A causal relationship mining method for high-dimensional time series data of gestational diabetes mellitus

Abstract

This invention relates to the field of data processing, specifically to a method for causal relationship mining of high-dimensional time-series data on gestational diabetes mellitus, comprising the following steps: S10: acquiring multivariate time-series data on gestational diabetes mellitus in pregnant women, and preprocessing the multivariate time-series data to form a time-series dataset for causal analysis; S20: based on the time-series dataset, performing feature modeling on the causal relationships between variables, calculating the causal strength and performing adaptive grouping low-rank modeling to construct a global causal probability graph; S30: based on the global causal probability graph, performing parameter alternation optimization on the time-series prediction model, and introducing time smoothing constraints during the optimization process, and using the optimized time-series prediction model to output the final time-series causal structure; this application can achieve stable modeling of high-dimensional time-series causal structures while ensuring the accuracy of causal discovery, providing a reliable basis for subsequent causal analysis and decision support.

Description

Technical Field

[0001] This invention relates to the field of data processing, and more specifically to a method for mining causal relationships in high-dimensional time-series data of gestational diabetes mellitus. Background Technology

[0002] With the development of the Internet of Things (IoT) in healthcare and wearable devices, multivariate time-series data are widely used in the clinical management of gestational diabetes mellitus (GDM). By analyzing time-series data such as continuous glucose monitoring (CGM), nutritional intake, and exercise indicators, the correlation between interventions and blood glucose fluctuations can be explored. However, traditional correlation analysis is insufficient to reveal the true causal mechanisms between physiological indicators; therefore, time-series causal relationship mining has become an important technology for achieving precise intervention in GDM.

[0003] Existing time-series causal analysis methods can achieve certain results under low-dimensional or sufficient sample conditions, but they still face many problems in high-dimensional medical time-series data scenarios. On the one hand, when the number of variables is large and the sample length is limited, traditional methods are prone to search space explosion, resulting in extremely high computational complexity. On the other hand, patients' physiological states change dynamically at different stages of pregnancy, and existing methods often struggle to simultaneously ensure the accuracy of the causal structure and the stability of the time dimension. In addition, high-dimensional causal modeling contains a large number of redundant parameters, affecting the model's generalization ability. Summary of the Invention

[0004] The purpose of this invention is to provide a causal relationship mining method for high-dimensional time-series data of gestational diabetes mellitus that can adapt to the characteristics of such data.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A method for causal relationship mining of high-dimensional time-series data on gestational diabetes mellitus includes the following steps: S10: Obtain multivariate time-series data of gestational diabetes mellitus in pregnant women, and preprocess the multivariate time-series data to form a time-series dataset for causal analysis; S20: Based on this time series dataset, feature modeling is performed on the causal relationships between variables, causal strength is calculated and adaptive grouping low-rank modeling is performed to construct a global causal probability graph; S30: Based on the global causal probability graph, the parameters of the time series prediction model are alternately optimized, and a time smoothing constraint is introduced during the optimization process. The optimized time series prediction model is used to output the final time series causal structure. Preferably, in step S10, the multivariate time series data is physical observation data with collection timestamps, gestational age information, or relative event time identifiers. The data types of the multivariate time series data include one or more of the following: dynamic blood glucose data, dietary nutrition data, exercise activity data, medication or intervention behavior data, vital sign monitoring data, laboratory test data, and prenatal examination or ultrasound measurement data.

[0006] Preferably, in step S20, the specific steps for constructing the global causal probability graph are as follows: The Granger causality test was performed on the multivariate time series data based on the bivariate vector autoregression model, the causal significance score was calculated, and the causal significance score was transformed by negative logarithm to generate the causal strength score. Based on the causal strength score, the relationship between variables is mapped to a causal probability prior matrix. The total causal strength is calculated based on the causal probability prior matrix and the causal strength score. The multiple variables are sorted and uniformly divided into multiple groups in sequence to generate a grouping matrix and an organization causal probability matrix. A global causal probability graph is obtained through adaptive grouping and low-rank modeling.

[0007] Preferably, in step S30, a message-passing graph neural network is used as the time-series prediction model, and the message-passing graph neural network is optimized in two stages: In the first stage, the network parameters of the message passing graph neural network are fixed, and the causal structure parameters are optimized using the Gumbel-Softmax estimator. In the second stage, a global causal probability graph is constructed with fixed causal structure parameters, and the network parameters of the message passing graph neural network are optimized. In the first stage of optimization, a time smoothing constraint is introduced to limit the drastic changes in the causal structure between adjacent optimization iterations.

