CN116434969B - Multi-center chronic disease prediction device based on causal structure invariance - Google Patents

Multi-center chronic disease prediction device based on causal structure invariance Download PDF

Info

Publication number
CN116434969B
CN116434969B CN202310701410.8A CN202310701410A CN116434969B CN 116434969 B CN116434969 B CN 116434969B CN 202310701410 A CN202310701410 A CN 202310701410A CN 116434969 B CN116434969 B CN 116434969B
Authority
CN
China
Prior art keywords
causal
center
patient
chronic disease
causal structure
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.)
Active
Application number
CN202310701410.8A
Other languages
Chinese (zh)
Other versions
CN116434969A (en
Inventor
王丰
李劲松
池胜强
谭笑
周天舒
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang Lab
Original Assignee
Zhejiang Lab
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Zhejiang Lab filed Critical Zhejiang Lab
Priority to CN202310701410.8A priority Critical patent/CN116434969B/en
Publication of CN116434969A publication Critical patent/CN116434969A/en
Application granted granted Critical
Publication of CN116434969B publication Critical patent/CN116434969B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

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
    • G16H50/70—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
    • 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
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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
    • G06N3/047—Probabilistic or stochastic 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
    • G06N3/048—Activation functions
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N7/00—Computing arrangements based on specific mathematical models
    • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
    • 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Public Health (AREA)
  • Medical Informatics (AREA)
  • Primary Health Care (AREA)
  • Probability & Statistics with Applications (AREA)
  • Epidemiology (AREA)
  • Pathology (AREA)
  • Computational Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Pure & Applied Mathematics (AREA)
  • Mathematical Optimization (AREA)
  • Mathematical Analysis (AREA)
  • Algebra (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

The invention discloses a multi-center chronic disease prediction device based on invariance of a causal structure, which comprises the following steps of establishing a causal relation of each center according to characteristic data of a single center patient and the results of the single center patient, fitting the characteristic data of the patient to the results of the single center patient under the causal relation, and jointly correcting the causal relation according to the difference of the fitting results of the characteristic data of the patient under the causal relation and the actual results and the difference between the causal relations of different medical centers, wherein when the fitting error and the difference between the causal relations of different centers reach preset requirements, the causal relation is subjected to noise treatment to obtain a stable causal structure under a multi-center scene; establishing a more visual causal structure diagram through a causal structure; and constructing a chronic disease outcome prediction model through a causal structure chart. The invention digs the deeper causal logic relationship in the data, solves the problem of insufficient interpretability of the traditional method, and provides more reasonable decision advice for chronic disease decision support.

