, and obtaining a prediction error through training and testing.
The disease case number prediction method based on the graph neural network and the transfer learning disclosed by the embodiment considers not only the historical case number of the region to be predicted, but also the mobility among the regions, so that the prediction result of the disease number is more accurate.
In addition, by using the information aggregation network comprising the graph convolution network and the graph attention mechanism, the node characteristics can be aggregated according to the difference of the importance of different nodes to the central node, so that the acquired output characteristics of the nodes are more accurate, and the accuracy of disease number prediction is further ensured.
In the embodiment, the problem that the prediction of the number of trained cases is inaccurate due to insufficient sample data for training the model in the country in the early stage of the disease is also considered, when the prediction of the model of the cases is performed, a transfer learning method is used, the data of the country in the stable period of the disease is used for model training, the parameters of the trained model are transferred to the prediction model of the country in the early stage of the disease, the model training is used for training the prediction model of the country in the early stage of the disease, the accuracy of the model training is ensured, and therefore the accurate prediction of the number of cases in the country in the early stage of the disease is achieved.
The performance of the disease case number prediction model disclosed in this example was verified using new coronary pneumonia data from four european countries (italy, uk, spain and france).
The details of the data set are shown in table 1, which shows the usage time interval of the data set of four countries, the number of regions each country contains, and the average number of newly added cases per day.
TABLE 1 details of the data set
To verify the predictive power of the method proposed in this example, a comparison was made with the following method, as follows:
1) LAST _ DAY uses the number of cases in the past few DAYs of the region as the predicted number of future cases.
2) AVG the average value of the daily number of cases in the area before prediction is used as the number of cases in the future.
3) AVG _ WINDOW compared to AVG, a time selection WINDOW is added to specify the average number of cases over a particular number of days in the past as the number of cases predicted in the future.
4) LSTM uses a two-layer long-short term memory network, and new node information is derived from the previous characteristics of the node. The LSTM used herein for comparison its input is the prediction of the previous week's case data.
5) PROPHET input is similar to ARIMA, and is suitable for prediction models of various time sequences.
6) ARIMA-autoregressive moving average model whose inputs are the entire time series data before prediction.
7) TL _ BASE was trained using the data set of 4 countries and tested on the data set of one of the countries. (Note: the training set includes countries for future testing)
To evaluate the performance of all models to predict the number of cases, they were tested on data sets in italy, british, spain and france. The results show the average error of each model prediction over 3 days, 7 days, and 14 days. The details are shown in table 2, table 3 and table 4.
TABLE 21-3 average errors for days
TABLE 31-7 days average error
TABLE mean error for days 41-14
It can be observed from the data shown in the above three tables that the disease case number prediction model provided in this embodiment achieves the optimal effect. This shows the effectiveness of the IAN computation framework and the transfer learning method on the case number prediction problem. Since the spread of epidemics has a high similarity in the short term, many models do not differ greatly in their effectiveness in short-term prediction. However, some conventional model methods show significant disadvantages as the prediction time is extended. In addition, since the time-series method does not consider the mobility between regions but only considers the time factor, a large performance difference is generated from the IAN proposed in this embodiment. Of the average prediction error results in four countries within 3 days, the IAN is 2.65, 1.99, 8.03 and 10.03 lower than that of the time series model with relatively good prediction results. The average prediction error results over 7 days were 2.01, 2.04, 7.02 and 7.87 lower, respectively. The average prediction error results over 14 days were 0.90, 0.97, 4.58 and 6.50 lower, respectively. It follows that the mobility between regions is a necessary consideration for predicting the number of cases.