[0008] Preferably, the total loss function is calculated using prediction error loss, graph sparsity regularization loss, and time smoothing constraint. After the message passing graph neural network training converges, Bernoulli sampling and threshold binarization are performed on the global causal probability graph.

[0009] A causal relationship mining system for high-dimensional time-series data of gestational diabetes mellitus includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the causal relationship mining method for high-dimensional time-series data of gestational diabetes mellitus as described above.

[0010] By adopting the aforementioned design scheme, the beneficial effects of the present invention are: by introducing a causal strength evaluation mechanism based on Granger causality test, reliable initial prior information can be provided for high-dimensional temporal causal structures; By combining adaptive grouping with low-rank modeling, redundant parameters in high-dimensional causal structures can be significantly reduced, improving the computational efficiency and stability of the model. By alternately optimizing causal structure learning and time series prediction models, and introducing time smoothing constraints during the optimization process, the discovered causal relationships maintain continuity and consistency across different time stages, thereby improving the robustness of causal discovery results in complex noise environments. This application can achieve stable modeling of high-dimensional temporal causal structures while ensuring the accuracy of causal discovery, providing a reliable basis for subsequent causal analysis and decision support. Attached Figure Description

[0011] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0013] The terms "first," "second," "third," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0014] A method for causal relationship mining of high-dimensional time-series data on gestational diabetes mellitus, such as Figure 1 As shown, it includes the following steps: S10: By connecting to continuous blood glucose monitoring devices, wearable sports monitoring devices, dietary recording terminals, vital sign monitoring devices, and hospital clinical information systems, the system acquires multivariate time-series data on gestational diabetes generated from the pregnant woman's actual physiological state, dietary behavior, exercise behavior, and medical testing process. The system then preprocesses the multivariate time-series data to form a time-series dataset for causal analysis.

[0015] In this embodiment, the multivariate time-series data of gestational diabetes mellitus is not simply generated by a mathematical model, but is generated or recorded by physical acquisition devices, patient interaction terminals, and hospital information systems in a real medical setting. The data acquisition devices or systems include one or more of the following: continuous glucose monitoring devices, fingertip blood glucose meters, wearable fitness monitoring devices, dietary recording terminals, blood pressure monitoring devices, weight monitoring devices, prenatal ultrasound equipment, hospital electronic medical record systems, laboratory information systems, and pregnancy nutrition management platforms.

[0016] The multivariate time-series data includes: interstitial fluid glucose levels, fingertip blood glucose levels, or pre- and post-meal blood glucose levels generated by continuous glucose monitoring devices; meal times, food types, intake amounts, energy, carbohydrate, protein, and fat intakes generated by dietary recording terminals; steps, exercise intensity, energy consumption, heart rate, sleep duration, and activity time generated by wearable fitness monitoring devices; oral glucose tolerance test results, glycated hemoglobin, blood lipids, liver and kidney function, inflammatory markers, gestational age, pre-pregnancy body mass index, blood pressure, medication dosage, and nutritional intervention records generated by medical information systems or laboratory information systems; and fetal biparietal diameter, head circumference, abdominal circumference, femur length, and estimated fetal weight generated by prenatal ultrasound equipment.

[0017] Each data point corresponds to the pregnant woman's physiological state, detection behavior, or intervention behavior in the physical world, and is accompanied by a collection timestamp, gestational age information, or relative event time identifier. The acquisition process includes: generating raw observation records through sensors, detection instruments, or patient interaction terminals; transmitting the raw observation records to a processor or memory via wired communication, wireless communication, or a database interface; and the processor parsing, de-identifying, and structuring the raw observation records according to patient identification, collection time, and variable type to form a multivariate time-series data matrix.

[0018] In this embodiment, the preprocessing of the multivariate time series data specifically includes: truncating and aligning the multivariate time series of gestational diabetes mellitus from different sources or with different sampling frequencies according to the collection time, gestational week or a unified time window to ensure that each variable is analyzed on a unified time axis. Missing values ​​in this multivariate time series data are labeled or processed to reduce the impact of missing data on the causal analysis results; Z-score standardization is performed on the multivariate time series data to ensure that the numerical distribution meets the requirements of a uniform scale, thereby forming a time series dataset for causal analysis. By performing uniform preprocessing and standardization on the multivariate time series data, the impact of differences in the scales of different variables on the results of causal analysis can be effectively reduced.