Description

Multi-center chronic disease prediction device based on causal structure invariance
Technical Field
The invention belongs to the technical field of medical health information, and particularly relates to a multi-center chronic disease prediction device based on invariance of a causal structure.
Background
In recent years, the incidence rate of chronic diseases has an ascending trend, and the health and the quality of life of residents are seriously affected, so that a huge disease burden is caused. According to the latest data of world health organization, the number of people dying from chronic diseases in 2019 worldwide is up to 4100 ten thousand, accounting for 74% of the number of people dying worldwide; and chronic diseases account for 80% of the first ten causes of death, accounting for 44% of deaths worldwide. The chronic diseases and morbidity of China also rise rapidly, and the premature death rate of residents in China in 2019 caused by four major chronic diseases such as cardiovascular and cerebrovascular diseases, cancers, chronic respiratory diseases and diabetes is 16.5 percent. Therefore, early prediction of the risk of chronic disease onset is important for the prevention and control of diseases by actively taking intervention measures.
With the continuous development of artificial intelligence technology, algorithms for machine learning and deep learning are widely applied to the field of health care. The existing chronic disease diagnosis and prediction methods mainly use machine learning-based methods such as a support vector machine, logistic regression, naive Bayes and the like. For example, the lung cancer stage situation is predicted by using data such as a sensor, age and the like through a support vector machine method; and (3) carrying out chronic hepatitis stage analysis and the like by using gene expression information through a random forest and k nearest neighbor method. However, in real life, data between medical centers has heterogeneity due to differences in regions, population, intervention types, data statistics methods, and the like. If the methods are directly used for carrying out cross-center analysis on the heterogeneity data, the prediction results of the same model aiming at different regions and different medical centers have larger difference. However, the inherent causal links of medical decision problems of the same scenario should be stable and should not vary with data distribution differences.
The traditional machine learning method is based on a correlation fitting function and data, and cannot distinguish whether the clinical features have causal or pseudo-correlation with the occurrence of diseases. The model system constructed based on the pseudo-correlation cannot explain the real reasons behind the prediction. Conventional interpretability methods, such as the variable importance method SHAP, can subdivide the predicted outcome by calculating the SHAP value and evaluate the extent to which each feature affects the outcome.
The Chinese patent of invention with bulletin number of CN106169165B discloses a symptom hierarchy association and prediction method for diagnosis and treatment data, which constructs a symptom topic hierarchy space based on a hierarchy topic model, adopts a maximum probability criterion to realize diagnosis and treatment record and hierarchical mapping of patients, and comprehensively considers various attribute information of diagnosis and treatment places, ages, sexes and time of the patients to realize dynamic prediction of diseases.
The Chinese patent application with publication number of CN113744870A discloses a main diagnosis and prediction system and method for a first page of a medical record, wherein the system extracts clinical data associated with patients from a database, performs data cleaning and feature construction on the clinical data associated with the patients, obtains the features associated with the patients, and stores the features in a modeling sample library; then, the first round of grouping is carried out on the patients according to the characteristics associated with the patients, and the first round of grouping identification associated with the patients is obtained; generating a corresponding main diagnosis prediction model of the first medical records page for each first round of grouping respectively; when the system receives a main diagnosis prediction request, according to a new patient identifier carried by the request, the system sequentially passes through a data organization module and a crowd grouping module to obtain a first round of grouping identifiers related to the new patient, and then invokes a main diagnosis prediction model corresponding to the first round of grouping identifiers to obtain a main diagnosis prediction result of the new patient.
These methods only account for the relevance of the influence of independent variables on outcomes, lacking an explanatory analysis of the causal structure between the variables and the outcomes.
Disclosure of Invention
In order to solve the defects of the prior art, the invention provides a multi-center chronic disease prediction device based on invariance of a causal structure, which is used for excavating a deeper causal logic relationship in data through a multi-layer graph neural network and carrying out visual display of the causal relationship through a causal structure representation module, so that the problem of insufficient interpretability of the traditional method is solved, and a more reasonable decision suggestion is provided for chronic disease decision support.
A causal structural invariance based multi-center chronic disease prediction apparatus comprising the steps of:
step 1, collecting data: collecting data of chronic patients in different medical centers;
step 2, data preprocessing: preprocessing the patient data obtained in the step 1, aligning the characteristics of different medical centers with the same semantic concepts and different expressions, and establishing a characteristic data set of a multi-center patient and a final set of the multi-center patient;
step 3, obtaining a stable causal structure matrix in a multi-center scene: establishing a causal relation of each center according to the characteristic data of a single center patient and the results thereof, fitting the characteristic data of the patient to the results thereof under the causal relation, correcting the causal relation together according to the difference of the fitting results of the characteristic data of the patient under the causal relation and the actual results and the difference between the causal relations of different medical centers, and obtaining a stable causal structure matrix under a multi-center scene after the causal relation is subjected to noise treatment when the fitting error and the difference between the causal relations of different centers reach preset requirements;