Example 2
In this embodiment, a disease case number prediction system based on a graph neural network and transfer learning is disclosed, which includes:
the data acquisition module is used for acquiring the historical time case number of each region of the country to be predicted and the moving number of people of each region in corresponding time;
the characteristic vector acquisition module of the node is used for inputting the historical time case number of each region and the moving number of each region in corresponding time into the constructed digraph to obtain the characteristic vector of each node of the digraph;
the case number prediction module is used for inputting the feature vectors of all the nodes into a trained case number prediction model to obtain a case number prediction result;
the case number prediction model adopts a multi-layer information aggregation network, the output of the previous layer of information aggregation network is used as the input of the current information aggregation network, the input of the first layer of information aggregation network is the feature vector of each node of the directed graph, and the feature vector of each node of the directed graph and the output of each layer of information aggregation network are connected through a full connection layer to obtain a case number prediction result.
Example 3
In this embodiment, an electronic device is disclosed, which includes a memory, a processor, and computer instructions stored in the memory and executed on the processor, wherein the computer instructions, when executed by the processor, perform the steps of the method for predicting the number of disease cases based on the neural network and the transfer learning disclosed in embodiment 1.
Example 4
In this embodiment, a computer readable storage medium is disclosed for storing computer instructions which, when executed by a processor, perform the steps of the disease case number prediction method based on graph neural network and transfer learning disclosed in embodiment 1.
As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the above embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the embodiments of the invention without departing from the spirit and scope of the invention, which is to be covered by the claims.
Claims (10)
1. The disease case number prediction method based on the graph neural network and the transfer learning is characterized by comprising the following steps of:
acquiring the historical time case number of each region of a country to be predicted and the moving number of each region within corresponding time;
inputting the historical time case number of each region and the moving number of each region in corresponding time into the constructed digraph to obtain the characteristic vector of each node of the digraph;
inputting the feature vector of each node into a trained case number prediction model to obtain a case number prediction result;
the case number prediction model adopts a multi-layer information aggregation network, the output of the previous layer of information aggregation network is used as the input of the current information aggregation network, the input of the first layer of information aggregation network is the feature vector of each node of the directed graph, and the feature vector of each node of the directed graph and the output of each layer of information aggregation network are connected through a full connection layer to obtain a case number prediction result.
2. The disease case number prediction method based on graph neural network and transfer learning of claim 1, wherein a series of directed graphs are created for each country according to different time, nodes of the directed graphs represent regions in the country, and edges of the directed graphs are the number of moving persons between the regions in the country.
3. The disease case number prediction method based on graph neural network and transfer learning of claim 1, wherein the information aggregation network adopts a graph convolution network, and adds a graph attention mechanism in the graph convolution network.
4. The disease case number prediction method based on graph neural network and transfer learning of claim 3, wherein the process of obtaining the output features of the information aggregation network nodes is as follows:
the method comprises the steps of obtaining the depth characteristics of each node of the graph convolution network, inputting the depth characteristics of each node into a graph attention mechanism, obtaining attention weight coefficients among the nodes, and obtaining the output characteristics of each node of the information aggregation network according to the depth characteristics and the attention weight coefficients of the nodes.
5. The disease case number prediction method based on graph neural network and transfer learning of claim 4, wherein when the attention weight coefficient is calculated, the attention mechanism of the neighboring node of a certain node is normalized by using a softmax function.
6. The disease case number prediction method based on graph neural network and transfer learning of claim 1, wherein a transfer learning method is added to the information aggregation network, and the information aggregation network is trained by using disease data of different countries.
7. The disease case number prediction method based on graph neural network and transfer learning of claim 6, wherein the specific process of training the information aggregation network is as follows:
selecting the disease data of one country as a meta-test set, selecting the disease data of the rest countries as meta-training sets, inputting the disease data of the meta-training sets into a group of information aggregation networks for training, initializing the information aggregation networks of the meta-test sets by using the parameters of the trained information aggregation networks, inputting the disease data of the meta-test sets into the initialized information aggregation networks, and training the information aggregation networks of the meta-test sets.