[0019] S20: Based on this time series dataset, feature modeling is performed on the causal relationships between variables, causal strength is calculated and adaptive grouping low-rank modeling is performed to construct a global causal probability graph. In this embodiment, a Granger causality test is performed on the multivariate time series data based on a bivariate vector autoregression (VAR) model to calculate the causal significance score. And perform a negative logarithmic transformation on it to generate a causal strength score. Causal significance score The calculation formula is as follows: ; in, The lag order is... The maximum lag order, and These represent the lag order as follows: At that time, the sum of squared residuals of the unconstrained and constrained vector autoregressive models, This represents the time series index corresponding to the causal variable. This represents the time series index corresponding to the target variable; Based on causal significance score Calculate the causal strength score : .

[0020] Based on the causality strength score Mapping the relationships between variables to a causal probability prior matrix R, assuming the following is used... Used to characterize the initial state of a potential causal relationship between variables. It can be expressed by the following formula: ; in, The maximum causal strength score after truncation; for insignificant causal strengths, set... .

[0021] Based on the causal probability prior matrix R and the causal strength score The total causal strength is calculated by ranking the multivariate variables and dividing them into L groups, generating a grouping matrix G and an organizational causal probability matrix Q. Through this adaptive grouping and low-rank modeling approach, a global causal probability map M is obtained. The elements of the organizational causal probability matrix Q are... The probability of the strongest connection between a node in the group and the target variable is defined by the following formula: ; in, Indicates the first Group 1 and the 2nd group Group-level causal probability among variables; Grouped matrix The elements in the array are used to represent the first element. Does the _th time series variable belong to the _th _th There are 1 group, among which... For grouped indexes, For indexing time series variables. If Then it means the first The time series variable belongs to the first Groups; if Then it means the first The time series variable does not belong to the first Grouped.

[0022] The global causal probability graph M is calculated by combining the grouping matrix G as follows: ; Where T represents the transpose symbol in matrix operations; S30: Based on the global causal probability graph The parameters of the time series prediction model are alternately optimized, and time smoothing constraints are introduced during the optimization process. The optimized time series prediction model is used to output the final time series causal structure. In this embodiment, a message-passing graph neural network (MPGNN) is used as the time-series prediction model, and a two-stage alternating optimization is performed: In the first stage, the network parameters of the message-passing graph neural network are fixed, and the causal structure parameters are optimized using a Gumbel-Softmax estimator. ,in The second stage involves constructing a global causal probability graph with fixed causal structure parameters. Optimize the network parameters of the message passing graph neural network. In the first stage of optimization, a time smoothing constraint is introduced. To limit the drastic changes in the causal structure between adjacent optimization iterations m-1 and m, it is expressed by the following formula: ; in, and They represent the first time. Second and third The causal graph obtained in the secondary structure optimization stage. denoted by Frobenius norm, used to measure the degree of difference between the causal graphs obtained in two adjacent structural optimization stages; The total loss function, calculated by combining prediction error loss, graph sparsity regularization loss, and time smoothing constraint, is expressed by the following formula: ; in, This represents the loss due to prediction error. This represents the graph sparsity regularization loss. This indicates a time smoothing constraint. and These represent the weight coefficients of the corresponding loss terms. The prediction error loss is expressed as: ; in, This represents the actual observed value of the j-th variable at time t. This indicates that the time series forecasting model outputs the predicted value of the j-th variable at time t. Indicates the observation mask. Indicates the total number of variables. This represents the total time length. The graph sparsity regularization loss is expressed as: ; in, Indicates the causal structure parameter. This represents the group-level causal probability matrix obtained by mapping causal structure parameters. express Norm.

[0023] After the time-series prediction model training converges, the global causal probability map M optimized by the prediction model is binarized by Bernoulli sampling to output the final stable time-series causal structure of gestational diabetes.

[0024] This embodiment also provides a system for implementing the above method.

[0025] A causal relationship mining system for high-dimensional time-series data of gestational diabetes mellitus includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the causal relationship mining method for high-dimensional time-series data of gestational diabetes mellitus as described above.