step 4, causal structure matrix transformation: establishing a more visual causal structure diagram through a causal structure matrix;
step 5, predicting the ending: and constructing a chronic disease ending prediction model through a causal structure chart, and inputting the characteristic data of the patient into the prediction model to obtain the chronic disease ending prediction.
Preferably, in step 3, a causal relationship of each center is established, the characteristic data of the patient is fitted to the results of the patient under the causal relationship, and the causal relationship is corrected together with the causal relationship of different medical centers according to the difference of the fitted results of the characteristic data of the patient under the causal relationship and the actual results, and the causal relationship is trained through a graph neural network.
Further, the graph neural network is of a two-layer network structure, the first-layer neural network convolves the characteristic data set of a single central patient with the central causality initialized randomly one by one, the first-layer network output is obtained after the central causality is activated, the second-layer graph neural network convolves the central causality with the first-layer network output, the second-layer network output is obtained after the central causality is activated, the two-layer graph neural network is connected with a fully-connected network, the fitting result is obtained after the central causality is activated, and network parameters and the causality are optimized by using an optimizer.
Further, the correction uses the formula:
wherein c is the number of medical centers,for the number of patients in the ith medical center,for the actual outcome of the jth patient of the ith medical center, a representation 1 occurs, no representation 0 occurs,fitting results after modeling for the data of the jth patient of the ith medical center,for a multi-center fit result, Y is the multi-center actual result,for multi-center fitting of differences in the outcomes and actual outcomes,representing causal relationships of different medical centersIs a function of the variance of (a),is the difference of different central causal relations.
Preferably, in step 3, when the difference between the fitting error and the causal relationship of different centers meets a preset requirement, the causal relationship is back-deduced through the heterogeneity mapping to obtain a causal structure matrix, wherein Representing causal relationships of different medical centers.
Preferably, in step 4, the specific method for creating a more intuitive causal structure map through the causal structure matrix is: the median of the causal structure matrix isIf the element of the ith row and jth column in the causal structure matrixThe representative feature j is the cause of feature i, and in the causal structure, the feature i and the feature j form a unidirectional edge with j pointing to i, ifThen feature i does not make a causal relationship with feature j.
Preferably, in step 5, the specific method for constructing the chronic disease outcome prediction model through the causal structure chart is as follows: building a global adjacency matrix of size N x NWherein if the feature i and the feature j in the causal structure diagram form a unidirectional edge with j pointing to i, thenThe element of the ith row and the jth column is 1, otherwise, the element is 0; using global adjacency matricesConstruction of a 2-layer graph convolutional neural network for predicting patient outcome in chronic diseaseThe situation is generated.
Further, global adjacency matrix is utilizedA 2-layer graph convolution neural network is constructed and used for predicting the occurrence of chronic disease ending of a patient, and the following formula is adopted:
wherein ,as a characteristic data of the patient,for the predicted outcome of chronic disease in patients, reLU ()In order to activate the function,、as a result of the trainable network parameters,the calculation formula of (2) is as follows:
is thatIs a matrix of degrees of (a),is a unit diagonal matrix.
Further, the method comprises the steps of. Through loss ofUpdating using cross entropy、Loss ofThe specific formula of (2) is:
wherein ,for the number of patients in the ith medical center,for the actual outcome of the jth patient of the ith medical center, a representation 1 occurs, no representation 0 occurs,fitting results after the data of the jth patient of the ith medical center is subjected to the model.
The invention also provides a multi-center chronic disease prediction device based on causal structure invariance, which comprises:
the data acquisition module is used for acquiring characteristic data of chronic patients in different medical centers;
the data preprocessing module is used for preprocessing patient data;
the multi-center graph structure learning module comprises a graph neural network training unit and a heterogeneity mapping unit, wherein the graph neural network training unit is used for fitting a multi-center patient outcome, acquiring a multi-center causal relationship through the difference between fitting errors and causal relationships of different medical centers, and acquiring a multi-center universal causal structure after the heterogeneity mapping unit carries out noise processing on the causal relationship;
the causal structure representation module is used for establishing a more visual causal structure diagram through a causal structure;
and the chronic disease prediction module is used for constructing a chronic disease outcome prediction model through the causal structure chart, and inputting the characteristic data of the patient into the prediction model to obtain the chronic disease outcome prediction.
The invention has the beneficial effects that:
aiming at the problem that the traditional prognosis prediction model is difficult to process heterogeneous data, the invention provides a multi-center chronic disease prediction device by utilizing the characteristic that a causal structure among different medical centers is unchanged, a heterogeneous relation of characteristics among different medical centers is simulated through a heterogeneous mapping module, the invariable causal structure is learned through a multi-layer graph neural network, the generalization of the model for different distribution samples is improved, and the heterogeneous clinical data can be subjected to chronic disease prognosis prediction through reliable causal relation.
Aiming at the problem that the traditional machine learning interpretability method only quantitatively describes the importance of variables and can not provide qualitative causal structural interpretation, the invention digs the deeper causal logic relationship in the data, solves the problem of insufficient interpretability of the traditional method, and provides more reasonable decision suggestion for chronic disease decision support.
Drawings
FIG. 1 is a flow chart of a causal structural invariance-based multi-center chronic disease prediction method provided by the invention.
Fig. 2 is a schematic representation of multi-center patient characterization data provided by the present invention.
Fig. 3 is a diagram of a multi-center diagram structure learning process according to the present invention.
Fig. 4 is a network structure of the neural network according to the present invention.
FIG. 5 is a causal block diagram provided by the present inventionIs a schematic diagram of the generation process.
Fig. 6 is a schematic block diagram of a chronic disease prediction apparatus provided by the present invention.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
FIG. 1 is a flowchart of a multi-center chronic disease prediction method based on causal structural invariance provided by the invention, as shown in FIG. 1, the multi-center chronic disease prediction method provided by the invention comprises the following steps:
step 1, collecting data: characteristic data of chronic patients in different medical centers are collected.
In an embodiment, the types of characteristic data collected from the patient include demographic, surgical, medication, laboratory, diagnostic, and daily monitoring data. The feature information specifically includes:
demographic information including gender, age, ethnicity, region, etc.; the medication information comprises complication medication and the like; assay information including blood routine, liver function, etc.; the diagnosis information comprises chronic disease diagnosis, complications and the like; daily monitoring information comprises blood pressure, weight and the like; based on the characteristic information, the original data of the chronic disease patient is collected.
Step 2, data preprocessing: and (3) preprocessing the patient data obtained in the step (1).
In the embodiment, based on the original data of the chronic disease patient, the same concept and different expression characteristics of different medical centers are unified, meanwhile, unreasonable data, error data, repeated data and the like in the original data are deleted, the missing continuous data are filled by adopting an average value, and the missing discrete data are filled by adopting a mode.
A schematic of multi-centered patient characterization data is shown in fig. 2. The set of all chronic disease related features of the medical center is denoted as F, wherein the number of feature types included is denoted as N. Representing the feature data set of the ith medical center patient as having a size ofIs a two-dimensional matrix of (2), wherein ,for the number of patients in the ith medical center,each 1 column represents the values of all types of characteristics of a patient at a 1 st medical center,every 1 row represents the values of all patients for a certain 1 type of feature in the ith medical center.
The feature set F also comprises the actual ending y of the patient, and when no ending event occurs, the value of y is 0; when a finalization event occurs, y takes a value of 1. Taking chronic kidney disease cardiovascular event prediction as an example, y represents whether a chronic kidney disease patient will have cardiovascular event within 1 year, when the value of y is 1, the patient will have cardiovascular event, and when the value of y is 0, the patient will not have cardiovascular event. Patient outcome sets for the ith medical center are represented as lengthVector of (3)，Is a target outcome event for a 1 st patient at the ith medical center. It will be appreciated that whenWhen one row of (a) is all 0 s,a set of characteristic data representing patients of the ith medical center chronically ill patient for whom a target outcome event has not occurred.
The number of medical centers participating in statistics is marked as c, and the characteristic data set of the multi-center patient is thatMultiple center patient outcome set as. Patient map data { X, Y } is constructed from the patient's feature data set and the outcome set.
Step 3, obtaining a stable causal structure matrix in a multi-center scene: according to the characteristic data of a single central patient and the results thereof, a causal relation of each center is established, the characteristic data of the patient is fitted to the results thereof under the causal relation, the causal relation is corrected together according to the difference of the fitting results of the characteristic data of the patient under the causal relation and the actual results and the difference between the causal relations of different medical centers, and when the fitting error and the difference between the causal relations of different centers reach preset requirements, a causal structure matrix with universal multi-centers is obtained after the causal relation is subjected to noise treatment.
In an embodiment, the causal relationship among N types of features in the feature set F is learned in a multi-center scene, and a stable causal structure matrix in the multi-center scene is learned based on graph neural network training and heterogeneity mapping. FIG. 3 is a diagram of a multi-center diagram structure learning process in which the causal relationship structure between features of the ith medical center is represented as a learnable relationship adjacency matrixThe multi-center causal relationship is that。
Fig. 4 shows a network structure of the neural network according to the present invention. The multi-center causality a is randomly initialized, and is learned by fitting to the multi-center patient outcome Y using a graph neural network. Specifically, the graph neural network is of a two-layer network structure, the first-layer neural network convolves the multi-center causality A with the characteristic data X of the multi-center patient, and then sigmoid activation is connected to obtain a first-layer network output X 1 Denoted as X 1 =sigmoid(AX)=sigmoid(A 1 X 1 ,…, A i X i ,…,A c X c ). The neural network of the second layer graph outputs the multi-center causality A and the first layer network output X 1 Convolving, then activating by sigmoid to obtain second layer network output X 2 Denoted as X 2 =sigmoid(AX 1 )= sigmoid(A 1 sigmoid(A 1 X 1 ),…,A i sigmoid(A i X i ),…,A c sigmoid(A c X c )). The two-layer graph neural network is connected with a full-connection network, and the fitting result is obtained after sigmoid is activated=sigmoid(X 2 W+b). Wherein W, B is a network parameter.
In the training process, the multi-center causal relationship A is corrected together according to the difference of the fitting result of the characteristic data of the patient under the causal relationship and the actual result and the difference between the causal relationships of different medical centers, and the correction adopts the formula:
wherein c is the number of medical centers,for the number of patients in the ith medical center,for the actual outcome of the jth patient of the ith medical center, a representation 1 occurs, no representation 0 occurs,fitting results after modeling for the data of the jth patient of the ith medical center,to fit the outcome, Y is the actual outcome,to fit the difference between the end and the actual end,representing different medical centersIs a function of the variance of (a),is the difference of different central causal relations.
Optimizing multi-center causal relationship a, network parameter W, and network parameter using an optimizer such as adam algorithm. And after the model is stably converged, the multi-center causal relationship A can be obtained.
The multi-center causality A is obtained through heterogeneity mapping back-extrusionCausal structure matrix. Specifically, the relationship adjacency matrix of the ith medical centerCan be expressed as a causal structure matrixHeterogeneity map of (a), i.e.). Multicenter causality a== Let noise be expressed as noise. Thus, multi-center causal relationship. Due toThe terms make the variance of different medical centers as small as possible, so the causal structure matrixThe total noise is minimal, i.e. causal structure matrix。
Step 4, causal structure matrix transformation: and a more visual causal structure diagram is established through the causal structure matrix.
In an embodiment, as shown in FIG. 5, a causal structure diagramWherein feature 1, feature 2, feature 3, feature 4 and feature 5 represent respectivelyIs a different type of 5 features. Causal structure diagramIs a directed graph showing the causal structure between features and outcomes. Specifically, a causal structure matrixIs of the median ofIf a causal structure matrixElements of row i and column j of the listThen the representative feature j is the cause of feature i, in the causal structure diagramThe feature i and the feature j form a unidirectional edge of j- & gt i, otherwise, ifNo causal relationship is constructed. The causal structure matrix as shown in fig. 5In which the element of row 2 and column 4 is 1, then feature 3 is the cause of feature 1, inThe middle feature 1 and the feature 3 form a unidirectional edge of the feature 3-the feature 1; if the element of row 4 and column 2 is 0, then feature 1 is not the cause of feature 3.
Step 5, predicting the ending: and constructing a chronic disease ending prediction model through a causal structure chart, and inputting the characteristic data of the patient into the prediction model to obtain the chronic disease ending prediction.
In an embodiment, a global adjacency matrix of size N is constructed. Wherein, if a causal structure diagramThe feature i and the feature j form a unidirectional edge of i- & gt jThe element of the j-th row and i-th column is 1, otherwise 0. Using global adjacency matricesA layer 2 graph roll-up neural network may be constructed for predicting the occurrence of chronic disease outcome events. Specifically, the formula is as follows:
wherein ,as a characteristic data of the patient,for the predicted outcome of chronic disease in patients, reLU ()In order to activate the function,、as a result of the trainable network parameters,the calculation formula of (2) is as follows:
is thatIs a matrix of degrees of (a),is a unit diagonal matrix.
Loss ofUsing cross entropy to update、The specific formula is as follows:
wherein ,for the number of patients in the ith medical center,for the actual outcome of the jth patient of the ith medical center, a representation 1 occurs and no representation 0 occurs.Fitting results after the data of the jth patient of the ith medical center is subjected to the model. Optimization using an optimizer such as adam algorithmNetwork parameters、. The model can be used for predicting the chronic disease outcome after stable convergence.
Based on the above method, the embodiment further provides a multi-center chronic disease prediction device based on invariance of a causal structure, which comprises a memory and one or more processors, wherein executable codes are stored in the memory, and the one or more processors execute the executable codes to realize the method, and the method comprises the following steps:
step 1, collecting data;
step 2, data preprocessing;
step 3, obtaining a stable causal structure matrix in a multi-center scene;
step 4, converting a causal structure matrix;
and 5, predicting the ending.
FIG. 6 also provides a multi-center chronic disease prediction apparatus based on causal structure invariance, comprising a data acquisition module, a data preprocessing module, a multi-center graph structure learning module, a causal structure representation module and a chronic disease prediction module.
The data acquisition module is used for acquiring characteristic data of chronic patients in different medical centers and transmitting the characteristic data to the data preprocessing module.
In particular, the types of characteristic data collected from a patient include demographic, surgical, medication, laboratory, diagnostic, and daily monitoring data. The feature information specifically includes: demographic information including gender, age, ethnicity, region, etc.; the medication information comprises complication medication and the like; assay information including blood routine, liver function, etc.; the diagnosis information comprises chronic disease diagnosis, complications and the like; daily monitoring data including blood pressure, body weight, etc.; based on the characteristic information, the original data of the chronic disease patient is collected.
The data preprocessing module preprocesses the characteristic data of the chronic disease patient and transmits the characteristic data to the multi-center structure learning module.
Specifically, the data preprocessing module preprocesses patient data obtained by the data acquisition module, aligns features of different medical centers with the same semantic concepts and different expressions, and establishes a feature data set X of the multi-center patient and a final set Y of the multi-center patient.
The multi-center structure learning module is used for fitting the multi-center patient outcome, and obtaining multi-center causal relation through the difference between fitting errors and causal relation of different medical centers, and the causal relation is processed by noise to obtain a stable causal structure in a multi-center scene.
Specifically, the multi-center learning module includes a graph neural network training unit and a heterogeneity mapping unit, the graph neural network training unit fitting the multi-center patient actual outcome Y to learn the multi-center causality a.
The multi-center causality a is first randomly initialized and learned by fitting to the multi-center patient outcome Y using a graph neural network. The graph neural network is of a two-layer network structure, the first-layer neural network convolves the multi-center causality A with the characteristic data set X of the multi-center patient, and then sigmoid activation is connected to obtain a first-layer network output X 1 Denoted as X 1 =sigmoid(AX)= sigmoid(A 1 X 1 ,…, A i X i ,…,A c X c ). The second layer graph neural network outputs X with the first layer network through the multi-center causality A 1 Convolving, then activating by sigmoid to obtain second layer network output X 2 Denoted as X 2 =sigmoid(AX 1 )=sigmoid(A 1 sigmoid(A 1 X 1 ),…,A i sigmoid(A i X i ),…, A c sigmoid(A c X c )). The two-layer graph neural network is connected with a full-connection network, and the fitting result is obtained after sigmoid is activated=sigmoid(X 2 W+b). Wherein W, B is a network parameter.
In the training process, the multi-causal relation A is corrected together according to the difference of the fitting outcome of the characteristic data of the patient under the causal relation and the actual outcome and the difference between the causal relations of different medical centers, wherein the specific formula is as follows:
wherein c is the number of medical centers,for the number of patients in the ith medical center,for the actual outcome of the jth patient of the ith medical center, a representation 1 occurs, no representation 0 occurs,fitting results after modeling for the data of the jth patient of the ith medical center,to fit the outcome, Y is the actual outcome,to fit the difference between the end and the actual end,representing different medical centersIs a function of the variance of (a),is the difference of different central causal relations.
Optimizing multi-center causal relationship a, network parameter W, and network parameter using an optimizer such as adam algorithm. And after the model is stably converged, the multi-center causal relationship A can be obtained.
The heterogeneity mapping unit reversely pushes the multi-center causality A to obtain a causality structure matrix. Multi-center causal relationship matrix a==Let noise be expressed as noise. Thus, a multi-center causal relationship matrix. Due toThe terms make the variance of different medical centers as small as possible, so the causal structure matrixThe total noise is minimal, i.e. causal structure matrix。
The causal structure representation module converts the causal structure matrix into a more intuitive causal structure mapImproving the interpretation of the causal structure of chronic diseases and reducing the difficulty in understanding the causal structureDegree.
Specifically, causal structure diagramIs a directed graph showing the causal structure between features and outcomes. Causal structure matrixIs of the median ofIf a causal structure matrixElements of row i and column j of the listThen the representative feature j is the cause of feature i, in the causal structure diagramThe feature i and the feature j form a unidirectional edge of j- & gt i, otherwise, ifNo causal relationship is constructed.
The chronic disease prediction module constructs a chronic disease outcome prediction model through a causal structure chart, and inputs the characteristic data of the patient into the prediction model to obtain the chronic disease outcome prediction.
Specifically, a global adjacency matrix of size N x N is constructed. Wherein, if a causal structure diagramThe feature i and the feature j form a unidirectional edge of i- & gt jThe element of the j-th row and i-th column is 1, otherwise 0. Using global adjacency matricesA layer 2 graph roll-up neural network may be constructed for predicting the occurrence of chronic disease outcome events. The formula is as follows:
wherein ,as a characteristic data of the patient,for the predicted outcome of chronic disease in patients, reLU ()In order to activate the function,、as a result of the trainable network parameters,the calculation formula of (2) is as follows:
is thatIs a matrix of degrees of (a),is a unit diagonal matrix.
Loss ofUsing cross entropy to update、The specific formula is as follows:
wherein ,for the number of patients in the ith medical center,for the actual outcome of the jth patient of the ith medical center, a representation 1 occurs and no representation 0 occurs.Fitting results after the data of the jth patient of the ith medical center is subjected to the model. Optimizing network parameters using an optimizer such as adam algorithm、. The model can be used for predicting the chronic disease outcome after stable convergence.
Although the present invention has been described with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications may be made to the embodiments described, or equivalents may be substituted for elements thereof, and any modifications, equivalents, improvements and changes may be made without departing from the spirit and principles of the present invention.

Claims (8)

1. A causal structure invariance based multi-center chronic disease prediction apparatus comprising a memory and one or more processors, said memory having executable code stored therein, said one or more processors executing said executable code to effect the steps of:
step 1, collecting data: collecting data of chronic patients in different medical centers;
step 2, data preprocessing: preprocessing the patient data obtained in the step 1, aligning the characteristics of different medical centers with the same semantic concepts and different expressions, and establishing a characteristic data set of a multi-center patient and a final set of the multi-center patient;
step 3, obtaining a stable causal structure matrix in a multi-center scene: establishing a causal relation of each center based on a pre-constructed graph neural network model according to the characteristic data of a single center patient and the ending thereof, fitting the characteristic data of the patient to the ending thereof under the causal relation, correcting the causal relation together according to the difference of the fitting ending of the characteristic data of the patient under the causal relation and the actual ending and the difference between the causal relations of different medical centers, obtaining a stable causal structure matrix under a multi-center scene after noise processing when the difference between the fitting error and the causal relations of different centers reaches the preset requirement,
the graph neural network is of a two-layer network structure, the first-layer neural network convolves the characteristic data set of a single central patient with the central causality initialized randomly one by one, the first-layer network output is obtained after the activation, the second-layer graph neural network convolves the central causality with the first-layer network output, the second-layer network output is obtained after the activation, the two-layer graph neural network is connected with a full-connection network, the fitting result is obtained after the activation, and the network parameters and the causality are optimized by using an optimizer;
step 4, causal structure matrix transformation: establishing a more visual causal structure diagram through a causal structure matrix;
step 5, predicting the ending: and constructing a chronic disease ending prediction model through a causal structure chart, and inputting the characteristic data of the patient into the prediction model to obtain the chronic disease ending prediction.
2. The causal structure invariance based multi-center chronic disease prediction apparatus of claim 1, wherein the correction uses the formula:
，
，
，
，
wherein c is the number of medical centers,for the ith medical center patient count, +.>For the actual outcome of the jth patient of the ith medical centre, the occurrence indicates 1 and the non-occurrence indicates 0,/is>Fitting outcome after modeling for data of the jth patient of the ith medical center,/th patient>Fitting the outcome for multicenter, Y for multicenter actual outcome, < >>Fitting differences between the outcomes and the actual outcomes for multiple centers,/->Representing causal relationship of different medical centers +.>Variance of->Is the difference of different central causal relations.
3. The causal structure invariance based multi-center chronic disease prediction apparatus according to claim 1, wherein in the step 3, when the difference between the fitting error and the different center causal relationships reaches a preset requirement, the causal relationships are back-deduced through heterogeneity mapping to obtain a causal structure matrix, wherein Representing causal relationships of different medical centers.
4. The causal structure invariance based multi-center chronic disease prediction apparatus according to claim 1, wherein in the step 4, the specific method for creating a more intuitive causal structure map through a causal structure matrix is as follows: the median of the causal structure matrix isIf the element of the ith row and jth column in the causal structure matrix +.>The representative feature j is the cause of feature i, and in the causal structure diagram, features i and j form a unidirectional edge with j pointing to i, if +.>Then feature i does not make a causal relationship with feature j.
5. The causal structure invariance based multi-center chronic disease prediction apparatus according to claim 4, wherein in the step 5, the specific method for constructing the chronic disease outcome prediction model through the causal structure chart is as follows: building a global adjacency matrix of size N x NWherein if the feature i and the feature j in the causal structure diagram form a unidirectional edge with j pointing to i, then +.>The element of the ith row and the jth column is 1, otherwise, the element is 0; using global adjacency matrix->And constructing a 2-layer graph convolution neural network for predicting the occurrence of chronic disease ending of the patient.
6. The causal structure invariance based multi-center chronic disease prediction apparatus of claim 5, wherein the global adjacency matrix is utilizedA 2-layer graph convolution neural network is constructed and used for predicting the occurrence of chronic disease ending of a patient, and the following formula is adopted:
，
，
wherein ,for the characteristic data of the patient->For the predictive outcome of chronic diseases in patients, reLU (),>to activate the function +.>、/>For trainable network parameters, +.>The calculation formula of (2) is as follows:
，
，
is->Degree matrix of>Is a unit diagonal matrix.
7. The causal structural invariance based multi-center chronic disease prediction apparatus of claim 6The device is characterized by thatUpdating +.>、/>Loss->The specific formula of (2) is:
，
，
wherein ,for the ith medical center patient count, +.>For the actual outcome of the jth patient of the ith medical centre, the occurrence indicates 1 and the non-occurrence indicates 0,/is>Fitting results after the data of the jth patient of the ith medical center is subjected to the model.
8. A causal structural invariance based multi-center chronic disease prediction apparatus, comprising:
the data acquisition module is used for acquiring characteristic data of chronic patients in different medical centers;
the data preprocessing module is used for preprocessing patient data;
the multi-center graph structure learning module comprises a graph neural network training unit and a heterogeneity mapping unit, wherein the graph neural network training unit is used for fitting a multi-center patient outcome, acquiring a multi-center causal relationship through the difference between fitting errors and causal relationships of different medical centers, and acquiring a multi-center universal causal structure after the heterogeneity mapping unit carries out noise processing on the causal relationship;
the causal structure representation module is used for establishing a more visual causal structure diagram through a causal structure;
and the chronic disease prediction module is used for constructing a chronic disease outcome prediction model through the causal structure chart, and inputting the characteristic data of the patient into the prediction model to obtain the chronic disease outcome prediction.
CN202310701410.8A 2023-06-14 2023-06-14 Multi-center chronic disease prediction device based on causal structure invariance Active CN116434969B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310701410.8A CN116434969B (en) 2023-06-14 2023-06-14 Multi-center chronic disease prediction device based on causal structure invariance