8. The disease case number prediction system based on the graph neural network and the transfer learning is characterized by comprising the following steps:
the data acquisition module is used for acquiring the historical time case number of each region of the country to be predicted and the moving number of people of each region in corresponding time;
the characteristic vector acquisition module of the node is used for inputting the historical time case number of each region and the moving number of each region in corresponding time into the constructed digraph to obtain the characteristic vector of each node of the digraph;
the case number prediction module is used for inputting the feature vectors of all the nodes into a trained case number prediction model to obtain a case number prediction result;
the case number prediction model adopts a multi-layer information aggregation network, the output of the previous layer of information aggregation network is used as the input of the current information aggregation network, the input of the first layer of information aggregation network is the feature vector of each node of the directed graph, and the feature vector of each node of the directed graph and the output of each layer of information aggregation network are connected through a full connection layer to obtain a case number prediction result.
9. An electronic device comprising a memory and a processor, and computer instructions stored on the memory and executed on the processor, wherein the computer instructions, when executed by the processor, perform the steps of the method for predicting the number of cases of disease based on neural network and transfer learning according to any one of claims 1 to 7.
10. A computer-readable storage medium storing computer instructions which, when executed by a processor, perform the steps of the method for disease case number prediction based on graph neural network and transfer learning of any one of claims 1 to 7.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110877916.5A CN113517076A (en) | 2021-07-30 | 2021-07-30 | Disease case number prediction method and system based on graph neural network and transfer learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110877916.5A CN113517076A (en) | 2021-07-30 | 2021-07-30 | Disease case number prediction method and system based on graph neural network and transfer learning |
Publications (1)
Publication Number | Publication Date |
---|---|
CN113517076A true CN113517076A (en) | 2021-10-19 |
Family
ID=78068907
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110877916.5A Pending CN113517076A (en) | 2021-07-30 | 2021-07-30 | Disease case number prediction method and system based on graph neural network and transfer learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113517076A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114362994A (en) * | 2021-11-26 | 2022-04-15 | 北京交通大学 | Multilayer different-granularity intelligent aggregation railway system operation behavior safety risk identification method |
CN115359870A (en) * | 2022-10-20 | 2022-11-18 | 之江实验室 | Disease diagnosis and treatment process abnormity identification system based on hierarchical graph neural network |
CN115862888A (en) * | 2023-02-17 | 2023-03-28 | 之江实验室 | Infectious disease infection prediction method, system, device and storage medium |
CN116013541A (en) * | 2023-01-03 | 2023-04-25 | 武汉博科国泰信息技术有限公司 | Resource allocation method and system for tuberculosis detection |
CN117747121A (en) * | 2023-12-19 | 2024-03-22 | 首都医科大学宣武医院 | Diabetes risk prediction system based on multiple models |
CN118016318A (en) * | 2024-04-08 | 2024-05-10 | 中国科学院地理科学与资源研究所 | Method for constructing zoonosis risk prediction model based on graph neural network |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112116155A (en) * | 2020-09-18 | 2020-12-22 | 平安科技(深圳)有限公司 | Population mobility prediction method and device based on intelligent decision and computer equipment |
CN112863696A (en) * | 2021-04-25 | 2021-05-28 | 浙江大学 | Drug sensitivity prediction method and device based on transfer learning and graph neural network |
CN113035368A (en) * | 2021-04-13 | 2021-06-25 | 桂林电子科技大学 | Disease propagation prediction method based on differential migration diagram neural network |
-
2021
- 2021-07-30 CN CN202110877916.5A patent/CN113517076A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN112116155A (en) * | 2020-09-18 | 2020-12-22 | 平安科技(深圳)有限公司 | Population mobility prediction method and device based on intelligent decision and computer equipment |