[0026] In summary, the method and algorithm designed in this application fully consider the characteristics of high-dimensional time-series data of gestational diabetes mellitus, such as the large number of variables (e.g., continuous glucose monitoring (CGM), multiple nutrient indicators, multidimensional exercise data, etc.), limited sample length, and dynamic changes in causal relationships. By introducing a causal strength assessment and initialization strategy based on Granger test (GCI), an adaptive grouping and low-rank modeling mechanism (CGL), and a dynamic refinement and smoothing constraint mechanism based on sparsity rate changes (SDR), this invention fundamentally solves the problems of exponentially increasing computational complexity and unstable results in existing time-series causal analysis methods when processing high-dimensional medical time-series data. This invention not only effectively reduces the model's consumption of computational resources but also improves the robustness of causal structure modeling, accurately revealing the deep causal mechanism between clinical interventions and blood glucose fluctuations, providing reliable technical support for the precise and dynamic diagnosis and treatment of gestational diabetes mellitus and clinical decision support.

[0027] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above description only uses the division of steps and functional modules as examples. In practical application scenarios (such as integration into a smart hospital pregnancy nutrition management system or a portable blood glucose monitoring device), the above functions can be completed by different functional modules as needed, that is, the system structure can be divided into different functional modules to complete all or part of the functions described above. The above functional modules can be implemented in hardware (such as a dedicated computing chip) or in software functional modules, and their specific implementation does not constitute a limitation on the scope of protection of this invention.

[0028] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for mining causal relationships in high-dimensional time-series data of gestational diabetes mellitus, characterized by: Includes the following steps: S10: Obtain multivariate time-series data of gestational diabetes mellitus in pregnant women, and preprocess the multivariate time-series data to form a time-series dataset for causal analysis; In step S10, the multivariate time series data is physical observation data with collection timestamps, gestational age information or relative event time identifiers. The data types of the multivariate time series data include one or more of the following: dynamic blood glucose data, dietary nutrition data, exercise activity data, medication dosage and nutritional intervention records, vital sign monitoring data, laboratory test data, prenatal examination or ultrasound measurement data. S20: Based on this time series dataset, feature modeling is performed on the causal relationships between variables, causal strength is calculated and adaptive grouping low-rank modeling is performed to construct a global causal probability graph; The specific steps for constructing a global causal probability graph are as follows: The Granger causality test was performed on the multivariate time series data based on the bivariate vector autoregression model, the causal significance score was calculated, and the causal significance score was transformed by negative logarithm to generate the causal strength score. Based on the causal strength score, the relationship between variables is mapped to a causal probability prior matrix. The total causal strength is calculated based on the causal probability prior matrix and the causal strength score. The multiple variables are sorted and uniformly divided into multiple groups in sequence to generate a grouping matrix and an organization causal probability matrix. A global causal probability map is obtained through adaptive grouping and low-rank modeling. The elements of the organization's causal probability matrix Q The probability of the strongest connection between a node in the group and the target variable is defined by the following formula: ； in, Indicates the first Group 1 and the 2nd group Group-level causal probability among variables; Grouped matrix The elements in the array are used to represent the first element. Does the _th time series variable belong to the _th _th There are 1 group, among which... For grouped indexes, For indexes of time series variables, if Then it means the first The time series variable belongs to the first Groups; if Then it means the first The time series variable does not belong to the first Groups, Used to characterize the initial state of potential causal relationships between variables; S30: Based on the global causal probability graph, the parameters of the time series prediction model are alternately optimized, and the Frobenius norm smoothing constraint is introduced during the optimization process. The optimized time series prediction model is used to output the final time series causal structure.

2. The causal relationship mining method for high-dimensional time-series data of gestational diabetes mellitus as described in claim 1, characterized in that: In step S30, a message-passing graph neural network is used as the time-series prediction model, and the message-passing graph neural network is optimized in two stages: In the first stage, the network parameters of the message passing graph neural network are fixed, and the causal structure parameters are optimized using the Gumbel-Softmax estimator. In the second stage, a global causal probability graph is constructed with fixed causal structure parameters, and the network parameters of the message passing graph neural network are optimized. In the first stage of optimization, a time smoothing constraint is introduced to limit the drastic changes in the causal structure between adjacent optimization iterations.

3. The causal relationship mining method for high-dimensional time-series data of gestational diabetes mellitus as described in claim 2, characterized in that: The total loss function is calculated using prediction error loss, graph sparsity regularization loss, and time smoothing constraint. After the message passing graph neural network training converges, Bernoulli sampling and threshold binarization are performed on the global causal probability graph.

4. A causal relationship mining system for high-dimensional time-series data of gestational diabetes mellitus, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the causal relationship mining method for high-dimensional time-series data of gestational diabetes mellitus as described in any one of claims 1-3.

Images