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310701410.8A CN116434969B (en) 2023-06-14 2023-06-14 Multi-center chronic disease prediction device based on causal structure invariance

Publications (2)

Publication Number Publication Date
CN116434969A CN116434969A (en) 2023-07-14
CN116434969B true CN116434969B (en) 2023-09-12

Family

ID=87080146

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310701410.8A Active CN116434969B (en) 2023-06-14 2023-06-14 Multi-center chronic disease prediction device based on causal structure invariance

Country Status (1)

Country Link
CN (1) CN116434969B (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112133420A (en) * 2019-06-06 2020-12-25 日本电气株式会社 Data processing method, device and computer readable storage medium
CN113409901A (en) * 2021-06-29 2021-09-17 南华大学 Causal inference method and system for cascade medical observation data
CN113871003A (en) * 2021-12-01 2021-12-31 浙江大学 Disease auxiliary differential diagnosis system based on causal medical knowledge graph
CN114664452A (en) * 2022-05-20 2022-06-24 之江实验室 General multi-disease prediction system based on causal verification data generation
CN115691652A (en) * 2022-11-17 2023-02-03 国科大杭州高等研究院 Causal relationship prediction method based on multi-sample data

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115698335A (en) * 2020-05-22 2023-02-03 因斯特罗公司 Predicting disease outcome using machine learning models

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112133420A (en) * 2019-06-06 2020-12-25 日本电气株式会社 Data processing method, device and computer readable storage medium
CN113409901A (en) * 2021-06-29 2021-09-17 南华大学 Causal inference method and system for cascade medical observation data
CN113871003A (en) * 2021-12-01 2021-12-31 浙江大学 Disease auxiliary differential diagnosis system based on causal medical knowledge graph
CN114664452A (en) * 2022-05-20 2022-06-24 之江实验室 General multi-disease prediction system based on causal verification data generation
CN115691652A (en) * 2022-11-17 2023-02-03 国科大杭州高等研究院 Causal relationship prediction method based on multi-sample data

Also Published As

Publication number Publication date
CN116434969A (en) 2023-07-14

Similar Documents

Publication Publication Date Title
Kaur et al. Medical diagnostic systems using artificial intelligence (ai) algorithms: Principles and perspectives
US20210343384A1 (en) Systems and methods for managing autoimmune conditions, disorders and diseases
De Bois et al. Adversarial multi-source transfer learning in healthcare: Application to glucose prediction for diabetic people
CN113421652A (en) Method for analyzing medical data, method for training model and analyzer
Jiang et al. The Research of Clinical Decision Support System Based on Three‐Layer Knowledge Base Model
Raeesi Vanani et al. IoT-based diseases prediction and diagnosis system for healthcare
CN111387938B (en) Patient heart failure death risk prediction system based on characteristic rearrangement one-dimensional convolutional neural network
D’Antoni et al. Auto-Regressive Time Delayed jump neural network for blood glucose levels forecasting
Vieira et al. Deep neural networks
CN112542242A (en) Data transformation/symptom scoring
AU2024204390A1 (en) Computer based object detection within a video or image
CN115239993A (en) Human body alopecia type and stage identification system based on cross-domain semi-supervised learning
CN116959715B (en) Disease prognosis prediction system based on time sequence evolution process explanation
Dutta Detecting Lung Cancer Using Machine Learning Techniques.
Yang et al. DeepMPM: a mortality risk prediction model using longitudinal EHR data
CN114611879A (en) Clinical risk prediction system based on multitask learning
WO2024027438A1 (en) Personalized state-space progression model-based assisted decision-making system for disease
Rahman et al. Ensemble of multiple models for robust intelligent heart disease prediction system
Yadav et al. Machine learning in expert systems for disease diagnostics in human healthcare
CN113990502A (en) ICU heart failure prediction system based on heterogeneous graph neural network
Mohammadi et al. A user-friendly approach for the diagnosis of diabetic retinopathy using ChatGPT and automated machine learning
Kalaivani et al. Heart disease diagnosis using optimized features of hybridized ALCSOGA algorithm and LSTM classifier
CN116434969B (en) Multi-center chronic disease prediction device based on causal structure invariance
CN117038096A (en) Chronic disease prediction method based on low-resource medical data and knowledge mining
Tavakoli et al. Predicting neural deterioration in patients with alzheimer’s disease using a convolutional neural network

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
GR01 Patent grant
GR01 Patent grant