CN113035368A (en) * | 2021-04-13 | 2021-06-25 | 桂林电子科技大学 | Disease propagation prediction method based on differential migration diagram neural network |
CN112863696A (en) * | 2021-04-25 | 2021-05-28 | 浙江大学 | Drug sensitivity prediction method and device based on transfer learning and graph neural network |
Non-Patent Citations (1)
Title |
---|
马帅: ""基于注意力实现的图神经网络"", 《计算机研究与发展》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114362994A (en) * | 2021-11-26 | 2022-04-15 | 北京交通大学 | Multilayer different-granularity intelligent aggregation railway system operation behavior safety risk identification method |
CN114362994B (en) * | 2021-11-26 | 2023-01-06 | 北京交通大学 | Multilayer different-granularity intelligent aggregation railway system operation behavior safety risk identification method |
CN115359870A (en) * | 2022-10-20 | 2022-11-18 | 之江实验室 | Disease diagnosis and treatment process abnormity identification system based on hierarchical graph neural network |
CN116013541A (en) * | 2023-01-03 | 2023-04-25 | 武汉博科国泰信息技术有限公司 | Resource allocation method and system for tuberculosis detection |
CN115862888A (en) * | 2023-02-17 | 2023-03-28 | 之江实验室 | Infectious disease infection prediction method, system, device and storage medium |
CN117747121A (en) * | 2023-12-19 | 2024-03-22 | 首都医科大学宣武医院 | Diabetes risk prediction system based on multiple models |
CN118016318A (en) * | 2024-04-08 | 2024-05-10 | 中国科学院地理科学与资源研究所 | Method for constructing zoonosis risk prediction model based on graph neural network |
CN118016318B (en) * | 2024-04-08 | 2024-06-14 | 中国科学院地理科学与资源研究所 | Method for constructing zoonosis risk prediction model based on graph neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN113517076A (en) | Disease case number prediction method and system based on graph neural network and transfer learning | |
Efimov et al. | On an interval prediction of COVID-19 development based on a SEIR epidemic model | |
La Gatta et al. | An epidemiological neural network exploiting dynamic graph structured data applied to the COVID-19 outbreak | |
Rochmawati et al. | Covid symptom severity using decision tree | |
Maathuis et al. | A Review of Some Recent Advances in Causal Inference. | |
CN109523021B (en) | Dynamic network structure prediction method based on long-time and short-time memory network | |
Lefèvre et al. | SIR-type epidemic models as block-structured Markov processes | |
CN113505924B (en) | Information propagation prediction method and system based on cascading space-time characteristics | |
De la Sen et al. | On an Sir Epidemic Model for the COVID‐19 Pandemic and the Logistic Equation | |
Paré et al. | Stability analysis and control of virus spread over time-varying networks | |
CN107590139A (en) | A kind of knowledge mapping based on circular matrix translation represents learning method | |
CN113035368A (en) | Disease propagation prediction method based on differential migration diagram neural network | |
Deng et al. | Concise summarization of heterogeneous treatment effect using total variation regularized regression | |
Shami et al. | Economic aspects of the detection of new strains in a multi-strain epidemiological–mathematical model | |
Lin et al. | Takagi-sugeno fuzzy model identification using coevolution particle swarm optimization with multi-strategy | |
Liang et al. | Discovering dynamic models of COVID‐19 transmission | |
Kuylen et al. | Social contact patterns in an individual-based simulator for the transmission of infectious diseases (stride) | |
CN114997504A (en) | Disease case number prediction method and system based on time sequence convolution network and graph convolution network | |
Hadiwibowo et al. | A policy strategy evaluation for covid-19 pandemic in the city of surabaya using vensim ventana dynamic system simulation | |
Ramesh et al. | Statistical model checking for dynamical processes on networks: a healthcare application | |
Saptarini et al. | Dengue haemorrhagic fever outbreak prediction using Elman Levenberg neural network and genetic algorithm | |
Palopoli et al. | A Markovian model for the spread of the SARS-CoV-2 virus | |
CN112185581A (en) | New coronary pneumonia SIT model inversion method based on Python and BP neural network | |
Burman et al. | A Flexible Agent-Based Model to Study COVID-19 Outbreak--A Generic Approach | |
CN118116614A (en) | Disease transmission dynamics method based on dynamic isolation |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |